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/raid/raid10.h>
23 #include <linux/raid/bitmap.h>
26 * RAID10 provides a combination of RAID0 and RAID1 functionality.
27 * The layout of data is defined by
30 * near_copies (stored in low byte of layout)
31 * far_copies (stored in second byte of layout)
32 * far_offset (stored in bit 16 of layout )
34 * The data to be stored is divided into chunks using chunksize.
35 * Each device is divided into far_copies sections.
36 * In each section, chunks are laid out in a style similar to raid0, but
37 * near_copies copies of each chunk is stored (each on a different drive).
38 * The starting device for each section is offset near_copies from the starting
39 * device of the previous section.
40 * Thus they are (near_copies*far_copies) of each chunk, and each is on a different
42 * near_copies and far_copies must be at least one, and their product is at most
45 * If far_offset is true, then the far_copies are handled a bit differently.
46 * The copies are still in different stripes, but instead of be very far apart
47 * on disk, there are adjacent stripes.
51 * Number of guaranteed r10bios in case of extreme VM load:
53 #define NR_RAID10_BIOS 256
55 static void unplug_slaves(mddev_t *mddev);
57 static void allow_barrier(conf_t *conf);
58 static void lower_barrier(conf_t *conf);
60 static void * r10bio_pool_alloc(gfp_t gfp_flags, void *data)
64 int size = offsetof(struct r10bio_s, devs[conf->copies]);
66 /* allocate a r10bio with room for raid_disks entries in the bios array */
67 r10_bio = kzalloc(size, gfp_flags);
69 unplug_slaves(conf->mddev);
74 static void r10bio_pool_free(void *r10_bio, void *data)
79 #define RESYNC_BLOCK_SIZE (64*1024)
80 //#define RESYNC_BLOCK_SIZE PAGE_SIZE
81 #define RESYNC_SECTORS (RESYNC_BLOCK_SIZE >> 9)
82 #define RESYNC_PAGES ((RESYNC_BLOCK_SIZE + PAGE_SIZE-1) / PAGE_SIZE)
83 #define RESYNC_WINDOW (2048*1024)
86 * When performing a resync, we need to read and compare, so
87 * we need as many pages are there are copies.
88 * When performing a recovery, we need 2 bios, one for read,
89 * one for write (we recover only one drive per r10buf)
92 static void * r10buf_pool_alloc(gfp_t gfp_flags, void *data)
101 r10_bio = r10bio_pool_alloc(gfp_flags, conf);
103 unplug_slaves(conf->mddev);
107 if (test_bit(MD_RECOVERY_SYNC, &conf->mddev->recovery))
108 nalloc = conf->copies; /* resync */
110 nalloc = 2; /* recovery */
115 for (j = nalloc ; j-- ; ) {
116 bio = bio_alloc(gfp_flags, RESYNC_PAGES);
119 r10_bio->devs[j].bio = bio;
122 * Allocate RESYNC_PAGES data pages and attach them
125 for (j = 0 ; j < nalloc; j++) {
126 bio = r10_bio->devs[j].bio;
127 for (i = 0; i < RESYNC_PAGES; i++) {
128 page = alloc_page(gfp_flags);
132 bio->bi_io_vec[i].bv_page = page;
140 safe_put_page(bio->bi_io_vec[i-1].bv_page);
142 for (i = 0; i < RESYNC_PAGES ; i++)
143 safe_put_page(r10_bio->devs[j].bio->bi_io_vec[i].bv_page);
146 while ( ++j < nalloc )
147 bio_put(r10_bio->devs[j].bio);
148 r10bio_pool_free(r10_bio, conf);
152 static void r10buf_pool_free(void *__r10_bio, void *data)
156 r10bio_t *r10bio = __r10_bio;
159 for (j=0; j < conf->copies; j++) {
160 struct bio *bio = r10bio->devs[j].bio;
162 for (i = 0; i < RESYNC_PAGES; i++) {
163 safe_put_page(bio->bi_io_vec[i].bv_page);
164 bio->bi_io_vec[i].bv_page = NULL;
169 r10bio_pool_free(r10bio, conf);
172 static void put_all_bios(conf_t *conf, r10bio_t *r10_bio)
176 for (i = 0; i < conf->copies; i++) {
177 struct bio **bio = & r10_bio->devs[i].bio;
178 if (*bio && *bio != IO_BLOCKED)
184 static void free_r10bio(r10bio_t *r10_bio)
186 conf_t *conf = mddev_to_conf(r10_bio->mddev);
189 * Wake up any possible resync thread that waits for the device
194 put_all_bios(conf, r10_bio);
195 mempool_free(r10_bio, conf->r10bio_pool);
198 static void put_buf(r10bio_t *r10_bio)
200 conf_t *conf = mddev_to_conf(r10_bio->mddev);
202 mempool_free(r10_bio, conf->r10buf_pool);
207 static void reschedule_retry(r10bio_t *r10_bio)
210 mddev_t *mddev = r10_bio->mddev;
211 conf_t *conf = mddev_to_conf(mddev);
213 spin_lock_irqsave(&conf->device_lock, flags);
214 list_add(&r10_bio->retry_list, &conf->retry_list);
216 spin_unlock_irqrestore(&conf->device_lock, flags);
218 md_wakeup_thread(mddev->thread);
222 * raid_end_bio_io() is called when we have finished servicing a mirrored
223 * operation and are ready to return a success/failure code to the buffer
226 static void raid_end_bio_io(r10bio_t *r10_bio)
228 struct bio *bio = r10_bio->master_bio;
231 test_bit(R10BIO_Uptodate, &r10_bio->state) ? 0 : -EIO);
232 free_r10bio(r10_bio);
236 * Update disk head position estimator based on IRQ completion info.
238 static inline void update_head_pos(int slot, r10bio_t *r10_bio)
240 conf_t *conf = mddev_to_conf(r10_bio->mddev);
242 conf->mirrors[r10_bio->devs[slot].devnum].head_position =
243 r10_bio->devs[slot].addr + (r10_bio->sectors);
246 static void raid10_end_read_request(struct bio *bio, int error)
248 int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
249 r10bio_t * r10_bio = (r10bio_t *)(bio->bi_private);
251 conf_t *conf = mddev_to_conf(r10_bio->mddev);
254 slot = r10_bio->read_slot;
255 dev = r10_bio->devs[slot].devnum;
257 * this branch is our 'one mirror IO has finished' event handler:
259 update_head_pos(slot, r10_bio);
263 * Set R10BIO_Uptodate in our master bio, so that
264 * we will return a good error code to the higher
265 * levels even if IO on some other mirrored buffer fails.
267 * The 'master' represents the composite IO operation to
268 * user-side. So if something waits for IO, then it will
269 * wait for the 'master' bio.
271 set_bit(R10BIO_Uptodate, &r10_bio->state);
272 raid_end_bio_io(r10_bio);
277 char b[BDEVNAME_SIZE];
278 if (printk_ratelimit())
279 printk(KERN_ERR "raid10: %s: rescheduling sector %llu\n",
280 bdevname(conf->mirrors[dev].rdev->bdev,b), (unsigned long long)r10_bio->sector);
281 reschedule_retry(r10_bio);
284 rdev_dec_pending(conf->mirrors[dev].rdev, conf->mddev);
287 static void raid10_end_write_request(struct bio *bio, int error)
289 int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
290 r10bio_t * r10_bio = (r10bio_t *)(bio->bi_private);
292 conf_t *conf = mddev_to_conf(r10_bio->mddev);
294 for (slot = 0; slot < conf->copies; slot++)
295 if (r10_bio->devs[slot].bio == bio)
297 dev = r10_bio->devs[slot].devnum;
300 * this branch is our 'one mirror IO has finished' event handler:
303 md_error(r10_bio->mddev, conf->mirrors[dev].rdev);
304 /* an I/O failed, we can't clear the bitmap */
305 set_bit(R10BIO_Degraded, &r10_bio->state);
308 * Set R10BIO_Uptodate in our master bio, so that
309 * we will return a good error code for to the higher
310 * levels even if IO on some other mirrored buffer fails.
312 * The 'master' represents the composite IO operation to
313 * user-side. So if something waits for IO, then it will
314 * wait for the 'master' bio.
