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;
230 bio_endio(bio, bio->bi_size,
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 int raid10_end_read_request(struct bio *bio, unsigned int bytes_done, 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);
256 slot = r10_bio->read_slot;
257 dev = r10_bio->devs[slot].devnum;
259 * this branch is our 'one mirror IO has finished' event handler:
261 update_head_pos(slot, r10_bio);
265 * Set R10BIO_Uptodate in our master bio, so that
266 * we will return a good error code to the higher
267 * levels even if IO on some other mirrored buffer fails.
269 * The 'master' represents the composite IO operation to
270 * user-side. So if something waits for IO, then it will
271 * wait for the 'master' bio.
273 set_bit(R10BIO_Uptodate, &r10_bio->state);
274 raid_end_bio_io(r10_bio);
279 char b[BDEVNAME_SIZE];
280 if (printk_ratelimit())
281 printk(KERN_ERR "raid10: %s: rescheduling sector %llu\n",
282 bdevname(conf->mirrors[dev].rdev->bdev,b), (unsigned long long)r10_bio->sector);
283 reschedule_retry(r10_bio);
286 rdev_dec_pending(conf->mirrors[dev].rdev, conf->mddev);
290 static int raid10_end_write_request(struct bio *bio, unsigned int bytes_done, int error)
292 int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
293 r10bio_t * r10_bio = (r10bio_t *)(bio->bi_private);
295 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);
347 * RAID10 layout manager
348 * Aswell as the chunksize and raid_disks count, there are two
349 * parameters: near_copies and far_copies.
350 * near_copies * far_copies must be <= raid_disks.
351 * Normally one of these will be 1.
352 * If both are 1, we get raid0.
353 * If near_copies == raid_disks, we get raid1.
355 * Chunks are layed out in raid0 style with near_copies copies of the
356 * first chunk, followed by near_copies copies of the next chunk and
358 * If far_copies > 1, then after 1/far_copies of the array has been assigned
359 * as described above, we start again with a device offset of near_copies.
360 * So we effectively have another copy of the whole array further down all
361 * the drives, but with blocks on different drives.
362 * With this layout, and block is never stored twice on the one device.
364 * raid10_find_phys finds the sector offset of a given virtual sector
365 * on each device that it is on.
367 * raid10_find_virt does the reverse mapping, from a device and a
368 * sector offset to a virtual address
371 static void raid10_find_phys(conf_t *conf, r10bio_t *r10bio)
381 /* now calculate first sector/dev */
382 chunk = r10bio->sector >> conf->chunk_shift;
383 sector = r10bio->sector & conf->chunk_mask;
385 chunk *= conf->near_copies;
387 dev = sector_div(stripe, conf->raid_disks);
388 if (conf->far_offset)
389 stripe *= conf->far_copies;
391 sector += stripe << conf->chunk_shift;
393 /* and calculate all the others */
394 for (n=0; n < conf->near_copies; n++) {
397 r10bio->devs[slot].addr = sector;
398 r10bio->devs[slot].devnum = d;
401 for (f = 1; f < conf->far_copies; f++) {
402 d += conf->near_copies;
403 if (d >= conf->raid_disks)
404 d -= conf->raid_disks;
406 r10bio->devs[slot].devnum = d;
407 r10bio->devs[slot].addr = s;
411 if (dev >= conf->raid_disks) {
413 sector += (conf->chunk_mask + 1);
416 BUG_ON(slot != conf->copies);
419 static sector_t raid10_find_virt(conf_t *conf, sector_t sector, int dev)
421 sector_t offset, chunk, vchunk;
423 offset = sector & conf->chunk_mask;
424 if (conf->far_offset) {
426 chunk = sector >> conf->chunk_shift;
427 fc = sector_div(chunk, conf->far_copies);
428 dev -= fc * conf->near_copies;
430 dev += conf->raid_disks;
432 while (sector > conf->stride) {
433 sector -= conf->stride;
434 if (dev < conf->near_copies)
435 dev += conf->raid_disks - conf->near_copies;
437 dev -= conf->near_copies;
439 chunk = sector >> conf->chunk_shift;
441 vchunk = chunk * conf->raid_disks + dev;
442 sector_div(vchunk, conf->near_copies);
443 return (vchunk << conf->chunk_shift) + offset;
447 * raid10_mergeable_bvec -- tell bio layer if a two requests can be merged
449 * @bio: the buffer head that's been built up so far
450 * @biovec: the request that could be merged to it.
452 * Return amount of bytes we can accept at this offset
453 * If near_copies == raid_disk, there are no striping issues,
454 * but in that case, the function isn't called at all.
456 static int raid10_mergeable_bvec(request_queue_t *q, struct bio *bio,
457 struct bio_vec *bio_vec)
459 mddev_t *mddev = q->queuedata;
460 sector_t sector = bio->bi_sector + get_start_sect(bio->bi_bdev);
462 unsigned int chunk_sectors = mddev->chunk_size >> 9;
463 unsigned int bio_sectors = bio->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 <= bio_vec->bv_len && bio_sectors == 0)
468 return bio_vec->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 */
549 for (nslot = slot; nslot < conf->copies; nslot++) {
550 int ndisk = r10_bio->devs[nslot].devnum;
553 if ((rdev=rcu_dereference(conf->mirrors[ndisk].rdev)) == NULL ||
554 r10_bio->devs[nslot].bio == IO_BLOCKED ||
555 !test_bit(In_sync, &rdev->flags))
558 /* This optimisation is debatable, and completely destroys
559 * sequential read speed for 'far copies' arrays. So only
560 * keep it for 'near' arrays, and review those later.
562 if (conf->near_copies > 1 && !atomic_read(&rdev->nr_pending)) {
567 new_distance = abs(r10_bio->devs[nslot].addr -
568 conf->mirrors[ndisk].head_position);
569 if (new_distance < current_distance) {
570 current_distance = new_distance;
577 r10_bio->read_slot = slot;
578 /* conf->next_seq_sect = this_sector + sectors;*/
580 if (disk >= 0 && (rdev=rcu_dereference(conf->mirrors[disk].rdev))!= NULL)
581 atomic_inc(&conf->mirrors[disk].rdev->nr_pending);
589 static void unplug_slaves(mddev_t *mddev)
591 conf_t *conf = mddev_to_conf(mddev);
595 for (i=0; i<mddev->raid_disks; i++) {
596 mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[i].rdev);
597 if (rdev && !test_bit(Faulty, &rdev->flags) && atomic_read(&rdev->nr_pending)) {
598 request_queue_t *r_queue = bdev_get_queue(rdev->bdev);
600 atomic_inc(&rdev->nr_pending);
603 if (r_queue->unplug_fn)
604 r_queue->unplug_fn(r_queue);
606 rdev_dec_pending(rdev, mddev);
613 static void raid10_unplug(request_queue_t *q)
615 mddev_t *mddev = q->queuedata;
617 unplug_slaves(q->queuedata);
618 md_wakeup_thread(mddev->thread);
621 static int raid10_issue_flush(request_queue_t *q, struct gendisk *disk,
622 sector_t *error_sector)
624 mddev_t *mddev = q->queuedata;
625 conf_t *conf = mddev_to_conf(mddev);
629 for (i=0; i<mddev->raid_disks && ret == 0; i++) {
630 mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[i].rdev);
631 if (rdev && !test_bit(Faulty, &rdev->flags)) {
632 struct block_device *bdev = rdev->bdev;
633 request_queue_t *r_queue = bdev_get_queue(bdev);
635 if (!r_queue->issue_flush_fn)
638 atomic_inc(&rdev->nr_pending);
640 ret = r_queue->issue_flush_fn(r_queue, bdev->bd_disk,
642 rdev_dec_pending(rdev, mddev);
652 * Sometimes we need to suspend IO while we do something else,
653 * either some resync/recovery, or reconfigure the array.
