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)) {
542 if (test_and_clear_bit(R5_SyncIO, &sh->dev[i].flags))
545 bi = &sh->dev[i].req;
546 rbi = &sh->dev[i].rreq; /* For writing to replacement */
551 bi->bi_end_io = raid5_end_write_request;
552 rbi->bi_end_io = raid5_end_write_request;
554 bi->bi_end_io = raid5_end_read_request;
557 rrdev = rcu_dereference(conf->disks[i].replacement);
558 smp_mb(); /* Ensure that if rrdev is NULL, rdev won't be */
559 rdev = rcu_dereference(conf->disks[i].rdev);
568 /* We raced and saw duplicates */
571 if (test_bit(R5_ReadRepl, &sh->dev[i].flags) && rrdev)
576 if (rdev && test_bit(Faulty, &rdev->flags))
579 atomic_inc(&rdev->nr_pending);
580 if (rrdev && test_bit(Faulty, &rrdev->flags))
583 atomic_inc(&rrdev->nr_pending);
586 /* We have already checked bad blocks for reads. Now
587 * need to check for writes. We never accept write errors
588 * on the replacement, so we don't to check rrdev.
590 while ((rw & WRITE) && rdev &&
591 test_bit(WriteErrorSeen, &rdev->flags)) {
594 int bad = is_badblock(rdev, sh->sector, STRIPE_SECTORS,
595 &first_bad, &bad_sectors);
600 set_bit(BlockedBadBlocks, &rdev->flags);
601 if (!conf->mddev->external &&
602 conf->mddev->flags) {
603 /* It is very unlikely, but we might
604 * still need to write out the
605 * bad block log - better give it
607 md_check_recovery(conf->mddev);
610 * Because md_wait_for_blocked_rdev
611 * will dec nr_pending, we must
612 * increment it first.
614 atomic_inc(&rdev->nr_pending);
615 md_wait_for_blocked_rdev(rdev, conf->mddev);
617 /* Acknowledged bad block - skip the write */
618 rdev_dec_pending(rdev, conf->mddev);
624 if (s->syncing || s->expanding || s->expanded
626 md_sync_acct(rdev->bdev, STRIPE_SECTORS);
628 set_bit(STRIPE_IO_STARTED, &sh->state);
630 bi->bi_bdev = rdev->bdev;
631 pr_debug("%s: for %llu schedule op %ld on disc %d\n",
632 __func__, (unsigned long long)sh->sector,
634 atomic_inc(&sh->count);
635 if (use_new_offset(conf, sh))
636 bi->bi_sector = (sh->sector
637 + rdev->new_data_offset);
639 bi->bi_sector = (sh->sector
640 + rdev->data_offset);
641 bi->bi_flags = 1 << BIO_UPTODATE;
643 bi->bi_io_vec[0].bv_len = STRIPE_SIZE;
644 bi->bi_io_vec[0].bv_offset = 0;
645 bi->bi_size = STRIPE_SIZE;
648 set_bit(R5_DOUBLE_LOCKED, &sh->dev[i].flags);
649 generic_make_request(bi);
652 if (s->syncing || s->expanding || s->expanded
654 md_sync_acct(rrdev->bdev, STRIPE_SECTORS);
656 set_bit(STRIPE_IO_STARTED, &sh->state);
658 rbi->bi_bdev = rrdev->bdev;
659 pr_debug("%s: for %llu schedule op %ld on "
660 "replacement disc %d\n",
661 __func__, (unsigned long long)sh->sector,
663 atomic_inc(&sh->count);
664 if (use_new_offset(conf, sh))
665 rbi->bi_sector = (sh->sector
666 + rrdev->new_data_offset);
668 rbi->bi_sector = (sh->sector
669 + rrdev->data_offset);
670 rbi->bi_flags = 1 << BIO_UPTODATE;
672 rbi->bi_io_vec[0].bv_len = STRIPE_SIZE;
673 rbi->bi_io_vec[0].bv_offset = 0;
674 rbi->bi_size = STRIPE_SIZE;
676 generic_make_request(rbi);
678 if (!rdev && !rrdev) {
680 set_bit(STRIPE_DEGRADED, &sh->state);
681 pr_debug("skip op %ld on disc %d for sector %llu\n",
682 bi->bi_rw, i, (unsigned long long)sh->sector);
683 clear_bit(R5_LOCKED, &sh->dev[i].flags);
684 set_bit(STRIPE_HANDLE, &sh->state);
689 static struct dma_async_tx_descriptor *
690 async_copy_data(int frombio, struct bio *bio, struct page *page,
691 sector_t sector, struct dma_async_tx_descriptor *tx)
694 struct page *bio_page;
697 struct async_submit_ctl submit;
698 enum async_tx_flags flags = 0;
700 if (bio->bi_sector >= sector)
701 page_offset = (signed)(bio->bi_sector - sector) * 512;
703 page_offset = (signed)(sector - bio->bi_sector) * -512;
706 flags |= ASYNC_TX_FENCE;
707 init_async_submit(&submit, flags, tx, NULL, NULL, NULL);
709 bio_for_each_segment(bvl, bio, i) {
710 int len = bvl->bv_len;
714 if (page_offset < 0) {
715 b_offset = -page_offset;
716 page_offset += b_offset;
720 if (len > 0 && page_offset + len > STRIPE_SIZE)
721 clen = STRIPE_SIZE - page_offset;
726 b_offset += bvl->bv_offset;
727 bio_page = bvl->bv_page;
729 tx = async_memcpy(page, bio_page, page_offset,
730 b_offset, clen, &submit);
732 tx = async_memcpy(bio_page, page, b_offset,
733 page_offset, clen, &submit);
735 /* chain the operations */
736 submit.depend_tx = tx;
738 if (clen < len) /* hit end of page */
746 static void ops_complete_biofill(void *stripe_head_ref)
748 struct stripe_head *sh = stripe_head_ref;
749 struct bio *return_bi = NULL;
750 struct r5conf *conf = sh->raid_conf;
753 pr_debug("%s: stripe %llu\n", __func__,
754 (unsigned long long)sh->sector);
756 /* clear completed biofills */
757 spin_lock_irq(&conf->device_lock);
758 for (i = sh->disks; i--; ) {
759 struct r5dev *dev = &sh->dev[i];
761 /* acknowledge completion of a biofill operation */
762 /* and check if we need to reply to a read request,
763 * new R5_Wantfill requests are held off until
764 * !STRIPE_BIOFILL_RUN
766 if (test_and_clear_bit(R5_Wantfill, &dev->flags)) {
767 struct bio *rbi, *rbi2;
772 while (rbi && rbi->bi_sector <
773 dev->sector + STRIPE_SECTORS) {
774 rbi2 = r5_next_bio(rbi, dev->sector);
775 if (!raid5_dec_bi_phys_segments(rbi)) {
776 rbi->bi_next = return_bi;
783 spin_unlock_irq(&conf->device_lock);
784 clear_bit(STRIPE_BIOFILL_RUN, &sh->state);
786 return_io(return_bi);
788 set_bit(STRIPE_HANDLE, &sh->state);
792 static void ops_run_biofill(struct stripe_head *sh)
794 struct dma_async_tx_descriptor *tx = NULL;
795 struct r5conf *conf = sh->raid_conf;
796 struct async_submit_ctl submit;
799 pr_debug("%s: stripe %llu\n", __func__,
800 (unsigned long long)sh->sector);
802 for (i = sh->disks; i--; ) {
803 struct r5dev *dev = &sh->dev[i];
804 if (test_bit(R5_Wantfill, &dev->flags)) {
806 spin_lock_irq(&conf->device_lock);
807 dev->read = rbi = dev->toread;
809 spin_unlock_irq(&conf->device_lock);
810 while (rbi && rbi->bi_sector <
811 dev->sector + STRIPE_SECTORS) {
812 tx = async_copy_data(0, rbi, dev->page,
814 rbi = r5_next_bio(rbi, dev->sector);
819 atomic_inc(&sh->count);
820 init_async_submit(&submit, ASYNC_TX_ACK, tx, ops_complete_biofill, sh, NULL);
821 async_trigger_callback(&submit);
824 static void mark_target_uptodate(struct stripe_head *sh, int target)
831 tgt = &sh->dev[target];
832 set_bit(R5_UPTODATE, &tgt->flags);
833 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
834 clear_bit(R5_Wantcompute, &tgt->flags);
837 static void ops_complete_compute(void *stripe_head_ref)
839 struct stripe_head *sh = stripe_head_ref;
841 pr_debug("%s: stripe %llu\n", __func__,
842 (unsigned long long)sh->sector);
844 /* mark the computed target(s) as uptodate */
845 mark_target_uptodate(sh, sh->ops.target);
846 mark_target_uptodate(sh, sh->ops.target2);
848 clear_bit(STRIPE_COMPUTE_RUN, &sh->state);
849 if (sh->check_state == check_state_compute_run)
850 sh->check_state = check_state_compute_result;
851 set_bit(STRIPE_HANDLE, &sh->state);
855 /* return a pointer to the address conversion region of the scribble buffer */
856 static addr_conv_t *to_addr_conv(struct stripe_head *sh,
857 struct raid5_percpu *percpu)
859 return percpu->scribble + sizeof(struct page *) * (sh->disks + 2);
862 static struct dma_async_tx_descriptor *
863 ops_run_compute5(struct stripe_head *sh, struct raid5_percpu *percpu)
865 int disks = sh->disks;
866 struct page **xor_srcs = percpu->scribble;
867 int target = sh->ops.target;
868 struct r5dev *tgt = &sh->dev[target];
869 struct page *xor_dest = tgt->page;
871 struct dma_async_tx_descriptor *tx;
872 struct async_submit_ctl submit;
875 pr_debug("%s: stripe %llu block: %d\n",
876 __func__, (unsigned long long)sh->sector, target);
877 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
879 for (i = disks; i--; )
881 xor_srcs[count++] = sh->dev[i].page;
883 atomic_inc(&sh->count);
885 init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST, NULL,
886 ops_complete_compute, sh, to_addr_conv(sh, percpu));
887 if (unlikely(count == 1))
888 tx = async_memcpy(xor_dest, xor_srcs[0], 0, 0, STRIPE_SIZE, &submit);
890 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
895 /* set_syndrome_sources - populate source buffers for gen_syndrome
896 * @srcs - (struct page *) array of size sh->disks
897 * @sh - stripe_head to parse
899 * Populates srcs in proper layout order for the stripe and returns the
900 * 'count' of sources to be used in a call to async_gen_syndrome. The P
901 * destination buffer is recorded in srcs[count] and the Q destination
902 * is recorded in srcs[count+1]].
904 static int set_syndrome_sources(struct page **srcs, struct stripe_head *sh)
906 int disks = sh->disks;
907 int syndrome_disks = sh->ddf_layout ? disks : (disks - 2);
908 int d0_idx = raid6_d0(sh);
912 for (i = 0; i < disks; i++)
918 int slot = raid6_idx_to_slot(i, sh, &count, syndrome_disks);
920 srcs[slot] = sh->dev[i].page;
921 i = raid6_next_disk(i, disks);
922 } while (i != d0_idx);
924 return syndrome_disks;
927 static struct dma_async_tx_descriptor *
928 ops_run_compute6_1(struct stripe_head *sh, struct raid5_percpu *percpu)
930 int disks = sh->disks;
931 struct page **blocks = percpu->scribble;
933 int qd_idx = sh->qd_idx;
934 struct dma_async_tx_descriptor *tx;
935 struct async_submit_ctl submit;
941 if (sh->ops.target < 0)
942 target = sh->ops.target2;
943 else if (sh->ops.target2 < 0)
944 target = sh->ops.target;
946 /* we should only have one valid target */
949 pr_debug("%s: stripe %llu block: %d\n",
950 __func__, (unsigned long long)sh->sector, target);
952 tgt = &sh->dev[target];
953 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
956 atomic_inc(&sh->count);
958 if (target == qd_idx) {
959 count = set_syndrome_sources(blocks, sh);
960 blocks[count] = NULL; /* regenerating p is not necessary */
961 BUG_ON(blocks[count+1] != dest); /* q should already be set */
962 init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
963 ops_complete_compute, sh,
964 to_addr_conv(sh, percpu));
965 tx = async_gen_syndrome(blocks, 0, count+2, STRIPE_SIZE, &submit);
967 /* Compute any data- or p-drive using XOR */
969 for (i = disks; i-- ; ) {
970 if (i == target || i == qd_idx)
972 blocks[count++] = sh->dev[i].page;
975 init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST,
976 NULL, ops_complete_compute, sh,
977 to_addr_conv(sh, percpu));
978 tx = async_xor(dest, blocks, 0, count, STRIPE_SIZE, &submit);
984 static struct dma_async_tx_descriptor *
985 ops_run_compute6_2(struct stripe_head *sh, struct raid5_percpu *percpu)
987 int i, count, disks = sh->disks;
988 int syndrome_disks = sh->ddf_layout ? disks : disks-2;
989 int d0_idx = raid6_d0(sh);
990 int faila = -1, failb = -1;
991 int target = sh->ops.target;
992 int target2 = sh->ops.target2;
993 struct r5dev *tgt = &sh->dev[target];
994 struct r5dev *tgt2 = &sh->dev[target2];
995 struct dma_async_tx_descriptor *tx;
996 struct page **blocks = percpu->scribble;
997 struct async_submit_ctl submit;
999 pr_debug("%s: stripe %llu block1: %d block2: %d\n",
1000 __func__, (unsigned long long)sh->sector, target, target2);
1001 BUG_ON(target < 0 || target2 < 0);
1002 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
1003 BUG_ON(!test_bit(R5_Wantcompute, &tgt2->flags));
1005 /* we need to open-code set_syndrome_sources to handle the
1006 * slot number conversion for 'faila' and 'failb'
1008 for (i = 0; i < disks ; i++)
1013 int slot = raid6_idx_to_slot(i, sh, &count, syndrome_disks);
1015 blocks[slot] = sh->dev[i].page;
1021 i = raid6_next_disk(i, disks);
1022 } while (i != d0_idx);
1024 BUG_ON(faila == failb);
1027 pr_debug("%s: stripe: %llu faila: %d failb: %d\n",
1028 __func__, (unsigned long long)sh->sector, faila, failb);
1030 atomic_inc(&sh->count);
1032 if (failb == syndrome_disks+1) {
1033 /* Q disk is one of the missing disks */
1034 if (faila == syndrome_disks) {
1035 /* Missing P+Q, just recompute */
1036 init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
1037 ops_complete_compute, sh,
1038 to_addr_conv(sh, percpu));
1039 return async_gen_syndrome(blocks, 0, syndrome_disks+2,
1040 STRIPE_SIZE, &submit);
1044 int qd_idx = sh->qd_idx;
1046 /* Missing D+Q: recompute D from P, then recompute Q */
1047 if (target == qd_idx)
1048 data_target = target2;
1050 data_target = target;
1053 for (i = disks; i-- ; ) {
1054 if (i == data_target || i == qd_idx)
1056 blocks[count++] = sh->dev[i].page;
1058 dest = sh->dev[data_target].page;
1059 init_async_submit(&submit,
1060 ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST,
1062 to_addr_conv(sh, percpu));
1063 tx = async_xor(dest, blocks, 0, count, STRIPE_SIZE,
1066 count = set_syndrome_sources(blocks, sh);
1067 init_async_submit(&submit, ASYNC_TX_FENCE, tx,
1068 ops_complete_compute, sh,
1069 to_addr_conv(sh, percpu));
1070 return async_gen_syndrome(blocks, 0, count+2,
1071 STRIPE_SIZE, &submit);
1074 init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
1075 ops_complete_compute, sh,
1076 to_addr_conv(sh, percpu));
1077 if (failb == syndrome_disks) {
1078 /* We're missing D+P. */
1079 return async_raid6_datap_recov(syndrome_disks+2,
1083 /* We're missing D+D. */
1084 return async_raid6_2data_recov(syndrome_disks+2,
1085 STRIPE_SIZE, faila, failb,
1092 static void ops_complete_prexor(void *stripe_head_ref)
1094 struct stripe_head *sh = stripe_head_ref;
1096 pr_debug("%s: stripe %llu\n", __func__,
1097 (unsigned long long)sh->sector);
1100 static struct dma_async_tx_descriptor *
1101 ops_run_prexor(struct stripe_head *sh, struct raid5_percpu *percpu,
1102 struct dma_async_tx_descriptor *tx)
1104 int disks = sh->disks;
1105 struct page **xor_srcs = percpu->scribble;
1106 int count = 0, pd_idx = sh->pd_idx, i;
1107 struct async_submit_ctl submit;
1109 /* existing parity data subtracted */
1110 struct page *xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
1112 pr_debug("%s: stripe %llu\n", __func__,
1113 (unsigned long long)sh->sector);
1115 for (i = disks; i--; ) {
1116 struct r5dev *dev = &sh->dev[i];
1117 /* Only process blocks that are known to be uptodate */
1118 if (test_bit(R5_Wantdrain, &dev->flags))
1119 xor_srcs[count++] = dev->page;
1122 init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_DROP_DST, tx,
1123 ops_complete_prexor, sh, to_addr_conv(sh, percpu));
1124 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
1129 static struct dma_async_tx_descriptor *
1130 ops_run_biodrain(struct stripe_head *sh, struct dma_async_tx_descriptor *tx)
1132 int disks = sh->disks;
1135 pr_debug("%s: stripe %llu\n", __func__,
1136 (unsigned long long)sh->sector);
1138 for (i = disks; i--; ) {
1139 struct r5dev *dev = &sh->dev[i];
1142 if (test_and_clear_bit(R5_Wantdrain, &dev->flags)) {
1145 spin_lock_irq(&sh->raid_conf->device_lock);
1146 chosen = dev->towrite;
1147 dev->towrite = NULL;
1148 BUG_ON(dev->written);
1149 wbi = dev->written = chosen;
1150 spin_unlock_irq(&sh->raid_conf->device_lock);
1152 while (wbi && wbi->bi_sector <
1153 dev->sector + STRIPE_SECTORS) {
1154 if (wbi->bi_rw & REQ_FUA)
1155 set_bit(R5_WantFUA, &dev->flags);
1156 if (wbi->bi_rw & REQ_SYNC)
1157 set_bit(R5_SyncIO, &dev->flags);
1158 tx = async_copy_data(1, wbi, dev->page,
1160 wbi = r5_next_bio(wbi, dev->sector);
1168 static void ops_complete_reconstruct(void *stripe_head_ref)
1170 struct stripe_head *sh = stripe_head_ref;
1171 int disks = sh->disks;
1172 int pd_idx = sh->pd_idx;
1173 int qd_idx = sh->qd_idx;
1175 bool fua = false, sync = false;
1177 pr_debug("%s: stripe %llu\n", __func__,
1178 (unsigned long long)sh->sector);
1180 for (i = disks; i--; ) {
1181 fua |= test_bit(R5_WantFUA, &sh->dev[i].flags);
1182 sync |= test_bit(R5_SyncIO, &sh->dev[i].flags);
1185 for (i = disks; i--; ) {
1186 struct r5dev *dev = &sh->dev[i];
1188 if (dev->written || i == pd_idx || i == qd_idx) {
1189 set_bit(R5_UPTODATE, &dev->flags);
1191 set_bit(R5_WantFUA, &dev->flags);
1193 set_bit(R5_SyncIO, &dev->flags);
1197 if (sh->reconstruct_state == reconstruct_state_drain_run)
1198 sh->reconstruct_state = reconstruct_state_drain_result;
1199 else if (sh->reconstruct_state == reconstruct_state_prexor_drain_run)
1200 sh->reconstruct_state = reconstruct_state_prexor_drain_result;
1202 BUG_ON(sh->reconstruct_state != reconstruct_state_run);
1203 sh->reconstruct_state = reconstruct_state_result;
1206 set_bit(STRIPE_HANDLE, &sh->state);
1211 ops_run_reconstruct5(struct stripe_head *sh, struct raid5_percpu *percpu,
1212 struct dma_async_tx_descriptor *tx)
1214 int disks = sh->disks;
1215 struct page **xor_srcs = percpu->scribble;
1216 struct async_submit_ctl submit;
1217 int count = 0, pd_idx = sh->pd_idx, i;
1218 struct page *xor_dest;
1220 unsigned long flags;
1222 pr_debug("%s: stripe %llu\n", __func__,
1223 (unsigned long long)sh->sector);
1225 /* check if prexor is active which means only process blocks
1226 * that are part of a read-modify-write (written)
1228 if (sh->reconstruct_state == reconstruct_state_prexor_drain_run) {
1230 xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
1231 for (i = disks; i--; ) {
1232 struct r5dev *dev = &sh->dev[i];
1234 xor_srcs[count++] = dev->page;
1237 xor_dest = sh->dev[pd_idx].page;
1238 for (i = disks; i--; ) {
1239 struct r5dev *dev = &sh->dev[i];
1241 xor_srcs[count++] = dev->page;
1245 /* 1/ if we prexor'd then the dest is reused as a source
1246 * 2/ if we did not prexor then we are redoing the parity
1247 * set ASYNC_TX_XOR_DROP_DST and ASYNC_TX_XOR_ZERO_DST
1248 * for the synchronous xor case
1250 flags = ASYNC_TX_ACK |
1251 (prexor ? ASYNC_TX_XOR_DROP_DST : ASYNC_TX_XOR_ZERO_DST);
1253 atomic_inc(&sh->count);
1255 init_async_submit(&submit, flags, tx, ops_complete_reconstruct, sh,
1256 to_addr_conv(sh, percpu));
1257 if (unlikely(count == 1))
1258 tx = async_memcpy(xor_dest, xor_srcs[0], 0, 0, STRIPE_SIZE, &submit);
1260 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
1264 ops_run_reconstruct6(struct stripe_head *sh, struct raid5_percpu *percpu,
1265 struct dma_async_tx_descriptor *tx)
1267 struct async_submit_ctl submit;
1268 struct page **blocks = percpu->scribble;
1271 pr_debug("%s: stripe %llu\n", __func__, (unsigned long long)sh->sector);
1273 count = set_syndrome_sources(blocks, sh);
1275 atomic_inc(&sh->count);
1277 init_async_submit(&submit, ASYNC_TX_ACK, tx, ops_complete_reconstruct,
1278 sh, to_addr_conv(sh, percpu));
1279 async_gen_syndrome(blocks, 0, count+2, STRIPE_SIZE, &submit);
1282 static void ops_complete_check(void *stripe_head_ref)
1284 struct stripe_head *sh = stripe_head_ref;
1286 pr_debug("%s: stripe %llu\n", __func__,
1287 (unsigned long long)sh->sector);
1289 sh->check_state = check_state_check_result;
1290 set_bit(STRIPE_HANDLE, &sh->state);
1294 static void ops_run_check_p(struct stripe_head *sh, struct raid5_percpu *percpu)
