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 !test_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
201 list_add_tail(&sh->lru, &conf->delayed_list);
202 else if (test_bit(STRIPE_BIT_DELAY, &sh->state) &&
203 sh->bm_seq - conf->seq_write > 0)
204 list_add_tail(&sh->lru, &conf->bitmap_list);
206 clear_bit(STRIPE_DELAYED, &sh->state);
207 clear_bit(STRIPE_BIT_DELAY, &sh->state);
208 list_add_tail(&sh->lru, &conf->handle_list);
210 md_wakeup_thread(conf->mddev->thread);
212 BUG_ON(stripe_operations_active(sh));
213 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
214 atomic_dec(&conf->preread_active_stripes);
215 if (atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD)
216 md_wakeup_thread(conf->mddev->thread);
218 atomic_dec(&conf->active_stripes);
219 if (!test_bit(STRIPE_EXPANDING, &sh->state)) {
220 list_add_tail(&sh->lru, &conf->inactive_list);
221 wake_up(&conf->wait_for_stripe);
222 if (conf->retry_read_aligned)
223 md_wakeup_thread(conf->mddev->thread);
229 static void release_stripe(struct stripe_head *sh)
231 struct r5conf *conf = sh->raid_conf;
234 spin_lock_irqsave(&conf->device_lock, flags);
235 __release_stripe(conf, sh);
236 spin_unlock_irqrestore(&conf->device_lock, flags);
239 static inline void remove_hash(struct stripe_head *sh)
241 pr_debug("remove_hash(), stripe %llu\n",
242 (unsigned long long)sh->sector);
244 hlist_del_init(&sh->hash);
247 static inline void insert_hash(struct r5conf *conf, struct stripe_head *sh)
249 struct hlist_head *hp = stripe_hash(conf, sh->sector);
251 pr_debug("insert_hash(), stripe %llu\n",
252 (unsigned long long)sh->sector);
254 hlist_add_head(&sh->hash, hp);
258 /* find an idle stripe, make sure it is unhashed, and return it. */
259 static struct stripe_head *get_free_stripe(struct r5conf *conf)
261 struct stripe_head *sh = NULL;
262 struct list_head *first;
264 if (list_empty(&conf->inactive_list))
266 first = conf->inactive_list.next;
267 sh = list_entry(first, struct stripe_head, lru);
268 list_del_init(first);
270 atomic_inc(&conf->active_stripes);
275 static void shrink_buffers(struct stripe_head *sh)
279 int num = sh->raid_conf->pool_size;
281 for (i = 0; i < num ; i++) {
285 sh->dev[i].page = NULL;
290 static int grow_buffers(struct stripe_head *sh)
293 int num = sh->raid_conf->pool_size;
295 for (i = 0; i < num; i++) {
298 if (!(page = alloc_page(GFP_KERNEL))) {
301 sh->dev[i].page = page;
306 static void raid5_build_block(struct stripe_head *sh, int i, int previous);
307 static void stripe_set_idx(sector_t stripe, struct r5conf *conf, int previous,
308 struct stripe_head *sh);
310 static void init_stripe(struct stripe_head *sh, sector_t sector, int previous)
312 struct r5conf *conf = sh->raid_conf;
315 BUG_ON(atomic_read(&sh->count) != 0);
316 BUG_ON(test_bit(STRIPE_HANDLE, &sh->state));
317 BUG_ON(stripe_operations_active(sh));
319 pr_debug("init_stripe called, stripe %llu\n",
320 (unsigned long long)sh->sector);
324 sh->generation = conf->generation - previous;
325 sh->disks = previous ? conf->previous_raid_disks : conf->raid_disks;
327 stripe_set_idx(sector, conf, previous, sh);
331 for (i = sh->disks; i--; ) {
332 struct r5dev *dev = &sh->dev[i];
334 if (dev->toread || dev->read || dev->towrite || dev->written ||
335 test_bit(R5_LOCKED, &dev->flags)) {
336 printk(KERN_ERR "sector=%llx i=%d %p %p %p %p %d\n",
337 (unsigned long long)sh->sector, i, dev->toread,
338 dev->read, dev->towrite, dev->written,
339 test_bit(R5_LOCKED, &dev->flags));
343 raid5_build_block(sh, i, previous);
345 insert_hash(conf, sh);
348 static struct stripe_head *__find_stripe(struct r5conf *conf, sector_t sector,
351 struct stripe_head *sh;
352 struct hlist_node *hn;
354 pr_debug("__find_stripe, sector %llu\n", (unsigned long long)sector);
355 hlist_for_each_entry(sh, hn, stripe_hash(conf, sector), hash)
356 if (sh->sector == sector && sh->generation == generation)
358 pr_debug("__stripe %llu not in cache\n", (unsigned long long)sector);
363 * Need to check if array has failed when deciding whether to:
365 * - remove non-faulty devices
368 * This determination is simple when no reshape is happening.
369 * However if there is a reshape, we need to carefully check
370 * both the before and after sections.
371 * This is because some failed devices may only affect one
372 * of the two sections, and some non-in_sync devices may
373 * be insync in the section most affected by failed devices.
375 static int has_failed(struct r5conf *conf)
379 if (conf->mddev->reshape_position == MaxSector)
380 return conf->mddev->degraded > conf->max_degraded;
384 for (i = 0; i < conf->previous_raid_disks; i++) {
385 struct md_rdev *rdev = rcu_dereference(conf->disks[i].rdev);
386 if (!rdev || test_bit(Faulty, &rdev->flags))
388 else if (test_bit(In_sync, &rdev->flags))
391 /* not in-sync or faulty.
392 * If the reshape increases the number of devices,
393 * this is being recovered by the reshape, so
394 * this 'previous' section is not in_sync.
395 * If the number of devices is being reduced however,
396 * the device can only be part of the array if
397 * we are reverting a reshape, so this section will
400 if (conf->raid_disks >= conf->previous_raid_disks)
404 if (degraded > conf->max_degraded)
408 for (i = 0; i < conf->raid_disks; i++) {
409 struct md_rdev *rdev = rcu_dereference(conf->disks[i].rdev);
410 if (!rdev || test_bit(Faulty, &rdev->flags))
412 else if (test_bit(In_sync, &rdev->flags))
415 /* not in-sync or faulty.
416 * If reshape increases the number of devices, this
417 * section has already been recovered, else it
418 * almost certainly hasn't.
420 if (conf->raid_disks <= conf->previous_raid_disks)
424 if (degraded > conf->max_degraded)
429 static struct stripe_head *
430 get_active_stripe(struct r5conf *conf, sector_t sector,
431 int previous, int noblock, int noquiesce)
433 struct stripe_head *sh;
435 pr_debug("get_stripe, sector %llu\n", (unsigned long long)sector);
437 spin_lock_irq(&conf->device_lock);
440 wait_event_lock_irq(conf->wait_for_stripe,
441 conf->quiesce == 0 || noquiesce,
442 conf->device_lock, /* nothing */);
443 sh = __find_stripe(conf, sector, conf->generation - previous);
445 if (!conf->inactive_blocked)
446 sh = get_free_stripe(conf);
447 if (noblock && sh == NULL)
450 conf->inactive_blocked = 1;
451 wait_event_lock_irq(conf->wait_for_stripe,
452 !list_empty(&conf->inactive_list) &&
453 (atomic_read(&conf->active_stripes)
454 < (conf->max_nr_stripes *3/4)
455 || !conf->inactive_blocked),
458 conf->inactive_blocked = 0;
460 init_stripe(sh, sector, previous);
462 if (atomic_read(&sh->count)) {
463 BUG_ON(!list_empty(&sh->lru)
464 && !test_bit(STRIPE_EXPANDING, &sh->state));
466 if (!test_bit(STRIPE_HANDLE, &sh->state))
467 atomic_inc(&conf->active_stripes);
468 if (list_empty(&sh->lru) &&
469 !test_bit(STRIPE_EXPANDING, &sh->state))
471 list_del_init(&sh->lru);
474 } while (sh == NULL);
477 atomic_inc(&sh->count);
479 spin_unlock_irq(&conf->device_lock);
484 raid5_end_read_request(struct bio *bi, int error);
486 raid5_end_write_request(struct bio *bi, int error);
488 static void ops_run_io(struct stripe_head *sh, struct stripe_head_state *s)
490 struct r5conf *conf = sh->raid_conf;
491 int i, disks = sh->disks;
495 for (i = disks; i--; ) {
498 struct md_rdev *rdev;
499 if (test_and_clear_bit(R5_Wantwrite, &sh->dev[i].flags)) {
500 if (test_and_clear_bit(R5_WantFUA, &sh->dev[i].flags))
504 } else if (test_and_clear_bit(R5_Wantread, &sh->dev[i].flags))
509 bi = &sh->dev[i].req;
513 bi->bi_end_io = raid5_end_write_request;
515 bi->bi_end_io = raid5_end_read_request;
518 rdev = rcu_dereference(conf->disks[i].rdev);
519 if (rdev && test_bit(Faulty, &rdev->flags))
522 atomic_inc(&rdev->nr_pending);
525 /* We have already checked bad blocks for reads. Now
526 * need to check for writes.
528 while ((rw & WRITE) && rdev &&
529 test_bit(WriteErrorSeen, &rdev->flags)) {
532 int bad = is_badblock(rdev, sh->sector, STRIPE_SECTORS,
533 &first_bad, &bad_sectors);
538 set_bit(BlockedBadBlocks, &rdev->flags);
539 if (!conf->mddev->external &&
540 conf->mddev->flags) {
541 /* It is very unlikely, but we might
542 * still need to write out the
543 * bad block log - better give it
545 md_check_recovery(conf->mddev);
548 * Because md_wait_for_blocked_rdev
549 * will dec nr_pending, we must
550 * increment it first.
552 atomic_inc(&rdev->nr_pending);
553 md_wait_for_blocked_rdev(rdev, conf->mddev);
555 /* Acknowledged bad block - skip the write */
556 rdev_dec_pending(rdev, conf->mddev);
562 if (s->syncing || s->expanding || s->expanded)
563 md_sync_acct(rdev->bdev, STRIPE_SECTORS);
565 set_bit(STRIPE_IO_STARTED, &sh->state);
567 bi->bi_bdev = rdev->bdev;
568 pr_debug("%s: for %llu schedule op %ld on disc %d\n",
569 __func__, (unsigned long long)sh->sector,
571 atomic_inc(&sh->count);
572 bi->bi_sector = sh->sector + rdev->data_offset;
573 bi->bi_flags = 1 << BIO_UPTODATE;
577 bi->bi_io_vec = &sh->dev[i].vec;
578 bi->bi_io_vec[0].bv_len = STRIPE_SIZE;
579 bi->bi_io_vec[0].bv_offset = 0;
580 bi->bi_size = STRIPE_SIZE;
582 generic_make_request(bi);
585 set_bit(STRIPE_DEGRADED, &sh->state);
586 pr_debug("skip op %ld on disc %d for sector %llu\n",
587 bi->bi_rw, i, (unsigned long long)sh->sector);
588 clear_bit(R5_LOCKED, &sh->dev[i].flags);
589 set_bit(STRIPE_HANDLE, &sh->state);
594 static struct dma_async_tx_descriptor *
595 async_copy_data(int frombio, struct bio *bio, struct page *page,
596 sector_t sector, struct dma_async_tx_descriptor *tx)
599 struct page *bio_page;
602 struct async_submit_ctl submit;
603 enum async_tx_flags flags = 0;
605 if (bio->bi_sector >= sector)
606 page_offset = (signed)(bio->bi_sector - sector) * 512;
608 page_offset = (signed)(sector - bio->bi_sector) * -512;
611 flags |= ASYNC_TX_FENCE;
612 init_async_submit(&submit, flags, tx, NULL, NULL, NULL);
614 bio_for_each_segment(bvl, bio, i) {
615 int len = bvl->bv_len;
619 if (page_offset < 0) {
620 b_offset = -page_offset;
621 page_offset += b_offset;
625 if (len > 0 && page_offset + len > STRIPE_SIZE)
626 clen = STRIPE_SIZE - page_offset;
631 b_offset += bvl->bv_offset;
632 bio_page = bvl->bv_page;
634 tx = async_memcpy(page, bio_page, page_offset,
635 b_offset, clen, &submit);
637 tx = async_memcpy(bio_page, page, b_offset,
638 page_offset, clen, &submit);
640 /* chain the operations */
641 submit.depend_tx = tx;
643 if (clen < len) /* hit end of page */
651 static void ops_complete_biofill(void *stripe_head_ref)
653 struct stripe_head *sh = stripe_head_ref;
654 struct bio *return_bi = NULL;
655 struct r5conf *conf = sh->raid_conf;
658 pr_debug("%s: stripe %llu\n", __func__,
659 (unsigned long long)sh->sector);
661 /* clear completed biofills */
662 spin_lock_irq(&conf->device_lock);
663 for (i = sh->disks; i--; ) {
664 struct r5dev *dev = &sh->dev[i];
666 /* acknowledge completion of a biofill operation */
667 /* and check if we need to reply to a read request,
668 * new R5_Wantfill requests are held off until
669 * !STRIPE_BIOFILL_RUN
671 if (test_and_clear_bit(R5_Wantfill, &dev->flags)) {
672 struct bio *rbi, *rbi2;
677 while (rbi && rbi->bi_sector <
678 dev->sector + STRIPE_SECTORS) {
679 rbi2 = r5_next_bio(rbi, dev->sector);
680 if (!raid5_dec_bi_phys_segments(rbi)) {
681 rbi->bi_next = return_bi;
688 spin_unlock_irq(&conf->device_lock);
689 clear_bit(STRIPE_BIOFILL_RUN, &sh->state);
691 return_io(return_bi);
693 set_bit(STRIPE_HANDLE, &sh->state);
697 static void ops_run_biofill(struct stripe_head *sh)
699 struct dma_async_tx_descriptor *tx = NULL;
700 struct r5conf *conf = sh->raid_conf;
701 struct async_submit_ctl submit;
704 pr_debug("%s: stripe %llu\n", __func__,
705 (unsigned long long)sh->sector);
707 for (i = sh->disks; i--; ) {
708 struct r5dev *dev = &sh->dev[i];
709 if (test_bit(R5_Wantfill, &dev->flags)) {
711 spin_lock_irq(&conf->device_lock);
712 dev->read = rbi = dev->toread;
714 spin_unlock_irq(&conf->device_lock);
715 while (rbi && rbi->bi_sector <
716 dev->sector + STRIPE_SECTORS) {
717 tx = async_copy_data(0, rbi, dev->page,
719 rbi = r5_next_bio(rbi, dev->sector);
724 atomic_inc(&sh->count);
725 init_async_submit(&submit, ASYNC_TX_ACK, tx, ops_complete_biofill, sh, NULL);
726 async_trigger_callback(&submit);
729 static void mark_target_uptodate(struct stripe_head *sh, int target)
736 tgt = &sh->dev[target];
737 set_bit(R5_UPTODATE, &tgt->flags);
738 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
739 clear_bit(R5_Wantcompute, &tgt->flags);
742 static void ops_complete_compute(void *stripe_head_ref)
744 struct stripe_head *sh = stripe_head_ref;
746 pr_debug("%s: stripe %llu\n", __func__,
747 (unsigned long long)sh->sector);
749 /* mark the computed target(s) as uptodate */
750 mark_target_uptodate(sh, sh->ops.target);
751 mark_target_uptodate(sh, sh->ops.target2);
753 clear_bit(STRIPE_COMPUTE_RUN, &sh->state);
754 if (sh->check_state == check_state_compute_run)
755 sh->check_state = check_state_compute_result;
756 set_bit(STRIPE_HANDLE, &sh->state);
760 /* return a pointer to the address conversion region of the scribble buffer */
761 static addr_conv_t *to_addr_conv(struct stripe_head *sh,
762 struct raid5_percpu *percpu)
764 return percpu->scribble + sizeof(struct page *) * (sh->disks + 2);
767 static struct dma_async_tx_descriptor *
768 ops_run_compute5(struct stripe_head *sh, struct raid5_percpu *percpu)
770 int disks = sh->disks;
771 struct page **xor_srcs = percpu->scribble;
772 int target = sh->ops.target;
773 struct r5dev *tgt = &sh->dev[target];
774 struct page *xor_dest = tgt->page;
776 struct dma_async_tx_descriptor *tx;
777 struct async_submit_ctl submit;
780 pr_debug("%s: stripe %llu block: %d\n",
781 __func__, (unsigned long long)sh->sector, target);
782 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
784 for (i = disks; i--; )
786 xor_srcs[count++] = sh->dev[i].page;
788 atomic_inc(&sh->count);
790 init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST, NULL,
791 ops_complete_compute, sh, to_addr_conv(sh, percpu));
792 if (unlikely(count == 1))
793 tx = async_memcpy(xor_dest, xor_srcs[0], 0, 0, STRIPE_SIZE, &submit);
795 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
800 /* set_syndrome_sources - populate source buffers for gen_syndrome
801 * @srcs - (struct page *) array of size sh->disks
802 * @sh - stripe_head to parse
804 * Populates srcs in proper layout order for the stripe and returns the
805 * 'count' of sources to be used in a call to async_gen_syndrome. The P
806 * destination buffer is recorded in srcs[count] and the Q destination
807 * is recorded in srcs[count+1]].
