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 futher 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 "dm-bio-list.h"
22 #include <linux/delay.h>
23 #include <linux/raid/raid10.h>
24 #include <linux/raid/bitmap.h>
27 * RAID10 provides a combination of RAID0 and RAID1 functionality.
28 * The layout of data is defined by
31 * near_copies (stored in low byte of layout)
32 * far_copies (stored in second byte of layout)
33 * far_offset (stored in bit 16 of layout )
35 * The data to be stored is divided into chunks using chunksize.
36 * Each device is divided into far_copies sections.
37 * In each section, chunks are laid out in a style similar to raid0, but
38 * near_copies copies of each chunk is stored (each on a different drive).
39 * The starting device for each section is offset near_copies from the starting
40 * device of the previous section.
41 * Thus they are (near_copies*far_copies) of each chunk, and each is on a different
43 * near_copies and far_copies must be at least one, and their product is at most
46 * If far_offset is true, then the far_copies are handled a bit differently.
47 * The copies are still in different stripes, but instead of be very far apart
48 * on disk, there are adjacent stripes.
52 * Number of guaranteed r10bios in case of extreme VM load:
54 #define NR_RAID10_BIOS 256
56 static void unplug_slaves(mddev_t *mddev);
58 static void allow_barrier(conf_t *conf);
59 static void lower_barrier(conf_t *conf);
61 static void * r10bio_pool_alloc(gfp_t gfp_flags, void *data)
65 int size = offsetof(struct r10bio_s, devs[conf->copies]);
67 /* allocate a r10bio with room for raid_disks entries in the bios array */
68 r10_bio = kzalloc(size, gfp_flags);
70 unplug_slaves(conf->mddev);
75 static void r10bio_pool_free(void *r10_bio, void *data)
80 /* Maximum size of each resync request */
81 #define RESYNC_BLOCK_SIZE (64*1024)
82 #define RESYNC_PAGES ((RESYNC_BLOCK_SIZE + PAGE_SIZE-1) / PAGE_SIZE)
83 /* amount of memory to reserve for resync requests */
84 #define RESYNC_WINDOW (1024*1024)
85 /* maximum number of concurrent requests, memory permitting */
86 #define RESYNC_DEPTH (32*1024*1024/RESYNC_BLOCK_SIZE)
89 * When performing a resync, we need to read and compare, so
90 * we need as many pages are there are copies.
91 * When performing a recovery, we need 2 bios, one for read,
92 * one for write (we recover only one drive per r10buf)
95 static void * r10buf_pool_alloc(gfp_t gfp_flags, void *data)
104 r10_bio = r10bio_pool_alloc(gfp_flags, conf);
106 unplug_slaves(conf->mddev);
110 if (test_bit(MD_RECOVERY_SYNC, &conf->mddev->recovery))
111 nalloc = conf->copies; /* resync */
113 nalloc = 2; /* recovery */
118 for (j = nalloc ; j-- ; ) {
119 bio = bio_alloc(gfp_flags, RESYNC_PAGES);
122 r10_bio->devs[j].bio = bio;
125 * Allocate RESYNC_PAGES data pages and attach them
128 for (j = 0 ; j < nalloc; j++) {
129 bio = r10_bio->devs[j].bio;
130 for (i = 0; i < RESYNC_PAGES; i++) {
131 page = alloc_page(gfp_flags);
135 bio->bi_io_vec[i].bv_page = page;
143 safe_put_page(bio->bi_io_vec[i-1].bv_page);
145 for (i = 0; i < RESYNC_PAGES ; i++)
146 safe_put_page(r10_bio->devs[j].bio->bi_io_vec[i].bv_page);
149 while ( ++j < nalloc )
150 bio_put(r10_bio->devs[j].bio);
151 r10bio_pool_free(r10_bio, conf);
155 static void r10buf_pool_free(void *__r10_bio, void *data)
159 r10bio_t *r10bio = __r10_bio;
162 for (j=0; j < conf->copies; j++) {
163 struct bio *bio = r10bio->devs[j].bio;
165 for (i = 0; i < RESYNC_PAGES; i++) {
166 safe_put_page(bio->bi_io_vec[i].bv_page);
167 bio->bi_io_vec[i].bv_page = NULL;
172 r10bio_pool_free(r10bio, conf);
175 static void put_all_bios(conf_t *conf, r10bio_t *r10_bio)
179 for (i = 0; i < conf->copies; i++) {
180 struct bio **bio = & r10_bio->devs[i].bio;
181 if (*bio && *bio != IO_BLOCKED)
187 static void free_r10bio(r10bio_t *r10_bio)
189 conf_t *conf = mddev_to_conf(r10_bio->mddev);
192 * Wake up any possible resync thread that waits for the device
197 put_all_bios(conf, r10_bio);
198 mempool_free(r10_bio, conf->r10bio_pool);
201 static void put_buf(r10bio_t *r10_bio)
203 conf_t *conf = mddev_to_conf(r10_bio->mddev);
205 mempool_free(r10_bio, conf->r10buf_pool);
210 static void reschedule_retry(r10bio_t *r10_bio)
213 mddev_t *mddev = r10_bio->mddev;
214 conf_t *conf = mddev_to_conf(mddev);
216 spin_lock_irqsave(&conf->device_lock, flags);
217 list_add(&r10_bio->retry_list, &conf->retry_list);
219 spin_unlock_irqrestore(&conf->device_lock, flags);
221 /* wake up frozen array... */
222 wake_up(&conf->wait_barrier);
224 md_wakeup_thread(mddev->thread);
228 * raid_end_bio_io() is called when we have finished servicing a mirrored
229 * operation and are ready to return a success/failure code to the buffer
232 static void raid_end_bio_io(r10bio_t *r10_bio)
234 struct bio *bio = r10_bio->master_bio;
237 test_bit(R10BIO_Uptodate, &r10_bio->state) ? 0 : -EIO);
238 free_r10bio(r10_bio);
242 * Update disk head position estimator based on IRQ completion info.
244 static inline void update_head_pos(int slot, r10bio_t *r10_bio)
246 conf_t *conf = mddev_to_conf(r10_bio->mddev);
248 conf->mirrors[r10_bio->devs[slot].devnum].head_position =
249 r10_bio->devs[slot].addr + (r10_bio->sectors);
252 static void raid10_end_read_request(struct bio *bio, int error)
254 int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
255 r10bio_t * r10_bio = (r10bio_t *)(bio->bi_private);
257 conf_t *conf = mddev_to_conf(r10_bio->mddev);
260 slot = r10_bio->read_slot;
261 dev = r10_bio->devs[slot].devnum;
263 * this branch is our 'one mirror IO has finished' event handler:
265 update_head_pos(slot, r10_bio);
269 * Set R10BIO_Uptodate in our master bio, so that
270 * we will return a good error code to the higher
271 * levels even if IO on some other mirrored buffer fails.
273 * The 'master' represents the composite IO operation to
274 * user-side. So if something waits for IO, then it will
275 * wait for the 'master' bio.
277 set_bit(R10BIO_Uptodate, &r10_bio->state);
278 raid_end_bio_io(r10_bio);
283 char b[BDEVNAME_SIZE];
284 if (printk_ratelimit())
285 printk(KERN_ERR "raid10: %s: rescheduling sector %llu\n",
286 bdevname(conf->mirrors[dev].rdev->bdev,b), (unsigned long long)r10_bio->sector);
287 reschedule_retry(r10_bio);
290 rdev_dec_pending(conf->mirrors[dev].rdev, conf->mddev);
293 static void raid10_end_write_request(struct bio *bio, int error)
295 int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
296 r10bio_t * r10_bio = (r10bio_t *)(bio->bi_private);
298 conf_t *conf = mddev_to_conf(r10_bio->mddev);
300 for (slot = 0; slot < conf->copies; slot++)
301 if (r10_bio->devs[slot].bio == bio)
303 dev = r10_bio->devs[slot].devnum;
306 * this branch is our 'one mirror IO has finished' event handler:
309 md_error(r10_bio->mddev, conf->mirrors[dev].rdev);
310 /* an I/O failed, we can't clear the bitmap */
311 set_bit(R10BIO_Degraded, &r10_bio->state);
314 * Set R10BIO_Uptodate in our master bio, so that
315 * we will return a good error code for to the higher
316 * levels even if IO on some other mirrored buffer fails.