654 * To do this we raise a 'barrier'.
655 * The 'barrier' is a counter that can be raised multiple times
656 * to count how many activities are happening which preclude
658 * We can only raise the barrier if there is no pending IO.
659 * i.e. if nr_pending == 0.
660 * We choose only to raise the barrier if no-one is waiting for the
661 * barrier to go down. This means that as soon as an IO request
662 * is ready, no other operations which require a barrier will start
663 * until the IO request has had a chance.
665 * So: regular IO calls 'wait_barrier'. When that returns there
666 * is no backgroup IO happening, It must arrange to call
667 * allow_barrier when it has finished its IO.
668 * backgroup IO calls must call raise_barrier. Once that returns
669 * there is no normal IO happeing. It must arrange to call
670 * lower_barrier when the particular background IO completes.
672 #define RESYNC_DEPTH 32
674 static void raise_barrier(conf_t *conf, int force)
676 BUG_ON(force && !conf->barrier);
677 spin_lock_irq(&conf->resync_lock);
679 /* Wait until no block IO is waiting (unless 'force') */
680 wait_event_lock_irq(conf->wait_barrier, force || !conf->nr_waiting,
682 raid10_unplug(conf->mddev->queue));
684 /* block any new IO from starting */
687 /* No wait for all pending IO to complete */
688 wait_event_lock_irq(conf->wait_barrier,
689 !conf->nr_pending && conf->barrier < RESYNC_DEPTH,
691 raid10_unplug(conf->mddev->queue));
693 spin_unlock_irq(&conf->resync_lock);
696 static void lower_barrier(conf_t *conf)
699 spin_lock_irqsave(&conf->resync_lock, flags);
701 spin_unlock_irqrestore(&conf->resync_lock, flags);
702 wake_up(&conf->wait_barrier);
705 static void wait_barrier(conf_t *conf)
707 spin_lock_irq(&conf->resync_lock);
710 wait_event_lock_irq(conf->wait_barrier, !conf->barrier,
712 raid10_unplug(conf->mddev->queue));
716 spin_unlock_irq(&conf->resync_lock);
719 static void allow_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 freeze_array(conf_t *conf)
730 /* stop syncio and normal IO and wait for everything to
732 * We increment barrier and nr_waiting, and then
733 * wait until barrier+nr_pending match nr_queued+2
735 spin_lock_irq(&conf->resync_lock);
738 wait_event_lock_irq(conf->wait_barrier,
739 conf->barrier+conf->nr_pending == conf->nr_queued+2,
741 raid10_unplug(conf->mddev->queue));
742 spin_unlock_irq(&conf->resync_lock);
745 static void unfreeze_array(conf_t *conf)
747 /* reverse the effect of the freeze */
748 spin_lock_irq(&conf->resync_lock);
751 wake_up(&conf->wait_barrier);
752 spin_unlock_irq(&conf->resync_lock);
755 static int make_request(request_queue_t *q, struct bio * bio)
757 mddev_t *mddev = q->queuedata;
758 conf_t *conf = mddev_to_conf(mddev);
759 mirror_info_t *mirror;
761 struct bio *read_bio;
763 int chunk_sects = conf->chunk_mask + 1;
764 const int rw = bio_data_dir(bio);
768 if (unlikely(bio_barrier(bio))) {
769 bio_endio(bio, bio->bi_size, -EOPNOTSUPP);
773 /* If this request crosses a chunk boundary, we need to
774 * split it. This will only happen for 1 PAGE (or less) requests.
776 if (unlikely( (bio->bi_sector & conf->chunk_mask) + (bio->bi_size >> 9)
778 conf->near_copies < conf->raid_disks)) {
780 /* Sanity check -- queue functions should prevent this happening */
781 if (bio->bi_vcnt != 1 ||
784 /* This is a one page bio that upper layers
785 * refuse to split for us, so we need to split it.
787 bp = bio_split(bio, bio_split_pool,
788 chunk_sects - (bio->bi_sector & (chunk_sects - 1)) );
789 if (make_request(q, &bp->bio1))
790 generic_make_request(&bp->bio1);
791 if (make_request(q, &bp->bio2))
792 generic_make_request(&bp->bio2);
794 bio_pair_release(bp);
797 printk("raid10_make_request bug: can't convert block across chunks"
798 " or bigger than %dk %llu %d\n", chunk_sects/2,
799 (unsigned long long)bio->bi_sector, bio->bi_size >> 10);
801 bio_io_error(bio, bio->bi_size);
805 md_write_start(mddev, bio);
808 * Register the new request and wait if the reconstruction
809 * thread has put up a bar for new requests.
810 * Continue immediately if no resync is active currently.