316 set_bit(R10BIO_Uptodate, &r10_bio->state);
318 update_head_pos(slot, r10_bio);
322 * Let's see if all mirrored write operations have finished
325 if (atomic_dec_and_test(&r10_bio->remaining)) {
326 /* clear the bitmap if all writes complete successfully */
327 bitmap_endwrite(r10_bio->mddev->bitmap, r10_bio->sector,
329 !test_bit(R10BIO_Degraded, &r10_bio->state),
331 md_write_end(r10_bio->mddev);
332 raid_end_bio_io(r10_bio);
335 rdev_dec_pending(conf->mirrors[dev].rdev, conf->mddev);
340 * RAID10 layout manager
341 * Aswell as the chunksize and raid_disks count, there are two
342 * parameters: near_copies and far_copies.
343 * near_copies * far_copies must be <= raid_disks.
344 * Normally one of these will be 1.
345 * If both are 1, we get raid0.
346 * If near_copies == raid_disks, we get raid1.
348 * Chunks are layed out in raid0 style with near_copies copies of the
349 * first chunk, followed by near_copies copies of the next chunk and
351 * If far_copies > 1, then after 1/far_copies of the array has been assigned
352 * as described above, we start again with a device offset of near_copies.
353 * So we effectively have another copy of the whole array further down all
354 * the drives, but with blocks on different drives.
355 * With this layout, and block is never stored twice on the one device.
357 * raid10_find_phys finds the sector offset of a given virtual sector
358 * on each device that it is on.
360 * raid10_find_virt does the reverse mapping, from a device and a
361 * sector offset to a virtual address
364 static void raid10_find_phys(conf_t *conf, r10bio_t *r10bio)
374 /* now calculate first sector/dev */
375 chunk = r10bio->sector >> conf->chunk_shift;
376 sector = r10bio->sector & conf->chunk_mask;
378 chunk *= conf->near_copies;
380 dev = sector_div(stripe, conf->raid_disks);
381 if (conf->far_offset)
382 stripe *= conf->far_copies;
384 sector += stripe << conf->chunk_shift;
386 /* and calculate all the others */
387 for (n=0; n < conf->near_copies; n++) {
390 r10bio->devs[slot].addr = sector;
391 r10bio->devs[slot].devnum = d;
394 for (f = 1; f < conf->far_copies; f++) {
395 d += conf->near_copies;
396 if (d >= conf->raid_disks)
397 d -= conf->raid_disks;
399 r10bio->devs[slot].devnum = d;
400 r10bio->devs[slot].addr = s;
404 if (dev >= conf->raid_disks) {
406 sector += (conf->chunk_mask + 1);
409 BUG_ON(slot != conf->copies);
412 static sector_t raid10_find_virt(conf_t *conf, sector_t sector, int dev)
414 sector_t offset, chunk, vchunk;
416 offset = sector & conf->chunk_mask;
417 if (conf->far_offset) {
419 chunk = sector >> conf->chunk_shift;
420 fc = sector_div(chunk, conf->far_copies);
421 dev -= fc * conf->near_copies;
423 dev += conf->raid_disks;
425 while (sector >= conf->stride) {
426 sector -= conf->stride;
427 if (dev < conf->near_copies)
428 dev += conf->raid_disks - conf->near_copies;
430 dev -= conf->near_copies;
432 chunk = sector >> conf->chunk_shift;
434 vchunk = chunk * conf->raid_disks + dev;
435 sector_div(vchunk, conf->near_copies);
436 return (vchunk << conf->chunk_shift) + offset;
440 * raid10_mergeable_bvec -- tell bio layer if a two requests can be merged
442 * @bio: the buffer head that's been built up so far
443 * @biovec: the request that could be merged to it.
445 * Return amount of bytes we can accept at this offset
446 * If near_copies == raid_disk, there are no striping issues,
447 * but in that case, the function isn't called at all.
449 static int raid10_mergeable_bvec(struct request_queue *q, struct bio *bio,
450 struct bio_vec *bio_vec)
452 mddev_t *mddev = q->queuedata;
453 sector_t sector = bio->bi_sector + get_start_sect(bio->bi_bdev);
455 unsigned int chunk_sectors = mddev->chunk_size >> 9;
456 unsigned int bio_sectors = bio->bi_size >> 9;
458 max = (chunk_sectors - ((sector & (chunk_sectors - 1)) + bio_sectors)) << 9;
459 if (max < 0) max = 0; /* bio_add cannot handle a negative return */
460 if (max <= bio_vec->bv_len && bio_sectors == 0)
461 return bio_vec->bv_len;
467 * This routine returns the disk from which the requested read should
468 * be done. There is a per-array 'next expected sequential IO' sector
469 * number - if this matches on the next IO then we use the last disk.
470 * There is also a per-disk 'last know head position' sector that is
471 * maintained from IRQ contexts, both the normal and the resync IO
472 * completion handlers update this position correctly. If there is no
473 * perfect sequential match then we pick the disk whose head is closest.
475 * If there are 2 mirrors in the same 2 devices, performance degrades
476 * because position is mirror, not device based.
478 * The rdev for the device selected will have nr_pending incremented.
482 * FIXME: possibly should rethink readbalancing and do it differently
483 * depending on near_copies / far_copies geometry.
485 static int read_balance(conf_t *conf, r10bio_t *r10_bio)
487 const unsigned long this_sector = r10_bio->sector;
488 int disk, slot, nslot;
489 const int sectors = r10_bio->sectors;
490 sector_t new_distance, current_distance;
493 raid10_find_phys(conf, r10_bio);
496 * Check if we can balance. We can balance on the whole
497 * device if no resync is going on (recovery is ok), or below
498 * the resync window. We take the first readable disk when
499 * above the resync window.
501 if (conf->mddev->recovery_cp < MaxSector
502 && (this_sector + sectors >= conf->next_resync)) {
503 /* make sure that disk is operational */
505 disk = r10_bio->devs[slot].devnum;
507 while ((rdev = rcu_dereference(conf->mirrors[disk].rdev)) == NULL ||
508 r10_bio->devs[slot].bio == IO_BLOCKED ||
509 !test_bit(In_sync, &rdev->flags)) {
511 if (slot == conf->copies) {
516 disk = r10_bio->devs[slot].devnum;
522 /* make sure the disk is operational */
524 disk = r10_bio->devs[slot].devnum;
525 while ((rdev=rcu_dereference(conf->mirrors[disk].rdev)) == NULL ||
526 r10_bio->devs[slot].bio == IO_BLOCKED ||
527 !test_bit(In_sync, &rdev->flags)) {
529 if (slot == conf->copies) {
533 disk = r10_bio->devs[slot].devnum;
537 current_distance = abs(r10_bio->devs[slot].addr -
538 conf->mirrors[disk].head_position);
540 /* Find the disk whose head is closest */
542 for (nslot = slot; nslot < conf->copies; nslot++) {
543 int ndisk = r10_bio->devs[nslot].devnum;
546 if ((rdev=rcu_dereference(conf->mirrors[ndisk].rdev)) == NULL ||
547 r10_bio->devs[nslot].bio == IO_BLOCKED ||
548 !test_bit(In_sync, &rdev->flags))
551 /* This optimisation is debatable, and completely destroys
552 * sequential read speed for 'far copies' arrays. So only
553 * keep it for 'near' arrays, and review those later.
