2 * raid5.c : Multiple Devices driver for Linux
3 * Copyright (C) 1996, 1997 Ingo Molnar, Miguel de Icaza, Gadi Oxman
4 * Copyright (C) 1999, 2000 Ingo Molnar
5 * Copyright (C) 2002, 2003 H. Peter Anvin
7 * RAID-4/5/6 management functions.
8 * Thanks to Penguin Computing for making the RAID-6 development possible
9 * by donating a test server!
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.
24 * The sequencing for updating the bitmap reliably is a little
25 * subtle (and I got it wrong the first time) so it deserves some
28 * We group bitmap updates into batches. Each batch has a number.
29 * We may write out several batches at once, but that isn't very important.
30 * conf->seq_write is the number of the last batch successfully written.
31 * conf->seq_flush is the number of the last batch that was closed to
33 * When we discover that we will need to write to any block in a stripe
34 * (in add_stripe_bio) we update the in-memory bitmap and record in sh->bm_seq
35 * the number of the batch it will be in. This is seq_flush+1.
36 * When we are ready to do a write, if that batch hasn't been written yet,
37 * we plug the array and queue the stripe for later.
38 * When an unplug happens, we increment bm_flush, thus closing the current
40 * When we notice that bm_flush > bm_write, we write out all pending updates
41 * to the bitmap, and advance bm_write to where bm_flush was.
42 * This may occasionally write a bit out twice, but is sure never to
46 #include <linux/blkdev.h>
47 #include <linux/kthread.h>
48 #include <linux/raid/pq.h>
49 #include <linux/async_tx.h>
50 #include <linux/module.h>
51 #include <linux/async.h>
52 #include <linux/seq_file.h>
53 #include <linux/cpu.h>
54 #include <linux/slab.h>
55 #include <linux/ratelimit.h>
65 #define NR_STRIPES 256
66 #define STRIPE_SIZE PAGE_SIZE
67 #define STRIPE_SHIFT (PAGE_SHIFT - 9)
68 #define STRIPE_SECTORS (STRIPE_SIZE>>9)
69 #define IO_THRESHOLD 1
70 #define BYPASS_THRESHOLD 1
71 #define NR_HASH (PAGE_SIZE / sizeof(struct hlist_head))
72 #define HASH_MASK (NR_HASH - 1)
74 static inline struct hlist_head *stripe_hash(struct r5conf *conf, sector_t sect)
76 int hash = (sect >> STRIPE_SHIFT) & HASH_MASK;
77 return &conf->stripe_hashtbl[hash];
80 /* bio's attached to a stripe+device for I/O are linked together in bi_sector
81 * order without overlap. There may be several bio's per stripe+device, and
82 * a bio could span several devices.
83 * When walking this list for a particular stripe+device, we must never proceed
84 * beyond a bio that extends past this device, as the next bio might no longer
86 * This function is used to determine the 'next' bio in the list, given the sector
87 * of the current stripe+device
89 static inline struct bio *r5_next_bio(struct bio *bio, sector_t sector)
91 int sectors = bio->bi_size >> 9;
92 if (bio->bi_sector + sectors < sector + STRIPE_SECTORS)
99 * We maintain a biased count of active stripes in the bottom 16 bits of
100 * bi_phys_segments, and a count of processed stripes in the upper 16 bits
102 static inline int raid5_bi_phys_segments(struct bio *bio)
104 return bio->bi_phys_segments & 0xffff;
107 static inline int raid5_bi_hw_segments(struct bio *bio)
109 return (bio->bi_phys_segments >> 16) & 0xffff;
112 static inline int raid5_dec_bi_phys_segments(struct bio *bio)
114 --bio->bi_phys_segments;
115 return raid5_bi_phys_segments(bio);
118 static inline int raid5_dec_bi_hw_segments(struct bio *bio)
120 unsigned short val = raid5_bi_hw_segments(bio);
123 bio->bi_phys_segments = (val << 16) | raid5_bi_phys_segments(bio);
127 static inline void raid5_set_bi_hw_segments(struct bio *bio, unsigned int cnt)
129 bio->bi_phys_segments = raid5_bi_phys_segments(bio) | (cnt << 16);
132 /* Find first data disk in a raid6 stripe */
133 static inline int raid6_d0(struct stripe_head *sh)
136 /* ddf always start from first device */
138 /* md starts just after Q block */
139 if (sh->qd_idx == sh->disks - 1)
142 return sh->qd_idx + 1;
144 static inline int raid6_next_disk(int disk, int raid_disks)
147 return (disk < raid_disks) ? disk : 0;
150 /* When walking through the disks in a raid5, starting at raid6_d0,
151 * We need to map each disk to a 'slot', where the data disks are slot
152 * 0 .. raid_disks-3, the parity disk is raid_disks-2 and the Q disk
153 * is raid_disks-1. This help does that mapping.
155 static int raid6_idx_to_slot(int idx, struct stripe_head *sh,
156 int *count, int syndrome_disks)
162 if (idx == sh->pd_idx)
163 return syndrome_disks;
164 if (idx == sh->qd_idx)
165 return syndrome_disks + 1;
171 static void return_io(struct bio *return_bi)
173 struct bio *bi = return_bi;
176 return_bi = bi->bi_next;
184 static void print_raid5_conf (struct r5conf *conf);
186 static int stripe_operations_active(struct stripe_head *sh)
188 return sh->check_state || sh->reconstruct_state ||
189 test_bit(STRIPE_BIOFILL_RUN, &sh->state) ||
190 test_bit(STRIPE_COMPUTE_RUN, &sh->state);
193 static void __release_stripe(struct r5conf *conf, struct stripe_head *sh)
195 if (atomic_dec_and_test(&sh->count)) {
196 BUG_ON(!list_empty(&sh->lru));
197 BUG_ON(atomic_read(&conf->active_stripes)==0);
198 if (test_bit(STRIPE_HANDLE, &sh->state)) {
199 if (test_bit(STRIPE_DELAYED, &sh->state))
200 list_add_tail(&sh->lru, &conf->delayed_list);
201 else if (test_bit(STRIPE_BIT_DELAY, &sh->state) &&
202 sh->bm_seq - conf->seq_write > 0)
203 list_add_tail(&sh->lru, &conf->bitmap_list);
205 clear_bit(STRIPE_BIT_DELAY, &sh->state);
206 list_add_tail(&sh->lru, &conf->handle_list);
208 md_wakeup_thread(conf->mddev->thread);
210 BUG_ON(stripe_operations_active(sh));
211 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
212 if (atomic_dec_return(&conf->preread_active_stripes)
214 md_wakeup_thread(conf->mddev->thread);
215 atomic_dec(&conf->active_stripes);
216 if (!test_bit(STRIPE_EXPANDING, &sh->state)) {
217 list_add_tail(&sh->lru, &conf->inactive_list);
218 wake_up(&conf->wait_for_stripe);
219 if (conf->retry_read_aligned)
220 md_wakeup_thread(conf->mddev->thread);
226 static void release_stripe(struct stripe_head *sh)
228 struct r5conf *conf = sh->raid_conf;
231 spin_lock_irqsave(&conf->device_lock, flags);
232 __release_stripe(conf, sh);
233 spin_unlock_irqrestore(&conf->device_lock, flags);
236 static inline void remove_hash(struct stripe_head *sh)
238 pr_debug("remove_hash(), stripe %llu\n",
239 (unsigned long long)sh->sector);
241 hlist_del_init(&sh->hash);
244 static inline void insert_hash(struct r5conf *conf, struct stripe_head *sh)
246 struct hlist_head *hp = stripe_hash(conf, sh->sector);
248 pr_debug("insert_hash(), stripe %llu\n",
249 (unsigned long long)sh->sector);
251 hlist_add_head(&sh->hash, hp);
255 /* find an idle stripe, make sure it is unhashed, and return it. */
256 static struct stripe_head *get_free_stripe(struct r5conf *conf)
258 struct stripe_head *sh = NULL;
259 struct list_head *first;
261 if (list_empty(&conf->inactive_list))
263 first = conf->inactive_list.next;
264 sh = list_entry(first, struct stripe_head, lru);
265 list_del_init(first);
267 atomic_inc(&conf->active_stripes);
272 static void shrink_buffers(struct stripe_head *sh)
276 int num = sh->raid_conf->pool_size;
278 for (i = 0; i < num ; i++) {
282 sh->dev[i].page = NULL;
287 static int grow_buffers(struct stripe_head *sh)
290 int num = sh->raid_conf->pool_size;
292 for (i = 0; i < num; i++) {
295 if (!(page = alloc_page(GFP_KERNEL))) {
298 sh->dev[i].page = page;
303 static void raid5_build_block(struct stripe_head *sh, int i, int previous);
304 static void stripe_set_idx(sector_t stripe, struct r5conf *conf, int previous,
305 struct stripe_head *sh);
307 static void init_stripe(struct stripe_head *sh, sector_t sector, int previous)
309 struct r5conf *conf = sh->raid_conf;
312 BUG_ON(atomic_read(&sh->count) != 0);
313 BUG_ON(test_bit(STRIPE_HANDLE, &sh->state));
314 BUG_ON(stripe_operations_active(sh));
316 pr_debug("init_stripe called, stripe %llu\n",
317 (unsigned long long)sh->sector);
321 sh->generation = conf->generation - previous;
322 sh->disks = previous ? conf->previous_raid_disks : conf->raid_disks;
324 stripe_set_idx(sector, conf, previous, sh);
328 for (i = sh->disks; i--; ) {
329 struct r5dev *dev = &sh->dev[i];
331 if (dev->toread || dev->read || dev->towrite || dev->written ||
332 test_bit(R5_LOCKED, &dev->flags)) {
333 printk(KERN_ERR "sector=%llx i=%d %p %p %p %p %d\n",
334 (unsigned long long)sh->sector, i, dev->toread,
335 dev->read, dev->towrite, dev->written,
336 test_bit(R5_LOCKED, &dev->flags));
340 raid5_build_block(sh, i, previous);
342 insert_hash(conf, sh);
345 static struct stripe_head *__find_stripe(struct r5conf *conf, sector_t sector,
348 struct stripe_head *sh;
349 struct hlist_node *hn;
351 pr_debug("__find_stripe, sector %llu\n", (unsigned long long)sector);
352 hlist_for_each_entry(sh, hn, stripe_hash(conf, sector), hash)
353 if (sh->sector == sector && sh->generation == generation)
355 pr_debug("__stripe %llu not in cache\n", (unsigned long long)sector);
360 * Need to check if array has failed when deciding whether to:
362 * - remove non-faulty devices
365 * This determination is simple when no reshape is happening.
366 * However if there is a reshape, we need to carefully check
367 * both the before and after sections.
368 * This is because some failed devices may only affect one
369 * of the two sections, and some non-in_sync devices may
370 * be insync in the section most affected by failed devices.
372 static int calc_degraded(struct r5conf *conf)
374 int degraded, degraded2;
379 for (i = 0; i < conf->previous_raid_disks; i++) {
380 struct md_rdev *rdev = rcu_dereference(conf->disks[i].rdev);
381 if (!rdev || test_bit(Faulty, &rdev->flags))
383 else if (test_bit(In_sync, &rdev->flags))
386 /* not in-sync or faulty.
387 * If the reshape increases the number of devices,
388 * this is being recovered by the reshape, so
389 * this 'previous' section is not in_sync.
390 * If the number of devices is being reduced however,
391 * the device can only be part of the array if
392 * we are reverting a reshape, so this section will
395 if (conf->raid_disks >= conf->previous_raid_disks)
399 if (conf->raid_disks == conf->previous_raid_disks)
403 for (i = 0; i < conf->raid_disks; i++) {
404 struct md_rdev *rdev = rcu_dereference(conf->disks[i].rdev);
405 if (!rdev || test_bit(Faulty, &rdev->flags))
407 else if (test_bit(In_sync, &rdev->flags))
410 /* not in-sync or faulty.
411 * If reshape increases the number of devices, this
412 * section has already been recovered, else it
413 * almost certainly hasn't.
415 if (conf->raid_disks <= conf->previous_raid_disks)
419 if (degraded2 > degraded)
424 static int has_failed(struct r5conf *conf)
428 if (conf->mddev->reshape_position == MaxSector)
429 return conf->mddev->degraded > conf->max_degraded;
431 degraded = calc_degraded(conf);
432 if (degraded > conf->max_degraded)
437 static struct stripe_head *
438 get_active_stripe(struct r5conf *conf, sector_t sector,
439 int previous, int noblock, int noquiesce)
441 struct stripe_head *sh;
443 pr_debug("get_stripe, sector %llu\n", (unsigned long long)sector);
445 spin_lock_irq(&conf->device_lock);
448 wait_event_lock_irq(conf->wait_for_stripe,
449 conf->quiesce == 0 || noquiesce,
450 conf->device_lock, /* nothing */);
451 sh = __find_stripe(conf, sector, conf->generation - previous);
453 if (!conf->inactive_blocked)
454 sh = get_free_stripe(conf);
455 if (noblock && sh == NULL)
458 conf->inactive_blocked = 1;
459 wait_event_lock_irq(conf->wait_for_stripe,
460 !list_empty(&conf->inactive_list) &&
461 (atomic_read(&conf->active_stripes)
462 < (conf->max_nr_stripes *3/4)
463 || !conf->inactive_blocked),
466 conf->inactive_blocked = 0;
468 init_stripe(sh, sector, previous);
470 if (atomic_read(&sh->count)) {
471 BUG_ON(!list_empty(&sh->lru)
472 && !test_bit(STRIPE_EXPANDING, &sh->state));
474 if (!test_bit(STRIPE_HANDLE, &sh->state))
475 atomic_inc(&conf->active_stripes);
476 if (list_empty(&sh->lru) &&
477 !test_bit(STRIPE_EXPANDING, &sh->state))
479 list_del_init(&sh->lru);
482 } while (sh == NULL);
485 atomic_inc(&sh->count);
487 spin_unlock_irq(&conf->device_lock);
491 /* Determine if 'data_offset' or 'new_data_offset' should be used
492 * in this stripe_head.
494 static int use_new_offset(struct r5conf *conf, struct stripe_head *sh)
496 sector_t progress = conf->reshape_progress;
497 /* Need a memory barrier to make sure we see the value
498 * of conf->generation, or ->data_offset that was set before
499 * reshape_progress was updated.
502 if (progress == MaxSector)
504 if (sh->generation == conf->generation - 1)
506 /* We are in a reshape, and this is a new-generation stripe,
507 * so use new_data_offset.
513 raid5_end_read_request(struct bio *bi, int error);
515 raid5_end_write_request(struct bio *bi, int error);
517 static void ops_run_io(struct stripe_head *sh, struct stripe_head_state *s)
519 struct r5conf *conf = sh->raid_conf;
520 int i, disks = sh->disks;
524 for (i = disks; i--; ) {
526 int replace_only = 0;
527 struct bio *bi, *rbi;
528 struct md_rdev *rdev, *rrdev = NULL;
529 if (test_and_clear_bit(R5_Wantwrite, &sh->dev[i].flags)) {
530 if (test_and_clear_bit(R5_WantFUA, &sh->dev[i].flags))
534 } else if (test_and_clear_bit(R5_Wantread, &sh->dev[i].flags))
536 else if (test_and_clear_bit(R5_WantReplace,
537 &sh->dev[i].flags)) {
543 bi = &sh->dev[i].req;
544 rbi = &sh->dev[i].rreq; /* For writing to replacement */
549 bi->bi_end_io = raid5_end_write_request;
550 rbi->bi_end_io = raid5_end_write_request;
552 bi->bi_end_io = raid5_end_read_request;
555 rrdev = rcu_dereference(conf->disks[i].replacement);
556 smp_mb(); /* Ensure that if rrdev is NULL, rdev won't be */
557 rdev = rcu_dereference(conf->disks[i].rdev);
566 /* We raced and saw duplicates */
569 if (test_bit(R5_ReadRepl, &sh->dev[i].flags) && rrdev)
574 if (rdev && test_bit(Faulty, &rdev->flags))
577 atomic_inc(&rdev->nr_pending);
578 if (rrdev && test_bit(Faulty, &rrdev->flags))
581 atomic_inc(&rrdev->nr_pending);
584 /* We have already checked bad blocks for reads. Now
585 * need to check for writes. We never accept write errors
586 * on the replacement, so we don't to check rrdev.
588 while ((rw & WRITE) && rdev &&
589 test_bit(WriteErrorSeen, &rdev->flags)) {
592 int bad = is_badblock(rdev, sh->sector, STRIPE_SECTORS,
593 &first_bad, &bad_sectors);
598 set_bit(BlockedBadBlocks, &rdev->flags);
599 if (!conf->mddev->external &&
600 conf->mddev->flags) {
601 /* It is very unlikely, but we might
602 * still need to write out the
603 * bad block log - better give it
605 md_check_recovery(conf->mddev);
607 md_wait_for_blocked_rdev(rdev, conf->mddev);
609 /* Acknowledged bad block - skip the write */
610 rdev_dec_pending(rdev, conf->mddev);
616 if (s->syncing || s->expanding || s->expanded
618 md_sync_acct(rdev->bdev, STRIPE_SECTORS);
620 set_bit(STRIPE_IO_STARTED, &sh->state);
622 bi->bi_bdev = rdev->bdev;
623 pr_debug("%s: for %llu schedule op %ld on disc %d\n",
624 __func__, (unsigned long long)sh->sector,
626 atomic_inc(&sh->count);
627 if (use_new_offset(conf, sh))
628 bi->bi_sector = (sh->sector
629 + rdev->new_data_offset);
631 bi->bi_sector = (sh->sector
632 + rdev->data_offset);
633 bi->bi_flags = 1 << BIO_UPTODATE;
635 bi->bi_io_vec[0].bv_len = STRIPE_SIZE;
636 bi->bi_io_vec[0].bv_offset = 0;
637 bi->bi_size = STRIPE_SIZE;
640 set_bit(R5_DOUBLE_LOCKED, &sh->dev[i].flags);
641 generic_make_request(bi);
644 if (s->syncing || s->expanding || s->expanded
646 md_sync_acct(rrdev->bdev, STRIPE_SECTORS);
648 set_bit(STRIPE_IO_STARTED, &sh->state);
650 rbi->bi_bdev = rrdev->bdev;
651 pr_debug("%s: for %llu schedule op %ld on "
652 "replacement disc %d\n",
653 __func__, (unsigned long long)sh->sector,
655 atomic_inc(&sh->count);
656 if (use_new_offset(conf, sh))
657 rbi->bi_sector = (sh->sector
658 + rrdev->new_data_offset);
660 rbi->bi_sector = (sh->sector
661 + rrdev->data_offset);
662 rbi->bi_flags = 1 << BIO_UPTODATE;
664 rbi->bi_io_vec[0].bv_len = STRIPE_SIZE;
665 rbi->bi_io_vec[0].bv_offset = 0;
666 rbi->bi_size = STRIPE_SIZE;
668 generic_make_request(rbi);
670 if (!rdev && !rrdev) {
672 set_bit(STRIPE_DEGRADED, &sh->state);
673 pr_debug("skip op %ld on disc %d for sector %llu\n",
674 bi->bi_rw, i, (unsigned long long)sh->sector);
675 clear_bit(R5_LOCKED, &sh->dev[i].flags);
676 set_bit(STRIPE_HANDLE, &sh->state);
681 static struct dma_async_tx_descriptor *
682 async_copy_data(int frombio, struct bio *bio, struct page *page,
683 sector_t sector, struct dma_async_tx_descriptor *tx)
686 struct page *bio_page;
689 struct async_submit_ctl submit;
690 enum async_tx_flags flags = 0;
692 if (bio->bi_sector >= sector)
693 page_offset = (signed)(bio->bi_sector - sector) * 512;
695 page_offset = (signed)(sector - bio->bi_sector) * -512;
698 flags |= ASYNC_TX_FENCE;
699 init_async_submit(&submit, flags, tx, NULL, NULL, NULL);
701 bio_for_each_segment(bvl, bio, i) {
702 int len = bvl->bv_len;
706 if (page_offset < 0) {
707 b_offset = -page_offset;
708 page_offset += b_offset;
712 if (len > 0 && page_offset + len > STRIPE_SIZE)
713 clen = STRIPE_SIZE - page_offset;
718 b_offset += bvl->bv_offset;
719 bio_page = bvl->bv_page;
721 tx = async_memcpy(page, bio_page, page_offset,
722 b_offset, clen, &submit);
724 tx = async_memcpy(bio_page, page, b_offset,
725 page_offset, clen, &submit);
727 /* chain the operations */
728 submit.depend_tx = tx;
730 if (clen < len) /* hit end of page */
738 static void ops_complete_biofill(void *stripe_head_ref)
740 struct stripe_head *sh = stripe_head_ref;
741 struct bio *return_bi = NULL;
742 struct r5conf *conf = sh->raid_conf;
745 pr_debug("%s: stripe %llu\n", __func__,
746 (unsigned long long)sh->sector);
748 /* clear completed biofills */
749 spin_lock_irq(&conf->device_lock);
750 for (i = sh->disks; i--; ) {
751 struct r5dev *dev = &sh->dev[i];
753 /* acknowledge completion of a biofill operation */
754 /* and check if we need to reply to a read request,
755 * new R5_Wantfill requests are held off until
756 * !STRIPE_BIOFILL_RUN
758 if (test_and_clear_bit(R5_Wantfill, &dev->flags)) {
759 struct bio *rbi, *rbi2;
764 while (rbi && rbi->bi_sector <
765 dev->sector + STRIPE_SECTORS) {
766 rbi2 = r5_next_bio(rbi, dev->sector);
767 if (!raid5_dec_bi_phys_segments(rbi)) {
768 rbi->bi_next = return_bi;
775 spin_unlock_irq(&conf->device_lock);
776 clear_bit(STRIPE_BIOFILL_RUN, &sh->state);
778 return_io(return_bi);
780 set_bit(STRIPE_HANDLE, &sh->state);
784 static void ops_run_biofill(struct stripe_head *sh)
786 struct dma_async_tx_descriptor *tx = NULL;
787 struct r5conf *conf = sh->raid_conf;
788 struct async_submit_ctl submit;
791 pr_debug("%s: stripe %llu\n", __func__,
792 (unsigned long long)sh->sector);
794 for (i = sh->disks; i--; ) {
795 struct r5dev *dev = &sh->dev[i];
796 if (test_bit(R5_Wantfill, &dev->flags)) {
798 spin_lock_irq(&conf->device_lock);
799 dev->read = rbi = dev->toread;
801 spin_unlock_irq(&conf->device_lock);
802 while (rbi && rbi->bi_sector <
803 dev->sector + STRIPE_SECTORS) {
804 tx = async_copy_data(0, rbi, dev->page,
806 rbi = r5_next_bio(rbi, dev->sector);
811 atomic_inc(&sh->count);
812 init_async_submit(&submit, ASYNC_TX_ACK, tx, ops_complete_biofill, sh, NULL);
813 async_trigger_callback(&submit);
816 static void mark_target_uptodate(struct stripe_head *sh, int target)
823 tgt = &sh->dev[target];
824 set_bit(R5_UPTODATE, &tgt->flags);
825 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
826 clear_bit(R5_Wantcompute, &tgt->flags);
829 static void ops_complete_compute(void *stripe_head_ref)
831 struct stripe_head *sh = stripe_head_ref;
833 pr_debug("%s: stripe %llu\n", __func__,
834 (unsigned long long)sh->sector);
836 /* mark the computed target(s) as uptodate */
837 mark_target_uptodate(sh, sh->ops.target);
838 mark_target_uptodate(sh, sh->ops.target2);
840 clear_bit(STRIPE_COMPUTE_RUN, &sh->state);
841 if (sh->check_state == check_state_compute_run)
842 sh->check_state = check_state_compute_result;
843 set_bit(STRIPE_HANDLE, &sh->state);
847 /* return a pointer to the address conversion region of the scribble buffer */
848 static addr_conv_t *to_addr_conv(struct stripe_head *sh,
849 struct raid5_percpu *percpu)
851 return percpu->scribble + sizeof(struct page *) * (sh->disks + 2);
854 static struct dma_async_tx_descriptor *
855 ops_run_compute5(struct stripe_head *sh, struct raid5_percpu *percpu)
857 int disks = sh->disks;
858 struct page **xor_srcs = percpu->scribble;
859 int target = sh->ops.target;
860 struct r5dev *tgt = &sh->dev[target];
861 struct page *xor_dest = tgt->page;
863 struct dma_async_tx_descriptor *tx;
864 struct async_submit_ctl submit;
867 pr_debug("%s: stripe %llu block: %d\n",
868 __func__, (unsigned long long)sh->sector, target);
869 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
871 for (i = disks; i--; )
873 xor_srcs[count++] = sh->dev[i].page;
875 atomic_inc(&sh->count);
877 init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST, NULL,
878 ops_complete_compute, sh, to_addr_conv(sh, percpu));
879 if (unlikely(count == 1))
880 tx = async_memcpy(xor_dest, xor_srcs[0], 0, 0, STRIPE_SIZE, &submit);
882 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
887 /* set_syndrome_sources - populate source buffers for gen_syndrome
888 * @srcs - (struct page *) array of size sh->disks
889 * @sh - stripe_head to parse
891 * Populates srcs in proper layout order for the stripe and returns the
892 * 'count' of sources to be used in a call to async_gen_syndrome. The P
893 * destination buffer is recorded in srcs[count] and the Q destination
894 * is recorded in srcs[count+1]].