809 static int set_syndrome_sources(struct page **srcs, struct stripe_head *sh)
811 int disks = sh->disks;
812 int syndrome_disks = sh->ddf_layout ? disks : (disks - 2);
813 int d0_idx = raid6_d0(sh);
817 for (i = 0; i < disks; i++)
823 int slot = raid6_idx_to_slot(i, sh, &count, syndrome_disks);
825 srcs[slot] = sh->dev[i].page;
826 i = raid6_next_disk(i, disks);
827 } while (i != d0_idx);
829 return syndrome_disks;
832 static struct dma_async_tx_descriptor *
833 ops_run_compute6_1(struct stripe_head *sh, struct raid5_percpu *percpu)
835 int disks = sh->disks;
836 struct page **blocks = percpu->scribble;
838 int qd_idx = sh->qd_idx;
839 struct dma_async_tx_descriptor *tx;
840 struct async_submit_ctl submit;
846 if (sh->ops.target < 0)
847 target = sh->ops.target2;
848 else if (sh->ops.target2 < 0)
849 target = sh->ops.target;
851 /* we should only have one valid target */
854 pr_debug("%s: stripe %llu block: %d\n",
855 __func__, (unsigned long long)sh->sector, target);
857 tgt = &sh->dev[target];
858 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
861 atomic_inc(&sh->count);
863 if (target == qd_idx) {
864 count = set_syndrome_sources(blocks, sh);
865 blocks[count] = NULL; /* regenerating p is not necessary */
866 BUG_ON(blocks[count+1] != dest); /* q should already be set */
867 init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
868 ops_complete_compute, sh,
869 to_addr_conv(sh, percpu));
870 tx = async_gen_syndrome(blocks, 0, count+2, STRIPE_SIZE, &submit);
872 /* Compute any data- or p-drive using XOR */
874 for (i = disks; i-- ; ) {
875 if (i == target || i == qd_idx)
877 blocks[count++] = sh->dev[i].page;
880 init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST,
881 NULL, ops_complete_compute, sh,
882 to_addr_conv(sh, percpu));
883 tx = async_xor(dest, blocks, 0, count, STRIPE_SIZE, &submit);
889 static struct dma_async_tx_descriptor *
890 ops_run_compute6_2(struct stripe_head *sh, struct raid5_percpu *percpu)
892 int i, count, disks = sh->disks;
893 int syndrome_disks = sh->ddf_layout ? disks : disks-2;
894 int d0_idx = raid6_d0(sh);
895 int faila = -1, failb = -1;
896 int target = sh->ops.target;
897 int target2 = sh->ops.target2;
898 struct r5dev *tgt = &sh->dev[target];
899 struct r5dev *tgt2 = &sh->dev[target2];
900 struct dma_async_tx_descriptor *tx;
901 struct page **blocks = percpu->scribble;
902 struct async_submit_ctl submit;
904 pr_debug("%s: stripe %llu block1: %d block2: %d\n",
905 __func__, (unsigned long long)sh->sector, target, target2);
906 BUG_ON(target < 0 || target2 < 0);
907 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
908 BUG_ON(!test_bit(R5_Wantcompute, &tgt2->flags));
910 /* we need to open-code set_syndrome_sources to handle the
911 * slot number conversion for 'faila' and 'failb'
913 for (i = 0; i < disks ; i++)
918 int slot = raid6_idx_to_slot(i, sh, &count, syndrome_disks);
920 blocks[slot] = sh->dev[i].page;
926 i = raid6_next_disk(i, disks);
927 } while (i != d0_idx);
929 BUG_ON(faila == failb);
932 pr_debug("%s: stripe: %llu faila: %d failb: %d\n",
933 __func__, (unsigned long long)sh->sector, faila, failb);
935 atomic_inc(&sh->count);
937 if (failb == syndrome_disks+1) {
938 /* Q disk is one of the missing disks */
939 if (faila == syndrome_disks) {
940 /* Missing P+Q, just recompute */
941 init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
942 ops_complete_compute, sh,
943 to_addr_conv(sh, percpu));
944 return async_gen_syndrome(blocks, 0, syndrome_disks+2,
945 STRIPE_SIZE, &submit);
949 int qd_idx = sh->qd_idx;
951 /* Missing D+Q: recompute D from P, then recompute Q */
952 if (target == qd_idx)
953 data_target = target2;
955 data_target = target;
958 for (i = disks; i-- ; ) {
959 if (i == data_target || i == qd_idx)
961 blocks[count++] = sh->dev[i].page;
963 dest = sh->dev[data_target].page;
964 init_async_submit(&submit,
965 ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST,
967 to_addr_conv(sh, percpu));
968 tx = async_xor(dest, blocks, 0, count, STRIPE_SIZE,
971 count = set_syndrome_sources(blocks, sh);
972 init_async_submit(&submit, ASYNC_TX_FENCE, tx,
973 ops_complete_compute, sh,
974 to_addr_conv(sh, percpu));
975 return async_gen_syndrome(blocks, 0, count+2,
976 STRIPE_SIZE, &submit);
979 init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
980 ops_complete_compute, sh,
981 to_addr_conv(sh, percpu));
982 if (failb == syndrome_disks) {
983 /* We're missing D+P. */
984 return async_raid6_datap_recov(syndrome_disks+2,
988 /* We're missing D+D. */
989 return async_raid6_2data_recov(syndrome_disks+2,
990 STRIPE_SIZE, faila, failb,
997 static void ops_complete_prexor(void *stripe_head_ref)
999 struct stripe_head *sh = stripe_head_ref;
1001 pr_debug("%s: stripe %llu\n", __func__,
1002 (unsigned long long)sh->sector);
1005 static struct dma_async_tx_descriptor *
1006 ops_run_prexor(struct stripe_head *sh, struct raid5_percpu *percpu,
1007 struct dma_async_tx_descriptor *tx)
1009 int disks = sh->disks;
1010 struct page **xor_srcs = percpu->scribble;
1011 int count = 0, pd_idx = sh->pd_idx, i;
1012 struct async_submit_ctl submit;
1014 /* existing parity data subtracted */
1015 struct page *xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
1017 pr_debug("%s: stripe %llu\n", __func__,
1018 (unsigned long long)sh->sector);
1020 for (i = disks; i--; ) {
1021 struct r5dev *dev = &sh->dev[i];
1022 /* Only process blocks that are known to be uptodate */
1023 if (test_bit(R5_Wantdrain, &dev->flags))
1024 xor_srcs[count++] = dev->page;
1027 init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_DROP_DST, tx,
1028 ops_complete_prexor, sh, to_addr_conv(sh, percpu));
1029 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
1034 static struct dma_async_tx_descriptor *
1035 ops_run_biodrain(struct stripe_head *sh, struct dma_async_tx_descriptor *tx)
1037 int disks = sh->disks;
1040 pr_debug("%s: stripe %llu\n", __func__,
1041 (unsigned long long)sh->sector);
1043 for (i = disks; i--; ) {
1044 struct r5dev *dev = &sh->dev[i];
1047 if (test_and_clear_bit(R5_Wantdrain, &dev->flags)) {
1050 spin_lock_irq(&sh->raid_conf->device_lock);
1051 chosen = dev->towrite;
1052 dev->towrite = NULL;
1053 BUG_ON(dev->written);
1054 wbi = dev->written = chosen;
1055 spin_unlock_irq(&sh->raid_conf->device_lock);
1057 while (wbi && wbi->bi_sector <
1058 dev->sector + STRIPE_SECTORS) {
1059 if (wbi->bi_rw & REQ_FUA)
1060 set_bit(R5_WantFUA, &dev->flags);
1061 tx = async_copy_data(1, wbi, dev->page,
1063 wbi = r5_next_bio(wbi, dev->sector);
1071 static void ops_complete_reconstruct(void *stripe_head_ref)
1073 struct stripe_head *sh = stripe_head_ref;
1074 int disks = sh->disks;
1075 int pd_idx = sh->pd_idx;
1076 int qd_idx = sh->qd_idx;
1080 pr_debug("%s: stripe %llu\n", __func__,
1081 (unsigned long long)sh->sector);
1083 for (i = disks; i--; )
1084 fua |= test_bit(R5_WantFUA, &sh->dev[i].flags);
1086 for (i = disks; i--; ) {
1087 struct r5dev *dev = &sh->dev[i];
1089 if (dev->written || i == pd_idx || i == qd_idx) {
1090 set_bit(R5_UPTODATE, &dev->flags);
1092 set_bit(R5_WantFUA, &dev->flags);
1096 if (sh->reconstruct_state == reconstruct_state_drain_run)
1097 sh->reconstruct_state = reconstruct_state_drain_result;
1098 else if (sh->reconstruct_state == reconstruct_state_prexor_drain_run)
1099 sh->reconstruct_state = reconstruct_state_prexor_drain_result;
1101 BUG_ON(sh->reconstruct_state != reconstruct_state_run);
1102 sh->reconstruct_state = reconstruct_state_result;
1105 set_bit(STRIPE_HANDLE, &sh->state);
1110 ops_run_reconstruct5(struct stripe_head *sh, struct raid5_percpu *percpu,
1111 struct dma_async_tx_descriptor *tx)
1113 int disks = sh->disks;
1114 struct page **xor_srcs = percpu->scribble;
1115 struct async_submit_ctl submit;
1116 int count = 0, pd_idx = sh->pd_idx, i;
1117 struct page *xor_dest;
1119 unsigned long flags;
1121 pr_debug("%s: stripe %llu\n", __func__,
1122 (unsigned long long)sh->sector);
1124 /* check if prexor is active which means only process blocks
1125 * that are part of a read-modify-write (written)
1127 if (sh->reconstruct_state == reconstruct_state_prexor_drain_run) {
1129 xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
1130 for (i = disks; i--; ) {
1131 struct r5dev *dev = &sh->dev[i];
1133 xor_srcs[count++] = dev->page;
1136 xor_dest = sh->dev[pd_idx].page;
1137 for (i = disks; i--; ) {
1138 struct r5dev *dev = &sh->dev[i];
1140 xor_srcs[count++] = dev->page;
1144 /* 1/ if we prexor'd then the dest is reused as a source
1145 * 2/ if we did not prexor then we are redoing the parity
1146 * set ASYNC_TX_XOR_DROP_DST and ASYNC_TX_XOR_ZERO_DST
1147 * for the synchronous xor case
1149 flags = ASYNC_TX_ACK |
1150 (prexor ? ASYNC_TX_XOR_DROP_DST : ASYNC_TX_XOR_ZERO_DST);
1152 atomic_inc(&sh->count);
1154 init_async_submit(&submit, flags, tx, ops_complete_reconstruct, sh,
1155 to_addr_conv(sh, percpu));
1156 if (unlikely(count == 1))
1157 tx = async_memcpy(xor_dest, xor_srcs[0], 0, 0, STRIPE_SIZE, &submit);
1159 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
1163 ops_run_reconstruct6(struct stripe_head *sh, struct raid5_percpu *percpu,
1164 struct dma_async_tx_descriptor *tx)
1166 struct async_submit_ctl submit;
1167 struct page **blocks = percpu->scribble;
1170 pr_debug("%s: stripe %llu\n", __func__, (unsigned long long)sh->sector);
1172 count = set_syndrome_sources(blocks, sh);
1174 atomic_inc(&sh->count);
1176 init_async_submit(&submit, ASYNC_TX_ACK, tx, ops_complete_reconstruct,
1177 sh, to_addr_conv(sh, percpu));
1178 async_gen_syndrome(blocks, 0, count+2, STRIPE_SIZE, &submit);
1181 static void ops_complete_check(void *stripe_head_ref)
1183 struct stripe_head *sh = stripe_head_ref;
1185 pr_debug("%s: stripe %llu\n", __func__,
1186 (unsigned long long)sh->sector);
1188 sh->check_state = check_state_check_result;
1189 set_bit(STRIPE_HANDLE, &sh->state);
1193 static void ops_run_check_p(struct stripe_head *sh, struct raid5_percpu *percpu)
1195 int disks = sh->disks;
1196 int pd_idx = sh->pd_idx;
1197 int qd_idx = sh->qd_idx;
1198 struct page *xor_dest;
1199 struct page **xor_srcs = percpu->scribble;
1200 struct dma_async_tx_descriptor *tx;
1201 struct async_submit_ctl submit;
1205 pr_debug("%s: stripe %llu\n", __func__,
1206 (unsigned long long)sh->sector);
1209 xor_dest = sh->dev[pd_idx].page;
1210 xor_srcs[count++] = xor_dest;
1211 for (i = disks; i--; ) {
1212 if (i == pd_idx || i == qd_idx)
1214 xor_srcs[count++] = sh->dev[i].page;
1217 init_async_submit(&submit, 0, NULL, NULL, NULL,
1218 to_addr_conv(sh, percpu));
1219 tx = async_xor_val(xor_dest, xor_srcs, 0, count, STRIPE_SIZE,
1220 &sh->ops.zero_sum_result, &submit);
1222 atomic_inc(&sh->count);
1223 init_async_submit(&submit, ASYNC_TX_ACK, tx, ops_complete_check, sh, NULL);
1224 tx = async_trigger_callback(&submit);
1227 static void ops_run_check_pq(struct stripe_head *sh, struct raid5_percpu *percpu, int checkp)
1229 struct page **srcs = percpu->scribble;
1230 struct async_submit_ctl submit;
1233 pr_debug("%s: stripe %llu checkp: %d\n", __func__,
1234 (unsigned long long)sh->sector, checkp);
1236 count = set_syndrome_sources(srcs, sh);
1240 atomic_inc(&sh->count);
1241 init_async_submit(&submit, ASYNC_TX_ACK, NULL, ops_complete_check,
1242 sh, to_addr_conv(sh, percpu));
1243 async_syndrome_val(srcs, 0, count+2, STRIPE_SIZE,
1244 &sh->ops.zero_sum_result, percpu->spare_page, &submit);
1247 static void __raid_run_ops(struct stripe_head *sh, unsigned long ops_request)
1249 int overlap_clear = 0, i, disks = sh->disks;
1250 struct dma_async_tx_descriptor *tx = NULL;
1251 struct r5conf *conf = sh->raid_conf;
1252 int level = conf->level;
1253 struct raid5_percpu *percpu;
1257 percpu = per_cpu_ptr(conf->percpu, cpu);
1258 if (test_bit(STRIPE_OP_BIOFILL, &ops_request)) {
1259 ops_run_biofill(sh);
1263 if (test_bit(STRIPE_OP_COMPUTE_BLK, &ops_request)) {
1265 tx = ops_run_compute5(sh, percpu);
1267 if (sh->ops.target2 < 0 || sh->ops.target < 0)
1268 tx = ops_run_compute6_1(sh, percpu);
1270 tx = ops_run_compute6_2(sh, percpu);
1272 /* terminate the chain if reconstruct is not set to be run */
1273 if (tx && !test_bit(STRIPE_OP_RECONSTRUCT, &ops_request))
1277 if (test_bit(STRIPE_OP_PREXOR, &ops_request))
1278 tx = ops_run_prexor(sh, percpu, tx);
1280 if (test_bit(STRIPE_OP_BIODRAIN, &ops_request)) {
1281 tx = ops_run_biodrain(sh, tx);
1285 if (test_bit(STRIPE_OP_RECONSTRUCT, &ops_request)) {
1287 ops_run_reconstruct5(sh, percpu, tx);
1289 ops_run_reconstruct6(sh, percpu, tx);
1292 if (test_bit(STRIPE_OP_CHECK, &ops_request)) {
1293 if (sh->check_state == check_state_run)
1294 ops_run_check_p(sh, percpu);
1295 else if (sh->check_state == check_state_run_q)
1296 ops_run_check_pq(sh, percpu, 0);
1297 else if (sh->check_state == check_state_run_pq)
1298 ops_run_check_pq(sh, percpu, 1);
1304 for (i = disks; i--; ) {
1305 struct r5dev *dev = &sh->dev[i];
1306 if (test_and_clear_bit(R5_Overlap, &dev->flags))
1307 wake_up(&sh->raid_conf->wait_for_overlap);
1312 #ifdef CONFIG_MULTICORE_RAID456
1313 static void async_run_ops(void *param, async_cookie_t cookie)
1315 struct stripe_head *sh = param;
1316 unsigned long ops_request = sh->ops.request;
1318 clear_bit_unlock(STRIPE_OPS_REQ_PENDING, &sh->state);
1319 wake_up(&sh->ops.wait_for_ops);
1321 __raid_run_ops(sh, ops_request);
1325 static void raid_run_ops(struct stripe_head *sh, unsigned long ops_request)
1327 /* since handle_stripe can be called outside of raid5d context
1328 * we need to ensure sh->ops.request is de-staged before another
1331 wait_event(sh->ops.wait_for_ops,
1332 !test_and_set_bit_lock(STRIPE_OPS_REQ_PENDING, &sh->state));
1333 sh->ops.request = ops_request;
1335 atomic_inc(&sh->count);
1336 async_schedule(async_run_ops, sh);
1339 #define raid_run_ops __raid_run_ops
1342 static int grow_one_stripe(struct r5conf *conf)
1344 struct stripe_head *sh;
1345 sh = kmem_cache_zalloc(conf->slab_cache, GFP_KERNEL);
1349 sh->raid_conf = conf;
1350 #ifdef CONFIG_MULTICORE_RAID456
1351 init_waitqueue_head(&sh->ops.wait_for_ops);
1354 if (grow_buffers(sh)) {
1356 kmem_cache_free(conf->slab_cache, sh);
1359 /* we just created an active stripe so... */
1360 atomic_set(&sh->count, 1);
1361 atomic_inc(&conf->active_stripes);
1362 INIT_LIST_HEAD(&sh->lru);
1367 static int grow_stripes(struct r5conf *conf, int num)
1369 struct kmem_cache *sc;
1370 int devs = max(conf->raid_disks, conf->previous_raid_disks);
1372 if (conf->mddev->gendisk)
1373 sprintf(conf->cache_name[0],
1374 "raid%d-%s", conf->level, mdname(conf->mddev));
1376 sprintf(conf->cache_name[0],
1377 "raid%d-%p", conf->level, conf->mddev);
1378 sprintf(conf->cache_name[1], "%s-alt", conf->cache_name[0]);
1380 conf->active_name = 0;
1381 sc = kmem_cache_create(conf->cache_name[conf->active_name],
1382 sizeof(struct stripe_head)+(devs-1)*sizeof(struct r5dev),
1386 conf->slab_cache = sc;
1387 conf->pool_size = devs;
1389 if (!grow_one_stripe(conf))
1395 * scribble_len - return the required size of the scribble region
1396 * @num - total number of disks in the array
1398 * The size must be enough to contain:
1399 * 1/ a struct page pointer for each device in the array +2
1400 * 2/ room to convert each entry in (1) to its corresponding dma
1401 * (dma_map_page()) or page (page_address()) address.
1403 * Note: the +2 is for the destination buffers of the ddf/raid6 case where we
1404 * calculate over all devices (not just the data blocks), using zeros in place
1405 * of the P and Q blocks.
1407 static size_t scribble_len(int num)
1411 len = sizeof(struct page *) * (num+2) + sizeof(addr_conv_t) * (num+2);
1416 static int resize_stripes(struct r5conf *conf, int newsize)
1418 /* Make all the stripes able to hold 'newsize' devices.
1419 * New slots in each stripe get 'page' set to a new page.
1421 * This happens in stages:
1422 * 1/ create a new kmem_cache and allocate the required number of
1424 * 2/ gather all the old stripe_heads and tranfer the pages across
1425 * to the new stripe_heads. This will have the side effect of
1426 * freezing the array as once all stripe_heads have been collected,
1427 * no IO will be possible. Old stripe heads are freed once their
1428 * pages have been transferred over, and the old kmem_cache is
1429 * freed when all stripes are done.
1430 * 3/ reallocate conf->disks to be suitable bigger. If this fails,
1431 * we simple return a failre status - no need to clean anything up.
1432 * 4/ allocate new pages for the new slots in the new stripe_heads.
1433 * If this fails, we don't bother trying the shrink the
1434 * stripe_heads down again, we just leave them as they are.
1435 * As each stripe_head is processed the new one is released into
1438 * Once step2 is started, we cannot afford to wait for a write,
1439 * so we use GFP_NOIO allocations.
1441 struct stripe_head *osh, *nsh;
1442 LIST_HEAD(newstripes);
1443 struct disk_info *ndisks;
1446 struct kmem_cache *sc;
1449 if (newsize <= conf->pool_size)
1450 return 0; /* never bother to shrink */
1452 err = md_allow_write(conf->mddev);
1457 sc = kmem_cache_create(conf->cache_name[1-conf->active_name],
1458 sizeof(struct stripe_head)+(newsize-1)*sizeof(struct r5dev),
1463 for (i = conf->max_nr_stripes; i; i--) {
1464 nsh = kmem_cache_zalloc(sc, GFP_KERNEL);
1468 nsh->raid_conf = conf;
1469 #ifdef CONFIG_MULTICORE_RAID456
1470 init_waitqueue_head(&nsh->ops.wait_for_ops);
1473 list_add(&nsh->lru, &newstripes);
1476 /* didn't get enough, give up */
1477 while (!list_empty(&newstripes)) {
1478 nsh = list_entry(newstripes.next, struct stripe_head, lru);
1479 list_del(&nsh->lru);
1480 kmem_cache_free(sc, nsh);
1482 kmem_cache_destroy(sc);
1485 /* Step 2 - Must use GFP_NOIO now.
1486 * OK, we have enough stripes, start collecting inactive
1487 * stripes and copying them over
1489 list_for_each_entry(nsh, &newstripes, lru) {
1490 spin_lock_irq(&conf->device_lock);
1491 wait_event_lock_irq(conf->wait_for_stripe,
1492 !list_empty(&conf->inactive_list),
1495 osh = get_free_stripe(conf);
1496 spin_unlock_irq(&conf->device_lock);
1497 atomic_set(&nsh->count, 1);
1498 for(i=0; i<conf->pool_size; i++)
1499 nsh->dev[i].page = osh->dev[i].page;
1500 for( ; i<newsize; i++)
1501 nsh->dev[i].page = NULL;
1502 kmem_cache_free(conf->slab_cache, osh);
1504 kmem_cache_destroy(conf->slab_cache);
1507 * At this point, we are holding all the stripes so the array
1508 * is completely stalled, so now is a good time to resize
1509 * conf->disks and the scribble region
1511 ndisks = kzalloc(newsize * sizeof(struct disk_info), GFP_NOIO);
1513 for (i=0; i<conf->raid_disks; i++)
1514 ndisks[i] = conf->disks[i];
1516 conf->disks = ndisks;
1521 conf->scribble_len = scribble_len(newsize);
1522 for_each_present_cpu(cpu) {
1523 struct raid5_percpu *percpu;
1526 percpu = per_cpu_ptr(conf->percpu, cpu);
1527 scribble = kmalloc(conf->scribble_len, GFP_NOIO);
1530 kfree(percpu->scribble);
1531 percpu->scribble = scribble;
1539 /* Step 4, return new stripes to service */
1540 while(!list_empty(&newstripes)) {
1541 nsh = list_entry(newstripes.next, struct stripe_head, lru);
1542 list_del_init(&nsh->lru);
1544 for (i=conf->raid_disks; i < newsize; i++)
1545 if (nsh->dev[i].page == NULL) {
1546 struct page *p = alloc_page(GFP_NOIO);
1547 nsh->dev[i].page = p;
1551 release_stripe(nsh);
1553 /* critical section pass, GFP_NOIO no longer needed */
1555 conf->slab_cache = sc;
1556 conf->active_name = 1-conf->active_name;
1557 conf->pool_size = newsize;
1561 static int drop_one_stripe(struct r5conf *conf)
1563 struct stripe_head *sh;
1565 spin_lock_irq(&conf->device_lock);
1566 sh = get_free_stripe(conf);
1567 spin_unlock_irq(&conf->device_lock);
1570 BUG_ON(atomic_read(&sh->count));
1572 kmem_cache_free(conf->slab_cache, sh);
1573 atomic_dec(&conf->active_stripes);
1577 static void shrink_stripes(struct r5conf *conf)
1579 while (drop_one_stripe(conf))
1582 if (conf->slab_cache)
1583 kmem_cache_destroy(conf->slab_cache);
1584 conf->slab_cache = NULL;
1587 static void raid5_end_read_request(struct bio * bi, int error)
1589 struct stripe_head *sh = bi->bi_private;
1590 struct r5conf *conf = sh->raid_conf;
1591 int disks = sh->disks, i;
1592 int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
1593 char b[BDEVNAME_SIZE];
1594 struct md_rdev *rdev;
1597 for (i=0 ; i<disks; i++)
1598 if (bi == &sh->dev[i].req)
1601 pr_debug("end_read_request %llu/%d, count: %d, uptodate %d.\n",
1602 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
1610 set_bit(R5_UPTODATE, &sh->dev[i].flags);
1611 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
1612 rdev = conf->disks[i].rdev;
1615 "md/raid:%s: read error corrected"
1616 " (%lu sectors at %llu on %s)\n",
1617 mdname(conf->mddev), STRIPE_SECTORS,
1618 (unsigned long long)(sh->sector
1619 + rdev->data_offset),
1620 bdevname(rdev->bdev, b));
1621 atomic_add(STRIPE_SECTORS, &rdev->corrected_errors);
1622 clear_bit(R5_ReadError, &sh->dev[i].flags);
1623 clear_bit(R5_ReWrite, &sh->dev[i].flags);
1625 if (atomic_read(&conf->disks[i].rdev->read_errors))
1626 atomic_set(&conf->disks[i].rdev->read_errors, 0);
1628 const char *bdn = bdevname(conf->disks[i].rdev->bdev, b);
1630 rdev = conf->disks[i].rdev;
1632 clear_bit(R5_UPTODATE, &sh->dev[i].flags);
1633 atomic_inc(&rdev->read_errors);
1634 if (conf->mddev->degraded >= conf->max_degraded)
1637 "md/raid:%s: read error not correctable "
1638 "(sector %llu on %s).\n",
1639 mdname(conf->mddev),
1640 (unsigned long long)(sh->sector
1641 + rdev->data_offset),
1643 else if (test_bit(R5_ReWrite, &sh->dev[i].flags))
1647 "md/raid:%s: read error NOT corrected!! "
1648 "(sector %llu on %s).\n",
1649 mdname(conf->mddev),
1650 (unsigned long long)(sh->sector
1651 + rdev->data_offset),
1653 else if (atomic_read(&rdev->read_errors)
1654 > conf->max_nr_stripes)
1656 "md/raid:%s: Too many read errors, failing device %s.\n",
1657 mdname(conf->mddev), bdn);
1661 set_bit(R5_ReadError, &sh->dev[i].flags);
1663 clear_bit(R5_ReadError, &sh->dev[i].flags);
1664 clear_bit(R5_ReWrite, &sh->dev[i].flags);
1665 md_error(conf->mddev, rdev);
1668 rdev_dec_pending(conf->disks[i].rdev, conf->mddev);
1669 clear_bit(R5_LOCKED, &sh->dev[i].flags);
1670 set_bit(STRIPE_HANDLE, &sh->state);
1674 static void raid5_end_write_request(struct bio *bi, int error)
1676 struct stripe_head *sh = bi->bi_private;
1677 struct r5conf *conf = sh->raid_conf;
1678 int disks = sh->disks, i;
1679 int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
1683 for (i=0 ; i<disks; i++)
1684 if (bi == &sh->dev[i].req)
1687 pr_debug("end_write_request %llu/%d, count %d, uptodate: %d.\n",
1688 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
1696 set_bit(STRIPE_DEGRADED, &sh->state);
1697 set_bit(WriteErrorSeen, &conf->disks[i].rdev->flags);
1698 set_bit(R5_WriteError, &sh->dev[i].flags);
1699 } else if (is_badblock(conf->disks[i].rdev, sh->sector, STRIPE_SECTORS,
1700 &first_bad, &bad_sectors))
1701 set_bit(R5_MadeGood, &sh->dev[i].flags);
1703 rdev_dec_pending(conf->disks[i].rdev, conf->mddev);
1705 clear_bit(R5_LOCKED, &sh->dev[i].flags);
1706 set_bit(STRIPE_HANDLE, &sh->state);
1711 static sector_t compute_blocknr(struct stripe_head *sh, int i, int previous);
1713 static void raid5_build_block(struct stripe_head *sh, int i, int previous)
1715 struct r5dev *dev = &sh->dev[i];
1717 bio_init(&dev->req);
1718 dev->req.bi_io_vec = &dev->vec;
1720 dev->req.bi_max_vecs++;
1721 dev->vec.bv_page = dev->page;
1722 dev->vec.bv_len = STRIPE_SIZE;
1723 dev->vec.bv_offset = 0;
1725 dev->req.bi_sector = sh->sector;
1726 dev->req.bi_private = sh;
1729 dev->sector = compute_blocknr(sh, i, previous);
1732 static void error(struct mddev *mddev, struct md_rdev *rdev)
1734 char b[BDEVNAME_SIZE];
1735 struct r5conf *conf = mddev->private;
1736 pr_debug("raid456: error called\n");
1738 if (test_and_clear_bit(In_sync, &rdev->flags)) {
1739 unsigned long flags;
1740 spin_lock_irqsave(&conf->device_lock, flags);
1742 spin_unlock_irqrestore(&conf->device_lock, flags);
1744 * if recovery was running, make sure it aborts.
