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 <linux/delay.h>
22 #include <linux/blkdev.h>
23 #include <linux/seq_file.h>
30 * RAID10 provides a combination of RAID0 and RAID1 functionality.
31 * The layout of data is defined by
34 * near_copies (stored in low byte of layout)
35 * far_copies (stored in second byte of layout)
36 * far_offset (stored in bit 16 of layout )
38 * The data to be stored is divided into chunks using chunksize.
39 * Each device is divided into far_copies sections.
40 * In each section, chunks are laid out in a style similar to raid0, but
41 * near_copies copies of each chunk is stored (each on a different drive).
42 * The starting device for each section is offset near_copies from the starting
43 * device of the previous section.
44 * Thus they are (near_copies*far_copies) of each chunk, and each is on a different
46 * near_copies and far_copies must be at least one, and their product is at most
49 * If far_offset is true, then the far_copies are handled a bit differently.
50 * The copies are still in different stripes, but instead of be very far apart
51 * on disk, there are adjacent stripes.
55 * Number of guaranteed r10bios in case of extreme VM load:
57 #define NR_RAID10_BIOS 256
59 static void unplug_slaves(mddev_t *mddev);
61 static void allow_barrier(conf_t *conf);
62 static void lower_barrier(conf_t *conf);
64 static void * r10bio_pool_alloc(gfp_t gfp_flags, void *data)
68 int size = offsetof(struct r10bio_s, devs[conf->copies]);
70 /* allocate a r10bio with room for raid_disks entries in the bios array */
71 r10_bio = kzalloc(size, gfp_flags);
72 if (!r10_bio && conf->mddev)
73 unplug_slaves(conf->mddev);
78 static void r10bio_pool_free(void *r10_bio, void *data)
83 /* Maximum size of each resync request */
84 #define RESYNC_BLOCK_SIZE (64*1024)
85 #define RESYNC_PAGES ((RESYNC_BLOCK_SIZE + PAGE_SIZE-1) / PAGE_SIZE)
86 /* amount of memory to reserve for resync requests */
87 #define RESYNC_WINDOW (1024*1024)
88 /* maximum number of concurrent requests, memory permitting */
89 #define RESYNC_DEPTH (32*1024*1024/RESYNC_BLOCK_SIZE)
92 * When performing a resync, we need to read and compare, so
93 * we need as many pages are there are copies.
94 * When performing a recovery, we need 2 bios, one for read,
95 * one for write (we recover only one drive per r10buf)
98 static void * r10buf_pool_alloc(gfp_t gfp_flags, void *data)
107 r10_bio = r10bio_pool_alloc(gfp_flags, conf);
109 unplug_slaves(conf->mddev);
113 if (test_bit(MD_RECOVERY_SYNC, &conf->mddev->recovery))
114 nalloc = conf->copies; /* resync */
116 nalloc = 2; /* recovery */
121 for (j = nalloc ; j-- ; ) {
122 bio = bio_alloc(gfp_flags, RESYNC_PAGES);
125 r10_bio->devs[j].bio = bio;
128 * Allocate RESYNC_PAGES data pages and attach them
131 for (j = 0 ; j < nalloc; j++) {
132 bio = r10_bio->devs[j].bio;
133 for (i = 0; i < RESYNC_PAGES; i++) {
134 page = alloc_page(gfp_flags);
138 bio->bi_io_vec[i].bv_page = page;
146 safe_put_page(bio->bi_io_vec[i-1].bv_page);
148 for (i = 0; i < RESYNC_PAGES ; i++)
149 safe_put_page(r10_bio->devs[j].bio->bi_io_vec[i].bv_page);
152 while ( ++j < nalloc )
153 bio_put(r10_bio->devs[j].bio);
154 r10bio_pool_free(r10_bio, conf);
158 static void r10buf_pool_free(void *__r10_bio, void *data)
162 r10bio_t *r10bio = __r10_bio;
165 for (j=0; j < conf->copies; j++) {
166 struct bio *bio = r10bio->devs[j].bio;
168 for (i = 0; i < RESYNC_PAGES; i++) {
169 safe_put_page(bio->bi_io_vec[i].bv_page);
170 bio->bi_io_vec[i].bv_page = NULL;
175 r10bio_pool_free(r10bio, conf);
178 static void put_all_bios(conf_t *conf, r10bio_t *r10_bio)
182 for (i = 0; i < conf->copies; i++) {
183 struct bio **bio = & r10_bio->devs[i].bio;
184 if (*bio && *bio != IO_BLOCKED)
190 static void free_r10bio(r10bio_t *r10_bio)
192 conf_t *conf = r10_bio->mddev->private;
195 * Wake up any possible resync thread that waits for the device
200 put_all_bios(conf, r10_bio);
201 mempool_free(r10_bio, conf->r10bio_pool);
204 static void put_buf(r10bio_t *r10_bio)
206 conf_t *conf = r10_bio->mddev->private;
208 mempool_free(r10_bio, conf->r10buf_pool);
213 static void reschedule_retry(r10bio_t *r10_bio)
216 mddev_t *mddev = r10_bio->mddev;
217 conf_t *conf = mddev->private;
219 spin_lock_irqsave(&conf->device_lock, flags);
220 list_add(&r10_bio->retry_list, &conf->retry_list);
222 spin_unlock_irqrestore(&conf->device_lock, flags);
224 /* wake up frozen array... */
225 wake_up(&conf->wait_barrier);
227 md_wakeup_thread(mddev->thread);
231 * raid_end_bio_io() is called when we have finished servicing a mirrored
232 * operation and are ready to return a success/failure code to the buffer
235 static void raid_end_bio_io(r10bio_t *r10_bio)
237 struct bio *bio = r10_bio->master_bio;
240 test_bit(R10BIO_Uptodate, &r10_bio->state) ? 0 : -EIO);
241 free_r10bio(r10_bio);
245 * Update disk head position estimator based on IRQ completion info.
247 static inline void update_head_pos(int slot, r10bio_t *r10_bio)
249 conf_t *conf = r10_bio->mddev->private;
251 conf->mirrors[r10_bio->devs[slot].devnum].head_position =
252 r10_bio->devs[slot].addr + (r10_bio->sectors);
255 static void raid10_end_read_request(struct bio *bio, int error)
257 int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
258 r10bio_t *r10_bio = bio->bi_private;
260 conf_t *conf = r10_bio->mddev->private;
263 slot = r10_bio->read_slot;
264 dev = r10_bio->devs[slot].devnum;
266 * this branch is our 'one mirror IO has finished' event handler:
268 update_head_pos(slot, r10_bio);
272 * Set R10BIO_Uptodate in our master bio, so that
273 * we will return a good error code to the higher
274 * levels even if IO on some other mirrored buffer fails.
276 * The 'master' represents the composite IO operation to
277 * user-side. So if something waits for IO, then it will
278 * wait for the 'master' bio.
280 set_bit(R10BIO_Uptodate, &r10_bio->state);
281 raid_end_bio_io(r10_bio);
286 char b[BDEVNAME_SIZE];
287 if (printk_ratelimit())
288 printk(KERN_ERR "md/raid10:%s: %s: rescheduling sector %llu\n",
290 bdevname(conf->mirrors[dev].rdev->bdev,b), (unsigned long long)r10_bio->sector);
291 reschedule_retry(r10_bio);
294 rdev_dec_pending(conf->mirrors[dev].rdev, conf->mddev);
297 static void raid10_end_write_request(struct bio *bio, int error)
299 int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
300 r10bio_t *r10_bio = bio->bi_private;
302 conf_t *conf = r10_bio->mddev->private;
304 for (slot = 0; slot < conf->copies; slot++)
305 if (r10_bio->devs[slot].bio == bio)
307 dev = r10_bio->devs[slot].devnum;
310 * this branch is our 'one mirror IO has finished' event handler:
313 md_error(r10_bio->mddev, conf->mirrors[dev].rdev);
314 /* an I/O failed, we can't clear the bitmap */
315 set_bit(R10BIO_Degraded, &r10_bio->state);
318 * Set R10BIO_Uptodate in our master bio, so that
319 * we will return a good error code for to the higher
320 * levels even if IO on some other mirrored buffer fails.
322 * The 'master' represents the composite IO operation to
323 * user-side. So if something waits for IO, then it will
324 * wait for the 'master' bio.
326 set_bit(R10BIO_Uptodate, &r10_bio->state);
328 update_head_pos(slot, r10_bio);
332 * Let's see if all mirrored write operations have finished
335 if (atomic_dec_and_test(&r10_bio->remaining)) {
336 /* clear the bitmap if all writes complete successfully */
337 bitmap_endwrite(r10_bio->mddev->bitmap, r10_bio->sector,
339 !test_bit(R10BIO_Degraded, &r10_bio->state),
341 md_write_end(r10_bio->mddev);
342 raid_end_bio_io(r10_bio);
345 rdev_dec_pending(conf->mirrors[dev].rdev, conf->mddev);
350 * RAID10 layout manager
351 * Aswell as the chunksize and raid_disks count, there are two
352 * parameters: near_copies and far_copies.
353 * near_copies * far_copies must be <= raid_disks.
354 * Normally one of these will be 1.
355 * If both are 1, we get raid0.