896 static int set_syndrome_sources(struct page **srcs, struct stripe_head *sh)
898 int disks = sh->disks;
899 int syndrome_disks = sh->ddf_layout ? disks : (disks - 2);
900 int d0_idx = raid6_d0(sh);
904 for (i = 0; i < disks; i++)
910 int slot = raid6_idx_to_slot(i, sh, &count, syndrome_disks);
912 srcs[slot] = sh->dev[i].page;
913 i = raid6_next_disk(i, disks);
914 } while (i != d0_idx);
916 return syndrome_disks;
919 static struct dma_async_tx_descriptor *
920 ops_run_compute6_1(struct stripe_head *sh, struct raid5_percpu *percpu)
922 int disks = sh->disks;
923 struct page **blocks = percpu->scribble;
925 int qd_idx = sh->qd_idx;
926 struct dma_async_tx_descriptor *tx;
927 struct async_submit_ctl submit;
933 if (sh->ops.target < 0)
934 target = sh->ops.target2;
935 else if (sh->ops.target2 < 0)
936 target = sh->ops.target;
938 /* we should only have one valid target */
941 pr_debug("%s: stripe %llu block: %d\n",
942 __func__, (unsigned long long)sh->sector, target);
944 tgt = &sh->dev[target];
945 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
948 atomic_inc(&sh->count);
950 if (target == qd_idx) {
951 count = set_syndrome_sources(blocks, sh);
952 blocks[count] = NULL; /* regenerating p is not necessary */
953 BUG_ON(blocks[count+1] != dest); /* q should already be set */
954 init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
955 ops_complete_compute, sh,
956 to_addr_conv(sh, percpu));
957 tx = async_gen_syndrome(blocks, 0, count+2, STRIPE_SIZE, &submit);
959 /* Compute any data- or p-drive using XOR */
961 for (i = disks; i-- ; ) {
962 if (i == target || i == qd_idx)
964 blocks[count++] = sh->dev[i].page;
967 init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST,
968 NULL, ops_complete_compute, sh,
969 to_addr_conv(sh, percpu));
970 tx = async_xor(dest, blocks, 0, count, STRIPE_SIZE, &submit);
976 static struct dma_async_tx_descriptor *
977 ops_run_compute6_2(struct stripe_head *sh, struct raid5_percpu *percpu)
979 int i, count, disks = sh->disks;
980 int syndrome_disks = sh->ddf_layout ? disks : disks-2;
981 int d0_idx = raid6_d0(sh);
982 int faila = -1, failb = -1;
983 int target = sh->ops.target;
984 int target2 = sh->ops.target2;
985 struct r5dev *tgt = &sh->dev[target];
986 struct r5dev *tgt2 = &sh->dev[target2];
987 struct dma_async_tx_descriptor *tx;
988 struct page **blocks = percpu->scribble;
989 struct async_submit_ctl submit;
991 pr_debug("%s: stripe %llu block1: %d block2: %d\n",
992 __func__, (unsigned long long)sh->sector, target, target2);
993 BUG_ON(target < 0 || target2 < 0);
994 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
995 BUG_ON(!test_bit(R5_Wantcompute, &tgt2->flags));
997 /* we need to open-code set_syndrome_sources to handle the
998 * slot number conversion for 'faila' and 'failb'
1000 for (i = 0; i < disks ; i++)
1005 int slot = raid6_idx_to_slot(i, sh, &count, syndrome_disks);
1007 blocks[slot] = sh->dev[i].page;
1013 i = raid6_next_disk(i, disks);
1014 } while (i != d0_idx);
1016 BUG_ON(faila == failb);
1019 pr_debug("%s: stripe: %llu faila: %d failb: %d\n",
1020 __func__, (unsigned long long)sh->sector, faila, failb);
1022 atomic_inc(&sh->count);
1024 if (failb == syndrome_disks+1) {
1025 /* Q disk is one of the missing disks */
1026 if (faila == syndrome_disks) {
1027 /* Missing P+Q, just recompute */
1028 init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
1029 ops_complete_compute, sh,
1030 to_addr_conv(sh, percpu));
1031 return async_gen_syndrome(blocks, 0, syndrome_disks+2,
1032 STRIPE_SIZE, &submit);
1036 int qd_idx = sh->qd_idx;
1038 /* Missing D+Q: recompute D from P, then recompute Q */
1039 if (target == qd_idx)
1040 data_target = target2;
1042 data_target = target;
1045 for (i = disks; i-- ; ) {
1046 if (i == data_target || i == qd_idx)
1048 blocks[count++] = sh->dev[i].page;
1050 dest = sh->dev[data_target].page;
1051 init_async_submit(&submit,
1052 ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST,
1054 to_addr_conv(sh, percpu));
1055 tx = async_xor(dest, blocks, 0, count, STRIPE_SIZE,
1058 count = set_syndrome_sources(blocks, sh);
1059 init_async_submit(&submit, ASYNC_TX_FENCE, tx,
1060 ops_complete_compute, sh,
1061 to_addr_conv(sh, percpu));
1062 return async_gen_syndrome(blocks, 0, count+2,
1063 STRIPE_SIZE, &submit);
1066 init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
1067 ops_complete_compute, sh,
1068 to_addr_conv(sh, percpu));
1069 if (failb == syndrome_disks) {
1070 /* We're missing D+P. */
1071 return async_raid6_datap_recov(syndrome_disks+2,
1075 /* We're missing D+D. */
1076 return async_raid6_2data_recov(syndrome_disks+2,
1077 STRIPE_SIZE, faila, failb,
1084 static void ops_complete_prexor(void *stripe_head_ref)
1086 struct stripe_head *sh = stripe_head_ref;
1088 pr_debug("%s: stripe %llu\n", __func__,
1089 (unsigned long long)sh->sector);
1092 static struct dma_async_tx_descriptor *
1093 ops_run_prexor(struct stripe_head *sh, struct raid5_percpu *percpu,
1094 struct dma_async_tx_descriptor *tx)
1096 int disks = sh->disks;
1097 struct page **xor_srcs = percpu->scribble;
1098 int count = 0, pd_idx = sh->pd_idx, i;
1099 struct async_submit_ctl submit;
1101 /* existing parity data subtracted */
1102 struct page *xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
1104 pr_debug("%s: stripe %llu\n", __func__,
1105 (unsigned long long)sh->sector);
1107 for (i = disks; i--; ) {
1108 struct r5dev *dev = &sh->dev[i];
1109 /* Only process blocks that are known to be uptodate */
1110 if (test_bit(R5_Wantdrain, &dev->flags))
1111 xor_srcs[count++] = dev->page;
1114 init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_DROP_DST, tx,
1115 ops_complete_prexor, sh, to_addr_conv(sh, percpu));
1116 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
1121 static struct dma_async_tx_descriptor *
1122 ops_run_biodrain(struct stripe_head *sh, struct dma_async_tx_descriptor *tx)
1124 int disks = sh->disks;
1127 pr_debug("%s: stripe %llu\n", __func__,
1128 (unsigned long long)sh->sector);
1130 for (i = disks; i--; ) {
1131 struct r5dev *dev = &sh->dev[i];
1134 if (test_and_clear_bit(R5_Wantdrain, &dev->flags)) {
1137 spin_lock_irq(&sh->raid_conf->device_lock);
1138 chosen = dev->towrite;
1139 dev->towrite = NULL;
1140 BUG_ON(dev->written);
1141 wbi = dev->written = chosen;
1142 spin_unlock_irq(&sh->raid_conf->device_lock);
1144 while (wbi && wbi->bi_sector <
1145 dev->sector + STRIPE_SECTORS) {
1146 if (wbi->bi_rw & REQ_FUA)
1147 set_bit(R5_WantFUA, &dev->flags);
1148 tx = async_copy_data(1, wbi, dev->page,
1150 wbi = r5_next_bio(wbi, dev->sector);
1158 static void ops_complete_reconstruct(void *stripe_head_ref)
1160 struct stripe_head *sh = stripe_head_ref;
1161 int disks = sh->disks;
1162 int pd_idx = sh->pd_idx;
1163 int qd_idx = sh->qd_idx;
1167 pr_debug("%s: stripe %llu\n", __func__,
1168 (unsigned long long)sh->sector);
1170 for (i = disks; i--; )
1171 fua |= test_bit(R5_WantFUA, &sh->dev[i].flags);
1173 for (i = disks; i--; ) {
1174 struct r5dev *dev = &sh->dev[i];
1176 if (dev->written || i == pd_idx || i == qd_idx) {
1177 set_bit(R5_UPTODATE, &dev->flags);
1179 set_bit(R5_WantFUA, &dev->flags);
1183 if (sh->reconstruct_state == reconstruct_state_drain_run)
1184 sh->reconstruct_state = reconstruct_state_drain_result;
1185 else if (sh->reconstruct_state == reconstruct_state_prexor_drain_run)
1186 sh->reconstruct_state = reconstruct_state_prexor_drain_result;
1188 BUG_ON(sh->reconstruct_state != reconstruct_state_run);
1189 sh->reconstruct_state = reconstruct_state_result;
1192 set_bit(STRIPE_HANDLE, &sh->state);
1197 ops_run_reconstruct5(struct stripe_head *sh, struct raid5_percpu *percpu,
1198 struct dma_async_tx_descriptor *tx)
1200 int disks = sh->disks;
1201 struct page **xor_srcs = percpu->scribble;
1202 struct async_submit_ctl submit;
1203 int count = 0, pd_idx = sh->pd_idx, i;
1204 struct page *xor_dest;
1206 unsigned long flags;
1208 pr_debug("%s: stripe %llu\n", __func__,
1209 (unsigned long long)sh->sector);
1211 /* check if prexor is active which means only process blocks
1212 * that are part of a read-modify-write (written)
1214 if (sh->reconstruct_state == reconstruct_state_prexor_drain_run) {
1216 xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
1217 for (i = disks; i--; ) {
1218 struct r5dev *dev = &sh->dev[i];
1220 xor_srcs[count++] = dev->page;
1223 xor_dest = sh->dev[pd_idx].page;
1224 for (i = disks; i--; ) {
1225 struct r5dev *dev = &sh->dev[i];
1227 xor_srcs[count++] = dev->page;
1231 /* 1/ if we prexor'd then the dest is reused as a source
1232 * 2/ if we did not prexor then we are redoing the parity
1233 * set ASYNC_TX_XOR_DROP_DST and ASYNC_TX_XOR_ZERO_DST
1234 * for the synchronous xor case
1236 flags = ASYNC_TX_ACK |
1237 (prexor ? ASYNC_TX_XOR_DROP_DST : ASYNC_TX_XOR_ZERO_DST);
1239 atomic_inc(&sh->count);
1241 init_async_submit(&submit, flags, tx, ops_complete_reconstruct, sh,
1242 to_addr_conv(sh, percpu));
1243 if (unlikely(count == 1))
1244 tx = async_memcpy(xor_dest, xor_srcs[0], 0, 0, STRIPE_SIZE, &submit);
1246 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
1250 ops_run_reconstruct6(struct stripe_head *sh, struct raid5_percpu *percpu,
1251 struct dma_async_tx_descriptor *tx)
1253 struct async_submit_ctl submit;
1254 struct page **blocks = percpu->scribble;
1257 pr_debug("%s: stripe %llu\n", __func__, (unsigned long long)sh->sector);
1259 count = set_syndrome_sources(blocks, sh);
1261 atomic_inc(&sh->count);
1263 init_async_submit(&submit, ASYNC_TX_ACK, tx, ops_complete_reconstruct,
1264 sh, to_addr_conv(sh, percpu));
1265 async_gen_syndrome(blocks, 0, count+2, STRIPE_SIZE, &submit);
1268 static void ops_complete_check(void *stripe_head_ref)
1270 struct stripe_head *sh = stripe_head_ref;
1272 pr_debug("%s: stripe %llu\n", __func__,
1273 (unsigned long long)sh->sector);
1275 sh->check_state = check_state_check_result;
1276 set_bit(STRIPE_HANDLE, &sh->state);
1280 static void ops_run_check_p(struct stripe_head *sh, struct raid5_percpu *percpu)
1282 int disks = sh->disks;
1283 int pd_idx = sh->pd_idx;
1284 int qd_idx = sh->qd_idx;
1285 struct page *xor_dest;
1286 struct page **xor_srcs = percpu->scribble;
1287 struct dma_async_tx_descriptor *tx;
1288 struct async_submit_ctl submit;
1292 pr_debug("%s: stripe %llu\n", __func__,
1293 (unsigned long long)sh->sector);
1296 xor_dest = sh->dev[pd_idx].page;
1297 xor_srcs[count++] = xor_dest;
1298 for (i = disks; i--; ) {
1299 if (i == pd_idx || i == qd_idx)
1301 xor_srcs[count++] = sh->dev[i].page;
1304 init_async_submit(&submit, 0, NULL, NULL, NULL,
1305 to_addr_conv(sh, percpu));
1306 tx = async_xor_val(xor_dest, xor_srcs, 0, count, STRIPE_SIZE,
1307 &sh->ops.zero_sum_result, &submit);
1309 atomic_inc(&sh->count);
1310 init_async_submit(&submit, ASYNC_TX_ACK, tx, ops_complete_check, sh, NULL);
1311 tx = async_trigger_callback(&submit);
1314 static void ops_run_check_pq(struct stripe_head *sh, struct raid5_percpu *percpu, int checkp)
1316 struct page **srcs = percpu->scribble;
1317 struct async_submit_ctl submit;
1320 pr_debug("%s: stripe %llu checkp: %d\n", __func__,
1321 (unsigned long long)sh->sector, checkp);
1323 count = set_syndrome_sources(srcs, sh);
1327 atomic_inc(&sh->count);
1328 init_async_submit(&submit, ASYNC_TX_ACK, NULL, ops_complete_check,
1329 sh, to_addr_conv(sh, percpu));
1330 async_syndrome_val(srcs, 0, count+2, STRIPE_SIZE,
1331 &sh->ops.zero_sum_result, percpu->spare_page, &submit);
1334 static void __raid_run_ops(struct stripe_head *sh, unsigned long ops_request)
1336 int overlap_clear = 0, i, disks = sh->disks;
1337 struct dma_async_tx_descriptor *tx = NULL;
1338 struct r5conf *conf = sh->raid_conf;
1339 int level = conf->level;
1340 struct raid5_percpu *percpu;
1344 percpu = per_cpu_ptr(conf->percpu, cpu);
1345 if (test_bit(STRIPE_OP_BIOFILL, &ops_request)) {
1346 ops_run_biofill(sh);
1350 if (test_bit(STRIPE_OP_COMPUTE_BLK, &ops_request)) {
1352 tx = ops_run_compute5(sh, percpu);
1354 if (sh->ops.target2 < 0 || sh->ops.target < 0)
1355 tx = ops_run_compute6_1(sh, percpu);
1357 tx = ops_run_compute6_2(sh, percpu);
1359 /* terminate the chain if reconstruct is not set to be run */
1360 if (tx && !test_bit(STRIPE_OP_RECONSTRUCT, &ops_request))
1364 if (test_bit(STRIPE_OP_PREXOR, &ops_request))
1365 tx = ops_run_prexor(sh, percpu, tx);
1367 if (test_bit(STRIPE_OP_BIODRAIN, &ops_request)) {
1368 tx = ops_run_biodrain(sh, tx);
1372 if (test_bit(STRIPE_OP_RECONSTRUCT, &ops_request)) {
1374 ops_run_reconstruct5(sh, percpu, tx);
1376 ops_run_reconstruct6(sh, percpu, tx);
1379 if (test_bit(STRIPE_OP_CHECK, &ops_request)) {
1380 if (sh->check_state == check_state_run)
1381 ops_run_check_p(sh, percpu);
1382 else if (sh->check_state == check_state_run_q)
1383 ops_run_check_pq(sh, percpu, 0);
1384 else if (sh->check_state == check_state_run_pq)
1385 ops_run_check_pq(sh, percpu, 1);
1391 for (i = disks; i--; ) {
1392 struct r5dev *dev = &sh->dev[i];
1393 if (test_and_clear_bit(R5_Overlap, &dev->flags))
1394 wake_up(&sh->raid_conf->wait_for_overlap);
1399 #ifdef CONFIG_MULTICORE_RAID456
1400 static void async_run_ops(void *param, async_cookie_t cookie)
1402 struct stripe_head *sh = param;
1403 unsigned long ops_request = sh->ops.request;
1405 clear_bit_unlock(STRIPE_OPS_REQ_PENDING, &sh->state);
1406 wake_up(&sh->ops.wait_for_ops);
1408 __raid_run_ops(sh, ops_request);
1412 static void raid_run_ops(struct stripe_head *sh, unsigned long ops_request)
1414 /* since handle_stripe can be called outside of raid5d context
1415 * we need to ensure sh->ops.request is de-staged before another
1418 wait_event(sh->ops.wait_for_ops,
1419 !test_and_set_bit_lock(STRIPE_OPS_REQ_PENDING, &sh->state));
1420 sh->ops.request = ops_request;
1422 atomic_inc(&sh->count);
1423 async_schedule(async_run_ops, sh);
1426 #define raid_run_ops __raid_run_ops
1429 static int grow_one_stripe(struct r5conf *conf)
1431 struct stripe_head *sh;
1432 sh = kmem_cache_zalloc(conf->slab_cache, GFP_KERNEL);
1436 sh->raid_conf = conf;
1437 #ifdef CONFIG_MULTICORE_RAID456
1438 init_waitqueue_head(&sh->ops.wait_for_ops);
1441 if (grow_buffers(sh)) {
1443 kmem_cache_free(conf->slab_cache, sh);
1446 /* we just created an active stripe so... */
1447 atomic_set(&sh->count, 1);
1448 atomic_inc(&conf->active_stripes);
1449 INIT_LIST_HEAD(&sh->lru);
1454 static int grow_stripes(struct r5conf *conf, int num)
1456 struct kmem_cache *sc;
1457 int devs = max(conf->raid_disks, conf->previous_raid_disks);
1459 if (conf->mddev->gendisk)
1460 sprintf(conf->cache_name[0],
1461 "raid%d-%s", conf->level, mdname(conf->mddev));
1463 sprintf(conf->cache_name[0],
1464 "raid%d-%p", conf->level, conf->mddev);
1465 sprintf(conf->cache_name[1], "%s-alt", conf->cache_name[0]);
1467 conf->active_name = 0;
1468 sc = kmem_cache_create(conf->cache_name[conf->active_name],
1469 sizeof(struct stripe_head)+(devs-1)*sizeof(struct r5dev),
1473 conf->slab_cache = sc;
1474 conf->pool_size = devs;
1476 if (!grow_one_stripe(conf))
1482 * scribble_len - return the required size of the scribble region
1483 * @num - total number of disks in the array
1485 * The size must be enough to contain:
1486 * 1/ a struct page pointer for each device in the array +2
1487 * 2/ room to convert each entry in (1) to its corresponding dma
1488 * (dma_map_page()) or page (page_address()) address.
1490 * Note: the +2 is for the destination buffers of the ddf/raid6 case where we
1491 * calculate over all devices (not just the data blocks), using zeros in place
1492 * of the P and Q blocks.
1494 static size_t scribble_len(int num)
1498 len = sizeof(struct page *) * (num+2) + sizeof(addr_conv_t) * (num+2);
1503 static int resize_stripes(struct r5conf *conf, int newsize)
1505 /* Make all the stripes able to hold 'newsize' devices.
1506 * New slots in each stripe get 'page' set to a new page.
1508 * This happens in stages:
1509 * 1/ create a new kmem_cache and allocate the required number of
1511 * 2/ gather all the old stripe_heads and tranfer the pages across
1512 * to the new stripe_heads. This will have the side effect of
1513 * freezing the array as once all stripe_heads have been collected,
1514 * no IO will be possible. Old stripe heads are freed once their
1515 * pages have been transferred over, and the old kmem_cache is
1516 * freed when all stripes are done.
1517 * 3/ reallocate conf->disks to be suitable bigger. If this fails,
1518 * we simple return a failre status - no need to clean anything up.
1519 * 4/ allocate new pages for the new slots in the new stripe_heads.
1520 * If this fails, we don't bother trying the shrink the
1521 * stripe_heads down again, we just leave them as they are.
1522 * As each stripe_head is processed the new one is released into
1525 * Once step2 is started, we cannot afford to wait for a write,
1526 * so we use GFP_NOIO allocations.
1528 struct stripe_head *osh, *nsh;
1529 LIST_HEAD(newstripes);
1530 struct disk_info *ndisks;
1533 struct kmem_cache *sc;
1536 if (newsize <= conf->pool_size)
1537 return 0; /* never bother to shrink */
1539 err = md_allow_write(conf->mddev);
1544 sc = kmem_cache_create(conf->cache_name[1-conf->active_name],
1545 sizeof(struct stripe_head)+(newsize-1)*sizeof(struct r5dev),
1550 for (i = conf->max_nr_stripes; i; i--) {
1551 nsh = kmem_cache_zalloc(sc, GFP_KERNEL);
1555 nsh->raid_conf = conf;
1556 #ifdef CONFIG_MULTICORE_RAID456
1557 init_waitqueue_head(&nsh->ops.wait_for_ops);
1560 list_add(&nsh->lru, &newstripes);
1563 /* didn't get enough, give up */
1564 while (!list_empty(&newstripes)) {
1565 nsh = list_entry(newstripes.next, struct stripe_head, lru);
1566 list_del(&nsh->lru);
1567 kmem_cache_free(sc, nsh);
1569 kmem_cache_destroy(sc);
1572 /* Step 2 - Must use GFP_NOIO now.