318 * The 'master' represents the composite IO operation to
319 * user-side. So if something waits for IO, then it will
320 * wait for the 'master' bio.
322 set_bit(R10BIO_Uptodate, &r10_bio->state);
324 update_head_pos(slot, r10_bio);
328 * Let's see if all mirrored write operations have finished
331 if (atomic_dec_and_test(&r10_bio->remaining)) {
332 /* clear the bitmap if all writes complete successfully */
333 bitmap_endwrite(r10_bio->mddev->bitmap, r10_bio->sector,
335 !test_bit(R10BIO_Degraded, &r10_bio->state),
337 md_write_end(r10_bio->mddev);
338 raid_end_bio_io(r10_bio);
341 rdev_dec_pending(conf->mirrors[dev].rdev, conf->mddev);
346 * RAID10 layout manager
347 * Aswell as the chunksize and raid_disks count, there are two
348 * parameters: near_copies and far_copies.
349 * near_copies * far_copies must be <= raid_disks.
350 * Normally one of these will be 1.
351 * If both are 1, we get raid0.
352 * If near_copies == raid_disks, we get raid1.
354 * Chunks are layed out in raid0 style with near_copies copies of the
355 * first chunk, followed by near_copies copies of the next chunk and
357 * If far_copies > 1, then after 1/far_copies of the array has been assigned
358 * as described above, we start again with a device offset of near_copies.
359 * So we effectively have another copy of the whole array further down all
360 * the drives, but with blocks on different drives.
361 * With this layout, and block is never stored twice on the one device.
363 * raid10_find_phys finds the sector offset of a given virtual sector
364 * on each device that it is on.
366 * raid10_find_virt does the reverse mapping, from a device and a
367 * sector offset to a virtual address
370 static void raid10_find_phys(conf_t *conf, r10bio_t *r10bio)
380 /* now calculate first sector/dev */
381 chunk = r10bio->sector >> conf->chunk_shift;
382 sector = r10bio->sector & conf->chunk_mask;
384 chunk *= conf->near_copies;
386 dev = sector_div(stripe, conf->raid_disks);
387 if (conf->far_offset)
388 stripe *= conf->far_copies;
390 sector += stripe << conf->chunk_shift;
392 /* and calculate all the others */
393 for (n=0; n < conf->near_copies; n++) {
396 r10bio->devs[slot].addr = sector;
397 r10bio->devs[slot].devnum = d;
400 for (f = 1; f < conf->far_copies; f++) {
401 d += conf->near_copies;
402 if (d >= conf->raid_disks)
403 d -= conf->raid_disks;
405 r10bio->devs[slot].devnum = d;
406 r10bio->devs[slot].addr = s;
410 if (dev >= conf->raid_disks) {
412 sector += (conf->chunk_mask + 1);
415 BUG_ON(slot != conf->copies);
418 static sector_t raid10_find_virt(conf_t *conf, sector_t sector, int dev)
420 sector_t offset, chunk, vchunk;
422 offset = sector & conf->chunk_mask;
423 if (conf->far_offset) {
425 chunk = sector >> conf->chunk_shift;
426 fc = sector_div(chunk, conf->far_copies);
427 dev -= fc * conf->near_copies;
429 dev += conf->raid_disks;
431 while (sector >= conf->stride) {
432 sector -= conf->stride;
433 if (dev < conf->near_copies)
434 dev += conf->raid_disks - conf->near_copies;
436 dev -= conf->near_copies;
438 chunk = sector >> conf->chunk_shift;
440 vchunk = chunk * conf->raid_disks + dev;
441 sector_div(vchunk, conf->near_copies);
442 return (vchunk << conf->chunk_shift) + offset;
446 * raid10_mergeable_bvec -- tell bio layer if a two requests can be merged
448 * @bvm: properties of new bio
449 * @biovec: the request that could be merged to it.
451 * Return amount of bytes we can accept at this offset
452 * If near_copies == raid_disk, there are no striping issues,
453 * but in that case, the function isn't called at all.
455 static int raid10_mergeable_bvec(struct request_queue *q,
456 struct bvec_merge_data *bvm,
457 struct bio_vec *biovec)
459 mddev_t *mddev = q->queuedata;
460 sector_t sector = bvm->bi_sector + get_start_sect(bvm->bi_bdev);
462 unsigned int chunk_sectors = mddev->chunk_size >> 9;
463 unsigned int bio_sectors = bvm->bi_size >> 9;
465 max = (chunk_sectors - ((sector & (chunk_sectors - 1)) + bio_sectors)) << 9;
466 if (max < 0) max = 0; /* bio_add cannot handle a negative return */
467 if (max <= biovec->bv_len && bio_sectors == 0)
468 return biovec->bv_len;
474 * This routine returns the disk from which the requested read should
475 * be done. There is a per-array 'next expected sequential IO' sector
476 * number - if this matches on the next IO then we use the last disk.
477 * There is also a per-disk 'last know head position' sector that is
478 * maintained from IRQ contexts, both the normal and the resync IO
479 * completion handlers update this position correctly. If there is no
480 * perfect sequential match then we pick the disk whose head is closest.
482 * If there are 2 mirrors in the same 2 devices, performance degrades
483 * because position is mirror, not device based.
485 * The rdev for the device selected will have nr_pending incremented.
489 * FIXME: possibly should rethink readbalancing and do it differently
490 * depending on near_copies / far_copies geometry.
492 static int read_balance(conf_t *conf, r10bio_t *r10_bio)
494 const unsigned long this_sector = r10_bio->sector;
495 int disk, slot, nslot;
496 const int sectors = r10_bio->sectors;
497 sector_t new_distance, current_distance;
500 raid10_find_phys(conf, r10_bio);
503 * Check if we can balance. We can balance on the whole
504 * device if no resync is going on (recovery is ok), or below
505 * the resync window. We take the first readable disk when
506 * above the resync window.
508 if (conf->mddev->recovery_cp < MaxSector
509 && (this_sector + sectors >= conf->next_resync)) {
510 /* make sure that disk is operational */
512 disk = r10_bio->devs[slot].devnum;
514 while ((rdev = rcu_dereference(conf->mirrors[disk].rdev)) == NULL ||
515 r10_bio->devs[slot].bio == IO_BLOCKED ||
516 !test_bit(In_sync, &rdev->flags)) {
518 if (slot == conf->copies) {
523 disk = r10_bio->devs[slot].devnum;
529 /* make sure the disk is operational */
531 disk = r10_bio->devs[slot].devnum;
532 while ((rdev=rcu_dereference(conf->mirrors[disk].rdev)) == NULL ||
533 r10_bio->devs[slot].bio == IO_BLOCKED ||
534 !test_bit(In_sync, &rdev->flags)) {
536 if (slot == conf->copies) {
540 disk = r10_bio->devs[slot].devnum;
544 current_distance = abs(r10_bio->devs[slot].addr -
545 conf->mirrors[disk].head_position);
547 /* Find the disk whose head is closest,
548 * or - for far > 1 - find the closest to partition beginning */
550 for (nslot = slot; nslot < conf->copies; nslot++) {
551 int ndisk = r10_bio->devs[nslot].devnum;
554 if ((rdev=rcu_dereference(conf->mirrors[ndisk].rdev)) == NULL ||
555 r10_bio->devs[nslot].bio == IO_BLOCKED ||
556 !test_bit(In_sync, &rdev->flags))
559 /* This optimisation is debatable, and completely destroys
560 * sequential read speed for 'far copies' arrays. So only
561 * keep it for 'near' arrays, and review those later.