814 disk_stat_inc(mddev->gendisk, ios[rw]);
815 disk_stat_add(mddev->gendisk, sectors[rw], bio_sectors(bio));
817 r10_bio = mempool_alloc(conf->r10bio_pool, GFP_NOIO);
819 r10_bio->master_bio = bio;
820 r10_bio->sectors = bio->bi_size >> 9;
822 r10_bio->mddev = mddev;
823 r10_bio->sector = bio->bi_sector;
828 * read balancing logic:
830 int disk = read_balance(conf, r10_bio);
831 int slot = r10_bio->read_slot;
833 raid_end_bio_io(r10_bio);
836 mirror = conf->mirrors + disk;
838 read_bio = bio_clone(bio, GFP_NOIO);
840 r10_bio->devs[slot].bio = read_bio;
842 read_bio->bi_sector = r10_bio->devs[slot].addr +
843 mirror->rdev->data_offset;
844 read_bio->bi_bdev = mirror->rdev->bdev;
845 read_bio->bi_end_io = raid10_end_read_request;
846 read_bio->bi_rw = READ;
847 read_bio->bi_private = r10_bio;
849 generic_make_request(read_bio);
856 /* first select target devices under spinlock and
857 * inc refcount on their rdev. Record them by setting
860 raid10_find_phys(conf, r10_bio);
862 for (i = 0; i < conf->copies; i++) {
863 int d = r10_bio->devs[i].devnum;
864 mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[d].rdev);
866 !test_bit(Faulty, &rdev->flags)) {
867 atomic_inc(&rdev->nr_pending);
868 r10_bio->devs[i].bio = bio;
870 r10_bio->devs[i].bio = NULL;
871 set_bit(R10BIO_Degraded, &r10_bio->state);
876 atomic_set(&r10_bio->remaining, 0);
879 for (i = 0; i < conf->copies; i++) {
881 int d = r10_bio->devs[i].devnum;
882 if (!r10_bio->devs[i].bio)
885 mbio = bio_clone(bio, GFP_NOIO);
886 r10_bio->devs[i].bio = mbio;
888 mbio->bi_sector = r10_bio->devs[i].addr+
889 conf->mirrors[d].rdev->data_offset;
890 mbio->bi_bdev = conf->mirrors[d].rdev->bdev;
891 mbio->bi_end_io = raid10_end_write_request;
893 mbio->bi_private = r10_bio;
895 atomic_inc(&r10_bio->remaining);
896 bio_list_add(&bl, mbio);
899 bitmap_startwrite(mddev->bitmap, bio->bi_sector, r10_bio->sectors, 0);
900 spin_lock_irqsave(&conf->device_lock, flags);
901 bio_list_merge(&conf->pending_bio_list, &bl);
902 blk_plug_device(mddev->queue);
903 spin_unlock_irqrestore(&conf->device_lock, flags);
908 static void status(struct seq_file *seq, mddev_t *mddev)
910 conf_t *conf = mddev_to_conf(mddev);
913 if (conf->near_copies < conf->raid_disks)
914 seq_printf(seq, " %dK chunks", mddev->chunk_size/1024);
915 if (conf->near_copies > 1)
916 seq_printf(seq, " %d near-copies", conf->near_copies);
917 if (conf->far_copies > 1) {
918 if (conf->far_offset)
919 seq_printf(seq, " %d offset-copies", conf->far_copies);
921 seq_printf(seq, " %d far-copies", conf->far_copies);
923 seq_printf(seq, " [%d/%d] [", conf->raid_disks,
924 conf->raid_disks - mddev->degraded);
925 for (i = 0; i < conf->raid_disks; i++)
926 seq_printf(seq, "%s",
927 conf->mirrors[i].rdev &&
928 test_bit(In_sync, &conf->mirrors[i].rdev->flags) ? "U" : "_");
929 seq_printf(seq, "]");
932 static void error(mddev_t *mddev, mdk_rdev_t *rdev)
934 char b[BDEVNAME_SIZE];
935 conf_t *conf = mddev_to_conf(mddev);
938 * If it is not operational, then we have already marked it as dead
939 * else if it is the last working disks, ignore the error, let the
940 * next level up know.
941 * else mark the drive as failed
943 if (test_bit(In_sync, &rdev->flags)
944 && conf->raid_disks-mddev->degraded == 1)
946 * Don't fail the drive, just return an IO error.
947 * The test should really be more sophisticated than
948 * "working_disks == 1", but it isn't critical, and
949 * can wait until we do more sophisticated "is the drive
950 * really dead" tests...
953 if (test_bit(In_sync, &rdev->flags)) {
956 * if recovery is running, make sure it aborts.
958 set_bit(MD_RECOVERY_ERR, &mddev->recovery);
960 clear_bit(In_sync, &rdev->flags);
961 set_bit(Faulty, &rdev->flags);
962 set_bit(MD_CHANGE_DEVS, &mddev->flags);
963 printk(KERN_ALERT "raid10: Disk failure on %s, disabling device. \n"
964 " Operation continuing on %d devices\n",
965 bdevname(rdev->bdev,b), conf->raid_disks - mddev->degraded);
968 static void print_conf(conf_t *conf)
973 printk("RAID10 conf printout:\n");
978 printk(" --- wd:%d rd:%d\n", conf->raid_disks - conf->mddev->degraded,
981 for (i = 0; i < conf->raid_disks; i++) {
982 char b[BDEVNAME_SIZE];
983 tmp = conf->mirrors + i;
985 printk(" disk %d, wo:%d, o:%d, dev:%s\n",
986 i, !test_bit(In_sync, &tmp->rdev->flags),
987 !test_bit(Faulty, &tmp->rdev->flags),
988 bdevname(tmp->rdev->bdev,b));
992 static void close_sync(conf_t *conf)
997 mempool_destroy(conf->r10buf_pool);
998 conf->r10buf_pool = NULL;
1001 /* check if there are enough drives for
1002 * every block to appear on atleast one
1004 static int enough(conf_t *conf)
1009 int n = conf->copies;
1012 if (conf->mirrors[first].rdev)
1014 first = (first+1) % conf->raid_disks;
1018 } while (first != 0);
1022 static int raid10_spare_active(mddev_t *mddev)
1025 conf_t *conf = mddev->private;
1029 * Find all non-in_sync disks within the RAID10 configuration
1030 * and mark them in_sync
1032 for (i = 0; i < conf->raid_disks; i++) {
1033 tmp = conf->mirrors + i;
1035 && !test_bit(Faulty, &tmp->rdev->flags)
1036 && !test_bit(In_sync, &tmp->rdev->flags)) {
1038 set_bit(In_sync, &tmp->rdev->flags);
1047 static int raid10_add_disk(mddev_t *mddev, mdk_rdev_t *rdev)
1049 conf_t *conf = mddev->private;
1054 if (mddev->recovery_cp < MaxSector)
1055 /* only hot-add to in-sync arrays, as recovery is
1056 * very different from resync
1062 if (rdev->saved_raid_disk >= 0 &&
1063 conf->mirrors[rdev->saved_raid_disk].rdev == NULL)
1064 mirror = rdev->saved_raid_disk;
1067 for ( ; mirror < mddev->raid_disks; mirror++)
1068 if ( !(p=conf->mirrors+mirror)->rdev) {
1070 blk_queue_stack_limits(mddev->queue,
1071 rdev->bdev->bd_disk->queue);
1072 /* as we don't honour merge_bvec_fn, we must never risk
1073 * violating it, so limit ->max_sector to one PAGE, as
1074 * a one page request is never in violation.
