2 * raid5.c : Multiple Devices driver for Linux
3 * Copyright (C) 1996, 1997 Ingo Molnar, Miguel de Icaza, Gadi Oxman
4 * Copyright (C) 1999, 2000 Ingo Molnar
5 * Copyright (C) 2002, 2003 H. Peter Anvin
7 * RAID-4/5/6 management functions.
8 * Thanks to Penguin Computing for making the RAID-6 development possible
9 * by donating a test server!
11 * This program is free software; you can redistribute it and/or modify
12 * it under the terms of the GNU General Public License as published by
13 * the Free Software Foundation; either version 2, or (at your option)
16 * You should have received a copy of the GNU General Public License
17 * (for example /usr/src/linux/COPYING); if not, write to the Free
18 * Software Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
24 * The sequencing for updating the bitmap reliably is a little
25 * subtle (and I got it wrong the first time) so it deserves some
28 * We group bitmap updates into batches. Each batch has a number.
29 * We may write out several batches at once, but that isn't very important.
30 * conf->seq_write is the number of the last batch successfully written.
31 * conf->seq_flush is the number of the last batch that was closed to
33 * When we discover that we will need to write to any block in a stripe
34 * (in add_stripe_bio) we update the in-memory bitmap and record in sh->bm_seq
35 * the number of the batch it will be in. This is seq_flush+1.
36 * When we are ready to do a write, if that batch hasn't been written yet,
37 * we plug the array and queue the stripe for later.
38 * When an unplug happens, we increment bm_flush, thus closing the current
40 * When we notice that bm_flush > bm_write, we write out all pending updates
41 * to the bitmap, and advance bm_write to where bm_flush was.
42 * This may occasionally write a bit out twice, but is sure never to
46 #include <linux/blkdev.h>
47 #include <linux/kthread.h>
48 #include <linux/raid/pq.h>
49 #include <linux/async_tx.h>
50 #include <linux/module.h>
51 #include <linux/async.h>
52 #include <linux/seq_file.h>
53 #include <linux/cpu.h>
54 #include <linux/slab.h>
55 #include <linux/ratelimit.h>
65 #define NR_STRIPES 256
66 #define STRIPE_SIZE PAGE_SIZE
67 #define STRIPE_SHIFT (PAGE_SHIFT - 9)
68 #define STRIPE_SECTORS (STRIPE_SIZE>>9)
69 #define IO_THRESHOLD 1
70 #define BYPASS_THRESHOLD 1
71 #define NR_HASH (PAGE_SIZE / sizeof(struct hlist_head))
72 #define HASH_MASK (NR_HASH - 1)
74 static inline struct hlist_head *stripe_hash(struct r5conf *conf, sector_t sect)
76 int hash = (sect >> STRIPE_SHIFT) & HASH_MASK;
77 return &conf->stripe_hashtbl[hash];
80 /* bio's attached to a stripe+device for I/O are linked together in bi_sector
81 * order without overlap. There may be several bio's per stripe+device, and
82 * a bio could span several devices.
83 * When walking this list for a particular stripe+device, we must never proceed
84 * beyond a bio that extends past this device, as the next bio might no longer
86 * This function is used to determine the 'next' bio in the list, given the sector
87 * of the current stripe+device
89 static inline struct bio *r5_next_bio(struct bio *bio, sector_t sector)
91 int sectors = bio->bi_size >> 9;
92 if (bio->bi_sector + sectors < sector + STRIPE_SECTORS)
99 * We maintain a biased count of active stripes in the bottom 16 bits of
100 * bi_phys_segments, and a count of processed stripes in the upper 16 bits
102 static inline int raid5_bi_phys_segments(struct bio *bio)
104 return bio->bi_phys_segments & 0xffff;
107 static inline int raid5_bi_hw_segments(struct bio *bio)
109 return (bio->bi_phys_segments >> 16) & 0xffff;
112 static inline int raid5_dec_bi_phys_segments(struct bio *bio)
114 --bio->bi_phys_segments;
115 return raid5_bi_phys_segments(bio);
118 static inline int raid5_dec_bi_hw_segments(struct bio *bio)
120 unsigned short val = raid5_bi_hw_segments(bio);
123 bio->bi_phys_segments = (val << 16) | raid5_bi_phys_segments(bio);
127 static inline void raid5_set_bi_hw_segments(struct bio *bio, unsigned int cnt)
129 bio->bi_phys_segments = raid5_bi_phys_segments(bio) | (cnt << 16);
132 /* Find first data disk in a raid6 stripe */
133 static inline int raid6_d0(struct stripe_head *sh)
136 /* ddf always start from first device */
138 /* md starts just after Q block */
139 if (sh->qd_idx == sh->disks - 1)
142 return sh->qd_idx + 1;
144 static inline int raid6_next_disk(int disk, int raid_disks)
147 return (disk < raid_disks) ? disk : 0;
150 /* When walking through the disks in a raid5, starting at raid6_d0,
151 * We need to map each disk to a 'slot', where the data disks are slot
152 * 0 .. raid_disks-3, the parity disk is raid_disks-2 and the Q disk
153 * is raid_disks-1. This help does that mapping.
155 static int raid6_idx_to_slot(int idx, struct stripe_head *sh,
156 int *count, int syndrome_disks)
162 if (idx == sh->pd_idx)
163 return syndrome_disks;
164 if (idx == sh->qd_idx)
165 return syndrome_disks + 1;
171 static void return_io(struct bio *return_bi)
173 struct bio *bi = return_bi;
176 return_bi = bi->bi_next;
184 static void print_raid5_conf (struct r5conf *conf);
186 static int stripe_operations_active(struct stripe_head *sh)
188 return sh->check_state || sh->reconstruct_state ||
189 test_bit(STRIPE_BIOFILL_RUN, &sh->state) ||
190 test_bit(STRIPE_COMPUTE_RUN, &sh->state);
193 static void __release_stripe(struct r5conf *conf, struct stripe_head *sh)
195 if (atomic_dec_and_test(&sh->count)) {
196 BUG_ON(!list_empty(&sh->lru));
197 BUG_ON(atomic_read(&conf->active_stripes)==0);
198 if (test_bit(STRIPE_HANDLE, &sh->state)) {
199 if (test_bit(STRIPE_DELAYED, &sh->state))
200 list_add_tail(&sh->lru, &conf->delayed_list);
201 else if (test_bit(STRIPE_BIT_DELAY, &sh->state) &&
202 sh->bm_seq - conf->seq_write > 0)
203 list_add_tail(&sh->lru, &conf->bitmap_list);
205 clear_bit(STRIPE_BIT_DELAY, &sh->state);
206 list_add_tail(&sh->lru, &conf->handle_list);
208 md_wakeup_thread(conf->mddev->thread);
210 BUG_ON(stripe_operations_active(sh));
211 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
212 atomic_dec(&conf->preread_active_stripes);
213 if (atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD)
214 md_wakeup_thread(conf->mddev->thread);
216 atomic_dec(&conf->active_stripes);
217 if (!test_bit(STRIPE_EXPANDING, &sh->state)) {
218 list_add_tail(&sh->lru, &conf->inactive_list);
219 wake_up(&conf->wait_for_stripe);
220 if (conf->retry_read_aligned)
221 md_wakeup_thread(conf->mddev->thread);
227 static void release_stripe(struct stripe_head *sh)
229 struct r5conf *conf = sh->raid_conf;
232 spin_lock_irqsave(&conf->device_lock, flags);
233 __release_stripe(conf, sh);
234 spin_unlock_irqrestore(&conf->device_lock, flags);
237 static inline void remove_hash(struct stripe_head *sh)
239 pr_debug("remove_hash(), stripe %llu\n",
240 (unsigned long long)sh->sector);
242 hlist_del_init(&sh->hash);
245 static inline void insert_hash(struct r5conf *conf, struct stripe_head *sh)
247 struct hlist_head *hp = stripe_hash(conf, sh->sector);
249 pr_debug("insert_hash(), stripe %llu\n",
250 (unsigned long long)sh->sector);
252 hlist_add_head(&sh->hash, hp);
256 /* find an idle stripe, make sure it is unhashed, and return it. */
257 static struct stripe_head *get_free_stripe(struct r5conf *conf)
259 struct stripe_head *sh = NULL;
260 struct list_head *first;
262 if (list_empty(&conf->inactive_list))
264 first = conf->inactive_list.next;
265 sh = list_entry(first, struct stripe_head, lru);
266 list_del_init(first);
268 atomic_inc(&conf->active_stripes);
273 static void shrink_buffers(struct stripe_head *sh)
277 int num = sh->raid_conf->pool_size;
279 for (i = 0; i < num ; i++) {
283 sh->dev[i].page = NULL;
288 static int grow_buffers(struct stripe_head *sh)
291 int num = sh->raid_conf->pool_size;
293 for (i = 0; i < num; i++) {
296 if (!(page = alloc_page(GFP_KERNEL))) {
299 sh->dev[i].page = page;
304 static void raid5_build_block(struct stripe_head *sh, int i, int previous);
305 static void stripe_set_idx(sector_t stripe, struct r5conf *conf, int previous,
306 struct stripe_head *sh);
308 static void init_stripe(struct stripe_head *sh, sector_t sector, int previous)
310 struct r5conf *conf = sh->raid_conf;
313 BUG_ON(atomic_read(&sh->count) != 0);
314 BUG_ON(test_bit(STRIPE_HANDLE, &sh->state));
315 BUG_ON(stripe_operations_active(sh));
317 pr_debug("init_stripe called, stripe %llu\n",
318 (unsigned long long)sh->sector);
322 sh->generation = conf->generation - previous;
323 sh->disks = previous ? conf->previous_raid_disks : conf->raid_disks;
325 stripe_set_idx(sector, conf, previous, sh);
329 for (i = sh->disks; i--; ) {
330 struct r5dev *dev = &sh->dev[i];
332 if (dev->toread || dev->read || dev->towrite || dev->written ||
333 test_bit(R5_LOCKED, &dev->flags)) {
334 printk(KERN_ERR "sector=%llx i=%d %p %p %p %p %d\n",
335 (unsigned long long)sh->sector, i, dev->toread,
336 dev->read, dev->towrite, dev->written,
337 test_bit(R5_LOCKED, &dev->flags));
341 raid5_build_block(sh, i, previous);
343 insert_hash(conf, sh);
346 static struct stripe_head *__find_stripe(struct r5conf *conf, sector_t sector,
349 struct stripe_head *sh;
350 struct hlist_node *hn;
352 pr_debug("__find_stripe, sector %llu\n", (unsigned long long)sector);
353 hlist_for_each_entry(sh, hn, stripe_hash(conf, sector), hash)
354 if (sh->sector == sector && sh->generation == generation)
356 pr_debug("__stripe %llu not in cache\n", (unsigned long long)sector);
361 * Need to check if array has failed when deciding whether to:
363 * - remove non-faulty devices
366 * This determination is simple when no reshape is happening.
367 * However if there is a reshape, we need to carefully check
368 * both the before and after sections.
369 * This is because some failed devices may only affect one
370 * of the two sections, and some non-in_sync devices may
371 * be insync in the section most affected by failed devices.
373 static int calc_degraded(struct r5conf *conf)
375 int degraded, degraded2;
380 for (i = 0; i < conf->previous_raid_disks; i++) {
381 struct md_rdev *rdev = rcu_dereference(conf->disks[i].rdev);
382 if (!rdev || test_bit(Faulty, &rdev->flags))
384 else if (test_bit(In_sync, &rdev->flags))
387 /* not in-sync or faulty.
388 * If the reshape increases the number of devices,
389 * this is being recovered by the reshape, so
390 * this 'previous' section is not in_sync.
391 * If the number of devices is being reduced however,
392 * the device can only be part of the array if
393 * we are reverting a reshape, so this section will
396 if (conf->raid_disks >= conf->previous_raid_disks)
400 if (conf->raid_disks == conf->previous_raid_disks)
404 for (i = 0; i < conf->raid_disks; i++) {
405 struct md_rdev *rdev = rcu_dereference(conf->disks[i].rdev);
406 if (!rdev || test_bit(Faulty, &rdev->flags))
408 else if (test_bit(In_sync, &rdev->flags))
411 /* not in-sync or faulty.
412 * If reshape increases the number of devices, this
413 * section has already been recovered, else it
414 * almost certainly hasn't.
416 if (conf->raid_disks <= conf->previous_raid_disks)
420 if (degraded2 > degraded)
425 static int has_failed(struct r5conf *conf)
429 if (conf->mddev->reshape_position == MaxSector)
430 return conf->mddev->degraded > conf->max_degraded;
432 degraded = calc_degraded(conf);
433 if (degraded > conf->max_degraded)
438 static struct stripe_head *
439 get_active_stripe(struct r5conf *conf, sector_t sector,
440 int previous, int noblock, int noquiesce)
442 struct stripe_head *sh;
444 pr_debug("get_stripe, sector %llu\n", (unsigned long long)sector);
446 spin_lock_irq(&conf->device_lock);
449 wait_event_lock_irq(conf->wait_for_stripe,
450 conf->quiesce == 0 || noquiesce,
451 conf->device_lock, /* nothing */);
452 sh = __find_stripe(conf, sector, conf->generation - previous);
454 if (!conf->inactive_blocked)
455 sh = get_free_stripe(conf);
456 if (noblock && sh == NULL)
459 conf->inactive_blocked = 1;
460 wait_event_lock_irq(conf->wait_for_stripe,
461 !list_empty(&conf->inactive_list) &&
462 (atomic_read(&conf->active_stripes)
463 < (conf->max_nr_stripes *3/4)
464 || !conf->inactive_blocked),
467 conf->inactive_blocked = 0;
469 init_stripe(sh, sector, previous);
471 if (atomic_read(&sh->count)) {
472 BUG_ON(!list_empty(&sh->lru)
473 && !test_bit(STRIPE_EXPANDING, &sh->state));
475 if (!test_bit(STRIPE_HANDLE, &sh->state))
476 atomic_inc(&conf->active_stripes);
477 if (list_empty(&sh->lru) &&
478 !test_bit(STRIPE_EXPANDING, &sh->state))
480 list_del_init(&sh->lru);
483 } while (sh == NULL);
486 atomic_inc(&sh->count);
488 spin_unlock_irq(&conf->device_lock);
493 raid5_end_read_request(struct bio *bi, int error);
495 raid5_end_write_request(struct bio *bi, int error);
497 static void ops_run_io(struct stripe_head *sh, struct stripe_head_state *s)
499 struct r5conf *conf = sh->raid_conf;
500 int i, disks = sh->disks;
504 for (i = disks; i--; ) {
507 struct md_rdev *rdev;
508 if (test_and_clear_bit(R5_Wantwrite, &sh->dev[i].flags)) {
509 if (test_and_clear_bit(R5_WantFUA, &sh->dev[i].flags))
513 } else if (test_and_clear_bit(R5_Wantread, &sh->dev[i].flags))
518 bi = &sh->dev[i].req;
522 bi->bi_end_io = raid5_end_write_request;
524 bi->bi_end_io = raid5_end_read_request;
528 test_bit(R5_ReadRepl, &sh->dev[i].flags))
529 rdev = rcu_dereference(conf->disks[i].replacement);
531 rdev = rcu_dereference(conf->disks[i].rdev);
532 if (rdev && test_bit(Faulty, &rdev->flags))
535 atomic_inc(&rdev->nr_pending);
538 /* We have already checked bad blocks for reads. Now
539 * need to check for writes.
541 while ((rw & WRITE) && rdev &&
542 test_bit(WriteErrorSeen, &rdev->flags)) {
545 int bad = is_badblock(rdev, sh->sector, STRIPE_SECTORS,
546 &first_bad, &bad_sectors);
551 set_bit(BlockedBadBlocks, &rdev->flags);
552 if (!conf->mddev->external &&
553 conf->mddev->flags) {
554 /* It is very unlikely, but we might
555 * still need to write out the
556 * bad block log - better give it
558 md_check_recovery(conf->mddev);
560 md_wait_for_blocked_rdev(rdev, conf->mddev);
562 /* Acknowledged bad block - skip the write */
563 rdev_dec_pending(rdev, conf->mddev);
569 if (s->syncing || s->expanding || s->expanded)
570 md_sync_acct(rdev->bdev, STRIPE_SECTORS);
572 set_bit(STRIPE_IO_STARTED, &sh->state);
574 bi->bi_bdev = rdev->bdev;
575 pr_debug("%s: for %llu schedule op %ld on disc %d\n",
576 __func__, (unsigned long long)sh->sector,
578 atomic_inc(&sh->count);
579 bi->bi_sector = sh->sector + rdev->data_offset;
580 bi->bi_flags = 1 << BIO_UPTODATE;
582 bi->bi_io_vec[0].bv_len = STRIPE_SIZE;
583 bi->bi_io_vec[0].bv_offset = 0;
584 bi->bi_size = STRIPE_SIZE;
586 generic_make_request(bi);
589 set_bit(STRIPE_DEGRADED, &sh->state);
590 pr_debug("skip op %ld on disc %d for sector %llu\n",
591 bi->bi_rw, i, (unsigned long long)sh->sector);
592 clear_bit(R5_LOCKED, &sh->dev[i].flags);
593 set_bit(STRIPE_HANDLE, &sh->state);
598 static struct dma_async_tx_descriptor *
599 async_copy_data(int frombio, struct bio *bio, struct page *page,
600 sector_t sector, struct dma_async_tx_descriptor *tx)
603 struct page *bio_page;
606 struct async_submit_ctl submit;
607 enum async_tx_flags flags = 0;
609 if (bio->bi_sector >= sector)
610 page_offset = (signed)(bio->bi_sector - sector) * 512;
612 page_offset = (signed)(sector - bio->bi_sector) * -512;
615 flags |= ASYNC_TX_FENCE;
616 init_async_submit(&submit, flags, tx, NULL, NULL, NULL);
618 bio_for_each_segment(bvl, bio, i) {
619 int len = bvl->bv_len;
623 if (page_offset < 0) {
624 b_offset = -page_offset;
625 page_offset += b_offset;
629 if (len > 0 && page_offset + len > STRIPE_SIZE)
630 clen = STRIPE_SIZE - page_offset;
635 b_offset += bvl->bv_offset;
636 bio_page = bvl->bv_page;
638 tx = async_memcpy(page, bio_page, page_offset,
639 b_offset, clen, &submit);
641 tx = async_memcpy(bio_page, page, b_offset,
642 page_offset, clen, &submit);
644 /* chain the operations */
645 submit.depend_tx = tx;
647 if (clen < len) /* hit end of page */
655 static void ops_complete_biofill(void *stripe_head_ref)
657 struct stripe_head *sh = stripe_head_ref;
658 struct bio *return_bi = NULL;
659 struct r5conf *conf = sh->raid_conf;
662 pr_debug("%s: stripe %llu\n", __func__,
663 (unsigned long long)sh->sector);
665 /* clear completed biofills */
666 spin_lock_irq(&conf->device_lock);
667 for (i = sh->disks; i--; ) {
668 struct r5dev *dev = &sh->dev[i];
670 /* acknowledge completion of a biofill operation */
671 /* and check if we need to reply to a read request,
672 * new R5_Wantfill requests are held off until
673 * !STRIPE_BIOFILL_RUN
675 if (test_and_clear_bit(R5_Wantfill, &dev->flags)) {
676 struct bio *rbi, *rbi2;
681 while (rbi && rbi->bi_sector <
682 dev->sector + STRIPE_SECTORS) {
683 rbi2 = r5_next_bio(rbi, dev->sector);
684 if (!raid5_dec_bi_phys_segments(rbi)) {
685 rbi->bi_next = return_bi;
692 spin_unlock_irq(&conf->device_lock);
693 clear_bit(STRIPE_BIOFILL_RUN, &sh->state);
695 return_io(return_bi);
697 set_bit(STRIPE_HANDLE, &sh->state);
701 static void ops_run_biofill(struct stripe_head *sh)
703 struct dma_async_tx_descriptor *tx = NULL;
704 struct r5conf *conf = sh->raid_conf;
705 struct async_submit_ctl submit;
708 pr_debug("%s: stripe %llu\n", __func__,
709 (unsigned long long)sh->sector);
711 for (i = sh->disks; i--; ) {
712 struct r5dev *dev = &sh->dev[i];
713 if (test_bit(R5_Wantfill, &dev->flags)) {
715 spin_lock_irq(&conf->device_lock);
716 dev->read = rbi = dev->toread;
718 spin_unlock_irq(&conf->device_lock);
719 while (rbi && rbi->bi_sector <
720 dev->sector + STRIPE_SECTORS) {
721 tx = async_copy_data(0, rbi, dev->page,
723 rbi = r5_next_bio(rbi, dev->sector);
728 atomic_inc(&sh->count);
729 init_async_submit(&submit, ASYNC_TX_ACK, tx, ops_complete_biofill, sh, NULL);
730 async_trigger_callback(&submit);
733 static void mark_target_uptodate(struct stripe_head *sh, int target)
740 tgt = &sh->dev[target];
741 set_bit(R5_UPTODATE, &tgt->flags);
742 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
743 clear_bit(R5_Wantcompute, &tgt->flags);
746 static void ops_complete_compute(void *stripe_head_ref)
748 struct stripe_head *sh = stripe_head_ref;
750 pr_debug("%s: stripe %llu\n", __func__,
751 (unsigned long long)sh->sector);
753 /* mark the computed target(s) as uptodate */
754 mark_target_uptodate(sh, sh->ops.target);
755 mark_target_uptodate(sh, sh->ops.target2);
757 clear_bit(STRIPE_COMPUTE_RUN, &sh->state);
758 if (sh->check_state == check_state_compute_run)
759 sh->check_state = check_state_compute_result;
760 set_bit(STRIPE_HANDLE, &sh->state);
764 /* return a pointer to the address conversion region of the scribble buffer */
765 static addr_conv_t *to_addr_conv(struct stripe_head *sh,
766 struct raid5_percpu *percpu)
768 return percpu->scribble + sizeof(struct page *) * (sh->disks + 2);
771 static struct dma_async_tx_descriptor *
772 ops_run_compute5(struct stripe_head *sh, struct raid5_percpu *percpu)
774 int disks = sh->disks;
775 struct page **xor_srcs = percpu->scribble;
776 int target = sh->ops.target;
777 struct r5dev *tgt = &sh->dev[target];
778 struct page *xor_dest = tgt->page;
780 struct dma_async_tx_descriptor *tx;
781 struct async_submit_ctl submit;
784 pr_debug("%s: stripe %llu block: %d\n",
785 __func__, (unsigned long long)sh->sector, target);
786 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
788 for (i = disks; i--; )
790 xor_srcs[count++] = sh->dev[i].page;
792 atomic_inc(&sh->count);
794 init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST, NULL,
795 ops_complete_compute, sh, to_addr_conv(sh, percpu));
796 if (unlikely(count == 1))
797 tx = async_memcpy(xor_dest, xor_srcs[0], 0, 0, STRIPE_SIZE, &submit);
799 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
804 /* set_syndrome_sources - populate source buffers for gen_syndrome
805 * @srcs - (struct page *) array of size sh->disks
806 * @sh - stripe_head to parse
808 * Populates srcs in proper layout order for the stripe and returns the
809 * 'count' of sources to be used in a call to async_gen_syndrome. The P
810 * destination buffer is recorded in srcs[count] and the Q destination
811 * is recorded in srcs[count+1]].
