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/async.h>
51 #include <linux/seq_file.h>
52 #include <linux/cpu.h>
53 #include <linux/slab.h>
54 #include <linux/ratelimit.h>
64 #define NR_STRIPES 256
65 #define STRIPE_SIZE PAGE_SIZE
66 #define STRIPE_SHIFT (PAGE_SHIFT - 9)
67 #define STRIPE_SECTORS (STRIPE_SIZE>>9)
68 #define IO_THRESHOLD 1
69 #define BYPASS_THRESHOLD 1
70 #define NR_HASH (PAGE_SIZE / sizeof(struct hlist_head))
71 #define HASH_MASK (NR_HASH - 1)
73 #define stripe_hash(conf, sect) (&((conf)->stripe_hashtbl[((sect) >> STRIPE_SHIFT) & HASH_MASK]))
75 /* bio's attached to a stripe+device for I/O are linked together in bi_sector
76 * order without overlap. There may be several bio's per stripe+device, and
77 * a bio could span several devices.
78 * When walking this list for a particular stripe+device, we must never proceed
79 * beyond a bio that extends past this device, as the next bio might no longer
81 * This macro is used to determine the 'next' bio in the list, given the sector
82 * of the current stripe+device
84 #define r5_next_bio(bio, sect) ( ( (bio)->bi_sector + ((bio)->bi_size>>9) < sect + STRIPE_SECTORS) ? (bio)->bi_next : NULL)
86 * The following can be used to debug the driver
88 #define RAID5_PARANOIA 1
89 #if RAID5_PARANOIA && defined(CONFIG_SMP)
90 # define CHECK_DEVLOCK() assert_spin_locked(&conf->device_lock)
92 # define CHECK_DEVLOCK()
101 * We maintain a biased count of active stripes in the bottom 16 bits of
102 * bi_phys_segments, and a count of processed stripes in the upper 16 bits
104 static inline int raid5_bi_phys_segments(struct bio *bio)
106 return bio->bi_phys_segments & 0xffff;
109 static inline int raid5_bi_hw_segments(struct bio *bio)
111 return (bio->bi_phys_segments >> 16) & 0xffff;
114 static inline int raid5_dec_bi_phys_segments(struct bio *bio)
116 --bio->bi_phys_segments;
117 return raid5_bi_phys_segments(bio);
120 static inline int raid5_dec_bi_hw_segments(struct bio *bio)
122 unsigned short val = raid5_bi_hw_segments(bio);
125 bio->bi_phys_segments = (val << 16) | raid5_bi_phys_segments(bio);
129 static inline void raid5_set_bi_hw_segments(struct bio *bio, unsigned int cnt)
131 bio->bi_phys_segments = raid5_bi_phys_segments(bio) | (cnt << 16);
134 /* Find first data disk in a raid6 stripe */
135 static inline int raid6_d0(struct stripe_head *sh)
138 /* ddf always start from first device */
140 /* md starts just after Q block */
141 if (sh->qd_idx == sh->disks - 1)
144 return sh->qd_idx + 1;
146 static inline int raid6_next_disk(int disk, int raid_disks)
149 return (disk < raid_disks) ? disk : 0;
152 /* When walking through the disks in a raid5, starting at raid6_d0,
153 * We need to map each disk to a 'slot', where the data disks are slot
154 * 0 .. raid_disks-3, the parity disk is raid_disks-2 and the Q disk
155 * is raid_disks-1. This help does that mapping.
157 static int raid6_idx_to_slot(int idx, struct stripe_head *sh,
158 int *count, int syndrome_disks)
164 if (idx == sh->pd_idx)
165 return syndrome_disks;
166 if (idx == sh->qd_idx)
167 return syndrome_disks + 1;
173 static void return_io(struct bio *return_bi)
175 struct bio *bi = return_bi;
178 return_bi = bi->bi_next;
186 static void print_raid5_conf (raid5_conf_t *conf);
188 static int stripe_operations_active(struct stripe_head *sh)
190 return sh->check_state || sh->reconstruct_state ||
191 test_bit(STRIPE_BIOFILL_RUN, &sh->state) ||
192 test_bit(STRIPE_COMPUTE_RUN, &sh->state);
195 static void __release_stripe(raid5_conf_t *conf, struct stripe_head *sh)
197 if (atomic_dec_and_test(&sh->count)) {
198 BUG_ON(!list_empty(&sh->lru));
199 BUG_ON(atomic_read(&conf->active_stripes)==0);
200 if (test_bit(STRIPE_HANDLE, &sh->state)) {
201 if (test_bit(STRIPE_DELAYED, &sh->state))
202 list_add_tail(&sh->lru, &conf->delayed_list);
203 else if (test_bit(STRIPE_BIT_DELAY, &sh->state) &&
204 sh->bm_seq - conf->seq_write > 0)
205 list_add_tail(&sh->lru, &conf->bitmap_list);
207 clear_bit(STRIPE_BIT_DELAY, &sh->state);
208 list_add_tail(&sh->lru, &conf->handle_list);
210 md_wakeup_thread(conf->mddev->thread);
212 BUG_ON(stripe_operations_active(sh));
213 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
214 atomic_dec(&conf->preread_active_stripes);
215 if (atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD)
216 md_wakeup_thread(conf->mddev->thread);
218 atomic_dec(&conf->active_stripes);
219 if (!test_bit(STRIPE_EXPANDING, &sh->state)) {
220 list_add_tail(&sh->lru, &conf->inactive_list);
221 wake_up(&conf->wait_for_stripe);
222 if (conf->retry_read_aligned)
223 md_wakeup_thread(conf->mddev->thread);
229 static void release_stripe(struct stripe_head *sh)
231 raid5_conf_t *conf = sh->raid_conf;
234 spin_lock_irqsave(&conf->device_lock, flags);
235 __release_stripe(conf, sh);
236 spin_unlock_irqrestore(&conf->device_lock, flags);
239 static inline void remove_hash(struct stripe_head *sh)
241 pr_debug("remove_hash(), stripe %llu\n",
242 (unsigned long long)sh->sector);
244 hlist_del_init(&sh->hash);
247 static inline void insert_hash(raid5_conf_t *conf, struct stripe_head *sh)
249 struct hlist_head *hp = stripe_hash(conf, sh->sector);
251 pr_debug("insert_hash(), stripe %llu\n",
252 (unsigned long long)sh->sector);
255 hlist_add_head(&sh->hash, hp);
259 /* find an idle stripe, make sure it is unhashed, and return it. */
260 static struct stripe_head *get_free_stripe(raid5_conf_t *conf)
262 struct stripe_head *sh = NULL;
263 struct list_head *first;
266 if (list_empty(&conf->inactive_list))
268 first = conf->inactive_list.next;
269 sh = list_entry(first, struct stripe_head, lru);
270 list_del_init(first);
272 atomic_inc(&conf->active_stripes);
277 static void shrink_buffers(struct stripe_head *sh)
281 int num = sh->raid_conf->pool_size;
283 for (i = 0; i < num ; i++) {
287 sh->dev[i].page = NULL;
292 static int grow_buffers(struct stripe_head *sh)
295 int num = sh->raid_conf->pool_size;
297 for (i = 0; i < num; i++) {
300 if (!(page = alloc_page(GFP_KERNEL))) {
303 sh->dev[i].page = page;
308 static void raid5_build_block(struct stripe_head *sh, int i, int previous);
309 static void stripe_set_idx(sector_t stripe, raid5_conf_t *conf, int previous,
310 struct stripe_head *sh);
312 static void init_stripe(struct stripe_head *sh, sector_t sector, int previous)
314 raid5_conf_t *conf = sh->raid_conf;
317 BUG_ON(atomic_read(&sh->count) != 0);
318 BUG_ON(test_bit(STRIPE_HANDLE, &sh->state));
319 BUG_ON(stripe_operations_active(sh));
322 pr_debug("init_stripe called, stripe %llu\n",
323 (unsigned long long)sh->sector);
327 sh->generation = conf->generation - previous;
328 sh->disks = previous ? conf->previous_raid_disks : conf->raid_disks;
330 stripe_set_idx(sector, conf, previous, sh);
334 for (i = sh->disks; i--; ) {
335 struct r5dev *dev = &sh->dev[i];
337 if (dev->toread || dev->read || dev->towrite || dev->written ||
338 test_bit(R5_LOCKED, &dev->flags)) {
339 printk(KERN_ERR "sector=%llx i=%d %p %p %p %p %d\n",
340 (unsigned long long)sh->sector, i, dev->toread,
341 dev->read, dev->towrite, dev->written,
342 test_bit(R5_LOCKED, &dev->flags));
346 raid5_build_block(sh, i, previous);
348 insert_hash(conf, sh);
351 static struct stripe_head *__find_stripe(raid5_conf_t *conf, sector_t sector,
354 struct stripe_head *sh;
355 struct hlist_node *hn;
358 pr_debug("__find_stripe, sector %llu\n", (unsigned long long)sector);
359 hlist_for_each_entry(sh, hn, stripe_hash(conf, sector), hash)
360 if (sh->sector == sector && sh->generation == generation)
362 pr_debug("__stripe %llu not in cache\n", (unsigned long long)sector);
367 * Need to check if array has failed when deciding whether to:
369 * - remove non-faulty devices
372 * This determination is simple when no reshape is happening.
373 * However if there is a reshape, we need to carefully check
374 * both the before and after sections.
375 * This is because some failed devices may only affect one
376 * of the two sections, and some non-in_sync devices may
377 * be insync in the section most affected by failed devices.
379 static int has_failed(raid5_conf_t *conf)
383 if (conf->mddev->reshape_position == MaxSector)
384 return conf->mddev->degraded > conf->max_degraded;
388 for (i = 0; i < conf->previous_raid_disks; i++) {
389 mdk_rdev_t *rdev = rcu_dereference(conf->disks[i].rdev);
390 if (!rdev || test_bit(Faulty, &rdev->flags))
392 else if (test_bit(In_sync, &rdev->flags))
395 /* not in-sync or faulty.
396 * If the reshape increases the number of devices,
397 * this is being recovered by the reshape, so
398 * this 'previous' section is not in_sync.
399 * If the number of devices is being reduced however,
400 * the device can only be part of the array if
401 * we are reverting a reshape, so this section will
404 if (conf->raid_disks >= conf->previous_raid_disks)
408 if (degraded > conf->max_degraded)
412 for (i = 0; i < conf->raid_disks; i++) {
413 mdk_rdev_t *rdev = rcu_dereference(conf->disks[i].rdev);
414 if (!rdev || test_bit(Faulty, &rdev->flags))
416 else if (test_bit(In_sync, &rdev->flags))
419 /* not in-sync or faulty.
420 * If reshape increases the number of devices, this
421 * section has already been recovered, else it
422 * almost certainly hasn't.
424 if (conf->raid_disks <= conf->previous_raid_disks)
428 if (degraded > conf->max_degraded)
433 static struct stripe_head *
434 get_active_stripe(raid5_conf_t *conf, sector_t sector,
435 int previous, int noblock, int noquiesce)
437 struct stripe_head *sh;
439 pr_debug("get_stripe, sector %llu\n", (unsigned long long)sector);
441 spin_lock_irq(&conf->device_lock);
444 wait_event_lock_irq(conf->wait_for_stripe,
445 conf->quiesce == 0 || noquiesce,
446 conf->device_lock, /* nothing */);
447 sh = __find_stripe(conf, sector, conf->generation - previous);
449 if (!conf->inactive_blocked)
450 sh = get_free_stripe(conf);
451 if (noblock && sh == NULL)
454 conf->inactive_blocked = 1;
455 wait_event_lock_irq(conf->wait_for_stripe,
456 !list_empty(&conf->inactive_list) &&
457 (atomic_read(&conf->active_stripes)
458 < (conf->max_nr_stripes *3/4)
459 || !conf->inactive_blocked),
462 conf->inactive_blocked = 0;
464 init_stripe(sh, sector, previous);
466 if (atomic_read(&sh->count)) {
467 BUG_ON(!list_empty(&sh->lru)
468 && !test_bit(STRIPE_EXPANDING, &sh->state));
470 if (!test_bit(STRIPE_HANDLE, &sh->state))
471 atomic_inc(&conf->active_stripes);
472 if (list_empty(&sh->lru) &&
473 !test_bit(STRIPE_EXPANDING, &sh->state))
475 list_del_init(&sh->lru);
478 } while (sh == NULL);
481 atomic_inc(&sh->count);
483 spin_unlock_irq(&conf->device_lock);
488 raid5_end_read_request(struct bio *bi, int error);
490 raid5_end_write_request(struct bio *bi, int error);
492 static void ops_run_io(struct stripe_head *sh, struct stripe_head_state *s)
494 raid5_conf_t *conf = sh->raid_conf;
495 int i, disks = sh->disks;
499 for (i = disks; i--; ) {
503 if (test_and_clear_bit(R5_Wantwrite, &sh->dev[i].flags)) {
504 if (test_and_clear_bit(R5_WantFUA, &sh->dev[i].flags))
508 } else if (test_and_clear_bit(R5_Wantread, &sh->dev[i].flags))
513 bi = &sh->dev[i].req;
517 bi->bi_end_io = raid5_end_write_request;
519 bi->bi_end_io = raid5_end_read_request;
522 rdev = rcu_dereference(conf->disks[i].rdev);
523 if (rdev && test_bit(Faulty, &rdev->flags))
526 atomic_inc(&rdev->nr_pending);
529 /* We have already checked bad blocks for reads. Now
530 * need to check for writes.
532 while ((rw & WRITE) && rdev &&
533 test_bit(WriteErrorSeen, &rdev->flags)) {
536 int bad = is_badblock(rdev, sh->sector, STRIPE_SECTORS,
537 &first_bad, &bad_sectors);
542 set_bit(BlockedBadBlocks, &rdev->flags);
543 if (!conf->mddev->external &&
544 conf->mddev->flags) {
545 /* It is very unlikely, but we might
546 * still need to write out the
547 * bad block log - better give it
549 md_check_recovery(conf->mddev);
551 md_wait_for_blocked_rdev(rdev, conf->mddev);
553 /* Acknowledged bad block - skip the write */
554 rdev_dec_pending(rdev, conf->mddev);
560 if (s->syncing || s->expanding || s->expanded)
561 md_sync_acct(rdev->bdev, STRIPE_SECTORS);
563 set_bit(STRIPE_IO_STARTED, &sh->state);
565 bi->bi_bdev = rdev->bdev;
566 pr_debug("%s: for %llu schedule op %ld on disc %d\n",
567 __func__, (unsigned long long)sh->sector,
569 atomic_inc(&sh->count);
570 bi->bi_sector = sh->sector + rdev->data_offset;
571 bi->bi_flags = 1 << BIO_UPTODATE;
575 bi->bi_io_vec = &sh->dev[i].vec;
576 bi->bi_io_vec[0].bv_len = STRIPE_SIZE;
577 bi->bi_io_vec[0].bv_offset = 0;
578 bi->bi_size = STRIPE_SIZE;
580 generic_make_request(bi);
583 set_bit(STRIPE_DEGRADED, &sh->state);
584 pr_debug("skip op %ld on disc %d for sector %llu\n",
585 bi->bi_rw, i, (unsigned long long)sh->sector);
586 clear_bit(R5_LOCKED, &sh->dev[i].flags);
587 set_bit(STRIPE_HANDLE, &sh->state);
592 static struct dma_async_tx_descriptor *
593 async_copy_data(int frombio, struct bio *bio, struct page *page,
594 sector_t sector, struct dma_async_tx_descriptor *tx)
597 struct page *bio_page;
600 struct async_submit_ctl submit;
601 enum async_tx_flags flags = 0;
603 if (bio->bi_sector >= sector)
604 page_offset = (signed)(bio->bi_sector - sector) * 512;
606 page_offset = (signed)(sector - bio->bi_sector) * -512;
609 flags |= ASYNC_TX_FENCE;
610 init_async_submit(&submit, flags, tx, NULL, NULL, NULL);
612 bio_for_each_segment(bvl, bio, i) {
613 int len = bvl->bv_len;
617 if (page_offset < 0) {
618 b_offset = -page_offset;
619 page_offset += b_offset;
623 if (len > 0 && page_offset + len > STRIPE_SIZE)
624 clen = STRIPE_SIZE - page_offset;
629 b_offset += bvl->bv_offset;
630 bio_page = bvl->bv_page;
632 tx = async_memcpy(page, bio_page, page_offset,
633 b_offset, clen, &submit);
635 tx = async_memcpy(bio_page, page, b_offset,
636 page_offset, clen, &submit);
638 /* chain the operations */
639 submit.depend_tx = tx;
641 if (clen < len) /* hit end of page */
649 static void ops_complete_biofill(void *stripe_head_ref)
651 struct stripe_head *sh = stripe_head_ref;
652 struct bio *return_bi = NULL;
653 raid5_conf_t *conf = sh->raid_conf;
656 pr_debug("%s: stripe %llu\n", __func__,
657 (unsigned long long)sh->sector);
659 /* clear completed biofills */
660 spin_lock_irq(&conf->device_lock);
661 for (i = sh->disks; i--; ) {
662 struct r5dev *dev = &sh->dev[i];
664 /* acknowledge completion of a biofill operation */
665 /* and check if we need to reply to a read request,
666 * new R5_Wantfill requests are held off until
667 * !STRIPE_BIOFILL_RUN
669 if (test_and_clear_bit(R5_Wantfill, &dev->flags)) {
670 struct bio *rbi, *rbi2;
675 while (rbi && rbi->bi_sector <
676 dev->sector + STRIPE_SECTORS) {
677 rbi2 = r5_next_bio(rbi, dev->sector);
678 if (!raid5_dec_bi_phys_segments(rbi)) {
679 rbi->bi_next = return_bi;
686 spin_unlock_irq(&conf->device_lock);
687 clear_bit(STRIPE_BIOFILL_RUN, &sh->state);
689 return_io(return_bi);
691 set_bit(STRIPE_HANDLE, &sh->state);
695 static void ops_run_biofill(struct stripe_head *sh)
697 struct dma_async_tx_descriptor *tx = NULL;
698 raid5_conf_t *conf = sh->raid_conf;
699 struct async_submit_ctl submit;
702 pr_debug("%s: stripe %llu\n", __func__,
703 (unsigned long long)sh->sector);
705 for (i = sh->disks; i--; ) {
706 struct r5dev *dev = &sh->dev[i];
707 if (test_bit(R5_Wantfill, &dev->flags)) {
709 spin_lock_irq(&conf->device_lock);
710 dev->read = rbi = dev->toread;
712 spin_unlock_irq(&conf->device_lock);
713 while (rbi && rbi->bi_sector <
714 dev->sector + STRIPE_SECTORS) {
715 tx = async_copy_data(0, rbi, dev->page,
717 rbi = r5_next_bio(rbi, dev->sector);
722 atomic_inc(&sh->count);
723 init_async_submit(&submit, ASYNC_TX_ACK, tx, ops_complete_biofill, sh, NULL);
724 async_trigger_callback(&submit);
727 static void mark_target_uptodate(struct stripe_head *sh, int target)
734 tgt = &sh->dev[target];
735 set_bit(R5_UPTODATE, &tgt->flags);
736 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
737 clear_bit(R5_Wantcompute, &tgt->flags);
740 static void ops_complete_compute(void *stripe_head_ref)
742 struct stripe_head *sh = stripe_head_ref;
744 pr_debug("%s: stripe %llu\n", __func__,
745 (unsigned long long)sh->sector);
747 /* mark the computed target(s) as uptodate */
748 mark_target_uptodate(sh, sh->ops.target);
749 mark_target_uptodate(sh, sh->ops.target2);
751 clear_bit(STRIPE_COMPUTE_RUN, &sh->state);
752 if (sh->check_state == check_state_compute_run)
753 sh->check_state = check_state_compute_result;
754 set_bit(STRIPE_HANDLE, &sh->state);
758 /* return a pointer to the address conversion region of the scribble buffer */
759 static addr_conv_t *to_addr_conv(struct stripe_head *sh,
760 struct raid5_percpu *percpu)
762 return percpu->scribble + sizeof(struct page *) * (sh->disks + 2);
765 static struct dma_async_tx_descriptor *
766 ops_run_compute5(struct stripe_head *sh, struct raid5_percpu *percpu)
768 int disks = sh->disks;
769 struct page **xor_srcs = percpu->scribble;
770 int target = sh->ops.target;
771 struct r5dev *tgt = &sh->dev[target];
772 struct page *xor_dest = tgt->page;
774 struct dma_async_tx_descriptor *tx;
775 struct async_submit_ctl submit;
778 pr_debug("%s: stripe %llu block: %d\n",
779 __func__, (unsigned long long)sh->sector, target);
780 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
782 for (i = disks; i--; )
784 xor_srcs[count++] = sh->dev[i].page;
786 atomic_inc(&sh->count);
788 init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST, NULL,
789 ops_complete_compute, sh, to_addr_conv(sh, percpu));
790 if (unlikely(count == 1))
791 tx = async_memcpy(xor_dest, xor_srcs[0], 0, 0, STRIPE_SIZE, &submit);
793 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
798 /* set_syndrome_sources - populate source buffers for gen_syndrome
799 * @srcs - (struct page *) array of size sh->disks
800 * @sh - stripe_head to parse
802 * Populates srcs in proper layout order for the stripe and returns the
803 * 'count' of sources to be used in a call to async_gen_syndrome. The P
804 * destination buffer is recorded in srcs[count] and the Q destination
805 * is recorded in srcs[count+1]].