563 if (conf->near_copies > 1 && !atomic_read(&rdev->nr_pending)) {
569 /* for far > 1 always use the lowest address */
570 if (conf->far_copies > 1)
571 new_distance = r10_bio->devs[nslot].addr;
573 new_distance = abs(r10_bio->devs[nslot].addr -
574 conf->mirrors[ndisk].head_position);
575 if (new_distance < current_distance) {
576 current_distance = new_distance;
583 r10_bio->read_slot = slot;
584 /* conf->next_seq_sect = this_sector + sectors;*/
586 if (disk >= 0 && (rdev=rcu_dereference(conf->mirrors[disk].rdev))!= NULL)
587 atomic_inc(&conf->mirrors[disk].rdev->nr_pending);
595 static void unplug_slaves(mddev_t *mddev)
597 conf_t *conf = mddev_to_conf(mddev);
601 for (i=0; i<mddev->raid_disks; i++) {
602 mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[i].rdev);
603 if (rdev && !test_bit(Faulty, &rdev->flags) && atomic_read(&rdev->nr_pending)) {
604 struct request_queue *r_queue = bdev_get_queue(rdev->bdev);
606 atomic_inc(&rdev->nr_pending);
611 rdev_dec_pending(rdev, mddev);
618 static void raid10_unplug(struct request_queue *q)
620 mddev_t *mddev = q->queuedata;
622 unplug_slaves(q->queuedata);
623 md_wakeup_thread(mddev->thread);
626 static int raid10_congested(void *data, int bits)
628 mddev_t *mddev = data;
629 conf_t *conf = mddev_to_conf(mddev);
633 for (i = 0; i < mddev->raid_disks && ret == 0; i++) {
634 mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[i].rdev);
635 if (rdev && !test_bit(Faulty, &rdev->flags)) {
636 struct request_queue *q = bdev_get_queue(rdev->bdev);
638 ret |= bdi_congested(&q->backing_dev_info, bits);
645 static int flush_pending_writes(conf_t *conf)
647 /* Any writes that have been queued but are awaiting
648 * bitmap updates get flushed here.
649 * We return 1 if any requests were actually submitted.
653 spin_lock_irq(&conf->device_lock);
655 if (conf->pending_bio_list.head) {
657 bio = bio_list_get(&conf->pending_bio_list);
658 blk_remove_plug(conf->mddev->queue);
659 spin_unlock_irq(&conf->device_lock);
660 /* flush any pending bitmap writes to disk
661 * before proceeding w/ I/O */
662 bitmap_unplug(conf->mddev->bitmap);
664 while (bio) { /* submit pending writes */
665 struct bio *next = bio->bi_next;
667 generic_make_request(bio);
672 spin_unlock_irq(&conf->device_lock);
676 * Sometimes we need to suspend IO while we do something else,
677 * either some resync/recovery, or reconfigure the array.
678 * To do this we raise a 'barrier'.
679 * The 'barrier' is a counter that can be raised multiple times
680 * to count how many activities are happening which preclude
682 * We can only raise the barrier if there is no pending IO.
683 * i.e. if nr_pending == 0.
684 * We choose only to raise the barrier if no-one is waiting for the
685 * barrier to go down. This means that as soon as an IO request
686 * is ready, no other operations which require a barrier will start
687 * until the IO request has had a chance.
689 * So: regular IO calls 'wait_barrier'. When that returns there
690 * is no backgroup IO happening, It must arrange to call
691 * allow_barrier when it has finished its IO.
692 * backgroup IO calls must call raise_barrier. Once that returns
693 * there is no normal IO happeing. It must arrange to call
694 * lower_barrier when the particular background IO completes.
697 static void raise_barrier(conf_t *conf, int force)
699 BUG_ON(force && !conf->barrier);
700 spin_lock_irq(&conf->resync_lock);
702 /* Wait until no block IO is waiting (unless 'force') */
703 wait_event_lock_irq(conf->wait_barrier, force || !conf->nr_waiting,
705 raid10_unplug(conf->mddev->queue));
707 /* block any new IO from starting */
710 /* No wait for all pending IO to complete */
711 wait_event_lock_irq(conf->wait_barrier,
712 !conf->nr_pending && conf->barrier < RESYNC_DEPTH,
714 raid10_unplug(conf->mddev->queue));
716 spin_unlock_irq(&conf->resync_lock);
719 static void lower_barrier(conf_t *conf)
722 spin_lock_irqsave(&conf->resync_lock, flags);
724 spin_unlock_irqrestore(&conf->resync_lock, flags);
725 wake_up(&conf->wait_barrier);
728 static void wait_barrier(conf_t *conf)
730 spin_lock_irq(&conf->resync_lock);
733 wait_event_lock_irq(conf->wait_barrier, !conf->barrier,
735 raid10_unplug(conf->mddev->queue));
739 spin_unlock_irq(&conf->resync_lock);
742 static void allow_barrier(conf_t *conf)
745 spin_lock_irqsave(&conf->resync_lock, flags);
747 spin_unlock_irqrestore(&conf->resync_lock, flags);
748 wake_up(&conf->wait_barrier);
751 static void freeze_array(conf_t *conf)
753 /* stop syncio and normal IO and wait for everything to
755 * We increment barrier and nr_waiting, and then
756 * wait until nr_pending match nr_queued+1
757 * This is called in the context of one normal IO request
758 * that has failed. Thus any sync request that might be pending
759 * will be blocked by nr_pending, and we need to wait for
760 * pending IO requests to complete or be queued for re-try.
761 * Thus the number queued (nr_queued) plus this request (1)
762 * must match the number of pending IOs (nr_pending) before
765 spin_lock_irq(&conf->resync_lock);
768 wait_event_lock_irq(conf->wait_barrier,
769 conf->nr_pending == conf->nr_queued+1,
771 ({ flush_pending_writes(conf);
772 raid10_unplug(conf->mddev->queue); }));
773 spin_unlock_irq(&conf->resync_lock);
776 static void unfreeze_array(conf_t *conf)
778 /* reverse the effect of the freeze */
779 spin_lock_irq(&conf->resync_lock);
782 wake_up(&conf->wait_barrier);
783 spin_unlock_irq(&conf->resync_lock);
786 static int make_request(struct request_queue *q, struct bio * bio)
788 mddev_t *mddev = q->queuedata;
789 conf_t *conf = mddev_to_conf(mddev);
790 mirror_info_t *mirror;
792 struct bio *read_bio;
795 int chunk_sects = conf->chunk_mask + 1;
796 const int rw = bio_data_dir(bio);
797 const int do_sync = bio_sync(bio);
800 mdk_rdev_t *blocked_rdev;
802 if (unlikely(bio_barrier(bio))) {
803 bio_endio(bio, -EOPNOTSUPP);
807 /* If this request crosses a chunk boundary, we need to
808 * split it. This will only happen for 1 PAGE (or less) requests.
810 if (unlikely( (bio->bi_sector & conf->chunk_mask) + (bio->bi_size >> 9)
812 conf->near_copies < conf->raid_disks)) {
814 /* Sanity check -- queue functions should prevent this happening */
815 if (bio->bi_vcnt != 1 ||
818 /* This is a one page bio that upper layers
819 * refuse to split for us, so we need to split it.
822 chunk_sects - (bio->bi_sector & (chunk_sects - 1)) );
823 if (make_request(q, &bp->bio1))
824 generic_make_request(&bp->bio1);
825 if (make_request(q, &bp->bio2))
826 generic_make_request(&bp->bio2);
828 bio_pair_release(bp);
831 printk("raid10_make_request bug: can't convert block across chunks"
832 " or bigger than %dk %llu %d\n", chunk_sects/2,
833 (unsigned long long)bio->bi_sector, bio->bi_size >> 10);
839 md_write_start(mddev, bio);
842 * Register the new request and wait if the reconstruction
843 * thread has put up a bar for new requests.
844 * Continue immediately if no resync is active currently.