1296 int disks = sh->disks;
1297 int pd_idx = sh->pd_idx;
1298 int qd_idx = sh->qd_idx;
1299 struct page *xor_dest;
1300 struct page **xor_srcs = percpu->scribble;
1301 struct dma_async_tx_descriptor *tx;
1302 struct async_submit_ctl submit;
1306 pr_debug("%s: stripe %llu\n", __func__,
1307 (unsigned long long)sh->sector);
1310 xor_dest = sh->dev[pd_idx].page;
1311 xor_srcs[count++] = xor_dest;
1312 for (i = disks; i--; ) {
1313 if (i == pd_idx || i == qd_idx)
1315 xor_srcs[count++] = sh->dev[i].page;
1318 init_async_submit(&submit, 0, NULL, NULL, NULL,
1319 to_addr_conv(sh, percpu));
1320 tx = async_xor_val(xor_dest, xor_srcs, 0, count, STRIPE_SIZE,
1321 &sh->ops.zero_sum_result, &submit);
1323 atomic_inc(&sh->count);
1324 init_async_submit(&submit, ASYNC_TX_ACK, tx, ops_complete_check, sh, NULL);
1325 tx = async_trigger_callback(&submit);
1328 static void ops_run_check_pq(struct stripe_head *sh, struct raid5_percpu *percpu, int checkp)
1330 struct page **srcs = percpu->scribble;
1331 struct async_submit_ctl submit;
1334 pr_debug("%s: stripe %llu checkp: %d\n", __func__,
1335 (unsigned long long)sh->sector, checkp);
1337 count = set_syndrome_sources(srcs, sh);
1341 atomic_inc(&sh->count);
1342 init_async_submit(&submit, ASYNC_TX_ACK, NULL, ops_complete_check,
1343 sh, to_addr_conv(sh, percpu));
1344 async_syndrome_val(srcs, 0, count+2, STRIPE_SIZE,
1345 &sh->ops.zero_sum_result, percpu->spare_page, &submit);
1348 static void __raid_run_ops(struct stripe_head *sh, unsigned long ops_request)
1350 int overlap_clear = 0, i, disks = sh->disks;
1351 struct dma_async_tx_descriptor *tx = NULL;
1352 struct r5conf *conf = sh->raid_conf;
1353 int level = conf->level;
1354 struct raid5_percpu *percpu;
1358 percpu = per_cpu_ptr(conf->percpu, cpu);
1359 if (test_bit(STRIPE_OP_BIOFILL, &ops_request)) {
1360 ops_run_biofill(sh);
1364 if (test_bit(STRIPE_OP_COMPUTE_BLK, &ops_request)) {
1366 tx = ops_run_compute5(sh, percpu);
1368 if (sh->ops.target2 < 0 || sh->ops.target < 0)
1369 tx = ops_run_compute6_1(sh, percpu);
1371 tx = ops_run_compute6_2(sh, percpu);
1373 /* terminate the chain if reconstruct is not set to be run */
1374 if (tx && !test_bit(STRIPE_OP_RECONSTRUCT, &ops_request))
1378 if (test_bit(STRIPE_OP_PREXOR, &ops_request))
1379 tx = ops_run_prexor(sh, percpu, tx);
1381 if (test_bit(STRIPE_OP_BIODRAIN, &ops_request)) {
1382 tx = ops_run_biodrain(sh, tx);
1386 if (test_bit(STRIPE_OP_RECONSTRUCT, &ops_request)) {
1388 ops_run_reconstruct5(sh, percpu, tx);
1390 ops_run_reconstruct6(sh, percpu, tx);
1393 if (test_bit(STRIPE_OP_CHECK, &ops_request)) {
1394 if (sh->check_state == check_state_run)
1395 ops_run_check_p(sh, percpu);
1396 else if (sh->check_state == check_state_run_q)
1397 ops_run_check_pq(sh, percpu, 0);
1398 else if (sh->check_state == check_state_run_pq)
1399 ops_run_check_pq(sh, percpu, 1);
1405 for (i = disks; i--; ) {
1406 struct r5dev *dev = &sh->dev[i];
1407 if (test_and_clear_bit(R5_Overlap, &dev->flags))
1408 wake_up(&sh->raid_conf->wait_for_overlap);
1413 #ifdef CONFIG_MULTICORE_RAID456
1414 static void async_run_ops(void *param, async_cookie_t cookie)
1416 struct stripe_head *sh = param;
1417 unsigned long ops_request = sh->ops.request;
1419 clear_bit_unlock(STRIPE_OPS_REQ_PENDING, &sh->state);
1420 wake_up(&sh->ops.wait_for_ops);
1422 __raid_run_ops(sh, ops_request);
1426 static void raid_run_ops(struct stripe_head *sh, unsigned long ops_request)
1428 /* since handle_stripe can be called outside of raid5d context
1429 * we need to ensure sh->ops.request is de-staged before another
1432 wait_event(sh->ops.wait_for_ops,
1433 !test_and_set_bit_lock(STRIPE_OPS_REQ_PENDING, &sh->state));
1434 sh->ops.request = ops_request;
1436 atomic_inc(&sh->count);
1437 async_schedule(async_run_ops, sh);
1440 #define raid_run_ops __raid_run_ops
1443 static int grow_one_stripe(struct r5conf *conf)
1445 struct stripe_head *sh;
1446 sh = kmem_cache_zalloc(conf->slab_cache, GFP_KERNEL);
1450 sh->raid_conf = conf;
1451 #ifdef CONFIG_MULTICORE_RAID456
1452 init_waitqueue_head(&sh->ops.wait_for_ops);
1455 if (grow_buffers(sh)) {
1457 kmem_cache_free(conf->slab_cache, sh);
1460 /* we just created an active stripe so... */
1461 atomic_set(&sh->count, 1);
1462 atomic_inc(&conf->active_stripes);
1463 INIT_LIST_HEAD(&sh->lru);
1468 static int grow_stripes(struct r5conf *conf, int num)
1470 struct kmem_cache *sc;
1471 int devs = max(conf->raid_disks, conf->previous_raid_disks);
1473 if (conf->mddev->gendisk)
1474 sprintf(conf->cache_name[0],
1475 "raid%d-%s", conf->level, mdname(conf->mddev));
1477 sprintf(conf->cache_name[0],
1478 "raid%d-%p", conf->level, conf->mddev);
1479 sprintf(conf->cache_name[1], "%s-alt", conf->cache_name[0]);
1481 conf->active_name = 0;
1482 sc = kmem_cache_create(conf->cache_name[conf->active_name],
1483 sizeof(struct stripe_head)+(devs-1)*sizeof(struct r5dev),
1487 conf->slab_cache = sc;
1488 conf->pool_size = devs;
1490 if (!grow_one_stripe(conf))
1496 * scribble_len - return the required size of the scribble region
1497 * @num - total number of disks in the array
1499 * The size must be enough to contain:
1500 * 1/ a struct page pointer for each device in the array +2
1501 * 2/ room to convert each entry in (1) to its corresponding dma
1502 * (dma_map_page()) or page (page_address()) address.
1504 * Note: the +2 is for the destination buffers of the ddf/raid6 case where we
1505 * calculate over all devices (not just the data blocks), using zeros in place
1506 * of the P and Q blocks.
1508 static size_t scribble_len(int num)
1512 len = sizeof(struct page *) * (num+2) + sizeof(addr_conv_t) * (num+2);
1517 static int resize_stripes(struct r5conf *conf, int newsize)
1519 /* Make all the stripes able to hold 'newsize' devices.
1520 * New slots in each stripe get 'page' set to a new page.
1522 * This happens in stages:
1523 * 1/ create a new kmem_cache and allocate the required number of
1525 * 2/ gather all the old stripe_heads and tranfer the pages across
1526 * to the new stripe_heads. This will have the side effect of
1527 * freezing the array as once all stripe_heads have been collected,
1528 * no IO will be possible. Old stripe heads are freed once their
1529 * pages have been transferred over, and the old kmem_cache is
1530 * freed when all stripes are done.
1531 * 3/ reallocate conf->disks to be suitable bigger. If this fails,
1532 * we simple return a failre status - no need to clean anything up.
1533 * 4/ allocate new pages for the new slots in the new stripe_heads.
1534 * If this fails, we don't bother trying the shrink the
1535 * stripe_heads down again, we just leave them as they are.
1536 * As each stripe_head is processed the new one is released into
1539 * Once step2 is started, we cannot afford to wait for a write,
1540 * so we use GFP_NOIO allocations.
1542 struct stripe_head *osh, *nsh;
1543 LIST_HEAD(newstripes);
1544 struct disk_info *ndisks;
1547 struct kmem_cache *sc;
1550 if (newsize <= conf->pool_size)
1551 return 0; /* never bother to shrink */
1553 err = md_allow_write(conf->mddev);
1558 sc = kmem_cache_create(conf->cache_name[1-conf->active_name],
1559 sizeof(struct stripe_head)+(newsize-1)*sizeof(struct r5dev),
1564 for (i = conf->max_nr_stripes; i; i--) {
1565 nsh = kmem_cache_zalloc(sc, GFP_KERNEL);
1569 nsh->raid_conf = conf;
1570 #ifdef CONFIG_MULTICORE_RAID456
1571 init_waitqueue_head(&nsh->ops.wait_for_ops);
1574 list_add(&nsh->lru, &newstripes);
1577 /* didn't get enough, give up */
1578 while (!list_empty(&newstripes)) {
1579 nsh = list_entry(newstripes.next, struct stripe_head, lru);
1580 list_del(&nsh->lru);
1581 kmem_cache_free(sc, nsh);
1583 kmem_cache_destroy(sc);
1586 /* Step 2 - Must use GFP_NOIO now.
1587 * OK, we have enough stripes, start collecting inactive
1588 * stripes and copying them over
1590 list_for_each_entry(nsh, &newstripes, lru) {
1591 spin_lock_irq(&conf->device_lock);
1592 wait_event_lock_irq(conf->wait_for_stripe,
1593 !list_empty(&conf->inactive_list),
1596 osh = get_free_stripe(conf);
1597 spin_unlock_irq(&conf->device_lock);
1598 atomic_set(&nsh->count, 1);
1599 for(i=0; i<conf->pool_size; i++)
1600 nsh->dev[i].page = osh->dev[i].page;
1601 for( ; i<newsize; i++)
1602 nsh->dev[i].page = NULL;
1603 kmem_cache_free(conf->slab_cache, osh);
1605 kmem_cache_destroy(conf->slab_cache);
1608 * At this point, we are holding all the stripes so the array
1609 * is completely stalled, so now is a good time to resize
1610 * conf->disks and the scribble region
1612 ndisks = kzalloc(newsize * sizeof(struct disk_info), GFP_NOIO);
1614 for (i=0; i<conf->raid_disks; i++)
1615 ndisks[i] = conf->disks[i];
1617 conf->disks = ndisks;
1622 conf->scribble_len = scribble_len(newsize);
1623 for_each_present_cpu(cpu) {
1624 struct raid5_percpu *percpu;
1627 percpu = per_cpu_ptr(conf->percpu, cpu);
1628 scribble = kmalloc(conf->scribble_len, GFP_NOIO);
1631 kfree(percpu->scribble);
1632 percpu->scribble = scribble;
1640 /* Step 4, return new stripes to service */
1641 while(!list_empty(&newstripes)) {
1642 nsh = list_entry(newstripes.next, struct stripe_head, lru);
1643 list_del_init(&nsh->lru);
1645 for (i=conf->raid_disks; i < newsize; i++)
1646 if (nsh->dev[i].page == NULL) {
1647 struct page *p = alloc_page(GFP_NOIO);
1648 nsh->dev[i].page = p;
1652 release_stripe(nsh);
1654 /* critical section pass, GFP_NOIO no longer needed */
1656 conf->slab_cache = sc;
1657 conf->active_name = 1-conf->active_name;
1658 conf->pool_size = newsize;
1662 static int drop_one_stripe(struct r5conf *conf)
1664 struct stripe_head *sh;
1666 spin_lock_irq(&conf->device_lock);
1667 sh = get_free_stripe(conf);
1668 spin_unlock_irq(&conf->device_lock);
1671 BUG_ON(atomic_read(&sh->count));
1673 kmem_cache_free(conf->slab_cache, sh);
1674 atomic_dec(&conf->active_stripes);
1678 static void shrink_stripes(struct r5conf *conf)
1680 while (drop_one_stripe(conf))
1683 if (conf->slab_cache)
1684 kmem_cache_destroy(conf->slab_cache);
1685 conf->slab_cache = NULL;
1688 static void raid5_end_read_request(struct bio * bi, int error)
1690 struct stripe_head *sh = bi->bi_private;
1691 struct r5conf *conf = sh->raid_conf;
1692 int disks = sh->disks, i;
1693 int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
1694 char b[BDEVNAME_SIZE];
1695 struct md_rdev *rdev = NULL;
1698 for (i=0 ; i<disks; i++)
1699 if (bi == &sh->dev[i].req)
1702 pr_debug("end_read_request %llu/%d, count: %d, uptodate %d.\n",
1703 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
1709 if (test_bit(R5_ReadRepl, &sh->dev[i].flags))
1710 /* If replacement finished while this request was outstanding,
1711 * 'replacement' might be NULL already.
1712 * In that case it moved down to 'rdev'.
1713 * rdev is not removed until all requests are finished.
1715 rdev = conf->disks[i].replacement;
1717 rdev = conf->disks[i].rdev;
1719 if (use_new_offset(conf, sh))
1720 s = sh->sector + rdev->new_data_offset;
1722 s = sh->sector + rdev->data_offset;
1724 set_bit(R5_UPTODATE, &sh->dev[i].flags);
1725 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
1726 /* Note that this cannot happen on a
1727 * replacement device. We just fail those on
1732 "md/raid:%s: read error corrected"
1733 " (%lu sectors at %llu on %s)\n",
1734 mdname(conf->mddev), STRIPE_SECTORS,
1735 (unsigned long long)s,
1736 bdevname(rdev->bdev, b));
1737 atomic_add(STRIPE_SECTORS, &rdev->corrected_errors);
1738 clear_bit(R5_ReadError, &sh->dev[i].flags);
1739 clear_bit(R5_ReWrite, &sh->dev[i].flags);
1741 if (atomic_read(&rdev->read_errors))
1742 atomic_set(&rdev->read_errors, 0);
1744 const char *bdn = bdevname(rdev->bdev, b);
1748 clear_bit(R5_UPTODATE, &sh->dev[i].flags);
1749 atomic_inc(&rdev->read_errors);
1750 if (test_bit(R5_ReadRepl, &sh->dev[i].flags))
1753 "md/raid:%s: read error on replacement device "
1754 "(sector %llu on %s).\n",
1755 mdname(conf->mddev),
1756 (unsigned long long)s,
1758 else if (conf->mddev->degraded >= conf->max_degraded) {
1762 "md/raid:%s: read error not correctable "
1763 "(sector %llu on %s).\n",
1764 mdname(conf->mddev),
1765 (unsigned long long)s,
1767 } else if (test_bit(R5_ReWrite, &sh->dev[i].flags)) {
1772 "md/raid:%s: read error NOT corrected!! "
1773 "(sector %llu on %s).\n",
1774 mdname(conf->mddev),
1775 (unsigned long long)s,
1777 } else if (atomic_read(&rdev->read_errors)
1778 > conf->max_nr_stripes)
1780 "md/raid:%s: Too many read errors, failing device %s.\n",
1781 mdname(conf->mddev), bdn);
1785 set_bit(R5_ReadError, &sh->dev[i].flags);
1787 clear_bit(R5_ReadError, &sh->dev[i].flags);
1788 clear_bit(R5_ReWrite, &sh->dev[i].flags);
1790 && test_bit(In_sync, &rdev->flags)
1791 && rdev_set_badblocks(
1792 rdev, sh->sector, STRIPE_SECTORS, 0)))
1793 md_error(conf->mddev, rdev);
1796 rdev_dec_pending(rdev, conf->mddev);
1797 clear_bit(R5_LOCKED, &sh->dev[i].flags);
1798 set_bit(STRIPE_HANDLE, &sh->state);
1802 static void raid5_end_write_request(struct bio *bi, int error)
1804 struct stripe_head *sh = bi->bi_private;
1805 struct r5conf *conf = sh->raid_conf;
1806 int disks = sh->disks, i;
1807 struct md_rdev *uninitialized_var(rdev);
1808 int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
1811 int replacement = 0;
1813 for (i = 0 ; i < disks; i++) {
1814 if (bi == &sh->dev[i].req) {
1815 rdev = conf->disks[i].rdev;
1818 if (bi == &sh->dev[i].rreq) {
1819 rdev = conf->disks[i].replacement;
1823 /* rdev was removed and 'replacement'
1824 * replaced it. rdev is not removed
1825 * until all requests are finished.
1827 rdev = conf->disks[i].rdev;
1831 pr_debug("end_write_request %llu/%d, count %d, uptodate: %d.\n",
1832 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
1841 md_error(conf->mddev, rdev);
1842 else if (is_badblock(rdev, sh->sector,
1844 &first_bad, &bad_sectors))
1845 set_bit(R5_MadeGoodRepl, &sh->dev[i].flags);
1848 set_bit(WriteErrorSeen, &rdev->flags);
1849 set_bit(R5_WriteError, &sh->dev[i].flags);
1850 if (!test_and_set_bit(WantReplacement, &rdev->flags))
1851 set_bit(MD_RECOVERY_NEEDED,
1852 &rdev->mddev->recovery);
1853 } else if (is_badblock(rdev, sh->sector,
1855 &first_bad, &bad_sectors))
1856 set_bit(R5_MadeGood, &sh->dev[i].flags);
1858 rdev_dec_pending(rdev, conf->mddev);
1860 if (!test_and_clear_bit(R5_DOUBLE_LOCKED, &sh->dev[i].flags))
1861 clear_bit(R5_LOCKED, &sh->dev[i].flags);
1862 set_bit(STRIPE_HANDLE, &sh->state);
1866 static sector_t compute_blocknr(struct stripe_head *sh, int i, int previous);
1868 static void raid5_build_block(struct stripe_head *sh, int i, int previous)
1870 struct r5dev *dev = &sh->dev[i];
1872 bio_init(&dev->req);
1873 dev->req.bi_io_vec = &dev->vec;
1875 dev->req.bi_max_vecs++;
1876 dev->req.bi_private = sh;
1877 dev->vec.bv_page = dev->page;
1879 bio_init(&dev->rreq);
1880 dev->rreq.bi_io_vec = &dev->rvec;
1881 dev->rreq.bi_vcnt++;
1882 dev->rreq.bi_max_vecs++;
1883 dev->rreq.bi_private = sh;
1884 dev->rvec.bv_page = dev->page;
1887 dev->sector = compute_blocknr(sh, i, previous);
1890 static void error(struct mddev *mddev, struct md_rdev *rdev)
1892 char b[BDEVNAME_SIZE];
1893 struct r5conf *conf = mddev->private;
1894 unsigned long flags;
1895 pr_debug("raid456: error called\n");
1897 spin_lock_irqsave(&conf->device_lock, flags);
1898 clear_bit(In_sync, &rdev->flags);
1899 mddev->degraded = calc_degraded(conf);
1900 spin_unlock_irqrestore(&conf->device_lock, flags);
1901 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1903 set_bit(Blocked, &rdev->flags);
1904 set_bit(Faulty, &rdev->flags);
1905 set_bit(MD_CHANGE_DEVS, &mddev->flags);
1907 "md/raid:%s: Disk failure on %s, disabling device.\n"
1908 "md/raid:%s: Operation continuing on %d devices.\n",
1910 bdevname(rdev->bdev, b),
1912 conf->raid_disks - mddev->degraded);
1916 * Input: a 'big' sector number,
1917 * Output: index of the data and parity disk, and the sector # in them.
1919 static sector_t raid5_compute_sector(struct r5conf *conf, sector_t r_sector,
1920 int previous, int *dd_idx,
1921 struct stripe_head *sh)
1923 sector_t stripe, stripe2;
1924 sector_t chunk_number;
1925 unsigned int chunk_offset;
1928 sector_t new_sector;
1929 int algorithm = previous ? conf->prev_algo
1931 int sectors_per_chunk = previous ? conf->prev_chunk_sectors
1932 : conf->chunk_sectors;
1933 int raid_disks = previous ? conf->previous_raid_disks
1935 int data_disks = raid_disks - conf->max_degraded;
1937 /* First compute the information on this sector */
1940 * Compute the chunk number and the sector offset inside the chunk
1942 chunk_offset = sector_div(r_sector, sectors_per_chunk);
1943 chunk_number = r_sector;
1946 * Compute the stripe number
1948 stripe = chunk_number;
1949 *dd_idx = sector_div(stripe, data_disks);
1952 * Select the parity disk based on the user selected algorithm.
1954 pd_idx = qd_idx = -1;
1955 switch(conf->level) {
1957 pd_idx = data_disks;
1960 switch (algorithm) {
1961 case ALGORITHM_LEFT_ASYMMETRIC:
1962 pd_idx = data_disks - sector_div(stripe2, raid_disks);
1963 if (*dd_idx >= pd_idx)
1966 case ALGORITHM_RIGHT_ASYMMETRIC:
1967 pd_idx = sector_div(stripe2, raid_disks);
1968 if (*dd_idx >= pd_idx)
1971 case ALGORITHM_LEFT_SYMMETRIC:
1972 pd_idx = data_disks - sector_div(stripe2, raid_disks);
1973 *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
1975 case ALGORITHM_RIGHT_SYMMETRIC:
1976 pd_idx = sector_div(stripe2, raid_disks);
1977 *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
1979 case ALGORITHM_PARITY_0:
1983 case ALGORITHM_PARITY_N:
1984 pd_idx = data_disks;
1992 switch (algorithm) {
1993 case ALGORITHM_LEFT_ASYMMETRIC:
1994 pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
1995 qd_idx = pd_idx + 1;
1996 if (pd_idx == raid_disks-1) {
1997 (*dd_idx)++; /* Q D D D P */
1999 } else if (*dd_idx >= pd_idx)
2000 (*dd_idx) += 2; /* D D P Q D */
2002 case ALGORITHM_RIGHT_ASYMMETRIC:
2003 pd_idx = sector_div(stripe2, raid_disks);
2004 qd_idx = pd_idx + 1;
2005 if (pd_idx == raid_disks-1) {
2006 (*dd_idx)++; /* Q D D D P */
2008 } else if (*dd_idx >= pd_idx)
2009 (*dd_idx) += 2; /* D D P Q D */
2011 case ALGORITHM_LEFT_SYMMETRIC:
2012 pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
2013 qd_idx = (pd_idx + 1) % raid_disks;
2014 *dd_idx = (pd_idx + 2 + *dd_idx) % raid_disks;
2016 case ALGORITHM_RIGHT_SYMMETRIC:
2017 pd_idx = sector_div(stripe2, raid_disks);
2018 qd_idx = (pd_idx + 1) % raid_disks;
2019 *dd_idx = (pd_idx + 2 + *dd_idx) % raid_disks;
2022 case ALGORITHM_PARITY_0:
2027 case ALGORITHM_PARITY_N:
2028 pd_idx = data_disks;
2029 qd_idx = data_disks + 1;
2032 case ALGORITHM_ROTATING_ZERO_RESTART:
2033 /* Exactly the same as RIGHT_ASYMMETRIC, but or
2034 * of blocks for computing Q is different.