555 if (conf->near_copies > 1 && !atomic_read(&rdev->nr_pending)) {
560 new_distance = abs(r10_bio->devs[nslot].addr -
561 conf->mirrors[ndisk].head_position);
562 if (new_distance < current_distance) {
563 current_distance = new_distance;
570 r10_bio->read_slot = slot;
571 /* conf->next_seq_sect = this_sector + sectors;*/
573 if (disk >= 0 && (rdev=rcu_dereference(conf->mirrors[disk].rdev))!= NULL)
574 atomic_inc(&conf->mirrors[disk].rdev->nr_pending);
582 static void unplug_slaves(mddev_t *mddev)
584 conf_t *conf = mddev_to_conf(mddev);
588 for (i=0; i<mddev->raid_disks; i++) {
589 mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[i].rdev);
590 if (rdev && !test_bit(Faulty, &rdev->flags) && atomic_read(&rdev->nr_pending)) {
591 struct request_queue *r_queue = bdev_get_queue(rdev->bdev);
593 atomic_inc(&rdev->nr_pending);
598 rdev_dec_pending(rdev, mddev);
605 static void raid10_unplug(struct request_queue *q)
607 mddev_t *mddev = q->queuedata;
609 unplug_slaves(q->queuedata);
610 md_wakeup_thread(mddev->thread);
613 static int raid10_congested(void *data, int bits)
615 mddev_t *mddev = data;
616 conf_t *conf = mddev_to_conf(mddev);
620 for (i = 0; i < mddev->raid_disks && ret == 0; i++) {
621 mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[i].rdev);
622 if (rdev && !test_bit(Faulty, &rdev->flags)) {
623 struct request_queue *q = bdev_get_queue(rdev->bdev);
625 ret |= bdi_congested(&q->backing_dev_info, bits);
632 static int flush_pending_writes(conf_t *conf)
634 /* Any writes that have been queued but are awaiting
635 * bitmap updates get flushed here.
636 * We return 1 if any requests were actually submitted.
640 spin_lock_irq(&conf->device_lock);
642 if (conf->pending_bio_list.head) {
644 bio = bio_list_get(&conf->pending_bio_list);
645 blk_remove_plug(conf->mddev->queue);
646 spin_unlock_irq(&conf->device_lock);
647 /* flush any pending bitmap writes to disk
648 * before proceeding w/ I/O */
649 bitmap_unplug(conf->mddev->bitmap);
651 while (bio) { /* submit pending writes */
652 struct bio *next = bio->bi_next;
654 generic_make_request(bio);
659 spin_unlock_irq(&conf->device_lock);
663 * Sometimes we need to suspend IO while we do something else,
664 * either some resync/recovery, or reconfigure the array.
665 * To do this we raise a 'barrier'.
666 * The 'barrier' is a counter that can be raised multiple times
667 * to count how many activities are happening which preclude
669 * We can only raise the barrier if there is no pending IO.
670 * i.e. if nr_pending == 0.
671 * We choose only to raise the barrier if no-one is waiting for the
672 * barrier to go down. This means that as soon as an IO request
673 * is ready, no other operations which require a barrier will start
674 * until the IO request has had a chance.
676 * So: regular IO calls 'wait_barrier'. When that returns there
677 * is no backgroup IO happening, It must arrange to call
678 * allow_barrier when it has finished its IO.
679 * backgroup IO calls must call raise_barrier. Once that returns
680 * there is no normal IO happeing. It must arrange to call
681 * lower_barrier when the particular background IO completes.
683 #define RESYNC_DEPTH 32
685 static void raise_barrier(conf_t *conf, int force)
687 BUG_ON(force && !conf->barrier);
688 spin_lock_irq(&conf->resync_lock);
690 /* Wait until no block IO is waiting (unless 'force') */
691 wait_event_lock_irq(conf->wait_barrier, force || !conf->nr_waiting,
693 raid10_unplug(conf->mddev->queue));
695 /* block any new IO from starting */
698 /* No wait for all pending IO to complete */
699 wait_event_lock_irq(conf->wait_barrier,
700 !conf->nr_pending && conf->barrier < RESYNC_DEPTH,
702 raid10_unplug(conf->mddev->queue));
704 spin_unlock_irq(&conf->resync_lock);
707 static void lower_barrier(conf_t *conf)
710 spin_lock_irqsave(&conf->resync_lock, flags);
712 spin_unlock_irqrestore(&conf->resync_lock, flags);
713 wake_up(&conf->wait_barrier);
716 static void wait_barrier(conf_t *conf)
718 spin_lock_irq(&conf->resync_lock);
721 wait_event_lock_irq(conf->wait_barrier, !conf->barrier,
723 raid10_unplug(conf->mddev->queue));
727 spin_unlock_irq(&conf->resync_lock);
730 static void allow_barrier(conf_t *conf)
733 spin_lock_irqsave(&conf->resync_lock, flags);
735 spin_unlock_irqrestore(&conf->resync_lock, flags);
736 wake_up(&conf->wait_barrier);
739 static void freeze_array(conf_t *conf)
741 /* stop syncio and normal IO and wait for everything to
743 * We increment barrier and nr_waiting, and then
744 * wait until barrier+nr_pending match nr_queued+2
746 spin_lock_irq(&conf->resync_lock);
749 wait_event_lock_irq(conf->wait_barrier,
750 conf->barrier+conf->nr_pending == conf->nr_queued+2,
752 ({ flush_pending_writes(conf);
753 raid10_unplug(conf->mddev->queue); }));
754 spin_unlock_irq(&conf->resync_lock);
757 static void unfreeze_array(conf_t *conf)
759 /* reverse the effect of the freeze */
760 spin_lock_irq(&conf->resync_lock);
763 wake_up(&conf->wait_barrier);
764 spin_unlock_irq(&conf->resync_lock);
767 static int make_request(struct request_queue *q, struct bio * bio)
769 mddev_t *mddev = q->queuedata;
770 conf_t *conf = mddev_to_conf(mddev);
771 mirror_info_t *mirror;
773 struct bio *read_bio;
775 int chunk_sects = conf->chunk_mask + 1;
776 const int rw = bio_data_dir(bio);
777 const int do_sync = bio_sync(bio);
781 if (unlikely(bio_barrier(bio))) {
782 bio_endio(bio, -EOPNOTSUPP);
786 /* If this request crosses a chunk boundary, we need to
787 * split it. This will only happen for 1 PAGE (or less) requests.
789 if (unlikely( (bio->bi_sector & conf->chunk_mask) + (bio->bi_size >> 9)
791 conf->near_copies < conf->raid_disks)) {
793 /* Sanity check -- queue functions should prevent this happening */
794 if (bio->bi_vcnt != 1 ||
797 /* This is a one page bio that upper layers
798 * refuse to split for us, so we need to split it.
800 bp = bio_split(bio, bio_split_pool,
801 chunk_sects - (bio->bi_sector & (chunk_sects - 1)) );
802 if (make_request(q, &bp->bio1))
803 generic_make_request(&bp->bio1);
804 if (make_request(q, &bp->bio2))
805 generic_make_request(&bp->bio2);
807 bio_pair_release(bp);
810 printk("raid10_make_request bug: can't convert block across chunks"
811 " or bigger than %dk %llu %d\n", chunk_sects/2,
812 (unsigned long long)bio->bi_sector, bio->bi_size >> 10);
818 md_write_start(mddev, bio);
821 * Register the new request and wait if the reconstruction
822 * thread has put up a bar for new requests.
823 * Continue immediately if no resync is active currently.