356 * If near_copies == raid_disks, we get raid1.
358 * Chunks are layed out in raid0 style with near_copies copies of the
359 * first chunk, followed by near_copies copies of the next chunk and
361 * If far_copies > 1, then after 1/far_copies of the array has been assigned
362 * as described above, we start again with a device offset of near_copies.
363 * So we effectively have another copy of the whole array further down all
364 * the drives, but with blocks on different drives.
365 * With this layout, and block is never stored twice on the one device.
367 * raid10_find_phys finds the sector offset of a given virtual sector
368 * on each device that it is on.
370 * raid10_find_virt does the reverse mapping, from a device and a
371 * sector offset to a virtual address
374 static void raid10_find_phys(conf_t *conf, r10bio_t *r10bio)
384 /* now calculate first sector/dev */
385 chunk = r10bio->sector >> conf->chunk_shift;
386 sector = r10bio->sector & conf->chunk_mask;
388 chunk *= conf->near_copies;
390 dev = sector_div(stripe, conf->raid_disks);
391 if (conf->far_offset)
392 stripe *= conf->far_copies;
394 sector += stripe << conf->chunk_shift;
396 /* and calculate all the others */
397 for (n=0; n < conf->near_copies; n++) {
400 r10bio->devs[slot].addr = sector;
401 r10bio->devs[slot].devnum = d;
404 for (f = 1; f < conf->far_copies; f++) {
405 d += conf->near_copies;
406 if (d >= conf->raid_disks)
407 d -= conf->raid_disks;
409 r10bio->devs[slot].devnum = d;
410 r10bio->devs[slot].addr = s;
414 if (dev >= conf->raid_disks) {
416 sector += (conf->chunk_mask + 1);
419 BUG_ON(slot != conf->copies);
422 static sector_t raid10_find_virt(conf_t *conf, sector_t sector, int dev)
424 sector_t offset, chunk, vchunk;
426 offset = sector & conf->chunk_mask;
427 if (conf->far_offset) {
429 chunk = sector >> conf->chunk_shift;
430 fc = sector_div(chunk, conf->far_copies);
431 dev -= fc * conf->near_copies;
433 dev += conf->raid_disks;
435 while (sector >= conf->stride) {
436 sector -= conf->stride;
437 if (dev < conf->near_copies)
438 dev += conf->raid_disks - conf->near_copies;
440 dev -= conf->near_copies;
442 chunk = sector >> conf->chunk_shift;
444 vchunk = chunk * conf->raid_disks + dev;
445 sector_div(vchunk, conf->near_copies);
446 return (vchunk << conf->chunk_shift) + offset;
450 * raid10_mergeable_bvec -- tell bio layer if a two requests can be merged
452 * @bvm: properties of new bio
453 * @biovec: the request that could be merged to it.
455 * Return amount of bytes we can accept at this offset
456 * If near_copies == raid_disk, there are no striping issues,
457 * but in that case, the function isn't called at all.
459 static int raid10_mergeable_bvec(struct request_queue *q,
460 struct bvec_merge_data *bvm,
461 struct bio_vec *biovec)
463 mddev_t *mddev = q->queuedata;
464 sector_t sector = bvm->bi_sector + get_start_sect(bvm->bi_bdev);
466 unsigned int chunk_sectors = mddev->chunk_sectors;
467 unsigned int bio_sectors = bvm->bi_size >> 9;
469 max = (chunk_sectors - ((sector & (chunk_sectors - 1)) + bio_sectors)) << 9;
470 if (max < 0) max = 0; /* bio_add cannot handle a negative return */
471 if (max <= biovec->bv_len && bio_sectors == 0)
472 return biovec->bv_len;
478 * This routine returns the disk from which the requested read should
479 * be done. There is a per-array 'next expected sequential IO' sector
480 * number - if this matches on the next IO then we use the last disk.
481 * There is also a per-disk 'last know head position' sector that is
482 * maintained from IRQ contexts, both the normal and the resync IO
483 * completion handlers update this position correctly. If there is no
484 * perfect sequential match then we pick the disk whose head is closest.
486 * If there are 2 mirrors in the same 2 devices, performance degrades
487 * because position is mirror, not device based.
489 * The rdev for the device selected will have nr_pending incremented.
493 * FIXME: possibly should rethink readbalancing and do it differently
494 * depending on near_copies / far_copies geometry.
496 static int read_balance(conf_t *conf, r10bio_t *r10_bio)
498 const sector_t this_sector = r10_bio->sector;
499 int disk, slot, nslot;
500 const int sectors = r10_bio->sectors;
501 sector_t new_distance, current_distance;
504 raid10_find_phys(conf, r10_bio);
507 * Check if we can balance. We can balance on the whole
508 * device if no resync is going on (recovery is ok), or below
509 * the resync window. We take the first readable disk when
510 * above the resync window.
512 if (conf->mddev->recovery_cp < MaxSector
513 && (this_sector + sectors >= conf->next_resync)) {
514 /* make sure that disk is operational */
516 disk = r10_bio->devs[slot].devnum;
518 while ((rdev = rcu_dereference(conf->mirrors[disk].rdev)) == NULL ||
519 r10_bio->devs[slot].bio == IO_BLOCKED ||
520 !test_bit(In_sync, &rdev->flags)) {
522 if (slot == conf->copies) {
527 disk = r10_bio->devs[slot].devnum;
533 /* make sure the disk is operational */
535 disk = r10_bio->devs[slot].devnum;
536 while ((rdev=rcu_dereference(conf->mirrors[disk].rdev)) == NULL ||
537 r10_bio->devs[slot].bio == IO_BLOCKED ||
538 !test_bit(In_sync, &rdev->flags)) {
540 if (slot == conf->copies) {
544 disk = r10_bio->devs[slot].devnum;
548 current_distance = abs(r10_bio->devs[slot].addr -
549 conf->mirrors[disk].head_position);
551 /* Find the disk whose head is closest,
552 * or - for far > 1 - find the closest to partition beginning */
554 for (nslot = slot; nslot < conf->copies; nslot++) {
555 int ndisk = r10_bio->devs[nslot].devnum;
558 if ((rdev=rcu_dereference(conf->mirrors[ndisk].rdev)) == NULL ||
559 r10_bio->devs[nslot].bio == IO_BLOCKED ||
560 !test_bit(In_sync, &rdev->flags))
563 /* This optimisation is debatable, and completely destroys
564 * sequential read speed for 'far copies' arrays. So only
565 * keep it for 'near' arrays, and review those later.
567 if (conf->near_copies > 1 && !atomic_read(&rdev->nr_pending)) {
573 /* for far > 1 always use the lowest address */
574 if (conf->far_copies > 1)
575 new_distance = r10_bio->devs[nslot].addr;
577 new_distance = abs(r10_bio->devs[nslot].addr -
578 conf->mirrors[ndisk].head_position);
579 if (new_distance < current_distance) {
580 current_distance = new_distance;
587 r10_bio->read_slot = slot;
588 /* conf->next_seq_sect = this_sector + sectors;*/
590 if (disk >= 0 && (rdev=rcu_dereference(conf->mirrors[disk].rdev))!= NULL)
591 atomic_inc(&conf->mirrors[disk].rdev->nr_pending);
599 static void unplug_slaves(mddev_t *mddev)
601 conf_t *conf = mddev->private;
605 for (i=0; i < conf->raid_disks; i++) {
606 mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[i].rdev);
607 if (rdev && !test_bit(Faulty, &rdev->flags) && atomic_read(&rdev->nr_pending)) {
608 struct request_queue *r_queue = bdev_get_queue(rdev->bdev);
610 atomic_inc(&rdev->nr_pending);
615 rdev_dec_pending(rdev, mddev);
622 static void raid10_unplug(struct request_queue *q)
624 mddev_t *mddev = q->queuedata;
626 unplug_slaves(q->queuedata);
627 md_wakeup_thread(mddev->thread);
630 static int raid10_congested(void *data, int bits)
632 mddev_t *mddev = data;
633 conf_t *conf = mddev->private;
636 if (mddev_congested(mddev, bits))
639 for (i = 0; i < conf->raid_disks && ret == 0; i++) {
640 mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[i].rdev);
641 if (rdev && !test_bit(Faulty, &rdev->flags)) {
642 struct request_queue *q = bdev_get_queue(rdev->bdev);
644 ret |= bdi_congested(&q->backing_dev_info, bits);
651 static int flush_pending_writes(conf_t *conf)
653 /* Any writes that have been queued but are awaiting
654 * bitmap updates get flushed here.
655 * We return 1 if any requests were actually submitted.
659 spin_lock_irq(&conf->device_lock);
661 if (conf->pending_bio_list.head) {
663 bio = bio_list_get(&conf->pending_bio_list);
664 blk_remove_plug(conf->mddev->queue);
665 spin_unlock_irq(&conf->device_lock);
666 /* flush any pending bitmap writes to disk
667 * before proceeding w/ I/O */
668 bitmap_unplug(conf->mddev->bitmap);
670 while (bio) { /* submit pending writes */
671 struct bio *next = bio->bi_next;
673 generic_make_request(bio);
678 spin_unlock_irq(&conf->device_lock);
682 * Sometimes we need to suspend IO while we do something else,
683 * either some resync/recovery, or reconfigure the array.