2036 pd_idx = sector_div(stripe2, raid_disks);
2037 qd_idx = pd_idx + 1;
2038 if (pd_idx == raid_disks-1) {
2039 (*dd_idx)++; /* Q D D D P */
2041 } else if (*dd_idx >= pd_idx)
2042 (*dd_idx) += 2; /* D D P Q D */
2046 case ALGORITHM_ROTATING_N_RESTART:
2047 /* Same a left_asymmetric, by first stripe is
2048 * D D D P Q rather than
2052 pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
2053 qd_idx = pd_idx + 1;
2054 if (pd_idx == raid_disks-1) {
2055 (*dd_idx)++; /* Q D D D P */
2057 } else if (*dd_idx >= pd_idx)
2058 (*dd_idx) += 2; /* D D P Q D */
2062 case ALGORITHM_ROTATING_N_CONTINUE:
2063 /* Same as left_symmetric but Q is before P */
2064 pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
2065 qd_idx = (pd_idx + raid_disks - 1) % raid_disks;
2066 *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
2070 case ALGORITHM_LEFT_ASYMMETRIC_6:
2071 /* RAID5 left_asymmetric, with Q on last device */
2072 pd_idx = data_disks - sector_div(stripe2, raid_disks-1);
2073 if (*dd_idx >= pd_idx)
2075 qd_idx = raid_disks - 1;
2078 case ALGORITHM_RIGHT_ASYMMETRIC_6:
2079 pd_idx = sector_div(stripe2, raid_disks-1);
2080 if (*dd_idx >= pd_idx)
2082 qd_idx = raid_disks - 1;
2085 case ALGORITHM_LEFT_SYMMETRIC_6:
2086 pd_idx = data_disks - sector_div(stripe2, raid_disks-1);
2087 *dd_idx = (pd_idx + 1 + *dd_idx) % (raid_disks-1);
2088 qd_idx = raid_disks - 1;
2091 case ALGORITHM_RIGHT_SYMMETRIC_6:
2092 pd_idx = sector_div(stripe2, raid_disks-1);
2093 *dd_idx = (pd_idx + 1 + *dd_idx) % (raid_disks-1);
2094 qd_idx = raid_disks - 1;
2097 case ALGORITHM_PARITY_0_6:
2100 qd_idx = raid_disks - 1;
2110 sh->pd_idx = pd_idx;
2111 sh->qd_idx = qd_idx;
2112 sh->ddf_layout = ddf_layout;
2115 * Finally, compute the new sector number
2117 new_sector = (sector_t)stripe * sectors_per_chunk + chunk_offset;
2122 static sector_t compute_blocknr(struct stripe_head *sh, int i, int previous)
2124 struct r5conf *conf = sh->raid_conf;
2125 int raid_disks = sh->disks;
2126 int data_disks = raid_disks - conf->max_degraded;
2127 sector_t new_sector = sh->sector, check;
2128 int sectors_per_chunk = previous ? conf->prev_chunk_sectors
2129 : conf->chunk_sectors;
2130 int algorithm = previous ? conf->prev_algo
2134 sector_t chunk_number;
2135 int dummy1, dd_idx = i;
2137 struct stripe_head sh2;
2140 chunk_offset = sector_div(new_sector, sectors_per_chunk);
2141 stripe = new_sector;
2143 if (i == sh->pd_idx)
2145 switch(conf->level) {
2148 switch (algorithm) {
2149 case ALGORITHM_LEFT_ASYMMETRIC:
2150 case ALGORITHM_RIGHT_ASYMMETRIC:
2154 case ALGORITHM_LEFT_SYMMETRIC:
2155 case ALGORITHM_RIGHT_SYMMETRIC:
2158 i -= (sh->pd_idx + 1);
2160 case ALGORITHM_PARITY_0:
2163 case ALGORITHM_PARITY_N:
2170 if (i == sh->qd_idx)
2171 return 0; /* It is the Q disk */
2172 switch (algorithm) {
2173 case ALGORITHM_LEFT_ASYMMETRIC:
2174 case ALGORITHM_RIGHT_ASYMMETRIC:
2175 case ALGORITHM_ROTATING_ZERO_RESTART:
2176 case ALGORITHM_ROTATING_N_RESTART:
2177 if (sh->pd_idx == raid_disks-1)
2178 i--; /* Q D D D P */
2179 else if (i > sh->pd_idx)
2180 i -= 2; /* D D P Q D */
2182 case ALGORITHM_LEFT_SYMMETRIC:
2183 case ALGORITHM_RIGHT_SYMMETRIC:
2184 if (sh->pd_idx == raid_disks-1)
2185 i--; /* Q D D D P */
2190 i -= (sh->pd_idx + 2);
2193 case ALGORITHM_PARITY_0:
2196 case ALGORITHM_PARITY_N:
2198 case ALGORITHM_ROTATING_N_CONTINUE:
2199 /* Like left_symmetric, but P is before Q */
2200 if (sh->pd_idx == 0)
2201 i--; /* P D D D Q */
2206 i -= (sh->pd_idx + 1);
2209 case ALGORITHM_LEFT_ASYMMETRIC_6:
2210 case ALGORITHM_RIGHT_ASYMMETRIC_6:
2214 case ALGORITHM_LEFT_SYMMETRIC_6:
2215 case ALGORITHM_RIGHT_SYMMETRIC_6:
2217 i += data_disks + 1;
2218 i -= (sh->pd_idx + 1);
2220 case ALGORITHM_PARITY_0_6:
2229 chunk_number = stripe * data_disks + i;
2230 r_sector = chunk_number * sectors_per_chunk + chunk_offset;
2232 check = raid5_compute_sector(conf, r_sector,
2233 previous, &dummy1, &sh2);
2234 if (check != sh->sector || dummy1 != dd_idx || sh2.pd_idx != sh->pd_idx
2235 || sh2.qd_idx != sh->qd_idx) {
2236 printk(KERN_ERR "md/raid:%s: compute_blocknr: map not correct\n",
2237 mdname(conf->mddev));
2245 schedule_reconstruction(struct stripe_head *sh, struct stripe_head_state *s,
2246 int rcw, int expand)
2248 int i, pd_idx = sh->pd_idx, disks = sh->disks;
2249 struct r5conf *conf = sh->raid_conf;
2250 int level = conf->level;
2253 /* if we are not expanding this is a proper write request, and
2254 * there will be bios with new data to be drained into the
2258 sh->reconstruct_state = reconstruct_state_drain_run;
2259 set_bit(STRIPE_OP_BIODRAIN, &s->ops_request);
2261 sh->reconstruct_state = reconstruct_state_run;
2263 set_bit(STRIPE_OP_RECONSTRUCT, &s->ops_request);
2265 for (i = disks; i--; ) {
2266 struct r5dev *dev = &sh->dev[i];
2269 set_bit(R5_LOCKED, &dev->flags);
2270 set_bit(R5_Wantdrain, &dev->flags);
2272 clear_bit(R5_UPTODATE, &dev->flags);
2276 if (s->locked + conf->max_degraded == disks)
2277 if (!test_and_set_bit(STRIPE_FULL_WRITE, &sh->state))
2278 atomic_inc(&conf->pending_full_writes);
2281 BUG_ON(!(test_bit(R5_UPTODATE, &sh->dev[pd_idx].flags) ||
2282 test_bit(R5_Wantcompute, &sh->dev[pd_idx].flags)));
2284 sh->reconstruct_state = reconstruct_state_prexor_drain_run;
2285 set_bit(STRIPE_OP_PREXOR, &s->ops_request);
2286 set_bit(STRIPE_OP_BIODRAIN, &s->ops_request);
2287 set_bit(STRIPE_OP_RECONSTRUCT, &s->ops_request);
2289 for (i = disks; i--; ) {
2290 struct r5dev *dev = &sh->dev[i];
2295 (test_bit(R5_UPTODATE, &dev->flags) ||
2296 test_bit(R5_Wantcompute, &dev->flags))) {
2297 set_bit(R5_Wantdrain, &dev->flags);
2298 set_bit(R5_LOCKED, &dev->flags);
2299 clear_bit(R5_UPTODATE, &dev->flags);
2305 /* keep the parity disk(s) locked while asynchronous operations
2308 set_bit(R5_LOCKED, &sh->dev[pd_idx].flags);
2309 clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
2313 int qd_idx = sh->qd_idx;
2314 struct r5dev *dev = &sh->dev[qd_idx];
2316 set_bit(R5_LOCKED, &dev->flags);
2317 clear_bit(R5_UPTODATE, &dev->flags);
2321 pr_debug("%s: stripe %llu locked: %d ops_request: %lx\n",
2322 __func__, (unsigned long long)sh->sector,
2323 s->locked, s->ops_request);
2327 * Each stripe/dev can have one or more bion attached.
2328 * toread/towrite point to the first in a chain.
2329 * The bi_next chain must be in order.
2331 static int add_stripe_bio(struct stripe_head *sh, struct bio *bi, int dd_idx, int forwrite)
2334 struct r5conf *conf = sh->raid_conf;
2337 pr_debug("adding bi b#%llu to stripe s#%llu\n",
2338 (unsigned long long)bi->bi_sector,
2339 (unsigned long long)sh->sector);
2342 spin_lock_irq(&conf->device_lock);
2344 bip = &sh->dev[dd_idx].towrite;
2345 if (*bip == NULL && sh->dev[dd_idx].written == NULL)
2348 bip = &sh->dev[dd_idx].toread;
2349 while (*bip && (*bip)->bi_sector < bi->bi_sector) {
2350 if ((*bip)->bi_sector + ((*bip)->bi_size >> 9) > bi->bi_sector)
2352 bip = & (*bip)->bi_next;
2354 if (*bip && (*bip)->bi_sector < bi->bi_sector + ((bi->bi_size)>>9))
2357 BUG_ON(*bip && bi->bi_next && (*bip) != bi->bi_next);
2361 bi->bi_phys_segments++;
2364 /* check if page is covered */
2365 sector_t sector = sh->dev[dd_idx].sector;
2366 for (bi=sh->dev[dd_idx].towrite;
2367 sector < sh->dev[dd_idx].sector + STRIPE_SECTORS &&
2368 bi && bi->bi_sector <= sector;
2369 bi = r5_next_bio(bi, sh->dev[dd_idx].sector)) {
2370 if (bi->bi_sector + (bi->bi_size>>9) >= sector)
2371 sector = bi->bi_sector + (bi->bi_size>>9);
2373 if (sector >= sh->dev[dd_idx].sector + STRIPE_SECTORS)
2374 set_bit(R5_OVERWRITE, &sh->dev[dd_idx].flags);
2376 spin_unlock_irq(&conf->device_lock);
2378 pr_debug("added bi b#%llu to stripe s#%llu, disk %d.\n",
2379 (unsigned long long)(*bip)->bi_sector,
2380 (unsigned long long)sh->sector, dd_idx);
2382 if (conf->mddev->bitmap && firstwrite) {
2383 bitmap_startwrite(conf->mddev->bitmap, sh->sector,
2385 sh->bm_seq = conf->seq_flush+1;
2386 set_bit(STRIPE_BIT_DELAY, &sh->state);
2391 set_bit(R5_Overlap, &sh->dev[dd_idx].flags);
2392 spin_unlock_irq(&conf->device_lock);
2396 static void end_reshape(struct r5conf *conf);
2398 static void stripe_set_idx(sector_t stripe, struct r5conf *conf, int previous,
2399 struct stripe_head *sh)
2401 int sectors_per_chunk =
2402 previous ? conf->prev_chunk_sectors : conf->chunk_sectors;
2404 int chunk_offset = sector_div(stripe, sectors_per_chunk);
2405 int disks = previous ? conf->previous_raid_disks : conf->raid_disks;
2407 raid5_compute_sector(conf,
2408 stripe * (disks - conf->max_degraded)
2409 *sectors_per_chunk + chunk_offset,
2415 handle_failed_stripe(struct r5conf *conf, struct stripe_head *sh,
2416 struct stripe_head_state *s, int disks,
2417 struct bio **return_bi)
2420 for (i = disks; i--; ) {
2424 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
2425 struct md_rdev *rdev;
2427 rdev = rcu_dereference(conf->disks[i].rdev);
2428 if (rdev && test_bit(In_sync, &rdev->flags))
2429 atomic_inc(&rdev->nr_pending);
2434 if (!rdev_set_badblocks(
2438 md_error(conf->mddev, rdev);
2439 rdev_dec_pending(rdev, conf->mddev);
2442 spin_lock_irq(&conf->device_lock);
2443 /* fail all writes first */
2444 bi = sh->dev[i].towrite;
2445 sh->dev[i].towrite = NULL;
2451 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
2452 wake_up(&conf->wait_for_overlap);
2454 while (bi && bi->bi_sector <
2455 sh->dev[i].sector + STRIPE_SECTORS) {
2456 struct bio *nextbi = r5_next_bio(bi, sh->dev[i].sector);
2457 clear_bit(BIO_UPTODATE, &bi->bi_flags);
2458 if (!raid5_dec_bi_phys_segments(bi)) {
2459 md_write_end(conf->mddev);
2460 bi->bi_next = *return_bi;
2465 /* and fail all 'written' */
2466 bi = sh->dev[i].written;
2467 sh->dev[i].written = NULL;
2468 if (bi) bitmap_end = 1;
2469 while (bi && bi->bi_sector <
2470 sh->dev[i].sector + STRIPE_SECTORS) {
2471 struct bio *bi2 = r5_next_bio(bi, sh->dev[i].sector);
2472 clear_bit(BIO_UPTODATE, &bi->bi_flags);
2473 if (!raid5_dec_bi_phys_segments(bi)) {
2474 md_write_end(conf->mddev);
2475 bi->bi_next = *return_bi;
2481 /* fail any reads if this device is non-operational and
2482 * the data has not reached the cache yet.
2484 if (!test_bit(R5_Wantfill, &sh->dev[i].flags) &&
2485 (!test_bit(R5_Insync, &sh->dev[i].flags) ||
2486 test_bit(R5_ReadError, &sh->dev[i].flags))) {
2487 bi = sh->dev[i].toread;
2488 sh->dev[i].toread = NULL;
2489 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
2490 wake_up(&conf->wait_for_overlap);
2491 if (bi) s->to_read--;
2492 while (bi && bi->bi_sector <
2493 sh->dev[i].sector + STRIPE_SECTORS) {
2494 struct bio *nextbi =
2495 r5_next_bio(bi, sh->dev[i].sector);
2496 clear_bit(BIO_UPTODATE, &bi->bi_flags);
2497 if (!raid5_dec_bi_phys_segments(bi)) {
2498 bi->bi_next = *return_bi;
2504 spin_unlock_irq(&conf->device_lock);
2506 bitmap_endwrite(conf->mddev->bitmap, sh->sector,
2507 STRIPE_SECTORS, 0, 0);
2508 /* If we were in the middle of a write the parity block might
2509 * still be locked - so just clear all R5_LOCKED flags
2511 clear_bit(R5_LOCKED, &sh->dev[i].flags);
2514 if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
2515 if (atomic_dec_and_test(&conf->pending_full_writes))
2516 md_wakeup_thread(conf->mddev->thread);
2520 handle_failed_sync(struct r5conf *conf, struct stripe_head *sh,
2521 struct stripe_head_state *s)
2526 clear_bit(STRIPE_SYNCING, &sh->state);
2529 /* There is nothing more to do for sync/check/repair.
2530 * Don't even need to abort as that is handled elsewhere
2531 * if needed, and not always wanted e.g. if there is a known
2533 * For recover/replace we need to record a bad block on all
2534 * non-sync devices, or abort the recovery
2536 if (test_bit(MD_RECOVERY_RECOVER, &conf->mddev->recovery)) {
2537 /* During recovery devices cannot be removed, so
2538 * locking and refcounting of rdevs is not needed
2540 for (i = 0; i < conf->raid_disks; i++) {
2541 struct md_rdev *rdev = conf->disks[i].rdev;
2543 && !test_bit(Faulty, &rdev->flags)
2544 && !test_bit(In_sync, &rdev->flags)
2545 && !rdev_set_badblocks(rdev, sh->sector,
2548 rdev = conf->disks[i].replacement;
2550 && !test_bit(Faulty, &rdev->flags)
2551 && !test_bit(In_sync, &rdev->flags)
2552 && !rdev_set_badblocks(rdev, sh->sector,
2557 conf->recovery_disabled =
2558 conf->mddev->recovery_disabled;
2560 md_done_sync(conf->mddev, STRIPE_SECTORS, !abort);
2563 static int want_replace(struct stripe_head *sh, int disk_idx)
2565 struct md_rdev *rdev;
2567 /* Doing recovery so rcu locking not required */
2568 rdev = sh->raid_conf->disks[disk_idx].replacement;
2570 && !test_bit(Faulty, &rdev->flags)
2571 && !test_bit(In_sync, &rdev->flags)
2572 && (rdev->recovery_offset <= sh->sector
2573 || rdev->mddev->recovery_cp <= sh->sector))
2579 /* fetch_block - checks the given member device to see if its data needs
2580 * to be read or computed to satisfy a request.
2582 * Returns 1 when no more member devices need to be checked, otherwise returns
2583 * 0 to tell the loop in handle_stripe_fill to continue
2585 static int fetch_block(struct stripe_head *sh, struct stripe_head_state *s,
2586 int disk_idx, int disks)
2588 struct r5dev *dev = &sh->dev[disk_idx];
2589 struct r5dev *fdev[2] = { &sh->dev[s->failed_num[0]],
2590 &sh->dev[s->failed_num[1]] };
2592 /* is the data in this block needed, and can we get it? */
2593 if (!test_bit(R5_LOCKED, &dev->flags) &&
2594 !test_bit(R5_UPTODATE, &dev->flags) &&
2596 (dev->towrite && !test_bit(R5_OVERWRITE, &dev->flags)) ||
2597 s->syncing || s->expanding ||
2598 (s->replacing && want_replace(sh, disk_idx)) ||
2599 (s->failed >= 1 && fdev[0]->toread) ||
2600 (s->failed >= 2 && fdev[1]->toread) ||
2601 (sh->raid_conf->level <= 5 && s->failed && fdev[0]->towrite &&
2602 !test_bit(R5_OVERWRITE, &fdev[0]->flags)) ||
2603 (sh->raid_conf->level == 6 && s->failed && s->to_write))) {
2604 /* we would like to get this block, possibly by computing it,
2605 * otherwise read it if the backing disk is insync
2607 BUG_ON(test_bit(R5_Wantcompute, &dev->flags));
2608 BUG_ON(test_bit(R5_Wantread, &dev->flags));
2609 if ((s->uptodate == disks - 1) &&
2610 (s->failed && (disk_idx == s->failed_num[0] ||
2611 disk_idx == s->failed_num[1]))) {
2612 /* have disk failed, and we're requested to fetch it;
2615 pr_debug("Computing stripe %llu block %d\n",
2616 (unsigned long long)sh->sector, disk_idx);
2617 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
2618 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
2619 set_bit(R5_Wantcompute, &dev->flags);
2620 sh->ops.target = disk_idx;
2621 sh->ops.target2 = -1; /* no 2nd target */
2623 /* Careful: from this point on 'uptodate' is in the eye
2624 * of raid_run_ops which services 'compute' operations
2625 * before writes. R5_Wantcompute flags a block that will
2626 * be R5_UPTODATE by the time it is needed for a
2627 * subsequent operation.
2631 } else if (s->uptodate == disks-2 && s->failed >= 2) {
2632 /* Computing 2-failure is *very* expensive; only
2633 * do it if failed >= 2
2636 for (other = disks; other--; ) {
2637 if (other == disk_idx)
2639 if (!test_bit(R5_UPTODATE,
2640 &sh->dev[other].flags))
2644 pr_debug("Computing stripe %llu blocks %d,%d\n",
2645 (unsigned long long)sh->sector,
2647 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
2648 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
2649 set_bit(R5_Wantcompute, &sh->dev[disk_idx].flags);
2650 set_bit(R5_Wantcompute, &sh->dev[other].flags);
2651 sh->ops.target = disk_idx;
2652 sh->ops.target2 = other;
2656 } else if (test_bit(R5_Insync, &dev->flags)) {
2657 set_bit(R5_LOCKED, &dev->flags);
2658 set_bit(R5_Wantread, &dev->flags);
2660 pr_debug("Reading block %d (sync=%d)\n",
2661 disk_idx, s->syncing);
2669 * handle_stripe_fill - read or compute data to satisfy pending requests.
2671 static void handle_stripe_fill(struct stripe_head *sh,
2672 struct stripe_head_state *s,
2677 /* look for blocks to read/compute, skip this if a compute
2678 * is already in flight, or if the stripe contents are in the
2679 * midst of changing due to a write
2681 if (!test_bit(STRIPE_COMPUTE_RUN, &sh->state) && !sh->check_state &&
2682 !sh->reconstruct_state)
2683 for (i = disks; i--; )
2684 if (fetch_block(sh, s, i, disks))
2686 set_bit(STRIPE_HANDLE, &sh->state);
2690 /* handle_stripe_clean_event
2691 * any written block on an uptodate or failed drive can be returned.
2692 * Note that if we 'wrote' to a failed drive, it will be UPTODATE, but
2693 * never LOCKED, so we don't need to test 'failed' directly.