1573 * OK, we have enough stripes, start collecting inactive
1574 * stripes and copying them over
1576 list_for_each_entry(nsh, &newstripes, lru) {
1577 spin_lock_irq(&conf->device_lock);
1578 wait_event_lock_irq(conf->wait_for_stripe,
1579 !list_empty(&conf->inactive_list),
1582 osh = get_free_stripe(conf);
1583 spin_unlock_irq(&conf->device_lock);
1584 atomic_set(&nsh->count, 1);
1585 for(i=0; i<conf->pool_size; i++)
1586 nsh->dev[i].page = osh->dev[i].page;
1587 for( ; i<newsize; i++)
1588 nsh->dev[i].page = NULL;
1589 kmem_cache_free(conf->slab_cache, osh);
1591 kmem_cache_destroy(conf->slab_cache);
1594 * At this point, we are holding all the stripes so the array
1595 * is completely stalled, so now is a good time to resize
1596 * conf->disks and the scribble region
1598 ndisks = kzalloc(newsize * sizeof(struct disk_info), GFP_NOIO);
1600 for (i=0; i<conf->raid_disks; i++)
1601 ndisks[i] = conf->disks[i];
1603 conf->disks = ndisks;
1608 conf->scribble_len = scribble_len(newsize);
1609 for_each_present_cpu(cpu) {
1610 struct raid5_percpu *percpu;
1613 percpu = per_cpu_ptr(conf->percpu, cpu);
1614 scribble = kmalloc(conf->scribble_len, GFP_NOIO);
1617 kfree(percpu->scribble);
1618 percpu->scribble = scribble;
1626 /* Step 4, return new stripes to service */
1627 while(!list_empty(&newstripes)) {
1628 nsh = list_entry(newstripes.next, struct stripe_head, lru);
1629 list_del_init(&nsh->lru);
1631 for (i=conf->raid_disks; i < newsize; i++)
1632 if (nsh->dev[i].page == NULL) {
1633 struct page *p = alloc_page(GFP_NOIO);
1634 nsh->dev[i].page = p;
1638 release_stripe(nsh);
1640 /* critical section pass, GFP_NOIO no longer needed */
1642 conf->slab_cache = sc;
1643 conf->active_name = 1-conf->active_name;
1644 conf->pool_size = newsize;
1648 static int drop_one_stripe(struct r5conf *conf)
1650 struct stripe_head *sh;
1652 spin_lock_irq(&conf->device_lock);
1653 sh = get_free_stripe(conf);
1654 spin_unlock_irq(&conf->device_lock);
1657 BUG_ON(atomic_read(&sh->count));
1659 kmem_cache_free(conf->slab_cache, sh);
1660 atomic_dec(&conf->active_stripes);
1664 static void shrink_stripes(struct r5conf *conf)
1666 while (drop_one_stripe(conf))
1669 if (conf->slab_cache)
1670 kmem_cache_destroy(conf->slab_cache);
1671 conf->slab_cache = NULL;
1674 static void raid5_end_read_request(struct bio * bi, int error)
1676 struct stripe_head *sh = bi->bi_private;
1677 struct r5conf *conf = sh->raid_conf;
1678 int disks = sh->disks, i;
1679 int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
1680 char b[BDEVNAME_SIZE];
1681 struct md_rdev *rdev = NULL;
1684 for (i=0 ; i<disks; i++)
1685 if (bi == &sh->dev[i].req)
1688 pr_debug("end_read_request %llu/%d, count: %d, uptodate %d.\n",
1689 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
1695 if (test_bit(R5_ReadRepl, &sh->dev[i].flags))
1696 /* If replacement finished while this request was outstanding,
1697 * 'replacement' might be NULL already.
1698 * In that case it moved down to 'rdev'.
1699 * rdev is not removed until all requests are finished.
1701 rdev = conf->disks[i].replacement;
1703 rdev = conf->disks[i].rdev;
1705 if (use_new_offset(conf, sh))
1706 s = sh->sector + rdev->new_data_offset;
1708 s = sh->sector + rdev->data_offset;
1710 set_bit(R5_UPTODATE, &sh->dev[i].flags);
1711 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
1712 /* Note that this cannot happen on a
1713 * replacement device. We just fail those on
1718 "md/raid:%s: read error corrected"
1719 " (%lu sectors at %llu on %s)\n",
1720 mdname(conf->mddev), STRIPE_SECTORS,
1721 (unsigned long long)s,
1722 bdevname(rdev->bdev, b));
1723 atomic_add(STRIPE_SECTORS, &rdev->corrected_errors);
1724 clear_bit(R5_ReadError, &sh->dev[i].flags);
1725 clear_bit(R5_ReWrite, &sh->dev[i].flags);
1727 if (atomic_read(&rdev->read_errors))
1728 atomic_set(&rdev->read_errors, 0);
1730 const char *bdn = bdevname(rdev->bdev, b);
1733 clear_bit(R5_UPTODATE, &sh->dev[i].flags);
1734 atomic_inc(&rdev->read_errors);
1735 if (test_bit(R5_ReadRepl, &sh->dev[i].flags))
1738 "md/raid:%s: read error on replacement device "
1739 "(sector %llu on %s).\n",
1740 mdname(conf->mddev),
1741 (unsigned long long)s,
1743 else if (conf->mddev->degraded >= conf->max_degraded)
1746 "md/raid:%s: read error not correctable "
1747 "(sector %llu on %s).\n",
1748 mdname(conf->mddev),
1749 (unsigned long long)s,
1751 else if (test_bit(R5_ReWrite, &sh->dev[i].flags))
1755 "md/raid:%s: read error NOT corrected!! "
1756 "(sector %llu on %s).\n",
1757 mdname(conf->mddev),
1758 (unsigned long long)s,
1760 else if (atomic_read(&rdev->read_errors)
1761 > conf->max_nr_stripes)
1763 "md/raid:%s: Too many read errors, failing device %s.\n",
1764 mdname(conf->mddev), bdn);
1768 set_bit(R5_ReadError, &sh->dev[i].flags);
1770 clear_bit(R5_ReadError, &sh->dev[i].flags);
1771 clear_bit(R5_ReWrite, &sh->dev[i].flags);
1772 md_error(conf->mddev, rdev);
1775 rdev_dec_pending(rdev, conf->mddev);
1776 clear_bit(R5_LOCKED, &sh->dev[i].flags);
1777 set_bit(STRIPE_HANDLE, &sh->state);
1781 static void raid5_end_write_request(struct bio *bi, int error)
1783 struct stripe_head *sh = bi->bi_private;
1784 struct r5conf *conf = sh->raid_conf;
1785 int disks = sh->disks, i;
1786 struct md_rdev *uninitialized_var(rdev);
1787 int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
1790 int replacement = 0;
1792 for (i = 0 ; i < disks; i++) {
1793 if (bi == &sh->dev[i].req) {
1794 rdev = conf->disks[i].rdev;
1797 if (bi == &sh->dev[i].rreq) {
1798 rdev = conf->disks[i].replacement;
1802 /* rdev was removed and 'replacement'
1803 * replaced it. rdev is not removed
1804 * until all requests are finished.
1806 rdev = conf->disks[i].rdev;
1810 pr_debug("end_write_request %llu/%d, count %d, uptodate: %d.\n",
1811 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
1820 md_error(conf->mddev, rdev);
1821 else if (is_badblock(rdev, sh->sector,
1823 &first_bad, &bad_sectors))
1824 set_bit(R5_MadeGoodRepl, &sh->dev[i].flags);
1827 set_bit(WriteErrorSeen, &rdev->flags);
1828 set_bit(R5_WriteError, &sh->dev[i].flags);
1829 if (!test_and_set_bit(WantReplacement, &rdev->flags))
1830 set_bit(MD_RECOVERY_NEEDED,
1831 &rdev->mddev->recovery);
1832 } else if (is_badblock(rdev, sh->sector,
1834 &first_bad, &bad_sectors))
1835 set_bit(R5_MadeGood, &sh->dev[i].flags);
1837 rdev_dec_pending(rdev, conf->mddev);
1839 if (!test_and_clear_bit(R5_DOUBLE_LOCKED, &sh->dev[i].flags))
1840 clear_bit(R5_LOCKED, &sh->dev[i].flags);
1841 set_bit(STRIPE_HANDLE, &sh->state);
1845 static sector_t compute_blocknr(struct stripe_head *sh, int i, int previous);
1847 static void raid5_build_block(struct stripe_head *sh, int i, int previous)
1849 struct r5dev *dev = &sh->dev[i];
1851 bio_init(&dev->req);
1852 dev->req.bi_io_vec = &dev->vec;
1854 dev->req.bi_max_vecs++;
1855 dev->req.bi_private = sh;
1856 dev->vec.bv_page = dev->page;
1858 bio_init(&dev->rreq);
1859 dev->rreq.bi_io_vec = &dev->rvec;
1860 dev->rreq.bi_vcnt++;
1861 dev->rreq.bi_max_vecs++;
1862 dev->rreq.bi_private = sh;
1863 dev->rvec.bv_page = dev->page;
1866 dev->sector = compute_blocknr(sh, i, previous);
1869 static void error(struct mddev *mddev, struct md_rdev *rdev)
1871 char b[BDEVNAME_SIZE];
1872 struct r5conf *conf = mddev->private;
1873 unsigned long flags;
1874 pr_debug("raid456: error called\n");
1876 spin_lock_irqsave(&conf->device_lock, flags);
1877 clear_bit(In_sync, &rdev->flags);
1878 mddev->degraded = calc_degraded(conf);
1879 spin_unlock_irqrestore(&conf->device_lock, flags);
1880 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1882 set_bit(Blocked, &rdev->flags);
1883 set_bit(Faulty, &rdev->flags);
1884 set_bit(MD_CHANGE_DEVS, &mddev->flags);
1886 "md/raid:%s: Disk failure on %s, disabling device.\n"
1887 "md/raid:%s: Operation continuing on %d devices.\n",
1889 bdevname(rdev->bdev, b),
1891 conf->raid_disks - mddev->degraded);
1895 * Input: a 'big' sector number,
1896 * Output: index of the data and parity disk, and the sector # in them.
1898 static sector_t raid5_compute_sector(struct r5conf *conf, sector_t r_sector,
1899 int previous, int *dd_idx,
1900 struct stripe_head *sh)
1902 sector_t stripe, stripe2;
1903 sector_t chunk_number;
1904 unsigned int chunk_offset;
1907 sector_t new_sector;
1908 int algorithm = previous ? conf->prev_algo
1910 int sectors_per_chunk = previous ? conf->prev_chunk_sectors
1911 : conf->chunk_sectors;
1912 int raid_disks = previous ? conf->previous_raid_disks
1914 int data_disks = raid_disks - conf->max_degraded;
1916 /* First compute the information on this sector */
1919 * Compute the chunk number and the sector offset inside the chunk
1921 chunk_offset = sector_div(r_sector, sectors_per_chunk);
1922 chunk_number = r_sector;
1925 * Compute the stripe number
1927 stripe = chunk_number;
1928 *dd_idx = sector_div(stripe, data_disks);
1931 * Select the parity disk based on the user selected algorithm.
1933 pd_idx = qd_idx = -1;
1934 switch(conf->level) {
1936 pd_idx = data_disks;
1939 switch (algorithm) {
1940 case ALGORITHM_LEFT_ASYMMETRIC:
1941 pd_idx = data_disks - sector_div(stripe2, raid_disks);
1942 if (*dd_idx >= pd_idx)
1945 case ALGORITHM_RIGHT_ASYMMETRIC:
1946 pd_idx = sector_div(stripe2, raid_disks);
1947 if (*dd_idx >= pd_idx)
1950 case ALGORITHM_LEFT_SYMMETRIC:
1951 pd_idx = data_disks - sector_div(stripe2, raid_disks);
1952 *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
1954 case ALGORITHM_RIGHT_SYMMETRIC:
1955 pd_idx = sector_div(stripe2, raid_disks);
1956 *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
1958 case ALGORITHM_PARITY_0:
1962 case ALGORITHM_PARITY_N:
1963 pd_idx = data_disks;
1971 switch (algorithm) {
1972 case ALGORITHM_LEFT_ASYMMETRIC:
1973 pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
1974 qd_idx = pd_idx + 1;
1975 if (pd_idx == raid_disks-1) {
1976 (*dd_idx)++; /* Q D D D P */
1978 } else if (*dd_idx >= pd_idx)
1979 (*dd_idx) += 2; /* D D P Q D */
1981 case ALGORITHM_RIGHT_ASYMMETRIC:
1982 pd_idx = sector_div(stripe2, raid_disks);
1983 qd_idx = pd_idx + 1;
1984 if (pd_idx == raid_disks-1) {
1985 (*dd_idx)++; /* Q D D D P */
1987 } else if (*dd_idx >= pd_idx)
1988 (*dd_idx) += 2; /* D D P Q D */
1990 case ALGORITHM_LEFT_SYMMETRIC:
1991 pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
1992 qd_idx = (pd_idx + 1) % raid_disks;
1993 *dd_idx = (pd_idx + 2 + *dd_idx) % raid_disks;
1995 case ALGORITHM_RIGHT_SYMMETRIC:
1996 pd_idx = sector_div(stripe2, raid_disks);
1997 qd_idx = (pd_idx + 1) % raid_disks;
1998 *dd_idx = (pd_idx + 2 + *dd_idx) % raid_disks;
2001 case ALGORITHM_PARITY_0:
2006 case ALGORITHM_PARITY_N:
2007 pd_idx = data_disks;
2008 qd_idx = data_disks + 1;
2011 case ALGORITHM_ROTATING_ZERO_RESTART:
2012 /* Exactly the same as RIGHT_ASYMMETRIC, but or
2013 * of blocks for computing Q is different.
2015 pd_idx = sector_div(stripe2, raid_disks);
2016 qd_idx = pd_idx + 1;
2017 if (pd_idx == raid_disks-1) {
2018 (*dd_idx)++; /* Q D D D P */
2020 } else if (*dd_idx >= pd_idx)
2021 (*dd_idx) += 2; /* D D P Q D */
2025 case ALGORITHM_ROTATING_N_RESTART:
2026 /* Same a left_asymmetric, by first stripe is
2027 * D D D P Q rather than
2031 pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
2032 qd_idx = pd_idx + 1;
2033 if (pd_idx == raid_disks-1) {
2034 (*dd_idx)++; /* Q D D D P */
2036 } else if (*dd_idx >= pd_idx)
2037 (*dd_idx) += 2; /* D D P Q D */
2041 case ALGORITHM_ROTATING_N_CONTINUE:
2042 /* Same as left_symmetric but Q is before P */
2043 pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
2044 qd_idx = (pd_idx + raid_disks - 1) % raid_disks;
2045 *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
2049 case ALGORITHM_LEFT_ASYMMETRIC_6:
2050 /* RAID5 left_asymmetric, with Q on last device */
2051 pd_idx = data_disks - sector_div(stripe2, raid_disks-1);
2052 if (*dd_idx >= pd_idx)
2054 qd_idx = raid_disks - 1;
2057 case ALGORITHM_RIGHT_ASYMMETRIC_6:
2058 pd_idx = sector_div(stripe2, raid_disks-1);
2059 if (*dd_idx >= pd_idx)
2061 qd_idx = raid_disks - 1;
2064 case ALGORITHM_LEFT_SYMMETRIC_6:
2065 pd_idx = data_disks - sector_div(stripe2, raid_disks-1);
2066 *dd_idx = (pd_idx + 1 + *dd_idx) % (raid_disks-1);
2067 qd_idx = raid_disks - 1;
2070 case ALGORITHM_RIGHT_SYMMETRIC_6:
2071 pd_idx = sector_div(stripe2, raid_disks-1);
2072 *dd_idx = (pd_idx + 1 + *dd_idx) % (raid_disks-1);
2073 qd_idx = raid_disks - 1;
2076 case ALGORITHM_PARITY_0_6:
2079 qd_idx = raid_disks - 1;
2089 sh->pd_idx = pd_idx;
2090 sh->qd_idx = qd_idx;
2091 sh->ddf_layout = ddf_layout;
2094 * Finally, compute the new sector number
2096 new_sector = (sector_t)stripe * sectors_per_chunk + chunk_offset;
2101 static sector_t compute_blocknr(struct stripe_head *sh, int i, int previous)
2103 struct r5conf *conf = sh->raid_conf;
2104 int raid_disks = sh->disks;
2105 int data_disks = raid_disks - conf->max_degraded;
2106 sector_t new_sector = sh->sector, check;
2107 int sectors_per_chunk = previous ? conf->prev_chunk_sectors
2108 : conf->chunk_sectors;
2109 int algorithm = previous ? conf->prev_algo
2113 sector_t chunk_number;
2114 int dummy1, dd_idx = i;
2116 struct stripe_head sh2;
2119 chunk_offset = sector_div(new_sector, sectors_per_chunk);
2120 stripe = new_sector;
2122 if (i == sh->pd_idx)
2124 switch(conf->level) {
2127 switch (algorithm) {
2128 case ALGORITHM_LEFT_ASYMMETRIC:
2129 case ALGORITHM_RIGHT_ASYMMETRIC:
2133 case ALGORITHM_LEFT_SYMMETRIC:
2134 case ALGORITHM_RIGHT_SYMMETRIC:
2137 i -= (sh->pd_idx + 1);
2139 case ALGORITHM_PARITY_0:
2142 case ALGORITHM_PARITY_N:
2149 if (i == sh->qd_idx)
2150 return 0; /* It is the Q disk */
2151 switch (algorithm) {
2152 case ALGORITHM_LEFT_ASYMMETRIC:
2153 case ALGORITHM_RIGHT_ASYMMETRIC:
2154 case ALGORITHM_ROTATING_ZERO_RESTART:
2155 case ALGORITHM_ROTATING_N_RESTART:
2156 if (sh->pd_idx == raid_disks-1)
2157 i--; /* Q D D D P */
2158 else if (i > sh->pd_idx)
2159 i -= 2; /* D D P Q D */
2161 case ALGORITHM_LEFT_SYMMETRIC:
2162 case ALGORITHM_RIGHT_SYMMETRIC:
2163 if (sh->pd_idx == raid_disks-1)
2164 i--; /* Q D D D P */
2169 i -= (sh->pd_idx + 2);
2172 case ALGORITHM_PARITY_0:
2175 case ALGORITHM_PARITY_N:
2177 case ALGORITHM_ROTATING_N_CONTINUE:
2178 /* Like left_symmetric, but P is before Q */
2179 if (sh->pd_idx == 0)
2180 i--; /* P D D D Q */
2185 i -= (sh->pd_idx + 1);
2188 case ALGORITHM_LEFT_ASYMMETRIC_6:
2189 case ALGORITHM_RIGHT_ASYMMETRIC_6:
2193 case ALGORITHM_LEFT_SYMMETRIC_6:
2194 case ALGORITHM_RIGHT_SYMMETRIC_6:
2196 i += data_disks + 1;
2197 i -= (sh->pd_idx + 1);
2199 case ALGORITHM_PARITY_0_6:
2208 chunk_number = stripe * data_disks + i;
2209 r_sector = chunk_number * sectors_per_chunk + chunk_offset;
2211 check = raid5_compute_sector(conf, r_sector,
2212 previous, &dummy1, &sh2);
2213 if (check != sh->sector || dummy1 != dd_idx || sh2.pd_idx != sh->pd_idx
2214 || sh2.qd_idx != sh->qd_idx) {
2215 printk(KERN_ERR "md/raid:%s: compute_blocknr: map not correct\n",
2216 mdname(conf->mddev));
2224 schedule_reconstruction(struct stripe_head *sh, struct stripe_head_state *s,
2225 int rcw, int expand)
2227 int i, pd_idx = sh->pd_idx, disks = sh->disks;
2228 struct r5conf *conf = sh->raid_conf;
2229 int level = conf->level;
2232 /* if we are not expanding this is a proper write request, and
2233 * there will be bios with new data to be drained into the
2237 sh->reconstruct_state = reconstruct_state_drain_run;
2238 set_bit(STRIPE_OP_BIODRAIN, &s->ops_request);
2240 sh->reconstruct_state = reconstruct_state_run;
2242 set_bit(STRIPE_OP_RECONSTRUCT, &s->ops_request);
2244 for (i = disks; i--; ) {
2245 struct r5dev *dev = &sh->dev[i];
2248 set_bit(R5_LOCKED, &dev->flags);
2249 set_bit(R5_Wantdrain, &dev->flags);
2251 clear_bit(R5_UPTODATE, &dev->flags);
2255 if (s->locked + conf->max_degraded == disks)
2256 if (!test_and_set_bit(STRIPE_FULL_WRITE, &sh->state))
2257 atomic_inc(&conf->pending_full_writes);
2260 BUG_ON(!(test_bit(R5_UPTODATE, &sh->dev[pd_idx].flags) ||
2261 test_bit(R5_Wantcompute, &sh->dev[pd_idx].flags)));
2263 sh->reconstruct_state = reconstruct_state_prexor_drain_run;
2264 set_bit(STRIPE_OP_PREXOR, &s->ops_request);
2265 set_bit(STRIPE_OP_BIODRAIN, &s->ops_request);
2266 set_bit(STRIPE_OP_RECONSTRUCT, &s->ops_request);
2268 for (i = disks; i--; ) {
2269 struct r5dev *dev = &sh->dev[i];
2274 (test_bit(R5_UPTODATE, &dev->flags) ||
2275 test_bit(R5_Wantcompute, &dev->flags))) {
2276 set_bit(R5_Wantdrain, &dev->flags);
2277 set_bit(R5_LOCKED, &dev->flags);
2278 clear_bit(R5_UPTODATE, &dev->flags);
2284 /* keep the parity disk(s) locked while asynchronous operations
2287 set_bit(R5_LOCKED, &sh->dev[pd_idx].flags);
2288 clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
2292 int qd_idx = sh->qd_idx;
2293 struct r5dev *dev = &sh->dev[qd_idx];
2295 set_bit(R5_LOCKED, &dev->flags);
2296 clear_bit(R5_UPTODATE, &dev->flags);
2300 pr_debug("%s: stripe %llu locked: %d ops_request: %lx\n",
2301 __func__, (unsigned long long)sh->sector,
2302 s->locked, s->ops_request);
2306 * Each stripe/dev can have one or more bion attached.
2307 * toread/towrite point to the first in a chain.
2308 * The bi_next chain must be in order.
2310 static int add_stripe_bio(struct stripe_head *sh, struct bio *bi, int dd_idx, int forwrite)
2313 struct r5conf *conf = sh->raid_conf;
2316 pr_debug("adding bi b#%llu to stripe s#%llu\n",
2317 (unsigned long long)bi->bi_sector,
2318 (unsigned long long)sh->sector);
2321 spin_lock_irq(&conf->device_lock);
2323 bip = &sh->dev[dd_idx].towrite;
2324 if (*bip == NULL && sh->dev[dd_idx].written == NULL)
2327 bip = &sh->dev[dd_idx].toread;
2328 while (*bip && (*bip)->bi_sector < bi->bi_sector) {
2329 if ((*bip)->bi_sector + ((*bip)->bi_size >> 9) > bi->bi_sector)
2331 bip = & (*bip)->bi_next;
2333 if (*bip && (*bip)->bi_sector < bi->bi_sector + ((bi->bi_size)>>9))
2336 BUG_ON(*bip && bi->bi_next && (*bip) != bi->bi_next);
2340 bi->bi_phys_segments++;
2343 /* check if page is covered */
2344 sector_t sector = sh->dev[dd_idx].sector;
2345 for (bi=sh->dev[dd_idx].towrite;
2346 sector < sh->dev[dd_idx].sector + STRIPE_SECTORS &&
2347 bi && bi->bi_sector <= sector;
2348 bi = r5_next_bio(bi, sh->dev[dd_idx].sector)) {
2349 if (bi->bi_sector + (bi->bi_size>>9) >= sector)
2350 sector = bi->bi_sector + (bi->bi_size>>9);
2352 if (sector >= sh->dev[dd_idx].sector + STRIPE_SECTORS)
2353 set_bit(R5_OVERWRITE, &sh->dev[dd_idx].flags);
2355 spin_unlock_irq(&conf->device_lock);
2357 pr_debug("added bi b#%llu to stripe s#%llu, disk %d.\n",
2358 (unsigned long long)(*bip)->bi_sector,
2359 (unsigned long long)sh->sector, dd_idx);
2361 if (conf->mddev->bitmap && firstwrite) {
2362 bitmap_startwrite(conf->mddev->bitmap, sh->sector,
2364 sh->bm_seq = conf->seq_flush+1;
2365 set_bit(STRIPE_BIT_DELAY, &sh->state);
2370 set_bit(R5_Overlap, &sh->dev[dd_idx].flags);
2371 spin_unlock_irq(&conf->device_lock);
2375 static void end_reshape(struct r5conf *conf);
2377 static void stripe_set_idx(sector_t stripe, struct r5conf *conf, int previous,
2378 struct stripe_head *sh)
2380 int sectors_per_chunk =
2381 previous ? conf->prev_chunk_sectors : conf->chunk_sectors;
2383 int chunk_offset = sector_div(stripe, sectors_per_chunk);
2384 int disks = previous ? conf->previous_raid_disks : conf->raid_disks;
2386 raid5_compute_sector(conf,
2387 stripe * (disks - conf->max_degraded)
2388 *sectors_per_chunk + chunk_offset,
2394 handle_failed_stripe(struct r5conf *conf, struct stripe_head *sh,
2395 struct stripe_head_state *s, int disks,
2396 struct bio **return_bi)
2399 for (i = disks; i--; ) {
2403 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
2404 struct md_rdev *rdev;
2406 rdev = rcu_dereference(conf->disks[i].rdev);
2407 if (rdev && test_bit(In_sync, &rdev->flags))
2408 atomic_inc(&rdev->nr_pending);
2413 if (!rdev_set_badblocks(
2417 md_error(conf->mddev, rdev);
2418 rdev_dec_pending(rdev, conf->mddev);
2421 spin_lock_irq(&conf->device_lock);
2422 /* fail all writes first */
2423 bi = sh->dev[i].towrite;
2424 sh->dev[i].towrite = NULL;
2430 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
2431 wake_up(&conf->wait_for_overlap);
2433 while (bi && bi->bi_sector <
2434 sh->dev[i].sector + STRIPE_SECTORS) {
2435 struct bio *nextbi = r5_next_bio(bi, sh->dev[i].sector);
2436 clear_bit(BIO_UPTODATE, &bi->bi_flags);
2437 if (!raid5_dec_bi_phys_segments(bi)) {
2438 md_write_end(conf->mddev);
2439 bi->bi_next = *return_bi;
2444 /* and fail all 'written' */
2445 bi = sh->dev[i].written;
2446 sh->dev[i].written = NULL;
2447 if (bi) bitmap_end = 1;
2448 while (bi && bi->bi_sector <
2449 sh->dev[i].sector + STRIPE_SECTORS) {
2450 struct bio *bi2 = r5_next_bio(bi, sh->dev[i].sector);
2451 clear_bit(BIO_UPTODATE, &bi->bi_flags);
2452 if (!raid5_dec_bi_phys_segments(bi)) {
2453 md_write_end(conf->mddev);
2454 bi->bi_next = *return_bi;
2460 /* fail any reads if this device is non-operational and
2461 * the data has not reached the cache yet.