1746 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1748 set_bit(Blocked, &rdev->flags);
1749 set_bit(Faulty, &rdev->flags);
1750 set_bit(MD_CHANGE_DEVS, &mddev->flags);
1752 "md/raid:%s: Disk failure on %s, disabling device.\n"
1753 "md/raid:%s: Operation continuing on %d devices.\n",
1755 bdevname(rdev->bdev, b),
1757 conf->raid_disks - mddev->degraded);
1761 * Input: a 'big' sector number,
1762 * Output: index of the data and parity disk, and the sector # in them.
1764 static sector_t raid5_compute_sector(struct r5conf *conf, sector_t r_sector,
1765 int previous, int *dd_idx,
1766 struct stripe_head *sh)
1768 sector_t stripe, stripe2;
1769 sector_t chunk_number;
1770 unsigned int chunk_offset;
1773 sector_t new_sector;
1774 int algorithm = previous ? conf->prev_algo
1776 int sectors_per_chunk = previous ? conf->prev_chunk_sectors
1777 : conf->chunk_sectors;
1778 int raid_disks = previous ? conf->previous_raid_disks
1780 int data_disks = raid_disks - conf->max_degraded;
1782 /* First compute the information on this sector */
1785 * Compute the chunk number and the sector offset inside the chunk
1787 chunk_offset = sector_div(r_sector, sectors_per_chunk);
1788 chunk_number = r_sector;
1791 * Compute the stripe number
1793 stripe = chunk_number;
1794 *dd_idx = sector_div(stripe, data_disks);
1797 * Select the parity disk based on the user selected algorithm.
1799 pd_idx = qd_idx = -1;
1800 switch(conf->level) {
1802 pd_idx = data_disks;
1805 switch (algorithm) {
1806 case ALGORITHM_LEFT_ASYMMETRIC:
1807 pd_idx = data_disks - sector_div(stripe2, raid_disks);
1808 if (*dd_idx >= pd_idx)
1811 case ALGORITHM_RIGHT_ASYMMETRIC:
1812 pd_idx = sector_div(stripe2, raid_disks);
1813 if (*dd_idx >= pd_idx)
1816 case ALGORITHM_LEFT_SYMMETRIC:
1817 pd_idx = data_disks - sector_div(stripe2, raid_disks);
1818 *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
1820 case ALGORITHM_RIGHT_SYMMETRIC:
1821 pd_idx = sector_div(stripe2, raid_disks);
1822 *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
1824 case ALGORITHM_PARITY_0:
1828 case ALGORITHM_PARITY_N:
1829 pd_idx = data_disks;
1837 switch (algorithm) {
1838 case ALGORITHM_LEFT_ASYMMETRIC:
1839 pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
1840 qd_idx = pd_idx + 1;
1841 if (pd_idx == raid_disks-1) {
1842 (*dd_idx)++; /* Q D D D P */
1844 } else if (*dd_idx >= pd_idx)
1845 (*dd_idx) += 2; /* D D P Q D */
1847 case ALGORITHM_RIGHT_ASYMMETRIC:
1848 pd_idx = sector_div(stripe2, raid_disks);
1849 qd_idx = pd_idx + 1;
1850 if (pd_idx == raid_disks-1) {
1851 (*dd_idx)++; /* Q D D D P */
1853 } else if (*dd_idx >= pd_idx)
1854 (*dd_idx) += 2; /* D D P Q D */
1856 case ALGORITHM_LEFT_SYMMETRIC:
1857 pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
1858 qd_idx = (pd_idx + 1) % raid_disks;
1859 *dd_idx = (pd_idx + 2 + *dd_idx) % raid_disks;
1861 case ALGORITHM_RIGHT_SYMMETRIC:
1862 pd_idx = sector_div(stripe2, raid_disks);
1863 qd_idx = (pd_idx + 1) % raid_disks;
1864 *dd_idx = (pd_idx + 2 + *dd_idx) % raid_disks;
1867 case ALGORITHM_PARITY_0:
1872 case ALGORITHM_PARITY_N:
1873 pd_idx = data_disks;
1874 qd_idx = data_disks + 1;
1877 case ALGORITHM_ROTATING_ZERO_RESTART:
1878 /* Exactly the same as RIGHT_ASYMMETRIC, but or
1879 * of blocks for computing Q is different.
1881 pd_idx = sector_div(stripe2, raid_disks);
1882 qd_idx = pd_idx + 1;
1883 if (pd_idx == raid_disks-1) {
1884 (*dd_idx)++; /* Q D D D P */
1886 } else if (*dd_idx >= pd_idx)
1887 (*dd_idx) += 2; /* D D P Q D */
1891 case ALGORITHM_ROTATING_N_RESTART:
1892 /* Same a left_asymmetric, by first stripe is
1893 * D D D P Q rather than
1897 pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
1898 qd_idx = pd_idx + 1;
1899 if (pd_idx == raid_disks-1) {
1900 (*dd_idx)++; /* Q D D D P */
1902 } else if (*dd_idx >= pd_idx)
1903 (*dd_idx) += 2; /* D D P Q D */
1907 case ALGORITHM_ROTATING_N_CONTINUE:
1908 /* Same as left_symmetric but Q is before P */
1909 pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
1910 qd_idx = (pd_idx + raid_disks - 1) % raid_disks;
1911 *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
1915 case ALGORITHM_LEFT_ASYMMETRIC_6:
1916 /* RAID5 left_asymmetric, with Q on last device */
1917 pd_idx = data_disks - sector_div(stripe2, raid_disks-1);
1918 if (*dd_idx >= pd_idx)
1920 qd_idx = raid_disks - 1;
1923 case ALGORITHM_RIGHT_ASYMMETRIC_6:
1924 pd_idx = sector_div(stripe2, raid_disks-1);
1925 if (*dd_idx >= pd_idx)
1927 qd_idx = raid_disks - 1;
1930 case ALGORITHM_LEFT_SYMMETRIC_6:
1931 pd_idx = data_disks - sector_div(stripe2, raid_disks-1);
1932 *dd_idx = (pd_idx + 1 + *dd_idx) % (raid_disks-1);
1933 qd_idx = raid_disks - 1;
1936 case ALGORITHM_RIGHT_SYMMETRIC_6:
1937 pd_idx = sector_div(stripe2, raid_disks-1);
1938 *dd_idx = (pd_idx + 1 + *dd_idx) % (raid_disks-1);
1939 qd_idx = raid_disks - 1;
1942 case ALGORITHM_PARITY_0_6:
1945 qd_idx = raid_disks - 1;
1955 sh->pd_idx = pd_idx;
1956 sh->qd_idx = qd_idx;
1957 sh->ddf_layout = ddf_layout;
1960 * Finally, compute the new sector number
1962 new_sector = (sector_t)stripe * sectors_per_chunk + chunk_offset;
1967 static sector_t compute_blocknr(struct stripe_head *sh, int i, int previous)
1969 struct r5conf *conf = sh->raid_conf;
1970 int raid_disks = sh->disks;
1971 int data_disks = raid_disks - conf->max_degraded;
1972 sector_t new_sector = sh->sector, check;
1973 int sectors_per_chunk = previous ? conf->prev_chunk_sectors
1974 : conf->chunk_sectors;
1975 int algorithm = previous ? conf->prev_algo
1979 sector_t chunk_number;
1980 int dummy1, dd_idx = i;
1982 struct stripe_head sh2;
1985 chunk_offset = sector_div(new_sector, sectors_per_chunk);
1986 stripe = new_sector;
1988 if (i == sh->pd_idx)
1990 switch(conf->level) {
1993 switch (algorithm) {
1994 case ALGORITHM_LEFT_ASYMMETRIC:
1995 case ALGORITHM_RIGHT_ASYMMETRIC:
1999 case ALGORITHM_LEFT_SYMMETRIC:
2000 case ALGORITHM_RIGHT_SYMMETRIC:
2003 i -= (sh->pd_idx + 1);
2005 case ALGORITHM_PARITY_0:
2008 case ALGORITHM_PARITY_N:
2015 if (i == sh->qd_idx)
2016 return 0; /* It is the Q disk */
2017 switch (algorithm) {
2018 case ALGORITHM_LEFT_ASYMMETRIC:
2019 case ALGORITHM_RIGHT_ASYMMETRIC:
2020 case ALGORITHM_ROTATING_ZERO_RESTART:
2021 case ALGORITHM_ROTATING_N_RESTART:
2022 if (sh->pd_idx == raid_disks-1)
2023 i--; /* Q D D D P */
2024 else if (i > sh->pd_idx)
2025 i -= 2; /* D D P Q D */
2027 case ALGORITHM_LEFT_SYMMETRIC:
2028 case ALGORITHM_RIGHT_SYMMETRIC:
2029 if (sh->pd_idx == raid_disks-1)
2030 i--; /* Q D D D P */
2035 i -= (sh->pd_idx + 2);
2038 case ALGORITHM_PARITY_0:
2041 case ALGORITHM_PARITY_N:
2043 case ALGORITHM_ROTATING_N_CONTINUE:
2044 /* Like left_symmetric, but P is before Q */
2045 if (sh->pd_idx == 0)
2046 i--; /* P D D D Q */
2051 i -= (sh->pd_idx + 1);
2054 case ALGORITHM_LEFT_ASYMMETRIC_6:
2055 case ALGORITHM_RIGHT_ASYMMETRIC_6:
2059 case ALGORITHM_LEFT_SYMMETRIC_6:
2060 case ALGORITHM_RIGHT_SYMMETRIC_6:
2062 i += data_disks + 1;
2063 i -= (sh->pd_idx + 1);
2065 case ALGORITHM_PARITY_0_6:
2074 chunk_number = stripe * data_disks + i;
2075 r_sector = chunk_number * sectors_per_chunk + chunk_offset;
2077 check = raid5_compute_sector(conf, r_sector,
2078 previous, &dummy1, &sh2);
2079 if (check != sh->sector || dummy1 != dd_idx || sh2.pd_idx != sh->pd_idx
2080 || sh2.qd_idx != sh->qd_idx) {
2081 printk(KERN_ERR "md/raid:%s: compute_blocknr: map not correct\n",
2082 mdname(conf->mddev));
2090 schedule_reconstruction(struct stripe_head *sh, struct stripe_head_state *s,
2091 int rcw, int expand)
2093 int i, pd_idx = sh->pd_idx, disks = sh->disks;
2094 struct r5conf *conf = sh->raid_conf;
2095 int level = conf->level;
2098 /* if we are not expanding this is a proper write request, and
2099 * there will be bios with new data to be drained into the
2103 sh->reconstruct_state = reconstruct_state_drain_run;
2104 set_bit(STRIPE_OP_BIODRAIN, &s->ops_request);
2106 sh->reconstruct_state = reconstruct_state_run;
2108 set_bit(STRIPE_OP_RECONSTRUCT, &s->ops_request);
2110 for (i = disks; i--; ) {
2111 struct r5dev *dev = &sh->dev[i];
2114 set_bit(R5_LOCKED, &dev->flags);
2115 set_bit(R5_Wantdrain, &dev->flags);
2117 clear_bit(R5_UPTODATE, &dev->flags);
2121 if (s->locked + conf->max_degraded == disks)
2122 if (!test_and_set_bit(STRIPE_FULL_WRITE, &sh->state))
2123 atomic_inc(&conf->pending_full_writes);
2126 BUG_ON(!(test_bit(R5_UPTODATE, &sh->dev[pd_idx].flags) ||
2127 test_bit(R5_Wantcompute, &sh->dev[pd_idx].flags)));
2129 sh->reconstruct_state = reconstruct_state_prexor_drain_run;
2130 set_bit(STRIPE_OP_PREXOR, &s->ops_request);
2131 set_bit(STRIPE_OP_BIODRAIN, &s->ops_request);
2132 set_bit(STRIPE_OP_RECONSTRUCT, &s->ops_request);
2134 for (i = disks; i--; ) {
2135 struct r5dev *dev = &sh->dev[i];
2140 (test_bit(R5_UPTODATE, &dev->flags) ||
2141 test_bit(R5_Wantcompute, &dev->flags))) {
2142 set_bit(R5_Wantdrain, &dev->flags);
2143 set_bit(R5_LOCKED, &dev->flags);
2144 clear_bit(R5_UPTODATE, &dev->flags);
2150 /* keep the parity disk(s) locked while asynchronous operations
2153 set_bit(R5_LOCKED, &sh->dev[pd_idx].flags);
2154 clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
2158 int qd_idx = sh->qd_idx;
2159 struct r5dev *dev = &sh->dev[qd_idx];
2161 set_bit(R5_LOCKED, &dev->flags);
2162 clear_bit(R5_UPTODATE, &dev->flags);
2166 pr_debug("%s: stripe %llu locked: %d ops_request: %lx\n",
2167 __func__, (unsigned long long)sh->sector,
2168 s->locked, s->ops_request);
2172 * Each stripe/dev can have one or more bion attached.
2173 * toread/towrite point to the first in a chain.
2174 * The bi_next chain must be in order.
2176 static int add_stripe_bio(struct stripe_head *sh, struct bio *bi, int dd_idx, int forwrite)
2179 struct r5conf *conf = sh->raid_conf;
2182 pr_debug("adding bi b#%llu to stripe s#%llu\n",
2183 (unsigned long long)bi->bi_sector,
2184 (unsigned long long)sh->sector);
2187 spin_lock_irq(&conf->device_lock);
2189 bip = &sh->dev[dd_idx].towrite;
2190 if (*bip == NULL && sh->dev[dd_idx].written == NULL)
2193 bip = &sh->dev[dd_idx].toread;
2194 while (*bip && (*bip)->bi_sector < bi->bi_sector) {
2195 if ((*bip)->bi_sector + ((*bip)->bi_size >> 9) > bi->bi_sector)
2197 bip = & (*bip)->bi_next;
2199 if (*bip && (*bip)->bi_sector < bi->bi_sector + ((bi->bi_size)>>9))
2202 BUG_ON(*bip && bi->bi_next && (*bip) != bi->bi_next);
2206 bi->bi_phys_segments++;
2209 /* check if page is covered */
2210 sector_t sector = sh->dev[dd_idx].sector;
2211 for (bi=sh->dev[dd_idx].towrite;
2212 sector < sh->dev[dd_idx].sector + STRIPE_SECTORS &&
2213 bi && bi->bi_sector <= sector;
2214 bi = r5_next_bio(bi, sh->dev[dd_idx].sector)) {
2215 if (bi->bi_sector + (bi->bi_size>>9) >= sector)
2216 sector = bi->bi_sector + (bi->bi_size>>9);
2218 if (sector >= sh->dev[dd_idx].sector + STRIPE_SECTORS)
2219 set_bit(R5_OVERWRITE, &sh->dev[dd_idx].flags);
2221 spin_unlock_irq(&conf->device_lock);
2223 pr_debug("added bi b#%llu to stripe s#%llu, disk %d.\n",
2224 (unsigned long long)(*bip)->bi_sector,
2225 (unsigned long long)sh->sector, dd_idx);
2227 if (conf->mddev->bitmap && firstwrite) {
2228 bitmap_startwrite(conf->mddev->bitmap, sh->sector,
2230 sh->bm_seq = conf->seq_flush+1;
2231 set_bit(STRIPE_BIT_DELAY, &sh->state);
2236 set_bit(R5_Overlap, &sh->dev[dd_idx].flags);
2237 spin_unlock_irq(&conf->device_lock);
2241 static void end_reshape(struct r5conf *conf);
2243 static void stripe_set_idx(sector_t stripe, struct r5conf *conf, int previous,
2244 struct stripe_head *sh)
2246 int sectors_per_chunk =
2247 previous ? conf->prev_chunk_sectors : conf->chunk_sectors;
2249 int chunk_offset = sector_div(stripe, sectors_per_chunk);
2250 int disks = previous ? conf->previous_raid_disks : conf->raid_disks;
2252 raid5_compute_sector(conf,
2253 stripe * (disks - conf->max_degraded)
2254 *sectors_per_chunk + chunk_offset,
2260 handle_failed_stripe(struct r5conf *conf, struct stripe_head *sh,
2261 struct stripe_head_state *s, int disks,
2262 struct bio **return_bi)
2265 for (i = disks; i--; ) {
2269 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
2270 struct md_rdev *rdev;
2272 rdev = rcu_dereference(conf->disks[i].rdev);
2273 if (rdev && test_bit(In_sync, &rdev->flags))
2274 atomic_inc(&rdev->nr_pending);
2279 if (!rdev_set_badblocks(
2283 md_error(conf->mddev, rdev);
2284 rdev_dec_pending(rdev, conf->mddev);
2287 spin_lock_irq(&conf->device_lock);
2288 /* fail all writes first */
2289 bi = sh->dev[i].towrite;
2290 sh->dev[i].towrite = NULL;
2296 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
2297 wake_up(&conf->wait_for_overlap);
2299 while (bi && bi->bi_sector <
2300 sh->dev[i].sector + STRIPE_SECTORS) {
2301 struct bio *nextbi = r5_next_bio(bi, sh->dev[i].sector);
2302 clear_bit(BIO_UPTODATE, &bi->bi_flags);
2303 if (!raid5_dec_bi_phys_segments(bi)) {
2304 md_write_end(conf->mddev);
2305 bi->bi_next = *return_bi;
2310 /* and fail all 'written' */
2311 bi = sh->dev[i].written;
2312 sh->dev[i].written = NULL;
2313 if (bi) bitmap_end = 1;
2314 while (bi && bi->bi_sector <
2315 sh->dev[i].sector + STRIPE_SECTORS) {
2316 struct bio *bi2 = r5_next_bio(bi, sh->dev[i].sector);
2317 clear_bit(BIO_UPTODATE, &bi->bi_flags);
2318 if (!raid5_dec_bi_phys_segments(bi)) {
2319 md_write_end(conf->mddev);
2320 bi->bi_next = *return_bi;
2326 /* fail any reads if this device is non-operational and
2327 * the data has not reached the cache yet.
2329 if (!test_bit(R5_Wantfill, &sh->dev[i].flags) &&
2330 (!test_bit(R5_Insync, &sh->dev[i].flags) ||
2331 test_bit(R5_ReadError, &sh->dev[i].flags))) {
2332 bi = sh->dev[i].toread;
2333 sh->dev[i].toread = NULL;
2334 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
2335 wake_up(&conf->wait_for_overlap);
2336 if (bi) s->to_read--;
2337 while (bi && bi->bi_sector <
2338 sh->dev[i].sector + STRIPE_SECTORS) {
2339 struct bio *nextbi =
2340 r5_next_bio(bi, sh->dev[i].sector);
2341 clear_bit(BIO_UPTODATE, &bi->bi_flags);
2342 if (!raid5_dec_bi_phys_segments(bi)) {
2343 bi->bi_next = *return_bi;
2349 spin_unlock_irq(&conf->device_lock);
2351 bitmap_endwrite(conf->mddev->bitmap, sh->sector,
2352 STRIPE_SECTORS, 0, 0);
2353 /* If we were in the middle of a write the parity block might
2354 * still be locked - so just clear all R5_LOCKED flags
2356 clear_bit(R5_LOCKED, &sh->dev[i].flags);
2359 if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
2360 if (atomic_dec_and_test(&conf->pending_full_writes))
2361 md_wakeup_thread(conf->mddev->thread);
2365 handle_failed_sync(struct r5conf *conf, struct stripe_head *sh,
2366 struct stripe_head_state *s)
2371 md_done_sync(conf->mddev, STRIPE_SECTORS, 0);
2372 clear_bit(STRIPE_SYNCING, &sh->state);
2374 /* There is nothing more to do for sync/check/repair.