848 cpu = part_stat_lock();
849 part_stat_inc(cpu, &mddev->gendisk->part0, ios[rw]);
850 part_stat_add(cpu, &mddev->gendisk->part0, sectors[rw],
854 r10_bio = mempool_alloc(conf->r10bio_pool, GFP_NOIO);
856 r10_bio->master_bio = bio;
857 r10_bio->sectors = bio->bi_size >> 9;
859 r10_bio->mddev = mddev;
860 r10_bio->sector = bio->bi_sector;
865 * read balancing logic:
867 int disk = read_balance(conf, r10_bio);
868 int slot = r10_bio->read_slot;
870 raid_end_bio_io(r10_bio);
873 mirror = conf->mirrors + disk;
875 read_bio = bio_clone(bio, GFP_NOIO);
877 r10_bio->devs[slot].bio = read_bio;
879 read_bio->bi_sector = r10_bio->devs[slot].addr +
880 mirror->rdev->data_offset;
881 read_bio->bi_bdev = mirror->rdev->bdev;
882 read_bio->bi_end_io = raid10_end_read_request;
883 read_bio->bi_rw = READ | do_sync;
884 read_bio->bi_private = r10_bio;
886 generic_make_request(read_bio);
893 /* first select target devices under rcu_lock and
894 * inc refcount on their rdev. Record them by setting
897 raid10_find_phys(conf, r10_bio);
901 for (i = 0; i < conf->copies; i++) {
902 int d = r10_bio->devs[i].devnum;
903 mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[d].rdev);
904 if (rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
905 atomic_inc(&rdev->nr_pending);
909 if (rdev && !test_bit(Faulty, &rdev->flags)) {
910 atomic_inc(&rdev->nr_pending);
911 r10_bio->devs[i].bio = bio;
913 r10_bio->devs[i].bio = NULL;
914 set_bit(R10BIO_Degraded, &r10_bio->state);
919 if (unlikely(blocked_rdev)) {
920 /* Have to wait for this device to get unblocked, then retry */
924 for (j = 0; j < i; j++)
925 if (r10_bio->devs[j].bio) {
926 d = r10_bio->devs[j].devnum;
927 rdev_dec_pending(conf->mirrors[d].rdev, mddev);
930 md_wait_for_blocked_rdev(blocked_rdev, mddev);
935 atomic_set(&r10_bio->remaining, 0);
938 for (i = 0; i < conf->copies; i++) {
940 int d = r10_bio->devs[i].devnum;
941 if (!r10_bio->devs[i].bio)
944 mbio = bio_clone(bio, GFP_NOIO);
945 r10_bio->devs[i].bio = mbio;
947 mbio->bi_sector = r10_bio->devs[i].addr+
948 conf->mirrors[d].rdev->data_offset;
949 mbio->bi_bdev = conf->mirrors[d].rdev->bdev;
950 mbio->bi_end_io = raid10_end_write_request;
951 mbio->bi_rw = WRITE | do_sync;
952 mbio->bi_private = r10_bio;
954 atomic_inc(&r10_bio->remaining);
955 bio_list_add(&bl, mbio);
958 if (unlikely(!atomic_read(&r10_bio->remaining))) {
959 /* the array is dead */
961 raid_end_bio_io(r10_bio);
965 bitmap_startwrite(mddev->bitmap, bio->bi_sector, r10_bio->sectors, 0);
966 spin_lock_irqsave(&conf->device_lock, flags);
967 bio_list_merge(&conf->pending_bio_list, &bl);
968 blk_plug_device(mddev->queue);
969 spin_unlock_irqrestore(&conf->device_lock, flags);
971 /* In case raid10d snuck in to freeze_array */
972 wake_up(&conf->wait_barrier);
975 md_wakeup_thread(mddev->thread);
980 static void status(struct seq_file *seq, mddev_t *mddev)
982 conf_t *conf = mddev_to_conf(mddev);
985 if (conf->near_copies < conf->raid_disks)
986 seq_printf(seq, " %dK chunks", mddev->chunk_size/1024);
987 if (conf->near_copies > 1)
988 seq_printf(seq, " %d near-copies", conf->near_copies);
989 if (conf->far_copies > 1) {
990 if (conf->far_offset)
991 seq_printf(seq, " %d offset-copies", conf->far_copies);
993 seq_printf(seq, " %d far-copies", conf->far_copies);
995 seq_printf(seq, " [%d/%d] [", conf->raid_disks,
996 conf->raid_disks - mddev->degraded);
997 for (i = 0; i < conf->raid_disks; i++)
998 seq_printf(seq, "%s",
999 conf->mirrors[i].rdev &&
1000 test_bit(In_sync, &conf->mirrors[i].rdev->flags) ? "U" : "_");
1001 seq_printf(seq, "]");
1004 static void error(mddev_t *mddev, mdk_rdev_t *rdev)
1006 char b[BDEVNAME_SIZE];
1007 conf_t *conf = mddev_to_conf(mddev);
1010 * If it is not operational, then we have already marked it as dead
1011 * else if it is the last working disks, ignore the error, let the
1012 * next level up know.
1013 * else mark the drive as failed
1015 if (test_bit(In_sync, &rdev->flags)
1016 && conf->raid_disks-mddev->degraded == 1)
1018 * Don't fail the drive, just return an IO error.
1019 * The test should really be more sophisticated than
1020 * "working_disks == 1", but it isn't critical, and
1021 * can wait until we do more sophisticated "is the drive
1022 * really dead" tests...
1025 if (test_and_clear_bit(In_sync, &rdev->flags)) {
1026 unsigned long flags;
1027 spin_lock_irqsave(&conf->device_lock, flags);
1029 spin_unlock_irqrestore(&conf->device_lock, flags);
1031 * if recovery is running, make sure it aborts.
1033 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1035 set_bit(Faulty, &rdev->flags);
1036 set_bit(MD_CHANGE_DEVS, &mddev->flags);
1037 printk(KERN_ALERT "raid10: Disk failure on %s, disabling device.\n"
1038 "raid10: Operation continuing on %d devices.\n",
1039 bdevname(rdev->bdev,b), conf->raid_disks - mddev->degraded);
1042 static void print_conf(conf_t *conf)
1047 printk("RAID10 conf printout:\n");
1049 printk("(!conf)\n");
1052 printk(" --- wd:%d rd:%d\n", conf->raid_disks - conf->mddev->degraded,
1055 for (i = 0; i < conf->raid_disks; i++) {
1056 char b[BDEVNAME_SIZE];
1057 tmp = conf->mirrors + i;
1059 printk(" disk %d, wo:%d, o:%d, dev:%s\n",
1060 i, !test_bit(In_sync, &tmp->rdev->flags),
1061 !test_bit(Faulty, &tmp->rdev->flags),
1062 bdevname(tmp->rdev->bdev,b));
1066 static void close_sync(conf_t *conf)
1069 allow_barrier(conf);
1071 mempool_destroy(conf->r10buf_pool);
1072 conf->r10buf_pool = NULL;
1075 /* check if there are enough drives for
1076 * every block to appear on atleast one
1078 static int enough(conf_t *conf)
1083 int n = conf->copies;
1086 if (conf->mirrors[first].rdev)
1088 first = (first+1) % conf->raid_disks;
1092 } while (first != 0);
1096 static int raid10_spare_active(mddev_t *mddev)
1099 conf_t *conf = mddev->private;
1103 * Find all non-in_sync disks within the RAID10 configuration
1104 * and mark them in_sync
1106 for (i = 0; i < conf->raid_disks; i++) {
1107 tmp = conf->mirrors + i;
1109 && !test_bit(Faulty, &tmp->rdev->flags)
1110 && !test_and_set_bit(In_sync, &tmp->rdev->flags)) {
1111 unsigned long flags;
1112 spin_lock_irqsave(&conf->device_lock, flags);
1114 spin_unlock_irqrestore(&conf->device_lock, flags);
1123 static int raid10_add_disk(mddev_t *mddev, mdk_rdev_t *rdev)
1125 conf_t *conf = mddev->private;
1130 int last = mddev->raid_disks - 1;
1132 if (mddev->recovery_cp < MaxSector)
1133 /* only hot-add to in-sync arrays, as recovery is
1134 * very different from resync
1140 if (rdev->raid_disk >= 0)
1141 first = last = rdev->raid_disk;
1143 if (rdev->saved_raid_disk >= 0 &&
1144 rdev->saved_raid_disk >= first &&
1145 conf->mirrors[rdev->saved_raid_disk].rdev == NULL)
1146 mirror = rdev->saved_raid_disk;
1149 for ( ; mirror <= last ; mirror++)
1150 if ( !(p=conf->mirrors+mirror)->rdev) {
1152 blk_queue_stack_limits(mddev->queue,
1153 rdev->bdev->bd_disk->queue);
1154 /* as we don't honour merge_bvec_fn, we must never risk
1155 * violating it, so limit ->max_sector to one PAGE, as
1156 * a one page request is never in violation.
