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->bm_write is the number of the last batch successfully written.
31 * conf->bm_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 bm_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/seq_file.h>
51 #include <linux/cpu.h>
60 #define NR_STRIPES 256
61 #define STRIPE_SIZE PAGE_SIZE
62 #define STRIPE_SHIFT (PAGE_SHIFT - 9)
63 #define STRIPE_SECTORS (STRIPE_SIZE>>9)
64 #define IO_THRESHOLD 1
65 #define BYPASS_THRESHOLD 1
66 #define NR_HASH (PAGE_SIZE / sizeof(struct hlist_head))
67 #define HASH_MASK (NR_HASH - 1)
69 #define stripe_hash(conf, sect) (&((conf)->stripe_hashtbl[((sect) >> STRIPE_SHIFT) & HASH_MASK]))
71 /* bio's attached to a stripe+device for I/O are linked together in bi_sector
72 * order without overlap. There may be several bio's per stripe+device, and
73 * a bio could span several devices.
74 * When walking this list for a particular stripe+device, we must never proceed
75 * beyond a bio that extends past this device, as the next bio might no longer
77 * This macro is used to determine the 'next' bio in the list, given the sector
78 * of the current stripe+device
80 #define r5_next_bio(bio, sect) ( ( (bio)->bi_sector + ((bio)->bi_size>>9) < sect + STRIPE_SECTORS) ? (bio)->bi_next : NULL)
82 * The following can be used to debug the driver
84 #define RAID5_PARANOIA 1
85 #if RAID5_PARANOIA && defined(CONFIG_SMP)
86 # define CHECK_DEVLOCK() assert_spin_locked(&conf->device_lock)
88 # define CHECK_DEVLOCK()
96 #define printk_rl(args...) ((void) (printk_ratelimit() && printk(args)))
99 * We maintain a biased count of active stripes in the bottom 16 bits of
100 * bi_phys_segments, and a count of processed stripes in the upper 16 bits
102 static inline int raid5_bi_phys_segments(struct bio *bio)
104 return bio->bi_phys_segments & 0xffff;
107 static inline int raid5_bi_hw_segments(struct bio *bio)
109 return (bio->bi_phys_segments >> 16) & 0xffff;
112 static inline int raid5_dec_bi_phys_segments(struct bio *bio)
114 --bio->bi_phys_segments;
115 return raid5_bi_phys_segments(bio);
118 static inline int raid5_dec_bi_hw_segments(struct bio *bio)
120 unsigned short val = raid5_bi_hw_segments(bio);
123 bio->bi_phys_segments = (val << 16) | raid5_bi_phys_segments(bio);
127 static inline void raid5_set_bi_hw_segments(struct bio *bio, unsigned int cnt)
129 bio->bi_phys_segments = raid5_bi_phys_segments(bio) || (cnt << 16);
132 /* Find first data disk in a raid6 stripe */
133 static inline int raid6_d0(struct stripe_head *sh)
136 /* ddf always start from first device */
138 /* md starts just after Q block */
139 if (sh->qd_idx == sh->disks - 1)
142 return sh->qd_idx + 1;
144 static inline int raid6_next_disk(int disk, int raid_disks)
147 return (disk < raid_disks) ? disk : 0;
150 /* When walking through the disks in a raid5, starting at raid6_d0,
151 * We need to map each disk to a 'slot', where the data disks are slot
152 * 0 .. raid_disks-3, the parity disk is raid_disks-2 and the Q disk
153 * is raid_disks-1. This help does that mapping.
155 static int raid6_idx_to_slot(int idx, struct stripe_head *sh,
156 int *count, int syndrome_disks)
160 if (idx == sh->pd_idx)
161 return syndrome_disks;
162 if (idx == sh->qd_idx)
163 return syndrome_disks + 1;
168 static void return_io(struct bio *return_bi)
170 struct bio *bi = return_bi;
173 return_bi = bi->bi_next;
181 static void print_raid5_conf (raid5_conf_t *conf);
183 static int stripe_operations_active(struct stripe_head *sh)
185 return sh->check_state || sh->reconstruct_state ||
186 test_bit(STRIPE_BIOFILL_RUN, &sh->state) ||
187 test_bit(STRIPE_COMPUTE_RUN, &sh->state);
190 static void __release_stripe(raid5_conf_t *conf, struct stripe_head *sh)
192 if (atomic_dec_and_test(&sh->count)) {
193 BUG_ON(!list_empty(&sh->lru));
194 BUG_ON(atomic_read(&conf->active_stripes)==0);
195 if (test_bit(STRIPE_HANDLE, &sh->state)) {
196 if (test_bit(STRIPE_DELAYED, &sh->state)) {
197 list_add_tail(&sh->lru, &conf->delayed_list);
198 blk_plug_device(conf->mddev->queue);
199 } else if (test_bit(STRIPE_BIT_DELAY, &sh->state) &&
200 sh->bm_seq - conf->seq_write > 0) {
201 list_add_tail(&sh->lru, &conf->bitmap_list);
202 blk_plug_device(conf->mddev->queue);
204 clear_bit(STRIPE_BIT_DELAY, &sh->state);
205 list_add_tail(&sh->lru, &conf->handle_list);
207 md_wakeup_thread(conf->mddev->thread);
209 BUG_ON(stripe_operations_active(sh));
210 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
211 atomic_dec(&conf->preread_active_stripes);
212 if (atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD)
213 md_wakeup_thread(conf->mddev->thread);
215 atomic_dec(&conf->active_stripes);
216 if (!test_bit(STRIPE_EXPANDING, &sh->state)) {
217 list_add_tail(&sh->lru, &conf->inactive_list);
218 wake_up(&conf->wait_for_stripe);
219 if (conf->retry_read_aligned)
220 md_wakeup_thread(conf->mddev->thread);
226 static void release_stripe(struct stripe_head *sh)
228 raid5_conf_t *conf = sh->raid_conf;
231 spin_lock_irqsave(&conf->device_lock, flags);
232 __release_stripe(conf, sh);
233 spin_unlock_irqrestore(&conf->device_lock, flags);
236 static inline void remove_hash(struct stripe_head *sh)
238 pr_debug("remove_hash(), stripe %llu\n",
239 (unsigned long long)sh->sector);
241 hlist_del_init(&sh->hash);
244 static inline void insert_hash(raid5_conf_t *conf, struct stripe_head *sh)
246 struct hlist_head *hp = stripe_hash(conf, sh->sector);
248 pr_debug("insert_hash(), stripe %llu\n",
249 (unsigned long long)sh->sector);
252 hlist_add_head(&sh->hash, hp);
256 /* find an idle stripe, make sure it is unhashed, and return it. */
257 static struct stripe_head *get_free_stripe(raid5_conf_t *conf)
259 struct stripe_head *sh = NULL;
260 struct list_head *first;
263 if (list_empty(&conf->inactive_list))
265 first = conf->inactive_list.next;
266 sh = list_entry(first, struct stripe_head, lru);
267 list_del_init(first);
269 atomic_inc(&conf->active_stripes);
274 static void shrink_buffers(struct stripe_head *sh, int num)
279 for (i=0; i<num ; i++) {
283 sh->dev[i].page = NULL;
288 static int grow_buffers(struct stripe_head *sh, int num)
292 for (i=0; i<num; i++) {
295 if (!(page = alloc_page(GFP_KERNEL))) {
298 sh->dev[i].page = page;
303 static void raid5_build_block(struct stripe_head *sh, int i, int previous);
304 static void stripe_set_idx(sector_t stripe, raid5_conf_t *conf, int previous,
305 struct stripe_head *sh);
307 static void init_stripe(struct stripe_head *sh, sector_t sector, int previous)
309 raid5_conf_t *conf = sh->raid_conf;
312 BUG_ON(atomic_read(&sh->count) != 0);
313 BUG_ON(test_bit(STRIPE_HANDLE, &sh->state));
314 BUG_ON(stripe_operations_active(sh));
317 pr_debug("init_stripe called, stripe %llu\n",
318 (unsigned long long)sh->sector);
322 sh->generation = conf->generation - previous;
323 sh->disks = previous ? conf->previous_raid_disks : conf->raid_disks;
325 stripe_set_idx(sector, conf, previous, sh);
329 for (i = sh->disks; i--; ) {
330 struct r5dev *dev = &sh->dev[i];
332 if (dev->toread || dev->read || dev->towrite || dev->written ||
333 test_bit(R5_LOCKED, &dev->flags)) {
334 printk(KERN_ERR "sector=%llx i=%d %p %p %p %p %d\n",
335 (unsigned long long)sh->sector, i, dev->toread,
336 dev->read, dev->towrite, dev->written,
337 test_bit(R5_LOCKED, &dev->flags));
341 raid5_build_block(sh, i, previous);
343 insert_hash(conf, sh);
346 static struct stripe_head *__find_stripe(raid5_conf_t *conf, sector_t sector,
349 struct stripe_head *sh;
350 struct hlist_node *hn;
353 pr_debug("__find_stripe, sector %llu\n", (unsigned long long)sector);
354 hlist_for_each_entry(sh, hn, stripe_hash(conf, sector), hash)
355 if (sh->sector == sector && sh->generation == generation)
357 pr_debug("__stripe %llu not in cache\n", (unsigned long long)sector);
361 static void unplug_slaves(mddev_t *mddev);
362 static void raid5_unplug_device(struct request_queue *q);
364 static struct stripe_head *
365 get_active_stripe(raid5_conf_t *conf, sector_t sector,