813 static int set_syndrome_sources(struct page **srcs, struct stripe_head *sh)
815 int disks = sh->disks;
816 int syndrome_disks = sh->ddf_layout ? disks : (disks - 2);
817 int d0_idx = raid6_d0(sh);
821 for (i = 0; i < disks; i++)
827 int slot = raid6_idx_to_slot(i, sh, &count, syndrome_disks);
829 srcs[slot] = sh->dev[i].page;
830 i = raid6_next_disk(i, disks);
831 } while (i != d0_idx);
833 return syndrome_disks;
836 static struct dma_async_tx_descriptor *
837 ops_run_compute6_1(struct stripe_head *sh, struct raid5_percpu *percpu)
839 int disks = sh->disks;
840 struct page **blocks = percpu->scribble;
842 int qd_idx = sh->qd_idx;
843 struct dma_async_tx_descriptor *tx;
844 struct async_submit_ctl submit;
850 if (sh->ops.target < 0)
851 target = sh->ops.target2;
852 else if (sh->ops.target2 < 0)
853 target = sh->ops.target;
855 /* we should only have one valid target */
858 pr_debug("%s: stripe %llu block: %d\n",
859 __func__, (unsigned long long)sh->sector, target);
861 tgt = &sh->dev[target];
862 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
865 atomic_inc(&sh->count);
867 if (target == qd_idx) {
868 count = set_syndrome_sources(blocks, sh);
869 blocks[count] = NULL; /* regenerating p is not necessary */
870 BUG_ON(blocks[count+1] != dest); /* q should already be set */
871 init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
872 ops_complete_compute, sh,
873 to_addr_conv(sh, percpu));
874 tx = async_gen_syndrome(blocks, 0, count+2, STRIPE_SIZE, &submit);
876 /* Compute any data- or p-drive using XOR */
878 for (i = disks; i-- ; ) {
879 if (i == target || i == qd_idx)
881 blocks[count++] = sh->dev[i].page;
884 init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST,
885 NULL, ops_complete_compute, sh,
886 to_addr_conv(sh, percpu));
887 tx = async_xor(dest, blocks, 0, count, STRIPE_SIZE, &submit);
893 static struct dma_async_tx_descriptor *
894 ops_run_compute6_2(struct stripe_head *sh, struct raid5_percpu *percpu)
896 int i, count, disks = sh->disks;
897 int syndrome_disks = sh->ddf_layout ? disks : disks-2;
898 int d0_idx = raid6_d0(sh);
899 int faila = -1, failb = -1;
900 int target = sh->ops.target;
901 int target2 = sh->ops.target2;
902 struct r5dev *tgt = &sh->dev[target];
903 struct r5dev *tgt2 = &sh->dev[target2];
904 struct dma_async_tx_descriptor *tx;
905 struct page **blocks = percpu->scribble;
906 struct async_submit_ctl submit;
908 pr_debug("%s: stripe %llu block1: %d block2: %d\n",
909 __func__, (unsigned long long)sh->sector, target, target2);
910 BUG_ON(target < 0 || target2 < 0);
911 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
912 BUG_ON(!test_bit(R5_Wantcompute, &tgt2->flags));
914 /* we need to open-code set_syndrome_sources to handle the
915 * slot number conversion for 'faila' and 'failb'
917 for (i = 0; i < disks ; i++)
922 int slot = raid6_idx_to_slot(i, sh, &count, syndrome_disks);
924 blocks[slot] = sh->dev[i].page;
930 i = raid6_next_disk(i, disks);
931 } while (i != d0_idx);
933 BUG_ON(faila == failb);
936 pr_debug("%s: stripe: %llu faila: %d failb: %d\n",
937 __func__, (unsigned long long)sh->sector, faila, failb);
939 atomic_inc(&sh->count);
941 if (failb == syndrome_disks+1) {
942 /* Q disk is one of the missing disks */
943 if (faila == syndrome_disks) {
944 /* Missing P+Q, just recompute */
945 init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
946 ops_complete_compute, sh,
947 to_addr_conv(sh, percpu));
948 return async_gen_syndrome(blocks, 0, syndrome_disks+2,
949 STRIPE_SIZE, &submit);
953 int qd_idx = sh->qd_idx;
955 /* Missing D+Q: recompute D from P, then recompute Q */
956 if (target == qd_idx)
957 data_target = target2;
959 data_target = target;
962 for (i = disks; i-- ; ) {
963 if (i == data_target || i == qd_idx)
965 blocks[count++] = sh->dev[i].page;
967 dest = sh->dev[data_target].page;
968 init_async_submit(&submit,
969 ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST,
971 to_addr_conv(sh, percpu));
972 tx = async_xor(dest, blocks, 0, count, STRIPE_SIZE,
975 count = set_syndrome_sources(blocks, sh);
976 init_async_submit(&submit, ASYNC_TX_FENCE, tx,
977 ops_complete_compute, sh,
978 to_addr_conv(sh, percpu));
979 return async_gen_syndrome(blocks, 0, count+2,
980 STRIPE_SIZE, &submit);
983 init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
984 ops_complete_compute, sh,
985 to_addr_conv(sh, percpu));
986 if (failb == syndrome_disks) {
987 /* We're missing D+P. */
988 return async_raid6_datap_recov(syndrome_disks+2,
992 /* We're missing D+D. */
993 return async_raid6_2data_recov(syndrome_disks+2,
994 STRIPE_SIZE, faila, failb,
1001 static void ops_complete_prexor(void *stripe_head_ref)
1003 struct stripe_head *sh = stripe_head_ref;
1005 pr_debug("%s: stripe %llu\n", __func__,
1006 (unsigned long long)sh->sector);
1009 static struct dma_async_tx_descriptor *
1010 ops_run_prexor(struct stripe_head *sh, struct raid5_percpu *percpu,
1011 struct dma_async_tx_descriptor *tx)
1013 int disks = sh->disks;
1014 struct page **xor_srcs = percpu->scribble;
1015 int count = 0, pd_idx = sh->pd_idx, i;
1016 struct async_submit_ctl submit;
1018 /* existing parity data subtracted */
1019 struct page *xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
1021 pr_debug("%s: stripe %llu\n", __func__,
1022 (unsigned long long)sh->sector);
1024 for (i = disks; i--; ) {
1025 struct r5dev *dev = &sh->dev[i];
1026 /* Only process blocks that are known to be uptodate */
1027 if (test_bit(R5_Wantdrain, &dev->flags))
1028 xor_srcs[count++] = dev->page;
1031 init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_DROP_DST, tx,
1032 ops_complete_prexor, sh, to_addr_conv(sh, percpu));
1033 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
1038 static struct dma_async_tx_descriptor *
1039 ops_run_biodrain(struct stripe_head *sh, struct dma_async_tx_descriptor *tx)
1041 int disks = sh->disks;
1044 pr_debug("%s: stripe %llu\n", __func__,
1045 (unsigned long long)sh->sector);
1047 for (i = disks; i--; ) {
1048 struct r5dev *dev = &sh->dev[i];
1051 if (test_and_clear_bit(R5_Wantdrain, &dev->flags)) {
1054 spin_lock_irq(&sh->raid_conf->device_lock);
1055 chosen = dev->towrite;
1056 dev->towrite = NULL;
1057 BUG_ON(dev->written);
1058 wbi = dev->written = chosen;
1059 spin_unlock_irq(&sh->raid_conf->device_lock);
1061 while (wbi && wbi->bi_sector <
1062 dev->sector + STRIPE_SECTORS) {
1063 if (wbi->bi_rw & REQ_FUA)
1064 set_bit(R5_WantFUA, &dev->flags);
1065 tx = async_copy_data(1, wbi, dev->page,
1067 wbi = r5_next_bio(wbi, dev->sector);
1075 static void ops_complete_reconstruct(void *stripe_head_ref)
1077 struct stripe_head *sh = stripe_head_ref;
1078 int disks = sh->disks;
1079 int pd_idx = sh->pd_idx;
1080 int qd_idx = sh->qd_idx;
1084 pr_debug("%s: stripe %llu\n", __func__,
1085 (unsigned long long)sh->sector);
1087 for (i = disks; i--; )
1088 fua |= test_bit(R5_WantFUA, &sh->dev[i].flags);
1090 for (i = disks; i--; ) {
1091 struct r5dev *dev = &sh->dev[i];
1093 if (dev->written || i == pd_idx || i == qd_idx) {
1094 set_bit(R5_UPTODATE, &dev->flags);
1096 set_bit(R5_WantFUA, &dev->flags);
1100 if (sh->reconstruct_state == reconstruct_state_drain_run)
1101 sh->reconstruct_state = reconstruct_state_drain_result;
1102 else if (sh->reconstruct_state == reconstruct_state_prexor_drain_run)
1103 sh->reconstruct_state = reconstruct_state_prexor_drain_result;
1105 BUG_ON(sh->reconstruct_state != reconstruct_state_run);
1106 sh->reconstruct_state = reconstruct_state_result;
1109 set_bit(STRIPE_HANDLE, &sh->state);
1114 ops_run_reconstruct5(struct stripe_head *sh, struct raid5_percpu *percpu,
1115 struct dma_async_tx_descriptor *tx)
1117 int disks = sh->disks;
1118 struct page **xor_srcs = percpu->scribble;
1119 struct async_submit_ctl submit;
1120 int count = 0, pd_idx = sh->pd_idx, i;
1121 struct page *xor_dest;
1123 unsigned long flags;
1125 pr_debug("%s: stripe %llu\n", __func__,
1126 (unsigned long long)sh->sector);
1128 /* check if prexor is active which means only process blocks
1129 * that are part of a read-modify-write (written)
1131 if (sh->reconstruct_state == reconstruct_state_prexor_drain_run) {
1133 xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
1134 for (i = disks; i--; ) {
1135 struct r5dev *dev = &sh->dev[i];
1137 xor_srcs[count++] = dev->page;
1140 xor_dest = sh->dev[pd_idx].page;
1141 for (i = disks; i--; ) {
1142 struct r5dev *dev = &sh->dev[i];
1144 xor_srcs[count++] = dev->page;
1148 /* 1/ if we prexor'd then the dest is reused as a source
1149 * 2/ if we did not prexor then we are redoing the parity
1150 * set ASYNC_TX_XOR_DROP_DST and ASYNC_TX_XOR_ZERO_DST
1151 * for the synchronous xor case
1153 flags = ASYNC_TX_ACK |
1154 (prexor ? ASYNC_TX_XOR_DROP_DST : ASYNC_TX_XOR_ZERO_DST);
1156 atomic_inc(&sh->count);
1158 init_async_submit(&submit, flags, tx, ops_complete_reconstruct, sh,
1159 to_addr_conv(sh, percpu));
1160 if (unlikely(count == 1))
1161 tx = async_memcpy(xor_dest, xor_srcs[0], 0, 0, STRIPE_SIZE, &submit);
1163 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
1167 ops_run_reconstruct6(struct stripe_head *sh, struct raid5_percpu *percpu,
1168 struct dma_async_tx_descriptor *tx)
1170 struct async_submit_ctl submit;
1171 struct page **blocks = percpu->scribble;
1174 pr_debug("%s: stripe %llu\n", __func__, (unsigned long long)sh->sector);
1176 count = set_syndrome_sources(blocks, sh);
1178 atomic_inc(&sh->count);
1180 init_async_submit(&submit, ASYNC_TX_ACK, tx, ops_complete_reconstruct,
1181 sh, to_addr_conv(sh, percpu));
1182 async_gen_syndrome(blocks, 0, count+2, STRIPE_SIZE, &submit);
1185 static void ops_complete_check(void *stripe_head_ref)
1187 struct stripe_head *sh = stripe_head_ref;
1189 pr_debug("%s: stripe %llu\n", __func__,
1190 (unsigned long long)sh->sector);
1192 sh->check_state = check_state_check_result;
1193 set_bit(STRIPE_HANDLE, &sh->state);
1197 static void ops_run_check_p(struct stripe_head *sh, struct raid5_percpu *percpu)
1199 int disks = sh->disks;
1200 int pd_idx = sh->pd_idx;
1201 int qd_idx = sh->qd_idx;
1202 struct page *xor_dest;
1203 struct page **xor_srcs = percpu->scribble;
1204 struct dma_async_tx_descriptor *tx;
1205 struct async_submit_ctl submit;
1209 pr_debug("%s: stripe %llu\n", __func__,
1210 (unsigned long long)sh->sector);
1213 xor_dest = sh->dev[pd_idx].page;
1214 xor_srcs[count++] = xor_dest;
1215 for (i = disks; i--; ) {
1216 if (i == pd_idx || i == qd_idx)
1218 xor_srcs[count++] = sh->dev[i].page;
1221 init_async_submit(&submit, 0, NULL, NULL, NULL,
1222 to_addr_conv(sh, percpu));
1223 tx = async_xor_val(xor_dest, xor_srcs, 0, count, STRIPE_SIZE,
1224 &sh->ops.zero_sum_result, &submit);
1226 atomic_inc(&sh->count);
1227 init_async_submit(&submit, ASYNC_TX_ACK, tx, ops_complete_check, sh, NULL);
1228 tx = async_trigger_callback(&submit);
1231 static void ops_run_check_pq(struct stripe_head *sh, struct raid5_percpu *percpu, int checkp)
1233 struct page **srcs = percpu->scribble;
1234 struct async_submit_ctl submit;
1237 pr_debug("%s: stripe %llu checkp: %d\n", __func__,
1238 (unsigned long long)sh->sector, checkp);
1240 count = set_syndrome_sources(srcs, sh);
1244 atomic_inc(&sh->count);
1245 init_async_submit(&submit, ASYNC_TX_ACK, NULL, ops_complete_check,
1246 sh, to_addr_conv(sh, percpu));
1247 async_syndrome_val(srcs, 0, count+2, STRIPE_SIZE,
1248 &sh->ops.zero_sum_result, percpu->spare_page, &submit);
1251 static void __raid_run_ops(struct stripe_head *sh, unsigned long ops_request)
1253 int overlap_clear = 0, i, disks = sh->disks;
1254 struct dma_async_tx_descriptor *tx = NULL;
1255 struct r5conf *conf = sh->raid_conf;
1256 int level = conf->level;
1257 struct raid5_percpu *percpu;
1261 percpu = per_cpu_ptr(conf->percpu, cpu);
1262 if (test_bit(STRIPE_OP_BIOFILL, &ops_request)) {
1263 ops_run_biofill(sh);
1267 if (test_bit(STRIPE_OP_COMPUTE_BLK, &ops_request)) {
1269 tx = ops_run_compute5(sh, percpu);
1271 if (sh->ops.target2 < 0 || sh->ops.target < 0)
1272 tx = ops_run_compute6_1(sh, percpu);
1274 tx = ops_run_compute6_2(sh, percpu);
1276 /* terminate the chain if reconstruct is not set to be run */
1277 if (tx && !test_bit(STRIPE_OP_RECONSTRUCT, &ops_request))
1281 if (test_bit(STRIPE_OP_PREXOR, &ops_request))
1282 tx = ops_run_prexor(sh, percpu, tx);
1284 if (test_bit(STRIPE_OP_BIODRAIN, &ops_request)) {
1285 tx = ops_run_biodrain(sh, tx);
1289 if (test_bit(STRIPE_OP_RECONSTRUCT, &ops_request)) {
1291 ops_run_reconstruct5(sh, percpu, tx);
1293 ops_run_reconstruct6(sh, percpu, tx);
1296 if (test_bit(STRIPE_OP_CHECK, &ops_request)) {
1297 if (sh->check_state == check_state_run)
1298 ops_run_check_p(sh, percpu);
1299 else if (sh->check_state == check_state_run_q)
1300 ops_run_check_pq(sh, percpu, 0);
1301 else if (sh->check_state == check_state_run_pq)
1302 ops_run_check_pq(sh, percpu, 1);
1308 for (i = disks; i--; ) {
1309 struct r5dev *dev = &sh->dev[i];
1310 if (test_and_clear_bit(R5_Overlap, &dev->flags))
1311 wake_up(&sh->raid_conf->wait_for_overlap);
1316 #ifdef CONFIG_MULTICORE_RAID456
1317 static void async_run_ops(void *param, async_cookie_t cookie)
1319 struct stripe_head *sh = param;
1320 unsigned long ops_request = sh->ops.request;
1322 clear_bit_unlock(STRIPE_OPS_REQ_PENDING, &sh->state);
1323 wake_up(&sh->ops.wait_for_ops);
1325 __raid_run_ops(sh, ops_request);
1329 static void raid_run_ops(struct stripe_head *sh, unsigned long ops_request)
1331 /* since handle_stripe can be called outside of raid5d context
1332 * we need to ensure sh->ops.request is de-staged before another
1335 wait_event(sh->ops.wait_for_ops,
1336 !test_and_set_bit_lock(STRIPE_OPS_REQ_PENDING, &sh->state));
1337 sh->ops.request = ops_request;
1339 atomic_inc(&sh->count);
1340 async_schedule(async_run_ops, sh);
1343 #define raid_run_ops __raid_run_ops
1346 static int grow_one_stripe(struct r5conf *conf)
1348 struct stripe_head *sh;
1349 sh = kmem_cache_zalloc(conf->slab_cache, GFP_KERNEL);
1353 sh->raid_conf = conf;
1354 #ifdef CONFIG_MULTICORE_RAID456
1355 init_waitqueue_head(&sh->ops.wait_for_ops);
1358 if (grow_buffers(sh)) {
1360 kmem_cache_free(conf->slab_cache, sh);
1363 /* we just created an active stripe so... */
1364 atomic_set(&sh->count, 1);
1365 atomic_inc(&conf->active_stripes);
1366 INIT_LIST_HEAD(&sh->lru);
1371 static int grow_stripes(struct r5conf *conf, int num)
1373 struct kmem_cache *sc;
1374 int devs = max(conf->raid_disks, conf->previous_raid_disks);
1376 if (conf->mddev->gendisk)
1377 sprintf(conf->cache_name[0],
1378 "raid%d-%s", conf->level, mdname(conf->mddev));
1380 sprintf(conf->cache_name[0],
1381 "raid%d-%p", conf->level, conf->mddev);
1382 sprintf(conf->cache_name[1], "%s-alt", conf->cache_name[0]);
1384 conf->active_name = 0;
1385 sc = kmem_cache_create(conf->cache_name[conf->active_name],
1386 sizeof(struct stripe_head)+(devs-1)*sizeof(struct r5dev),
1390 conf->slab_cache = sc;
1391 conf->pool_size = devs;
1393 if (!grow_one_stripe(conf))
1399 * scribble_len - return the required size of the scribble region
1400 * @num - total number of disks in the array
1402 * The size must be enough to contain:
1403 * 1/ a struct page pointer for each device in the array +2
1404 * 2/ room to convert each entry in (1) to its corresponding dma
1405 * (dma_map_page()) or page (page_address()) address.
1407 * Note: the +2 is for the destination buffers of the ddf/raid6 case where we
1408 * calculate over all devices (not just the data blocks), using zeros in place
1409 * of the P and Q blocks.
1411 static size_t scribble_len(int num)
1415 len = sizeof(struct page *) * (num+2) + sizeof(addr_conv_t) * (num+2);
1420 static int resize_stripes(struct r5conf *conf, int newsize)
1422 /* Make all the stripes able to hold 'newsize' devices.
1423 * New slots in each stripe get 'page' set to a new page.
1425 * This happens in stages:
1426 * 1/ create a new kmem_cache and allocate the required number of
1428 * 2/ gather all the old stripe_heads and tranfer the pages across
1429 * to the new stripe_heads. This will have the side effect of
1430 * freezing the array as once all stripe_heads have been collected,
1431 * no IO will be possible. Old stripe heads are freed once their
1432 * pages have been transferred over, and the old kmem_cache is
1433 * freed when all stripes are done.
1434 * 3/ reallocate conf->disks to be suitable bigger. If this fails,
1435 * we simple return a failre status - no need to clean anything up.
1436 * 4/ allocate new pages for the new slots in the new stripe_heads.
1437 * If this fails, we don't bother trying the shrink the
1438 * stripe_heads down again, we just leave them as they are.
1439 * As each stripe_head is processed the new one is released into
1442 * Once step2 is started, we cannot afford to wait for a write,
1443 * so we use GFP_NOIO allocations.
1445 struct stripe_head *osh, *nsh;
1446 LIST_HEAD(newstripes);
1447 struct disk_info *ndisks;
1450 struct kmem_cache *sc;
1453 if (newsize <= conf->pool_size)
1454 return 0; /* never bother to shrink */
1456 err = md_allow_write(conf->mddev);
1461 sc = kmem_cache_create(conf->cache_name[1-conf->active_name],
1462 sizeof(struct stripe_head)+(newsize-1)*sizeof(struct r5dev),
1467 for (i = conf->max_nr_stripes; i; i--) {
1468 nsh = kmem_cache_zalloc(sc, GFP_KERNEL);
1472 nsh->raid_conf = conf;
1473 #ifdef CONFIG_MULTICORE_RAID456
1474 init_waitqueue_head(&nsh->ops.wait_for_ops);
1477 list_add(&nsh->lru, &newstripes);
1480 /* didn't get enough, give up */
1481 while (!list_empty(&newstripes)) {
1482 nsh = list_entry(newstripes.next, struct stripe_head, lru);
1483 list_del(&nsh->lru);
1484 kmem_cache_free(sc, nsh);
1486 kmem_cache_destroy(sc);
1489 /* Step 2 - Must use GFP_NOIO now.