807 static int set_syndrome_sources(struct page **srcs, struct stripe_head *sh)
809 int disks = sh->disks;
810 int syndrome_disks = sh->ddf_layout ? disks : (disks - 2);
811 int d0_idx = raid6_d0(sh);
815 for (i = 0; i < disks; i++)
821 int slot = raid6_idx_to_slot(i, sh, &count, syndrome_disks);
823 srcs[slot] = sh->dev[i].page;
824 i = raid6_next_disk(i, disks);
825 } while (i != d0_idx);
827 return syndrome_disks;
830 static struct dma_async_tx_descriptor *
831 ops_run_compute6_1(struct stripe_head *sh, struct raid5_percpu *percpu)
833 int disks = sh->disks;
834 struct page **blocks = percpu->scribble;
836 int qd_idx = sh->qd_idx;
837 struct dma_async_tx_descriptor *tx;
838 struct async_submit_ctl submit;
844 if (sh->ops.target < 0)
845 target = sh->ops.target2;
846 else if (sh->ops.target2 < 0)
847 target = sh->ops.target;
849 /* we should only have one valid target */
852 pr_debug("%s: stripe %llu block: %d\n",
853 __func__, (unsigned long long)sh->sector, target);
855 tgt = &sh->dev[target];
856 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
859 atomic_inc(&sh->count);
861 if (target == qd_idx) {
862 count = set_syndrome_sources(blocks, sh);
863 blocks[count] = NULL; /* regenerating p is not necessary */
864 BUG_ON(blocks[count+1] != dest); /* q should already be set */
865 init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
866 ops_complete_compute, sh,
867 to_addr_conv(sh, percpu));
868 tx = async_gen_syndrome(blocks, 0, count+2, STRIPE_SIZE, &submit);
870 /* Compute any data- or p-drive using XOR */
872 for (i = disks; i-- ; ) {
873 if (i == target || i == qd_idx)
875 blocks[count++] = sh->dev[i].page;
878 init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST,
879 NULL, ops_complete_compute, sh,
880 to_addr_conv(sh, percpu));
881 tx = async_xor(dest, blocks, 0, count, STRIPE_SIZE, &submit);
887 static struct dma_async_tx_descriptor *
888 ops_run_compute6_2(struct stripe_head *sh, struct raid5_percpu *percpu)
890 int i, count, disks = sh->disks;
891 int syndrome_disks = sh->ddf_layout ? disks : disks-2;
892 int d0_idx = raid6_d0(sh);
893 int faila = -1, failb = -1;
894 int target = sh->ops.target;
895 int target2 = sh->ops.target2;
896 struct r5dev *tgt = &sh->dev[target];
897 struct r5dev *tgt2 = &sh->dev[target2];
898 struct dma_async_tx_descriptor *tx;
899 struct page **blocks = percpu->scribble;
900 struct async_submit_ctl submit;
902 pr_debug("%s: stripe %llu block1: %d block2: %d\n",
903 __func__, (unsigned long long)sh->sector, target, target2);
904 BUG_ON(target < 0 || target2 < 0);
905 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
906 BUG_ON(!test_bit(R5_Wantcompute, &tgt2->flags));
908 /* we need to open-code set_syndrome_sources to handle the
909 * slot number conversion for 'faila' and 'failb'
911 for (i = 0; i < disks ; i++)
916 int slot = raid6_idx_to_slot(i, sh, &count, syndrome_disks);
918 blocks[slot] = sh->dev[i].page;
924 i = raid6_next_disk(i, disks);
925 } while (i != d0_idx);
927 BUG_ON(faila == failb);
930 pr_debug("%s: stripe: %llu faila: %d failb: %d\n",
931 __func__, (unsigned long long)sh->sector, faila, failb);
933 atomic_inc(&sh->count);
935 if (failb == syndrome_disks+1) {
936 /* Q disk is one of the missing disks */
937 if (faila == syndrome_disks) {
938 /* Missing P+Q, just recompute */
939 init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
940 ops_complete_compute, sh,
941 to_addr_conv(sh, percpu));
942 return async_gen_syndrome(blocks, 0, syndrome_disks+2,
943 STRIPE_SIZE, &submit);
947 int qd_idx = sh->qd_idx;
949 /* Missing D+Q: recompute D from P, then recompute Q */
950 if (target == qd_idx)
951 data_target = target2;
953 data_target = target;
956 for (i = disks; i-- ; ) {
957 if (i == data_target || i == qd_idx)
959 blocks[count++] = sh->dev[i].page;
961 dest = sh->dev[data_target].page;
962 init_async_submit(&submit,
963 ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST,
965 to_addr_conv(sh, percpu));
966 tx = async_xor(dest, blocks, 0, count, STRIPE_SIZE,
969 count = set_syndrome_sources(blocks, sh);
970 init_async_submit(&submit, ASYNC_TX_FENCE, tx,
971 ops_complete_compute, sh,
972 to_addr_conv(sh, percpu));
973 return async_gen_syndrome(blocks, 0, count+2,
974 STRIPE_SIZE, &submit);
977 init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
978 ops_complete_compute, sh,
979 to_addr_conv(sh, percpu));
980 if (failb == syndrome_disks) {
981 /* We're missing D+P. */
982 return async_raid6_datap_recov(syndrome_disks+2,
986 /* We're missing D+D. */
987 return async_raid6_2data_recov(syndrome_disks+2,
988 STRIPE_SIZE, faila, failb,
995 static void ops_complete_prexor(void *stripe_head_ref)
997 struct stripe_head *sh = stripe_head_ref;
999 pr_debug("%s: stripe %llu\n", __func__,
1000 (unsigned long long)sh->sector);
1003 static struct dma_async_tx_descriptor *
1004 ops_run_prexor(struct stripe_head *sh, struct raid5_percpu *percpu,
1005 struct dma_async_tx_descriptor *tx)
1007 int disks = sh->disks;
1008 struct page **xor_srcs = percpu->scribble;
1009 int count = 0, pd_idx = sh->pd_idx, i;
1010 struct async_submit_ctl submit;
1012 /* existing parity data subtracted */
1013 struct page *xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
1015 pr_debug("%s: stripe %llu\n", __func__,
1016 (unsigned long long)sh->sector);
1018 for (i = disks; i--; ) {
1019 struct r5dev *dev = &sh->dev[i];
1020 /* Only process blocks that are known to be uptodate */
1021 if (test_bit(R5_Wantdrain, &dev->flags))
1022 xor_srcs[count++] = dev->page;
1025 init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_DROP_DST, tx,
1026 ops_complete_prexor, sh, to_addr_conv(sh, percpu));
1027 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
1032 static struct dma_async_tx_descriptor *
1033 ops_run_biodrain(struct stripe_head *sh, struct dma_async_tx_descriptor *tx)
1035 int disks = sh->disks;
1038 pr_debug("%s: stripe %llu\n", __func__,
1039 (unsigned long long)sh->sector);
1041 for (i = disks; i--; ) {
1042 struct r5dev *dev = &sh->dev[i];
1045 if (test_and_clear_bit(R5_Wantdrain, &dev->flags)) {
1048 spin_lock_irq(&sh->raid_conf->device_lock);
1049 chosen = dev->towrite;
1050 dev->towrite = NULL;
1051 BUG_ON(dev->written);
1052 wbi = dev->written = chosen;
1053 spin_unlock_irq(&sh->raid_conf->device_lock);
1055 while (wbi && wbi->bi_sector <
1056 dev->sector + STRIPE_SECTORS) {
1057 if (wbi->bi_rw & REQ_FUA)
1058 set_bit(R5_WantFUA, &dev->flags);
1059 tx = async_copy_data(1, wbi, dev->page,
1061 wbi = r5_next_bio(wbi, dev->sector);
1069 static void ops_complete_reconstruct(void *stripe_head_ref)
1071 struct stripe_head *sh = stripe_head_ref;
1072 int disks = sh->disks;
1073 int pd_idx = sh->pd_idx;
1074 int qd_idx = sh->qd_idx;
1078 pr_debug("%s: stripe %llu\n", __func__,
1079 (unsigned long long)sh->sector);
1081 for (i = disks; i--; )
1082 fua |= test_bit(R5_WantFUA, &sh->dev[i].flags);
1084 for (i = disks; i--; ) {
1085 struct r5dev *dev = &sh->dev[i];
1087 if (dev->written || i == pd_idx || i == qd_idx) {
1088 set_bit(R5_UPTODATE, &dev->flags);
1090 set_bit(R5_WantFUA, &dev->flags);
1094 if (sh->reconstruct_state == reconstruct_state_drain_run)
1095 sh->reconstruct_state = reconstruct_state_drain_result;
1096 else if (sh->reconstruct_state == reconstruct_state_prexor_drain_run)
1097 sh->reconstruct_state = reconstruct_state_prexor_drain_result;
1099 BUG_ON(sh->reconstruct_state != reconstruct_state_run);
1100 sh->reconstruct_state = reconstruct_state_result;
1103 set_bit(STRIPE_HANDLE, &sh->state);
1108 ops_run_reconstruct5(struct stripe_head *sh, struct raid5_percpu *percpu,
1109 struct dma_async_tx_descriptor *tx)
1111 int disks = sh->disks;
1112 struct page **xor_srcs = percpu->scribble;
1113 struct async_submit_ctl submit;
1114 int count = 0, pd_idx = sh->pd_idx, i;
1115 struct page *xor_dest;
1117 unsigned long flags;
1119 pr_debug("%s: stripe %llu\n", __func__,
1120 (unsigned long long)sh->sector);
1122 /* check if prexor is active which means only process blocks
1123 * that are part of a read-modify-write (written)
1125 if (sh->reconstruct_state == reconstruct_state_prexor_drain_run) {
1127 xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
1128 for (i = disks; i--; ) {
1129 struct r5dev *dev = &sh->dev[i];
1131 xor_srcs[count++] = dev->page;
1134 xor_dest = sh->dev[pd_idx].page;
1135 for (i = disks; i--; ) {
1136 struct r5dev *dev = &sh->dev[i];
1138 xor_srcs[count++] = dev->page;
1142 /* 1/ if we prexor'd then the dest is reused as a source
1143 * 2/ if we did not prexor then we are redoing the parity
1144 * set ASYNC_TX_XOR_DROP_DST and ASYNC_TX_XOR_ZERO_DST
1145 * for the synchronous xor case
1147 flags = ASYNC_TX_ACK |
1148 (prexor ? ASYNC_TX_XOR_DROP_DST : ASYNC_TX_XOR_ZERO_DST);
1150 atomic_inc(&sh->count);
1152 init_async_submit(&submit, flags, tx, ops_complete_reconstruct, sh,
1153 to_addr_conv(sh, percpu));
1154 if (unlikely(count == 1))
1155 tx = async_memcpy(xor_dest, xor_srcs[0], 0, 0, STRIPE_SIZE, &submit);
1157 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
1161 ops_run_reconstruct6(struct stripe_head *sh, struct raid5_percpu *percpu,
1162 struct dma_async_tx_descriptor *tx)
1164 struct async_submit_ctl submit;
1165 struct page **blocks = percpu->scribble;
1168 pr_debug("%s: stripe %llu\n", __func__, (unsigned long long)sh->sector);
1170 count = set_syndrome_sources(blocks, sh);
1172 atomic_inc(&sh->count);
1174 init_async_submit(&submit, ASYNC_TX_ACK, tx, ops_complete_reconstruct,
1175 sh, to_addr_conv(sh, percpu));
1176 async_gen_syndrome(blocks, 0, count+2, STRIPE_SIZE, &submit);
1179 static void ops_complete_check(void *stripe_head_ref)
1181 struct stripe_head *sh = stripe_head_ref;
1183 pr_debug("%s: stripe %llu\n", __func__,
1184 (unsigned long long)sh->sector);
1186 sh->check_state = check_state_check_result;
1187 set_bit(STRIPE_HANDLE, &sh->state);
1191 static void ops_run_check_p(struct stripe_head *sh, struct raid5_percpu *percpu)
1193 int disks = sh->disks;
1194 int pd_idx = sh->pd_idx;
1195 int qd_idx = sh->qd_idx;
1196 struct page *xor_dest;
1197 struct page **xor_srcs = percpu->scribble;
1198 struct dma_async_tx_descriptor *tx;
1199 struct async_submit_ctl submit;
1203 pr_debug("%s: stripe %llu\n", __func__,
1204 (unsigned long long)sh->sector);
1207 xor_dest = sh->dev[pd_idx].page;
1208 xor_srcs[count++] = xor_dest;
1209 for (i = disks; i--; ) {
1210 if (i == pd_idx || i == qd_idx)
1212 xor_srcs[count++] = sh->dev[i].page;
1215 init_async_submit(&submit, 0, NULL, NULL, NULL,
1216 to_addr_conv(sh, percpu));
1217 tx = async_xor_val(xor_dest, xor_srcs, 0, count, STRIPE_SIZE,
1218 &sh->ops.zero_sum_result, &submit);
1220 atomic_inc(&sh->count);
1221 init_async_submit(&submit, ASYNC_TX_ACK, tx, ops_complete_check, sh, NULL);
1222 tx = async_trigger_callback(&submit);
1225 static void ops_run_check_pq(struct stripe_head *sh, struct raid5_percpu *percpu, int checkp)
1227 struct page **srcs = percpu->scribble;
1228 struct async_submit_ctl submit;
1231 pr_debug("%s: stripe %llu checkp: %d\n", __func__,
1232 (unsigned long long)sh->sector, checkp);
1234 count = set_syndrome_sources(srcs, sh);
1238 atomic_inc(&sh->count);
1239 init_async_submit(&submit, ASYNC_TX_ACK, NULL, ops_complete_check,
1240 sh, to_addr_conv(sh, percpu));
1241 async_syndrome_val(srcs, 0, count+2, STRIPE_SIZE,
1242 &sh->ops.zero_sum_result, percpu->spare_page, &submit);
1245 static void __raid_run_ops(struct stripe_head *sh, unsigned long ops_request)
1247 int overlap_clear = 0, i, disks = sh->disks;
1248 struct dma_async_tx_descriptor *tx = NULL;
1249 raid5_conf_t *conf = sh->raid_conf;
1250 int level = conf->level;
1251 struct raid5_percpu *percpu;
1255 percpu = per_cpu_ptr(conf->percpu, cpu);
1256 if (test_bit(STRIPE_OP_BIOFILL, &ops_request)) {
1257 ops_run_biofill(sh);
1261 if (test_bit(STRIPE_OP_COMPUTE_BLK, &ops_request)) {
1263 tx = ops_run_compute5(sh, percpu);
1265 if (sh->ops.target2 < 0 || sh->ops.target < 0)
1266 tx = ops_run_compute6_1(sh, percpu);
1268 tx = ops_run_compute6_2(sh, percpu);
1270 /* terminate the chain if reconstruct is not set to be run */
1271 if (tx && !test_bit(STRIPE_OP_RECONSTRUCT, &ops_request))
1275 if (test_bit(STRIPE_OP_PREXOR, &ops_request))
1276 tx = ops_run_prexor(sh, percpu, tx);
1278 if (test_bit(STRIPE_OP_BIODRAIN, &ops_request)) {
1279 tx = ops_run_biodrain(sh, tx);
1283 if (test_bit(STRIPE_OP_RECONSTRUCT, &ops_request)) {
1285 ops_run_reconstruct5(sh, percpu, tx);
1287 ops_run_reconstruct6(sh, percpu, tx);
1290 if (test_bit(STRIPE_OP_CHECK, &ops_request)) {
1291 if (sh->check_state == check_state_run)
1292 ops_run_check_p(sh, percpu);
1293 else if (sh->check_state == check_state_run_q)
1294 ops_run_check_pq(sh, percpu, 0);
1295 else if (sh->check_state == check_state_run_pq)
1296 ops_run_check_pq(sh, percpu, 1);
1302 for (i = disks; i--; ) {
1303 struct r5dev *dev = &sh->dev[i];
1304 if (test_and_clear_bit(R5_Overlap, &dev->flags))
1305 wake_up(&sh->raid_conf->wait_for_overlap);
1310 #ifdef CONFIG_MULTICORE_RAID456
1311 static void async_run_ops(void *param, async_cookie_t cookie)
1313 struct stripe_head *sh = param;
1314 unsigned long ops_request = sh->ops.request;
1316 clear_bit_unlock(STRIPE_OPS_REQ_PENDING, &sh->state);
1317 wake_up(&sh->ops.wait_for_ops);
1319 __raid_run_ops(sh, ops_request);
1323 static void raid_run_ops(struct stripe_head *sh, unsigned long ops_request)
1325 /* since handle_stripe can be called outside of raid5d context
1326 * we need to ensure sh->ops.request is de-staged before another
1329 wait_event(sh->ops.wait_for_ops,
1330 !test_and_set_bit_lock(STRIPE_OPS_REQ_PENDING, &sh->state));
1331 sh->ops.request = ops_request;
1333 atomic_inc(&sh->count);
1334 async_schedule(async_run_ops, sh);
1337 #define raid_run_ops __raid_run_ops
1340 static int grow_one_stripe(raid5_conf_t *conf)
1342 struct stripe_head *sh;
1343 sh = kmem_cache_zalloc(conf->slab_cache, GFP_KERNEL);
1347 sh->raid_conf = conf;
1348 #ifdef CONFIG_MULTICORE_RAID456
1349 init_waitqueue_head(&sh->ops.wait_for_ops);
1352 if (grow_buffers(sh)) {
1354 kmem_cache_free(conf->slab_cache, sh);
1357 /* we just created an active stripe so... */
1358 atomic_set(&sh->count, 1);
1359 atomic_inc(&conf->active_stripes);
1360 INIT_LIST_HEAD(&sh->lru);
1365 static int grow_stripes(raid5_conf_t *conf, int num)
1367 struct kmem_cache *sc;
1368 int devs = max(conf->raid_disks, conf->previous_raid_disks);
1370 if (conf->mddev->gendisk)
1371 sprintf(conf->cache_name[0],
1372 "raid%d-%s", conf->level, mdname(conf->mddev));
1374 sprintf(conf->cache_name[0],
1375 "raid%d-%p", conf->level, conf->mddev);
1376 sprintf(conf->cache_name[1], "%s-alt", conf->cache_name[0]);
1378 conf->active_name = 0;
1379 sc = kmem_cache_create(conf->cache_name[conf->active_name],
1380 sizeof(struct stripe_head)+(devs-1)*sizeof(struct r5dev),
1384 conf->slab_cache = sc;
1385 conf->pool_size = devs;
1387 if (!grow_one_stripe(conf))
1393 * scribble_len - return the required size of the scribble region
1394 * @num - total number of disks in the array
1396 * The size must be enough to contain:
1397 * 1/ a struct page pointer for each device in the array +2
1398 * 2/ room to convert each entry in (1) to its corresponding dma
1399 * (dma_map_page()) or page (page_address()) address.
1401 * Note: the +2 is for the destination buffers of the ddf/raid6 case where we
1402 * calculate over all devices (not just the data blocks), using zeros in place
1403 * of the P and Q blocks.
1405 static size_t scribble_len(int num)
1409 len = sizeof(struct page *) * (num+2) + sizeof(addr_conv_t) * (num+2);
1414 static int resize_stripes(raid5_conf_t *conf, int newsize)
1416 /* Make all the stripes able to hold 'newsize' devices.
1417 * New slots in each stripe get 'page' set to a new page.
1419 * This happens in stages:
1420 * 1/ create a new kmem_cache and allocate the required number of
1422 * 2/ gather all the old stripe_heads and tranfer the pages across
1423 * to the new stripe_heads. This will have the side effect of
1424 * freezing the array as once all stripe_heads have been collected,
1425 * no IO will be possible. Old stripe heads are freed once their
1426 * pages have been transferred over, and the old kmem_cache is
1427 * freed when all stripes are done.
1428 * 3/ reallocate conf->disks to be suitable bigger. If this fails,
1429 * we simple return a failre status - no need to clean anything up.
1430 * 4/ allocate new pages for the new slots in the new stripe_heads.
1431 * If this fails, we don't bother trying the shrink the
1432 * stripe_heads down again, we just leave them as they are.
1433 * As each stripe_head is processed the new one is released into
1436 * Once step2 is started, we cannot afford to wait for a write,
1437 * so we use GFP_NOIO allocations.
1439 struct stripe_head *osh, *nsh;
1440 LIST_HEAD(newstripes);
1441 struct disk_info *ndisks;
1444 struct kmem_cache *sc;
1447 if (newsize <= conf->pool_size)
1448 return 0; /* never bother to shrink */
1450 err = md_allow_write(conf->mddev);
1455 sc = kmem_cache_create(conf->cache_name[1-conf->active_name],
1456 sizeof(struct stripe_head)+(newsize-1)*sizeof(struct r5dev),
1461 for (i = conf->max_nr_stripes; i; i--) {
1462 nsh = kmem_cache_zalloc(sc, GFP_KERNEL);
1466 nsh->raid_conf = conf;
1467 #ifdef CONFIG_MULTICORE_RAID456
1468 init_waitqueue_head(&nsh->ops.wait_for_ops);
1471 list_add(&nsh->lru, &newstripes);
1474 /* didn't get enough, give up */
1475 while (!list_empty(&newstripes)) {
1476 nsh = list_entry(newstripes.next, struct stripe_head, lru);
1477 list_del(&nsh->lru);
1478 kmem_cache_free(sc, nsh);
1480 kmem_cache_destroy(sc);
1483 /* Step 2 - Must use GFP_NOIO now.
