2 * raid5.c : Multiple Devices driver for Linux
3 * Copyright (C) 1996, 1997 Ingo Molnar, Miguel de Icaza, Gadi Oxman
4 * Copyright (C) 1999, 2000 Ingo Molnar
5 * Copyright (C) 2002, 2003 H. Peter Anvin
7 * RAID-4/5/6 management functions.
8 * Thanks to Penguin Computing for making the RAID-6 development possible
9 * by donating a test server!
11 * This program is free software; you can redistribute it and/or modify
12 * it under the terms of the GNU General Public License as published by
13 * the Free Software Foundation; either version 2, or (at your option)
16 * You should have received a copy of the GNU General Public License
17 * (for example /usr/src/linux/COPYING); if not, write to the Free
18 * Software Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
24 * The sequencing for updating the bitmap reliably is a little
25 * subtle (and I got it wrong the first time) so it deserves some
28 * We group bitmap updates into batches. Each batch has a number.
29 * We may write out several batches at once, but that isn't very important.
30 * conf->seq_write is the number of the last batch successfully written.
31 * conf->seq_flush is the number of the last batch that was closed to
33 * When we discover that we will need to write to any block in a stripe
34 * (in add_stripe_bio) we update the in-memory bitmap and record in sh->bm_seq
35 * the number of the batch it will be in. This is seq_flush+1.
36 * When we are ready to do a write, if that batch hasn't been written yet,
37 * we plug the array and queue the stripe for later.
38 * When an unplug happens, we increment bm_flush, thus closing the current
40 * When we notice that bm_flush > bm_write, we write out all pending updates
41 * to the bitmap, and advance bm_write to where bm_flush was.
42 * This may occasionally write a bit out twice, but is sure never to
46 #include <linux/blkdev.h>
47 #include <linux/kthread.h>
48 #include <linux/raid/pq.h>
49 #include <linux/async_tx.h>
50 #include <linux/async.h>
51 #include <linux/seq_file.h>
52 #include <linux/cpu.h>
53 #include <linux/slab.h>
54 #include <linux/ratelimit.h>
64 #define NR_STRIPES 256
65 #define STRIPE_SIZE PAGE_SIZE
66 #define STRIPE_SHIFT (PAGE_SHIFT - 9)
67 #define STRIPE_SECTORS (STRIPE_SIZE>>9)
68 #define IO_THRESHOLD 1
69 #define BYPASS_THRESHOLD 1
70 #define NR_HASH (PAGE_SIZE / sizeof(struct hlist_head))
71 #define HASH_MASK (NR_HASH - 1)
73 #define stripe_hash(conf, sect) (&((conf)->stripe_hashtbl[((sect) >> STRIPE_SHIFT) & HASH_MASK]))
75 /* bio's attached to a stripe+device for I/O are linked together in bi_sector
76 * order without overlap. There may be several bio's per stripe+device, and
77 * a bio could span several devices.
78 * When walking this list for a particular stripe+device, we must never proceed
79 * beyond a bio that extends past this device, as the next bio might no longer
81 * This macro is used to determine the 'next' bio in the list, given the sector
82 * of the current stripe+device
84 #define r5_next_bio(bio, sect) ( ( (bio)->bi_sector + ((bio)->bi_size>>9) < sect + STRIPE_SECTORS) ? (bio)->bi_next : NULL)
86 * The following can be used to debug the driver
88 #define RAID5_PARANOIA 1
89 #if RAID5_PARANOIA && defined(CONFIG_SMP)
90 # define CHECK_DEVLOCK() assert_spin_locked(&conf->device_lock)
92 # define CHECK_DEVLOCK()
101 * We maintain a biased count of active stripes in the bottom 16 bits of
102 * bi_phys_segments, and a count of processed stripes in the upper 16 bits
104 static inline int raid5_bi_phys_segments(struct bio *bio)
106 return bio->bi_phys_segments & 0xffff;
109 static inline int raid5_bi_hw_segments(struct bio *bio)
111 return (bio->bi_phys_segments >> 16) & 0xffff;
114 static inline int raid5_dec_bi_phys_segments(struct bio *bio)
116 --bio->bi_phys_segments;
117 return raid5_bi_phys_segments(bio);
120 static inline int raid5_dec_bi_hw_segments(struct bio *bio)
122 unsigned short val = raid5_bi_hw_segments(bio);
125 bio->bi_phys_segments = (val << 16) | raid5_bi_phys_segments(bio);
129 static inline void raid5_set_bi_hw_segments(struct bio *bio, unsigned int cnt)
131 bio->bi_phys_segments = raid5_bi_phys_segments(bio) | (cnt << 16);
134 /* Find first data disk in a raid6 stripe */
135 static inline int raid6_d0(struct stripe_head *sh)
138 /* ddf always start from first device */
140 /* md starts just after Q block */
141 if (sh->qd_idx == sh->disks - 1)
144 return sh->qd_idx + 1;
146 static inline int raid6_next_disk(int disk, int raid_disks)
149 return (disk < raid_disks) ? disk : 0;
152 /* When walking through the disks in a raid5, starting at raid6_d0,
153 * We need to map each disk to a 'slot', where the data disks are slot
154 * 0 .. raid_disks-3, the parity disk is raid_disks-2 and the Q disk
155 * is raid_disks-1. This help does that mapping.
157 static int raid6_idx_to_slot(int idx, struct stripe_head *sh,
158 int *count, int syndrome_disks)
164 if (idx == sh->pd_idx)
165 return syndrome_disks;
166 if (idx == sh->qd_idx)
167 return syndrome_disks + 1;
173 static void return_io(struct bio *return_bi)
175 struct bio *bi = return_bi;
178 return_bi = bi->bi_next;
186 static void print_raid5_conf (raid5_conf_t *conf);
188 static int stripe_operations_active(struct stripe_head *sh)
190 return sh->check_state || sh->reconstruct_state ||
191 test_bit(STRIPE_BIOFILL_RUN, &sh->state) ||
192 test_bit(STRIPE_COMPUTE_RUN, &sh->state);
195 static void __release_stripe(raid5_conf_t *conf, struct stripe_head *sh)
197 if (atomic_dec_and_test(&sh->count)) {
198 BUG_ON(!list_empty(&sh->lru));
199 BUG_ON(atomic_read(&conf->active_stripes)==0);
200 if (test_bit(STRIPE_HANDLE, &sh->state)) {
201 if (test_bit(STRIPE_DELAYED, &sh->state))
202 list_add_tail(&sh->lru, &conf->delayed_list);
203 else if (test_bit(STRIPE_BIT_DELAY, &sh->state) &&
204 sh->bm_seq - conf->seq_write > 0)
205 list_add_tail(&sh->lru, &conf->bitmap_list);
207 clear_bit(STRIPE_BIT_DELAY, &sh->state);
208 list_add_tail(&sh->lru, &conf->handle_list);
210 md_wakeup_thread(conf->mddev->thread);
212 BUG_ON(stripe_operations_active(sh));
213 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
214 atomic_dec(&conf->preread_active_stripes);
215 if (atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD)
216 md_wakeup_thread(conf->mddev->thread);
218 atomic_dec(&conf->active_stripes);
219 if (!test_bit(STRIPE_EXPANDING, &sh->state)) {
220 list_add_tail(&sh->lru, &conf->inactive_list);
221 wake_up(&conf->wait_for_stripe);
222 if (conf->retry_read_aligned)
223 md_wakeup_thread(conf->mddev->thread);
229 static void release_stripe(struct stripe_head *sh)
231 raid5_conf_t *conf = sh->raid_conf;
234 spin_lock_irqsave(&conf->device_lock, flags);
235 __release_stripe(conf, sh);
236 spin_unlock_irqrestore(&conf->device_lock, flags);
239 static inline void remove_hash(struct stripe_head *sh)
241 pr_debug("remove_hash(), stripe %llu\n",
242 (unsigned long long)sh->sector);
244 hlist_del_init(&sh->hash);
247 static inline void insert_hash(raid5_conf_t *conf, struct stripe_head *sh)
249 struct hlist_head *hp = stripe_hash(conf, sh->sector);
251 pr_debug("insert_hash(), stripe %llu\n",
252 (unsigned long long)sh->sector);
255 hlist_add_head(&sh->hash, hp);
259 /* find an idle stripe, make sure it is unhashed, and return it. */
260 static struct stripe_head *get_free_stripe(raid5_conf_t *conf)
262 struct stripe_head *sh = NULL;
263 struct list_head *first;
266 if (list_empty(&conf->inactive_list))
268 first = conf->inactive_list.next;
269 sh = list_entry(first, struct stripe_head, lru);
270 list_del_init(first);
272 atomic_inc(&conf->active_stripes);
277 static void shrink_buffers(struct stripe_head *sh)
281 int num = sh->raid_conf->pool_size;
283 for (i = 0; i < num ; i++) {
287 sh->dev[i].page = NULL;
292 static int grow_buffers(struct stripe_head *sh)
295 int num = sh->raid_conf->pool_size;
297 for (i = 0; i < num; i++) {
300 if (!(page = alloc_page(GFP_KERNEL))) {
303 sh->dev[i].page = page;
308 static void raid5_build_block(struct stripe_head *sh, int i, int previous);
309 static void stripe_set_idx(sector_t stripe, raid5_conf_t *conf, int previous,
310 struct stripe_head *sh);
312 static void init_stripe(struct stripe_head *sh, sector_t sector, int previous)
314 raid5_conf_t *conf = sh->raid_conf;
317 BUG_ON(atomic_read(&sh->count) != 0);
318 BUG_ON(test_bit(STRIPE_HANDLE, &sh->state));
319 BUG_ON(stripe_operations_active(sh));
322 pr_debug("init_stripe called, stripe %llu\n",
323 (unsigned long long)sh->sector);
327 sh->generation = conf->generation - previous;
328 sh->disks = previous ? conf->previous_raid_disks : conf->raid_disks;
330 stripe_set_idx(sector, conf, previous, sh);
334 for (i = sh->disks; i--; ) {
335 struct r5dev *dev = &sh->dev[i];
337 if (dev->toread || dev->read || dev->towrite || dev->written ||
338 test_bit(R5_LOCKED, &dev->flags)) {
339 printk(KERN_ERR "sector=%llx i=%d %p %p %p %p %d\n",
340 (unsigned long long)sh->sector, i, dev->toread,
341 dev->read, dev->towrite, dev->written,
342 test_bit(R5_LOCKED, &dev->flags));
346 raid5_build_block(sh, i, previous);
348 insert_hash(conf, sh);
351 static struct stripe_head *__find_stripe(raid5_conf_t *conf, sector_t sector,
354 struct stripe_head *sh;
355 struct hlist_node *hn;
358 pr_debug("__find_stripe, sector %llu\n", (unsigned long long)sector);
359 hlist_for_each_entry(sh, hn, stripe_hash(conf, sector), hash)
360 if (sh->sector == sector && sh->generation == generation)
362 pr_debug("__stripe %llu not in cache\n", (unsigned long long)sector);
367 * Need to check if array has failed when deciding whether to:
369 * - remove non-faulty devices
372 * This determination is simple when no reshape is happening.
373 * However if there is a reshape, we need to carefully check
374 * both the before and after sections.
375 * This is because some failed devices may only affect one
376 * of the two sections, and some non-in_sync devices may
377 * be insync in the section most affected by failed devices.
379 static int has_failed(raid5_conf_t *conf)
383 if (conf->mddev->reshape_position == MaxSector)
384 return conf->mddev->degraded > conf->max_degraded;
388 for (i = 0; i < conf->previous_raid_disks; i++) {
389 mdk_rdev_t *rdev = rcu_dereference(conf->disks[i].rdev);
390 if (!rdev || test_bit(Faulty, &rdev->flags))
392 else if (test_bit(In_sync, &rdev->flags))
395 /* not in-sync or faulty.
396 * If the reshape increases the number of devices,
397 * this is being recovered by the reshape, so
398 * this 'previous' section is not in_sync.
399 * If the number of devices is being reduced however,
400 * the device can only be part of the array if
401 * we are reverting a reshape, so this section will
404 if (conf->raid_disks >= conf->previous_raid_disks)
408 if (degraded > conf->max_degraded)
412 for (i = 0; i < conf->raid_disks; i++) {
413 mdk_rdev_t *rdev = rcu_dereference(conf->disks[i].rdev);
414 if (!rdev || test_bit(Faulty, &rdev->flags))
416 else if (test_bit(In_sync, &rdev->flags))
419 /* not in-sync or faulty.
420 * If reshape increases the number of devices, this
421 * section has already been recovered, else it
422 * almost certainly hasn't.
424 if (conf->raid_disks <= conf->previous_raid_disks)
428 if (degraded > conf->max_degraded)
433 static struct stripe_head *
434 get_active_stripe(raid5_conf_t *conf, sector_t sector,
435 int previous, int noblock, int noquiesce)
437 struct stripe_head *sh;
439 pr_debug("get_stripe, sector %llu\n", (unsigned long long)sector);
441 spin_lock_irq(&conf->device_lock);
444 wait_event_lock_irq(conf->wait_for_stripe,
445 conf->quiesce == 0 || noquiesce,
446 conf->device_lock, /* nothing */);
447 sh = __find_stripe(conf, sector, conf->generation - previous);
449 if (!conf->inactive_blocked)
450 sh = get_free_stripe(conf);
451 if (noblock && sh == NULL)
454 conf->inactive_blocked = 1;
455 wait_event_lock_irq(conf->wait_for_stripe,
456 !list_empty(&conf->inactive_list) &&
457 (atomic_read(&conf->active_stripes)
458 < (conf->max_nr_stripes *3/4)
459 || !conf->inactive_blocked),
462 conf->inactive_blocked = 0;
464 init_stripe(sh, sector, previous);
466 if (atomic_read(&sh->count)) {
467 BUG_ON(!list_empty(&sh->lru)
468 && !test_bit(STRIPE_EXPANDING, &sh->state));
470 if (!test_bit(STRIPE_HANDLE, &sh->state))
471 atomic_inc(&conf->active_stripes);
472 if (list_empty(&sh->lru) &&
473 !test_bit(STRIPE_EXPANDING, &sh->state))
475 list_del_init(&sh->lru);
478 } while (sh == NULL);
481 atomic_inc(&sh->count);
483 spin_unlock_irq(&conf->device_lock);
488 raid5_end_read_request(struct bio *bi, int error);
490 raid5_end_write_request(struct bio *bi, int error);
492 static void ops_run_io(struct stripe_head *sh, struct stripe_head_state *s)
494 raid5_conf_t *conf = sh->raid_conf;
495 int i, disks = sh->disks;
499 for (i = disks; i--; ) {
503 if (test_and_clear_bit(R5_Wantwrite, &sh->dev[i].flags)) {
504 if (test_and_clear_bit(R5_WantFUA, &sh->dev[i].flags))
508 } else if (test_and_clear_bit(R5_Wantread, &sh->dev[i].flags))
513 bi = &sh->dev[i].req;
517 bi->bi_end_io = raid5_end_write_request;
519 bi->bi_end_io = raid5_end_read_request;
522 rdev = rcu_dereference(conf->disks[i].rdev);
523 if (rdev && test_bit(Faulty, &rdev->flags))
526 atomic_inc(&rdev->nr_pending);
530 if (s->syncing || s->expanding || s->expanded)
531 md_sync_acct(rdev->bdev, STRIPE_SECTORS);
533 set_bit(STRIPE_IO_STARTED, &sh->state);
535 bi->bi_bdev = rdev->bdev;
536 pr_debug("%s: for %llu schedule op %ld on disc %d\n",
537 __func__, (unsigned long long)sh->sector,
539 atomic_inc(&sh->count);
540 bi->bi_sector = sh->sector + rdev->data_offset;
541 bi->bi_flags = 1 << BIO_UPTODATE;
545 bi->bi_io_vec = &sh->dev[i].vec;
546 bi->bi_io_vec[0].bv_len = STRIPE_SIZE;
547 bi->bi_io_vec[0].bv_offset = 0;
548 bi->bi_size = STRIPE_SIZE;
550 generic_make_request(bi);
553 set_bit(STRIPE_DEGRADED, &sh->state);
554 pr_debug("skip op %ld on disc %d for sector %llu\n",
555 bi->bi_rw, i, (unsigned long long)sh->sector);
556 clear_bit(R5_LOCKED, &sh->dev[i].flags);
557 set_bit(STRIPE_HANDLE, &sh->state);
562 static struct dma_async_tx_descriptor *
563 async_copy_data(int frombio, struct bio *bio, struct page *page,
564 sector_t sector, struct dma_async_tx_descriptor *tx)
567 struct page *bio_page;
570 struct async_submit_ctl submit;
571 enum async_tx_flags flags = 0;
573 if (bio->bi_sector >= sector)
574 page_offset = (signed)(bio->bi_sector - sector) * 512;
576 page_offset = (signed)(sector - bio->bi_sector) * -512;
579 flags |= ASYNC_TX_FENCE;
580 init_async_submit(&submit, flags, tx, NULL, NULL, NULL);
582 bio_for_each_segment(bvl, bio, i) {
583 int len = bvl->bv_len;
587 if (page_offset < 0) {
588 b_offset = -page_offset;
589 page_offset += b_offset;
593 if (len > 0 && page_offset + len > STRIPE_SIZE)
594 clen = STRIPE_SIZE - page_offset;
599 b_offset += bvl->bv_offset;
600 bio_page = bvl->bv_page;
602 tx = async_memcpy(page, bio_page, page_offset,
603 b_offset, clen, &submit);
605 tx = async_memcpy(bio_page, page, b_offset,
606 page_offset, clen, &submit);
608 /* chain the operations */
609 submit.depend_tx = tx;
611 if (clen < len) /* hit end of page */
619 static void ops_complete_biofill(void *stripe_head_ref)
621 struct stripe_head *sh = stripe_head_ref;
622 struct bio *return_bi = NULL;
623 raid5_conf_t *conf = sh->raid_conf;
626 pr_debug("%s: stripe %llu\n", __func__,
627 (unsigned long long)sh->sector);
629 /* clear completed biofills */
630 spin_lock_irq(&conf->device_lock);
631 for (i = sh->disks; i--; ) {
632 struct r5dev *dev = &sh->dev[i];
634 /* acknowledge completion of a biofill operation */
635 /* and check if we need to reply to a read request,
636 * new R5_Wantfill requests are held off until
637 * !STRIPE_BIOFILL_RUN
639 if (test_and_clear_bit(R5_Wantfill, &dev->flags)) {
640 struct bio *rbi, *rbi2;
645 while (rbi && rbi->bi_sector <
646 dev->sector + STRIPE_SECTORS) {
647 rbi2 = r5_next_bio(rbi, dev->sector);
648 if (!raid5_dec_bi_phys_segments(rbi)) {
649 rbi->bi_next = return_bi;
656 spin_unlock_irq(&conf->device_lock);
657 clear_bit(STRIPE_BIOFILL_RUN, &sh->state);
659 return_io(return_bi);
661 set_bit(STRIPE_HANDLE, &sh->state);
665 static void ops_run_biofill(struct stripe_head *sh)
667 struct dma_async_tx_descriptor *tx = NULL;
668 raid5_conf_t *conf = sh->raid_conf;
669 struct async_submit_ctl submit;
672 pr_debug("%s: stripe %llu\n", __func__,
673 (unsigned long long)sh->sector);
675 for (i = sh->disks; i--; ) {
676 struct r5dev *dev = &sh->dev[i];
677 if (test_bit(R5_Wantfill, &dev->flags)) {
679 spin_lock_irq(&conf->device_lock);
680 dev->read = rbi = dev->toread;
682 spin_unlock_irq(&conf->device_lock);
683 while (rbi && rbi->bi_sector <
684 dev->sector + STRIPE_SECTORS) {
685 tx = async_copy_data(0, rbi, dev->page,
687 rbi = r5_next_bio(rbi, dev->sector);
692 atomic_inc(&sh->count);
693 init_async_submit(&submit, ASYNC_TX_ACK, tx, ops_complete_biofill, sh, NULL);
694 async_trigger_callback(&submit);
697 static void mark_target_uptodate(struct stripe_head *sh, int target)
704 tgt = &sh->dev[target];
705 set_bit(R5_UPTODATE, &tgt->flags);
706 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
707 clear_bit(R5_Wantcompute, &tgt->flags);
710 static void ops_complete_compute(void *stripe_head_ref)
712 struct stripe_head *sh = stripe_head_ref;
714 pr_debug("%s: stripe %llu\n", __func__,
715 (unsigned long long)sh->sector);
717 /* mark the computed target(s) as uptodate */
718 mark_target_uptodate(sh, sh->ops.target);
719 mark_target_uptodate(sh, sh->ops.target2);
721 clear_bit(STRIPE_COMPUTE_RUN, &sh->state);
722 if (sh->check_state == check_state_compute_run)
723 sh->check_state = check_state_compute_result;
724 set_bit(STRIPE_HANDLE, &sh->state);
728 /* return a pointer to the address conversion region of the scribble buffer */
729 static addr_conv_t *to_addr_conv(struct stripe_head *sh,
730 struct raid5_percpu *percpu)
732 return percpu->scribble + sizeof(struct page *) * (sh->disks + 2);
735 static struct dma_async_tx_descriptor *
736 ops_run_compute5(struct stripe_head *sh, struct raid5_percpu *percpu)
738 int disks = sh->disks;
739 struct page **xor_srcs = percpu->scribble;
740 int target = sh->ops.target;
741 struct r5dev *tgt = &sh->dev[target];
742 struct page *xor_dest = tgt->page;
744 struct dma_async_tx_descriptor *tx;
745 struct async_submit_ctl submit;
748 pr_debug("%s: stripe %llu block: %d\n",
749 __func__, (unsigned long long)sh->sector, target);
750 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
752 for (i = disks; i--; )
754 xor_srcs[count++] = sh->dev[i].page;
756 atomic_inc(&sh->count);
758 init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST, NULL,
759 ops_complete_compute, sh, to_addr_conv(sh, percpu));
760 if (unlikely(count == 1))
761 tx = async_memcpy(xor_dest, xor_srcs[0], 0, 0, STRIPE_SIZE, &submit);
763 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
768 /* set_syndrome_sources - populate source buffers for gen_syndrome
769 * @srcs - (struct page *) array of size sh->disks
770 * @sh - stripe_head to parse
772 * Populates srcs in proper layout order for the stripe and returns the
773 * 'count' of sources to be used in a call to async_gen_syndrome. The P
774 * destination buffer is recorded in srcs[count] and the Q destination
775 * is recorded in srcs[count+1]].