2375 * For recover we need to record a bad block on all
2376 * non-sync devices, or abort the recovery
2378 if (!test_bit(MD_RECOVERY_RECOVER, &conf->mddev->recovery))
2380 /* During recovery devices cannot be removed, so locking and
2381 * refcounting of rdevs is not needed
2383 for (i = 0; i < conf->raid_disks; i++) {
2384 struct md_rdev *rdev = conf->disks[i].rdev;
2386 || test_bit(Faulty, &rdev->flags)
2387 || test_bit(In_sync, &rdev->flags))
2389 if (!rdev_set_badblocks(rdev, sh->sector,
2394 conf->recovery_disabled = conf->mddev->recovery_disabled;
2395 set_bit(MD_RECOVERY_INTR, &conf->mddev->recovery);
2399 /* fetch_block - checks the given member device to see if its data needs
2400 * to be read or computed to satisfy a request.
2402 * Returns 1 when no more member devices need to be checked, otherwise returns
2403 * 0 to tell the loop in handle_stripe_fill to continue
2405 static int fetch_block(struct stripe_head *sh, struct stripe_head_state *s,
2406 int disk_idx, int disks)
2408 struct r5dev *dev = &sh->dev[disk_idx];
2409 struct r5dev *fdev[2] = { &sh->dev[s->failed_num[0]],
2410 &sh->dev[s->failed_num[1]] };
2412 /* is the data in this block needed, and can we get it? */
2413 if (!test_bit(R5_LOCKED, &dev->flags) &&
2414 !test_bit(R5_UPTODATE, &dev->flags) &&
2416 (dev->towrite && !test_bit(R5_OVERWRITE, &dev->flags)) ||
2417 s->syncing || s->expanding ||
2418 (s->failed >= 1 && fdev[0]->toread) ||
2419 (s->failed >= 2 && fdev[1]->toread) ||
2420 (sh->raid_conf->level <= 5 && s->failed && fdev[0]->towrite &&
2421 !test_bit(R5_OVERWRITE, &fdev[0]->flags)) ||
2422 (sh->raid_conf->level == 6 && s->failed && s->to_write))) {
2423 /* we would like to get this block, possibly by computing it,
2424 * otherwise read it if the backing disk is insync
2426 BUG_ON(test_bit(R5_Wantcompute, &dev->flags));
2427 BUG_ON(test_bit(R5_Wantread, &dev->flags));
2428 if ((s->uptodate == disks - 1) &&
2429 (s->failed && (disk_idx == s->failed_num[0] ||
2430 disk_idx == s->failed_num[1]))) {
2431 /* have disk failed, and we're requested to fetch it;
2434 pr_debug("Computing stripe %llu block %d\n",
2435 (unsigned long long)sh->sector, disk_idx);
2436 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
2437 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
2438 set_bit(R5_Wantcompute, &dev->flags);
2439 sh->ops.target = disk_idx;
2440 sh->ops.target2 = -1; /* no 2nd target */
2442 /* Careful: from this point on 'uptodate' is in the eye
2443 * of raid_run_ops which services 'compute' operations
2444 * before writes. R5_Wantcompute flags a block that will
2445 * be R5_UPTODATE by the time it is needed for a
2446 * subsequent operation.
2450 } else if (s->uptodate == disks-2 && s->failed >= 2) {
2451 /* Computing 2-failure is *very* expensive; only
2452 * do it if failed >= 2
2455 for (other = disks; other--; ) {
2456 if (other == disk_idx)
2458 if (!test_bit(R5_UPTODATE,
2459 &sh->dev[other].flags))
2463 pr_debug("Computing stripe %llu blocks %d,%d\n",
2464 (unsigned long long)sh->sector,
2466 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
2467 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
2468 set_bit(R5_Wantcompute, &sh->dev[disk_idx].flags);
2469 set_bit(R5_Wantcompute, &sh->dev[other].flags);
2470 sh->ops.target = disk_idx;
2471 sh->ops.target2 = other;
2475 } else if (test_bit(R5_Insync, &dev->flags)) {
2476 set_bit(R5_LOCKED, &dev->flags);
2477 set_bit(R5_Wantread, &dev->flags);
2479 pr_debug("Reading block %d (sync=%d)\n",
2480 disk_idx, s->syncing);
2488 * handle_stripe_fill - read or compute data to satisfy pending requests.
2490 static void handle_stripe_fill(struct stripe_head *sh,
2491 struct stripe_head_state *s,
2496 /* look for blocks to read/compute, skip this if a compute
2497 * is already in flight, or if the stripe contents are in the
2498 * midst of changing due to a write
2500 if (!test_bit(STRIPE_COMPUTE_RUN, &sh->state) && !sh->check_state &&
2501 !sh->reconstruct_state)
2502 for (i = disks; i--; )
2503 if (fetch_block(sh, s, i, disks))
2505 set_bit(STRIPE_HANDLE, &sh->state);
2509 /* handle_stripe_clean_event
2510 * any written block on an uptodate or failed drive can be returned.
2511 * Note that if we 'wrote' to a failed drive, it will be UPTODATE, but
2512 * never LOCKED, so we don't need to test 'failed' directly.
2514 static void handle_stripe_clean_event(struct r5conf *conf,
2515 struct stripe_head *sh, int disks, struct bio **return_bi)
2520 for (i = disks; i--; )
2521 if (sh->dev[i].written) {
2523 if (!test_bit(R5_LOCKED, &dev->flags) &&
2524 test_bit(R5_UPTODATE, &dev->flags)) {
2525 /* We can return any write requests */
2526 struct bio *wbi, *wbi2;
2528 pr_debug("Return write for disc %d\n", i);
2529 spin_lock_irq(&conf->device_lock);
2531 dev->written = NULL;
2532 while (wbi && wbi->bi_sector <
2533 dev->sector + STRIPE_SECTORS) {
2534 wbi2 = r5_next_bio(wbi, dev->sector);
2535 if (!raid5_dec_bi_phys_segments(wbi)) {
2536 md_write_end(conf->mddev);
2537 wbi->bi_next = *return_bi;
2542 if (dev->towrite == NULL)
2544 spin_unlock_irq(&conf->device_lock);
2546 bitmap_endwrite(conf->mddev->bitmap,
2549 !test_bit(STRIPE_DEGRADED, &sh->state),
2554 if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
2555 if (atomic_dec_and_test(&conf->pending_full_writes))
2556 md_wakeup_thread(conf->mddev->thread);
2559 static void handle_stripe_dirtying(struct r5conf *conf,
2560 struct stripe_head *sh,
2561 struct stripe_head_state *s,
2564 int rmw = 0, rcw = 0, i;
2565 if (conf->max_degraded == 2) {
2566 /* RAID6 requires 'rcw' in current implementation
2567 * Calculate the real rcw later - for now fake it
2568 * look like rcw is cheaper
2571 } else for (i = disks; i--; ) {
2572 /* would I have to read this buffer for read_modify_write */
2573 struct r5dev *dev = &sh->dev[i];
2574 if ((dev->towrite || i == sh->pd_idx) &&
2575 !test_bit(R5_LOCKED, &dev->flags) &&
2576 !(test_bit(R5_UPTODATE, &dev->flags) ||
2577 test_bit(R5_Wantcompute, &dev->flags))) {
2578 if (test_bit(R5_Insync, &dev->flags))
2581 rmw += 2*disks; /* cannot read it */
2583 /* Would I have to read this buffer for reconstruct_write */
2584 if (!test_bit(R5_OVERWRITE, &dev->flags) && i != sh->pd_idx &&
2585 !test_bit(R5_LOCKED, &dev->flags) &&
2586 !(test_bit(R5_UPTODATE, &dev->flags) ||
2587 test_bit(R5_Wantcompute, &dev->flags))) {
2588 if (test_bit(R5_Insync, &dev->flags)) rcw++;
2593 pr_debug("for sector %llu, rmw=%d rcw=%d\n",
2594 (unsigned long long)sh->sector, rmw, rcw);
2595 set_bit(STRIPE_HANDLE, &sh->state);
2596 if (rmw < rcw && rmw > 0)
2597 /* prefer read-modify-write, but need to get some data */
2598 for (i = disks; i--; ) {
2599 struct r5dev *dev = &sh->dev[i];
2600 if ((dev->towrite || i == sh->pd_idx) &&
2601 !test_bit(R5_LOCKED, &dev->flags) &&
2602 !(test_bit(R5_UPTODATE, &dev->flags) ||
2603 test_bit(R5_Wantcompute, &dev->flags)) &&
2604 test_bit(R5_Insync, &dev->flags)) {
2606 test_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2607 pr_debug("Read_old block "
2608 "%d for r-m-w\n", i);
2609 set_bit(R5_LOCKED, &dev->flags);
2610 set_bit(R5_Wantread, &dev->flags);
2613 set_bit(STRIPE_DELAYED, &sh->state);
2614 set_bit(STRIPE_HANDLE, &sh->state);
2618 if (rcw <= rmw && rcw > 0) {
2619 /* want reconstruct write, but need to get some data */
2621 for (i = disks; i--; ) {
2622 struct r5dev *dev = &sh->dev[i];
2623 if (!test_bit(R5_OVERWRITE, &dev->flags) &&
2624 i != sh->pd_idx && i != sh->qd_idx &&
2625 !test_bit(R5_LOCKED, &dev->flags) &&
2626 !(test_bit(R5_UPTODATE, &dev->flags) ||
2627 test_bit(R5_Wantcompute, &dev->flags))) {
2629 if (!test_bit(R5_Insync, &dev->flags))
2630 continue; /* it's a failed drive */
2632 test_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2633 pr_debug("Read_old block "
2634 "%d for Reconstruct\n", i);
2635 set_bit(R5_LOCKED, &dev->flags);
2636 set_bit(R5_Wantread, &dev->flags);
2639 set_bit(STRIPE_DELAYED, &sh->state);
2640 set_bit(STRIPE_HANDLE, &sh->state);
2645 /* now if nothing is locked, and if we have enough data,
2646 * we can start a write request
2648 /* since handle_stripe can be called at any time we need to handle the
2649 * case where a compute block operation has been submitted and then a
2650 * subsequent call wants to start a write request. raid_run_ops only
2651 * handles the case where compute block and reconstruct are requested
2652 * simultaneously. If this is not the case then new writes need to be
2653 * held off until the compute completes.
2655 if ((s->req_compute || !test_bit(STRIPE_COMPUTE_RUN, &sh->state)) &&
2656 (s->locked == 0 && (rcw == 0 || rmw == 0) &&
2657 !test_bit(STRIPE_BIT_DELAY, &sh->state)))
2658 schedule_reconstruction(sh, s, rcw == 0, 0);
2661 static void handle_parity_checks5(struct r5conf *conf, struct stripe_head *sh,
2662 struct stripe_head_state *s, int disks)
2664 struct r5dev *dev = NULL;
2666 set_bit(STRIPE_HANDLE, &sh->state);
2668 switch (sh->check_state) {
2669 case check_state_idle:
2670 /* start a new check operation if there are no failures */
2671 if (s->failed == 0) {
2672 BUG_ON(s->uptodate != disks);
2673 sh->check_state = check_state_run;
2674 set_bit(STRIPE_OP_CHECK, &s->ops_request);
2675 clear_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags);
2679 dev = &sh->dev[s->failed_num[0]];
2681 case check_state_compute_result:
2682 sh->check_state = check_state_idle;
2684 dev = &sh->dev[sh->pd_idx];
2686 /* check that a write has not made the stripe insync */
2687 if (test_bit(STRIPE_INSYNC, &sh->state))
2690 /* either failed parity check, or recovery is happening */
2691 BUG_ON(!test_bit(R5_UPTODATE, &dev->flags));
2692 BUG_ON(s->uptodate != disks);
2694 set_bit(R5_LOCKED, &dev->flags);
2696 set_bit(R5_Wantwrite, &dev->flags);
2698 clear_bit(STRIPE_DEGRADED, &sh->state);
2699 set_bit(STRIPE_INSYNC, &sh->state);
2701 case check_state_run:
2702 break; /* we will be called again upon completion */
2703 case check_state_check_result:
2704 sh->check_state = check_state_idle;
2706 /* if a failure occurred during the check operation, leave
2707 * STRIPE_INSYNC not set and let the stripe be handled again
2712 /* handle a successful check operation, if parity is correct
2713 * we are done. Otherwise update the mismatch count and repair
2714 * parity if !MD_RECOVERY_CHECK
2716 if ((sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) == 0)
2717 /* parity is correct (on disc,
2718 * not in buffer any more)
2720 set_bit(STRIPE_INSYNC, &sh->state);
2722 conf->mddev->resync_mismatches += STRIPE_SECTORS;
2723 if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery))
2724 /* don't try to repair!! */
2725 set_bit(STRIPE_INSYNC, &sh->state);
2727 sh->check_state = check_state_compute_run;
2728 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
2729 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
2730 set_bit(R5_Wantcompute,
2731 &sh->dev[sh->pd_idx].flags);
2732 sh->ops.target = sh->pd_idx;
2733 sh->ops.target2 = -1;
2738 case check_state_compute_run:
2741 printk(KERN_ERR "%s: unknown check_state: %d sector: %llu\n",
2742 __func__, sh->check_state,
2743 (unsigned long long) sh->sector);
2749 static void handle_parity_checks6(struct r5conf *conf, struct stripe_head *sh,
2750 struct stripe_head_state *s,
2753 int pd_idx = sh->pd_idx;
2754 int qd_idx = sh->qd_idx;
2757 set_bit(STRIPE_HANDLE, &sh->state);
2759 BUG_ON(s->failed > 2);
2761 /* Want to check and possibly repair P and Q.
2762 * However there could be one 'failed' device, in which
2763 * case we can only check one of them, possibly using the
2764 * other to generate missing data
2767 switch (sh->check_state) {
2768 case check_state_idle:
2769 /* start a new check operation if there are < 2 failures */
2770 if (s->failed == s->q_failed) {
2771 /* The only possible failed device holds Q, so it
2772 * makes sense to check P (If anything else were failed,
2773 * we would have used P to recreate it).
2775 sh->check_state = check_state_run;
2777 if (!s->q_failed && s->failed < 2) {
2778 /* Q is not failed, and we didn't use it to generate
2779 * anything, so it makes sense to check it
2781 if (sh->check_state == check_state_run)
2782 sh->check_state = check_state_run_pq;
2784 sh->check_state = check_state_run_q;
2787 /* discard potentially stale zero_sum_result */
2788 sh->ops.zero_sum_result = 0;
2790 if (sh->check_state == check_state_run) {
2791 /* async_xor_zero_sum destroys the contents of P */
2792 clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
2795 if (sh->check_state >= check_state_run &&
2796 sh->check_state <= check_state_run_pq) {
2797 /* async_syndrome_zero_sum preserves P and Q, so
2798 * no need to mark them !uptodate here
2800 set_bit(STRIPE_OP_CHECK, &s->ops_request);
2804 /* we have 2-disk failure */
2805 BUG_ON(s->failed != 2);
2807 case check_state_compute_result:
2808 sh->check_state = check_state_idle;
2810 /* check that a write has not made the stripe insync */
2811 if (test_bit(STRIPE_INSYNC, &sh->state))
2814 /* now write out any block on a failed drive,
2815 * or P or Q if they were recomputed
2817 BUG_ON(s->uptodate < disks - 1); /* We don't need Q to recover */
2818 if (s->failed == 2) {
2819 dev = &sh->dev[s->failed_num[1]];
2821 set_bit(R5_LOCKED, &dev->flags);
2822 set_bit(R5_Wantwrite, &dev->flags);
2824 if (s->failed >= 1) {
2825 dev = &sh->dev[s->failed_num[0]];
2827 set_bit(R5_LOCKED, &dev->flags);
2828 set_bit(R5_Wantwrite, &dev->flags);
2830 if (sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) {
2831 dev = &sh->dev[pd_idx];
2833 set_bit(R5_LOCKED, &dev->flags);
2834 set_bit(R5_Wantwrite, &dev->flags);
2836 if (sh->ops.zero_sum_result & SUM_CHECK_Q_RESULT) {
2837 dev = &sh->dev[qd_idx];
2839 set_bit(R5_LOCKED, &dev->flags);
2840 set_bit(R5_Wantwrite, &dev->flags);
2842 clear_bit(STRIPE_DEGRADED, &sh->state);
2844 set_bit(STRIPE_INSYNC, &sh->state);
2846 case check_state_run:
2847 case check_state_run_q:
2848 case check_state_run_pq:
2849 break; /* we will be called again upon completion */
2850 case check_state_check_result:
2851 sh->check_state = check_state_idle;
2853 /* handle a successful check operation, if parity is correct
2854 * we are done. Otherwise update the mismatch count and repair
2855 * parity if !MD_RECOVERY_CHECK
2857 if (sh->ops.zero_sum_result == 0) {
2858 /* both parities are correct */
2860 set_bit(STRIPE_INSYNC, &sh->state);
2862 /* in contrast to the raid5 case we can validate
2863 * parity, but still have a failure to write
2866 sh->check_state = check_state_compute_result;
2867 /* Returning at this point means that we may go
2868 * off and bring p and/or q uptodate again so
2869 * we make sure to check zero_sum_result again
2870 * to verify if p or q need writeback
2874 conf->mddev->resync_mismatches += STRIPE_SECTORS;
2875 if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery))
2876 /* don't try to repair!! */
2877 set_bit(STRIPE_INSYNC, &sh->state);
2879 int *target = &sh->ops.target;
2881 sh->ops.target = -1;
2882 sh->ops.target2 = -1;
2883 sh->check_state = check_state_compute_run;
2884 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
2885 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
2886 if (sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) {
2887 set_bit(R5_Wantcompute,
2888 &sh->dev[pd_idx].flags);
2890 target = &sh->ops.target2;
2893 if (sh->ops.zero_sum_result & SUM_CHECK_Q_RESULT) {
2894 set_bit(R5_Wantcompute,
2895 &sh->dev[qd_idx].flags);
2902 case check_state_compute_run:
2905 printk(KERN_ERR "%s: unknown check_state: %d sector: %llu\n",
2906 __func__, sh->check_state,
2907 (unsigned long long) sh->sector);
2912 static void handle_stripe_expansion(struct r5conf *conf, struct stripe_head *sh)
2916 /* We have read all the blocks in this stripe and now we need to
2917 * copy some of them into a target stripe for expand.
2919 struct dma_async_tx_descriptor *tx = NULL;
2920 clear_bit(STRIPE_EXPAND_SOURCE, &sh->state);
2921 for (i = 0; i < sh->disks; i++)
2922 if (i != sh->pd_idx && i != sh->qd_idx) {
2924 struct stripe_head *sh2;
2925 struct async_submit_ctl submit;
2927 sector_t bn = compute_blocknr(sh, i, 1);
2928 sector_t s = raid5_compute_sector(conf, bn, 0,
2930 sh2 = get_active_stripe(conf, s, 0, 1, 1);
2932 /* so far only the early blocks of this stripe
2933 * have been requested. When later blocks
2934 * get requested, we will try again
2937 if (!test_bit(STRIPE_EXPANDING, &sh2->state) ||
2938 test_bit(R5_Expanded, &sh2->dev[dd_idx].flags)) {
2939 /* must have already done this block */
2940 release_stripe(sh2);
2944 /* place all the copies on one channel */
2945 init_async_submit(&submit, 0, tx, NULL, NULL, NULL);
2946 tx = async_memcpy(sh2->dev[dd_idx].page,
2947 sh->dev[i].page, 0, 0, STRIPE_SIZE,
2950 set_bit(R5_Expanded, &sh2->dev[dd_idx].flags);
2951 set_bit(R5_UPTODATE, &sh2->dev[dd_idx].flags);
2952 for (j = 0; j < conf->raid_disks; j++)
2953 if (j != sh2->pd_idx &&
2955 !test_bit(R5_Expanded, &sh2->dev[j].flags))
2957 if (j == conf->raid_disks) {
2958 set_bit(STRIPE_EXPAND_READY, &sh2->state);
2959 set_bit(STRIPE_HANDLE, &sh2->state);
2961 release_stripe(sh2);
2964 /* done submitting copies, wait for them to complete */
2967 dma_wait_for_async_tx(tx);
2973 * handle_stripe - do things to a stripe.
2975 * We lock the stripe and then examine the state of various bits
2976 * to see what needs to be done.
2978 * return some read request which now have data
2979 * return some write requests which are safely on disc
2980 * schedule a read on some buffers
2981 * schedule a write of some buffers
2982 * return confirmation of parity correctness
2984 * buffers are taken off read_list or write_list, and bh_cache buffers
2985 * get BH_Lock set before the stripe lock is released.
2989 static void analyse_stripe(struct stripe_head *sh, struct stripe_head_state *s)
2991 struct r5conf *conf = sh->raid_conf;
2992 int disks = sh->disks;
2996 memset(s, 0, sizeof(*s));
2998 s->syncing = test_bit(STRIPE_SYNCING, &sh->state);
2999 s->expanding = test_bit(STRIPE_EXPAND_SOURCE, &sh->state);
3000 s->expanded = test_bit(STRIPE_EXPAND_READY, &sh->state);
3001 s->failed_num[0] = -1;
3002 s->failed_num[1] = -1;
3004 /* Now to look around and see what can be done */
3006 spin_lock_irq(&conf->device_lock);
3007 for (i=disks; i--; ) {
3008 struct md_rdev *rdev;
3015 pr_debug("check %d: state 0x%lx read %p write %p written %p\n",
3016 i, dev->flags, dev->toread, dev->towrite, dev->written);
3017 /* maybe we can reply to a read
3019 * new wantfill requests are only permitted while
3020 * ops_complete_biofill is guaranteed to be inactive
3022 if (test_bit(R5_UPTODATE, &dev->flags) && dev->toread &&
3023 !test_bit(STRIPE_BIOFILL_RUN, &sh->state))
3024 set_bit(R5_Wantfill, &dev->flags);
3026 /* now count some things */
3027 if (test_bit(R5_LOCKED, &dev->flags))
3029 if (test_bit(R5_UPTODATE, &dev->flags))
3031 if (test_bit(R5_Wantcompute, &dev->flags)) {
3033 BUG_ON(s->compute > 2);
3036 if (test_bit(R5_Wantfill, &dev->flags))
3038 else if (dev->toread)
3042 if (!test_bit(R5_OVERWRITE, &dev->flags))
3047 rdev = rcu_dereference(conf->disks[i].rdev);
3048 if (rdev && test_bit(Faulty, &rdev->flags))
3051 is_bad = is_badblock(rdev, sh->sector, STRIPE_SECTORS,
3052 &first_bad, &bad_sectors);
3053 if (s->blocked_rdev == NULL
3054 && (test_bit(Blocked, &rdev->flags)
3057 set_bit(BlockedBadBlocks,
3059 s->blocked_rdev = rdev;
3060 atomic_inc(&rdev->nr_pending);
3063 clear_bit(R5_Insync, &dev->flags);
3067 /* also not in-sync */
3068 if (!test_bit(WriteErrorSeen, &rdev->flags)) {
3069 /* treat as in-sync, but with a read error
3070 * which we can now try to correct
3072 set_bit(R5_Insync, &dev->flags);
3073 set_bit(R5_ReadError, &dev->flags);
3075 } else if (test_bit(In_sync, &rdev->flags))
3076 set_bit(R5_Insync, &dev->flags);
3077 else if (sh->sector + STRIPE_SECTORS <= rdev->recovery_offset)
3078 /* in sync if before recovery_offset */
3079 set_bit(R5_Insync, &dev->flags);
3080 else if (test_bit(R5_UPTODATE, &dev->flags) &&
3081 test_bit(R5_Expanded, &dev->flags))
3082 /* If we've reshaped into here, we assume it is Insync.