1490 * OK, we have enough stripes, start collecting inactive
1491 * stripes and copying them over
1493 list_for_each_entry(nsh, &newstripes, lru) {
1494 spin_lock_irq(&conf->device_lock);
1495 wait_event_lock_irq(conf->wait_for_stripe,
1496 !list_empty(&conf->inactive_list),
1499 osh = get_free_stripe(conf);
1500 spin_unlock_irq(&conf->device_lock);
1501 atomic_set(&nsh->count, 1);
1502 for(i=0; i<conf->pool_size; i++)
1503 nsh->dev[i].page = osh->dev[i].page;
1504 for( ; i<newsize; i++)
1505 nsh->dev[i].page = NULL;
1506 kmem_cache_free(conf->slab_cache, osh);
1508 kmem_cache_destroy(conf->slab_cache);
1511 * At this point, we are holding all the stripes so the array
1512 * is completely stalled, so now is a good time to resize
1513 * conf->disks and the scribble region
1515 ndisks = kzalloc(newsize * sizeof(struct disk_info), GFP_NOIO);
1517 for (i=0; i<conf->raid_disks; i++)
1518 ndisks[i] = conf->disks[i];
1520 conf->disks = ndisks;
1525 conf->scribble_len = scribble_len(newsize);
1526 for_each_present_cpu(cpu) {
1527 struct raid5_percpu *percpu;
1530 percpu = per_cpu_ptr(conf->percpu, cpu);
1531 scribble = kmalloc(conf->scribble_len, GFP_NOIO);
1534 kfree(percpu->scribble);
1535 percpu->scribble = scribble;
1543 /* Step 4, return new stripes to service */
1544 while(!list_empty(&newstripes)) {
1545 nsh = list_entry(newstripes.next, struct stripe_head, lru);
1546 list_del_init(&nsh->lru);
1548 for (i=conf->raid_disks; i < newsize; i++)
1549 if (nsh->dev[i].page == NULL) {
1550 struct page *p = alloc_page(GFP_NOIO);
1551 nsh->dev[i].page = p;
1555 release_stripe(nsh);
1557 /* critical section pass, GFP_NOIO no longer needed */
1559 conf->slab_cache = sc;
1560 conf->active_name = 1-conf->active_name;
1561 conf->pool_size = newsize;
1565 static int drop_one_stripe(struct r5conf *conf)
1567 struct stripe_head *sh;
1569 spin_lock_irq(&conf->device_lock);
1570 sh = get_free_stripe(conf);
1571 spin_unlock_irq(&conf->device_lock);
1574 BUG_ON(atomic_read(&sh->count));
1576 kmem_cache_free(conf->slab_cache, sh);
1577 atomic_dec(&conf->active_stripes);
1581 static void shrink_stripes(struct r5conf *conf)
1583 while (drop_one_stripe(conf))
1586 if (conf->slab_cache)
1587 kmem_cache_destroy(conf->slab_cache);
1588 conf->slab_cache = NULL;
1591 static void raid5_end_read_request(struct bio * bi, int error)
1593 struct stripe_head *sh = bi->bi_private;
1594 struct r5conf *conf = sh->raid_conf;
1595 int disks = sh->disks, i;
1596 int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
1597 char b[BDEVNAME_SIZE];
1598 struct md_rdev *rdev;
1601 for (i=0 ; i<disks; i++)
1602 if (bi == &sh->dev[i].req)
1605 pr_debug("end_read_request %llu/%d, count: %d, uptodate %d.\n",
1606 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
1612 if (test_bit(R5_ReadRepl, &sh->dev[i].flags))
1613 rdev = conf->disks[i].replacement;
1615 rdev = conf->disks[i].rdev;
1618 set_bit(R5_UPTODATE, &sh->dev[i].flags);
1619 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
1620 /* Note that this cannot happen on a
1621 * replacement device. We just fail those on
1626 "md/raid:%s: read error corrected"
1627 " (%lu sectors at %llu on %s)\n",
1628 mdname(conf->mddev), STRIPE_SECTORS,
1629 (unsigned long long)(sh->sector
1630 + rdev->data_offset),
1631 bdevname(rdev->bdev, b));
1632 atomic_add(STRIPE_SECTORS, &rdev->corrected_errors);
1633 clear_bit(R5_ReadError, &sh->dev[i].flags);
1634 clear_bit(R5_ReWrite, &sh->dev[i].flags);
1636 if (atomic_read(&rdev->read_errors))
1637 atomic_set(&rdev->read_errors, 0);
1639 const char *bdn = bdevname(rdev->bdev, b);
1642 clear_bit(R5_UPTODATE, &sh->dev[i].flags);
1643 atomic_inc(&rdev->read_errors);
1644 if (test_bit(R5_ReadRepl, &sh->dev[i].flags))
1647 "md/raid:%s: read error on replacement device "
1648 "(sector %llu on %s).\n",
1649 mdname(conf->mddev),
1650 (unsigned long long)(sh->sector
1651 + rdev->data_offset),
1653 else if (conf->mddev->degraded >= conf->max_degraded)
1656 "md/raid:%s: read error not correctable "
1657 "(sector %llu on %s).\n",
1658 mdname(conf->mddev),
1659 (unsigned long long)(sh->sector
1660 + rdev->data_offset),
1662 else if (test_bit(R5_ReWrite, &sh->dev[i].flags))
1666 "md/raid:%s: read error NOT corrected!! "
1667 "(sector %llu on %s).\n",
1668 mdname(conf->mddev),
1669 (unsigned long long)(sh->sector
1670 + rdev->data_offset),
1672 else if (atomic_read(&rdev->read_errors)
1673 > conf->max_nr_stripes)
1675 "md/raid:%s: Too many read errors, failing device %s.\n",
1676 mdname(conf->mddev), bdn);
1680 set_bit(R5_ReadError, &sh->dev[i].flags);
1682 clear_bit(R5_ReadError, &sh->dev[i].flags);
1683 clear_bit(R5_ReWrite, &sh->dev[i].flags);
1684 md_error(conf->mddev, rdev);
1687 rdev_dec_pending(rdev, conf->mddev);
1688 clear_bit(R5_LOCKED, &sh->dev[i].flags);
1689 set_bit(STRIPE_HANDLE, &sh->state);
1693 static void raid5_end_write_request(struct bio *bi, int error)
1695 struct stripe_head *sh = bi->bi_private;
1696 struct r5conf *conf = sh->raid_conf;
1697 int disks = sh->disks, i;
1698 int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
1702 for (i=0 ; i<disks; i++)
1703 if (bi == &sh->dev[i].req)
1706 pr_debug("end_write_request %llu/%d, count %d, uptodate: %d.\n",
1707 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
1715 set_bit(WriteErrorSeen, &conf->disks[i].rdev->flags);
1716 set_bit(R5_WriteError, &sh->dev[i].flags);
1717 } else if (is_badblock(conf->disks[i].rdev, sh->sector, STRIPE_SECTORS,
1718 &first_bad, &bad_sectors))
1719 set_bit(R5_MadeGood, &sh->dev[i].flags);
1721 rdev_dec_pending(conf->disks[i].rdev, conf->mddev);
1723 clear_bit(R5_LOCKED, &sh->dev[i].flags);
1724 set_bit(STRIPE_HANDLE, &sh->state);
1729 static sector_t compute_blocknr(struct stripe_head *sh, int i, int previous);
1731 static void raid5_build_block(struct stripe_head *sh, int i, int previous)
1733 struct r5dev *dev = &sh->dev[i];
1735 bio_init(&dev->req);
1736 dev->req.bi_io_vec = &dev->vec;
1738 dev->req.bi_max_vecs++;
1739 dev->req.bi_private = sh;
1740 dev->vec.bv_page = dev->page;
1743 dev->sector = compute_blocknr(sh, i, previous);
1746 static void error(struct mddev *mddev, struct md_rdev *rdev)
1748 char b[BDEVNAME_SIZE];
1749 struct r5conf *conf = mddev->private;
1750 unsigned long flags;
1751 pr_debug("raid456: error called\n");
1753 spin_lock_irqsave(&conf->device_lock, flags);
1754 clear_bit(In_sync, &rdev->flags);
1755 mddev->degraded = calc_degraded(conf);
1756 spin_unlock_irqrestore(&conf->device_lock, flags);
1757 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1759 set_bit(Blocked, &rdev->flags);
1760 set_bit(Faulty, &rdev->flags);
1761 set_bit(MD_CHANGE_DEVS, &mddev->flags);
1763 "md/raid:%s: Disk failure on %s, disabling device.\n"
1764 "md/raid:%s: Operation continuing on %d devices.\n",
1766 bdevname(rdev->bdev, b),
1768 conf->raid_disks - mddev->degraded);
1772 * Input: a 'big' sector number,
1773 * Output: index of the data and parity disk, and the sector # in them.
1775 static sector_t raid5_compute_sector(struct r5conf *conf, sector_t r_sector,
1776 int previous, int *dd_idx,
1777 struct stripe_head *sh)
1779 sector_t stripe, stripe2;
1780 sector_t chunk_number;
1781 unsigned int chunk_offset;
1784 sector_t new_sector;
1785 int algorithm = previous ? conf->prev_algo
1787 int sectors_per_chunk = previous ? conf->prev_chunk_sectors
1788 : conf->chunk_sectors;
1789 int raid_disks = previous ? conf->previous_raid_disks
1791 int data_disks = raid_disks - conf->max_degraded;
1793 /* First compute the information on this sector */
1796 * Compute the chunk number and the sector offset inside the chunk
1798 chunk_offset = sector_div(r_sector, sectors_per_chunk);
1799 chunk_number = r_sector;
1802 * Compute the stripe number
1804 stripe = chunk_number;
1805 *dd_idx = sector_div(stripe, data_disks);
1808 * Select the parity disk based on the user selected algorithm.
1810 pd_idx = qd_idx = -1;
1811 switch(conf->level) {
1813 pd_idx = data_disks;
1816 switch (algorithm) {
1817 case ALGORITHM_LEFT_ASYMMETRIC:
1818 pd_idx = data_disks - sector_div(stripe2, raid_disks);
1819 if (*dd_idx >= pd_idx)
1822 case ALGORITHM_RIGHT_ASYMMETRIC:
1823 pd_idx = sector_div(stripe2, raid_disks);
1824 if (*dd_idx >= pd_idx)
1827 case ALGORITHM_LEFT_SYMMETRIC:
1828 pd_idx = data_disks - sector_div(stripe2, raid_disks);
1829 *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
1831 case ALGORITHM_RIGHT_SYMMETRIC:
1832 pd_idx = sector_div(stripe2, raid_disks);
1833 *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
1835 case ALGORITHM_PARITY_0:
1839 case ALGORITHM_PARITY_N:
1840 pd_idx = data_disks;
1848 switch (algorithm) {
1849 case ALGORITHM_LEFT_ASYMMETRIC:
1850 pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
1851 qd_idx = pd_idx + 1;
1852 if (pd_idx == raid_disks-1) {
1853 (*dd_idx)++; /* Q D D D P */
1855 } else if (*dd_idx >= pd_idx)
1856 (*dd_idx) += 2; /* D D P Q D */
1858 case ALGORITHM_RIGHT_ASYMMETRIC:
1859 pd_idx = sector_div(stripe2, raid_disks);
1860 qd_idx = pd_idx + 1;
1861 if (pd_idx == raid_disks-1) {
1862 (*dd_idx)++; /* Q D D D P */
1864 } else if (*dd_idx >= pd_idx)
1865 (*dd_idx) += 2; /* D D P Q D */
1867 case ALGORITHM_LEFT_SYMMETRIC:
1868 pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
1869 qd_idx = (pd_idx + 1) % raid_disks;
1870 *dd_idx = (pd_idx + 2 + *dd_idx) % raid_disks;
1872 case ALGORITHM_RIGHT_SYMMETRIC:
1873 pd_idx = sector_div(stripe2, raid_disks);
1874 qd_idx = (pd_idx + 1) % raid_disks;
1875 *dd_idx = (pd_idx + 2 + *dd_idx) % raid_disks;
1878 case ALGORITHM_PARITY_0:
1883 case ALGORITHM_PARITY_N:
1884 pd_idx = data_disks;
1885 qd_idx = data_disks + 1;
1888 case ALGORITHM_ROTATING_ZERO_RESTART:
1889 /* Exactly the same as RIGHT_ASYMMETRIC, but or
1890 * of blocks for computing Q is different.
1892 pd_idx = sector_div(stripe2, raid_disks);
1893 qd_idx = pd_idx + 1;
1894 if (pd_idx == raid_disks-1) {
1895 (*dd_idx)++; /* Q D D D P */
1897 } else if (*dd_idx >= pd_idx)
1898 (*dd_idx) += 2; /* D D P Q D */
1902 case ALGORITHM_ROTATING_N_RESTART:
1903 /* Same a left_asymmetric, by first stripe is
1904 * D D D P Q rather than
1908 pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
1909 qd_idx = pd_idx + 1;
1910 if (pd_idx == raid_disks-1) {
1911 (*dd_idx)++; /* Q D D D P */
1913 } else if (*dd_idx >= pd_idx)
1914 (*dd_idx) += 2; /* D D P Q D */
1918 case ALGORITHM_ROTATING_N_CONTINUE:
1919 /* Same as left_symmetric but Q is before P */
1920 pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
1921 qd_idx = (pd_idx + raid_disks - 1) % raid_disks;
1922 *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
1926 case ALGORITHM_LEFT_ASYMMETRIC_6:
1927 /* RAID5 left_asymmetric, with Q on last device */
1928 pd_idx = data_disks - sector_div(stripe2, raid_disks-1);
1929 if (*dd_idx >= pd_idx)
1931 qd_idx = raid_disks - 1;
1934 case ALGORITHM_RIGHT_ASYMMETRIC_6:
1935 pd_idx = sector_div(stripe2, raid_disks-1);
1936 if (*dd_idx >= pd_idx)
1938 qd_idx = raid_disks - 1;
1941 case ALGORITHM_LEFT_SYMMETRIC_6:
1942 pd_idx = data_disks - sector_div(stripe2, raid_disks-1);
1943 *dd_idx = (pd_idx + 1 + *dd_idx) % (raid_disks-1);
1944 qd_idx = raid_disks - 1;
1947 case ALGORITHM_RIGHT_SYMMETRIC_6:
1948 pd_idx = sector_div(stripe2, raid_disks-1);
1949 *dd_idx = (pd_idx + 1 + *dd_idx) % (raid_disks-1);
1950 qd_idx = raid_disks - 1;
1953 case ALGORITHM_PARITY_0_6:
1956 qd_idx = raid_disks - 1;
1966 sh->pd_idx = pd_idx;
1967 sh->qd_idx = qd_idx;
1968 sh->ddf_layout = ddf_layout;
1971 * Finally, compute the new sector number
1973 new_sector = (sector_t)stripe * sectors_per_chunk + chunk_offset;
1978 static sector_t compute_blocknr(struct stripe_head *sh, int i, int previous)
1980 struct r5conf *conf = sh->raid_conf;
1981 int raid_disks = sh->disks;
1982 int data_disks = raid_disks - conf->max_degraded;
1983 sector_t new_sector = sh->sector, check;
1984 int sectors_per_chunk = previous ? conf->prev_chunk_sectors
1985 : conf->chunk_sectors;
1986 int algorithm = previous ? conf->prev_algo
1990 sector_t chunk_number;
1991 int dummy1, dd_idx = i;
1993 struct stripe_head sh2;
1996 chunk_offset = sector_div(new_sector, sectors_per_chunk);
1997 stripe = new_sector;
1999 if (i == sh->pd_idx)
2001 switch(conf->level) {
2004 switch (algorithm) {
2005 case ALGORITHM_LEFT_ASYMMETRIC:
2006 case ALGORITHM_RIGHT_ASYMMETRIC:
2010 case ALGORITHM_LEFT_SYMMETRIC:
2011 case ALGORITHM_RIGHT_SYMMETRIC:
2014 i -= (sh->pd_idx + 1);
2016 case ALGORITHM_PARITY_0:
2019 case ALGORITHM_PARITY_N:
2026 if (i == sh->qd_idx)
2027 return 0; /* It is the Q disk */
2028 switch (algorithm) {
2029 case ALGORITHM_LEFT_ASYMMETRIC:
2030 case ALGORITHM_RIGHT_ASYMMETRIC:
2031 case ALGORITHM_ROTATING_ZERO_RESTART:
2032 case ALGORITHM_ROTATING_N_RESTART:
2033 if (sh->pd_idx == raid_disks-1)
2034 i--; /* Q D D D P */
2035 else if (i > sh->pd_idx)
2036 i -= 2; /* D D P Q D */
2038 case ALGORITHM_LEFT_SYMMETRIC:
2039 case ALGORITHM_RIGHT_SYMMETRIC:
2040 if (sh->pd_idx == raid_disks-1)
2041 i--; /* Q D D D P */
2046 i -= (sh->pd_idx + 2);
2049 case ALGORITHM_PARITY_0:
2052 case ALGORITHM_PARITY_N:
2054 case ALGORITHM_ROTATING_N_CONTINUE:
2055 /* Like left_symmetric, but P is before Q */
2056 if (sh->pd_idx == 0)
2057 i--; /* P D D D Q */
2062 i -= (sh->pd_idx + 1);
2065 case ALGORITHM_LEFT_ASYMMETRIC_6:
2066 case ALGORITHM_RIGHT_ASYMMETRIC_6:
2070 case ALGORITHM_LEFT_SYMMETRIC_6:
2071 case ALGORITHM_RIGHT_SYMMETRIC_6:
2073 i += data_disks + 1;
2074 i -= (sh->pd_idx + 1);
2076 case ALGORITHM_PARITY_0_6:
2085 chunk_number = stripe * data_disks + i;
2086 r_sector = chunk_number * sectors_per_chunk + chunk_offset;
2088 check = raid5_compute_sector(conf, r_sector,
2089 previous, &dummy1, &sh2);
2090 if (check != sh->sector || dummy1 != dd_idx || sh2.pd_idx != sh->pd_idx
2091 || sh2.qd_idx != sh->qd_idx) {
2092 printk(KERN_ERR "md/raid:%s: compute_blocknr: map not correct\n",
2093 mdname(conf->mddev));
2101 schedule_reconstruction(struct stripe_head *sh, struct stripe_head_state *s,
2102 int rcw, int expand)
2104 int i, pd_idx = sh->pd_idx, disks = sh->disks;
2105 struct r5conf *conf = sh->raid_conf;
2106 int level = conf->level;
2109 /* if we are not expanding this is a proper write request, and
2110 * there will be bios with new data to be drained into the
2114 sh->reconstruct_state = reconstruct_state_drain_run;
2115 set_bit(STRIPE_OP_BIODRAIN, &s->ops_request);
2117 sh->reconstruct_state = reconstruct_state_run;
2119 set_bit(STRIPE_OP_RECONSTRUCT, &s->ops_request);
2121 for (i = disks; i--; ) {
2122 struct r5dev *dev = &sh->dev[i];
2125 set_bit(R5_LOCKED, &dev->flags);
2126 set_bit(R5_Wantdrain, &dev->flags);
2128 clear_bit(R5_UPTODATE, &dev->flags);
2132 if (s->locked + conf->max_degraded == disks)
2133 if (!test_and_set_bit(STRIPE_FULL_WRITE, &sh->state))
2134 atomic_inc(&conf->pending_full_writes);
2137 BUG_ON(!(test_bit(R5_UPTODATE, &sh->dev[pd_idx].flags) ||
2138 test_bit(R5_Wantcompute, &sh->dev[pd_idx].flags)));
2140 sh->reconstruct_state = reconstruct_state_prexor_drain_run;
2141 set_bit(STRIPE_OP_PREXOR, &s->ops_request);
2142 set_bit(STRIPE_OP_BIODRAIN, &s->ops_request);
2143 set_bit(STRIPE_OP_RECONSTRUCT, &s->ops_request);
2145 for (i = disks; i--; ) {
2146 struct r5dev *dev = &sh->dev[i];
2151 (test_bit(R5_UPTODATE, &dev->flags) ||
2152 test_bit(R5_Wantcompute, &dev->flags))) {
2153 set_bit(R5_Wantdrain, &dev->flags);
2154 set_bit(R5_LOCKED, &dev->flags);
2155 clear_bit(R5_UPTODATE, &dev->flags);
2161 /* keep the parity disk(s) locked while asynchronous operations
2164 set_bit(R5_LOCKED, &sh->dev[pd_idx].flags);
2165 clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
2169 int qd_idx = sh->qd_idx;
2170 struct r5dev *dev = &sh->dev[qd_idx];
2172 set_bit(R5_LOCKED, &dev->flags);
2173 clear_bit(R5_UPTODATE, &dev->flags);
2177 pr_debug("%s: stripe %llu locked: %d ops_request: %lx\n",
2178 __func__, (unsigned long long)sh->sector,
2179 s->locked, s->ops_request);
2183 * Each stripe/dev can have one or more bion attached.
2184 * toread/towrite point to the first in a chain.
2185 * The bi_next chain must be in order.
2187 static int add_stripe_bio(struct stripe_head *sh, struct bio *bi, int dd_idx, int forwrite)
2190 struct r5conf *conf = sh->raid_conf;
2193 pr_debug("adding bi b#%llu to stripe s#%llu\n",
2194 (unsigned long long)bi->bi_sector,
2195 (unsigned long long)sh->sector);
2198 spin_lock_irq(&conf->device_lock);
2200 bip = &sh->dev[dd_idx].towrite;
2201 if (*bip == NULL && sh->dev[dd_idx].written == NULL)
2204 bip = &sh->dev[dd_idx].toread;
2205 while (*bip && (*bip)->bi_sector < bi->bi_sector) {
2206 if ((*bip)->bi_sector + ((*bip)->bi_size >> 9) > bi->bi_sector)
2208 bip = & (*bip)->bi_next;
2210 if (*bip && (*bip)->bi_sector < bi->bi_sector + ((bi->bi_size)>>9))
2213 BUG_ON(*bip && bi->bi_next && (*bip) != bi->bi_next);
2217 bi->bi_phys_segments++;
2220 /* check if page is covered */
2221 sector_t sector = sh->dev[dd_idx].sector;
2222 for (bi=sh->dev[dd_idx].towrite;
2223 sector < sh->dev[dd_idx].sector + STRIPE_SECTORS &&
2224 bi && bi->bi_sector <= sector;
2225 bi = r5_next_bio(bi, sh->dev[dd_idx].sector)) {
2226 if (bi->bi_sector + (bi->bi_size>>9) >= sector)
2227 sector = bi->bi_sector + (bi->bi_size>>9);
2229 if (sector >= sh->dev[dd_idx].sector + STRIPE_SECTORS)
2230 set_bit(R5_OVERWRITE, &sh->dev[dd_idx].flags);
2232 spin_unlock_irq(&conf->device_lock);
2234 pr_debug("added bi b#%llu to stripe s#%llu, disk %d.\n",
2235 (unsigned long long)(*bip)->bi_sector,
2236 (unsigned long long)sh->sector, dd_idx);
2238 if (conf->mddev->bitmap && firstwrite) {
2239 bitmap_startwrite(conf->mddev->bitmap, sh->sector,
2241 sh->bm_seq = conf->seq_flush+1;
2242 set_bit(STRIPE_BIT_DELAY, &sh->state);
2247 set_bit(R5_Overlap, &sh->dev[dd_idx].flags);
2248 spin_unlock_irq(&conf->device_lock);
2252 static void end_reshape(struct r5conf *conf);
2254 static void stripe_set_idx(sector_t stripe, struct r5conf *conf, int previous,
2255 struct stripe_head *sh)
2257 int sectors_per_chunk =
2258 previous ? conf->prev_chunk_sectors : conf->chunk_sectors;
2260 int chunk_offset = sector_div(stripe, sectors_per_chunk);
2261 int disks = previous ? conf->previous_raid_disks : conf->raid_disks;
2263 raid5_compute_sector(conf,
2264 stripe * (disks - conf->max_degraded)
2265 *sectors_per_chunk + chunk_offset,
2271 handle_failed_stripe(struct r5conf *conf, struct stripe_head *sh,
2272 struct stripe_head_state *s, int disks,
2273 struct bio **return_bi)
2276 for (i = disks; i--; ) {
2280 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
2281 struct md_rdev *rdev;
2283 rdev = rcu_dereference(conf->disks[i].rdev);
2284 if (rdev && test_bit(In_sync, &rdev->flags))
2285 atomic_inc(&rdev->nr_pending);
2290 if (!rdev_set_badblocks(
2294 md_error(conf->mddev, rdev);
2295 rdev_dec_pending(rdev, conf->mddev);
2298 spin_lock_irq(&conf->device_lock);
2299 /* fail all writes first */
2300 bi = sh->dev[i].towrite;
2301 sh->dev[i].towrite = NULL;
2307 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
2308 wake_up(&conf->wait_for_overlap);
2310 while (bi && bi->bi_sector <
2311 sh->dev[i].sector + STRIPE_SECTORS) {
2312 struct bio *nextbi = r5_next_bio(bi, sh->dev[i].sector);
2313 clear_bit(BIO_UPTODATE, &bi->bi_flags);
2314 if (!raid5_dec_bi_phys_segments(bi)) {
2315 md_write_end(conf->mddev);
2316 bi->bi_next = *return_bi;
2321 /* and fail all 'written' */
2322 bi = sh->dev[i].written;
2323 sh->dev[i].written = NULL;
2324 if (bi) bitmap_end = 1;
2325 while (bi && bi->bi_sector <
2326 sh->dev[i].sector + STRIPE_SECTORS) {
2327 struct bio *bi2 = r5_next_bio(bi, sh->dev[i].sector);
2328 clear_bit(BIO_UPTODATE, &bi->bi_flags);
2329 if (!raid5_dec_bi_phys_segments(bi)) {
2330 md_write_end(conf->mddev);
2331 bi->bi_next = *return_bi;
2337 /* fail any reads if this device is non-operational and
2338 * the data has not reached the cache yet.