2463 if (!test_bit(R5_Wantfill, &sh->dev[i].flags) &&
2464 (!test_bit(R5_Insync, &sh->dev[i].flags) ||
2465 test_bit(R5_ReadError, &sh->dev[i].flags))) {
2466 bi = sh->dev[i].toread;
2467 sh->dev[i].toread = NULL;
2468 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
2469 wake_up(&conf->wait_for_overlap);
2470 if (bi) s->to_read--;
2471 while (bi && bi->bi_sector <
2472 sh->dev[i].sector + STRIPE_SECTORS) {
2473 struct bio *nextbi =
2474 r5_next_bio(bi, sh->dev[i].sector);
2475 clear_bit(BIO_UPTODATE, &bi->bi_flags);
2476 if (!raid5_dec_bi_phys_segments(bi)) {
2477 bi->bi_next = *return_bi;
2483 spin_unlock_irq(&conf->device_lock);
2485 bitmap_endwrite(conf->mddev->bitmap, sh->sector,
2486 STRIPE_SECTORS, 0, 0);
2487 /* If we were in the middle of a write the parity block might
2488 * still be locked - so just clear all R5_LOCKED flags
2490 clear_bit(R5_LOCKED, &sh->dev[i].flags);
2493 if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
2494 if (atomic_dec_and_test(&conf->pending_full_writes))
2495 md_wakeup_thread(conf->mddev->thread);
2499 handle_failed_sync(struct r5conf *conf, struct stripe_head *sh,
2500 struct stripe_head_state *s)
2505 clear_bit(STRIPE_SYNCING, &sh->state);
2508 /* There is nothing more to do for sync/check/repair.
2509 * Don't even need to abort as that is handled elsewhere
2510 * if needed, and not always wanted e.g. if there is a known
2512 * For recover/replace we need to record a bad block on all
2513 * non-sync devices, or abort the recovery
2515 if (test_bit(MD_RECOVERY_RECOVER, &conf->mddev->recovery)) {
2516 /* During recovery devices cannot be removed, so
2517 * locking and refcounting of rdevs is not needed
2519 for (i = 0; i < conf->raid_disks; i++) {
2520 struct md_rdev *rdev = conf->disks[i].rdev;
2522 && !test_bit(Faulty, &rdev->flags)
2523 && !test_bit(In_sync, &rdev->flags)
2524 && !rdev_set_badblocks(rdev, sh->sector,
2527 rdev = conf->disks[i].replacement;
2529 && !test_bit(Faulty, &rdev->flags)
2530 && !test_bit(In_sync, &rdev->flags)
2531 && !rdev_set_badblocks(rdev, sh->sector,
2536 conf->recovery_disabled =
2537 conf->mddev->recovery_disabled;
2539 md_done_sync(conf->mddev, STRIPE_SECTORS, !abort);
2542 static int want_replace(struct stripe_head *sh, int disk_idx)
2544 struct md_rdev *rdev;
2546 /* Doing recovery so rcu locking not required */
2547 rdev = sh->raid_conf->disks[disk_idx].replacement;
2549 && !test_bit(Faulty, &rdev->flags)
2550 && !test_bit(In_sync, &rdev->flags)
2551 && (rdev->recovery_offset <= sh->sector
2552 || rdev->mddev->recovery_cp <= sh->sector))
2558 /* fetch_block - checks the given member device to see if its data needs
2559 * to be read or computed to satisfy a request.
2561 * Returns 1 when no more member devices need to be checked, otherwise returns
2562 * 0 to tell the loop in handle_stripe_fill to continue
2564 static int fetch_block(struct stripe_head *sh, struct stripe_head_state *s,
2565 int disk_idx, int disks)
2567 struct r5dev *dev = &sh->dev[disk_idx];
2568 struct r5dev *fdev[2] = { &sh->dev[s->failed_num[0]],
2569 &sh->dev[s->failed_num[1]] };
2571 /* is the data in this block needed, and can we get it? */
2572 if (!test_bit(R5_LOCKED, &dev->flags) &&
2573 !test_bit(R5_UPTODATE, &dev->flags) &&
2575 (dev->towrite && !test_bit(R5_OVERWRITE, &dev->flags)) ||
2576 s->syncing || s->expanding ||
2577 (s->replacing && want_replace(sh, disk_idx)) ||
2578 (s->failed >= 1 && fdev[0]->toread) ||
2579 (s->failed >= 2 && fdev[1]->toread) ||
2580 (sh->raid_conf->level <= 5 && s->failed && fdev[0]->towrite &&
2581 !test_bit(R5_OVERWRITE, &fdev[0]->flags)) ||
2582 (sh->raid_conf->level == 6 && s->failed && s->to_write))) {
2583 /* we would like to get this block, possibly by computing it,
2584 * otherwise read it if the backing disk is insync
2586 BUG_ON(test_bit(R5_Wantcompute, &dev->flags));
2587 BUG_ON(test_bit(R5_Wantread, &dev->flags));
2588 if ((s->uptodate == disks - 1) &&
2589 (s->failed && (disk_idx == s->failed_num[0] ||
2590 disk_idx == s->failed_num[1]))) {
2591 /* have disk failed, and we're requested to fetch it;
2594 pr_debug("Computing stripe %llu block %d\n",
2595 (unsigned long long)sh->sector, disk_idx);
2596 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
2597 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
2598 set_bit(R5_Wantcompute, &dev->flags);
2599 sh->ops.target = disk_idx;
2600 sh->ops.target2 = -1; /* no 2nd target */
2602 /* Careful: from this point on 'uptodate' is in the eye
2603 * of raid_run_ops which services 'compute' operations
2604 * before writes. R5_Wantcompute flags a block that will
2605 * be R5_UPTODATE by the time it is needed for a
2606 * subsequent operation.
2610 } else if (s->uptodate == disks-2 && s->failed >= 2) {
2611 /* Computing 2-failure is *very* expensive; only
2612 * do it if failed >= 2
2615 for (other = disks; other--; ) {
2616 if (other == disk_idx)
2618 if (!test_bit(R5_UPTODATE,
2619 &sh->dev[other].flags))
2623 pr_debug("Computing stripe %llu blocks %d,%d\n",
2624 (unsigned long long)sh->sector,
2626 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
2627 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
2628 set_bit(R5_Wantcompute, &sh->dev[disk_idx].flags);
2629 set_bit(R5_Wantcompute, &sh->dev[other].flags);
2630 sh->ops.target = disk_idx;
2631 sh->ops.target2 = other;
2635 } else if (test_bit(R5_Insync, &dev->flags)) {
2636 set_bit(R5_LOCKED, &dev->flags);
2637 set_bit(R5_Wantread, &dev->flags);
2639 pr_debug("Reading block %d (sync=%d)\n",
2640 disk_idx, s->syncing);
2648 * handle_stripe_fill - read or compute data to satisfy pending requests.
2650 static void handle_stripe_fill(struct stripe_head *sh,
2651 struct stripe_head_state *s,
2656 /* look for blocks to read/compute, skip this if a compute
2657 * is already in flight, or if the stripe contents are in the
2658 * midst of changing due to a write
2660 if (!test_bit(STRIPE_COMPUTE_RUN, &sh->state) && !sh->check_state &&
2661 !sh->reconstruct_state)
2662 for (i = disks; i--; )
2663 if (fetch_block(sh, s, i, disks))
2665 set_bit(STRIPE_HANDLE, &sh->state);
2669 /* handle_stripe_clean_event
2670 * any written block on an uptodate or failed drive can be returned.
2671 * Note that if we 'wrote' to a failed drive, it will be UPTODATE, but
2672 * never LOCKED, so we don't need to test 'failed' directly.
2674 static void handle_stripe_clean_event(struct r5conf *conf,
2675 struct stripe_head *sh, int disks, struct bio **return_bi)
2680 for (i = disks; i--; )
2681 if (sh->dev[i].written) {
2683 if (!test_bit(R5_LOCKED, &dev->flags) &&
2684 test_bit(R5_UPTODATE, &dev->flags)) {
2685 /* We can return any write requests */
2686 struct bio *wbi, *wbi2;
2688 pr_debug("Return write for disc %d\n", i);
2689 spin_lock_irq(&conf->device_lock);
2691 dev->written = NULL;
2692 while (wbi && wbi->bi_sector <
2693 dev->sector + STRIPE_SECTORS) {
2694 wbi2 = r5_next_bio(wbi, dev->sector);
2695 if (!raid5_dec_bi_phys_segments(wbi)) {
2696 md_write_end(conf->mddev);
2697 wbi->bi_next = *return_bi;
2702 if (dev->towrite == NULL)
2704 spin_unlock_irq(&conf->device_lock);
2706 bitmap_endwrite(conf->mddev->bitmap,
2709 !test_bit(STRIPE_DEGRADED, &sh->state),
2714 if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
2715 if (atomic_dec_and_test(&conf->pending_full_writes))
2716 md_wakeup_thread(conf->mddev->thread);
2719 static void handle_stripe_dirtying(struct r5conf *conf,
2720 struct stripe_head *sh,
2721 struct stripe_head_state *s,
2724 int rmw = 0, rcw = 0, i;
2725 if (conf->max_degraded == 2) {
2726 /* RAID6 requires 'rcw' in current implementation
2727 * Calculate the real rcw later - for now fake it
2728 * look like rcw is cheaper
2731 } else for (i = disks; i--; ) {
2732 /* would I have to read this buffer for read_modify_write */
2733 struct r5dev *dev = &sh->dev[i];
2734 if ((dev->towrite || i == sh->pd_idx) &&
2735 !test_bit(R5_LOCKED, &dev->flags) &&
2736 !(test_bit(R5_UPTODATE, &dev->flags) ||
2737 test_bit(R5_Wantcompute, &dev->flags))) {
2738 if (test_bit(R5_Insync, &dev->flags))
2741 rmw += 2*disks; /* cannot read it */
2743 /* Would I have to read this buffer for reconstruct_write */
2744 if (!test_bit(R5_OVERWRITE, &dev->flags) && i != sh->pd_idx &&
2745 !test_bit(R5_LOCKED, &dev->flags) &&
2746 !(test_bit(R5_UPTODATE, &dev->flags) ||
2747 test_bit(R5_Wantcompute, &dev->flags))) {
2748 if (test_bit(R5_Insync, &dev->flags)) rcw++;
2753 pr_debug("for sector %llu, rmw=%d rcw=%d\n",
2754 (unsigned long long)sh->sector, rmw, rcw);
2755 set_bit(STRIPE_HANDLE, &sh->state);
2756 if (rmw < rcw && rmw > 0)
2757 /* prefer read-modify-write, but need to get some data */
2758 for (i = disks; i--; ) {
2759 struct r5dev *dev = &sh->dev[i];
2760 if ((dev->towrite || i == sh->pd_idx) &&
2761 !test_bit(R5_LOCKED, &dev->flags) &&
2762 !(test_bit(R5_UPTODATE, &dev->flags) ||
2763 test_bit(R5_Wantcompute, &dev->flags)) &&
2764 test_bit(R5_Insync, &dev->flags)) {
2766 test_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2767 pr_debug("Read_old block "
2768 "%d for r-m-w\n", i);
2769 set_bit(R5_LOCKED, &dev->flags);
2770 set_bit(R5_Wantread, &dev->flags);
2773 set_bit(STRIPE_DELAYED, &sh->state);
2774 set_bit(STRIPE_HANDLE, &sh->state);
2778 if (rcw <= rmw && rcw > 0) {
2779 /* want reconstruct write, but need to get some data */
2781 for (i = disks; i--; ) {
2782 struct r5dev *dev = &sh->dev[i];
2783 if (!test_bit(R5_OVERWRITE, &dev->flags) &&
2784 i != sh->pd_idx && i != sh->qd_idx &&
2785 !test_bit(R5_LOCKED, &dev->flags) &&
2786 !(test_bit(R5_UPTODATE, &dev->flags) ||
2787 test_bit(R5_Wantcompute, &dev->flags))) {
2789 if (!test_bit(R5_Insync, &dev->flags))
2790 continue; /* it's a failed drive */
2792 test_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2793 pr_debug("Read_old block "
2794 "%d for Reconstruct\n", i);
2795 set_bit(R5_LOCKED, &dev->flags);
2796 set_bit(R5_Wantread, &dev->flags);
2799 set_bit(STRIPE_DELAYED, &sh->state);
2800 set_bit(STRIPE_HANDLE, &sh->state);
2805 /* now if nothing is locked, and if we have enough data,
2806 * we can start a write request
2808 /* since handle_stripe can be called at any time we need to handle the
2809 * case where a compute block operation has been submitted and then a
2810 * subsequent call wants to start a write request. raid_run_ops only
2811 * handles the case where compute block and reconstruct are requested
2812 * simultaneously. If this is not the case then new writes need to be
2813 * held off until the compute completes.
2815 if ((s->req_compute || !test_bit(STRIPE_COMPUTE_RUN, &sh->state)) &&
2816 (s->locked == 0 && (rcw == 0 || rmw == 0) &&
2817 !test_bit(STRIPE_BIT_DELAY, &sh->state)))
2818 schedule_reconstruction(sh, s, rcw == 0, 0);
2821 static void handle_parity_checks5(struct r5conf *conf, struct stripe_head *sh,
2822 struct stripe_head_state *s, int disks)
2824 struct r5dev *dev = NULL;
2826 set_bit(STRIPE_HANDLE, &sh->state);
2828 switch (sh->check_state) {
2829 case check_state_idle:
2830 /* start a new check operation if there are no failures */
2831 if (s->failed == 0) {
2832 BUG_ON(s->uptodate != disks);
2833 sh->check_state = check_state_run;
2834 set_bit(STRIPE_OP_CHECK, &s->ops_request);
2835 clear_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags);
2839 dev = &sh->dev[s->failed_num[0]];
2841 case check_state_compute_result:
2842 sh->check_state = check_state_idle;
2844 dev = &sh->dev[sh->pd_idx];
2846 /* check that a write has not made the stripe insync */
2847 if (test_bit(STRIPE_INSYNC, &sh->state))
2850 /* either failed parity check, or recovery is happening */
2851 BUG_ON(!test_bit(R5_UPTODATE, &dev->flags));
2852 BUG_ON(s->uptodate != disks);
2854 set_bit(R5_LOCKED, &dev->flags);
2856 set_bit(R5_Wantwrite, &dev->flags);
2858 clear_bit(STRIPE_DEGRADED, &sh->state);
2859 set_bit(STRIPE_INSYNC, &sh->state);
2861 case check_state_run:
2862 break; /* we will be called again upon completion */
2863 case check_state_check_result:
2864 sh->check_state = check_state_idle;
2866 /* if a failure occurred during the check operation, leave
2867 * STRIPE_INSYNC not set and let the stripe be handled again
2872 /* handle a successful check operation, if parity is correct
2873 * we are done. Otherwise update the mismatch count and repair
2874 * parity if !MD_RECOVERY_CHECK
2876 if ((sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) == 0)
2877 /* parity is correct (on disc,
2878 * not in buffer any more)
2880 set_bit(STRIPE_INSYNC, &sh->state);
2882 conf->mddev->resync_mismatches += STRIPE_SECTORS;
2883 if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery))
2884 /* don't try to repair!! */
2885 set_bit(STRIPE_INSYNC, &sh->state);
2887 sh->check_state = check_state_compute_run;
2888 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
2889 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
2890 set_bit(R5_Wantcompute,
2891 &sh->dev[sh->pd_idx].flags);
2892 sh->ops.target = sh->pd_idx;
2893 sh->ops.target2 = -1;
2898 case check_state_compute_run:
2901 printk(KERN_ERR "%s: unknown check_state: %d sector: %llu\n",
2902 __func__, sh->check_state,
2903 (unsigned long long) sh->sector);
2909 static void handle_parity_checks6(struct r5conf *conf, struct stripe_head *sh,
2910 struct stripe_head_state *s,
2913 int pd_idx = sh->pd_idx;
2914 int qd_idx = sh->qd_idx;
2917 set_bit(STRIPE_HANDLE, &sh->state);
2919 BUG_ON(s->failed > 2);
2921 /* Want to check and possibly repair P and Q.
2922 * However there could be one 'failed' device, in which
2923 * case we can only check one of them, possibly using the
2924 * other to generate missing data
2927 switch (sh->check_state) {
2928 case check_state_idle:
2929 /* start a new check operation if there are < 2 failures */
2930 if (s->failed == s->q_failed) {
2931 /* The only possible failed device holds Q, so it
2932 * makes sense to check P (If anything else were failed,
2933 * we would have used P to recreate it).
2935 sh->check_state = check_state_run;
2937 if (!s->q_failed && s->failed < 2) {
2938 /* Q is not failed, and we didn't use it to generate
2939 * anything, so it makes sense to check it
2941 if (sh->check_state == check_state_run)
2942 sh->check_state = check_state_run_pq;
2944 sh->check_state = check_state_run_q;
2947 /* discard potentially stale zero_sum_result */
2948 sh->ops.zero_sum_result = 0;
2950 if (sh->check_state == check_state_run) {
2951 /* async_xor_zero_sum destroys the contents of P */
2952 clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
2955 if (sh->check_state >= check_state_run &&
2956 sh->check_state <= check_state_run_pq) {
2957 /* async_syndrome_zero_sum preserves P and Q, so
2958 * no need to mark them !uptodate here
2960 set_bit(STRIPE_OP_CHECK, &s->ops_request);
2964 /* we have 2-disk failure */
2965 BUG_ON(s->failed != 2);
2967 case check_state_compute_result:
2968 sh->check_state = check_state_idle;
2970 /* check that a write has not made the stripe insync */
2971 if (test_bit(STRIPE_INSYNC, &sh->state))
2974 /* now write out any block on a failed drive,
2975 * or P or Q if they were recomputed
2977 BUG_ON(s->uptodate < disks - 1); /* We don't need Q to recover */
2978 if (s->failed == 2) {
2979 dev = &sh->dev[s->failed_num[1]];
2981 set_bit(R5_LOCKED, &dev->flags);
2982 set_bit(R5_Wantwrite, &dev->flags);
2984 if (s->failed >= 1) {
2985 dev = &sh->dev[s->failed_num[0]];
2987 set_bit(R5_LOCKED, &dev->flags);
2988 set_bit(R5_Wantwrite, &dev->flags);
2990 if (sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) {
2991 dev = &sh->dev[pd_idx];
2993 set_bit(R5_LOCKED, &dev->flags);
2994 set_bit(R5_Wantwrite, &dev->flags);
2996 if (sh->ops.zero_sum_result & SUM_CHECK_Q_RESULT) {
2997 dev = &sh->dev[qd_idx];
2999 set_bit(R5_LOCKED, &dev->flags);
3000 set_bit(R5_Wantwrite, &dev->flags);
3002 clear_bit(STRIPE_DEGRADED, &sh->state);
3004 set_bit(STRIPE_INSYNC, &sh->state);
3006 case check_state_run:
3007 case check_state_run_q:
3008 case check_state_run_pq:
3009 break; /* we will be called again upon completion */
3010 case check_state_check_result:
3011 sh->check_state = check_state_idle;
3013 /* handle a successful check operation, if parity is correct
3014 * we are done. Otherwise update the mismatch count and repair
3015 * parity if !MD_RECOVERY_CHECK
3017 if (sh->ops.zero_sum_result == 0) {
3018 /* both parities are correct */
3020 set_bit(STRIPE_INSYNC, &sh->state);
3022 /* in contrast to the raid5 case we can validate
3023 * parity, but still have a failure to write
3026 sh->check_state = check_state_compute_result;
3027 /* Returning at this point means that we may go
3028 * off and bring p and/or q uptodate again so
3029 * we make sure to check zero_sum_result again
3030 * to verify if p or q need writeback
3034 conf->mddev->resync_mismatches += STRIPE_SECTORS;
3035 if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery))
3036 /* don't try to repair!! */
3037 set_bit(STRIPE_INSYNC, &sh->state);
3039 int *target = &sh->ops.target;
3041 sh->ops.target = -1;
3042 sh->ops.target2 = -1;
3043 sh->check_state = check_state_compute_run;
3044 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
3045 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
3046 if (sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) {
3047 set_bit(R5_Wantcompute,
3048 &sh->dev[pd_idx].flags);
3050 target = &sh->ops.target2;
3053 if (sh->ops.zero_sum_result & SUM_CHECK_Q_RESULT) {
3054 set_bit(R5_Wantcompute,
3055 &sh->dev[qd_idx].flags);
3062 case check_state_compute_run:
3065 printk(KERN_ERR "%s: unknown check_state: %d sector: %llu\n",
3066 __func__, sh->check_state,
3067 (unsigned long long) sh->sector);
3072 static void handle_stripe_expansion(struct r5conf *conf, struct stripe_head *sh)
3076 /* We have read all the blocks in this stripe and now we need to
3077 * copy some of them into a target stripe for expand.
3079 struct dma_async_tx_descriptor *tx = NULL;
3080 clear_bit(STRIPE_EXPAND_SOURCE, &sh->state);
3081 for (i = 0; i < sh->disks; i++)
3082 if (i != sh->pd_idx && i != sh->qd_idx) {
3084 struct stripe_head *sh2;
3085 struct async_submit_ctl submit;
3087 sector_t bn = compute_blocknr(sh, i, 1);
3088 sector_t s = raid5_compute_sector(conf, bn, 0,
3090 sh2 = get_active_stripe(conf, s, 0, 1, 1);
3092 /* so far only the early blocks of this stripe
3093 * have been requested. When later blocks
3094 * get requested, we will try again
3097 if (!test_bit(STRIPE_EXPANDING, &sh2->state) ||
3098 test_bit(R5_Expanded, &sh2->dev[dd_idx].flags)) {
3099 /* must have already done this block */
3100 release_stripe(sh2);
3104 /* place all the copies on one channel */
3105 init_async_submit(&submit, 0, tx, NULL, NULL, NULL);
3106 tx = async_memcpy(sh2->dev[dd_idx].page,
3107 sh->dev[i].page, 0, 0, STRIPE_SIZE,
3110 set_bit(R5_Expanded, &sh2->dev[dd_idx].flags);
3111 set_bit(R5_UPTODATE, &sh2->dev[dd_idx].flags);
3112 for (j = 0; j < conf->raid_disks; j++)
3113 if (j != sh2->pd_idx &&
3115 !test_bit(R5_Expanded, &sh2->dev[j].flags))
3117 if (j == conf->raid_disks) {
3118 set_bit(STRIPE_EXPAND_READY, &sh2->state);
3119 set_bit(STRIPE_HANDLE, &sh2->state);
3121 release_stripe(sh2);
3124 /* done submitting copies, wait for them to complete */
3127 dma_wait_for_async_tx(tx);
3132 * handle_stripe - do things to a stripe.
3134 * We lock the stripe by setting STRIPE_ACTIVE and then examine the
3135 * state of various bits to see what needs to be done.