827 disk_stat_inc(mddev->gendisk, ios[rw]);
828 disk_stat_add(mddev->gendisk, sectors[rw], bio_sectors(bio));
830 r10_bio = mempool_alloc(conf->r10bio_pool, GFP_NOIO);
832 r10_bio->master_bio = bio;
833 r10_bio->sectors = bio->bi_size >> 9;
835 r10_bio->mddev = mddev;
836 r10_bio->sector = bio->bi_sector;
841 * read balancing logic:
843 int disk = read_balance(conf, r10_bio);
844 int slot = r10_bio->read_slot;
846 raid_end_bio_io(r10_bio);
849 mirror = conf->mirrors + disk;
851 read_bio = bio_clone(bio, GFP_NOIO);
853 r10_bio->devs[slot].bio = read_bio;
855 read_bio->bi_sector = r10_bio->devs[slot].addr +
856 mirror->rdev->data_offset;
857 read_bio->bi_bdev = mirror->rdev->bdev;
858 read_bio->bi_end_io = raid10_end_read_request;
859 read_bio->bi_rw = READ | do_sync;
860 read_bio->bi_private = r10_bio;
862 generic_make_request(read_bio);
869 /* first select target devices under spinlock and
870 * inc refcount on their rdev. Record them by setting
873 raid10_find_phys(conf, r10_bio);
875 for (i = 0; i < conf->copies; i++) {
876 int d = r10_bio->devs[i].devnum;
877 mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[d].rdev);
879 !test_bit(Faulty, &rdev->flags)) {
880 atomic_inc(&rdev->nr_pending);
881 r10_bio->devs[i].bio = bio;
883 r10_bio->devs[i].bio = NULL;
884 set_bit(R10BIO_Degraded, &r10_bio->state);
889 atomic_set(&r10_bio->remaining, 0);
892 for (i = 0; i < conf->copies; i++) {
894 int d = r10_bio->devs[i].devnum;
895 if (!r10_bio->devs[i].bio)
898 mbio = bio_clone(bio, GFP_NOIO);
899 r10_bio->devs[i].bio = mbio;
901 mbio->bi_sector = r10_bio->devs[i].addr+
902 conf->mirrors[d].rdev->data_offset;
903 mbio->bi_bdev = conf->mirrors[d].rdev->bdev;
904 mbio->bi_end_io = raid10_end_write_request;
905 mbio->bi_rw = WRITE | do_sync;
906 mbio->bi_private = r10_bio;
908 atomic_inc(&r10_bio->remaining);
909 bio_list_add(&bl, mbio);
912 if (unlikely(!atomic_read(&r10_bio->remaining))) {
913 /* the array is dead */
915 raid_end_bio_io(r10_bio);
919 bitmap_startwrite(mddev->bitmap, bio->bi_sector, r10_bio->sectors, 0);
920 spin_lock_irqsave(&conf->device_lock, flags);
921 bio_list_merge(&conf->pending_bio_list, &bl);
922 blk_plug_device(mddev->queue);
923 spin_unlock_irqrestore(&conf->device_lock, flags);
925 /* In case raid10d snuck in to freeze_array */
926 wake_up(&conf->wait_barrier);
929 md_wakeup_thread(mddev->thread);
934 static void status(struct seq_file *seq, mddev_t *mddev)
936 conf_t *conf = mddev_to_conf(mddev);
939 if (conf->near_copies < conf->raid_disks)
940 seq_printf(seq, " %dK chunks", mddev->chunk_size/1024);
941 if (conf->near_copies > 1)
942 seq_printf(seq, " %d near-copies", conf->near_copies);
943 if (conf->far_copies > 1) {
944 if (conf->far_offset)
945 seq_printf(seq, " %d offset-copies", conf->far_copies);
947 seq_printf(seq, " %d far-copies", conf->far_copies);
949 seq_printf(seq, " [%d/%d] [", conf->raid_disks,
950 conf->raid_disks - mddev->degraded);
951 for (i = 0; i < conf->raid_disks; i++)
952 seq_printf(seq, "%s",
953 conf->mirrors[i].rdev &&
954 test_bit(In_sync, &conf->mirrors[i].rdev->flags) ? "U" : "_");
955 seq_printf(seq, "]");
958 static void error(mddev_t *mddev, mdk_rdev_t *rdev)
960 char b[BDEVNAME_SIZE];
961 conf_t *conf = mddev_to_conf(mddev);
964 * If it is not operational, then we have already marked it as dead
965 * else if it is the last working disks, ignore the error, let the
966 * next level up know.
967 * else mark the drive as failed
969 if (test_bit(In_sync, &rdev->flags)
970 && conf->raid_disks-mddev->degraded == 1)
972 * Don't fail the drive, just return an IO error.
973 * The test should really be more sophisticated than
974 * "working_disks == 1", but it isn't critical, and
975 * can wait until we do more sophisticated "is the drive
976 * really dead" tests...
979 if (test_and_clear_bit(In_sync, &rdev->flags)) {
981 spin_lock_irqsave(&conf->device_lock, flags);
983 spin_unlock_irqrestore(&conf->device_lock, flags);
985 * if recovery is running, make sure it aborts.
987 set_bit(MD_RECOVERY_ERR, &mddev->recovery);
989 set_bit(Faulty, &rdev->flags);
990 set_bit(MD_CHANGE_DEVS, &mddev->flags);
991 printk(KERN_ALERT "raid10: Disk failure on %s, disabling device. \n"
992 " Operation continuing on %d devices\n",
993 bdevname(rdev->bdev,b), conf->raid_disks - mddev->degraded);
996 static void print_conf(conf_t *conf)
1001 printk("RAID10 conf printout:\n");
1003 printk("(!conf)\n");
1006 printk(" --- wd:%d rd:%d\n", conf->raid_disks - conf->mddev->degraded,
1009 for (i = 0; i < conf->raid_disks; i++) {
1010 char b[BDEVNAME_SIZE];
1011 tmp = conf->mirrors + i;
1013 printk(" disk %d, wo:%d, o:%d, dev:%s\n",
1014 i, !test_bit(In_sync, &tmp->rdev->flags),
1015 !test_bit(Faulty, &tmp->rdev->flags),
1016 bdevname(tmp->rdev->bdev,b));
1020 static void close_sync(conf_t *conf)
1023 allow_barrier(conf);
1025 mempool_destroy(conf->r10buf_pool);
1026 conf->r10buf_pool = NULL;
1029 /* check if there are enough drives for
1030 * every block to appear on atleast one
1032 static int enough(conf_t *conf)
1037 int n = conf->copies;
1040 if (conf->mirrors[first].rdev)
1042 first = (first+1) % conf->raid_disks;
1046 } while (first != 0);
1050 static int raid10_spare_active(mddev_t *mddev)
1053 conf_t *conf = mddev->private;
1057 * Find all non-in_sync disks within the RAID10 configuration
1058 * and mark them in_sync
1060 for (i = 0; i < conf->raid_disks; i++) {
1061 tmp = conf->mirrors + i;
1063 && !test_bit(Faulty, &tmp->rdev->flags)
1064 && !test_and_set_bit(In_sync, &tmp->rdev->flags)) {
1065 unsigned long flags;
1066 spin_lock_irqsave(&conf->device_lock, flags);
1068 spin_unlock_irqrestore(&conf->device_lock, flags);
1077 static int raid10_add_disk(mddev_t *mddev, mdk_rdev_t *rdev)
1079 conf_t *conf = mddev->private;
1084 if (mddev->recovery_cp < MaxSector)
1085 /* only hot-add to in-sync arrays, as recovery is
1086 * very different from resync
1092 if (rdev->saved_raid_disk >= 0 &&
1093 conf->mirrors[rdev->saved_raid_disk].rdev == NULL)
1094 mirror = rdev->saved_raid_disk;
1097 for ( ; mirror < mddev->raid_disks; mirror++)
1098 if ( !(p=conf->mirrors+mirror)->rdev) {
1100 blk_queue_stack_limits(mddev->queue,
1101 rdev->bdev->bd_disk->queue);
1102 /* as we don't honour merge_bvec_fn, we must never risk
1103 * violating it, so limit ->max_sector to one PAGE, as
1104 * a one page request is never in violation.
1106 if (rdev->bdev->bd_disk->queue->merge_bvec_fn &&
1107 mddev->queue->max_sectors > (PAGE_SIZE>>9))
1108 mddev->queue->max_sectors = (PAGE_SIZE>>9);
1110 p->head_position = 0;
1111 rdev->raid_disk = mirror;
1113 if (rdev->saved_raid_disk != mirror)
1115 rcu_assign_pointer(p->rdev, rdev);
1123 static int raid10_remove_disk(mddev_t *mddev, int number)
1125 conf_t *conf = mddev->private;
1128 mirror_info_t *p = conf->mirrors+ number;
1133 if (test_bit(In_sync, &rdev->flags) ||
1134 atomic_read(&rdev->nr_pending)) {
1140 if (atomic_read(&rdev->nr_pending)) {
1141 /* lost the race, try later */
1153 static void end_sync_read(struct bio *bio, int error)
1155 r10bio_t * r10_bio = (r10bio_t *)(bio->bi_private);
1156 conf_t *conf = mddev_to_conf(r10_bio->mddev);
1159 for (i=0; i<conf->copies; i++)
1160 if (r10_bio->devs[i].bio == bio)
1162 BUG_ON(i == conf->copies);
1163 update_head_pos(i, r10_bio);
1164 d = r10_bio->devs[i].devnum;
1166 if (test_bit(BIO_UPTODATE, &bio->bi_flags))
1167 set_bit(R10BIO_Uptodate, &r10_bio->state);
1169 atomic_add(r10_bio->sectors,
1170 &conf->mirrors[d].rdev->corrected_errors);
1171 if (!test_bit(MD_RECOVERY_SYNC, &conf->mddev->recovery))
1172 md_error(r10_bio->mddev,
1173 conf->mirrors[d].rdev);
1176 /* for reconstruct, we always reschedule after a read.