3083 * We will shortly update recovery_offset to make
3086 set_bit(R5_Insync, &dev->flags);
3088 if (test_bit(R5_WriteError, &dev->flags)) {
3089 clear_bit(R5_Insync, &dev->flags);
3090 if (!test_bit(Faulty, &rdev->flags)) {
3091 s->handle_bad_blocks = 1;
3092 atomic_inc(&rdev->nr_pending);
3094 clear_bit(R5_WriteError, &dev->flags);
3096 if (test_bit(R5_MadeGood, &dev->flags)) {
3097 if (!test_bit(Faulty, &rdev->flags)) {
3098 s->handle_bad_blocks = 1;
3099 atomic_inc(&rdev->nr_pending);
3101 clear_bit(R5_MadeGood, &dev->flags);
3103 if (!test_bit(R5_Insync, &dev->flags)) {
3104 /* The ReadError flag will just be confusing now */
3105 clear_bit(R5_ReadError, &dev->flags);
3106 clear_bit(R5_ReWrite, &dev->flags);
3108 if (test_bit(R5_ReadError, &dev->flags))
3109 clear_bit(R5_Insync, &dev->flags);
3110 if (!test_bit(R5_Insync, &dev->flags)) {
3112 s->failed_num[s->failed] = i;
3116 spin_unlock_irq(&conf->device_lock);
3120 static void handle_stripe(struct stripe_head *sh)
3122 struct stripe_head_state s;
3123 struct r5conf *conf = sh->raid_conf;
3126 int disks = sh->disks;
3127 struct r5dev *pdev, *qdev;
3129 clear_bit(STRIPE_HANDLE, &sh->state);
3130 if (test_and_set_bit_lock(STRIPE_ACTIVE, &sh->state)) {
3131 /* already being handled, ensure it gets handled
3132 * again when current action finishes */
3133 set_bit(STRIPE_HANDLE, &sh->state);
3137 if (test_and_clear_bit(STRIPE_SYNC_REQUESTED, &sh->state)) {
3138 set_bit(STRIPE_SYNCING, &sh->state);
3139 clear_bit(STRIPE_INSYNC, &sh->state);
3141 clear_bit(STRIPE_DELAYED, &sh->state);
3143 pr_debug("handling stripe %llu, state=%#lx cnt=%d, "
3144 "pd_idx=%d, qd_idx=%d\n, check:%d, reconstruct:%d\n",
3145 (unsigned long long)sh->sector, sh->state,
3146 atomic_read(&sh->count), sh->pd_idx, sh->qd_idx,
3147 sh->check_state, sh->reconstruct_state);
3149 analyse_stripe(sh, &s);
3151 if (s.handle_bad_blocks) {
3152 set_bit(STRIPE_HANDLE, &sh->state);
3156 if (unlikely(s.blocked_rdev)) {
3157 if (s.syncing || s.expanding || s.expanded ||
3158 s.to_write || s.written) {
3159 set_bit(STRIPE_HANDLE, &sh->state);
3162 /* There is nothing for the blocked_rdev to block */
3163 rdev_dec_pending(s.blocked_rdev, conf->mddev);
3164 s.blocked_rdev = NULL;
3167 if (s.to_fill && !test_bit(STRIPE_BIOFILL_RUN, &sh->state)) {
3168 set_bit(STRIPE_OP_BIOFILL, &s.ops_request);
3169 set_bit(STRIPE_BIOFILL_RUN, &sh->state);
3172 pr_debug("locked=%d uptodate=%d to_read=%d"
3173 " to_write=%d failed=%d failed_num=%d,%d\n",
3174 s.locked, s.uptodate, s.to_read, s.to_write, s.failed,
3175 s.failed_num[0], s.failed_num[1]);
3176 /* check if the array has lost more than max_degraded devices and,
3177 * if so, some requests might need to be failed.
3179 if (s.failed > conf->max_degraded) {
3180 sh->check_state = 0;
3181 sh->reconstruct_state = 0;
3182 if (s.to_read+s.to_write+s.written)
3183 handle_failed_stripe(conf, sh, &s, disks, &s.return_bi);
3185 handle_failed_sync(conf, sh, &s);
3189 * might be able to return some write requests if the parity blocks
3190 * are safe, or on a failed drive
3192 pdev = &sh->dev[sh->pd_idx];
3193 s.p_failed = (s.failed >= 1 && s.failed_num[0] == sh->pd_idx)
3194 || (s.failed >= 2 && s.failed_num[1] == sh->pd_idx);
3195 qdev = &sh->dev[sh->qd_idx];
3196 s.q_failed = (s.failed >= 1 && s.failed_num[0] == sh->qd_idx)
3197 || (s.failed >= 2 && s.failed_num[1] == sh->qd_idx)
3201 (s.p_failed || ((test_bit(R5_Insync, &pdev->flags)
3202 && !test_bit(R5_LOCKED, &pdev->flags)
3203 && test_bit(R5_UPTODATE, &pdev->flags)))) &&
3204 (s.q_failed || ((test_bit(R5_Insync, &qdev->flags)
3205 && !test_bit(R5_LOCKED, &qdev->flags)
3206 && test_bit(R5_UPTODATE, &qdev->flags)))))
3207 handle_stripe_clean_event(conf, sh, disks, &s.return_bi);
3209 /* Now we might consider reading some blocks, either to check/generate
3210 * parity, or to satisfy requests
3211 * or to load a block that is being partially written.
3213 if (s.to_read || s.non_overwrite
3214 || (conf->level == 6 && s.to_write && s.failed)
3215 || (s.syncing && (s.uptodate + s.compute < disks)) || s.expanding)
3216 handle_stripe_fill(sh, &s, disks);
3218 /* Now we check to see if any write operations have recently
3222 if (sh->reconstruct_state == reconstruct_state_prexor_drain_result)
3224 if (sh->reconstruct_state == reconstruct_state_drain_result ||
3225 sh->reconstruct_state == reconstruct_state_prexor_drain_result) {
3226 sh->reconstruct_state = reconstruct_state_idle;
3228 /* All the 'written' buffers and the parity block are ready to
3229 * be written back to disk
3231 BUG_ON(!test_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags));
3232 BUG_ON(sh->qd_idx >= 0 &&
3233 !test_bit(R5_UPTODATE, &sh->dev[sh->qd_idx].flags));
3234 for (i = disks; i--; ) {
3235 struct r5dev *dev = &sh->dev[i];
3236 if (test_bit(R5_LOCKED, &dev->flags) &&
3237 (i == sh->pd_idx || i == sh->qd_idx ||
3239 pr_debug("Writing block %d\n", i);
3240 set_bit(R5_Wantwrite, &dev->flags);
3245 if (!test_bit(R5_Insync, &dev->flags) ||
3246 ((i == sh->pd_idx || i == sh->qd_idx) &&
3248 set_bit(STRIPE_INSYNC, &sh->state);
3251 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
3252 s.dec_preread_active = 1;
3255 /* Now to consider new write requests and what else, if anything
3256 * should be read. We do not handle new writes when:
3257 * 1/ A 'write' operation (copy+xor) is already in flight.
3258 * 2/ A 'check' operation is in flight, as it may clobber the parity
3261 if (s.to_write && !sh->reconstruct_state && !sh->check_state)
3262 handle_stripe_dirtying(conf, sh, &s, disks);
3264 /* maybe we need to check and possibly fix the parity for this stripe
3265 * Any reads will already have been scheduled, so we just see if enough
3266 * data is available. The parity check is held off while parity
3267 * dependent operations are in flight.
3269 if (sh->check_state ||
3270 (s.syncing && s.locked == 0 &&
3271 !test_bit(STRIPE_COMPUTE_RUN, &sh->state) &&
3272 !test_bit(STRIPE_INSYNC, &sh->state))) {
3273 if (conf->level == 6)
3274 handle_parity_checks6(conf, sh, &s, disks);
3276 handle_parity_checks5(conf, sh, &s, disks);
3279 if (s.syncing && s.locked == 0 && test_bit(STRIPE_INSYNC, &sh->state)) {
3280 md_done_sync(conf->mddev, STRIPE_SECTORS, 1);
3281 clear_bit(STRIPE_SYNCING, &sh->state);
3284 /* If the failed drives are just a ReadError, then we might need
3285 * to progress the repair/check process
3287 if (s.failed <= conf->max_degraded && !conf->mddev->ro)
3288 for (i = 0; i < s.failed; i++) {
3289 struct r5dev *dev = &sh->dev[s.failed_num[i]];
3290 if (test_bit(R5_ReadError, &dev->flags)
3291 && !test_bit(R5_LOCKED, &dev->flags)
3292 && test_bit(R5_UPTODATE, &dev->flags)
3294 if (!test_bit(R5_ReWrite, &dev->flags)) {
3295 set_bit(R5_Wantwrite, &dev->flags);
3296 set_bit(R5_ReWrite, &dev->flags);
3297 set_bit(R5_LOCKED, &dev->flags);
3300 /* let's read it back */
3301 set_bit(R5_Wantread, &dev->flags);
3302 set_bit(R5_LOCKED, &dev->flags);
3309 /* Finish reconstruct operations initiated by the expansion process */
3310 if (sh->reconstruct_state == reconstruct_state_result) {
3311 struct stripe_head *sh_src
3312 = get_active_stripe(conf, sh->sector, 1, 1, 1);
3313 if (sh_src && test_bit(STRIPE_EXPAND_SOURCE, &sh_src->state)) {
3314 /* sh cannot be written until sh_src has been read.
3315 * so arrange for sh to be delayed a little
3317 set_bit(STRIPE_DELAYED, &sh->state);
3318 set_bit(STRIPE_HANDLE, &sh->state);
3319 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE,
3321 atomic_inc(&conf->preread_active_stripes);
3322 release_stripe(sh_src);
3326 release_stripe(sh_src);
3328 sh->reconstruct_state = reconstruct_state_idle;
3329 clear_bit(STRIPE_EXPANDING, &sh->state);
3330 for (i = conf->raid_disks; i--; ) {
3331 set_bit(R5_Wantwrite, &sh->dev[i].flags);
3332 set_bit(R5_LOCKED, &sh->dev[i].flags);
3337 if (s.expanded && test_bit(STRIPE_EXPANDING, &sh->state) &&
3338 !sh->reconstruct_state) {
3339 /* Need to write out all blocks after computing parity */
3340 sh->disks = conf->raid_disks;
3341 stripe_set_idx(sh->sector, conf, 0, sh);
3342 schedule_reconstruction(sh, &s, 1, 1);
3343 } else if (s.expanded && !sh->reconstruct_state && s.locked == 0) {
3344 clear_bit(STRIPE_EXPAND_READY, &sh->state);
3345 atomic_dec(&conf->reshape_stripes);
3346 wake_up(&conf->wait_for_overlap);
3347 md_done_sync(conf->mddev, STRIPE_SECTORS, 1);
3350 if (s.expanding && s.locked == 0 &&
3351 !test_bit(STRIPE_COMPUTE_RUN, &sh->state))
3352 handle_stripe_expansion(conf, sh);
3355 /* wait for this device to become unblocked */
3356 if (conf->mddev->external && unlikely(s.blocked_rdev))
3357 md_wait_for_blocked_rdev(s.blocked_rdev, conf->mddev);
3359 if (s.handle_bad_blocks)
3360 for (i = disks; i--; ) {
3361 struct md_rdev *rdev;
3362 struct r5dev *dev = &sh->dev[i];
3363 if (test_and_clear_bit(R5_WriteError, &dev->flags)) {
3364 /* We own a safe reference to the rdev */
3365 rdev = conf->disks[i].rdev;
3366 if (!rdev_set_badblocks(rdev, sh->sector,
3368 md_error(conf->mddev, rdev);
3369 rdev_dec_pending(rdev, conf->mddev);
3371 if (test_and_clear_bit(R5_MadeGood, &dev->flags)) {
3372 rdev = conf->disks[i].rdev;
3373 rdev_clear_badblocks(rdev, sh->sector,
3375 rdev_dec_pending(rdev, conf->mddev);
3380 raid_run_ops(sh, s.ops_request);
3384 if (s.dec_preread_active) {
3385 /* We delay this until after ops_run_io so that if make_request
3386 * is waiting on a flush, it won't continue until the writes
3387 * have actually been submitted.
3389 atomic_dec(&conf->preread_active_stripes);
3390 if (atomic_read(&conf->preread_active_stripes) <
3392 md_wakeup_thread(conf->mddev->thread);
3395 return_io(s.return_bi);
3397 clear_bit_unlock(STRIPE_ACTIVE, &sh->state);
3400 static void raid5_activate_delayed(struct r5conf *conf)
3402 if (atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD) {
3403 while (!list_empty(&conf->delayed_list)) {
3404 struct list_head *l = conf->delayed_list.next;
3405 struct stripe_head *sh;
3406 sh = list_entry(l, struct stripe_head, lru);
3408 clear_bit(STRIPE_DELAYED, &sh->state);
3409 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
3410 atomic_inc(&conf->preread_active_stripes);
3411 list_add_tail(&sh->lru, &conf->hold_list);
3416 static void activate_bit_delay(struct r5conf *conf)
3418 /* device_lock is held */
3419 struct list_head head;
3420 list_add(&head, &conf->bitmap_list);
3421 list_del_init(&conf->bitmap_list);
3422 while (!list_empty(&head)) {
3423 struct stripe_head *sh = list_entry(head.next, struct stripe_head, lru);
3424 list_del_init(&sh->lru);
3425 atomic_inc(&sh->count);
3426 __release_stripe(conf, sh);
3430 int md_raid5_congested(struct mddev *mddev, int bits)
3432 struct r5conf *conf = mddev->private;
3434 /* No difference between reads and writes. Just check
3435 * how busy the stripe_cache is
3438 if (conf->inactive_blocked)
3442 if (list_empty_careful(&conf->inactive_list))
3447 EXPORT_SYMBOL_GPL(md_raid5_congested);
3449 static int raid5_congested(void *data, int bits)
3451 struct mddev *mddev = data;
3453 return mddev_congested(mddev, bits) ||
3454 md_raid5_congested(mddev, bits);
3457 /* We want read requests to align with chunks where possible,
3458 * but write requests don't need to.
3460 static int raid5_mergeable_bvec(struct request_queue *q,
3461 struct bvec_merge_data *bvm,
3462 struct bio_vec *biovec)
3464 struct mddev *mddev = q->queuedata;
3465 sector_t sector = bvm->bi_sector + get_start_sect(bvm->bi_bdev);
3467 unsigned int chunk_sectors = mddev->chunk_sectors;
3468 unsigned int bio_sectors = bvm->bi_size >> 9;
3470 if ((bvm->bi_rw & 1) == WRITE)
3471 return biovec->bv_len; /* always allow writes to be mergeable */
3473 if (mddev->new_chunk_sectors < mddev->chunk_sectors)
3474 chunk_sectors = mddev->new_chunk_sectors;
3475 max = (chunk_sectors - ((sector & (chunk_sectors - 1)) + bio_sectors)) << 9;
3476 if (max < 0) max = 0;
3477 if (max <= biovec->bv_len && bio_sectors == 0)
3478 return biovec->bv_len;
3484 static int in_chunk_boundary(struct mddev *mddev, struct bio *bio)
3486 sector_t sector = bio->bi_sector + get_start_sect(bio->bi_bdev);
3487 unsigned int chunk_sectors = mddev->chunk_sectors;
3488 unsigned int bio_sectors = bio->bi_size >> 9;
3490 if (mddev->new_chunk_sectors < mddev->chunk_sectors)
3491 chunk_sectors = mddev->new_chunk_sectors;
3492 return chunk_sectors >=
3493 ((sector & (chunk_sectors - 1)) + bio_sectors);
3497 * add bio to the retry LIFO ( in O(1) ... we are in interrupt )
3498 * later sampled by raid5d.
3500 static void add_bio_to_retry(struct bio *bi,struct r5conf *conf)
3502 unsigned long flags;
3504 spin_lock_irqsave(&conf->device_lock, flags);
3506 bi->bi_next = conf->retry_read_aligned_list;
3507 conf->retry_read_aligned_list = bi;
3509 spin_unlock_irqrestore(&conf->device_lock, flags);
3510 md_wakeup_thread(conf->mddev->thread);
3514 static struct bio *remove_bio_from_retry(struct r5conf *conf)
3518 bi = conf->retry_read_aligned;
3520 conf->retry_read_aligned = NULL;
3523 bi = conf->retry_read_aligned_list;
3525 conf->retry_read_aligned_list = bi->bi_next;
3528 * this sets the active strip count to 1 and the processed
3529 * strip count to zero (upper 8 bits)
3531 bi->bi_phys_segments = 1; /* biased count of active stripes */
3539 * The "raid5_align_endio" should check if the read succeeded and if it
3540 * did, call bio_endio on the original bio (having bio_put the new bio
3542 * If the read failed..