2695 static void handle_stripe_clean_event(struct r5conf *conf,
2696 struct stripe_head *sh, int disks, struct bio **return_bi)
2701 for (i = disks; i--; )
2702 if (sh->dev[i].written) {
2704 if (!test_bit(R5_LOCKED, &dev->flags) &&
2705 test_bit(R5_UPTODATE, &dev->flags)) {
2706 /* We can return any write requests */
2707 struct bio *wbi, *wbi2;
2709 pr_debug("Return write for disc %d\n", i);
2710 spin_lock_irq(&conf->device_lock);
2712 dev->written = NULL;
2713 while (wbi && wbi->bi_sector <
2714 dev->sector + STRIPE_SECTORS) {
2715 wbi2 = r5_next_bio(wbi, dev->sector);
2716 if (!raid5_dec_bi_phys_segments(wbi)) {
2717 md_write_end(conf->mddev);
2718 wbi->bi_next = *return_bi;
2723 if (dev->towrite == NULL)
2725 spin_unlock_irq(&conf->device_lock);
2727 bitmap_endwrite(conf->mddev->bitmap,
2730 !test_bit(STRIPE_DEGRADED, &sh->state),
2735 if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
2736 if (atomic_dec_and_test(&conf->pending_full_writes))
2737 md_wakeup_thread(conf->mddev->thread);
2740 static void handle_stripe_dirtying(struct r5conf *conf,
2741 struct stripe_head *sh,
2742 struct stripe_head_state *s,
2745 int rmw = 0, rcw = 0, i;
2746 if (conf->max_degraded == 2) {
2747 /* RAID6 requires 'rcw' in current implementation
2748 * Calculate the real rcw later - for now fake it
2749 * look like rcw is cheaper
2752 } else for (i = disks; i--; ) {
2753 /* would I have to read this buffer for read_modify_write */
2754 struct r5dev *dev = &sh->dev[i];
2755 if ((dev->towrite || i == sh->pd_idx) &&
2756 !test_bit(R5_LOCKED, &dev->flags) &&
2757 !(test_bit(R5_UPTODATE, &dev->flags) ||
2758 test_bit(R5_Wantcompute, &dev->flags))) {
2759 if (test_bit(R5_Insync, &dev->flags))
2762 rmw += 2*disks; /* cannot read it */
2764 /* Would I have to read this buffer for reconstruct_write */
2765 if (!test_bit(R5_OVERWRITE, &dev->flags) && i != sh->pd_idx &&
2766 !test_bit(R5_LOCKED, &dev->flags) &&
2767 !(test_bit(R5_UPTODATE, &dev->flags) ||
2768 test_bit(R5_Wantcompute, &dev->flags))) {
2769 if (test_bit(R5_Insync, &dev->flags)) rcw++;
2774 pr_debug("for sector %llu, rmw=%d rcw=%d\n",
2775 (unsigned long long)sh->sector, rmw, rcw);
2776 set_bit(STRIPE_HANDLE, &sh->state);
2777 if (rmw < rcw && rmw > 0)
2778 /* prefer read-modify-write, but need to get some data */
2779 for (i = disks; i--; ) {
2780 struct r5dev *dev = &sh->dev[i];
2781 if ((dev->towrite || i == sh->pd_idx) &&
2782 !test_bit(R5_LOCKED, &dev->flags) &&
2783 !(test_bit(R5_UPTODATE, &dev->flags) ||
2784 test_bit(R5_Wantcompute, &dev->flags)) &&
2785 test_bit(R5_Insync, &dev->flags)) {
2787 test_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2788 pr_debug("Read_old block "
2789 "%d for r-m-w\n", i);
2790 set_bit(R5_LOCKED, &dev->flags);
2791 set_bit(R5_Wantread, &dev->flags);
2794 set_bit(STRIPE_DELAYED, &sh->state);
2795 set_bit(STRIPE_HANDLE, &sh->state);
2799 if (rcw <= rmw && rcw > 0) {
2800 /* want reconstruct write, but need to get some data */
2802 for (i = disks; i--; ) {
2803 struct r5dev *dev = &sh->dev[i];
2804 if (!test_bit(R5_OVERWRITE, &dev->flags) &&
2805 i != sh->pd_idx && i != sh->qd_idx &&
2806 !test_bit(R5_LOCKED, &dev->flags) &&
2807 !(test_bit(R5_UPTODATE, &dev->flags) ||
2808 test_bit(R5_Wantcompute, &dev->flags))) {
2810 if (!test_bit(R5_Insync, &dev->flags))
2811 continue; /* it's a failed drive */
2813 test_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2814 pr_debug("Read_old block "
2815 "%d for Reconstruct\n", i);
2816 set_bit(R5_LOCKED, &dev->flags);
2817 set_bit(R5_Wantread, &dev->flags);
2820 set_bit(STRIPE_DELAYED, &sh->state);
2821 set_bit(STRIPE_HANDLE, &sh->state);
2826 /* now if nothing is locked, and if we have enough data,
2827 * we can start a write request
2829 /* since handle_stripe can be called at any time we need to handle the
2830 * case where a compute block operation has been submitted and then a
2831 * subsequent call wants to start a write request. raid_run_ops only
2832 * handles the case where compute block and reconstruct are requested
2833 * simultaneously. If this is not the case then new writes need to be
2834 * held off until the compute completes.
2836 if ((s->req_compute || !test_bit(STRIPE_COMPUTE_RUN, &sh->state)) &&
2837 (s->locked == 0 && (rcw == 0 || rmw == 0) &&
2838 !test_bit(STRIPE_BIT_DELAY, &sh->state)))
2839 schedule_reconstruction(sh, s, rcw == 0, 0);
2842 static void handle_parity_checks5(struct r5conf *conf, struct stripe_head *sh,
2843 struct stripe_head_state *s, int disks)
2845 struct r5dev *dev = NULL;
2847 set_bit(STRIPE_HANDLE, &sh->state);
2849 switch (sh->check_state) {
2850 case check_state_idle:
2851 /* start a new check operation if there are no failures */
2852 if (s->failed == 0) {
2853 BUG_ON(s->uptodate != disks);
2854 sh->check_state = check_state_run;
2855 set_bit(STRIPE_OP_CHECK, &s->ops_request);
2856 clear_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags);
2860 dev = &sh->dev[s->failed_num[0]];
2862 case check_state_compute_result:
2863 sh->check_state = check_state_idle;
2865 dev = &sh->dev[sh->pd_idx];
2867 /* check that a write has not made the stripe insync */
2868 if (test_bit(STRIPE_INSYNC, &sh->state))
2871 /* either failed parity check, or recovery is happening */
2872 BUG_ON(!test_bit(R5_UPTODATE, &dev->flags));
2873 BUG_ON(s->uptodate != disks);
2875 set_bit(R5_LOCKED, &dev->flags);
2877 set_bit(R5_Wantwrite, &dev->flags);
2879 clear_bit(STRIPE_DEGRADED, &sh->state);
2880 set_bit(STRIPE_INSYNC, &sh->state);
2882 case check_state_run:
2883 break; /* we will be called again upon completion */
2884 case check_state_check_result:
2885 sh->check_state = check_state_idle;
2887 /* if a failure occurred during the check operation, leave
2888 * STRIPE_INSYNC not set and let the stripe be handled again
2893 /* handle a successful check operation, if parity is correct
2894 * we are done. Otherwise update the mismatch count and repair
2895 * parity if !MD_RECOVERY_CHECK
2897 if ((sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) == 0)
2898 /* parity is correct (on disc,
2899 * not in buffer any more)
2901 set_bit(STRIPE_INSYNC, &sh->state);
2903 conf->mddev->resync_mismatches += STRIPE_SECTORS;
2904 if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery))
2905 /* don't try to repair!! */
2906 set_bit(STRIPE_INSYNC, &sh->state);
2908 sh->check_state = check_state_compute_run;
2909 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
2910 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
2911 set_bit(R5_Wantcompute,
2912 &sh->dev[sh->pd_idx].flags);
2913 sh->ops.target = sh->pd_idx;
2914 sh->ops.target2 = -1;
2919 case check_state_compute_run:
2922 printk(KERN_ERR "%s: unknown check_state: %d sector: %llu\n",
2923 __func__, sh->check_state,
2924 (unsigned long long) sh->sector);
2930 static void handle_parity_checks6(struct r5conf *conf, struct stripe_head *sh,
2931 struct stripe_head_state *s,
2934 int pd_idx = sh->pd_idx;
2935 int qd_idx = sh->qd_idx;
2938 set_bit(STRIPE_HANDLE, &sh->state);
2940 BUG_ON(s->failed > 2);
2942 /* Want to check and possibly repair P and Q.
2943 * However there could be one 'failed' device, in which
2944 * case we can only check one of them, possibly using the
2945 * other to generate missing data
2948 switch (sh->check_state) {
2949 case check_state_idle:
2950 /* start a new check operation if there are < 2 failures */
2951 if (s->failed == s->q_failed) {
2952 /* The only possible failed device holds Q, so it
2953 * makes sense to check P (If anything else were failed,
2954 * we would have used P to recreate it).
2956 sh->check_state = check_state_run;
2958 if (!s->q_failed && s->failed < 2) {
2959 /* Q is not failed, and we didn't use it to generate
2960 * anything, so it makes sense to check it
2962 if (sh->check_state == check_state_run)
2963 sh->check_state = check_state_run_pq;
2965 sh->check_state = check_state_run_q;
2968 /* discard potentially stale zero_sum_result */
2969 sh->ops.zero_sum_result = 0;
2971 if (sh->check_state == check_state_run) {
2972 /* async_xor_zero_sum destroys the contents of P */
2973 clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
2976 if (sh->check_state >= check_state_run &&
2977 sh->check_state <= check_state_run_pq) {
2978 /* async_syndrome_zero_sum preserves P and Q, so
2979 * no need to mark them !uptodate here
2981 set_bit(STRIPE_OP_CHECK, &s->ops_request);
2985 /* we have 2-disk failure */
2986 BUG_ON(s->failed != 2);
2988 case check_state_compute_result:
2989 sh->check_state = check_state_idle;
2991 /* check that a write has not made the stripe insync */
2992 if (test_bit(STRIPE_INSYNC, &sh->state))
2995 /* now write out any block on a failed drive,
2996 * or P or Q if they were recomputed
2998 BUG_ON(s->uptodate < disks - 1); /* We don't need Q to recover */
2999 if (s->failed == 2) {
3000 dev = &sh->dev[s->failed_num[1]];
3002 set_bit(R5_LOCKED, &dev->flags);
3003 set_bit(R5_Wantwrite, &dev->flags);
3005 if (s->failed >= 1) {
3006 dev = &sh->dev[s->failed_num[0]];
3008 set_bit(R5_LOCKED, &dev->flags);
3009 set_bit(R5_Wantwrite, &dev->flags);
3011 if (sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) {
3012 dev = &sh->dev[pd_idx];
3014 set_bit(R5_LOCKED, &dev->flags);
3015 set_bit(R5_Wantwrite, &dev->flags);
3017 if (sh->ops.zero_sum_result & SUM_CHECK_Q_RESULT) {
3018 dev = &sh->dev[qd_idx];
3020 set_bit(R5_LOCKED, &dev->flags);
3021 set_bit(R5_Wantwrite, &dev->flags);
3023 clear_bit(STRIPE_DEGRADED, &sh->state);
3025 set_bit(STRIPE_INSYNC, &sh->state);
3027 case check_state_run:
3028 case check_state_run_q:
3029 case check_state_run_pq:
3030 break; /* we will be called again upon completion */
3031 case check_state_check_result:
3032 sh->check_state = check_state_idle;
3034 /* handle a successful check operation, if parity is correct
3035 * we are done. Otherwise update the mismatch count and repair
3036 * parity if !MD_RECOVERY_CHECK
3038 if (sh->ops.zero_sum_result == 0) {
3039 /* both parities are correct */
3041 set_bit(STRIPE_INSYNC, &sh->state);
3043 /* in contrast to the raid5 case we can validate
3044 * parity, but still have a failure to write
3047 sh->check_state = check_state_compute_result;
3048 /* Returning at this point means that we may go
3049 * off and bring p and/or q uptodate again so
3050 * we make sure to check zero_sum_result again
3051 * to verify if p or q need writeback
3055 conf->mddev->resync_mismatches += STRIPE_SECTORS;
3056 if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery))
3057 /* don't try to repair!! */
3058 set_bit(STRIPE_INSYNC, &sh->state);
3060 int *target = &sh->ops.target;
3062 sh->ops.target = -1;
3063 sh->ops.target2 = -1;
3064 sh->check_state = check_state_compute_run;
3065 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
3066 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
3067 if (sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) {
3068 set_bit(R5_Wantcompute,
3069 &sh->dev[pd_idx].flags);
3071 target = &sh->ops.target2;
3074 if (sh->ops.zero_sum_result & SUM_CHECK_Q_RESULT) {
3075 set_bit(R5_Wantcompute,
3076 &sh->dev[qd_idx].flags);
3083 case check_state_compute_run:
3086 printk(KERN_ERR "%s: unknown check_state: %d sector: %llu\n",
3087 __func__, sh->check_state,
3088 (unsigned long long) sh->sector);
3093 static void handle_stripe_expansion(struct r5conf *conf, struct stripe_head *sh)
3097 /* We have read all the blocks in this stripe and now we need to
3098 * copy some of them into a target stripe for expand.
3100 struct dma_async_tx_descriptor *tx = NULL;
3101 clear_bit(STRIPE_EXPAND_SOURCE, &sh->state);
3102 for (i = 0; i < sh->disks; i++)
3103 if (i != sh->pd_idx && i != sh->qd_idx) {
3105 struct stripe_head *sh2;
3106 struct async_submit_ctl submit;
3108 sector_t bn = compute_blocknr(sh, i, 1);
3109 sector_t s = raid5_compute_sector(conf, bn, 0,
3111 sh2 = get_active_stripe(conf, s, 0, 1, 1);
3113 /* so far only the early blocks of this stripe
3114 * have been requested. When later blocks
3115 * get requested, we will try again
3118 if (!test_bit(STRIPE_EXPANDING, &sh2->state) ||
3119 test_bit(R5_Expanded, &sh2->dev[dd_idx].flags)) {
3120 /* must have already done this block */
3121 release_stripe(sh2);
3125 /* place all the copies on one channel */
3126 init_async_submit(&submit, 0, tx, NULL, NULL, NULL);
3127 tx = async_memcpy(sh2->dev[dd_idx].page,
3128 sh->dev[i].page, 0, 0, STRIPE_SIZE,
3131 set_bit(R5_Expanded, &sh2->dev[dd_idx].flags);
3132 set_bit(R5_UPTODATE, &sh2->dev[dd_idx].flags);
3133 for (j = 0; j < conf->raid_disks; j++)
3134 if (j != sh2->pd_idx &&
3136 !test_bit(R5_Expanded, &sh2->dev[j].flags))
3138 if (j == conf->raid_disks) {
3139 set_bit(STRIPE_EXPAND_READY, &sh2->state);
3140 set_bit(STRIPE_HANDLE, &sh2->state);
3142 release_stripe(sh2);
3145 /* done submitting copies, wait for them to complete */
3148 dma_wait_for_async_tx(tx);
3153 * handle_stripe - do things to a stripe.
3155 * We lock the stripe by setting STRIPE_ACTIVE and then examine the
3156 * state of various bits to see what needs to be done.
3158 * return some read requests which now have data
3159 * return some write requests which are safely on storage
3160 * schedule a read on some buffers
3161 * schedule a write of some buffers
3162 * return confirmation of parity correctness
3166 static void analyse_stripe(struct stripe_head *sh, struct stripe_head_state *s)
3168 struct r5conf *conf = sh->raid_conf;
3169 int disks = sh->disks;
3172 int do_recovery = 0;
3174 memset(s, 0, sizeof(*s));
3176 s->expanding = test_bit(STRIPE_EXPAND_SOURCE, &sh->state);
3177 s->expanded = test_bit(STRIPE_EXPAND_READY, &sh->state);
3178 s->failed_num[0] = -1;
3179 s->failed_num[1] = -1;
3181 /* Now to look around and see what can be done */
3183 spin_lock_irq(&conf->device_lock);
3184 for (i=disks; i--; ) {
3185 struct md_rdev *rdev;
3192 pr_debug("check %d: state 0x%lx read %p write %p written %p\n",
3194 dev->toread, dev->towrite, dev->written);
3195 /* maybe we can reply to a read
3197 * new wantfill requests are only permitted while
3198 * ops_complete_biofill is guaranteed to be inactive
3200 if (test_bit(R5_UPTODATE, &dev->flags) && dev->toread &&
3201 !test_bit(STRIPE_BIOFILL_RUN, &sh->state))
3202 set_bit(R5_Wantfill, &dev->flags);
3204 /* now count some things */
3205 if (test_bit(R5_LOCKED, &dev->flags))
3207 if (test_bit(R5_UPTODATE, &dev->flags))
3209 if (test_bit(R5_Wantcompute, &dev->flags)) {
3211 BUG_ON(s->compute > 2);
3214 if (test_bit(R5_Wantfill, &dev->flags))
3216 else if (dev->toread)
3220 if (!test_bit(R5_OVERWRITE, &dev->flags))
3225 /* Prefer to use the replacement for reads, but only
3226 * if it is recovered enough and has no bad blocks.
3228 rdev = rcu_dereference(conf->disks[i].replacement);
3229 if (rdev && !test_bit(Faulty, &rdev->flags) &&
3230 rdev->recovery_offset >= sh->sector + STRIPE_SECTORS &&
3231 !is_badblock(rdev, sh->sector, STRIPE_SECTORS,
3232 &first_bad, &bad_sectors))
3233 set_bit(R5_ReadRepl, &dev->flags);
3236 set_bit(R5_NeedReplace, &dev->flags);
3237 rdev = rcu_dereference(conf->disks[i].rdev);
3238 clear_bit(R5_ReadRepl, &dev->flags);
3240 if (rdev && test_bit(Faulty, &rdev->flags))
3243 is_bad = is_badblock(rdev, sh->sector, STRIPE_SECTORS,
3244 &first_bad, &bad_sectors);
3245 if (s->blocked_rdev == NULL
3246 && (test_bit(Blocked, &rdev->flags)
3249 set_bit(BlockedBadBlocks,
3251 s->blocked_rdev = rdev;
3252 atomic_inc(&rdev->nr_pending);
3255 clear_bit(R5_Insync, &dev->flags);
3259 /* also not in-sync */
3260 if (!test_bit(WriteErrorSeen, &rdev->flags) &&
3261 test_bit(R5_UPTODATE, &dev->flags)) {
3262 /* treat as in-sync, but with a read error
3263 * which we can now try to correct
3265 set_bit(R5_Insync, &dev->flags);
3266 set_bit(R5_ReadError, &dev->flags);
3268 } else if (test_bit(In_sync, &rdev->flags))
3269 set_bit(R5_Insync, &dev->flags);
3270 else if (sh->sector + STRIPE_SECTORS <= rdev->recovery_offset)
3271 /* in sync if before recovery_offset */
3272 set_bit(R5_Insync, &dev->flags);
3273 else if (test_bit(R5_UPTODATE, &dev->flags) &&
3274 test_bit(R5_Expanded, &dev->flags))
3275 /* If we've reshaped into here, we assume it is Insync.
3276 * We will shortly update recovery_offset to make
3279 set_bit(R5_Insync, &dev->flags);
3281 if (rdev && test_bit(R5_WriteError, &dev->flags)) {
3282 /* This flag does not apply to '.replacement'
3283 * only to .rdev, so make sure to check that*/
3284 struct md_rdev *rdev2 = rcu_dereference(
3285 conf->disks[i].rdev);
3287 clear_bit(R5_Insync, &dev->flags);
3288 if (rdev2 && !test_bit(Faulty, &rdev2->flags)) {
3289 s->handle_bad_blocks = 1;
3290 atomic_inc(&rdev2->nr_pending);
3292 clear_bit(R5_WriteError, &dev->flags);
3294 if (rdev && test_bit(R5_MadeGood, &dev->flags)) {
3295 /* This flag does not apply to '.replacement'
3296 * only to .rdev, so make sure to check that*/
3297 struct md_rdev *rdev2 = rcu_dereference(
3298 conf->disks[i].rdev);
3299 if (rdev2 && !test_bit(Faulty, &rdev2->flags)) {
3300 s->handle_bad_blocks = 1;
3301 atomic_inc(&rdev2->nr_pending);
3303 clear_bit(R5_MadeGood, &dev->flags);
3305 if (test_bit(R5_MadeGoodRepl, &dev->flags)) {
3306 struct md_rdev *rdev2 = rcu_dereference(
3307 conf->disks[i].replacement);
3308 if (rdev2 && !test_bit(Faulty, &rdev2->flags)) {
3309 s->handle_bad_blocks = 1;
3310 atomic_inc(&rdev2->nr_pending);
3312 clear_bit(R5_MadeGoodRepl, &dev->flags);
3314 if (!test_bit(R5_Insync, &dev->flags)) {
3315 /* The ReadError flag will just be confusing now */
3316 clear_bit(R5_ReadError, &dev->flags);
3317 clear_bit(R5_ReWrite, &dev->flags);
3319 if (test_bit(R5_ReadError, &dev->flags))
3320 clear_bit(R5_Insync, &dev->flags);
3321 if (!test_bit(R5_Insync, &dev->flags)) {
3323 s->failed_num[s->failed] = i;
3325 if (rdev && !test_bit(Faulty, &rdev->flags))
3329 spin_unlock_irq(&conf->device_lock);
3330 if (test_bit(STRIPE_SYNCING, &sh->state)) {
3331 /* If there is a failed device being replaced,
3332 * we must be recovering.
3333 * else if we are after recovery_cp, we must be syncing
3334 * else if MD_RECOVERY_REQUESTED is set, we also are syncing.
3335 * else we can only be replacing
3336 * sync and recovery both need to read all devices, and so
3337 * use the same flag.
3340 sh->sector >= conf->mddev->recovery_cp ||
3341 test_bit(MD_RECOVERY_REQUESTED, &(conf->mddev->recovery)))
3349 static void handle_stripe(struct stripe_head *sh)
3351 struct stripe_head_state s;
3352 struct r5conf *conf = sh->raid_conf;
3355 int disks = sh->disks;
3356 struct r5dev *pdev, *qdev;
3358 clear_bit(STRIPE_HANDLE, &sh->state);
3359 if (test_and_set_bit_lock(STRIPE_ACTIVE, &sh->state)) {
3360 /* already being handled, ensure it gets handled
3361 * again when current action finishes */
3362 set_bit(STRIPE_HANDLE, &sh->state);
3366 if (test_and_clear_bit(STRIPE_SYNC_REQUESTED, &sh->state)) {
3367 set_bit(STRIPE_SYNCING, &sh->state);
3368 clear_bit(STRIPE_INSYNC, &sh->state);
3370 clear_bit(STRIPE_DELAYED, &sh->state);
3372 pr_debug("handling stripe %llu, state=%#lx cnt=%d, "
3373 "pd_idx=%d, qd_idx=%d\n, check:%d, reconstruct:%d\n",
3374 (unsigned long long)sh->sector, sh->state,
3375 atomic_read(&sh->count), sh->pd_idx, sh->qd_idx,
3376 sh->check_state, sh->reconstruct_state);
3378 analyse_stripe(sh, &s);
3380 if (s.handle_bad_blocks) {
3381 set_bit(STRIPE_HANDLE, &sh->state);
3385 if (unlikely(s.blocked_rdev)) {
3386 if (s.syncing || s.expanding || s.expanded ||
3387 s.replacing || s.to_write || s.written) {
3388 set_bit(STRIPE_HANDLE, &sh->state);
3391 /* There is nothing for the blocked_rdev to block */
3392 rdev_dec_pending(s.blocked_rdev, conf->mddev);
3393 s.blocked_rdev = NULL;
3396 if (s.to_fill && !test_bit(STRIPE_BIOFILL_RUN, &sh->state)) {
3397 set_bit(STRIPE_OP_BIOFILL, &s.ops_request);
3398 set_bit(STRIPE_BIOFILL_RUN, &sh->state);
3401 pr_debug("locked=%d uptodate=%d to_read=%d"
3402 " to_write=%d failed=%d failed_num=%d,%d\n",
3403 s.locked, s.uptodate, s.to_read, s.to_write, s.failed,
3404 s.failed_num[0], s.failed_num[1]);
3405 /* check if the array has lost more than max_degraded devices and,
3406 * if so, some requests might need to be failed.