366 int previous, int noblock)
368 struct stripe_head *sh;
370 pr_debug("get_stripe, sector %llu\n", (unsigned long long)sector);
372 spin_lock_irq(&conf->device_lock);
375 wait_event_lock_irq(conf->wait_for_stripe,
377 conf->device_lock, /* nothing */);
378 sh = __find_stripe(conf, sector, conf->generation - previous);
380 if (!conf->inactive_blocked)
381 sh = get_free_stripe(conf);
382 if (noblock && sh == NULL)
385 conf->inactive_blocked = 1;
386 wait_event_lock_irq(conf->wait_for_stripe,
387 !list_empty(&conf->inactive_list) &&
388 (atomic_read(&conf->active_stripes)
389 < (conf->max_nr_stripes *3/4)
390 || !conf->inactive_blocked),
392 raid5_unplug_device(conf->mddev->queue)
394 conf->inactive_blocked = 0;
396 init_stripe(sh, sector, previous);
398 if (atomic_read(&sh->count)) {
399 BUG_ON(!list_empty(&sh->lru)
400 && !test_bit(STRIPE_EXPANDING, &sh->state));
402 if (!test_bit(STRIPE_HANDLE, &sh->state))
403 atomic_inc(&conf->active_stripes);
404 if (list_empty(&sh->lru) &&
405 !test_bit(STRIPE_EXPANDING, &sh->state))
407 list_del_init(&sh->lru);
410 } while (sh == NULL);
413 atomic_inc(&sh->count);
415 spin_unlock_irq(&conf->device_lock);
420 raid5_end_read_request(struct bio *bi, int error);
422 raid5_end_write_request(struct bio *bi, int error);
424 static void ops_run_io(struct stripe_head *sh, struct stripe_head_state *s)
426 raid5_conf_t *conf = sh->raid_conf;
427 int i, disks = sh->disks;
431 for (i = disks; i--; ) {
435 if (test_and_clear_bit(R5_Wantwrite, &sh->dev[i].flags))
437 else if (test_and_clear_bit(R5_Wantread, &sh->dev[i].flags))
442 bi = &sh->dev[i].req;
446 bi->bi_end_io = raid5_end_write_request;
448 bi->bi_end_io = raid5_end_read_request;
451 rdev = rcu_dereference(conf->disks[i].rdev);
452 if (rdev && test_bit(Faulty, &rdev->flags))
455 atomic_inc(&rdev->nr_pending);
459 if (s->syncing || s->expanding || s->expanded)
460 md_sync_acct(rdev->bdev, STRIPE_SECTORS);
462 set_bit(STRIPE_IO_STARTED, &sh->state);
464 bi->bi_bdev = rdev->bdev;
465 pr_debug("%s: for %llu schedule op %ld on disc %d\n",
466 __func__, (unsigned long long)sh->sector,
468 atomic_inc(&sh->count);
469 bi->bi_sector = sh->sector + rdev->data_offset;
470 bi->bi_flags = 1 << BIO_UPTODATE;
474 bi->bi_io_vec = &sh->dev[i].vec;
475 bi->bi_io_vec[0].bv_len = STRIPE_SIZE;
476 bi->bi_io_vec[0].bv_offset = 0;
477 bi->bi_size = STRIPE_SIZE;
480 test_bit(R5_ReWrite, &sh->dev[i].flags))
481 atomic_add(STRIPE_SECTORS,
482 &rdev->corrected_errors);
483 generic_make_request(bi);
486 set_bit(STRIPE_DEGRADED, &sh->state);
487 pr_debug("skip op %ld on disc %d for sector %llu\n",
488 bi->bi_rw, i, (unsigned long long)sh->sector);
489 clear_bit(R5_LOCKED, &sh->dev[i].flags);
490 set_bit(STRIPE_HANDLE, &sh->state);
495 static struct dma_async_tx_descriptor *
496 async_copy_data(int frombio, struct bio *bio, struct page *page,
497 sector_t sector, struct dma_async_tx_descriptor *tx)
500 struct page *bio_page;
503 struct async_submit_ctl submit;
505 if (bio->bi_sector >= sector)
506 page_offset = (signed)(bio->bi_sector - sector) * 512;
508 page_offset = (signed)(sector - bio->bi_sector) * -512;
510 init_async_submit(&submit, 0, tx, NULL, NULL, NULL);
511 bio_for_each_segment(bvl, bio, i) {
512 int len = bio_iovec_idx(bio, i)->bv_len;
516 if (page_offset < 0) {
517 b_offset = -page_offset;
518 page_offset += b_offset;
522 if (len > 0 && page_offset + len > STRIPE_SIZE)
523 clen = STRIPE_SIZE - page_offset;
528 b_offset += bio_iovec_idx(bio, i)->bv_offset;
529 bio_page = bio_iovec_idx(bio, i)->bv_page;
531 tx = async_memcpy(page, bio_page, page_offset,
532 b_offset, clen, &submit);
534 tx = async_memcpy(bio_page, page, b_offset,
535 page_offset, clen, &submit);
537 /* chain the operations */
538 submit.depend_tx = tx;
540 if (clen < len) /* hit end of page */
548 static void ops_complete_biofill(void *stripe_head_ref)
550 struct stripe_head *sh = stripe_head_ref;
551 struct bio *return_bi = NULL;
552 raid5_conf_t *conf = sh->raid_conf;
555 pr_debug("%s: stripe %llu\n", __func__,
556 (unsigned long long)sh->sector);
558 /* clear completed biofills */
559 spin_lock_irq(&conf->device_lock);
560 for (i = sh->disks; i--; ) {
561 struct r5dev *dev = &sh->dev[i];
563 /* acknowledge completion of a biofill operation */
564 /* and check if we need to reply to a read request,
565 * new R5_Wantfill requests are held off until
566 * !STRIPE_BIOFILL_RUN
568 if (test_and_clear_bit(R5_Wantfill, &dev->flags)) {
569 struct bio *rbi, *rbi2;
574 while (rbi && rbi->bi_sector <
575 dev->sector + STRIPE_SECTORS) {
576 rbi2 = r5_next_bio(rbi, dev->sector);
577 if (!raid5_dec_bi_phys_segments(rbi)) {
578 rbi->bi_next = return_bi;
585 spin_unlock_irq(&conf->device_lock);
586 clear_bit(STRIPE_BIOFILL_RUN, &sh->state);
588 return_io(return_bi);
590 set_bit(STRIPE_HANDLE, &sh->state);
594 static void ops_run_biofill(struct stripe_head *sh)
596 struct dma_async_tx_descriptor *tx = NULL;
597 raid5_conf_t *conf = sh->raid_conf;
598 struct async_submit_ctl submit;
601 pr_debug("%s: stripe %llu\n", __func__,
602 (unsigned long long)sh->sector);
604 for (i = sh->disks; i--; ) {
605 struct r5dev *dev = &sh->dev[i];
606 if (test_bit(R5_Wantfill, &dev->flags)) {
608 spin_lock_irq(&conf->device_lock);
609 dev->read = rbi = dev->toread;
611 spin_unlock_irq(&conf->device_lock);
612 while (rbi && rbi->bi_sector <
613 dev->sector + STRIPE_SECTORS) {
614 tx = async_copy_data(0, rbi, dev->page,
616 rbi = r5_next_bio(rbi, dev->sector);
621 atomic_inc(&sh->count);
622 init_async_submit(&submit, ASYNC_TX_ACK, tx, ops_complete_biofill, sh, NULL);
623 async_trigger_callback(&submit);
626 static void mark_target_uptodate(struct stripe_head *sh, int target)
633 tgt = &sh->dev[target];
634 set_bit(R5_UPTODATE, &tgt->flags);
635 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
636 clear_bit(R5_Wantcompute, &tgt->flags);
639 static void ops_complete_compute(void *stripe_head_ref)
641 struct stripe_head *sh = stripe_head_ref;
643 pr_debug("%s: stripe %llu\n", __func__,
644 (unsigned long long)sh->sector);
646 /* mark the computed target(s) as uptodate */
647 mark_target_uptodate(sh, sh->ops.target);
648 mark_target_uptodate(sh, sh->ops.target2);
650 clear_bit(STRIPE_COMPUTE_RUN, &sh->state);
651 if (sh->check_state == check_state_compute_run)
652 sh->check_state = check_state_compute_result;
653 set_bit(STRIPE_HANDLE, &sh->state);
657 /* return a pointer to the address conversion region of the scribble buffer */
658 static addr_conv_t *to_addr_conv(struct stripe_head *sh,
659 struct raid5_percpu *percpu)
661 return percpu->scribble + sizeof(struct page *) * (sh->disks + 2);
664 static struct dma_async_tx_descriptor *
665 ops_run_compute5(struct stripe_head *sh, struct raid5_percpu *percpu)
667 int disks = sh->disks;
668 struct page **xor_srcs = percpu->scribble;
669 int target = sh->ops.target;
670 struct r5dev *tgt = &sh->dev[target];
671 struct page *xor_dest = tgt->page;
673 struct dma_async_tx_descriptor *tx;
674 struct async_submit_ctl submit;
677 pr_debug("%s: stripe %llu block: %d\n",
678 __func__, (unsigned long long)sh->sector, target);
679 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
681 for (i = disks; i--; )
683 xor_srcs[count++] = sh->dev[i].page;
685 atomic_inc(&sh->count);
687 init_async_submit(&submit, ASYNC_TX_XOR_ZERO_DST, NULL,
688 ops_complete_compute, sh, to_addr_conv(sh, percpu));
689 if (unlikely(count == 1))
690 tx = async_memcpy(xor_dest, xor_srcs[0], 0, 0, STRIPE_SIZE, &submit);
692 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
697 /* set_syndrome_sources - populate source buffers for gen_syndrome
698 * @srcs - (struct page *) array of size sh->disks
699 * @sh - stripe_head to parse
701 * Populates srcs in proper layout order for the stripe and returns the
702 * 'count' of sources to be used in a call to async_gen_syndrome. The P
703 * destination buffer is recorded in srcs[count] and the Q destination
704 * is recorded in srcs[count+1]].