1076 if (rdev->bdev->bd_disk->queue->merge_bvec_fn &&
1077 mddev->queue->max_sectors > (PAGE_SIZE>>9))
1078 mddev->queue->max_sectors = (PAGE_SIZE>>9);
1080 p->head_position = 0;
1081 rdev->raid_disk = mirror;
1083 if (rdev->saved_raid_disk != mirror)
1085 rcu_assign_pointer(p->rdev, rdev);
1093 static int raid10_remove_disk(mddev_t *mddev, int number)
1095 conf_t *conf = mddev->private;
1098 mirror_info_t *p = conf->mirrors+ number;
1103 if (test_bit(In_sync, &rdev->flags) ||
1104 atomic_read(&rdev->nr_pending)) {
1110 if (atomic_read(&rdev->nr_pending)) {
1111 /* lost the race, try later */
1123 static int end_sync_read(struct bio *bio, unsigned int bytes_done, int error)
1125 r10bio_t * r10_bio = (r10bio_t *)(bio->bi_private);
1126 conf_t *conf = mddev_to_conf(r10_bio->mddev);
1132 for (i=0; i<conf->copies; i++)
1133 if (r10_bio->devs[i].bio == bio)
1135 BUG_ON(i == conf->copies);
1136 update_head_pos(i, r10_bio);
1137 d = r10_bio->devs[i].devnum;
1139 if (test_bit(BIO_UPTODATE, &bio->bi_flags))
1140 set_bit(R10BIO_Uptodate, &r10_bio->state);
1142 atomic_add(r10_bio->sectors,
1143 &conf->mirrors[d].rdev->corrected_errors);
1144 if (!test_bit(MD_RECOVERY_SYNC, &conf->mddev->recovery))
1145 md_error(r10_bio->mddev,
1146 conf->mirrors[d].rdev);
1149 /* for reconstruct, we always reschedule after a read.
1150 * for resync, only after all reads
1152 if (test_bit(R10BIO_IsRecover, &r10_bio->state) ||
1153 atomic_dec_and_test(&r10_bio->remaining)) {
1154 /* we have read all the blocks,
1155 * do the comparison in process context in raid10d
1157 reschedule_retry(r10_bio);
1159 rdev_dec_pending(conf->mirrors[d].rdev, conf->mddev);
1163 static int end_sync_write(struct bio *bio, unsigned int bytes_done, int error)
1165 int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
1166 r10bio_t * r10_bio = (r10bio_t *)(bio->bi_private);
1167 mddev_t *mddev = r10_bio->mddev;
1168 conf_t *conf = mddev_to_conf(mddev);
1174 for (i = 0; i < conf->copies; i++)
1175 if (r10_bio->devs[i].bio == bio)
1177 d = r10_bio->devs[i].devnum;
1180 md_error(mddev, conf->mirrors[d].rdev);
1181 update_head_pos(i, r10_bio);
1183 while (atomic_dec_and_test(&r10_bio->remaining)) {
1184 if (r10_bio->master_bio == NULL) {
1185 /* the primary of several recovery bios */
1186 md_done_sync(mddev, r10_bio->sectors, 1);
1190 r10bio_t *r10_bio2 = (r10bio_t *)r10_bio->master_bio;
1195 rdev_dec_pending(conf->mirrors[d].rdev, mddev);
1200 * Note: sync and recover and handled very differently for raid10
1201 * This code is for resync.
1202 * For resync, we read through virtual addresses and read all blocks.
1203 * If there is any error, we schedule a write. The lowest numbered
1204 * drive is authoritative.
1205 * However requests come for physical address, so we need to map.
1206 * For every physical address there are raid_disks/copies virtual addresses,
1207 * which is always are least one, but is not necessarly an integer.
1208 * This means that a physical address can span multiple chunks, so we may
1209 * have to submit multiple io requests for a single sync request.
1212 * We check if all blocks are in-sync and only write to blocks that
1215 static void sync_request_write(mddev_t *mddev, r10bio_t *r10_bio)
1217 conf_t *conf = mddev_to_conf(mddev);
1219 struct bio *tbio, *fbio;
1221 atomic_set(&r10_bio->remaining, 1);
1223 /* find the first device with a block */
1224 for (i=0; i<conf->copies; i++)
1225 if (test_bit(BIO_UPTODATE, &r10_bio->devs[i].bio->bi_flags))
1228 if (i == conf->copies)
1232 fbio = r10_bio->devs[i].bio;
1234 /* now find blocks with errors */
1235 for (i=0 ; i < conf->copies ; i++) {
1237 int vcnt = r10_bio->sectors >> (PAGE_SHIFT-9);
1239 tbio = r10_bio->devs[i].bio;
1241 if (tbio->bi_end_io != end_sync_read)
1245 if (test_bit(BIO_UPTODATE, &r10_bio->devs[i].bio->bi_flags)) {
1246 /* We know that the bi_io_vec layout is the same for
1247 * both 'first' and 'i', so we just compare them.
1248 * All vec entries are PAGE_SIZE;
1250 for (j = 0; j < vcnt; j++)
1251 if (memcmp(page_address(fbio->bi_io_vec[j].bv_page),
1252 page_address(tbio->bi_io_vec[j].bv_page),
1257 mddev->resync_mismatches += r10_bio->sectors;
1259 if (test_bit(MD_RECOVERY_CHECK, &mddev->recovery))
1260 /* Don't fix anything. */
1262 /* Ok, we need to write this bio
1263 * First we need to fixup bv_offset, bv_len and
1264 * bi_vecs, as the read request might have corrupted these
1266 tbio->bi_vcnt = vcnt;
1267 tbio->bi_size = r10_bio->sectors << 9;
1269 tbio->bi_phys_segments = 0;
1270 tbio->bi_hw_segments = 0;
1271 tbio->bi_hw_front_size = 0;
1272 tbio->bi_hw_back_size = 0;
1273 tbio->bi_flags &= ~(BIO_POOL_MASK - 1);
1274 tbio->bi_flags |= 1 << BIO_UPTODATE;
1275 tbio->bi_next = NULL;
1276 tbio->bi_rw = WRITE;
1277 tbio->bi_private = r10_bio;
1278 tbio->bi_sector = r10_bio->devs[i].addr;
1280 for (j=0; j < vcnt ; j++) {
1281 tbio->bi_io_vec[j].bv_offset = 0;
1282 tbio->bi_io_vec[j].bv_len = PAGE_SIZE;
1284 memcpy(page_address(tbio->bi_io_vec[j].bv_page),
1285 page_address(fbio->bi_io_vec[j].bv_page),
1288 tbio->bi_end_io = end_sync_write;
1290 d = r10_bio->devs[i].devnum;
1291 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
1292 atomic_inc(&r10_bio->remaining);
1293 md_sync_acct(conf->mirrors[d].rdev->bdev, tbio->bi_size >> 9);
1295 tbio->bi_sector += conf->mirrors[d].rdev->data_offset;
1296 tbio->bi_bdev = conf->mirrors[d].rdev->bdev;
1297 generic_make_request(tbio);
1301 if (atomic_dec_and_test(&r10_bio->remaining)) {
1302 md_done_sync(mddev, r10_bio->sectors, 1);
1308 * Now for the recovery code.