1484 * OK, we have enough stripes, start collecting inactive
1485 * stripes and copying them over
1487 list_for_each_entry(nsh, &newstripes, lru) {
1488 spin_lock_irq(&conf->device_lock);
1489 wait_event_lock_irq(conf->wait_for_stripe,
1490 !list_empty(&conf->inactive_list),
1493 osh = get_free_stripe(conf);
1494 spin_unlock_irq(&conf->device_lock);
1495 atomic_set(&nsh->count, 1);
1496 for(i=0; i<conf->pool_size; i++)
1497 nsh->dev[i].page = osh->dev[i].page;
1498 for( ; i<newsize; i++)
1499 nsh->dev[i].page = NULL;
1500 kmem_cache_free(conf->slab_cache, osh);
1502 kmem_cache_destroy(conf->slab_cache);
1505 * At this point, we are holding all the stripes so the array
1506 * is completely stalled, so now is a good time to resize
1507 * conf->disks and the scribble region
1509 ndisks = kzalloc(newsize * sizeof(struct disk_info), GFP_NOIO);
1511 for (i=0; i<conf->raid_disks; i++)
1512 ndisks[i] = conf->disks[i];
1514 conf->disks = ndisks;
1519 conf->scribble_len = scribble_len(newsize);
1520 for_each_present_cpu(cpu) {
1521 struct raid5_percpu *percpu;
1524 percpu = per_cpu_ptr(conf->percpu, cpu);
1525 scribble = kmalloc(conf->scribble_len, GFP_NOIO);
1528 kfree(percpu->scribble);
1529 percpu->scribble = scribble;
1537 /* Step 4, return new stripes to service */
1538 while(!list_empty(&newstripes)) {
1539 nsh = list_entry(newstripes.next, struct stripe_head, lru);
1540 list_del_init(&nsh->lru);
1542 for (i=conf->raid_disks; i < newsize; i++)
1543 if (nsh->dev[i].page == NULL) {
1544 struct page *p = alloc_page(GFP_NOIO);
1545 nsh->dev[i].page = p;
1549 release_stripe(nsh);
1551 /* critical section pass, GFP_NOIO no longer needed */
1553 conf->slab_cache = sc;
1554 conf->active_name = 1-conf->active_name;
1555 conf->pool_size = newsize;
1559 static int drop_one_stripe(raid5_conf_t *conf)
1561 struct stripe_head *sh;
1563 spin_lock_irq(&conf->device_lock);
1564 sh = get_free_stripe(conf);
1565 spin_unlock_irq(&conf->device_lock);
1568 BUG_ON(atomic_read(&sh->count));
1570 kmem_cache_free(conf->slab_cache, sh);
1571 atomic_dec(&conf->active_stripes);
1575 static void shrink_stripes(raid5_conf_t *conf)
1577 while (drop_one_stripe(conf))
1580 if (conf->slab_cache)
1581 kmem_cache_destroy(conf->slab_cache);
1582 conf->slab_cache = NULL;
1585 static void raid5_end_read_request(struct bio * bi, int error)
1587 struct stripe_head *sh = bi->bi_private;
1588 raid5_conf_t *conf = sh->raid_conf;
1589 int disks = sh->disks, i;
1590 int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
1591 char b[BDEVNAME_SIZE];
1595 for (i=0 ; i<disks; i++)
1596 if (bi == &sh->dev[i].req)
1599 pr_debug("end_read_request %llu/%d, count: %d, uptodate %d.\n",
1600 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
1608 set_bit(R5_UPTODATE, &sh->dev[i].flags);
1609 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
1610 rdev = conf->disks[i].rdev;
1613 "md/raid:%s: read error corrected"
1614 " (%lu sectors at %llu on %s)\n",
1615 mdname(conf->mddev), STRIPE_SECTORS,
1616 (unsigned long long)(sh->sector
1617 + rdev->data_offset),
1618 bdevname(rdev->bdev, b));
1619 atomic_add(STRIPE_SECTORS, &rdev->corrected_errors);
1620 clear_bit(R5_ReadError, &sh->dev[i].flags);
1621 clear_bit(R5_ReWrite, &sh->dev[i].flags);
1623 if (atomic_read(&conf->disks[i].rdev->read_errors))
1624 atomic_set(&conf->disks[i].rdev->read_errors, 0);
1626 const char *bdn = bdevname(conf->disks[i].rdev->bdev, b);
1628 rdev = conf->disks[i].rdev;
1630 clear_bit(R5_UPTODATE, &sh->dev[i].flags);
1631 atomic_inc(&rdev->read_errors);
1632 if (conf->mddev->degraded >= conf->max_degraded)
1635 "md/raid:%s: read error not correctable "
1636 "(sector %llu on %s).\n",
1637 mdname(conf->mddev),
1638 (unsigned long long)(sh->sector
1639 + rdev->data_offset),
1641 else if (test_bit(R5_ReWrite, &sh->dev[i].flags))
1645 "md/raid:%s: read error NOT corrected!! "
1646 "(sector %llu on %s).\n",
1647 mdname(conf->mddev),
1648 (unsigned long long)(sh->sector
1649 + rdev->data_offset),
1651 else if (atomic_read(&rdev->read_errors)
1652 > conf->max_nr_stripes)
1654 "md/raid:%s: Too many read errors, failing device %s.\n",
1655 mdname(conf->mddev), bdn);
1659 set_bit(R5_ReadError, &sh->dev[i].flags);
1661 clear_bit(R5_ReadError, &sh->dev[i].flags);
1662 clear_bit(R5_ReWrite, &sh->dev[i].flags);
1663 md_error(conf->mddev, rdev);
1666 rdev_dec_pending(conf->disks[i].rdev, conf->mddev);
1667 clear_bit(R5_LOCKED, &sh->dev[i].flags);
1668 set_bit(STRIPE_HANDLE, &sh->state);
1672 static void raid5_end_write_request(struct bio *bi, int error)
1674 struct stripe_head *sh = bi->bi_private;
1675 raid5_conf_t *conf = sh->raid_conf;
1676 int disks = sh->disks, i;
1677 int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
1679 for (i=0 ; i<disks; i++)
1680 if (bi == &sh->dev[i].req)
1683 pr_debug("end_write_request %llu/%d, count %d, uptodate: %d.\n",
1684 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
1692 set_bit(WriteErrorSeen, &conf->disks[i].rdev->flags);
1693 set_bit(R5_WriteError, &sh->dev[i].flags);
1696 rdev_dec_pending(conf->disks[i].rdev, conf->mddev);
1698 clear_bit(R5_LOCKED, &sh->dev[i].flags);
1699 set_bit(STRIPE_HANDLE, &sh->state);
1704 static sector_t compute_blocknr(struct stripe_head *sh, int i, int previous);
1706 static void raid5_build_block(struct stripe_head *sh, int i, int previous)
1708 struct r5dev *dev = &sh->dev[i];
1710 bio_init(&dev->req);
1711 dev->req.bi_io_vec = &dev->vec;
1713 dev->req.bi_max_vecs++;
1714 dev->vec.bv_page = dev->page;
1715 dev->vec.bv_len = STRIPE_SIZE;
1716 dev->vec.bv_offset = 0;
1718 dev->req.bi_sector = sh->sector;
1719 dev->req.bi_private = sh;
1722 dev->sector = compute_blocknr(sh, i, previous);
1725 static void error(mddev_t *mddev, mdk_rdev_t *rdev)
1727 char b[BDEVNAME_SIZE];
1728 raid5_conf_t *conf = mddev->private;
1729 pr_debug("raid456: error called\n");
1731 if (test_and_clear_bit(In_sync, &rdev->flags)) {
1732 unsigned long flags;
1733 spin_lock_irqsave(&conf->device_lock, flags);
1735 spin_unlock_irqrestore(&conf->device_lock, flags);
1737 * if recovery was running, make sure it aborts.
1739 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1741 set_bit(Blocked, &rdev->flags);
1742 set_bit(Faulty, &rdev->flags);
1743 set_bit(MD_CHANGE_DEVS, &mddev->flags);
1745 "md/raid:%s: Disk failure on %s, disabling device.\n"
1746 "md/raid:%s: Operation continuing on %d devices.\n",
1748 bdevname(rdev->bdev, b),
1750 conf->raid_disks - mddev->degraded);
1754 * Input: a 'big' sector number,
1755 * Output: index of the data and parity disk, and the sector # in them.
1757 static sector_t raid5_compute_sector(raid5_conf_t *conf, sector_t r_sector,
1758 int previous, int *dd_idx,
1759 struct stripe_head *sh)
1761 sector_t stripe, stripe2;
1762 sector_t chunk_number;
1763 unsigned int chunk_offset;
1766 sector_t new_sector;
1767 int algorithm = previous ? conf->prev_algo
1769 int sectors_per_chunk = previous ? conf->prev_chunk_sectors
1770 : conf->chunk_sectors;
1771 int raid_disks = previous ? conf->previous_raid_disks
1773 int data_disks = raid_disks - conf->max_degraded;
1775 /* First compute the information on this sector */
1778 * Compute the chunk number and the sector offset inside the chunk
1780 chunk_offset = sector_div(r_sector, sectors_per_chunk);
1781 chunk_number = r_sector;
1784 * Compute the stripe number
1786 stripe = chunk_number;
1787 *dd_idx = sector_div(stripe, data_disks);
1790 * Select the parity disk based on the user selected algorithm.
1792 pd_idx = qd_idx = -1;
1793 switch(conf->level) {
1795 pd_idx = data_disks;
1798 switch (algorithm) {
1799 case ALGORITHM_LEFT_ASYMMETRIC:
1800 pd_idx = data_disks - sector_div(stripe2, raid_disks);
1801 if (*dd_idx >= pd_idx)
1804 case ALGORITHM_RIGHT_ASYMMETRIC:
1805 pd_idx = sector_div(stripe2, raid_disks);
1806 if (*dd_idx >= pd_idx)
1809 case ALGORITHM_LEFT_SYMMETRIC:
1810 pd_idx = data_disks - sector_div(stripe2, raid_disks);
1811 *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
1813 case ALGORITHM_RIGHT_SYMMETRIC:
1814 pd_idx = sector_div(stripe2, raid_disks);
1815 *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
1817 case ALGORITHM_PARITY_0:
1821 case ALGORITHM_PARITY_N:
1822 pd_idx = data_disks;
1830 switch (algorithm) {
1831 case ALGORITHM_LEFT_ASYMMETRIC:
1832 pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
1833 qd_idx = pd_idx + 1;
1834 if (pd_idx == raid_disks-1) {
1835 (*dd_idx)++; /* Q D D D P */
1837 } else if (*dd_idx >= pd_idx)
1838 (*dd_idx) += 2; /* D D P Q D */
1840 case ALGORITHM_RIGHT_ASYMMETRIC:
1841 pd_idx = sector_div(stripe2, raid_disks);
1842 qd_idx = pd_idx + 1;
1843 if (pd_idx == raid_disks-1) {
1844 (*dd_idx)++; /* Q D D D P */
1846 } else if (*dd_idx >= pd_idx)
1847 (*dd_idx) += 2; /* D D P Q D */
1849 case ALGORITHM_LEFT_SYMMETRIC:
1850 pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
1851 qd_idx = (pd_idx + 1) % raid_disks;
1852 *dd_idx = (pd_idx + 2 + *dd_idx) % raid_disks;
1854 case ALGORITHM_RIGHT_SYMMETRIC:
1855 pd_idx = sector_div(stripe2, raid_disks);
1856 qd_idx = (pd_idx + 1) % raid_disks;
1857 *dd_idx = (pd_idx + 2 + *dd_idx) % raid_disks;
1860 case ALGORITHM_PARITY_0:
1865 case ALGORITHM_PARITY_N:
1866 pd_idx = data_disks;
1867 qd_idx = data_disks + 1;
1870 case ALGORITHM_ROTATING_ZERO_RESTART:
1871 /* Exactly the same as RIGHT_ASYMMETRIC, but or
1872 * of blocks for computing Q is different.
1874 pd_idx = sector_div(stripe2, raid_disks);
1875 qd_idx = pd_idx + 1;
1876 if (pd_idx == raid_disks-1) {
1877 (*dd_idx)++; /* Q D D D P */
1879 } else if (*dd_idx >= pd_idx)
1880 (*dd_idx) += 2; /* D D P Q D */
1884 case ALGORITHM_ROTATING_N_RESTART:
1885 /* Same a left_asymmetric, by first stripe is
1886 * D D D P Q rather than
1890 pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
1891 qd_idx = pd_idx + 1;
1892 if (pd_idx == raid_disks-1) {
1893 (*dd_idx)++; /* Q D D D P */
1895 } else if (*dd_idx >= pd_idx)
1896 (*dd_idx) += 2; /* D D P Q D */
1900 case ALGORITHM_ROTATING_N_CONTINUE:
1901 /* Same as left_symmetric but Q is before P */
1902 pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
1903 qd_idx = (pd_idx + raid_disks - 1) % raid_disks;
1904 *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
1908 case ALGORITHM_LEFT_ASYMMETRIC_6:
1909 /* RAID5 left_asymmetric, with Q on last device */
1910 pd_idx = data_disks - sector_div(stripe2, raid_disks-1);
1911 if (*dd_idx >= pd_idx)
1913 qd_idx = raid_disks - 1;
1916 case ALGORITHM_RIGHT_ASYMMETRIC_6:
1917 pd_idx = sector_div(stripe2, raid_disks-1);
1918 if (*dd_idx >= pd_idx)
1920 qd_idx = raid_disks - 1;
1923 case ALGORITHM_LEFT_SYMMETRIC_6:
1924 pd_idx = data_disks - sector_div(stripe2, raid_disks-1);
1925 *dd_idx = (pd_idx + 1 + *dd_idx) % (raid_disks-1);
1926 qd_idx = raid_disks - 1;
1929 case ALGORITHM_RIGHT_SYMMETRIC_6:
1930 pd_idx = sector_div(stripe2, raid_disks-1);
1931 *dd_idx = (pd_idx + 1 + *dd_idx) % (raid_disks-1);
1932 qd_idx = raid_disks - 1;
1935 case ALGORITHM_PARITY_0_6:
1938 qd_idx = raid_disks - 1;
1948 sh->pd_idx = pd_idx;
1949 sh->qd_idx = qd_idx;
1950 sh->ddf_layout = ddf_layout;
1953 * Finally, compute the new sector number
1955 new_sector = (sector_t)stripe * sectors_per_chunk + chunk_offset;
1960 static sector_t compute_blocknr(struct stripe_head *sh, int i, int previous)
1962 raid5_conf_t *conf = sh->raid_conf;
1963 int raid_disks = sh->disks;
1964 int data_disks = raid_disks - conf->max_degraded;
1965 sector_t new_sector = sh->sector, check;
1966 int sectors_per_chunk = previous ? conf->prev_chunk_sectors
1967 : conf->chunk_sectors;
1968 int algorithm = previous ? conf->prev_algo
1972 sector_t chunk_number;
1973 int dummy1, dd_idx = i;
1975 struct stripe_head sh2;
1978 chunk_offset = sector_div(new_sector, sectors_per_chunk);
1979 stripe = new_sector;
1981 if (i == sh->pd_idx)
1983 switch(conf->level) {
1986 switch (algorithm) {
1987 case ALGORITHM_LEFT_ASYMMETRIC:
1988 case ALGORITHM_RIGHT_ASYMMETRIC:
1992 case ALGORITHM_LEFT_SYMMETRIC:
1993 case ALGORITHM_RIGHT_SYMMETRIC:
1996 i -= (sh->pd_idx + 1);
1998 case ALGORITHM_PARITY_0:
2001 case ALGORITHM_PARITY_N:
2008 if (i == sh->qd_idx)
2009 return 0; /* It is the Q disk */
2010 switch (algorithm) {
2011 case ALGORITHM_LEFT_ASYMMETRIC:
2012 case ALGORITHM_RIGHT_ASYMMETRIC:
2013 case ALGORITHM_ROTATING_ZERO_RESTART:
2014 case ALGORITHM_ROTATING_N_RESTART:
2015 if (sh->pd_idx == raid_disks-1)
2016 i--; /* Q D D D P */
2017 else if (i > sh->pd_idx)
2018 i -= 2; /* D D P Q D */
2020 case ALGORITHM_LEFT_SYMMETRIC:
2021 case ALGORITHM_RIGHT_SYMMETRIC:
2022 if (sh->pd_idx == raid_disks-1)
2023 i--; /* Q D D D P */
2028 i -= (sh->pd_idx + 2);
2031 case ALGORITHM_PARITY_0:
2034 case ALGORITHM_PARITY_N:
2036 case ALGORITHM_ROTATING_N_CONTINUE:
2037 /* Like left_symmetric, but P is before Q */
2038 if (sh->pd_idx == 0)
2039 i--; /* P D D D Q */
2044 i -= (sh->pd_idx + 1);
2047 case ALGORITHM_LEFT_ASYMMETRIC_6:
2048 case ALGORITHM_RIGHT_ASYMMETRIC_6:
2052 case ALGORITHM_LEFT_SYMMETRIC_6:
2053 case ALGORITHM_RIGHT_SYMMETRIC_6:
2055 i += data_disks + 1;
2056 i -= (sh->pd_idx + 1);
2058 case ALGORITHM_PARITY_0_6:
2067 chunk_number = stripe * data_disks + i;
2068 r_sector = chunk_number * sectors_per_chunk + chunk_offset;
2070 check = raid5_compute_sector(conf, r_sector,
2071 previous, &dummy1, &sh2);
2072 if (check != sh->sector || dummy1 != dd_idx || sh2.pd_idx != sh->pd_idx
2073 || sh2.qd_idx != sh->qd_idx) {
2074 printk(KERN_ERR "md/raid:%s: compute_blocknr: map not correct\n",
2075 mdname(conf->mddev));
2083 schedule_reconstruction(struct stripe_head *sh, struct stripe_head_state *s,
2084 int rcw, int expand)
2086 int i, pd_idx = sh->pd_idx, disks = sh->disks;
2087 raid5_conf_t *conf = sh->raid_conf;
2088 int level = conf->level;
2091 /* if we are not expanding this is a proper write request, and
2092 * there will be bios with new data to be drained into the
2096 sh->reconstruct_state = reconstruct_state_drain_run;
2097 set_bit(STRIPE_OP_BIODRAIN, &s->ops_request);
2099 sh->reconstruct_state = reconstruct_state_run;
2101 set_bit(STRIPE_OP_RECONSTRUCT, &s->ops_request);
2103 for (i = disks; i--; ) {
2104 struct r5dev *dev = &sh->dev[i];
2107 set_bit(R5_LOCKED, &dev->flags);
2108 set_bit(R5_Wantdrain, &dev->flags);
2110 clear_bit(R5_UPTODATE, &dev->flags);
2114 if (s->locked + conf->max_degraded == disks)
2115 if (!test_and_set_bit(STRIPE_FULL_WRITE, &sh->state))
2116 atomic_inc(&conf->pending_full_writes);
2119 BUG_ON(!(test_bit(R5_UPTODATE, &sh->dev[pd_idx].flags) ||
2120 test_bit(R5_Wantcompute, &sh->dev[pd_idx].flags)));
2122 sh->reconstruct_state = reconstruct_state_prexor_drain_run;
2123 set_bit(STRIPE_OP_PREXOR, &s->ops_request);
2124 set_bit(STRIPE_OP_BIODRAIN, &s->ops_request);
2125 set_bit(STRIPE_OP_RECONSTRUCT, &s->ops_request);
2127 for (i = disks; i--; ) {
2128 struct r5dev *dev = &sh->dev[i];
2133 (test_bit(R5_UPTODATE, &dev->flags) ||
2134 test_bit(R5_Wantcompute, &dev->flags))) {
2135 set_bit(R5_Wantdrain, &dev->flags);
2136 set_bit(R5_LOCKED, &dev->flags);
2137 clear_bit(R5_UPTODATE, &dev->flags);
2143 /* keep the parity disk(s) locked while asynchronous operations
2146 set_bit(R5_LOCKED, &sh->dev[pd_idx].flags);
2147 clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
2151 int qd_idx = sh->qd_idx;
2152 struct r5dev *dev = &sh->dev[qd_idx];
2154 set_bit(R5_LOCKED, &dev->flags);
2155 clear_bit(R5_UPTODATE, &dev->flags);
2159 pr_debug("%s: stripe %llu locked: %d ops_request: %lx\n",
2160 __func__, (unsigned long long)sh->sector,
2161 s->locked, s->ops_request);
2165 * Each stripe/dev can have one or more bion attached.
2166 * toread/towrite point to the first in a chain.
2167 * The bi_next chain must be in order.
2169 static int add_stripe_bio(struct stripe_head *sh, struct bio *bi, int dd_idx, int forwrite)
2172 raid5_conf_t *conf = sh->raid_conf;
2175 pr_debug("adding bi b#%llu to stripe s#%llu\n",
2176 (unsigned long long)bi->bi_sector,
2177 (unsigned long long)sh->sector);
2180 spin_lock_irq(&conf->device_lock);
2182 bip = &sh->dev[dd_idx].towrite;
2183 if (*bip == NULL && sh->dev[dd_idx].written == NULL)
2186 bip = &sh->dev[dd_idx].toread;
2187 while (*bip && (*bip)->bi_sector < bi->bi_sector) {
2188 if ((*bip)->bi_sector + ((*bip)->bi_size >> 9) > bi->bi_sector)
2190 bip = & (*bip)->bi_next;
2192 if (*bip && (*bip)->bi_sector < bi->bi_sector + ((bi->bi_size)>>9))
2195 BUG_ON(*bip && bi->bi_next && (*bip) != bi->bi_next);
2199 bi->bi_phys_segments++;
2202 /* check if page is covered */
2203 sector_t sector = sh->dev[dd_idx].sector;
2204 for (bi=sh->dev[dd_idx].towrite;
2205 sector < sh->dev[dd_idx].sector + STRIPE_SECTORS &&
2206 bi && bi->bi_sector <= sector;
2207 bi = r5_next_bio(bi, sh->dev[dd_idx].sector)) {
2208 if (bi->bi_sector + (bi->bi_size>>9) >= sector)
2209 sector = bi->bi_sector + (bi->bi_size>>9);
2211 if (sector >= sh->dev[dd_idx].sector + STRIPE_SECTORS)
2212 set_bit(R5_OVERWRITE, &sh->dev[dd_idx].flags);
2214 spin_unlock_irq(&conf->device_lock);
2216 pr_debug("added bi b#%llu to stripe s#%llu, disk %d.\n",
2217 (unsigned long long)(*bip)->bi_sector,
2218 (unsigned long long)sh->sector, dd_idx);
2220 if (conf->mddev->bitmap && firstwrite) {
2221 bitmap_startwrite(conf->mddev->bitmap, sh->sector,
2223 sh->bm_seq = conf->seq_flush+1;
2224 set_bit(STRIPE_BIT_DELAY, &sh->state);
2229 set_bit(R5_Overlap, &sh->dev[dd_idx].flags);
2230 spin_unlock_irq(&conf->device_lock);
2234 static void end_reshape(raid5_conf_t *conf);
2236 static void stripe_set_idx(sector_t stripe, raid5_conf_t *conf, int previous,
2237 struct stripe_head *sh)
2239 int sectors_per_chunk =
2240 previous ? conf->prev_chunk_sectors : conf->chunk_sectors;
2242 int chunk_offset = sector_div(stripe, sectors_per_chunk);
2243 int disks = previous ? conf->previous_raid_disks : conf->raid_disks;
2245 raid5_compute_sector(conf,
2246 stripe * (disks - conf->max_degraded)
2247 *sectors_per_chunk + chunk_offset,
2253 handle_failed_stripe(raid5_conf_t *conf, struct stripe_head *sh,
2254 struct stripe_head_state *s, int disks,
2255 struct bio **return_bi)
2258 for (i = disks; i--; ) {
2262 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
2265 rdev = rcu_dereference(conf->disks[i].rdev);
2266 if (rdev && test_bit(In_sync, &rdev->flags))
2267 atomic_inc(&rdev->nr_pending);
2272 if (!rdev_set_badblocks(
2276 md_error(conf->mddev, rdev);
2277 rdev_dec_pending(rdev, conf->mddev);
2280 spin_lock_irq(&conf->device_lock);
2281 /* fail all writes first */
2282 bi = sh->dev[i].towrite;
2283 sh->dev[i].towrite = NULL;
2289 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
2290 wake_up(&conf->wait_for_overlap);
2292 while (bi && bi->bi_sector <
2293 sh->dev[i].sector + STRIPE_SECTORS) {
2294 struct bio *nextbi = r5_next_bio(bi, sh->dev[i].sector);
2295 clear_bit(BIO_UPTODATE, &bi->bi_flags);
2296 if (!raid5_dec_bi_phys_segments(bi)) {
2297 md_write_end(conf->mddev);
2298 bi->bi_next = *return_bi;
2303 /* and fail all 'written' */
2304 bi = sh->dev[i].written;
2305 sh->dev[i].written = NULL;
2306 if (bi) bitmap_end = 1;
2307 while (bi && bi->bi_sector <
2308 sh->dev[i].sector + STRIPE_SECTORS) {
2309 struct bio *bi2 = r5_next_bio(bi, sh->dev[i].sector);
2310 clear_bit(BIO_UPTODATE, &bi->bi_flags);
2311 if (!raid5_dec_bi_phys_segments(bi)) {
2312 md_write_end(conf->mddev);
2313 bi->bi_next = *return_bi;
2319 /* fail any reads if this device is non-operational and
2320 * the data has not reached the cache yet.