706 static int set_syndrome_sources(struct page **srcs, struct stripe_head *sh)
708 int disks = sh->disks;
709 int syndrome_disks = sh->ddf_layout ? disks : (disks - 2);
710 int d0_idx = raid6_d0(sh);
714 for (i = 0; i < disks; i++)
715 srcs[i] = (void *)raid6_empty_zero_page;
720 int slot = raid6_idx_to_slot(i, sh, &count, syndrome_disks);
722 srcs[slot] = sh->dev[i].page;
723 i = raid6_next_disk(i, disks);
724 } while (i != d0_idx);
725 BUG_ON(count != syndrome_disks);
730 static struct dma_async_tx_descriptor *
731 ops_run_compute6_1(struct stripe_head *sh, struct raid5_percpu *percpu)
733 int disks = sh->disks;
734 struct page **blocks = percpu->scribble;
736 int qd_idx = sh->qd_idx;
737 struct dma_async_tx_descriptor *tx;
738 struct async_submit_ctl submit;
744 if (sh->ops.target < 0)
745 target = sh->ops.target2;
746 else if (sh->ops.target2 < 0)
747 target = sh->ops.target;
749 /* we should only have one valid target */
752 pr_debug("%s: stripe %llu block: %d\n",
753 __func__, (unsigned long long)sh->sector, target);
755 tgt = &sh->dev[target];
756 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
759 atomic_inc(&sh->count);
761 if (target == qd_idx) {
762 count = set_syndrome_sources(blocks, sh);
763 blocks[count] = NULL; /* regenerating p is not necessary */
764 BUG_ON(blocks[count+1] != dest); /* q should already be set */
765 init_async_submit(&submit, 0, NULL, ops_complete_compute, sh,
766 to_addr_conv(sh, percpu));
767 tx = async_gen_syndrome(blocks, 0, count+2, STRIPE_SIZE, &submit);
769 /* Compute any data- or p-drive using XOR */
771 for (i = disks; i-- ; ) {
772 if (i == target || i == qd_idx)
774 blocks[count++] = sh->dev[i].page;
777 init_async_submit(&submit, ASYNC_TX_XOR_ZERO_DST, NULL,
778 ops_complete_compute, sh,
779 to_addr_conv(sh, percpu));
780 tx = async_xor(dest, blocks, 0, count, STRIPE_SIZE, &submit);
786 static struct dma_async_tx_descriptor *
787 ops_run_compute6_2(struct stripe_head *sh, struct raid5_percpu *percpu)
789 int i, count, disks = sh->disks;
790 int syndrome_disks = sh->ddf_layout ? disks : disks-2;
791 int d0_idx = raid6_d0(sh);
792 int faila = -1, failb = -1;
793 int target = sh->ops.target;
794 int target2 = sh->ops.target2;
795 struct r5dev *tgt = &sh->dev[target];
796 struct r5dev *tgt2 = &sh->dev[target2];
797 struct dma_async_tx_descriptor *tx;
798 struct page **blocks = percpu->scribble;
799 struct async_submit_ctl submit;
801 pr_debug("%s: stripe %llu block1: %d block2: %d\n",
802 __func__, (unsigned long long)sh->sector, target, target2);
803 BUG_ON(target < 0 || target2 < 0);
804 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
805 BUG_ON(!test_bit(R5_Wantcompute, &tgt2->flags));
807 /* we need to open-code set_syndrome_sources to handle to the
808 * slot number conversion for 'faila' and 'failb'
810 for (i = 0; i < disks ; i++)
811 blocks[i] = (void *)raid6_empty_zero_page;
815 int slot = raid6_idx_to_slot(i, sh, &count, syndrome_disks);
817 blocks[slot] = sh->dev[i].page;
823 i = raid6_next_disk(i, disks);
824 } while (i != d0_idx);
825 BUG_ON(count != syndrome_disks);
827 BUG_ON(faila == failb);
830 pr_debug("%s: stripe: %llu faila: %d failb: %d\n",
831 __func__, (unsigned long long)sh->sector, faila, failb);
833 atomic_inc(&sh->count);
835 if (failb == syndrome_disks+1) {
836 /* Q disk is one of the missing disks */
837 if (faila == syndrome_disks) {
838 /* Missing P+Q, just recompute */
839 init_async_submit(&submit, 0, NULL, ops_complete_compute,
840 sh, to_addr_conv(sh, percpu));
841 return async_gen_syndrome(blocks, 0, count+2,
842 STRIPE_SIZE, &submit);
846 int qd_idx = sh->qd_idx;
848 /* Missing D+Q: recompute D from P, then recompute Q */
849 if (target == qd_idx)
850 data_target = target2;
852 data_target = target;
855 for (i = disks; i-- ; ) {
856 if (i == data_target || i == qd_idx)
858 blocks[count++] = sh->dev[i].page;
860 dest = sh->dev[data_target].page;
861 init_async_submit(&submit, ASYNC_TX_XOR_ZERO_DST, NULL,
862 NULL, NULL, to_addr_conv(sh, percpu));
863 tx = async_xor(dest, blocks, 0, count, STRIPE_SIZE,
866 count = set_syndrome_sources(blocks, sh);
867 init_async_submit(&submit, 0, tx, ops_complete_compute,
868 sh, to_addr_conv(sh, percpu));
869 return async_gen_syndrome(blocks, 0, count+2,
870 STRIPE_SIZE, &submit);
874 init_async_submit(&submit, 0, NULL, ops_complete_compute, sh,
875 to_addr_conv(sh, percpu));
876 if (failb == syndrome_disks) {
877 /* We're missing D+P. */
878 return async_raid6_datap_recov(syndrome_disks+2, STRIPE_SIZE,
879 faila, blocks, &submit);
881 /* We're missing D+D. */
882 return async_raid6_2data_recov(syndrome_disks+2, STRIPE_SIZE,
883 faila, failb, blocks, &submit);
888 static void ops_complete_prexor(void *stripe_head_ref)
890 struct stripe_head *sh = stripe_head_ref;
892 pr_debug("%s: stripe %llu\n", __func__,
893 (unsigned long long)sh->sector);
896 static struct dma_async_tx_descriptor *
897 ops_run_prexor(struct stripe_head *sh, struct raid5_percpu *percpu,
898 struct dma_async_tx_descriptor *tx)
900 int disks = sh->disks;
901 struct page **xor_srcs = percpu->scribble;
902 int count = 0, pd_idx = sh->pd_idx, i;
903 struct async_submit_ctl submit;
905 /* existing parity data subtracted */
906 struct page *xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
908 pr_debug("%s: stripe %llu\n", __func__,
909 (unsigned long long)sh->sector);
911 for (i = disks; i--; ) {
912 struct r5dev *dev = &sh->dev[i];
913 /* Only process blocks that are known to be uptodate */
914 if (test_bit(R5_Wantdrain, &dev->flags))
915 xor_srcs[count++] = dev->page;
918 init_async_submit(&submit, ASYNC_TX_XOR_DROP_DST, tx,
919 ops_complete_prexor, sh, to_addr_conv(sh, percpu));
920 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
925 static struct dma_async_tx_descriptor *
926 ops_run_biodrain(struct stripe_head *sh, struct dma_async_tx_descriptor *tx)
928 int disks = sh->disks;
931 pr_debug("%s: stripe %llu\n", __func__,
932 (unsigned long long)sh->sector);
934 for (i = disks; i--; ) {
935 struct r5dev *dev = &sh->dev[i];
938 if (test_and_clear_bit(R5_Wantdrain, &dev->flags)) {
941 spin_lock(&sh->lock);
942 chosen = dev->towrite;
944 BUG_ON(dev->written);
945 wbi = dev->written = chosen;
946 spin_unlock(&sh->lock);
948 while (wbi && wbi->bi_sector <
949 dev->sector + STRIPE_SECTORS) {
950 tx = async_copy_data(1, wbi, dev->page,
952 wbi = r5_next_bio(wbi, dev->sector);
960 static void ops_complete_reconstruct(void *stripe_head_ref)
962 struct stripe_head *sh = stripe_head_ref;
963 int disks = sh->disks;
964 int pd_idx = sh->pd_idx;
965 int qd_idx = sh->qd_idx;
968 pr_debug("%s: stripe %llu\n", __func__,
969 (unsigned long long)sh->sector);
971 for (i = disks; i--; ) {
972 struct r5dev *dev = &sh->dev[i];
974 if (dev->written || i == pd_idx || i == qd_idx)
975 set_bit(R5_UPTODATE, &dev->flags);
978 if (sh->reconstruct_state == reconstruct_state_drain_run)
979 sh->reconstruct_state = reconstruct_state_drain_result;
980 else if (sh->reconstruct_state == reconstruct_state_prexor_drain_run)
981 sh->reconstruct_state = reconstruct_state_prexor_drain_result;
983 BUG_ON(sh->reconstruct_state != reconstruct_state_run);
984 sh->reconstruct_state = reconstruct_state_result;