1309 * Recovery happens across physical sectors.
1310 * We recover all non-is_sync drives by finding the virtual address of
1311 * each, and then choose a working drive that also has that virt address.
1312 * There is a separate r10_bio for each non-in_sync drive.
1313 * Only the first two slots are in use. The first for reading,
1314 * The second for writing.
1318 static void recovery_request_write(mddev_t *mddev, r10bio_t *r10_bio)
1320 conf_t *conf = mddev_to_conf(mddev);
1322 struct bio *bio, *wbio;
1325 /* move the pages across to the second bio
1326 * and submit the write request
1328 bio = r10_bio->devs[0].bio;
1329 wbio = r10_bio->devs[1].bio;
1330 for (i=0; i < wbio->bi_vcnt; i++) {
1331 struct page *p = bio->bi_io_vec[i].bv_page;
1332 bio->bi_io_vec[i].bv_page = wbio->bi_io_vec[i].bv_page;
1333 wbio->bi_io_vec[i].bv_page = p;
1335 d = r10_bio->devs[1].devnum;
1337 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
1338 md_sync_acct(conf->mirrors[d].rdev->bdev, wbio->bi_size >> 9);
1339 if (test_bit(R10BIO_Uptodate, &r10_bio->state))
1340 generic_make_request(wbio);
1342 bio_endio(wbio, wbio->bi_size, -EIO);
1347 * This is a kernel thread which:
1349 * 1. Retries failed read operations on working mirrors.
1350 * 2. Updates the raid superblock when problems encounter.
1351 * 3. Performs writes following reads for array synchronising.
1354 static void fix_read_error(conf_t *conf, mddev_t *mddev, r10bio_t *r10_bio)
1356 int sect = 0; /* Offset from r10_bio->sector */
1357 int sectors = r10_bio->sectors;
1361 int sl = r10_bio->read_slot;
1365 if (s > (PAGE_SIZE>>9))
1370 int d = r10_bio->devs[sl].devnum;
1371 rdev = rcu_dereference(conf->mirrors[d].rdev);
1373 test_bit(In_sync, &rdev->flags)) {
1374 atomic_inc(&rdev->nr_pending);
1376 success = sync_page_io(rdev->bdev,
1377 r10_bio->devs[sl].addr +
1378 sect + rdev->data_offset,
1380 conf->tmppage, READ);
1381 rdev_dec_pending(rdev, mddev);
1387 if (sl == conf->copies)
1389 } while (!success && sl != r10_bio->read_slot);
1393 /* Cannot read from anywhere -- bye bye array */
1394 int dn = r10_bio->devs[r10_bio->read_slot].devnum;
1395 md_error(mddev, conf->mirrors[dn].rdev);
1400 /* write it back and re-read */
1402 while (sl != r10_bio->read_slot) {
1407 d = r10_bio->devs[sl].devnum;
1408 rdev = rcu_dereference(conf->mirrors[d].rdev);
1410 test_bit(In_sync, &rdev->flags)) {
1411 atomic_inc(&rdev->nr_pending);
1413 atomic_add(s, &rdev->corrected_errors);
1414 if (sync_page_io(rdev->bdev,
1415 r10_bio->devs[sl].addr +
1416 sect + rdev->data_offset,
1417 s<<9, conf->tmppage, WRITE)
1419 /* Well, this device is dead */
1420 md_error(mddev, rdev);
1421 rdev_dec_pending(rdev, mddev);
1426 while (sl != r10_bio->read_slot) {
1431 d = r10_bio->devs[sl].devnum;
1432 rdev = rcu_dereference(conf->mirrors[d].rdev);
1434 test_bit(In_sync, &rdev->flags)) {
1435 char b[BDEVNAME_SIZE];
1436 atomic_inc(&rdev->nr_pending);
1438 if (sync_page_io(rdev->bdev,
1439 r10_bio->devs[sl].addr +
1440 sect + rdev->data_offset,
1441 s<<9, conf->tmppage, READ) == 0)
1442 /* Well, this device is dead */
1443 md_error(mddev, rdev);
1446 "raid10:%s: read error corrected"
1447 " (%d sectors at %llu on %s)\n",
1449 (unsigned long long)sect+
1451 bdevname(rdev->bdev, b));
1453 rdev_dec_pending(rdev, mddev);
1464 static void raid10d(mddev_t *mddev)
1468 unsigned long flags;
1469 conf_t *conf = mddev_to_conf(mddev);
1470 struct list_head *head = &conf->retry_list;
1474 md_check_recovery(mddev);
1477 char b[BDEVNAME_SIZE];
1478 spin_lock_irqsave(&conf->device_lock, flags);
1480 if (conf->pending_bio_list.head) {
1481 bio = bio_list_get(&conf->pending_bio_list);
1482 blk_remove_plug(mddev->queue);
1483 spin_unlock_irqrestore(&conf->device_lock, flags);
1484 /* flush any pending bitmap writes to disk before proceeding w/ I/O */
1485 if (bitmap_unplug(mddev->bitmap) != 0)
1486 printk("%s: bitmap file write failed!\n", mdname(mddev));
1488 while (bio) { /* submit pending writes */
1489 struct bio *next = bio->bi_next;
1490 bio->bi_next = NULL;
1491 generic_make_request(bio);
1499 if (list_empty(head))
1501 r10_bio = list_entry(head->prev, r10bio_t, retry_list);
1502 list_del(head->prev);
1504 spin_unlock_irqrestore(&conf->device_lock, flags);
1506 mddev = r10_bio->mddev;
1507 conf = mddev_to_conf(mddev);
1508 if (test_bit(R10BIO_IsSync, &r10_bio->state)) {
1509 sync_request_write(mddev, r10_bio);
1511 } else if (test_bit(R10BIO_IsRecover, &r10_bio->state)) {
1512 recovery_request_write(mddev, r10_bio);
1516 /* we got a read error. Maybe the drive is bad. Maybe just
1517 * the block and we can fix it.
1518 * We freeze all other IO, and try reading the block from
1519 * other devices. When we find one, we re-write
1520 * and check it that fixes the read error.