2340 if (!test_bit(R5_Wantfill, &sh->dev[i].flags) &&
2341 (!test_bit(R5_Insync, &sh->dev[i].flags) ||
2342 test_bit(R5_ReadError, &sh->dev[i].flags))) {
2343 bi = sh->dev[i].toread;
2344 sh->dev[i].toread = NULL;
2345 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
2346 wake_up(&conf->wait_for_overlap);
2347 if (bi) s->to_read--;
2348 while (bi && bi->bi_sector <
2349 sh->dev[i].sector + STRIPE_SECTORS) {
2350 struct bio *nextbi =
2351 r5_next_bio(bi, sh->dev[i].sector);
2352 clear_bit(BIO_UPTODATE, &bi->bi_flags);
2353 if (!raid5_dec_bi_phys_segments(bi)) {
2354 bi->bi_next = *return_bi;
2360 spin_unlock_irq(&conf->device_lock);
2362 bitmap_endwrite(conf->mddev->bitmap, sh->sector,
2363 STRIPE_SECTORS, 0, 0);
2364 /* If we were in the middle of a write the parity block might
2365 * still be locked - so just clear all R5_LOCKED flags
2367 clear_bit(R5_LOCKED, &sh->dev[i].flags);
2370 if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
2371 if (atomic_dec_and_test(&conf->pending_full_writes))
2372 md_wakeup_thread(conf->mddev->thread);
2376 handle_failed_sync(struct r5conf *conf, struct stripe_head *sh,
2377 struct stripe_head_state *s)
2382 md_done_sync(conf->mddev, STRIPE_SECTORS, 0);
2383 clear_bit(STRIPE_SYNCING, &sh->state);
2385 /* There is nothing more to do for sync/check/repair.
2386 * For recover we need to record a bad block on all
2387 * non-sync devices, or abort the recovery
2389 if (!test_bit(MD_RECOVERY_RECOVER, &conf->mddev->recovery))
2391 /* During recovery devices cannot be removed, so locking and
2392 * refcounting of rdevs is not needed
2394 for (i = 0; i < conf->raid_disks; i++) {
2395 struct md_rdev *rdev = conf->disks[i].rdev;
2397 || test_bit(Faulty, &rdev->flags)
2398 || test_bit(In_sync, &rdev->flags))
2400 if (!rdev_set_badblocks(rdev, sh->sector,
2405 conf->recovery_disabled = conf->mddev->recovery_disabled;
2406 set_bit(MD_RECOVERY_INTR, &conf->mddev->recovery);
2410 /* fetch_block - checks the given member device to see if its data needs
2411 * to be read or computed to satisfy a request.
2413 * Returns 1 when no more member devices need to be checked, otherwise returns
2414 * 0 to tell the loop in handle_stripe_fill to continue
2416 static int fetch_block(struct stripe_head *sh, struct stripe_head_state *s,
2417 int disk_idx, int disks)
2419 struct r5dev *dev = &sh->dev[disk_idx];
2420 struct r5dev *fdev[2] = { &sh->dev[s->failed_num[0]],
2421 &sh->dev[s->failed_num[1]] };
2423 /* is the data in this block needed, and can we get it? */
2424 if (!test_bit(R5_LOCKED, &dev->flags) &&
2425 !test_bit(R5_UPTODATE, &dev->flags) &&
2427 (dev->towrite && !test_bit(R5_OVERWRITE, &dev->flags)) ||
2428 s->syncing || s->expanding ||
2429 (s->failed >= 1 && fdev[0]->toread) ||
2430 (s->failed >= 2 && fdev[1]->toread) ||
2431 (sh->raid_conf->level <= 5 && s->failed && fdev[0]->towrite &&
2432 !test_bit(R5_OVERWRITE, &fdev[0]->flags)) ||
2433 (sh->raid_conf->level == 6 && s->failed && s->to_write))) {
2434 /* we would like to get this block, possibly by computing it,
2435 * otherwise read it if the backing disk is insync
2437 BUG_ON(test_bit(R5_Wantcompute, &dev->flags));
2438 BUG_ON(test_bit(R5_Wantread, &dev->flags));
2439 if ((s->uptodate == disks - 1) &&
2440 (s->failed && (disk_idx == s->failed_num[0] ||
2441 disk_idx == s->failed_num[1]))) {
2442 /* have disk failed, and we're requested to fetch it;
2445 pr_debug("Computing stripe %llu block %d\n",
2446 (unsigned long long)sh->sector, disk_idx);
2447 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
2448 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
2449 set_bit(R5_Wantcompute, &dev->flags);
2450 sh->ops.target = disk_idx;
2451 sh->ops.target2 = -1; /* no 2nd target */
2453 /* Careful: from this point on 'uptodate' is in the eye
2454 * of raid_run_ops which services 'compute' operations
2455 * before writes. R5_Wantcompute flags a block that will
2456 * be R5_UPTODATE by the time it is needed for a
2457 * subsequent operation.
2461 } else if (s->uptodate == disks-2 && s->failed >= 2) {
2462 /* Computing 2-failure is *very* expensive; only
2463 * do it if failed >= 2
2466 for (other = disks; other--; ) {
2467 if (other == disk_idx)
2469 if (!test_bit(R5_UPTODATE,
2470 &sh->dev[other].flags))
2474 pr_debug("Computing stripe %llu blocks %d,%d\n",
2475 (unsigned long long)sh->sector,
2477 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
2478 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
2479 set_bit(R5_Wantcompute, &sh->dev[disk_idx].flags);
2480 set_bit(R5_Wantcompute, &sh->dev[other].flags);
2481 sh->ops.target = disk_idx;
2482 sh->ops.target2 = other;
2486 } else if (test_bit(R5_Insync, &dev->flags)) {
2487 set_bit(R5_LOCKED, &dev->flags);
2488 set_bit(R5_Wantread, &dev->flags);
2490 pr_debug("Reading block %d (sync=%d)\n",
2491 disk_idx, s->syncing);
2499 * handle_stripe_fill - read or compute data to satisfy pending requests.
2501 static void handle_stripe_fill(struct stripe_head *sh,
2502 struct stripe_head_state *s,
2507 /* look for blocks to read/compute, skip this if a compute
2508 * is already in flight, or if the stripe contents are in the
2509 * midst of changing due to a write
2511 if (!test_bit(STRIPE_COMPUTE_RUN, &sh->state) && !sh->check_state &&
2512 !sh->reconstruct_state)
2513 for (i = disks; i--; )
2514 if (fetch_block(sh, s, i, disks))
2516 set_bit(STRIPE_HANDLE, &sh->state);
2520 /* handle_stripe_clean_event
2521 * any written block on an uptodate or failed drive can be returned.
2522 * Note that if we 'wrote' to a failed drive, it will be UPTODATE, but
2523 * never LOCKED, so we don't need to test 'failed' directly.
2525 static void handle_stripe_clean_event(struct r5conf *conf,
2526 struct stripe_head *sh, int disks, struct bio **return_bi)
2531 for (i = disks; i--; )
2532 if (sh->dev[i].written) {
2534 if (!test_bit(R5_LOCKED, &dev->flags) &&
2535 test_bit(R5_UPTODATE, &dev->flags)) {
2536 /* We can return any write requests */
2537 struct bio *wbi, *wbi2;
2539 pr_debug("Return write for disc %d\n", i);
2540 spin_lock_irq(&conf->device_lock);
2542 dev->written = NULL;
2543 while (wbi && wbi->bi_sector <
2544 dev->sector + STRIPE_SECTORS) {
2545 wbi2 = r5_next_bio(wbi, dev->sector);
2546 if (!raid5_dec_bi_phys_segments(wbi)) {
2547 md_write_end(conf->mddev);
2548 wbi->bi_next = *return_bi;
2553 if (dev->towrite == NULL)
2555 spin_unlock_irq(&conf->device_lock);
2557 bitmap_endwrite(conf->mddev->bitmap,
2560 !test_bit(STRIPE_DEGRADED, &sh->state),
2565 if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
2566 if (atomic_dec_and_test(&conf->pending_full_writes))
2567 md_wakeup_thread(conf->mddev->thread);
2570 static void handle_stripe_dirtying(struct r5conf *conf,
2571 struct stripe_head *sh,
2572 struct stripe_head_state *s,
2575 int rmw = 0, rcw = 0, i;
2576 if (conf->max_degraded == 2) {
2577 /* RAID6 requires 'rcw' in current implementation
2578 * Calculate the real rcw later - for now fake it
2579 * look like rcw is cheaper
2582 } else for (i = disks; i--; ) {
2583 /* would I have to read this buffer for read_modify_write */
2584 struct r5dev *dev = &sh->dev[i];
2585 if ((dev->towrite || i == sh->pd_idx) &&
2586 !test_bit(R5_LOCKED, &dev->flags) &&
2587 !(test_bit(R5_UPTODATE, &dev->flags) ||
2588 test_bit(R5_Wantcompute, &dev->flags))) {
2589 if (test_bit(R5_Insync, &dev->flags))
2592 rmw += 2*disks; /* cannot read it */
2594 /* Would I have to read this buffer for reconstruct_write */
2595 if (!test_bit(R5_OVERWRITE, &dev->flags) && i != sh->pd_idx &&
2596 !test_bit(R5_LOCKED, &dev->flags) &&
2597 !(test_bit(R5_UPTODATE, &dev->flags) ||
2598 test_bit(R5_Wantcompute, &dev->flags))) {
2599 if (test_bit(R5_Insync, &dev->flags)) rcw++;
2604 pr_debug("for sector %llu, rmw=%d rcw=%d\n",
2605 (unsigned long long)sh->sector, rmw, rcw);
2606 set_bit(STRIPE_HANDLE, &sh->state);
2607 if (rmw < rcw && rmw > 0)
2608 /* prefer read-modify-write, but need to get some data */
2609 for (i = disks; i--; ) {
2610 struct r5dev *dev = &sh->dev[i];
2611 if ((dev->towrite || i == sh->pd_idx) &&
2612 !test_bit(R5_LOCKED, &dev->flags) &&
2613 !(test_bit(R5_UPTODATE, &dev->flags) ||
2614 test_bit(R5_Wantcompute, &dev->flags)) &&
2615 test_bit(R5_Insync, &dev->flags)) {
2617 test_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2618 pr_debug("Read_old block "
2619 "%d for r-m-w\n", i);
2620 set_bit(R5_LOCKED, &dev->flags);
2621 set_bit(R5_Wantread, &dev->flags);
2624 set_bit(STRIPE_DELAYED, &sh->state);
2625 set_bit(STRIPE_HANDLE, &sh->state);
2629 if (rcw <= rmw && rcw > 0) {
2630 /* want reconstruct write, but need to get some data */
2632 for (i = disks; i--; ) {
2633 struct r5dev *dev = &sh->dev[i];
2634 if (!test_bit(R5_OVERWRITE, &dev->flags) &&
2635 i != sh->pd_idx && i != sh->qd_idx &&
2636 !test_bit(R5_LOCKED, &dev->flags) &&
2637 !(test_bit(R5_UPTODATE, &dev->flags) ||
2638 test_bit(R5_Wantcompute, &dev->flags))) {
2640 if (!test_bit(R5_Insync, &dev->flags))
2641 continue; /* it's a failed drive */
2643 test_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2644 pr_debug("Read_old block "
2645 "%d for Reconstruct\n", i);
2646 set_bit(R5_LOCKED, &dev->flags);
2647 set_bit(R5_Wantread, &dev->flags);
2650 set_bit(STRIPE_DELAYED, &sh->state);
2651 set_bit(STRIPE_HANDLE, &sh->state);
2656 /* now if nothing is locked, and if we have enough data,
2657 * we can start a write request
2659 /* since handle_stripe can be called at any time we need to handle the
2660 * case where a compute block operation has been submitted and then a
2661 * subsequent call wants to start a write request. raid_run_ops only
2662 * handles the case where compute block and reconstruct are requested
2663 * simultaneously. If this is not the case then new writes need to be
2664 * held off until the compute completes.
2666 if ((s->req_compute || !test_bit(STRIPE_COMPUTE_RUN, &sh->state)) &&
2667 (s->locked == 0 && (rcw == 0 || rmw == 0) &&
2668 !test_bit(STRIPE_BIT_DELAY, &sh->state)))
2669 schedule_reconstruction(sh, s, rcw == 0, 0);
2672 static void handle_parity_checks5(struct r5conf *conf, struct stripe_head *sh,
2673 struct stripe_head_state *s, int disks)
2675 struct r5dev *dev = NULL;
2677 set_bit(STRIPE_HANDLE, &sh->state);
2679 switch (sh->check_state) {
2680 case check_state_idle:
2681 /* start a new check operation if there are no failures */
2682 if (s->failed == 0) {
2683 BUG_ON(s->uptodate != disks);
2684 sh->check_state = check_state_run;
2685 set_bit(STRIPE_OP_CHECK, &s->ops_request);
2686 clear_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags);
2690 dev = &sh->dev[s->failed_num[0]];
2692 case check_state_compute_result:
2693 sh->check_state = check_state_idle;
2695 dev = &sh->dev[sh->pd_idx];
2697 /* check that a write has not made the stripe insync */
2698 if (test_bit(STRIPE_INSYNC, &sh->state))
2701 /* either failed parity check, or recovery is happening */
2702 BUG_ON(!test_bit(R5_UPTODATE, &dev->flags));
2703 BUG_ON(s->uptodate != disks);
2705 set_bit(R5_LOCKED, &dev->flags);
2707 set_bit(R5_Wantwrite, &dev->flags);
2709 clear_bit(STRIPE_DEGRADED, &sh->state);
2710 set_bit(STRIPE_INSYNC, &sh->state);
2712 case check_state_run:
2713 break; /* we will be called again upon completion */
2714 case check_state_check_result:
2715 sh->check_state = check_state_idle;
2717 /* if a failure occurred during the check operation, leave
2718 * STRIPE_INSYNC not set and let the stripe be handled again
2723 /* handle a successful check operation, if parity is correct
2724 * we are done. Otherwise update the mismatch count and repair
2725 * parity if !MD_RECOVERY_CHECK
2727 if ((sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) == 0)
2728 /* parity is correct (on disc,
2729 * not in buffer any more)
2731 set_bit(STRIPE_INSYNC, &sh->state);
2733 conf->mddev->resync_mismatches += STRIPE_SECTORS;
2734 if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery))
2735 /* don't try to repair!! */
2736 set_bit(STRIPE_INSYNC, &sh->state);
2738 sh->check_state = check_state_compute_run;
2739 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
2740 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
2741 set_bit(R5_Wantcompute,
2742 &sh->dev[sh->pd_idx].flags);
2743 sh->ops.target = sh->pd_idx;
2744 sh->ops.target2 = -1;
2749 case check_state_compute_run:
2752 printk(KERN_ERR "%s: unknown check_state: %d sector: %llu\n",
2753 __func__, sh->check_state,
2754 (unsigned long long) sh->sector);
2760 static void handle_parity_checks6(struct r5conf *conf, struct stripe_head *sh,
2761 struct stripe_head_state *s,
2764 int pd_idx = sh->pd_idx;
2765 int qd_idx = sh->qd_idx;
2768 set_bit(STRIPE_HANDLE, &sh->state);
2770 BUG_ON(s->failed > 2);
2772 /* Want to check and possibly repair P and Q.
2773 * However there could be one 'failed' device, in which
2774 * case we can only check one of them, possibly using the
2775 * other to generate missing data
2778 switch (sh->check_state) {
2779 case check_state_idle:
2780 /* start a new check operation if there are < 2 failures */
2781 if (s->failed == s->q_failed) {
2782 /* The only possible failed device holds Q, so it
2783 * makes sense to check P (If anything else were failed,
2784 * we would have used P to recreate it).
2786 sh->check_state = check_state_run;
2788 if (!s->q_failed && s->failed < 2) {
2789 /* Q is not failed, and we didn't use it to generate
2790 * anything, so it makes sense to check it
2792 if (sh->check_state == check_state_run)
2793 sh->check_state = check_state_run_pq;
2795 sh->check_state = check_state_run_q;
2798 /* discard potentially stale zero_sum_result */
2799 sh->ops.zero_sum_result = 0;
2801 if (sh->check_state == check_state_run) {
2802 /* async_xor_zero_sum destroys the contents of P */
2803 clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
2806 if (sh->check_state >= check_state_run &&
2807 sh->check_state <= check_state_run_pq) {
2808 /* async_syndrome_zero_sum preserves P and Q, so
2809 * no need to mark them !uptodate here
2811 set_bit(STRIPE_OP_CHECK, &s->ops_request);
2815 /* we have 2-disk failure */
2816 BUG_ON(s->failed != 2);
2818 case check_state_compute_result:
2819 sh->check_state = check_state_idle;
2821 /* check that a write has not made the stripe insync */
2822 if (test_bit(STRIPE_INSYNC, &sh->state))
2825 /* now write out any block on a failed drive,
2826 * or P or Q if they were recomputed
2828 BUG_ON(s->uptodate < disks - 1); /* We don't need Q to recover */
2829 if (s->failed == 2) {
2830 dev = &sh->dev[s->failed_num[1]];
2832 set_bit(R5_LOCKED, &dev->flags);
2833 set_bit(R5_Wantwrite, &dev->flags);
2835 if (s->failed >= 1) {
2836 dev = &sh->dev[s->failed_num[0]];
2838 set_bit(R5_LOCKED, &dev->flags);
2839 set_bit(R5_Wantwrite, &dev->flags);
2841 if (sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) {
2842 dev = &sh->dev[pd_idx];
2844 set_bit(R5_LOCKED, &dev->flags);
2845 set_bit(R5_Wantwrite, &dev->flags);
2847 if (sh->ops.zero_sum_result & SUM_CHECK_Q_RESULT) {
2848 dev = &sh->dev[qd_idx];
2850 set_bit(R5_LOCKED, &dev->flags);
2851 set_bit(R5_Wantwrite, &dev->flags);
2853 clear_bit(STRIPE_DEGRADED, &sh->state);
2855 set_bit(STRIPE_INSYNC, &sh->state);
2857 case check_state_run:
2858 case check_state_run_q:
2859 case check_state_run_pq:
2860 break; /* we will be called again upon completion */
2861 case check_state_check_result:
2862 sh->check_state = check_state_idle;
2864 /* handle a successful check operation, if parity is correct
2865 * we are done. Otherwise update the mismatch count and repair
2866 * parity if !MD_RECOVERY_CHECK
2868 if (sh->ops.zero_sum_result == 0) {
2869 /* both parities are correct */
2871 set_bit(STRIPE_INSYNC, &sh->state);
2873 /* in contrast to the raid5 case we can validate
2874 * parity, but still have a failure to write
2877 sh->check_state = check_state_compute_result;
2878 /* Returning at this point means that we may go
2879 * off and bring p and/or q uptodate again so
2880 * we make sure to check zero_sum_result again
2881 * to verify if p or q need writeback
2885 conf->mddev->resync_mismatches += STRIPE_SECTORS;
2886 if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery))
2887 /* don't try to repair!! */
2888 set_bit(STRIPE_INSYNC, &sh->state);
2890 int *target = &sh->ops.target;
2892 sh->ops.target = -1;
2893 sh->ops.target2 = -1;
2894 sh->check_state = check_state_compute_run;
2895 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
2896 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
2897 if (sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) {
2898 set_bit(R5_Wantcompute,
2899 &sh->dev[pd_idx].flags);
2901 target = &sh->ops.target2;
2904 if (sh->ops.zero_sum_result & SUM_CHECK_Q_RESULT) {
2905 set_bit(R5_Wantcompute,
2906 &sh->dev[qd_idx].flags);
2913 case check_state_compute_run:
2916 printk(KERN_ERR "%s: unknown check_state: %d sector: %llu\n",
2917 __func__, sh->check_state,
2918 (unsigned long long) sh->sector);
2923 static void handle_stripe_expansion(struct r5conf *conf, struct stripe_head *sh)
2927 /* We have read all the blocks in this stripe and now we need to
2928 * copy some of them into a target stripe for expand.
2930 struct dma_async_tx_descriptor *tx = NULL;
2931 clear_bit(STRIPE_EXPAND_SOURCE, &sh->state);
2932 for (i = 0; i < sh->disks; i++)
2933 if (i != sh->pd_idx && i != sh->qd_idx) {
2935 struct stripe_head *sh2;
2936 struct async_submit_ctl submit;
2938 sector_t bn = compute_blocknr(sh, i, 1);
2939 sector_t s = raid5_compute_sector(conf, bn, 0,
2941 sh2 = get_active_stripe(conf, s, 0, 1, 1);
2943 /* so far only the early blocks of this stripe
2944 * have been requested. When later blocks
2945 * get requested, we will try again
2948 if (!test_bit(STRIPE_EXPANDING, &sh2->state) ||
2949 test_bit(R5_Expanded, &sh2->dev[dd_idx].flags)) {
2950 /* must have already done this block */
2951 release_stripe(sh2);
2955 /* place all the copies on one channel */
2956 init_async_submit(&submit, 0, tx, NULL, NULL, NULL);
2957 tx = async_memcpy(sh2->dev[dd_idx].page,
2958 sh->dev[i].page, 0, 0, STRIPE_SIZE,
2961 set_bit(R5_Expanded, &sh2->dev[dd_idx].flags);
2962 set_bit(R5_UPTODATE, &sh2->dev[dd_idx].flags);
2963 for (j = 0; j < conf->raid_disks; j++)
2964 if (j != sh2->pd_idx &&
2966 !test_bit(R5_Expanded, &sh2->dev[j].flags))
2968 if (j == conf->raid_disks) {
2969 set_bit(STRIPE_EXPAND_READY, &sh2->state);
2970 set_bit(STRIPE_HANDLE, &sh2->state);
2972 release_stripe(sh2);
2975 /* done submitting copies, wait for them to complete */
2978 dma_wait_for_async_tx(tx);
2984 * handle_stripe - do things to a stripe.