1158 if (rdev->bdev->bd_disk->queue->merge_bvec_fn &&
1159 mddev->queue->max_sectors > (PAGE_SIZE>>9))
1160 mddev->queue->max_sectors = (PAGE_SIZE>>9);
1162 p->head_position = 0;
1163 rdev->raid_disk = mirror;
1165 if (rdev->saved_raid_disk != mirror)
1167 rcu_assign_pointer(p->rdev, rdev);
1175 static int raid10_remove_disk(mddev_t *mddev, int number)
1177 conf_t *conf = mddev->private;
1180 mirror_info_t *p = conf->mirrors+ number;
1185 if (test_bit(In_sync, &rdev->flags) ||
1186 atomic_read(&rdev->nr_pending)) {
1190 /* Only remove faulty devices in recovery
1193 if (!test_bit(Faulty, &rdev->flags) &&
1200 if (atomic_read(&rdev->nr_pending)) {
1201 /* lost the race, try later */
1213 static void end_sync_read(struct bio *bio, int error)
1215 r10bio_t * r10_bio = (r10bio_t *)(bio->bi_private);
1216 conf_t *conf = mddev_to_conf(r10_bio->mddev);
1219 for (i=0; i<conf->copies; i++)
1220 if (r10_bio->devs[i].bio == bio)
1222 BUG_ON(i == conf->copies);
1223 update_head_pos(i, r10_bio);
1224 d = r10_bio->devs[i].devnum;
1226 if (test_bit(BIO_UPTODATE, &bio->bi_flags))
1227 set_bit(R10BIO_Uptodate, &r10_bio->state);
1229 atomic_add(r10_bio->sectors,
1230 &conf->mirrors[d].rdev->corrected_errors);
1231 if (!test_bit(MD_RECOVERY_SYNC, &conf->mddev->recovery))
1232 md_error(r10_bio->mddev,
1233 conf->mirrors[d].rdev);
1236 /* for reconstruct, we always reschedule after a read.
1237 * for resync, only after all reads
1239 if (test_bit(R10BIO_IsRecover, &r10_bio->state) ||
1240 atomic_dec_and_test(&r10_bio->remaining)) {
1241 /* we have read all the blocks,
1242 * do the comparison in process context in raid10d
1244 reschedule_retry(r10_bio);
1246 rdev_dec_pending(conf->mirrors[d].rdev, conf->mddev);
1249 static void end_sync_write(struct bio *bio, int error)
1251 int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
1252 r10bio_t * r10_bio = (r10bio_t *)(bio->bi_private);
1253 mddev_t *mddev = r10_bio->mddev;
1254 conf_t *conf = mddev_to_conf(mddev);
1257 for (i = 0; i < conf->copies; i++)
1258 if (r10_bio->devs[i].bio == bio)
1260 d = r10_bio->devs[i].devnum;
1263 md_error(mddev, conf->mirrors[d].rdev);
1265 update_head_pos(i, r10_bio);
1267 while (atomic_dec_and_test(&r10_bio->remaining)) {
1268 if (r10_bio->master_bio == NULL) {
1269 /* the primary of several recovery bios */
1270 md_done_sync(mddev, r10_bio->sectors, 1);
1274 r10bio_t *r10_bio2 = (r10bio_t *)r10_bio->master_bio;
1279 rdev_dec_pending(conf->mirrors[d].rdev, mddev);
1283 * Note: sync and recover and handled very differently for raid10
1284 * This code is for resync.
1285 * For resync, we read through virtual addresses and read all blocks.
1286 * If there is any error, we schedule a write. The lowest numbered
1287 * drive is authoritative.
1288 * However requests come for physical address, so we need to map.
1289 * For every physical address there are raid_disks/copies virtual addresses,
1290 * which is always are least one, but is not necessarly an integer.
1291 * This means that a physical address can span multiple chunks, so we may
1292 * have to submit multiple io requests for a single sync request.
1295 * We check if all blocks are in-sync and only write to blocks that
1298 static void sync_request_write(mddev_t *mddev, r10bio_t *r10_bio)
1300 conf_t *conf = mddev_to_conf(mddev);
1302 struct bio *tbio, *fbio;
1304 atomic_set(&r10_bio->remaining, 1);
1306 /* find the first device with a block */
1307 for (i=0; i<conf->copies; i++)
1308 if (test_bit(BIO_UPTODATE, &r10_bio->devs[i].bio->bi_flags))
1311 if (i == conf->copies)
1315 fbio = r10_bio->devs[i].bio;
1317 /* now find blocks with errors */
1318 for (i=0 ; i < conf->copies ; i++) {
1320 int vcnt = r10_bio->sectors >> (PAGE_SHIFT-9);
1322 tbio = r10_bio->devs[i].bio;
1324 if (tbio->bi_end_io != end_sync_read)
1328 if (test_bit(BIO_UPTODATE, &r10_bio->devs[i].bio->bi_flags)) {
1329 /* We know that the bi_io_vec layout is the same for
1330 * both 'first' and 'i', so we just compare them.
1331 * All vec entries are PAGE_SIZE;
1333 for (j = 0; j < vcnt; j++)
1334 if (memcmp(page_address(fbio->bi_io_vec[j].bv_page),
1335 page_address(tbio->bi_io_vec[j].bv_page),
1340 mddev->resync_mismatches += r10_bio->sectors;
1342 if (test_bit(MD_RECOVERY_CHECK, &mddev->recovery))
1343 /* Don't fix anything. */
1345 /* Ok, we need to write this bio
1346 * First we need to fixup bv_offset, bv_len and
1347 * bi_vecs, as the read request might have corrupted these
1349 tbio->bi_vcnt = vcnt;
1350 tbio->bi_size = r10_bio->sectors << 9;
1352 tbio->bi_phys_segments = 0;
1353 tbio->bi_flags &= ~(BIO_POOL_MASK - 1);
1354 tbio->bi_flags |= 1 << BIO_UPTODATE;
1355 tbio->bi_next = NULL;
1356 tbio->bi_rw = WRITE;
1357 tbio->bi_private = r10_bio;
1358 tbio->bi_sector = r10_bio->devs[i].addr;
1360 for (j=0; j < vcnt ; j++) {
1361 tbio->bi_io_vec[j].bv_offset = 0;
1362 tbio->bi_io_vec[j].bv_len = PAGE_SIZE;
1364 memcpy(page_address(tbio->bi_io_vec[j].bv_page),
1365 page_address(fbio->bi_io_vec[j].bv_page),
1368 tbio->bi_end_io = end_sync_write;
1370 d = r10_bio->devs[i].devnum;
1371 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
1372 atomic_inc(&r10_bio->remaining);
1373 md_sync_acct(conf->mirrors[d].rdev->bdev, tbio->bi_size >> 9);
1375 tbio->bi_sector += conf->mirrors[d].rdev->data_offset;
1376 tbio->bi_bdev = conf->mirrors[d].rdev->bdev;
1377 generic_make_request(tbio);
1381 if (atomic_dec_and_test(&r10_bio->remaining)) {
1382 md_done_sync(mddev, r10_bio->sectors, 1);
1388 * Now for the recovery code.
1389 * Recovery happens across physical sectors.
1390 * We recover all non-is_sync drives by finding the virtual address of
1391 * each, and then choose a working drive that also has that virt address.
1392 * There is a separate r10_bio for each non-in_sync drive.
1393 * Only the first two slots are in use. The first for reading,
1394 * The second for writing.
1398 static void recovery_request_write(mddev_t *mddev, r10bio_t *r10_bio)
1400 conf_t *conf = mddev_to_conf(mddev);
1402 struct bio *bio, *wbio;
1405 /* move the pages across to the second bio
1406 * and submit the write request
1408 bio = r10_bio->devs[0].bio;
1409 wbio = r10_bio->devs[1].bio;
1410 for (i=0; i < wbio->bi_vcnt; i++) {
1411 struct page *p = bio->bi_io_vec[i].bv_page;
1412 bio->bi_io_vec[i].bv_page = wbio->bi_io_vec[i].bv_page;
1413 wbio->bi_io_vec[i].bv_page = p;
1415 d = r10_bio->devs[1].devnum;
1417 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
1418 md_sync_acct(conf->mirrors[d].rdev->bdev, wbio->bi_size >> 9);
1419 if (test_bit(R10BIO_Uptodate, &r10_bio->state))
1420 generic_make_request(wbio);
1422 bio_endio(wbio, -EIO);
1427 * This is a kernel thread which:
1429 * 1. Retries failed read operations on working mirrors.
1430 * 2. Updates the raid superblock when problems encounter.
1431 * 3. Performs writes following reads for array synchronising.