777 static int set_syndrome_sources(struct page **srcs, struct stripe_head *sh)
779 int disks = sh->disks;
780 int syndrome_disks = sh->ddf_layout ? disks : (disks - 2);
781 int d0_idx = raid6_d0(sh);
785 for (i = 0; i < disks; i++)
791 int slot = raid6_idx_to_slot(i, sh, &count, syndrome_disks);
793 srcs[slot] = sh->dev[i].page;
794 i = raid6_next_disk(i, disks);
795 } while (i != d0_idx);
797 return syndrome_disks;
800 static struct dma_async_tx_descriptor *
801 ops_run_compute6_1(struct stripe_head *sh, struct raid5_percpu *percpu)
803 int disks = sh->disks;
804 struct page **blocks = percpu->scribble;
806 int qd_idx = sh->qd_idx;
807 struct dma_async_tx_descriptor *tx;
808 struct async_submit_ctl submit;
814 if (sh->ops.target < 0)
815 target = sh->ops.target2;
816 else if (sh->ops.target2 < 0)
817 target = sh->ops.target;
819 /* we should only have one valid target */
822 pr_debug("%s: stripe %llu block: %d\n",
823 __func__, (unsigned long long)sh->sector, target);
825 tgt = &sh->dev[target];
826 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
829 atomic_inc(&sh->count);
831 if (target == qd_idx) {
832 count = set_syndrome_sources(blocks, sh);
833 blocks[count] = NULL; /* regenerating p is not necessary */
834 BUG_ON(blocks[count+1] != dest); /* q should already be set */
835 init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
836 ops_complete_compute, sh,
837 to_addr_conv(sh, percpu));
838 tx = async_gen_syndrome(blocks, 0, count+2, STRIPE_SIZE, &submit);
840 /* Compute any data- or p-drive using XOR */
842 for (i = disks; i-- ; ) {
843 if (i == target || i == qd_idx)
845 blocks[count++] = sh->dev[i].page;
848 init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST,
849 NULL, ops_complete_compute, sh,
850 to_addr_conv(sh, percpu));
851 tx = async_xor(dest, blocks, 0, count, STRIPE_SIZE, &submit);
857 static struct dma_async_tx_descriptor *
858 ops_run_compute6_2(struct stripe_head *sh, struct raid5_percpu *percpu)
860 int i, count, disks = sh->disks;
861 int syndrome_disks = sh->ddf_layout ? disks : disks-2;
862 int d0_idx = raid6_d0(sh);
863 int faila = -1, failb = -1;
864 int target = sh->ops.target;
865 int target2 = sh->ops.target2;
866 struct r5dev *tgt = &sh->dev[target];
867 struct r5dev *tgt2 = &sh->dev[target2];
868 struct dma_async_tx_descriptor *tx;
869 struct page **blocks = percpu->scribble;
870 struct async_submit_ctl submit;
872 pr_debug("%s: stripe %llu block1: %d block2: %d\n",
873 __func__, (unsigned long long)sh->sector, target, target2);
874 BUG_ON(target < 0 || target2 < 0);
875 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
876 BUG_ON(!test_bit(R5_Wantcompute, &tgt2->flags));
878 /* we need to open-code set_syndrome_sources to handle the
879 * slot number conversion for 'faila' and 'failb'
881 for (i = 0; i < disks ; i++)
886 int slot = raid6_idx_to_slot(i, sh, &count, syndrome_disks);
888 blocks[slot] = sh->dev[i].page;
894 i = raid6_next_disk(i, disks);
895 } while (i != d0_idx);
897 BUG_ON(faila == failb);
900 pr_debug("%s: stripe: %llu faila: %d failb: %d\n",
901 __func__, (unsigned long long)sh->sector, faila, failb);
903 atomic_inc(&sh->count);
905 if (failb == syndrome_disks+1) {
906 /* Q disk is one of the missing disks */
907 if (faila == syndrome_disks) {
908 /* Missing P+Q, just recompute */
909 init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
910 ops_complete_compute, sh,
911 to_addr_conv(sh, percpu));
912 return async_gen_syndrome(blocks, 0, syndrome_disks+2,
913 STRIPE_SIZE, &submit);
917 int qd_idx = sh->qd_idx;
919 /* Missing D+Q: recompute D from P, then recompute Q */
920 if (target == qd_idx)
921 data_target = target2;
923 data_target = target;
926 for (i = disks; i-- ; ) {
927 if (i == data_target || i == qd_idx)
929 blocks[count++] = sh->dev[i].page;
931 dest = sh->dev[data_target].page;
932 init_async_submit(&submit,
933 ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST,
935 to_addr_conv(sh, percpu));
936 tx = async_xor(dest, blocks, 0, count, STRIPE_SIZE,
939 count = set_syndrome_sources(blocks, sh);
940 init_async_submit(&submit, ASYNC_TX_FENCE, tx,
941 ops_complete_compute, sh,
942 to_addr_conv(sh, percpu));
943 return async_gen_syndrome(blocks, 0, count+2,
944 STRIPE_SIZE, &submit);
947 init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
948 ops_complete_compute, sh,
949 to_addr_conv(sh, percpu));
950 if (failb == syndrome_disks) {
951 /* We're missing D+P. */
952 return async_raid6_datap_recov(syndrome_disks+2,
956 /* We're missing D+D. */
957 return async_raid6_2data_recov(syndrome_disks+2,
958 STRIPE_SIZE, faila, failb,
965 static void ops_complete_prexor(void *stripe_head_ref)
967 struct stripe_head *sh = stripe_head_ref;
969 pr_debug("%s: stripe %llu\n", __func__,
970 (unsigned long long)sh->sector);
973 static struct dma_async_tx_descriptor *
974 ops_run_prexor(struct stripe_head *sh, struct raid5_percpu *percpu,
975 struct dma_async_tx_descriptor *tx)
977 int disks = sh->disks;
978 struct page **xor_srcs = percpu->scribble;
979 int count = 0, pd_idx = sh->pd_idx, i;
980 struct async_submit_ctl submit;
982 /* existing parity data subtracted */
983 struct page *xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
985 pr_debug("%s: stripe %llu\n", __func__,
986 (unsigned long long)sh->sector);
988 for (i = disks; i--; ) {
989 struct r5dev *dev = &sh->dev[i];
990 /* Only process blocks that are known to be uptodate */
991 if (test_bit(R5_Wantdrain, &dev->flags))
992 xor_srcs[count++] = dev->page;
995 init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_DROP_DST, tx,
996 ops_complete_prexor, sh, to_addr_conv(sh, percpu));
997 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
1002 static struct dma_async_tx_descriptor *
1003 ops_run_biodrain(struct stripe_head *sh, struct dma_async_tx_descriptor *tx)
1005 int disks = sh->disks;
1008 pr_debug("%s: stripe %llu\n", __func__,
1009 (unsigned long long)sh->sector);
1011 for (i = disks; i--; ) {
1012 struct r5dev *dev = &sh->dev[i];
1015 if (test_and_clear_bit(R5_Wantdrain, &dev->flags)) {
1018 spin_lock_irq(&sh->raid_conf->device_lock);
1019 chosen = dev->towrite;
1020 dev->towrite = NULL;
1021 BUG_ON(dev->written);
1022 wbi = dev->written = chosen;
1023 spin_unlock_irq(&sh->raid_conf->device_lock);
1025 while (wbi && wbi->bi_sector <
1026 dev->sector + STRIPE_SECTORS) {
1027 if (wbi->bi_rw & REQ_FUA)
1028 set_bit(R5_WantFUA, &dev->flags);
1029 tx = async_copy_data(1, wbi, dev->page,
1031 wbi = r5_next_bio(wbi, dev->sector);
1039 static void ops_complete_reconstruct(void *stripe_head_ref)
1041 struct stripe_head *sh = stripe_head_ref;
1042 int disks = sh->disks;
1043 int pd_idx = sh->pd_idx;
1044 int qd_idx = sh->qd_idx;
1048 pr_debug("%s: stripe %llu\n", __func__,
1049 (unsigned long long)sh->sector);
1051 for (i = disks; i--; )
1052 fua |= test_bit(R5_WantFUA, &sh->dev[i].flags);
1054 for (i = disks; i--; ) {
1055 struct r5dev *dev = &sh->dev[i];
1057 if (dev->written || i == pd_idx || i == qd_idx) {
1058 set_bit(R5_UPTODATE, &dev->flags);
1060 set_bit(R5_WantFUA, &dev->flags);
1064 if (sh->reconstruct_state == reconstruct_state_drain_run)
1065 sh->reconstruct_state = reconstruct_state_drain_result;
1066 else if (sh->reconstruct_state == reconstruct_state_prexor_drain_run)
1067 sh->reconstruct_state = reconstruct_state_prexor_drain_result;
1069 BUG_ON(sh->reconstruct_state != reconstruct_state_run);
1070 sh->reconstruct_state = reconstruct_state_result;
1073 set_bit(STRIPE_HANDLE, &sh->state);
1078 ops_run_reconstruct5(struct stripe_head *sh, struct raid5_percpu *percpu,
1079 struct dma_async_tx_descriptor *tx)
1081 int disks = sh->disks;
1082 struct page **xor_srcs = percpu->scribble;
1083 struct async_submit_ctl submit;
1084 int count = 0, pd_idx = sh->pd_idx, i;
1085 struct page *xor_dest;
1087 unsigned long flags;
1089 pr_debug("%s: stripe %llu\n", __func__,
1090 (unsigned long long)sh->sector);
1092 /* check if prexor is active which means only process blocks
1093 * that are part of a read-modify-write (written)
1095 if (sh->reconstruct_state == reconstruct_state_prexor_drain_run) {
1097 xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
1098 for (i = disks; i--; ) {
1099 struct r5dev *dev = &sh->dev[i];
1101 xor_srcs[count++] = dev->page;
1104 xor_dest = sh->dev[pd_idx].page;
1105 for (i = disks; i--; ) {
1106 struct r5dev *dev = &sh->dev[i];
1108 xor_srcs[count++] = dev->page;
1112 /* 1/ if we prexor'd then the dest is reused as a source
1113 * 2/ if we did not prexor then we are redoing the parity
1114 * set ASYNC_TX_XOR_DROP_DST and ASYNC_TX_XOR_ZERO_DST
1115 * for the synchronous xor case
1117 flags = ASYNC_TX_ACK |
1118 (prexor ? ASYNC_TX_XOR_DROP_DST : ASYNC_TX_XOR_ZERO_DST);
1120 atomic_inc(&sh->count);
1122 init_async_submit(&submit, flags, tx, ops_complete_reconstruct, sh,
1123 to_addr_conv(sh, percpu));
1124 if (unlikely(count == 1))
1125 tx = async_memcpy(xor_dest, xor_srcs[0], 0, 0, STRIPE_SIZE, &submit);
1127 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
1131 ops_run_reconstruct6(struct stripe_head *sh, struct raid5_percpu *percpu,
1132 struct dma_async_tx_descriptor *tx)
1134 struct async_submit_ctl submit;
1135 struct page **blocks = percpu->scribble;
1138 pr_debug("%s: stripe %llu\n", __func__, (unsigned long long)sh->sector);
1140 count = set_syndrome_sources(blocks, sh);
1142 atomic_inc(&sh->count);
1144 init_async_submit(&submit, ASYNC_TX_ACK, tx, ops_complete_reconstruct,
1145 sh, to_addr_conv(sh, percpu));
1146 async_gen_syndrome(blocks, 0, count+2, STRIPE_SIZE, &submit);
1149 static void ops_complete_check(void *stripe_head_ref)
1151 struct stripe_head *sh = stripe_head_ref;
1153 pr_debug("%s: stripe %llu\n", __func__,
1154 (unsigned long long)sh->sector);
1156 sh->check_state = check_state_check_result;
1157 set_bit(STRIPE_HANDLE, &sh->state);
1161 static void ops_run_check_p(struct stripe_head *sh, struct raid5_percpu *percpu)
1163 int disks = sh->disks;
1164 int pd_idx = sh->pd_idx;
1165 int qd_idx = sh->qd_idx;
1166 struct page *xor_dest;
1167 struct page **xor_srcs = percpu->scribble;
1168 struct dma_async_tx_descriptor *tx;
1169 struct async_submit_ctl submit;
1173 pr_debug("%s: stripe %llu\n", __func__,
1174 (unsigned long long)sh->sector);
1177 xor_dest = sh->dev[pd_idx].page;
1178 xor_srcs[count++] = xor_dest;
1179 for (i = disks; i--; ) {
1180 if (i == pd_idx || i == qd_idx)
1182 xor_srcs[count++] = sh->dev[i].page;
1185 init_async_submit(&submit, 0, NULL, NULL, NULL,
1186 to_addr_conv(sh, percpu));
1187 tx = async_xor_val(xor_dest, xor_srcs, 0, count, STRIPE_SIZE,
1188 &sh->ops.zero_sum_result, &submit);
1190 atomic_inc(&sh->count);
1191 init_async_submit(&submit, ASYNC_TX_ACK, tx, ops_complete_check, sh, NULL);
1192 tx = async_trigger_callback(&submit);
1195 static void ops_run_check_pq(struct stripe_head *sh, struct raid5_percpu *percpu, int checkp)
1197 struct page **srcs = percpu->scribble;
1198 struct async_submit_ctl submit;
1201 pr_debug("%s: stripe %llu checkp: %d\n", __func__,
1202 (unsigned long long)sh->sector, checkp);
1204 count = set_syndrome_sources(srcs, sh);
1208 atomic_inc(&sh->count);
1209 init_async_submit(&submit, ASYNC_TX_ACK, NULL, ops_complete_check,
1210 sh, to_addr_conv(sh, percpu));
1211 async_syndrome_val(srcs, 0, count+2, STRIPE_SIZE,
1212 &sh->ops.zero_sum_result, percpu->spare_page, &submit);
1215 static void __raid_run_ops(struct stripe_head *sh, unsigned long ops_request)
1217 int overlap_clear = 0, i, disks = sh->disks;
1218 struct dma_async_tx_descriptor *tx = NULL;
1219 raid5_conf_t *conf = sh->raid_conf;
1220 int level = conf->level;
1221 struct raid5_percpu *percpu;
1225 percpu = per_cpu_ptr(conf->percpu, cpu);
1226 if (test_bit(STRIPE_OP_BIOFILL, &ops_request)) {
1227 ops_run_biofill(sh);
1231 if (test_bit(STRIPE_OP_COMPUTE_BLK, &ops_request)) {
1233 tx = ops_run_compute5(sh, percpu);
1235 if (sh->ops.target2 < 0 || sh->ops.target < 0)
1236 tx = ops_run_compute6_1(sh, percpu);
1238 tx = ops_run_compute6_2(sh, percpu);
1240 /* terminate the chain if reconstruct is not set to be run */
1241 if (tx && !test_bit(STRIPE_OP_RECONSTRUCT, &ops_request))
1245 if (test_bit(STRIPE_OP_PREXOR, &ops_request))
1246 tx = ops_run_prexor(sh, percpu, tx);
1248 if (test_bit(STRIPE_OP_BIODRAIN, &ops_request)) {
1249 tx = ops_run_biodrain(sh, tx);
1253 if (test_bit(STRIPE_OP_RECONSTRUCT, &ops_request)) {
1255 ops_run_reconstruct5(sh, percpu, tx);
1257 ops_run_reconstruct6(sh, percpu, tx);
1260 if (test_bit(STRIPE_OP_CHECK, &ops_request)) {
1261 if (sh->check_state == check_state_run)
1262 ops_run_check_p(sh, percpu);
1263 else if (sh->check_state == check_state_run_q)
1264 ops_run_check_pq(sh, percpu, 0);
1265 else if (sh->check_state == check_state_run_pq)
1266 ops_run_check_pq(sh, percpu, 1);
1272 for (i = disks; i--; ) {
1273 struct r5dev *dev = &sh->dev[i];
1274 if (test_and_clear_bit(R5_Overlap, &dev->flags))
1275 wake_up(&sh->raid_conf->wait_for_overlap);
1280 #ifdef CONFIG_MULTICORE_RAID456
1281 static void async_run_ops(void *param, async_cookie_t cookie)
1283 struct stripe_head *sh = param;
1284 unsigned long ops_request = sh->ops.request;
1286 clear_bit_unlock(STRIPE_OPS_REQ_PENDING, &sh->state);
1287 wake_up(&sh->ops.wait_for_ops);
1289 __raid_run_ops(sh, ops_request);
1293 static void raid_run_ops(struct stripe_head *sh, unsigned long ops_request)
1295 /* since handle_stripe can be called outside of raid5d context
1296 * we need to ensure sh->ops.request is de-staged before another
1299 wait_event(sh->ops.wait_for_ops,
1300 !test_and_set_bit_lock(STRIPE_OPS_REQ_PENDING, &sh->state));
1301 sh->ops.request = ops_request;
1303 atomic_inc(&sh->count);
1304 async_schedule(async_run_ops, sh);
1307 #define raid_run_ops __raid_run_ops
1310 static int grow_one_stripe(raid5_conf_t *conf)
1312 struct stripe_head *sh;
1313 sh = kmem_cache_zalloc(conf->slab_cache, GFP_KERNEL);
1317 sh->raid_conf = conf;
1318 #ifdef CONFIG_MULTICORE_RAID456
1319 init_waitqueue_head(&sh->ops.wait_for_ops);
1322 if (grow_buffers(sh)) {
1324 kmem_cache_free(conf->slab_cache, sh);
1327 /* we just created an active stripe so... */
1328 atomic_set(&sh->count, 1);
1329 atomic_inc(&conf->active_stripes);
1330 INIT_LIST_HEAD(&sh->lru);
1335 static int grow_stripes(raid5_conf_t *conf, int num)
1337 struct kmem_cache *sc;
1338 int devs = max(conf->raid_disks, conf->previous_raid_disks);
1340 if (conf->mddev->gendisk)
1341 sprintf(conf->cache_name[0],
1342 "raid%d-%s", conf->level, mdname(conf->mddev));
1344 sprintf(conf->cache_name[0],
1345 "raid%d-%p", conf->level, conf->mddev);
1346 sprintf(conf->cache_name[1], "%s-alt", conf->cache_name[0]);
1348 conf->active_name = 0;
1349 sc = kmem_cache_create(conf->cache_name[conf->active_name],
1350 sizeof(struct stripe_head)+(devs-1)*sizeof(struct r5dev),
1354 conf->slab_cache = sc;
1355 conf->pool_size = devs;
1357 if (!grow_one_stripe(conf))
1363 * scribble_len - return the required size of the scribble region
1364 * @num - total number of disks in the array
1366 * The size must be enough to contain:
1367 * 1/ a struct page pointer for each device in the array +2
1368 * 2/ room to convert each entry in (1) to its corresponding dma
1369 * (dma_map_page()) or page (page_address()) address.
1371 * Note: the +2 is for the destination buffers of the ddf/raid6 case where we
1372 * calculate over all devices (not just the data blocks), using zeros in place
1373 * of the P and Q blocks.
1375 static size_t scribble_len(int num)
1379 len = sizeof(struct page *) * (num+2) + sizeof(addr_conv_t) * (num+2);
1384 static int resize_stripes(raid5_conf_t *conf, int newsize)
1386 /* Make all the stripes able to hold 'newsize' devices.
1387 * New slots in each stripe get 'page' set to a new page.
1389 * This happens in stages:
1390 * 1/ create a new kmem_cache and allocate the required number of
1392 * 2/ gather all the old stripe_heads and tranfer the pages across
1393 * to the new stripe_heads. This will have the side effect of
1394 * freezing the array as once all stripe_heads have been collected,
1395 * no IO will be possible. Old stripe heads are freed once their
1396 * pages have been transferred over, and the old kmem_cache is
1397 * freed when all stripes are done.
1398 * 3/ reallocate conf->disks to be suitable bigger. If this fails,
1399 * we simple return a failre status - no need to clean anything up.
1400 * 4/ allocate new pages for the new slots in the new stripe_heads.
1401 * If this fails, we don't bother trying the shrink the
1402 * stripe_heads down again, we just leave them as they are.
1403 * As each stripe_head is processed the new one is released into
1406 * Once step2 is started, we cannot afford to wait for a write,
1407 * so we use GFP_NOIO allocations.
1409 struct stripe_head *osh, *nsh;
1410 LIST_HEAD(newstripes);
1411 struct disk_info *ndisks;
1414 struct kmem_cache *sc;
1417 if (newsize <= conf->pool_size)
1418 return 0; /* never bother to shrink */
1420 err = md_allow_write(conf->mddev);
1425 sc = kmem_cache_create(conf->cache_name[1-conf->active_name],
1426 sizeof(struct stripe_head)+(newsize-1)*sizeof(struct r5dev),
1431 for (i = conf->max_nr_stripes; i; i--) {
1432 nsh = kmem_cache_zalloc(sc, GFP_KERNEL);
1436 nsh->raid_conf = conf;
1437 #ifdef CONFIG_MULTICORE_RAID456
1438 init_waitqueue_head(&nsh->ops.wait_for_ops);
1441 list_add(&nsh->lru, &newstripes);
1444 /* didn't get enough, give up */
1445 while (!list_empty(&newstripes)) {
1446 nsh = list_entry(newstripes.next, struct stripe_head, lru);
1447 list_del(&nsh->lru);
1448 kmem_cache_free(sc, nsh);
1450 kmem_cache_destroy(sc);
1453 /* Step 2 - Must use GFP_NOIO now.