3408 if (s.failed > conf->max_degraded) {
3409 sh->check_state = 0;
3410 sh->reconstruct_state = 0;
3411 if (s.to_read+s.to_write+s.written)
3412 handle_failed_stripe(conf, sh, &s, disks, &s.return_bi);
3413 if (s.syncing + s.replacing)
3414 handle_failed_sync(conf, sh, &s);
3418 * might be able to return some write requests if the parity blocks
3419 * are safe, or on a failed drive
3421 pdev = &sh->dev[sh->pd_idx];
3422 s.p_failed = (s.failed >= 1 && s.failed_num[0] == sh->pd_idx)
3423 || (s.failed >= 2 && s.failed_num[1] == sh->pd_idx);
3424 qdev = &sh->dev[sh->qd_idx];
3425 s.q_failed = (s.failed >= 1 && s.failed_num[0] == sh->qd_idx)
3426 || (s.failed >= 2 && s.failed_num[1] == sh->qd_idx)
3430 (s.p_failed || ((test_bit(R5_Insync, &pdev->flags)
3431 && !test_bit(R5_LOCKED, &pdev->flags)
3432 && test_bit(R5_UPTODATE, &pdev->flags)))) &&
3433 (s.q_failed || ((test_bit(R5_Insync, &qdev->flags)
3434 && !test_bit(R5_LOCKED, &qdev->flags)
3435 && test_bit(R5_UPTODATE, &qdev->flags)))))
3436 handle_stripe_clean_event(conf, sh, disks, &s.return_bi);
3438 /* Now we might consider reading some blocks, either to check/generate
3439 * parity, or to satisfy requests
3440 * or to load a block that is being partially written.
3442 if (s.to_read || s.non_overwrite
3443 || (conf->level == 6 && s.to_write && s.failed)
3444 || (s.syncing && (s.uptodate + s.compute < disks))
3447 handle_stripe_fill(sh, &s, disks);
3449 /* Now we check to see if any write operations have recently
3453 if (sh->reconstruct_state == reconstruct_state_prexor_drain_result)
3455 if (sh->reconstruct_state == reconstruct_state_drain_result ||
3456 sh->reconstruct_state == reconstruct_state_prexor_drain_result) {
3457 sh->reconstruct_state = reconstruct_state_idle;
3459 /* All the 'written' buffers and the parity block are ready to
3460 * be written back to disk
3462 BUG_ON(!test_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags));
3463 BUG_ON(sh->qd_idx >= 0 &&
3464 !test_bit(R5_UPTODATE, &sh->dev[sh->qd_idx].flags));
3465 for (i = disks; i--; ) {
3466 struct r5dev *dev = &sh->dev[i];
3467 if (test_bit(R5_LOCKED, &dev->flags) &&
3468 (i == sh->pd_idx || i == sh->qd_idx ||
3470 pr_debug("Writing block %d\n", i);
3471 set_bit(R5_Wantwrite, &dev->flags);
3474 if (!test_bit(R5_Insync, &dev->flags) ||
3475 ((i == sh->pd_idx || i == sh->qd_idx) &&
3477 set_bit(STRIPE_INSYNC, &sh->state);
3480 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
3481 s.dec_preread_active = 1;
3484 /* Now to consider new write requests and what else, if anything
3485 * should be read. We do not handle new writes when:
3486 * 1/ A 'write' operation (copy+xor) is already in flight.
3487 * 2/ A 'check' operation is in flight, as it may clobber the parity
3490 if (s.to_write && !sh->reconstruct_state && !sh->check_state)
3491 handle_stripe_dirtying(conf, sh, &s, disks);
3493 /* maybe we need to check and possibly fix the parity for this stripe
3494 * Any reads will already have been scheduled, so we just see if enough
3495 * data is available. The parity check is held off while parity
3496 * dependent operations are in flight.
3498 if (sh->check_state ||
3499 (s.syncing && s.locked == 0 &&
3500 !test_bit(STRIPE_COMPUTE_RUN, &sh->state) &&
3501 !test_bit(STRIPE_INSYNC, &sh->state))) {
3502 if (conf->level == 6)
3503 handle_parity_checks6(conf, sh, &s, disks);
3505 handle_parity_checks5(conf, sh, &s, disks);
3508 if (s.replacing && s.locked == 0
3509 && !test_bit(STRIPE_INSYNC, &sh->state)) {
3510 /* Write out to replacement devices where possible */
3511 for (i = 0; i < conf->raid_disks; i++)
3512 if (test_bit(R5_UPTODATE, &sh->dev[i].flags) &&
3513 test_bit(R5_NeedReplace, &sh->dev[i].flags)) {
3514 set_bit(R5_WantReplace, &sh->dev[i].flags);
3515 set_bit(R5_LOCKED, &sh->dev[i].flags);
3518 set_bit(STRIPE_INSYNC, &sh->state);
3520 if ((s.syncing || s.replacing) && s.locked == 0 &&
3521 test_bit(STRIPE_INSYNC, &sh->state)) {
3522 md_done_sync(conf->mddev, STRIPE_SECTORS, 1);
3523 clear_bit(STRIPE_SYNCING, &sh->state);
3526 /* If the failed drives are just a ReadError, then we might need
3527 * to progress the repair/check process
3529 if (s.failed <= conf->max_degraded && !conf->mddev->ro)
3530 for (i = 0; i < s.failed; i++) {
3531 struct r5dev *dev = &sh->dev[s.failed_num[i]];
3532 if (test_bit(R5_ReadError, &dev->flags)
3533 && !test_bit(R5_LOCKED, &dev->flags)
3534 && test_bit(R5_UPTODATE, &dev->flags)
3536 if (!test_bit(R5_ReWrite, &dev->flags)) {
3537 set_bit(R5_Wantwrite, &dev->flags);
3538 set_bit(R5_ReWrite, &dev->flags);
3539 set_bit(R5_LOCKED, &dev->flags);
3542 /* let's read it back */
3543 set_bit(R5_Wantread, &dev->flags);
3544 set_bit(R5_LOCKED, &dev->flags);
3551 /* Finish reconstruct operations initiated by the expansion process */
3552 if (sh->reconstruct_state == reconstruct_state_result) {
3553 struct stripe_head *sh_src
3554 = get_active_stripe(conf, sh->sector, 1, 1, 1);
3555 if (sh_src && test_bit(STRIPE_EXPAND_SOURCE, &sh_src->state)) {
3556 /* sh cannot be written until sh_src has been read.
3557 * so arrange for sh to be delayed a little
3559 set_bit(STRIPE_DELAYED, &sh->state);
3560 set_bit(STRIPE_HANDLE, &sh->state);
3561 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE,
3563 atomic_inc(&conf->preread_active_stripes);
3564 release_stripe(sh_src);
3568 release_stripe(sh_src);
3570 sh->reconstruct_state = reconstruct_state_idle;
3571 clear_bit(STRIPE_EXPANDING, &sh->state);
3572 for (i = conf->raid_disks; i--; ) {
3573 set_bit(R5_Wantwrite, &sh->dev[i].flags);
3574 set_bit(R5_LOCKED, &sh->dev[i].flags);
3579 if (s.expanded && test_bit(STRIPE_EXPANDING, &sh->state) &&
3580 !sh->reconstruct_state) {
3581 /* Need to write out all blocks after computing parity */
3582 sh->disks = conf->raid_disks;
3583 stripe_set_idx(sh->sector, conf, 0, sh);
3584 schedule_reconstruction(sh, &s, 1, 1);
3585 } else if (s.expanded && !sh->reconstruct_state && s.locked == 0) {
3586 clear_bit(STRIPE_EXPAND_READY, &sh->state);
3587 atomic_dec(&conf->reshape_stripes);
3588 wake_up(&conf->wait_for_overlap);
3589 md_done_sync(conf->mddev, STRIPE_SECTORS, 1);
3592 if (s.expanding && s.locked == 0 &&
3593 !test_bit(STRIPE_COMPUTE_RUN, &sh->state))
3594 handle_stripe_expansion(conf, sh);
3597 /* wait for this device to become unblocked */
3598 if (unlikely(s.blocked_rdev)) {
3599 if (conf->mddev->external)
3600 md_wait_for_blocked_rdev(s.blocked_rdev,
3603 /* Internal metadata will immediately
3604 * be written by raid5d, so we don't
3605 * need to wait here.
3607 rdev_dec_pending(s.blocked_rdev,
3611 if (s.handle_bad_blocks)
3612 for (i = disks; i--; ) {
3613 struct md_rdev *rdev;
3614 struct r5dev *dev = &sh->dev[i];
3615 if (test_and_clear_bit(R5_WriteError, &dev->flags)) {
3616 /* We own a safe reference to the rdev */
3617 rdev = conf->disks[i].rdev;
3618 if (!rdev_set_badblocks(rdev, sh->sector,
3620 md_error(conf->mddev, rdev);
3621 rdev_dec_pending(rdev, conf->mddev);
3623 if (test_and_clear_bit(R5_MadeGood, &dev->flags)) {
3624 rdev = conf->disks[i].rdev;
3625 rdev_clear_badblocks(rdev, sh->sector,
3627 rdev_dec_pending(rdev, conf->mddev);
3629 if (test_and_clear_bit(R5_MadeGoodRepl, &dev->flags)) {
3630 rdev = conf->disks[i].replacement;
3632 /* rdev have been moved down */
3633 rdev = conf->disks[i].rdev;
3634 rdev_clear_badblocks(rdev, sh->sector,
3636 rdev_dec_pending(rdev, conf->mddev);
3641 raid_run_ops(sh, s.ops_request);
3645 if (s.dec_preread_active) {
3646 /* We delay this until after ops_run_io so that if make_request
3647 * is waiting on a flush, it won't continue until the writes
3648 * have actually been submitted.
3650 atomic_dec(&conf->preread_active_stripes);
3651 if (atomic_read(&conf->preread_active_stripes) <
3653 md_wakeup_thread(conf->mddev->thread);
3656 return_io(s.return_bi);
3658 clear_bit_unlock(STRIPE_ACTIVE, &sh->state);
3661 static void raid5_activate_delayed(struct r5conf *conf)
3663 if (atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD) {
3664 while (!list_empty(&conf->delayed_list)) {
3665 struct list_head *l = conf->delayed_list.next;
3666 struct stripe_head *sh;
3667 sh = list_entry(l, struct stripe_head, lru);
3669 clear_bit(STRIPE_DELAYED, &sh->state);
3670 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
3671 atomic_inc(&conf->preread_active_stripes);
3672 list_add_tail(&sh->lru, &conf->hold_list);
3677 static void activate_bit_delay(struct r5conf *conf)
3679 /* device_lock is held */
3680 struct list_head head;
3681 list_add(&head, &conf->bitmap_list);
3682 list_del_init(&conf->bitmap_list);
3683 while (!list_empty(&head)) {
3684 struct stripe_head *sh = list_entry(head.next, struct stripe_head, lru);
3685 list_del_init(&sh->lru);
3686 atomic_inc(&sh->count);
3687 __release_stripe(conf, sh);
3691 int md_raid5_congested(struct mddev *mddev, int bits)
3693 struct r5conf *conf = mddev->private;
3695 /* No difference between reads and writes. Just check
3696 * how busy the stripe_cache is
3699 if (conf->inactive_blocked)
3703 if (list_empty_careful(&conf->inactive_list))
3708 EXPORT_SYMBOL_GPL(md_raid5_congested);
3710 static int raid5_congested(void *data, int bits)
3712 struct mddev *mddev = data;
3714 return mddev_congested(mddev, bits) ||
3715 md_raid5_congested(mddev, bits);
3718 /* We want read requests to align with chunks where possible,
3719 * but write requests don't need to.
3721 static int raid5_mergeable_bvec(struct request_queue *q,
3722 struct bvec_merge_data *bvm,
3723 struct bio_vec *biovec)
3725 struct mddev *mddev = q->queuedata;
3726 sector_t sector = bvm->bi_sector + get_start_sect(bvm->bi_bdev);
3728 unsigned int chunk_sectors = mddev->chunk_sectors;
3729 unsigned int bio_sectors = bvm->bi_size >> 9;
3731 if ((bvm->bi_rw & 1) == WRITE)
3732 return biovec->bv_len; /* always allow writes to be mergeable */
3734 if (mddev->new_chunk_sectors < mddev->chunk_sectors)
3735 chunk_sectors = mddev->new_chunk_sectors;
3736 max = (chunk_sectors - ((sector & (chunk_sectors - 1)) + bio_sectors)) << 9;
3737 if (max < 0) max = 0;
3738 if (max <= biovec->bv_len && bio_sectors == 0)
3739 return biovec->bv_len;
3745 static int in_chunk_boundary(struct mddev *mddev, struct bio *bio)
3747 sector_t sector = bio->bi_sector + get_start_sect(bio->bi_bdev);
3748 unsigned int chunk_sectors = mddev->chunk_sectors;
3749 unsigned int bio_sectors = bio->bi_size >> 9;
3751 if (mddev->new_chunk_sectors < mddev->chunk_sectors)
3752 chunk_sectors = mddev->new_chunk_sectors;
3753 return chunk_sectors >=
3754 ((sector & (chunk_sectors - 1)) + bio_sectors);
3758 * add bio to the retry LIFO ( in O(1) ... we are in interrupt )
3759 * later sampled by raid5d.
3761 static void add_bio_to_retry(struct bio *bi,struct r5conf *conf)
3763 unsigned long flags;
3765 spin_lock_irqsave(&conf->device_lock, flags);
3767 bi->bi_next = conf->retry_read_aligned_list;
3768 conf->retry_read_aligned_list = bi;
3770 spin_unlock_irqrestore(&conf->device_lock, flags);
3771 md_wakeup_thread(conf->mddev->thread);
3775 static struct bio *remove_bio_from_retry(struct r5conf *conf)
3779 bi = conf->retry_read_aligned;
3781 conf->retry_read_aligned = NULL;
3784 bi = conf->retry_read_aligned_list;
3786 conf->retry_read_aligned_list = bi->bi_next;
3789 * this sets the active strip count to 1 and the processed
3790 * strip count to zero (upper 8 bits)
3792 bi->bi_phys_segments = 1; /* biased count of active stripes */
3800 * The "raid5_align_endio" should check if the read succeeded and if it
3801 * did, call bio_endio on the original bio (having bio_put the new bio
3803 * If the read failed..
3805 static void raid5_align_endio(struct bio *bi, int error)
3807 struct bio* raid_bi = bi->bi_private;
3808 struct mddev *mddev;
3809 struct r5conf *conf;
3810 int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
3811 struct md_rdev *rdev;
3815 rdev = (void*)raid_bi->bi_next;
3816 raid_bi->bi_next = NULL;
3817 mddev = rdev->mddev;
3818 conf = mddev->private;
3820 rdev_dec_pending(rdev, conf->mddev);
3822 if (!error && uptodate) {
3823 bio_endio(raid_bi, 0);
3824 if (atomic_dec_and_test(&conf->active_aligned_reads))
3825 wake_up(&conf->wait_for_stripe);
3830 pr_debug("raid5_align_endio : io error...handing IO for a retry\n");
3832 add_bio_to_retry(raid_bi, conf);
3835 static int bio_fits_rdev(struct bio *bi)
3837 struct request_queue *q = bdev_get_queue(bi->bi_bdev);
3839 if ((bi->bi_size>>9) > queue_max_sectors(q))
3841 blk_recount_segments(q, bi);
3842 if (bi->bi_phys_segments > queue_max_segments(q))
3845 if (q->merge_bvec_fn)
3846 /* it's too hard to apply the merge_bvec_fn at this stage,
3855 static int chunk_aligned_read(struct mddev *mddev, struct bio * raid_bio)
3857 struct r5conf *conf = mddev->private;
3859 struct bio* align_bi;
3860 struct md_rdev *rdev;
3861 sector_t end_sector;
3863 if (!in_chunk_boundary(mddev, raid_bio)) {
3864 pr_debug("chunk_aligned_read : non aligned\n");
3868 * use bio_clone_mddev to make a copy of the bio
3870 align_bi = bio_clone_mddev(raid_bio, GFP_NOIO, mddev);
3874 * set bi_end_io to a new function, and set bi_private to the
3877 align_bi->bi_end_io = raid5_align_endio;
3878 align_bi->bi_private = raid_bio;
3882 align_bi->bi_sector = raid5_compute_sector(conf, raid_bio->bi_sector,
3886 end_sector = align_bi->bi_sector + (align_bi->bi_size >> 9);
3888 rdev = rcu_dereference(conf->disks[dd_idx].replacement);
3889 if (!rdev || test_bit(Faulty, &rdev->flags) ||
3890 rdev->recovery_offset < end_sector) {
3891 rdev = rcu_dereference(conf->disks[dd_idx].rdev);
3893 (test_bit(Faulty, &rdev->flags) ||
3894 !(test_bit(In_sync, &rdev->flags) ||
3895 rdev->recovery_offset >= end_sector)))
3902 atomic_inc(&rdev->nr_pending);
3904 raid_bio->bi_next = (void*)rdev;
3905 align_bi->bi_bdev = rdev->bdev;
3906 align_bi->bi_flags &= ~(1 << BIO_SEG_VALID);
3908 if (!bio_fits_rdev(align_bi) ||
3909 is_badblock(rdev, align_bi->bi_sector, align_bi->bi_size>>9,
3910 &first_bad, &bad_sectors)) {
3911 /* too big in some way, or has a known bad block */
3913 rdev_dec_pending(rdev, mddev);
3917 /* No reshape active, so we can trust rdev->data_offset */
3918 align_bi->bi_sector += rdev->data_offset;
3920 spin_lock_irq(&conf->device_lock);
3921 wait_event_lock_irq(conf->wait_for_stripe,
3923 conf->device_lock, /* nothing */);
3924 atomic_inc(&conf->active_aligned_reads);
3925 spin_unlock_irq(&conf->device_lock);
3927 generic_make_request(align_bi);
3936 /* __get_priority_stripe - get the next stripe to process
3938 * Full stripe writes are allowed to pass preread active stripes up until
3939 * the bypass_threshold is exceeded. In general the bypass_count
3940 * increments when the handle_list is handled before the hold_list; however, it
3941 * will not be incremented when STRIPE_IO_STARTED is sampled set signifying a
3942 * stripe with in flight i/o. The bypass_count will be reset when the
3943 * head of the hold_list has changed, i.e. the head was promoted to the
3946 static struct stripe_head *__get_priority_stripe(struct r5conf *conf)
3948 struct stripe_head *sh;
3950 pr_debug("%s: handle: %s hold: %s full_writes: %d bypass_count: %d\n",
3952 list_empty(&conf->handle_list) ? "empty" : "busy",
3953 list_empty(&conf->hold_list) ? "empty" : "busy",
3954 atomic_read(&conf->pending_full_writes), conf->bypass_count);
3956 if (!list_empty(&conf->handle_list)) {
3957 sh = list_entry(conf->handle_list.next, typeof(*sh), lru);
3959 if (list_empty(&conf->hold_list))
3960 conf->bypass_count = 0;
3961 else if (!test_bit(STRIPE_IO_STARTED, &sh->state)) {
3962 if (conf->hold_list.next == conf->last_hold)
3963 conf->bypass_count++;
3965 conf->last_hold = conf->hold_list.next;
3966 conf->bypass_count -= conf->bypass_threshold;
3967 if (conf->bypass_count < 0)
3968 conf->bypass_count = 0;
3971 } else if (!list_empty(&conf->hold_list) &&
3972 ((conf->bypass_threshold &&
3973 conf->bypass_count > conf->bypass_threshold) ||
3974 atomic_read(&conf->pending_full_writes) == 0)) {
3975 sh = list_entry(conf->hold_list.next,
3977 conf->bypass_count -= conf->bypass_threshold;
3978 if (conf->bypass_count < 0)
3979 conf->bypass_count = 0;
3983 list_del_init(&sh->lru);
3984 atomic_inc(&sh->count);
3985 BUG_ON(atomic_read(&sh->count) != 1);
3989 static void make_request(struct mddev *mddev, struct bio * bi)
3991 struct r5conf *conf = mddev->private;
3993 sector_t new_sector;
3994 sector_t logical_sector, last_sector;
3995 struct stripe_head *sh;
3996 const int rw = bio_data_dir(bi);
4000 if (unlikely(bi->bi_rw & REQ_FLUSH)) {
4001 md_flush_request(mddev, bi);
4005 md_write_start(mddev, bi);
4008 mddev->reshape_position == MaxSector &&
4009 chunk_aligned_read(mddev,bi))
4012 logical_sector = bi->bi_sector & ~((sector_t)STRIPE_SECTORS-1);
4013 last_sector = bi->bi_sector + (bi->bi_size>>9);
4015 bi->bi_phys_segments = 1; /* over-loaded to count active stripes */
4017 plugged = mddev_check_plugged(mddev);
4018 for (;logical_sector < last_sector; logical_sector += STRIPE_SECTORS) {
4024 prepare_to_wait(&conf->wait_for_overlap, &w, TASK_UNINTERRUPTIBLE);
4025 if (unlikely(conf->reshape_progress != MaxSector)) {
4026 /* spinlock is needed as reshape_progress may be
4027 * 64bit on a 32bit platform, and so it might be
4028 * possible to see a half-updated value
4029 * Of course reshape_progress could change after
4030 * the lock is dropped, so once we get a reference
4031 * to the stripe that we think it is, we will have
4034 spin_lock_irq(&conf->device_lock);
4035 if (mddev->reshape_backwards
4036 ? logical_sector < conf->reshape_progress
4037 : logical_sector >= conf->reshape_progress) {
4040 if (mddev->reshape_backwards
4041 ? logical_sector < conf->reshape_safe
4042 : logical_sector >= conf->reshape_safe) {
4043 spin_unlock_irq(&conf->device_lock);
4048 spin_unlock_irq(&conf->device_lock);
4051 new_sector = raid5_compute_sector(conf, logical_sector,
4054 pr_debug("raid456: make_request, sector %llu logical %llu\n",
4055 (unsigned long long)new_sector,
4056 (unsigned long long)logical_sector);
4058 sh = get_active_stripe(conf, new_sector, previous,
4059 (bi->bi_rw&RWA_MASK), 0);
4061 if (unlikely(previous)) {
4062 /* expansion might have moved on while waiting for a
4063 * stripe, so we must do the range check again.