987 set_bit(STRIPE_HANDLE, &sh->state);
992 ops_run_reconstruct5(struct stripe_head *sh, struct raid5_percpu *percpu,
993 struct dma_async_tx_descriptor *tx)
995 int disks = sh->disks;
996 struct page **xor_srcs = percpu->scribble;
997 struct async_submit_ctl submit;
998 int count = 0, pd_idx = sh->pd_idx, i;
999 struct page *xor_dest;
1001 unsigned long flags;
1003 pr_debug("%s: stripe %llu\n", __func__,
1004 (unsigned long long)sh->sector);
1006 /* check if prexor is active which means only process blocks
1007 * that are part of a read-modify-write (written)
1009 if (sh->reconstruct_state == reconstruct_state_prexor_drain_run) {
1011 xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
1012 for (i = disks; i--; ) {
1013 struct r5dev *dev = &sh->dev[i];
1015 xor_srcs[count++] = dev->page;
1018 xor_dest = sh->dev[pd_idx].page;
1019 for (i = disks; i--; ) {
1020 struct r5dev *dev = &sh->dev[i];
1022 xor_srcs[count++] = dev->page;
1026 /* 1/ if we prexor'd then the dest is reused as a source
1027 * 2/ if we did not prexor then we are redoing the parity
1028 * set ASYNC_TX_XOR_DROP_DST and ASYNC_TX_XOR_ZERO_DST
1029 * for the synchronous xor case
1031 flags = ASYNC_TX_ACK |
1032 (prexor ? ASYNC_TX_XOR_DROP_DST : ASYNC_TX_XOR_ZERO_DST);
1034 atomic_inc(&sh->count);
1036 init_async_submit(&submit, flags, tx, ops_complete_reconstruct, sh,
1037 to_addr_conv(sh, percpu));
1038 if (unlikely(count == 1))
1039 tx = async_memcpy(xor_dest, xor_srcs[0], 0, 0, STRIPE_SIZE, &submit);
1041 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
1045 ops_run_reconstruct6(struct stripe_head *sh, struct raid5_percpu *percpu,
1046 struct dma_async_tx_descriptor *tx)
1048 struct async_submit_ctl submit;
1049 struct page **blocks = percpu->scribble;
1052 pr_debug("%s: stripe %llu\n", __func__, (unsigned long long)sh->sector);
1054 count = set_syndrome_sources(blocks, sh);
1056 atomic_inc(&sh->count);
1058 init_async_submit(&submit, ASYNC_TX_ACK, tx, ops_complete_reconstruct,
1059 sh, to_addr_conv(sh, percpu));
1060 async_gen_syndrome(blocks, 0, count+2, STRIPE_SIZE, &submit);
1063 static void ops_complete_check(void *stripe_head_ref)
1065 struct stripe_head *sh = stripe_head_ref;
1067 pr_debug("%s: stripe %llu\n", __func__,
1068 (unsigned long long)sh->sector);
1070 sh->check_state = check_state_check_result;
1071 set_bit(STRIPE_HANDLE, &sh->state);
1075 static void ops_run_check_p(struct stripe_head *sh, struct raid5_percpu *percpu)
1077 int disks = sh->disks;
1078 int pd_idx = sh->pd_idx;
1079 int qd_idx = sh->qd_idx;
1080 struct page *xor_dest;
1081 struct page **xor_srcs = percpu->scribble;
1082 struct dma_async_tx_descriptor *tx;
1083 struct async_submit_ctl submit;
1087 pr_debug("%s: stripe %llu\n", __func__,
1088 (unsigned long long)sh->sector);
1091 xor_dest = sh->dev[pd_idx].page;
1092 xor_srcs[count++] = xor_dest;
1093 for (i = disks; i--; ) {
1094 if (i == pd_idx || i == qd_idx)
1096 xor_srcs[count++] = sh->dev[i].page;
1099 init_async_submit(&submit, 0, NULL, NULL, NULL,
1100 to_addr_conv(sh, percpu));
1101 tx = async_xor_val(xor_dest, xor_srcs, 0, count, STRIPE_SIZE,
1102 &sh->ops.zero_sum_result, &submit);
1104 atomic_inc(&sh->count);
1105 init_async_submit(&submit, ASYNC_TX_ACK, tx, ops_complete_check, sh, NULL);
1106 tx = async_trigger_callback(&submit);
1109 static void ops_run_check_pq(struct stripe_head *sh, struct raid5_percpu *percpu, int checkp)
1111 struct page **srcs = percpu->scribble;
1112 struct async_submit_ctl submit;
1115 pr_debug("%s: stripe %llu checkp: %d\n", __func__,
1116 (unsigned long long)sh->sector, checkp);
1118 count = set_syndrome_sources(srcs, sh);
1122 atomic_inc(&sh->count);
1123 init_async_submit(&submit, ASYNC_TX_ACK, NULL, ops_complete_check,
1124 sh, to_addr_conv(sh, percpu));
1125 async_syndrome_val(srcs, 0, count+2, STRIPE_SIZE,
1126 &sh->ops.zero_sum_result, percpu->spare_page, &submit);
1129 static void raid_run_ops(struct stripe_head *sh, unsigned long ops_request)
1131 int overlap_clear = 0, i, disks = sh->disks;
1132 struct dma_async_tx_descriptor *tx = NULL;
1133 raid5_conf_t *conf = sh->raid_conf;
1134 int level = conf->level;
1135 struct raid5_percpu *percpu;
1139 percpu = per_cpu_ptr(conf->percpu, cpu);
1140 if (test_bit(STRIPE_OP_BIOFILL, &ops_request)) {
1141 ops_run_biofill(sh);
1145 if (test_bit(STRIPE_OP_COMPUTE_BLK, &ops_request)) {
1147 tx = ops_run_compute5(sh, percpu);
1149 if (sh->ops.target2 < 0 || sh->ops.target < 0)
1150 tx = ops_run_compute6_1(sh, percpu);
1152 tx = ops_run_compute6_2(sh, percpu);
1154 /* terminate the chain if reconstruct is not set to be run */
1155 if (tx && !test_bit(STRIPE_OP_RECONSTRUCT, &ops_request))
1159 if (test_bit(STRIPE_OP_PREXOR, &ops_request))
1160 tx = ops_run_prexor(sh, percpu, tx);
1162 if (test_bit(STRIPE_OP_BIODRAIN, &ops_request)) {
1163 tx = ops_run_biodrain(sh, tx);
1167 if (test_bit(STRIPE_OP_RECONSTRUCT, &ops_request)) {
1169 ops_run_reconstruct5(sh, percpu, tx);
1171 ops_run_reconstruct6(sh, percpu, tx);
1174 if (test_bit(STRIPE_OP_CHECK, &ops_request)) {
1175 if (sh->check_state == check_state_run)
1176 ops_run_check_p(sh, percpu);
1177 else if (sh->check_state == check_state_run_q)
1178 ops_run_check_pq(sh, percpu, 0);
1179 else if (sh->check_state == check_state_run_pq)
1180 ops_run_check_pq(sh, percpu, 1);
1186 for (i = disks; i--; ) {
1187 struct r5dev *dev = &sh->dev[i];
1188 if (test_and_clear_bit(R5_Overlap, &dev->flags))
1189 wake_up(&sh->raid_conf->wait_for_overlap);
1194 static int grow_one_stripe(raid5_conf_t *conf)
1196 struct stripe_head *sh;
1197 sh = kmem_cache_alloc(conf->slab_cache, GFP_KERNEL);
1200 memset(sh, 0, sizeof(*sh) + (conf->raid_disks-1)*sizeof(struct r5dev));
1201 sh->raid_conf = conf;
1202 spin_lock_init(&sh->lock);
1204 if (grow_buffers(sh, conf->raid_disks)) {
1205 shrink_buffers(sh, conf->raid_disks);
1206 kmem_cache_free(conf->slab_cache, sh);
1209 sh->disks = conf->raid_disks;
1210 /* we just created an active stripe so... */
1211 atomic_set(&sh->count, 1);
1212 atomic_inc(&conf->active_stripes);
1213 INIT_LIST_HEAD(&sh->lru);
1218 static int grow_stripes(raid5_conf_t *conf, int num)
1220 struct kmem_cache *sc;
1221 int devs = conf->raid_disks;
1223 sprintf(conf->cache_name[0],
1224 "raid%d-%s", conf->level, mdname(conf->mddev));
1225 sprintf(conf->cache_name[1],
1226 "raid%d-%s-alt", conf->level, mdname(conf->mddev));
1227 conf->active_name = 0;
1228 sc = kmem_cache_create(conf->cache_name[conf->active_name],
1229 sizeof(struct stripe_head)+(devs-1)*sizeof(struct r5dev),
1233 conf->slab_cache = sc;
1234 conf->pool_size = devs;
1236 if (!grow_one_stripe(conf))
1242 * scribble_len - return the required size of the scribble region
1243 * @num - total number of disks in the array
1245 * The size must be enough to contain:
1246 * 1/ a struct page pointer for each device in the array +2
1247 * 2/ room to convert each entry in (1) to its corresponding dma
1248 * (dma_map_page()) or page (page_address()) address.
1250 * Note: the +2 is for the destination buffers of the ddf/raid6 case where we
1251 * calculate over all devices (not just the data blocks), using zeros in place
1252 * of the P and Q blocks.