684 * To do this we raise a 'barrier'.
685 * The 'barrier' is a counter that can be raised multiple times
686 * to count how many activities are happening which preclude
688 * We can only raise the barrier if there is no pending IO.
689 * i.e. if nr_pending == 0.
690 * We choose only to raise the barrier if no-one is waiting for the
691 * barrier to go down. This means that as soon as an IO request
692 * is ready, no other operations which require a barrier will start
693 * until the IO request has had a chance.
695 * So: regular IO calls 'wait_barrier'. When that returns there
696 * is no backgroup IO happening, It must arrange to call
697 * allow_barrier when it has finished its IO.
698 * backgroup IO calls must call raise_barrier. Once that returns
699 * there is no normal IO happeing. It must arrange to call
700 * lower_barrier when the particular background IO completes.
703 static void raise_barrier(conf_t *conf, int force)
705 BUG_ON(force && !conf->barrier);
706 spin_lock_irq(&conf->resync_lock);
708 /* Wait until no block IO is waiting (unless 'force') */
709 wait_event_lock_irq(conf->wait_barrier, force || !conf->nr_waiting,
711 raid10_unplug(conf->mddev->queue));
713 /* block any new IO from starting */
716 /* No wait for all pending IO to complete */
717 wait_event_lock_irq(conf->wait_barrier,
718 !conf->nr_pending && conf->barrier < RESYNC_DEPTH,
720 raid10_unplug(conf->mddev->queue));
722 spin_unlock_irq(&conf->resync_lock);
725 static void lower_barrier(conf_t *conf)
728 spin_lock_irqsave(&conf->resync_lock, flags);
730 spin_unlock_irqrestore(&conf->resync_lock, flags);
731 wake_up(&conf->wait_barrier);
734 static void wait_barrier(conf_t *conf)
736 spin_lock_irq(&conf->resync_lock);
739 wait_event_lock_irq(conf->wait_barrier, !conf->barrier,
741 raid10_unplug(conf->mddev->queue));
745 spin_unlock_irq(&conf->resync_lock);
748 static void allow_barrier(conf_t *conf)
751 spin_lock_irqsave(&conf->resync_lock, flags);
753 spin_unlock_irqrestore(&conf->resync_lock, flags);
754 wake_up(&conf->wait_barrier);
757 static void freeze_array(conf_t *conf)
759 /* stop syncio and normal IO and wait for everything to
761 * We increment barrier and nr_waiting, and then
762 * wait until nr_pending match nr_queued+1
763 * This is called in the context of one normal IO request
764 * that has failed. Thus any sync request that might be pending
765 * will be blocked by nr_pending, and we need to wait for
766 * pending IO requests to complete or be queued for re-try.
767 * Thus the number queued (nr_queued) plus this request (1)
768 * must match the number of pending IOs (nr_pending) before
771 spin_lock_irq(&conf->resync_lock);
774 wait_event_lock_irq(conf->wait_barrier,
775 conf->nr_pending == conf->nr_queued+1,
777 ({ flush_pending_writes(conf);
778 raid10_unplug(conf->mddev->queue); }));
779 spin_unlock_irq(&conf->resync_lock);
782 static void unfreeze_array(conf_t *conf)
784 /* reverse the effect of the freeze */
785 spin_lock_irq(&conf->resync_lock);
788 wake_up(&conf->wait_barrier);
789 spin_unlock_irq(&conf->resync_lock);
792 static int make_request(mddev_t *mddev, struct bio * bio)
794 conf_t *conf = mddev->private;
795 mirror_info_t *mirror;
797 struct bio *read_bio;
799 int chunk_sects = conf->chunk_mask + 1;
800 const int rw = bio_data_dir(bio);
801 const bool do_sync = bio_rw_flagged(bio, BIO_RW_SYNCIO);
804 mdk_rdev_t *blocked_rdev;
806 if (unlikely(bio_rw_flagged(bio, BIO_RW_BARRIER))) {
807 md_barrier_request(mddev, bio);
811 /* If this request crosses a chunk boundary, we need to
812 * split it. This will only happen for 1 PAGE (or less) requests.
814 if (unlikely( (bio->bi_sector & conf->chunk_mask) + (bio->bi_size >> 9)
816 conf->near_copies < conf->raid_disks)) {
818 /* Sanity check -- queue functions should prevent this happening */
819 if (bio->bi_vcnt != 1 ||
822 /* This is a one page bio that upper layers
823 * refuse to split for us, so we need to split it.
826 chunk_sects - (bio->bi_sector & (chunk_sects - 1)) );
827 if (make_request(mddev, &bp->bio1))
828 generic_make_request(&bp->bio1);
829 if (make_request(mddev, &bp->bio2))
830 generic_make_request(&bp->bio2);
832 bio_pair_release(bp);
835 printk("md/raid10:%s: make_request bug: can't convert block across chunks"
836 " or bigger than %dk %llu %d\n", mdname(mddev), chunk_sects/2,
837 (unsigned long long)bio->bi_sector, bio->bi_size >> 10);
843 md_write_start(mddev, bio);
846 * Register the new request and wait if the reconstruction
847 * thread has put up a bar for new requests.
848 * Continue immediately if no resync is active currently.
852 r10_bio = mempool_alloc(conf->r10bio_pool, GFP_NOIO);
854 r10_bio->master_bio = bio;
855 r10_bio->sectors = bio->bi_size >> 9;
857 r10_bio->mddev = mddev;
858 r10_bio->sector = bio->bi_sector;
863 * read balancing logic:
865 int disk = read_balance(conf, r10_bio);
866 int slot = r10_bio->read_slot;
868 raid_end_bio_io(r10_bio);
871 mirror = conf->mirrors + disk;
873 read_bio = bio_clone(bio, GFP_NOIO);
875 r10_bio->devs[slot].bio = read_bio;
877 read_bio->bi_sector = r10_bio->devs[slot].addr +
878 mirror->rdev->data_offset;
879 read_bio->bi_bdev = mirror->rdev->bdev;
880 read_bio->bi_end_io = raid10_end_read_request;
881 read_bio->bi_rw = READ | (do_sync << BIO_RW_SYNCIO);
882 read_bio->bi_private = r10_bio;
884 generic_make_request(read_bio);
891 /* first select target devices under rcu_lock and
892 * inc refcount on their rdev. Record them by setting
895 raid10_find_phys(conf, r10_bio);
899 for (i = 0; i < conf->copies; i++) {
900 int d = r10_bio->devs[i].devnum;
901 mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[d].rdev);
902 if (rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
903 atomic_inc(&rdev->nr_pending);
907 if (rdev && !test_bit(Faulty, &rdev->flags)) {
908 atomic_inc(&rdev->nr_pending);
909 r10_bio->devs[i].bio = bio;
911 r10_bio->devs[i].bio = NULL;
912 set_bit(R10BIO_Degraded, &r10_bio->state);
917 if (unlikely(blocked_rdev)) {
918 /* Have to wait for this device to get unblocked, then retry */
922 for (j = 0; j < i; j++)
923 if (r10_bio->devs[j].bio) {
924 d = r10_bio->devs[j].devnum;
925 rdev_dec_pending(conf->mirrors[d].rdev, mddev);
928 md_wait_for_blocked_rdev(blocked_rdev, mddev);
933 atomic_set(&r10_bio->remaining, 0);
936 for (i = 0; i < conf->copies; i++) {
938 int d = r10_bio->devs[i].devnum;
939 if (!r10_bio->devs[i].bio)
942 mbio = bio_clone(bio, GFP_NOIO);
943 r10_bio->devs[i].bio = mbio;
945 mbio->bi_sector = r10_bio->devs[i].addr+
946 conf->mirrors[d].rdev->data_offset;
947 mbio->bi_bdev = conf->mirrors[d].rdev->bdev;
948 mbio->bi_end_io = raid10_end_write_request;
949 mbio->bi_rw = WRITE | (do_sync << BIO_RW_SYNCIO);
950 mbio->bi_private = r10_bio;
952 atomic_inc(&r10_bio->remaining);
953 bio_list_add(&bl, mbio);
956 if (unlikely(!atomic_read(&r10_bio->remaining))) {
957 /* the array is dead */
959 raid_end_bio_io(r10_bio);
963 bitmap_startwrite(mddev->bitmap, bio->bi_sector, r10_bio->sectors, 0);
964 spin_lock_irqsave(&conf->device_lock, flags);
965 bio_list_merge(&conf->pending_bio_list, &bl);
966 blk_plug_device(mddev->queue);
967 spin_unlock_irqrestore(&conf->device_lock, flags);
969 /* In case raid10d snuck in to freeze_array */
970 wake_up(&conf->wait_barrier);
973 md_wakeup_thread(mddev->thread);
978 static void status(struct seq_file *seq, mddev_t *mddev)
980 conf_t *conf = mddev->private;
983 if (conf->near_copies < conf->raid_disks)
984 seq_printf(seq, " %dK chunks", mddev->chunk_sectors / 2);
985 if (conf->near_copies > 1)
986 seq_printf(seq, " %d near-copies", conf->near_copies);
987 if (conf->far_copies > 1) {
988 if (conf->far_offset)
989 seq_printf(seq, " %d offset-copies", conf->far_copies);
991 seq_printf(seq, " %d far-copies", conf->far_copies);
993 seq_printf(seq, " [%d/%d] [", conf->raid_disks,
994 conf->raid_disks - mddev->degraded);
995 for (i = 0; i < conf->raid_disks; i++)
996 seq_printf(seq, "%s",
997 conf->mirrors[i].rdev &&
998 test_bit(In_sync, &conf->mirrors[i].rdev->flags) ? "U" : "_");
999 seq_printf(seq, "]");
1002 static void error(mddev_t *mddev, mdk_rdev_t *rdev)
1004 char b[BDEVNAME_SIZE];
1005 conf_t *conf = mddev->private;
1008 * If it is not operational, then we have already marked it as dead
1009 * else if it is the last working disks, ignore the error, let the
1010 * next level up know.