4064 * Expansion could still move past after this
4065 * test, but as we are holding a reference to
4066 * 'sh', we know that if that happens,
4067 * STRIPE_EXPANDING will get set and the expansion
4068 * won't proceed until we finish with the stripe.
4071 spin_lock_irq(&conf->device_lock);
4072 if (mddev->reshape_backwards
4073 ? logical_sector >= conf->reshape_progress
4074 : logical_sector < conf->reshape_progress)
4075 /* mismatch, need to try again */
4077 spin_unlock_irq(&conf->device_lock);
4086 logical_sector >= mddev->suspend_lo &&
4087 logical_sector < mddev->suspend_hi) {
4089 /* As the suspend_* range is controlled by
4090 * userspace, we want an interruptible
4093 flush_signals(current);
4094 prepare_to_wait(&conf->wait_for_overlap,
4095 &w, TASK_INTERRUPTIBLE);
4096 if (logical_sector >= mddev->suspend_lo &&
4097 logical_sector < mddev->suspend_hi)
4102 if (test_bit(STRIPE_EXPANDING, &sh->state) ||
4103 !add_stripe_bio(sh, bi, dd_idx, rw)) {
4104 /* Stripe is busy expanding or
4105 * add failed due to overlap. Flush everything
4108 md_wakeup_thread(mddev->thread);
4113 finish_wait(&conf->wait_for_overlap, &w);
4114 set_bit(STRIPE_HANDLE, &sh->state);
4115 clear_bit(STRIPE_DELAYED, &sh->state);
4116 if ((bi->bi_rw & REQ_SYNC) &&
4117 !test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
4118 atomic_inc(&conf->preread_active_stripes);
4121 /* cannot get stripe for read-ahead, just give-up */
4122 clear_bit(BIO_UPTODATE, &bi->bi_flags);
4123 finish_wait(&conf->wait_for_overlap, &w);
4129 md_wakeup_thread(mddev->thread);
4131 spin_lock_irq(&conf->device_lock);
4132 remaining = raid5_dec_bi_phys_segments(bi);
4133 spin_unlock_irq(&conf->device_lock);
4134 if (remaining == 0) {
4137 md_write_end(mddev);
4143 static sector_t raid5_size(struct mddev *mddev, sector_t sectors, int raid_disks);
4145 static sector_t reshape_request(struct mddev *mddev, sector_t sector_nr, int *skipped)
4147 /* reshaping is quite different to recovery/resync so it is
4148 * handled quite separately ... here.
4150 * On each call to sync_request, we gather one chunk worth of
4151 * destination stripes and flag them as expanding.
4152 * Then we find all the source stripes and request reads.
4153 * As the reads complete, handle_stripe will copy the data
4154 * into the destination stripe and release that stripe.
4156 struct r5conf *conf = mddev->private;
4157 struct stripe_head *sh;
4158 sector_t first_sector, last_sector;
4159 int raid_disks = conf->previous_raid_disks;
4160 int data_disks = raid_disks - conf->max_degraded;
4161 int new_data_disks = conf->raid_disks - conf->max_degraded;
4164 sector_t writepos, readpos, safepos;
4165 sector_t stripe_addr;
4166 int reshape_sectors;
4167 struct list_head stripes;
4169 if (sector_nr == 0) {
4170 /* If restarting in the middle, skip the initial sectors */
4171 if (mddev->reshape_backwards &&
4172 conf->reshape_progress < raid5_size(mddev, 0, 0)) {
4173 sector_nr = raid5_size(mddev, 0, 0)
4174 - conf->reshape_progress;
4175 } else if (!mddev->reshape_backwards &&
4176 conf->reshape_progress > 0)
4177 sector_nr = conf->reshape_progress;
4178 sector_div(sector_nr, new_data_disks);
4180 mddev->curr_resync_completed = sector_nr;
4181 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
4187 /* We need to process a full chunk at a time.
4188 * If old and new chunk sizes differ, we need to process the
4191 if (mddev->new_chunk_sectors > mddev->chunk_sectors)
4192 reshape_sectors = mddev->new_chunk_sectors;
4194 reshape_sectors = mddev->chunk_sectors;
4196 /* We update the metadata at least every 10 seconds, or when
4197 * the data about to be copied would over-write the source of
4198 * the data at the front of the range. i.e. one new_stripe
4199 * along from reshape_progress new_maps to after where
4200 * reshape_safe old_maps to
4202 writepos = conf->reshape_progress;
4203 sector_div(writepos, new_data_disks);
4204 readpos = conf->reshape_progress;
4205 sector_div(readpos, data_disks);
4206 safepos = conf->reshape_safe;
4207 sector_div(safepos, data_disks);
4208 if (mddev->reshape_backwards) {
4209 writepos -= min_t(sector_t, reshape_sectors, writepos);
4210 readpos += reshape_sectors;
4211 safepos += reshape_sectors;
4213 writepos += reshape_sectors;
4214 readpos -= min_t(sector_t, reshape_sectors, readpos);
4215 safepos -= min_t(sector_t, reshape_sectors, safepos);
4218 /* Having calculated the 'writepos' possibly use it
4219 * to set 'stripe_addr' which is where we will write to.
4221 if (mddev->reshape_backwards) {
4222 BUG_ON(conf->reshape_progress == 0);
4223 stripe_addr = writepos;
4224 BUG_ON((mddev->dev_sectors &
4225 ~((sector_t)reshape_sectors - 1))
4226 - reshape_sectors - stripe_addr
4229 BUG_ON(writepos != sector_nr + reshape_sectors);
4230 stripe_addr = sector_nr;
4233 /* 'writepos' is the most advanced device address we might write.
4234 * 'readpos' is the least advanced device address we might read.
4235 * 'safepos' is the least address recorded in the metadata as having
4237 * If there is a min_offset_diff, these are adjusted either by
4238 * increasing the safepos/readpos if diff is negative, or
4239 * increasing writepos if diff is positive.
4240 * If 'readpos' is then behind 'writepos', there is no way that we can
4241 * ensure safety in the face of a crash - that must be done by userspace
4242 * making a backup of the data. So in that case there is no particular
4243 * rush to update metadata.
4244 * Otherwise if 'safepos' is behind 'writepos', then we really need to
4245 * update the metadata to advance 'safepos' to match 'readpos' so that
4246 * we can be safe in the event of a crash.
4247 * So we insist on updating metadata if safepos is behind writepos and
4248 * readpos is beyond writepos.
4249 * In any case, update the metadata every 10 seconds.
4250 * Maybe that number should be configurable, but I'm not sure it is
4251 * worth it.... maybe it could be a multiple of safemode_delay???
4253 if (conf->min_offset_diff < 0) {
4254 safepos += -conf->min_offset_diff;
4255 readpos += -conf->min_offset_diff;
4257 writepos += conf->min_offset_diff;
4259 if ((mddev->reshape_backwards
4260 ? (safepos > writepos && readpos < writepos)
4261 : (safepos < writepos && readpos > writepos)) ||
4262 time_after(jiffies, conf->reshape_checkpoint + 10*HZ)) {
4263 /* Cannot proceed until we've updated the superblock... */
4264 wait_event(conf->wait_for_overlap,
4265 atomic_read(&conf->reshape_stripes)==0);
4266 mddev->reshape_position = conf->reshape_progress;
4267 mddev->curr_resync_completed = sector_nr;
4268 conf->reshape_checkpoint = jiffies;
4269 set_bit(MD_CHANGE_DEVS, &mddev->flags);
4270 md_wakeup_thread(mddev->thread);
4271 wait_event(mddev->sb_wait, mddev->flags == 0 ||
4272 kthread_should_stop());
4273 spin_lock_irq(&conf->device_lock);
4274 conf->reshape_safe = mddev->reshape_position;
4275 spin_unlock_irq(&conf->device_lock);
4276 wake_up(&conf->wait_for_overlap);
4277 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
4280 INIT_LIST_HEAD(&stripes);
4281 for (i = 0; i < reshape_sectors; i += STRIPE_SECTORS) {
4283 int skipped_disk = 0;
4284 sh = get_active_stripe(conf, stripe_addr+i, 0, 0, 1);
4285 set_bit(STRIPE_EXPANDING, &sh->state);
4286 atomic_inc(&conf->reshape_stripes);
4287 /* If any of this stripe is beyond the end of the old
4288 * array, then we need to zero those blocks
4290 for (j=sh->disks; j--;) {
4292 if (j == sh->pd_idx)
4294 if (conf->level == 6 &&
4297 s = compute_blocknr(sh, j, 0);
4298 if (s < raid5_size(mddev, 0, 0)) {
4302 memset(page_address(sh->dev[j].page), 0, STRIPE_SIZE);
4303 set_bit(R5_Expanded, &sh->dev[j].flags);
4304 set_bit(R5_UPTODATE, &sh->dev[j].flags);
4306 if (!skipped_disk) {
4307 set_bit(STRIPE_EXPAND_READY, &sh->state);
4308 set_bit(STRIPE_HANDLE, &sh->state);
4310 list_add(&sh->lru, &stripes);
4312 spin_lock_irq(&conf->device_lock);
4313 if (mddev->reshape_backwards)
4314 conf->reshape_progress -= reshape_sectors * new_data_disks;
4316 conf->reshape_progress += reshape_sectors * new_data_disks;
4317 spin_unlock_irq(&conf->device_lock);
4318 /* Ok, those stripe are ready. We can start scheduling
4319 * reads on the source stripes.
4320 * The source stripes are determined by mapping the first and last
4321 * block on the destination stripes.
4324 raid5_compute_sector(conf, stripe_addr*(new_data_disks),
4327 raid5_compute_sector(conf, ((stripe_addr+reshape_sectors)
4328 * new_data_disks - 1),
4330 if (last_sector >= mddev->dev_sectors)
4331 last_sector = mddev->dev_sectors - 1;
4332 while (first_sector <= last_sector) {
4333 sh = get_active_stripe(conf, first_sector, 1, 0, 1);
4334 set_bit(STRIPE_EXPAND_SOURCE, &sh->state);
4335 set_bit(STRIPE_HANDLE, &sh->state);
4337 first_sector += STRIPE_SECTORS;
4339 /* Now that the sources are clearly marked, we can release
4340 * the destination stripes
4342 while (!list_empty(&stripes)) {
4343 sh = list_entry(stripes.next, struct stripe_head, lru);
4344 list_del_init(&sh->lru);
4347 /* If this takes us to the resync_max point where we have to pause,
4348 * then we need to write out the superblock.
4350 sector_nr += reshape_sectors;
4351 if ((sector_nr - mddev->curr_resync_completed) * 2
4352 >= mddev->resync_max - mddev->curr_resync_completed) {
4353 /* Cannot proceed until we've updated the superblock... */
4354 wait_event(conf->wait_for_overlap,
4355 atomic_read(&conf->reshape_stripes) == 0);
4356 mddev->reshape_position = conf->reshape_progress;
4357 mddev->curr_resync_completed = sector_nr;
4358 conf->reshape_checkpoint = jiffies;
4359 set_bit(MD_CHANGE_DEVS, &mddev->flags);
4360 md_wakeup_thread(mddev->thread);
4361 wait_event(mddev->sb_wait,
4362 !test_bit(MD_CHANGE_DEVS, &mddev->flags)
4363 || kthread_should_stop());
4364 spin_lock_irq(&conf->device_lock);
4365 conf->reshape_safe = mddev->reshape_position;
4366 spin_unlock_irq(&conf->device_lock);
4367 wake_up(&conf->wait_for_overlap);
4368 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
4370 return reshape_sectors;
4373 /* FIXME go_faster isn't used */
4374 static inline sector_t sync_request(struct mddev *mddev, sector_t sector_nr, int *skipped, int go_faster)
4376 struct r5conf *conf = mddev->private;
4377 struct stripe_head *sh;
4378 sector_t max_sector = mddev->dev_sectors;
4379 sector_t sync_blocks;
4380 int still_degraded = 0;
4383 if (sector_nr >= max_sector) {
4384 /* just being told to finish up .. nothing much to do */
4386 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery)) {
4391 if (mddev->curr_resync < max_sector) /* aborted */
4392 bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
4394 else /* completed sync */
4396 bitmap_close_sync(mddev->bitmap);
4401 /* Allow raid5_quiesce to complete */
4402 wait_event(conf->wait_for_overlap, conf->quiesce != 2);
4404 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
4405 return reshape_request(mddev, sector_nr, skipped);
4407 /* No need to check resync_max as we never do more than one
4408 * stripe, and as resync_max will always be on a chunk boundary,
4409 * if the check in md_do_sync didn't fire, there is no chance
4410 * of overstepping resync_max here
4413 /* if there is too many failed drives and we are trying
4414 * to resync, then assert that we are finished, because there is
4415 * nothing we can do.
4417 if (mddev->degraded >= conf->max_degraded &&
4418 test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
4419 sector_t rv = mddev->dev_sectors - sector_nr;
4423 if (!bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
4424 !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
4425 !conf->fullsync && sync_blocks >= STRIPE_SECTORS) {
4426 /* we can skip this block, and probably more */
4427 sync_blocks /= STRIPE_SECTORS;
4429 return sync_blocks * STRIPE_SECTORS; /* keep things rounded to whole stripes */
4432 bitmap_cond_end_sync(mddev->bitmap, sector_nr);
4434 sh = get_active_stripe(conf, sector_nr, 0, 1, 0);
4436 sh = get_active_stripe(conf, sector_nr, 0, 0, 0);
4437 /* make sure we don't swamp the stripe cache if someone else
4438 * is trying to get access
4440 schedule_timeout_uninterruptible(1);
4442 /* Need to check if array will still be degraded after recovery/resync
4443 * We don't need to check the 'failed' flag as when that gets set,
4446 for (i = 0; i < conf->raid_disks; i++)
4447 if (conf->disks[i].rdev == NULL)
4450 bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, still_degraded);
4452 set_bit(STRIPE_SYNC_REQUESTED, &sh->state);
4457 return STRIPE_SECTORS;
4460 static int retry_aligned_read(struct r5conf *conf, struct bio *raid_bio)
4462 /* We may not be able to submit a whole bio at once as there
4463 * may not be enough stripe_heads available.
4464 * We cannot pre-allocate enough stripe_heads as we may need
4465 * more than exist in the cache (if we allow ever large chunks).
4466 * So we do one stripe head at a time and record in
4467 * ->bi_hw_segments how many have been done.
4469 * We *know* that this entire raid_bio is in one chunk, so
4470 * it will be only one 'dd_idx' and only need one call to raid5_compute_sector.
4472 struct stripe_head *sh;
4474 sector_t sector, logical_sector, last_sector;
4479 logical_sector = raid_bio->bi_sector & ~((sector_t)STRIPE_SECTORS-1);
4480 sector = raid5_compute_sector(conf, logical_sector,
4482 last_sector = raid_bio->bi_sector + (raid_bio->bi_size>>9);
4484 for (; logical_sector < last_sector;
4485 logical_sector += STRIPE_SECTORS,
4486 sector += STRIPE_SECTORS,
4489 if (scnt < raid5_bi_hw_segments(raid_bio))
4490 /* already done this stripe */
4493 sh = get_active_stripe(conf, sector, 0, 1, 0);
4496 /* failed to get a stripe - must wait */
4497 raid5_set_bi_hw_segments(raid_bio, scnt);
4498 conf->retry_read_aligned = raid_bio;
4502 if (!add_stripe_bio(sh, raid_bio, dd_idx, 0)) {
4504 raid5_set_bi_hw_segments(raid_bio, scnt);
4505 conf->retry_read_aligned = raid_bio;
4513 spin_lock_irq(&conf->device_lock);
4514 remaining = raid5_dec_bi_phys_segments(raid_bio);
4515 spin_unlock_irq(&conf->device_lock);
4517 bio_endio(raid_bio, 0);
4518 if (atomic_dec_and_test(&conf->active_aligned_reads))
4519 wake_up(&conf->wait_for_stripe);
4525 * This is our raid5 kernel thread.
4527 * We scan the hash table for stripes which can be handled now.
4528 * During the scan, completed stripes are saved for us by the interrupt
4529 * handler, so that they will not have to wait for our next wakeup.