3544 static void raid5_align_endio(struct bio *bi, int error)
3546 struct bio* raid_bi = bi->bi_private;
3547 struct mddev *mddev;
3548 struct r5conf *conf;
3549 int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
3550 struct md_rdev *rdev;
3554 rdev = (void*)raid_bi->bi_next;
3555 raid_bi->bi_next = NULL;
3556 mddev = rdev->mddev;
3557 conf = mddev->private;
3559 rdev_dec_pending(rdev, conf->mddev);
3561 if (!error && uptodate) {
3562 bio_endio(raid_bi, 0);
3563 if (atomic_dec_and_test(&conf->active_aligned_reads))
3564 wake_up(&conf->wait_for_stripe);
3569 pr_debug("raid5_align_endio : io error...handing IO for a retry\n");
3571 add_bio_to_retry(raid_bi, conf);
3574 static int bio_fits_rdev(struct bio *bi)
3576 struct request_queue *q = bdev_get_queue(bi->bi_bdev);
3578 if ((bi->bi_size>>9) > queue_max_sectors(q))
3580 blk_recount_segments(q, bi);
3581 if (bi->bi_phys_segments > queue_max_segments(q))
3584 if (q->merge_bvec_fn)
3585 /* it's too hard to apply the merge_bvec_fn at this stage,
3594 static int chunk_aligned_read(struct mddev *mddev, struct bio * raid_bio)
3596 struct r5conf *conf = mddev->private;
3598 struct bio* align_bi;
3599 struct md_rdev *rdev;
3601 if (!in_chunk_boundary(mddev, raid_bio)) {
3602 pr_debug("chunk_aligned_read : non aligned\n");
3606 * use bio_clone_mddev to make a copy of the bio
3608 align_bi = bio_clone_mddev(raid_bio, GFP_NOIO, mddev);
3612 * set bi_end_io to a new function, and set bi_private to the
3615 align_bi->bi_end_io = raid5_align_endio;
3616 align_bi->bi_private = raid_bio;
3620 align_bi->bi_sector = raid5_compute_sector(conf, raid_bio->bi_sector,
3625 rdev = rcu_dereference(conf->disks[dd_idx].rdev);
3626 if (rdev && test_bit(In_sync, &rdev->flags)) {
3630 atomic_inc(&rdev->nr_pending);
3632 raid_bio->bi_next = (void*)rdev;
3633 align_bi->bi_bdev = rdev->bdev;
3634 align_bi->bi_flags &= ~(1 << BIO_SEG_VALID);
3636 if (!bio_fits_rdev(align_bi) ||
3637 is_badblock(rdev, align_bi->bi_sector, align_bi->bi_size>>9,
3638 &first_bad, &bad_sectors)) {
3639 /* too big in some way, or has a known bad block */
3641 rdev_dec_pending(rdev, mddev);
3645 /* No reshape active, so we can trust rdev->data_offset */
3646 align_bi->bi_sector += rdev->data_offset;
3648 spin_lock_irq(&conf->device_lock);
3649 wait_event_lock_irq(conf->wait_for_stripe,
3651 conf->device_lock, /* nothing */);
3652 atomic_inc(&conf->active_aligned_reads);
3653 spin_unlock_irq(&conf->device_lock);
3655 generic_make_request(align_bi);
3664 /* __get_priority_stripe - get the next stripe to process
3666 * Full stripe writes are allowed to pass preread active stripes up until
3667 * the bypass_threshold is exceeded. In general the bypass_count
3668 * increments when the handle_list is handled before the hold_list; however, it
3669 * will not be incremented when STRIPE_IO_STARTED is sampled set signifying a
3670 * stripe with in flight i/o. The bypass_count will be reset when the
3671 * head of the hold_list has changed, i.e. the head was promoted to the
3674 static struct stripe_head *__get_priority_stripe(struct r5conf *conf)
3676 struct stripe_head *sh;
3678 pr_debug("%s: handle: %s hold: %s full_writes: %d bypass_count: %d\n",
3680 list_empty(&conf->handle_list) ? "empty" : "busy",
3681 list_empty(&conf->hold_list) ? "empty" : "busy",
3682 atomic_read(&conf->pending_full_writes), conf->bypass_count);
3684 if (!list_empty(&conf->handle_list)) {
3685 sh = list_entry(conf->handle_list.next, typeof(*sh), lru);
3687 if (list_empty(&conf->hold_list))
3688 conf->bypass_count = 0;
3689 else if (!test_bit(STRIPE_IO_STARTED, &sh->state)) {
3690 if (conf->hold_list.next == conf->last_hold)
3691 conf->bypass_count++;
3693 conf->last_hold = conf->hold_list.next;
3694 conf->bypass_count -= conf->bypass_threshold;
3695 if (conf->bypass_count < 0)
3696 conf->bypass_count = 0;
3699 } else if (!list_empty(&conf->hold_list) &&
3700 ((conf->bypass_threshold &&
3701 conf->bypass_count > conf->bypass_threshold) ||
3702 atomic_read(&conf->pending_full_writes) == 0)) {
3703 sh = list_entry(conf->hold_list.next,
3705 conf->bypass_count -= conf->bypass_threshold;
3706 if (conf->bypass_count < 0)
3707 conf->bypass_count = 0;
3711 list_del_init(&sh->lru);
3712 atomic_inc(&sh->count);
3713 BUG_ON(atomic_read(&sh->count) != 1);
3717 static void make_request(struct mddev *mddev, struct bio * bi)
3719 struct r5conf *conf = mddev->private;
3721 sector_t new_sector;
3722 sector_t logical_sector, last_sector;
3723 struct stripe_head *sh;
3724 const int rw = bio_data_dir(bi);
3728 if (unlikely(bi->bi_rw & REQ_FLUSH)) {
3729 md_flush_request(mddev, bi);
3733 md_write_start(mddev, bi);
3736 mddev->reshape_position == MaxSector &&
3737 chunk_aligned_read(mddev,bi))
3740 logical_sector = bi->bi_sector & ~((sector_t)STRIPE_SECTORS-1);
3741 last_sector = bi->bi_sector + (bi->bi_size>>9);
3743 bi->bi_phys_segments = 1; /* over-loaded to count active stripes */
3745 plugged = mddev_check_plugged(mddev);
3746 for (;logical_sector < last_sector; logical_sector += STRIPE_SECTORS) {
3748 int disks, data_disks;
3753 disks = conf->raid_disks;
3754 prepare_to_wait(&conf->wait_for_overlap, &w, TASK_UNINTERRUPTIBLE);
3755 if (unlikely(conf->reshape_progress != MaxSector)) {
3756 /* spinlock is needed as reshape_progress may be
3757 * 64bit on a 32bit platform, and so it might be
3758 * possible to see a half-updated value
3759 * Of course reshape_progress could change after
3760 * the lock is dropped, so once we get a reference
3761 * to the stripe that we think it is, we will have
3764 spin_lock_irq(&conf->device_lock);
3765 if (mddev->delta_disks < 0
3766 ? logical_sector < conf->reshape_progress
3767 : logical_sector >= conf->reshape_progress) {
3768 disks = conf->previous_raid_disks;
3771 if (mddev->delta_disks < 0
3772 ? logical_sector < conf->reshape_safe
3773 : logical_sector >= conf->reshape_safe) {
3774 spin_unlock_irq(&conf->device_lock);
3779 spin_unlock_irq(&conf->device_lock);
3781 data_disks = disks - conf->max_degraded;
3783 new_sector = raid5_compute_sector(conf, logical_sector,
3786 pr_debug("raid456: make_request, sector %llu logical %llu\n",
3787 (unsigned long long)new_sector,
3788 (unsigned long long)logical_sector);
3790 sh = get_active_stripe(conf, new_sector, previous,
3791 (bi->bi_rw&RWA_MASK), 0);
3793 if (unlikely(previous)) {
3794 /* expansion might have moved on while waiting for a
3795 * stripe, so we must do the range check again.
3796 * Expansion could still move past after this
3797 * test, but as we are holding a reference to
3798 * 'sh', we know that if that happens,
3799 * STRIPE_EXPANDING will get set and the expansion
3800 * won't proceed until we finish with the stripe.
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 /* mismatch, need to try again */
3809 spin_unlock_irq(&conf->device_lock);
3818 logical_sector >= mddev->suspend_lo &&
3819 logical_sector < mddev->suspend_hi) {
3821 /* As the suspend_* range is controlled by
3822 * userspace, we want an interruptible
3825 flush_signals(current);
3826 prepare_to_wait(&conf->wait_for_overlap,
3827 &w, TASK_INTERRUPTIBLE);
3828 if (logical_sector >= mddev->suspend_lo &&
3829 logical_sector < mddev->suspend_hi)
3834 if (test_bit(STRIPE_EXPANDING, &sh->state) ||
3835 !add_stripe_bio(sh, bi, dd_idx, rw)) {
3836 /* Stripe is busy expanding or
3837 * add failed due to overlap. Flush everything
3840 md_wakeup_thread(mddev->thread);
3845 finish_wait(&conf->wait_for_overlap, &w);
3846 set_bit(STRIPE_HANDLE, &sh->state);
3847 clear_bit(STRIPE_DELAYED, &sh->state);
3848 if ((bi->bi_rw & REQ_SYNC) &&
3849 !test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
3850 atomic_inc(&conf->preread_active_stripes);
3853 /* cannot get stripe for read-ahead, just give-up */
3854 clear_bit(BIO_UPTODATE, &bi->bi_flags);
3855 finish_wait(&conf->wait_for_overlap, &w);
3861 md_wakeup_thread(mddev->thread);
3863 spin_lock_irq(&conf->device_lock);
3864 remaining = raid5_dec_bi_phys_segments(bi);
3865 spin_unlock_irq(&conf->device_lock);
3866 if (remaining == 0) {
3869 md_write_end(mddev);
3875 static sector_t raid5_size(struct mddev *mddev, sector_t sectors, int raid_disks);
3877 static sector_t reshape_request(struct mddev *mddev, sector_t sector_nr, int *skipped)
3879 /* reshaping is quite different to recovery/resync so it is
3880 * handled quite separately ... here.
3882 * On each call to sync_request, we gather one chunk worth of
3883 * destination stripes and flag them as expanding.
3884 * Then we find all the source stripes and request reads.
3885 * As the reads complete, handle_stripe will copy the data
3886 * into the destination stripe and release that stripe.
3888 struct r5conf *conf = mddev->private;
3889 struct stripe_head *sh;
3890 sector_t first_sector, last_sector;
3891 int raid_disks = conf->previous_raid_disks;
3892 int data_disks = raid_disks - conf->max_degraded;
3893 int new_data_disks = conf->raid_disks - conf->max_degraded;
3896 sector_t writepos, readpos, safepos;
3897 sector_t stripe_addr;
3898 int reshape_sectors;
3899 struct list_head stripes;
3901 if (sector_nr == 0) {
3902 /* If restarting in the middle, skip the initial sectors */
3903 if (mddev->delta_disks < 0 &&
3904 conf->reshape_progress < raid5_size(mddev, 0, 0)) {
3905 sector_nr = raid5_size(mddev, 0, 0)
3906 - conf->reshape_progress;
3907 } else if (mddev->delta_disks >= 0 &&
3908 conf->reshape_progress > 0)
3909 sector_nr = conf->reshape_progress;
3910 sector_div(sector_nr, new_data_disks);
3912 mddev->curr_resync_completed = sector_nr;
3913 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
3919 /* We need to process a full chunk at a time.
3920 * If old and new chunk sizes differ, we need to process the
3923 if (mddev->new_chunk_sectors > mddev->chunk_sectors)
3924 reshape_sectors = mddev->new_chunk_sectors;
3926 reshape_sectors = mddev->chunk_sectors;
3928 /* we update the metadata when there is more than 3Meg
3929 * in the block range (that is rather arbitrary, should
3930 * probably be time based) or when the data about to be
3931 * copied would over-write the source of the data at
3932 * the front of the range.
3933 * i.e. one new_stripe along from reshape_progress new_maps
3934 * to after where reshape_safe old_maps to
3936 writepos = conf->reshape_progress;
3937 sector_div(writepos, new_data_disks);
3938 readpos = conf->reshape_progress;
3939 sector_div(readpos, data_disks);
3940 safepos = conf->reshape_safe;
3941 sector_div(safepos, data_disks);
3942 if (mddev->delta_disks < 0) {
3943 writepos -= min_t(sector_t, reshape_sectors, writepos);
3944 readpos += reshape_sectors;
3945 safepos += reshape_sectors;
3947 writepos += reshape_sectors;
3948 readpos -= min_t(sector_t, reshape_sectors, readpos);
3949 safepos -= min_t(sector_t, reshape_sectors, safepos);
3952 /* 'writepos' is the most advanced device address we might write.
3953 * 'readpos' is the least advanced device address we might read.
3954 * 'safepos' is the least address recorded in the metadata as having
3956 * If 'readpos' is behind 'writepos', then there is no way that we can
3957 * ensure safety in the face of a crash - that must be done by userspace
3958 * making a backup of the data. So in that case there is no particular
3959 * rush to update metadata.
3960 * Otherwise if 'safepos' is behind 'writepos', then we really need to
3961 * update the metadata to advance 'safepos' to match 'readpos' so that
3962 * we can be safe in the event of a crash.
3963 * So we insist on updating metadata if safepos is behind writepos and
3964 * readpos is beyond writepos.
3965 * In any case, update the metadata every 10 seconds.
3966 * Maybe that number should be configurable, but I'm not sure it is
3967 * worth it.... maybe it could be a multiple of safemode_delay???
3969 if ((mddev->delta_disks < 0
3970 ? (safepos > writepos && readpos < writepos)
3971 : (safepos < writepos && readpos > writepos)) ||
3972 time_after(jiffies, conf->reshape_checkpoint + 10*HZ)) {
3973 /* Cannot proceed until we've updated the superblock... */
3974 wait_event(conf->wait_for_overlap,
3975 atomic_read(&conf->reshape_stripes)==0);
3976 mddev->reshape_position = conf->reshape_progress;
3977 mddev->curr_resync_completed = sector_nr;
3978 conf->reshape_checkpoint = jiffies;
3979 set_bit(MD_CHANGE_DEVS, &mddev->flags);
3980 md_wakeup_thread(mddev->thread);
3981 wait_event(mddev->sb_wait, mddev->flags == 0 ||
3982 kthread_should_stop());
3983 spin_lock_irq(&conf->device_lock);
3984 conf->reshape_safe = mddev->reshape_position;
3985 spin_unlock_irq(&conf->device_lock);
3986 wake_up(&conf->wait_for_overlap);
3987 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
3990 if (mddev->delta_disks < 0) {
3991 BUG_ON(conf->reshape_progress == 0);
3992 stripe_addr = writepos;
3993 BUG_ON((mddev->dev_sectors &
3994 ~((sector_t)reshape_sectors - 1))
3995 - reshape_sectors - stripe_addr
3998 BUG_ON(writepos != sector_nr + reshape_sectors);
3999 stripe_addr = sector_nr;
4001 INIT_LIST_HEAD(&stripes);
4002 for (i = 0; i < reshape_sectors; i += STRIPE_SECTORS) {
4004 int skipped_disk = 0;
4005 sh = get_active_stripe(conf, stripe_addr+i, 0, 0, 1);
4006 set_bit(STRIPE_EXPANDING, &sh->state);
4007 atomic_inc(&conf->reshape_stripes);
4008 /* If any of this stripe is beyond the end of the old
4009 * array, then we need to zero those blocks
4011 for (j=sh->disks; j--;) {
4013 if (j == sh->pd_idx)
4015 if (conf->level == 6 &&
4018 s = compute_blocknr(sh, j, 0);
4019 if (s < raid5_size(mddev, 0, 0)) {
4023 memset(page_address(sh->dev[j].page), 0, STRIPE_SIZE);
4024 set_bit(R5_Expanded, &sh->dev[j].flags);
4025 set_bit(R5_UPTODATE, &sh->dev[j].flags);
4027 if (!skipped_disk) {
4028 set_bit(STRIPE_EXPAND_READY, &sh->state);
4029 set_bit(STRIPE_HANDLE, &sh->state);
4031 list_add(&sh->lru, &stripes);
4033 spin_lock_irq(&conf->device_lock);
4034 if (mddev->delta_disks < 0)
4035 conf->reshape_progress -= reshape_sectors * new_data_disks;
4037 conf->reshape_progress += reshape_sectors * new_data_disks;
4038 spin_unlock_irq(&conf->device_lock);
4039 /* Ok, those stripe are ready. We can start scheduling
4040 * reads on the source stripes.
4041 * The source stripes are determined by mapping the first and last
4042 * block on the destination stripes.
4045 raid5_compute_sector(conf, stripe_addr*(new_data_disks),
4048 raid5_compute_sector(conf, ((stripe_addr+reshape_sectors)
4049 * new_data_disks - 1),
4051 if (last_sector >= mddev->dev_sectors)
4052 last_sector = mddev->dev_sectors - 1;
4053 while (first_sector <= last_sector) {
4054 sh = get_active_stripe(conf, first_sector, 1, 0, 1);
4055 set_bit(STRIPE_EXPAND_SOURCE, &sh->state);
4056 set_bit(STRIPE_HANDLE, &sh->state);
4058 first_sector += STRIPE_SECTORS;
4060 /* Now that the sources are clearly marked, we can release
4061 * the destination stripes
4063 while (!list_empty(&stripes)) {
4064 sh = list_entry(stripes.next, struct stripe_head, lru);
4065 list_del_init(&sh->lru);
4068 /* If this takes us to the resync_max point where we have to pause,
4069 * then we need to write out the superblock.
4071 sector_nr += reshape_sectors;
4072 if ((sector_nr - mddev->curr_resync_completed) * 2
4073 >= mddev->resync_max - mddev->curr_resync_completed) {
4074 /* Cannot proceed until we've updated the superblock... */
4075 wait_event(conf->wait_for_overlap,
4076 atomic_read(&conf->reshape_stripes) == 0);
4077 mddev->reshape_position = conf->reshape_progress;
4078 mddev->curr_resync_completed = sector_nr;
4079 conf->reshape_checkpoint = jiffies;
4080 set_bit(MD_CHANGE_DEVS, &mddev->flags);
4081 md_wakeup_thread(mddev->thread);
4082 wait_event(mddev->sb_wait,
4083 !test_bit(MD_CHANGE_DEVS, &mddev->flags)
4084 || kthread_should_stop());
4085 spin_lock_irq(&conf->device_lock);
4086 conf->reshape_safe = mddev->reshape_position;
4087 spin_unlock_irq(&conf->device_lock);
4088 wake_up(&conf->wait_for_overlap);
4089 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
4091 return reshape_sectors;
4094 /* FIXME go_faster isn't used */
4095 static inline sector_t sync_request(struct mddev *mddev, sector_t sector_nr, int *skipped, int go_faster)
4097 struct r5conf *conf = mddev->private;
4098 struct stripe_head *sh;
4099 sector_t max_sector = mddev->dev_sectors;
4100 sector_t sync_blocks;
4101 int still_degraded = 0;
4104 if (sector_nr >= max_sector) {
4105 /* just being told to finish up .. nothing much to do */
4107 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery)) {
4112 if (mddev->curr_resync < max_sector) /* aborted */
4113 bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
4115 else /* completed sync */
4117 bitmap_close_sync(mddev->bitmap);
4122 /* Allow raid5_quiesce to complete */
4123 wait_event(conf->wait_for_overlap, conf->quiesce != 2);
4125 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
4126 return reshape_request(mddev, sector_nr, skipped);
4128 /* No need to check resync_max as we never do more than one
4129 * stripe, and as resync_max will always be on a chunk boundary,
4130 * if the check in md_do_sync didn't fire, there is no chance
4131 * of overstepping resync_max here
4134 /* if there is too many failed drives and we are trying
4135 * to resync, then assert that we are finished, because there is
4136 * nothing we can do.
4138 if (mddev->degraded >= conf->max_degraded &&
4139 test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
4140 sector_t rv = mddev->dev_sectors - sector_nr;
4144 if (!bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
4145 !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
4146 !conf->fullsync && sync_blocks >= STRIPE_SECTORS) {
4147 /* we can skip this block, and probably more */
4148 sync_blocks /= STRIPE_SECTORS;
4150 return sync_blocks * STRIPE_SECTORS; /* keep things rounded to whole stripes */
4154 bitmap_cond_end_sync(mddev->bitmap, sector_nr);
4156 sh = get_active_stripe(conf, sector_nr, 0, 1, 0);
4158 sh = get_active_stripe(conf, sector_nr, 0, 0, 0);
4159 /* make sure we don't swamp the stripe cache if someone else
4160 * is trying to get access
4162 schedule_timeout_uninterruptible(1);
4164 /* Need to check if array will still be degraded after recovery/resync
4165 * We don't need to check the 'failed' flag as when that gets set,
4168 for (i = 0; i < conf->raid_disks; i++)
4169 if (conf->disks[i].rdev == NULL)
4172 bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, still_degraded);
4174 set_bit(STRIPE_SYNC_REQUESTED, &sh->state);
4179 return STRIPE_SECTORS;
4182 static int retry_aligned_read(struct r5conf *conf, struct bio *raid_bio)
4184 /* We may not be able to submit a whole bio at once as there
4185 * may not be enough stripe_heads available.
4186 * We cannot pre-allocate enough stripe_heads as we may need
4187 * more than exist in the cache (if we allow ever large chunks).
4188 * So we do one stripe head at a time and record in
4189 * ->bi_hw_segments how many have been done.
4191 * We *know* that this entire raid_bio is in one chunk, so
4192 * it will be only one 'dd_idx' and only need one call to raid5_compute_sector.
4194 struct stripe_head *sh;
4196 sector_t sector, logical_sector, last_sector;
4201 logical_sector = raid_bio->bi_sector & ~((sector_t)STRIPE_SECTORS-1);
4202 sector = raid5_compute_sector(conf, logical_sector,
4204 last_sector = raid_bio->bi_sector + (raid_bio->bi_size>>9);
4206 for (; logical_sector < last_sector;
4207 logical_sector += STRIPE_SECTORS,
4208 sector += STRIPE_SECTORS,
4211 if (scnt < raid5_bi_hw_segments(raid_bio))
4212 /* already done this stripe */
4215 sh = get_active_stripe(conf, sector, 0, 1, 0);
4218 /* failed to get a stripe - must wait */
4219 raid5_set_bi_hw_segments(raid_bio, scnt);
4220 conf->retry_read_aligned = raid_bio;
4224 set_bit(R5_ReadError, &sh->dev[dd_idx].flags);
4225 if (!add_stripe_bio(sh, raid_bio, dd_idx, 0)) {
4227 raid5_set_bi_hw_segments(raid_bio, scnt);
4228 conf->retry_read_aligned = raid_bio;
4236 spin_lock_irq(&conf->device_lock);
4237 remaining = raid5_dec_bi_phys_segments(raid_bio);
4238 spin_unlock_irq(&conf->device_lock);
4240 bio_endio(raid_bio, 0);
4241 if (atomic_dec_and_test(&conf->active_aligned_reads))
4242 wake_up(&conf->wait_for_stripe);
4248 * This is our raid5 kernel thread.
4250 * We scan the hash table for stripes which can be handled now.
4251 * During the scan, completed stripes are saved for us by the interrupt
4252 * handler, so that they will not have to wait for our next wakeup.