1011 * else mark the drive as failed
1013 if (test_bit(In_sync, &rdev->flags)
1014 && conf->raid_disks-mddev->degraded == 1)
1016 * Don't fail the drive, just return an IO error.
1017 * The test should really be more sophisticated than
1018 * "working_disks == 1", but it isn't critical, and
1019 * can wait until we do more sophisticated "is the drive
1020 * really dead" tests...
1023 if (test_and_clear_bit(In_sync, &rdev->flags)) {
1024 unsigned long flags;
1025 spin_lock_irqsave(&conf->device_lock, flags);
1027 spin_unlock_irqrestore(&conf->device_lock, flags);
1029 * if recovery is running, make sure it aborts.
1031 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1033 set_bit(Faulty, &rdev->flags);
1034 set_bit(MD_CHANGE_DEVS, &mddev->flags);
1035 printk(KERN_ALERT "md/raid10:%s: Disk failure on %s, disabling device.\n"
1036 KERN_ALERT "md/raid10:%s: Operation continuing on %d devices.\n",
1037 mdname(mddev), bdevname(rdev->bdev, b),
1038 mdname(mddev), conf->raid_disks - mddev->degraded);
1041 static void print_conf(conf_t *conf)
1046 printk(KERN_DEBUG "RAID10 conf printout:\n");
1048 printk(KERN_DEBUG "(!conf)\n");
1051 printk(KERN_DEBUG " --- wd:%d rd:%d\n", conf->raid_disks - conf->mddev->degraded,
1054 for (i = 0; i < conf->raid_disks; i++) {
1055 char b[BDEVNAME_SIZE];
1056 tmp = conf->mirrors + i;
1058 printk(KERN_DEBUG " disk %d, wo:%d, o:%d, dev:%s\n",
1059 i, !test_bit(In_sync, &tmp->rdev->flags),
1060 !test_bit(Faulty, &tmp->rdev->flags),
1061 bdevname(tmp->rdev->bdev,b));
1065 static void close_sync(conf_t *conf)
1068 allow_barrier(conf);
1070 mempool_destroy(conf->r10buf_pool);
1071 conf->r10buf_pool = NULL;
1074 /* check if there are enough drives for
1075 * every block to appear on atleast one
1077 static int enough(conf_t *conf)
1082 int n = conf->copies;
1085 if (conf->mirrors[first].rdev)
1087 first = (first+1) % conf->raid_disks;
1091 } while (first != 0);
1095 static int raid10_spare_active(mddev_t *mddev)
1098 conf_t *conf = mddev->private;
1102 * Find all non-in_sync disks within the RAID10 configuration
1103 * and mark them in_sync
1105 for (i = 0; i < conf->raid_disks; i++) {
1106 tmp = conf->mirrors + i;
1108 && !test_bit(Faulty, &tmp->rdev->flags)
1109 && !test_and_set_bit(In_sync, &tmp->rdev->flags)) {
1110 unsigned long flags;
1111 spin_lock_irqsave(&conf->device_lock, flags);
1113 spin_unlock_irqrestore(&conf->device_lock, flags);
1122 static int raid10_add_disk(mddev_t *mddev, mdk_rdev_t *rdev)
1124 conf_t *conf = mddev->private;
1129 int last = conf->raid_disks - 1;
1131 if (mddev->recovery_cp < MaxSector)
1132 /* only hot-add to in-sync arrays, as recovery is
1133 * very different from resync
1139 if (rdev->raid_disk >= 0)
1140 first = last = rdev->raid_disk;
1142 if (rdev->saved_raid_disk >= 0 &&
1143 rdev->saved_raid_disk >= first &&
1144 conf->mirrors[rdev->saved_raid_disk].rdev == NULL)
1145 mirror = rdev->saved_raid_disk;
1148 for ( ; mirror <= last ; mirror++)
1149 if ( !(p=conf->mirrors+mirror)->rdev) {
1151 disk_stack_limits(mddev->gendisk, rdev->bdev,
1152 rdev->data_offset << 9);
1153 /* as we don't honour merge_bvec_fn, we must
1154 * never risk violating it, so limit
1155 * ->max_segments to one lying with a single
1156 * page, as a one page request is never in
1159 if (rdev->bdev->bd_disk->queue->merge_bvec_fn) {
1160 blk_queue_max_segments(mddev->queue, 1);
1161 blk_queue_segment_boundary(mddev->queue,
1162 PAGE_CACHE_SIZE - 1);
1165 p->head_position = 0;
1166 rdev->raid_disk = mirror;
1168 if (rdev->saved_raid_disk != mirror)
1170 rcu_assign_pointer(p->rdev, rdev);
1174 md_integrity_add_rdev(rdev, mddev);
1179 static int raid10_remove_disk(mddev_t *mddev, int number)
1181 conf_t *conf = mddev->private;
1184 mirror_info_t *p = conf->mirrors+ number;
1189 if (test_bit(In_sync, &rdev->flags) ||
1190 atomic_read(&rdev->nr_pending)) {
1194 /* Only remove faulty devices in recovery
1197 if (!test_bit(Faulty, &rdev->flags) &&
1204 if (atomic_read(&rdev->nr_pending)) {
1205 /* lost the race, try later */
1210 md_integrity_register(mddev);
1219 static void end_sync_read(struct bio *bio, int error)
1221 r10bio_t *r10_bio = bio->bi_private;
1222 conf_t *conf = r10_bio->mddev->private;
1225 for (i=0; i<conf->copies; i++)
1226 if (r10_bio->devs[i].bio == bio)
1228 BUG_ON(i == conf->copies);
1229 update_head_pos(i, r10_bio);
1230 d = r10_bio->devs[i].devnum;
1232 if (test_bit(BIO_UPTODATE, &bio->bi_flags))
1233 set_bit(R10BIO_Uptodate, &r10_bio->state);
1235 atomic_add(r10_bio->sectors,
1236 &conf->mirrors[d].rdev->corrected_errors);
1237 if (!test_bit(MD_RECOVERY_SYNC, &conf->mddev->recovery))
1238 md_error(r10_bio->mddev,
1239 conf->mirrors[d].rdev);
1242 /* for reconstruct, we always reschedule after a read.
1243 * for resync, only after all reads
1245 rdev_dec_pending(conf->mirrors[d].rdev, conf->mddev);
1246 if (test_bit(R10BIO_IsRecover, &r10_bio->state) ||
1247 atomic_dec_and_test(&r10_bio->remaining)) {
1248 /* we have read all the blocks,
1249 * do the comparison in process context in raid10d
1251 reschedule_retry(r10_bio);
1255 static void end_sync_write(struct bio *bio, int error)
1257 int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
1258 r10bio_t *r10_bio = bio->bi_private;
1259 mddev_t *mddev = r10_bio->mddev;
1260 conf_t *conf = mddev->private;
1263 for (i = 0; i < conf->copies; i++)
1264 if (r10_bio->devs[i].bio == bio)
1266 d = r10_bio->devs[i].devnum;
1269 md_error(mddev, conf->mirrors[d].rdev);
1271 update_head_pos(i, r10_bio);
1273 rdev_dec_pending(conf->mirrors[d].rdev, mddev);
1274 while (atomic_dec_and_test(&r10_bio->remaining)) {
1275 if (r10_bio->master_bio == NULL) {
1276 /* the primary of several recovery bios */
1277 sector_t s = r10_bio->sectors;
1279 md_done_sync(mddev, s, 1);
1282 r10bio_t *r10_bio2 = (r10bio_t *)r10_bio->master_bio;
1290 * Note: sync and recover and handled very differently for raid10
1291 * This code is for resync.
1292 * For resync, we read through virtual addresses and read all blocks.
1293 * If there is any error, we schedule a write. The lowest numbered
1294 * drive is authoritative.
1295 * However requests come for physical address, so we need to map.
1296 * For every physical address there are raid_disks/copies virtual addresses,
1297 * which is always are least one, but is not necessarly an integer.
1298 * This means that a physical address can span multiple chunks, so we may
1299 * have to submit multiple io requests for a single sync request.