2322 if (!test_bit(R5_Wantfill, &sh->dev[i].flags) &&
2323 (!test_bit(R5_Insync, &sh->dev[i].flags) ||
2324 test_bit(R5_ReadError, &sh->dev[i].flags))) {
2325 bi = sh->dev[i].toread;
2326 sh->dev[i].toread = NULL;
2327 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
2328 wake_up(&conf->wait_for_overlap);
2329 if (bi) s->to_read--;
2330 while (bi && bi->bi_sector <
2331 sh->dev[i].sector + STRIPE_SECTORS) {
2332 struct bio *nextbi =
2333 r5_next_bio(bi, sh->dev[i].sector);
2334 clear_bit(BIO_UPTODATE, &bi->bi_flags);
2335 if (!raid5_dec_bi_phys_segments(bi)) {
2336 bi->bi_next = *return_bi;
2342 spin_unlock_irq(&conf->device_lock);
2344 bitmap_endwrite(conf->mddev->bitmap, sh->sector,
2345 STRIPE_SECTORS, 0, 0);
2346 /* If we were in the middle of a write the parity block might
2347 * still be locked - so just clear all R5_LOCKED flags
2349 clear_bit(R5_LOCKED, &sh->dev[i].flags);
2352 if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
2353 if (atomic_dec_and_test(&conf->pending_full_writes))
2354 md_wakeup_thread(conf->mddev->thread);
2358 handle_failed_sync(raid5_conf_t *conf, struct stripe_head *sh,
2359 struct stripe_head_state *s)
2364 md_done_sync(conf->mddev, STRIPE_SECTORS, 0);
2365 clear_bit(STRIPE_SYNCING, &sh->state);
2367 /* There is nothing more to do for sync/check/repair.
2368 * For recover we need to record a bad block on all
2369 * non-sync devices, or abort the recovery
2371 if (!test_bit(MD_RECOVERY_RECOVER, &conf->mddev->recovery))
2373 /* During recovery devices cannot be removed, so locking and
2374 * refcounting of rdevs is not needed
2376 for (i = 0; i < conf->raid_disks; i++) {
2377 mdk_rdev_t *rdev = conf->disks[i].rdev;
2379 || test_bit(Faulty, &rdev->flags)
2380 || test_bit(In_sync, &rdev->flags))
2382 if (!rdev_set_badblocks(rdev, sh->sector,
2387 conf->recovery_disabled = conf->mddev->recovery_disabled;
2388 set_bit(MD_RECOVERY_INTR, &conf->mddev->recovery);
2392 /* fetch_block - checks the given member device to see if its data needs
2393 * to be read or computed to satisfy a request.
2395 * Returns 1 when no more member devices need to be checked, otherwise returns
2396 * 0 to tell the loop in handle_stripe_fill to continue
2398 static int fetch_block(struct stripe_head *sh, struct stripe_head_state *s,
2399 int disk_idx, int disks)
2401 struct r5dev *dev = &sh->dev[disk_idx];
2402 struct r5dev *fdev[2] = { &sh->dev[s->failed_num[0]],
2403 &sh->dev[s->failed_num[1]] };
2405 /* is the data in this block needed, and can we get it? */
2406 if (!test_bit(R5_LOCKED, &dev->flags) &&
2407 !test_bit(R5_UPTODATE, &dev->flags) &&
2409 (dev->towrite && !test_bit(R5_OVERWRITE, &dev->flags)) ||
2410 s->syncing || s->expanding ||
2411 (s->failed >= 1 && fdev[0]->toread) ||
2412 (s->failed >= 2 && fdev[1]->toread) ||
2413 (sh->raid_conf->level <= 5 && s->failed && fdev[0]->towrite &&
2414 !test_bit(R5_OVERWRITE, &fdev[0]->flags)) ||
2415 (sh->raid_conf->level == 6 && s->failed && s->to_write))) {
2416 /* we would like to get this block, possibly by computing it,
2417 * otherwise read it if the backing disk is insync
2419 BUG_ON(test_bit(R5_Wantcompute, &dev->flags));
2420 BUG_ON(test_bit(R5_Wantread, &dev->flags));
2421 if ((s->uptodate == disks - 1) &&
2422 (s->failed && (disk_idx == s->failed_num[0] ||
2423 disk_idx == s->failed_num[1]))) {
2424 /* have disk failed, and we're requested to fetch it;
2427 pr_debug("Computing stripe %llu block %d\n",
2428 (unsigned long long)sh->sector, disk_idx);
2429 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
2430 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
2431 set_bit(R5_Wantcompute, &dev->flags);
2432 sh->ops.target = disk_idx;
2433 sh->ops.target2 = -1; /* no 2nd target */
2435 /* Careful: from this point on 'uptodate' is in the eye
2436 * of raid_run_ops which services 'compute' operations
2437 * before writes. R5_Wantcompute flags a block that will
2438 * be R5_UPTODATE by the time it is needed for a
2439 * subsequent operation.
2443 } else if (s->uptodate == disks-2 && s->failed >= 2) {
2444 /* Computing 2-failure is *very* expensive; only
2445 * do it if failed >= 2
2448 for (other = disks; other--; ) {
2449 if (other == disk_idx)
2451 if (!test_bit(R5_UPTODATE,
2452 &sh->dev[other].flags))
2456 pr_debug("Computing stripe %llu blocks %d,%d\n",
2457 (unsigned long long)sh->sector,
2459 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
2460 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
2461 set_bit(R5_Wantcompute, &sh->dev[disk_idx].flags);
2462 set_bit(R5_Wantcompute, &sh->dev[other].flags);
2463 sh->ops.target = disk_idx;
2464 sh->ops.target2 = other;
2468 } else if (test_bit(R5_Insync, &dev->flags)) {
2469 set_bit(R5_LOCKED, &dev->flags);
2470 set_bit(R5_Wantread, &dev->flags);
2472 pr_debug("Reading block %d (sync=%d)\n",
2473 disk_idx, s->syncing);
2481 * handle_stripe_fill - read or compute data to satisfy pending requests.
2483 static void handle_stripe_fill(struct stripe_head *sh,
2484 struct stripe_head_state *s,
2489 /* look for blocks to read/compute, skip this if a compute
2490 * is already in flight, or if the stripe contents are in the
2491 * midst of changing due to a write
2493 if (!test_bit(STRIPE_COMPUTE_RUN, &sh->state) && !sh->check_state &&
2494 !sh->reconstruct_state)
2495 for (i = disks; i--; )
2496 if (fetch_block(sh, s, i, disks))
2498 set_bit(STRIPE_HANDLE, &sh->state);
2502 /* handle_stripe_clean_event
2503 * any written block on an uptodate or failed drive can be returned.
2504 * Note that if we 'wrote' to a failed drive, it will be UPTODATE, but
2505 * never LOCKED, so we don't need to test 'failed' directly.
2507 static void handle_stripe_clean_event(raid5_conf_t *conf,
2508 struct stripe_head *sh, int disks, struct bio **return_bi)
2513 for (i = disks; i--; )
2514 if (sh->dev[i].written) {
2516 if (!test_bit(R5_LOCKED, &dev->flags) &&
2517 test_bit(R5_UPTODATE, &dev->flags)) {
2518 /* We can return any write requests */
2519 struct bio *wbi, *wbi2;
2521 pr_debug("Return write for disc %d\n", i);
2522 spin_lock_irq(&conf->device_lock);
2524 dev->written = NULL;
2525 while (wbi && wbi->bi_sector <
2526 dev->sector + STRIPE_SECTORS) {
2527 wbi2 = r5_next_bio(wbi, dev->sector);
2528 if (!raid5_dec_bi_phys_segments(wbi)) {
2529 md_write_end(conf->mddev);
2530 wbi->bi_next = *return_bi;
2535 if (dev->towrite == NULL)
2537 spin_unlock_irq(&conf->device_lock);
2539 bitmap_endwrite(conf->mddev->bitmap,
2542 !test_bit(STRIPE_DEGRADED, &sh->state),
2547 if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
2548 if (atomic_dec_and_test(&conf->pending_full_writes))
2549 md_wakeup_thread(conf->mddev->thread);
2552 static void handle_stripe_dirtying(raid5_conf_t *conf,
2553 struct stripe_head *sh,
2554 struct stripe_head_state *s,
2557 int rmw = 0, rcw = 0, i;
2558 if (conf->max_degraded == 2) {
2559 /* RAID6 requires 'rcw' in current implementation
2560 * Calculate the real rcw later - for now fake it
2561 * look like rcw is cheaper
2564 } else for (i = disks; i--; ) {
2565 /* would I have to read this buffer for read_modify_write */
2566 struct r5dev *dev = &sh->dev[i];
2567 if ((dev->towrite || i == sh->pd_idx) &&
2568 !test_bit(R5_LOCKED, &dev->flags) &&
2569 !(test_bit(R5_UPTODATE, &dev->flags) ||
2570 test_bit(R5_Wantcompute, &dev->flags))) {
2571 if (test_bit(R5_Insync, &dev->flags))
2574 rmw += 2*disks; /* cannot read it */
2576 /* Would I have to read this buffer for reconstruct_write */
2577 if (!test_bit(R5_OVERWRITE, &dev->flags) && i != sh->pd_idx &&
2578 !test_bit(R5_LOCKED, &dev->flags) &&
2579 !(test_bit(R5_UPTODATE, &dev->flags) ||
2580 test_bit(R5_Wantcompute, &dev->flags))) {
2581 if (test_bit(R5_Insync, &dev->flags)) rcw++;
2586 pr_debug("for sector %llu, rmw=%d rcw=%d\n",
2587 (unsigned long long)sh->sector, rmw, rcw);
2588 set_bit(STRIPE_HANDLE, &sh->state);
2589 if (rmw < rcw && rmw > 0)
2590 /* prefer read-modify-write, but need to get some data */
2591 for (i = disks; i--; ) {
2592 struct r5dev *dev = &sh->dev[i];
2593 if ((dev->towrite || i == sh->pd_idx) &&
2594 !test_bit(R5_LOCKED, &dev->flags) &&
2595 !(test_bit(R5_UPTODATE, &dev->flags) ||
2596 test_bit(R5_Wantcompute, &dev->flags)) &&
2597 test_bit(R5_Insync, &dev->flags)) {
2599 test_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2600 pr_debug("Read_old block "
2601 "%d for r-m-w\n", i);
2602 set_bit(R5_LOCKED, &dev->flags);
2603 set_bit(R5_Wantread, &dev->flags);
2606 set_bit(STRIPE_DELAYED, &sh->state);
2607 set_bit(STRIPE_HANDLE, &sh->state);
2611 if (rcw <= rmw && rcw > 0) {
2612 /* want reconstruct write, but need to get some data */
2614 for (i = disks; i--; ) {
2615 struct r5dev *dev = &sh->dev[i];
2616 if (!test_bit(R5_OVERWRITE, &dev->flags) &&
2617 i != sh->pd_idx && i != sh->qd_idx &&
2618 !test_bit(R5_LOCKED, &dev->flags) &&
2619 !(test_bit(R5_UPTODATE, &dev->flags) ||
2620 test_bit(R5_Wantcompute, &dev->flags))) {
2622 if (!test_bit(R5_Insync, &dev->flags))
2623 continue; /* it's a failed drive */
2625 test_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2626 pr_debug("Read_old block "
2627 "%d for Reconstruct\n", i);
2628 set_bit(R5_LOCKED, &dev->flags);
2629 set_bit(R5_Wantread, &dev->flags);
2632 set_bit(STRIPE_DELAYED, &sh->state);
2633 set_bit(STRIPE_HANDLE, &sh->state);
2638 /* now if nothing is locked, and if we have enough data,
2639 * we can start a write request
2641 /* since handle_stripe can be called at any time we need to handle the
2642 * case where a compute block operation has been submitted and then a
2643 * subsequent call wants to start a write request. raid_run_ops only
2644 * handles the case where compute block and reconstruct are requested
2645 * simultaneously. If this is not the case then new writes need to be
2646 * held off until the compute completes.
2648 if ((s->req_compute || !test_bit(STRIPE_COMPUTE_RUN, &sh->state)) &&
2649 (s->locked == 0 && (rcw == 0 || rmw == 0) &&
2650 !test_bit(STRIPE_BIT_DELAY, &sh->state)))
2651 schedule_reconstruction(sh, s, rcw == 0, 0);
2654 static void handle_parity_checks5(raid5_conf_t *conf, struct stripe_head *sh,
2655 struct stripe_head_state *s, int disks)
2657 struct r5dev *dev = NULL;
2659 set_bit(STRIPE_HANDLE, &sh->state);
2661 switch (sh->check_state) {
2662 case check_state_idle:
2663 /* start a new check operation if there are no failures */
2664 if (s->failed == 0) {
2665 BUG_ON(s->uptodate != disks);
2666 sh->check_state = check_state_run;
2667 set_bit(STRIPE_OP_CHECK, &s->ops_request);
2668 clear_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags);
2672 dev = &sh->dev[s->failed_num[0]];
2674 case check_state_compute_result:
2675 sh->check_state = check_state_idle;
2677 dev = &sh->dev[sh->pd_idx];
2679 /* check that a write has not made the stripe insync */
2680 if (test_bit(STRIPE_INSYNC, &sh->state))
2683 /* either failed parity check, or recovery is happening */
2684 BUG_ON(!test_bit(R5_UPTODATE, &dev->flags));
2685 BUG_ON(s->uptodate != disks);
2687 set_bit(R5_LOCKED, &dev->flags);
2689 set_bit(R5_Wantwrite, &dev->flags);
2691 clear_bit(STRIPE_DEGRADED, &sh->state);
2692 set_bit(STRIPE_INSYNC, &sh->state);
2694 case check_state_run:
2695 break; /* we will be called again upon completion */
2696 case check_state_check_result:
2697 sh->check_state = check_state_idle;
2699 /* if a failure occurred during the check operation, leave
2700 * STRIPE_INSYNC not set and let the stripe be handled again
2705 /* handle a successful check operation, if parity is correct
2706 * we are done. Otherwise update the mismatch count and repair
2707 * parity if !MD_RECOVERY_CHECK
2709 if ((sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) == 0)
2710 /* parity is correct (on disc,
2711 * not in buffer any more)
2713 set_bit(STRIPE_INSYNC, &sh->state);
2715 conf->mddev->resync_mismatches += STRIPE_SECTORS;
2716 if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery))
2717 /* don't try to repair!! */
2718 set_bit(STRIPE_INSYNC, &sh->state);
2720 sh->check_state = check_state_compute_run;
2721 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
2722 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
2723 set_bit(R5_Wantcompute,
2724 &sh->dev[sh->pd_idx].flags);
2725 sh->ops.target = sh->pd_idx;
2726 sh->ops.target2 = -1;
2731 case check_state_compute_run:
2734 printk(KERN_ERR "%s: unknown check_state: %d sector: %llu\n",
2735 __func__, sh->check_state,
2736 (unsigned long long) sh->sector);
2742 static void handle_parity_checks6(raid5_conf_t *conf, struct stripe_head *sh,
2743 struct stripe_head_state *s,
2746 int pd_idx = sh->pd_idx;
2747 int qd_idx = sh->qd_idx;
2750 set_bit(STRIPE_HANDLE, &sh->state);
2752 BUG_ON(s->failed > 2);
2754 /* Want to check and possibly repair P and Q.
2755 * However there could be one 'failed' device, in which
2756 * case we can only check one of them, possibly using the
2757 * other to generate missing data
2760 switch (sh->check_state) {
2761 case check_state_idle:
2762 /* start a new check operation if there are < 2 failures */
2763 if (s->failed == s->q_failed) {
2764 /* The only possible failed device holds Q, so it
2765 * makes sense to check P (If anything else were failed,
2766 * we would have used P to recreate it).
2768 sh->check_state = check_state_run;
2770 if (!s->q_failed && s->failed < 2) {
2771 /* Q is not failed, and we didn't use it to generate
2772 * anything, so it makes sense to check it
2774 if (sh->check_state == check_state_run)
2775 sh->check_state = check_state_run_pq;
2777 sh->check_state = check_state_run_q;
2780 /* discard potentially stale zero_sum_result */
2781 sh->ops.zero_sum_result = 0;
2783 if (sh->check_state == check_state_run) {
2784 /* async_xor_zero_sum destroys the contents of P */
2785 clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
2788 if (sh->check_state >= check_state_run &&
2789 sh->check_state <= check_state_run_pq) {
2790 /* async_syndrome_zero_sum preserves P and Q, so
2791 * no need to mark them !uptodate here
2793 set_bit(STRIPE_OP_CHECK, &s->ops_request);
2797 /* we have 2-disk failure */
2798 BUG_ON(s->failed != 2);
2800 case check_state_compute_result:
2801 sh->check_state = check_state_idle;
2803 /* check that a write has not made the stripe insync */
2804 if (test_bit(STRIPE_INSYNC, &sh->state))
2807 /* now write out any block on a failed drive,
2808 * or P or Q if they were recomputed
2810 BUG_ON(s->uptodate < disks - 1); /* We don't need Q to recover */
2811 if (s->failed == 2) {
2812 dev = &sh->dev[s->failed_num[1]];
2814 set_bit(R5_LOCKED, &dev->flags);
2815 set_bit(R5_Wantwrite, &dev->flags);
2817 if (s->failed >= 1) {
2818 dev = &sh->dev[s->failed_num[0]];
2820 set_bit(R5_LOCKED, &dev->flags);
2821 set_bit(R5_Wantwrite, &dev->flags);
2823 if (sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) {
2824 dev = &sh->dev[pd_idx];
2826 set_bit(R5_LOCKED, &dev->flags);
2827 set_bit(R5_Wantwrite, &dev->flags);
2829 if (sh->ops.zero_sum_result & SUM_CHECK_Q_RESULT) {
2830 dev = &sh->dev[qd_idx];
2832 set_bit(R5_LOCKED, &dev->flags);
2833 set_bit(R5_Wantwrite, &dev->flags);
2835 clear_bit(STRIPE_DEGRADED, &sh->state);
2837 set_bit(STRIPE_INSYNC, &sh->state);
2839 case check_state_run:
2840 case check_state_run_q:
2841 case check_state_run_pq:
2842 break; /* we will be called again upon completion */
2843 case check_state_check_result:
2844 sh->check_state = check_state_idle;
2846 /* handle a successful check operation, if parity is correct
2847 * we are done. Otherwise update the mismatch count and repair
2848 * parity if !MD_RECOVERY_CHECK
2850 if (sh->ops.zero_sum_result == 0) {
2851 /* both parities are correct */
2853 set_bit(STRIPE_INSYNC, &sh->state);
2855 /* in contrast to the raid5 case we can validate
2856 * parity, but still have a failure to write
2859 sh->check_state = check_state_compute_result;
2860 /* Returning at this point means that we may go
2861 * off and bring p and/or q uptodate again so
2862 * we make sure to check zero_sum_result again
2863 * to verify if p or q need writeback
2867 conf->mddev->resync_mismatches += STRIPE_SECTORS;
2868 if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery))
2869 /* don't try to repair!! */
2870 set_bit(STRIPE_INSYNC, &sh->state);
2872 int *target = &sh->ops.target;
2874 sh->ops.target = -1;
2875 sh->ops.target2 = -1;
2876 sh->check_state = check_state_compute_run;
2877 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
2878 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
2879 if (sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) {
2880 set_bit(R5_Wantcompute,
2881 &sh->dev[pd_idx].flags);
2883 target = &sh->ops.target2;
2886 if (sh->ops.zero_sum_result & SUM_CHECK_Q_RESULT) {
2887 set_bit(R5_Wantcompute,
2888 &sh->dev[qd_idx].flags);
2895 case check_state_compute_run:
2898 printk(KERN_ERR "%s: unknown check_state: %d sector: %llu\n",
2899 __func__, sh->check_state,
2900 (unsigned long long) sh->sector);
2905 static void handle_stripe_expansion(raid5_conf_t *conf, struct stripe_head *sh)
2909 /* We have read all the blocks in this stripe and now we need to
2910 * copy some of them into a target stripe for expand.
2912 struct dma_async_tx_descriptor *tx = NULL;
2913 clear_bit(STRIPE_EXPAND_SOURCE, &sh->state);
2914 for (i = 0; i < sh->disks; i++)
2915 if (i != sh->pd_idx && i != sh->qd_idx) {
2917 struct stripe_head *sh2;
2918 struct async_submit_ctl submit;
2920 sector_t bn = compute_blocknr(sh, i, 1);
2921 sector_t s = raid5_compute_sector(conf, bn, 0,
2923 sh2 = get_active_stripe(conf, s, 0, 1, 1);
2925 /* so far only the early blocks of this stripe
2926 * have been requested. When later blocks
2927 * get requested, we will try again
2930 if (!test_bit(STRIPE_EXPANDING, &sh2->state) ||
2931 test_bit(R5_Expanded, &sh2->dev[dd_idx].flags)) {
2932 /* must have already done this block */
2933 release_stripe(sh2);
2937 /* place all the copies on one channel */
2938 init_async_submit(&submit, 0, tx, NULL, NULL, NULL);
2939 tx = async_memcpy(sh2->dev[dd_idx].page,
2940 sh->dev[i].page, 0, 0, STRIPE_SIZE,
2943 set_bit(R5_Expanded, &sh2->dev[dd_idx].flags);
2944 set_bit(R5_UPTODATE, &sh2->dev[dd_idx].flags);
2945 for (j = 0; j < conf->raid_disks; j++)
2946 if (j != sh2->pd_idx &&
2948 !test_bit(R5_Expanded, &sh2->dev[j].flags))
2950 if (j == conf->raid_disks) {
2951 set_bit(STRIPE_EXPAND_READY, &sh2->state);
2952 set_bit(STRIPE_HANDLE, &sh2->state);
2954 release_stripe(sh2);
2957 /* done submitting copies, wait for them to complete */
2960 dma_wait_for_async_tx(tx);
2966 * handle_stripe - do things to a stripe.
2968 * We lock the stripe and then examine the state of various bits
2969 * to see what needs to be done.