1521 * This is all done synchronously while the array is
1524 if (mddev->ro == 0) {
1526 fix_read_error(conf, mddev, r10_bio);
1527 unfreeze_array(conf);
1530 bio = r10_bio->devs[r10_bio->read_slot].bio;
1531 r10_bio->devs[r10_bio->read_slot].bio =
1532 mddev->ro ? IO_BLOCKED : NULL;
1534 mirror = read_balance(conf, r10_bio);
1536 printk(KERN_ALERT "raid10: %s: unrecoverable I/O"
1537 " read error for block %llu\n",
1538 bdevname(bio->bi_bdev,b),
1539 (unsigned long long)r10_bio->sector);
1540 raid_end_bio_io(r10_bio);
1542 rdev = conf->mirrors[mirror].rdev;
1543 if (printk_ratelimit())
1544 printk(KERN_ERR "raid10: %s: redirecting sector %llu to"
1545 " another mirror\n",
1546 bdevname(rdev->bdev,b),
1547 (unsigned long long)r10_bio->sector);
1548 bio = bio_clone(r10_bio->master_bio, GFP_NOIO);
1549 r10_bio->devs[r10_bio->read_slot].bio = bio;
1550 bio->bi_sector = r10_bio->devs[r10_bio->read_slot].addr
1551 + rdev->data_offset;
1552 bio->bi_bdev = rdev->bdev;
1554 bio->bi_private = r10_bio;
1555 bio->bi_end_io = raid10_end_read_request;
1557 generic_make_request(bio);
1561 spin_unlock_irqrestore(&conf->device_lock, flags);
1563 unplug_slaves(mddev);
1567 static int init_resync(conf_t *conf)
1571 buffs = RESYNC_WINDOW / RESYNC_BLOCK_SIZE;
1572 BUG_ON(conf->r10buf_pool);
1573 conf->r10buf_pool = mempool_create(buffs, r10buf_pool_alloc, r10buf_pool_free, conf);
1574 if (!conf->r10buf_pool)
1576 conf->next_resync = 0;
1581 * perform a "sync" on one "block"
1583 * We need to make sure that no normal I/O request - particularly write
1584 * requests - conflict with active sync requests.
1586 * This is achieved by tracking pending requests and a 'barrier' concept
1587 * that can be installed to exclude normal IO requests.
1589 * Resync and recovery are handled very differently.
1590 * We differentiate by looking at MD_RECOVERY_SYNC in mddev->recovery.
1592 * For resync, we iterate over virtual addresses, read all copies,
1593 * and update if there are differences. If only one copy is live,
1595 * For recovery, we iterate over physical addresses, read a good
1596 * value for each non-in_sync drive, and over-write.
1598 * So, for recovery we may have several outstanding complex requests for a
1599 * given address, one for each out-of-sync device. We model this by allocating
1600 * a number of r10_bio structures, one for each out-of-sync device.
1601 * As we setup these structures, we collect all bio's together into a list
1602 * which we then process collectively to add pages, and then process again
1603 * to pass to generic_make_request.
1605 * The r10_bio structures are linked using a borrowed master_bio pointer.
1606 * This link is counted in ->remaining. When the r10_bio that points to NULL
1607 * has its remaining count decremented to 0, the whole complex operation
1612 static sector_t sync_request(mddev_t *mddev, sector_t sector_nr, int *skipped, int go_faster)
1614 conf_t *conf = mddev_to_conf(mddev);
1616 struct bio *biolist = NULL, *bio;
1617 sector_t max_sector, nr_sectors;
1623 sector_t sectors_skipped = 0;
1624 int chunks_skipped = 0;
1626 if (!conf->r10buf_pool)
1627 if (init_resync(conf))
1631 max_sector = mddev->size << 1;
1632 if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery))
1633 max_sector = mddev->resync_max_sectors;
1634 if (sector_nr >= max_sector) {
1635 /* If we aborted, we need to abort the
1636 * sync on the 'current' bitmap chucks (there can
1637 * be several when recovering multiple devices).
1638 * as we may have started syncing it but not finished.
1639 * We can find the current address in
1640 * mddev->curr_resync, but for recovery,
1641 * we need to convert that to several
1642 * virtual addresses.
1644 if (mddev->curr_resync < max_sector) { /* aborted */
1645 if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery))
1646 bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
1648 else for (i=0; i<conf->raid_disks; i++) {
1650 raid10_find_virt(conf, mddev->curr_resync, i);
1651 bitmap_end_sync(mddev->bitmap, sect,
1654 } else /* completed sync */
1657 bitmap_close_sync(mddev->bitmap);
1660 return sectors_skipped;
1662 if (chunks_skipped >= conf->raid_disks) {
1663 /* if there has been nothing to do on any drive,
1664 * then there is nothing to do at all..
1667 return (max_sector - sector_nr) + sectors_skipped;
1670 /* make sure whole request will fit in a chunk - if chunks
1673 if (conf->near_copies < conf->raid_disks &&
1674 max_sector > (sector_nr | conf->chunk_mask))
1675 max_sector = (sector_nr | conf->chunk_mask) + 1;
1677 * If there is non-resync activity waiting for us then
1678 * put in a delay to throttle resync.
1680 if (!go_faster && conf->nr_waiting)
1681 msleep_interruptible(1000);
1683 /* Again, very different code for resync and recovery.
1684 * Both must result in an r10bio with a list of bios that
1685 * have bi_end_io, bi_sector, bi_bdev set,
1686 * and bi_private set to the r10bio.
1687 * For recovery, we may actually create several r10bios
1688 * with 2 bios in each, that correspond to the bios in the main one.
1689 * In this case, the subordinate r10bios link back through a
1690 * borrowed master_bio pointer, and the counter in the master
1691 * includes a ref from each subordinate.
1693 /* First, we decide what to do and set ->bi_end_io
1694 * To end_sync_read if we want to read, and
1695 * end_sync_write if we will want to write.