1302 * We check if all blocks are in-sync and only write to blocks that
1305 static void sync_request_write(mddev_t *mddev, r10bio_t *r10_bio)
1307 conf_t *conf = mddev->private;
1309 struct bio *tbio, *fbio;
1311 atomic_set(&r10_bio->remaining, 1);
1313 /* find the first device with a block */
1314 for (i=0; i<conf->copies; i++)
1315 if (test_bit(BIO_UPTODATE, &r10_bio->devs[i].bio->bi_flags))
1318 if (i == conf->copies)
1322 fbio = r10_bio->devs[i].bio;
1324 /* now find blocks with errors */
1325 for (i=0 ; i < conf->copies ; i++) {
1327 int vcnt = r10_bio->sectors >> (PAGE_SHIFT-9);
1329 tbio = r10_bio->devs[i].bio;
1331 if (tbio->bi_end_io != end_sync_read)
1335 if (test_bit(BIO_UPTODATE, &r10_bio->devs[i].bio->bi_flags)) {
1336 /* We know that the bi_io_vec layout is the same for
1337 * both 'first' and 'i', so we just compare them.
1338 * All vec entries are PAGE_SIZE;
1340 for (j = 0; j < vcnt; j++)
1341 if (memcmp(page_address(fbio->bi_io_vec[j].bv_page),
1342 page_address(tbio->bi_io_vec[j].bv_page),
1347 mddev->resync_mismatches += r10_bio->sectors;
1349 if (test_bit(MD_RECOVERY_CHECK, &mddev->recovery))
1350 /* Don't fix anything. */
1352 /* Ok, we need to write this bio
1353 * First we need to fixup bv_offset, bv_len and
1354 * bi_vecs, as the read request might have corrupted these
1356 tbio->bi_vcnt = vcnt;
1357 tbio->bi_size = r10_bio->sectors << 9;
1359 tbio->bi_phys_segments = 0;
1360 tbio->bi_flags &= ~(BIO_POOL_MASK - 1);
1361 tbio->bi_flags |= 1 << BIO_UPTODATE;
1362 tbio->bi_next = NULL;
1363 tbio->bi_rw = WRITE;
1364 tbio->bi_private = r10_bio;
1365 tbio->bi_sector = r10_bio->devs[i].addr;
1367 for (j=0; j < vcnt ; j++) {
1368 tbio->bi_io_vec[j].bv_offset = 0;
1369 tbio->bi_io_vec[j].bv_len = PAGE_SIZE;
1371 memcpy(page_address(tbio->bi_io_vec[j].bv_page),
1372 page_address(fbio->bi_io_vec[j].bv_page),
1375 tbio->bi_end_io = end_sync_write;
1377 d = r10_bio->devs[i].devnum;
1378 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
1379 atomic_inc(&r10_bio->remaining);
1380 md_sync_acct(conf->mirrors[d].rdev->bdev, tbio->bi_size >> 9);
1382 tbio->bi_sector += conf->mirrors[d].rdev->data_offset;
1383 tbio->bi_bdev = conf->mirrors[d].rdev->bdev;
1384 generic_make_request(tbio);
1388 if (atomic_dec_and_test(&r10_bio->remaining)) {
1389 md_done_sync(mddev, r10_bio->sectors, 1);
1395 * Now for the recovery code.
1396 * Recovery happens across physical sectors.
1397 * We recover all non-is_sync drives by finding the virtual address of
1398 * each, and then choose a working drive that also has that virt address.
1399 * There is a separate r10_bio for each non-in_sync drive.
1400 * Only the first two slots are in use. The first for reading,
1401 * The second for writing.
1405 static void recovery_request_write(mddev_t *mddev, r10bio_t *r10_bio)
1407 conf_t *conf = mddev->private;
1409 struct bio *bio, *wbio;
1412 /* move the pages across to the second bio
1413 * and submit the write request
1415 bio = r10_bio->devs[0].bio;
1416 wbio = r10_bio->devs[1].bio;
1417 for (i=0; i < wbio->bi_vcnt; i++) {
1418 struct page *p = bio->bi_io_vec[i].bv_page;
1419 bio->bi_io_vec[i].bv_page = wbio->bi_io_vec[i].bv_page;
1420 wbio->bi_io_vec[i].bv_page = p;
1422 d = r10_bio->devs[1].devnum;
1424 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
1425 md_sync_acct(conf->mirrors[d].rdev->bdev, wbio->bi_size >> 9);
1426 if (test_bit(R10BIO_Uptodate, &r10_bio->state))
1427 generic_make_request(wbio);
1429 bio_endio(wbio, -EIO);
1434 * Used by fix_read_error() to decay the per rdev read_errors.
1435 * We halve the read error count for every hour that has elapsed
1436 * since the last recorded read error.
1439 static void check_decay_read_errors(mddev_t *mddev, mdk_rdev_t *rdev)
1441 struct timespec cur_time_mon;
1442 unsigned long hours_since_last;
1443 unsigned int read_errors = atomic_read(&rdev->read_errors);
1445 ktime_get_ts(&cur_time_mon);
1447 if (rdev->last_read_error.tv_sec == 0 &&
1448 rdev->last_read_error.tv_nsec == 0) {
1449 /* first time we've seen a read error */
1450 rdev->last_read_error = cur_time_mon;
1454 hours_since_last = (cur_time_mon.tv_sec -
1455 rdev->last_read_error.tv_sec) / 3600;
1457 rdev->last_read_error = cur_time_mon;
1460 * if hours_since_last is > the number of bits in read_errors
1461 * just set read errors to 0. We do this to avoid
1462 * overflowing the shift of read_errors by hours_since_last.
1464 if (hours_since_last >= 8 * sizeof(read_errors))
1465 atomic_set(&rdev->read_errors, 0);
1467 atomic_set(&rdev->read_errors, read_errors >> hours_since_last);
1471 * This is a kernel thread which:
1473 * 1. Retries failed read operations on working mirrors.
1474 * 2. Updates the raid superblock when problems encounter.
1475 * 3. Performs writes following reads for array synchronising.
1478 static void fix_read_error(conf_t *conf, mddev_t *mddev, r10bio_t *r10_bio)
1480 int sect = 0; /* Offset from r10_bio->sector */
1481 int sectors = r10_bio->sectors;
1483 int max_read_errors = atomic_read(&mddev->max_corr_read_errors);
1487 int d = r10_bio->devs[r10_bio->read_slot].devnum;
1488 char b[BDEVNAME_SIZE];
1489 int cur_read_error_count = 0;
1491 rdev = rcu_dereference(conf->mirrors[d].rdev);
1492 bdevname(rdev->bdev, b);
1494 if (test_bit(Faulty, &rdev->flags)) {
1496 /* drive has already been failed, just ignore any
1497 more fix_read_error() attempts */
1501 check_decay_read_errors(mddev, rdev);
1502 atomic_inc(&rdev->read_errors);
1503 cur_read_error_count = atomic_read(&rdev->read_errors);
1504 if (cur_read_error_count > max_read_errors) {
1507 "md/raid10:%s: %s: Raid device exceeded "
1508 "read_error threshold "
1509 "[cur %d:max %d]\n",
1511 b, cur_read_error_count, max_read_errors);
1513 "md/raid10:%s: %s: Failing raid "
1514 "device\n", mdname(mddev), b);
1515 md_error(mddev, conf->mirrors[d].rdev);
1523 int sl = r10_bio->read_slot;
1527 if (s > (PAGE_SIZE>>9))
1532 int d = r10_bio->devs[sl].devnum;
1533 rdev = rcu_dereference(conf->mirrors[d].rdev);
1535 test_bit(In_sync, &rdev->flags)) {
1536 atomic_inc(&rdev->nr_pending);
1538 success = sync_page_io(rdev->bdev,
1539 r10_bio->devs[sl].addr +
1540 sect + rdev->data_offset,
1542 conf->tmppage, READ);
1543 rdev_dec_pending(rdev, mddev);
1549 if (sl == conf->copies)
1551 } while (!success && sl != r10_bio->read_slot);
1555 /* Cannot read from anywhere -- bye bye array */
1556 int dn = r10_bio->devs[r10_bio->read_slot].devnum;
1557 md_error(mddev, conf->mirrors[dn].rdev);
1562 /* write it back and re-read */
1564 while (sl != r10_bio->read_slot) {
1565 char b[BDEVNAME_SIZE];
1570 d = r10_bio->devs[sl].devnum;
1571 rdev = rcu_dereference(conf->mirrors[d].rdev);
1573 test_bit(In_sync, &rdev->flags)) {
1574 atomic_inc(&rdev->nr_pending);
1576 atomic_add(s, &rdev->corrected_errors);
1577 if (sync_page_io(rdev->bdev,
1578 r10_bio->devs[sl].addr +
1579 sect + rdev->data_offset,
1580 s<<9, conf->tmppage, WRITE)
1582 /* Well, this device is dead */
1584 "md/raid10:%s: read correction "
1586 " (%d sectors at %llu on %s)\n",
1588 (unsigned long long)(sect+
1590 bdevname(rdev->bdev, b));
1591 printk(KERN_NOTICE "md/raid10:%s: %s: failing "
1594 bdevname(rdev->bdev, b));
1595 md_error(mddev, rdev);
1597 rdev_dec_pending(rdev, mddev);
1602 while (sl != r10_bio->read_slot) {
1607 d = r10_bio->devs[sl].devnum;
1608 rdev = rcu_dereference(conf->mirrors[d].rdev);
1610 test_bit(In_sync, &rdev->flags)) {
1611 char b[BDEVNAME_SIZE];
1612 atomic_inc(&rdev->nr_pending);
1614 if (sync_page_io(rdev->bdev,
1615 r10_bio->devs[sl].addr +
1616 sect + rdev->data_offset,
1617 s<<9, conf->tmppage,
1619 /* Well, this device is dead */
1621 "md/raid10:%s: unable to read back "
1623 " (%d sectors at %llu on %s)\n",
1625 (unsigned long long)(sect+
1627 bdevname(rdev->bdev, b));
1628 printk(KERN_NOTICE "md/raid10:%s: %s: failing drive\n",
1630 bdevname(rdev->bdev, b));
1632 md_error(mddev, rdev);
1635 "md/raid10:%s: read error corrected"
1636 " (%d sectors at %llu on %s)\n",
1638 (unsigned long long)(sect+
1640 bdevname(rdev->bdev, b));
1643 rdev_dec_pending(rdev, mddev);
1654 static void raid10d(mddev_t *mddev)
1658 unsigned long flags;
1659 conf_t *conf = mddev->private;
1660 struct list_head *head = &conf->retry_list;
1664 md_check_recovery(mddev);
1667 char b[BDEVNAME_SIZE];
1669 unplug += flush_pending_writes(conf);
1671 spin_lock_irqsave(&conf->device_lock, flags);
1672 if (list_empty(head)) {
1673 spin_unlock_irqrestore(&conf->device_lock, flags);
1676 r10_bio = list_entry(head->prev, r10bio_t, retry_list);
1677 list_del(head->prev);
1679 spin_unlock_irqrestore(&conf->device_lock, flags);
1681 mddev = r10_bio->mddev;
1682 conf = mddev->private;
1683 if (test_bit(R10BIO_IsSync, &r10_bio->state)) {
1684 sync_request_write(mddev, r10_bio);
1686 } else if (test_bit(R10BIO_IsRecover, &r10_bio->state)) {
1687 recovery_request_write(mddev, r10_bio);
1691 /* we got a read error. Maybe the drive is bad. Maybe just
1692 * the block and we can fix it.