1454 * OK, we have enough stripes, start collecting inactive
1455 * stripes and copying them over
1457 list_for_each_entry(nsh, &newstripes, lru) {
1458 spin_lock_irq(&conf->device_lock);
1459 wait_event_lock_irq(conf->wait_for_stripe,
1460 !list_empty(&conf->inactive_list),
1463 osh = get_free_stripe(conf);
1464 spin_unlock_irq(&conf->device_lock);
1465 atomic_set(&nsh->count, 1);
1466 for(i=0; i<conf->pool_size; i++)
1467 nsh->dev[i].page = osh->dev[i].page;
1468 for( ; i<newsize; i++)
1469 nsh->dev[i].page = NULL;
1470 kmem_cache_free(conf->slab_cache, osh);
1472 kmem_cache_destroy(conf->slab_cache);
1475 * At this point, we are holding all the stripes so the array
1476 * is completely stalled, so now is a good time to resize
1477 * conf->disks and the scribble region
1479 ndisks = kzalloc(newsize * sizeof(struct disk_info), GFP_NOIO);
1481 for (i=0; i<conf->raid_disks; i++)
1482 ndisks[i] = conf->disks[i];
1484 conf->disks = ndisks;
1489 conf->scribble_len = scribble_len(newsize);
1490 for_each_present_cpu(cpu) {
1491 struct raid5_percpu *percpu;
1494 percpu = per_cpu_ptr(conf->percpu, cpu);
1495 scribble = kmalloc(conf->scribble_len, GFP_NOIO);
1498 kfree(percpu->scribble);
1499 percpu->scribble = scribble;
1507 /* Step 4, return new stripes to service */
1508 while(!list_empty(&newstripes)) {
1509 nsh = list_entry(newstripes.next, struct stripe_head, lru);
1510 list_del_init(&nsh->lru);
1512 for (i=conf->raid_disks; i < newsize; i++)
1513 if (nsh->dev[i].page == NULL) {
1514 struct page *p = alloc_page(GFP_NOIO);
1515 nsh->dev[i].page = p;
1519 release_stripe(nsh);
1521 /* critical section pass, GFP_NOIO no longer needed */
1523 conf->slab_cache = sc;
1524 conf->active_name = 1-conf->active_name;
1525 conf->pool_size = newsize;
1529 static int drop_one_stripe(raid5_conf_t *conf)
1531 struct stripe_head *sh;
1533 spin_lock_irq(&conf->device_lock);
1534 sh = get_free_stripe(conf);
1535 spin_unlock_irq(&conf->device_lock);
1538 BUG_ON(atomic_read(&sh->count));
1540 kmem_cache_free(conf->slab_cache, sh);
1541 atomic_dec(&conf->active_stripes);
1545 static void shrink_stripes(raid5_conf_t *conf)
1547 while (drop_one_stripe(conf))
1550 if (conf->slab_cache)
1551 kmem_cache_destroy(conf->slab_cache);
1552 conf->slab_cache = NULL;
1555 static void raid5_end_read_request(struct bio * bi, int error)
1557 struct stripe_head *sh = bi->bi_private;
1558 raid5_conf_t *conf = sh->raid_conf;
1559 int disks = sh->disks, i;
1560 int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
1561 char b[BDEVNAME_SIZE];
1565 for (i=0 ; i<disks; i++)
1566 if (bi == &sh->dev[i].req)
1569 pr_debug("end_read_request %llu/%d, count: %d, uptodate %d.\n",
1570 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
1578 set_bit(R5_UPTODATE, &sh->dev[i].flags);
1579 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
1580 rdev = conf->disks[i].rdev;
1583 "md/raid:%s: read error corrected"
1584 " (%lu sectors at %llu on %s)\n",
1585 mdname(conf->mddev), STRIPE_SECTORS,
1586 (unsigned long long)(sh->sector
1587 + rdev->data_offset),
1588 bdevname(rdev->bdev, b));
1589 atomic_add(STRIPE_SECTORS, &rdev->corrected_errors);
1590 clear_bit(R5_ReadError, &sh->dev[i].flags);
1591 clear_bit(R5_ReWrite, &sh->dev[i].flags);
1593 if (atomic_read(&conf->disks[i].rdev->read_errors))
1594 atomic_set(&conf->disks[i].rdev->read_errors, 0);
1596 const char *bdn = bdevname(conf->disks[i].rdev->bdev, b);
1598 rdev = conf->disks[i].rdev;
1600 clear_bit(R5_UPTODATE, &sh->dev[i].flags);
1601 atomic_inc(&rdev->read_errors);
1602 if (conf->mddev->degraded >= conf->max_degraded)
1605 "md/raid:%s: read error not correctable "
1606 "(sector %llu on %s).\n",
1607 mdname(conf->mddev),
1608 (unsigned long long)(sh->sector
1609 + rdev->data_offset),
1611 else if (test_bit(R5_ReWrite, &sh->dev[i].flags))
1615 "md/raid:%s: read error NOT corrected!! "
1616 "(sector %llu on %s).\n",
1617 mdname(conf->mddev),
1618 (unsigned long long)(sh->sector
1619 + rdev->data_offset),
1621 else if (atomic_read(&rdev->read_errors)
1622 > conf->max_nr_stripes)
1624 "md/raid:%s: Too many read errors, failing device %s.\n",
1625 mdname(conf->mddev), bdn);
1629 set_bit(R5_ReadError, &sh->dev[i].flags);
1631 clear_bit(R5_ReadError, &sh->dev[i].flags);
1632 clear_bit(R5_ReWrite, &sh->dev[i].flags);
1633 md_error(conf->mddev, rdev);
1636 rdev_dec_pending(conf->disks[i].rdev, conf->mddev);
1637 clear_bit(R5_LOCKED, &sh->dev[i].flags);
1638 set_bit(STRIPE_HANDLE, &sh->state);
1642 static void raid5_end_write_request(struct bio *bi, int error)
1644 struct stripe_head *sh = bi->bi_private;
1645 raid5_conf_t *conf = sh->raid_conf;
1646 int disks = sh->disks, i;
1647 int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
1649 for (i=0 ; i<disks; i++)
1650 if (bi == &sh->dev[i].req)
1653 pr_debug("end_write_request %llu/%d, count %d, uptodate: %d.\n",
1654 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
1662 md_error(conf->mddev, conf->disks[i].rdev);
1664 rdev_dec_pending(conf->disks[i].rdev, conf->mddev);
1666 clear_bit(R5_LOCKED, &sh->dev[i].flags);
1667 set_bit(STRIPE_HANDLE, &sh->state);
1672 static sector_t compute_blocknr(struct stripe_head *sh, int i, int previous);
1674 static void raid5_build_block(struct stripe_head *sh, int i, int previous)
1676 struct r5dev *dev = &sh->dev[i];
1678 bio_init(&dev->req);
1679 dev->req.bi_io_vec = &dev->vec;
1681 dev->req.bi_max_vecs++;
1682 dev->vec.bv_page = dev->page;
1683 dev->vec.bv_len = STRIPE_SIZE;
1684 dev->vec.bv_offset = 0;
1686 dev->req.bi_sector = sh->sector;
1687 dev->req.bi_private = sh;
1690 dev->sector = compute_blocknr(sh, i, previous);
1693 static void error(mddev_t *mddev, mdk_rdev_t *rdev)
1695 char b[BDEVNAME_SIZE];
1696 raid5_conf_t *conf = mddev->private;
1697 pr_debug("raid456: error called\n");
1699 if (test_and_clear_bit(In_sync, &rdev->flags)) {
1700 unsigned long flags;
1701 spin_lock_irqsave(&conf->device_lock, flags);
1703 spin_unlock_irqrestore(&conf->device_lock, flags);
1705 * if recovery was running, make sure it aborts.
1707 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1709 set_bit(Blocked, &rdev->flags);
1710 set_bit(Faulty, &rdev->flags);
1711 set_bit(MD_CHANGE_DEVS, &mddev->flags);
1713 "md/raid:%s: Disk failure on %s, disabling device.\n"
1714 "md/raid:%s: Operation continuing on %d devices.\n",
1716 bdevname(rdev->bdev, b),
1718 conf->raid_disks - mddev->degraded);
1722 * Input: a 'big' sector number,
1723 * Output: index of the data and parity disk, and the sector # in them.
1725 static sector_t raid5_compute_sector(raid5_conf_t *conf, sector_t r_sector,
1726 int previous, int *dd_idx,
1727 struct stripe_head *sh)
1729 sector_t stripe, stripe2;
1730 sector_t chunk_number;
1731 unsigned int chunk_offset;
1734 sector_t new_sector;
1735 int algorithm = previous ? conf->prev_algo
1737 int sectors_per_chunk = previous ? conf->prev_chunk_sectors
1738 : conf->chunk_sectors;
1739 int raid_disks = previous ? conf->previous_raid_disks
1741 int data_disks = raid_disks - conf->max_degraded;
1743 /* First compute the information on this sector */
1746 * Compute the chunk number and the sector offset inside the chunk
1748 chunk_offset = sector_div(r_sector, sectors_per_chunk);
1749 chunk_number = r_sector;
1752 * Compute the stripe number
1754 stripe = chunk_number;
1755 *dd_idx = sector_div(stripe, data_disks);
1758 * Select the parity disk based on the user selected algorithm.
1760 pd_idx = qd_idx = -1;
1761 switch(conf->level) {
1763 pd_idx = data_disks;
1766 switch (algorithm) {
1767 case ALGORITHM_LEFT_ASYMMETRIC:
1768 pd_idx = data_disks - sector_div(stripe2, raid_disks);
1769 if (*dd_idx >= pd_idx)
1772 case ALGORITHM_RIGHT_ASYMMETRIC:
1773 pd_idx = sector_div(stripe2, raid_disks);
1774 if (*dd_idx >= pd_idx)
1777 case ALGORITHM_LEFT_SYMMETRIC:
1778 pd_idx = data_disks - sector_div(stripe2, raid_disks);
1779 *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
1781 case ALGORITHM_RIGHT_SYMMETRIC:
1782 pd_idx = sector_div(stripe2, raid_disks);
1783 *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
1785 case ALGORITHM_PARITY_0:
1789 case ALGORITHM_PARITY_N:
1790 pd_idx = data_disks;
1798 switch (algorithm) {
1799 case ALGORITHM_LEFT_ASYMMETRIC:
1800 pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
1801 qd_idx = pd_idx + 1;
1802 if (pd_idx == raid_disks-1) {
1803 (*dd_idx)++; /* Q D D D P */
1805 } else if (*dd_idx >= pd_idx)
1806 (*dd_idx) += 2; /* D D P Q D */
1808 case ALGORITHM_RIGHT_ASYMMETRIC:
1809 pd_idx = sector_div(stripe2, raid_disks);
1810 qd_idx = pd_idx + 1;
1811 if (pd_idx == raid_disks-1) {
1812 (*dd_idx)++; /* Q D D D P */
1814 } else if (*dd_idx >= pd_idx)
1815 (*dd_idx) += 2; /* D D P Q D */
1817 case ALGORITHM_LEFT_SYMMETRIC:
1818 pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
1819 qd_idx = (pd_idx + 1) % raid_disks;
1820 *dd_idx = (pd_idx + 2 + *dd_idx) % raid_disks;
1822 case ALGORITHM_RIGHT_SYMMETRIC:
1823 pd_idx = sector_div(stripe2, raid_disks);
1824 qd_idx = (pd_idx + 1) % raid_disks;
1825 *dd_idx = (pd_idx + 2 + *dd_idx) % raid_disks;
1828 case ALGORITHM_PARITY_0:
1833 case ALGORITHM_PARITY_N:
1834 pd_idx = data_disks;
1835 qd_idx = data_disks + 1;
1838 case ALGORITHM_ROTATING_ZERO_RESTART:
1839 /* Exactly the same as RIGHT_ASYMMETRIC, but or
1840 * of blocks for computing Q is different.
1842 pd_idx = sector_div(stripe2, raid_disks);
1843 qd_idx = pd_idx + 1;
1844 if (pd_idx == raid_disks-1) {
1845 (*dd_idx)++; /* Q D D D P */
1847 } else if (*dd_idx >= pd_idx)
1848 (*dd_idx) += 2; /* D D P Q D */
1852 case ALGORITHM_ROTATING_N_RESTART:
1853 /* Same a left_asymmetric, by first stripe is
1854 * D D D P Q rather than
1858 pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
1859 qd_idx = pd_idx + 1;
1860 if (pd_idx == raid_disks-1) {
1861 (*dd_idx)++; /* Q D D D P */
1863 } else if (*dd_idx >= pd_idx)
1864 (*dd_idx) += 2; /* D D P Q D */
1868 case ALGORITHM_ROTATING_N_CONTINUE:
1869 /* Same as left_symmetric but Q is before P */
1870 pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
1871 qd_idx = (pd_idx + raid_disks - 1) % raid_disks;
1872 *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
1876 case ALGORITHM_LEFT_ASYMMETRIC_6:
1877 /* RAID5 left_asymmetric, with Q on last device */
1878 pd_idx = data_disks - sector_div(stripe2, raid_disks-1);
1879 if (*dd_idx >= pd_idx)
1881 qd_idx = raid_disks - 1;
1884 case ALGORITHM_RIGHT_ASYMMETRIC_6:
1885 pd_idx = sector_div(stripe2, raid_disks-1);
1886 if (*dd_idx >= pd_idx)
1888 qd_idx = raid_disks - 1;
1891 case ALGORITHM_LEFT_SYMMETRIC_6:
1892 pd_idx = data_disks - sector_div(stripe2, raid_disks-1);
1893 *dd_idx = (pd_idx + 1 + *dd_idx) % (raid_disks-1);
1894 qd_idx = raid_disks - 1;
1897 case ALGORITHM_RIGHT_SYMMETRIC_6:
1898 pd_idx = sector_div(stripe2, raid_disks-1);
1899 *dd_idx = (pd_idx + 1 + *dd_idx) % (raid_disks-1);
1900 qd_idx = raid_disks - 1;
1903 case ALGORITHM_PARITY_0_6:
1906 qd_idx = raid_disks - 1;
1916 sh->pd_idx = pd_idx;
1917 sh->qd_idx = qd_idx;
1918 sh->ddf_layout = ddf_layout;
1921 * Finally, compute the new sector number
1923 new_sector = (sector_t)stripe * sectors_per_chunk + chunk_offset;
1928 static sector_t compute_blocknr(struct stripe_head *sh, int i, int previous)
1930 raid5_conf_t *conf = sh->raid_conf;
1931 int raid_disks = sh->disks;
1932 int data_disks = raid_disks - conf->max_degraded;
1933 sector_t new_sector = sh->sector, check;
1934 int sectors_per_chunk = previous ? conf->prev_chunk_sectors
1935 : conf->chunk_sectors;
1936 int algorithm = previous ? conf->prev_algo
1940 sector_t chunk_number;
1941 int dummy1, dd_idx = i;
1943 struct stripe_head sh2;
1946 chunk_offset = sector_div(new_sector, sectors_per_chunk);
1947 stripe = new_sector;
1949 if (i == sh->pd_idx)
1951 switch(conf->level) {
1954 switch (algorithm) {
1955 case ALGORITHM_LEFT_ASYMMETRIC:
1956 case ALGORITHM_RIGHT_ASYMMETRIC:
1960 case ALGORITHM_LEFT_SYMMETRIC:
1961 case ALGORITHM_RIGHT_SYMMETRIC:
1964 i -= (sh->pd_idx + 1);
1966 case ALGORITHM_PARITY_0:
1969 case ALGORITHM_PARITY_N:
1976 if (i == sh->qd_idx)
1977 return 0; /* It is the Q disk */
1978 switch (algorithm) {
1979 case ALGORITHM_LEFT_ASYMMETRIC:
1980 case ALGORITHM_RIGHT_ASYMMETRIC:
1981 case ALGORITHM_ROTATING_ZERO_RESTART:
1982 case ALGORITHM_ROTATING_N_RESTART:
1983 if (sh->pd_idx == raid_disks-1)
1984 i--; /* Q D D D P */
1985 else if (i > sh->pd_idx)
1986 i -= 2; /* D D P Q D */
1988 case ALGORITHM_LEFT_SYMMETRIC:
1989 case ALGORITHM_RIGHT_SYMMETRIC:
1990 if (sh->pd_idx == raid_disks-1)
1991 i--; /* Q D D D P */
1996 i -= (sh->pd_idx + 2);
1999 case ALGORITHM_PARITY_0:
2002 case ALGORITHM_PARITY_N:
2004 case ALGORITHM_ROTATING_N_CONTINUE:
2005 /* Like left_symmetric, but P is before Q */
2006 if (sh->pd_idx == 0)
2007 i--; /* P D D D Q */
2012 i -= (sh->pd_idx + 1);
2015 case ALGORITHM_LEFT_ASYMMETRIC_6:
2016 case ALGORITHM_RIGHT_ASYMMETRIC_6:
2020 case ALGORITHM_LEFT_SYMMETRIC_6:
2021 case ALGORITHM_RIGHT_SYMMETRIC_6:
2023 i += data_disks + 1;
2024 i -= (sh->pd_idx + 1);
2026 case ALGORITHM_PARITY_0_6:
2035 chunk_number = stripe * data_disks + i;
2036 r_sector = chunk_number * sectors_per_chunk + chunk_offset;
2038 check = raid5_compute_sector(conf, r_sector,
2039 previous, &dummy1, &sh2);
2040 if (check != sh->sector || dummy1 != dd_idx || sh2.pd_idx != sh->pd_idx
2041 || sh2.qd_idx != sh->qd_idx) {
2042 printk(KERN_ERR "md/raid:%s: compute_blocknr: map not correct\n",
2043 mdname(conf->mddev));
2051 schedule_reconstruction(struct stripe_head *sh, struct stripe_head_state *s,
2052 int rcw, int expand)
2054 int i, pd_idx = sh->pd_idx, disks = sh->disks;
2055 raid5_conf_t *conf = sh->raid_conf;
2056 int level = conf->level;
2059 /* if we are not expanding this is a proper write request, and
2060 * there will be bios with new data to be drained into the
2064 sh->reconstruct_state = reconstruct_state_drain_run;
2065 set_bit(STRIPE_OP_BIODRAIN, &s->ops_request);
2067 sh->reconstruct_state = reconstruct_state_run;
2069 set_bit(STRIPE_OP_RECONSTRUCT, &s->ops_request);
2071 for (i = disks; i--; ) {
2072 struct r5dev *dev = &sh->dev[i];
2075 set_bit(R5_LOCKED, &dev->flags);
2076 set_bit(R5_Wantdrain, &dev->flags);
2078 clear_bit(R5_UPTODATE, &dev->flags);
2082 if (s->locked + conf->max_degraded == disks)
2083 if (!test_and_set_bit(STRIPE_FULL_WRITE, &sh->state))
2084 atomic_inc(&conf->pending_full_writes);
2087 BUG_ON(!(test_bit(R5_UPTODATE, &sh->dev[pd_idx].flags) ||
2088 test_bit(R5_Wantcompute, &sh->dev[pd_idx].flags)));
2090 sh->reconstruct_state = reconstruct_state_prexor_drain_run;
2091 set_bit(STRIPE_OP_PREXOR, &s->ops_request);
2092 set_bit(STRIPE_OP_BIODRAIN, &s->ops_request);
2093 set_bit(STRIPE_OP_RECONSTRUCT, &s->ops_request);
2095 for (i = disks; i--; ) {
2096 struct r5dev *dev = &sh->dev[i];
2101 (test_bit(R5_UPTODATE, &dev->flags) ||
2102 test_bit(R5_Wantcompute, &dev->flags))) {
2103 set_bit(R5_Wantdrain, &dev->flags);
2104 set_bit(R5_LOCKED, &dev->flags);
2105 clear_bit(R5_UPTODATE, &dev->flags);
2111 /* keep the parity disk(s) locked while asynchronous operations
2114 set_bit(R5_LOCKED, &sh->dev[pd_idx].flags);
2115 clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
2119 int qd_idx = sh->qd_idx;
2120 struct r5dev *dev = &sh->dev[qd_idx];
2122 set_bit(R5_LOCKED, &dev->flags);
2123 clear_bit(R5_UPTODATE, &dev->flags);
2127 pr_debug("%s: stripe %llu locked: %d ops_request: %lx\n",
2128 __func__, (unsigned long long)sh->sector,
2129 s->locked, s->ops_request);
2133 * Each stripe/dev can have one or more bion attached.
2134 * toread/towrite point to the first in a chain.
2135 * The bi_next chain must be in order.
2137 static int add_stripe_bio(struct stripe_head *sh, struct bio *bi, int dd_idx, int forwrite)
2140 raid5_conf_t *conf = sh->raid_conf;
2143 pr_debug("adding bi b#%llu to stripe s#%llu\n",
2144 (unsigned long long)bi->bi_sector,
2145 (unsigned long long)sh->sector);
2148 spin_lock_irq(&conf->device_lock);
2150 bip = &sh->dev[dd_idx].towrite;
2151 if (*bip == NULL && sh->dev[dd_idx].written == NULL)
2154 bip = &sh->dev[dd_idx].toread;
2155 while (*bip && (*bip)->bi_sector < bi->bi_sector) {
2156 if ((*bip)->bi_sector + ((*bip)->bi_size >> 9) > bi->bi_sector)
2158 bip = & (*bip)->bi_next;
2160 if (*bip && (*bip)->bi_sector < bi->bi_sector + ((bi->bi_size)>>9))
2163 BUG_ON(*bip && bi->bi_next && (*bip) != bi->bi_next);
2167 bi->bi_phys_segments++;
2170 /* check if page is covered */
2171 sector_t sector = sh->dev[dd_idx].sector;
2172 for (bi=sh->dev[dd_idx].towrite;
2173 sector < sh->dev[dd_idx].sector + STRIPE_SECTORS &&
2174 bi && bi->bi_sector <= sector;
2175 bi = r5_next_bio(bi, sh->dev[dd_idx].sector)) {
2176 if (bi->bi_sector + (bi->bi_size>>9) >= sector)
2177 sector = bi->bi_sector + (bi->bi_size>>9);
2179 if (sector >= sh->dev[dd_idx].sector + STRIPE_SECTORS)
2180 set_bit(R5_OVERWRITE, &sh->dev[dd_idx].flags);
2182 spin_unlock_irq(&conf->device_lock);
2184 pr_debug("added bi b#%llu to stripe s#%llu, disk %d.\n",
2185 (unsigned long long)(*bip)->bi_sector,
2186 (unsigned long long)sh->sector, dd_idx);
2188 if (conf->mddev->bitmap && firstwrite) {
2189 bitmap_startwrite(conf->mddev->bitmap, sh->sector,
2191 sh->bm_seq = conf->seq_flush+1;
2192 set_bit(STRIPE_BIT_DELAY, &sh->state);
2197 set_bit(R5_Overlap, &sh->dev[dd_idx].flags);
2198 spin_unlock_irq(&conf->device_lock);
2202 static void end_reshape(raid5_conf_t *conf);
2204 static void stripe_set_idx(sector_t stripe, raid5_conf_t *conf, int previous,
2205 struct stripe_head *sh)
2207 int sectors_per_chunk =
2208 previous ? conf->prev_chunk_sectors : conf->chunk_sectors;
2210 int chunk_offset = sector_div(stripe, sectors_per_chunk);
2211 int disks = previous ? conf->previous_raid_disks : conf->raid_disks;
2213 raid5_compute_sector(conf,
2214 stripe * (disks - conf->max_degraded)
2215 *sectors_per_chunk + chunk_offset,
2221 handle_failed_stripe(raid5_conf_t *conf, struct stripe_head *sh,
2222 struct stripe_head_state *s, int disks,
2223 struct bio **return_bi)
2226 for (i = disks; i--; ) {
2230 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
2233 rdev = rcu_dereference(conf->disks[i].rdev);
2234 if (rdev && test_bit(In_sync, &rdev->flags))
2235 /* multiple read failures in one stripe */
2236 md_error(conf->mddev, rdev);
2239 spin_lock_irq(&conf->device_lock);
2240 /* fail all writes first */
2241 bi = sh->dev[i].towrite;
2242 sh->dev[i].towrite = NULL;
2248 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
2249 wake_up(&conf->wait_for_overlap);
2251 while (bi && bi->bi_sector <
2252 sh->dev[i].sector + STRIPE_SECTORS) {
2253 struct bio *nextbi = r5_next_bio(bi, sh->dev[i].sector);
2254 clear_bit(BIO_UPTODATE, &bi->bi_flags);
2255 if (!raid5_dec_bi_phys_segments(bi)) {
2256 md_write_end(conf->mddev);
2257 bi->bi_next = *return_bi;
2262 /* and fail all 'written' */
2263 bi = sh->dev[i].written;
2264 sh->dev[i].written = NULL;
2265 if (bi) bitmap_end = 1;
2266 while (bi && bi->bi_sector <
2267 sh->dev[i].sector + STRIPE_SECTORS) {
2268 struct bio *bi2 = r5_next_bio(bi, sh->dev[i].sector);
2269 clear_bit(BIO_UPTODATE, &bi->bi_flags);
2270 if (!raid5_dec_bi_phys_segments(bi)) {
2271 md_write_end(conf->mddev);
2272 bi->bi_next = *return_bi;
2278 /* fail any reads if this device is non-operational and
2279 * the data has not reached the cache yet.