3137 * return some read requests which now have data
3138 * return some write requests which are safely on storage
3139 * schedule a read on some buffers
3140 * schedule a write of some buffers
3141 * return confirmation of parity correctness
3145 static void analyse_stripe(struct stripe_head *sh, struct stripe_head_state *s)
3147 struct r5conf *conf = sh->raid_conf;
3148 int disks = sh->disks;
3151 int do_recovery = 0;
3153 memset(s, 0, sizeof(*s));
3155 s->expanding = test_bit(STRIPE_EXPAND_SOURCE, &sh->state);
3156 s->expanded = test_bit(STRIPE_EXPAND_READY, &sh->state);
3157 s->failed_num[0] = -1;
3158 s->failed_num[1] = -1;
3160 /* Now to look around and see what can be done */
3162 spin_lock_irq(&conf->device_lock);
3163 for (i=disks; i--; ) {
3164 struct md_rdev *rdev;
3171 pr_debug("check %d: state 0x%lx read %p write %p written %p\n",
3173 dev->toread, dev->towrite, dev->written);
3174 /* maybe we can reply to a read
3176 * new wantfill requests are only permitted while
3177 * ops_complete_biofill is guaranteed to be inactive
3179 if (test_bit(R5_UPTODATE, &dev->flags) && dev->toread &&
3180 !test_bit(STRIPE_BIOFILL_RUN, &sh->state))
3181 set_bit(R5_Wantfill, &dev->flags);
3183 /* now count some things */
3184 if (test_bit(R5_LOCKED, &dev->flags))
3186 if (test_bit(R5_UPTODATE, &dev->flags))
3188 if (test_bit(R5_Wantcompute, &dev->flags)) {
3190 BUG_ON(s->compute > 2);
3193 if (test_bit(R5_Wantfill, &dev->flags))
3195 else if (dev->toread)
3199 if (!test_bit(R5_OVERWRITE, &dev->flags))
3204 /* Prefer to use the replacement for reads, but only
3205 * if it is recovered enough and has no bad blocks.
3207 rdev = rcu_dereference(conf->disks[i].replacement);
3208 if (rdev && !test_bit(Faulty, &rdev->flags) &&
3209 rdev->recovery_offset >= sh->sector + STRIPE_SECTORS &&
3210 !is_badblock(rdev, sh->sector, STRIPE_SECTORS,
3211 &first_bad, &bad_sectors))
3212 set_bit(R5_ReadRepl, &dev->flags);
3215 set_bit(R5_NeedReplace, &dev->flags);
3216 rdev = rcu_dereference(conf->disks[i].rdev);
3217 clear_bit(R5_ReadRepl, &dev->flags);
3219 if (rdev && test_bit(Faulty, &rdev->flags))
3222 is_bad = is_badblock(rdev, sh->sector, STRIPE_SECTORS,
3223 &first_bad, &bad_sectors);
3224 if (s->blocked_rdev == NULL
3225 && (test_bit(Blocked, &rdev->flags)
3228 set_bit(BlockedBadBlocks,
3230 s->blocked_rdev = rdev;
3231 atomic_inc(&rdev->nr_pending);
3234 clear_bit(R5_Insync, &dev->flags);
3238 /* also not in-sync */
3239 if (!test_bit(WriteErrorSeen, &rdev->flags) &&
3240 test_bit(R5_UPTODATE, &dev->flags)) {
3241 /* treat as in-sync, but with a read error
3242 * which we can now try to correct
3244 set_bit(R5_Insync, &dev->flags);
3245 set_bit(R5_ReadError, &dev->flags);
3247 } else if (test_bit(In_sync, &rdev->flags))
3248 set_bit(R5_Insync, &dev->flags);
3249 else if (sh->sector + STRIPE_SECTORS <= rdev->recovery_offset)
3250 /* in sync if before recovery_offset */
3251 set_bit(R5_Insync, &dev->flags);
3252 else if (test_bit(R5_UPTODATE, &dev->flags) &&
3253 test_bit(R5_Expanded, &dev->flags))
3254 /* If we've reshaped into here, we assume it is Insync.
3255 * We will shortly update recovery_offset to make
3258 set_bit(R5_Insync, &dev->flags);
3260 if (rdev && test_bit(R5_WriteError, &dev->flags)) {
3261 /* This flag does not apply to '.replacement'
3262 * only to .rdev, so make sure to check that*/
3263 struct md_rdev *rdev2 = rcu_dereference(
3264 conf->disks[i].rdev);
3266 clear_bit(R5_Insync, &dev->flags);
3267 if (rdev2 && !test_bit(Faulty, &rdev2->flags)) {
3268 s->handle_bad_blocks = 1;
3269 atomic_inc(&rdev2->nr_pending);
3271 clear_bit(R5_WriteError, &dev->flags);
3273 if (rdev && test_bit(R5_MadeGood, &dev->flags)) {
3274 /* This flag does not apply to '.replacement'
3275 * only to .rdev, so make sure to check that*/
3276 struct md_rdev *rdev2 = rcu_dereference(
3277 conf->disks[i].rdev);
3278 if (rdev2 && !test_bit(Faulty, &rdev2->flags)) {
3279 s->handle_bad_blocks = 1;
3280 atomic_inc(&rdev2->nr_pending);
3282 clear_bit(R5_MadeGood, &dev->flags);
3284 if (test_bit(R5_MadeGoodRepl, &dev->flags)) {
3285 struct md_rdev *rdev2 = rcu_dereference(
3286 conf->disks[i].replacement);
3287 if (rdev2 && !test_bit(Faulty, &rdev2->flags)) {
3288 s->handle_bad_blocks = 1;
3289 atomic_inc(&rdev2->nr_pending);
3291 clear_bit(R5_MadeGoodRepl, &dev->flags);
3293 if (!test_bit(R5_Insync, &dev->flags)) {
3294 /* The ReadError flag will just be confusing now */
3295 clear_bit(R5_ReadError, &dev->flags);
3296 clear_bit(R5_ReWrite, &dev->flags);
3298 if (test_bit(R5_ReadError, &dev->flags))
3299 clear_bit(R5_Insync, &dev->flags);
3300 if (!test_bit(R5_Insync, &dev->flags)) {
3302 s->failed_num[s->failed] = i;
3304 if (rdev && !test_bit(Faulty, &rdev->flags))
3308 spin_unlock_irq(&conf->device_lock);
3309 if (test_bit(STRIPE_SYNCING, &sh->state)) {
3310 /* If there is a failed device being replaced,
3311 * we must be recovering.
3312 * else if we are after recovery_cp, we must be syncing
3313 * else if MD_RECOVERY_REQUESTED is set, we also are syncing.
3314 * else we can only be replacing
3315 * sync and recovery both need to read all devices, and so
3316 * use the same flag.
3319 sh->sector >= conf->mddev->recovery_cp ||
3320 test_bit(MD_RECOVERY_REQUESTED, &(conf->mddev->recovery)))
3328 static void handle_stripe(struct stripe_head *sh)
3330 struct stripe_head_state s;
3331 struct r5conf *conf = sh->raid_conf;
3334 int disks = sh->disks;
3335 struct r5dev *pdev, *qdev;
3337 clear_bit(STRIPE_HANDLE, &sh->state);
3338 if (test_and_set_bit_lock(STRIPE_ACTIVE, &sh->state)) {
3339 /* already being handled, ensure it gets handled
3340 * again when current action finishes */
3341 set_bit(STRIPE_HANDLE, &sh->state);
3345 if (test_and_clear_bit(STRIPE_SYNC_REQUESTED, &sh->state)) {
3346 set_bit(STRIPE_SYNCING, &sh->state);
3347 clear_bit(STRIPE_INSYNC, &sh->state);
3349 clear_bit(STRIPE_DELAYED, &sh->state);
3351 pr_debug("handling stripe %llu, state=%#lx cnt=%d, "
3352 "pd_idx=%d, qd_idx=%d\n, check:%d, reconstruct:%d\n",
3353 (unsigned long long)sh->sector, sh->state,
3354 atomic_read(&sh->count), sh->pd_idx, sh->qd_idx,
3355 sh->check_state, sh->reconstruct_state);
3357 analyse_stripe(sh, &s);
3359 if (s.handle_bad_blocks) {
3360 set_bit(STRIPE_HANDLE, &sh->state);
3364 if (unlikely(s.blocked_rdev)) {
3365 if (s.syncing || s.expanding || s.expanded ||
3366 s.replacing || s.to_write || s.written) {
3367 set_bit(STRIPE_HANDLE, &sh->state);
3370 /* There is nothing for the blocked_rdev to block */
3371 rdev_dec_pending(s.blocked_rdev, conf->mddev);
3372 s.blocked_rdev = NULL;
3375 if (s.to_fill && !test_bit(STRIPE_BIOFILL_RUN, &sh->state)) {
3376 set_bit(STRIPE_OP_BIOFILL, &s.ops_request);
3377 set_bit(STRIPE_BIOFILL_RUN, &sh->state);
3380 pr_debug("locked=%d uptodate=%d to_read=%d"
3381 " to_write=%d failed=%d failed_num=%d,%d\n",
3382 s.locked, s.uptodate, s.to_read, s.to_write, s.failed,
3383 s.failed_num[0], s.failed_num[1]);
3384 /* check if the array has lost more than max_degraded devices and,
3385 * if so, some requests might need to be failed.
3387 if (s.failed > conf->max_degraded) {
3388 sh->check_state = 0;
3389 sh->reconstruct_state = 0;
3390 if (s.to_read+s.to_write+s.written)
3391 handle_failed_stripe(conf, sh, &s, disks, &s.return_bi);
3392 if (s.syncing + s.replacing)
3393 handle_failed_sync(conf, sh, &s);
3397 * might be able to return some write requests if the parity blocks
3398 * are safe, or on a failed drive
3400 pdev = &sh->dev[sh->pd_idx];
3401 s.p_failed = (s.failed >= 1 && s.failed_num[0] == sh->pd_idx)
3402 || (s.failed >= 2 && s.failed_num[1] == sh->pd_idx);
3403 qdev = &sh->dev[sh->qd_idx];
3404 s.q_failed = (s.failed >= 1 && s.failed_num[0] == sh->qd_idx)
3405 || (s.failed >= 2 && s.failed_num[1] == sh->qd_idx)
3409 (s.p_failed || ((test_bit(R5_Insync, &pdev->flags)
3410 && !test_bit(R5_LOCKED, &pdev->flags)
3411 && test_bit(R5_UPTODATE, &pdev->flags)))) &&
3412 (s.q_failed || ((test_bit(R5_Insync, &qdev->flags)
3413 && !test_bit(R5_LOCKED, &qdev->flags)
3414 && test_bit(R5_UPTODATE, &qdev->flags)))))
3415 handle_stripe_clean_event(conf, sh, disks, &s.return_bi);
3417 /* Now we might consider reading some blocks, either to check/generate
3418 * parity, or to satisfy requests
3419 * or to load a block that is being partially written.
3421 if (s.to_read || s.non_overwrite
3422 || (conf->level == 6 && s.to_write && s.failed)
3423 || (s.syncing && (s.uptodate + s.compute < disks))
3426 handle_stripe_fill(sh, &s, disks);
3428 /* Now we check to see if any write operations have recently
3432 if (sh->reconstruct_state == reconstruct_state_prexor_drain_result)
3434 if (sh->reconstruct_state == reconstruct_state_drain_result ||
3435 sh->reconstruct_state == reconstruct_state_prexor_drain_result) {
3436 sh->reconstruct_state = reconstruct_state_idle;
3438 /* All the 'written' buffers and the parity block are ready to
3439 * be written back to disk
3441 BUG_ON(!test_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags));
3442 BUG_ON(sh->qd_idx >= 0 &&
3443 !test_bit(R5_UPTODATE, &sh->dev[sh->qd_idx].flags));
3444 for (i = disks; i--; ) {
3445 struct r5dev *dev = &sh->dev[i];
3446 if (test_bit(R5_LOCKED, &dev->flags) &&
3447 (i == sh->pd_idx || i == sh->qd_idx ||
3449 pr_debug("Writing block %d\n", i);
3450 set_bit(R5_Wantwrite, &dev->flags);
3453 if (!test_bit(R5_Insync, &dev->flags) ||
3454 ((i == sh->pd_idx || i == sh->qd_idx) &&
3456 set_bit(STRIPE_INSYNC, &sh->state);
3459 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
3460 s.dec_preread_active = 1;
3463 /* Now to consider new write requests and what else, if anything
3464 * should be read. We do not handle new writes when:
3465 * 1/ A 'write' operation (copy+xor) is already in flight.
3466 * 2/ A 'check' operation is in flight, as it may clobber the parity
3469 if (s.to_write && !sh->reconstruct_state && !sh->check_state)
3470 handle_stripe_dirtying(conf, sh, &s, disks);
3472 /* maybe we need to check and possibly fix the parity for this stripe
3473 * Any reads will already have been scheduled, so we just see if enough
3474 * data is available. The parity check is held off while parity
3475 * dependent operations are in flight.
3477 if (sh->check_state ||
3478 (s.syncing && s.locked == 0 &&
3479 !test_bit(STRIPE_COMPUTE_RUN, &sh->state) &&
3480 !test_bit(STRIPE_INSYNC, &sh->state))) {
3481 if (conf->level == 6)
3482 handle_parity_checks6(conf, sh, &s, disks);
3484 handle_parity_checks5(conf, sh, &s, disks);
3487 if (s.replacing && s.locked == 0
3488 && !test_bit(STRIPE_INSYNC, &sh->state)) {
3489 /* Write out to replacement devices where possible */
3490 for (i = 0; i < conf->raid_disks; i++)
3491 if (test_bit(R5_UPTODATE, &sh->dev[i].flags) &&
3492 test_bit(R5_NeedReplace, &sh->dev[i].flags)) {
3493 set_bit(R5_WantReplace, &sh->dev[i].flags);
3494 set_bit(R5_LOCKED, &sh->dev[i].flags);
3497 set_bit(STRIPE_INSYNC, &sh->state);
3499 if ((s.syncing || s.replacing) && s.locked == 0 &&
3500 test_bit(STRIPE_INSYNC, &sh->state)) {
3501 md_done_sync(conf->mddev, STRIPE_SECTORS, 1);
3502 clear_bit(STRIPE_SYNCING, &sh->state);
3505 /* If the failed drives are just a ReadError, then we might need
3506 * to progress the repair/check process
3508 if (s.failed <= conf->max_degraded && !conf->mddev->ro)
3509 for (i = 0; i < s.failed; i++) {
3510 struct r5dev *dev = &sh->dev[s.failed_num[i]];
3511 if (test_bit(R5_ReadError, &dev->flags)
3512 && !test_bit(R5_LOCKED, &dev->flags)
3513 && test_bit(R5_UPTODATE, &dev->flags)
3515 if (!test_bit(R5_ReWrite, &dev->flags)) {
3516 set_bit(R5_Wantwrite, &dev->flags);
3517 set_bit(R5_ReWrite, &dev->flags);
3518 set_bit(R5_LOCKED, &dev->flags);
3521 /* let's read it back */
3522 set_bit(R5_Wantread, &dev->flags);
3523 set_bit(R5_LOCKED, &dev->flags);
3530 /* Finish reconstruct operations initiated by the expansion process */
3531 if (sh->reconstruct_state == reconstruct_state_result) {
3532 struct stripe_head *sh_src
3533 = get_active_stripe(conf, sh->sector, 1, 1, 1);
3534 if (sh_src && test_bit(STRIPE_EXPAND_SOURCE, &sh_src->state)) {
3535 /* sh cannot be written until sh_src has been read.
3536 * so arrange for sh to be delayed a little
3538 set_bit(STRIPE_DELAYED, &sh->state);
3539 set_bit(STRIPE_HANDLE, &sh->state);
3540 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE,
3542 atomic_inc(&conf->preread_active_stripes);
3543 release_stripe(sh_src);
3547 release_stripe(sh_src);
3549 sh->reconstruct_state = reconstruct_state_idle;
3550 clear_bit(STRIPE_EXPANDING, &sh->state);
3551 for (i = conf->raid_disks; i--; ) {
3552 set_bit(R5_Wantwrite, &sh->dev[i].flags);
3553 set_bit(R5_LOCKED, &sh->dev[i].flags);
3558 if (s.expanded && test_bit(STRIPE_EXPANDING, &sh->state) &&
3559 !sh->reconstruct_state) {
3560 /* Need to write out all blocks after computing parity */
3561 sh->disks = conf->raid_disks;
3562 stripe_set_idx(sh->sector, conf, 0, sh);
3563 schedule_reconstruction(sh, &s, 1, 1);
3564 } else if (s.expanded && !sh->reconstruct_state && s.locked == 0) {
3565 clear_bit(STRIPE_EXPAND_READY, &sh->state);
3566 atomic_dec(&conf->reshape_stripes);
3567 wake_up(&conf->wait_for_overlap);
3568 md_done_sync(conf->mddev, STRIPE_SECTORS, 1);
3571 if (s.expanding && s.locked == 0 &&
3572 !test_bit(STRIPE_COMPUTE_RUN, &sh->state))
3573 handle_stripe_expansion(conf, sh);
3576 /* wait for this device to become unblocked */
3577 if (conf->mddev->external && unlikely(s.blocked_rdev))
3578 md_wait_for_blocked_rdev(s.blocked_rdev, conf->mddev);
3580 if (s.handle_bad_blocks)
3581 for (i = disks; i--; ) {
3582 struct md_rdev *rdev;
3583 struct r5dev *dev = &sh->dev[i];
3584 if (test_and_clear_bit(R5_WriteError, &dev->flags)) {
3585 /* We own a safe reference to the rdev */
3586 rdev = conf->disks[i].rdev;
3587 if (!rdev_set_badblocks(rdev, sh->sector,
3589 md_error(conf->mddev, rdev);
3590 rdev_dec_pending(rdev, conf->mddev);
3592 if (test_and_clear_bit(R5_MadeGood, &dev->flags)) {
3593 rdev = conf->disks[i].rdev;
3594 rdev_clear_badblocks(rdev, sh->sector,
3596 rdev_dec_pending(rdev, conf->mddev);
3598 if (test_and_clear_bit(R5_MadeGoodRepl, &dev->flags)) {
3599 rdev = conf->disks[i].replacement;
3601 /* rdev have been moved down */
3602 rdev = conf->disks[i].rdev;
3603 rdev_clear_badblocks(rdev, sh->sector,
3605 rdev_dec_pending(rdev, conf->mddev);
3610 raid_run_ops(sh, s.ops_request);
3614 if (s.dec_preread_active) {
3615 /* We delay this until after ops_run_io so that if make_request
3616 * is waiting on a flush, it won't continue until the writes
3617 * have actually been submitted.
3619 atomic_dec(&conf->preread_active_stripes);
3620 if (atomic_read(&conf->preread_active_stripes) <
3622 md_wakeup_thread(conf->mddev->thread);
3625 return_io(s.return_bi);
3627 clear_bit_unlock(STRIPE_ACTIVE, &sh->state);
3630 static void raid5_activate_delayed(struct r5conf *conf)
3632 if (atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD) {
3633 while (!list_empty(&conf->delayed_list)) {
3634 struct list_head *l = conf->delayed_list.next;
3635 struct stripe_head *sh;
3636 sh = list_entry(l, struct stripe_head, lru);
3638 clear_bit(STRIPE_DELAYED, &sh->state);
3639 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
3640 atomic_inc(&conf->preread_active_stripes);
3641 list_add_tail(&sh->lru, &conf->hold_list);
3646 static void activate_bit_delay(struct r5conf *conf)
3648 /* device_lock is held */
3649 struct list_head head;
3650 list_add(&head, &conf->bitmap_list);
3651 list_del_init(&conf->bitmap_list);
3652 while (!list_empty(&head)) {
3653 struct stripe_head *sh = list_entry(head.next, struct stripe_head, lru);
3654 list_del_init(&sh->lru);
3655 atomic_inc(&sh->count);
3656 __release_stripe(conf, sh);
3660 int md_raid5_congested(struct mddev *mddev, int bits)
3662 struct r5conf *conf = mddev->private;
3664 /* No difference between reads and writes. Just check
3665 * how busy the stripe_cache is
3668 if (conf->inactive_blocked)
3672 if (list_empty_careful(&conf->inactive_list))
3677 EXPORT_SYMBOL_GPL(md_raid5_congested);
3679 static int raid5_congested(void *data, int bits)
3681 struct mddev *mddev = data;
3683 return mddev_congested(mddev, bits) ||
3684 md_raid5_congested(mddev, bits);
3687 /* We want read requests to align with chunks where possible,
3688 * but write requests don't need to.
3690 static int raid5_mergeable_bvec(struct request_queue *q,
3691 struct bvec_merge_data *bvm,
3692 struct bio_vec *biovec)
3694 struct mddev *mddev = q->queuedata;
3695 sector_t sector = bvm->bi_sector + get_start_sect(bvm->bi_bdev);
3697 unsigned int chunk_sectors = mddev->chunk_sectors;
3698 unsigned int bio_sectors = bvm->bi_size >> 9;
3700 if ((bvm->bi_rw & 1) == WRITE)
3701 return biovec->bv_len; /* always allow writes to be mergeable */
3703 if (mddev->new_chunk_sectors < mddev->chunk_sectors)
3704 chunk_sectors = mddev->new_chunk_sectors;
3705 max = (chunk_sectors - ((sector & (chunk_sectors - 1)) + bio_sectors)) << 9;
3706 if (max < 0) max = 0;
3707 if (max <= biovec->bv_len && bio_sectors == 0)
3708 return biovec->bv_len;
3714 static int in_chunk_boundary(struct mddev *mddev, struct bio *bio)
3716 sector_t sector = bio->bi_sector + get_start_sect(bio->bi_bdev);
3717 unsigned int chunk_sectors = mddev->chunk_sectors;
3718 unsigned int bio_sectors = bio->bi_size >> 9;
3720 if (mddev->new_chunk_sectors < mddev->chunk_sectors)
3721 chunk_sectors = mddev->new_chunk_sectors;
3722 return chunk_sectors >=
3723 ((sector & (chunk_sectors - 1)) + bio_sectors);
3727 * add bio to the retry LIFO ( in O(1) ... we are in interrupt )
3728 * later sampled by raid5d.
3730 static void add_bio_to_retry(struct bio *bi,struct r5conf *conf)
3732 unsigned long flags;
3734 spin_lock_irqsave(&conf->device_lock, flags);
3736 bi->bi_next = conf->retry_read_aligned_list;
3737 conf->retry_read_aligned_list = bi;
3739 spin_unlock_irqrestore(&conf->device_lock, flags);
3740 md_wakeup_thread(conf->mddev->thread);
3744 static struct bio *remove_bio_from_retry(struct r5conf *conf)
3748 bi = conf->retry_read_aligned;
3750 conf->retry_read_aligned = NULL;
3753 bi = conf->retry_read_aligned_list;
3755 conf->retry_read_aligned_list = bi->bi_next;
3758 * this sets the active strip count to 1 and the processed
3759 * strip count to zero (upper 8 bits)
3761 bi->bi_phys_segments = 1; /* biased count of active stripes */
3769 * The "raid5_align_endio" should check if the read succeeded and if it
3770 * did, call bio_endio on the original bio (having bio_put the new bio
3772 * If the read failed..