1434 static void fix_read_error(conf_t *conf, mddev_t *mddev, r10bio_t *r10_bio)
1436 int sect = 0; /* Offset from r10_bio->sector */
1437 int sectors = r10_bio->sectors;
1441 int sl = r10_bio->read_slot;
1445 if (s > (PAGE_SIZE>>9))
1450 int d = r10_bio->devs[sl].devnum;
1451 rdev = rcu_dereference(conf->mirrors[d].rdev);
1453 test_bit(In_sync, &rdev->flags)) {
1454 atomic_inc(&rdev->nr_pending);
1456 success = sync_page_io(rdev->bdev,
1457 r10_bio->devs[sl].addr +
1458 sect + rdev->data_offset,
1460 conf->tmppage, READ);
1461 rdev_dec_pending(rdev, mddev);
1467 if (sl == conf->copies)
1469 } while (!success && sl != r10_bio->read_slot);
1473 /* Cannot read from anywhere -- bye bye array */
1474 int dn = r10_bio->devs[r10_bio->read_slot].devnum;
1475 md_error(mddev, conf->mirrors[dn].rdev);
1480 /* write it back and re-read */
1482 while (sl != r10_bio->read_slot) {
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 atomic_add(s, &rdev->corrected_errors);
1494 if (sync_page_io(rdev->bdev,
1495 r10_bio->devs[sl].addr +
1496 sect + rdev->data_offset,
1497 s<<9, conf->tmppage, WRITE)
1499 /* Well, this device is dead */
1500 md_error(mddev, rdev);
1501 rdev_dec_pending(rdev, mddev);
1506 while (sl != r10_bio->read_slot) {
1511 d = r10_bio->devs[sl].devnum;
1512 rdev = rcu_dereference(conf->mirrors[d].rdev);
1514 test_bit(In_sync, &rdev->flags)) {
1515 char b[BDEVNAME_SIZE];
1516 atomic_inc(&rdev->nr_pending);
1518 if (sync_page_io(rdev->bdev,
1519 r10_bio->devs[sl].addr +
1520 sect + rdev->data_offset,
1521 s<<9, conf->tmppage, READ) == 0)
1522 /* Well, this device is dead */
1523 md_error(mddev, rdev);
1526 "raid10:%s: read error corrected"
1527 " (%d sectors at %llu on %s)\n",
1529 (unsigned long long)(sect+
1531 bdevname(rdev->bdev, b));
1533 rdev_dec_pending(rdev, mddev);
1544 static void raid10d(mddev_t *mddev)
1548 unsigned long flags;
1549 conf_t *conf = mddev_to_conf(mddev);
1550 struct list_head *head = &conf->retry_list;
1554 md_check_recovery(mddev);
1557 char b[BDEVNAME_SIZE];
1559 unplug += flush_pending_writes(conf);
1561 spin_lock_irqsave(&conf->device_lock, flags);
1562 if (list_empty(head)) {
1563 spin_unlock_irqrestore(&conf->device_lock, flags);
1566 r10_bio = list_entry(head->prev, r10bio_t, retry_list);
1567 list_del(head->prev);
1569 spin_unlock_irqrestore(&conf->device_lock, flags);
1571 mddev = r10_bio->mddev;
1572 conf = mddev_to_conf(mddev);
1573 if (test_bit(R10BIO_IsSync, &r10_bio->state)) {
1574 sync_request_write(mddev, r10_bio);
1576 } else if (test_bit(R10BIO_IsRecover, &r10_bio->state)) {
1577 recovery_request_write(mddev, r10_bio);
1581 /* we got a read error. Maybe the drive is bad. Maybe just
1582 * the block and we can fix it.
1583 * We freeze all other IO, and try reading the block from
1584 * other devices. When we find one, we re-write
1585 * and check it that fixes the read error.
1586 * This is all done synchronously while the array is
1589 if (mddev->ro == 0) {
1591 fix_read_error(conf, mddev, r10_bio);
1592 unfreeze_array(conf);
1595 bio = r10_bio->devs[r10_bio->read_slot].bio;
1596 r10_bio->devs[r10_bio->read_slot].bio =
1597 mddev->ro ? IO_BLOCKED : NULL;
1598 mirror = read_balance(conf, r10_bio);
1600 printk(KERN_ALERT "raid10: %s: unrecoverable I/O"
1601 " read error for block %llu\n",
1602 bdevname(bio->bi_bdev,b),
1603 (unsigned long long)r10_bio->sector);
1604 raid_end_bio_io(r10_bio);
1607 const int do_sync = bio_sync(r10_bio->master_bio);
1609 rdev = conf->mirrors[mirror].rdev;
1610 if (printk_ratelimit())
1611 printk(KERN_ERR "raid10: %s: redirecting sector %llu to"
1612 " another mirror\n",
1613 bdevname(rdev->bdev,b),
1614 (unsigned long long)r10_bio->sector);
1615 bio = bio_clone(r10_bio->master_bio, GFP_NOIO);
1616 r10_bio->devs[r10_bio->read_slot].bio = bio;
1617 bio->bi_sector = r10_bio->devs[r10_bio->read_slot].addr
1618 + rdev->data_offset;
1619 bio->bi_bdev = rdev->bdev;
1620 bio->bi_rw = READ | do_sync;
1621 bio->bi_private = r10_bio;
1622 bio->bi_end_io = raid10_end_read_request;
1624 generic_make_request(bio);
1629 unplug_slaves(mddev);
1633 static int init_resync(conf_t *conf)
1637 buffs = RESYNC_WINDOW / RESYNC_BLOCK_SIZE;
1638 BUG_ON(conf->r10buf_pool);
1639 conf->r10buf_pool = mempool_create(buffs, r10buf_pool_alloc, r10buf_pool_free, conf);
1640 if (!conf->r10buf_pool)
1642 conf->next_resync = 0;
1647 * perform a "sync" on one "block"
1649 * We need to make sure that no normal I/O request - particularly write
1650 * requests - conflict with active sync requests.
1652 * This is achieved by tracking pending requests and a 'barrier' concept
1653 * that can be installed to exclude normal IO requests.
1655 * Resync and recovery are handled very differently.
1656 * We differentiate by looking at MD_RECOVERY_SYNC in mddev->recovery.
1658 * For resync, we iterate over virtual addresses, read all copies,
1659 * and update if there are differences. If only one copy is live,
1661 * For recovery, we iterate over physical addresses, read a good
1662 * value for each non-in_sync drive, and over-write.
1664 * So, for recovery we may have several outstanding complex requests for a
1665 * given address, one for each out-of-sync device. We model this by allocating
1666 * a number of r10_bio structures, one for each out-of-sync device.
1667 * As we setup these structures, we collect all bio's together into a list
1668 * which we then process collectively to add pages, and then process again
1669 * to pass to generic_make_request.
1671 * The r10_bio structures are linked using a borrowed master_bio pointer.
1672 * This link is counted in ->remaining. When the r10_bio that points to NULL
1673 * has its remaining count decremented to 0, the whole complex operation
1678 static sector_t sync_request(mddev_t *mddev, sector_t sector_nr, int *skipped, int go_faster)
1680 conf_t *conf = mddev_to_conf(mddev);
1682 struct bio *biolist = NULL, *bio;
1683 sector_t max_sector, nr_sectors;
1689 sector_t sectors_skipped = 0;
1690 int chunks_skipped = 0;
1692 if (!conf->r10buf_pool)
1693 if (init_resync(conf))
1697 max_sector = mddev->size << 1;
1698 if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery))
1699 max_sector = mddev->resync_max_sectors;
1700 if (sector_nr >= max_sector) {
1701 /* If we aborted, we need to abort the
1702 * sync on the 'current' bitmap chucks (there can
1703 * be several when recovering multiple devices).
1704 * as we may have started syncing it but not finished.
1705 * We can find the current address in
1706 * mddev->curr_resync, but for recovery,
1707 * we need to convert that to several
1708 * virtual addresses.
1710 if (mddev->curr_resync < max_sector) { /* aborted */
1711 if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery))
1712 bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
1714 else for (i=0; i<conf->raid_disks; i++) {
1716 raid10_find_virt(conf, mddev->curr_resync, i);
1717 bitmap_end_sync(mddev->bitmap, sect,
1720 } else /* completed sync */
1723 bitmap_close_sync(mddev->bitmap);
1726 return sectors_skipped;
1728 if (chunks_skipped >= conf->raid_disks) {
1729 /* if there has been nothing to do on any drive,
1730 * then there is nothing to do at all..