1177 * for resync, only after all reads
1179 if (test_bit(R10BIO_IsRecover, &r10_bio->state) ||
1180 atomic_dec_and_test(&r10_bio->remaining)) {
1181 /* we have read all the blocks,
1182 * do the comparison in process context in raid10d
1184 reschedule_retry(r10_bio);
1186 rdev_dec_pending(conf->mirrors[d].rdev, conf->mddev);
1189 static void end_sync_write(struct bio *bio, int error)
1191 int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
1192 r10bio_t * r10_bio = (r10bio_t *)(bio->bi_private);
1193 mddev_t *mddev = r10_bio->mddev;
1194 conf_t *conf = mddev_to_conf(mddev);
1197 for (i = 0; i < conf->copies; i++)
1198 if (r10_bio->devs[i].bio == bio)
1200 d = r10_bio->devs[i].devnum;
1203 md_error(mddev, conf->mirrors[d].rdev);
1204 update_head_pos(i, r10_bio);
1206 while (atomic_dec_and_test(&r10_bio->remaining)) {
1207 if (r10_bio->master_bio == NULL) {
1208 /* the primary of several recovery bios */
1209 md_done_sync(mddev, r10_bio->sectors, 1);
1213 r10bio_t *r10_bio2 = (r10bio_t *)r10_bio->master_bio;
1218 rdev_dec_pending(conf->mirrors[d].rdev, mddev);
1222 * Note: sync and recover and handled very differently for raid10
1223 * This code is for resync.
1224 * For resync, we read through virtual addresses and read all blocks.
1225 * If there is any error, we schedule a write. The lowest numbered
1226 * drive is authoritative.
1227 * However requests come for physical address, so we need to map.
1228 * For every physical address there are raid_disks/copies virtual addresses,
1229 * which is always are least one, but is not necessarly an integer.
1230 * This means that a physical address can span multiple chunks, so we may
1231 * have to submit multiple io requests for a single sync request.
1234 * We check if all blocks are in-sync and only write to blocks that
1237 static void sync_request_write(mddev_t *mddev, r10bio_t *r10_bio)
1239 conf_t *conf = mddev_to_conf(mddev);
1241 struct bio *tbio, *fbio;
1243 atomic_set(&r10_bio->remaining, 1);
1245 /* find the first device with a block */
1246 for (i=0; i<conf->copies; i++)
1247 if (test_bit(BIO_UPTODATE, &r10_bio->devs[i].bio->bi_flags))
1250 if (i == conf->copies)
1254 fbio = r10_bio->devs[i].bio;
1256 /* now find blocks with errors */
1257 for (i=0 ; i < conf->copies ; i++) {
1259 int vcnt = r10_bio->sectors >> (PAGE_SHIFT-9);
1261 tbio = r10_bio->devs[i].bio;
1263 if (tbio->bi_end_io != end_sync_read)
1267 if (test_bit(BIO_UPTODATE, &r10_bio->devs[i].bio->bi_flags)) {
1268 /* We know that the bi_io_vec layout is the same for
1269 * both 'first' and 'i', so we just compare them.
1270 * All vec entries are PAGE_SIZE;
1272 for (j = 0; j < vcnt; j++)
1273 if (memcmp(page_address(fbio->bi_io_vec[j].bv_page),
1274 page_address(tbio->bi_io_vec[j].bv_page),
1279 mddev->resync_mismatches += r10_bio->sectors;
1281 if (test_bit(MD_RECOVERY_CHECK, &mddev->recovery))
1282 /* Don't fix anything. */
1284 /* Ok, we need to write this bio
1285 * First we need to fixup bv_offset, bv_len and
1286 * bi_vecs, as the read request might have corrupted these
1288 tbio->bi_vcnt = vcnt;
1289 tbio->bi_size = r10_bio->sectors << 9;
1291 tbio->bi_phys_segments = 0;
1292 tbio->bi_hw_segments = 0;
1293 tbio->bi_hw_front_size = 0;
1294 tbio->bi_hw_back_size = 0;
1295 tbio->bi_flags &= ~(BIO_POOL_MASK - 1);
1296 tbio->bi_flags |= 1 << BIO_UPTODATE;
1297 tbio->bi_next = NULL;
1298 tbio->bi_rw = WRITE;
1299 tbio->bi_private = r10_bio;
1300 tbio->bi_sector = r10_bio->devs[i].addr;
1302 for (j=0; j < vcnt ; j++) {
1303 tbio->bi_io_vec[j].bv_offset = 0;
1304 tbio->bi_io_vec[j].bv_len = PAGE_SIZE;
1306 memcpy(page_address(tbio->bi_io_vec[j].bv_page),
1307 page_address(fbio->bi_io_vec[j].bv_page),
1310 tbio->bi_end_io = end_sync_write;
1312 d = r10_bio->devs[i].devnum;
1313 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
1314 atomic_inc(&r10_bio->remaining);
1315 md_sync_acct(conf->mirrors[d].rdev->bdev, tbio->bi_size >> 9);
1317 tbio->bi_sector += conf->mirrors[d].rdev->data_offset;
1318 tbio->bi_bdev = conf->mirrors[d].rdev->bdev;
1319 generic_make_request(tbio);
1323 if (atomic_dec_and_test(&r10_bio->remaining)) {
1324 md_done_sync(mddev, r10_bio->sectors, 1);
1330 * Now for the recovery code.
1331 * Recovery happens across physical sectors.
1332 * We recover all non-is_sync drives by finding the virtual address of
1333 * each, and then choose a working drive that also has that virt address.
1334 * There is a separate r10_bio for each non-in_sync drive.
1335 * Only the first two slots are in use. The first for reading,
1336 * The second for writing.
1340 static void recovery_request_write(mddev_t *mddev, r10bio_t *r10_bio)
1342 conf_t *conf = mddev_to_conf(mddev);
1344 struct bio *bio, *wbio;
1347 /* move the pages across to the second bio
1348 * and submit the write request
1350 bio = r10_bio->devs[0].bio;
1351 wbio = r10_bio->devs[1].bio;
1352 for (i=0; i < wbio->bi_vcnt; i++) {
1353 struct page *p = bio->bi_io_vec[i].bv_page;
1354 bio->bi_io_vec[i].bv_page = wbio->bi_io_vec[i].bv_page;
1355 wbio->bi_io_vec[i].bv_page = p;
1357 d = r10_bio->devs[1].devnum;
1359 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
1360 md_sync_acct(conf->mirrors[d].rdev->bdev, wbio->bi_size >> 9);
1361 if (test_bit(R10BIO_Uptodate, &r10_bio->state))
1362 generic_make_request(wbio);
1364 bio_endio(wbio, -EIO);
1369 * This is a kernel thread which:
1371 * 1. Retries failed read operations on working mirrors.
1372 * 2. Updates the raid superblock when problems encounter.
1373 * 3. Performs writes following reads for array synchronising.