3774 static void raid5_align_endio(struct bio *bi, int error)
3776 struct bio* raid_bi = bi->bi_private;
3777 struct mddev *mddev;
3778 struct r5conf *conf;
3779 int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
3780 struct md_rdev *rdev;
3784 rdev = (void*)raid_bi->bi_next;
3785 raid_bi->bi_next = NULL;
3786 mddev = rdev->mddev;
3787 conf = mddev->private;
3789 rdev_dec_pending(rdev, conf->mddev);
3791 if (!error && uptodate) {
3792 bio_endio(raid_bi, 0);
3793 if (atomic_dec_and_test(&conf->active_aligned_reads))
3794 wake_up(&conf->wait_for_stripe);
3799 pr_debug("raid5_align_endio : io error...handing IO for a retry\n");
3801 add_bio_to_retry(raid_bi, conf);
3804 static int bio_fits_rdev(struct bio *bi)
3806 struct request_queue *q = bdev_get_queue(bi->bi_bdev);
3808 if ((bi->bi_size>>9) > queue_max_sectors(q))
3810 blk_recount_segments(q, bi);
3811 if (bi->bi_phys_segments > queue_max_segments(q))
3814 if (q->merge_bvec_fn)
3815 /* it's too hard to apply the merge_bvec_fn at this stage,
3824 static int chunk_aligned_read(struct mddev *mddev, struct bio * raid_bio)
3826 struct r5conf *conf = mddev->private;
3828 struct bio* align_bi;
3829 struct md_rdev *rdev;
3830 sector_t end_sector;
3832 if (!in_chunk_boundary(mddev, raid_bio)) {
3833 pr_debug("chunk_aligned_read : non aligned\n");
3837 * use bio_clone_mddev to make a copy of the bio
3839 align_bi = bio_clone_mddev(raid_bio, GFP_NOIO, mddev);
3843 * set bi_end_io to a new function, and set bi_private to the
3846 align_bi->bi_end_io = raid5_align_endio;
3847 align_bi->bi_private = raid_bio;
3851 align_bi->bi_sector = raid5_compute_sector(conf, raid_bio->bi_sector,
3855 end_sector = align_bi->bi_sector + (align_bi->bi_size >> 9);
3857 rdev = rcu_dereference(conf->disks[dd_idx].replacement);
3858 if (!rdev || test_bit(Faulty, &rdev->flags) ||
3859 rdev->recovery_offset < end_sector) {
3860 rdev = rcu_dereference(conf->disks[dd_idx].rdev);
3862 (test_bit(Faulty, &rdev->flags) ||
3863 !(test_bit(In_sync, &rdev->flags) ||
3864 rdev->recovery_offset >= end_sector)))
3871 atomic_inc(&rdev->nr_pending);
3873 raid_bio->bi_next = (void*)rdev;
3874 align_bi->bi_bdev = rdev->bdev;
3875 align_bi->bi_flags &= ~(1 << BIO_SEG_VALID);
3876 /* No reshape active, so we can trust rdev->data_offset */
3877 align_bi->bi_sector += rdev->data_offset;
3879 if (!bio_fits_rdev(align_bi) ||
3880 is_badblock(rdev, align_bi->bi_sector, align_bi->bi_size>>9,
3881 &first_bad, &bad_sectors)) {
3882 /* too big in some way, or has a known bad block */
3884 rdev_dec_pending(rdev, mddev);
3888 spin_lock_irq(&conf->device_lock);
3889 wait_event_lock_irq(conf->wait_for_stripe,
3891 conf->device_lock, /* nothing */);
3892 atomic_inc(&conf->active_aligned_reads);
3893 spin_unlock_irq(&conf->device_lock);
3895 generic_make_request(align_bi);
3904 /* __get_priority_stripe - get the next stripe to process
3906 * Full stripe writes are allowed to pass preread active stripes up until
3907 * the bypass_threshold is exceeded. In general the bypass_count
3908 * increments when the handle_list is handled before the hold_list; however, it
3909 * will not be incremented when STRIPE_IO_STARTED is sampled set signifying a
3910 * stripe with in flight i/o. The bypass_count will be reset when the
3911 * head of the hold_list has changed, i.e. the head was promoted to the
3914 static struct stripe_head *__get_priority_stripe(struct r5conf *conf)
3916 struct stripe_head *sh;
3918 pr_debug("%s: handle: %s hold: %s full_writes: %d bypass_count: %d\n",
3920 list_empty(&conf->handle_list) ? "empty" : "busy",
3921 list_empty(&conf->hold_list) ? "empty" : "busy",
3922 atomic_read(&conf->pending_full_writes), conf->bypass_count);
3924 if (!list_empty(&conf->handle_list)) {
3925 sh = list_entry(conf->handle_list.next, typeof(*sh), lru);
3927 if (list_empty(&conf->hold_list))
3928 conf->bypass_count = 0;
3929 else if (!test_bit(STRIPE_IO_STARTED, &sh->state)) {
3930 if (conf->hold_list.next == conf->last_hold)
3931 conf->bypass_count++;
3933 conf->last_hold = conf->hold_list.next;
3934 conf->bypass_count -= conf->bypass_threshold;
3935 if (conf->bypass_count < 0)
3936 conf->bypass_count = 0;
3939 } else if (!list_empty(&conf->hold_list) &&
3940 ((conf->bypass_threshold &&
3941 conf->bypass_count > conf->bypass_threshold) ||
3942 atomic_read(&conf->pending_full_writes) == 0)) {
3943 sh = list_entry(conf->hold_list.next,
3945 conf->bypass_count -= conf->bypass_threshold;
3946 if (conf->bypass_count < 0)
3947 conf->bypass_count = 0;
3951 list_del_init(&sh->lru);
3952 atomic_inc(&sh->count);
3953 BUG_ON(atomic_read(&sh->count) != 1);
3957 static void make_request(struct mddev *mddev, struct bio * bi)
3959 struct r5conf *conf = mddev->private;
3961 sector_t new_sector;
3962 sector_t logical_sector, last_sector;
3963 struct stripe_head *sh;
3964 const int rw = bio_data_dir(bi);
3968 if (unlikely(bi->bi_rw & REQ_FLUSH)) {
3969 md_flush_request(mddev, bi);
3973 md_write_start(mddev, bi);
3976 mddev->reshape_position == MaxSector &&
3977 chunk_aligned_read(mddev,bi))
3980 logical_sector = bi->bi_sector & ~((sector_t)STRIPE_SECTORS-1);
3981 last_sector = bi->bi_sector + (bi->bi_size>>9);
3983 bi->bi_phys_segments = 1; /* over-loaded to count active stripes */
3985 plugged = mddev_check_plugged(mddev);
3986 for (;logical_sector < last_sector; logical_sector += STRIPE_SECTORS) {
3988 int disks, data_disks;
3993 disks = conf->raid_disks;
3994 prepare_to_wait(&conf->wait_for_overlap, &w, TASK_UNINTERRUPTIBLE);
3995 if (unlikely(conf->reshape_progress != MaxSector)) {
3996 /* spinlock is needed as reshape_progress may be
3997 * 64bit on a 32bit platform, and so it might be
3998 * possible to see a half-updated value
3999 * Of course reshape_progress could change after
4000 * the lock is dropped, so once we get a reference
4001 * to the stripe that we think it is, we will have
4004 spin_lock_irq(&conf->device_lock);
4005 if (mddev->reshape_backwards
4006 ? logical_sector < conf->reshape_progress
4007 : logical_sector >= conf->reshape_progress) {
4008 disks = conf->previous_raid_disks;
4011 if (mddev->reshape_backwards
4012 ? logical_sector < conf->reshape_safe
4013 : logical_sector >= conf->reshape_safe) {
4014 spin_unlock_irq(&conf->device_lock);
4019 spin_unlock_irq(&conf->device_lock);
4021 data_disks = disks - conf->max_degraded;
4023 new_sector = raid5_compute_sector(conf, logical_sector,
4026 pr_debug("raid456: make_request, sector %llu logical %llu\n",
4027 (unsigned long long)new_sector,
4028 (unsigned long long)logical_sector);
4030 sh = get_active_stripe(conf, new_sector, previous,
4031 (bi->bi_rw&RWA_MASK), 0);
4033 if (unlikely(previous)) {
4034 /* expansion might have moved on while waiting for a
4035 * stripe, so we must do the range check again.
4036 * Expansion could still move past after this
4037 * test, but as we are holding a reference to
4038 * 'sh', we know that if that happens,
4039 * STRIPE_EXPANDING will get set and the expansion
4040 * won't proceed until we finish with the stripe.
4043 spin_lock_irq(&conf->device_lock);
4044 if (mddev->reshape_backwards
4045 ? logical_sector >= conf->reshape_progress
4046 : logical_sector < conf->reshape_progress)
4047 /* mismatch, need to try again */
4049 spin_unlock_irq(&conf->device_lock);
4058 logical_sector >= mddev->suspend_lo &&
4059 logical_sector < mddev->suspend_hi) {
4061 /* As the suspend_* range is controlled by
4062 * userspace, we want an interruptible
4065 flush_signals(current);
4066 prepare_to_wait(&conf->wait_for_overlap,
4067 &w, TASK_INTERRUPTIBLE);
4068 if (logical_sector >= mddev->suspend_lo &&
4069 logical_sector < mddev->suspend_hi)
4074 if (test_bit(STRIPE_EXPANDING, &sh->state) ||
4075 !add_stripe_bio(sh, bi, dd_idx, rw)) {
4076 /* Stripe is busy expanding or
4077 * add failed due to overlap. Flush everything
4080 md_wakeup_thread(mddev->thread);
4085 finish_wait(&conf->wait_for_overlap, &w);
4086 set_bit(STRIPE_HANDLE, &sh->state);
4087 clear_bit(STRIPE_DELAYED, &sh->state);
4088 if ((bi->bi_rw & REQ_SYNC) &&
4089 !test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
4090 atomic_inc(&conf->preread_active_stripes);
4093 /* cannot get stripe for read-ahead, just give-up */
4094 clear_bit(BIO_UPTODATE, &bi->bi_flags);
4095 finish_wait(&conf->wait_for_overlap, &w);
4101 md_wakeup_thread(mddev->thread);
4103 spin_lock_irq(&conf->device_lock);
4104 remaining = raid5_dec_bi_phys_segments(bi);
4105 spin_unlock_irq(&conf->device_lock);
4106 if (remaining == 0) {
4109 md_write_end(mddev);
4115 static sector_t raid5_size(struct mddev *mddev, sector_t sectors, int raid_disks);
4117 static sector_t reshape_request(struct mddev *mddev, sector_t sector_nr, int *skipped)
4119 /* reshaping is quite different to recovery/resync so it is
4120 * handled quite separately ... here.
4122 * On each call to sync_request, we gather one chunk worth of
4123 * destination stripes and flag them as expanding.
4124 * Then we find all the source stripes and request reads.
4125 * As the reads complete, handle_stripe will copy the data
4126 * into the destination stripe and release that stripe.
4128 struct r5conf *conf = mddev->private;
4129 struct stripe_head *sh;
4130 sector_t first_sector, last_sector;
4131 int raid_disks = conf->previous_raid_disks;
4132 int data_disks = raid_disks - conf->max_degraded;
4133 int new_data_disks = conf->raid_disks - conf->max_degraded;
4136 sector_t writepos, readpos, safepos;
4137 sector_t stripe_addr;
4138 int reshape_sectors;
4139 struct list_head stripes;
4141 if (sector_nr == 0) {
4142 /* If restarting in the middle, skip the initial sectors */
4143 if (mddev->reshape_backwards &&
4144 conf->reshape_progress < raid5_size(mddev, 0, 0)) {
4145 sector_nr = raid5_size(mddev, 0, 0)
4146 - conf->reshape_progress;
4147 } else if (!mddev->reshape_backwards &&
4148 conf->reshape_progress > 0)
4149 sector_nr = conf->reshape_progress;
4150 sector_div(sector_nr, new_data_disks);
4152 mddev->curr_resync_completed = sector_nr;
4153 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
4159 /* We need to process a full chunk at a time.
4160 * If old and new chunk sizes differ, we need to process the
4163 if (mddev->new_chunk_sectors > mddev->chunk_sectors)
4164 reshape_sectors = mddev->new_chunk_sectors;
4166 reshape_sectors = mddev->chunk_sectors;
4168 /* we update the metadata when there is more than 3Meg
4169 * in the block range (that is rather arbitrary, should
4170 * probably be time based) or when the data about to be
4171 * copied would over-write the source of the data at
4172 * the front of the range.
4173 * i.e. one new_stripe along from reshape_progress new_maps
4174 * to after where reshape_safe old_maps to
4176 writepos = conf->reshape_progress;
4177 sector_div(writepos, new_data_disks);
4178 readpos = conf->reshape_progress;
4179 sector_div(readpos, data_disks);
4180 safepos = conf->reshape_safe;
4181 sector_div(safepos, data_disks);
4182 if (mddev->reshape_backwards) {
4183 writepos -= min_t(sector_t, reshape_sectors, writepos);
4184 readpos += reshape_sectors;
4185 safepos += reshape_sectors;
4187 writepos += reshape_sectors;
4188 readpos -= min_t(sector_t, reshape_sectors, readpos);
4189 safepos -= min_t(sector_t, reshape_sectors, safepos);
4192 /* 'writepos' is the most advanced device address we might write.
4193 * 'readpos' is the least advanced device address we might read.
4194 * 'safepos' is the least address recorded in the metadata as having
4196 * If 'readpos' is behind 'writepos', then there is no way that we can
4197 * ensure safety in the face of a crash - that must be done by userspace
4198 * making a backup of the data. So in that case there is no particular
4199 * rush to update metadata.
4200 * Otherwise if 'safepos' is behind 'writepos', then we really need to
4201 * update the metadata to advance 'safepos' to match 'readpos' so that
4202 * we can be safe in the event of a crash.
4203 * So we insist on updating metadata if safepos is behind writepos and
4204 * readpos is beyond writepos.
4205 * In any case, update the metadata every 10 seconds.
4206 * Maybe that number should be configurable, but I'm not sure it is
4207 * worth it.... maybe it could be a multiple of safemode_delay???
4209 if ((mddev->reshape_backwards
4210 ? (safepos > writepos && readpos < writepos)
4211 : (safepos < writepos && readpos > writepos)) ||
4212 time_after(jiffies, conf->reshape_checkpoint + 10*HZ)) {
4213 /* Cannot proceed until we've updated the superblock... */
4214 wait_event(conf->wait_for_overlap,
4215 atomic_read(&conf->reshape_stripes)==0);
4216 mddev->reshape_position = conf->reshape_progress;
4217 mddev->curr_resync_completed = sector_nr;
4218 conf->reshape_checkpoint = jiffies;
4219 set_bit(MD_CHANGE_DEVS, &mddev->flags);
4220 md_wakeup_thread(mddev->thread);
4221 wait_event(mddev->sb_wait, mddev->flags == 0 ||
4222 kthread_should_stop());
4223 spin_lock_irq(&conf->device_lock);
4224 conf->reshape_safe = mddev->reshape_position;
4225 spin_unlock_irq(&conf->device_lock);
4226 wake_up(&conf->wait_for_overlap);
4227 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
4230 if (mddev->reshape_backwards) {
4231 BUG_ON(conf->reshape_progress == 0);
4232 stripe_addr = writepos;
4233 BUG_ON((mddev->dev_sectors &
4234 ~((sector_t)reshape_sectors - 1))
4235 - reshape_sectors - stripe_addr
4238 BUG_ON(writepos != sector_nr + reshape_sectors);
4239 stripe_addr = sector_nr;
4241 INIT_LIST_HEAD(&stripes);
4242 for (i = 0; i < reshape_sectors; i += STRIPE_SECTORS) {
4244 int skipped_disk = 0;
4245 sh = get_active_stripe(conf, stripe_addr+i, 0, 0, 1);
4246 set_bit(STRIPE_EXPANDING, &sh->state);
4247 atomic_inc(&conf->reshape_stripes);
4248 /* If any of this stripe is beyond the end of the old
4249 * array, then we need to zero those blocks
4251 for (j=sh->disks; j--;) {
4253 if (j == sh->pd_idx)
4255 if (conf->level == 6 &&
4258 s = compute_blocknr(sh, j, 0);
4259 if (s < raid5_size(mddev, 0, 0)) {
4263 memset(page_address(sh->dev[j].page), 0, STRIPE_SIZE);
4264 set_bit(R5_Expanded, &sh->dev[j].flags);
4265 set_bit(R5_UPTODATE, &sh->dev[j].flags);
4267 if (!skipped_disk) {
4268 set_bit(STRIPE_EXPAND_READY, &sh->state);
4269 set_bit(STRIPE_HANDLE, &sh->state);
4271 list_add(&sh->lru, &stripes);
4273 spin_lock_irq(&conf->device_lock);
4274 if (mddev->reshape_backwards)
4275 conf->reshape_progress -= reshape_sectors * new_data_disks;
4277 conf->reshape_progress += reshape_sectors * new_data_disks;
4278 spin_unlock_irq(&conf->device_lock);
4279 /* Ok, those stripe are ready. We can start scheduling
4280 * reads on the source stripes.
4281 * The source stripes are determined by mapping the first and last
4282 * block on the destination stripes.
4285 raid5_compute_sector(conf, stripe_addr*(new_data_disks),
4288 raid5_compute_sector(conf, ((stripe_addr+reshape_sectors)
4289 * new_data_disks - 1),
4291 if (last_sector >= mddev->dev_sectors)
4292 last_sector = mddev->dev_sectors - 1;
4293 while (first_sector <= last_sector) {
4294 sh = get_active_stripe(conf, first_sector, 1, 0, 1);
4295 set_bit(STRIPE_EXPAND_SOURCE, &sh->state);
4296 set_bit(STRIPE_HANDLE, &sh->state);
4298 first_sector += STRIPE_SECTORS;
4300 /* Now that the sources are clearly marked, we can release
4301 * the destination stripes
4303 while (!list_empty(&stripes)) {
4304 sh = list_entry(stripes.next, struct stripe_head, lru);
4305 list_del_init(&sh->lru);
4308 /* If this takes us to the resync_max point where we have to pause,
4309 * then we need to write out the superblock.
4311 sector_nr += reshape_sectors;
4312 if ((sector_nr - mddev->curr_resync_completed) * 2
4313 >= mddev->resync_max - mddev->curr_resync_completed) {
4314 /* Cannot proceed until we've updated the superblock... */
4315 wait_event(conf->wait_for_overlap,
4316 atomic_read(&conf->reshape_stripes) == 0);
4317 mddev->reshape_position = conf->reshape_progress;
4318 mddev->curr_resync_completed = sector_nr;
4319 conf->reshape_checkpoint = jiffies;
4320 set_bit(MD_CHANGE_DEVS, &mddev->flags);
4321 md_wakeup_thread(mddev->thread);
4322 wait_event(mddev->sb_wait,
4323 !test_bit(MD_CHANGE_DEVS, &mddev->flags)
4324 || kthread_should_stop());
4325 spin_lock_irq(&conf->device_lock);
4326 conf->reshape_safe = mddev->reshape_position;
4327 spin_unlock_irq(&conf->device_lock);
4328 wake_up(&conf->wait_for_overlap);
4329 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
4331 return reshape_sectors;
4334 /* FIXME go_faster isn't used */
4335 static inline sector_t sync_request(struct mddev *mddev, sector_t sector_nr, int *skipped, int go_faster)
4337 struct r5conf *conf = mddev->private;
4338 struct stripe_head *sh;
4339 sector_t max_sector = mddev->dev_sectors;
4340 sector_t sync_blocks;
4341 int still_degraded = 0;
4344 if (sector_nr >= max_sector) {
4345 /* just being told to finish up .. nothing much to do */
4347 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery)) {
4352 if (mddev->curr_resync < max_sector) /* aborted */
4353 bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
4355 else /* completed sync */
4357 bitmap_close_sync(mddev->bitmap);
4362 /* Allow raid5_quiesce to complete */
4363 wait_event(conf->wait_for_overlap, conf->quiesce != 2);
4365 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
4366 return reshape_request(mddev, sector_nr, skipped);
4368 /* No need to check resync_max as we never do more than one
4369 * stripe, and as resync_max will always be on a chunk boundary,
4370 * if the check in md_do_sync didn't fire, there is no chance
4371 * of overstepping resync_max here
4374 /* if there is too many failed drives and we are trying
4375 * to resync, then assert that we are finished, because there is
4376 * nothing we can do.
4378 if (mddev->degraded >= conf->max_degraded &&
4379 test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
4380 sector_t rv = mddev->dev_sectors - sector_nr;
4384 if (!bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
4385 !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
4386 !conf->fullsync && sync_blocks >= STRIPE_SECTORS) {
4387 /* we can skip this block, and probably more */
4388 sync_blocks /= STRIPE_SECTORS;
4390 return sync_blocks * STRIPE_SECTORS; /* keep things rounded to whole stripes */
4393 bitmap_cond_end_sync(mddev->bitmap, sector_nr);
4395 sh = get_active_stripe(conf, sector_nr, 0, 1, 0);
4397 sh = get_active_stripe(conf, sector_nr, 0, 0, 0);
4398 /* make sure we don't swamp the stripe cache if someone else
4399 * is trying to get access
4401 schedule_timeout_uninterruptible(1);
4403 /* Need to check if array will still be degraded after recovery/resync
4404 * We don't need to check the 'failed' flag as when that gets set,
4407 for (i = 0; i < conf->raid_disks; i++)
4408 if (conf->disks[i].rdev == NULL)
4411 bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, still_degraded);
4413 set_bit(STRIPE_SYNC_REQUESTED, &sh->state);
4418 return STRIPE_SECTORS;
4421 static int retry_aligned_read(struct r5conf *conf, struct bio *raid_bio)
4423 /* We may not be able to submit a whole bio at once as there
4424 * may not be enough stripe_heads available.
4425 * We cannot pre-allocate enough stripe_heads as we may need
4426 * more than exist in the cache (if we allow ever large chunks).
4427 * So we do one stripe head at a time and record in
4428 * ->bi_hw_segments how many have been done.
4430 * We *know* that this entire raid_bio is in one chunk, so
4431 * it will be only one 'dd_idx' and only need one call to raid5_compute_sector.
4433 struct stripe_head *sh;
4435 sector_t sector, logical_sector, last_sector;
4440 logical_sector = raid_bio->bi_sector & ~((sector_t)STRIPE_SECTORS-1);
4441 sector = raid5_compute_sector(conf, logical_sector,
4443 last_sector = raid_bio->bi_sector + (raid_bio->bi_size>>9);
4445 for (; logical_sector < last_sector;
4446 logical_sector += STRIPE_SECTORS,
4447 sector += STRIPE_SECTORS,
4450 if (scnt < raid5_bi_hw_segments(raid_bio))
4451 /* already done this stripe */
4454 sh = get_active_stripe(conf, sector, 0, 1, 0);
4457 /* failed to get a stripe - must wait */
4458 raid5_set_bi_hw_segments(raid_bio, scnt);
4459 conf->retry_read_aligned = raid_bio;
4463 if (!add_stripe_bio(sh, raid_bio, dd_idx, 0)) {
4465 raid5_set_bi_hw_segments(raid_bio, scnt);
4466 conf->retry_read_aligned = raid_bio;
4474 spin_lock_irq(&conf->device_lock);
4475 remaining = raid5_dec_bi_phys_segments(raid_bio);
4476 spin_unlock_irq(&conf->device_lock);
4478 bio_endio(raid_bio, 0);
4479 if (atomic_dec_and_test(&conf->active_aligned_reads))
4480 wake_up(&conf->wait_for_stripe);
4486 * This is our raid5 kernel thread.
4488 * We scan the hash table for stripes which can be handled now.
4489 * During the scan, completed stripes are saved for us by the interrupt
4490 * handler, so that they will not have to wait for our next wakeup.