2281 if (!test_bit(R5_Wantfill, &sh->dev[i].flags) &&
2282 (!test_bit(R5_Insync, &sh->dev[i].flags) ||
2283 test_bit(R5_ReadError, &sh->dev[i].flags))) {
2284 bi = sh->dev[i].toread;
2285 sh->dev[i].toread = NULL;
2286 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
2287 wake_up(&conf->wait_for_overlap);
2288 if (bi) s->to_read--;
2289 while (bi && bi->bi_sector <
2290 sh->dev[i].sector + STRIPE_SECTORS) {
2291 struct bio *nextbi =
2292 r5_next_bio(bi, sh->dev[i].sector);
2293 clear_bit(BIO_UPTODATE, &bi->bi_flags);
2294 if (!raid5_dec_bi_phys_segments(bi)) {
2295 bi->bi_next = *return_bi;
2301 spin_unlock_irq(&conf->device_lock);
2303 bitmap_endwrite(conf->mddev->bitmap, sh->sector,
2304 STRIPE_SECTORS, 0, 0);
2305 /* If we were in the middle of a write the parity block might
2306 * still be locked - so just clear all R5_LOCKED flags
2308 clear_bit(R5_LOCKED, &sh->dev[i].flags);
2311 if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
2312 if (atomic_dec_and_test(&conf->pending_full_writes))
2313 md_wakeup_thread(conf->mddev->thread);
2316 /* fetch_block - checks the given member device to see if its data needs
2317 * to be read or computed to satisfy a request.
2319 * Returns 1 when no more member devices need to be checked, otherwise returns
2320 * 0 to tell the loop in handle_stripe_fill to continue
2322 static int fetch_block(struct stripe_head *sh, struct stripe_head_state *s,
2323 int disk_idx, int disks)
2325 struct r5dev *dev = &sh->dev[disk_idx];
2326 struct r5dev *fdev[2] = { &sh->dev[s->failed_num[0]],
2327 &sh->dev[s->failed_num[1]] };
2329 /* is the data in this block needed, and can we get it? */
2330 if (!test_bit(R5_LOCKED, &dev->flags) &&
2331 !test_bit(R5_UPTODATE, &dev->flags) &&
2333 (dev->towrite && !test_bit(R5_OVERWRITE, &dev->flags)) ||
2334 s->syncing || s->expanding ||
2335 (s->failed >= 1 && fdev[0]->toread) ||
2336 (s->failed >= 2 && fdev[1]->toread) ||
2337 (sh->raid_conf->level <= 5 && s->failed && fdev[0]->towrite &&
2338 !test_bit(R5_OVERWRITE, &fdev[0]->flags)) ||
2339 (sh->raid_conf->level == 6 && s->failed && s->to_write))) {
2340 /* we would like to get this block, possibly by computing it,
2341 * otherwise read it if the backing disk is insync
2343 BUG_ON(test_bit(R5_Wantcompute, &dev->flags));
2344 BUG_ON(test_bit(R5_Wantread, &dev->flags));
2345 if ((s->uptodate == disks - 1) &&
2346 (s->failed && (disk_idx == s->failed_num[0] ||
2347 disk_idx == s->failed_num[1]))) {
2348 /* have disk failed, and we're requested to fetch it;
2351 pr_debug("Computing stripe %llu block %d\n",
2352 (unsigned long long)sh->sector, disk_idx);
2353 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
2354 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
2355 set_bit(R5_Wantcompute, &dev->flags);
2356 sh->ops.target = disk_idx;
2357 sh->ops.target2 = -1; /* no 2nd target */
2359 /* Careful: from this point on 'uptodate' is in the eye
2360 * of raid_run_ops which services 'compute' operations
2361 * before writes. R5_Wantcompute flags a block that will
2362 * be R5_UPTODATE by the time it is needed for a
2363 * subsequent operation.
2367 } else if (s->uptodate == disks-2 && s->failed >= 2) {
2368 /* Computing 2-failure is *very* expensive; only
2369 * do it if failed >= 2
2372 for (other = disks; other--; ) {
2373 if (other == disk_idx)
2375 if (!test_bit(R5_UPTODATE,
2376 &sh->dev[other].flags))
2380 pr_debug("Computing stripe %llu blocks %d,%d\n",
2381 (unsigned long long)sh->sector,
2383 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
2384 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
2385 set_bit(R5_Wantcompute, &sh->dev[disk_idx].flags);
2386 set_bit(R5_Wantcompute, &sh->dev[other].flags);
2387 sh->ops.target = disk_idx;
2388 sh->ops.target2 = other;
2392 } else if (test_bit(R5_Insync, &dev->flags)) {
2393 set_bit(R5_LOCKED, &dev->flags);
2394 set_bit(R5_Wantread, &dev->flags);
2396 pr_debug("Reading block %d (sync=%d)\n",
2397 disk_idx, s->syncing);
2405 * handle_stripe_fill - read or compute data to satisfy pending requests.
2407 static void handle_stripe_fill(struct stripe_head *sh,
2408 struct stripe_head_state *s,
2413 /* look for blocks to read/compute, skip this if a compute
2414 * is already in flight, or if the stripe contents are in the
2415 * midst of changing due to a write
2417 if (!test_bit(STRIPE_COMPUTE_RUN, &sh->state) && !sh->check_state &&
2418 !sh->reconstruct_state)
2419 for (i = disks; i--; )
2420 if (fetch_block(sh, s, i, disks))
2422 set_bit(STRIPE_HANDLE, &sh->state);
2426 /* handle_stripe_clean_event
2427 * any written block on an uptodate or failed drive can be returned.
2428 * Note that if we 'wrote' to a failed drive, it will be UPTODATE, but
2429 * never LOCKED, so we don't need to test 'failed' directly.
2431 static void handle_stripe_clean_event(raid5_conf_t *conf,
2432 struct stripe_head *sh, int disks, struct bio **return_bi)
2437 for (i = disks; i--; )
2438 if (sh->dev[i].written) {
2440 if (!test_bit(R5_LOCKED, &dev->flags) &&
2441 test_bit(R5_UPTODATE, &dev->flags)) {
2442 /* We can return any write requests */
2443 struct bio *wbi, *wbi2;
2445 pr_debug("Return write for disc %d\n", i);
2446 spin_lock_irq(&conf->device_lock);
2448 dev->written = NULL;
2449 while (wbi && wbi->bi_sector <
2450 dev->sector + STRIPE_SECTORS) {
2451 wbi2 = r5_next_bio(wbi, dev->sector);
2452 if (!raid5_dec_bi_phys_segments(wbi)) {
2453 md_write_end(conf->mddev);
2454 wbi->bi_next = *return_bi;
2459 if (dev->towrite == NULL)
2461 spin_unlock_irq(&conf->device_lock);
2463 bitmap_endwrite(conf->mddev->bitmap,
2466 !test_bit(STRIPE_DEGRADED, &sh->state),
2471 if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
2472 if (atomic_dec_and_test(&conf->pending_full_writes))
2473 md_wakeup_thread(conf->mddev->thread);
2476 static void handle_stripe_dirtying(raid5_conf_t *conf,
2477 struct stripe_head *sh,
2478 struct stripe_head_state *s,
2481 int rmw = 0, rcw = 0, i;
2482 if (conf->max_degraded == 2) {
2483 /* RAID6 requires 'rcw' in current implementation
2484 * Calculate the real rcw later - for now fake it
2485 * look like rcw is cheaper
2488 } else for (i = disks; i--; ) {
2489 /* would I have to read this buffer for read_modify_write */
2490 struct r5dev *dev = &sh->dev[i];
2491 if ((dev->towrite || i == sh->pd_idx) &&
2492 !test_bit(R5_LOCKED, &dev->flags) &&
2493 !(test_bit(R5_UPTODATE, &dev->flags) ||
2494 test_bit(R5_Wantcompute, &dev->flags))) {
2495 if (test_bit(R5_Insync, &dev->flags))
2498 rmw += 2*disks; /* cannot read it */
2500 /* Would I have to read this buffer for reconstruct_write */
2501 if (!test_bit(R5_OVERWRITE, &dev->flags) && i != sh->pd_idx &&
2502 !test_bit(R5_LOCKED, &dev->flags) &&
2503 !(test_bit(R5_UPTODATE, &dev->flags) ||
2504 test_bit(R5_Wantcompute, &dev->flags))) {
2505 if (test_bit(R5_Insync, &dev->flags)) rcw++;
2510 pr_debug("for sector %llu, rmw=%d rcw=%d\n",
2511 (unsigned long long)sh->sector, rmw, rcw);
2512 set_bit(STRIPE_HANDLE, &sh->state);
2513 if (rmw < rcw && rmw > 0)
2514 /* prefer read-modify-write, but need to get some data */
2515 for (i = disks; i--; ) {
2516 struct r5dev *dev = &sh->dev[i];
2517 if ((dev->towrite || i == sh->pd_idx) &&
2518 !test_bit(R5_LOCKED, &dev->flags) &&
2519 !(test_bit(R5_UPTODATE, &dev->flags) ||
2520 test_bit(R5_Wantcompute, &dev->flags)) &&
2521 test_bit(R5_Insync, &dev->flags)) {
2523 test_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2524 pr_debug("Read_old block "
2525 "%d for r-m-w\n", i);
2526 set_bit(R5_LOCKED, &dev->flags);
2527 set_bit(R5_Wantread, &dev->flags);
2530 set_bit(STRIPE_DELAYED, &sh->state);
2531 set_bit(STRIPE_HANDLE, &sh->state);
2535 if (rcw <= rmw && rcw > 0) {
2536 /* want reconstruct write, but need to get some data */
2538 for (i = disks; i--; ) {
2539 struct r5dev *dev = &sh->dev[i];
2540 if (!test_bit(R5_OVERWRITE, &dev->flags) &&
2541 i != sh->pd_idx && i != sh->qd_idx &&
2542 !test_bit(R5_LOCKED, &dev->flags) &&
2543 !(test_bit(R5_UPTODATE, &dev->flags) ||
2544 test_bit(R5_Wantcompute, &dev->flags))) {
2546 if (!test_bit(R5_Insync, &dev->flags))
2547 continue; /* it's a failed drive */
2549 test_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2550 pr_debug("Read_old block "
2551 "%d for Reconstruct\n", i);
2552 set_bit(R5_LOCKED, &dev->flags);
2553 set_bit(R5_Wantread, &dev->flags);
2556 set_bit(STRIPE_DELAYED, &sh->state);
2557 set_bit(STRIPE_HANDLE, &sh->state);
2562 /* now if nothing is locked, and if we have enough data,
2563 * we can start a write request
2565 /* since handle_stripe can be called at any time we need to handle the
2566 * case where a compute block operation has been submitted and then a
2567 * subsequent call wants to start a write request. raid_run_ops only
2568 * handles the case where compute block and reconstruct are requested
2569 * simultaneously. If this is not the case then new writes need to be
2570 * held off until the compute completes.
2572 if ((s->req_compute || !test_bit(STRIPE_COMPUTE_RUN, &sh->state)) &&
2573 (s->locked == 0 && (rcw == 0 || rmw == 0) &&
2574 !test_bit(STRIPE_BIT_DELAY, &sh->state)))
2575 schedule_reconstruction(sh, s, rcw == 0, 0);
2578 static void handle_parity_checks5(raid5_conf_t *conf, struct stripe_head *sh,
2579 struct stripe_head_state *s, int disks)
2581 struct r5dev *dev = NULL;
2583 set_bit(STRIPE_HANDLE, &sh->state);
2585 switch (sh->check_state) {
2586 case check_state_idle:
2587 /* start a new check operation if there are no failures */
2588 if (s->failed == 0) {
2589 BUG_ON(s->uptodate != disks);
2590 sh->check_state = check_state_run;
2591 set_bit(STRIPE_OP_CHECK, &s->ops_request);
2592 clear_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags);
2596 dev = &sh->dev[s->failed_num[0]];
2598 case check_state_compute_result:
2599 sh->check_state = check_state_idle;
2601 dev = &sh->dev[sh->pd_idx];
2603 /* check that a write has not made the stripe insync */
2604 if (test_bit(STRIPE_INSYNC, &sh->state))
2607 /* either failed parity check, or recovery is happening */
2608 BUG_ON(!test_bit(R5_UPTODATE, &dev->flags));
2609 BUG_ON(s->uptodate != disks);
2611 set_bit(R5_LOCKED, &dev->flags);
2613 set_bit(R5_Wantwrite, &dev->flags);
2615 clear_bit(STRIPE_DEGRADED, &sh->state);
2616 set_bit(STRIPE_INSYNC, &sh->state);
2618 case check_state_run:
2619 break; /* we will be called again upon completion */
2620 case check_state_check_result:
2621 sh->check_state = check_state_idle;
2623 /* if a failure occurred during the check operation, leave
2624 * STRIPE_INSYNC not set and let the stripe be handled again
2629 /* handle a successful check operation, if parity is correct
2630 * we are done. Otherwise update the mismatch count and repair
2631 * parity if !MD_RECOVERY_CHECK
2633 if ((sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) == 0)
2634 /* parity is correct (on disc,
2635 * not in buffer any more)
2637 set_bit(STRIPE_INSYNC, &sh->state);
2639 conf->mddev->resync_mismatches += STRIPE_SECTORS;
2640 if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery))
2641 /* don't try to repair!! */
2642 set_bit(STRIPE_INSYNC, &sh->state);
2644 sh->check_state = check_state_compute_run;
2645 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
2646 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
2647 set_bit(R5_Wantcompute,
2648 &sh->dev[sh->pd_idx].flags);
2649 sh->ops.target = sh->pd_idx;
2650 sh->ops.target2 = -1;
2655 case check_state_compute_run:
2658 printk(KERN_ERR "%s: unknown check_state: %d sector: %llu\n",
2659 __func__, sh->check_state,
2660 (unsigned long long) sh->sector);
2666 static void handle_parity_checks6(raid5_conf_t *conf, struct stripe_head *sh,
2667 struct stripe_head_state *s,
2670 int pd_idx = sh->pd_idx;
2671 int qd_idx = sh->qd_idx;
2674 set_bit(STRIPE_HANDLE, &sh->state);
2676 BUG_ON(s->failed > 2);
2678 /* Want to check and possibly repair P and Q.
2679 * However there could be one 'failed' device, in which
2680 * case we can only check one of them, possibly using the
2681 * other to generate missing data
2684 switch (sh->check_state) {
2685 case check_state_idle:
2686 /* start a new check operation if there are < 2 failures */
2687 if (s->failed == s->q_failed) {
2688 /* The only possible failed device holds Q, so it
2689 * makes sense to check P (If anything else were failed,
2690 * we would have used P to recreate it).
2692 sh->check_state = check_state_run;
2694 if (!s->q_failed && s->failed < 2) {
2695 /* Q is not failed, and we didn't use it to generate
2696 * anything, so it makes sense to check it
2698 if (sh->check_state == check_state_run)
2699 sh->check_state = check_state_run_pq;
2701 sh->check_state = check_state_run_q;
2704 /* discard potentially stale zero_sum_result */
2705 sh->ops.zero_sum_result = 0;
2707 if (sh->check_state == check_state_run) {
2708 /* async_xor_zero_sum destroys the contents of P */
2709 clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
2712 if (sh->check_state >= check_state_run &&
2713 sh->check_state <= check_state_run_pq) {
2714 /* async_syndrome_zero_sum preserves P and Q, so
2715 * no need to mark them !uptodate here
2717 set_bit(STRIPE_OP_CHECK, &s->ops_request);
2721 /* we have 2-disk failure */
2722 BUG_ON(s->failed != 2);
2724 case check_state_compute_result:
2725 sh->check_state = check_state_idle;
2727 /* check that a write has not made the stripe insync */
2728 if (test_bit(STRIPE_INSYNC, &sh->state))
2731 /* now write out any block on a failed drive,
2732 * or P or Q if they were recomputed
2734 BUG_ON(s->uptodate < disks - 1); /* We don't need Q to recover */
2735 if (s->failed == 2) {
2736 dev = &sh->dev[s->failed_num[1]];
2738 set_bit(R5_LOCKED, &dev->flags);
2739 set_bit(R5_Wantwrite, &dev->flags);
2741 if (s->failed >= 1) {
2742 dev = &sh->dev[s->failed_num[0]];
2744 set_bit(R5_LOCKED, &dev->flags);
2745 set_bit(R5_Wantwrite, &dev->flags);
2747 if (sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) {
2748 dev = &sh->dev[pd_idx];
2750 set_bit(R5_LOCKED, &dev->flags);
2751 set_bit(R5_Wantwrite, &dev->flags);
2753 if (sh->ops.zero_sum_result & SUM_CHECK_Q_RESULT) {
2754 dev = &sh->dev[qd_idx];
2756 set_bit(R5_LOCKED, &dev->flags);
2757 set_bit(R5_Wantwrite, &dev->flags);
2759 clear_bit(STRIPE_DEGRADED, &sh->state);
2761 set_bit(STRIPE_INSYNC, &sh->state);
2763 case check_state_run:
2764 case check_state_run_q:
2765 case check_state_run_pq:
2766 break; /* we will be called again upon completion */
2767 case check_state_check_result:
2768 sh->check_state = check_state_idle;
2770 /* handle a successful check operation, if parity is correct
2771 * we are done. Otherwise update the mismatch count and repair
2772 * parity if !MD_RECOVERY_CHECK
2774 if (sh->ops.zero_sum_result == 0) {
2775 /* both parities are correct */
2777 set_bit(STRIPE_INSYNC, &sh->state);
2779 /* in contrast to the raid5 case we can validate
2780 * parity, but still have a failure to write
2783 sh->check_state = check_state_compute_result;
2784 /* Returning at this point means that we may go
2785 * off and bring p and/or q uptodate again so
2786 * we make sure to check zero_sum_result again
2787 * to verify if p or q need writeback
2791 conf->mddev->resync_mismatches += STRIPE_SECTORS;
2792 if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery))
2793 /* don't try to repair!! */
2794 set_bit(STRIPE_INSYNC, &sh->state);
2796 int *target = &sh->ops.target;
2798 sh->ops.target = -1;
2799 sh->ops.target2 = -1;
2800 sh->check_state = check_state_compute_run;
2801 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
2802 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
2803 if (sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) {
2804 set_bit(R5_Wantcompute,
2805 &sh->dev[pd_idx].flags);
2807 target = &sh->ops.target2;
2810 if (sh->ops.zero_sum_result & SUM_CHECK_Q_RESULT) {
2811 set_bit(R5_Wantcompute,
2812 &sh->dev[qd_idx].flags);
2819 case check_state_compute_run:
2822 printk(KERN_ERR "%s: unknown check_state: %d sector: %llu\n",
2823 __func__, sh->check_state,
2824 (unsigned long long) sh->sector);
2829 static void handle_stripe_expansion(raid5_conf_t *conf, struct stripe_head *sh)
2833 /* We have read all the blocks in this stripe and now we need to
2834 * copy some of them into a target stripe for expand.
2836 struct dma_async_tx_descriptor *tx = NULL;
2837 clear_bit(STRIPE_EXPAND_SOURCE, &sh->state);
2838 for (i = 0; i < sh->disks; i++)
2839 if (i != sh->pd_idx && i != sh->qd_idx) {
2841 struct stripe_head *sh2;
2842 struct async_submit_ctl submit;
2844 sector_t bn = compute_blocknr(sh, i, 1);
2845 sector_t s = raid5_compute_sector(conf, bn, 0,
2847 sh2 = get_active_stripe(conf, s, 0, 1, 1);
2849 /* so far only the early blocks of this stripe
2850 * have been requested. When later blocks
2851 * get requested, we will try again
2854 if (!test_bit(STRIPE_EXPANDING, &sh2->state) ||
2855 test_bit(R5_Expanded, &sh2->dev[dd_idx].flags)) {
2856 /* must have already done this block */
2857 release_stripe(sh2);
2861 /* place all the copies on one channel */
2862 init_async_submit(&submit, 0, tx, NULL, NULL, NULL);
2863 tx = async_memcpy(sh2->dev[dd_idx].page,
2864 sh->dev[i].page, 0, 0, STRIPE_SIZE,
2867 set_bit(R5_Expanded, &sh2->dev[dd_idx].flags);
2868 set_bit(R5_UPTODATE, &sh2->dev[dd_idx].flags);
2869 for (j = 0; j < conf->raid_disks; j++)
2870 if (j != sh2->pd_idx &&
2872 !test_bit(R5_Expanded, &sh2->dev[j].flags))
2874 if (j == conf->raid_disks) {
2875 set_bit(STRIPE_EXPAND_READY, &sh2->state);
2876 set_bit(STRIPE_HANDLE, &sh2->state);
2878 release_stripe(sh2);
2881 /* done submitting copies, wait for them to complete */
2884 dma_wait_for_async_tx(tx);
2890 * handle_stripe - do things to a stripe.