1698 max_sync = RESYNC_PAGES << (PAGE_SHIFT-9);
1699 if (!test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
1700 /* recovery... the complicated one */
1704 for (i=0 ; i<conf->raid_disks; i++)
1705 if (conf->mirrors[i].rdev &&
1706 !test_bit(In_sync, &conf->mirrors[i].rdev->flags)) {
1707 int still_degraded = 0;
1708 /* want to reconstruct this device */
1709 r10bio_t *rb2 = r10_bio;
1710 sector_t sect = raid10_find_virt(conf, sector_nr, i);
1712 /* Unless we are doing a full sync, we only need
1713 * to recover the block if it is set in the bitmap
1715 must_sync = bitmap_start_sync(mddev->bitmap, sect,
1717 if (sync_blocks < max_sync)
1718 max_sync = sync_blocks;
1721 /* yep, skip the sync_blocks here, but don't assume
1722 * that there will never be anything to do here
1724 chunks_skipped = -1;
1728 r10_bio = mempool_alloc(conf->r10buf_pool, GFP_NOIO);
1729 raise_barrier(conf, rb2 != NULL);
1730 atomic_set(&r10_bio->remaining, 0);
1732 r10_bio->master_bio = (struct bio*)rb2;
1734 atomic_inc(&rb2->remaining);
1735 r10_bio->mddev = mddev;
1736 set_bit(R10BIO_IsRecover, &r10_bio->state);
1737 r10_bio->sector = sect;
1739 raid10_find_phys(conf, r10_bio);
1740 /* Need to check if this section will still be
1743 for (j=0; j<conf->copies;j++) {
1744 int d = r10_bio->devs[j].devnum;
1745 if (conf->mirrors[d].rdev == NULL ||
1746 test_bit(Faulty, &conf->mirrors[d].rdev->flags)) {
1751 must_sync = bitmap_start_sync(mddev->bitmap, sect,
1752 &sync_blocks, still_degraded);
1754 for (j=0; j<conf->copies;j++) {
1755 int d = r10_bio->devs[j].devnum;
1756 if (conf->mirrors[d].rdev &&
1757 test_bit(In_sync, &conf->mirrors[d].rdev->flags)) {
1758 /* This is where we read from */
1759 bio = r10_bio->devs[0].bio;
1760 bio->bi_next = biolist;
1762 bio->bi_private = r10_bio;
1763 bio->bi_end_io = end_sync_read;
1765 bio->bi_sector = r10_bio->devs[j].addr +
1766 conf->mirrors[d].rdev->data_offset;
1767 bio->bi_bdev = conf->mirrors[d].rdev->bdev;
1768 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
1769 atomic_inc(&r10_bio->remaining);
1770 /* and we write to 'i' */
1772 for (k=0; k<conf->copies; k++)
1773 if (r10_bio->devs[k].devnum == i)
1775 bio = r10_bio->devs[1].bio;
1776 bio->bi_next = biolist;
1778 bio->bi_private = r10_bio;
1779 bio->bi_end_io = end_sync_write;
1781 bio->bi_sector = r10_bio->devs[k].addr +
1782 conf->mirrors[i].rdev->data_offset;
1783 bio->bi_bdev = conf->mirrors[i].rdev->bdev;
1785 r10_bio->devs[0].devnum = d;
1786 r10_bio->devs[1].devnum = i;
1791 if (j == conf->copies) {
1792 /* Cannot recover, so abort the recovery */
1795 if (!test_and_set_bit(MD_RECOVERY_ERR, &mddev->recovery))
1796 printk(KERN_INFO "raid10: %s: insufficient working devices for recovery.\n",
1801 if (biolist == NULL) {
1803 r10bio_t *rb2 = r10_bio;
1804 r10_bio = (r10bio_t*) rb2->master_bio;
1805 rb2->master_bio = NULL;
1811 /* resync. Schedule a read for every block at this virt offset */
1814 if (!bitmap_start_sync(mddev->bitmap, sector_nr,
1815 &sync_blocks, mddev->degraded) &&
1816 !conf->fullsync && !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
1817 /* We can skip this block */
1819 return sync_blocks + sectors_skipped;
1821 if (sync_blocks < max_sync)
1822 max_sync = sync_blocks;
1823 r10_bio = mempool_alloc(conf->r10buf_pool, GFP_NOIO);
1825 r10_bio->mddev = mddev;
1826 atomic_set(&r10_bio->remaining, 0);
1827 raise_barrier(conf, 0);
1828 conf->next_resync = sector_nr;
1830 r10_bio->master_bio = NULL;
1831 r10_bio->sector = sector_nr;
1832 set_bit(R10BIO_IsSync, &r10_bio->state);
1833 raid10_find_phys(conf, r10_bio);
1834 r10_bio->sectors = (sector_nr | conf->chunk_mask) - sector_nr +1;
1836 for (i=0; i<conf->copies; i++) {
1837 int d = r10_bio->devs[i].devnum;
1838 bio = r10_bio->devs[i].bio;
1839 bio->bi_end_io = NULL;
1840 if (conf->mirrors[d].rdev == NULL ||
1841 test_bit(Faulty, &conf->mirrors[d].rdev->flags))
1843 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
1844 atomic_inc(&r10_bio->remaining);
1845 bio->bi_next = biolist;
1847 bio->bi_private = r10_bio;
1848 bio->bi_end_io = end_sync_read;
1850 bio->bi_sector = r10_bio->devs[i].addr +
1851 conf->mirrors[d].rdev->data_offset;
1852 bio->bi_bdev = conf->mirrors[d].rdev->bdev;
1857 for (i=0; i<conf->copies; i++) {
1858 int d = r10_bio->devs[i].devnum;
1859 if (r10_bio->devs[i].bio->bi_end_io)
1860 rdev_dec_pending(conf->mirrors[d].rdev, mddev);
1868 for (bio = biolist; bio ; bio=bio->bi_next) {
1870 bio->bi_flags &= ~(BIO_POOL_MASK - 1);
1872 bio->bi_flags |= 1 << BIO_UPTODATE;
1875 bio->bi_phys_segments = 0;
1876 bio->bi_hw_segments = 0;
1881 if (sector_nr + max_sync < max_sector)
1882 max_sector = sector_nr + max_sync;
1885 int len = PAGE_SIZE;
1887 if (sector_nr + (len>>9) > max_sector)
1888 len = (max_sector - sector_nr) << 9;
1891 for (bio= biolist ; bio ; bio=bio->bi_next) {
1892 page = bio->bi_io_vec[bio->bi_vcnt].bv_page;
1893 if (bio_add_page(bio, page, len, 0) == 0) {
1896 bio->bi_io_vec[bio->bi_vcnt].bv_page = page;
1897 for (bio2 = biolist; bio2 && bio2 != bio; bio2 = bio2->bi_next) {
1898 /* remove last page from this bio */
1900 bio2->bi_size -= len;
1901 bio2->bi_flags &= ~(1<< BIO_SEG_VALID);
1907 nr_sectors += len>>9;
1908 sector_nr += len>>9;
1909 } while (biolist->bi_vcnt < RESYNC_PAGES);
1911 r10_bio->sectors = nr_sectors;
1915 biolist = biolist->bi_next;
1917 bio->bi_next = NULL;
1918 r10_bio = bio->bi_private;
1919 r10_bio->sectors = nr_sectors;
1921 if (bio->bi_end_io == end_sync_read) {
1922 md_sync_acct(bio->bi_bdev, nr_sectors);
1923 generic_make_request(bio);
1927 if (sectors_skipped)
1928 /* pretend they weren't skipped, it makes
1929 * no important difference in this case
1931 md_done_sync(mddev, sectors_skipped, 1);
1933 return sectors_skipped + nr_sectors;
1935 /* There is nowhere to write, so all non-sync
1936 * drives must be failed, so try the next chunk...