4492 static void raid5d(struct mddev *mddev)
4494 struct stripe_head *sh;
4495 struct r5conf *conf = mddev->private;
4497 struct blk_plug plug;
4499 pr_debug("+++ raid5d active\n");
4501 md_check_recovery(mddev);
4503 blk_start_plug(&plug);
4505 spin_lock_irq(&conf->device_lock);
4509 if (atomic_read(&mddev->plug_cnt) == 0 &&
4510 !list_empty(&conf->bitmap_list)) {
4511 /* Now is a good time to flush some bitmap updates */
4513 spin_unlock_irq(&conf->device_lock);
4514 bitmap_unplug(mddev->bitmap);
4515 spin_lock_irq(&conf->device_lock);
4516 conf->seq_write = conf->seq_flush;
4517 activate_bit_delay(conf);
4519 if (atomic_read(&mddev->plug_cnt) == 0)
4520 raid5_activate_delayed(conf);
4522 while ((bio = remove_bio_from_retry(conf))) {
4524 spin_unlock_irq(&conf->device_lock);
4525 ok = retry_aligned_read(conf, bio);
4526 spin_lock_irq(&conf->device_lock);
4532 sh = __get_priority_stripe(conf);
4536 spin_unlock_irq(&conf->device_lock);
4543 if (mddev->flags & ~(1<<MD_CHANGE_PENDING))
4544 md_check_recovery(mddev);
4546 spin_lock_irq(&conf->device_lock);
4548 pr_debug("%d stripes handled\n", handled);
4550 spin_unlock_irq(&conf->device_lock);
4552 async_tx_issue_pending_all();
4553 blk_finish_plug(&plug);
4555 pr_debug("--- raid5d inactive\n");
4559 raid5_show_stripe_cache_size(struct mddev *mddev, char *page)
4561 struct r5conf *conf = mddev->private;
4563 return sprintf(page, "%d\n", conf->max_nr_stripes);
4569 raid5_set_cache_size(struct mddev *mddev, int size)
4571 struct r5conf *conf = mddev->private;
4574 if (size <= 16 || size > 32768)
4576 while (size < conf->max_nr_stripes) {
4577 if (drop_one_stripe(conf))
4578 conf->max_nr_stripes--;
4582 err = md_allow_write(mddev);
4585 while (size > conf->max_nr_stripes) {
4586 if (grow_one_stripe(conf))
4587 conf->max_nr_stripes++;
4592 EXPORT_SYMBOL(raid5_set_cache_size);
4595 raid5_store_stripe_cache_size(struct mddev *mddev, const char *page, size_t len)
4597 struct r5conf *conf = mddev->private;
4601 if (len >= PAGE_SIZE)
4606 if (strict_strtoul(page, 10, &new))
4608 err = raid5_set_cache_size(mddev, new);
4614 static struct md_sysfs_entry
4615 raid5_stripecache_size = __ATTR(stripe_cache_size, S_IRUGO | S_IWUSR,
4616 raid5_show_stripe_cache_size,
4617 raid5_store_stripe_cache_size);
4620 raid5_show_preread_threshold(struct mddev *mddev, char *page)
4622 struct r5conf *conf = mddev->private;
4624 return sprintf(page, "%d\n", conf->bypass_threshold);
4630 raid5_store_preread_threshold(struct mddev *mddev, const char *page, size_t len)
4632 struct r5conf *conf = mddev->private;
4634 if (len >= PAGE_SIZE)
4639 if (strict_strtoul(page, 10, &new))
4641 if (new > conf->max_nr_stripes)
4643 conf->bypass_threshold = new;
4647 static struct md_sysfs_entry
4648 raid5_preread_bypass_threshold = __ATTR(preread_bypass_threshold,
4650 raid5_show_preread_threshold,
4651 raid5_store_preread_threshold);
4654 stripe_cache_active_show(struct mddev *mddev, char *page)
4656 struct r5conf *conf = mddev->private;
4658 return sprintf(page, "%d\n", atomic_read(&conf->active_stripes));
4663 static struct md_sysfs_entry
4664 raid5_stripecache_active = __ATTR_RO(stripe_cache_active);
4666 static struct attribute *raid5_attrs[] = {
4667 &raid5_stripecache_size.attr,
4668 &raid5_stripecache_active.attr,
4669 &raid5_preread_bypass_threshold.attr,
4672 static struct attribute_group raid5_attrs_group = {
4674 .attrs = raid5_attrs,
4678 raid5_size(struct mddev *mddev, sector_t sectors, int raid_disks)
4680 struct r5conf *conf = mddev->private;
4683 sectors = mddev->dev_sectors;
4685 /* size is defined by the smallest of previous and new size */
4686 raid_disks = min(conf->raid_disks, conf->previous_raid_disks);
4688 sectors &= ~((sector_t)mddev->chunk_sectors - 1);
4689 sectors &= ~((sector_t)mddev->new_chunk_sectors - 1);
4690 return sectors * (raid_disks - conf->max_degraded);
4693 static void raid5_free_percpu(struct r5conf *conf)
4695 struct raid5_percpu *percpu;
4702 for_each_possible_cpu(cpu) {
4703 percpu = per_cpu_ptr(conf->percpu, cpu);
4704 safe_put_page(percpu->spare_page);
4705 kfree(percpu->scribble);
4707 #ifdef CONFIG_HOTPLUG_CPU
4708 unregister_cpu_notifier(&conf->cpu_notify);
4712 free_percpu(conf->percpu);
4715 static void free_conf(struct r5conf *conf)
4717 shrink_stripes(conf);
4718 raid5_free_percpu(conf);
4720 kfree(conf->stripe_hashtbl);
4724 #ifdef CONFIG_HOTPLUG_CPU
4725 static int raid456_cpu_notify(struct notifier_block *nfb, unsigned long action,
4728 struct r5conf *conf = container_of(nfb, struct r5conf, cpu_notify);
4729 long cpu = (long)hcpu;
4730 struct raid5_percpu *percpu = per_cpu_ptr(conf->percpu, cpu);
4733 case CPU_UP_PREPARE:
4734 case CPU_UP_PREPARE_FROZEN:
4735 if (conf->level == 6 && !percpu->spare_page)
4736 percpu->spare_page = alloc_page(GFP_KERNEL);
4737 if (!percpu->scribble)
4738 percpu->scribble = kmalloc(conf->scribble_len, GFP_KERNEL);
4740 if (!percpu->scribble ||
4741 (conf->level == 6 && !percpu->spare_page)) {
4742 safe_put_page(percpu->spare_page);
4743 kfree(percpu->scribble);
4744 pr_err("%s: failed memory allocation for cpu%ld\n",
4746 return notifier_from_errno(-ENOMEM);
4750 case CPU_DEAD_FROZEN:
4751 safe_put_page(percpu->spare_page);
4752 kfree(percpu->scribble);
4753 percpu->spare_page = NULL;
4754 percpu->scribble = NULL;
4763 static int raid5_alloc_percpu(struct r5conf *conf)
4766 struct page *spare_page;
4767 struct raid5_percpu __percpu *allcpus;
4771 allcpus = alloc_percpu(struct raid5_percpu);
4774 conf->percpu = allcpus;
4778 for_each_present_cpu(cpu) {
4779 if (conf->level == 6) {
4780 spare_page = alloc_page(GFP_KERNEL);
4785 per_cpu_ptr(conf->percpu, cpu)->spare_page = spare_page;
4787 scribble = kmalloc(conf->scribble_len, GFP_KERNEL);
4792 per_cpu_ptr(conf->percpu, cpu)->scribble = scribble;
4794 #ifdef CONFIG_HOTPLUG_CPU
4795 conf->cpu_notify.notifier_call = raid456_cpu_notify;
4796 conf->cpu_notify.priority = 0;
4798 err = register_cpu_notifier(&conf->cpu_notify);
4805 static struct r5conf *setup_conf(struct mddev *mddev)
4807 struct r5conf *conf;
4808 int raid_disk, memory, max_disks;
4809 struct md_rdev *rdev;
4810 struct disk_info *disk;
4812 if (mddev->new_level != 5
4813 && mddev->new_level != 4
4814 && mddev->new_level != 6) {
4815 printk(KERN_ERR "md/raid:%s: raid level not set to 4/5/6 (%d)\n",
4816 mdname(mddev), mddev->new_level);
4817 return ERR_PTR(-EIO);
4819 if ((mddev->new_level == 5
4820 && !algorithm_valid_raid5(mddev->new_layout)) ||
4821 (mddev->new_level == 6
4822 && !algorithm_valid_raid6(mddev->new_layout))) {
4823 printk(KERN_ERR "md/raid:%s: layout %d not supported\n",
4824 mdname(mddev), mddev->new_layout);
4825 return ERR_PTR(-EIO);
4827 if (mddev->new_level == 6 && mddev->raid_disks < 4) {
4828 printk(KERN_ERR "md/raid:%s: not enough configured devices (%d, minimum 4)\n",
4829 mdname(mddev), mddev->raid_disks);
4830 return ERR_PTR(-EINVAL);
4833 if (!mddev->new_chunk_sectors ||
4834 (mddev->new_chunk_sectors << 9) % PAGE_SIZE ||
4835 !is_power_of_2(mddev->new_chunk_sectors)) {
4836 printk(KERN_ERR "md/raid:%s: invalid chunk size %d\n",
4837 mdname(mddev), mddev->new_chunk_sectors << 9);
4838 return ERR_PTR(-EINVAL);
4841 conf = kzalloc(sizeof(struct r5conf), GFP_KERNEL);
4844 spin_lock_init(&conf->device_lock);
4845 init_waitqueue_head(&conf->wait_for_stripe);
4846 init_waitqueue_head(&conf->wait_for_overlap);
4847 INIT_LIST_HEAD(&conf->handle_list);
4848 INIT_LIST_HEAD(&conf->hold_list);
4849 INIT_LIST_HEAD(&conf->delayed_list);
4850 INIT_LIST_HEAD(&conf->bitmap_list);
4851 INIT_LIST_HEAD(&conf->inactive_list);
4852 atomic_set(&conf->active_stripes, 0);
4853 atomic_set(&conf->preread_active_stripes, 0);
4854 atomic_set(&conf->active_aligned_reads, 0);
4855 conf->bypass_threshold = BYPASS_THRESHOLD;
4856 conf->recovery_disabled = mddev->recovery_disabled - 1;
4858 conf->raid_disks = mddev->raid_disks;
4859 if (mddev->reshape_position == MaxSector)
4860 conf->previous_raid_disks = mddev->raid_disks;
4862 conf->previous_raid_disks = mddev->raid_disks - mddev->delta_disks;
4863 max_disks = max(conf->raid_disks, conf->previous_raid_disks);
4864 conf->scribble_len = scribble_len(max_disks);
4866 conf->disks = kzalloc(max_disks * sizeof(struct disk_info),
4871 conf->mddev = mddev;
4873 if ((conf->stripe_hashtbl = kzalloc(PAGE_SIZE, GFP_KERNEL)) == NULL)
4876 conf->level = mddev->new_level;
4877 if (raid5_alloc_percpu(conf) != 0)
4880 pr_debug("raid456: run(%s) called.\n", mdname(mddev));
4882 rdev_for_each(rdev, mddev) {
4883 raid_disk = rdev->raid_disk;
4884 if (raid_disk >= max_disks
4887 disk = conf->disks + raid_disk;
4889 if (test_bit(Replacement, &rdev->flags)) {
4890 if (disk->replacement)
4892 disk->replacement = rdev;
4899 if (test_bit(In_sync, &rdev->flags)) {
4900 char b[BDEVNAME_SIZE];
4901 printk(KERN_INFO "md/raid:%s: device %s operational as raid"
4903 mdname(mddev), bdevname(rdev->bdev, b), raid_disk);
4904 } else if (rdev->saved_raid_disk != raid_disk)
4905 /* Cannot rely on bitmap to complete recovery */
4909 conf->chunk_sectors = mddev->new_chunk_sectors;
4910 conf->level = mddev->new_level;
4911 if (conf->level == 6)
4912 conf->max_degraded = 2;
4914 conf->max_degraded = 1;
4915 conf->algorithm = mddev->new_layout;
4916 conf->max_nr_stripes = NR_STRIPES;
4917 conf->reshape_progress = mddev->reshape_position;
4918 if (conf->reshape_progress != MaxSector) {
4919 conf->prev_chunk_sectors = mddev->chunk_sectors;
4920 conf->prev_algo = mddev->layout;
4923 memory = conf->max_nr_stripes * (sizeof(struct stripe_head) +
4924 max_disks * ((sizeof(struct bio) + PAGE_SIZE))) / 1024;
4925 if (grow_stripes(conf, conf->max_nr_stripes)) {
4927 "md/raid:%s: couldn't allocate %dkB for buffers\n",
4928 mdname(mddev), memory);
4931 printk(KERN_INFO "md/raid:%s: allocated %dkB\n",
4932 mdname(mddev), memory);
4934 conf->thread = md_register_thread(raid5d, mddev, NULL);
4935 if (!conf->thread) {
4937 "md/raid:%s: couldn't allocate thread.\n",
4947 return ERR_PTR(-EIO);
4949 return ERR_PTR(-ENOMEM);
4953 static int only_parity(int raid_disk, int algo, int raid_disks, int max_degraded)
4956 case ALGORITHM_PARITY_0:
4957 if (raid_disk < max_degraded)
4960 case ALGORITHM_PARITY_N:
4961 if (raid_disk >= raid_disks - max_degraded)
4964 case ALGORITHM_PARITY_0_6:
4965 if (raid_disk == 0 ||
4966 raid_disk == raid_disks - 1)
4969 case ALGORITHM_LEFT_ASYMMETRIC_6:
4970 case ALGORITHM_RIGHT_ASYMMETRIC_6:
4971 case ALGORITHM_LEFT_SYMMETRIC_6:
4972 case ALGORITHM_RIGHT_SYMMETRIC_6:
4973 if (raid_disk == raid_disks - 1)
4979 static int run(struct mddev *mddev)
4981 struct r5conf *conf;
4982 int working_disks = 0;
4983 int dirty_parity_disks = 0;
4984 struct md_rdev *rdev;
4985 sector_t reshape_offset = 0;
4988 if (mddev->recovery_cp != MaxSector)
4989 printk(KERN_NOTICE "md/raid:%s: not clean"
4990 " -- starting background reconstruction\n",
4992 if (mddev->reshape_position != MaxSector) {
4993 /* Check that we can continue the reshape.
4994 * Currently only disks can change, it must
4995 * increase, and we must be past the point where
4996 * a stripe over-writes itself
4998 sector_t here_new, here_old;
5000 int max_degraded = (mddev->level == 6 ? 2 : 1);
5002 if (mddev->new_level != mddev->level) {
5003 printk(KERN_ERR "md/raid:%s: unsupported reshape "
5004 "required - aborting.\n",
5008 old_disks = mddev->raid_disks - mddev->delta_disks;
5009 /* reshape_position must be on a new-stripe boundary, and one
5010 * further up in new geometry must map after here in old
5013 here_new = mddev->reshape_position;
5014 if (sector_div(here_new, mddev->new_chunk_sectors *
5015 (mddev->raid_disks - max_degraded))) {
5016 printk(KERN_ERR "md/raid:%s: reshape_position not "
5017 "on a stripe boundary\n", mdname(mddev));
5020 reshape_offset = here_new * mddev->new_chunk_sectors;
5021 /* here_new is the stripe we will write to */
5022 here_old = mddev->reshape_position;
5023 sector_div(here_old, mddev->chunk_sectors *
5024 (old_disks-max_degraded));
5025 /* here_old is the first stripe that we might need to read
5027 if (mddev->delta_disks == 0) {
5028 /* We cannot be sure it is safe to start an in-place
5029 * reshape. It is only safe if user-space if monitoring
5030 * and taking constant backups.
5031 * mdadm always starts a situation like this in
5032 * readonly mode so it can take control before
5033 * allowing any writes. So just check for that.
5035 if ((here_new * mddev->new_chunk_sectors !=
5036 here_old * mddev->chunk_sectors) ||
5038 printk(KERN_ERR "md/raid:%s: in-place reshape must be started"
5039 " in read-only mode - aborting\n",
5043 } else if (mddev->reshape_backwards
5044 ? (here_new * mddev->new_chunk_sectors <=
5045 here_old * mddev->chunk_sectors)
5046 : (here_new * mddev->new_chunk_sectors >=
5047 here_old * mddev->chunk_sectors)) {
5048 /* Reading from the same stripe as writing to - bad */
5049 printk(KERN_ERR "md/raid:%s: reshape_position too early for "
5050 "auto-recovery - aborting.\n",
5054 printk(KERN_INFO "md/raid:%s: reshape will continue\n",
5056 /* OK, we should be able to continue; */
5058 BUG_ON(mddev->level != mddev->new_level);
5059 BUG_ON(mddev->layout != mddev->new_layout);
5060 BUG_ON(mddev->chunk_sectors != mddev->new_chunk_sectors);
5061 BUG_ON(mddev->delta_disks != 0);
5064 if (mddev->private == NULL)
5065 conf = setup_conf(mddev);
5067 conf = mddev->private;
5070 return PTR_ERR(conf);
5072 mddev->thread = conf->thread;
5073 conf->thread = NULL;
5074 mddev->private = conf;
5076 for (i = 0; i < conf->raid_disks && conf->previous_raid_disks;
5078 rdev = conf->disks[i].rdev;
5079 if (!rdev && conf->disks[i].replacement) {
5080 /* The replacement is all we have yet */
5081 rdev = conf->disks[i].replacement;
5082 conf->disks[i].replacement = NULL;
5083 clear_bit(Replacement, &rdev->flags);
5084 conf->disks[i].rdev = rdev;
5088 if (conf->disks[i].replacement &&
5089 conf->reshape_progress != MaxSector) {
5090 /* replacements and reshape simply do not mix. */
5091 printk(KERN_ERR "md: cannot handle concurrent "
5092 "replacement and reshape.\n");
5095 if (test_bit(In_sync, &rdev->flags)) {
5099 /* This disc is not fully in-sync. However if it
5100 * just stored parity (beyond the recovery_offset),
5101 * when we don't need to be concerned about the
5102 * array being dirty.
5103 * When reshape goes 'backwards', we never have
5104 * partially completed devices, so we only need
5105 * to worry about reshape going forwards.
5107 /* Hack because v0.91 doesn't store recovery_offset properly. */
5108 if (mddev->major_version == 0 &&
5109 mddev->minor_version > 90)
5110 rdev->recovery_offset = reshape_offset;
5112 if (rdev->recovery_offset < reshape_offset) {
5113 /* We need to check old and new layout */
5114 if (!only_parity(rdev->raid_disk,
5117 conf->max_degraded))
5120 if (!only_parity(rdev->raid_disk,
5122 conf->previous_raid_disks,
5123 conf->max_degraded))
5125 dirty_parity_disks++;
5129 * 0 for a fully functional array, 1 or 2 for a degraded array.
5131 mddev->degraded = calc_degraded(conf);
5133 if (has_failed(conf)) {
5134 printk(KERN_ERR "md/raid:%s: not enough operational devices"
5135 " (%d/%d failed)\n",
5136 mdname(mddev), mddev->degraded, conf->raid_disks);
5140 /* device size must be a multiple of chunk size */
5141 mddev->dev_sectors &= ~(mddev->chunk_sectors - 1);
5142 mddev->resync_max_sectors = mddev->dev_sectors;
5144 if (mddev->degraded > dirty_parity_disks &&
5145 mddev->recovery_cp != MaxSector) {
5146 if (mddev->ok_start_degraded)
5148 "md/raid:%s: starting dirty degraded array"
5149 " - data corruption possible.\n",
5153 "md/raid:%s: cannot start dirty degraded array.\n",
5159 if (mddev->degraded == 0)
5160 printk(KERN_INFO "md/raid:%s: raid level %d active with %d out of %d"
5161 " devices, algorithm %d\n", mdname(mddev), conf->level,
5162 mddev->raid_disks-mddev->degraded, mddev->raid_disks,
5165 printk(KERN_ALERT "md/raid:%s: raid level %d active with %d"
5166 " out of %d devices, algorithm %d\n",
5167 mdname(mddev), conf->level,
5168 mddev->raid_disks - mddev->degraded,
5169 mddev->raid_disks, mddev->new_layout);
5171 print_raid5_conf(conf);
5173 if (conf->reshape_progress != MaxSector) {
5174 conf->reshape_safe = conf->reshape_progress;
5175 atomic_set(&conf->reshape_stripes, 0);
5176 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
5177 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
5178 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
5179 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
5180 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
5185 /* Ok, everything is just fine now */
5186 if (mddev->to_remove == &raid5_attrs_group)
5187 mddev->to_remove = NULL;
5188 else if (mddev->kobj.sd &&
5189 sysfs_create_group(&mddev->kobj, &raid5_attrs_group))
5191 "raid5: failed to create sysfs attributes for %s\n",
5193 md_set_array_sectors(mddev, raid5_size(mddev, 0, 0));
5197 /* read-ahead size must cover two whole stripes, which
5198 * is 2 * (datadisks) * chunksize where 'n' is the
5199 * number of raid devices
5201 int data_disks = conf->previous_raid_disks - conf->max_degraded;
5202 int stripe = data_disks *
5203 ((mddev->chunk_sectors << 9) / PAGE_SIZE);
5204 if (mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
5205 mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
5207 blk_queue_merge_bvec(mddev->queue, raid5_mergeable_bvec);
5209 mddev->queue->backing_dev_info.congested_data = mddev;
5210 mddev->queue->backing_dev_info.congested_fn = raid5_congested;
5212 chunk_size = mddev->chunk_sectors << 9;
5213 blk_queue_io_min(mddev->queue, chunk_size);
5214 blk_queue_io_opt(mddev->queue, chunk_size *
5215 (conf->raid_disks - conf->max_degraded));
5217 rdev_for_each(rdev, mddev) {
5218 disk_stack_limits(mddev->gendisk, rdev->bdev,
5219 rdev->data_offset << 9);
5220 disk_stack_limits(mddev->gendisk, rdev->bdev,
5221 rdev->new_data_offset << 9);
5227 md_unregister_thread(&mddev->thread);
5228 print_raid5_conf(conf);
5230 mddev->private = NULL;
5231 printk(KERN_ALERT "md/raid:%s: failed to run raid set.\n", mdname(mddev));
5235 static int stop(struct mddev *mddev)
5237 struct r5conf *conf = mddev->private;
5239 md_unregister_thread(&mddev->thread);
5241 mddev->queue->backing_dev_info.congested_fn = NULL;
5243 mddev->private = NULL;
5244 mddev->to_remove = &raid5_attrs_group;
5248 static void status(struct seq_file *seq, struct mddev *mddev)
5250 struct r5conf *conf = mddev->private;
5253 seq_printf(seq, " level %d, %dk chunk, algorithm %d", mddev->level,
5254 mddev->chunk_sectors / 2, mddev->layout);
5255 seq_printf (seq, " [%d/%d] [", conf->raid_disks, conf->raid_disks - mddev->degraded);
5256 for (i = 0; i < conf->raid_disks; i++)
5257 seq_printf (seq, "%s",
5258 conf->disks[i].rdev &&
5259 test_bit(In_sync, &conf->disks[i].rdev->flags) ? "U" : "_");
5260 seq_printf (seq, "]");
5263 static void print_raid5_conf (struct r5conf *conf)
5266 struct disk_info *tmp;
5268 printk(KERN_DEBUG "RAID conf printout:\n");
5270 printk("(conf==NULL)\n");
5273 printk(KERN_DEBUG " --- level:%d rd:%d wd:%d\n", conf->level,
5275 conf->raid_disks - conf->mddev->degraded);
5277 for (i = 0; i < conf->raid_disks; i++) {
5278 char b[BDEVNAME_SIZE];
5279 tmp = conf->disks + i;
5281 printk(KERN_DEBUG " disk %d, o:%d, dev:%s\n",
5282 i, !test_bit(Faulty, &tmp->rdev->flags),
5283 bdevname(tmp->rdev->bdev, b));
5287 static int raid5_spare_active(struct mddev *mddev)
5290 struct r5conf *conf = mddev->private;
5291 struct disk_info *tmp;
5293 unsigned long flags;
5295 for (i = 0; i < conf->raid_disks; i++) {
5296 tmp = conf->disks + i;
5297 if (tmp->replacement
5298 && tmp->replacement->recovery_offset == MaxSector
5299 && !test_bit(Faulty, &tmp->replacement->flags)
5300 && !test_and_set_bit(In_sync, &tmp->replacement->flags)) {
5301 /* Replacement has just become active. */
5303 || !test_and_clear_bit(In_sync, &tmp->rdev->flags))
5306 /* Replaced device not technically faulty,
5307 * but we need to be sure it gets removed
5308 * and never re-added.