2892 * We lock the stripe and then examine the state of various bits
2893 * to see what needs to be done.
2895 * return some read request which now have data
2896 * return some write requests which are safely on disc
2897 * schedule a read on some buffers
2898 * schedule a write of some buffers
2899 * return confirmation of parity correctness
2901 * buffers are taken off read_list or write_list, and bh_cache buffers
2902 * get BH_Lock set before the stripe lock is released.
2906 static void analyse_stripe(struct stripe_head *sh, struct stripe_head_state *s)
2908 raid5_conf_t *conf = sh->raid_conf;
2909 int disks = sh->disks;
2913 memset(s, 0, sizeof(*s));
2915 s->syncing = test_bit(STRIPE_SYNCING, &sh->state);
2916 s->expanding = test_bit(STRIPE_EXPAND_SOURCE, &sh->state);
2917 s->expanded = test_bit(STRIPE_EXPAND_READY, &sh->state);
2918 s->failed_num[0] = -1;
2919 s->failed_num[1] = -1;
2921 /* Now to look around and see what can be done */
2923 spin_lock_irq(&conf->device_lock);
2924 for (i=disks; i--; ) {
2932 pr_debug("check %d: state 0x%lx read %p write %p written %p\n",
2933 i, dev->flags, dev->toread, dev->towrite, dev->written);
2934 /* maybe we can reply to a read
2936 * new wantfill requests are only permitted while
2937 * ops_complete_biofill is guaranteed to be inactive
2939 if (test_bit(R5_UPTODATE, &dev->flags) && dev->toread &&
2940 !test_bit(STRIPE_BIOFILL_RUN, &sh->state))
2941 set_bit(R5_Wantfill, &dev->flags);
2943 /* now count some things */
2944 if (test_bit(R5_LOCKED, &dev->flags))
2946 if (test_bit(R5_UPTODATE, &dev->flags))
2948 if (test_bit(R5_Wantcompute, &dev->flags)) {
2950 BUG_ON(s->compute > 2);
2953 if (test_bit(R5_Wantfill, &dev->flags))
2955 else if (dev->toread)
2959 if (!test_bit(R5_OVERWRITE, &dev->flags))
2964 rdev = rcu_dereference(conf->disks[i].rdev);
2966 is_bad = is_badblock(rdev, sh->sector, STRIPE_SECTORS,
2967 &first_bad, &bad_sectors);
2968 if (s->blocked_rdev == NULL
2969 && (test_bit(Blocked, &rdev->flags)
2972 set_bit(BlockedBadBlocks,
2974 s->blocked_rdev = rdev;
2975 atomic_inc(&rdev->nr_pending);
2978 clear_bit(R5_Insync, &dev->flags);
2982 /* also not in-sync */
2983 if (!test_bit(WriteErrorSeen, &rdev->flags)) {
2984 /* treat as in-sync, but with a read error
2985 * which we can now try to correct
2987 set_bit(R5_Insync, &dev->flags);
2988 set_bit(R5_ReadError, &dev->flags);
2990 } else if (test_bit(In_sync, &rdev->flags))
2991 set_bit(R5_Insync, &dev->flags);
2993 /* in sync if before recovery_offset */
2994 if (sh->sector + STRIPE_SECTORS <= rdev->recovery_offset)
2995 set_bit(R5_Insync, &dev->flags);
2997 if (!test_bit(R5_Insync, &dev->flags)) {
2998 /* The ReadError flag will just be confusing now */
2999 clear_bit(R5_ReadError, &dev->flags);
3000 clear_bit(R5_ReWrite, &dev->flags);
3002 if (test_bit(R5_ReadError, &dev->flags))
3003 clear_bit(R5_Insync, &dev->flags);
3004 if (!test_bit(R5_Insync, &dev->flags)) {
3006 s->failed_num[s->failed] = i;
3010 spin_unlock_irq(&conf->device_lock);
3014 static void handle_stripe(struct stripe_head *sh)
3016 struct stripe_head_state s;
3017 raid5_conf_t *conf = sh->raid_conf;
3020 int disks = sh->disks;
3021 struct r5dev *pdev, *qdev;
3023 clear_bit(STRIPE_HANDLE, &sh->state);
3024 if (test_and_set_bit(STRIPE_ACTIVE, &sh->state)) {
3025 /* already being handled, ensure it gets handled
3026 * again when current action finishes */
3027 set_bit(STRIPE_HANDLE, &sh->state);
3031 if (test_and_clear_bit(STRIPE_SYNC_REQUESTED, &sh->state)) {
3032 set_bit(STRIPE_SYNCING, &sh->state);
3033 clear_bit(STRIPE_INSYNC, &sh->state);
3035 clear_bit(STRIPE_DELAYED, &sh->state);
3037 pr_debug("handling stripe %llu, state=%#lx cnt=%d, "
3038 "pd_idx=%d, qd_idx=%d\n, check:%d, reconstruct:%d\n",
3039 (unsigned long long)sh->sector, sh->state,
3040 atomic_read(&sh->count), sh->pd_idx, sh->qd_idx,
3041 sh->check_state, sh->reconstruct_state);
3043 analyse_stripe(sh, &s);
3045 if (unlikely(s.blocked_rdev)) {
3046 if (s.syncing || s.expanding || s.expanded ||
3047 s.to_write || s.written) {
3048 set_bit(STRIPE_HANDLE, &sh->state);
3051 /* There is nothing for the blocked_rdev to block */
3052 rdev_dec_pending(s.blocked_rdev, conf->mddev);
3053 s.blocked_rdev = NULL;
3056 if (s.to_fill && !test_bit(STRIPE_BIOFILL_RUN, &sh->state)) {
3057 set_bit(STRIPE_OP_BIOFILL, &s.ops_request);
3058 set_bit(STRIPE_BIOFILL_RUN, &sh->state);
3061 pr_debug("locked=%d uptodate=%d to_read=%d"
3062 " to_write=%d failed=%d failed_num=%d,%d\n",
3063 s.locked, s.uptodate, s.to_read, s.to_write, s.failed,
3064 s.failed_num[0], s.failed_num[1]);
3065 /* check if the array has lost more than max_degraded devices and,
3066 * if so, some requests might need to be failed.
3068 if (s.failed > conf->max_degraded && s.to_read+s.to_write+s.written)
3069 handle_failed_stripe(conf, sh, &s, disks, &s.return_bi);
3070 if (s.failed > conf->max_degraded && s.syncing) {
3071 md_done_sync(conf->mddev, STRIPE_SECTORS, 0);
3072 clear_bit(STRIPE_SYNCING, &sh->state);
3077 * might be able to return some write requests if the parity blocks
3078 * are safe, or on a failed drive
3080 pdev = &sh->dev[sh->pd_idx];
3081 s.p_failed = (s.failed >= 1 && s.failed_num[0] == sh->pd_idx)
3082 || (s.failed >= 2 && s.failed_num[1] == sh->pd_idx);
3083 qdev = &sh->dev[sh->qd_idx];
3084 s.q_failed = (s.failed >= 1 && s.failed_num[0] == sh->qd_idx)
3085 || (s.failed >= 2 && s.failed_num[1] == sh->qd_idx)
3089 (s.p_failed || ((test_bit(R5_Insync, &pdev->flags)
3090 && !test_bit(R5_LOCKED, &pdev->flags)
3091 && test_bit(R5_UPTODATE, &pdev->flags)))) &&
3092 (s.q_failed || ((test_bit(R5_Insync, &qdev->flags)
3093 && !test_bit(R5_LOCKED, &qdev->flags)
3094 && test_bit(R5_UPTODATE, &qdev->flags)))))
3095 handle_stripe_clean_event(conf, sh, disks, &s.return_bi);
3097 /* Now we might consider reading some blocks, either to check/generate
3098 * parity, or to satisfy requests
3099 * or to load a block that is being partially written.
3101 if (s.to_read || s.non_overwrite
3102 || (conf->level == 6 && s.to_write && s.failed)
3103 || (s.syncing && (s.uptodate + s.compute < disks)) || s.expanding)
3104 handle_stripe_fill(sh, &s, disks);
3106 /* Now we check to see if any write operations have recently
3110 if (sh->reconstruct_state == reconstruct_state_prexor_drain_result)
3112 if (sh->reconstruct_state == reconstruct_state_drain_result ||
3113 sh->reconstruct_state == reconstruct_state_prexor_drain_result) {
3114 sh->reconstruct_state = reconstruct_state_idle;
3116 /* All the 'written' buffers and the parity block are ready to
3117 * be written back to disk
3119 BUG_ON(!test_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags));
3120 BUG_ON(sh->qd_idx >= 0 &&
3121 !test_bit(R5_UPTODATE, &sh->dev[sh->qd_idx].flags));
3122 for (i = disks; i--; ) {
3123 struct r5dev *dev = &sh->dev[i];
3124 if (test_bit(R5_LOCKED, &dev->flags) &&
3125 (i == sh->pd_idx || i == sh->qd_idx ||
3127 pr_debug("Writing block %d\n", i);
3128 set_bit(R5_Wantwrite, &dev->flags);
3131 if (!test_bit(R5_Insync, &dev->flags) ||
3132 ((i == sh->pd_idx || i == sh->qd_idx) &&
3134 set_bit(STRIPE_INSYNC, &sh->state);
3137 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
3138 s.dec_preread_active = 1;
3141 /* Now to consider new write requests and what else, if anything
3142 * should be read. We do not handle new writes when:
3143 * 1/ A 'write' operation (copy+xor) is already in flight.
3144 * 2/ A 'check' operation is in flight, as it may clobber the parity
3147 if (s.to_write && !sh->reconstruct_state && !sh->check_state)
3148 handle_stripe_dirtying(conf, sh, &s, disks);
3150 /* maybe we need to check and possibly fix the parity for this stripe
3151 * Any reads will already have been scheduled, so we just see if enough
3152 * data is available. The parity check is held off while parity
3153 * dependent operations are in flight.
3155 if (sh->check_state ||
3156 (s.syncing && s.locked == 0 &&
3157 !test_bit(STRIPE_COMPUTE_RUN, &sh->state) &&
3158 !test_bit(STRIPE_INSYNC, &sh->state))) {
3159 if (conf->level == 6)
3160 handle_parity_checks6(conf, sh, &s, disks);
3162 handle_parity_checks5(conf, sh, &s, disks);
3165 if (s.syncing && s.locked == 0 && test_bit(STRIPE_INSYNC, &sh->state)) {
3166 md_done_sync(conf->mddev, STRIPE_SECTORS, 1);
3167 clear_bit(STRIPE_SYNCING, &sh->state);
3170 /* If the failed drives are just a ReadError, then we might need
3171 * to progress the repair/check process
3173 if (s.failed <= conf->max_degraded && !conf->mddev->ro)
3174 for (i = 0; i < s.failed; i++) {
3175 struct r5dev *dev = &sh->dev[s.failed_num[i]];
3176 if (test_bit(R5_ReadError, &dev->flags)
3177 && !test_bit(R5_LOCKED, &dev->flags)
3178 && test_bit(R5_UPTODATE, &dev->flags)
3180 if (!test_bit(R5_ReWrite, &dev->flags)) {
3181 set_bit(R5_Wantwrite, &dev->flags);
3182 set_bit(R5_ReWrite, &dev->flags);
3183 set_bit(R5_LOCKED, &dev->flags);
3186 /* let's read it back */
3187 set_bit(R5_Wantread, &dev->flags);
3188 set_bit(R5_LOCKED, &dev->flags);
3195 /* Finish reconstruct operations initiated by the expansion process */
3196 if (sh->reconstruct_state == reconstruct_state_result) {
3197 struct stripe_head *sh_src
3198 = get_active_stripe(conf, sh->sector, 1, 1, 1);
3199 if (sh_src && test_bit(STRIPE_EXPAND_SOURCE, &sh_src->state)) {
3200 /* sh cannot be written until sh_src has been read.
3201 * so arrange for sh to be delayed a little
3203 set_bit(STRIPE_DELAYED, &sh->state);
3204 set_bit(STRIPE_HANDLE, &sh->state);
3205 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE,
3207 atomic_inc(&conf->preread_active_stripes);
3208 release_stripe(sh_src);
3212 release_stripe(sh_src);
3214 sh->reconstruct_state = reconstruct_state_idle;
3215 clear_bit(STRIPE_EXPANDING, &sh->state);
3216 for (i = conf->raid_disks; i--; ) {
3217 set_bit(R5_Wantwrite, &sh->dev[i].flags);
3218 set_bit(R5_LOCKED, &sh->dev[i].flags);
3223 if (s.expanded && test_bit(STRIPE_EXPANDING, &sh->state) &&
3224 !sh->reconstruct_state) {
3225 /* Need to write out all blocks after computing parity */
3226 sh->disks = conf->raid_disks;
3227 stripe_set_idx(sh->sector, conf, 0, sh);
3228 schedule_reconstruction(sh, &s, 1, 1);
3229 } else if (s.expanded && !sh->reconstruct_state && s.locked == 0) {
3230 clear_bit(STRIPE_EXPAND_READY, &sh->state);
3231 atomic_dec(&conf->reshape_stripes);
3232 wake_up(&conf->wait_for_overlap);
3233 md_done_sync(conf->mddev, STRIPE_SECTORS, 1);
3236 if (s.expanding && s.locked == 0 &&
3237 !test_bit(STRIPE_COMPUTE_RUN, &sh->state))
3238 handle_stripe_expansion(conf, sh);
3241 /* wait for this device to become unblocked */
3242 if (unlikely(s.blocked_rdev))
3243 md_wait_for_blocked_rdev(s.blocked_rdev, conf->mddev);
3246 raid_run_ops(sh, s.ops_request);
3251 if (s.dec_preread_active) {
3252 /* We delay this until after ops_run_io so that if make_request
3253 * is waiting on a flush, it won't continue until the writes
3254 * have actually been submitted.
3256 atomic_dec(&conf->preread_active_stripes);
3257 if (atomic_read(&conf->preread_active_stripes) <
3259 md_wakeup_thread(conf->mddev->thread);
3262 return_io(s.return_bi);
3264 clear_bit(STRIPE_ACTIVE, &sh->state);
3267 static void raid5_activate_delayed(raid5_conf_t *conf)
3269 if (atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD) {
3270 while (!list_empty(&conf->delayed_list)) {
3271 struct list_head *l = conf->delayed_list.next;
3272 struct stripe_head *sh;
3273 sh = list_entry(l, struct stripe_head, lru);
3275 clear_bit(STRIPE_DELAYED, &sh->state);
3276 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
3277 atomic_inc(&conf->preread_active_stripes);
3278 list_add_tail(&sh->lru, &conf->hold_list);
3283 static void activate_bit_delay(raid5_conf_t *conf)
3285 /* device_lock is held */
3286 struct list_head head;
3287 list_add(&head, &conf->bitmap_list);
3288 list_del_init(&conf->bitmap_list);
3289 while (!list_empty(&head)) {
3290 struct stripe_head *sh = list_entry(head.next, struct stripe_head, lru);
3291 list_del_init(&sh->lru);
3292 atomic_inc(&sh->count);
3293 __release_stripe(conf, sh);
3297 int md_raid5_congested(mddev_t *mddev, int bits)
3299 raid5_conf_t *conf = mddev->private;
3301 /* No difference between reads and writes. Just check
3302 * how busy the stripe_cache is
3305 if (conf->inactive_blocked)
3309 if (list_empty_careful(&conf->inactive_list))
3314 EXPORT_SYMBOL_GPL(md_raid5_congested);
3316 static int raid5_congested(void *data, int bits)
3318 mddev_t *mddev = data;
3320 return mddev_congested(mddev, bits) ||
3321 md_raid5_congested(mddev, bits);
3324 /* We want read requests to align with chunks where possible,
3325 * but write requests don't need to.
3327 static int raid5_mergeable_bvec(struct request_queue *q,
3328 struct bvec_merge_data *bvm,
3329 struct bio_vec *biovec)
3331 mddev_t *mddev = q->queuedata;
3332 sector_t sector = bvm->bi_sector + get_start_sect(bvm->bi_bdev);
3334 unsigned int chunk_sectors = mddev->chunk_sectors;
3335 unsigned int bio_sectors = bvm->bi_size >> 9;
3337 if ((bvm->bi_rw & 1) == WRITE)
3338 return biovec->bv_len; /* always allow writes to be mergeable */
3340 if (mddev->new_chunk_sectors < mddev->chunk_sectors)
3341 chunk_sectors = mddev->new_chunk_sectors;
3342 max = (chunk_sectors - ((sector & (chunk_sectors - 1)) + bio_sectors)) << 9;
3343 if (max < 0) max = 0;
3344 if (max <= biovec->bv_len && bio_sectors == 0)
3345 return biovec->bv_len;
3351 static int in_chunk_boundary(mddev_t *mddev, struct bio *bio)
3353 sector_t sector = bio->bi_sector + get_start_sect(bio->bi_bdev);
3354 unsigned int chunk_sectors = mddev->chunk_sectors;
3355 unsigned int bio_sectors = bio->bi_size >> 9;
3357 if (mddev->new_chunk_sectors < mddev->chunk_sectors)
3358 chunk_sectors = mddev->new_chunk_sectors;
3359 return chunk_sectors >=
3360 ((sector & (chunk_sectors - 1)) + bio_sectors);
3364 * add bio to the retry LIFO ( in O(1) ... we are in interrupt )
3365 * later sampled by raid5d.
3367 static void add_bio_to_retry(struct bio *bi,raid5_conf_t *conf)
3369 unsigned long flags;
3371 spin_lock_irqsave(&conf->device_lock, flags);
3373 bi->bi_next = conf->retry_read_aligned_list;
3374 conf->retry_read_aligned_list = bi;
3376 spin_unlock_irqrestore(&conf->device_lock, flags);
3377 md_wakeup_thread(conf->mddev->thread);
3381 static struct bio *remove_bio_from_retry(raid5_conf_t *conf)
3385 bi = conf->retry_read_aligned;
3387 conf->retry_read_aligned = NULL;
3390 bi = conf->retry_read_aligned_list;
3392 conf->retry_read_aligned_list = bi->bi_next;
3395 * this sets the active strip count to 1 and the processed
3396 * strip count to zero (upper 8 bits)
3398 bi->bi_phys_segments = 1; /* biased count of active stripes */
3406 * The "raid5_align_endio" should check if the read succeeded and if it
3407 * did, call bio_endio on the original bio (having bio_put the new bio
3409 * If the read failed..
3411 static void raid5_align_endio(struct bio *bi, int error)
3413 struct bio* raid_bi = bi->bi_private;
3416 int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
3421 rdev = (void*)raid_bi->bi_next;
3422 raid_bi->bi_next = NULL;
3423 mddev = rdev->mddev;
3424 conf = mddev->private;
3426 rdev_dec_pending(rdev, conf->mddev);
3428 if (!error && uptodate) {
3429 bio_endio(raid_bi, 0);
3430 if (atomic_dec_and_test(&conf->active_aligned_reads))
3431 wake_up(&conf->wait_for_stripe);
3436 pr_debug("raid5_align_endio : io error...handing IO for a retry\n");
3438 add_bio_to_retry(raid_bi, conf);
3441 static int bio_fits_rdev(struct bio *bi)
3443 struct request_queue *q = bdev_get_queue(bi->bi_bdev);
3445 if ((bi->bi_size>>9) > queue_max_sectors(q))
3447 blk_recount_segments(q, bi);
3448 if (bi->bi_phys_segments > queue_max_segments(q))
3451 if (q->merge_bvec_fn)
3452 /* it's too hard to apply the merge_bvec_fn at this stage,
3461 static int chunk_aligned_read(mddev_t *mddev, struct bio * raid_bio)
3463 raid5_conf_t *conf = mddev->private;
3465 struct bio* align_bi;
3468 if (!in_chunk_boundary(mddev, raid_bio)) {
3469 pr_debug("chunk_aligned_read : non aligned\n");
3473 * use bio_clone_mddev to make a copy of the bio
3475 align_bi = bio_clone_mddev(raid_bio, GFP_NOIO, mddev);
3479 * set bi_end_io to a new function, and set bi_private to the
3482 align_bi->bi_end_io = raid5_align_endio;
3483 align_bi->bi_private = raid_bio;
3487 align_bi->bi_sector = raid5_compute_sector(conf, raid_bio->bi_sector,
3492 rdev = rcu_dereference(conf->disks[dd_idx].rdev);
3493 if (rdev && test_bit(In_sync, &rdev->flags)) {
3497 atomic_inc(&rdev->nr_pending);
3499 raid_bio->bi_next = (void*)rdev;
3500 align_bi->bi_bdev = rdev->bdev;
3501 align_bi->bi_flags &= ~(1 << BIO_SEG_VALID);
3502 align_bi->bi_sector += rdev->data_offset;
3504 if (!bio_fits_rdev(align_bi) ||
3505 is_badblock(rdev, align_bi->bi_sector, align_bi->bi_size>>9,
3506 &first_bad, &bad_sectors)) {
3507 /* too big in some way, or has a known bad block */
3509 rdev_dec_pending(rdev, mddev);
3513 spin_lock_irq(&conf->device_lock);
3514 wait_event_lock_irq(conf->wait_for_stripe,
3516 conf->device_lock, /* nothing */);
3517 atomic_inc(&conf->active_aligned_reads);
3518 spin_unlock_irq(&conf->device_lock);
3520 generic_make_request(align_bi);
3529 /* __get_priority_stripe - get the next stripe to process
3531 * Full stripe writes are allowed to pass preread active stripes up until
3532 * the bypass_threshold is exceeded. In general the bypass_count
3533 * increments when the handle_list is handled before the hold_list; however, it
3534 * will not be incremented when STRIPE_IO_STARTED is sampled set signifying a
3535 * stripe with in flight i/o. The bypass_count will be reset when the
3536 * head of the hold_list has changed, i.e. the head was promoted to the
3539 static struct stripe_head *__get_priority_stripe(raid5_conf_t *conf)
3541 struct stripe_head *sh;
3543 pr_debug("%s: handle: %s hold: %s full_writes: %d bypass_count: %d\n",
3545 list_empty(&conf->handle_list) ? "empty" : "busy",
3546 list_empty(&conf->hold_list) ? "empty" : "busy",
3547 atomic_read(&conf->pending_full_writes), conf->bypass_count);
3549 if (!list_empty(&conf->handle_list)) {
3550 sh = list_entry(conf->handle_list.next, typeof(*sh), lru);
3552 if (list_empty(&conf->hold_list))
3553 conf->bypass_count = 0;
3554 else if (!test_bit(STRIPE_IO_STARTED, &sh->state)) {
3555 if (conf->hold_list.next == conf->last_hold)
3556 conf->bypass_count++;
3558 conf->last_hold = conf->hold_list.next;
3559 conf->bypass_count -= conf->bypass_threshold;
3560 if (conf->bypass_count < 0)
3561 conf->bypass_count = 0;
3564 } else if (!list_empty(&conf->hold_list) &&
3565 ((conf->bypass_threshold &&
3566 conf->bypass_count > conf->bypass_threshold) ||
3567 atomic_read(&conf->pending_full_writes) == 0)) {
3568 sh = list_entry(conf->hold_list.next,
3570 conf->bypass_count -= conf->bypass_threshold;
3571 if (conf->bypass_count < 0)
3572 conf->bypass_count = 0;
3576 list_del_init(&sh->lru);
3577 atomic_inc(&sh->count);
3578 BUG_ON(atomic_read(&sh->count) != 1);
3582 static int make_request(mddev_t *mddev, struct bio * bi)
3584 raid5_conf_t *conf = mddev->private;
3586 sector_t new_sector;
3587 sector_t logical_sector, last_sector;
3588 struct stripe_head *sh;
3589 const int rw = bio_data_dir(bi);
3593 if (unlikely(bi->bi_rw & REQ_FLUSH)) {
3594 md_flush_request(mddev, bi);
3598 md_write_start(mddev, bi);
3601 mddev->reshape_position == MaxSector &&
3602 chunk_aligned_read(mddev,bi))
3605 logical_sector = bi->bi_sector & ~((sector_t)STRIPE_SECTORS-1);
3606 last_sector = bi->bi_sector + (bi->bi_size>>9);
3608 bi->bi_phys_segments = 1; /* over-loaded to count active stripes */
3610 plugged = mddev_check_plugged(mddev);
3611 for (;logical_sector < last_sector; logical_sector += STRIPE_SECTORS) {
3613 int disks, data_disks;
3618 disks = conf->raid_disks;
3619 prepare_to_wait(&conf->wait_for_overlap, &w, TASK_UNINTERRUPTIBLE);
3620 if (unlikely(conf->reshape_progress != MaxSector)) {
3621 /* spinlock is needed as reshape_progress may be
3622 * 64bit on a 32bit platform, and so it might be
3623 * possible to see a half-updated value
3624 * Of course reshape_progress could change after
3625 * the lock is dropped, so once we get a reference
3626 * to the stripe that we think it is, we will have
3629 spin_lock_irq(&conf->device_lock);
3630 if (mddev->delta_disks < 0
3631 ? logical_sector < conf->reshape_progress
3632 : logical_sector >= conf->reshape_progress) {
3633 disks = conf->previous_raid_disks;
3636 if (mddev->delta_disks < 0
3637 ? logical_sector < conf->reshape_safe
3638 : logical_sector >= conf->reshape_safe) {
3639 spin_unlock_irq(&conf->device_lock);
3644 spin_unlock_irq(&conf->device_lock);
3646 data_disks = disks - conf->max_degraded;
3648 new_sector = raid5_compute_sector(conf, logical_sector,
3651 pr_debug("raid456: make_request, sector %llu logical %llu\n",
3652 (unsigned long long)new_sector,
3653 (unsigned long long)logical_sector);
3655 sh = get_active_stripe(conf, new_sector, previous,
3656 (bi->bi_rw&RWA_MASK), 0);
3658 if (unlikely(previous)) {
3659 /* expansion might have moved on while waiting for a
3660 * stripe, so we must do the range check again.