1376 static void fix_read_error(conf_t *conf, mddev_t *mddev, r10bio_t *r10_bio)
1378 int sect = 0; /* Offset from r10_bio->sector */
1379 int sectors = r10_bio->sectors;
1383 int sl = r10_bio->read_slot;
1387 if (s > (PAGE_SIZE>>9))
1392 int d = r10_bio->devs[sl].devnum;
1393 rdev = rcu_dereference(conf->mirrors[d].rdev);
1395 test_bit(In_sync, &rdev->flags)) {
1396 atomic_inc(&rdev->nr_pending);
1398 success = sync_page_io(rdev->bdev,
1399 r10_bio->devs[sl].addr +
1400 sect + rdev->data_offset,
1402 conf->tmppage, READ);
1403 rdev_dec_pending(rdev, mddev);
1409 if (sl == conf->copies)
1411 } while (!success && sl != r10_bio->read_slot);
1415 /* Cannot read from anywhere -- bye bye array */
1416 int dn = r10_bio->devs[r10_bio->read_slot].devnum;
1417 md_error(mddev, conf->mirrors[dn].rdev);
1422 /* write it back and re-read */
1424 while (sl != r10_bio->read_slot) {
1429 d = r10_bio->devs[sl].devnum;
1430 rdev = rcu_dereference(conf->mirrors[d].rdev);
1432 test_bit(In_sync, &rdev->flags)) {
1433 atomic_inc(&rdev->nr_pending);
1435 atomic_add(s, &rdev->corrected_errors);
1436 if (sync_page_io(rdev->bdev,
1437 r10_bio->devs[sl].addr +
1438 sect + rdev->data_offset,
1439 s<<9, conf->tmppage, WRITE)
1441 /* Well, this device is dead */
1442 md_error(mddev, rdev);
1443 rdev_dec_pending(rdev, mddev);
1448 while (sl != r10_bio->read_slot) {
1453 d = r10_bio->devs[sl].devnum;
1454 rdev = rcu_dereference(conf->mirrors[d].rdev);
1456 test_bit(In_sync, &rdev->flags)) {
1457 char b[BDEVNAME_SIZE];
1458 atomic_inc(&rdev->nr_pending);
1460 if (sync_page_io(rdev->bdev,
1461 r10_bio->devs[sl].addr +
1462 sect + rdev->data_offset,
1463 s<<9, conf->tmppage, READ) == 0)
1464 /* Well, this device is dead */
1465 md_error(mddev, rdev);
1468 "raid10:%s: read error corrected"
1469 " (%d sectors at %llu on %s)\n",
1471 (unsigned long long)(sect+
1473 bdevname(rdev->bdev, b));
1475 rdev_dec_pending(rdev, mddev);
1486 static void raid10d(mddev_t *mddev)
1490 unsigned long flags;
1491 conf_t *conf = mddev_to_conf(mddev);
1492 struct list_head *head = &conf->retry_list;
1496 md_check_recovery(mddev);
1499 char b[BDEVNAME_SIZE];
1501 unplug += flush_pending_writes(conf);
1503 spin_lock_irqsave(&conf->device_lock, flags);
1504 if (list_empty(head)) {
1505 spin_unlock_irqrestore(&conf->device_lock, flags);
1508 r10_bio = list_entry(head->prev, r10bio_t, retry_list);
1509 list_del(head->prev);
1511 spin_unlock_irqrestore(&conf->device_lock, flags);
1513 mddev = r10_bio->mddev;
1514 conf = mddev_to_conf(mddev);
1515 if (test_bit(R10BIO_IsSync, &r10_bio->state)) {
1516 sync_request_write(mddev, r10_bio);
1518 } else if (test_bit(R10BIO_IsRecover, &r10_bio->state)) {
1519 recovery_request_write(mddev, r10_bio);
1523 /* we got a read error. Maybe the drive is bad. Maybe just
1524 * the block and we can fix it.
1525 * We freeze all other IO, and try reading the block from
1526 * other devices. When we find one, we re-write
1527 * and check it that fixes the read error.
1528 * This is all done synchronously while the array is
1531 if (mddev->ro == 0) {
1533 fix_read_error(conf, mddev, r10_bio);
1534 unfreeze_array(conf);
1537 bio = r10_bio->devs[r10_bio->read_slot].bio;
1538 r10_bio->devs[r10_bio->read_slot].bio =
1539 mddev->ro ? IO_BLOCKED : NULL;
1540 mirror = read_balance(conf, r10_bio);
1542 printk(KERN_ALERT "raid10: %s: unrecoverable I/O"
1543 " read error for block %llu\n",
1544 bdevname(bio->bi_bdev,b),
1545 (unsigned long long)r10_bio->sector);
1546 raid_end_bio_io(r10_bio);
1549 const int do_sync = bio_sync(r10_bio->master_bio);
1551 rdev = conf->mirrors[mirror].rdev;
1552 if (printk_ratelimit())
1553 printk(KERN_ERR "raid10: %s: redirecting sector %llu to"
1554 " another mirror\n",
1555 bdevname(rdev->bdev,b),
1556 (unsigned long long)r10_bio->sector);
1557 bio = bio_clone(r10_bio->master_bio, GFP_NOIO);
1558 r10_bio->devs[r10_bio->read_slot].bio = bio;
1559 bio->bi_sector = r10_bio->devs[r10_bio->read_slot].addr
1560 + rdev->data_offset;
1561 bio->bi_bdev = rdev->bdev;
1562 bio->bi_rw = READ | do_sync;
1563 bio->bi_private = r10_bio;
1564 bio->bi_end_io = raid10_end_read_request;
1566 generic_make_request(bio);
1571 unplug_slaves(mddev);
1575 static int init_resync(conf_t *conf)
1579 buffs = RESYNC_WINDOW / RESYNC_BLOCK_SIZE;
1580 BUG_ON(conf->r10buf_pool);
1581 conf->r10buf_pool = mempool_create(buffs, r10buf_pool_alloc, r10buf_pool_free, conf);
1582 if (!conf->r10buf_pool)
1584 conf->next_resync = 0;
1589 * perform a "sync" on one "block"
1591 * We need to make sure that no normal I/O request - particularly write
1592 * requests - conflict with active sync requests.
1594 * This is achieved by tracking pending requests and a 'barrier' concept
1595 * that can be installed to exclude normal IO requests.
1597 * Resync and recovery are handled very differently.
1598 * We differentiate by looking at MD_RECOVERY_SYNC in mddev->recovery.
1600 * For resync, we iterate over virtual addresses, read all copies,
1601 * and update if there are differences. If only one copy is live,
1603 * For recovery, we iterate over physical addresses, read a good
1604 * value for each non-in_sync drive, and over-write.
1606 * So, for recovery we may have several outstanding complex requests for a
1607 * given address, one for each out-of-sync device. We model this by allocating
1608 * a number of r10_bio structures, one for each out-of-sync device.
1609 * As we setup these structures, we collect all bio's together into a list
1610 * which we then process collectively to add pages, and then process again
1611 * to pass to generic_make_request.
1613 * The r10_bio structures are linked using a borrowed master_bio pointer.
1614 * This link is counted in ->remaining. When the r10_bio that points to NULL
1615 * has its remaining count decremented to 0, the whole complex operation
1620 static sector_t sync_request(mddev_t *mddev, sector_t sector_nr, int *skipped, int go_faster)
1622 conf_t *conf = mddev_to_conf(mddev);
1624 struct bio *biolist = NULL, *bio;
1625 sector_t max_sector, nr_sectors;
1631 sector_t sectors_skipped = 0;
1632 int chunks_skipped = 0;
1634 if (!conf->r10buf_pool)
1635 if (init_resync(conf))
1639 max_sector = mddev->size << 1;
1640 if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery))
1641 max_sector = mddev->resync_max_sectors;
1642 if (sector_nr >= max_sector) {
1643 /* If we aborted, we need to abort the
1644 * sync on the 'current' bitmap chucks (there can
1645 * be several when recovering multiple devices).
1646 * as we may have started syncing it but not finished.
1647 * We can find the current address in
1648 * mddev->curr_resync, but for recovery,
1649 * we need to convert that to several
1650 * virtual addresses.
1652 if (mddev->curr_resync < max_sector) { /* aborted */
1653 if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery))
1654 bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
1656 else for (i=0; i<conf->raid_disks; i++) {
1658 raid10_find_virt(conf, mddev->curr_resync, i);
1659 bitmap_end_sync(mddev->bitmap, sect,
1662 } else /* completed sync */
1665 bitmap_close_sync(mddev->bitmap);
1668 return sectors_skipped;
1670 if (chunks_skipped >= conf->raid_disks) {
1671 /* if there has been nothing to do on any drive,
1672 * then there is nothing to do at all..
1675 return (max_sector - sector_nr) + sectors_skipped;
1678 if (max_sector > mddev->resync_max)
1679 max_sector = mddev->resync_max; /* Don't do IO beyond here */
1681 /* make sure whole request will fit in a chunk - if chunks
1684 if (conf->near_copies < conf->raid_disks &&
1685 max_sector > (sector_nr | conf->chunk_mask))
1686 max_sector = (sector_nr | conf->chunk_mask) + 1;
1688 * If there is non-resync activity waiting for us then
1689 * put in a delay to throttle resync.