5310 set_bit(Faulty, &tmp->rdev->flags);
5311 sysfs_notify_dirent_safe(
5312 tmp->rdev->sysfs_state);
5314 sysfs_notify_dirent_safe(tmp->replacement->sysfs_state);
5315 } else if (tmp->rdev
5316 && tmp->rdev->recovery_offset == MaxSector
5317 && !test_bit(Faulty, &tmp->rdev->flags)
5318 && !test_and_set_bit(In_sync, &tmp->rdev->flags)) {
5320 sysfs_notify_dirent_safe(tmp->rdev->sysfs_state);
5323 spin_lock_irqsave(&conf->device_lock, flags);
5324 mddev->degraded = calc_degraded(conf);
5325 spin_unlock_irqrestore(&conf->device_lock, flags);
5326 print_raid5_conf(conf);
5330 static int raid5_remove_disk(struct mddev *mddev, struct md_rdev *rdev)
5332 struct r5conf *conf = mddev->private;
5334 int number = rdev->raid_disk;
5335 struct md_rdev **rdevp;
5336 struct disk_info *p = conf->disks + number;
5338 print_raid5_conf(conf);
5339 if (rdev == p->rdev)
5341 else if (rdev == p->replacement)
5342 rdevp = &p->replacement;
5346 if (number >= conf->raid_disks &&
5347 conf->reshape_progress == MaxSector)
5348 clear_bit(In_sync, &rdev->flags);
5350 if (test_bit(In_sync, &rdev->flags) ||
5351 atomic_read(&rdev->nr_pending)) {
5355 /* Only remove non-faulty devices if recovery
5358 if (!test_bit(Faulty, &rdev->flags) &&
5359 mddev->recovery_disabled != conf->recovery_disabled &&
5360 !has_failed(conf) &&
5361 (!p->replacement || p->replacement == rdev) &&
5362 number < conf->raid_disks) {
5368 if (atomic_read(&rdev->nr_pending)) {
5369 /* lost the race, try later */
5372 } else if (p->replacement) {
5373 /* We must have just cleared 'rdev' */
5374 p->rdev = p->replacement;
5375 clear_bit(Replacement, &p->replacement->flags);
5376 smp_mb(); /* Make sure other CPUs may see both as identical
5377 * but will never see neither - if they are careful
5379 p->replacement = NULL;
5380 clear_bit(WantReplacement, &rdev->flags);
5382 /* We might have just removed the Replacement as faulty-
5383 * clear the bit just in case
5385 clear_bit(WantReplacement, &rdev->flags);
5388 print_raid5_conf(conf);
5392 static int raid5_add_disk(struct mddev *mddev, struct md_rdev *rdev)
5394 struct r5conf *conf = mddev->private;
5397 struct disk_info *p;
5399 int last = conf->raid_disks - 1;
5401 if (mddev->recovery_disabled == conf->recovery_disabled)
5404 if (rdev->saved_raid_disk < 0 && has_failed(conf))
5405 /* no point adding a device */
5408 if (rdev->raid_disk >= 0)
5409 first = last = rdev->raid_disk;
5412 * find the disk ... but prefer rdev->saved_raid_disk
5415 if (rdev->saved_raid_disk >= 0 &&
5416 rdev->saved_raid_disk >= first &&
5417 conf->disks[rdev->saved_raid_disk].rdev == NULL)
5418 disk = rdev->saved_raid_disk;
5421 for ( ; disk <= last ; disk++) {
5422 p = conf->disks + disk;
5423 if (p->rdev == NULL) {
5424 clear_bit(In_sync, &rdev->flags);
5425 rdev->raid_disk = disk;
5427 if (rdev->saved_raid_disk != disk)
5429 rcu_assign_pointer(p->rdev, rdev);
5432 if (test_bit(WantReplacement, &p->rdev->flags) &&
5433 p->replacement == NULL) {
5434 clear_bit(In_sync, &rdev->flags);
5435 set_bit(Replacement, &rdev->flags);
5436 rdev->raid_disk = disk;
5439 rcu_assign_pointer(p->replacement, rdev);
5443 print_raid5_conf(conf);
5447 static int raid5_resize(struct mddev *mddev, sector_t sectors)
5449 /* no resync is happening, and there is enough space
5450 * on all devices, so we can resize.
5451 * We need to make sure resync covers any new space.
5452 * If the array is shrinking we should possibly wait until
5453 * any io in the removed space completes, but it hardly seems
5456 sectors &= ~((sector_t)mddev->chunk_sectors - 1);
5457 md_set_array_sectors(mddev, raid5_size(mddev, sectors,
5458 mddev->raid_disks));
5459 if (mddev->array_sectors >
5460 raid5_size(mddev, sectors, mddev->raid_disks))
5462 set_capacity(mddev->gendisk, mddev->array_sectors);
5463 revalidate_disk(mddev->gendisk);
5464 if (sectors > mddev->dev_sectors &&
5465 mddev->recovery_cp > mddev->dev_sectors) {
5466 mddev->recovery_cp = mddev->dev_sectors;
5467 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
5469 mddev->dev_sectors = sectors;
5470 mddev->resync_max_sectors = sectors;
5474 static int check_stripe_cache(struct mddev *mddev)
5476 /* Can only proceed if there are plenty of stripe_heads.
5477 * We need a minimum of one full stripe,, and for sensible progress
5478 * it is best to have about 4 times that.
5479 * If we require 4 times, then the default 256 4K stripe_heads will
5480 * allow for chunk sizes up to 256K, which is probably OK.
5481 * If the chunk size is greater, user-space should request more
5482 * stripe_heads first.
5484 struct r5conf *conf = mddev->private;
5485 if (((mddev->chunk_sectors << 9) / STRIPE_SIZE) * 4
5486 > conf->max_nr_stripes ||
5487 ((mddev->new_chunk_sectors << 9) / STRIPE_SIZE) * 4
5488 > conf->max_nr_stripes) {
5489 printk(KERN_WARNING "md/raid:%s: reshape: not enough stripes. Needed %lu\n",
5491 ((max(mddev->chunk_sectors, mddev->new_chunk_sectors) << 9)
5498 static int check_reshape(struct mddev *mddev)
5500 struct r5conf *conf = mddev->private;
5502 if (mddev->delta_disks == 0 &&
5503 mddev->new_layout == mddev->layout &&
5504 mddev->new_chunk_sectors == mddev->chunk_sectors)
5505 return 0; /* nothing to do */
5507 /* Cannot grow a bitmap yet */
5509 if (has_failed(conf))
5511 if (mddev->delta_disks < 0) {
5512 /* We might be able to shrink, but the devices must
5513 * be made bigger first.
5514 * For raid6, 4 is the minimum size.
5515 * Otherwise 2 is the minimum
5518 if (mddev->level == 6)
5520 if (mddev->raid_disks + mddev->delta_disks < min)
5524 if (!check_stripe_cache(mddev))
5527 return resize_stripes(conf, conf->raid_disks + mddev->delta_disks);
5530 static int raid5_start_reshape(struct mddev *mddev)
5532 struct r5conf *conf = mddev->private;
5533 struct md_rdev *rdev;
5535 unsigned long flags;
5537 if (test_bit(MD_RECOVERY_RUNNING, &mddev->recovery))
5540 if (!check_stripe_cache(mddev))
5543 rdev_for_each(rdev, mddev) {
5544 /* Don't support changing data_offset yet */
5545 if (rdev->new_data_offset != rdev->data_offset)
5547 if (!test_bit(In_sync, &rdev->flags)
5548 && !test_bit(Faulty, &rdev->flags))
5552 if (spares - mddev->degraded < mddev->delta_disks - conf->max_degraded)
5553 /* Not enough devices even to make a degraded array
5558 /* Refuse to reduce size of the array. Any reductions in
5559 * array size must be through explicit setting of array_size
5562 if (raid5_size(mddev, 0, conf->raid_disks + mddev->delta_disks)
5563 < mddev->array_sectors) {
5564 printk(KERN_ERR "md/raid:%s: array size must be reduced "
5565 "before number of disks\n", mdname(mddev));
5569 atomic_set(&conf->reshape_stripes, 0);
5570 spin_lock_irq(&conf->device_lock);
5571 conf->previous_raid_disks = conf->raid_disks;
5572 conf->raid_disks += mddev->delta_disks;
5573 conf->prev_chunk_sectors = conf->chunk_sectors;
5574 conf->chunk_sectors = mddev->new_chunk_sectors;
5575 conf->prev_algo = conf->algorithm;
5576 conf->algorithm = mddev->new_layout;
5578 /* Code that selects data_offset needs to see the generation update
5579 * if reshape_progress has been set - so a memory barrier needed.
5582 if (mddev->reshape_backwards)
5583 conf->reshape_progress = raid5_size(mddev, 0, 0);
5585 conf->reshape_progress = 0;
5586 conf->reshape_safe = conf->reshape_progress;
5587 spin_unlock_irq(&conf->device_lock);
5589 /* Add some new drives, as many as will fit.
5590 * We know there are enough to make the newly sized array work.
5591 * Don't add devices if we are reducing the number of
5592 * devices in the array. This is because it is not possible
5593 * to correctly record the "partially reconstructed" state of
5594 * such devices during the reshape and confusion could result.
5596 if (mddev->delta_disks >= 0) {
5597 rdev_for_each(rdev, mddev)
5598 if (rdev->raid_disk < 0 &&
5599 !test_bit(Faulty, &rdev->flags)) {
5600 if (raid5_add_disk(mddev, rdev) == 0) {
5602 >= conf->previous_raid_disks)
5603 set_bit(In_sync, &rdev->flags);
5605 rdev->recovery_offset = 0;
5607 if (sysfs_link_rdev(mddev, rdev))
5608 /* Failure here is OK */;
5610 } else if (rdev->raid_disk >= conf->previous_raid_disks
5611 && !test_bit(Faulty, &rdev->flags)) {
5612 /* This is a spare that was manually added */
5613 set_bit(In_sync, &rdev->flags);
5616 /* When a reshape changes the number of devices,
5617 * ->degraded is measured against the larger of the
5618 * pre and post number of devices.
5620 spin_lock_irqsave(&conf->device_lock, flags);
5621 mddev->degraded = calc_degraded(conf);
5622 spin_unlock_irqrestore(&conf->device_lock, flags);
5624 mddev->raid_disks = conf->raid_disks;
5625 mddev->reshape_position = conf->reshape_progress;
5626 set_bit(MD_CHANGE_DEVS, &mddev->flags);
5628 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
5629 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
5630 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
5631 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
5632 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
5634 if (!mddev->sync_thread) {
5635 mddev->recovery = 0;
5636 spin_lock_irq(&conf->device_lock);
5637 mddev->raid_disks = conf->raid_disks = conf->previous_raid_disks;
5638 rdev_for_each(rdev, mddev)
5639 rdev->new_data_offset = rdev->data_offset;
5641 conf->reshape_progress = MaxSector;
5642 mddev->reshape_position = MaxSector;
5643 spin_unlock_irq(&conf->device_lock);
5646 conf->reshape_checkpoint = jiffies;
5647 md_wakeup_thread(mddev->sync_thread);
5648 md_new_event(mddev);
5652 /* This is called from the reshape thread and should make any
5653 * changes needed in 'conf'
5655 static void end_reshape(struct r5conf *conf)
5658 if (!test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery)) {
5659 struct md_rdev *rdev;
5661 spin_lock_irq(&conf->device_lock);
5662 conf->previous_raid_disks = conf->raid_disks;
5663 rdev_for_each(rdev, conf->mddev)
5664 rdev->data_offset = rdev->new_data_offset;
5666 conf->reshape_progress = MaxSector;
5667 spin_unlock_irq(&conf->device_lock);
5668 wake_up(&conf->wait_for_overlap);
5670 /* read-ahead size must cover two whole stripes, which is
5671 * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
5673 if (conf->mddev->queue) {
5674 int data_disks = conf->raid_disks - conf->max_degraded;
5675 int stripe = data_disks * ((conf->chunk_sectors << 9)
5677 if (conf->mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
5678 conf->mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
5683 /* This is called from the raid5d thread with mddev_lock held.
5684 * It makes config changes to the device.
5686 static void raid5_finish_reshape(struct mddev *mddev)
5688 struct r5conf *conf = mddev->private;
5690 if (!test_bit(MD_RECOVERY_INTR, &mddev->recovery)) {
5692 if (mddev->delta_disks > 0) {
5693 md_set_array_sectors(mddev, raid5_size(mddev, 0, 0));
5694 set_capacity(mddev->gendisk, mddev->array_sectors);
5695 revalidate_disk(mddev->gendisk);
5698 spin_lock_irq(&conf->device_lock);
5699 mddev->degraded = calc_degraded(conf);
5700 spin_unlock_irq(&conf->device_lock);
5701 for (d = conf->raid_disks ;
5702 d < conf->raid_disks - mddev->delta_disks;
5704 struct md_rdev *rdev = conf->disks[d].rdev;
5706 raid5_remove_disk(mddev, rdev) == 0) {
5707 sysfs_unlink_rdev(mddev, rdev);
5708 rdev->raid_disk = -1;
5712 mddev->layout = conf->algorithm;
5713 mddev->chunk_sectors = conf->chunk_sectors;
5714 mddev->reshape_position = MaxSector;
5715 mddev->delta_disks = 0;
5716 mddev->reshape_backwards = 0;
5720 static void raid5_quiesce(struct mddev *mddev, int state)
5722 struct r5conf *conf = mddev->private;
5725 case 2: /* resume for a suspend */
5726 wake_up(&conf->wait_for_overlap);
5729 case 1: /* stop all writes */
5730 spin_lock_irq(&conf->device_lock);
5731 /* '2' tells resync/reshape to pause so that all
5732 * active stripes can drain
5735 wait_event_lock_irq(conf->wait_for_stripe,
5736 atomic_read(&conf->active_stripes) == 0 &&
5737 atomic_read(&conf->active_aligned_reads) == 0,
5738 conf->device_lock, /* nothing */);
5740 spin_unlock_irq(&conf->device_lock);
5741 /* allow reshape to continue */
5742 wake_up(&conf->wait_for_overlap);
5745 case 0: /* re-enable writes */
5746 spin_lock_irq(&conf->device_lock);
5748 wake_up(&conf->wait_for_stripe);
5749 wake_up(&conf->wait_for_overlap);
5750 spin_unlock_irq(&conf->device_lock);
5756 static void *raid45_takeover_raid0(struct mddev *mddev, int level)
5758 struct r0conf *raid0_conf = mddev->private;
5761 /* for raid0 takeover only one zone is supported */
5762 if (raid0_conf->nr_strip_zones > 1) {
5763 printk(KERN_ERR "md/raid:%s: cannot takeover raid0 with more than one zone.\n",
5765 return ERR_PTR(-EINVAL);
5768 sectors = raid0_conf->strip_zone[0].zone_end;
5769 sector_div(sectors, raid0_conf->strip_zone[0].nb_dev);
5770 mddev->dev_sectors = sectors;
5771 mddev->new_level = level;
5772 mddev->new_layout = ALGORITHM_PARITY_N;
5773 mddev->new_chunk_sectors = mddev->chunk_sectors;
5774 mddev->raid_disks += 1;
5775 mddev->delta_disks = 1;
5776 /* make sure it will be not marked as dirty */
5777 mddev->recovery_cp = MaxSector;
5779 return setup_conf(mddev);
5783 static void *raid5_takeover_raid1(struct mddev *mddev)
5787 if (mddev->raid_disks != 2 ||
5788 mddev->degraded > 1)
5789 return ERR_PTR(-EINVAL);
5791 /* Should check if there are write-behind devices? */
5793 chunksect = 64*2; /* 64K by default */
5795 /* The array must be an exact multiple of chunksize */
5796 while (chunksect && (mddev->array_sectors & (chunksect-1)))
5799 if ((chunksect<<9) < STRIPE_SIZE)
5800 /* array size does not allow a suitable chunk size */
5801 return ERR_PTR(-EINVAL);
5803 mddev->new_level = 5;
5804 mddev->new_layout = ALGORITHM_LEFT_SYMMETRIC;
5805 mddev->new_chunk_sectors = chunksect;
5807 return setup_conf(mddev);
5810 static void *raid5_takeover_raid6(struct mddev *mddev)
5814 switch (mddev->layout) {
5815 case ALGORITHM_LEFT_ASYMMETRIC_6:
5816 new_layout = ALGORITHM_LEFT_ASYMMETRIC;
5818 case ALGORITHM_RIGHT_ASYMMETRIC_6:
5819 new_layout = ALGORITHM_RIGHT_ASYMMETRIC;
5821 case ALGORITHM_LEFT_SYMMETRIC_6:
5822 new_layout = ALGORITHM_LEFT_SYMMETRIC;
5824 case ALGORITHM_RIGHT_SYMMETRIC_6:
5825 new_layout = ALGORITHM_RIGHT_SYMMETRIC;
5827 case ALGORITHM_PARITY_0_6:
5828 new_layout = ALGORITHM_PARITY_0;
5830 case ALGORITHM_PARITY_N:
5831 new_layout = ALGORITHM_PARITY_N;
5834 return ERR_PTR(-EINVAL);
5836 mddev->new_level = 5;
5837 mddev->new_layout = new_layout;
5838 mddev->delta_disks = -1;
5839 mddev->raid_disks -= 1;
5840 return setup_conf(mddev);
5844 static int raid5_check_reshape(struct mddev *mddev)
5846 /* For a 2-drive array, the layout and chunk size can be changed
5847 * immediately as not restriping is needed.
5848 * For larger arrays we record the new value - after validation
5849 * to be used by a reshape pass.
5851 struct r5conf *conf = mddev->private;
5852 int new_chunk = mddev->new_chunk_sectors;
5854 if (mddev->new_layout >= 0 && !algorithm_valid_raid5(mddev->new_layout))
5856 if (new_chunk > 0) {
5857 if (!is_power_of_2(new_chunk))
5859 if (new_chunk < (PAGE_SIZE>>9))
5861 if (mddev->array_sectors & (new_chunk-1))
5862 /* not factor of array size */
5866 /* They look valid */
5868 if (mddev->raid_disks == 2) {
5869 /* can make the change immediately */
5870 if (mddev->new_layout >= 0) {
5871 conf->algorithm = mddev->new_layout;
5872 mddev->layout = mddev->new_layout;
5874 if (new_chunk > 0) {
5875 conf->chunk_sectors = new_chunk ;
5876 mddev->chunk_sectors = new_chunk;
5878 set_bit(MD_CHANGE_DEVS, &mddev->flags);
5879 md_wakeup_thread(mddev->thread);
5881 return check_reshape(mddev);
5884 static int raid6_check_reshape(struct mddev *mddev)
5886 int new_chunk = mddev->new_chunk_sectors;
5888 if (mddev->new_layout >= 0 && !algorithm_valid_raid6(mddev->new_layout))
5890 if (new_chunk > 0) {
5891 if (!is_power_of_2(new_chunk))
5893 if (new_chunk < (PAGE_SIZE >> 9))
5895 if (mddev->array_sectors & (new_chunk-1))
5896 /* not factor of array size */
5900 /* They look valid */
5901 return check_reshape(mddev);
5904 static void *raid5_takeover(struct mddev *mddev)
5906 /* raid5 can take over:
5907 * raid0 - if there is only one strip zone - make it a raid4 layout
5908 * raid1 - if there are two drives. We need to know the chunk size
5909 * raid4 - trivial - just use a raid4 layout.
5910 * raid6 - Providing it is a *_6 layout
5912 if (mddev->level == 0)
5913 return raid45_takeover_raid0(mddev, 5);
5914 if (mddev->level == 1)
5915 return raid5_takeover_raid1(mddev);
5916 if (mddev->level == 4) {
5917 mddev->new_layout = ALGORITHM_PARITY_N;
5918 mddev->new_level = 5;
5919 return setup_conf(mddev);
5921 if (mddev->level == 6)
5922 return raid5_takeover_raid6(mddev);
5924 return ERR_PTR(-EINVAL);
5927 static void *raid4_takeover(struct mddev *mddev)
5929 /* raid4 can take over:
5930 * raid0 - if there is only one strip zone
5931 * raid5 - if layout is right
5933 if (mddev->level == 0)
5934 return raid45_takeover_raid0(mddev, 4);
5935 if (mddev->level == 5 &&
5936 mddev->layout == ALGORITHM_PARITY_N) {
5937 mddev->new_layout = 0;
5938 mddev->new_level = 4;
5939 return setup_conf(mddev);
5941 return ERR_PTR(-EINVAL);
5944 static struct md_personality raid5_personality;
5946 static void *raid6_takeover(struct mddev *mddev)
5948 /* Currently can only take over a raid5. We map the
5949 * personality to an equivalent raid6 personality
5950 * with the Q block at the end.
5954 if (mddev->pers != &raid5_personality)
5955 return ERR_PTR(-EINVAL);
5956 if (mddev->degraded > 1)
5957 return ERR_PTR(-EINVAL);
5958 if (mddev->raid_disks > 253)
5959 return ERR_PTR(-EINVAL);
5960 if (mddev->raid_disks < 3)
5961 return ERR_PTR(-EINVAL);
5963 switch (mddev->layout) {
5964 case ALGORITHM_LEFT_ASYMMETRIC:
5965 new_layout = ALGORITHM_LEFT_ASYMMETRIC_6;
5967 case ALGORITHM_RIGHT_ASYMMETRIC:
5968 new_layout = ALGORITHM_RIGHT_ASYMMETRIC_6;
5970 case ALGORITHM_LEFT_SYMMETRIC:
5971 new_layout = ALGORITHM_LEFT_SYMMETRIC_6;
5973 case ALGORITHM_RIGHT_SYMMETRIC:
5974 new_layout = ALGORITHM_RIGHT_SYMMETRIC_6;
5976 case ALGORITHM_PARITY_0:
5977 new_layout = ALGORITHM_PARITY_0_6;
5979 case ALGORITHM_PARITY_N:
5980 new_layout = ALGORITHM_PARITY_N;
5983 return ERR_PTR(-EINVAL);
5985 mddev->new_level = 6;
5986 mddev->new_layout = new_layout;
5987 mddev->delta_disks = 1;
5988 mddev->raid_disks += 1;
5989 return setup_conf(mddev);
5993 static struct md_personality raid6_personality =
5997 .owner = THIS_MODULE,
5998 .make_request = make_request,
6002 .error_handler = error,
6003 .hot_add_disk = raid5_add_disk,
6004 .hot_remove_disk= raid5_remove_disk,
6005 .spare_active = raid5_spare_active,
6006 .sync_request = sync_request,
6007 .resize = raid5_resize,
6009 .check_reshape = raid6_check_reshape,
6010 .start_reshape = raid5_start_reshape,
6011 .finish_reshape = raid5_finish_reshape,
6012 .quiesce = raid5_quiesce,
6013 .takeover = raid6_takeover,
6015 static struct md_personality raid5_personality =
6019 .owner = THIS_MODULE,
6020 .make_request = make_request,
6024 .error_handler = error,
6025 .hot_add_disk = raid5_add_disk,
6026 .hot_remove_disk= raid5_remove_disk,
6027 .spare_active = raid5_spare_active,
6028 .sync_request = sync_request,
6029 .resize = raid5_resize,
6031 .check_reshape = raid5_check_reshape,
6032 .start_reshape = raid5_start_reshape,
6033 .finish_reshape = raid5_finish_reshape,
6034 .quiesce = raid5_quiesce,
6035 .takeover = raid5_takeover,
6038 static struct md_personality raid4_personality =
6042 .owner = THIS_MODULE,
6043 .make_request = make_request,
6047 .error_handler = error,
6048 .hot_add_disk = raid5_add_disk,
6049 .hot_remove_disk= raid5_remove_disk,
6050 .spare_active = raid5_spare_active,
6051 .sync_request = sync_request,
6052 .resize = raid5_resize,
6054 .check_reshape = raid5_check_reshape,
6055 .start_reshape = raid5_start_reshape,
6056 .finish_reshape = raid5_finish_reshape,
6057 .quiesce = raid5_quiesce,
6058 .takeover = raid4_takeover,
6061 static int __init raid5_init(void)
6063 register_md_personality(&raid6_personality);
6064 register_md_personality(&raid5_personality);
6065 register_md_personality(&raid4_personality);
6069 static void raid5_exit(void)
6071 unregister_md_personality(&raid6_personality);
6072 unregister_md_personality(&raid5_personality);
6073 unregister_md_personality(&raid4_personality);
6076 module_init(raid5_init);
6077 module_exit(raid5_exit);
6078 MODULE_LICENSE("GPL");
6079 MODULE_DESCRIPTION("RAID4/5/6 (striping with parity) personality for MD");
6080 MODULE_ALIAS("md-personality-4"); /* RAID5 */
6081 MODULE_ALIAS("md-raid5");
6082 MODULE_ALIAS("md-raid4");
6083 MODULE_ALIAS("md-level-5");
6084 MODULE_ALIAS("md-level-4");
6085 MODULE_ALIAS("md-personality-8"); /* RAID6 */
6086 MODULE_ALIAS("md-raid6");
6087 MODULE_ALIAS("md-level-6");
6089 /* This used to be two separate modules, they were: */
6090 MODULE_ALIAS("raid5");
6091 MODULE_ALIAS("raid6");