4254 static void raid5d(struct mddev *mddev)
4256 struct stripe_head *sh;
4257 struct r5conf *conf = mddev->private;
4259 struct blk_plug plug;
4261 pr_debug("+++ raid5d active\n");
4263 md_check_recovery(mddev);
4265 blk_start_plug(&plug);
4267 spin_lock_irq(&conf->device_lock);
4271 if (atomic_read(&mddev->plug_cnt) == 0 &&
4272 !list_empty(&conf->bitmap_list)) {
4273 /* Now is a good time to flush some bitmap updates */
4275 spin_unlock_irq(&conf->device_lock);
4276 bitmap_unplug(mddev->bitmap);
4277 spin_lock_irq(&conf->device_lock);
4278 conf->seq_write = conf->seq_flush;
4279 activate_bit_delay(conf);
4281 if (atomic_read(&mddev->plug_cnt) == 0)
4282 raid5_activate_delayed(conf);
4284 while ((bio = remove_bio_from_retry(conf))) {
4286 spin_unlock_irq(&conf->device_lock);
4287 ok = retry_aligned_read(conf, bio);
4288 spin_lock_irq(&conf->device_lock);
4294 sh = __get_priority_stripe(conf);
4298 spin_unlock_irq(&conf->device_lock);
4305 if (mddev->flags & ~(1<<MD_CHANGE_PENDING))
4306 md_check_recovery(mddev);
4308 spin_lock_irq(&conf->device_lock);
4310 pr_debug("%d stripes handled\n", handled);
4312 spin_unlock_irq(&conf->device_lock);
4314 async_tx_issue_pending_all();
4315 blk_finish_plug(&plug);
4317 pr_debug("--- raid5d inactive\n");
4321 raid5_show_stripe_cache_size(struct mddev *mddev, char *page)
4323 struct r5conf *conf = mddev->private;
4325 return sprintf(page, "%d\n", conf->max_nr_stripes);
4331 raid5_set_cache_size(struct mddev *mddev, int size)
4333 struct r5conf *conf = mddev->private;
4336 if (size <= 16 || size > 32768)
4338 while (size < conf->max_nr_stripes) {
4339 if (drop_one_stripe(conf))
4340 conf->max_nr_stripes--;
4344 err = md_allow_write(mddev);
4347 while (size > conf->max_nr_stripes) {
4348 if (grow_one_stripe(conf))
4349 conf->max_nr_stripes++;
4354 EXPORT_SYMBOL(raid5_set_cache_size);
4357 raid5_store_stripe_cache_size(struct mddev *mddev, const char *page, size_t len)
4359 struct r5conf *conf = mddev->private;
4363 if (len >= PAGE_SIZE)
4368 if (strict_strtoul(page, 10, &new))
4370 err = raid5_set_cache_size(mddev, new);
4376 static struct md_sysfs_entry
4377 raid5_stripecache_size = __ATTR(stripe_cache_size, S_IRUGO | S_IWUSR,
4378 raid5_show_stripe_cache_size,
4379 raid5_store_stripe_cache_size);
4382 raid5_show_preread_threshold(struct mddev *mddev, char *page)
4384 struct r5conf *conf = mddev->private;
4386 return sprintf(page, "%d\n", conf->bypass_threshold);
4392 raid5_store_preread_threshold(struct mddev *mddev, const char *page, size_t len)
4394 struct r5conf *conf = mddev->private;
4396 if (len >= PAGE_SIZE)
4401 if (strict_strtoul(page, 10, &new))
4403 if (new > conf->max_nr_stripes)
4405 conf->bypass_threshold = new;
4409 static struct md_sysfs_entry
4410 raid5_preread_bypass_threshold = __ATTR(preread_bypass_threshold,
4412 raid5_show_preread_threshold,
4413 raid5_store_preread_threshold);
4416 stripe_cache_active_show(struct mddev *mddev, char *page)
4418 struct r5conf *conf = mddev->private;
4420 return sprintf(page, "%d\n", atomic_read(&conf->active_stripes));
4425 static struct md_sysfs_entry
4426 raid5_stripecache_active = __ATTR_RO(stripe_cache_active);
4428 static struct attribute *raid5_attrs[] = {
4429 &raid5_stripecache_size.attr,
4430 &raid5_stripecache_active.attr,
4431 &raid5_preread_bypass_threshold.attr,
4434 static struct attribute_group raid5_attrs_group = {
4436 .attrs = raid5_attrs,
4440 raid5_size(struct mddev *mddev, sector_t sectors, int raid_disks)
4442 struct r5conf *conf = mddev->private;
4445 sectors = mddev->dev_sectors;
4447 /* size is defined by the smallest of previous and new size */
4448 raid_disks = min(conf->raid_disks, conf->previous_raid_disks);
4450 sectors &= ~((sector_t)mddev->chunk_sectors - 1);
4451 sectors &= ~((sector_t)mddev->new_chunk_sectors - 1);
4452 return sectors * (raid_disks - conf->max_degraded);
4455 static void free_scratch_buffer(struct r5conf *conf, struct raid5_percpu *percpu)
4457 safe_put_page(percpu->spare_page);
4458 kfree(percpu->scribble);
4459 percpu->spare_page = NULL;
4460 percpu->scribble = NULL;
4463 static int alloc_scratch_buffer(struct r5conf *conf, struct raid5_percpu *percpu)
4465 if (conf->level == 6 && !percpu->spare_page)
4466 percpu->spare_page = alloc_page(GFP_KERNEL);
4467 if (!percpu->scribble)
4468 percpu->scribble = kmalloc(conf->scribble_len, GFP_KERNEL);
4470 if (!percpu->scribble || (conf->level == 6 && !percpu->spare_page)) {
4471 free_scratch_buffer(conf, percpu);
4478 static void raid5_free_percpu(struct r5conf *conf)
4485 #ifdef CONFIG_HOTPLUG_CPU
4486 unregister_cpu_notifier(&conf->cpu_notify);
4490 for_each_possible_cpu(cpu)
4491 free_scratch_buffer(conf, per_cpu_ptr(conf->percpu, cpu));
4494 free_percpu(conf->percpu);
4497 static void free_conf(struct r5conf *conf)
4499 shrink_stripes(conf);
4500 raid5_free_percpu(conf);
4502 kfree(conf->stripe_hashtbl);
4506 #ifdef CONFIG_HOTPLUG_CPU
4507 static int raid456_cpu_notify(struct notifier_block *nfb, unsigned long action,
4510 struct r5conf *conf = container_of(nfb, struct r5conf, cpu_notify);
4511 long cpu = (long)hcpu;
4512 struct raid5_percpu *percpu = per_cpu_ptr(conf->percpu, cpu);
4515 case CPU_UP_PREPARE:
4516 case CPU_UP_PREPARE_FROZEN:
4517 if (alloc_scratch_buffer(conf, percpu)) {
4518 pr_err("%s: failed memory allocation for cpu%ld\n",
4520 return notifier_from_errno(-ENOMEM);
4524 case CPU_DEAD_FROZEN:
4525 free_scratch_buffer(conf, per_cpu_ptr(conf->percpu, cpu));
4534 static int raid5_alloc_percpu(struct r5conf *conf)
4539 conf->percpu = alloc_percpu(struct raid5_percpu);
4543 #ifdef CONFIG_HOTPLUG_CPU
4544 conf->cpu_notify.notifier_call = raid456_cpu_notify;
4545 conf->cpu_notify.priority = 0;
4546 err = register_cpu_notifier(&conf->cpu_notify);
4552 for_each_present_cpu(cpu) {
4553 err = alloc_scratch_buffer(conf, per_cpu_ptr(conf->percpu, cpu));
4555 pr_err("%s: failed memory allocation for cpu%ld\n",
4565 static struct r5conf *setup_conf(struct mddev *mddev)
4567 struct r5conf *conf;
4568 int raid_disk, memory, max_disks;
4569 struct md_rdev *rdev;
4570 struct disk_info *disk;
4572 if (mddev->new_level != 5
4573 && mddev->new_level != 4
4574 && mddev->new_level != 6) {
4575 printk(KERN_ERR "md/raid:%s: raid level not set to 4/5/6 (%d)\n",
4576 mdname(mddev), mddev->new_level);
4577 return ERR_PTR(-EIO);
4579 if ((mddev->new_level == 5
4580 && !algorithm_valid_raid5(mddev->new_layout)) ||
4581 (mddev->new_level == 6
4582 && !algorithm_valid_raid6(mddev->new_layout))) {
4583 printk(KERN_ERR "md/raid:%s: layout %d not supported\n",
4584 mdname(mddev), mddev->new_layout);
4585 return ERR_PTR(-EIO);
4587 if (mddev->new_level == 6 && mddev->raid_disks < 4) {
4588 printk(KERN_ERR "md/raid:%s: not enough configured devices (%d, minimum 4)\n",
4589 mdname(mddev), mddev->raid_disks);
4590 return ERR_PTR(-EINVAL);
4593 if (!mddev->new_chunk_sectors ||
4594 (mddev->new_chunk_sectors << 9) % PAGE_SIZE ||
4595 !is_power_of_2(mddev->new_chunk_sectors)) {
4596 printk(KERN_ERR "md/raid:%s: invalid chunk size %d\n",
4597 mdname(mddev), mddev->new_chunk_sectors << 9);
4598 return ERR_PTR(-EINVAL);
4601 conf = kzalloc(sizeof(struct r5conf), GFP_KERNEL);
4604 spin_lock_init(&conf->device_lock);
4605 init_waitqueue_head(&conf->wait_for_stripe);
4606 init_waitqueue_head(&conf->wait_for_overlap);
4607 INIT_LIST_HEAD(&conf->handle_list);
4608 INIT_LIST_HEAD(&conf->hold_list);
4609 INIT_LIST_HEAD(&conf->delayed_list);
4610 INIT_LIST_HEAD(&conf->bitmap_list);
4611 INIT_LIST_HEAD(&conf->inactive_list);
4612 atomic_set(&conf->active_stripes, 0);
4613 atomic_set(&conf->preread_active_stripes, 0);
4614 atomic_set(&conf->active_aligned_reads, 0);
4615 conf->bypass_threshold = BYPASS_THRESHOLD;
4616 conf->recovery_disabled = mddev->recovery_disabled - 1;
4618 conf->raid_disks = mddev->raid_disks;
4619 if (mddev->reshape_position == MaxSector)
4620 conf->previous_raid_disks = mddev->raid_disks;
4622 conf->previous_raid_disks = mddev->raid_disks - mddev->delta_disks;
4623 max_disks = max(conf->raid_disks, conf->previous_raid_disks);
4624 conf->scribble_len = scribble_len(max_disks);
4626 conf->disks = kzalloc(max_disks * sizeof(struct disk_info),
4631 conf->mddev = mddev;
4633 if ((conf->stripe_hashtbl = kzalloc(PAGE_SIZE, GFP_KERNEL)) == NULL)
4636 conf->level = mddev->new_level;
4637 if (raid5_alloc_percpu(conf) != 0)
4640 pr_debug("raid456: run(%s) called.\n", mdname(mddev));
4642 list_for_each_entry(rdev, &mddev->disks, same_set) {
4643 raid_disk = rdev->raid_disk;
4644 if (raid_disk >= max_disks
4647 disk = conf->disks + raid_disk;
4651 if (test_bit(In_sync, &rdev->flags)) {
4652 char b[BDEVNAME_SIZE];
4653 printk(KERN_INFO "md/raid:%s: device %s operational as raid"
4655 mdname(mddev), bdevname(rdev->bdev, b), raid_disk);
4656 } else if (rdev->saved_raid_disk != raid_disk)
4657 /* Cannot rely on bitmap to complete recovery */
4661 conf->chunk_sectors = mddev->new_chunk_sectors;
4662 conf->level = mddev->new_level;
4663 if (conf->level == 6)
4664 conf->max_degraded = 2;
4666 conf->max_degraded = 1;
4667 conf->algorithm = mddev->new_layout;
4668 conf->max_nr_stripes = NR_STRIPES;
4669 conf->reshape_progress = mddev->reshape_position;
4670 if (conf->reshape_progress != MaxSector) {
4671 conf->prev_chunk_sectors = mddev->chunk_sectors;
4672 conf->prev_algo = mddev->layout;
4675 memory = conf->max_nr_stripes * (sizeof(struct stripe_head) +
4676 max_disks * ((sizeof(struct bio) + PAGE_SIZE))) / 1024;
4677 if (grow_stripes(conf, conf->max_nr_stripes)) {
4679 "md/raid:%s: couldn't allocate %dkB for buffers\n",
4680 mdname(mddev), memory);
4683 printk(KERN_INFO "md/raid:%s: allocated %dkB\n",
4684 mdname(mddev), memory);
4686 conf->thread = md_register_thread(raid5d, mddev, NULL);
4687 if (!conf->thread) {
4689 "md/raid:%s: couldn't allocate thread.\n",
4699 return ERR_PTR(-EIO);
4701 return ERR_PTR(-ENOMEM);
4705 static int only_parity(int raid_disk, int algo, int raid_disks, int max_degraded)
4708 case ALGORITHM_PARITY_0:
4709 if (raid_disk < max_degraded)
4712 case ALGORITHM_PARITY_N:
4713 if (raid_disk >= raid_disks - max_degraded)
4716 case ALGORITHM_PARITY_0_6:
4717 if (raid_disk == 0 ||
4718 raid_disk == raid_disks - 1)
4721 case ALGORITHM_LEFT_ASYMMETRIC_6:
4722 case ALGORITHM_RIGHT_ASYMMETRIC_6:
4723 case ALGORITHM_LEFT_SYMMETRIC_6:
4724 case ALGORITHM_RIGHT_SYMMETRIC_6:
4725 if (raid_disk == raid_disks - 1)
4731 static int run(struct mddev *mddev)
4733 struct r5conf *conf;
4734 int working_disks = 0;
4735 int dirty_parity_disks = 0;
4736 struct md_rdev *rdev;
4737 sector_t reshape_offset = 0;
4739 if (mddev->recovery_cp != MaxSector)
4740 printk(KERN_NOTICE "md/raid:%s: not clean"
4741 " -- starting background reconstruction\n",
4743 if (mddev->reshape_position != MaxSector) {
4744 /* Check that we can continue the reshape.
4745 * Currently only disks can change, it must
4746 * increase, and we must be past the point where
4747 * a stripe over-writes itself
4749 sector_t here_new, here_old;
4751 int max_degraded = (mddev->level == 6 ? 2 : 1);
4753 if (mddev->new_level != mddev->level) {
4754 printk(KERN_ERR "md/raid:%s: unsupported reshape "
4755 "required - aborting.\n",
4759 old_disks = mddev->raid_disks - mddev->delta_disks;
4760 /* reshape_position must be on a new-stripe boundary, and one
4761 * further up in new geometry must map after here in old
4764 here_new = mddev->reshape_position;
4765 if (sector_div(here_new, mddev->new_chunk_sectors *
4766 (mddev->raid_disks - max_degraded))) {
4767 printk(KERN_ERR "md/raid:%s: reshape_position not "
4768 "on a stripe boundary\n", mdname(mddev));
4771 reshape_offset = here_new * mddev->new_chunk_sectors;
4772 /* here_new is the stripe we will write to */
4773 here_old = mddev->reshape_position;
4774 sector_div(here_old, mddev->chunk_sectors *
4775 (old_disks-max_degraded));
4776 /* here_old is the first stripe that we might need to read
4778 if (mddev->delta_disks == 0) {
4779 /* We cannot be sure it is safe to start an in-place
4780 * reshape. It is only safe if user-space if monitoring
4781 * and taking constant backups.
4782 * mdadm always starts a situation like this in
4783 * readonly mode so it can take control before
4784 * allowing any writes. So just check for that.
4786 if ((here_new * mddev->new_chunk_sectors !=
4787 here_old * mddev->chunk_sectors) ||
4789 printk(KERN_ERR "md/raid:%s: in-place reshape must be started"
4790 " in read-only mode - aborting\n",
4794 } else if (mddev->delta_disks < 0
4795 ? (here_new * mddev->new_chunk_sectors <=
4796 here_old * mddev->chunk_sectors)
4797 : (here_new * mddev->new_chunk_sectors >=
4798 here_old * mddev->chunk_sectors)) {
4799 /* Reading from the same stripe as writing to - bad */
4800 printk(KERN_ERR "md/raid:%s: reshape_position too early for "
4801 "auto-recovery - aborting.\n",
4805 printk(KERN_INFO "md/raid:%s: reshape will continue\n",
4807 /* OK, we should be able to continue; */
4809 BUG_ON(mddev->level != mddev->new_level);
4810 BUG_ON(mddev->layout != mddev->new_layout);
4811 BUG_ON(mddev->chunk_sectors != mddev->new_chunk_sectors);
4812 BUG_ON(mddev->delta_disks != 0);
4815 if (mddev->private == NULL)
4816 conf = setup_conf(mddev);
4818 conf = mddev->private;
4821 return PTR_ERR(conf);
4823 mddev->thread = conf->thread;
4824 conf->thread = NULL;
4825 mddev->private = conf;
4828 * 0 for a fully functional array, 1 or 2 for a degraded array.
4830 list_for_each_entry(rdev, &mddev->disks, same_set) {
4831 if (rdev->raid_disk < 0)
4833 if (test_bit(In_sync, &rdev->flags)) {
4837 /* This disc is not fully in-sync. However if it
4838 * just stored parity (beyond the recovery_offset),
4839 * when we don't need to be concerned about the
4840 * array being dirty.
4841 * When reshape goes 'backwards', we never have
4842 * partially completed devices, so we only need
4843 * to worry about reshape going forwards.
4845 /* Hack because v0.91 doesn't store recovery_offset properly. */
4846 if (mddev->major_version == 0 &&
4847 mddev->minor_version > 90)
4848 rdev->recovery_offset = reshape_offset;
4850 if (rdev->recovery_offset < reshape_offset) {
4851 /* We need to check old and new layout */
4852 if (!only_parity(rdev->raid_disk,
4855 conf->max_degraded))
4858 if (!only_parity(rdev->raid_disk,
4860 conf->previous_raid_disks,
4861 conf->max_degraded))
4863 dirty_parity_disks++;
4866 mddev->degraded = (max(conf->raid_disks, conf->previous_raid_disks)
4869 if (has_failed(conf)) {
4870 printk(KERN_ERR "md/raid:%s: not enough operational devices"
4871 " (%d/%d failed)\n",
4872 mdname(mddev), mddev->degraded, conf->raid_disks);
4876 /* device size must be a multiple of chunk size */
4877 mddev->dev_sectors &= ~(mddev->chunk_sectors - 1);
4878 mddev->resync_max_sectors = mddev->dev_sectors;
4880 if (mddev->degraded > dirty_parity_disks &&
4881 mddev->recovery_cp != MaxSector) {
4882 if (mddev->ok_start_degraded)
4884 "md/raid:%s: starting dirty degraded array"
4885 " - data corruption possible.\n",
4889 "md/raid:%s: cannot start dirty degraded array.\n",
4895 if (mddev->degraded == 0)
4896 printk(KERN_INFO "md/raid:%s: raid level %d active with %d out of %d"
4897 " devices, algorithm %d\n", mdname(mddev), conf->level,
4898 mddev->raid_disks-mddev->degraded, mddev->raid_disks,
4901 printk(KERN_ALERT "md/raid:%s: raid level %d active with %d"
4902 " out of %d devices, algorithm %d\n",
4903 mdname(mddev), conf->level,
4904 mddev->raid_disks - mddev->degraded,
4905 mddev->raid_disks, mddev->new_layout);
4907 print_raid5_conf(conf);
4909 if (conf->reshape_progress != MaxSector) {
4910 conf->reshape_safe = conf->reshape_progress;
4911 atomic_set(&conf->reshape_stripes, 0);
4912 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
4913 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
4914 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
4915 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
4916 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
4921 /* Ok, everything is just fine now */
4922 if (mddev->to_remove == &raid5_attrs_group)
4923 mddev->to_remove = NULL;
4924 else if (mddev->kobj.sd &&
4925 sysfs_create_group(&mddev->kobj, &raid5_attrs_group))
4927 "raid5: failed to create sysfs attributes for %s\n",
4929 md_set_array_sectors(mddev, raid5_size(mddev, 0, 0));
4933 /* read-ahead size must cover two whole stripes, which
4934 * is 2 * (datadisks) * chunksize where 'n' is the
4935 * number of raid devices
4937 int data_disks = conf->previous_raid_disks - conf->max_degraded;
4938 int stripe = data_disks *
4939 ((mddev->chunk_sectors << 9) / PAGE_SIZE);
4940 if (mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
4941 mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
4943 blk_queue_merge_bvec(mddev->queue, raid5_mergeable_bvec);
4945 mddev->queue->backing_dev_info.congested_data = mddev;
4946 mddev->queue->backing_dev_info.congested_fn = raid5_congested;
4948 chunk_size = mddev->chunk_sectors << 9;
4949 blk_queue_io_min(mddev->queue, chunk_size);
4950 blk_queue_io_opt(mddev->queue, chunk_size *
4951 (conf->raid_disks - conf->max_degraded));
4953 list_for_each_entry(rdev, &mddev->disks, same_set)
4954 disk_stack_limits(mddev->gendisk, rdev->bdev,
4955 rdev->data_offset << 9);
4960 md_unregister_thread(&mddev->thread);
4961 print_raid5_conf(conf);
4963 mddev->private = NULL;
4964 printk(KERN_ALERT "md/raid:%s: failed to run raid set.\n", mdname(mddev));
4968 static int stop(struct mddev *mddev)
4970 struct r5conf *conf = mddev->private;
4972 md_unregister_thread(&mddev->thread);
4974 mddev->queue->backing_dev_info.congested_fn = NULL;
4976 mddev->private = NULL;
4977 mddev->to_remove = &raid5_attrs_group;
4981 static void status(struct seq_file *seq, struct mddev *mddev)
4983 struct r5conf *conf = mddev->private;
4986 seq_printf(seq, " level %d, %dk chunk, algorithm %d", mddev->level,
4987 mddev->chunk_sectors / 2, mddev->layout);
4988 seq_printf (seq, " [%d/%d] [", conf->raid_disks, conf->raid_disks - mddev->degraded);
4989 for (i = 0; i < conf->raid_disks; i++)
4990 seq_printf (seq, "%s",
4991 conf->disks[i].rdev &&
4992 test_bit(In_sync, &conf->disks[i].rdev->flags) ? "U" : "_");
4993 seq_printf (seq, "]");
4996 static void print_raid5_conf (struct r5conf *conf)
4999 struct disk_info *tmp;
5001 printk(KERN_DEBUG "RAID conf printout:\n");
5003 printk("(conf==NULL)\n");
5006 printk(KERN_DEBUG " --- level:%d rd:%d wd:%d\n", conf->level,
5008 conf->raid_disks - conf->mddev->degraded);
5010 for (i = 0; i < conf->raid_disks; i++) {
5011 char b[BDEVNAME_SIZE];
5012 tmp = conf->disks + i;
5014 printk(KERN_DEBUG " disk %d, o:%d, dev:%s\n",
5015 i, !test_bit(Faulty, &tmp->rdev->flags),
5016 bdevname(tmp->rdev->bdev, b));
5020 static int raid5_spare_active(struct mddev *mddev)
5023 struct r5conf *conf = mddev->private;
5024 struct disk_info *tmp;
5026 unsigned long flags;
5028 for (i = 0; i < conf->raid_disks; i++) {
5029 tmp = conf->disks + i;
5031 && tmp->rdev->recovery_offset == MaxSector
5032 && !test_bit(Faulty, &tmp->rdev->flags)
5033 && !test_and_set_bit(In_sync, &tmp->rdev->flags)) {
5035 sysfs_notify_dirent_safe(tmp->rdev->sysfs_state);
5038 spin_lock_irqsave(&conf->device_lock, flags);
5039 mddev->degraded -= count;
5040 spin_unlock_irqrestore(&conf->device_lock, flags);
5041 print_raid5_conf(conf);
5045 static int raid5_remove_disk(struct mddev *mddev, int number)
5047 struct r5conf *conf = mddev->private;
5049 struct md_rdev *rdev;
5050 struct disk_info *p = conf->disks + number;
5052 print_raid5_conf(conf);
5055 if (number >= conf->raid_disks &&
5056 conf->reshape_progress == MaxSector)
5057 clear_bit(In_sync, &rdev->flags);
5059 if (test_bit(In_sync, &rdev->flags) ||
5060 atomic_read(&rdev->nr_pending)) {
5064 /* Only remove non-faulty devices if recovery
5067 if (!test_bit(Faulty, &rdev->flags) &&
5068 mddev->recovery_disabled != conf->recovery_disabled &&
5069 !has_failed(conf) &&
5070 number < conf->raid_disks) {
5076 if (atomic_read(&rdev->nr_pending)) {
5077 /* lost the race, try later */
5084 print_raid5_conf(conf);
5088 static int raid5_add_disk(struct mddev *mddev, struct md_rdev *rdev)
5090 struct r5conf *conf = mddev->private;
5093 struct disk_info *p;
5095 int last = conf->raid_disks - 1;
5097 if (mddev->recovery_disabled == conf->recovery_disabled)
5100 if (has_failed(conf))
5101 /* no point adding a device */
5104 if (rdev->raid_disk >= 0)
5105 first = last = rdev->raid_disk;
5108 * find the disk ... but prefer rdev->saved_raid_disk
5111 if (rdev->saved_raid_disk >= 0 &&
5112 rdev->saved_raid_disk >= first &&
5113 conf->disks[rdev->saved_raid_disk].rdev == NULL)
5114 disk = rdev->saved_raid_disk;
5117 for ( ; disk <= last ; disk++)
5118 if ((p=conf->disks + disk)->rdev == NULL) {
5119 clear_bit(In_sync, &rdev->flags);
5120 rdev->raid_disk = disk;
5122 if (rdev->saved_raid_disk != disk)
5124 rcu_assign_pointer(p->rdev, rdev);
5127 print_raid5_conf(conf);
5131 static int raid5_resize(struct mddev *mddev, sector_t sectors)
5133 /* no resync is happening, and there is enough space
5134 * on all devices, so we can resize.