2971 * return some read request which now have data
2972 * return some write requests which are safely on disc
2973 * schedule a read on some buffers
2974 * schedule a write of some buffers
2975 * return confirmation of parity correctness
2977 * buffers are taken off read_list or write_list, and bh_cache buffers
2978 * get BH_Lock set before the stripe lock is released.
2982 static void analyse_stripe(struct stripe_head *sh, struct stripe_head_state *s)
2984 raid5_conf_t *conf = sh->raid_conf;
2985 int disks = sh->disks;
2989 memset(s, 0, sizeof(*s));
2991 s->syncing = test_bit(STRIPE_SYNCING, &sh->state);
2992 s->expanding = test_bit(STRIPE_EXPAND_SOURCE, &sh->state);
2993 s->expanded = test_bit(STRIPE_EXPAND_READY, &sh->state);
2994 s->failed_num[0] = -1;
2995 s->failed_num[1] = -1;
2997 /* Now to look around and see what can be done */
2999 spin_lock_irq(&conf->device_lock);
3000 for (i=disks; i--; ) {
3008 pr_debug("check %d: state 0x%lx read %p write %p written %p\n",
3009 i, dev->flags, dev->toread, dev->towrite, dev->written);
3010 /* maybe we can reply to a read
3012 * new wantfill requests are only permitted while
3013 * ops_complete_biofill is guaranteed to be inactive
3015 if (test_bit(R5_UPTODATE, &dev->flags) && dev->toread &&
3016 !test_bit(STRIPE_BIOFILL_RUN, &sh->state))
3017 set_bit(R5_Wantfill, &dev->flags);
3019 /* now count some things */
3020 if (test_bit(R5_LOCKED, &dev->flags))
3022 if (test_bit(R5_UPTODATE, &dev->flags))
3024 if (test_bit(R5_Wantcompute, &dev->flags)) {
3026 BUG_ON(s->compute > 2);
3029 if (test_bit(R5_Wantfill, &dev->flags))
3031 else if (dev->toread)
3035 if (!test_bit(R5_OVERWRITE, &dev->flags))
3040 rdev = rcu_dereference(conf->disks[i].rdev);
3042 is_bad = is_badblock(rdev, sh->sector, STRIPE_SECTORS,
3043 &first_bad, &bad_sectors);
3044 if (s->blocked_rdev == NULL
3045 && (test_bit(Blocked, &rdev->flags)
3048 set_bit(BlockedBadBlocks,
3050 s->blocked_rdev = rdev;
3051 atomic_inc(&rdev->nr_pending);
3054 clear_bit(R5_Insync, &dev->flags);
3058 /* also not in-sync */
3059 if (!test_bit(WriteErrorSeen, &rdev->flags)) {
3060 /* treat as in-sync, but with a read error
3061 * which we can now try to correct
3063 set_bit(R5_Insync, &dev->flags);
3064 set_bit(R5_ReadError, &dev->flags);
3066 } else if (test_bit(In_sync, &rdev->flags))
3067 set_bit(R5_Insync, &dev->flags);
3069 /* in sync if before recovery_offset */
3070 if (sh->sector + STRIPE_SECTORS <= rdev->recovery_offset)
3071 set_bit(R5_Insync, &dev->flags);
3073 if (test_bit(R5_WriteError, &dev->flags)) {
3074 clear_bit(R5_Insync, &dev->flags);
3075 if (!test_bit(Faulty, &rdev->flags)) {
3076 s->handle_bad_blocks = 1;
3077 atomic_inc(&rdev->nr_pending);
3079 clear_bit(R5_WriteError, &dev->flags);
3081 if (!test_bit(R5_Insync, &dev->flags)) {
3082 /* The ReadError flag will just be confusing now */
3083 clear_bit(R5_ReadError, &dev->flags);
3084 clear_bit(R5_ReWrite, &dev->flags);
3086 if (test_bit(R5_ReadError, &dev->flags))
3087 clear_bit(R5_Insync, &dev->flags);
3088 if (!test_bit(R5_Insync, &dev->flags)) {
3090 s->failed_num[s->failed] = i;
3094 spin_unlock_irq(&conf->device_lock);
3098 static void handle_stripe(struct stripe_head *sh)
3100 struct stripe_head_state s;
3101 raid5_conf_t *conf = sh->raid_conf;
3104 int disks = sh->disks;
3105 struct r5dev *pdev, *qdev;
3107 clear_bit(STRIPE_HANDLE, &sh->state);
3108 if (test_and_set_bit(STRIPE_ACTIVE, &sh->state)) {
3109 /* already being handled, ensure it gets handled
3110 * again when current action finishes */
3111 set_bit(STRIPE_HANDLE, &sh->state);
3115 if (test_and_clear_bit(STRIPE_SYNC_REQUESTED, &sh->state)) {
3116 set_bit(STRIPE_SYNCING, &sh->state);
3117 clear_bit(STRIPE_INSYNC, &sh->state);
3119 clear_bit(STRIPE_DELAYED, &sh->state);
3121 pr_debug("handling stripe %llu, state=%#lx cnt=%d, "
3122 "pd_idx=%d, qd_idx=%d\n, check:%d, reconstruct:%d\n",
3123 (unsigned long long)sh->sector, sh->state,
3124 atomic_read(&sh->count), sh->pd_idx, sh->qd_idx,
3125 sh->check_state, sh->reconstruct_state);
3127 analyse_stripe(sh, &s);
3129 if (s.handle_bad_blocks) {
3130 set_bit(STRIPE_HANDLE, &sh->state);
3134 if (unlikely(s.blocked_rdev)) {
3135 if (s.syncing || s.expanding || s.expanded ||
3136 s.to_write || s.written) {
3137 set_bit(STRIPE_HANDLE, &sh->state);
3140 /* There is nothing for the blocked_rdev to block */
3141 rdev_dec_pending(s.blocked_rdev, conf->mddev);
3142 s.blocked_rdev = NULL;
3145 if (s.to_fill && !test_bit(STRIPE_BIOFILL_RUN, &sh->state)) {
3146 set_bit(STRIPE_OP_BIOFILL, &s.ops_request);
3147 set_bit(STRIPE_BIOFILL_RUN, &sh->state);
3150 pr_debug("locked=%d uptodate=%d to_read=%d"
3151 " to_write=%d failed=%d failed_num=%d,%d\n",
3152 s.locked, s.uptodate, s.to_read, s.to_write, s.failed,
3153 s.failed_num[0], s.failed_num[1]);
3154 /* check if the array has lost more than max_degraded devices and,
3155 * if so, some requests might need to be failed.
3157 if (s.failed > conf->max_degraded && s.to_read+s.to_write+s.written)
3158 handle_failed_stripe(conf, sh, &s, disks, &s.return_bi);
3159 if (s.failed > conf->max_degraded && s.syncing)
3160 handle_failed_sync(conf, sh, &s);
3163 * might be able to return some write requests if the parity blocks
3164 * are safe, or on a failed drive
3166 pdev = &sh->dev[sh->pd_idx];
3167 s.p_failed = (s.failed >= 1 && s.failed_num[0] == sh->pd_idx)
3168 || (s.failed >= 2 && s.failed_num[1] == sh->pd_idx);
3169 qdev = &sh->dev[sh->qd_idx];
3170 s.q_failed = (s.failed >= 1 && s.failed_num[0] == sh->qd_idx)
3171 || (s.failed >= 2 && s.failed_num[1] == sh->qd_idx)
3175 (s.p_failed || ((test_bit(R5_Insync, &pdev->flags)
3176 && !test_bit(R5_LOCKED, &pdev->flags)
3177 && test_bit(R5_UPTODATE, &pdev->flags)))) &&
3178 (s.q_failed || ((test_bit(R5_Insync, &qdev->flags)
3179 && !test_bit(R5_LOCKED, &qdev->flags)
3180 && test_bit(R5_UPTODATE, &qdev->flags)))))
3181 handle_stripe_clean_event(conf, sh, disks, &s.return_bi);
3183 /* Now we might consider reading some blocks, either to check/generate
3184 * parity, or to satisfy requests
3185 * or to load a block that is being partially written.
3187 if (s.to_read || s.non_overwrite
3188 || (conf->level == 6 && s.to_write && s.failed)
3189 || (s.syncing && (s.uptodate + s.compute < disks)) || s.expanding)
3190 handle_stripe_fill(sh, &s, disks);
3192 /* Now we check to see if any write operations have recently
3196 if (sh->reconstruct_state == reconstruct_state_prexor_drain_result)
3198 if (sh->reconstruct_state == reconstruct_state_drain_result ||
3199 sh->reconstruct_state == reconstruct_state_prexor_drain_result) {
3200 sh->reconstruct_state = reconstruct_state_idle;
3202 /* All the 'written' buffers and the parity block are ready to
3203 * be written back to disk
3205 BUG_ON(!test_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags));
3206 BUG_ON(sh->qd_idx >= 0 &&
3207 !test_bit(R5_UPTODATE, &sh->dev[sh->qd_idx].flags));
3208 for (i = disks; i--; ) {
3209 struct r5dev *dev = &sh->dev[i];
3210 if (test_bit(R5_LOCKED, &dev->flags) &&
3211 (i == sh->pd_idx || i == sh->qd_idx ||
3213 pr_debug("Writing block %d\n", i);
3214 set_bit(R5_Wantwrite, &dev->flags);
3217 if (!test_bit(R5_Insync, &dev->flags) ||
3218 ((i == sh->pd_idx || i == sh->qd_idx) &&
3220 set_bit(STRIPE_INSYNC, &sh->state);
3223 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
3224 s.dec_preread_active = 1;
3227 /* Now to consider new write requests and what else, if anything
3228 * should be read. We do not handle new writes when:
3229 * 1/ A 'write' operation (copy+xor) is already in flight.
3230 * 2/ A 'check' operation is in flight, as it may clobber the parity
3233 if (s.to_write && !sh->reconstruct_state && !sh->check_state)
3234 handle_stripe_dirtying(conf, sh, &s, disks);
3236 /* maybe we need to check and possibly fix the parity for this stripe
3237 * Any reads will already have been scheduled, so we just see if enough
3238 * data is available. The parity check is held off while parity
3239 * dependent operations are in flight.
3241 if (sh->check_state ||
3242 (s.syncing && s.locked == 0 &&
3243 !test_bit(STRIPE_COMPUTE_RUN, &sh->state) &&
3244 !test_bit(STRIPE_INSYNC, &sh->state))) {
3245 if (conf->level == 6)
3246 handle_parity_checks6(conf, sh, &s, disks);
3248 handle_parity_checks5(conf, sh, &s, disks);
3251 if (s.syncing && s.locked == 0 && test_bit(STRIPE_INSYNC, &sh->state)) {
3252 md_done_sync(conf->mddev, STRIPE_SECTORS, 1);
3253 clear_bit(STRIPE_SYNCING, &sh->state);
3256 /* If the failed drives are just a ReadError, then we might need
3257 * to progress the repair/check process
3259 if (s.failed <= conf->max_degraded && !conf->mddev->ro)
3260 for (i = 0; i < s.failed; i++) {
3261 struct r5dev *dev = &sh->dev[s.failed_num[i]];
3262 if (test_bit(R5_ReadError, &dev->flags)
3263 && !test_bit(R5_LOCKED, &dev->flags)
3264 && test_bit(R5_UPTODATE, &dev->flags)
3266 if (!test_bit(R5_ReWrite, &dev->flags)) {
3267 set_bit(R5_Wantwrite, &dev->flags);
3268 set_bit(R5_ReWrite, &dev->flags);
3269 set_bit(R5_LOCKED, &dev->flags);
3272 /* let's read it back */
3273 set_bit(R5_Wantread, &dev->flags);
3274 set_bit(R5_LOCKED, &dev->flags);
3281 /* Finish reconstruct operations initiated by the expansion process */
3282 if (sh->reconstruct_state == reconstruct_state_result) {
3283 struct stripe_head *sh_src
3284 = get_active_stripe(conf, sh->sector, 1, 1, 1);
3285 if (sh_src && test_bit(STRIPE_EXPAND_SOURCE, &sh_src->state)) {
3286 /* sh cannot be written until sh_src has been read.
3287 * so arrange for sh to be delayed a little
3289 set_bit(STRIPE_DELAYED, &sh->state);
3290 set_bit(STRIPE_HANDLE, &sh->state);
3291 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE,
3293 atomic_inc(&conf->preread_active_stripes);
3294 release_stripe(sh_src);
3298 release_stripe(sh_src);
3300 sh->reconstruct_state = reconstruct_state_idle;
3301 clear_bit(STRIPE_EXPANDING, &sh->state);
3302 for (i = conf->raid_disks; i--; ) {
3303 set_bit(R5_Wantwrite, &sh->dev[i].flags);
3304 set_bit(R5_LOCKED, &sh->dev[i].flags);
3309 if (s.expanded && test_bit(STRIPE_EXPANDING, &sh->state) &&
3310 !sh->reconstruct_state) {
3311 /* Need to write out all blocks after computing parity */
3312 sh->disks = conf->raid_disks;
3313 stripe_set_idx(sh->sector, conf, 0, sh);
3314 schedule_reconstruction(sh, &s, 1, 1);
3315 } else if (s.expanded && !sh->reconstruct_state && s.locked == 0) {
3316 clear_bit(STRIPE_EXPAND_READY, &sh->state);
3317 atomic_dec(&conf->reshape_stripes);
3318 wake_up(&conf->wait_for_overlap);
3319 md_done_sync(conf->mddev, STRIPE_SECTORS, 1);
3322 if (s.expanding && s.locked == 0 &&
3323 !test_bit(STRIPE_COMPUTE_RUN, &sh->state))
3324 handle_stripe_expansion(conf, sh);
3327 /* wait for this device to become unblocked */
3328 if (unlikely(s.blocked_rdev))
3329 md_wait_for_blocked_rdev(s.blocked_rdev, conf->mddev);
3331 if (s.handle_bad_blocks)
3332 for (i = disks; i--; ) {
3334 struct r5dev *dev = &sh->dev[i];
3335 if (test_and_clear_bit(R5_WriteError, &dev->flags)) {
3336 /* We own a safe reference to the rdev */
3337 rdev = conf->disks[i].rdev;
3338 if (!rdev_set_badblocks(rdev, sh->sector,
3340 md_error(conf->mddev, rdev);
3341 rdev_dec_pending(rdev, conf->mddev);
3346 raid_run_ops(sh, s.ops_request);
3350 if (s.dec_preread_active) {
3351 /* We delay this until after ops_run_io so that if make_request
3352 * is waiting on a flush, it won't continue until the writes
3353 * have actually been submitted.
3355 atomic_dec(&conf->preread_active_stripes);
3356 if (atomic_read(&conf->preread_active_stripes) <
3358 md_wakeup_thread(conf->mddev->thread);
3361 return_io(s.return_bi);
3363 clear_bit(STRIPE_ACTIVE, &sh->state);
3366 static void raid5_activate_delayed(raid5_conf_t *conf)
3368 if (atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD) {
3369 while (!list_empty(&conf->delayed_list)) {
3370 struct list_head *l = conf->delayed_list.next;
3371 struct stripe_head *sh;
3372 sh = list_entry(l, struct stripe_head, lru);
3374 clear_bit(STRIPE_DELAYED, &sh->state);
3375 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
3376 atomic_inc(&conf->preread_active_stripes);
3377 list_add_tail(&sh->lru, &conf->hold_list);
3382 static void activate_bit_delay(raid5_conf_t *conf)
3384 /* device_lock is held */
3385 struct list_head head;
3386 list_add(&head, &conf->bitmap_list);
3387 list_del_init(&conf->bitmap_list);
3388 while (!list_empty(&head)) {
3389 struct stripe_head *sh = list_entry(head.next, struct stripe_head, lru);
3390 list_del_init(&sh->lru);
3391 atomic_inc(&sh->count);
3392 __release_stripe(conf, sh);
3396 int md_raid5_congested(mddev_t *mddev, int bits)
3398 raid5_conf_t *conf = mddev->private;
3400 /* No difference between reads and writes. Just check
3401 * how busy the stripe_cache is
3404 if (conf->inactive_blocked)
3408 if (list_empty_careful(&conf->inactive_list))
3413 EXPORT_SYMBOL_GPL(md_raid5_congested);
3415 static int raid5_congested(void *data, int bits)
3417 mddev_t *mddev = data;
3419 return mddev_congested(mddev, bits) ||
3420 md_raid5_congested(mddev, bits);
3423 /* We want read requests to align with chunks where possible,
3424 * but write requests don't need to.
3426 static int raid5_mergeable_bvec(struct request_queue *q,
3427 struct bvec_merge_data *bvm,
3428 struct bio_vec *biovec)
3430 mddev_t *mddev = q->queuedata;
3431 sector_t sector = bvm->bi_sector + get_start_sect(bvm->bi_bdev);
3433 unsigned int chunk_sectors = mddev->chunk_sectors;
3434 unsigned int bio_sectors = bvm->bi_size >> 9;
3436 if ((bvm->bi_rw & 1) == WRITE)
3437 return biovec->bv_len; /* always allow writes to be mergeable */
3439 if (mddev->new_chunk_sectors < mddev->chunk_sectors)
3440 chunk_sectors = mddev->new_chunk_sectors;
3441 max = (chunk_sectors - ((sector & (chunk_sectors - 1)) + bio_sectors)) << 9;
3442 if (max < 0) max = 0;
3443 if (max <= biovec->bv_len && bio_sectors == 0)
3444 return biovec->bv_len;
3450 static int in_chunk_boundary(mddev_t *mddev, struct bio *bio)
3452 sector_t sector = bio->bi_sector + get_start_sect(bio->bi_bdev);
3453 unsigned int chunk_sectors = mddev->chunk_sectors;
3454 unsigned int bio_sectors = bio->bi_size >> 9;
3456 if (mddev->new_chunk_sectors < mddev->chunk_sectors)
3457 chunk_sectors = mddev->new_chunk_sectors;
3458 return chunk_sectors >=
3459 ((sector & (chunk_sectors - 1)) + bio_sectors);
3463 * add bio to the retry LIFO ( in O(1) ... we are in interrupt )
3464 * later sampled by raid5d.
3466 static void add_bio_to_retry(struct bio *bi,raid5_conf_t *conf)
3468 unsigned long flags;
3470 spin_lock_irqsave(&conf->device_lock, flags);
3472 bi->bi_next = conf->retry_read_aligned_list;
3473 conf->retry_read_aligned_list = bi;
3475 spin_unlock_irqrestore(&conf->device_lock, flags);
3476 md_wakeup_thread(conf->mddev->thread);
3480 static struct bio *remove_bio_from_retry(raid5_conf_t *conf)
3484 bi = conf->retry_read_aligned;
3486 conf->retry_read_aligned = NULL;
3489 bi = conf->retry_read_aligned_list;
3491 conf->retry_read_aligned_list = bi->bi_next;
3494 * this sets the active strip count to 1 and the processed
3495 * strip count to zero (upper 8 bits)
3497 bi->bi_phys_segments = 1; /* biased count of active stripes */
3505 * The "raid5_align_endio" should check if the read succeeded and if it
3506 * did, call bio_endio on the original bio (having bio_put the new bio
3508 * If the read failed..