1693 * We freeze all other IO, and try reading the block from
1694 * other devices. When we find one, we re-write
1695 * and check it that fixes the read error.
1696 * This is all done synchronously while the array is
1699 if (mddev->ro == 0) {
1701 fix_read_error(conf, mddev, r10_bio);
1702 unfreeze_array(conf);
1705 bio = r10_bio->devs[r10_bio->read_slot].bio;
1706 r10_bio->devs[r10_bio->read_slot].bio =
1707 mddev->ro ? IO_BLOCKED : NULL;
1708 mirror = read_balance(conf, r10_bio);
1710 printk(KERN_ALERT "md/raid10:%s: %s: unrecoverable I/O"
1711 " read error for block %llu\n",
1713 bdevname(bio->bi_bdev,b),
1714 (unsigned long long)r10_bio->sector);
1715 raid_end_bio_io(r10_bio);
1718 const bool do_sync = bio_rw_flagged(r10_bio->master_bio, BIO_RW_SYNCIO);
1720 rdev = conf->mirrors[mirror].rdev;
1721 if (printk_ratelimit())
1722 printk(KERN_ERR "md/raid10:%s: %s: redirecting sector %llu to"
1723 " another mirror\n",
1725 bdevname(rdev->bdev,b),
1726 (unsigned long long)r10_bio->sector);
1727 bio = bio_clone(r10_bio->master_bio, GFP_NOIO);
1728 r10_bio->devs[r10_bio->read_slot].bio = bio;
1729 bio->bi_sector = r10_bio->devs[r10_bio->read_slot].addr
1730 + rdev->data_offset;
1731 bio->bi_bdev = rdev->bdev;
1732 bio->bi_rw = READ | (do_sync << BIO_RW_SYNCIO);
1733 bio->bi_private = r10_bio;
1734 bio->bi_end_io = raid10_end_read_request;
1736 generic_make_request(bio);
1742 unplug_slaves(mddev);
1746 static int init_resync(conf_t *conf)
1750 buffs = RESYNC_WINDOW / RESYNC_BLOCK_SIZE;
1751 BUG_ON(conf->r10buf_pool);
1752 conf->r10buf_pool = mempool_create(buffs, r10buf_pool_alloc, r10buf_pool_free, conf);
1753 if (!conf->r10buf_pool)
1755 conf->next_resync = 0;
1760 * perform a "sync" on one "block"
1762 * We need to make sure that no normal I/O request - particularly write
1763 * requests - conflict with active sync requests.
1765 * This is achieved by tracking pending requests and a 'barrier' concept
1766 * that can be installed to exclude normal IO requests.
1768 * Resync and recovery are handled very differently.
1769 * We differentiate by looking at MD_RECOVERY_SYNC in mddev->recovery.
1771 * For resync, we iterate over virtual addresses, read all copies,
1772 * and update if there are differences. If only one copy is live,
1774 * For recovery, we iterate over physical addresses, read a good
1775 * value for each non-in_sync drive, and over-write.
1777 * So, for recovery we may have several outstanding complex requests for a
1778 * given address, one for each out-of-sync device. We model this by allocating
1779 * a number of r10_bio structures, one for each out-of-sync device.
1780 * As we setup these structures, we collect all bio's together into a list
1781 * which we then process collectively to add pages, and then process again
1782 * to pass to generic_make_request.
1784 * The r10_bio structures are linked using a borrowed master_bio pointer.
1785 * This link is counted in ->remaining. When the r10_bio that points to NULL
1786 * has its remaining count decremented to 0, the whole complex operation
1791 static sector_t sync_request(mddev_t *mddev, sector_t sector_nr, int *skipped, int go_faster)
1793 conf_t *conf = mddev->private;
1795 struct bio *biolist = NULL, *bio;
1796 sector_t max_sector, nr_sectors;
1802 sector_t sectors_skipped = 0;
1803 int chunks_skipped = 0;
1805 if (!conf->r10buf_pool)
1806 if (init_resync(conf))
1810 max_sector = mddev->dev_sectors;
1811 if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery))
1812 max_sector = mddev->resync_max_sectors;
1813 if (sector_nr >= max_sector) {
1814 /* If we aborted, we need to abort the
1815 * sync on the 'current' bitmap chucks (there can
1816 * be several when recovering multiple devices).
1817 * as we may have started syncing it but not finished.
1818 * We can find the current address in
1819 * mddev->curr_resync, but for recovery,
1820 * we need to convert that to several
1821 * virtual addresses.
1823 if (mddev->curr_resync < max_sector) { /* aborted */
1824 if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery))
1825 bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
1827 else for (i=0; i<conf->raid_disks; i++) {
1829 raid10_find_virt(conf, mddev->curr_resync, i);
1830 bitmap_end_sync(mddev->bitmap, sect,
1833 } else /* completed sync */
1836 bitmap_close_sync(mddev->bitmap);
1839 return sectors_skipped;
1841 if (chunks_skipped >= conf->raid_disks) {
1842 /* if there has been nothing to do on any drive,
1843 * then there is nothing to do at all..
1846 return (max_sector - sector_nr) + sectors_skipped;
1849 if (max_sector > mddev->resync_max)
1850 max_sector = mddev->resync_max; /* Don't do IO beyond here */
1852 /* make sure whole request will fit in a chunk - if chunks
1855 if (conf->near_copies < conf->raid_disks &&
1856 max_sector > (sector_nr | conf->chunk_mask))
1857 max_sector = (sector_nr | conf->chunk_mask) + 1;
1859 * If there is non-resync activity waiting for us then
1860 * put in a delay to throttle resync.
1862 if (!go_faster && conf->nr_waiting)
1863 msleep_interruptible(1000);
1865 /* Again, very different code for resync and recovery.
1866 * Both must result in an r10bio with a list of bios that
1867 * have bi_end_io, bi_sector, bi_bdev set,
1868 * and bi_private set to the r10bio.
1869 * For recovery, we may actually create several r10bios
1870 * with 2 bios in each, that correspond to the bios in the main one.
1871 * In this case, the subordinate r10bios link back through a
1872 * borrowed master_bio pointer, and the counter in the master
1873 * includes a ref from each subordinate.
1875 /* First, we decide what to do and set ->bi_end_io
1876 * To end_sync_read if we want to read, and
1877 * end_sync_write if we will want to write.