5135 * We need to make sure resync covers any new space.
5136 * If the array is shrinking we should possibly wait until
5137 * any io in the removed space completes, but it hardly seems
5140 sectors &= ~((sector_t)mddev->chunk_sectors - 1);
5141 md_set_array_sectors(mddev, raid5_size(mddev, sectors,
5142 mddev->raid_disks));
5143 if (mddev->array_sectors >
5144 raid5_size(mddev, sectors, mddev->raid_disks))
5146 set_capacity(mddev->gendisk, mddev->array_sectors);
5147 revalidate_disk(mddev->gendisk);
5148 if (sectors > mddev->dev_sectors &&
5149 mddev->recovery_cp > mddev->dev_sectors) {
5150 mddev->recovery_cp = mddev->dev_sectors;
5151 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
5153 mddev->dev_sectors = sectors;
5154 mddev->resync_max_sectors = sectors;
5158 static int check_stripe_cache(struct mddev *mddev)
5160 /* Can only proceed if there are plenty of stripe_heads.
5161 * We need a minimum of one full stripe,, and for sensible progress
5162 * it is best to have about 4 times that.
5163 * If we require 4 times, then the default 256 4K stripe_heads will
5164 * allow for chunk sizes up to 256K, which is probably OK.
5165 * If the chunk size is greater, user-space should request more
5166 * stripe_heads first.
5168 struct r5conf *conf = mddev->private;
5169 if (((mddev->chunk_sectors << 9) / STRIPE_SIZE) * 4
5170 > conf->max_nr_stripes ||
5171 ((mddev->new_chunk_sectors << 9) / STRIPE_SIZE) * 4
5172 > conf->max_nr_stripes) {
5173 printk(KERN_WARNING "md/raid:%s: reshape: not enough stripes. Needed %lu\n",
5175 ((max(mddev->chunk_sectors, mddev->new_chunk_sectors) << 9)
5182 static int check_reshape(struct mddev *mddev)
5184 struct r5conf *conf = mddev->private;
5186 if (mddev->delta_disks == 0 &&
5187 mddev->new_layout == mddev->layout &&
5188 mddev->new_chunk_sectors == mddev->chunk_sectors)
5189 return 0; /* nothing to do */
5191 /* Cannot grow a bitmap yet */
5193 if (has_failed(conf))
5195 if (mddev->delta_disks < 0) {
5196 /* We might be able to shrink, but the devices must
5197 * be made bigger first.
5198 * For raid6, 4 is the minimum size.
5199 * Otherwise 2 is the minimum
5202 if (mddev->level == 6)
5204 if (mddev->raid_disks + mddev->delta_disks < min)
5208 if (!check_stripe_cache(mddev))
5211 return resize_stripes(conf, conf->raid_disks + mddev->delta_disks);
5214 static int raid5_start_reshape(struct mddev *mddev)
5216 struct r5conf *conf = mddev->private;
5217 struct md_rdev *rdev;
5219 unsigned long flags;
5221 if (test_bit(MD_RECOVERY_RUNNING, &mddev->recovery))
5224 if (!check_stripe_cache(mddev))
5227 list_for_each_entry(rdev, &mddev->disks, same_set)
5228 if (!test_bit(In_sync, &rdev->flags)
5229 && !test_bit(Faulty, &rdev->flags))
5232 if (spares - mddev->degraded < mddev->delta_disks - conf->max_degraded)
5233 /* Not enough devices even to make a degraded array
5238 /* Refuse to reduce size of the array. Any reductions in
5239 * array size must be through explicit setting of array_size
5242 if (raid5_size(mddev, 0, conf->raid_disks + mddev->delta_disks)
5243 < mddev->array_sectors) {
5244 printk(KERN_ERR "md/raid:%s: array size must be reduced "
5245 "before number of disks\n", mdname(mddev));
5249 atomic_set(&conf->reshape_stripes, 0);
5250 spin_lock_irq(&conf->device_lock);
5251 conf->previous_raid_disks = conf->raid_disks;
5252 conf->raid_disks += mddev->delta_disks;
5253 conf->prev_chunk_sectors = conf->chunk_sectors;
5254 conf->chunk_sectors = mddev->new_chunk_sectors;
5255 conf->prev_algo = conf->algorithm;
5256 conf->algorithm = mddev->new_layout;
5257 if (mddev->delta_disks < 0)
5258 conf->reshape_progress = raid5_size(mddev, 0, 0);
5260 conf->reshape_progress = 0;
5261 conf->reshape_safe = conf->reshape_progress;
5263 spin_unlock_irq(&conf->device_lock);
5265 /* Add some new drives, as many as will fit.
5266 * We know there are enough to make the newly sized array work.
5267 * Don't add devices if we are reducing the number of
5268 * devices in the array. This is because it is not possible
5269 * to correctly record the "partially reconstructed" state of
5270 * such devices during the reshape and confusion could result.
5272 if (mddev->delta_disks >= 0) {
5273 int added_devices = 0;
5274 list_for_each_entry(rdev, &mddev->disks, same_set)
5275 if (rdev->raid_disk < 0 &&
5276 !test_bit(Faulty, &rdev->flags)) {
5277 if (raid5_add_disk(mddev, rdev) == 0) {
5279 >= conf->previous_raid_disks) {
5280 set_bit(In_sync, &rdev->flags);
5283 rdev->recovery_offset = 0;
5285 if (sysfs_link_rdev(mddev, rdev))
5286 /* Failure here is OK */;
5288 } else if (rdev->raid_disk >= conf->previous_raid_disks
5289 && !test_bit(Faulty, &rdev->flags)) {
5290 /* This is a spare that was manually added */
5291 set_bit(In_sync, &rdev->flags);
5295 /* When a reshape changes the number of devices,
5296 * ->degraded is measured against the larger of the
5297 * pre and post number of devices.
5299 spin_lock_irqsave(&conf->device_lock, flags);
5300 mddev->degraded += (conf->raid_disks - conf->previous_raid_disks)
5302 spin_unlock_irqrestore(&conf->device_lock, flags);
5304 mddev->raid_disks = conf->raid_disks;
5305 mddev->reshape_position = conf->reshape_progress;
5306 set_bit(MD_CHANGE_DEVS, &mddev->flags);
5308 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
5309 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
5310 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
5311 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
5312 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
5314 if (!mddev->sync_thread) {
5315 mddev->recovery = 0;
5316 spin_lock_irq(&conf->device_lock);
5317 mddev->raid_disks = conf->raid_disks = conf->previous_raid_disks;
5318 conf->reshape_progress = MaxSector;
5319 spin_unlock_irq(&conf->device_lock);
5322 conf->reshape_checkpoint = jiffies;
5323 md_wakeup_thread(mddev->sync_thread);
5324 md_new_event(mddev);
5328 /* This is called from the reshape thread and should make any
5329 * changes needed in 'conf'
5331 static void end_reshape(struct r5conf *conf)
5334 if (!test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery)) {
5336 spin_lock_irq(&conf->device_lock);
5337 conf->previous_raid_disks = conf->raid_disks;
5338 conf->reshape_progress = MaxSector;
5339 spin_unlock_irq(&conf->device_lock);
5340 wake_up(&conf->wait_for_overlap);
5342 /* read-ahead size must cover two whole stripes, which is
5343 * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
5345 if (conf->mddev->queue) {
5346 int data_disks = conf->raid_disks - conf->max_degraded;
5347 int stripe = data_disks * ((conf->chunk_sectors << 9)
5349 if (conf->mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
5350 conf->mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
5355 /* This is called from the raid5d thread with mddev_lock held.
5356 * It makes config changes to the device.
5358 static void raid5_finish_reshape(struct mddev *mddev)
5360 struct r5conf *conf = mddev->private;
5362 if (!test_bit(MD_RECOVERY_INTR, &mddev->recovery)) {
5364 if (mddev->delta_disks > 0) {
5365 md_set_array_sectors(mddev, raid5_size(mddev, 0, 0));
5366 set_capacity(mddev->gendisk, mddev->array_sectors);
5367 revalidate_disk(mddev->gendisk);
5370 mddev->degraded = conf->raid_disks;
5371 for (d = 0; d < conf->raid_disks ; d++)
5372 if (conf->disks[d].rdev &&
5374 &conf->disks[d].rdev->flags))
5376 for (d = conf->raid_disks ;
5377 d < conf->raid_disks - mddev->delta_disks;
5379 struct md_rdev *rdev = conf->disks[d].rdev;
5380 if (rdev && raid5_remove_disk(mddev, d) == 0) {
5381 sysfs_unlink_rdev(mddev, rdev);
5382 rdev->raid_disk = -1;
5386 mddev->layout = conf->algorithm;
5387 mddev->chunk_sectors = conf->chunk_sectors;
5388 mddev->reshape_position = MaxSector;
5389 mddev->delta_disks = 0;
5393 static void raid5_quiesce(struct mddev *mddev, int state)
5395 struct r5conf *conf = mddev->private;
5398 case 2: /* resume for a suspend */
5399 wake_up(&conf->wait_for_overlap);
5402 case 1: /* stop all writes */
5403 spin_lock_irq(&conf->device_lock);
5404 /* '2' tells resync/reshape to pause so that all
5405 * active stripes can drain
5408 wait_event_lock_irq(conf->wait_for_stripe,
5409 atomic_read(&conf->active_stripes) == 0 &&
5410 atomic_read(&conf->active_aligned_reads) == 0,
5411 conf->device_lock, /* nothing */);
5413 spin_unlock_irq(&conf->device_lock);
5414 /* allow reshape to continue */
5415 wake_up(&conf->wait_for_overlap);
5418 case 0: /* re-enable writes */
5419 spin_lock_irq(&conf->device_lock);
5421 wake_up(&conf->wait_for_stripe);
5422 wake_up(&conf->wait_for_overlap);
5423 spin_unlock_irq(&conf->device_lock);
5429 static void *raid45_takeover_raid0(struct mddev *mddev, int level)
5431 struct r0conf *raid0_conf = mddev->private;
5434 /* for raid0 takeover only one zone is supported */
5435 if (raid0_conf->nr_strip_zones > 1) {
5436 printk(KERN_ERR "md/raid:%s: cannot takeover raid0 with more than one zone.\n",
5438 return ERR_PTR(-EINVAL);
5441 sectors = raid0_conf->strip_zone[0].zone_end;
5442 sector_div(sectors, raid0_conf->strip_zone[0].nb_dev);
5443 mddev->dev_sectors = sectors;
5444 mddev->new_level = level;
5445 mddev->new_layout = ALGORITHM_PARITY_N;
5446 mddev->new_chunk_sectors = mddev->chunk_sectors;
5447 mddev->raid_disks += 1;
5448 mddev->delta_disks = 1;
5449 /* make sure it will be not marked as dirty */
5450 mddev->recovery_cp = MaxSector;
5452 return setup_conf(mddev);
5456 static void *raid5_takeover_raid1(struct mddev *mddev)
5460 if (mddev->raid_disks != 2 ||
5461 mddev->degraded > 1)
5462 return ERR_PTR(-EINVAL);
5464 /* Should check if there are write-behind devices? */
5466 chunksect = 64*2; /* 64K by default */
5468 /* The array must be an exact multiple of chunksize */
5469 while (chunksect && (mddev->array_sectors & (chunksect-1)))
5472 if ((chunksect<<9) < STRIPE_SIZE)
5473 /* array size does not allow a suitable chunk size */
5474 return ERR_PTR(-EINVAL);
5476 mddev->new_level = 5;
5477 mddev->new_layout = ALGORITHM_LEFT_SYMMETRIC;
5478 mddev->new_chunk_sectors = chunksect;
5480 return setup_conf(mddev);
5483 static void *raid5_takeover_raid6(struct mddev *mddev)
5487 switch (mddev->layout) {
5488 case ALGORITHM_LEFT_ASYMMETRIC_6:
5489 new_layout = ALGORITHM_LEFT_ASYMMETRIC;
5491 case ALGORITHM_RIGHT_ASYMMETRIC_6:
5492 new_layout = ALGORITHM_RIGHT_ASYMMETRIC;
5494 case ALGORITHM_LEFT_SYMMETRIC_6:
5495 new_layout = ALGORITHM_LEFT_SYMMETRIC;
5497 case ALGORITHM_RIGHT_SYMMETRIC_6:
5498 new_layout = ALGORITHM_RIGHT_SYMMETRIC;
5500 case ALGORITHM_PARITY_0_6:
5501 new_layout = ALGORITHM_PARITY_0;
5503 case ALGORITHM_PARITY_N:
5504 new_layout = ALGORITHM_PARITY_N;
5507 return ERR_PTR(-EINVAL);
5509 mddev->new_level = 5;
5510 mddev->new_layout = new_layout;
5511 mddev->delta_disks = -1;
5512 mddev->raid_disks -= 1;
5513 return setup_conf(mddev);
5517 static int raid5_check_reshape(struct mddev *mddev)
5519 /* For a 2-drive array, the layout and chunk size can be changed
5520 * immediately as not restriping is needed.
5521 * For larger arrays we record the new value - after validation
5522 * to be used by a reshape pass.
5524 struct r5conf *conf = mddev->private;
5525 int new_chunk = mddev->new_chunk_sectors;
5527 if (mddev->new_layout >= 0 && !algorithm_valid_raid5(mddev->new_layout))
5529 if (new_chunk > 0) {
5530 if (!is_power_of_2(new_chunk))
5532 if (new_chunk < (PAGE_SIZE>>9))
5534 if (mddev->array_sectors & (new_chunk-1))
5535 /* not factor of array size */
5539 /* They look valid */
5541 if (mddev->raid_disks == 2) {
5542 /* can make the change immediately */
5543 if (mddev->new_layout >= 0) {
5544 conf->algorithm = mddev->new_layout;
5545 mddev->layout = mddev->new_layout;
5547 if (new_chunk > 0) {
5548 conf->chunk_sectors = new_chunk ;
5549 mddev->chunk_sectors = new_chunk;
5551 set_bit(MD_CHANGE_DEVS, &mddev->flags);
5552 md_wakeup_thread(mddev->thread);
5554 return check_reshape(mddev);
5557 static int raid6_check_reshape(struct mddev *mddev)
5559 int new_chunk = mddev->new_chunk_sectors;
5561 if (mddev->new_layout >= 0 && !algorithm_valid_raid6(mddev->new_layout))
5563 if (new_chunk > 0) {
5564 if (!is_power_of_2(new_chunk))
5566 if (new_chunk < (PAGE_SIZE >> 9))
5568 if (mddev->array_sectors & (new_chunk-1))
5569 /* not factor of array size */
5573 /* They look valid */
5574 return check_reshape(mddev);
5577 static void *raid5_takeover(struct mddev *mddev)
5579 /* raid5 can take over:
5580 * raid0 - if there is only one strip zone - make it a raid4 layout
5581 * raid1 - if there are two drives. We need to know the chunk size
5582 * raid4 - trivial - just use a raid4 layout.
5583 * raid6 - Providing it is a *_6 layout
5585 if (mddev->level == 0)
5586 return raid45_takeover_raid0(mddev, 5);
5587 if (mddev->level == 1)
5588 return raid5_takeover_raid1(mddev);
5589 if (mddev->level == 4) {
5590 mddev->new_layout = ALGORITHM_PARITY_N;
5591 mddev->new_level = 5;
5592 return setup_conf(mddev);
5594 if (mddev->level == 6)
5595 return raid5_takeover_raid6(mddev);
5597 return ERR_PTR(-EINVAL);
5600 static void *raid4_takeover(struct mddev *mddev)
5602 /* raid4 can take over:
5603 * raid0 - if there is only one strip zone
5604 * raid5 - if layout is right
5606 if (mddev->level == 0)
5607 return raid45_takeover_raid0(mddev, 4);
5608 if (mddev->level == 5 &&
5609 mddev->layout == ALGORITHM_PARITY_N) {
5610 mddev->new_layout = 0;
5611 mddev->new_level = 4;
5612 return setup_conf(mddev);
5614 return ERR_PTR(-EINVAL);
5617 static struct md_personality raid5_personality;
5619 static void *raid6_takeover(struct mddev *mddev)
5621 /* Currently can only take over a raid5. We map the
5622 * personality to an equivalent raid6 personality
5623 * with the Q block at the end.
5627 if (mddev->pers != &raid5_personality)
5628 return ERR_PTR(-EINVAL);
5629 if (mddev->degraded > 1)
5630 return ERR_PTR(-EINVAL);
5631 if (mddev->raid_disks > 253)
5632 return ERR_PTR(-EINVAL);
5633 if (mddev->raid_disks < 3)
5634 return ERR_PTR(-EINVAL);
5636 switch (mddev->layout) {
5637 case ALGORITHM_LEFT_ASYMMETRIC:
5638 new_layout = ALGORITHM_LEFT_ASYMMETRIC_6;
5640 case ALGORITHM_RIGHT_ASYMMETRIC:
5641 new_layout = ALGORITHM_RIGHT_ASYMMETRIC_6;
5643 case ALGORITHM_LEFT_SYMMETRIC:
5644 new_layout = ALGORITHM_LEFT_SYMMETRIC_6;
5646 case ALGORITHM_RIGHT_SYMMETRIC:
5647 new_layout = ALGORITHM_RIGHT_SYMMETRIC_6;
5649 case ALGORITHM_PARITY_0:
5650 new_layout = ALGORITHM_PARITY_0_6;
5652 case ALGORITHM_PARITY_N:
5653 new_layout = ALGORITHM_PARITY_N;
5656 return ERR_PTR(-EINVAL);
5658 mddev->new_level = 6;
5659 mddev->new_layout = new_layout;
5660 mddev->delta_disks = 1;
5661 mddev->raid_disks += 1;
5662 return setup_conf(mddev);
5666 static struct md_personality raid6_personality =
5670 .owner = THIS_MODULE,
5671 .make_request = make_request,
5675 .error_handler = error,
5676 .hot_add_disk = raid5_add_disk,
5677 .hot_remove_disk= raid5_remove_disk,
5678 .spare_active = raid5_spare_active,
5679 .sync_request = sync_request,
5680 .resize = raid5_resize,
5682 .check_reshape = raid6_check_reshape,
5683 .start_reshape = raid5_start_reshape,
5684 .finish_reshape = raid5_finish_reshape,
5685 .quiesce = raid5_quiesce,
5686 .takeover = raid6_takeover,
5688 static struct md_personality raid5_personality =
5692 .owner = THIS_MODULE,
5693 .make_request = make_request,
5697 .error_handler = error,
5698 .hot_add_disk = raid5_add_disk,
5699 .hot_remove_disk= raid5_remove_disk,
5700 .spare_active = raid5_spare_active,
5701 .sync_request = sync_request,
5702 .resize = raid5_resize,
5704 .check_reshape = raid5_check_reshape,
5705 .start_reshape = raid5_start_reshape,
5706 .finish_reshape = raid5_finish_reshape,
5707 .quiesce = raid5_quiesce,
5708 .takeover = raid5_takeover,
5711 static struct md_personality raid4_personality =
5715 .owner = THIS_MODULE,
5716 .make_request = make_request,
5720 .error_handler = error,
5721 .hot_add_disk = raid5_add_disk,
5722 .hot_remove_disk= raid5_remove_disk,
5723 .spare_active = raid5_spare_active,
5724 .sync_request = sync_request,
5725 .resize = raid5_resize,
5727 .check_reshape = raid5_check_reshape,
5728 .start_reshape = raid5_start_reshape,
5729 .finish_reshape = raid5_finish_reshape,
5730 .quiesce = raid5_quiesce,
5731 .takeover = raid4_takeover,
5734 static int __init raid5_init(void)
5736 register_md_personality(&raid6_personality);
5737 register_md_personality(&raid5_personality);
5738 register_md_personality(&raid4_personality);
5742 static void raid5_exit(void)
5744 unregister_md_personality(&raid6_personality);
5745 unregister_md_personality(&raid5_personality);
5746 unregister_md_personality(&raid4_personality);
5749 module_init(raid5_init);
5750 module_exit(raid5_exit);
5751 MODULE_LICENSE("GPL");
5752 MODULE_DESCRIPTION("RAID4/5/6 (striping with parity) personality for MD");
5753 MODULE_ALIAS("md-personality-4"); /* RAID5 */
5754 MODULE_ALIAS("md-raid5");
5755 MODULE_ALIAS("md-raid4");
5756 MODULE_ALIAS("md-level-5");
5757 MODULE_ALIAS("md-level-4");
5758 MODULE_ALIAS("md-personality-8"); /* RAID6 */
5759 MODULE_ALIAS("md-raid6");
5760 MODULE_ALIAS("md-level-6");
5762 /* This used to be two separate modules, they were: */
5763 MODULE_ALIAS("raid5");
5764 MODULE_ALIAS("raid6");