3661 * Expansion could still move past after this
3662 * test, but as we are holding a reference to
3663 * 'sh', we know that if that happens,
3664 * STRIPE_EXPANDING will get set and the expansion
3665 * won't proceed until we finish with the stripe.
3668 spin_lock_irq(&conf->device_lock);
3669 if (mddev->delta_disks < 0
3670 ? logical_sector >= conf->reshape_progress
3671 : logical_sector < conf->reshape_progress)
3672 /* mismatch, need to try again */
3674 spin_unlock_irq(&conf->device_lock);
3683 logical_sector >= mddev->suspend_lo &&
3684 logical_sector < mddev->suspend_hi) {
3686 /* As the suspend_* range is controlled by
3687 * userspace, we want an interruptible
3690 flush_signals(current);
3691 prepare_to_wait(&conf->wait_for_overlap,
3692 &w, TASK_INTERRUPTIBLE);
3693 if (logical_sector >= mddev->suspend_lo &&
3694 logical_sector < mddev->suspend_hi)
3699 if (test_bit(STRIPE_EXPANDING, &sh->state) ||
3700 !add_stripe_bio(sh, bi, dd_idx, rw)) {
3701 /* Stripe is busy expanding or
3702 * add failed due to overlap. Flush everything
3705 md_wakeup_thread(mddev->thread);
3710 finish_wait(&conf->wait_for_overlap, &w);
3711 set_bit(STRIPE_HANDLE, &sh->state);
3712 clear_bit(STRIPE_DELAYED, &sh->state);
3713 if ((bi->bi_rw & REQ_SYNC) &&
3714 !test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
3715 atomic_inc(&conf->preread_active_stripes);
3718 /* cannot get stripe for read-ahead, just give-up */
3719 clear_bit(BIO_UPTODATE, &bi->bi_flags);
3720 finish_wait(&conf->wait_for_overlap, &w);
3726 md_wakeup_thread(mddev->thread);
3728 spin_lock_irq(&conf->device_lock);
3729 remaining = raid5_dec_bi_phys_segments(bi);
3730 spin_unlock_irq(&conf->device_lock);
3731 if (remaining == 0) {
3734 md_write_end(mddev);
3742 static sector_t raid5_size(mddev_t *mddev, sector_t sectors, int raid_disks);
3744 static sector_t reshape_request(mddev_t *mddev, sector_t sector_nr, int *skipped)
3746 /* reshaping is quite different to recovery/resync so it is
3747 * handled quite separately ... here.
3749 * On each call to sync_request, we gather one chunk worth of
3750 * destination stripes and flag them as expanding.
3751 * Then we find all the source stripes and request reads.
3752 * As the reads complete, handle_stripe will copy the data
3753 * into the destination stripe and release that stripe.
3755 raid5_conf_t *conf = mddev->private;
3756 struct stripe_head *sh;
3757 sector_t first_sector, last_sector;
3758 int raid_disks = conf->previous_raid_disks;
3759 int data_disks = raid_disks - conf->max_degraded;
3760 int new_data_disks = conf->raid_disks - conf->max_degraded;
3763 sector_t writepos, readpos, safepos;
3764 sector_t stripe_addr;
3765 int reshape_sectors;
3766 struct list_head stripes;
3768 if (sector_nr == 0) {
3769 /* If restarting in the middle, skip the initial sectors */
3770 if (mddev->delta_disks < 0 &&
3771 conf->reshape_progress < raid5_size(mddev, 0, 0)) {
3772 sector_nr = raid5_size(mddev, 0, 0)
3773 - conf->reshape_progress;
3774 } else if (mddev->delta_disks >= 0 &&
3775 conf->reshape_progress > 0)
3776 sector_nr = conf->reshape_progress;
3777 sector_div(sector_nr, new_data_disks);
3779 mddev->curr_resync_completed = sector_nr;
3780 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
3786 /* We need to process a full chunk at a time.
3787 * If old and new chunk sizes differ, we need to process the
3790 if (mddev->new_chunk_sectors > mddev->chunk_sectors)
3791 reshape_sectors = mddev->new_chunk_sectors;
3793 reshape_sectors = mddev->chunk_sectors;
3795 /* we update the metadata when there is more than 3Meg
3796 * in the block range (that is rather arbitrary, should
3797 * probably be time based) or when the data about to be
3798 * copied would over-write the source of the data at
3799 * the front of the range.
3800 * i.e. one new_stripe along from reshape_progress new_maps
3801 * to after where reshape_safe old_maps to
3803 writepos = conf->reshape_progress;
3804 sector_div(writepos, new_data_disks);
3805 readpos = conf->reshape_progress;
3806 sector_div(readpos, data_disks);
3807 safepos = conf->reshape_safe;
3808 sector_div(safepos, data_disks);
3809 if (mddev->delta_disks < 0) {
3810 writepos -= min_t(sector_t, reshape_sectors, writepos);
3811 readpos += reshape_sectors;
3812 safepos += reshape_sectors;
3814 writepos += reshape_sectors;
3815 readpos -= min_t(sector_t, reshape_sectors, readpos);
3816 safepos -= min_t(sector_t, reshape_sectors, safepos);
3819 /* 'writepos' is the most advanced device address we might write.
3820 * 'readpos' is the least advanced device address we might read.
3821 * 'safepos' is the least address recorded in the metadata as having
3823 * If 'readpos' is behind 'writepos', then there is no way that we can
3824 * ensure safety in the face of a crash - that must be done by userspace
3825 * making a backup of the data. So in that case there is no particular
3826 * rush to update metadata.
3827 * Otherwise if 'safepos' is behind 'writepos', then we really need to
3828 * update the metadata to advance 'safepos' to match 'readpos' so that
3829 * we can be safe in the event of a crash.
3830 * So we insist on updating metadata if safepos is behind writepos and
3831 * readpos is beyond writepos.
3832 * In any case, update the metadata every 10 seconds.
3833 * Maybe that number should be configurable, but I'm not sure it is
3834 * worth it.... maybe it could be a multiple of safemode_delay???
3836 if ((mddev->delta_disks < 0
3837 ? (safepos > writepos && readpos < writepos)
3838 : (safepos < writepos && readpos > writepos)) ||
3839 time_after(jiffies, conf->reshape_checkpoint + 10*HZ)) {
3840 /* Cannot proceed until we've updated the superblock... */
3841 wait_event(conf->wait_for_overlap,
3842 atomic_read(&conf->reshape_stripes)==0);
3843 mddev->reshape_position = conf->reshape_progress;
3844 mddev->curr_resync_completed = sector_nr;
3845 conf->reshape_checkpoint = jiffies;
3846 set_bit(MD_CHANGE_DEVS, &mddev->flags);
3847 md_wakeup_thread(mddev->thread);
3848 wait_event(mddev->sb_wait, mddev->flags == 0 ||
3849 kthread_should_stop());
3850 spin_lock_irq(&conf->device_lock);
3851 conf->reshape_safe = mddev->reshape_position;
3852 spin_unlock_irq(&conf->device_lock);
3853 wake_up(&conf->wait_for_overlap);
3854 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
3857 if (mddev->delta_disks < 0) {
3858 BUG_ON(conf->reshape_progress == 0);
3859 stripe_addr = writepos;
3860 BUG_ON((mddev->dev_sectors &
3861 ~((sector_t)reshape_sectors - 1))
3862 - reshape_sectors - stripe_addr
3865 BUG_ON(writepos != sector_nr + reshape_sectors);
3866 stripe_addr = sector_nr;
3868 INIT_LIST_HEAD(&stripes);
3869 for (i = 0; i < reshape_sectors; i += STRIPE_SECTORS) {
3871 int skipped_disk = 0;
3872 sh = get_active_stripe(conf, stripe_addr+i, 0, 0, 1);
3873 set_bit(STRIPE_EXPANDING, &sh->state);
3874 atomic_inc(&conf->reshape_stripes);
3875 /* If any of this stripe is beyond the end of the old
3876 * array, then we need to zero those blocks
3878 for (j=sh->disks; j--;) {
3880 if (j == sh->pd_idx)
3882 if (conf->level == 6 &&
3885 s = compute_blocknr(sh, j, 0);
3886 if (s < raid5_size(mddev, 0, 0)) {
3890 memset(page_address(sh->dev[j].page), 0, STRIPE_SIZE);
3891 set_bit(R5_Expanded, &sh->dev[j].flags);
3892 set_bit(R5_UPTODATE, &sh->dev[j].flags);
3894 if (!skipped_disk) {
3895 set_bit(STRIPE_EXPAND_READY, &sh->state);
3896 set_bit(STRIPE_HANDLE, &sh->state);
3898 list_add(&sh->lru, &stripes);
3900 spin_lock_irq(&conf->device_lock);
3901 if (mddev->delta_disks < 0)
3902 conf->reshape_progress -= reshape_sectors * new_data_disks;
3904 conf->reshape_progress += reshape_sectors * new_data_disks;
3905 spin_unlock_irq(&conf->device_lock);
3906 /* Ok, those stripe are ready. We can start scheduling
3907 * reads on the source stripes.
3908 * The source stripes are determined by mapping the first and last
3909 * block on the destination stripes.
3912 raid5_compute_sector(conf, stripe_addr*(new_data_disks),
3915 raid5_compute_sector(conf, ((stripe_addr+reshape_sectors)
3916 * new_data_disks - 1),
3918 if (last_sector >= mddev->dev_sectors)
3919 last_sector = mddev->dev_sectors - 1;
3920 while (first_sector <= last_sector) {
3921 sh = get_active_stripe(conf, first_sector, 1, 0, 1);
3922 set_bit(STRIPE_EXPAND_SOURCE, &sh->state);
3923 set_bit(STRIPE_HANDLE, &sh->state);
3925 first_sector += STRIPE_SECTORS;
3927 /* Now that the sources are clearly marked, we can release
3928 * the destination stripes
3930 while (!list_empty(&stripes)) {
3931 sh = list_entry(stripes.next, struct stripe_head, lru);
3932 list_del_init(&sh->lru);
3935 /* If this takes us to the resync_max point where we have to pause,
3936 * then we need to write out the superblock.
3938 sector_nr += reshape_sectors;
3939 if ((sector_nr - mddev->curr_resync_completed) * 2
3940 >= mddev->resync_max - mddev->curr_resync_completed) {
3941 /* Cannot proceed until we've updated the superblock... */
3942 wait_event(conf->wait_for_overlap,
3943 atomic_read(&conf->reshape_stripes) == 0);
3944 mddev->reshape_position = conf->reshape_progress;
3945 mddev->curr_resync_completed = sector_nr;
3946 conf->reshape_checkpoint = jiffies;
3947 set_bit(MD_CHANGE_DEVS, &mddev->flags);
3948 md_wakeup_thread(mddev->thread);
3949 wait_event(mddev->sb_wait,
3950 !test_bit(MD_CHANGE_DEVS, &mddev->flags)
3951 || kthread_should_stop());
3952 spin_lock_irq(&conf->device_lock);
3953 conf->reshape_safe = mddev->reshape_position;
3954 spin_unlock_irq(&conf->device_lock);
3955 wake_up(&conf->wait_for_overlap);
3956 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
3958 return reshape_sectors;
3961 /* FIXME go_faster isn't used */
3962 static inline sector_t sync_request(mddev_t *mddev, sector_t sector_nr, int *skipped, int go_faster)
3964 raid5_conf_t *conf = mddev->private;
3965 struct stripe_head *sh;
3966 sector_t max_sector = mddev->dev_sectors;
3967 sector_t sync_blocks;
3968 int still_degraded = 0;
3971 if (sector_nr >= max_sector) {
3972 /* just being told to finish up .. nothing much to do */
3974 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery)) {
3979 if (mddev->curr_resync < max_sector) /* aborted */
3980 bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
3982 else /* completed sync */
3984 bitmap_close_sync(mddev->bitmap);
3989 /* Allow raid5_quiesce to complete */
3990 wait_event(conf->wait_for_overlap, conf->quiesce != 2);
3992 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
3993 return reshape_request(mddev, sector_nr, skipped);
3995 /* No need to check resync_max as we never do more than one
3996 * stripe, and as resync_max will always be on a chunk boundary,
3997 * if the check in md_do_sync didn't fire, there is no chance
3998 * of overstepping resync_max here
4001 /* if there is too many failed drives and we are trying
4002 * to resync, then assert that we are finished, because there is
4003 * nothing we can do.
4005 if (mddev->degraded >= conf->max_degraded &&
4006 test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
4007 sector_t rv = mddev->dev_sectors - sector_nr;
4011 if (!bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
4012 !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
4013 !conf->fullsync && sync_blocks >= STRIPE_SECTORS) {
4014 /* we can skip this block, and probably more */
4015 sync_blocks /= STRIPE_SECTORS;
4017 return sync_blocks * STRIPE_SECTORS; /* keep things rounded to whole stripes */
4021 bitmap_cond_end_sync(mddev->bitmap, sector_nr);
4023 sh = get_active_stripe(conf, sector_nr, 0, 1, 0);
4025 sh = get_active_stripe(conf, sector_nr, 0, 0, 0);
4026 /* make sure we don't swamp the stripe cache if someone else
4027 * is trying to get access
4029 schedule_timeout_uninterruptible(1);
4031 /* Need to check if array will still be degraded after recovery/resync
4032 * We don't need to check the 'failed' flag as when that gets set,
4035 for (i = 0; i < conf->raid_disks; i++)
4036 if (conf->disks[i].rdev == NULL)
4039 bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, still_degraded);
4041 set_bit(STRIPE_SYNC_REQUESTED, &sh->state);
4046 return STRIPE_SECTORS;
4049 static int retry_aligned_read(raid5_conf_t *conf, struct bio *raid_bio)
4051 /* We may not be able to submit a whole bio at once as there
4052 * may not be enough stripe_heads available.
4053 * We cannot pre-allocate enough stripe_heads as we may need
4054 * more than exist in the cache (if we allow ever large chunks).
4055 * So we do one stripe head at a time and record in
4056 * ->bi_hw_segments how many have been done.
4058 * We *know* that this entire raid_bio is in one chunk, so
4059 * it will be only one 'dd_idx' and only need one call to raid5_compute_sector.
4061 struct stripe_head *sh;
4063 sector_t sector, logical_sector, last_sector;
4068 logical_sector = raid_bio->bi_sector & ~((sector_t)STRIPE_SECTORS-1);
4069 sector = raid5_compute_sector(conf, logical_sector,
4071 last_sector = raid_bio->bi_sector + (raid_bio->bi_size>>9);
4073 for (; logical_sector < last_sector;
4074 logical_sector += STRIPE_SECTORS,
4075 sector += STRIPE_SECTORS,
4078 if (scnt < raid5_bi_hw_segments(raid_bio))
4079 /* already done this stripe */
4082 sh = get_active_stripe(conf, sector, 0, 1, 0);
4085 /* failed to get a stripe - must wait */
4086 raid5_set_bi_hw_segments(raid_bio, scnt);
4087 conf->retry_read_aligned = raid_bio;
4091 set_bit(R5_ReadError, &sh->dev[dd_idx].flags);
4092 if (!add_stripe_bio(sh, raid_bio, dd_idx, 0)) {
4094 raid5_set_bi_hw_segments(raid_bio, scnt);
4095 conf->retry_read_aligned = raid_bio;
4103 spin_lock_irq(&conf->device_lock);
4104 remaining = raid5_dec_bi_phys_segments(raid_bio);
4105 spin_unlock_irq(&conf->device_lock);
4107 bio_endio(raid_bio, 0);
4108 if (atomic_dec_and_test(&conf->active_aligned_reads))
4109 wake_up(&conf->wait_for_stripe);
4115 * This is our raid5 kernel thread.
4117 * We scan the hash table for stripes which can be handled now.
4118 * During the scan, completed stripes are saved for us by the interrupt
4119 * handler, so that they will not have to wait for our next wakeup.