1880 max_sync = RESYNC_PAGES << (PAGE_SHIFT-9);
1881 if (!test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
1882 /* recovery... the complicated one */
1886 for (i=0 ; i<conf->raid_disks; i++)
1887 if (conf->mirrors[i].rdev &&
1888 !test_bit(In_sync, &conf->mirrors[i].rdev->flags)) {
1889 int still_degraded = 0;
1890 /* want to reconstruct this device */
1891 r10bio_t *rb2 = r10_bio;
1892 sector_t sect = raid10_find_virt(conf, sector_nr, i);
1894 /* Unless we are doing a full sync, we only need
1895 * to recover the block if it is set in the bitmap
1897 must_sync = bitmap_start_sync(mddev->bitmap, sect,
1899 if (sync_blocks < max_sync)
1900 max_sync = sync_blocks;
1903 /* yep, skip the sync_blocks here, but don't assume
1904 * that there will never be anything to do here
1906 chunks_skipped = -1;
1910 r10_bio = mempool_alloc(conf->r10buf_pool, GFP_NOIO);
1911 raise_barrier(conf, rb2 != NULL);
1912 atomic_set(&r10_bio->remaining, 0);
1914 r10_bio->master_bio = (struct bio*)rb2;
1916 atomic_inc(&rb2->remaining);
1917 r10_bio->mddev = mddev;
1918 set_bit(R10BIO_IsRecover, &r10_bio->state);
1919 r10_bio->sector = sect;
1921 raid10_find_phys(conf, r10_bio);
1923 /* Need to check if the array will still be
1926 for (j=0; j<conf->raid_disks; j++)
1927 if (conf->mirrors[j].rdev == NULL ||
1928 test_bit(Faulty, &conf->mirrors[j].rdev->flags)) {
1933 must_sync = bitmap_start_sync(mddev->bitmap, sect,
1934 &sync_blocks, still_degraded);
1936 for (j=0; j<conf->copies;j++) {
1937 int d = r10_bio->devs[j].devnum;
1938 if (conf->mirrors[d].rdev &&
1939 test_bit(In_sync, &conf->mirrors[d].rdev->flags)) {
1940 /* This is where we read from */
1941 bio = r10_bio->devs[0].bio;
1942 bio->bi_next = biolist;
1944 bio->bi_private = r10_bio;
1945 bio->bi_end_io = end_sync_read;
1947 bio->bi_sector = r10_bio->devs[j].addr +
1948 conf->mirrors[d].rdev->data_offset;
1949 bio->bi_bdev = conf->mirrors[d].rdev->bdev;
1950 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
1951 atomic_inc(&r10_bio->remaining);
1952 /* and we write to 'i' */
1954 for (k=0; k<conf->copies; k++)
1955 if (r10_bio->devs[k].devnum == i)
1957 BUG_ON(k == conf->copies);
1958 bio = r10_bio->devs[1].bio;
1959 bio->bi_next = biolist;
1961 bio->bi_private = r10_bio;
1962 bio->bi_end_io = end_sync_write;
1964 bio->bi_sector = r10_bio->devs[k].addr +
1965 conf->mirrors[i].rdev->data_offset;
1966 bio->bi_bdev = conf->mirrors[i].rdev->bdev;
1968 r10_bio->devs[0].devnum = d;
1969 r10_bio->devs[1].devnum = i;
1974 if (j == conf->copies) {
1975 /* Cannot recover, so abort the recovery */
1978 atomic_dec(&rb2->remaining);
1980 if (!test_and_set_bit(MD_RECOVERY_INTR,
1982 printk(KERN_INFO "md/raid10:%s: insufficient "
1983 "working devices for recovery.\n",
1988 if (biolist == NULL) {
1990 r10bio_t *rb2 = r10_bio;
1991 r10_bio = (r10bio_t*) rb2->master_bio;
1992 rb2->master_bio = NULL;
1998 /* resync. Schedule a read for every block at this virt offset */
2001 bitmap_cond_end_sync(mddev->bitmap, sector_nr);
2003 if (!bitmap_start_sync(mddev->bitmap, sector_nr,
2004 &sync_blocks, mddev->degraded) &&
2005 !conf->fullsync && !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
2006 /* We can skip this block */
2008 return sync_blocks + sectors_skipped;
2010 if (sync_blocks < max_sync)
2011 max_sync = sync_blocks;
2012 r10_bio = mempool_alloc(conf->r10buf_pool, GFP_NOIO);
2014 r10_bio->mddev = mddev;
2015 atomic_set(&r10_bio->remaining, 0);
2016 raise_barrier(conf, 0);
2017 conf->next_resync = sector_nr;
2019 r10_bio->master_bio = NULL;
2020 r10_bio->sector = sector_nr;
2021 set_bit(R10BIO_IsSync, &r10_bio->state);
2022 raid10_find_phys(conf, r10_bio);
2023 r10_bio->sectors = (sector_nr | conf->chunk_mask) - sector_nr +1;
2025 for (i=0; i<conf->copies; i++) {
2026 int d = r10_bio->devs[i].devnum;
2027 bio = r10_bio->devs[i].bio;
2028 bio->bi_end_io = NULL;
2029 clear_bit(BIO_UPTODATE, &bio->bi_flags);
2030 if (conf->mirrors[d].rdev == NULL ||
2031 test_bit(Faulty, &conf->mirrors[d].rdev->flags))
2033 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
2034 atomic_inc(&r10_bio->remaining);
2035 bio->bi_next = biolist;
2037 bio->bi_private = r10_bio;
2038 bio->bi_end_io = end_sync_read;
2040 bio->bi_sector = r10_bio->devs[i].addr +
2041 conf->mirrors[d].rdev->data_offset;
2042 bio->bi_bdev = conf->mirrors[d].rdev->bdev;
2047 for (i=0; i<conf->copies; i++) {
2048 int d = r10_bio->devs[i].devnum;
2049 if (r10_bio->devs[i].bio->bi_end_io)
2050 rdev_dec_pending(conf->mirrors[d].rdev, mddev);
2058 for (bio = biolist; bio ; bio=bio->bi_next) {
2060 bio->bi_flags &= ~(BIO_POOL_MASK - 1);
2062 bio->bi_flags |= 1 << BIO_UPTODATE;
2065 bio->bi_phys_segments = 0;
2070 if (sector_nr + max_sync < max_sector)
2071 max_sector = sector_nr + max_sync;
2074 int len = PAGE_SIZE;
2076 if (sector_nr + (len>>9) > max_sector)
2077 len = (max_sector - sector_nr) << 9;
2080 for (bio= biolist ; bio ; bio=bio->bi_next) {
2081 page = bio->bi_io_vec[bio->bi_vcnt].bv_page;
2082 if (bio_add_page(bio, page, len, 0) == 0) {
2085 bio->bi_io_vec[bio->bi_vcnt].bv_page = page;
2086 for (bio2 = biolist; bio2 && bio2 != bio; bio2 = bio2->bi_next) {
2087 /* remove last page from this bio */
2089 bio2->bi_size -= len;
2090 bio2->bi_flags &= ~(1<< BIO_SEG_VALID);
2096 nr_sectors += len>>9;
2097 sector_nr += len>>9;
2098 } while (biolist->bi_vcnt < RESYNC_PAGES);
2100 r10_bio->sectors = nr_sectors;
2104 biolist = biolist->bi_next;
2106 bio->bi_next = NULL;
2107 r10_bio = bio->bi_private;
2108 r10_bio->sectors = nr_sectors;
2110 if (bio->bi_end_io == end_sync_read) {
2111 md_sync_acct(bio->bi_bdev, nr_sectors);
2112 generic_make_request(bio);
2116 if (sectors_skipped)
2117 /* pretend they weren't skipped, it makes
2118 * no important difference in this case
2120 md_done_sync(mddev, sectors_skipped, 1);
2122 return sectors_skipped + nr_sectors;
2124 /* There is nowhere to write, so all non-sync
2125 * drives must be failed, so try the next chunk...
2127 if (sector_nr + max_sync < max_sector)
2128 max_sector = sector_nr + max_sync;
2130 sectors_skipped += (max_sector - sector_nr);
2132 sector_nr = max_sector;
2137 raid10_size(mddev_t *mddev, sector_t sectors, int raid_disks)
2140 conf_t *conf = mddev->private;
2143 raid_disks = conf->raid_disks;
2145 sectors = conf->dev_sectors;
2147 size = sectors >> conf->chunk_shift;
2148 sector_div(size, conf->far_copies);
2149 size = size * raid_disks;
2150 sector_div(size, conf->near_copies);
2152 return size << conf->chunk_shift;
2156 static conf_t *setup_conf(mddev_t *mddev)
2158 conf_t *conf = NULL;
2160 sector_t stride, size;
2163 if (mddev->chunk_sectors < (PAGE_SIZE >> 9) ||
2164 !is_power_of_2(mddev->chunk_sectors)) {
2165 printk(KERN_ERR "md/raid10:%s: chunk size must be "
2166 "at least PAGE_SIZE(%ld) and be a power of 2.\n",
2167 mdname(mddev), PAGE_SIZE);
2171 nc = mddev->layout & 255;
2172 fc = (mddev->layout >> 8) & 255;
2173 fo = mddev->layout & (1<<16);
2175 if ((nc*fc) <2 || (nc*fc) > mddev->raid_disks ||
2176 (mddev->layout >> 17)) {
2177 printk(KERN_ERR "md/raid10:%s: unsupported raid10 layout: 0x%8x\n",
2178 mdname(mddev), mddev->layout);
2183 conf = kzalloc(sizeof(conf_t), GFP_KERNEL);
2187 conf->mirrors = kzalloc(sizeof(struct mirror_info)*mddev->raid_disks,
2192 conf->tmppage = alloc_page(GFP_KERNEL);
2197 conf->raid_disks = mddev->raid_disks;
2198 conf->near_copies = nc;
2199 conf->far_copies = fc;
2200 conf->copies = nc*fc;
2201 conf->far_offset = fo;
2202 conf->chunk_mask = mddev->new_chunk_sectors - 1;
2203 conf->chunk_shift = ffz(~mddev->new_chunk_sectors);
2205 conf->r10bio_pool = mempool_create(NR_RAID10_BIOS, r10bio_pool_alloc,
2206 r10bio_pool_free, conf);
2207 if (!conf->r10bio_pool)
2210 size = mddev->dev_sectors >> conf->chunk_shift;
2211 sector_div(size, fc);
2212 size = size * conf->raid_disks;
2213 sector_div(size, nc);
2214 /* 'size' is now the number of chunks in the array */
2215 /* calculate "used chunks per device" in 'stride' */
2216 stride = size * conf->copies;
2218 /* We need to round up when dividing by raid_disks to
2219 * get the stride size.