1939 sector_t sec = max_sector - sector_nr;
1940 sectors_skipped += sec;
1942 sector_nr = max_sector;
1947 static int run(mddev_t *mddev)
1951 mirror_info_t *disk;
1953 struct list_head *tmp;
1955 sector_t stride, size;
1957 if (mddev->chunk_size == 0) {
1958 printk(KERN_ERR "md/raid10: non-zero chunk size required.\n");
1962 nc = mddev->layout & 255;
1963 fc = (mddev->layout >> 8) & 255;
1964 fo = mddev->layout & (1<<16);
1965 if ((nc*fc) <2 || (nc*fc) > mddev->raid_disks ||
1966 (mddev->layout >> 17)) {
1967 printk(KERN_ERR "raid10: %s: unsupported raid10 layout: 0x%8x\n",
1968 mdname(mddev), mddev->layout);
1972 * copy the already verified devices into our private RAID10
1973 * bookkeeping area. [whatever we allocate in run(),
1974 * should be freed in stop()]
1976 conf = kzalloc(sizeof(conf_t), GFP_KERNEL);
1977 mddev->private = conf;
1979 printk(KERN_ERR "raid10: couldn't allocate memory for %s\n",
1983 conf->mirrors = kzalloc(sizeof(struct mirror_info)*mddev->raid_disks,
1985 if (!conf->mirrors) {
1986 printk(KERN_ERR "raid10: couldn't allocate memory for %s\n",
1991 conf->tmppage = alloc_page(GFP_KERNEL);
1995 conf->near_copies = nc;
1996 conf->far_copies = fc;
1997 conf->copies = nc*fc;
1998 conf->far_offset = fo;
1999 conf->chunk_mask = (sector_t)(mddev->chunk_size>>9)-1;
2000 conf->chunk_shift = ffz(~mddev->chunk_size) - 9;
2002 conf->stride = 1 << conf->chunk_shift;
2004 stride = mddev->size >> (conf->chunk_shift-1);
2005 sector_div(stride, fc);
2006 conf->stride = stride << conf->chunk_shift;
2008 conf->r10bio_pool = mempool_create(NR_RAID10_BIOS, r10bio_pool_alloc,
2009 r10bio_pool_free, conf);
2010 if (!conf->r10bio_pool) {
2011 printk(KERN_ERR "raid10: couldn't allocate memory for %s\n",
2016 ITERATE_RDEV(mddev, rdev, tmp) {
2017 disk_idx = rdev->raid_disk;
2018 if (disk_idx >= mddev->raid_disks
2021 disk = conf->mirrors + disk_idx;
2025 blk_queue_stack_limits(mddev->queue,
2026 rdev->bdev->bd_disk->queue);
2027 /* as we don't honour merge_bvec_fn, we must never risk
2028 * violating it, so limit ->max_sector to one PAGE, as
2029 * a one page request is never in violation.
2031 if (rdev->bdev->bd_disk->queue->merge_bvec_fn &&
2032 mddev->queue->max_sectors > (PAGE_SIZE>>9))
2033 mddev->queue->max_sectors = (PAGE_SIZE>>9);
2035 disk->head_position = 0;
2037 conf->raid_disks = mddev->raid_disks;
2038 conf->mddev = mddev;
2039 spin_lock_init(&conf->device_lock);
2040 INIT_LIST_HEAD(&conf->retry_list);
2042 spin_lock_init(&conf->resync_lock);
2043 init_waitqueue_head(&conf->wait_barrier);
2045 /* need to check that every block has at least one working mirror */
2046 if (!enough(conf)) {
2047 printk(KERN_ERR "raid10: not enough operational mirrors for %s\n",
2052 mddev->degraded = 0;
2053 for (i = 0; i < conf->raid_disks; i++) {
2055 disk = conf->mirrors + i;
2058 !test_bit(In_sync, &rdev->flags)) {
2059 disk->head_position = 0;
2065 mddev->thread = md_register_thread(raid10d, mddev, "%s_raid10");
2066 if (!mddev->thread) {
2068 "raid10: couldn't allocate thread for %s\n",
2074 "raid10: raid set %s active with %d out of %d devices\n",
2075 mdname(mddev), mddev->raid_disks - mddev->degraded,
2078 * Ok, everything is just fine now
2080 if (conf->far_offset) {
2081 size = mddev->size >> (conf->chunk_shift-1);
2082 size *= conf->raid_disks;
2083 size <<= conf->chunk_shift;
2084 sector_div(size, conf->far_copies);
2086 size = conf->stride * conf->raid_disks;
2087 sector_div(size, conf->near_copies);
2088 mddev->array_size = size/2;
2089 mddev->resync_max_sectors = size;
2091 mddev->queue->unplug_fn = raid10_unplug;
2092 mddev->queue->issue_flush_fn = raid10_issue_flush;
2094 /* Calculate max read-ahead size.
2095 * We need to readahead at least twice a whole stripe....
2099 int stripe = conf->raid_disks * (mddev->chunk_size / PAGE_SIZE);
2100 stripe /= conf->near_copies;
2101 if (mddev->queue->backing_dev_info.ra_pages < 2* stripe)
2102 mddev->queue->backing_dev_info.ra_pages = 2* stripe;
2105 if (conf->near_copies < mddev->raid_disks)
2106 blk_queue_merge_bvec(mddev->queue, raid10_mergeable_bvec);
2110 if (conf->r10bio_pool)
2111 mempool_destroy(conf->r10bio_pool);
2112 safe_put_page(conf->tmppage);
2113 kfree(conf->mirrors);
2115 mddev->private = NULL;
2120 static int stop(mddev_t *mddev)
2122 conf_t *conf = mddev_to_conf(mddev);
2124 md_unregister_thread(mddev->thread);
2125 mddev->thread = NULL;
2126 blk_sync_queue(mddev->queue); /* the unplug fn references 'conf'*/
2127 if (conf->r10bio_pool)
2128 mempool_destroy(conf->r10bio_pool);
2129 kfree(conf->mirrors);
2131 mddev->private = NULL;
2135 static void raid10_quiesce(mddev_t *mddev, int state)
2137 conf_t *conf = mddev_to_conf(mddev);
2141 raise_barrier(conf, 0);
2144 lower_barrier(conf);
2147 if (mddev->thread) {
2149 mddev->thread->timeout = mddev->bitmap->daemon_sleep * HZ;
2151 mddev->thread->timeout = MAX_SCHEDULE_TIMEOUT;
2152 md_wakeup_thread(mddev->thread);
2156 static struct mdk_personality raid10_personality =
2160 .owner = THIS_MODULE,
2161 .make_request = make_request,
2165 .error_handler = error,
2166 .hot_add_disk = raid10_add_disk,
2167 .hot_remove_disk= raid10_remove_disk,
2168 .spare_active = raid10_spare_active,
2169 .sync_request = sync_request,
2170 .quiesce = raid10_quiesce,
2173 static int __init raid_init(void)
2175 return register_md_personality(&raid10_personality);
2178 static void raid_exit(void)
2180 unregister_md_personality(&raid10_personality);
2183 module_init(raid_init);
2184 module_exit(raid_exit);
2185 MODULE_LICENSE("GPL");
2186 MODULE_ALIAS("md-personality-9"); /* RAID10 */
2187 MODULE_ALIAS("md-raid10");
2188 MODULE_ALIAS("md-level-10");