4121 static void raid5d(mddev_t *mddev)
4123 struct stripe_head *sh;
4124 raid5_conf_t *conf = mddev->private;
4126 struct blk_plug plug;
4128 pr_debug("+++ raid5d active\n");
4130 md_check_recovery(mddev);
4132 blk_start_plug(&plug);
4134 spin_lock_irq(&conf->device_lock);
4138 if (atomic_read(&mddev->plug_cnt) == 0 &&
4139 !list_empty(&conf->bitmap_list)) {
4140 /* Now is a good time to flush some bitmap updates */
4142 spin_unlock_irq(&conf->device_lock);
4143 bitmap_unplug(mddev->bitmap);
4144 spin_lock_irq(&conf->device_lock);
4145 conf->seq_write = conf->seq_flush;
4146 activate_bit_delay(conf);
4148 if (atomic_read(&mddev->plug_cnt) == 0)
4149 raid5_activate_delayed(conf);
4151 while ((bio = remove_bio_from_retry(conf))) {
4153 spin_unlock_irq(&conf->device_lock);
4154 ok = retry_aligned_read(conf, bio);
4155 spin_lock_irq(&conf->device_lock);
4161 sh = __get_priority_stripe(conf);
4165 spin_unlock_irq(&conf->device_lock);
4172 if (mddev->flags & ~(1<<MD_CHANGE_PENDING))
4173 md_check_recovery(mddev);
4175 spin_lock_irq(&conf->device_lock);
4177 pr_debug("%d stripes handled\n", handled);
4179 spin_unlock_irq(&conf->device_lock);
4181 async_tx_issue_pending_all();
4182 blk_finish_plug(&plug);
4184 pr_debug("--- raid5d inactive\n");
4188 raid5_show_stripe_cache_size(mddev_t *mddev, char *page)
4190 raid5_conf_t *conf = mddev->private;
4192 return sprintf(page, "%d\n", conf->max_nr_stripes);
4198 raid5_set_cache_size(mddev_t *mddev, int size)
4200 raid5_conf_t *conf = mddev->private;
4203 if (size <= 16 || size > 32768)
4205 while (size < conf->max_nr_stripes) {
4206 if (drop_one_stripe(conf))
4207 conf->max_nr_stripes--;
4211 err = md_allow_write(mddev);
4214 while (size > conf->max_nr_stripes) {
4215 if (grow_one_stripe(conf))
4216 conf->max_nr_stripes++;
4221 EXPORT_SYMBOL(raid5_set_cache_size);
4224 raid5_store_stripe_cache_size(mddev_t *mddev, const char *page, size_t len)
4226 raid5_conf_t *conf = mddev->private;
4230 if (len >= PAGE_SIZE)
4235 if (strict_strtoul(page, 10, &new))
4237 err = raid5_set_cache_size(mddev, new);
4243 static struct md_sysfs_entry
4244 raid5_stripecache_size = __ATTR(stripe_cache_size, S_IRUGO | S_IWUSR,
4245 raid5_show_stripe_cache_size,
4246 raid5_store_stripe_cache_size);
4249 raid5_show_preread_threshold(mddev_t *mddev, char *page)
4251 raid5_conf_t *conf = mddev->private;
4253 return sprintf(page, "%d\n", conf->bypass_threshold);
4259 raid5_store_preread_threshold(mddev_t *mddev, const char *page, size_t len)
4261 raid5_conf_t *conf = mddev->private;
4263 if (len >= PAGE_SIZE)
4268 if (strict_strtoul(page, 10, &new))
4270 if (new > conf->max_nr_stripes)
4272 conf->bypass_threshold = new;
4276 static struct md_sysfs_entry
4277 raid5_preread_bypass_threshold = __ATTR(preread_bypass_threshold,
4279 raid5_show_preread_threshold,
4280 raid5_store_preread_threshold);
4283 stripe_cache_active_show(mddev_t *mddev, char *page)
4285 raid5_conf_t *conf = mddev->private;
4287 return sprintf(page, "%d\n", atomic_read(&conf->active_stripes));
4292 static struct md_sysfs_entry
4293 raid5_stripecache_active = __ATTR_RO(stripe_cache_active);
4295 static struct attribute *raid5_attrs[] = {
4296 &raid5_stripecache_size.attr,
4297 &raid5_stripecache_active.attr,
4298 &raid5_preread_bypass_threshold.attr,
4301 static struct attribute_group raid5_attrs_group = {
4303 .attrs = raid5_attrs,
4307 raid5_size(mddev_t *mddev, sector_t sectors, int raid_disks)
4309 raid5_conf_t *conf = mddev->private;
4312 sectors = mddev->dev_sectors;
4314 /* size is defined by the smallest of previous and new size */
4315 raid_disks = min(conf->raid_disks, conf->previous_raid_disks);
4317 sectors &= ~((sector_t)mddev->chunk_sectors - 1);
4318 sectors &= ~((sector_t)mddev->new_chunk_sectors - 1);
4319 return sectors * (raid_disks - conf->max_degraded);
4322 static void raid5_free_percpu(raid5_conf_t *conf)
4324 struct raid5_percpu *percpu;
4331 for_each_possible_cpu(cpu) {
4332 percpu = per_cpu_ptr(conf->percpu, cpu);
4333 safe_put_page(percpu->spare_page);
4334 kfree(percpu->scribble);
4336 #ifdef CONFIG_HOTPLUG_CPU
4337 unregister_cpu_notifier(&conf->cpu_notify);
4341 free_percpu(conf->percpu);
4344 static void free_conf(raid5_conf_t *conf)
4346 shrink_stripes(conf);
4347 raid5_free_percpu(conf);
4349 kfree(conf->stripe_hashtbl);
4353 #ifdef CONFIG_HOTPLUG_CPU
4354 static int raid456_cpu_notify(struct notifier_block *nfb, unsigned long action,
4357 raid5_conf_t *conf = container_of(nfb, raid5_conf_t, cpu_notify);
4358 long cpu = (long)hcpu;
4359 struct raid5_percpu *percpu = per_cpu_ptr(conf->percpu, cpu);
4362 case CPU_UP_PREPARE:
4363 case CPU_UP_PREPARE_FROZEN:
4364 if (conf->level == 6 && !percpu->spare_page)
4365 percpu->spare_page = alloc_page(GFP_KERNEL);
4366 if (!percpu->scribble)
4367 percpu->scribble = kmalloc(conf->scribble_len, GFP_KERNEL);
4369 if (!percpu->scribble ||
4370 (conf->level == 6 && !percpu->spare_page)) {
4371 safe_put_page(percpu->spare_page);
4372 kfree(percpu->scribble);
4373 pr_err("%s: failed memory allocation for cpu%ld\n",
4375 return notifier_from_errno(-ENOMEM);
4379 case CPU_DEAD_FROZEN:
4380 safe_put_page(percpu->spare_page);
4381 kfree(percpu->scribble);
4382 percpu->spare_page = NULL;
4383 percpu->scribble = NULL;
4392 static int raid5_alloc_percpu(raid5_conf_t *conf)
4395 struct page *spare_page;
4396 struct raid5_percpu __percpu *allcpus;
4400 allcpus = alloc_percpu(struct raid5_percpu);
4403 conf->percpu = allcpus;
4407 for_each_present_cpu(cpu) {
4408 if (conf->level == 6) {
4409 spare_page = alloc_page(GFP_KERNEL);
4414 per_cpu_ptr(conf->percpu, cpu)->spare_page = spare_page;
4416 scribble = kmalloc(conf->scribble_len, GFP_KERNEL);
4421 per_cpu_ptr(conf->percpu, cpu)->scribble = scribble;
4423 #ifdef CONFIG_HOTPLUG_CPU
4424 conf->cpu_notify.notifier_call = raid456_cpu_notify;
4425 conf->cpu_notify.priority = 0;
4427 err = register_cpu_notifier(&conf->cpu_notify);
4434 static raid5_conf_t *setup_conf(mddev_t *mddev)
4437 int raid_disk, memory, max_disks;
4439 struct disk_info *disk;
4441 if (mddev->new_level != 5
4442 && mddev->new_level != 4
4443 && mddev->new_level != 6) {
4444 printk(KERN_ERR "md/raid:%s: raid level not set to 4/5/6 (%d)\n",
4445 mdname(mddev), mddev->new_level);
4446 return ERR_PTR(-EIO);
4448 if ((mddev->new_level == 5
4449 && !algorithm_valid_raid5(mddev->new_layout)) ||
4450 (mddev->new_level == 6
4451 && !algorithm_valid_raid6(mddev->new_layout))) {
4452 printk(KERN_ERR "md/raid:%s: layout %d not supported\n",
4453 mdname(mddev), mddev->new_layout);
4454 return ERR_PTR(-EIO);
4456 if (mddev->new_level == 6 && mddev->raid_disks < 4) {
4457 printk(KERN_ERR "md/raid:%s: not enough configured devices (%d, minimum 4)\n",
4458 mdname(mddev), mddev->raid_disks);
4459 return ERR_PTR(-EINVAL);
4462 if (!mddev->new_chunk_sectors ||
4463 (mddev->new_chunk_sectors << 9) % PAGE_SIZE ||
4464 !is_power_of_2(mddev->new_chunk_sectors)) {
4465 printk(KERN_ERR "md/raid:%s: invalid chunk size %d\n",
4466 mdname(mddev), mddev->new_chunk_sectors << 9);
4467 return ERR_PTR(-EINVAL);
4470 conf = kzalloc(sizeof(raid5_conf_t), GFP_KERNEL);
4473 spin_lock_init(&conf->device_lock);
4474 init_waitqueue_head(&conf->wait_for_stripe);
4475 init_waitqueue_head(&conf->wait_for_overlap);
4476 INIT_LIST_HEAD(&conf->handle_list);
4477 INIT_LIST_HEAD(&conf->hold_list);
4478 INIT_LIST_HEAD(&conf->delayed_list);
4479 INIT_LIST_HEAD(&conf->bitmap_list);
4480 INIT_LIST_HEAD(&conf->inactive_list);
4481 atomic_set(&conf->active_stripes, 0);
4482 atomic_set(&conf->preread_active_stripes, 0);
4483 atomic_set(&conf->active_aligned_reads, 0);
4484 conf->bypass_threshold = BYPASS_THRESHOLD;
4486 conf->raid_disks = mddev->raid_disks;
4487 if (mddev->reshape_position == MaxSector)
4488 conf->previous_raid_disks = mddev->raid_disks;
4490 conf->previous_raid_disks = mddev->raid_disks - mddev->delta_disks;
4491 max_disks = max(conf->raid_disks, conf->previous_raid_disks);
4492 conf->scribble_len = scribble_len(max_disks);
4494 conf->disks = kzalloc(max_disks * sizeof(struct disk_info),
4499 conf->mddev = mddev;
4501 if ((conf->stripe_hashtbl = kzalloc(PAGE_SIZE, GFP_KERNEL)) == NULL)
4504 conf->level = mddev->new_level;
4505 if (raid5_alloc_percpu(conf) != 0)
4508 pr_debug("raid456: run(%s) called.\n", mdname(mddev));
4510 list_for_each_entry(rdev, &mddev->disks, same_set) {
4511 raid_disk = rdev->raid_disk;
4512 if (raid_disk >= max_disks
4515 disk = conf->disks + raid_disk;
4519 if (test_bit(In_sync, &rdev->flags)) {
4520 char b[BDEVNAME_SIZE];
4521 printk(KERN_INFO "md/raid:%s: device %s operational as raid"
4523 mdname(mddev), bdevname(rdev->bdev, b), raid_disk);
4524 } else if (rdev->saved_raid_disk != raid_disk)
4525 /* Cannot rely on bitmap to complete recovery */
4529 conf->chunk_sectors = mddev->new_chunk_sectors;
4530 conf->level = mddev->new_level;
4531 if (conf->level == 6)
4532 conf->max_degraded = 2;
4534 conf->max_degraded = 1;
4535 conf->algorithm = mddev->new_layout;
4536 conf->max_nr_stripes = NR_STRIPES;
4537 conf->reshape_progress = mddev->reshape_position;
4538 if (conf->reshape_progress != MaxSector) {
4539 conf->prev_chunk_sectors = mddev->chunk_sectors;
4540 conf->prev_algo = mddev->layout;
4543 memory = conf->max_nr_stripes * (sizeof(struct stripe_head) +
4544 max_disks * ((sizeof(struct bio) + PAGE_SIZE))) / 1024;
4545 if (grow_stripes(conf, conf->max_nr_stripes)) {
4547 "md/raid:%s: couldn't allocate %dkB for buffers\n",
4548 mdname(mddev), memory);
4551 printk(KERN_INFO "md/raid:%s: allocated %dkB\n",
4552 mdname(mddev), memory);
4554 conf->thread = md_register_thread(raid5d, mddev, NULL);
4555 if (!conf->thread) {
4557 "md/raid:%s: couldn't allocate thread.\n",
4567 return ERR_PTR(-EIO);
4569 return ERR_PTR(-ENOMEM);
4573 static int only_parity(int raid_disk, int algo, int raid_disks, int max_degraded)
4576 case ALGORITHM_PARITY_0:
4577 if (raid_disk < max_degraded)
4580 case ALGORITHM_PARITY_N:
4581 if (raid_disk >= raid_disks - max_degraded)
4584 case ALGORITHM_PARITY_0_6:
4585 if (raid_disk == 0 ||
4586 raid_disk == raid_disks - 1)
4589 case ALGORITHM_LEFT_ASYMMETRIC_6:
4590 case ALGORITHM_RIGHT_ASYMMETRIC_6:
4591 case ALGORITHM_LEFT_SYMMETRIC_6:
4592 case ALGORITHM_RIGHT_SYMMETRIC_6:
4593 if (raid_disk == raid_disks - 1)
4599 static int run(mddev_t *mddev)
4602 int working_disks = 0;
4603 int dirty_parity_disks = 0;
4605 sector_t reshape_offset = 0;
4607 if (mddev->recovery_cp != MaxSector)
4608 printk(KERN_NOTICE "md/raid:%s: not clean"
4609 " -- starting background reconstruction\n",
4611 if (mddev->reshape_position != MaxSector) {
4612 /* Check that we can continue the reshape.
4613 * Currently only disks can change, it must
4614 * increase, and we must be past the point where
4615 * a stripe over-writes itself
4617 sector_t here_new, here_old;
4619 int max_degraded = (mddev->level == 6 ? 2 : 1);
4621 if (mddev->new_level != mddev->level) {
4622 printk(KERN_ERR "md/raid:%s: unsupported reshape "
4623 "required - aborting.\n",
4627 old_disks = mddev->raid_disks - mddev->delta_disks;
4628 /* reshape_position must be on a new-stripe boundary, and one
4629 * further up in new geometry must map after here in old
4632 here_new = mddev->reshape_position;
4633 if (sector_div(here_new, mddev->new_chunk_sectors *
4634 (mddev->raid_disks - max_degraded))) {
4635 printk(KERN_ERR "md/raid:%s: reshape_position not "
4636 "on a stripe boundary\n", mdname(mddev));
4639 reshape_offset = here_new * mddev->new_chunk_sectors;
4640 /* here_new is the stripe we will write to */
4641 here_old = mddev->reshape_position;
4642 sector_div(here_old, mddev->chunk_sectors *
4643 (old_disks-max_degraded));
4644 /* here_old is the first stripe that we might need to read
4646 if (mddev->delta_disks == 0) {
4647 /* We cannot be sure it is safe to start an in-place
4648 * reshape. It is only safe if user-space if monitoring
4649 * and taking constant backups.
4650 * mdadm always starts a situation like this in
4651 * readonly mode so it can take control before
4652 * allowing any writes. So just check for that.
4654 if ((here_new * mddev->new_chunk_sectors !=
4655 here_old * mddev->chunk_sectors) ||
4657 printk(KERN_ERR "md/raid:%s: in-place reshape must be started"
4658 " in read-only mode - aborting\n",
4662 } else if (mddev->delta_disks < 0
4663 ? (here_new * mddev->new_chunk_sectors <=
4664 here_old * mddev->chunk_sectors)
4665 : (here_new * mddev->new_chunk_sectors >=
4666 here_old * mddev->chunk_sectors)) {
4667 /* Reading from the same stripe as writing to - bad */
4668 printk(KERN_ERR "md/raid:%s: reshape_position too early for "
4669 "auto-recovery - aborting.\n",
4673 printk(KERN_INFO "md/raid:%s: reshape will continue\n",
4675 /* OK, we should be able to continue; */
4677 BUG_ON(mddev->level != mddev->new_level);
4678 BUG_ON(mddev->layout != mddev->new_layout);
4679 BUG_ON(mddev->chunk_sectors != mddev->new_chunk_sectors);
4680 BUG_ON(mddev->delta_disks != 0);
4683 if (mddev->private == NULL)
4684 conf = setup_conf(mddev);
4686 conf = mddev->private;
4689 return PTR_ERR(conf);
4691 mddev->thread = conf->thread;
4692 conf->thread = NULL;
4693 mddev->private = conf;
4696 * 0 for a fully functional array, 1 or 2 for a degraded array.
4698 list_for_each_entry(rdev, &mddev->disks, same_set) {
4699 if (rdev->raid_disk < 0)
4701 if (test_bit(In_sync, &rdev->flags)) {
4705 /* This disc is not fully in-sync. However if it
4706 * just stored parity (beyond the recovery_offset),
4707 * when we don't need to be concerned about the
4708 * array being dirty.
4709 * When reshape goes 'backwards', we never have
4710 * partially completed devices, so we only need
4711 * to worry about reshape going forwards.
4713 /* Hack because v0.91 doesn't store recovery_offset properly. */
4714 if (mddev->major_version == 0 &&
4715 mddev->minor_version > 90)
4716 rdev->recovery_offset = reshape_offset;
4718 if (rdev->recovery_offset < reshape_offset) {
4719 /* We need to check old and new layout */
4720 if (!only_parity(rdev->raid_disk,
4723 conf->max_degraded))
4726 if (!only_parity(rdev->raid_disk,
4728 conf->previous_raid_disks,
4729 conf->max_degraded))
4731 dirty_parity_disks++;
4734 mddev->degraded = (max(conf->raid_disks, conf->previous_raid_disks)
4737 if (has_failed(conf)) {
4738 printk(KERN_ERR "md/raid:%s: not enough operational devices"
4739 " (%d/%d failed)\n",
4740 mdname(mddev), mddev->degraded, conf->raid_disks);
4744 /* device size must be a multiple of chunk size */
4745 mddev->dev_sectors &= ~(mddev->chunk_sectors - 1);
4746 mddev->resync_max_sectors = mddev->dev_sectors;
4748 if (mddev->degraded > dirty_parity_disks &&
4749 mddev->recovery_cp != MaxSector) {
4750 if (mddev->ok_start_degraded)
4752 "md/raid:%s: starting dirty degraded array"
4753 " - data corruption possible.\n",
4757 "md/raid:%s: cannot start dirty degraded array.\n",
4763 if (mddev->degraded == 0)
4764 printk(KERN_INFO "md/raid:%s: raid level %d active with %d out of %d"
4765 " devices, algorithm %d\n", mdname(mddev), conf->level,
4766 mddev->raid_disks-mddev->degraded, mddev->raid_disks,
4769 printk(KERN_ALERT "md/raid:%s: raid level %d active with %d"
4770 " out of %d devices, algorithm %d\n",
4771 mdname(mddev), conf->level,
4772 mddev->raid_disks - mddev->degraded,
4773 mddev->raid_disks, mddev->new_layout);
4775 print_raid5_conf(conf);
4777 if (conf->reshape_progress != MaxSector) {
4778 conf->reshape_safe = conf->reshape_progress;
4779 atomic_set(&conf->reshape_stripes, 0);
4780 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
4781 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
4782 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
4783 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
4784 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
4789 /* Ok, everything is just fine now */
4790 if (mddev->to_remove == &raid5_attrs_group)
4791 mddev->to_remove = NULL;
4792 else if (mddev->kobj.sd &&
4793 sysfs_create_group(&mddev->kobj, &raid5_attrs_group))
4795 "raid5: failed to create sysfs attributes for %s\n",
4797 md_set_array_sectors(mddev, raid5_size(mddev, 0, 0));
4801 /* read-ahead size must cover two whole stripes, which
4802 * is 2 * (datadisks) * chunksize where 'n' is the
4803 * number of raid devices
4805 int data_disks = conf->previous_raid_disks - conf->max_degraded;
4806 int stripe = data_disks *
4807 ((mddev->chunk_sectors << 9) / PAGE_SIZE);
4808 if (mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
4809 mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
4811 blk_queue_merge_bvec(mddev->queue, raid5_mergeable_bvec);
4813 mddev->queue->backing_dev_info.congested_data = mddev;
4814 mddev->queue->backing_dev_info.congested_fn = raid5_congested;
4816 chunk_size = mddev->chunk_sectors << 9;
4817 blk_queue_io_min(mddev->queue, chunk_size);
4818 blk_queue_io_opt(mddev->queue, chunk_size *
4819 (conf->raid_disks - conf->max_degraded));
4821 list_for_each_entry(rdev, &mddev->disks, same_set)
4822 disk_stack_limits(mddev->gendisk, rdev->bdev,
4823 rdev->data_offset << 9);
4828 md_unregister_thread(mddev->thread);
4829 mddev->thread = NULL;
4831 print_raid5_conf(conf);
4834 mddev->private = NULL;
4835 printk(KERN_ALERT "md/raid:%s: failed to run raid set.\n", mdname(mddev));
4839 static int stop(mddev_t *mddev)
4841 raid5_conf_t *conf = mddev->private;
4843 md_unregister_thread(mddev->thread);
4844 mddev->thread = NULL;
4846 mddev->queue->backing_dev_info.congested_fn = NULL;
4848 mddev->private = NULL;
4849 mddev->to_remove = &raid5_attrs_group;
4854 static void print_sh(struct seq_file *seq, struct stripe_head *sh)
4858 seq_printf(seq, "sh %llu, pd_idx %d, state %ld.\n",
4859 (unsigned long long)sh->sector, sh->pd_idx, sh->state);
4860 seq_printf(seq, "sh %llu, count %d.\n",
4861 (unsigned long long)sh->sector, atomic_read(&sh->count));
4862 seq_printf(seq, "sh %llu, ", (unsigned long long)sh->sector);
4863 for (i = 0; i < sh->disks; i++) {
4864 seq_printf(seq, "(cache%d: %p %ld) ",
4865 i, sh->dev[i].page, sh->dev[i].flags);
4867 seq_printf(seq, "\n");
4870 static void printall(struct seq_file *seq, raid5_conf_t *conf)
4872 struct stripe_head *sh;
4873 struct hlist_node *hn;
4876 spin_lock_irq(&conf->device_lock);
4877 for (i = 0; i < NR_HASH; i++) {
4878 hlist_for_each_entry(sh, hn, &conf->stripe_hashtbl[i], hash) {
4879 if (sh->raid_conf != conf)
4884 spin_unlock_irq(&conf->device_lock);
4888 static void status(struct seq_file *seq, mddev_t *mddev)
4890 raid5_conf_t *conf = mddev->private;
4893 seq_printf(seq, " level %d, %dk chunk, algorithm %d", mddev->level,
4894 mddev->chunk_sectors / 2, mddev->layout);
4895 seq_printf (seq, " [%d/%d] [", conf->raid_disks, conf->raid_disks - mddev->degraded);
4896 for (i = 0; i < conf->raid_disks; i++)
4897 seq_printf (seq, "%s",
4898 conf->disks[i].rdev &&
4899 test_bit(In_sync, &conf->disks[i].rdev->flags) ? "U" : "_");
4900 seq_printf (seq, "]");
4902 seq_printf (seq, "\n");
4903 printall(seq, conf);
4907 static void print_raid5_conf (raid5_conf_t *conf)
4910 struct disk_info *tmp;
4912 printk(KERN_DEBUG "RAID conf printout:\n");
4914 printk("(conf==NULL)\n");
4917 printk(KERN_DEBUG " --- level:%d rd:%d wd:%d\n", conf->level,
4919 conf->raid_disks - conf->mddev->degraded);
4921 for (i = 0; i < conf->raid_disks; i++) {
4922 char b[BDEVNAME_SIZE];
4923 tmp = conf->disks + i;
4925 printk(KERN_DEBUG " disk %d, o:%d, dev:%s\n",
4926 i, !test_bit(Faulty, &tmp->rdev->flags),
4927 bdevname(tmp->rdev->bdev, b));
4931 static int raid5_spare_active(mddev_t *mddev)
4934 raid5_conf_t *conf = mddev->private;
4935 struct disk_info *tmp;
4937 unsigned long flags;
4939 for (i = 0; i < conf->raid_disks; i++) {
4940 tmp = conf->disks + i;
4942 && tmp->rdev->recovery_offset == MaxSector
4943 && !test_bit(Faulty, &tmp->rdev->flags)
4944 && !test_and_set_bit(In_sync, &tmp->rdev->flags)) {
4946 sysfs_notify_dirent_safe(tmp->rdev->sysfs_state);
4949 spin_lock_irqsave(&conf->device_lock, flags);
4950 mddev->degraded -= count;
4951 spin_unlock_irqrestore(&conf->device_lock, flags);
4952 print_raid5_conf(conf);
4956 static int raid5_remove_disk(mddev_t *mddev, int number)
4958 raid5_conf_t *conf = mddev->private;
4961 struct disk_info *p = conf->disks + number;
4963 print_raid5_conf(conf);
4966 if (number >= conf->raid_disks &&
4967 conf->reshape_progress == MaxSector)
4968 clear_bit(In_sync, &rdev->flags);
4970 if (test_bit(In_sync, &rdev->flags) ||
4971 atomic_read(&rdev->nr_pending)) {
4975 /* Only remove non-faulty devices if recovery
4978 if (!test_bit(Faulty, &rdev->flags) &&
4979 !has_failed(conf) &&
4980 number < conf->raid_disks) {
4986 if (atomic_read(&rdev->nr_pending)) {
4987 /* lost the race, try later */
4994 print_raid5_conf(conf);
4998 static int raid5_add_disk(mddev_t *mddev, mdk_rdev_t *rdev)
5000 raid5_conf_t *conf = mddev->private;
5003 struct disk_info *p;
5005 int last = conf->raid_disks - 1;
5007 if (has_failed(conf))
5008 /* no point adding a device */
5011 if (rdev->raid_disk >= 0)
5012 first = last = rdev->raid_disk;
5015 * find the disk ... but prefer rdev->saved_raid_disk
5018 if (rdev->saved_raid_disk >= 0 &&
5019 rdev->saved_raid_disk >= first &&
5020 conf->disks[rdev->saved_raid_disk].rdev == NULL)
5021 disk = rdev->saved_raid_disk;
5024 for ( ; disk <= last ; disk++)
5025 if ((p=conf->disks + disk)->rdev == NULL) {
5026 clear_bit(In_sync, &rdev->flags);
5027 rdev->raid_disk = disk;
5029 if (rdev->saved_raid_disk != disk)
5031 rcu_assign_pointer(p->rdev, rdev);
5034 print_raid5_conf(conf);
5038 static int raid5_resize(mddev_t *mddev, sector_t sectors)
5040 /* no resync is happening, and there is enough space
5041 * on all devices, so we can resize.