2986 * We lock the stripe and then examine the state of various bits
2987 * to see what needs to be done.
2989 * return some read request which now have data
2990 * return some write requests which are safely on disc
2991 * schedule a read on some buffers
2992 * schedule a write of some buffers
2993 * return confirmation of parity correctness
2995 * buffers are taken off read_list or write_list, and bh_cache buffers
2996 * get BH_Lock set before the stripe lock is released.
3000 static void analyse_stripe(struct stripe_head *sh, struct stripe_head_state *s)
3002 struct r5conf *conf = sh->raid_conf;
3003 int disks = sh->disks;
3007 memset(s, 0, sizeof(*s));
3009 s->syncing = test_bit(STRIPE_SYNCING, &sh->state);
3010 s->expanding = test_bit(STRIPE_EXPAND_SOURCE, &sh->state);
3011 s->expanded = test_bit(STRIPE_EXPAND_READY, &sh->state);
3012 s->failed_num[0] = -1;
3013 s->failed_num[1] = -1;
3015 /* Now to look around and see what can be done */
3017 spin_lock_irq(&conf->device_lock);
3018 for (i=disks; i--; ) {
3019 struct md_rdev *rdev;
3026 pr_debug("check %d: state 0x%lx read %p write %p written %p\n",
3027 i, dev->flags, dev->toread, dev->towrite, dev->written);
3028 /* maybe we can reply to a read
3030 * new wantfill requests are only permitted while
3031 * ops_complete_biofill is guaranteed to be inactive
3033 if (test_bit(R5_UPTODATE, &dev->flags) && dev->toread &&
3034 !test_bit(STRIPE_BIOFILL_RUN, &sh->state))
3035 set_bit(R5_Wantfill, &dev->flags);
3037 /* now count some things */
3038 if (test_bit(R5_LOCKED, &dev->flags))
3040 if (test_bit(R5_UPTODATE, &dev->flags))
3042 if (test_bit(R5_Wantcompute, &dev->flags)) {
3044 BUG_ON(s->compute > 2);
3047 if (test_bit(R5_Wantfill, &dev->flags))
3049 else if (dev->toread)
3053 if (!test_bit(R5_OVERWRITE, &dev->flags))
3058 /* Prefer to use the replacement for reads, but only
3059 * if it is recovered enough and has no bad blocks.
3061 rdev = rcu_dereference(conf->disks[i].replacement);
3062 if (rdev && !test_bit(Faulty, &rdev->flags) &&
3063 rdev->recovery_offset >= sh->sector + STRIPE_SECTORS &&
3064 !is_badblock(rdev, sh->sector, STRIPE_SECTORS,
3065 &first_bad, &bad_sectors))
3066 set_bit(R5_ReadRepl, &dev->flags);
3068 rdev = rcu_dereference(conf->disks[i].rdev);
3069 clear_bit(R5_ReadRepl, &dev->flags);
3071 if (rdev && test_bit(Faulty, &rdev->flags))
3074 is_bad = is_badblock(rdev, sh->sector, STRIPE_SECTORS,
3075 &first_bad, &bad_sectors);
3076 if (s->blocked_rdev == NULL
3077 && (test_bit(Blocked, &rdev->flags)
3080 set_bit(BlockedBadBlocks,
3082 s->blocked_rdev = rdev;
3083 atomic_inc(&rdev->nr_pending);
3086 clear_bit(R5_Insync, &dev->flags);
3090 /* also not in-sync */
3091 if (!test_bit(WriteErrorSeen, &rdev->flags)) {
3092 /* treat as in-sync, but with a read error
3093 * which we can now try to correct
3095 set_bit(R5_Insync, &dev->flags);
3096 set_bit(R5_ReadError, &dev->flags);
3098 } else if (test_bit(In_sync, &rdev->flags))
3099 set_bit(R5_Insync, &dev->flags);
3100 else if (sh->sector + STRIPE_SECTORS <= rdev->recovery_offset)
3101 /* in sync if before recovery_offset */
3102 set_bit(R5_Insync, &dev->flags);
3103 else if (test_bit(R5_UPTODATE, &dev->flags) &&
3104 test_bit(R5_Expanded, &dev->flags))
3105 /* If we've reshaped into here, we assume it is Insync.
3106 * We will shortly update recovery_offset to make
3109 set_bit(R5_Insync, &dev->flags);
3111 if (rdev && test_bit(R5_WriteError, &dev->flags)) {
3112 /* This flag does not apply to '.replacement'
3113 * only to .rdev, so make sure to check that*/
3114 struct md_rdev *rdev2 = rcu_dereference(
3115 conf->disks[i].rdev);
3117 clear_bit(R5_Insync, &dev->flags);
3118 if (rdev2 && !test_bit(Faulty, &rdev2->flags)) {
3119 s->handle_bad_blocks = 1;
3120 atomic_inc(&rdev2->nr_pending);
3122 clear_bit(R5_WriteError, &dev->flags);
3124 if (rdev && test_bit(R5_MadeGood, &dev->flags)) {
3125 /* This flag does not apply to '.replacement'
3126 * only to .rdev, so make sure to check that*/
3127 struct md_rdev *rdev2 = rcu_dereference(
3128 conf->disks[i].rdev);
3129 if (rdev2 && !test_bit(Faulty, &rdev2->flags)) {
3130 s->handle_bad_blocks = 1;
3131 atomic_inc(&rdev2->nr_pending);
3133 clear_bit(R5_MadeGood, &dev->flags);
3135 if (!test_bit(R5_Insync, &dev->flags)) {
3136 /* The ReadError flag will just be confusing now */
3137 clear_bit(R5_ReadError, &dev->flags);
3138 clear_bit(R5_ReWrite, &dev->flags);
3140 if (test_bit(R5_ReadError, &dev->flags))
3141 clear_bit(R5_Insync, &dev->flags);
3142 if (!test_bit(R5_Insync, &dev->flags)) {
3144 s->failed_num[s->failed] = i;
3148 spin_unlock_irq(&conf->device_lock);
3152 static void handle_stripe(struct stripe_head *sh)
3154 struct stripe_head_state s;
3155 struct r5conf *conf = sh->raid_conf;
3158 int disks = sh->disks;
3159 struct r5dev *pdev, *qdev;
3161 clear_bit(STRIPE_HANDLE, &sh->state);
3162 if (test_and_set_bit_lock(STRIPE_ACTIVE, &sh->state)) {
3163 /* already being handled, ensure it gets handled
3164 * again when current action finishes */
3165 set_bit(STRIPE_HANDLE, &sh->state);
3169 if (test_and_clear_bit(STRIPE_SYNC_REQUESTED, &sh->state)) {
3170 set_bit(STRIPE_SYNCING, &sh->state);
3171 clear_bit(STRIPE_INSYNC, &sh->state);
3173 clear_bit(STRIPE_DELAYED, &sh->state);
3175 pr_debug("handling stripe %llu, state=%#lx cnt=%d, "
3176 "pd_idx=%d, qd_idx=%d\n, check:%d, reconstruct:%d\n",
3177 (unsigned long long)sh->sector, sh->state,
3178 atomic_read(&sh->count), sh->pd_idx, sh->qd_idx,
3179 sh->check_state, sh->reconstruct_state);
3181 analyse_stripe(sh, &s);
3183 if (s.handle_bad_blocks) {
3184 set_bit(STRIPE_HANDLE, &sh->state);
3188 if (unlikely(s.blocked_rdev)) {
3189 if (s.syncing || s.expanding || s.expanded ||
3190 s.to_write || s.written) {
3191 set_bit(STRIPE_HANDLE, &sh->state);
3194 /* There is nothing for the blocked_rdev to block */
3195 rdev_dec_pending(s.blocked_rdev, conf->mddev);
3196 s.blocked_rdev = NULL;
3199 if (s.to_fill && !test_bit(STRIPE_BIOFILL_RUN, &sh->state)) {
3200 set_bit(STRIPE_OP_BIOFILL, &s.ops_request);
3201 set_bit(STRIPE_BIOFILL_RUN, &sh->state);
3204 pr_debug("locked=%d uptodate=%d to_read=%d"
3205 " to_write=%d failed=%d failed_num=%d,%d\n",
3206 s.locked, s.uptodate, s.to_read, s.to_write, s.failed,
3207 s.failed_num[0], s.failed_num[1]);
3208 /* check if the array has lost more than max_degraded devices and,
3209 * if so, some requests might need to be failed.
3211 if (s.failed > conf->max_degraded) {
3212 sh->check_state = 0;
3213 sh->reconstruct_state = 0;
3214 if (s.to_read+s.to_write+s.written)
3215 handle_failed_stripe(conf, sh, &s, disks, &s.return_bi);
3217 handle_failed_sync(conf, sh, &s);
3221 * might be able to return some write requests if the parity blocks
3222 * are safe, or on a failed drive
3224 pdev = &sh->dev[sh->pd_idx];
3225 s.p_failed = (s.failed >= 1 && s.failed_num[0] == sh->pd_idx)
3226 || (s.failed >= 2 && s.failed_num[1] == sh->pd_idx);
3227 qdev = &sh->dev[sh->qd_idx];
3228 s.q_failed = (s.failed >= 1 && s.failed_num[0] == sh->qd_idx)
3229 || (s.failed >= 2 && s.failed_num[1] == sh->qd_idx)
3233 (s.p_failed || ((test_bit(R5_Insync, &pdev->flags)
3234 && !test_bit(R5_LOCKED, &pdev->flags)
3235 && test_bit(R5_UPTODATE, &pdev->flags)))) &&
3236 (s.q_failed || ((test_bit(R5_Insync, &qdev->flags)
3237 && !test_bit(R5_LOCKED, &qdev->flags)
3238 && test_bit(R5_UPTODATE, &qdev->flags)))))
3239 handle_stripe_clean_event(conf, sh, disks, &s.return_bi);
3241 /* Now we might consider reading some blocks, either to check/generate
3242 * parity, or to satisfy requests
3243 * or to load a block that is being partially written.
3245 if (s.to_read || s.non_overwrite
3246 || (conf->level == 6 && s.to_write && s.failed)
3247 || (s.syncing && (s.uptodate + s.compute < disks)) || s.expanding)
3248 handle_stripe_fill(sh, &s, disks);
3250 /* Now we check to see if any write operations have recently
3254 if (sh->reconstruct_state == reconstruct_state_prexor_drain_result)
3256 if (sh->reconstruct_state == reconstruct_state_drain_result ||
3257 sh->reconstruct_state == reconstruct_state_prexor_drain_result) {
3258 sh->reconstruct_state = reconstruct_state_idle;
3260 /* All the 'written' buffers and the parity block are ready to
3261 * be written back to disk
3263 BUG_ON(!test_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags));
3264 BUG_ON(sh->qd_idx >= 0 &&
3265 !test_bit(R5_UPTODATE, &sh->dev[sh->qd_idx].flags));
3266 for (i = disks; i--; ) {
3267 struct r5dev *dev = &sh->dev[i];
3268 if (test_bit(R5_LOCKED, &dev->flags) &&
3269 (i == sh->pd_idx || i == sh->qd_idx ||
3271 pr_debug("Writing block %d\n", i);
3272 set_bit(R5_Wantwrite, &dev->flags);
3275 if (!test_bit(R5_Insync, &dev->flags) ||
3276 ((i == sh->pd_idx || i == sh->qd_idx) &&
3278 set_bit(STRIPE_INSYNC, &sh->state);
3281 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
3282 s.dec_preread_active = 1;
3285 /* Now to consider new write requests and what else, if anything
3286 * should be read. We do not handle new writes when:
3287 * 1/ A 'write' operation (copy+xor) is already in flight.
3288 * 2/ A 'check' operation is in flight, as it may clobber the parity
3291 if (s.to_write && !sh->reconstruct_state && !sh->check_state)
3292 handle_stripe_dirtying(conf, sh, &s, disks);
3294 /* maybe we need to check and possibly fix the parity for this stripe
3295 * Any reads will already have been scheduled, so we just see if enough
3296 * data is available. The parity check is held off while parity
3297 * dependent operations are in flight.
3299 if (sh->check_state ||
3300 (s.syncing && s.locked == 0 &&
3301 !test_bit(STRIPE_COMPUTE_RUN, &sh->state) &&
3302 !test_bit(STRIPE_INSYNC, &sh->state))) {
3303 if (conf->level == 6)
3304 handle_parity_checks6(conf, sh, &s, disks);
3306 handle_parity_checks5(conf, sh, &s, disks);
3309 if (s.syncing && s.locked == 0 && test_bit(STRIPE_INSYNC, &sh->state)) {
3310 md_done_sync(conf->mddev, STRIPE_SECTORS, 1);
3311 clear_bit(STRIPE_SYNCING, &sh->state);
3314 /* If the failed drives are just a ReadError, then we might need
3315 * to progress the repair/check process
3317 if (s.failed <= conf->max_degraded && !conf->mddev->ro)
3318 for (i = 0; i < s.failed; i++) {
3319 struct r5dev *dev = &sh->dev[s.failed_num[i]];
3320 if (test_bit(R5_ReadError, &dev->flags)
3321 && !test_bit(R5_LOCKED, &dev->flags)
3322 && test_bit(R5_UPTODATE, &dev->flags)
3324 if (!test_bit(R5_ReWrite, &dev->flags)) {
3325 set_bit(R5_Wantwrite, &dev->flags);
3326 set_bit(R5_ReWrite, &dev->flags);
3327 set_bit(R5_LOCKED, &dev->flags);
3330 /* let's read it back */
3331 set_bit(R5_Wantread, &dev->flags);
3332 set_bit(R5_LOCKED, &dev->flags);
3339 /* Finish reconstruct operations initiated by the expansion process */
3340 if (sh->reconstruct_state == reconstruct_state_result) {
3341 struct stripe_head *sh_src
3342 = get_active_stripe(conf, sh->sector, 1, 1, 1);
3343 if (sh_src && test_bit(STRIPE_EXPAND_SOURCE, &sh_src->state)) {
3344 /* sh cannot be written until sh_src has been read.
3345 * so arrange for sh to be delayed a little
3347 set_bit(STRIPE_DELAYED, &sh->state);
3348 set_bit(STRIPE_HANDLE, &sh->state);
3349 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE,
3351 atomic_inc(&conf->preread_active_stripes);
3352 release_stripe(sh_src);
3356 release_stripe(sh_src);
3358 sh->reconstruct_state = reconstruct_state_idle;
3359 clear_bit(STRIPE_EXPANDING, &sh->state);
3360 for (i = conf->raid_disks; i--; ) {
3361 set_bit(R5_Wantwrite, &sh->dev[i].flags);
3362 set_bit(R5_LOCKED, &sh->dev[i].flags);
3367 if (s.expanded && test_bit(STRIPE_EXPANDING, &sh->state) &&
3368 !sh->reconstruct_state) {
3369 /* Need to write out all blocks after computing parity */
3370 sh->disks = conf->raid_disks;
3371 stripe_set_idx(sh->sector, conf, 0, sh);
3372 schedule_reconstruction(sh, &s, 1, 1);
3373 } else if (s.expanded && !sh->reconstruct_state && s.locked == 0) {
3374 clear_bit(STRIPE_EXPAND_READY, &sh->state);
3375 atomic_dec(&conf->reshape_stripes);
3376 wake_up(&conf->wait_for_overlap);
3377 md_done_sync(conf->mddev, STRIPE_SECTORS, 1);
3380 if (s.expanding && s.locked == 0 &&
3381 !test_bit(STRIPE_COMPUTE_RUN, &sh->state))
3382 handle_stripe_expansion(conf, sh);
3385 /* wait for this device to become unblocked */
3386 if (conf->mddev->external && unlikely(s.blocked_rdev))
3387 md_wait_for_blocked_rdev(s.blocked_rdev, conf->mddev);
3389 if (s.handle_bad_blocks)
3390 for (i = disks; i--; ) {
3391 struct md_rdev *rdev;
3392 struct r5dev *dev = &sh->dev[i];
3393 if (test_and_clear_bit(R5_WriteError, &dev->flags)) {
3394 /* We own a safe reference to the rdev */
3395 rdev = conf->disks[i].rdev;
3396 if (!rdev_set_badblocks(rdev, sh->sector,
3398 md_error(conf->mddev, rdev);
3399 rdev_dec_pending(rdev, conf->mddev);
3401 if (test_and_clear_bit(R5_MadeGood, &dev->flags)) {
3402 rdev = conf->disks[i].rdev;
3403 rdev_clear_badblocks(rdev, sh->sector,
3405 rdev_dec_pending(rdev, conf->mddev);
3410 raid_run_ops(sh, s.ops_request);
3414 if (s.dec_preread_active) {
3415 /* We delay this until after ops_run_io so that if make_request
3416 * is waiting on a flush, it won't continue until the writes
3417 * have actually been submitted.
3419 atomic_dec(&conf->preread_active_stripes);
3420 if (atomic_read(&conf->preread_active_stripes) <
3422 md_wakeup_thread(conf->mddev->thread);
3425 return_io(s.return_bi);
3427 clear_bit_unlock(STRIPE_ACTIVE, &sh->state);
3430 static void raid5_activate_delayed(struct r5conf *conf)
3432 if (atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD) {
3433 while (!list_empty(&conf->delayed_list)) {
3434 struct list_head *l = conf->delayed_list.next;
3435 struct stripe_head *sh;
3436 sh = list_entry(l, struct stripe_head, lru);
3438 clear_bit(STRIPE_DELAYED, &sh->state);
3439 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
3440 atomic_inc(&conf->preread_active_stripes);
3441 list_add_tail(&sh->lru, &conf->hold_list);
3446 static void activate_bit_delay(struct r5conf *conf)
3448 /* device_lock is held */
3449 struct list_head head;
3450 list_add(&head, &conf->bitmap_list);
3451 list_del_init(&conf->bitmap_list);
3452 while (!list_empty(&head)) {
3453 struct stripe_head *sh = list_entry(head.next, struct stripe_head, lru);
3454 list_del_init(&sh->lru);
3455 atomic_inc(&sh->count);
3456 __release_stripe(conf, sh);
3460 int md_raid5_congested(struct mddev *mddev, int bits)
3462 struct r5conf *conf = mddev->private;
3464 /* No difference between reads and writes. Just check
3465 * how busy the stripe_cache is
3468 if (conf->inactive_blocked)
3472 if (list_empty_careful(&conf->inactive_list))
3477 EXPORT_SYMBOL_GPL(md_raid5_congested);
3479 static int raid5_congested(void *data, int bits)
3481 struct mddev *mddev = data;
3483 return mddev_congested(mddev, bits) ||
3484 md_raid5_congested(mddev, bits);
3487 /* We want read requests to align with chunks where possible,
3488 * but write requests don't need to.
3490 static int raid5_mergeable_bvec(struct request_queue *q,
3491 struct bvec_merge_data *bvm,
3492 struct bio_vec *biovec)
3494 struct mddev *mddev = q->queuedata;
3495 sector_t sector = bvm->bi_sector + get_start_sect(bvm->bi_bdev);
3497 unsigned int chunk_sectors = mddev->chunk_sectors;
3498 unsigned int bio_sectors = bvm->bi_size >> 9;
3500 if ((bvm->bi_rw & 1) == WRITE)
3501 return biovec->bv_len; /* always allow writes to be mergeable */
3503 if (mddev->new_chunk_sectors < mddev->chunk_sectors)
3504 chunk_sectors = mddev->new_chunk_sectors;
3505 max = (chunk_sectors - ((sector & (chunk_sectors - 1)) + bio_sectors)) << 9;
3506 if (max < 0) max = 0;
3507 if (max <= biovec->bv_len && bio_sectors == 0)
3508 return biovec->bv_len;
3514 static int in_chunk_boundary(struct mddev *mddev, struct bio *bio)
3516 sector_t sector = bio->bi_sector + get_start_sect(bio->bi_bdev);
3517 unsigned int chunk_sectors = mddev->chunk_sectors;
3518 unsigned int bio_sectors = bio->bi_size >> 9;
3520 if (mddev->new_chunk_sectors < mddev->chunk_sectors)
3521 chunk_sectors = mddev->new_chunk_sectors;
3522 return chunk_sectors >=
3523 ((sector & (chunk_sectors - 1)) + bio_sectors);
3527 * add bio to the retry LIFO ( in O(1) ... we are in interrupt )
3528 * later sampled by raid5d.
3530 static void add_bio_to_retry(struct bio *bi,struct r5conf *conf)
3532 unsigned long flags;
3534 spin_lock_irqsave(&conf->device_lock, flags);
3536 bi->bi_next = conf->retry_read_aligned_list;
3537 conf->retry_read_aligned_list = bi;
3539 spin_unlock_irqrestore(&conf->device_lock, flags);
3540 md_wakeup_thread(conf->mddev->thread);
3544 static struct bio *remove_bio_from_retry(struct r5conf *conf)
3548 bi = conf->retry_read_aligned;
3550 conf->retry_read_aligned = NULL;
3553 bi = conf->retry_read_aligned_list;
3555 conf->retry_read_aligned_list = bi->bi_next;
3558 * this sets the active strip count to 1 and the processed
3559 * strip count to zero (upper 8 bits)
3561 bi->bi_phys_segments = 1; /* biased count of active stripes */
3569 * The "raid5_align_endio" should check if the read succeeded and if it
3570 * did, call bio_endio on the original bio (having bio_put the new bio
3572 * If the read failed..