3510 static void raid5_align_endio(struct bio *bi, int error)
3512 struct bio* raid_bi = bi->bi_private;
3515 int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
3520 rdev = (void*)raid_bi->bi_next;
3521 raid_bi->bi_next = NULL;
3522 mddev = rdev->mddev;
3523 conf = mddev->private;
3525 rdev_dec_pending(rdev, conf->mddev);
3527 if (!error && uptodate) {
3528 bio_endio(raid_bi, 0);
3529 if (atomic_dec_and_test(&conf->active_aligned_reads))
3530 wake_up(&conf->wait_for_stripe);
3535 pr_debug("raid5_align_endio : io error...handing IO for a retry\n");
3537 add_bio_to_retry(raid_bi, conf);
3540 static int bio_fits_rdev(struct bio *bi)
3542 struct request_queue *q = bdev_get_queue(bi->bi_bdev);
3544 if ((bi->bi_size>>9) > queue_max_sectors(q))
3546 blk_recount_segments(q, bi);
3547 if (bi->bi_phys_segments > queue_max_segments(q))
3550 if (q->merge_bvec_fn)
3551 /* it's too hard to apply the merge_bvec_fn at this stage,
3560 static int chunk_aligned_read(mddev_t *mddev, struct bio * raid_bio)
3562 raid5_conf_t *conf = mddev->private;
3564 struct bio* align_bi;
3567 if (!in_chunk_boundary(mddev, raid_bio)) {
3568 pr_debug("chunk_aligned_read : non aligned\n");
3572 * use bio_clone_mddev to make a copy of the bio
3574 align_bi = bio_clone_mddev(raid_bio, GFP_NOIO, mddev);
3578 * set bi_end_io to a new function, and set bi_private to the
3581 align_bi->bi_end_io = raid5_align_endio;
3582 align_bi->bi_private = raid_bio;
3586 align_bi->bi_sector = raid5_compute_sector(conf, raid_bio->bi_sector,
3591 rdev = rcu_dereference(conf->disks[dd_idx].rdev);
3592 if (rdev && test_bit(In_sync, &rdev->flags)) {
3596 atomic_inc(&rdev->nr_pending);
3598 raid_bio->bi_next = (void*)rdev;
3599 align_bi->bi_bdev = rdev->bdev;
3600 align_bi->bi_flags &= ~(1 << BIO_SEG_VALID);
3601 align_bi->bi_sector += rdev->data_offset;
3603 if (!bio_fits_rdev(align_bi) ||
3604 is_badblock(rdev, align_bi->bi_sector, align_bi->bi_size>>9,
3605 &first_bad, &bad_sectors)) {
3606 /* too big in some way, or has a known bad block */
3608 rdev_dec_pending(rdev, mddev);
3612 spin_lock_irq(&conf->device_lock);
3613 wait_event_lock_irq(conf->wait_for_stripe,
3615 conf->device_lock, /* nothing */);
3616 atomic_inc(&conf->active_aligned_reads);
3617 spin_unlock_irq(&conf->device_lock);
3619 generic_make_request(align_bi);
3628 /* __get_priority_stripe - get the next stripe to process
3630 * Full stripe writes are allowed to pass preread active stripes up until
3631 * the bypass_threshold is exceeded. In general the bypass_count
3632 * increments when the handle_list is handled before the hold_list; however, it
3633 * will not be incremented when STRIPE_IO_STARTED is sampled set signifying a
3634 * stripe with in flight i/o. The bypass_count will be reset when the
3635 * head of the hold_list has changed, i.e. the head was promoted to the
3638 static struct stripe_head *__get_priority_stripe(raid5_conf_t *conf)
3640 struct stripe_head *sh;
3642 pr_debug("%s: handle: %s hold: %s full_writes: %d bypass_count: %d\n",
3644 list_empty(&conf->handle_list) ? "empty" : "busy",
3645 list_empty(&conf->hold_list) ? "empty" : "busy",
3646 atomic_read(&conf->pending_full_writes), conf->bypass_count);
3648 if (!list_empty(&conf->handle_list)) {
3649 sh = list_entry(conf->handle_list.next, typeof(*sh), lru);
3651 if (list_empty(&conf->hold_list))
3652 conf->bypass_count = 0;
3653 else if (!test_bit(STRIPE_IO_STARTED, &sh->state)) {
3654 if (conf->hold_list.next == conf->last_hold)
3655 conf->bypass_count++;
3657 conf->last_hold = conf->hold_list.next;
3658 conf->bypass_count -= conf->bypass_threshold;
3659 if (conf->bypass_count < 0)
3660 conf->bypass_count = 0;
3663 } else if (!list_empty(&conf->hold_list) &&
3664 ((conf->bypass_threshold &&
3665 conf->bypass_count > conf->bypass_threshold) ||
3666 atomic_read(&conf->pending_full_writes) == 0)) {
3667 sh = list_entry(conf->hold_list.next,
3669 conf->bypass_count -= conf->bypass_threshold;
3670 if (conf->bypass_count < 0)
3671 conf->bypass_count = 0;
3675 list_del_init(&sh->lru);
3676 atomic_inc(&sh->count);
3677 BUG_ON(atomic_read(&sh->count) != 1);
3681 static int make_request(mddev_t *mddev, struct bio * bi)
3683 raid5_conf_t *conf = mddev->private;
3685 sector_t new_sector;
3686 sector_t logical_sector, last_sector;
3687 struct stripe_head *sh;
3688 const int rw = bio_data_dir(bi);
3692 if (unlikely(bi->bi_rw & REQ_FLUSH)) {
3693 md_flush_request(mddev, bi);
3697 md_write_start(mddev, bi);
3700 mddev->reshape_position == MaxSector &&
3701 chunk_aligned_read(mddev,bi))
3704 logical_sector = bi->bi_sector & ~((sector_t)STRIPE_SECTORS-1);
3705 last_sector = bi->bi_sector + (bi->bi_size>>9);
3707 bi->bi_phys_segments = 1; /* over-loaded to count active stripes */
3709 plugged = mddev_check_plugged(mddev);
3710 for (;logical_sector < last_sector; logical_sector += STRIPE_SECTORS) {
3712 int disks, data_disks;
3717 disks = conf->raid_disks;
3718 prepare_to_wait(&conf->wait_for_overlap, &w, TASK_UNINTERRUPTIBLE);
3719 if (unlikely(conf->reshape_progress != MaxSector)) {
3720 /* spinlock is needed as reshape_progress may be
3721 * 64bit on a 32bit platform, and so it might be
3722 * possible to see a half-updated value
3723 * Of course reshape_progress could change after
3724 * the lock is dropped, so once we get a reference
3725 * to the stripe that we think it is, we will have
3728 spin_lock_irq(&conf->device_lock);
3729 if (mddev->delta_disks < 0
3730 ? logical_sector < conf->reshape_progress
3731 : logical_sector >= conf->reshape_progress) {
3732 disks = conf->previous_raid_disks;
3735 if (mddev->delta_disks < 0
3736 ? logical_sector < conf->reshape_safe
3737 : logical_sector >= conf->reshape_safe) {
3738 spin_unlock_irq(&conf->device_lock);
3743 spin_unlock_irq(&conf->device_lock);
3745 data_disks = disks - conf->max_degraded;
3747 new_sector = raid5_compute_sector(conf, logical_sector,
3750 pr_debug("raid456: make_request, sector %llu logical %llu\n",
3751 (unsigned long long)new_sector,
3752 (unsigned long long)logical_sector);
3754 sh = get_active_stripe(conf, new_sector, previous,
3755 (bi->bi_rw&RWA_MASK), 0);
3757 if (unlikely(previous)) {
3758 /* expansion might have moved on while waiting for a
3759 * stripe, so we must do the range check again.
3760 * Expansion could still move past after this
3761 * test, but as we are holding a reference to
3762 * 'sh', we know that if that happens,
3763 * STRIPE_EXPANDING will get set and the expansion
3764 * won't proceed until we finish with the stripe.
3767 spin_lock_irq(&conf->device_lock);
3768 if (mddev->delta_disks < 0
3769 ? logical_sector >= conf->reshape_progress
3770 : logical_sector < conf->reshape_progress)
3771 /* mismatch, need to try again */
3773 spin_unlock_irq(&conf->device_lock);
3782 logical_sector >= mddev->suspend_lo &&
3783 logical_sector < mddev->suspend_hi) {
3785 /* As the suspend_* range is controlled by
3786 * userspace, we want an interruptible
3789 flush_signals(current);
3790 prepare_to_wait(&conf->wait_for_overlap,
3791 &w, TASK_INTERRUPTIBLE);
3792 if (logical_sector >= mddev->suspend_lo &&
3793 logical_sector < mddev->suspend_hi)
3798 if (test_bit(STRIPE_EXPANDING, &sh->state) ||
3799 !add_stripe_bio(sh, bi, dd_idx, rw)) {
3800 /* Stripe is busy expanding or
3801 * add failed due to overlap. Flush everything
3804 md_wakeup_thread(mddev->thread);
3809 finish_wait(&conf->wait_for_overlap, &w);
3810 set_bit(STRIPE_HANDLE, &sh->state);
3811 clear_bit(STRIPE_DELAYED, &sh->state);
3812 if ((bi->bi_rw & REQ_SYNC) &&
3813 !test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
3814 atomic_inc(&conf->preread_active_stripes);
3817 /* cannot get stripe for read-ahead, just give-up */
3818 clear_bit(BIO_UPTODATE, &bi->bi_flags);
3819 finish_wait(&conf->wait_for_overlap, &w);
3825 md_wakeup_thread(mddev->thread);
3827 spin_lock_irq(&conf->device_lock);
3828 remaining = raid5_dec_bi_phys_segments(bi);
3829 spin_unlock_irq(&conf->device_lock);
3830 if (remaining == 0) {
3833 md_write_end(mddev);
3841 static sector_t raid5_size(mddev_t *mddev, sector_t sectors, int raid_disks);
3843 static sector_t reshape_request(mddev_t *mddev, sector_t sector_nr, int *skipped)
3845 /* reshaping is quite different to recovery/resync so it is
3846 * handled quite separately ... here.
3848 * On each call to sync_request, we gather one chunk worth of
3849 * destination stripes and flag them as expanding.
3850 * Then we find all the source stripes and request reads.
3851 * As the reads complete, handle_stripe will copy the data
3852 * into the destination stripe and release that stripe.
3854 raid5_conf_t *conf = mddev->private;
3855 struct stripe_head *sh;
3856 sector_t first_sector, last_sector;
3857 int raid_disks = conf->previous_raid_disks;
3858 int data_disks = raid_disks - conf->max_degraded;
3859 int new_data_disks = conf->raid_disks - conf->max_degraded;
3862 sector_t writepos, readpos, safepos;
3863 sector_t stripe_addr;
3864 int reshape_sectors;
3865 struct list_head stripes;
3867 if (sector_nr == 0) {
3868 /* If restarting in the middle, skip the initial sectors */
3869 if (mddev->delta_disks < 0 &&
3870 conf->reshape_progress < raid5_size(mddev, 0, 0)) {
3871 sector_nr = raid5_size(mddev, 0, 0)
3872 - conf->reshape_progress;
3873 } else if (mddev->delta_disks >= 0 &&
3874 conf->reshape_progress > 0)
3875 sector_nr = conf->reshape_progress;
3876 sector_div(sector_nr, new_data_disks);
3878 mddev->curr_resync_completed = sector_nr;
3879 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
3885 /* We need to process a full chunk at a time.
3886 * If old and new chunk sizes differ, we need to process the
3889 if (mddev->new_chunk_sectors > mddev->chunk_sectors)
3890 reshape_sectors = mddev->new_chunk_sectors;
3892 reshape_sectors = mddev->chunk_sectors;
3894 /* we update the metadata when there is more than 3Meg
3895 * in the block range (that is rather arbitrary, should
3896 * probably be time based) or when the data about to be
3897 * copied would over-write the source of the data at
3898 * the front of the range.
3899 * i.e. one new_stripe along from reshape_progress new_maps
3900 * to after where reshape_safe old_maps to
3902 writepos = conf->reshape_progress;
3903 sector_div(writepos, new_data_disks);
3904 readpos = conf->reshape_progress;
3905 sector_div(readpos, data_disks);
3906 safepos = conf->reshape_safe;
3907 sector_div(safepos, data_disks);
3908 if (mddev->delta_disks < 0) {
3909 writepos -= min_t(sector_t, reshape_sectors, writepos);
3910 readpos += reshape_sectors;
3911 safepos += reshape_sectors;
3913 writepos += reshape_sectors;
3914 readpos -= min_t(sector_t, reshape_sectors, readpos);
3915 safepos -= min_t(sector_t, reshape_sectors, safepos);
3918 /* 'writepos' is the most advanced device address we might write.
3919 * 'readpos' is the least advanced device address we might read.
3920 * 'safepos' is the least address recorded in the metadata as having
3922 * If 'readpos' is behind 'writepos', then there is no way that we can
3923 * ensure safety in the face of a crash - that must be done by userspace
3924 * making a backup of the data. So in that case there is no particular
3925 * rush to update metadata.
3926 * Otherwise if 'safepos' is behind 'writepos', then we really need to
3927 * update the metadata to advance 'safepos' to match 'readpos' so that
3928 * we can be safe in the event of a crash.
3929 * So we insist on updating metadata if safepos is behind writepos and
3930 * readpos is beyond writepos.
3931 * In any case, update the metadata every 10 seconds.
3932 * Maybe that number should be configurable, but I'm not sure it is
3933 * worth it.... maybe it could be a multiple of safemode_delay???
3935 if ((mddev->delta_disks < 0
3936 ? (safepos > writepos && readpos < writepos)
3937 : (safepos < writepos && readpos > writepos)) ||
3938 time_after(jiffies, conf->reshape_checkpoint + 10*HZ)) {
3939 /* Cannot proceed until we've updated the superblock... */
3940 wait_event(conf->wait_for_overlap,
3941 atomic_read(&conf->reshape_stripes)==0);
3942 mddev->reshape_position = conf->reshape_progress;
3943 mddev->curr_resync_completed = sector_nr;
3944 conf->reshape_checkpoint = jiffies;
3945 set_bit(MD_CHANGE_DEVS, &mddev->flags);
3946 md_wakeup_thread(mddev->thread);
3947 wait_event(mddev->sb_wait, mddev->flags == 0 ||
3948 kthread_should_stop());
3949 spin_lock_irq(&conf->device_lock);
3950 conf->reshape_safe = mddev->reshape_position;
3951 spin_unlock_irq(&conf->device_lock);
3952 wake_up(&conf->wait_for_overlap);
3953 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
3956 if (mddev->delta_disks < 0) {
3957 BUG_ON(conf->reshape_progress == 0);
3958 stripe_addr = writepos;
3959 BUG_ON((mddev->dev_sectors &
3960 ~((sector_t)reshape_sectors - 1))
3961 - reshape_sectors - stripe_addr
3964 BUG_ON(writepos != sector_nr + reshape_sectors);
3965 stripe_addr = sector_nr;
3967 INIT_LIST_HEAD(&stripes);
3968 for (i = 0; i < reshape_sectors; i += STRIPE_SECTORS) {
3970 int skipped_disk = 0;
3971 sh = get_active_stripe(conf, stripe_addr+i, 0, 0, 1);
3972 set_bit(STRIPE_EXPANDING, &sh->state);
3973 atomic_inc(&conf->reshape_stripes);
3974 /* If any of this stripe is beyond the end of the old
3975 * array, then we need to zero those blocks
3977 for (j=sh->disks; j--;) {
3979 if (j == sh->pd_idx)
3981 if (conf->level == 6 &&
3984 s = compute_blocknr(sh, j, 0);
3985 if (s < raid5_size(mddev, 0, 0)) {
3989 memset(page_address(sh->dev[j].page), 0, STRIPE_SIZE);
3990 set_bit(R5_Expanded, &sh->dev[j].flags);
3991 set_bit(R5_UPTODATE, &sh->dev[j].flags);
3993 if (!skipped_disk) {
3994 set_bit(STRIPE_EXPAND_READY, &sh->state);
3995 set_bit(STRIPE_HANDLE, &sh->state);
3997 list_add(&sh->lru, &stripes);
3999 spin_lock_irq(&conf->device_lock);
4000 if (mddev->delta_disks < 0)
4001 conf->reshape_progress -= reshape_sectors * new_data_disks;
4003 conf->reshape_progress += reshape_sectors * new_data_disks;
4004 spin_unlock_irq(&conf->device_lock);
4005 /* Ok, those stripe are ready. We can start scheduling
4006 * reads on the source stripes.
4007 * The source stripes are determined by mapping the first and last
4008 * block on the destination stripes.
4011 raid5_compute_sector(conf, stripe_addr*(new_data_disks),
4014 raid5_compute_sector(conf, ((stripe_addr+reshape_sectors)
4015 * new_data_disks - 1),
4017 if (last_sector >= mddev->dev_sectors)
4018 last_sector = mddev->dev_sectors - 1;
4019 while (first_sector <= last_sector) {
4020 sh = get_active_stripe(conf, first_sector, 1, 0, 1);
4021 set_bit(STRIPE_EXPAND_SOURCE, &sh->state);
4022 set_bit(STRIPE_HANDLE, &sh->state);
4024 first_sector += STRIPE_SECTORS;
4026 /* Now that the sources are clearly marked, we can release
4027 * the destination stripes
4029 while (!list_empty(&stripes)) {
4030 sh = list_entry(stripes.next, struct stripe_head, lru);
4031 list_del_init(&sh->lru);
4034 /* If this takes us to the resync_max point where we have to pause,
4035 * then we need to write out the superblock.
4037 sector_nr += reshape_sectors;
4038 if ((sector_nr - mddev->curr_resync_completed) * 2
4039 >= mddev->resync_max - mddev->curr_resync_completed) {
4040 /* Cannot proceed until we've updated the superblock... */
4041 wait_event(conf->wait_for_overlap,
4042 atomic_read(&conf->reshape_stripes) == 0);
4043 mddev->reshape_position = conf->reshape_progress;
4044 mddev->curr_resync_completed = sector_nr;
4045 conf->reshape_checkpoint = jiffies;
4046 set_bit(MD_CHANGE_DEVS, &mddev->flags);
4047 md_wakeup_thread(mddev->thread);
4048 wait_event(mddev->sb_wait,
4049 !test_bit(MD_CHANGE_DEVS, &mddev->flags)
4050 || kthread_should_stop());
4051 spin_lock_irq(&conf->device_lock);
4052 conf->reshape_safe = mddev->reshape_position;
4053 spin_unlock_irq(&conf->device_lock);
4054 wake_up(&conf->wait_for_overlap);
4055 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
4057 return reshape_sectors;
4060 /* FIXME go_faster isn't used */
4061 static inline sector_t sync_request(mddev_t *mddev, sector_t sector_nr, int *skipped, int go_faster)
4063 raid5_conf_t *conf = mddev->private;
4064 struct stripe_head *sh;
4065 sector_t max_sector = mddev->dev_sectors;
4066 sector_t sync_blocks;
4067 int still_degraded = 0;
4070 if (sector_nr >= max_sector) {
4071 /* just being told to finish up .. nothing much to do */
4073 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery)) {
4078 if (mddev->curr_resync < max_sector) /* aborted */
4079 bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
4081 else /* completed sync */
4083 bitmap_close_sync(mddev->bitmap);
4088 /* Allow raid5_quiesce to complete */
4089 wait_event(conf->wait_for_overlap, conf->quiesce != 2);
4091 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
4092 return reshape_request(mddev, sector_nr, skipped);
4094 /* No need to check resync_max as we never do more than one
4095 * stripe, and as resync_max will always be on a chunk boundary,
4096 * if the check in md_do_sync didn't fire, there is no chance
4097 * of overstepping resync_max here
4100 /* if there is too many failed drives and we are trying
4101 * to resync, then assert that we are finished, because there is
4102 * nothing we can do.
4104 if (mddev->degraded >= conf->max_degraded &&
4105 test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
4106 sector_t rv = mddev->dev_sectors - sector_nr;
4110 if (!bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
4111 !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
4112 !conf->fullsync && sync_blocks >= STRIPE_SECTORS) {
4113 /* we can skip this block, and probably more */
4114 sync_blocks /= STRIPE_SECTORS;
4116 return sync_blocks * STRIPE_SECTORS; /* keep things rounded to whole stripes */
4120 bitmap_cond_end_sync(mddev->bitmap, sector_nr);
4122 sh = get_active_stripe(conf, sector_nr, 0, 1, 0);
4124 sh = get_active_stripe(conf, sector_nr, 0, 0, 0);
4125 /* make sure we don't swamp the stripe cache if someone else
4126 * is trying to get access
4128 schedule_timeout_uninterruptible(1);
4130 /* Need to check if array will still be degraded after recovery/resync
4131 * We don't need to check the 'failed' flag as when that gets set,
4134 for (i = 0; i < conf->raid_disks; i++)
4135 if (conf->disks[i].rdev == NULL)
4138 bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, still_degraded);
4140 set_bit(STRIPE_SYNC_REQUESTED, &sh->state);
4145 return STRIPE_SECTORS;
4148 static int retry_aligned_read(raid5_conf_t *conf, struct bio *raid_bio)
4150 /* We may not be able to submit a whole bio at once as there
4151 * may not be enough stripe_heads available.
4152 * We cannot pre-allocate enough stripe_heads as we may need
4153 * more than exist in the cache (if we allow ever large chunks).
4154 * So we do one stripe head at a time and record in
4155 * ->bi_hw_segments how many have been done.
4157 * We *know* that this entire raid_bio is in one chunk, so
4158 * it will be only one 'dd_idx' and only need one call to raid5_compute_sector.
4160 struct stripe_head *sh;
4162 sector_t sector, logical_sector, last_sector;
4167 logical_sector = raid_bio->bi_sector & ~((sector_t)STRIPE_SECTORS-1);
4168 sector = raid5_compute_sector(conf, logical_sector,
4170 last_sector = raid_bio->bi_sector + (raid_bio->bi_size>>9);
4172 for (; logical_sector < last_sector;
4173 logical_sector += STRIPE_SECTORS,
4174 sector += STRIPE_SECTORS,
4177 if (scnt < raid5_bi_hw_segments(raid_bio))
4178 /* already done this stripe */
4181 sh = get_active_stripe(conf, sector, 0, 1, 0);
4184 /* failed to get a stripe - must wait */
4185 raid5_set_bi_hw_segments(raid_bio, scnt);
4186 conf->retry_read_aligned = raid_bio;
4190 set_bit(R5_ReadError, &sh->dev[dd_idx].flags);
4191 if (!add_stripe_bio(sh, raid_bio, dd_idx, 0)) {
4193 raid5_set_bi_hw_segments(raid_bio, scnt);
4194 conf->retry_read_aligned = raid_bio;
4202 spin_lock_irq(&conf->device_lock);
4203 remaining = raid5_dec_bi_phys_segments(raid_bio);
4204 spin_unlock_irq(&conf->device_lock);
4206 bio_endio(raid_bio, 0);
4207 if (atomic_dec_and_test(&conf->active_aligned_reads))
4208 wake_up(&conf->wait_for_stripe);
4214 * This is our raid5 kernel thread.
4216 * We scan the hash table for stripes which can be handled now.
4217 * During the scan, completed stripes are saved for us by the interrupt
4218 * handler, so that they will not have to wait for our next wakeup.