1733 return (max_sector - sector_nr) + sectors_skipped;
1736 if (max_sector > mddev->resync_max)
1737 max_sector = mddev->resync_max; /* Don't do IO beyond here */
1739 /* make sure whole request will fit in a chunk - if chunks
1742 if (conf->near_copies < conf->raid_disks &&
1743 max_sector > (sector_nr | conf->chunk_mask))
1744 max_sector = (sector_nr | conf->chunk_mask) + 1;
1746 * If there is non-resync activity waiting for us then
1747 * put in a delay to throttle resync.
1749 if (!go_faster && conf->nr_waiting)
1750 msleep_interruptible(1000);
1752 bitmap_cond_end_sync(mddev->bitmap, sector_nr);
1754 /* Again, very different code for resync and recovery.
1755 * Both must result in an r10bio with a list of bios that
1756 * have bi_end_io, bi_sector, bi_bdev set,
1757 * and bi_private set to the r10bio.
1758 * For recovery, we may actually create several r10bios
1759 * with 2 bios in each, that correspond to the bios in the main one.
1760 * In this case, the subordinate r10bios link back through a
1761 * borrowed master_bio pointer, and the counter in the master
1762 * includes a ref from each subordinate.
1764 /* First, we decide what to do and set ->bi_end_io
1765 * To end_sync_read if we want to read, and
1766 * end_sync_write if we will want to write.
1769 max_sync = RESYNC_PAGES << (PAGE_SHIFT-9);
1770 if (!test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
1771 /* recovery... the complicated one */
1775 for (i=0 ; i<conf->raid_disks; i++)
1776 if (conf->mirrors[i].rdev &&
1777 !test_bit(In_sync, &conf->mirrors[i].rdev->flags)) {
1778 int still_degraded = 0;
1779 /* want to reconstruct this device */
1780 r10bio_t *rb2 = r10_bio;
1781 sector_t sect = raid10_find_virt(conf, sector_nr, i);
1783 /* Unless we are doing a full sync, we only need
1784 * to recover the block if it is set in the bitmap
1786 must_sync = bitmap_start_sync(mddev->bitmap, sect,
1788 if (sync_blocks < max_sync)
1789 max_sync = sync_blocks;
1792 /* yep, skip the sync_blocks here, but don't assume
1793 * that there will never be anything to do here
1795 chunks_skipped = -1;
1799 r10_bio = mempool_alloc(conf->r10buf_pool, GFP_NOIO);
1800 raise_barrier(conf, rb2 != NULL);
1801 atomic_set(&r10_bio->remaining, 0);
1803 r10_bio->master_bio = (struct bio*)rb2;
1805 atomic_inc(&rb2->remaining);
1806 r10_bio->mddev = mddev;
1807 set_bit(R10BIO_IsRecover, &r10_bio->state);
1808 r10_bio->sector = sect;
1810 raid10_find_phys(conf, r10_bio);
1811 /* Need to check if this section will still be
1814 for (j=0; j<conf->copies;j++) {
1815 int d = r10_bio->devs[j].devnum;
1816 if (conf->mirrors[d].rdev == NULL ||
1817 test_bit(Faulty, &conf->mirrors[d].rdev->flags)) {
1822 must_sync = bitmap_start_sync(mddev->bitmap, sect,
1823 &sync_blocks, still_degraded);
1825 for (j=0; j<conf->copies;j++) {
1826 int d = r10_bio->devs[j].devnum;
1827 if (conf->mirrors[d].rdev &&
1828 test_bit(In_sync, &conf->mirrors[d].rdev->flags)) {
1829 /* This is where we read from */
1830 bio = r10_bio->devs[0].bio;
1831 bio->bi_next = biolist;
1833 bio->bi_private = r10_bio;
1834 bio->bi_end_io = end_sync_read;
1836 bio->bi_sector = r10_bio->devs[j].addr +
1837 conf->mirrors[d].rdev->data_offset;
1838 bio->bi_bdev = conf->mirrors[d].rdev->bdev;
1839 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
1840 atomic_inc(&r10_bio->remaining);
1841 /* and we write to 'i' */
1843 for (k=0; k<conf->copies; k++)
1844 if (r10_bio->devs[k].devnum == i)
1846 BUG_ON(k == conf->copies);
1847 bio = r10_bio->devs[1].bio;
1848 bio->bi_next = biolist;
1850 bio->bi_private = r10_bio;
1851 bio->bi_end_io = end_sync_write;
1853 bio->bi_sector = r10_bio->devs[k].addr +
1854 conf->mirrors[i].rdev->data_offset;
1855 bio->bi_bdev = conf->mirrors[i].rdev->bdev;
1857 r10_bio->devs[0].devnum = d;
1858 r10_bio->devs[1].devnum = i;
1863 if (j == conf->copies) {
1864 /* Cannot recover, so abort the recovery */
1867 atomic_dec(&rb2->remaining);
1869 if (!test_and_set_bit(MD_RECOVERY_INTR,
1871 printk(KERN_INFO "raid10: %s: insufficient working devices for recovery.\n",
1876 if (biolist == NULL) {
1878 r10bio_t *rb2 = r10_bio;
1879 r10_bio = (r10bio_t*) rb2->master_bio;
1880 rb2->master_bio = NULL;
1886 /* resync. Schedule a read for every block at this virt offset */
1889 if (!bitmap_start_sync(mddev->bitmap, sector_nr,
1890 &sync_blocks, mddev->degraded) &&
1891 !conf->fullsync && !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
1892 /* We can skip this block */
1894 return sync_blocks + sectors_skipped;
1896 if (sync_blocks < max_sync)
1897 max_sync = sync_blocks;
1898 r10_bio = mempool_alloc(conf->r10buf_pool, GFP_NOIO);
1900 r10_bio->mddev = mddev;
1901 atomic_set(&r10_bio->remaining, 0);
1902 raise_barrier(conf, 0);
1903 conf->next_resync = sector_nr;
1905 r10_bio->master_bio = NULL;
1906 r10_bio->sector = sector_nr;
1907 set_bit(R10BIO_IsSync, &r10_bio->state);
1908 raid10_find_phys(conf, r10_bio);
1909 r10_bio->sectors = (sector_nr | conf->chunk_mask) - sector_nr +1;
1911 for (i=0; i<conf->copies; i++) {
1912 int d = r10_bio->devs[i].devnum;
1913 bio = r10_bio->devs[i].bio;
1914 bio->bi_end_io = NULL;
1915 clear_bit(BIO_UPTODATE, &bio->bi_flags);
1916 if (conf->mirrors[d].rdev == NULL ||
1917 test_bit(Faulty, &conf->mirrors[d].rdev->flags))
1919 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
1920 atomic_inc(&r10_bio->remaining);
1921 bio->bi_next = biolist;
1923 bio->bi_private = r10_bio;
1924 bio->bi_end_io = end_sync_read;
1926 bio->bi_sector = r10_bio->devs[i].addr +
1927 conf->mirrors[d].rdev->data_offset;
1928 bio->bi_bdev = conf->mirrors[d].rdev->bdev;
1933 for (i=0; i<conf->copies; i++) {
1934 int d = r10_bio->devs[i].devnum;
1935 if (r10_bio->devs[i].bio->bi_end_io)
1936 rdev_dec_pending(conf->mirrors[d].rdev, mddev);
1944 for (bio = biolist; bio ; bio=bio->bi_next) {
1946 bio->bi_flags &= ~(BIO_POOL_MASK - 1);
1948 bio->bi_flags |= 1 << BIO_UPTODATE;
1951 bio->bi_phys_segments = 0;
1956 if (sector_nr + max_sync < max_sector)
1957 max_sector = sector_nr + max_sync;
1960 int len = PAGE_SIZE;
1962 if (sector_nr + (len>>9) > max_sector)
1963 len = (max_sector - sector_nr) << 9;
1966 for (bio= biolist ; bio ; bio=bio->bi_next) {
1967 page = bio->bi_io_vec[bio->bi_vcnt].bv_page;
1968 if (bio_add_page(bio, page, len, 0) == 0) {
1971 bio->bi_io_vec[bio->bi_vcnt].bv_page = page;
1972 for (bio2 = biolist; bio2 && bio2 != bio; bio2 = bio2->bi_next) {
1973 /* remove last page from this bio */
1975 bio2->bi_size -= len;
1976 bio2->bi_flags &= ~(1<< BIO_SEG_VALID);
1982 nr_sectors += len>>9;
1983 sector_nr += len>>9;
1984 } while (biolist->bi_vcnt < RESYNC_PAGES);
1986 r10_bio->sectors = nr_sectors;
1990 biolist = biolist->bi_next;
1992 bio->bi_next = NULL;
1993 r10_bio = bio->bi_private;
1994 r10_bio->sectors = nr_sectors;
1996 if (bio->bi_end_io == end_sync_read) {
1997 md_sync_acct(bio->bi_bdev, nr_sectors);
1998 generic_make_request(bio);
2002 if (sectors_skipped)
2003 /* pretend they weren't skipped, it makes
2004 * no important difference in this case
2006 md_done_sync(mddev, sectors_skipped, 1);
2008 return sectors_skipped + nr_sectors;
2010 /* There is nowhere to write, so all non-sync
2011 * drives must be failed, so try the next chunk...