5042 * We need to make sure resync covers any new space.
5043 * If the array is shrinking we should possibly wait until
5044 * any io in the removed space completes, but it hardly seems
5047 sectors &= ~((sector_t)mddev->chunk_sectors - 1);
5048 md_set_array_sectors(mddev, raid5_size(mddev, sectors,
5049 mddev->raid_disks));
5050 if (mddev->array_sectors >
5051 raid5_size(mddev, sectors, mddev->raid_disks))
5053 set_capacity(mddev->gendisk, mddev->array_sectors);
5054 revalidate_disk(mddev->gendisk);
5055 if (sectors > mddev->dev_sectors &&
5056 mddev->recovery_cp > mddev->dev_sectors) {
5057 mddev->recovery_cp = mddev->dev_sectors;
5058 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
5060 mddev->dev_sectors = sectors;
5061 mddev->resync_max_sectors = sectors;
5065 static int check_stripe_cache(mddev_t *mddev)
5067 /* Can only proceed if there are plenty of stripe_heads.
5068 * We need a minimum of one full stripe,, and for sensible progress
5069 * it is best to have about 4 times that.
5070 * If we require 4 times, then the default 256 4K stripe_heads will
5071 * allow for chunk sizes up to 256K, which is probably OK.
5072 * If the chunk size is greater, user-space should request more
5073 * stripe_heads first.
5075 raid5_conf_t *conf = mddev->private;
5076 if (((mddev->chunk_sectors << 9) / STRIPE_SIZE) * 4
5077 > conf->max_nr_stripes ||
5078 ((mddev->new_chunk_sectors << 9) / STRIPE_SIZE) * 4
5079 > conf->max_nr_stripes) {
5080 printk(KERN_WARNING "md/raid:%s: reshape: not enough stripes. Needed %lu\n",
5082 ((max(mddev->chunk_sectors, mddev->new_chunk_sectors) << 9)
5089 static int check_reshape(mddev_t *mddev)
5091 raid5_conf_t *conf = mddev->private;
5093 if (mddev->delta_disks == 0 &&
5094 mddev->new_layout == mddev->layout &&
5095 mddev->new_chunk_sectors == mddev->chunk_sectors)
5096 return 0; /* nothing to do */
5098 /* Cannot grow a bitmap yet */
5100 if (has_failed(conf))
5102 if (mddev->delta_disks < 0) {
5103 /* We might be able to shrink, but the devices must
5104 * be made bigger first.
5105 * For raid6, 4 is the minimum size.
5106 * Otherwise 2 is the minimum
5109 if (mddev->level == 6)
5111 if (mddev->raid_disks + mddev->delta_disks < min)
5115 if (!check_stripe_cache(mddev))
5118 return resize_stripes(conf, conf->raid_disks + mddev->delta_disks);
5121 static int raid5_start_reshape(mddev_t *mddev)
5123 raid5_conf_t *conf = mddev->private;
5126 unsigned long flags;
5128 if (test_bit(MD_RECOVERY_RUNNING, &mddev->recovery))
5131 if (!check_stripe_cache(mddev))
5134 list_for_each_entry(rdev, &mddev->disks, same_set)
5135 if (!test_bit(In_sync, &rdev->flags)
5136 && !test_bit(Faulty, &rdev->flags))
5139 if (spares - mddev->degraded < mddev->delta_disks - conf->max_degraded)
5140 /* Not enough devices even to make a degraded array
5145 /* Refuse to reduce size of the array. Any reductions in
5146 * array size must be through explicit setting of array_size
5149 if (raid5_size(mddev, 0, conf->raid_disks + mddev->delta_disks)
5150 < mddev->array_sectors) {
5151 printk(KERN_ERR "md/raid:%s: array size must be reduced "
5152 "before number of disks\n", mdname(mddev));
5156 atomic_set(&conf->reshape_stripes, 0);
5157 spin_lock_irq(&conf->device_lock);
5158 conf->previous_raid_disks = conf->raid_disks;
5159 conf->raid_disks += mddev->delta_disks;
5160 conf->prev_chunk_sectors = conf->chunk_sectors;
5161 conf->chunk_sectors = mddev->new_chunk_sectors;
5162 conf->prev_algo = conf->algorithm;
5163 conf->algorithm = mddev->new_layout;
5164 if (mddev->delta_disks < 0)
5165 conf->reshape_progress = raid5_size(mddev, 0, 0);
5167 conf->reshape_progress = 0;
5168 conf->reshape_safe = conf->reshape_progress;
5170 spin_unlock_irq(&conf->device_lock);
5172 /* Add some new drives, as many as will fit.
5173 * We know there are enough to make the newly sized array work.
5174 * Don't add devices if we are reducing the number of
5175 * devices in the array. This is because it is not possible
5176 * to correctly record the "partially reconstructed" state of
5177 * such devices during the reshape and confusion could result.
5179 if (mddev->delta_disks >= 0) {
5180 int added_devices = 0;
5181 list_for_each_entry(rdev, &mddev->disks, same_set)
5182 if (rdev->raid_disk < 0 &&
5183 !test_bit(Faulty, &rdev->flags)) {
5184 if (raid5_add_disk(mddev, rdev) == 0) {
5186 >= conf->previous_raid_disks) {
5187 set_bit(In_sync, &rdev->flags);
5190 rdev->recovery_offset = 0;
5192 if (sysfs_link_rdev(mddev, rdev))
5193 /* Failure here is OK */;
5195 } else if (rdev->raid_disk >= conf->previous_raid_disks
5196 && !test_bit(Faulty, &rdev->flags)) {
5197 /* This is a spare that was manually added */
5198 set_bit(In_sync, &rdev->flags);
5202 /* When a reshape changes the number of devices,
5203 * ->degraded is measured against the larger of the
5204 * pre and post number of devices.
5206 spin_lock_irqsave(&conf->device_lock, flags);
5207 mddev->degraded += (conf->raid_disks - conf->previous_raid_disks)
5209 spin_unlock_irqrestore(&conf->device_lock, flags);
5211 mddev->raid_disks = conf->raid_disks;
5212 mddev->reshape_position = conf->reshape_progress;
5213 set_bit(MD_CHANGE_DEVS, &mddev->flags);
5215 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
5216 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
5217 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
5218 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
5219 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
5221 if (!mddev->sync_thread) {
5222 mddev->recovery = 0;
5223 spin_lock_irq(&conf->device_lock);
5224 mddev->raid_disks = conf->raid_disks = conf->previous_raid_disks;
5225 conf->reshape_progress = MaxSector;
5226 spin_unlock_irq(&conf->device_lock);
5229 conf->reshape_checkpoint = jiffies;
5230 md_wakeup_thread(mddev->sync_thread);
5231 md_new_event(mddev);
5235 /* This is called from the reshape thread and should make any
5236 * changes needed in 'conf'
5238 static void end_reshape(raid5_conf_t *conf)
5241 if (!test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery)) {
5243 spin_lock_irq(&conf->device_lock);
5244 conf->previous_raid_disks = conf->raid_disks;
5245 conf->reshape_progress = MaxSector;
5246 spin_unlock_irq(&conf->device_lock);
5247 wake_up(&conf->wait_for_overlap);
5249 /* read-ahead size must cover two whole stripes, which is
5250 * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
5252 if (conf->mddev->queue) {
5253 int data_disks = conf->raid_disks - conf->max_degraded;
5254 int stripe = data_disks * ((conf->chunk_sectors << 9)
5256 if (conf->mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
5257 conf->mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
5262 /* This is called from the raid5d thread with mddev_lock held.
5263 * It makes config changes to the device.
5265 static void raid5_finish_reshape(mddev_t *mddev)
5267 raid5_conf_t *conf = mddev->private;
5269 if (!test_bit(MD_RECOVERY_INTR, &mddev->recovery)) {
5271 if (mddev->delta_disks > 0) {
5272 md_set_array_sectors(mddev, raid5_size(mddev, 0, 0));
5273 set_capacity(mddev->gendisk, mddev->array_sectors);
5274 revalidate_disk(mddev->gendisk);
5277 mddev->degraded = conf->raid_disks;
5278 for (d = 0; d < conf->raid_disks ; d++)
5279 if (conf->disks[d].rdev &&
5281 &conf->disks[d].rdev->flags))
5283 for (d = conf->raid_disks ;
5284 d < conf->raid_disks - mddev->delta_disks;
5286 mdk_rdev_t *rdev = conf->disks[d].rdev;
5287 if (rdev && raid5_remove_disk(mddev, d) == 0) {
5288 sysfs_unlink_rdev(mddev, rdev);
5289 rdev->raid_disk = -1;
5293 mddev->layout = conf->algorithm;
5294 mddev->chunk_sectors = conf->chunk_sectors;
5295 mddev->reshape_position = MaxSector;
5296 mddev->delta_disks = 0;
5300 static void raid5_quiesce(mddev_t *mddev, int state)
5302 raid5_conf_t *conf = mddev->private;
5305 case 2: /* resume for a suspend */
5306 wake_up(&conf->wait_for_overlap);
5309 case 1: /* stop all writes */
5310 spin_lock_irq(&conf->device_lock);
5311 /* '2' tells resync/reshape to pause so that all
5312 * active stripes can drain
5315 wait_event_lock_irq(conf->wait_for_stripe,
5316 atomic_read(&conf->active_stripes) == 0 &&
5317 atomic_read(&conf->active_aligned_reads) == 0,
5318 conf->device_lock, /* nothing */);
5320 spin_unlock_irq(&conf->device_lock);
5321 /* allow reshape to continue */
5322 wake_up(&conf->wait_for_overlap);
5325 case 0: /* re-enable writes */
5326 spin_lock_irq(&conf->device_lock);
5328 wake_up(&conf->wait_for_stripe);
5329 wake_up(&conf->wait_for_overlap);
5330 spin_unlock_irq(&conf->device_lock);
5336 static void *raid45_takeover_raid0(mddev_t *mddev, int level)
5338 struct raid0_private_data *raid0_priv = mddev->private;
5341 /* for raid0 takeover only one zone is supported */
5342 if (raid0_priv->nr_strip_zones > 1) {
5343 printk(KERN_ERR "md/raid:%s: cannot takeover raid0 with more than one zone.\n",
5345 return ERR_PTR(-EINVAL);
5348 sectors = raid0_priv->strip_zone[0].zone_end;
5349 sector_div(sectors, raid0_priv->strip_zone[0].nb_dev);
5350 mddev->dev_sectors = sectors;
5351 mddev->new_level = level;
5352 mddev->new_layout = ALGORITHM_PARITY_N;
5353 mddev->new_chunk_sectors = mddev->chunk_sectors;
5354 mddev->raid_disks += 1;
5355 mddev->delta_disks = 1;
5356 /* make sure it will be not marked as dirty */
5357 mddev->recovery_cp = MaxSector;
5359 return setup_conf(mddev);
5363 static void *raid5_takeover_raid1(mddev_t *mddev)
5367 if (mddev->raid_disks != 2 ||
5368 mddev->degraded > 1)
5369 return ERR_PTR(-EINVAL);
5371 /* Should check if there are write-behind devices? */
5373 chunksect = 64*2; /* 64K by default */
5375 /* The array must be an exact multiple of chunksize */
5376 while (chunksect && (mddev->array_sectors & (chunksect-1)))
5379 if ((chunksect<<9) < STRIPE_SIZE)
5380 /* array size does not allow a suitable chunk size */
5381 return ERR_PTR(-EINVAL);
5383 mddev->new_level = 5;
5384 mddev->new_layout = ALGORITHM_LEFT_SYMMETRIC;
5385 mddev->new_chunk_sectors = chunksect;
5387 return setup_conf(mddev);
5390 static void *raid5_takeover_raid6(mddev_t *mddev)
5394 switch (mddev->layout) {
5395 case ALGORITHM_LEFT_ASYMMETRIC_6:
5396 new_layout = ALGORITHM_LEFT_ASYMMETRIC;
5398 case ALGORITHM_RIGHT_ASYMMETRIC_6:
5399 new_layout = ALGORITHM_RIGHT_ASYMMETRIC;
5401 case ALGORITHM_LEFT_SYMMETRIC_6:
5402 new_layout = ALGORITHM_LEFT_SYMMETRIC;
5404 case ALGORITHM_RIGHT_SYMMETRIC_6:
5405 new_layout = ALGORITHM_RIGHT_SYMMETRIC;
5407 case ALGORITHM_PARITY_0_6:
5408 new_layout = ALGORITHM_PARITY_0;
5410 case ALGORITHM_PARITY_N:
5411 new_layout = ALGORITHM_PARITY_N;
5414 return ERR_PTR(-EINVAL);
5416 mddev->new_level = 5;
5417 mddev->new_layout = new_layout;
5418 mddev->delta_disks = -1;
5419 mddev->raid_disks -= 1;
5420 return setup_conf(mddev);
5424 static int raid5_check_reshape(mddev_t *mddev)
5426 /* For a 2-drive array, the layout and chunk size can be changed
5427 * immediately as not restriping is needed.
5428 * For larger arrays we record the new value - after validation
5429 * to be used by a reshape pass.
5431 raid5_conf_t *conf = mddev->private;
5432 int new_chunk = mddev->new_chunk_sectors;
5434 if (mddev->new_layout >= 0 && !algorithm_valid_raid5(mddev->new_layout))
5436 if (new_chunk > 0) {
5437 if (!is_power_of_2(new_chunk))
5439 if (new_chunk < (PAGE_SIZE>>9))
5441 if (mddev->array_sectors & (new_chunk-1))
5442 /* not factor of array size */
5446 /* They look valid */
5448 if (mddev->raid_disks == 2) {
5449 /* can make the change immediately */
5450 if (mddev->new_layout >= 0) {
5451 conf->algorithm = mddev->new_layout;
5452 mddev->layout = mddev->new_layout;
5454 if (new_chunk > 0) {
5455 conf->chunk_sectors = new_chunk ;
5456 mddev->chunk_sectors = new_chunk;
5458 set_bit(MD_CHANGE_DEVS, &mddev->flags);
5459 md_wakeup_thread(mddev->thread);
5461 return check_reshape(mddev);
5464 static int raid6_check_reshape(mddev_t *mddev)
5466 int new_chunk = mddev->new_chunk_sectors;
5468 if (mddev->new_layout >= 0 && !algorithm_valid_raid6(mddev->new_layout))
5470 if (new_chunk > 0) {
5471 if (!is_power_of_2(new_chunk))
5473 if (new_chunk < (PAGE_SIZE >> 9))
5475 if (mddev->array_sectors & (new_chunk-1))
5476 /* not factor of array size */
5480 /* They look valid */
5481 return check_reshape(mddev);
5484 static void *raid5_takeover(mddev_t *mddev)
5486 /* raid5 can take over:
5487 * raid0 - if there is only one strip zone - make it a raid4 layout
5488 * raid1 - if there are two drives. We need to know the chunk size
5489 * raid4 - trivial - just use a raid4 layout.
5490 * raid6 - Providing it is a *_6 layout
5492 if (mddev->level == 0)
5493 return raid45_takeover_raid0(mddev, 5);
5494 if (mddev->level == 1)
5495 return raid5_takeover_raid1(mddev);
5496 if (mddev->level == 4) {
5497 mddev->new_layout = ALGORITHM_PARITY_N;
5498 mddev->new_level = 5;
5499 return setup_conf(mddev);
5501 if (mddev->level == 6)
5502 return raid5_takeover_raid6(mddev);
5504 return ERR_PTR(-EINVAL);
5507 static void *raid4_takeover(mddev_t *mddev)
5509 /* raid4 can take over:
5510 * raid0 - if there is only one strip zone
5511 * raid5 - if layout is right
5513 if (mddev->level == 0)
5514 return raid45_takeover_raid0(mddev, 4);
5515 if (mddev->level == 5 &&
5516 mddev->layout == ALGORITHM_PARITY_N) {
5517 mddev->new_layout = 0;
5518 mddev->new_level = 4;
5519 return setup_conf(mddev);
5521 return ERR_PTR(-EINVAL);
5524 static struct mdk_personality raid5_personality;
5526 static void *raid6_takeover(mddev_t *mddev)
5528 /* Currently can only take over a raid5. We map the
5529 * personality to an equivalent raid6 personality
5530 * with the Q block at the end.
5534 if (mddev->pers != &raid5_personality)
5535 return ERR_PTR(-EINVAL);
5536 if (mddev->degraded > 1)
5537 return ERR_PTR(-EINVAL);
5538 if (mddev->raid_disks > 253)
5539 return ERR_PTR(-EINVAL);
5540 if (mddev->raid_disks < 3)
5541 return ERR_PTR(-EINVAL);
5543 switch (mddev->layout) {
5544 case ALGORITHM_LEFT_ASYMMETRIC:
5545 new_layout = ALGORITHM_LEFT_ASYMMETRIC_6;
5547 case ALGORITHM_RIGHT_ASYMMETRIC:
5548 new_layout = ALGORITHM_RIGHT_ASYMMETRIC_6;
5550 case ALGORITHM_LEFT_SYMMETRIC:
5551 new_layout = ALGORITHM_LEFT_SYMMETRIC_6;
5553 case ALGORITHM_RIGHT_SYMMETRIC:
5554 new_layout = ALGORITHM_RIGHT_SYMMETRIC_6;
5556 case ALGORITHM_PARITY_0:
5557 new_layout = ALGORITHM_PARITY_0_6;
5559 case ALGORITHM_PARITY_N:
5560 new_layout = ALGORITHM_PARITY_N;
5563 return ERR_PTR(-EINVAL);
5565 mddev->new_level = 6;
5566 mddev->new_layout = new_layout;
5567 mddev->delta_disks = 1;
5568 mddev->raid_disks += 1;
5569 return setup_conf(mddev);
5573 static struct mdk_personality raid6_personality =
5577 .owner = THIS_MODULE,
5578 .make_request = make_request,
5582 .error_handler = error,
5583 .hot_add_disk = raid5_add_disk,
5584 .hot_remove_disk= raid5_remove_disk,
5585 .spare_active = raid5_spare_active,
5586 .sync_request = sync_request,
5587 .resize = raid5_resize,
5589 .check_reshape = raid6_check_reshape,
5590 .start_reshape = raid5_start_reshape,
5591 .finish_reshape = raid5_finish_reshape,
5592 .quiesce = raid5_quiesce,
5593 .takeover = raid6_takeover,
5595 static struct mdk_personality raid5_personality =
5599 .owner = THIS_MODULE,
5600 .make_request = make_request,
5604 .error_handler = error,
5605 .hot_add_disk = raid5_add_disk,
5606 .hot_remove_disk= raid5_remove_disk,
5607 .spare_active = raid5_spare_active,
5608 .sync_request = sync_request,
5609 .resize = raid5_resize,
5611 .check_reshape = raid5_check_reshape,
5612 .start_reshape = raid5_start_reshape,
5613 .finish_reshape = raid5_finish_reshape,
5614 .quiesce = raid5_quiesce,
5615 .takeover = raid5_takeover,
5618 static struct mdk_personality raid4_personality =
5622 .owner = THIS_MODULE,
5623 .make_request = make_request,
5627 .error_handler = error,
5628 .hot_add_disk = raid5_add_disk,
5629 .hot_remove_disk= raid5_remove_disk,
5630 .spare_active = raid5_spare_active,
5631 .sync_request = sync_request,
5632 .resize = raid5_resize,
5634 .check_reshape = raid5_check_reshape,
5635 .start_reshape = raid5_start_reshape,
5636 .finish_reshape = raid5_finish_reshape,
5637 .quiesce = raid5_quiesce,
5638 .takeover = raid4_takeover,
5641 static int __init raid5_init(void)
5643 register_md_personality(&raid6_personality);
5644 register_md_personality(&raid5_personality);
5645 register_md_personality(&raid4_personality);
5649 static void raid5_exit(void)
5651 unregister_md_personality(&raid6_personality);
5652 unregister_md_personality(&raid5_personality);
5653 unregister_md_personality(&raid4_personality);
5656 module_init(raid5_init);
5657 module_exit(raid5_exit);
5658 MODULE_LICENSE("GPL");
5659 MODULE_DESCRIPTION("RAID4/5/6 (striping with parity) personality for MD");
5660 MODULE_ALIAS("md-personality-4"); /* RAID5 */
5661 MODULE_ALIAS("md-raid5");
5662 MODULE_ALIAS("md-raid4");
5663 MODULE_ALIAS("md-level-5");
5664 MODULE_ALIAS("md-level-4");
5665 MODULE_ALIAS("md-personality-8"); /* RAID6 */
5666 MODULE_ALIAS("md-raid6");
5667 MODULE_ALIAS("md-level-6");
5669 /* This used to be two separate modules, they were: */
5670 MODULE_ALIAS("raid5");
5671 MODULE_ALIAS("raid6");