4531 static void raid5d(struct mddev *mddev)
4533 struct stripe_head *sh;
4534 struct r5conf *conf = mddev->private;
4536 struct blk_plug plug;
4538 pr_debug("+++ raid5d active\n");
4540 md_check_recovery(mddev);
4542 blk_start_plug(&plug);
4544 spin_lock_irq(&conf->device_lock);
4548 if (atomic_read(&mddev->plug_cnt) == 0 &&
4549 !list_empty(&conf->bitmap_list)) {
4550 /* Now is a good time to flush some bitmap updates */
4552 spin_unlock_irq(&conf->device_lock);
4553 bitmap_unplug(mddev->bitmap);
4554 spin_lock_irq(&conf->device_lock);
4555 conf->seq_write = conf->seq_flush;
4556 activate_bit_delay(conf);
4558 if (atomic_read(&mddev->plug_cnt) == 0)
4559 raid5_activate_delayed(conf);
4561 while ((bio = remove_bio_from_retry(conf))) {
4563 spin_unlock_irq(&conf->device_lock);
4564 ok = retry_aligned_read(conf, bio);
4565 spin_lock_irq(&conf->device_lock);
4571 sh = __get_priority_stripe(conf);
4575 spin_unlock_irq(&conf->device_lock);
4582 if (mddev->flags & ~(1<<MD_CHANGE_PENDING))
4583 md_check_recovery(mddev);
4585 spin_lock_irq(&conf->device_lock);
4587 pr_debug("%d stripes handled\n", handled);
4589 spin_unlock_irq(&conf->device_lock);
4591 async_tx_issue_pending_all();
4592 blk_finish_plug(&plug);
4594 pr_debug("--- raid5d inactive\n");
4598 raid5_show_stripe_cache_size(struct mddev *mddev, char *page)
4600 struct r5conf *conf = mddev->private;
4602 return sprintf(page, "%d\n", conf->max_nr_stripes);
4608 raid5_set_cache_size(struct mddev *mddev, int size)
4610 struct r5conf *conf = mddev->private;
4613 if (size <= 16 || size > 32768)
4615 while (size < conf->max_nr_stripes) {
4616 if (drop_one_stripe(conf))
4617 conf->max_nr_stripes--;
4621 err = md_allow_write(mddev);
4624 while (size > conf->max_nr_stripes) {
4625 if (grow_one_stripe(conf))
4626 conf->max_nr_stripes++;
4631 EXPORT_SYMBOL(raid5_set_cache_size);
4634 raid5_store_stripe_cache_size(struct mddev *mddev, const char *page, size_t len)
4636 struct r5conf *conf = mddev->private;
4640 if (len >= PAGE_SIZE)
4645 if (strict_strtoul(page, 10, &new))
4647 err = raid5_set_cache_size(mddev, new);
4653 static struct md_sysfs_entry
4654 raid5_stripecache_size = __ATTR(stripe_cache_size, S_IRUGO | S_IWUSR,
4655 raid5_show_stripe_cache_size,
4656 raid5_store_stripe_cache_size);
4659 raid5_show_preread_threshold(struct mddev *mddev, char *page)
4661 struct r5conf *conf = mddev->private;
4663 return sprintf(page, "%d\n", conf->bypass_threshold);
4669 raid5_store_preread_threshold(struct mddev *mddev, const char *page, size_t len)
4671 struct r5conf *conf = mddev->private;
4673 if (len >= PAGE_SIZE)
4678 if (strict_strtoul(page, 10, &new))
4680 if (new > conf->max_nr_stripes)
4682 conf->bypass_threshold = new;
4686 static struct md_sysfs_entry
4687 raid5_preread_bypass_threshold = __ATTR(preread_bypass_threshold,
4689 raid5_show_preread_threshold,
4690 raid5_store_preread_threshold);
4693 stripe_cache_active_show(struct mddev *mddev, char *page)
4695 struct r5conf *conf = mddev->private;
4697 return sprintf(page, "%d\n", atomic_read(&conf->active_stripes));
4702 static struct md_sysfs_entry
4703 raid5_stripecache_active = __ATTR_RO(stripe_cache_active);
4705 static struct attribute *raid5_attrs[] = {
4706 &raid5_stripecache_size.attr,
4707 &raid5_stripecache_active.attr,
4708 &raid5_preread_bypass_threshold.attr,
4711 static struct attribute_group raid5_attrs_group = {
4713 .attrs = raid5_attrs,
4717 raid5_size(struct mddev *mddev, sector_t sectors, int raid_disks)
4719 struct r5conf *conf = mddev->private;
4722 sectors = mddev->dev_sectors;
4724 /* size is defined by the smallest of previous and new size */
4725 raid_disks = min(conf->raid_disks, conf->previous_raid_disks);
4727 sectors &= ~((sector_t)mddev->chunk_sectors - 1);
4728 sectors &= ~((sector_t)mddev->new_chunk_sectors - 1);
4729 return sectors * (raid_disks - conf->max_degraded);
4732 static void raid5_free_percpu(struct r5conf *conf)
4734 struct raid5_percpu *percpu;
4741 for_each_possible_cpu(cpu) {
4742 percpu = per_cpu_ptr(conf->percpu, cpu);
4743 safe_put_page(percpu->spare_page);
4744 kfree(percpu->scribble);
4746 #ifdef CONFIG_HOTPLUG_CPU
4747 unregister_cpu_notifier(&conf->cpu_notify);
4751 free_percpu(conf->percpu);
4754 static void free_conf(struct r5conf *conf)
4756 shrink_stripes(conf);
4757 raid5_free_percpu(conf);
4759 kfree(conf->stripe_hashtbl);
4763 #ifdef CONFIG_HOTPLUG_CPU
4764 static int raid456_cpu_notify(struct notifier_block *nfb, unsigned long action,
4767 struct r5conf *conf = container_of(nfb, struct r5conf, cpu_notify);
4768 long cpu = (long)hcpu;
4769 struct raid5_percpu *percpu = per_cpu_ptr(conf->percpu, cpu);
4772 case CPU_UP_PREPARE:
4773 case CPU_UP_PREPARE_FROZEN:
4774 if (conf->level == 6 && !percpu->spare_page)
4775 percpu->spare_page = alloc_page(GFP_KERNEL);
4776 if (!percpu->scribble)
4777 percpu->scribble = kmalloc(conf->scribble_len, GFP_KERNEL);
4779 if (!percpu->scribble ||
4780 (conf->level == 6 && !percpu->spare_page)) {
4781 safe_put_page(percpu->spare_page);
4782 kfree(percpu->scribble);
4783 pr_err("%s: failed memory allocation for cpu%ld\n",
4785 return notifier_from_errno(-ENOMEM);
4789 case CPU_DEAD_FROZEN:
4790 safe_put_page(percpu->spare_page);
4791 kfree(percpu->scribble);
4792 percpu->spare_page = NULL;
4793 percpu->scribble = NULL;
4802 static int raid5_alloc_percpu(struct r5conf *conf)
4805 struct page *spare_page;
4806 struct raid5_percpu __percpu *allcpus;
4810 allcpus = alloc_percpu(struct raid5_percpu);
4813 conf->percpu = allcpus;
4817 for_each_present_cpu(cpu) {
4818 if (conf->level == 6) {
4819 spare_page = alloc_page(GFP_KERNEL);
4824 per_cpu_ptr(conf->percpu, cpu)->spare_page = spare_page;
4826 scribble = kmalloc(conf->scribble_len, GFP_KERNEL);
4831 per_cpu_ptr(conf->percpu, cpu)->scribble = scribble;
4833 #ifdef CONFIG_HOTPLUG_CPU
4834 conf->cpu_notify.notifier_call = raid456_cpu_notify;
4835 conf->cpu_notify.priority = 0;
4837 err = register_cpu_notifier(&conf->cpu_notify);
4844 static struct r5conf *setup_conf(struct mddev *mddev)
4846 struct r5conf *conf;
4847 int raid_disk, memory, max_disks;
4848 struct md_rdev *rdev;
4849 struct disk_info *disk;
4852 if (mddev->new_level != 5
4853 && mddev->new_level != 4
4854 && mddev->new_level != 6) {
4855 printk(KERN_ERR "md/raid:%s: raid level not set to 4/5/6 (%d)\n",
4856 mdname(mddev), mddev->new_level);
4857 return ERR_PTR(-EIO);
4859 if ((mddev->new_level == 5
4860 && !algorithm_valid_raid5(mddev->new_layout)) ||
4861 (mddev->new_level == 6
4862 && !algorithm_valid_raid6(mddev->new_layout))) {
4863 printk(KERN_ERR "md/raid:%s: layout %d not supported\n",
4864 mdname(mddev), mddev->new_layout);
4865 return ERR_PTR(-EIO);
4867 if (mddev->new_level == 6 && mddev->raid_disks < 4) {
4868 printk(KERN_ERR "md/raid:%s: not enough configured devices (%d, minimum 4)\n",
4869 mdname(mddev), mddev->raid_disks);
4870 return ERR_PTR(-EINVAL);
4873 if (!mddev->new_chunk_sectors ||
4874 (mddev->new_chunk_sectors << 9) % PAGE_SIZE ||
4875 !is_power_of_2(mddev->new_chunk_sectors)) {
4876 printk(KERN_ERR "md/raid:%s: invalid chunk size %d\n",
4877 mdname(mddev), mddev->new_chunk_sectors << 9);
4878 return ERR_PTR(-EINVAL);
4881 conf = kzalloc(sizeof(struct r5conf), GFP_KERNEL);
4884 spin_lock_init(&conf->device_lock);
4885 init_waitqueue_head(&conf->wait_for_stripe);
4886 init_waitqueue_head(&conf->wait_for_overlap);
4887 INIT_LIST_HEAD(&conf->handle_list);
4888 INIT_LIST_HEAD(&conf->hold_list);
4889 INIT_LIST_HEAD(&conf->delayed_list);
4890 INIT_LIST_HEAD(&conf->bitmap_list);
4891 INIT_LIST_HEAD(&conf->inactive_list);
4892 atomic_set(&conf->active_stripes, 0);
4893 atomic_set(&conf->preread_active_stripes, 0);
4894 atomic_set(&conf->active_aligned_reads, 0);
4895 conf->bypass_threshold = BYPASS_THRESHOLD;
4896 conf->recovery_disabled = mddev->recovery_disabled - 1;
4898 conf->raid_disks = mddev->raid_disks;
4899 if (mddev->reshape_position == MaxSector)
4900 conf->previous_raid_disks = mddev->raid_disks;
4902 conf->previous_raid_disks = mddev->raid_disks - mddev->delta_disks;
4903 max_disks = max(conf->raid_disks, conf->previous_raid_disks);
4904 conf->scribble_len = scribble_len(max_disks);
4906 conf->disks = kzalloc(max_disks * sizeof(struct disk_info),
4911 conf->mddev = mddev;
4913 if ((conf->stripe_hashtbl = kzalloc(PAGE_SIZE, GFP_KERNEL)) == NULL)
4916 conf->level = mddev->new_level;
4917 if (raid5_alloc_percpu(conf) != 0)
4920 pr_debug("raid456: run(%s) called.\n", mdname(mddev));
4922 rdev_for_each(rdev, mddev) {
4923 raid_disk = rdev->raid_disk;
4924 if (raid_disk >= max_disks
4927 disk = conf->disks + raid_disk;
4929 if (test_bit(Replacement, &rdev->flags)) {
4930 if (disk->replacement)
4932 disk->replacement = rdev;
4939 if (test_bit(In_sync, &rdev->flags)) {
4940 char b[BDEVNAME_SIZE];
4941 printk(KERN_INFO "md/raid:%s: device %s operational as raid"
4943 mdname(mddev), bdevname(rdev->bdev, b), raid_disk);
4944 } else if (rdev->saved_raid_disk != raid_disk)
4945 /* Cannot rely on bitmap to complete recovery */
4949 conf->chunk_sectors = mddev->new_chunk_sectors;
4950 conf->level = mddev->new_level;
4951 if (conf->level == 6)
4952 conf->max_degraded = 2;
4954 conf->max_degraded = 1;
4955 conf->algorithm = mddev->new_layout;
4956 conf->max_nr_stripes = NR_STRIPES;
4957 conf->reshape_progress = mddev->reshape_position;
4958 if (conf->reshape_progress != MaxSector) {
4959 conf->prev_chunk_sectors = mddev->chunk_sectors;
4960 conf->prev_algo = mddev->layout;
4963 memory = conf->max_nr_stripes * (sizeof(struct stripe_head) +
4964 max_disks * ((sizeof(struct bio) + PAGE_SIZE))) / 1024;
4965 if (grow_stripes(conf, conf->max_nr_stripes)) {
4967 "md/raid:%s: couldn't allocate %dkB for buffers\n",
4968 mdname(mddev), memory);
4971 printk(KERN_INFO "md/raid:%s: allocated %dkB\n",
4972 mdname(mddev), memory);
4974 sprintf(pers_name, "raid%d", mddev->new_level);
4975 conf->thread = md_register_thread(raid5d, mddev, pers_name);
4976 if (!conf->thread) {
4978 "md/raid:%s: couldn't allocate thread.\n",
4988 return ERR_PTR(-EIO);
4990 return ERR_PTR(-ENOMEM);
4994 static int only_parity(int raid_disk, int algo, int raid_disks, int max_degraded)
4997 case ALGORITHM_PARITY_0:
4998 if (raid_disk < max_degraded)
5001 case ALGORITHM_PARITY_N:
5002 if (raid_disk >= raid_disks - max_degraded)
5005 case ALGORITHM_PARITY_0_6:
5006 if (raid_disk == 0 ||
5007 raid_disk == raid_disks - 1)
5010 case ALGORITHM_LEFT_ASYMMETRIC_6:
5011 case ALGORITHM_RIGHT_ASYMMETRIC_6:
5012 case ALGORITHM_LEFT_SYMMETRIC_6:
5013 case ALGORITHM_RIGHT_SYMMETRIC_6:
5014 if (raid_disk == raid_disks - 1)
5020 static int run(struct mddev *mddev)
5022 struct r5conf *conf;
5023 int working_disks = 0;
5024 int dirty_parity_disks = 0;
5025 struct md_rdev *rdev;
5026 sector_t reshape_offset = 0;
5028 long long min_offset_diff = 0;
5031 if (mddev->recovery_cp != MaxSector)
5032 printk(KERN_NOTICE "md/raid:%s: not clean"
5033 " -- starting background reconstruction\n",
5036 rdev_for_each(rdev, mddev) {
5038 if (rdev->raid_disk < 0)
5040 diff = (rdev->new_data_offset - rdev->data_offset);
5042 min_offset_diff = diff;
5044 } else if (mddev->reshape_backwards &&
5045 diff < min_offset_diff)
5046 min_offset_diff = diff;
5047 else if (!mddev->reshape_backwards &&
5048 diff > min_offset_diff)
5049 min_offset_diff = diff;
5052 if (mddev->reshape_position != MaxSector) {
5053 /* Check that we can continue the reshape.
5054 * Difficulties arise if the stripe we would write to
5055 * next is at or after the stripe we would read from next.
5056 * For a reshape that changes the number of devices, this
5057 * is only possible for a very short time, and mdadm makes
5058 * sure that time appears to have past before assembling
5059 * the array. So we fail if that time hasn't passed.
5060 * For a reshape that keeps the number of devices the same
5061 * mdadm must be monitoring the reshape can keeping the
5062 * critical areas read-only and backed up. It will start
5063 * the array in read-only mode, so we check for that.
5065 sector_t here_new, here_old;
5067 int max_degraded = (mddev->level == 6 ? 2 : 1);
5069 if (mddev->new_level != mddev->level) {
5070 printk(KERN_ERR "md/raid:%s: unsupported reshape "
5071 "required - aborting.\n",
5075 old_disks = mddev->raid_disks - mddev->delta_disks;
5076 /* reshape_position must be on a new-stripe boundary, and one
5077 * further up in new geometry must map after here in old
5080 here_new = mddev->reshape_position;
5081 if (sector_div(here_new, mddev->new_chunk_sectors *
5082 (mddev->raid_disks - max_degraded))) {
5083 printk(KERN_ERR "md/raid:%s: reshape_position not "
5084 "on a stripe boundary\n", mdname(mddev));
5087 reshape_offset = here_new * mddev->new_chunk_sectors;
5088 /* here_new is the stripe we will write to */
5089 here_old = mddev->reshape_position;
5090 sector_div(here_old, mddev->chunk_sectors *
5091 (old_disks-max_degraded));
5092 /* here_old is the first stripe that we might need to read
5094 if (mddev->delta_disks == 0) {
5095 if ((here_new * mddev->new_chunk_sectors !=
5096 here_old * mddev->chunk_sectors)) {
5097 printk(KERN_ERR "md/raid:%s: reshape position is"
5098 " confused - aborting\n", mdname(mddev));
5101 /* We cannot be sure it is safe to start an in-place
5102 * reshape. It is only safe if user-space is monitoring
5103 * and taking constant backups.
5104 * mdadm always starts a situation like this in
5105 * readonly mode so it can take control before
5106 * allowing any writes. So just check for that.
5108 if (abs(min_offset_diff) >= mddev->chunk_sectors &&
5109 abs(min_offset_diff) >= mddev->new_chunk_sectors)
5110 /* not really in-place - so OK */;
5111 else if (mddev->ro == 0) {
5112 printk(KERN_ERR "md/raid:%s: in-place reshape "
5113 "must be started in read-only mode "
5118 } else if (mddev->reshape_backwards
5119 ? (here_new * mddev->new_chunk_sectors + min_offset_diff <=
5120 here_old * mddev->chunk_sectors)
5121 : (here_new * mddev->new_chunk_sectors >=
5122 here_old * mddev->chunk_sectors + (-min_offset_diff))) {
5123 /* Reading from the same stripe as writing to - bad */
5124 printk(KERN_ERR "md/raid:%s: reshape_position too early for "
5125 "auto-recovery - aborting.\n",
5129 printk(KERN_INFO "md/raid:%s: reshape will continue\n",
5131 /* OK, we should be able to continue; */
5133 BUG_ON(mddev->level != mddev->new_level);
5134 BUG_ON(mddev->layout != mddev->new_layout);
5135 BUG_ON(mddev->chunk_sectors != mddev->new_chunk_sectors);
5136 BUG_ON(mddev->delta_disks != 0);
5139 if (mddev->private == NULL)
5140 conf = setup_conf(mddev);
5142 conf = mddev->private;
5145 return PTR_ERR(conf);
5147 conf->min_offset_diff = min_offset_diff;
5148 mddev->thread = conf->thread;
5149 conf->thread = NULL;
5150 mddev->private = conf;
5152 for (i = 0; i < conf->raid_disks && conf->previous_raid_disks;
5154 rdev = conf->disks[i].rdev;
5155 if (!rdev && conf->disks[i].replacement) {
5156 /* The replacement is all we have yet */
5157 rdev = conf->disks[i].replacement;
5158 conf->disks[i].replacement = NULL;
5159 clear_bit(Replacement, &rdev->flags);
5160 conf->disks[i].rdev = rdev;
5164 if (conf->disks[i].replacement &&
5165 conf->reshape_progress != MaxSector) {
5166 /* replacements and reshape simply do not mix. */
5167 printk(KERN_ERR "md: cannot handle concurrent "
5168 "replacement and reshape.\n");
5171 if (test_bit(In_sync, &rdev->flags)) {
5175 /* This disc is not fully in-sync. However if it
5176 * just stored parity (beyond the recovery_offset),
5177 * when we don't need to be concerned about the
5178 * array being dirty.
5179 * When reshape goes 'backwards', we never have
5180 * partially completed devices, so we only need
5181 * to worry about reshape going forwards.
5183 /* Hack because v0.91 doesn't store recovery_offset properly. */
5184 if (mddev->major_version == 0 &&
5185 mddev->minor_version > 90)
5186 rdev->recovery_offset = reshape_offset;
5188 if (rdev->recovery_offset < reshape_offset) {
5189 /* We need to check old and new layout */
5190 if (!only_parity(rdev->raid_disk,
5193 conf->max_degraded))
5196 if (!only_parity(rdev->raid_disk,
5198 conf->previous_raid_disks,
5199 conf->max_degraded))
5201 dirty_parity_disks++;
5205 * 0 for a fully functional array, 1 or 2 for a degraded array.
5207 mddev->degraded = calc_degraded(conf);
5209 if (has_failed(conf)) {
5210 printk(KERN_ERR "md/raid:%s: not enough operational devices"
5211 " (%d/%d failed)\n",
5212 mdname(mddev), mddev->degraded, conf->raid_disks);
5216 /* device size must be a multiple of chunk size */
5217 mddev->dev_sectors &= ~(mddev->chunk_sectors - 1);
5218 mddev->resync_max_sectors = mddev->dev_sectors;
5220 if (mddev->degraded > dirty_parity_disks &&
5221 mddev->recovery_cp != MaxSector) {
5222 if (mddev->ok_start_degraded)
5224 "md/raid:%s: starting dirty degraded array"
5225 " - data corruption possible.\n",
5229 "md/raid:%s: cannot start dirty degraded array.\n",
5235 if (mddev->degraded == 0)
5236 printk(KERN_INFO "md/raid:%s: raid level %d active with %d out of %d"
5237 " devices, algorithm %d\n", mdname(mddev), conf->level,
5238 mddev->raid_disks-mddev->degraded, mddev->raid_disks,
5241 printk(KERN_ALERT "md/raid:%s: raid level %d active with %d"
5242 " out of %d devices, algorithm %d\n",
5243 mdname(mddev), conf->level,
5244 mddev->raid_disks - mddev->degraded,
5245 mddev->raid_disks, mddev->new_layout);
5247 print_raid5_conf(conf);
5249 if (conf->reshape_progress != MaxSector) {
5250 conf->reshape_safe = conf->reshape_progress;
5251 atomic_set(&conf->reshape_stripes, 0);
5252 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
5253 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
5254 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
5255 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
5256 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
5261 /* Ok, everything is just fine now */
5262 if (mddev->to_remove == &raid5_attrs_group)
5263 mddev->to_remove = NULL;
5264 else if (mddev->kobj.sd &&
5265 sysfs_create_group(&mddev->kobj, &raid5_attrs_group))
5267 "raid5: failed to create sysfs attributes for %s\n",
5269 md_set_array_sectors(mddev, raid5_size(mddev, 0, 0));
5273 /* read-ahead size must cover two whole stripes, which
5274 * is 2 * (datadisks) * chunksize where 'n' is the
5275 * number of raid devices
5277 int data_disks = conf->previous_raid_disks - conf->max_degraded;
5278 int stripe = data_disks *
5279 ((mddev->chunk_sectors << 9) / PAGE_SIZE);
5280 if (mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
5281 mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
5283 blk_queue_merge_bvec(mddev->queue, raid5_mergeable_bvec);
5285 mddev->queue->backing_dev_info.congested_data = mddev;
5286 mddev->queue->backing_dev_info.congested_fn = raid5_congested;
5288 chunk_size = mddev->chunk_sectors << 9;
5289 blk_queue_io_min(mddev->queue, chunk_size);
5290 blk_queue_io_opt(mddev->queue, chunk_size *
5291 (conf->raid_disks - conf->max_degraded));
5293 rdev_for_each(rdev, mddev) {
5294 disk_stack_limits(mddev->gendisk, rdev->bdev,
5295 rdev->data_offset << 9);
5296 disk_stack_limits(mddev->gendisk, rdev->bdev,
5297 rdev->new_data_offset << 9);
5303 md_unregister_thread(&mddev->thread);
5304 print_raid5_conf(conf);
5306 mddev->private = NULL;
5307 printk(KERN_ALERT "md/raid:%s: failed to run raid set.\n", mdname(mddev));
5311 static int stop(struct mddev *mddev)
5313 struct r5conf *conf = mddev->private;
5315 md_unregister_thread(&mddev->thread);
5317 mddev->queue->backing_dev_info.congested_fn = NULL;
5319 mddev->private = NULL;
5320 mddev->to_remove = &raid5_attrs_group;
5324 static void status(struct seq_file *seq, struct mddev *mddev)
5326 struct r5conf *conf = mddev->private;
5329 seq_printf(seq, " level %d, %dk chunk, algorithm %d", mddev->level,
5330 mddev->chunk_sectors / 2, mddev->layout);
5331 seq_printf (seq, " [%d/%d] [", conf->raid_disks, conf->raid_disks - mddev->degraded);
5332 for (i = 0; i < conf->raid_disks; i++)
5333 seq_printf (seq, "%s",
5334 conf->disks[i].rdev &&
5335 test_bit(In_sync, &conf->disks[i].rdev->flags) ? "U" : "_");
5336 seq_printf (seq, "]");
5339 static void print_raid5_conf (struct r5conf *conf)
5342 struct disk_info *tmp;
5344 printk(KERN_DEBUG "RAID conf printout:\n");
5346 printk("(conf==NULL)\n");
5349 printk(KERN_DEBUG " --- level:%d rd:%d wd:%d\n", conf->level,
5351 conf->raid_disks - conf->mddev->degraded);
5353 for (i = 0; i < conf->raid_disks; i++) {
5354 char b[BDEVNAME_SIZE];
5355 tmp = conf->disks + i;
5357 printk(KERN_DEBUG " disk %d, o:%d, dev:%s\n",
5358 i, !test_bit(Faulty, &tmp->rdev->flags),
5359 bdevname(tmp->rdev->bdev, b));
5363 static int raid5_spare_active(struct mddev *mddev)
5366 struct r5conf *conf = mddev->private;
5367 struct disk_info *tmp;
5369 unsigned long flags;
5371 for (i = 0; i < conf->raid_disks; i++) {
5372 tmp = conf->disks + i;
5373 if (tmp->replacement
5374 && tmp->replacement->recovery_offset == MaxSector
5375 && !test_bit(Faulty, &tmp->replacement->flags)
5376 && !test_and_set_bit(In_sync, &tmp->replacement->flags)) {
5377 /* Replacement has just become active. */
5379 || !test_and_clear_bit(In_sync, &tmp->rdev->flags))
5382 /* Replaced device not technically faulty,
5383 * but we need to be sure it gets removed
5384 * and never re-added.
5386 set_bit(Faulty, &tmp->rdev->flags);
5387 sysfs_notify_dirent_safe(
5388 tmp->rdev->sysfs_state);
5390 sysfs_notify_dirent_safe(tmp->replacement->sysfs_state);
5391 } else if (tmp->rdev
5392 && tmp->rdev->recovery_offset == MaxSector
5393 && !test_bit(Faulty, &tmp->rdev->flags)
5394 && !test_and_set_bit(In_sync, &tmp->rdev->flags)) {
5396 sysfs_notify_dirent_safe(tmp->rdev->sysfs_state);
5399 spin_lock_irqsave(&conf->device_lock, flags);
5400 mddev->degraded = calc_degraded(conf);
5401 spin_unlock_irqrestore(&conf->device_lock, flags);
5402 print_raid5_conf(conf);
5406 static int raid5_remove_disk(struct mddev *mddev, struct md_rdev *rdev)
5408 struct r5conf *conf = mddev->private;
5410 int number = rdev->raid_disk;
5411 struct md_rdev **rdevp;
5412 struct disk_info *p = conf->disks + number;
5414 print_raid5_conf(conf);
5415 if (rdev == p->rdev)
5417 else if (rdev == p->replacement)
5418 rdevp = &p->replacement;
5422 if (number >= conf->raid_disks &&
5423 conf->reshape_progress == MaxSector)
5424 clear_bit(In_sync, &rdev->flags);
5426 if (test_bit(In_sync, &rdev->flags) ||
5427 atomic_read(&rdev->nr_pending)) {
5431 /* Only remove non-faulty devices if recovery
5434 if (!test_bit(Faulty, &rdev->flags) &&
5435 mddev->recovery_disabled != conf->recovery_disabled &&
5436 !has_failed(conf) &&
5437 (!p->replacement || p->replacement == rdev) &&
5438 number < conf->raid_disks) {
5444 if (atomic_read(&rdev->nr_pending)) {
5445 /* lost the race, try later */
5448 } else if (p->replacement) {
5449 /* We must have just cleared 'rdev' */
5450 p->rdev = p->replacement;
5451 clear_bit(Replacement, &p->replacement->flags);
5452 smp_mb(); /* Make sure other CPUs may see both as identical
5453 * but will never see neither - if they are careful
5455 p->replacement = NULL;
5456 clear_bit(WantReplacement, &rdev->flags);
5458 /* We might have just removed the Replacement as faulty-
5459 * clear the bit just in case
5461 clear_bit(WantReplacement, &rdev->flags);
5464 print_raid5_conf(conf);
5468 static int raid5_add_disk(struct mddev *mddev, struct md_rdev *rdev)
5470 struct r5conf *conf = mddev->private;
5473 struct disk_info *p;
5475 int last = conf->raid_disks - 1;
5477 if (mddev->recovery_disabled == conf->recovery_disabled)
5480 if (rdev->saved_raid_disk < 0 && has_failed(conf))
5481 /* no point adding a device */
5484 if (rdev->raid_disk >= 0)
5485 first = last = rdev->raid_disk;
5488 * find the disk ... but prefer rdev->saved_raid_disk
5491 if (rdev->saved_raid_disk >= 0 &&
5492 rdev->saved_raid_disk >= first &&
5493 conf->disks[rdev->saved_raid_disk].rdev == NULL)
5494 first = rdev->saved_raid_disk;
5496 for (disk = first; disk <= last; disk++) {
5497 p = conf->disks + disk;
5498 if (p->rdev == NULL) {
5499 clear_bit(In_sync, &rdev->flags);
5500 rdev->raid_disk = disk;
5502 if (rdev->saved_raid_disk != disk)
5504 rcu_assign_pointer(p->rdev, rdev);
5508 for (disk = first; disk <= last; disk++) {
5509 p = conf->disks + disk;
5510 if (test_bit(WantReplacement, &p->rdev->flags) &&
5511 p->replacement == NULL) {
5512 clear_bit(In_sync, &rdev->flags);
5513 set_bit(Replacement, &rdev->flags);
5514 rdev->raid_disk = disk;
5517 rcu_assign_pointer(p->replacement, rdev);
5522 print_raid5_conf(conf);
5526 static int raid5_resize(struct mddev *mddev, sector_t sectors)
5528 /* no resync is happening, and there is enough space
5529 * on all devices, so we can resize.