4220 static void raid5d(mddev_t *mddev)
4222 struct stripe_head *sh;
4223 raid5_conf_t *conf = mddev->private;
4225 struct blk_plug plug;
4227 pr_debug("+++ raid5d active\n");
4229 md_check_recovery(mddev);
4231 blk_start_plug(&plug);
4233 spin_lock_irq(&conf->device_lock);
4237 if (atomic_read(&mddev->plug_cnt) == 0 &&
4238 !list_empty(&conf->bitmap_list)) {
4239 /* Now is a good time to flush some bitmap updates */
4241 spin_unlock_irq(&conf->device_lock);
4242 bitmap_unplug(mddev->bitmap);
4243 spin_lock_irq(&conf->device_lock);
4244 conf->seq_write = conf->seq_flush;
4245 activate_bit_delay(conf);
4247 if (atomic_read(&mddev->plug_cnt) == 0)
4248 raid5_activate_delayed(conf);
4250 while ((bio = remove_bio_from_retry(conf))) {
4252 spin_unlock_irq(&conf->device_lock);
4253 ok = retry_aligned_read(conf, bio);
4254 spin_lock_irq(&conf->device_lock);
4260 sh = __get_priority_stripe(conf);
4264 spin_unlock_irq(&conf->device_lock);
4271 if (mddev->flags & ~(1<<MD_CHANGE_PENDING))
4272 md_check_recovery(mddev);
4274 spin_lock_irq(&conf->device_lock);
4276 pr_debug("%d stripes handled\n", handled);
4278 spin_unlock_irq(&conf->device_lock);
4280 async_tx_issue_pending_all();
4281 blk_finish_plug(&plug);
4283 pr_debug("--- raid5d inactive\n");
4287 raid5_show_stripe_cache_size(mddev_t *mddev, char *page)
4289 raid5_conf_t *conf = mddev->private;
4291 return sprintf(page, "%d\n", conf->max_nr_stripes);
4297 raid5_set_cache_size(mddev_t *mddev, int size)
4299 raid5_conf_t *conf = mddev->private;
4302 if (size <= 16 || size > 32768)
4304 while (size < conf->max_nr_stripes) {
4305 if (drop_one_stripe(conf))
4306 conf->max_nr_stripes--;
4310 err = md_allow_write(mddev);
4313 while (size > conf->max_nr_stripes) {
4314 if (grow_one_stripe(conf))
4315 conf->max_nr_stripes++;
4320 EXPORT_SYMBOL(raid5_set_cache_size);
4323 raid5_store_stripe_cache_size(mddev_t *mddev, const char *page, size_t len)
4325 raid5_conf_t *conf = mddev->private;
4329 if (len >= PAGE_SIZE)
4334 if (strict_strtoul(page, 10, &new))
4336 err = raid5_set_cache_size(mddev, new);
4342 static struct md_sysfs_entry
4343 raid5_stripecache_size = __ATTR(stripe_cache_size, S_IRUGO | S_IWUSR,
4344 raid5_show_stripe_cache_size,
4345 raid5_store_stripe_cache_size);
4348 raid5_show_preread_threshold(mddev_t *mddev, char *page)
4350 raid5_conf_t *conf = mddev->private;
4352 return sprintf(page, "%d\n", conf->bypass_threshold);
4358 raid5_store_preread_threshold(mddev_t *mddev, const char *page, size_t len)
4360 raid5_conf_t *conf = mddev->private;
4362 if (len >= PAGE_SIZE)
4367 if (strict_strtoul(page, 10, &new))
4369 if (new > conf->max_nr_stripes)
4371 conf->bypass_threshold = new;
4375 static struct md_sysfs_entry
4376 raid5_preread_bypass_threshold = __ATTR(preread_bypass_threshold,
4378 raid5_show_preread_threshold,
4379 raid5_store_preread_threshold);
4382 stripe_cache_active_show(mddev_t *mddev, char *page)
4384 raid5_conf_t *conf = mddev->private;
4386 return sprintf(page, "%d\n", atomic_read(&conf->active_stripes));
4391 static struct md_sysfs_entry
4392 raid5_stripecache_active = __ATTR_RO(stripe_cache_active);
4394 static struct attribute *raid5_attrs[] = {
4395 &raid5_stripecache_size.attr,
4396 &raid5_stripecache_active.attr,
4397 &raid5_preread_bypass_threshold.attr,
4400 static struct attribute_group raid5_attrs_group = {
4402 .attrs = raid5_attrs,
4406 raid5_size(mddev_t *mddev, sector_t sectors, int raid_disks)
4408 raid5_conf_t *conf = mddev->private;
4411 sectors = mddev->dev_sectors;
4413 /* size is defined by the smallest of previous and new size */
4414 raid_disks = min(conf->raid_disks, conf->previous_raid_disks);
4416 sectors &= ~((sector_t)mddev->chunk_sectors - 1);
4417 sectors &= ~((sector_t)mddev->new_chunk_sectors - 1);
4418 return sectors * (raid_disks - conf->max_degraded);
4421 static void raid5_free_percpu(raid5_conf_t *conf)
4423 struct raid5_percpu *percpu;
4430 for_each_possible_cpu(cpu) {
4431 percpu = per_cpu_ptr(conf->percpu, cpu);
4432 safe_put_page(percpu->spare_page);
4433 kfree(percpu->scribble);
4435 #ifdef CONFIG_HOTPLUG_CPU
4436 unregister_cpu_notifier(&conf->cpu_notify);
4440 free_percpu(conf->percpu);
4443 static void free_conf(raid5_conf_t *conf)
4445 shrink_stripes(conf);
4446 raid5_free_percpu(conf);
4448 kfree(conf->stripe_hashtbl);
4452 #ifdef CONFIG_HOTPLUG_CPU
4453 static int raid456_cpu_notify(struct notifier_block *nfb, unsigned long action,
4456 raid5_conf_t *conf = container_of(nfb, raid5_conf_t, cpu_notify);
4457 long cpu = (long)hcpu;
4458 struct raid5_percpu *percpu = per_cpu_ptr(conf->percpu, cpu);
4461 case CPU_UP_PREPARE:
4462 case CPU_UP_PREPARE_FROZEN:
4463 if (conf->level == 6 && !percpu->spare_page)
4464 percpu->spare_page = alloc_page(GFP_KERNEL);
4465 if (!percpu->scribble)
4466 percpu->scribble = kmalloc(conf->scribble_len, GFP_KERNEL);
4468 if (!percpu->scribble ||
4469 (conf->level == 6 && !percpu->spare_page)) {
4470 safe_put_page(percpu->spare_page);
4471 kfree(percpu->scribble);
4472 pr_err("%s: failed memory allocation for cpu%ld\n",
4474 return notifier_from_errno(-ENOMEM);
4478 case CPU_DEAD_FROZEN:
4479 safe_put_page(percpu->spare_page);
4480 kfree(percpu->scribble);
4481 percpu->spare_page = NULL;
4482 percpu->scribble = NULL;
4491 static int raid5_alloc_percpu(raid5_conf_t *conf)
4494 struct page *spare_page;
4495 struct raid5_percpu __percpu *allcpus;
4499 allcpus = alloc_percpu(struct raid5_percpu);
4502 conf->percpu = allcpus;
4506 for_each_present_cpu(cpu) {
4507 if (conf->level == 6) {
4508 spare_page = alloc_page(GFP_KERNEL);
4513 per_cpu_ptr(conf->percpu, cpu)->spare_page = spare_page;
4515 scribble = kmalloc(conf->scribble_len, GFP_KERNEL);
4520 per_cpu_ptr(conf->percpu, cpu)->scribble = scribble;
4522 #ifdef CONFIG_HOTPLUG_CPU
4523 conf->cpu_notify.notifier_call = raid456_cpu_notify;
4524 conf->cpu_notify.priority = 0;
4526 err = register_cpu_notifier(&conf->cpu_notify);
4533 static raid5_conf_t *setup_conf(mddev_t *mddev)
4536 int raid_disk, memory, max_disks;
4538 struct disk_info *disk;
4540 if (mddev->new_level != 5
4541 && mddev->new_level != 4
4542 && mddev->new_level != 6) {
4543 printk(KERN_ERR "md/raid:%s: raid level not set to 4/5/6 (%d)\n",
4544 mdname(mddev), mddev->new_level);
4545 return ERR_PTR(-EIO);
4547 if ((mddev->new_level == 5
4548 && !algorithm_valid_raid5(mddev->new_layout)) ||
4549 (mddev->new_level == 6
4550 && !algorithm_valid_raid6(mddev->new_layout))) {
4551 printk(KERN_ERR "md/raid:%s: layout %d not supported\n",
4552 mdname(mddev), mddev->new_layout);
4553 return ERR_PTR(-EIO);
4555 if (mddev->new_level == 6 && mddev->raid_disks < 4) {
4556 printk(KERN_ERR "md/raid:%s: not enough configured devices (%d, minimum 4)\n",
4557 mdname(mddev), mddev->raid_disks);
4558 return ERR_PTR(-EINVAL);
4561 if (!mddev->new_chunk_sectors ||
4562 (mddev->new_chunk_sectors << 9) % PAGE_SIZE ||
4563 !is_power_of_2(mddev->new_chunk_sectors)) {
4564 printk(KERN_ERR "md/raid:%s: invalid chunk size %d\n",
4565 mdname(mddev), mddev->new_chunk_sectors << 9);
4566 return ERR_PTR(-EINVAL);
4569 conf = kzalloc(sizeof(raid5_conf_t), GFP_KERNEL);
4572 spin_lock_init(&conf->device_lock);
4573 init_waitqueue_head(&conf->wait_for_stripe);
4574 init_waitqueue_head(&conf->wait_for_overlap);
4575 INIT_LIST_HEAD(&conf->handle_list);
4576 INIT_LIST_HEAD(&conf->hold_list);
4577 INIT_LIST_HEAD(&conf->delayed_list);
4578 INIT_LIST_HEAD(&conf->bitmap_list);
4579 INIT_LIST_HEAD(&conf->inactive_list);
4580 atomic_set(&conf->active_stripes, 0);
4581 atomic_set(&conf->preread_active_stripes, 0);
4582 atomic_set(&conf->active_aligned_reads, 0);
4583 conf->bypass_threshold = BYPASS_THRESHOLD;
4585 conf->raid_disks = mddev->raid_disks;
4586 if (mddev->reshape_position == MaxSector)
4587 conf->previous_raid_disks = mddev->raid_disks;
4589 conf->previous_raid_disks = mddev->raid_disks - mddev->delta_disks;
4590 max_disks = max(conf->raid_disks, conf->previous_raid_disks);
4591 conf->scribble_len = scribble_len(max_disks);
4593 conf->disks = kzalloc(max_disks * sizeof(struct disk_info),
4598 conf->mddev = mddev;
4600 if ((conf->stripe_hashtbl = kzalloc(PAGE_SIZE, GFP_KERNEL)) == NULL)
4603 conf->level = mddev->new_level;
4604 if (raid5_alloc_percpu(conf) != 0)
4607 pr_debug("raid456: run(%s) called.\n", mdname(mddev));
4609 list_for_each_entry(rdev, &mddev->disks, same_set) {
4610 raid_disk = rdev->raid_disk;
4611 if (raid_disk >= max_disks
4614 disk = conf->disks + raid_disk;
4618 if (test_bit(In_sync, &rdev->flags)) {
4619 char b[BDEVNAME_SIZE];
4620 printk(KERN_INFO "md/raid:%s: device %s operational as raid"
4622 mdname(mddev), bdevname(rdev->bdev, b), raid_disk);
4623 } else if (rdev->saved_raid_disk != raid_disk)
4624 /* Cannot rely on bitmap to complete recovery */
4628 conf->chunk_sectors = mddev->new_chunk_sectors;
4629 conf->level = mddev->new_level;
4630 if (conf->level == 6)
4631 conf->max_degraded = 2;
4633 conf->max_degraded = 1;
4634 conf->algorithm = mddev->new_layout;
4635 conf->max_nr_stripes = NR_STRIPES;
4636 conf->reshape_progress = mddev->reshape_position;
4637 if (conf->reshape_progress != MaxSector) {
4638 conf->prev_chunk_sectors = mddev->chunk_sectors;
4639 conf->prev_algo = mddev->layout;
4642 memory = conf->max_nr_stripes * (sizeof(struct stripe_head) +
4643 max_disks * ((sizeof(struct bio) + PAGE_SIZE))) / 1024;
4644 if (grow_stripes(conf, conf->max_nr_stripes)) {
4646 "md/raid:%s: couldn't allocate %dkB for buffers\n",
4647 mdname(mddev), memory);
4650 printk(KERN_INFO "md/raid:%s: allocated %dkB\n",
4651 mdname(mddev), memory);
4653 conf->thread = md_register_thread(raid5d, mddev, NULL);
4654 if (!conf->thread) {
4656 "md/raid:%s: couldn't allocate thread.\n",
4666 return ERR_PTR(-EIO);
4668 return ERR_PTR(-ENOMEM);
4672 static int only_parity(int raid_disk, int algo, int raid_disks, int max_degraded)
4675 case ALGORITHM_PARITY_0:
4676 if (raid_disk < max_degraded)
4679 case ALGORITHM_PARITY_N:
4680 if (raid_disk >= raid_disks - max_degraded)
4683 case ALGORITHM_PARITY_0_6:
4684 if (raid_disk == 0 ||
4685 raid_disk == raid_disks - 1)
4688 case ALGORITHM_LEFT_ASYMMETRIC_6:
4689 case ALGORITHM_RIGHT_ASYMMETRIC_6:
4690 case ALGORITHM_LEFT_SYMMETRIC_6:
4691 case ALGORITHM_RIGHT_SYMMETRIC_6:
4692 if (raid_disk == raid_disks - 1)
4698 static int run(mddev_t *mddev)
4701 int working_disks = 0;
4702 int dirty_parity_disks = 0;
4704 sector_t reshape_offset = 0;
4706 if (mddev->recovery_cp != MaxSector)
4707 printk(KERN_NOTICE "md/raid:%s: not clean"
4708 " -- starting background reconstruction\n",
4710 if (mddev->reshape_position != MaxSector) {
4711 /* Check that we can continue the reshape.
4712 * Currently only disks can change, it must
4713 * increase, and we must be past the point where
4714 * a stripe over-writes itself
4716 sector_t here_new, here_old;
4718 int max_degraded = (mddev->level == 6 ? 2 : 1);
4720 if (mddev->new_level != mddev->level) {
4721 printk(KERN_ERR "md/raid:%s: unsupported reshape "
4722 "required - aborting.\n",
4726 old_disks = mddev->raid_disks - mddev->delta_disks;
4727 /* reshape_position must be on a new-stripe boundary, and one
4728 * further up in new geometry must map after here in old
4731 here_new = mddev->reshape_position;
4732 if (sector_div(here_new, mddev->new_chunk_sectors *
4733 (mddev->raid_disks - max_degraded))) {
4734 printk(KERN_ERR "md/raid:%s: reshape_position not "
4735 "on a stripe boundary\n", mdname(mddev));
4738 reshape_offset = here_new * mddev->new_chunk_sectors;
4739 /* here_new is the stripe we will write to */
4740 here_old = mddev->reshape_position;
4741 sector_div(here_old, mddev->chunk_sectors *
4742 (old_disks-max_degraded));
4743 /* here_old is the first stripe that we might need to read
4745 if (mddev->delta_disks == 0) {
4746 /* We cannot be sure it is safe to start an in-place
4747 * reshape. It is only safe if user-space if monitoring
4748 * and taking constant backups.
4749 * mdadm always starts a situation like this in
4750 * readonly mode so it can take control before
4751 * allowing any writes. So just check for that.
4753 if ((here_new * mddev->new_chunk_sectors !=
4754 here_old * mddev->chunk_sectors) ||
4756 printk(KERN_ERR "md/raid:%s: in-place reshape must be started"
4757 " in read-only mode - aborting\n",
4761 } else if (mddev->delta_disks < 0
4762 ? (here_new * mddev->new_chunk_sectors <=
4763 here_old * mddev->chunk_sectors)
4764 : (here_new * mddev->new_chunk_sectors >=
4765 here_old * mddev->chunk_sectors)) {
4766 /* Reading from the same stripe as writing to - bad */
4767 printk(KERN_ERR "md/raid:%s: reshape_position too early for "
4768 "auto-recovery - aborting.\n",
4772 printk(KERN_INFO "md/raid:%s: reshape will continue\n",
4774 /* OK, we should be able to continue; */
4776 BUG_ON(mddev->level != mddev->new_level);
4777 BUG_ON(mddev->layout != mddev->new_layout);
4778 BUG_ON(mddev->chunk_sectors != mddev->new_chunk_sectors);
4779 BUG_ON(mddev->delta_disks != 0);
4782 if (mddev->private == NULL)
4783 conf = setup_conf(mddev);
4785 conf = mddev->private;
4788 return PTR_ERR(conf);
4790 mddev->thread = conf->thread;
4791 conf->thread = NULL;
4792 mddev->private = conf;
4795 * 0 for a fully functional array, 1 or 2 for a degraded array.
4797 list_for_each_entry(rdev, &mddev->disks, same_set) {
4798 if (rdev->raid_disk < 0)
4800 if (test_bit(In_sync, &rdev->flags)) {
4804 /* This disc is not fully in-sync. However if it
4805 * just stored parity (beyond the recovery_offset),
4806 * when we don't need to be concerned about the
4807 * array being dirty.
4808 * When reshape goes 'backwards', we never have
4809 * partially completed devices, so we only need
4810 * to worry about reshape going forwards.
4812 /* Hack because v0.91 doesn't store recovery_offset properly. */
4813 if (mddev->major_version == 0 &&
4814 mddev->minor_version > 90)
4815 rdev->recovery_offset = reshape_offset;
4817 if (rdev->recovery_offset < reshape_offset) {
4818 /* We need to check old and new layout */
4819 if (!only_parity(rdev->raid_disk,
4822 conf->max_degraded))
4825 if (!only_parity(rdev->raid_disk,
4827 conf->previous_raid_disks,
4828 conf->max_degraded))
4830 dirty_parity_disks++;
4833 mddev->degraded = (max(conf->raid_disks, conf->previous_raid_disks)
4836 if (has_failed(conf)) {
4837 printk(KERN_ERR "md/raid:%s: not enough operational devices"
4838 " (%d/%d failed)\n",
4839 mdname(mddev), mddev->degraded, conf->raid_disks);
4843 /* device size must be a multiple of chunk size */
4844 mddev->dev_sectors &= ~(mddev->chunk_sectors - 1);
4845 mddev->resync_max_sectors = mddev->dev_sectors;
4847 if (mddev->degraded > dirty_parity_disks &&
4848 mddev->recovery_cp != MaxSector) {
4849 if (mddev->ok_start_degraded)
4851 "md/raid:%s: starting dirty degraded array"
4852 " - data corruption possible.\n",
4856 "md/raid:%s: cannot start dirty degraded array.\n",
4862 if (mddev->degraded == 0)
4863 printk(KERN_INFO "md/raid:%s: raid level %d active with %d out of %d"
4864 " devices, algorithm %d\n", mdname(mddev), conf->level,
4865 mddev->raid_disks-mddev->degraded, mddev->raid_disks,
4868 printk(KERN_ALERT "md/raid:%s: raid level %d active with %d"
4869 " out of %d devices, algorithm %d\n",
4870 mdname(mddev), conf->level,
4871 mddev->raid_disks - mddev->degraded,
4872 mddev->raid_disks, mddev->new_layout);
4874 print_raid5_conf(conf);
4876 if (conf->reshape_progress != MaxSector) {
4877 conf->reshape_safe = conf->reshape_progress;
4878 atomic_set(&conf->reshape_stripes, 0);
4879 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
4880 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
4881 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
4882 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
4883 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
4888 /* Ok, everything is just fine now */
4889 if (mddev->to_remove == &raid5_attrs_group)
4890 mddev->to_remove = NULL;
4891 else if (mddev->kobj.sd &&
4892 sysfs_create_group(&mddev->kobj, &raid5_attrs_group))
4894 "raid5: failed to create sysfs attributes for %s\n",
4896 md_set_array_sectors(mddev, raid5_size(mddev, 0, 0));
4900 /* read-ahead size must cover two whole stripes, which
4901 * is 2 * (datadisks) * chunksize where 'n' is the
4902 * number of raid devices
4904 int data_disks = conf->previous_raid_disks - conf->max_degraded;
4905 int stripe = data_disks *
4906 ((mddev->chunk_sectors << 9) / PAGE_SIZE);
4907 if (mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
4908 mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
4910 blk_queue_merge_bvec(mddev->queue, raid5_mergeable_bvec);
4912 mddev->queue->backing_dev_info.congested_data = mddev;
4913 mddev->queue->backing_dev_info.congested_fn = raid5_congested;
4915 chunk_size = mddev->chunk_sectors << 9;
4916 blk_queue_io_min(mddev->queue, chunk_size);
4917 blk_queue_io_opt(mddev->queue, chunk_size *
4918 (conf->raid_disks - conf->max_degraded));
4920 list_for_each_entry(rdev, &mddev->disks, same_set)
4921 disk_stack_limits(mddev->gendisk, rdev->bdev,
4922 rdev->data_offset << 9);
4927 md_unregister_thread(mddev->thread);
4928 mddev->thread = NULL;
4930 print_raid5_conf(conf);
4933 mddev->private = NULL;
4934 printk(KERN_ALERT "md/raid:%s: failed to run raid set.\n", mdname(mddev));
4938 static int stop(mddev_t *mddev)
4940 raid5_conf_t *conf = mddev->private;
4942 md_unregister_thread(mddev->thread);
4943 mddev->thread = NULL;
4945 mddev->queue->backing_dev_info.congested_fn = NULL;
4947 mddev->private = NULL;
4948 mddev->to_remove = &raid5_attrs_group;
4953 static void print_sh(struct seq_file *seq, struct stripe_head *sh)
4957 seq_printf(seq, "sh %llu, pd_idx %d, state %ld.\n",
4958 (unsigned long long)sh->sector, sh->pd_idx, sh->state);
4959 seq_printf(seq, "sh %llu, count %d.\n",
4960 (unsigned long long)sh->sector, atomic_read(&sh->count));
4961 seq_printf(seq, "sh %llu, ", (unsigned long long)sh->sector);
4962 for (i = 0; i < sh->disks; i++) {
4963 seq_printf(seq, "(cache%d: %p %ld) ",
4964 i, sh->dev[i].page, sh->dev[i].flags);
4966 seq_printf(seq, "\n");
4969 static void printall(struct seq_file *seq, raid5_conf_t *conf)
4971 struct stripe_head *sh;
4972 struct hlist_node *hn;
4975 spin_lock_irq(&conf->device_lock);
4976 for (i = 0; i < NR_HASH; i++) {
4977 hlist_for_each_entry(sh, hn, &conf->stripe_hashtbl[i], hash) {
4978 if (sh->raid_conf != conf)
4983 spin_unlock_irq(&conf->device_lock);
4987 static void status(struct seq_file *seq, mddev_t *mddev)
4989 raid5_conf_t *conf = mddev->private;
4992 seq_printf(seq, " level %d, %dk chunk, algorithm %d", mddev->level,
4993 mddev->chunk_sectors / 2, mddev->layout);
4994 seq_printf (seq, " [%d/%d] [", conf->raid_disks, conf->raid_disks - mddev->degraded);
4995 for (i = 0; i < conf->raid_disks; i++)
4996 seq_printf (seq, "%s",
4997 conf->disks[i].rdev &&
4998 test_bit(In_sync, &conf->disks[i].rdev->flags) ? "U" : "_");
4999 seq_printf (seq, "]");
5001 seq_printf (seq, "\n");
5002 printall(seq, conf);
5006 static void print_raid5_conf (raid5_conf_t *conf)
5009 struct disk_info *tmp;
5011 printk(KERN_DEBUG "RAID conf printout:\n");
5013 printk("(conf==NULL)\n");
5016 printk(KERN_DEBUG " --- level:%d rd:%d wd:%d\n", conf->level,
5018 conf->raid_disks - conf->mddev->degraded);
5020 for (i = 0; i < conf->raid_disks; i++) {
5021 char b[BDEVNAME_SIZE];
5022 tmp = conf->disks + i;
5024 printk(KERN_DEBUG " disk %d, o:%d, dev:%s\n",
5025 i, !test_bit(Faulty, &tmp->rdev->flags),
5026 bdevname(tmp->rdev->bdev, b));
5030 static int raid5_spare_active(mddev_t *mddev)
5033 raid5_conf_t *conf = mddev->private;
5034 struct disk_info *tmp;
5036 unsigned long flags;
5038 for (i = 0; i < conf->raid_disks; i++) {
5039 tmp = conf->disks + i;
5041 && tmp->rdev->recovery_offset == MaxSector
5042 && !test_bit(Faulty, &tmp->rdev->flags)
5043 && !test_and_set_bit(In_sync, &tmp->rdev->flags)) {
5045 sysfs_notify_dirent_safe(tmp->rdev->sysfs_state);
5048 spin_lock_irqsave(&conf->device_lock, flags);
5049 mddev->degraded -= count;
5050 spin_unlock_irqrestore(&conf->device_lock, flags);
5051 print_raid5_conf(conf);
5055 static int raid5_remove_disk(mddev_t *mddev, int number)
5057 raid5_conf_t *conf = mddev->private;
5060 struct disk_info *p = conf->disks + number;
5062 print_raid5_conf(conf);
5065 if (number >= conf->raid_disks &&
5066 conf->reshape_progress == MaxSector)
5067 clear_bit(In_sync, &rdev->flags);
5069 if (test_bit(In_sync, &rdev->flags) ||
5070 atomic_read(&rdev->nr_pending)) {
5074 /* Only remove non-faulty devices if recovery
5077 if (!test_bit(Faulty, &rdev->flags) &&
5078 mddev->recovery_disabled != conf->recovery_disabled &&
5079 !has_failed(conf) &&
5080 number < conf->raid_disks) {
5086 if (atomic_read(&rdev->nr_pending)) {
5087 /* lost the race, try later */
5094 print_raid5_conf(conf);
5098 static int raid5_add_disk(mddev_t *mddev, mdk_rdev_t *rdev)
5100 raid5_conf_t *conf = mddev->private;
5103 struct disk_info *p;
5105 int last = conf->raid_disks - 1;
5107 if (mddev->recovery_disabled == conf->recovery_disabled)
5110 if (has_failed(conf))
5111 /* no point adding a device */
5114 if (rdev->raid_disk >= 0)
5115 first = last = rdev->raid_disk;
5118 * find the disk ... but prefer rdev->saved_raid_disk
5121 if (rdev->saved_raid_disk >= 0 &&
5122 rdev->saved_raid_disk >= first &&
5123 conf->disks[rdev->saved_raid_disk].rdev == NULL)
5124 disk = rdev->saved_raid_disk;
5127 for ( ; disk <= last ; disk++)
5128 if ((p=conf->disks + disk)->rdev == NULL) {
5129 clear_bit(In_sync, &rdev->flags);
5130 rdev->raid_disk = disk;
5132 if (rdev->saved_raid_disk != disk)
5134 rcu_assign_pointer(p->rdev, rdev);
5137 print_raid5_conf(conf);
5141 static int raid5_resize(mddev_t *mddev, sector_t sectors)
5143 /* no resync is happening, and there is enough space
5144 * on all devices, so we can resize.