1691 if (!go_faster && conf->nr_waiting)
1692 msleep_interruptible(1000);
1694 bitmap_cond_end_sync(mddev->bitmap, sector_nr);
1696 /* Again, very different code for resync and recovery.
1697 * Both must result in an r10bio with a list of bios that
1698 * have bi_end_io, bi_sector, bi_bdev set,
1699 * and bi_private set to the r10bio.
1700 * For recovery, we may actually create several r10bios
1701 * with 2 bios in each, that correspond to the bios in the main one.
1702 * In this case, the subordinate r10bios link back through a
1703 * borrowed master_bio pointer, and the counter in the master
1704 * includes a ref from each subordinate.
1706 /* First, we decide what to do and set ->bi_end_io
1707 * To end_sync_read if we want to read, and
1708 * end_sync_write if we will want to write.
1711 max_sync = RESYNC_PAGES << (PAGE_SHIFT-9);
1712 if (!test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
1713 /* recovery... the complicated one */
1717 for (i=0 ; i<conf->raid_disks; i++)
1718 if (conf->mirrors[i].rdev &&
1719 !test_bit(In_sync, &conf->mirrors[i].rdev->flags)) {
1720 int still_degraded = 0;
1721 /* want to reconstruct this device */
1722 r10bio_t *rb2 = r10_bio;
1723 sector_t sect = raid10_find_virt(conf, sector_nr, i);
1725 /* Unless we are doing a full sync, we only need
1726 * to recover the block if it is set in the bitmap
1728 must_sync = bitmap_start_sync(mddev->bitmap, sect,
1730 if (sync_blocks < max_sync)
1731 max_sync = sync_blocks;
1734 /* yep, skip the sync_blocks here, but don't assume
1735 * that there will never be anything to do here
1737 chunks_skipped = -1;
1741 r10_bio = mempool_alloc(conf->r10buf_pool, GFP_NOIO);
1742 raise_barrier(conf, rb2 != NULL);
1743 atomic_set(&r10_bio->remaining, 0);
1745 r10_bio->master_bio = (struct bio*)rb2;
1747 atomic_inc(&rb2->remaining);
1748 r10_bio->mddev = mddev;
1749 set_bit(R10BIO_IsRecover, &r10_bio->state);
1750 r10_bio->sector = sect;
1752 raid10_find_phys(conf, r10_bio);
1753 /* Need to check if this section will still be
1756 for (j=0; j<conf->copies;j++) {
1757 int d = r10_bio->devs[j].devnum;
1758 if (conf->mirrors[d].rdev == NULL ||
1759 test_bit(Faulty, &conf->mirrors[d].rdev->flags)) {
1764 must_sync = bitmap_start_sync(mddev->bitmap, sect,
1765 &sync_blocks, still_degraded);
1767 for (j=0; j<conf->copies;j++) {
1768 int d = r10_bio->devs[j].devnum;
1769 if (conf->mirrors[d].rdev &&
1770 test_bit(In_sync, &conf->mirrors[d].rdev->flags)) {
1771 /* This is where we read from */
1772 bio = r10_bio->devs[0].bio;
1773 bio->bi_next = biolist;
1775 bio->bi_private = r10_bio;
1776 bio->bi_end_io = end_sync_read;
1778 bio->bi_sector = r10_bio->devs[j].addr +
1779 conf->mirrors[d].rdev->data_offset;
1780 bio->bi_bdev = conf->mirrors[d].rdev->bdev;
1781 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
1782 atomic_inc(&r10_bio->remaining);
1783 /* and we write to 'i' */
1785 for (k=0; k<conf->copies; k++)
1786 if (r10_bio->devs[k].devnum == i)
1788 BUG_ON(k == conf->copies);
1789 bio = r10_bio->devs[1].bio;
1790 bio->bi_next = biolist;
1792 bio->bi_private = r10_bio;
1793 bio->bi_end_io = end_sync_write;
1795 bio->bi_sector = r10_bio->devs[k].addr +
1796 conf->mirrors[i].rdev->data_offset;
1797 bio->bi_bdev = conf->mirrors[i].rdev->bdev;
1799 r10_bio->devs[0].devnum = d;
1800 r10_bio->devs[1].devnum = i;
1805 if (j == conf->copies) {
1806 /* Cannot recover, so abort the recovery */
1809 if (!test_and_set_bit(MD_RECOVERY_ERR, &mddev->recovery))
1810 printk(KERN_INFO "raid10: %s: insufficient working devices for recovery.\n",
1815 if (biolist == NULL) {
1817 r10bio_t *rb2 = r10_bio;
1818 r10_bio = (r10bio_t*) rb2->master_bio;
1819 rb2->master_bio = NULL;
1825 /* resync. Schedule a read for every block at this virt offset */
1828 if (!bitmap_start_sync(mddev->bitmap, sector_nr,
1829 &sync_blocks, mddev->degraded) &&
1830 !conf->fullsync && !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
1831 /* We can skip this block */
1833 return sync_blocks + sectors_skipped;
1835 if (sync_blocks < max_sync)
1836 max_sync = sync_blocks;
1837 r10_bio = mempool_alloc(conf->r10buf_pool, GFP_NOIO);
1839 r10_bio->mddev = mddev;
1840 atomic_set(&r10_bio->remaining, 0);
1841 raise_barrier(conf, 0);
1842 conf->next_resync = sector_nr;
1844 r10_bio->master_bio = NULL;
1845 r10_bio->sector = sector_nr;
1846 set_bit(R10BIO_IsSync, &r10_bio->state);
1847 raid10_find_phys(conf, r10_bio);
1848 r10_bio->sectors = (sector_nr | conf->chunk_mask) - sector_nr +1;
1850 for (i=0; i<conf->copies; i++) {
1851 int d = r10_bio->devs[i].devnum;
1852 bio = r10_bio->devs[i].bio;
1853 bio->bi_end_io = NULL;
1854 clear_bit(BIO_UPTODATE, &bio->bi_flags);
1855 if (conf->mirrors[d].rdev == NULL ||
1856 test_bit(Faulty, &conf->mirrors[d].rdev->flags))
1858 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
1859 atomic_inc(&r10_bio->remaining);
1860 bio->bi_next = biolist;
1862 bio->bi_private = r10_bio;
1863 bio->bi_end_io = end_sync_read;
1865 bio->bi_sector = r10_bio->devs[i].addr +
1866 conf->mirrors[d].rdev->data_offset;
1867 bio->bi_bdev = conf->mirrors[d].rdev->bdev;
1872 for (i=0; i<conf->copies; i++) {
1873 int d = r10_bio->devs[i].devnum;
1874 if (r10_bio->devs[i].bio->bi_end_io)
1875 rdev_dec_pending(conf->mirrors[d].rdev, mddev);
1883 for (bio = biolist; bio ; bio=bio->bi_next) {
1885 bio->bi_flags &= ~(BIO_POOL_MASK - 1);
1887 bio->bi_flags |= 1 << BIO_UPTODATE;
1890 bio->bi_phys_segments = 0;
1891 bio->bi_hw_segments = 0;
1896 if (sector_nr + max_sync < max_sector)
1897 max_sector = sector_nr + max_sync;
1900 int len = PAGE_SIZE;
1902 if (sector_nr + (len>>9) > max_sector)
1903 len = (max_sector - sector_nr) << 9;
1906 for (bio= biolist ; bio ; bio=bio->bi_next) {
1907 page = bio->bi_io_vec[bio->bi_vcnt].bv_page;
1908 if (bio_add_page(bio, page, len, 0) == 0) {
1911 bio->bi_io_vec[bio->bi_vcnt].bv_page = page;
1912 for (bio2 = biolist; bio2 && bio2 != bio; bio2 = bio2->bi_next) {
1913 /* remove last page from this bio */
1915 bio2->bi_size -= len;
1916 bio2->bi_flags &= ~(1<< BIO_SEG_VALID);
1922 nr_sectors += len>>9;
1923 sector_nr += len>>9;
1924 } while (biolist->bi_vcnt < RESYNC_PAGES);
1926 r10_bio->sectors = nr_sectors;
1930 biolist = biolist->bi_next;
1932 bio->bi_next = NULL;
1933 r10_bio = bio->bi_private;
1934 r10_bio->sectors = nr_sectors;
1936 if (bio->bi_end_io == end_sync_read) {
1937 md_sync_acct(bio->bi_bdev, nr_sectors);
1938 generic_make_request(bio);
1942 if (sectors_skipped)
1943 /* pretend they weren't skipped, it makes
1944 * no important difference in this case
1946 md_done_sync(mddev, sectors_skipped, 1);
1948 return sectors_skipped + nr_sectors;
1950 /* There is nowhere to write, so all non-sync
1951 * drives must be failed, so try the next chunk...