1254 static size_t scribble_len(int num)
1258 len = sizeof(struct page *) * (num+2) + sizeof(addr_conv_t) * (num+2);
1263 static int resize_stripes(raid5_conf_t *conf, int newsize)
1265 /* Make all the stripes able to hold 'newsize' devices.
1266 * New slots in each stripe get 'page' set to a new page.
1268 * This happens in stages:
1269 * 1/ create a new kmem_cache and allocate the required number of
1271 * 2/ gather all the old stripe_heads and tranfer the pages across
1272 * to the new stripe_heads. This will have the side effect of
1273 * freezing the array as once all stripe_heads have been collected,
1274 * no IO will be possible. Old stripe heads are freed once their
1275 * pages have been transferred over, and the old kmem_cache is
1276 * freed when all stripes are done.
1277 * 3/ reallocate conf->disks to be suitable bigger. If this fails,
1278 * we simple return a failre status - no need to clean anything up.
1279 * 4/ allocate new pages for the new slots in the new stripe_heads.
1280 * If this fails, we don't bother trying the shrink the
1281 * stripe_heads down again, we just leave them as they are.
1282 * As each stripe_head is processed the new one is released into
1285 * Once step2 is started, we cannot afford to wait for a write,
1286 * so we use GFP_NOIO allocations.
1288 struct stripe_head *osh, *nsh;
1289 LIST_HEAD(newstripes);
1290 struct disk_info *ndisks;
1293 struct kmem_cache *sc;
1296 if (newsize <= conf->pool_size)
1297 return 0; /* never bother to shrink */
1299 err = md_allow_write(conf->mddev);
1304 sc = kmem_cache_create(conf->cache_name[1-conf->active_name],
1305 sizeof(struct stripe_head)+(newsize-1)*sizeof(struct r5dev),
1310 for (i = conf->max_nr_stripes; i; i--) {
1311 nsh = kmem_cache_alloc(sc, GFP_KERNEL);
1315 memset(nsh, 0, sizeof(*nsh) + (newsize-1)*sizeof(struct r5dev));
1317 nsh->raid_conf = conf;
1318 spin_lock_init(&nsh->lock);
1320 list_add(&nsh->lru, &newstripes);
1323 /* didn't get enough, give up */
1324 while (!list_empty(&newstripes)) {
1325 nsh = list_entry(newstripes.next, struct stripe_head, lru);
1326 list_del(&nsh->lru);
1327 kmem_cache_free(sc, nsh);
1329 kmem_cache_destroy(sc);
1332 /* Step 2 - Must use GFP_NOIO now.
1333 * OK, we have enough stripes, start collecting inactive
1334 * stripes and copying them over
1336 list_for_each_entry(nsh, &newstripes, lru) {
1337 spin_lock_irq(&conf->device_lock);
1338 wait_event_lock_irq(conf->wait_for_stripe,
1339 !list_empty(&conf->inactive_list),
1341 unplug_slaves(conf->mddev)
1343 osh = get_free_stripe(conf);
1344 spin_unlock_irq(&conf->device_lock);
1345 atomic_set(&nsh->count, 1);
1346 for(i=0; i<conf->pool_size; i++)
1347 nsh->dev[i].page = osh->dev[i].page;
1348 for( ; i<newsize; i++)
1349 nsh->dev[i].page = NULL;
1350 kmem_cache_free(conf->slab_cache, osh);
1352 kmem_cache_destroy(conf->slab_cache);
1355 * At this point, we are holding all the stripes so the array
1356 * is completely stalled, so now is a good time to resize
1357 * conf->disks and the scribble region
1359 ndisks = kzalloc(newsize * sizeof(struct disk_info), GFP_NOIO);
1361 for (i=0; i<conf->raid_disks; i++)
1362 ndisks[i] = conf->disks[i];
1364 conf->disks = ndisks;
1369 conf->scribble_len = scribble_len(newsize);
1370 for_each_present_cpu(cpu) {
1371 struct raid5_percpu *percpu;
1374 percpu = per_cpu_ptr(conf->percpu, cpu);
1375 scribble = kmalloc(conf->scribble_len, GFP_NOIO);
1378 kfree(percpu->scribble);
1379 percpu->scribble = scribble;
1387 /* Step 4, return new stripes to service */
1388 while(!list_empty(&newstripes)) {
1389 nsh = list_entry(newstripes.next, struct stripe_head, lru);
1390 list_del_init(&nsh->lru);
1392 for (i=conf->raid_disks; i < newsize; i++)
1393 if (nsh->dev[i].page == NULL) {
1394 struct page *p = alloc_page(GFP_NOIO);
1395 nsh->dev[i].page = p;
1399 release_stripe(nsh);
1401 /* critical section pass, GFP_NOIO no longer needed */
1403 conf->slab_cache = sc;
1404 conf->active_name = 1-conf->active_name;
1405 conf->pool_size = newsize;
1409 static int drop_one_stripe(raid5_conf_t *conf)
1411 struct stripe_head *sh;
1413 spin_lock_irq(&conf->device_lock);
1414 sh = get_free_stripe(conf);
1415 spin_unlock_irq(&conf->device_lock);
1418 BUG_ON(atomic_read(&sh->count));
1419 shrink_buffers(sh, conf->pool_size);
1420 kmem_cache_free(conf->slab_cache, sh);
1421 atomic_dec(&conf->active_stripes);
1425 static void shrink_stripes(raid5_conf_t *conf)
1427 while (drop_one_stripe(conf))
1430 if (conf->slab_cache)
1431 kmem_cache_destroy(conf->slab_cache);
1432 conf->slab_cache = NULL;
1435 static void raid5_end_read_request(struct bio * bi, int error)
1437 struct stripe_head *sh = bi->bi_private;
1438 raid5_conf_t *conf = sh->raid_conf;
1439 int disks = sh->disks, i;
1440 int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
1441 char b[BDEVNAME_SIZE];
1445 for (i=0 ; i<disks; i++)
1446 if (bi == &sh->dev[i].req)
1449 pr_debug("end_read_request %llu/%d, count: %d, uptodate %d.\n",
1450 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
1458 set_bit(R5_UPTODATE, &sh->dev[i].flags);
1459 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
1460 rdev = conf->disks[i].rdev;
1461 printk_rl(KERN_INFO "raid5:%s: read error corrected"
1462 " (%lu sectors at %llu on %s)\n",
1463 mdname(conf->mddev), STRIPE_SECTORS,
1464 (unsigned long long)(sh->sector
1465 + rdev->data_offset),
1466 bdevname(rdev->bdev, b));
1467 clear_bit(R5_ReadError, &sh->dev[i].flags);
1468 clear_bit(R5_ReWrite, &sh->dev[i].flags);
1470 if (atomic_read(&conf->disks[i].rdev->read_errors))
1471 atomic_set(&conf->disks[i].rdev->read_errors, 0);
1473 const char *bdn = bdevname(conf->disks[i].rdev->bdev, b);
1475 rdev = conf->disks[i].rdev;
1477 clear_bit(R5_UPTODATE, &sh->dev[i].flags);
1478 atomic_inc(&rdev->read_errors);
1479 if (conf->mddev->degraded)
1480 printk_rl(KERN_WARNING
1481 "raid5:%s: read error not correctable "
1482 "(sector %llu on %s).\n",
1483 mdname(conf->mddev),
1484 (unsigned long long)(sh->sector
1485 + rdev->data_offset),
1487 else if (test_bit(R5_ReWrite, &sh->dev[i].flags))
1489 printk_rl(KERN_WARNING
1490 "raid5:%s: read error NOT corrected!! "
1491 "(sector %llu on %s).\n",
1492 mdname(conf->mddev),
1493 (unsigned long long)(sh->sector
1494 + rdev->data_offset),
1496 else if (atomic_read(&rdev->read_errors)
1497 > conf->max_nr_stripes)
1499 "raid5:%s: Too many read errors, failing device %s.\n",
1500 mdname(conf->mddev), bdn);
1504 set_bit(R5_ReadError, &sh->dev[i].flags);
1506 clear_bit(R5_ReadError, &sh->dev[i].flags);
1507 clear_bit(R5_ReWrite, &sh->dev[i].flags);
1508 md_error(conf->mddev, rdev);
1511 rdev_dec_pending(conf->disks[i].rdev, conf->mddev);
1512 clear_bit(R5_LOCKED, &sh->dev[i].flags);
1513 set_bit(STRIPE_HANDLE, &sh->state);
1517 static void raid5_end_write_request(struct bio *bi, int error)
1519 struct stripe_head *sh = bi->bi_private;
1520 raid5_conf_t *conf = sh->raid_conf;
1521 int disks = sh->disks, i;
1522 int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
1524 for (i=0 ; i<disks; i++)
1525 if (bi == &sh->dev[i].req)
1528 pr_debug("end_write_request %llu/%d, count %d, uptodate: %d.\n",
1529 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
1537 md_error(conf->mddev, conf->disks[i].rdev);
1539 rdev_dec_pending(conf->disks[i].rdev, conf->mddev);
1541 clear_bit(R5_LOCKED, &sh->dev[i].flags);
1542 set_bit(STRIPE_HANDLE, &sh->state);
1547 static sector_t compute_blocknr(struct stripe_head *sh, int i, int previous);
1549 static void raid5_build_block(struct stripe_head *sh, int i, int previous)
1551 struct r5dev *dev = &sh->dev[i];
1553 bio_init(&dev->req);
1554 dev->req.bi_io_vec = &dev->vec;
1556 dev->req.bi_max_vecs++;
1557 dev->vec.bv_page = dev->page;
1558 dev->vec.bv_len = STRIPE_SIZE;
1559 dev->vec.bv_offset = 0;
1561 dev->req.bi_sector = sh->sector;
1562 dev->req.bi_private = sh;
1565 dev->sector = compute_blocknr(sh, i, previous);
1568 static void error(mddev_t *mddev, mdk_rdev_t *rdev)
1570 char b[BDEVNAME_SIZE];
1571 raid5_conf_t *conf = (raid5_conf_t *) mddev->private;
1572 pr_debug("raid5: error called\n");
1574 if (!test_bit(Faulty, &rdev->flags)) {
1575 set_bit(MD_CHANGE_DEVS, &mddev->flags);
1576 if (test_and_clear_bit(In_sync, &rdev->flags)) {
1577 unsigned long flags;
1578 spin_lock_irqsave(&conf->device_lock, flags);
1580 spin_unlock_irqrestore(&conf->device_lock, flags);
1582 * if recovery was running, make sure it aborts.