5530 * We need to make sure resync covers any new space.
5531 * If the array is shrinking we should possibly wait until
5532 * any io in the removed space completes, but it hardly seems
5536 sectors &= ~((sector_t)mddev->chunk_sectors - 1);
5537 newsize = raid5_size(mddev, sectors, mddev->raid_disks);
5538 if (mddev->external_size &&
5539 mddev->array_sectors > newsize)
5541 if (mddev->bitmap) {
5542 int ret = bitmap_resize(mddev->bitmap, sectors, 0, 0);
5546 md_set_array_sectors(mddev, newsize);
5547 set_capacity(mddev->gendisk, mddev->array_sectors);
5548 revalidate_disk(mddev->gendisk);
5549 if (sectors > mddev->dev_sectors &&
5550 mddev->recovery_cp > mddev->dev_sectors) {
5551 mddev->recovery_cp = mddev->dev_sectors;
5552 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
5554 mddev->dev_sectors = sectors;
5555 mddev->resync_max_sectors = sectors;
5559 static int check_stripe_cache(struct mddev *mddev)
5561 /* Can only proceed if there are plenty of stripe_heads.
5562 * We need a minimum of one full stripe,, and for sensible progress
5563 * it is best to have about 4 times that.
5564 * If we require 4 times, then the default 256 4K stripe_heads will
5565 * allow for chunk sizes up to 256K, which is probably OK.
5566 * If the chunk size is greater, user-space should request more
5567 * stripe_heads first.
5569 struct r5conf *conf = mddev->private;
5570 if (((mddev->chunk_sectors << 9) / STRIPE_SIZE) * 4
5571 > conf->max_nr_stripes ||
5572 ((mddev->new_chunk_sectors << 9) / STRIPE_SIZE) * 4
5573 > conf->max_nr_stripes) {
5574 printk(KERN_WARNING "md/raid:%s: reshape: not enough stripes. Needed %lu\n",
5576 ((max(mddev->chunk_sectors, mddev->new_chunk_sectors) << 9)
5583 static int check_reshape(struct mddev *mddev)
5585 struct r5conf *conf = mddev->private;
5587 if (mddev->delta_disks == 0 &&
5588 mddev->new_layout == mddev->layout &&
5589 mddev->new_chunk_sectors == mddev->chunk_sectors)
5590 return 0; /* nothing to do */
5591 if (has_failed(conf))
5593 if (mddev->delta_disks < 0) {
5594 /* We might be able to shrink, but the devices must
5595 * be made bigger first.
5596 * For raid6, 4 is the minimum size.
5597 * Otherwise 2 is the minimum
5600 if (mddev->level == 6)
5602 if (mddev->raid_disks + mddev->delta_disks < min)
5606 if (!check_stripe_cache(mddev))
5609 return resize_stripes(conf, conf->raid_disks + mddev->delta_disks);
5612 static int raid5_start_reshape(struct mddev *mddev)
5614 struct r5conf *conf = mddev->private;
5615 struct md_rdev *rdev;
5617 unsigned long flags;
5619 if (test_bit(MD_RECOVERY_RUNNING, &mddev->recovery))
5622 if (!check_stripe_cache(mddev))
5625 if (has_failed(conf))
5628 rdev_for_each(rdev, mddev) {
5629 if (!test_bit(In_sync, &rdev->flags)
5630 && !test_bit(Faulty, &rdev->flags))
5634 if (spares - mddev->degraded < mddev->delta_disks - conf->max_degraded)
5635 /* Not enough devices even to make a degraded array
5640 /* Refuse to reduce size of the array. Any reductions in
5641 * array size must be through explicit setting of array_size
5644 if (raid5_size(mddev, 0, conf->raid_disks + mddev->delta_disks)
5645 < mddev->array_sectors) {
5646 printk(KERN_ERR "md/raid:%s: array size must be reduced "
5647 "before number of disks\n", mdname(mddev));
5651 atomic_set(&conf->reshape_stripes, 0);
5652 spin_lock_irq(&conf->device_lock);
5653 conf->previous_raid_disks = conf->raid_disks;
5654 conf->raid_disks += mddev->delta_disks;
5655 conf->prev_chunk_sectors = conf->chunk_sectors;
5656 conf->chunk_sectors = mddev->new_chunk_sectors;
5657 conf->prev_algo = conf->algorithm;
5658 conf->algorithm = mddev->new_layout;
5660 /* Code that selects data_offset needs to see the generation update
5661 * if reshape_progress has been set - so a memory barrier needed.
5664 if (mddev->reshape_backwards)
5665 conf->reshape_progress = raid5_size(mddev, 0, 0);
5667 conf->reshape_progress = 0;
5668 conf->reshape_safe = conf->reshape_progress;
5669 spin_unlock_irq(&conf->device_lock);
5671 /* Add some new drives, as many as will fit.
5672 * We know there are enough to make the newly sized array work.
5673 * Don't add devices if we are reducing the number of
5674 * devices in the array. This is because it is not possible
5675 * to correctly record the "partially reconstructed" state of
5676 * such devices during the reshape and confusion could result.
5678 if (mddev->delta_disks >= 0) {
5679 rdev_for_each(rdev, mddev)
5680 if (rdev->raid_disk < 0 &&
5681 !test_bit(Faulty, &rdev->flags)) {
5682 if (raid5_add_disk(mddev, rdev) == 0) {
5684 >= conf->previous_raid_disks)
5685 set_bit(In_sync, &rdev->flags);
5687 rdev->recovery_offset = 0;
5689 if (sysfs_link_rdev(mddev, rdev))
5690 /* Failure here is OK */;
5692 } else if (rdev->raid_disk >= conf->previous_raid_disks
5693 && !test_bit(Faulty, &rdev->flags)) {
5694 /* This is a spare that was manually added */
5695 set_bit(In_sync, &rdev->flags);
5698 /* When a reshape changes the number of devices,
5699 * ->degraded is measured against the larger of the
5700 * pre and post number of devices.
5702 spin_lock_irqsave(&conf->device_lock, flags);
5703 mddev->degraded = calc_degraded(conf);
5704 spin_unlock_irqrestore(&conf->device_lock, flags);
5706 mddev->raid_disks = conf->raid_disks;
5707 mddev->reshape_position = conf->reshape_progress;
5708 set_bit(MD_CHANGE_DEVS, &mddev->flags);
5710 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
5711 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
5712 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
5713 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
5714 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
5716 if (!mddev->sync_thread) {
5717 mddev->recovery = 0;
5718 spin_lock_irq(&conf->device_lock);
5719 mddev->raid_disks = conf->raid_disks = conf->previous_raid_disks;
5720 rdev_for_each(rdev, mddev)
5721 rdev->new_data_offset = rdev->data_offset;
5723 conf->reshape_progress = MaxSector;
5724 mddev->reshape_position = MaxSector;
5725 spin_unlock_irq(&conf->device_lock);
5728 conf->reshape_checkpoint = jiffies;
5729 md_wakeup_thread(mddev->sync_thread);
5730 md_new_event(mddev);
5734 /* This is called from the reshape thread and should make any
5735 * changes needed in 'conf'
5737 static void end_reshape(struct r5conf *conf)
5740 if (!test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery)) {
5741 struct md_rdev *rdev;
5743 spin_lock_irq(&conf->device_lock);
5744 conf->previous_raid_disks = conf->raid_disks;
5745 rdev_for_each(rdev, conf->mddev)
5746 rdev->data_offset = rdev->new_data_offset;
5748 conf->reshape_progress = MaxSector;
5749 spin_unlock_irq(&conf->device_lock);
5750 wake_up(&conf->wait_for_overlap);
5752 /* read-ahead size must cover two whole stripes, which is
5753 * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
5755 if (conf->mddev->queue) {
5756 int data_disks = conf->raid_disks - conf->max_degraded;
5757 int stripe = data_disks * ((conf->chunk_sectors << 9)
5759 if (conf->mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
5760 conf->mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
5765 /* This is called from the raid5d thread with mddev_lock held.
5766 * It makes config changes to the device.
5768 static void raid5_finish_reshape(struct mddev *mddev)
5770 struct r5conf *conf = mddev->private;
5772 if (!test_bit(MD_RECOVERY_INTR, &mddev->recovery)) {
5774 if (mddev->delta_disks > 0) {
5775 md_set_array_sectors(mddev, raid5_size(mddev, 0, 0));
5776 set_capacity(mddev->gendisk, mddev->array_sectors);
5777 revalidate_disk(mddev->gendisk);
5780 spin_lock_irq(&conf->device_lock);
5781 mddev->degraded = calc_degraded(conf);
5782 spin_unlock_irq(&conf->device_lock);
5783 for (d = conf->raid_disks ;
5784 d < conf->raid_disks - mddev->delta_disks;
5786 struct md_rdev *rdev = conf->disks[d].rdev;
5788 clear_bit(In_sync, &rdev->flags);
5789 rdev = conf->disks[d].replacement;
5791 clear_bit(In_sync, &rdev->flags);
5794 mddev->layout = conf->algorithm;
5795 mddev->chunk_sectors = conf->chunk_sectors;
5796 mddev->reshape_position = MaxSector;
5797 mddev->delta_disks = 0;
5798 mddev->reshape_backwards = 0;
5802 static void raid5_quiesce(struct mddev *mddev, int state)
5804 struct r5conf *conf = mddev->private;
5807 case 2: /* resume for a suspend */
5808 wake_up(&conf->wait_for_overlap);
5811 case 1: /* stop all writes */
5812 spin_lock_irq(&conf->device_lock);
5813 /* '2' tells resync/reshape to pause so that all
5814 * active stripes can drain
5817 wait_event_lock_irq(conf->wait_for_stripe,
5818 atomic_read(&conf->active_stripes) == 0 &&
5819 atomic_read(&conf->active_aligned_reads) == 0,
5820 conf->device_lock, /* nothing */);
5822 spin_unlock_irq(&conf->device_lock);
5823 /* allow reshape to continue */
5824 wake_up(&conf->wait_for_overlap);
5827 case 0: /* re-enable writes */
5828 spin_lock_irq(&conf->device_lock);
5830 wake_up(&conf->wait_for_stripe);
5831 wake_up(&conf->wait_for_overlap);
5832 spin_unlock_irq(&conf->device_lock);
5838 static void *raid45_takeover_raid0(struct mddev *mddev, int level)
5840 struct r0conf *raid0_conf = mddev->private;
5843 /* for raid0 takeover only one zone is supported */
5844 if (raid0_conf->nr_strip_zones > 1) {
5845 printk(KERN_ERR "md/raid:%s: cannot takeover raid0 with more than one zone.\n",
5847 return ERR_PTR(-EINVAL);
5850 sectors = raid0_conf->strip_zone[0].zone_end;
5851 sector_div(sectors, raid0_conf->strip_zone[0].nb_dev);
5852 mddev->dev_sectors = sectors;
5853 mddev->new_level = level;
5854 mddev->new_layout = ALGORITHM_PARITY_N;
5855 mddev->new_chunk_sectors = mddev->chunk_sectors;
5856 mddev->raid_disks += 1;
5857 mddev->delta_disks = 1;
5858 /* make sure it will be not marked as dirty */
5859 mddev->recovery_cp = MaxSector;
5861 return setup_conf(mddev);
5865 static void *raid5_takeover_raid1(struct mddev *mddev)
5869 if (mddev->raid_disks != 2 ||
5870 mddev->degraded > 1)
5871 return ERR_PTR(-EINVAL);
5873 /* Should check if there are write-behind devices? */
5875 chunksect = 64*2; /* 64K by default */
5877 /* The array must be an exact multiple of chunksize */
5878 while (chunksect && (mddev->array_sectors & (chunksect-1)))
5881 if ((chunksect<<9) < STRIPE_SIZE)
5882 /* array size does not allow a suitable chunk size */
5883 return ERR_PTR(-EINVAL);
5885 mddev->new_level = 5;
5886 mddev->new_layout = ALGORITHM_LEFT_SYMMETRIC;
5887 mddev->new_chunk_sectors = chunksect;
5889 return setup_conf(mddev);
5892 static void *raid5_takeover_raid6(struct mddev *mddev)
5896 switch (mddev->layout) {
5897 case ALGORITHM_LEFT_ASYMMETRIC_6:
5898 new_layout = ALGORITHM_LEFT_ASYMMETRIC;
5900 case ALGORITHM_RIGHT_ASYMMETRIC_6:
5901 new_layout = ALGORITHM_RIGHT_ASYMMETRIC;
5903 case ALGORITHM_LEFT_SYMMETRIC_6:
5904 new_layout = ALGORITHM_LEFT_SYMMETRIC;
5906 case ALGORITHM_RIGHT_SYMMETRIC_6:
5907 new_layout = ALGORITHM_RIGHT_SYMMETRIC;
5909 case ALGORITHM_PARITY_0_6:
5910 new_layout = ALGORITHM_PARITY_0;
5912 case ALGORITHM_PARITY_N:
5913 new_layout = ALGORITHM_PARITY_N;
5916 return ERR_PTR(-EINVAL);
5918 mddev->new_level = 5;
5919 mddev->new_layout = new_layout;
5920 mddev->delta_disks = -1;
5921 mddev->raid_disks -= 1;
5922 return setup_conf(mddev);
5926 static int raid5_check_reshape(struct mddev *mddev)
5928 /* For a 2-drive array, the layout and chunk size can be changed
5929 * immediately as not restriping is needed.
5930 * For larger arrays we record the new value - after validation
5931 * to be used by a reshape pass.
5933 struct r5conf *conf = mddev->private;
5934 int new_chunk = mddev->new_chunk_sectors;
5936 if (mddev->new_layout >= 0 && !algorithm_valid_raid5(mddev->new_layout))
5938 if (new_chunk > 0) {
5939 if (!is_power_of_2(new_chunk))
5941 if (new_chunk < (PAGE_SIZE>>9))
5943 if (mddev->array_sectors & (new_chunk-1))
5944 /* not factor of array size */
5948 /* They look valid */
5950 if (mddev->raid_disks == 2) {
5951 /* can make the change immediately */
5952 if (mddev->new_layout >= 0) {
5953 conf->algorithm = mddev->new_layout;
5954 mddev->layout = mddev->new_layout;
5956 if (new_chunk > 0) {
5957 conf->chunk_sectors = new_chunk ;
5958 mddev->chunk_sectors = new_chunk;
5960 set_bit(MD_CHANGE_DEVS, &mddev->flags);
5961 md_wakeup_thread(mddev->thread);
5963 return check_reshape(mddev);
5966 static int raid6_check_reshape(struct mddev *mddev)
5968 int new_chunk = mddev->new_chunk_sectors;
5970 if (mddev->new_layout >= 0 && !algorithm_valid_raid6(mddev->new_layout))
5972 if (new_chunk > 0) {
5973 if (!is_power_of_2(new_chunk))
5975 if (new_chunk < (PAGE_SIZE >> 9))
5977 if (mddev->array_sectors & (new_chunk-1))
5978 /* not factor of array size */
5982 /* They look valid */
5983 return check_reshape(mddev);
5986 static void *raid5_takeover(struct mddev *mddev)
5988 /* raid5 can take over:
5989 * raid0 - if there is only one strip zone - make it a raid4 layout
5990 * raid1 - if there are two drives. We need to know the chunk size
5991 * raid4 - trivial - just use a raid4 layout.
5992 * raid6 - Providing it is a *_6 layout
5994 if (mddev->level == 0)
5995 return raid45_takeover_raid0(mddev, 5);
5996 if (mddev->level == 1)
5997 return raid5_takeover_raid1(mddev);
5998 if (mddev->level == 4) {
5999 mddev->new_layout = ALGORITHM_PARITY_N;
6000 mddev->new_level = 5;
6001 return setup_conf(mddev);
6003 if (mddev->level == 6)
6004 return raid5_takeover_raid6(mddev);
6006 return ERR_PTR(-EINVAL);
6009 static void *raid4_takeover(struct mddev *mddev)
6011 /* raid4 can take over:
6012 * raid0 - if there is only one strip zone
6013 * raid5 - if layout is right
6015 if (mddev->level == 0)
6016 return raid45_takeover_raid0(mddev, 4);
6017 if (mddev->level == 5 &&
6018 mddev->layout == ALGORITHM_PARITY_N) {
6019 mddev->new_layout = 0;
6020 mddev->new_level = 4;
6021 return setup_conf(mddev);
6023 return ERR_PTR(-EINVAL);
6026 static struct md_personality raid5_personality;
6028 static void *raid6_takeover(struct mddev *mddev)
6030 /* Currently can only take over a raid5. We map the
6031 * personality to an equivalent raid6 personality
6032 * with the Q block at the end.
6036 if (mddev->pers != &raid5_personality)
6037 return ERR_PTR(-EINVAL);
6038 if (mddev->degraded > 1)
6039 return ERR_PTR(-EINVAL);
6040 if (mddev->raid_disks > 253)
6041 return ERR_PTR(-EINVAL);
6042 if (mddev->raid_disks < 3)
6043 return ERR_PTR(-EINVAL);
6045 switch (mddev->layout) {
6046 case ALGORITHM_LEFT_ASYMMETRIC:
6047 new_layout = ALGORITHM_LEFT_ASYMMETRIC_6;
6049 case ALGORITHM_RIGHT_ASYMMETRIC:
6050 new_layout = ALGORITHM_RIGHT_ASYMMETRIC_6;
6052 case ALGORITHM_LEFT_SYMMETRIC:
6053 new_layout = ALGORITHM_LEFT_SYMMETRIC_6;
6055 case ALGORITHM_RIGHT_SYMMETRIC:
6056 new_layout = ALGORITHM_RIGHT_SYMMETRIC_6;
6058 case ALGORITHM_PARITY_0:
6059 new_layout = ALGORITHM_PARITY_0_6;
6061 case ALGORITHM_PARITY_N:
6062 new_layout = ALGORITHM_PARITY_N;
6065 return ERR_PTR(-EINVAL);
6067 mddev->new_level = 6;
6068 mddev->new_layout = new_layout;
6069 mddev->delta_disks = 1;
6070 mddev->raid_disks += 1;
6071 return setup_conf(mddev);
6075 static struct md_personality raid6_personality =
6079 .owner = THIS_MODULE,
6080 .make_request = make_request,
6084 .error_handler = error,
6085 .hot_add_disk = raid5_add_disk,
6086 .hot_remove_disk= raid5_remove_disk,
6087 .spare_active = raid5_spare_active,
6088 .sync_request = sync_request,
6089 .resize = raid5_resize,
6091 .check_reshape = raid6_check_reshape,
6092 .start_reshape = raid5_start_reshape,
6093 .finish_reshape = raid5_finish_reshape,
6094 .quiesce = raid5_quiesce,
6095 .takeover = raid6_takeover,
6097 static struct md_personality raid5_personality =
6101 .owner = THIS_MODULE,
6102 .make_request = make_request,
6106 .error_handler = error,
6107 .hot_add_disk = raid5_add_disk,
6108 .hot_remove_disk= raid5_remove_disk,
6109 .spare_active = raid5_spare_active,
6110 .sync_request = sync_request,
6111 .resize = raid5_resize,
6113 .check_reshape = raid5_check_reshape,
6114 .start_reshape = raid5_start_reshape,
6115 .finish_reshape = raid5_finish_reshape,
6116 .quiesce = raid5_quiesce,
6117 .takeover = raid5_takeover,
6120 static struct md_personality raid4_personality =
6124 .owner = THIS_MODULE,
6125 .make_request = make_request,
6129 .error_handler = error,
6130 .hot_add_disk = raid5_add_disk,
6131 .hot_remove_disk= raid5_remove_disk,
6132 .spare_active = raid5_spare_active,
6133 .sync_request = sync_request,
6134 .resize = raid5_resize,
6136 .check_reshape = raid5_check_reshape,
6137 .start_reshape = raid5_start_reshape,
6138 .finish_reshape = raid5_finish_reshape,
6139 .quiesce = raid5_quiesce,
6140 .takeover = raid4_takeover,
6143 static int __init raid5_init(void)
6145 register_md_personality(&raid6_personality);
6146 register_md_personality(&raid5_personality);
6147 register_md_personality(&raid4_personality);
6151 static void raid5_exit(void)
6153 unregister_md_personality(&raid6_personality);
6154 unregister_md_personality(&raid5_personality);
6155 unregister_md_personality(&raid4_personality);
6158 module_init(raid5_init);
6159 module_exit(raid5_exit);
6160 MODULE_LICENSE("GPL");
6161 MODULE_DESCRIPTION("RAID4/5/6 (striping with parity) personality for MD");
6162 MODULE_ALIAS("md-personality-4"); /* RAID5 */
6163 MODULE_ALIAS("md-raid5");
6164 MODULE_ALIAS("md-raid4");
6165 MODULE_ALIAS("md-level-5");
6166 MODULE_ALIAS("md-level-4");
6167 MODULE_ALIAS("md-personality-8"); /* RAID6 */
6168 MODULE_ALIAS("md-raid6");
6169 MODULE_ALIAS("md-level-6");
6171 /* This used to be two separate modules, they were: */
6172 MODULE_ALIAS("raid5");
6173 MODULE_ALIAS("raid6");