3574 static void raid5_align_endio(struct bio *bi, int error)
3576 struct bio* raid_bi = bi->bi_private;
3577 struct mddev *mddev;
3578 struct r5conf *conf;
3579 int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
3580 struct md_rdev *rdev;
3584 rdev = (void*)raid_bi->bi_next;
3585 raid_bi->bi_next = NULL;
3586 mddev = rdev->mddev;
3587 conf = mddev->private;
3589 rdev_dec_pending(rdev, conf->mddev);
3591 if (!error && uptodate) {
3592 bio_endio(raid_bi, 0);
3593 if (atomic_dec_and_test(&conf->active_aligned_reads))
3594 wake_up(&conf->wait_for_stripe);
3599 pr_debug("raid5_align_endio : io error...handing IO for a retry\n");
3601 add_bio_to_retry(raid_bi, conf);
3604 static int bio_fits_rdev(struct bio *bi)
3606 struct request_queue *q = bdev_get_queue(bi->bi_bdev);
3608 if ((bi->bi_size>>9) > queue_max_sectors(q))
3610 blk_recount_segments(q, bi);
3611 if (bi->bi_phys_segments > queue_max_segments(q))
3614 if (q->merge_bvec_fn)
3615 /* it's too hard to apply the merge_bvec_fn at this stage,
3624 static int chunk_aligned_read(struct mddev *mddev, struct bio * raid_bio)
3626 struct r5conf *conf = mddev->private;
3628 struct bio* align_bi;
3629 struct md_rdev *rdev;
3630 sector_t end_sector;
3632 if (!in_chunk_boundary(mddev, raid_bio)) {
3633 pr_debug("chunk_aligned_read : non aligned\n");
3637 * use bio_clone_mddev to make a copy of the bio
3639 align_bi = bio_clone_mddev(raid_bio, GFP_NOIO, mddev);
3643 * set bi_end_io to a new function, and set bi_private to the
3646 align_bi->bi_end_io = raid5_align_endio;
3647 align_bi->bi_private = raid_bio;
3651 align_bi->bi_sector = raid5_compute_sector(conf, raid_bio->bi_sector,
3655 end_sector = align_bi->bi_sector + (align_bi->bi_size >> 9);
3657 rdev = rcu_dereference(conf->disks[dd_idx].replacement);
3658 if (!rdev || test_bit(Faulty, &rdev->flags) ||
3659 rdev->recovery_offset < end_sector) {
3660 rdev = rcu_dereference(conf->disks[dd_idx].rdev);
3662 (test_bit(Faulty, &rdev->flags) ||
3663 !(test_bit(In_sync, &rdev->flags) ||
3664 rdev->recovery_offset >= end_sector)))
3671 atomic_inc(&rdev->nr_pending);
3673 raid_bio->bi_next = (void*)rdev;
3674 align_bi->bi_bdev = rdev->bdev;
3675 align_bi->bi_flags &= ~(1 << BIO_SEG_VALID);
3676 align_bi->bi_sector += rdev->data_offset;
3678 if (!bio_fits_rdev(align_bi) ||
3679 is_badblock(rdev, align_bi->bi_sector, align_bi->bi_size>>9,
3680 &first_bad, &bad_sectors)) {
3681 /* too big in some way, or has a known bad block */
3683 rdev_dec_pending(rdev, mddev);
3687 spin_lock_irq(&conf->device_lock);
3688 wait_event_lock_irq(conf->wait_for_stripe,
3690 conf->device_lock, /* nothing */);
3691 atomic_inc(&conf->active_aligned_reads);
3692 spin_unlock_irq(&conf->device_lock);
3694 generic_make_request(align_bi);
3703 /* __get_priority_stripe - get the next stripe to process
3705 * Full stripe writes are allowed to pass preread active stripes up until
3706 * the bypass_threshold is exceeded. In general the bypass_count
3707 * increments when the handle_list is handled before the hold_list; however, it
3708 * will not be incremented when STRIPE_IO_STARTED is sampled set signifying a
3709 * stripe with in flight i/o. The bypass_count will be reset when the
3710 * head of the hold_list has changed, i.e. the head was promoted to the
3713 static struct stripe_head *__get_priority_stripe(struct r5conf *conf)
3715 struct stripe_head *sh;
3717 pr_debug("%s: handle: %s hold: %s full_writes: %d bypass_count: %d\n",
3719 list_empty(&conf->handle_list) ? "empty" : "busy",
3720 list_empty(&conf->hold_list) ? "empty" : "busy",
3721 atomic_read(&conf->pending_full_writes), conf->bypass_count);
3723 if (!list_empty(&conf->handle_list)) {
3724 sh = list_entry(conf->handle_list.next, typeof(*sh), lru);
3726 if (list_empty(&conf->hold_list))
3727 conf->bypass_count = 0;
3728 else if (!test_bit(STRIPE_IO_STARTED, &sh->state)) {
3729 if (conf->hold_list.next == conf->last_hold)
3730 conf->bypass_count++;
3732 conf->last_hold = conf->hold_list.next;
3733 conf->bypass_count -= conf->bypass_threshold;
3734 if (conf->bypass_count < 0)
3735 conf->bypass_count = 0;
3738 } else if (!list_empty(&conf->hold_list) &&
3739 ((conf->bypass_threshold &&
3740 conf->bypass_count > conf->bypass_threshold) ||
3741 atomic_read(&conf->pending_full_writes) == 0)) {
3742 sh = list_entry(conf->hold_list.next,
3744 conf->bypass_count -= conf->bypass_threshold;
3745 if (conf->bypass_count < 0)
3746 conf->bypass_count = 0;
3750 list_del_init(&sh->lru);
3751 atomic_inc(&sh->count);
3752 BUG_ON(atomic_read(&sh->count) != 1);
3756 static void make_request(struct mddev *mddev, struct bio * bi)
3758 struct r5conf *conf = mddev->private;
3760 sector_t new_sector;
3761 sector_t logical_sector, last_sector;
3762 struct stripe_head *sh;
3763 const int rw = bio_data_dir(bi);
3767 if (unlikely(bi->bi_rw & REQ_FLUSH)) {
3768 md_flush_request(mddev, bi);
3772 md_write_start(mddev, bi);
3775 mddev->reshape_position == MaxSector &&
3776 chunk_aligned_read(mddev,bi))
3779 logical_sector = bi->bi_sector & ~((sector_t)STRIPE_SECTORS-1);
3780 last_sector = bi->bi_sector + (bi->bi_size>>9);
3782 bi->bi_phys_segments = 1; /* over-loaded to count active stripes */
3784 plugged = mddev_check_plugged(mddev);
3785 for (;logical_sector < last_sector; logical_sector += STRIPE_SECTORS) {
3787 int disks, data_disks;
3792 disks = conf->raid_disks;
3793 prepare_to_wait(&conf->wait_for_overlap, &w, TASK_UNINTERRUPTIBLE);
3794 if (unlikely(conf->reshape_progress != MaxSector)) {
3795 /* spinlock is needed as reshape_progress may be
3796 * 64bit on a 32bit platform, and so it might be
3797 * possible to see a half-updated value
3798 * Of course reshape_progress could change after
3799 * the lock is dropped, so once we get a reference
3800 * to the stripe that we think it is, we will have
3803 spin_lock_irq(&conf->device_lock);
3804 if (mddev->delta_disks < 0
3805 ? logical_sector < conf->reshape_progress
3806 : logical_sector >= conf->reshape_progress) {
3807 disks = conf->previous_raid_disks;
3810 if (mddev->delta_disks < 0
3811 ? logical_sector < conf->reshape_safe
3812 : logical_sector >= conf->reshape_safe) {
3813 spin_unlock_irq(&conf->device_lock);
3818 spin_unlock_irq(&conf->device_lock);
3820 data_disks = disks - conf->max_degraded;
3822 new_sector = raid5_compute_sector(conf, logical_sector,
3825 pr_debug("raid456: make_request, sector %llu logical %llu\n",
3826 (unsigned long long)new_sector,
3827 (unsigned long long)logical_sector);
3829 sh = get_active_stripe(conf, new_sector, previous,
3830 (bi->bi_rw&RWA_MASK), 0);
3832 if (unlikely(previous)) {
3833 /* expansion might have moved on while waiting for a
3834 * stripe, so we must do the range check again.
3835 * Expansion could still move past after this
3836 * test, but as we are holding a reference to
3837 * 'sh', we know that if that happens,
3838 * STRIPE_EXPANDING will get set and the expansion
3839 * won't proceed until we finish with the stripe.
3842 spin_lock_irq(&conf->device_lock);
3843 if (mddev->delta_disks < 0
3844 ? logical_sector >= conf->reshape_progress
3845 : logical_sector < conf->reshape_progress)
3846 /* mismatch, need to try again */
3848 spin_unlock_irq(&conf->device_lock);
3857 logical_sector >= mddev->suspend_lo &&
3858 logical_sector < mddev->suspend_hi) {
3860 /* As the suspend_* range is controlled by
3861 * userspace, we want an interruptible
3864 flush_signals(current);
3865 prepare_to_wait(&conf->wait_for_overlap,
3866 &w, TASK_INTERRUPTIBLE);
3867 if (logical_sector >= mddev->suspend_lo &&
3868 logical_sector < mddev->suspend_hi)
3873 if (test_bit(STRIPE_EXPANDING, &sh->state) ||
3874 !add_stripe_bio(sh, bi, dd_idx, rw)) {
3875 /* Stripe is busy expanding or
3876 * add failed due to overlap. Flush everything
3879 md_wakeup_thread(mddev->thread);
3884 finish_wait(&conf->wait_for_overlap, &w);
3885 set_bit(STRIPE_HANDLE, &sh->state);
3886 clear_bit(STRIPE_DELAYED, &sh->state);
3887 if ((bi->bi_rw & REQ_SYNC) &&
3888 !test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
3889 atomic_inc(&conf->preread_active_stripes);
3892 /* cannot get stripe for read-ahead, just give-up */
3893 clear_bit(BIO_UPTODATE, &bi->bi_flags);
3894 finish_wait(&conf->wait_for_overlap, &w);
3900 md_wakeup_thread(mddev->thread);
3902 spin_lock_irq(&conf->device_lock);
3903 remaining = raid5_dec_bi_phys_segments(bi);
3904 spin_unlock_irq(&conf->device_lock);
3905 if (remaining == 0) {
3908 md_write_end(mddev);
3914 static sector_t raid5_size(struct mddev *mddev, sector_t sectors, int raid_disks);
3916 static sector_t reshape_request(struct mddev *mddev, sector_t sector_nr, int *skipped)
3918 /* reshaping is quite different to recovery/resync so it is
3919 * handled quite separately ... here.
3921 * On each call to sync_request, we gather one chunk worth of
3922 * destination stripes and flag them as expanding.
3923 * Then we find all the source stripes and request reads.
3924 * As the reads complete, handle_stripe will copy the data
3925 * into the destination stripe and release that stripe.
3927 struct r5conf *conf = mddev->private;
3928 struct stripe_head *sh;
3929 sector_t first_sector, last_sector;
3930 int raid_disks = conf->previous_raid_disks;
3931 int data_disks = raid_disks - conf->max_degraded;
3932 int new_data_disks = conf->raid_disks - conf->max_degraded;
3935 sector_t writepos, readpos, safepos;
3936 sector_t stripe_addr;
3937 int reshape_sectors;
3938 struct list_head stripes;
3940 if (sector_nr == 0) {
3941 /* If restarting in the middle, skip the initial sectors */
3942 if (mddev->delta_disks < 0 &&
3943 conf->reshape_progress < raid5_size(mddev, 0, 0)) {
3944 sector_nr = raid5_size(mddev, 0, 0)
3945 - conf->reshape_progress;
3946 } else if (mddev->delta_disks >= 0 &&
3947 conf->reshape_progress > 0)
3948 sector_nr = conf->reshape_progress;
3949 sector_div(sector_nr, new_data_disks);
3951 mddev->curr_resync_completed = sector_nr;
3952 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
3958 /* We need to process a full chunk at a time.
3959 * If old and new chunk sizes differ, we need to process the
3962 if (mddev->new_chunk_sectors > mddev->chunk_sectors)
3963 reshape_sectors = mddev->new_chunk_sectors;
3965 reshape_sectors = mddev->chunk_sectors;
3967 /* we update the metadata when there is more than 3Meg
3968 * in the block range (that is rather arbitrary, should
3969 * probably be time based) or when the data about to be
3970 * copied would over-write the source of the data at
3971 * the front of the range.
3972 * i.e. one new_stripe along from reshape_progress new_maps
3973 * to after where reshape_safe old_maps to
3975 writepos = conf->reshape_progress;
3976 sector_div(writepos, new_data_disks);
3977 readpos = conf->reshape_progress;
3978 sector_div(readpos, data_disks);
3979 safepos = conf->reshape_safe;
3980 sector_div(safepos, data_disks);
3981 if (mddev->delta_disks < 0) {
3982 writepos -= min_t(sector_t, reshape_sectors, writepos);
3983 readpos += reshape_sectors;
3984 safepos += reshape_sectors;
3986 writepos += reshape_sectors;
3987 readpos -= min_t(sector_t, reshape_sectors, readpos);
3988 safepos -= min_t(sector_t, reshape_sectors, safepos);
3991 /* 'writepos' is the most advanced device address we might write.
3992 * 'readpos' is the least advanced device address we might read.
3993 * 'safepos' is the least address recorded in the metadata as having
3995 * If 'readpos' is behind 'writepos', then there is no way that we can
3996 * ensure safety in the face of a crash - that must be done by userspace
3997 * making a backup of the data. So in that case there is no particular
3998 * rush to update metadata.
3999 * Otherwise if 'safepos' is behind 'writepos', then we really need to
4000 * update the metadata to advance 'safepos' to match 'readpos' so that
4001 * we can be safe in the event of a crash.
4002 * So we insist on updating metadata if safepos is behind writepos and
4003 * readpos is beyond writepos.
4004 * In any case, update the metadata every 10 seconds.
4005 * Maybe that number should be configurable, but I'm not sure it is
4006 * worth it.... maybe it could be a multiple of safemode_delay???
4008 if ((mddev->delta_disks < 0
4009 ? (safepos > writepos && readpos < writepos)
4010 : (safepos < writepos && readpos > writepos)) ||
4011 time_after(jiffies, conf->reshape_checkpoint + 10*HZ)) {
4012 /* Cannot proceed until we've updated the superblock... */
4013 wait_event(conf->wait_for_overlap,
4014 atomic_read(&conf->reshape_stripes)==0);
4015 mddev->reshape_position = conf->reshape_progress;
4016 mddev->curr_resync_completed = sector_nr;
4017 conf->reshape_checkpoint = jiffies;
4018 set_bit(MD_CHANGE_DEVS, &mddev->flags);
4019 md_wakeup_thread(mddev->thread);
4020 wait_event(mddev->sb_wait, mddev->flags == 0 ||
4021 kthread_should_stop());
4022 spin_lock_irq(&conf->device_lock);
4023 conf->reshape_safe = mddev->reshape_position;
4024 spin_unlock_irq(&conf->device_lock);
4025 wake_up(&conf->wait_for_overlap);
4026 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
4029 if (mddev->delta_disks < 0) {
4030 BUG_ON(conf->reshape_progress == 0);
4031 stripe_addr = writepos;
4032 BUG_ON((mddev->dev_sectors &
4033 ~((sector_t)reshape_sectors - 1))
4034 - reshape_sectors - stripe_addr
4037 BUG_ON(writepos != sector_nr + reshape_sectors);
4038 stripe_addr = sector_nr;
4040 INIT_LIST_HEAD(&stripes);
4041 for (i = 0; i < reshape_sectors; i += STRIPE_SECTORS) {
4043 int skipped_disk = 0;
4044 sh = get_active_stripe(conf, stripe_addr+i, 0, 0, 1);
4045 set_bit(STRIPE_EXPANDING, &sh->state);
4046 atomic_inc(&conf->reshape_stripes);
4047 /* If any of this stripe is beyond the end of the old
4048 * array, then we need to zero those blocks
4050 for (j=sh->disks; j--;) {
4052 if (j == sh->pd_idx)
4054 if (conf->level == 6 &&
4057 s = compute_blocknr(sh, j, 0);
4058 if (s < raid5_size(mddev, 0, 0)) {
4062 memset(page_address(sh->dev[j].page), 0, STRIPE_SIZE);
4063 set_bit(R5_Expanded, &sh->dev[j].flags);
4064 set_bit(R5_UPTODATE, &sh->dev[j].flags);
4066 if (!skipped_disk) {
4067 set_bit(STRIPE_EXPAND_READY, &sh->state);
4068 set_bit(STRIPE_HANDLE, &sh->state);
4070 list_add(&sh->lru, &stripes);
4072 spin_lock_irq(&conf->device_lock);
4073 if (mddev->delta_disks < 0)
4074 conf->reshape_progress -= reshape_sectors * new_data_disks;
4076 conf->reshape_progress += reshape_sectors * new_data_disks;
4077 spin_unlock_irq(&conf->device_lock);
4078 /* Ok, those stripe are ready. We can start scheduling
4079 * reads on the source stripes.
4080 * The source stripes are determined by mapping the first and last
4081 * block on the destination stripes.
4084 raid5_compute_sector(conf, stripe_addr*(new_data_disks),
4087 raid5_compute_sector(conf, ((stripe_addr+reshape_sectors)
4088 * new_data_disks - 1),
4090 if (last_sector >= mddev->dev_sectors)
4091 last_sector = mddev->dev_sectors - 1;
4092 while (first_sector <= last_sector) {
4093 sh = get_active_stripe(conf, first_sector, 1, 0, 1);
4094 set_bit(STRIPE_EXPAND_SOURCE, &sh->state);
4095 set_bit(STRIPE_HANDLE, &sh->state);
4097 first_sector += STRIPE_SECTORS;
4099 /* Now that the sources are clearly marked, we can release
4100 * the destination stripes
4102 while (!list_empty(&stripes)) {
4103 sh = list_entry(stripes.next, struct stripe_head, lru);
4104 list_del_init(&sh->lru);
4107 /* If this takes us to the resync_max point where we have to pause,
4108 * then we need to write out the superblock.
4110 sector_nr += reshape_sectors;
4111 if ((sector_nr - mddev->curr_resync_completed) * 2
4112 >= mddev->resync_max - mddev->curr_resync_completed) {
4113 /* Cannot proceed until we've updated the superblock... */
4114 wait_event(conf->wait_for_overlap,
4115 atomic_read(&conf->reshape_stripes) == 0);
4116 mddev->reshape_position = conf->reshape_progress;
4117 mddev->curr_resync_completed = sector_nr;
4118 conf->reshape_checkpoint = jiffies;
4119 set_bit(MD_CHANGE_DEVS, &mddev->flags);
4120 md_wakeup_thread(mddev->thread);
4121 wait_event(mddev->sb_wait,
4122 !test_bit(MD_CHANGE_DEVS, &mddev->flags)
4123 || kthread_should_stop());
4124 spin_lock_irq(&conf->device_lock);
4125 conf->reshape_safe = mddev->reshape_position;
4126 spin_unlock_irq(&conf->device_lock);
4127 wake_up(&conf->wait_for_overlap);
4128 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
4130 return reshape_sectors;
4133 /* FIXME go_faster isn't used */
4134 static inline sector_t sync_request(struct mddev *mddev, sector_t sector_nr, int *skipped, int go_faster)
4136 struct r5conf *conf = mddev->private;
4137 struct stripe_head *sh;
4138 sector_t max_sector = mddev->dev_sectors;
4139 sector_t sync_blocks;
4140 int still_degraded = 0;
4143 if (sector_nr >= max_sector) {
4144 /* just being told to finish up .. nothing much to do */
4146 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery)) {
4151 if (mddev->curr_resync < max_sector) /* aborted */
4152 bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
4154 else /* completed sync */
4156 bitmap_close_sync(mddev->bitmap);
4161 /* Allow raid5_quiesce to complete */
4162 wait_event(conf->wait_for_overlap, conf->quiesce != 2);
4164 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
4165 return reshape_request(mddev, sector_nr, skipped);
4167 /* No need to check resync_max as we never do more than one
4168 * stripe, and as resync_max will always be on a chunk boundary,
4169 * if the check in md_do_sync didn't fire, there is no chance
4170 * of overstepping resync_max here
4173 /* if there is too many failed drives and we are trying
4174 * to resync, then assert that we are finished, because there is
4175 * nothing we can do.
4177 if (mddev->degraded >= conf->max_degraded &&
4178 test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
4179 sector_t rv = mddev->dev_sectors - sector_nr;
4183 if (!bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
4184 !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
4185 !conf->fullsync && sync_blocks >= STRIPE_SECTORS) {
4186 /* we can skip this block, and probably more */
4187 sync_blocks /= STRIPE_SECTORS;
4189 return sync_blocks * STRIPE_SECTORS; /* keep things rounded to whole stripes */
4193 bitmap_cond_end_sync(mddev->bitmap, sector_nr);
4195 sh = get_active_stripe(conf, sector_nr, 0, 1, 0);
4197 sh = get_active_stripe(conf, sector_nr, 0, 0, 0);
4198 /* make sure we don't swamp the stripe cache if someone else
4199 * is trying to get access
4201 schedule_timeout_uninterruptible(1);
4203 /* Need to check if array will still be degraded after recovery/resync
4204 * We don't need to check the 'failed' flag as when that gets set,
4207 for (i = 0; i < conf->raid_disks; i++)
4208 if (conf->disks[i].rdev == NULL)
4211 bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, still_degraded);
4213 set_bit(STRIPE_SYNC_REQUESTED, &sh->state);
4218 return STRIPE_SECTORS;
4221 static int retry_aligned_read(struct r5conf *conf, struct bio *raid_bio)
4223 /* We may not be able to submit a whole bio at once as there
4224 * may not be enough stripe_heads available.
4225 * We cannot pre-allocate enough stripe_heads as we may need
4226 * more than exist in the cache (if we allow ever large chunks).
4227 * So we do one stripe head at a time and record in
4228 * ->bi_hw_segments how many have been done.
4230 * We *know* that this entire raid_bio is in one chunk, so
4231 * it will be only one 'dd_idx' and only need one call to raid5_compute_sector.
4233 struct stripe_head *sh;
4235 sector_t sector, logical_sector, last_sector;
4240 logical_sector = raid_bio->bi_sector & ~((sector_t)STRIPE_SECTORS-1);
4241 sector = raid5_compute_sector(conf, logical_sector,
4243 last_sector = raid_bio->bi_sector + (raid_bio->bi_size>>9);
4245 for (; logical_sector < last_sector;
4246 logical_sector += STRIPE_SECTORS,
4247 sector += STRIPE_SECTORS,
4250 if (scnt < raid5_bi_hw_segments(raid_bio))
4251 /* already done this stripe */
4254 sh = get_active_stripe(conf, sector, 0, 1, 0);
4257 /* failed to get a stripe - must wait */
4258 raid5_set_bi_hw_segments(raid_bio, scnt);
4259 conf->retry_read_aligned = raid_bio;
4263 if (!add_stripe_bio(sh, raid_bio, dd_idx, 0)) {
4265 raid5_set_bi_hw_segments(raid_bio, scnt);
4266 conf->retry_read_aligned = raid_bio;
4274 spin_lock_irq(&conf->device_lock);
4275 remaining = raid5_dec_bi_phys_segments(raid_bio);
4276 spin_unlock_irq(&conf->device_lock);
4278 bio_endio(raid_bio, 0);
4279 if (atomic_dec_and_test(&conf->active_aligned_reads))
4280 wake_up(&conf->wait_for_stripe);
4286 * This is our raid5 kernel thread.
4288 * We scan the hash table for stripes which can be handled now.
4289 * During the scan, completed stripes are saved for us by the interrupt
4290 * handler, so that they will not have to wait for our next wakeup.