1954 sector_t sec = max_sector - sector_nr;
1955 sectors_skipped += sec;
1957 sector_nr = max_sector;
1962 static int run(mddev_t *mddev)
1966 mirror_info_t *disk;
1968 struct list_head *tmp;
1970 sector_t stride, size;
1972 if (mddev->chunk_size == 0) {
1973 printk(KERN_ERR "md/raid10: non-zero chunk size required.\n");
1977 nc = mddev->layout & 255;
1978 fc = (mddev->layout >> 8) & 255;
1979 fo = mddev->layout & (1<<16);
1980 if ((nc*fc) <2 || (nc*fc) > mddev->raid_disks ||
1981 (mddev->layout >> 17)) {
1982 printk(KERN_ERR "raid10: %s: unsupported raid10 layout: 0x%8x\n",
1983 mdname(mddev), mddev->layout);
1987 * copy the already verified devices into our private RAID10
1988 * bookkeeping area. [whatever we allocate in run(),
1989 * should be freed in stop()]
1991 conf = kzalloc(sizeof(conf_t), GFP_KERNEL);
1992 mddev->private = conf;
1994 printk(KERN_ERR "raid10: couldn't allocate memory for %s\n",
1998 conf->mirrors = kzalloc(sizeof(struct mirror_info)*mddev->raid_disks,
2000 if (!conf->mirrors) {
2001 printk(KERN_ERR "raid10: couldn't allocate memory for %s\n",
2006 conf->tmppage = alloc_page(GFP_KERNEL);
2010 conf->mddev = mddev;
2011 conf->raid_disks = mddev->raid_disks;
2012 conf->near_copies = nc;
2013 conf->far_copies = fc;
2014 conf->copies = nc*fc;
2015 conf->far_offset = fo;
2016 conf->chunk_mask = (sector_t)(mddev->chunk_size>>9)-1;
2017 conf->chunk_shift = ffz(~mddev->chunk_size) - 9;
2018 size = mddev->size >> (conf->chunk_shift-1);
2019 sector_div(size, fc);
2020 size = size * conf->raid_disks;
2021 sector_div(size, nc);
2022 /* 'size' is now the number of chunks in the array */
2023 /* calculate "used chunks per device" in 'stride' */
2024 stride = size * conf->copies;
2026 /* We need to round up when dividing by raid_disks to
2027 * get the stride size.
2029 stride += conf->raid_disks - 1;
2030 sector_div(stride, conf->raid_disks);
2031 mddev->size = stride << (conf->chunk_shift-1);
2036 sector_div(stride, fc);
2037 conf->stride = stride << conf->chunk_shift;
2039 conf->r10bio_pool = mempool_create(NR_RAID10_BIOS, r10bio_pool_alloc,
2040 r10bio_pool_free, conf);
2041 if (!conf->r10bio_pool) {
2042 printk(KERN_ERR "raid10: couldn't allocate memory for %s\n",
2047 rdev_for_each(rdev, tmp, mddev) {
2048 disk_idx = rdev->raid_disk;
2049 if (disk_idx >= mddev->raid_disks
2052 disk = conf->mirrors + disk_idx;
2056 blk_queue_stack_limits(mddev->queue,
2057 rdev->bdev->bd_disk->queue);
2058 /* as we don't honour merge_bvec_fn, we must never risk
2059 * violating it, so limit ->max_sector to one PAGE, as
2060 * a one page request is never in violation.
2062 if (rdev->bdev->bd_disk->queue->merge_bvec_fn &&
2063 mddev->queue->max_sectors > (PAGE_SIZE>>9))
2064 mddev->queue->max_sectors = (PAGE_SIZE>>9);
2066 disk->head_position = 0;
2068 spin_lock_init(&conf->device_lock);
2069 INIT_LIST_HEAD(&conf->retry_list);
2071 spin_lock_init(&conf->resync_lock);
2072 init_waitqueue_head(&conf->wait_barrier);
2074 /* need to check that every block has at least one working mirror */
2075 if (!enough(conf)) {
2076 printk(KERN_ERR "raid10: not enough operational mirrors for %s\n",
2081 mddev->degraded = 0;
2082 for (i = 0; i < conf->raid_disks; i++) {
2084 disk = conf->mirrors + i;
2087 !test_bit(In_sync, &disk->rdev->flags)) {
2088 disk->head_position = 0;
2094 mddev->thread = md_register_thread(raid10d, mddev, "%s_raid10");
2095 if (!mddev->thread) {
2097 "raid10: couldn't allocate thread for %s\n",
2103 "raid10: raid set %s active with %d out of %d devices\n",
2104 mdname(mddev), mddev->raid_disks - mddev->degraded,
2107 * Ok, everything is just fine now
2109 mddev->array_size = size << (conf->chunk_shift-1);
2110 mddev->resync_max_sectors = size << conf->chunk_shift;
2112 mddev->queue->unplug_fn = raid10_unplug;
2113 mddev->queue->backing_dev_info.congested_fn = raid10_congested;
2114 mddev->queue->backing_dev_info.congested_data = mddev;
2116 /* Calculate max read-ahead size.
2117 * We need to readahead at least twice a whole stripe....
2121 int stripe = conf->raid_disks * (mddev->chunk_size / PAGE_SIZE);
2122 stripe /= conf->near_copies;
2123 if (mddev->queue->backing_dev_info.ra_pages < 2* stripe)
2124 mddev->queue->backing_dev_info.ra_pages = 2* stripe;
2127 if (conf->near_copies < mddev->raid_disks)
2128 blk_queue_merge_bvec(mddev->queue, raid10_mergeable_bvec);
2132 if (conf->r10bio_pool)
2133 mempool_destroy(conf->r10bio_pool);
2134 safe_put_page(conf->tmppage);
2135 kfree(conf->mirrors);
2137 mddev->private = NULL;
2142 static int stop(mddev_t *mddev)
2144 conf_t *conf = mddev_to_conf(mddev);
2146 md_unregister_thread(mddev->thread);
2147 mddev->thread = NULL;
2148 blk_sync_queue(mddev->queue); /* the unplug fn references 'conf'*/
2149 if (conf->r10bio_pool)
2150 mempool_destroy(conf->r10bio_pool);
2151 kfree(conf->mirrors);
2153 mddev->private = NULL;
2157 static void raid10_quiesce(mddev_t *mddev, int state)
2159 conf_t *conf = mddev_to_conf(mddev);
2163 raise_barrier(conf, 0);
2166 lower_barrier(conf);
2169 if (mddev->thread) {
2171 mddev->thread->timeout = mddev->bitmap->daemon_sleep * HZ;
2173 mddev->thread->timeout = MAX_SCHEDULE_TIMEOUT;
2174 md_wakeup_thread(mddev->thread);
2178 static struct mdk_personality raid10_personality =
2182 .owner = THIS_MODULE,
2183 .make_request = make_request,
2187 .error_handler = error,
2188 .hot_add_disk = raid10_add_disk,
2189 .hot_remove_disk= raid10_remove_disk,
2190 .spare_active = raid10_spare_active,
2191 .sync_request = sync_request,
2192 .quiesce = raid10_quiesce,
2195 static int __init raid_init(void)
2197 return register_md_personality(&raid10_personality);
2200 static void raid_exit(void)
2202 unregister_md_personality(&raid10_personality);
2205 module_init(raid_init);
2206 module_exit(raid_exit);
2207 MODULE_LICENSE("GPL");
2208 MODULE_ALIAS("md-personality-9"); /* RAID10 */
2209 MODULE_ALIAS("md-raid10");
2210 MODULE_ALIAS("md-level-10");