2014 sector_t sec = max_sector - sector_nr;
2015 sectors_skipped += sec;
2017 sector_nr = max_sector;
2022 static int run(mddev_t *mddev)
2026 mirror_info_t *disk;
2029 sector_t stride, size;
2031 if (mddev->chunk_size < PAGE_SIZE) {
2032 printk(KERN_ERR "md/raid10: chunk size must be "
2033 "at least PAGE_SIZE(%ld).\n", PAGE_SIZE);
2037 nc = mddev->layout & 255;
2038 fc = (mddev->layout >> 8) & 255;
2039 fo = mddev->layout & (1<<16);
2040 if ((nc*fc) <2 || (nc*fc) > mddev->raid_disks ||
2041 (mddev->layout >> 17)) {
2042 printk(KERN_ERR "raid10: %s: unsupported raid10 layout: 0x%8x\n",
2043 mdname(mddev), mddev->layout);
2047 * copy the already verified devices into our private RAID10
2048 * bookkeeping area. [whatever we allocate in run(),
2049 * should be freed in stop()]
2051 conf = kzalloc(sizeof(conf_t), GFP_KERNEL);
2052 mddev->private = conf;
2054 printk(KERN_ERR "raid10: couldn't allocate memory for %s\n",
2058 conf->mirrors = kzalloc(sizeof(struct mirror_info)*mddev->raid_disks,
2060 if (!conf->mirrors) {
2061 printk(KERN_ERR "raid10: couldn't allocate memory for %s\n",
2066 conf->tmppage = alloc_page(GFP_KERNEL);
2070 conf->mddev = mddev;
2071 conf->raid_disks = mddev->raid_disks;
2072 conf->near_copies = nc;
2073 conf->far_copies = fc;
2074 conf->copies = nc*fc;
2075 conf->far_offset = fo;
2076 conf->chunk_mask = (sector_t)(mddev->chunk_size>>9)-1;
2077 conf->chunk_shift = ffz(~mddev->chunk_size) - 9;
2078 size = mddev->size >> (conf->chunk_shift-1);
2079 sector_div(size, fc);
2080 size = size * conf->raid_disks;
2081 sector_div(size, nc);
2082 /* 'size' is now the number of chunks in the array */
2083 /* calculate "used chunks per device" in 'stride' */
2084 stride = size * conf->copies;
2086 /* We need to round up when dividing by raid_disks to
2087 * get the stride size.
2089 stride += conf->raid_disks - 1;
2090 sector_div(stride, conf->raid_disks);
2091 mddev->size = stride << (conf->chunk_shift-1);
2096 sector_div(stride, fc);
2097 conf->stride = stride << conf->chunk_shift;
2099 conf->r10bio_pool = mempool_create(NR_RAID10_BIOS, r10bio_pool_alloc,
2100 r10bio_pool_free, conf);
2101 if (!conf->r10bio_pool) {
2102 printk(KERN_ERR "raid10: couldn't allocate memory for %s\n",
2107 spin_lock_init(&conf->device_lock);
2108 mddev->queue->queue_lock = &conf->device_lock;
2110 list_for_each_entry(rdev, &mddev->disks, same_set) {
2111 disk_idx = rdev->raid_disk;
2112 if (disk_idx >= mddev->raid_disks
2115 disk = conf->mirrors + disk_idx;
2119 blk_queue_stack_limits(mddev->queue,
2120 rdev->bdev->bd_disk->queue);
2121 /* as we don't honour merge_bvec_fn, we must never risk
2122 * violating it, so limit ->max_sector to one PAGE, as
2123 * a one page request is never in violation.
2125 if (rdev->bdev->bd_disk->queue->merge_bvec_fn &&
2126 mddev->queue->max_sectors > (PAGE_SIZE>>9))
2127 mddev->queue->max_sectors = (PAGE_SIZE>>9);
2129 disk->head_position = 0;
2131 INIT_LIST_HEAD(&conf->retry_list);
2133 spin_lock_init(&conf->resync_lock);
2134 init_waitqueue_head(&conf->wait_barrier);
2136 /* need to check that every block has at least one working mirror */
2137 if (!enough(conf)) {
2138 printk(KERN_ERR "raid10: not enough operational mirrors for %s\n",
2143 mddev->degraded = 0;
2144 for (i = 0; i < conf->raid_disks; i++) {
2146 disk = conf->mirrors + i;
2149 !test_bit(In_sync, &disk->rdev->flags)) {
2150 disk->head_position = 0;
2158 mddev->thread = md_register_thread(raid10d, mddev, "%s_raid10");
2159 if (!mddev->thread) {
2161 "raid10: couldn't allocate thread for %s\n",
2167 "raid10: raid set %s active with %d out of %d devices\n",
2168 mdname(mddev), mddev->raid_disks - mddev->degraded,
2171 * Ok, everything is just fine now
2173 mddev->array_sectors = size << conf->chunk_shift;
2174 mddev->resync_max_sectors = size << conf->chunk_shift;
2176 mddev->queue->unplug_fn = raid10_unplug;
2177 mddev->queue->backing_dev_info.congested_fn = raid10_congested;
2178 mddev->queue->backing_dev_info.congested_data = mddev;
2180 /* Calculate max read-ahead size.
2181 * We need to readahead at least twice a whole stripe....
2185 int stripe = conf->raid_disks * (mddev->chunk_size / PAGE_SIZE);
2186 stripe /= conf->near_copies;
2187 if (mddev->queue->backing_dev_info.ra_pages < 2* stripe)
2188 mddev->queue->backing_dev_info.ra_pages = 2* stripe;
2191 if (conf->near_copies < mddev->raid_disks)
2192 blk_queue_merge_bvec(mddev->queue, raid10_mergeable_bvec);
2196 if (conf->r10bio_pool)
2197 mempool_destroy(conf->r10bio_pool);
2198 safe_put_page(conf->tmppage);
2199 kfree(conf->mirrors);
2201 mddev->private = NULL;
2206 static int stop(mddev_t *mddev)
2208 conf_t *conf = mddev_to_conf(mddev);
2210 md_unregister_thread(mddev->thread);
2211 mddev->thread = NULL;
2212 blk_sync_queue(mddev->queue); /* the unplug fn references 'conf'*/
2213 if (conf->r10bio_pool)
2214 mempool_destroy(conf->r10bio_pool);
2215 kfree(conf->mirrors);
2217 mddev->private = NULL;
2221 static void raid10_quiesce(mddev_t *mddev, int state)
2223 conf_t *conf = mddev_to_conf(mddev);
2227 raise_barrier(conf, 0);
2230 lower_barrier(conf);
2233 if (mddev->thread) {
2235 mddev->thread->timeout = mddev->bitmap->daemon_sleep * HZ;
2237 mddev->thread->timeout = MAX_SCHEDULE_TIMEOUT;
2238 md_wakeup_thread(mddev->thread);
2242 static struct mdk_personality raid10_personality =
2246 .owner = THIS_MODULE,
2247 .make_request = make_request,
2251 .error_handler = error,
2252 .hot_add_disk = raid10_add_disk,
2253 .hot_remove_disk= raid10_remove_disk,
2254 .spare_active = raid10_spare_active,
2255 .sync_request = sync_request,
2256 .quiesce = raid10_quiesce,
2259 static int __init raid_init(void)
2261 return register_md_personality(&raid10_personality);
2264 static void raid_exit(void)
2266 unregister_md_personality(&raid10_personality);
2269 module_init(raid_init);
2270 module_exit(raid_exit);
2271 MODULE_LICENSE("GPL");
2272 MODULE_ALIAS("md-personality-9"); /* RAID10 */
2273 MODULE_ALIAS("md-raid10");
2274 MODULE_ALIAS("md-level-10");