5145 * We need to make sure resync covers any new space.
5146 * If the array is shrinking we should possibly wait until
5147 * any io in the removed space completes, but it hardly seems
5150 sectors &= ~((sector_t)mddev->chunk_sectors - 1);
5151 md_set_array_sectors(mddev, raid5_size(mddev, sectors,
5152 mddev->raid_disks));
5153 if (mddev->array_sectors >
5154 raid5_size(mddev, sectors, mddev->raid_disks))
5156 set_capacity(mddev->gendisk, mddev->array_sectors);
5157 revalidate_disk(mddev->gendisk);
5158 if (sectors > mddev->dev_sectors &&
5159 mddev->recovery_cp > mddev->dev_sectors) {
5160 mddev->recovery_cp = mddev->dev_sectors;
5161 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
5163 mddev->dev_sectors = sectors;
5164 mddev->resync_max_sectors = sectors;
5168 static int check_stripe_cache(mddev_t *mddev)
5170 /* Can only proceed if there are plenty of stripe_heads.
5171 * We need a minimum of one full stripe,, and for sensible progress
5172 * it is best to have about 4 times that.
5173 * If we require 4 times, then the default 256 4K stripe_heads will
5174 * allow for chunk sizes up to 256K, which is probably OK.
5175 * If the chunk size is greater, user-space should request more
5176 * stripe_heads first.
5178 raid5_conf_t *conf = mddev->private;
5179 if (((mddev->chunk_sectors << 9) / STRIPE_SIZE) * 4
5180 > conf->max_nr_stripes ||
5181 ((mddev->new_chunk_sectors << 9) / STRIPE_SIZE) * 4
5182 > conf->max_nr_stripes) {
5183 printk(KERN_WARNING "md/raid:%s: reshape: not enough stripes. Needed %lu\n",
5185 ((max(mddev->chunk_sectors, mddev->new_chunk_sectors) << 9)
5192 static int check_reshape(mddev_t *mddev)
5194 raid5_conf_t *conf = mddev->private;
5196 if (mddev->delta_disks == 0 &&
5197 mddev->new_layout == mddev->layout &&
5198 mddev->new_chunk_sectors == mddev->chunk_sectors)
5199 return 0; /* nothing to do */
5201 /* Cannot grow a bitmap yet */
5203 if (has_failed(conf))
5205 if (mddev->delta_disks < 0) {
5206 /* We might be able to shrink, but the devices must
5207 * be made bigger first.
5208 * For raid6, 4 is the minimum size.
5209 * Otherwise 2 is the minimum
5212 if (mddev->level == 6)
5214 if (mddev->raid_disks + mddev->delta_disks < min)
5218 if (!check_stripe_cache(mddev))
5221 return resize_stripes(conf, conf->raid_disks + mddev->delta_disks);
5224 static int raid5_start_reshape(mddev_t *mddev)
5226 raid5_conf_t *conf = mddev->private;
5229 unsigned long flags;
5231 if (test_bit(MD_RECOVERY_RUNNING, &mddev->recovery))
5234 if (!check_stripe_cache(mddev))
5237 list_for_each_entry(rdev, &mddev->disks, same_set)
5238 if (!test_bit(In_sync, &rdev->flags)
5239 && !test_bit(Faulty, &rdev->flags))
5242 if (spares - mddev->degraded < mddev->delta_disks - conf->max_degraded)
5243 /* Not enough devices even to make a degraded array
5248 /* Refuse to reduce size of the array. Any reductions in
5249 * array size must be through explicit setting of array_size
5252 if (raid5_size(mddev, 0, conf->raid_disks + mddev->delta_disks)
5253 < mddev->array_sectors) {
5254 printk(KERN_ERR "md/raid:%s: array size must be reduced "
5255 "before number of disks\n", mdname(mddev));
5259 atomic_set(&conf->reshape_stripes, 0);
5260 spin_lock_irq(&conf->device_lock);
5261 conf->previous_raid_disks = conf->raid_disks;
5262 conf->raid_disks += mddev->delta_disks;
5263 conf->prev_chunk_sectors = conf->chunk_sectors;
5264 conf->chunk_sectors = mddev->new_chunk_sectors;
5265 conf->prev_algo = conf->algorithm;
5266 conf->algorithm = mddev->new_layout;
5267 if (mddev->delta_disks < 0)
5268 conf->reshape_progress = raid5_size(mddev, 0, 0);
5270 conf->reshape_progress = 0;
5271 conf->reshape_safe = conf->reshape_progress;
5273 spin_unlock_irq(&conf->device_lock);
5275 /* Add some new drives, as many as will fit.
5276 * We know there are enough to make the newly sized array work.
5277 * Don't add devices if we are reducing the number of
5278 * devices in the array. This is because it is not possible
5279 * to correctly record the "partially reconstructed" state of
5280 * such devices during the reshape and confusion could result.
5282 if (mddev->delta_disks >= 0) {
5283 int added_devices = 0;
5284 list_for_each_entry(rdev, &mddev->disks, same_set)
5285 if (rdev->raid_disk < 0 &&
5286 !test_bit(Faulty, &rdev->flags)) {
5287 if (raid5_add_disk(mddev, rdev) == 0) {
5289 >= conf->previous_raid_disks) {
5290 set_bit(In_sync, &rdev->flags);
5293 rdev->recovery_offset = 0;
5295 if (sysfs_link_rdev(mddev, rdev))
5296 /* Failure here is OK */;
5298 } else if (rdev->raid_disk >= conf->previous_raid_disks
5299 && !test_bit(Faulty, &rdev->flags)) {
5300 /* This is a spare that was manually added */
5301 set_bit(In_sync, &rdev->flags);
5305 /* When a reshape changes the number of devices,
5306 * ->degraded is measured against the larger of the
5307 * pre and post number of devices.
5309 spin_lock_irqsave(&conf->device_lock, flags);
5310 mddev->degraded += (conf->raid_disks - conf->previous_raid_disks)
5312 spin_unlock_irqrestore(&conf->device_lock, flags);
5314 mddev->raid_disks = conf->raid_disks;
5315 mddev->reshape_position = conf->reshape_progress;
5316 set_bit(MD_CHANGE_DEVS, &mddev->flags);
5318 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
5319 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
5320 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
5321 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
5322 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
5324 if (!mddev->sync_thread) {
5325 mddev->recovery = 0;
5326 spin_lock_irq(&conf->device_lock);
5327 mddev->raid_disks = conf->raid_disks = conf->previous_raid_disks;
5328 conf->reshape_progress = MaxSector;
5329 spin_unlock_irq(&conf->device_lock);
5332 conf->reshape_checkpoint = jiffies;
5333 md_wakeup_thread(mddev->sync_thread);
5334 md_new_event(mddev);
5338 /* This is called from the reshape thread and should make any
5339 * changes needed in 'conf'
5341 static void end_reshape(raid5_conf_t *conf)
5344 if (!test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery)) {
5346 spin_lock_irq(&conf->device_lock);
5347 conf->previous_raid_disks = conf->raid_disks;
5348 conf->reshape_progress = MaxSector;
5349 spin_unlock_irq(&conf->device_lock);
5350 wake_up(&conf->wait_for_overlap);
5352 /* read-ahead size must cover two whole stripes, which is
5353 * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
5355 if (conf->mddev->queue) {
5356 int data_disks = conf->raid_disks - conf->max_degraded;
5357 int stripe = data_disks * ((conf->chunk_sectors << 9)
5359 if (conf->mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
5360 conf->mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
5365 /* This is called from the raid5d thread with mddev_lock held.
5366 * It makes config changes to the device.
5368 static void raid5_finish_reshape(mddev_t *mddev)
5370 raid5_conf_t *conf = mddev->private;
5372 if (!test_bit(MD_RECOVERY_INTR, &mddev->recovery)) {
5374 if (mddev->delta_disks > 0) {
5375 md_set_array_sectors(mddev, raid5_size(mddev, 0, 0));
5376 set_capacity(mddev->gendisk, mddev->array_sectors);
5377 revalidate_disk(mddev->gendisk);
5380 mddev->degraded = conf->raid_disks;
5381 for (d = 0; d < conf->raid_disks ; d++)
5382 if (conf->disks[d].rdev &&
5384 &conf->disks[d].rdev->flags))
5386 for (d = conf->raid_disks ;
5387 d < conf->raid_disks - mddev->delta_disks;
5389 mdk_rdev_t *rdev = conf->disks[d].rdev;
5390 if (rdev && raid5_remove_disk(mddev, d) == 0) {
5391 sysfs_unlink_rdev(mddev, rdev);
5392 rdev->raid_disk = -1;
5396 mddev->layout = conf->algorithm;
5397 mddev->chunk_sectors = conf->chunk_sectors;
5398 mddev->reshape_position = MaxSector;
5399 mddev->delta_disks = 0;
5403 static void raid5_quiesce(mddev_t *mddev, int state)
5405 raid5_conf_t *conf = mddev->private;
5408 case 2: /* resume for a suspend */
5409 wake_up(&conf->wait_for_overlap);
5412 case 1: /* stop all writes */
5413 spin_lock_irq(&conf->device_lock);
5414 /* '2' tells resync/reshape to pause so that all
5415 * active stripes can drain
5418 wait_event_lock_irq(conf->wait_for_stripe,
5419 atomic_read(&conf->active_stripes) == 0 &&
5420 atomic_read(&conf->active_aligned_reads) == 0,
5421 conf->device_lock, /* nothing */);
5423 spin_unlock_irq(&conf->device_lock);
5424 /* allow reshape to continue */
5425 wake_up(&conf->wait_for_overlap);
5428 case 0: /* re-enable writes */
5429 spin_lock_irq(&conf->device_lock);
5431 wake_up(&conf->wait_for_stripe);
5432 wake_up(&conf->wait_for_overlap);
5433 spin_unlock_irq(&conf->device_lock);
5439 static void *raid45_takeover_raid0(mddev_t *mddev, int level)
5441 struct raid0_private_data *raid0_priv = mddev->private;
5444 /* for raid0 takeover only one zone is supported */
5445 if (raid0_priv->nr_strip_zones > 1) {
5446 printk(KERN_ERR "md/raid:%s: cannot takeover raid0 with more than one zone.\n",
5448 return ERR_PTR(-EINVAL);
5451 sectors = raid0_priv->strip_zone[0].zone_end;
5452 sector_div(sectors, raid0_priv->strip_zone[0].nb_dev);
5453 mddev->dev_sectors = sectors;
5454 mddev->new_level = level;
5455 mddev->new_layout = ALGORITHM_PARITY_N;
5456 mddev->new_chunk_sectors = mddev->chunk_sectors;
5457 mddev->raid_disks += 1;
5458 mddev->delta_disks = 1;
5459 /* make sure it will be not marked as dirty */
5460 mddev->recovery_cp = MaxSector;
5462 return setup_conf(mddev);
5466 static void *raid5_takeover_raid1(mddev_t *mddev)
5470 if (mddev->raid_disks != 2 ||
5471 mddev->degraded > 1)
5472 return ERR_PTR(-EINVAL);
5474 /* Should check if there are write-behind devices? */
5476 chunksect = 64*2; /* 64K by default */
5478 /* The array must be an exact multiple of chunksize */
5479 while (chunksect && (mddev->array_sectors & (chunksect-1)))
5482 if ((chunksect<<9) < STRIPE_SIZE)
5483 /* array size does not allow a suitable chunk size */
5484 return ERR_PTR(-EINVAL);
5486 mddev->new_level = 5;
5487 mddev->new_layout = ALGORITHM_LEFT_SYMMETRIC;
5488 mddev->new_chunk_sectors = chunksect;
5490 return setup_conf(mddev);
5493 static void *raid5_takeover_raid6(mddev_t *mddev)
5497 switch (mddev->layout) {
5498 case ALGORITHM_LEFT_ASYMMETRIC_6:
5499 new_layout = ALGORITHM_LEFT_ASYMMETRIC;
5501 case ALGORITHM_RIGHT_ASYMMETRIC_6:
5502 new_layout = ALGORITHM_RIGHT_ASYMMETRIC;
5504 case ALGORITHM_LEFT_SYMMETRIC_6:
5505 new_layout = ALGORITHM_LEFT_SYMMETRIC;
5507 case ALGORITHM_RIGHT_SYMMETRIC_6:
5508 new_layout = ALGORITHM_RIGHT_SYMMETRIC;
5510 case ALGORITHM_PARITY_0_6:
5511 new_layout = ALGORITHM_PARITY_0;
5513 case ALGORITHM_PARITY_N:
5514 new_layout = ALGORITHM_PARITY_N;
5517 return ERR_PTR(-EINVAL);
5519 mddev->new_level = 5;
5520 mddev->new_layout = new_layout;
5521 mddev->delta_disks = -1;
5522 mddev->raid_disks -= 1;
5523 return setup_conf(mddev);
5527 static int raid5_check_reshape(mddev_t *mddev)
5529 /* For a 2-drive array, the layout and chunk size can be changed
5530 * immediately as not restriping is needed.
5531 * For larger arrays we record the new value - after validation
5532 * to be used by a reshape pass.
5534 raid5_conf_t *conf = mddev->private;
5535 int new_chunk = mddev->new_chunk_sectors;
5537 if (mddev->new_layout >= 0 && !algorithm_valid_raid5(mddev->new_layout))
5539 if (new_chunk > 0) {
5540 if (!is_power_of_2(new_chunk))
5542 if (new_chunk < (PAGE_SIZE>>9))
5544 if (mddev->array_sectors & (new_chunk-1))
5545 /* not factor of array size */
5549 /* They look valid */
5551 if (mddev->raid_disks == 2) {
5552 /* can make the change immediately */
5553 if (mddev->new_layout >= 0) {
5554 conf->algorithm = mddev->new_layout;
5555 mddev->layout = mddev->new_layout;
5557 if (new_chunk > 0) {
5558 conf->chunk_sectors = new_chunk ;
5559 mddev->chunk_sectors = new_chunk;
5561 set_bit(MD_CHANGE_DEVS, &mddev->flags);
5562 md_wakeup_thread(mddev->thread);
5564 return check_reshape(mddev);
5567 static int raid6_check_reshape(mddev_t *mddev)
5569 int new_chunk = mddev->new_chunk_sectors;
5571 if (mddev->new_layout >= 0 && !algorithm_valid_raid6(mddev->new_layout))
5573 if (new_chunk > 0) {
5574 if (!is_power_of_2(new_chunk))
5576 if (new_chunk < (PAGE_SIZE >> 9))
5578 if (mddev->array_sectors & (new_chunk-1))
5579 /* not factor of array size */
5583 /* They look valid */
5584 return check_reshape(mddev);
5587 static void *raid5_takeover(mddev_t *mddev)
5589 /* raid5 can take over:
5590 * raid0 - if there is only one strip zone - make it a raid4 layout
5591 * raid1 - if there are two drives. We need to know the chunk size
5592 * raid4 - trivial - just use a raid4 layout.
5593 * raid6 - Providing it is a *_6 layout
5595 if (mddev->level == 0)
5596 return raid45_takeover_raid0(mddev, 5);
5597 if (mddev->level == 1)
5598 return raid5_takeover_raid1(mddev);
5599 if (mddev->level == 4) {
5600 mddev->new_layout = ALGORITHM_PARITY_N;
5601 mddev->new_level = 5;
5602 return setup_conf(mddev);
5604 if (mddev->level == 6)
5605 return raid5_takeover_raid6(mddev);
5607 return ERR_PTR(-EINVAL);
5610 static void *raid4_takeover(mddev_t *mddev)
5612 /* raid4 can take over:
5613 * raid0 - if there is only one strip zone
5614 * raid5 - if layout is right
5616 if (mddev->level == 0)
5617 return raid45_takeover_raid0(mddev, 4);
5618 if (mddev->level == 5 &&
5619 mddev->layout == ALGORITHM_PARITY_N) {
5620 mddev->new_layout = 0;
5621 mddev->new_level = 4;
5622 return setup_conf(mddev);
5624 return ERR_PTR(-EINVAL);
5627 static struct mdk_personality raid5_personality;
5629 static void *raid6_takeover(mddev_t *mddev)
5631 /* Currently can only take over a raid5. We map the
5632 * personality to an equivalent raid6 personality
5633 * with the Q block at the end.
5637 if (mddev->pers != &raid5_personality)
5638 return ERR_PTR(-EINVAL);
5639 if (mddev->degraded > 1)
5640 return ERR_PTR(-EINVAL);
5641 if (mddev->raid_disks > 253)
5642 return ERR_PTR(-EINVAL);
5643 if (mddev->raid_disks < 3)
5644 return ERR_PTR(-EINVAL);
5646 switch (mddev->layout) {
5647 case ALGORITHM_LEFT_ASYMMETRIC:
5648 new_layout = ALGORITHM_LEFT_ASYMMETRIC_6;
5650 case ALGORITHM_RIGHT_ASYMMETRIC:
5651 new_layout = ALGORITHM_RIGHT_ASYMMETRIC_6;
5653 case ALGORITHM_LEFT_SYMMETRIC:
5654 new_layout = ALGORITHM_LEFT_SYMMETRIC_6;
5656 case ALGORITHM_RIGHT_SYMMETRIC:
5657 new_layout = ALGORITHM_RIGHT_SYMMETRIC_6;
5659 case ALGORITHM_PARITY_0:
5660 new_layout = ALGORITHM_PARITY_0_6;
5662 case ALGORITHM_PARITY_N:
5663 new_layout = ALGORITHM_PARITY_N;
5666 return ERR_PTR(-EINVAL);
5668 mddev->new_level = 6;
5669 mddev->new_layout = new_layout;
5670 mddev->delta_disks = 1;
5671 mddev->raid_disks += 1;
5672 return setup_conf(mddev);
5676 static struct mdk_personality raid6_personality =
5680 .owner = THIS_MODULE,
5681 .make_request = make_request,
5685 .error_handler = error,
5686 .hot_add_disk = raid5_add_disk,
5687 .hot_remove_disk= raid5_remove_disk,
5688 .spare_active = raid5_spare_active,
5689 .sync_request = sync_request,
5690 .resize = raid5_resize,
5692 .check_reshape = raid6_check_reshape,
5693 .start_reshape = raid5_start_reshape,
5694 .finish_reshape = raid5_finish_reshape,
5695 .quiesce = raid5_quiesce,
5696 .takeover = raid6_takeover,
5698 static struct mdk_personality raid5_personality =
5702 .owner = THIS_MODULE,
5703 .make_request = make_request,
5707 .error_handler = error,
5708 .hot_add_disk = raid5_add_disk,
5709 .hot_remove_disk= raid5_remove_disk,
5710 .spare_active = raid5_spare_active,
5711 .sync_request = sync_request,
5712 .resize = raid5_resize,
5714 .check_reshape = raid5_check_reshape,
5715 .start_reshape = raid5_start_reshape,
5716 .finish_reshape = raid5_finish_reshape,
5717 .quiesce = raid5_quiesce,
5718 .takeover = raid5_takeover,
5721 static struct mdk_personality raid4_personality =
5725 .owner = THIS_MODULE,
5726 .make_request = make_request,
5730 .error_handler = error,
5731 .hot_add_disk = raid5_add_disk,
5732 .hot_remove_disk= raid5_remove_disk,
5733 .spare_active = raid5_spare_active,
5734 .sync_request = sync_request,
5735 .resize = raid5_resize,
5737 .check_reshape = raid5_check_reshape,
5738 .start_reshape = raid5_start_reshape,
5739 .finish_reshape = raid5_finish_reshape,
5740 .quiesce = raid5_quiesce,
5741 .takeover = raid4_takeover,
5744 static int __init raid5_init(void)
5746 register_md_personality(&raid6_personality);
5747 register_md_personality(&raid5_personality);
5748 register_md_personality(&raid4_personality);
5752 static void raid5_exit(void)
5754 unregister_md_personality(&raid6_personality);
5755 unregister_md_personality(&raid5_personality);
5756 unregister_md_personality(&raid4_personality);
5759 module_init(raid5_init);
5760 module_exit(raid5_exit);
5761 MODULE_LICENSE("GPL");
5762 MODULE_DESCRIPTION("RAID4/5/6 (striping with parity) personality for MD");
5763 MODULE_ALIAS("md-personality-4"); /* RAID5 */
5764 MODULE_ALIAS("md-raid5");
5765 MODULE_ALIAS("md-raid4");
5766 MODULE_ALIAS("md-level-5");
5767 MODULE_ALIAS("md-level-4");
5768 MODULE_ALIAS("md-personality-8"); /* RAID6 */
5769 MODULE_ALIAS("md-raid6");
5770 MODULE_ALIAS("md-level-6");
5772 /* This used to be two separate modules, they were: */
5773 MODULE_ALIAS("raid5");
5774 MODULE_ALIAS("raid6");