1584 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1586 set_bit(Faulty, &rdev->flags);
1588 "raid5: Disk failure on %s, disabling device.\n"
1589 "raid5: Operation continuing on %d devices.\n",
1590 bdevname(rdev->bdev,b), conf->raid_disks - mddev->degraded);
1595 * Input: a 'big' sector number,
1596 * Output: index of the data and parity disk, and the sector # in them.
1598 static sector_t raid5_compute_sector(raid5_conf_t *conf, sector_t r_sector,
1599 int previous, int *dd_idx,
1600 struct stripe_head *sh)
1603 unsigned long chunk_number;
1604 unsigned int chunk_offset;
1607 sector_t new_sector;
1608 int algorithm = previous ? conf->prev_algo
1610 int sectors_per_chunk = previous ? (conf->prev_chunk >> 9)
1611 : (conf->chunk_size >> 9);
1612 int raid_disks = previous ? conf->previous_raid_disks
1614 int data_disks = raid_disks - conf->max_degraded;
1616 /* First compute the information on this sector */
1619 * Compute the chunk number and the sector offset inside the chunk
1621 chunk_offset = sector_div(r_sector, sectors_per_chunk);
1622 chunk_number = r_sector;
1623 BUG_ON(r_sector != chunk_number);
1626 * Compute the stripe number
1628 stripe = chunk_number / data_disks;
1631 * Compute the data disk and parity disk indexes inside the stripe
1633 *dd_idx = chunk_number % data_disks;
1636 * Select the parity disk based on the user selected algorithm.
1638 pd_idx = qd_idx = ~0;
1639 switch(conf->level) {
1641 pd_idx = data_disks;
1644 switch (algorithm) {
1645 case ALGORITHM_LEFT_ASYMMETRIC:
1646 pd_idx = data_disks - stripe % raid_disks;
1647 if (*dd_idx >= pd_idx)
1650 case ALGORITHM_RIGHT_ASYMMETRIC:
1651 pd_idx = stripe % raid_disks;
1652 if (*dd_idx >= pd_idx)
1655 case ALGORITHM_LEFT_SYMMETRIC:
1656 pd_idx = data_disks - stripe % raid_disks;
1657 *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
1659 case ALGORITHM_RIGHT_SYMMETRIC:
1660 pd_idx = stripe % raid_disks;
1661 *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
1663 case ALGORITHM_PARITY_0:
1667 case ALGORITHM_PARITY_N:
1668 pd_idx = data_disks;
1671 printk(KERN_ERR "raid5: unsupported algorithm %d\n",
1678 switch (algorithm) {
1679 case ALGORITHM_LEFT_ASYMMETRIC:
1680 pd_idx = raid_disks - 1 - (stripe % raid_disks);
1681 qd_idx = pd_idx + 1;
1682 if (pd_idx == raid_disks-1) {
1683 (*dd_idx)++; /* Q D D D P */
1685 } else if (*dd_idx >= pd_idx)
1686 (*dd_idx) += 2; /* D D P Q D */
1688 case ALGORITHM_RIGHT_ASYMMETRIC:
1689 pd_idx = stripe % raid_disks;
1690 qd_idx = pd_idx + 1;
1691 if (pd_idx == raid_disks-1) {
1692 (*dd_idx)++; /* Q D D D P */
1694 } else if (*dd_idx >= pd_idx)
1695 (*dd_idx) += 2; /* D D P Q D */
1697 case ALGORITHM_LEFT_SYMMETRIC:
1698 pd_idx = raid_disks - 1 - (stripe % raid_disks);
1699 qd_idx = (pd_idx + 1) % raid_disks;
1700 *dd_idx = (pd_idx + 2 + *dd_idx) % raid_disks;
1702 case ALGORITHM_RIGHT_SYMMETRIC:
1703 pd_idx = stripe % raid_disks;
1704 qd_idx = (pd_idx + 1) % raid_disks;
1705 *dd_idx = (pd_idx + 2 + *dd_idx) % raid_disks;
1708 case ALGORITHM_PARITY_0:
1713 case ALGORITHM_PARITY_N:
1714 pd_idx = data_disks;
1715 qd_idx = data_disks + 1;
1718 case ALGORITHM_ROTATING_ZERO_RESTART:
1719 /* Exactly the same as RIGHT_ASYMMETRIC, but or
1720 * of blocks for computing Q is different.
1722 pd_idx = stripe % raid_disks;
1723 qd_idx = pd_idx + 1;
1724 if (pd_idx == raid_disks-1) {
1725 (*dd_idx)++; /* Q D D D P */
1727 } else if (*dd_idx >= pd_idx)
1728 (*dd_idx) += 2; /* D D P Q D */
1732 case ALGORITHM_ROTATING_N_RESTART:
1733 /* Same a left_asymmetric, by first stripe is
1734 * D D D P Q rather than
1737 pd_idx = raid_disks - 1 - ((stripe + 1) % raid_disks);
1738 qd_idx = pd_idx + 1;
1739 if (pd_idx == raid_disks-1) {
1740 (*dd_idx)++; /* Q D D D P */
1742 } else if (*dd_idx >= pd_idx)
1743 (*dd_idx) += 2; /* D D P Q D */
1747 case ALGORITHM_ROTATING_N_CONTINUE:
1748 /* Same as left_symmetric but Q is before P */
1749 pd_idx = raid_disks - 1 - (stripe % raid_disks);
1750 qd_idx = (pd_idx + raid_disks - 1) % raid_disks;
1751 *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
1755 case ALGORITHM_LEFT_ASYMMETRIC_6:
1756 /* RAID5 left_asymmetric, with Q on last device */
1757 pd_idx = data_disks - stripe % (raid_disks-1);
1758 if (*dd_idx >= pd_idx)
1760 qd_idx = raid_disks - 1;
1763 case ALGORITHM_RIGHT_ASYMMETRIC_6:
1764 pd_idx = stripe % (raid_disks-1);
1765 if (*dd_idx >= pd_idx)
1767 qd_idx = raid_disks - 1;
1770 case ALGORITHM_LEFT_SYMMETRIC_6:
1771 pd_idx = data_disks - stripe % (raid_disks-1);
1772 *dd_idx = (pd_idx + 1 + *dd_idx) % (raid_disks-1);
1773 qd_idx = raid_disks - 1;
1776 case ALGORITHM_RIGHT_SYMMETRIC_6:
1777 pd_idx = stripe % (raid_disks-1);
1778 *dd_idx = (pd_idx + 1 + *dd_idx) % (raid_disks-1);
1779 qd_idx = raid_disks - 1;
1782 case ALGORITHM_PARITY_0_6:
1785 qd_idx = raid_disks - 1;
1790 printk(KERN_CRIT "raid6: unsupported algorithm %d\n",
1798 sh->pd_idx = pd_idx;
1799 sh->qd_idx = qd_idx;
1800 sh->ddf_layout = ddf_layout;
1803 * Finally, compute the new sector number
1805 new_sector = (sector_t)stripe * sectors_per_chunk + chunk_offset;
1810 static sector_t compute_blocknr(struct stripe_head *sh, int i, int previous)
1812 raid5_conf_t *conf = sh->raid_conf;
1813 int raid_disks = sh->disks;
1814 int data_disks = raid_disks - conf->max_degraded;
1815 sector_t new_sector = sh->sector, check;
1816 int sectors_per_chunk = previous ? (conf->prev_chunk >> 9)
1817 : (conf->chunk_size >> 9);
1818 int algorithm = previous ? conf->prev_algo
1822 int chunk_number, dummy1, dd_idx = i;
1824 struct stripe_head sh2;
1827 chunk_offset = sector_div(new_sector, sectors_per_chunk);
1828 stripe = new_sector;
1829 BUG_ON(new_sector != stripe);
1831 if (i == sh->pd_idx)
1833 switch(conf->level) {
1836 switch (algorithm) {
1837 case ALGORITHM_LEFT_ASYMMETRIC:
1838 case ALGORITHM_RIGHT_ASYMMETRIC:
1842 case ALGORITHM_LEFT_SYMMETRIC:
1843 case ALGORITHM_RIGHT_SYMMETRIC:
1846 i -= (sh->pd_idx + 1);
1848 case ALGORITHM_PARITY_0:
1851 case ALGORITHM_PARITY_N:
1854 printk(KERN_ERR "raid5: unsupported algorithm %d\n",
1860 if (i == sh->qd_idx)
1861 return 0; /* It is the Q disk */
1862 switch (algorithm) {
1863 case ALGORITHM_LEFT_ASYMMETRIC:
1864 case ALGORITHM_RIGHT_ASYMMETRIC:
1865 case ALGORITHM_ROTATING_ZERO_RESTART:
1866 case ALGORITHM_ROTATING_N_RESTART:
1867 if (sh->pd_idx == raid_disks-1)
1868 i--; /* Q D D D P */
1869 else if (i > sh->pd_idx)
1870 i -= 2; /* D D P Q D */
1872 case ALGORITHM_LEFT_SYMMETRIC:
1873 case ALGORITHM_RIGHT_SYMMETRIC:
1874 if (sh->pd_idx == raid_disks-1)
1875 i--; /* Q D D D P */
1880 i -= (sh->pd_idx + 2);
1883 case ALGORITHM_PARITY_0:
1886 case ALGORITHM_PARITY_N:
1888 case ALGORITHM_ROTATING_N_CONTINUE:
1889 if (sh->pd_idx == 0)
1890 i--; /* P D D D Q */
1891 else if (i > sh->pd_idx)
1892 i -= 2; /* D D Q P D */
1894 case ALGORITHM_LEFT_ASYMMETRIC_6:
1895 case ALGORITHM_RIGHT_ASYMMETRIC_6:
1899 case ALGORITHM_LEFT_SYMMETRIC_6:
1900 case ALGORITHM_RIGHT_SYMMETRIC_6:
1902 i += data_disks + 1;
1903 i -= (sh->pd_idx + 1);
1905 case ALGORITHM_PARITY_0_6:
1909 printk(KERN_CRIT "raid6: unsupported algorithm %d\n",
1916 chunk_number = stripe * data_disks + i;
1917 r_sector = (sector_t)chunk_number * sectors_per_chunk + chunk_offset;
1919 check = raid5_compute_sector(conf, r_sector,
1920 previous, &dummy1, &sh2);
1921 if (check != sh->sector || dummy1 != dd_idx || sh2.pd_idx != sh->pd_idx
1922 || sh2.qd_idx != sh->qd_idx) {
1923 printk(KERN_ERR "compute_blocknr: map not correct\n");
1932 * Copy data between a page in the stripe cache, and one or more bion
1933 * The page could align with the middle of the bio, or there could be
1934 * several bion, each with several bio_vecs, which cover part of the page
1935 * Multiple bion are linked together on bi_next. There may be extras
1936 * at the end of this list. We ignore them.