2221 stride += conf->raid_disks - 1;
2222 sector_div(stride, conf->raid_disks);
2224 conf->dev_sectors = stride << conf->chunk_shift;
2229 sector_div(stride, fc);
2230 conf->stride = stride << conf->chunk_shift;
2233 spin_lock_init(&conf->device_lock);
2234 INIT_LIST_HEAD(&conf->retry_list);
2236 spin_lock_init(&conf->resync_lock);
2237 init_waitqueue_head(&conf->wait_barrier);
2239 conf->thread = md_register_thread(raid10d, mddev, NULL);
2243 conf->scale_disks = 0;
2244 conf->mddev = mddev;
2248 printk(KERN_ERR "md/raid10:%s: couldn't allocate memory.\n",
2251 if (conf->r10bio_pool)
2252 mempool_destroy(conf->r10bio_pool);
2253 kfree(conf->mirrors);
2254 safe_put_page(conf->tmppage);
2257 return ERR_PTR(err);
2260 static int run(mddev_t *mddev)
2263 int i, disk_idx, chunk_size;
2264 mirror_info_t *disk;
2269 * copy the already verified devices into our private RAID10
2270 * bookkeeping area. [whatever we allocate in run(),
2271 * should be freed in stop()]
2274 if (mddev->private == NULL) {
2275 conf = setup_conf(mddev);
2277 return PTR_ERR(conf);
2278 mddev->private = conf;
2280 conf = mddev->private;
2284 mddev->queue->queue_lock = &conf->device_lock;
2286 mddev->thread = conf->thread;
2287 conf->thread = NULL;
2289 chunk_size = mddev->chunk_sectors << 9;
2290 blk_queue_io_min(mddev->queue, chunk_size);
2291 if (conf->raid_disks % conf->near_copies)
2292 blk_queue_io_opt(mddev->queue, chunk_size * conf->raid_disks);
2294 blk_queue_io_opt(mddev->queue, chunk_size *
2295 (conf->raid_disks / conf->near_copies));
2297 list_for_each_entry(rdev, &mddev->disks, same_set) {
2298 disk_idx = rdev->raid_disk;
2299 if (disk_idx >= conf->raid_disks
2302 if (conf->scale_disks) {
2303 disk_idx *= conf->scale_disks;
2304 rdev->raid_disk = disk_idx;
2305 /* MOVE 'rd%d' link !! */
2307 disk = conf->mirrors + disk_idx;
2310 disk_stack_limits(mddev->gendisk, rdev->bdev,
2311 rdev->data_offset << 9);
2312 /* as we don't honour merge_bvec_fn, we must never risk
2313 * violating it, so limit max_segments to 1 lying
2314 * within a single page.
2316 if (rdev->bdev->bd_disk->queue->merge_bvec_fn) {
2317 blk_queue_max_segments(mddev->queue, 1);
2318 blk_queue_segment_boundary(mddev->queue,
2319 PAGE_CACHE_SIZE - 1);
2322 disk->head_position = 0;
2324 /* need to check that every block has at least one working mirror */
2325 if (!enough(conf)) {
2326 printk(KERN_ERR "md/raid10:%s: not enough operational mirrors.\n",
2331 mddev->degraded = 0;
2332 for (i = 0; i < conf->raid_disks; i++) {
2334 disk = conf->mirrors + i;
2337 !test_bit(In_sync, &disk->rdev->flags)) {
2338 disk->head_position = 0;
2345 if (mddev->recovery_cp != MaxSector)
2346 printk(KERN_NOTICE "md/raid10:%s: not clean"
2347 " -- starting background reconstruction\n",
2350 "md/raid10:%s: active with %d out of %d devices\n",
2351 mdname(mddev), conf->raid_disks - mddev->degraded,
2354 * Ok, everything is just fine now
2356 mddev->dev_sectors = conf->dev_sectors;
2357 size = raid10_size(mddev, 0, 0);
2358 md_set_array_sectors(mddev, size);
2359 mddev->resync_max_sectors = size;
2361 mddev->queue->unplug_fn = raid10_unplug;
2362 mddev->queue->backing_dev_info.congested_fn = raid10_congested;
2363 mddev->queue->backing_dev_info.congested_data = mddev;
2365 /* Calculate max read-ahead size.
2366 * We need to readahead at least twice a whole stripe....
2370 int stripe = conf->raid_disks *
2371 ((mddev->chunk_sectors << 9) / PAGE_SIZE);
2372 stripe /= conf->near_copies;
2373 if (mddev->queue->backing_dev_info.ra_pages < 2* stripe)
2374 mddev->queue->backing_dev_info.ra_pages = 2* stripe;
2377 if (conf->near_copies < conf->raid_disks)
2378 blk_queue_merge_bvec(mddev->queue, raid10_mergeable_bvec);
2379 md_integrity_register(mddev);
2383 if (conf->r10bio_pool)
2384 mempool_destroy(conf->r10bio_pool);
2385 safe_put_page(conf->tmppage);
2386 kfree(conf->mirrors);
2388 mddev->private = NULL;
2389 md_unregister_thread(mddev->thread);
2394 static int stop(mddev_t *mddev)
2396 conf_t *conf = mddev->private;
2398 raise_barrier(conf, 0);
2399 lower_barrier(conf);
2401 md_unregister_thread(mddev->thread);
2402 mddev->thread = NULL;
2403 blk_sync_queue(mddev->queue); /* the unplug fn references 'conf'*/
2404 if (conf->r10bio_pool)
2405 mempool_destroy(conf->r10bio_pool);
2406 kfree(conf->mirrors);
2408 mddev->private = NULL;
2412 static void raid10_quiesce(mddev_t *mddev, int state)
2414 conf_t *conf = mddev->private;
2418 raise_barrier(conf, 0);
2421 lower_barrier(conf);
2426 static void *raid10_takeover_raid0(mddev_t *mddev)
2431 if (mddev->degraded > 0) {
2432 printk(KERN_ERR "md/raid10:%s: Error: degraded raid0!\n",
2434 return ERR_PTR(-EINVAL);
2437 /* Update slot numbers to obtain
2438 * degraded raid10 with missing mirrors
2440 list_for_each_entry(rdev, &mddev->disks, same_set) {
2441 rdev->raid_disk *= 2;
2444 /* Set new parameters */
2445 mddev->new_level = 10;
2446 /* new layout: far_copies = 1, near_copies = 2 */
2447 mddev->new_layout = (1<<8) + 2;
2448 mddev->new_chunk_sectors = mddev->chunk_sectors;
2449 mddev->delta_disks = mddev->raid_disks;
2450 mddev->degraded = mddev->raid_disks;
2451 mddev->raid_disks *= 2;
2452 /* make sure it will be not marked as dirty */
2453 mddev->recovery_cp = MaxSector;
2455 conf = setup_conf(mddev);
2456 conf->scale_disks = 2;
2460 static void *raid10_takeover(mddev_t *mddev)
2462 struct raid0_private_data *raid0_priv;
2464 /* raid10 can take over:
2465 * raid0 - providing it has only two drives
2467 if (mddev->level == 0) {
2468 /* for raid0 takeover only one zone is supported */
2469 raid0_priv = mddev->private;
2470 if (raid0_priv->nr_strip_zones > 1) {
2471 printk(KERN_ERR "md/raid10:%s: cannot takeover raid 0"
2472 " with more than one zone.\n",
2474 return ERR_PTR(-EINVAL);
2476 return raid10_takeover_raid0(mddev);
2478 return ERR_PTR(-EINVAL);
2481 static struct mdk_personality raid10_personality =
2485 .owner = THIS_MODULE,
2486 .make_request = make_request,
2490 .error_handler = error,
2491 .hot_add_disk = raid10_add_disk,
2492 .hot_remove_disk= raid10_remove_disk,
2493 .spare_active = raid10_spare_active,
2494 .sync_request = sync_request,
2495 .quiesce = raid10_quiesce,
2496 .size = raid10_size,
2497 .takeover = raid10_takeover,
2500 static int __init raid_init(void)
2502 return register_md_personality(&raid10_personality);
2505 static void raid_exit(void)
2507 unregister_md_personality(&raid10_personality);
2510 module_init(raid_init);
2511 module_exit(raid_exit);
2512 MODULE_LICENSE("GPL");
2513 MODULE_DESCRIPTION("RAID10 (striped mirror) personality for MD");
2514 MODULE_ALIAS("md-personality-9"); /* RAID10 */
2515 MODULE_ALIAS("md-raid10");
2516 MODULE_ALIAS("md-level-10");