4292 static void raid5d(struct mddev *mddev)
4294 struct stripe_head *sh;
4295 struct r5conf *conf = mddev->private;
4297 struct blk_plug plug;
4299 pr_debug("+++ raid5d active\n");
4301 md_check_recovery(mddev);
4303 blk_start_plug(&plug);
4305 spin_lock_irq(&conf->device_lock);
4309 if (atomic_read(&mddev->plug_cnt) == 0 &&
4310 !list_empty(&conf->bitmap_list)) {
4311 /* Now is a good time to flush some bitmap updates */
4313 spin_unlock_irq(&conf->device_lock);
4314 bitmap_unplug(mddev->bitmap);
4315 spin_lock_irq(&conf->device_lock);
4316 conf->seq_write = conf->seq_flush;
4317 activate_bit_delay(conf);
4319 if (atomic_read(&mddev->plug_cnt) == 0)
4320 raid5_activate_delayed(conf);
4322 while ((bio = remove_bio_from_retry(conf))) {
4324 spin_unlock_irq(&conf->device_lock);
4325 ok = retry_aligned_read(conf, bio);
4326 spin_lock_irq(&conf->device_lock);
4332 sh = __get_priority_stripe(conf);
4336 spin_unlock_irq(&conf->device_lock);
4343 if (mddev->flags & ~(1<<MD_CHANGE_PENDING))
4344 md_check_recovery(mddev);
4346 spin_lock_irq(&conf->device_lock);
4348 pr_debug("%d stripes handled\n", handled);
4350 spin_unlock_irq(&conf->device_lock);
4352 async_tx_issue_pending_all();
4353 blk_finish_plug(&plug);
4355 pr_debug("--- raid5d inactive\n");
4359 raid5_show_stripe_cache_size(struct mddev *mddev, char *page)
4361 struct r5conf *conf = mddev->private;
4363 return sprintf(page, "%d\n", conf->max_nr_stripes);
4369 raid5_set_cache_size(struct mddev *mddev, int size)
4371 struct r5conf *conf = mddev->private;
4374 if (size <= 16 || size > 32768)
4376 while (size < conf->max_nr_stripes) {
4377 if (drop_one_stripe(conf))
4378 conf->max_nr_stripes--;
4382 err = md_allow_write(mddev);
4385 while (size > conf->max_nr_stripes) {
4386 if (grow_one_stripe(conf))
4387 conf->max_nr_stripes++;
4392 EXPORT_SYMBOL(raid5_set_cache_size);
4395 raid5_store_stripe_cache_size(struct mddev *mddev, const char *page, size_t len)
4397 struct r5conf *conf = mddev->private;
4401 if (len >= PAGE_SIZE)
4406 if (strict_strtoul(page, 10, &new))
4408 err = raid5_set_cache_size(mddev, new);
4414 static struct md_sysfs_entry
4415 raid5_stripecache_size = __ATTR(stripe_cache_size, S_IRUGO | S_IWUSR,
4416 raid5_show_stripe_cache_size,
4417 raid5_store_stripe_cache_size);
4420 raid5_show_preread_threshold(struct mddev *mddev, char *page)
4422 struct r5conf *conf = mddev->private;
4424 return sprintf(page, "%d\n", conf->bypass_threshold);
4430 raid5_store_preread_threshold(struct mddev *mddev, const char *page, size_t len)
4432 struct r5conf *conf = mddev->private;
4434 if (len >= PAGE_SIZE)
4439 if (strict_strtoul(page, 10, &new))
4441 if (new > conf->max_nr_stripes)
4443 conf->bypass_threshold = new;
4447 static struct md_sysfs_entry
4448 raid5_preread_bypass_threshold = __ATTR(preread_bypass_threshold,
4450 raid5_show_preread_threshold,
4451 raid5_store_preread_threshold);
4454 stripe_cache_active_show(struct mddev *mddev, char *page)
4456 struct r5conf *conf = mddev->private;
4458 return sprintf(page, "%d\n", atomic_read(&conf->active_stripes));
4463 static struct md_sysfs_entry
4464 raid5_stripecache_active = __ATTR_RO(stripe_cache_active);
4466 static struct attribute *raid5_attrs[] = {
4467 &raid5_stripecache_size.attr,
4468 &raid5_stripecache_active.attr,
4469 &raid5_preread_bypass_threshold.attr,
4472 static struct attribute_group raid5_attrs_group = {
4474 .attrs = raid5_attrs,
4478 raid5_size(struct mddev *mddev, sector_t sectors, int raid_disks)
4480 struct r5conf *conf = mddev->private;
4483 sectors = mddev->dev_sectors;
4485 /* size is defined by the smallest of previous and new size */
4486 raid_disks = min(conf->raid_disks, conf->previous_raid_disks);
4488 sectors &= ~((sector_t)mddev->chunk_sectors - 1);
4489 sectors &= ~((sector_t)mddev->new_chunk_sectors - 1);
4490 return sectors * (raid_disks - conf->max_degraded);
4493 static void raid5_free_percpu(struct r5conf *conf)
4495 struct raid5_percpu *percpu;
4502 for_each_possible_cpu(cpu) {
4503 percpu = per_cpu_ptr(conf->percpu, cpu);
4504 safe_put_page(percpu->spare_page);
4505 kfree(percpu->scribble);
4507 #ifdef CONFIG_HOTPLUG_CPU
4508 unregister_cpu_notifier(&conf->cpu_notify);
4512 free_percpu(conf->percpu);
4515 static void free_conf(struct r5conf *conf)
4517 shrink_stripes(conf);
4518 raid5_free_percpu(conf);
4520 kfree(conf->stripe_hashtbl);
4524 #ifdef CONFIG_HOTPLUG_CPU
4525 static int raid456_cpu_notify(struct notifier_block *nfb, unsigned long action,
4528 struct r5conf *conf = container_of(nfb, struct r5conf, cpu_notify);
4529 long cpu = (long)hcpu;
4530 struct raid5_percpu *percpu = per_cpu_ptr(conf->percpu, cpu);
4533 case CPU_UP_PREPARE:
4534 case CPU_UP_PREPARE_FROZEN:
4535 if (conf->level == 6 && !percpu->spare_page)
4536 percpu->spare_page = alloc_page(GFP_KERNEL);
4537 if (!percpu->scribble)
4538 percpu->scribble = kmalloc(conf->scribble_len, GFP_KERNEL);
4540 if (!percpu->scribble ||
4541 (conf->level == 6 && !percpu->spare_page)) {
4542 safe_put_page(percpu->spare_page);
4543 kfree(percpu->scribble);
4544 pr_err("%s: failed memory allocation for cpu%ld\n",
4546 return notifier_from_errno(-ENOMEM);
4550 case CPU_DEAD_FROZEN:
4551 safe_put_page(percpu->spare_page);
4552 kfree(percpu->scribble);
4553 percpu->spare_page = NULL;
4554 percpu->scribble = NULL;
4563 static int raid5_alloc_percpu(struct r5conf *conf)
4566 struct page *spare_page;
4567 struct raid5_percpu __percpu *allcpus;
4571 allcpus = alloc_percpu(struct raid5_percpu);
4574 conf->percpu = allcpus;
4578 for_each_present_cpu(cpu) {
4579 if (conf->level == 6) {
4580 spare_page = alloc_page(GFP_KERNEL);
4585 per_cpu_ptr(conf->percpu, cpu)->spare_page = spare_page;
4587 scribble = kmalloc(conf->scribble_len, GFP_KERNEL);
4592 per_cpu_ptr(conf->percpu, cpu)->scribble = scribble;
4594 #ifdef CONFIG_HOTPLUG_CPU
4595 conf->cpu_notify.notifier_call = raid456_cpu_notify;
4596 conf->cpu_notify.priority = 0;
4598 err = register_cpu_notifier(&conf->cpu_notify);
4605 static struct r5conf *setup_conf(struct mddev *mddev)
4607 struct r5conf *conf;
4608 int raid_disk, memory, max_disks;
4609 struct md_rdev *rdev;
4610 struct disk_info *disk;
4612 if (mddev->new_level != 5
4613 && mddev->new_level != 4
4614 && mddev->new_level != 6) {
4615 printk(KERN_ERR "md/raid:%s: raid level not set to 4/5/6 (%d)\n",
4616 mdname(mddev), mddev->new_level);
4617 return ERR_PTR(-EIO);
4619 if ((mddev->new_level == 5
4620 && !algorithm_valid_raid5(mddev->new_layout)) ||
4621 (mddev->new_level == 6
4622 && !algorithm_valid_raid6(mddev->new_layout))) {
4623 printk(KERN_ERR "md/raid:%s: layout %d not supported\n",
4624 mdname(mddev), mddev->new_layout);
4625 return ERR_PTR(-EIO);
4627 if (mddev->new_level == 6 && mddev->raid_disks < 4) {
4628 printk(KERN_ERR "md/raid:%s: not enough configured devices (%d, minimum 4)\n",
4629 mdname(mddev), mddev->raid_disks);
4630 return ERR_PTR(-EINVAL);
4633 if (!mddev->new_chunk_sectors ||
4634 (mddev->new_chunk_sectors << 9) % PAGE_SIZE ||
4635 !is_power_of_2(mddev->new_chunk_sectors)) {
4636 printk(KERN_ERR "md/raid:%s: invalid chunk size %d\n",
4637 mdname(mddev), mddev->new_chunk_sectors << 9);
4638 return ERR_PTR(-EINVAL);
4641 conf = kzalloc(sizeof(struct r5conf), GFP_KERNEL);
4644 spin_lock_init(&conf->device_lock);
4645 init_waitqueue_head(&conf->wait_for_stripe);
4646 init_waitqueue_head(&conf->wait_for_overlap);
4647 INIT_LIST_HEAD(&conf->handle_list);
4648 INIT_LIST_HEAD(&conf->hold_list);
4649 INIT_LIST_HEAD(&conf->delayed_list);
4650 INIT_LIST_HEAD(&conf->bitmap_list);
4651 INIT_LIST_HEAD(&conf->inactive_list);
4652 atomic_set(&conf->active_stripes, 0);
4653 atomic_set(&conf->preread_active_stripes, 0);
4654 atomic_set(&conf->active_aligned_reads, 0);
4655 conf->bypass_threshold = BYPASS_THRESHOLD;
4656 conf->recovery_disabled = mddev->recovery_disabled - 1;
4658 conf->raid_disks = mddev->raid_disks;
4659 if (mddev->reshape_position == MaxSector)
4660 conf->previous_raid_disks = mddev->raid_disks;
4662 conf->previous_raid_disks = mddev->raid_disks - mddev->delta_disks;
4663 max_disks = max(conf->raid_disks, conf->previous_raid_disks);
4664 conf->scribble_len = scribble_len(max_disks);
4666 conf->disks = kzalloc(max_disks * sizeof(struct disk_info),
4671 conf->mddev = mddev;
4673 if ((conf->stripe_hashtbl = kzalloc(PAGE_SIZE, GFP_KERNEL)) == NULL)
4676 conf->level = mddev->new_level;
4677 if (raid5_alloc_percpu(conf) != 0)
4680 pr_debug("raid456: run(%s) called.\n", mdname(mddev));
4682 list_for_each_entry(rdev, &mddev->disks, same_set) {
4683 raid_disk = rdev->raid_disk;
4684 if (raid_disk >= max_disks
4687 disk = conf->disks + raid_disk;
4691 if (test_bit(In_sync, &rdev->flags)) {
4692 char b[BDEVNAME_SIZE];
4693 printk(KERN_INFO "md/raid:%s: device %s operational as raid"
4695 mdname(mddev), bdevname(rdev->bdev, b), raid_disk);
4696 } else if (rdev->saved_raid_disk != raid_disk)
4697 /* Cannot rely on bitmap to complete recovery */
4701 conf->chunk_sectors = mddev->new_chunk_sectors;
4702 conf->level = mddev->new_level;
4703 if (conf->level == 6)
4704 conf->max_degraded = 2;
4706 conf->max_degraded = 1;
4707 conf->algorithm = mddev->new_layout;
4708 conf->max_nr_stripes = NR_STRIPES;
4709 conf->reshape_progress = mddev->reshape_position;
4710 if (conf->reshape_progress != MaxSector) {
4711 conf->prev_chunk_sectors = mddev->chunk_sectors;
4712 conf->prev_algo = mddev->layout;
4715 memory = conf->max_nr_stripes * (sizeof(struct stripe_head) +
4716 max_disks * ((sizeof(struct bio) + PAGE_SIZE))) / 1024;
4717 if (grow_stripes(conf, conf->max_nr_stripes)) {
4719 "md/raid:%s: couldn't allocate %dkB for buffers\n",
4720 mdname(mddev), memory);
4723 printk(KERN_INFO "md/raid:%s: allocated %dkB\n",
4724 mdname(mddev), memory);
4726 conf->thread = md_register_thread(raid5d, mddev, NULL);
4727 if (!conf->thread) {
4729 "md/raid:%s: couldn't allocate thread.\n",
4739 return ERR_PTR(-EIO);
4741 return ERR_PTR(-ENOMEM);
4745 static int only_parity(int raid_disk, int algo, int raid_disks, int max_degraded)
4748 case ALGORITHM_PARITY_0:
4749 if (raid_disk < max_degraded)
4752 case ALGORITHM_PARITY_N:
4753 if (raid_disk >= raid_disks - max_degraded)
4756 case ALGORITHM_PARITY_0_6:
4757 if (raid_disk == 0 ||
4758 raid_disk == raid_disks - 1)
4761 case ALGORITHM_LEFT_ASYMMETRIC_6:
4762 case ALGORITHM_RIGHT_ASYMMETRIC_6:
4763 case ALGORITHM_LEFT_SYMMETRIC_6:
4764 case ALGORITHM_RIGHT_SYMMETRIC_6:
4765 if (raid_disk == raid_disks - 1)
4771 static int run(struct mddev *mddev)
4773 struct r5conf *conf;
4774 int working_disks = 0;
4775 int dirty_parity_disks = 0;
4776 struct md_rdev *rdev;
4777 sector_t reshape_offset = 0;
4779 if (mddev->recovery_cp != MaxSector)
4780 printk(KERN_NOTICE "md/raid:%s: not clean"
4781 " -- starting background reconstruction\n",
4783 if (mddev->reshape_position != MaxSector) {
4784 /* Check that we can continue the reshape.
4785 * Currently only disks can change, it must
4786 * increase, and we must be past the point where
4787 * a stripe over-writes itself
4789 sector_t here_new, here_old;
4791 int max_degraded = (mddev->level == 6 ? 2 : 1);
4793 if (mddev->new_level != mddev->level) {
4794 printk(KERN_ERR "md/raid:%s: unsupported reshape "
4795 "required - aborting.\n",
4799 old_disks = mddev->raid_disks - mddev->delta_disks;
4800 /* reshape_position must be on a new-stripe boundary, and one
4801 * further up in new geometry must map after here in old
4804 here_new = mddev->reshape_position;
4805 if (sector_div(here_new, mddev->new_chunk_sectors *
4806 (mddev->raid_disks - max_degraded))) {
4807 printk(KERN_ERR "md/raid:%s: reshape_position not "
4808 "on a stripe boundary\n", mdname(mddev));
4811 reshape_offset = here_new * mddev->new_chunk_sectors;
4812 /* here_new is the stripe we will write to */
4813 here_old = mddev->reshape_position;
4814 sector_div(here_old, mddev->chunk_sectors *
4815 (old_disks-max_degraded));
4816 /* here_old is the first stripe that we might need to read
4818 if (mddev->delta_disks == 0) {
4819 /* We cannot be sure it is safe to start an in-place
4820 * reshape. It is only safe if user-space if monitoring
4821 * and taking constant backups.
4822 * mdadm always starts a situation like this in
4823 * readonly mode so it can take control before
4824 * allowing any writes. So just check for that.
4826 if ((here_new * mddev->new_chunk_sectors !=
4827 here_old * mddev->chunk_sectors) ||
4829 printk(KERN_ERR "md/raid:%s: in-place reshape must be started"
4830 " in read-only mode - aborting\n",
4834 } else if (mddev->delta_disks < 0
4835 ? (here_new * mddev->new_chunk_sectors <=
4836 here_old * mddev->chunk_sectors)
4837 : (here_new * mddev->new_chunk_sectors >=
4838 here_old * mddev->chunk_sectors)) {
4839 /* Reading from the same stripe as writing to - bad */
4840 printk(KERN_ERR "md/raid:%s: reshape_position too early for "
4841 "auto-recovery - aborting.\n",
4845 printk(KERN_INFO "md/raid:%s: reshape will continue\n",
4847 /* OK, we should be able to continue; */
4849 BUG_ON(mddev->level != mddev->new_level);
4850 BUG_ON(mddev->layout != mddev->new_layout);
4851 BUG_ON(mddev->chunk_sectors != mddev->new_chunk_sectors);
4852 BUG_ON(mddev->delta_disks != 0);
4855 if (mddev->private == NULL)
4856 conf = setup_conf(mddev);
4858 conf = mddev->private;
4861 return PTR_ERR(conf);
4863 mddev->thread = conf->thread;
4864 conf->thread = NULL;
4865 mddev->private = conf;
4868 * 0 for a fully functional array, 1 or 2 for a degraded array.
4870 list_for_each_entry(rdev, &mddev->disks, same_set) {
4871 if (rdev->raid_disk < 0)
4873 if (test_bit(In_sync, &rdev->flags)) {
4877 /* This disc is not fully in-sync. However if it
4878 * just stored parity (beyond the recovery_offset),
4879 * when we don't need to be concerned about the
4880 * array being dirty.
4881 * When reshape goes 'backwards', we never have
4882 * partially completed devices, so we only need
4883 * to worry about reshape going forwards.
4885 /* Hack because v0.91 doesn't store recovery_offset properly. */
4886 if (mddev->major_version == 0 &&
4887 mddev->minor_version > 90)
4888 rdev->recovery_offset = reshape_offset;
4890 if (rdev->recovery_offset < reshape_offset) {
4891 /* We need to check old and new layout */
4892 if (!only_parity(rdev->raid_disk,
4895 conf->max_degraded))
4898 if (!only_parity(rdev->raid_disk,
4900 conf->previous_raid_disks,
4901 conf->max_degraded))
4903 dirty_parity_disks++;
4906 mddev->degraded = calc_degraded(conf);
4908 if (has_failed(conf)) {
4909 printk(KERN_ERR "md/raid:%s: not enough operational devices"
4910 " (%d/%d failed)\n",
4911 mdname(mddev), mddev->degraded, conf->raid_disks);
4915 /* device size must be a multiple of chunk size */
4916 mddev->dev_sectors &= ~(mddev->chunk_sectors - 1);
4917 mddev->resync_max_sectors = mddev->dev_sectors;
4919 if (mddev->degraded > dirty_parity_disks &&
4920 mddev->recovery_cp != MaxSector) {
4921 if (mddev->ok_start_degraded)
4923 "md/raid:%s: starting dirty degraded array"
4924 " - data corruption possible.\n",
4928 "md/raid:%s: cannot start dirty degraded array.\n",
4934 if (mddev->degraded == 0)
4935 printk(KERN_INFO "md/raid:%s: raid level %d active with %d out of %d"
4936 " devices, algorithm %d\n", mdname(mddev), conf->level,
4937 mddev->raid_disks-mddev->degraded, mddev->raid_disks,
4940 printk(KERN_ALERT "md/raid:%s: raid level %d active with %d"
4941 " out of %d devices, algorithm %d\n",
4942 mdname(mddev), conf->level,
4943 mddev->raid_disks - mddev->degraded,
4944 mddev->raid_disks, mddev->new_layout);
4946 print_raid5_conf(conf);
4948 if (conf->reshape_progress != MaxSector) {
4949 conf->reshape_safe = conf->reshape_progress;
4950 atomic_set(&conf->reshape_stripes, 0);
4951 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
4952 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
4953 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
4954 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
4955 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
4960 /* Ok, everything is just fine now */
4961 if (mddev->to_remove == &raid5_attrs_group)
4962 mddev->to_remove = NULL;
4963 else if (mddev->kobj.sd &&
4964 sysfs_create_group(&mddev->kobj, &raid5_attrs_group))
4966 "raid5: failed to create sysfs attributes for %s\n",
4968 md_set_array_sectors(mddev, raid5_size(mddev, 0, 0));
4972 /* read-ahead size must cover two whole stripes, which
4973 * is 2 * (datadisks) * chunksize where 'n' is the
4974 * number of raid devices
4976 int data_disks = conf->previous_raid_disks - conf->max_degraded;
4977 int stripe = data_disks *
4978 ((mddev->chunk_sectors << 9) / PAGE_SIZE);
4979 if (mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
4980 mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
4982 blk_queue_merge_bvec(mddev->queue, raid5_mergeable_bvec);
4984 mddev->queue->backing_dev_info.congested_data = mddev;
4985 mddev->queue->backing_dev_info.congested_fn = raid5_congested;
4987 chunk_size = mddev->chunk_sectors << 9;
4988 blk_queue_io_min(mddev->queue, chunk_size);
4989 blk_queue_io_opt(mddev->queue, chunk_size *
4990 (conf->raid_disks - conf->max_degraded));
4992 list_for_each_entry(rdev, &mddev->disks, same_set)
4993 disk_stack_limits(mddev->gendisk, rdev->bdev,
4994 rdev->data_offset << 9);
4999 md_unregister_thread(&mddev->thread);
5000 print_raid5_conf(conf);
5002 mddev->private = NULL;
5003 printk(KERN_ALERT "md/raid:%s: failed to run raid set.\n", mdname(mddev));
5007 static int stop(struct mddev *mddev)
5009 struct r5conf *conf = mddev->private;
5011 md_unregister_thread(&mddev->thread);
5013 mddev->queue->backing_dev_info.congested_fn = NULL;
5015 mddev->private = NULL;
5016 mddev->to_remove = &raid5_attrs_group;
5020 static void status(struct seq_file *seq, struct mddev *mddev)
5022 struct r5conf *conf = mddev->private;
5025 seq_printf(seq, " level %d, %dk chunk, algorithm %d", mddev->level,
5026 mddev->chunk_sectors / 2, mddev->layout);
5027 seq_printf (seq, " [%d/%d] [", conf->raid_disks, conf->raid_disks - mddev->degraded);
5028 for (i = 0; i < conf->raid_disks; i++)
5029 seq_printf (seq, "%s",
5030 conf->disks[i].rdev &&
5031 test_bit(In_sync, &conf->disks[i].rdev->flags) ? "U" : "_");
5032 seq_printf (seq, "]");
5035 static void print_raid5_conf (struct r5conf *conf)
5038 struct disk_info *tmp;
5040 printk(KERN_DEBUG "RAID conf printout:\n");
5042 printk("(conf==NULL)\n");
5045 printk(KERN_DEBUG " --- level:%d rd:%d wd:%d\n", conf->level,
5047 conf->raid_disks - conf->mddev->degraded);
5049 for (i = 0; i < conf->raid_disks; i++) {
5050 char b[BDEVNAME_SIZE];
5051 tmp = conf->disks + i;
5053 printk(KERN_DEBUG " disk %d, o:%d, dev:%s\n",
5054 i, !test_bit(Faulty, &tmp->rdev->flags),
5055 bdevname(tmp->rdev->bdev, b));
5059 static int raid5_spare_active(struct mddev *mddev)
5062 struct r5conf *conf = mddev->private;
5063 struct disk_info *tmp;
5065 unsigned long flags;
5067 for (i = 0; i < conf->raid_disks; i++) {
5068 tmp = conf->disks + i;
5070 && tmp->rdev->recovery_offset == MaxSector
5071 && !test_bit(Faulty, &tmp->rdev->flags)
5072 && !test_and_set_bit(In_sync, &tmp->rdev->flags)) {
5074 sysfs_notify_dirent_safe(tmp->rdev->sysfs_state);
5077 spin_lock_irqsave(&conf->device_lock, flags);
5078 mddev->degraded = calc_degraded(conf);
5079 spin_unlock_irqrestore(&conf->device_lock, flags);
5080 print_raid5_conf(conf);
5084 static int raid5_remove_disk(struct mddev *mddev, struct md_rdev *rdev)
5086 struct r5conf *conf = mddev->private;
5088 int number = rdev->raid_disk;
5089 struct disk_info *p = conf->disks + number;
5091 print_raid5_conf(conf);
5092 if (rdev == p->rdev) {
5093 if (number >= conf->raid_disks &&
5094 conf->reshape_progress == MaxSector)
5095 clear_bit(In_sync, &rdev->flags);
5097 if (test_bit(In_sync, &rdev->flags) ||
5098 atomic_read(&rdev->nr_pending)) {
5102 /* Only remove non-faulty devices if recovery
5105 if (!test_bit(Faulty, &rdev->flags) &&
5106 mddev->recovery_disabled != conf->recovery_disabled &&
5107 !has_failed(conf) &&
5108 number < conf->raid_disks) {
5114 if (atomic_read(&rdev->nr_pending)) {
5115 /* lost the race, try later */
5122 print_raid5_conf(conf);
5126 static int raid5_add_disk(struct mddev *mddev, struct md_rdev *rdev)
5128 struct r5conf *conf = mddev->private;
5131 struct disk_info *p;
5133 int last = conf->raid_disks - 1;
5135 if (mddev->recovery_disabled == conf->recovery_disabled)
5138 if (has_failed(conf))
5139 /* no point adding a device */
5142 if (rdev->raid_disk >= 0)
5143 first = last = rdev->raid_disk;
5146 * find the disk ... but prefer rdev->saved_raid_disk
5149 if (rdev->saved_raid_disk >= 0 &&
5150 rdev->saved_raid_disk >= first &&
5151 conf->disks[rdev->saved_raid_disk].rdev == NULL)
5152 disk = rdev->saved_raid_disk;
5155 for ( ; disk <= last ; disk++)
5156 if ((p=conf->disks + disk)->rdev == NULL) {
5157 clear_bit(In_sync, &rdev->flags);
5158 rdev->raid_disk = disk;
5160 if (rdev->saved_raid_disk != disk)
5162 rcu_assign_pointer(p->rdev, rdev);
5165 print_raid5_conf(conf);
5169 static int raid5_resize(struct mddev *mddev, sector_t sectors)
5171 /* no resync is happening, and there is enough space
5172 * on all devices, so we can resize.