1938 static void copy_data(int frombio, struct bio *bio,
1942 char *pa = page_address(page);
1943 struct bio_vec *bvl;
1947 if (bio->bi_sector >= sector)
1948 page_offset = (signed)(bio->bi_sector - sector) * 512;
1950 page_offset = (signed)(sector - bio->bi_sector) * -512;
1951 bio_for_each_segment(bvl, bio, i) {
1952 int len = bio_iovec_idx(bio,i)->bv_len;
1956 if (page_offset < 0) {
1957 b_offset = -page_offset;
1958 page_offset += b_offset;
1962 if (len > 0 && page_offset + len > STRIPE_SIZE)
1963 clen = STRIPE_SIZE - page_offset;
1967 char *ba = __bio_kmap_atomic(bio, i, KM_USER0);
1969 memcpy(pa+page_offset, ba+b_offset, clen);
1971 memcpy(ba+b_offset, pa+page_offset, clen);
1972 __bio_kunmap_atomic(ba, KM_USER0);
1974 if (clen < len) /* hit end of page */
1980 #define check_xor() do { \
1981 if (count == MAX_XOR_BLOCKS) { \
1982 xor_blocks(count, STRIPE_SIZE, dest, ptr);\
1987 static void compute_parity6(struct stripe_head *sh, int method)
1989 raid5_conf_t *conf = sh->raid_conf;
1990 int i, pd_idx, qd_idx, d0_idx, disks = sh->disks, count;
1991 int syndrome_disks = sh->ddf_layout ? disks : (disks - 2);
1993 /**** FIX THIS: This could be very bad if disks is close to 256 ****/
1994 void *ptrs[syndrome_disks+2];
1996 pd_idx = sh->pd_idx;
1997 qd_idx = sh->qd_idx;
1998 d0_idx = raid6_d0(sh);
2000 pr_debug("compute_parity, stripe %llu, method %d\n",
2001 (unsigned long long)sh->sector, method);
2004 case READ_MODIFY_WRITE:
2005 BUG(); /* READ_MODIFY_WRITE N/A for RAID-6 */
2006 case RECONSTRUCT_WRITE:
2007 for (i= disks; i-- ;)
2008 if ( i != pd_idx && i != qd_idx && sh->dev[i].towrite ) {
2009 chosen = sh->dev[i].towrite;
2010 sh->dev[i].towrite = NULL;
2012 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
2013 wake_up(&conf->wait_for_overlap);
2015 BUG_ON(sh->dev[i].written);
2016 sh->dev[i].written = chosen;
2020 BUG(); /* Not implemented yet */
2023 for (i = disks; i--;)
2024 if (sh->dev[i].written) {
2025 sector_t sector = sh->dev[i].sector;
2026 struct bio *wbi = sh->dev[i].written;
2027 while (wbi && wbi->bi_sector < sector + STRIPE_SECTORS) {
2028 copy_data(1, wbi, sh->dev[i].page, sector);
2029 wbi = r5_next_bio(wbi, sector);
2032 set_bit(R5_LOCKED, &sh->dev[i].flags);
2033 set_bit(R5_UPTODATE, &sh->dev[i].flags);
2036 /* Note that unlike RAID-5, the ordering of the disks matters greatly.*/
2038 for (i = 0; i < disks; i++)
2039 ptrs[i] = (void *)raid6_empty_zero_page;
2044 int slot = raid6_idx_to_slot(i, sh, &count, syndrome_disks);
2046 ptrs[slot] = page_address(sh->dev[i].page);
2047 if (slot < syndrome_disks &&
2048 !test_bit(R5_UPTODATE, &sh->dev[i].flags)) {
2049 printk(KERN_ERR "block %d/%d not uptodate "
2050 "on parity calc\n", i, count);
2054 i = raid6_next_disk(i, disks);
2055 } while (i != d0_idx);
2056 BUG_ON(count != syndrome_disks);
2058 raid6_call.gen_syndrome(syndrome_disks+2, STRIPE_SIZE, ptrs);
2061 case RECONSTRUCT_WRITE:
2062 set_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
2063 set_bit(R5_UPTODATE, &sh->dev[qd_idx].flags);
2064 set_bit(R5_LOCKED, &sh->dev[pd_idx].flags);
2065 set_bit(R5_LOCKED, &sh->dev[qd_idx].flags);
2068 set_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
2069 set_bit(R5_UPTODATE, &sh->dev[qd_idx].flags);
2075 /* Compute one missing block */
2076 static void compute_block_1(struct stripe_head *sh, int dd_idx, int nozero)
2078 int i, count, disks = sh->disks;
2079 void *ptr[MAX_XOR_BLOCKS], *dest, *p;
2080 int qd_idx = sh->qd_idx;
2082 pr_debug("compute_block_1, stripe %llu, idx %d\n",
2083 (unsigned long long)sh->sector, dd_idx);
2085 if ( dd_idx == qd_idx ) {
2086 /* We're actually computing the Q drive */
2087 compute_parity6(sh, UPDATE_PARITY);
2089 dest = page_address(sh->dev[dd_idx].page);
2090 if (!nozero) memset(dest, 0, STRIPE_SIZE);
2092 for (i = disks ; i--; ) {
2093 if (i == dd_idx || i == qd_idx)
2095 p = page_address(sh->dev[i].page);
2096 if (test_bit(R5_UPTODATE, &sh->dev[i].flags))
2099 printk("compute_block() %d, stripe %llu, %d"
2100 " not present\n", dd_idx,
2101 (unsigned long long)sh->sector, i);
2106 xor_blocks(count, STRIPE_SIZE, dest, ptr);
2107 if (!nozero) set_bit(R5_UPTODATE, &sh->dev[dd_idx].flags);
2108 else clear_bit(R5_UPTODATE, &sh->dev[dd_idx].flags);
2112 /* Compute two missing blocks */
2113 static void compute_block_2(struct stripe_head *sh, int dd_idx1, int dd_idx2)
2115 int i, count, disks = sh->disks;
2116 int syndrome_disks = sh->ddf_layout ? disks : disks-2;
2117 int d0_idx = raid6_d0(sh);
2118 int faila = -1, failb = -1;
2119 /**** FIX THIS: This could be very bad if disks is close to 256 ****/
2120 void *ptrs[syndrome_disks+2];
2122 for (i = 0; i < disks ; i++)
2123 ptrs[i] = (void *)raid6_empty_zero_page;
2127 int slot = raid6_idx_to_slot(i, sh, &count, syndrome_disks);
2129 ptrs[slot] = page_address(sh->dev[i].page);
2135 i = raid6_next_disk(i, disks);
2136 } while (i != d0_idx);
2137 BUG_ON(count != syndrome_disks);
2139 BUG_ON(faila == failb);
2140 if ( failb < faila ) { int tmp = faila; faila = failb; failb = tmp; }
2142 pr_debug("compute_block_2, stripe %llu, idx %d,%d (%d,%d)\n",
2143 (unsigned long long)sh->sector, dd_idx1, dd_idx2,
2146 if (failb == syndrome_disks+1) {
2147 /* Q disk is one of the missing disks */
2148 if (faila == syndrome_disks) {
2149 /* Missing P+Q, just recompute */
2150 compute_parity6(sh, UPDATE_PARITY);
2153 /* We're missing D+Q; recompute D from P */
2154 compute_block_1(sh, ((dd_idx1 == sh->qd_idx) ?