5173 * We need to make sure resync covers any new space.
5174 * If the array is shrinking we should possibly wait until
5175 * any io in the removed space completes, but it hardly seems
5178 sectors &= ~((sector_t)mddev->chunk_sectors - 1);
5179 md_set_array_sectors(mddev, raid5_size(mddev, sectors,
5180 mddev->raid_disks));
5181 if (mddev->array_sectors >
5182 raid5_size(mddev, sectors, mddev->raid_disks))
5184 set_capacity(mddev->gendisk, mddev->array_sectors);
5185 revalidate_disk(mddev->gendisk);
5186 if (sectors > mddev->dev_sectors &&
5187 mddev->recovery_cp > mddev->dev_sectors) {
5188 mddev->recovery_cp = mddev->dev_sectors;
5189 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
5191 mddev->dev_sectors = sectors;
5192 mddev->resync_max_sectors = sectors;
5196 static int check_stripe_cache(struct mddev *mddev)
5198 /* Can only proceed if there are plenty of stripe_heads.
5199 * We need a minimum of one full stripe,, and for sensible progress
5200 * it is best to have about 4 times that.
5201 * If we require 4 times, then the default 256 4K stripe_heads will
5202 * allow for chunk sizes up to 256K, which is probably OK.
5203 * If the chunk size is greater, user-space should request more
5204 * stripe_heads first.
5206 struct r5conf *conf = mddev->private;
5207 if (((mddev->chunk_sectors << 9) / STRIPE_SIZE) * 4
5208 > conf->max_nr_stripes ||
5209 ((mddev->new_chunk_sectors << 9) / STRIPE_SIZE) * 4
5210 > conf->max_nr_stripes) {
5211 printk(KERN_WARNING "md/raid:%s: reshape: not enough stripes. Needed %lu\n",
5213 ((max(mddev->chunk_sectors, mddev->new_chunk_sectors) << 9)
5220 static int check_reshape(struct mddev *mddev)
5222 struct r5conf *conf = mddev->private;
5224 if (mddev->delta_disks == 0 &&
5225 mddev->new_layout == mddev->layout &&
5226 mddev->new_chunk_sectors == mddev->chunk_sectors)
5227 return 0; /* nothing to do */
5229 /* Cannot grow a bitmap yet */
5231 if (has_failed(conf))
5233 if (mddev->delta_disks < 0) {
5234 /* We might be able to shrink, but the devices must
5235 * be made bigger first.
5236 * For raid6, 4 is the minimum size.
5237 * Otherwise 2 is the minimum
5240 if (mddev->level == 6)
5242 if (mddev->raid_disks + mddev->delta_disks < min)
5246 if (!check_stripe_cache(mddev))
5249 return resize_stripes(conf, conf->raid_disks + mddev->delta_disks);
5252 static int raid5_start_reshape(struct mddev *mddev)
5254 struct r5conf *conf = mddev->private;
5255 struct md_rdev *rdev;
5257 unsigned long flags;
5259 if (test_bit(MD_RECOVERY_RUNNING, &mddev->recovery))
5262 if (!check_stripe_cache(mddev))
5265 list_for_each_entry(rdev, &mddev->disks, same_set)
5266 if (!test_bit(In_sync, &rdev->flags)
5267 && !test_bit(Faulty, &rdev->flags))
5270 if (spares - mddev->degraded < mddev->delta_disks - conf->max_degraded)
5271 /* Not enough devices even to make a degraded array
5276 /* Refuse to reduce size of the array. Any reductions in
5277 * array size must be through explicit setting of array_size
5280 if (raid5_size(mddev, 0, conf->raid_disks + mddev->delta_disks)
5281 < mddev->array_sectors) {
5282 printk(KERN_ERR "md/raid:%s: array size must be reduced "
5283 "before number of disks\n", mdname(mddev));
5287 atomic_set(&conf->reshape_stripes, 0);
5288 spin_lock_irq(&conf->device_lock);
5289 conf->previous_raid_disks = conf->raid_disks;
5290 conf->raid_disks += mddev->delta_disks;
5291 conf->prev_chunk_sectors = conf->chunk_sectors;
5292 conf->chunk_sectors = mddev->new_chunk_sectors;
5293 conf->prev_algo = conf->algorithm;
5294 conf->algorithm = mddev->new_layout;
5295 if (mddev->delta_disks < 0)
5296 conf->reshape_progress = raid5_size(mddev, 0, 0);
5298 conf->reshape_progress = 0;
5299 conf->reshape_safe = conf->reshape_progress;
5301 spin_unlock_irq(&conf->device_lock);
5303 /* Add some new drives, as many as will fit.
5304 * We know there are enough to make the newly sized array work.
5305 * Don't add devices if we are reducing the number of
5306 * devices in the array. This is because it is not possible
5307 * to correctly record the "partially reconstructed" state of
5308 * such devices during the reshape and confusion could result.
5310 if (mddev->delta_disks >= 0) {
5311 int added_devices = 0;
5312 list_for_each_entry(rdev, &mddev->disks, same_set)
5313 if (rdev->raid_disk < 0 &&
5314 !test_bit(Faulty, &rdev->flags)) {
5315 if (raid5_add_disk(mddev, rdev) == 0) {
5317 >= conf->previous_raid_disks) {
5318 set_bit(In_sync, &rdev->flags);
5321 rdev->recovery_offset = 0;
5323 if (sysfs_link_rdev(mddev, rdev))
5324 /* Failure here is OK */;
5326 } else if (rdev->raid_disk >= conf->previous_raid_disks
5327 && !test_bit(Faulty, &rdev->flags)) {
5328 /* This is a spare that was manually added */
5329 set_bit(In_sync, &rdev->flags);
5333 /* When a reshape changes the number of devices,
5334 * ->degraded is measured against the larger of the
5335 * pre and post number of devices.
5337 spin_lock_irqsave(&conf->device_lock, flags);
5338 mddev->degraded = calc_degraded(conf);
5339 spin_unlock_irqrestore(&conf->device_lock, flags);
5341 mddev->raid_disks = conf->raid_disks;
5342 mddev->reshape_position = conf->reshape_progress;
5343 set_bit(MD_CHANGE_DEVS, &mddev->flags);
5345 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
5346 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
5347 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
5348 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
5349 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
5351 if (!mddev->sync_thread) {
5352 mddev->recovery = 0;
5353 spin_lock_irq(&conf->device_lock);
5354 mddev->raid_disks = conf->raid_disks = conf->previous_raid_disks;
5355 conf->reshape_progress = MaxSector;
5356 spin_unlock_irq(&conf->device_lock);
5359 conf->reshape_checkpoint = jiffies;
5360 md_wakeup_thread(mddev->sync_thread);
5361 md_new_event(mddev);
5365 /* This is called from the reshape thread and should make any
5366 * changes needed in 'conf'
5368 static void end_reshape(struct r5conf *conf)
5371 if (!test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery)) {
5373 spin_lock_irq(&conf->device_lock);
5374 conf->previous_raid_disks = conf->raid_disks;
5375 conf->reshape_progress = MaxSector;
5376 spin_unlock_irq(&conf->device_lock);
5377 wake_up(&conf->wait_for_overlap);
5379 /* read-ahead size must cover two whole stripes, which is
5380 * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
5382 if (conf->mddev->queue) {
5383 int data_disks = conf->raid_disks - conf->max_degraded;
5384 int stripe = data_disks * ((conf->chunk_sectors << 9)
5386 if (conf->mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
5387 conf->mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
5392 /* This is called from the raid5d thread with mddev_lock held.
5393 * It makes config changes to the device.
5395 static void raid5_finish_reshape(struct mddev *mddev)
5397 struct r5conf *conf = mddev->private;
5399 if (!test_bit(MD_RECOVERY_INTR, &mddev->recovery)) {
5401 if (mddev->delta_disks > 0) {
5402 md_set_array_sectors(mddev, raid5_size(mddev, 0, 0));
5403 set_capacity(mddev->gendisk, mddev->array_sectors);
5404 revalidate_disk(mddev->gendisk);
5407 spin_lock_irq(&conf->device_lock);
5408 mddev->degraded = calc_degraded(conf);
5409 spin_unlock_irq(&conf->device_lock);
5410 for (d = conf->raid_disks ;
5411 d < conf->raid_disks - mddev->delta_disks;
5413 struct md_rdev *rdev = conf->disks[d].rdev;
5415 raid5_remove_disk(mddev, rdev) == 0) {
5416 sysfs_unlink_rdev(mddev, rdev);
5417 rdev->raid_disk = -1;
5421 mddev->layout = conf->algorithm;
5422 mddev->chunk_sectors = conf->chunk_sectors;
5423 mddev->reshape_position = MaxSector;
5424 mddev->delta_disks = 0;
5428 static void raid5_quiesce(struct mddev *mddev, int state)
5430 struct r5conf *conf = mddev->private;
5433 case 2: /* resume for a suspend */
5434 wake_up(&conf->wait_for_overlap);
5437 case 1: /* stop all writes */
5438 spin_lock_irq(&conf->device_lock);
5439 /* '2' tells resync/reshape to pause so that all
5440 * active stripes can drain
5443 wait_event_lock_irq(conf->wait_for_stripe,
5444 atomic_read(&conf->active_stripes) == 0 &&
5445 atomic_read(&conf->active_aligned_reads) == 0,
5446 conf->device_lock, /* nothing */);
5448 spin_unlock_irq(&conf->device_lock);
5449 /* allow reshape to continue */
5450 wake_up(&conf->wait_for_overlap);
5453 case 0: /* re-enable writes */
5454 spin_lock_irq(&conf->device_lock);
5456 wake_up(&conf->wait_for_stripe);
5457 wake_up(&conf->wait_for_overlap);
5458 spin_unlock_irq(&conf->device_lock);
5464 static void *raid45_takeover_raid0(struct mddev *mddev, int level)
5466 struct r0conf *raid0_conf = mddev->private;
5469 /* for raid0 takeover only one zone is supported */
5470 if (raid0_conf->nr_strip_zones > 1) {
5471 printk(KERN_ERR "md/raid:%s: cannot takeover raid0 with more than one zone.\n",
5473 return ERR_PTR(-EINVAL);
5476 sectors = raid0_conf->strip_zone[0].zone_end;
5477 sector_div(sectors, raid0_conf->strip_zone[0].nb_dev);
5478 mddev->dev_sectors = sectors;
5479 mddev->new_level = level;
5480 mddev->new_layout = ALGORITHM_PARITY_N;
5481 mddev->new_chunk_sectors = mddev->chunk_sectors;
5482 mddev->raid_disks += 1;
5483 mddev->delta_disks = 1;
5484 /* make sure it will be not marked as dirty */
5485 mddev->recovery_cp = MaxSector;
5487 return setup_conf(mddev);
5491 static void *raid5_takeover_raid1(struct mddev *mddev)
5495 if (mddev->raid_disks != 2 ||
5496 mddev->degraded > 1)
5497 return ERR_PTR(-EINVAL);
5499 /* Should check if there are write-behind devices? */
5501 chunksect = 64*2; /* 64K by default */
5503 /* The array must be an exact multiple of chunksize */
5504 while (chunksect && (mddev->array_sectors & (chunksect-1)))
5507 if ((chunksect<<9) < STRIPE_SIZE)
5508 /* array size does not allow a suitable chunk size */
5509 return ERR_PTR(-EINVAL);
5511 mddev->new_level = 5;
5512 mddev->new_layout = ALGORITHM_LEFT_SYMMETRIC;
5513 mddev->new_chunk_sectors = chunksect;
5515 return setup_conf(mddev);
5518 static void *raid5_takeover_raid6(struct mddev *mddev)
5522 switch (mddev->layout) {
5523 case ALGORITHM_LEFT_ASYMMETRIC_6:
5524 new_layout = ALGORITHM_LEFT_ASYMMETRIC;
5526 case ALGORITHM_RIGHT_ASYMMETRIC_6:
5527 new_layout = ALGORITHM_RIGHT_ASYMMETRIC;
5529 case ALGORITHM_LEFT_SYMMETRIC_6:
5530 new_layout = ALGORITHM_LEFT_SYMMETRIC;
5532 case ALGORITHM_RIGHT_SYMMETRIC_6:
5533 new_layout = ALGORITHM_RIGHT_SYMMETRIC;
5535 case ALGORITHM_PARITY_0_6:
5536 new_layout = ALGORITHM_PARITY_0;
5538 case ALGORITHM_PARITY_N:
5539 new_layout = ALGORITHM_PARITY_N;
5542 return ERR_PTR(-EINVAL);
5544 mddev->new_level = 5;
5545 mddev->new_layout = new_layout;
5546 mddev->delta_disks = -1;
5547 mddev->raid_disks -= 1;
5548 return setup_conf(mddev);
5552 static int raid5_check_reshape(struct mddev *mddev)
5554 /* For a 2-drive array, the layout and chunk size can be changed
5555 * immediately as not restriping is needed.
5556 * For larger arrays we record the new value - after validation
5557 * to be used by a reshape pass.
5559 struct r5conf *conf = mddev->private;
5560 int new_chunk = mddev->new_chunk_sectors;
5562 if (mddev->new_layout >= 0 && !algorithm_valid_raid5(mddev->new_layout))
5564 if (new_chunk > 0) {
5565 if (!is_power_of_2(new_chunk))
5567 if (new_chunk < (PAGE_SIZE>>9))
5569 if (mddev->array_sectors & (new_chunk-1))
5570 /* not factor of array size */
5574 /* They look valid */
5576 if (mddev->raid_disks == 2) {
5577 /* can make the change immediately */
5578 if (mddev->new_layout >= 0) {
5579 conf->algorithm = mddev->new_layout;
5580 mddev->layout = mddev->new_layout;
5582 if (new_chunk > 0) {
5583 conf->chunk_sectors = new_chunk ;
5584 mddev->chunk_sectors = new_chunk;
5586 set_bit(MD_CHANGE_DEVS, &mddev->flags);
5587 md_wakeup_thread(mddev->thread);
5589 return check_reshape(mddev);
5592 static int raid6_check_reshape(struct mddev *mddev)
5594 int new_chunk = mddev->new_chunk_sectors;
5596 if (mddev->new_layout >= 0 && !algorithm_valid_raid6(mddev->new_layout))
5598 if (new_chunk > 0) {
5599 if (!is_power_of_2(new_chunk))
5601 if (new_chunk < (PAGE_SIZE >> 9))
5603 if (mddev->array_sectors & (new_chunk-1))
5604 /* not factor of array size */
5608 /* They look valid */
5609 return check_reshape(mddev);
5612 static void *raid5_takeover(struct mddev *mddev)
5614 /* raid5 can take over:
5615 * raid0 - if there is only one strip zone - make it a raid4 layout
5616 * raid1 - if there are two drives. We need to know the chunk size
5617 * raid4 - trivial - just use a raid4 layout.
5618 * raid6 - Providing it is a *_6 layout
5620 if (mddev->level == 0)
5621 return raid45_takeover_raid0(mddev, 5);
5622 if (mddev->level == 1)
5623 return raid5_takeover_raid1(mddev);
5624 if (mddev->level == 4) {
5625 mddev->new_layout = ALGORITHM_PARITY_N;
5626 mddev->new_level = 5;
5627 return setup_conf(mddev);
5629 if (mddev->level == 6)
5630 return raid5_takeover_raid6(mddev);
5632 return ERR_PTR(-EINVAL);
5635 static void *raid4_takeover(struct mddev *mddev)
5637 /* raid4 can take over:
5638 * raid0 - if there is only one strip zone
5639 * raid5 - if layout is right
5641 if (mddev->level == 0)
5642 return raid45_takeover_raid0(mddev, 4);
5643 if (mddev->level == 5 &&
5644 mddev->layout == ALGORITHM_PARITY_N) {
5645 mddev->new_layout = 0;
5646 mddev->new_level = 4;
5647 return setup_conf(mddev);
5649 return ERR_PTR(-EINVAL);
5652 static struct md_personality raid5_personality;
5654 static void *raid6_takeover(struct mddev *mddev)
5656 /* Currently can only take over a raid5. We map the
5657 * personality to an equivalent raid6 personality
5658 * with the Q block at the end.
5662 if (mddev->pers != &raid5_personality)
5663 return ERR_PTR(-EINVAL);
5664 if (mddev->degraded > 1)
5665 return ERR_PTR(-EINVAL);
5666 if (mddev->raid_disks > 253)
5667 return ERR_PTR(-EINVAL);
5668 if (mddev->raid_disks < 3)
5669 return ERR_PTR(-EINVAL);
5671 switch (mddev->layout) {
5672 case ALGORITHM_LEFT_ASYMMETRIC:
5673 new_layout = ALGORITHM_LEFT_ASYMMETRIC_6;
5675 case ALGORITHM_RIGHT_ASYMMETRIC:
5676 new_layout = ALGORITHM_RIGHT_ASYMMETRIC_6;
5678 case ALGORITHM_LEFT_SYMMETRIC:
5679 new_layout = ALGORITHM_LEFT_SYMMETRIC_6;
5681 case ALGORITHM_RIGHT_SYMMETRIC:
5682 new_layout = ALGORITHM_RIGHT_SYMMETRIC_6;
5684 case ALGORITHM_PARITY_0:
5685 new_layout = ALGORITHM_PARITY_0_6;
5687 case ALGORITHM_PARITY_N:
5688 new_layout = ALGORITHM_PARITY_N;
5691 return ERR_PTR(-EINVAL);
5693 mddev->new_level = 6;
5694 mddev->new_layout = new_layout;
5695 mddev->delta_disks = 1;
5696 mddev->raid_disks += 1;
5697 return setup_conf(mddev);
5701 static struct md_personality raid6_personality =
5705 .owner = THIS_MODULE,
5706 .make_request = make_request,
5710 .error_handler = error,
5711 .hot_add_disk = raid5_add_disk,
5712 .hot_remove_disk= raid5_remove_disk,
5713 .spare_active = raid5_spare_active,
5714 .sync_request = sync_request,
5715 .resize = raid5_resize,
5717 .check_reshape = raid6_check_reshape,
5718 .start_reshape = raid5_start_reshape,
5719 .finish_reshape = raid5_finish_reshape,
5720 .quiesce = raid5_quiesce,
5721 .takeover = raid6_takeover,
5723 static struct md_personality raid5_personality =
5727 .owner = THIS_MODULE,
5728 .make_request = make_request,
5732 .error_handler = error,
5733 .hot_add_disk = raid5_add_disk,
5734 .hot_remove_disk= raid5_remove_disk,
5735 .spare_active = raid5_spare_active,
5736 .sync_request = sync_request,
5737 .resize = raid5_resize,
5739 .check_reshape = raid5_check_reshape,
5740 .start_reshape = raid5_start_reshape,
5741 .finish_reshape = raid5_finish_reshape,
5742 .quiesce = raid5_quiesce,
5743 .takeover = raid5_takeover,
5746 static struct md_personality raid4_personality =
5750 .owner = THIS_MODULE,
5751 .make_request = make_request,
5755 .error_handler = error,
5756 .hot_add_disk = raid5_add_disk,
5757 .hot_remove_disk= raid5_remove_disk,
5758 .spare_active = raid5_spare_active,
5759 .sync_request = sync_request,
5760 .resize = raid5_resize,
5762 .check_reshape = raid5_check_reshape,
5763 .start_reshape = raid5_start_reshape,
5764 .finish_reshape = raid5_finish_reshape,
5765 .quiesce = raid5_quiesce,
5766 .takeover = raid4_takeover,
5769 static int __init raid5_init(void)
5771 register_md_personality(&raid6_personality);
5772 register_md_personality(&raid5_personality);
5773 register_md_personality(&raid4_personality);
5777 static void raid5_exit(void)
5779 unregister_md_personality(&raid6_personality);
5780 unregister_md_personality(&raid5_personality);
5781 unregister_md_personality(&raid4_personality);
5784 module_init(raid5_init);
5785 module_exit(raid5_exit);
5786 MODULE_LICENSE("GPL");
5787 MODULE_DESCRIPTION("RAID4/5/6 (striping with parity) personality for MD");
5788 MODULE_ALIAS("md-personality-4"); /* RAID5 */
5789 MODULE_ALIAS("md-raid5");
5790 MODULE_ALIAS("md-raid4");
5791 MODULE_ALIAS("md-level-5");
5792 MODULE_ALIAS("md-level-4");
5793 MODULE_ALIAS("md-personality-8"); /* RAID6 */
5794 MODULE_ALIAS("md-raid6");
5795 MODULE_ALIAS("md-level-6");
5797 /* This used to be two separate modules, they were: */
5798 MODULE_ALIAS("raid5");
5799 MODULE_ALIAS("raid6");