2157 compute_parity6(sh, UPDATE_PARITY); /* Is this necessary? */
2162 /* We're missing D+P or D+D; */
2163 if (failb == syndrome_disks) {
2164 /* We're missing D+P. */
2165 raid6_datap_recov(syndrome_disks+2, STRIPE_SIZE, faila, ptrs);
2167 /* We're missing D+D. */
2168 raid6_2data_recov(syndrome_disks+2, STRIPE_SIZE, faila, failb,
2172 /* Both the above update both missing blocks */
2173 set_bit(R5_UPTODATE, &sh->dev[dd_idx1].flags);
2174 set_bit(R5_UPTODATE, &sh->dev[dd_idx2].flags);
2178 schedule_reconstruction(struct stripe_head *sh, struct stripe_head_state *s,
2179 int rcw, int expand)
2181 int i, pd_idx = sh->pd_idx, disks = sh->disks;
2182 raid5_conf_t *conf = sh->raid_conf;
2183 int level = conf->level;
2186 /* if we are not expanding this is a proper write request, and
2187 * there will be bios with new data to be drained into the
2191 sh->reconstruct_state = reconstruct_state_drain_run;
2192 set_bit(STRIPE_OP_BIODRAIN, &s->ops_request);
2194 sh->reconstruct_state = reconstruct_state_run;
2196 set_bit(STRIPE_OP_RECONSTRUCT, &s->ops_request);
2198 for (i = disks; i--; ) {
2199 struct r5dev *dev = &sh->dev[i];
2202 set_bit(R5_LOCKED, &dev->flags);
2203 set_bit(R5_Wantdrain, &dev->flags);
2205 clear_bit(R5_UPTODATE, &dev->flags);
2209 if (s->locked + conf->max_degraded == disks)
2210 if (!test_and_set_bit(STRIPE_FULL_WRITE, &sh->state))
2211 atomic_inc(&conf->pending_full_writes);
2214 BUG_ON(!(test_bit(R5_UPTODATE, &sh->dev[pd_idx].flags) ||
2215 test_bit(R5_Wantcompute, &sh->dev[pd_idx].flags)));
2217 sh->reconstruct_state = reconstruct_state_prexor_drain_run;
2218 set_bit(STRIPE_OP_PREXOR, &s->ops_request);
2219 set_bit(STRIPE_OP_BIODRAIN, &s->ops_request);
2220 set_bit(STRIPE_OP_RECONSTRUCT, &s->ops_request);
2222 for (i = disks; i--; ) {
2223 struct r5dev *dev = &sh->dev[i];
2228 (test_bit(R5_UPTODATE, &dev->flags) ||
2229 test_bit(R5_Wantcompute, &dev->flags))) {
2230 set_bit(R5_Wantdrain, &dev->flags);
2231 set_bit(R5_LOCKED, &dev->flags);
2232 clear_bit(R5_UPTODATE, &dev->flags);
2238 /* keep the parity disk(s) locked while asynchronous operations
2241 set_bit(R5_LOCKED, &sh->dev[pd_idx].flags);
2242 clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
2246 int qd_idx = sh->qd_idx;
2247 struct r5dev *dev = &sh->dev[qd_idx];
2249 set_bit(R5_LOCKED, &dev->flags);
2250 clear_bit(R5_UPTODATE, &dev->flags);
2254 pr_debug("%s: stripe %llu locked: %d ops_request: %lx\n",
2255 __func__, (unsigned long long)sh->sector,
2256 s->locked, s->ops_request);
2260 * Each stripe/dev can have one or more bion attached.
2261 * toread/towrite point to the first in a chain.
2262 * The bi_next chain must be in order.
2264 static int add_stripe_bio(struct stripe_head *sh, struct bio *bi, int dd_idx, int forwrite)
2267 raid5_conf_t *conf = sh->raid_conf;
2270 pr_debug("adding bh b#%llu to stripe s#%llu\n",
2271 (unsigned long long)bi->bi_sector,
2272 (unsigned long long)sh->sector);
2275 spin_lock(&sh->lock);
2276 spin_lock_irq(&conf->device_lock);
2278 bip = &sh->dev[dd_idx].towrite;
2279 if (*bip == NULL && sh->dev[dd_idx].written == NULL)
2282 bip = &sh->dev[dd_idx].toread;
2283 while (*bip && (*bip)->bi_sector < bi->bi_sector) {
2284 if ((*bip)->bi_sector + ((*bip)->bi_size >> 9) > bi->bi_sector)
2286 bip = & (*bip)->bi_next;
2288 if (*bip && (*bip)->bi_sector < bi->bi_sector + ((bi->bi_size)>>9))
2291 BUG_ON(*bip && bi->bi_next && (*bip) != bi->bi_next);
2295 bi->bi_phys_segments++;
2296 spin_unlock_irq(&conf->device_lock);
2297 spin_unlock(&sh->lock);
2299 pr_debug("added bi b#%llu to stripe s#%llu, disk %d.\n",
2300 (unsigned long long)bi->bi_sector,
2301 (unsigned long long)sh->sector, dd_idx);
2303 if (conf->mddev->bitmap && firstwrite) {
2304 bitmap_startwrite(conf->mddev->bitmap, sh->sector,
2306 sh->bm_seq = conf->seq_flush+1;
2307 set_bit(STRIPE_BIT_DELAY, &sh->state);
2311 /* check if page is covered */
2312 sector_t sector = sh->dev[dd_idx].sector;
2313 for (bi=sh->dev[dd_idx].towrite;
2314 sector < sh->dev[dd_idx].sector + STRIPE_SECTORS &&
2315 bi && bi->bi_sector <= sector;
2316 bi = r5_next_bio(bi, sh->dev[dd_idx].sector)) {
2317 if (bi->bi_sector + (bi->bi_size>>9) >= sector)
2318 sector = bi->bi_sector + (bi->bi_size>>9);
2320 if (sector >= sh->dev[dd_idx].sector + STRIPE_SECTORS)
2321 set_bit(R5_OVERWRITE, &sh->dev[dd_idx].flags);
2326 set_bit(R5_Overlap, &sh->dev[dd_idx].flags);
2327 spin_unlock_irq(&conf->device_lock);
2328 spin_unlock(&sh->lock);
2332 static void end_reshape(raid5_conf_t *conf);
2334 static int page_is_zero(struct page *p)
2336 char *a = page_address(p);
2337 return ((*(u32*)a) == 0 &&
2338 memcmp(a, a+4, STRIPE_SIZE-4)==0);
2341 static void stripe_set_idx(sector_t stripe, raid5_conf_t *conf, int previous,
2342 struct stripe_head *sh)
2344 int sectors_per_chunk =
2345 previous ? (conf->prev_chunk >> 9)
2346 : (conf->chunk_size >> 9);
2348 int chunk_offset = sector_div(stripe, sectors_per_chunk);
2349 int disks = previous ? conf->previous_raid_disks : conf->raid_disks;
2351 raid5_compute_sector(conf,
2352 stripe * (disks - conf->max_degraded)
2353 *sectors_per_chunk + chunk_offset,
2359 handle_failed_stripe(raid5_conf_t *conf, struct stripe_head *sh,
2360 struct stripe_head_state *s, int disks,
2361 struct bio **return_bi)
2364 for (i = disks; i--; ) {
2368 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
2371 rdev = rcu_dereference(conf->disks[i].rdev);
2372 if (rdev && test_bit(In_sync, &rdev->flags))
2373 /* multiple read failures in one stripe */
2374 md_error(conf->mddev, rdev);
2377 spin_lock_irq(&conf->device_lock);
2378 /* fail all writes first */
2379 bi = sh->dev[i].towrite;
2380 sh->dev[i].towrite = NULL;
2386 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
2387 wake_up(&conf->wait_for_overlap);
2389 while (bi && bi->bi_sector <
2390 sh->dev[i].sector + STRIPE_SECTORS) {
2391 struct bio *nextbi = r5_next_bio(bi, sh->dev[i].sector);
2392 clear_bit(BIO_UPTODATE, &bi->bi_flags);
2393 if (!raid5_dec_bi_phys_segments(bi)) {
2394 md_write_end(conf->mddev);
2395 bi->bi_next = *return_bi;
2400 /* and fail all 'written' */
2401 bi = sh->dev[i].written;
2402 sh->dev[i].written = NULL;
2403 if (bi) bitmap_end = 1;
2404 while (bi && bi->bi_sector <
2405 sh->dev[i].sector + STRIPE_SECTORS) {
2406 struct bio *bi2 = r5_next_bio(bi, sh->dev[i].sector);
2407 clear_bit(BIO_UPTODATE, &bi->bi_flags);
2408 if (!raid5_dec_bi_phys_segments(bi)) {
2409 md_write_end(conf->mddev);
2410 bi->bi_next = *return_bi;
2416 /* fail any reads if this device is non-operational and
2417 * the data has not reached the cache yet.
2419 if (!test_bit(R5_Wantfill, &sh->dev[i].flags) &&
2420 (!test_bit(R5_Insync, &sh->dev[i].flags) ||
2421 test_bit(R5_ReadError, &sh->dev[i].flags))) {
2422 bi = sh->dev[i].toread;
2423 sh->dev[i].toread = NULL;
2424 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
2425 wake_up(&conf->wait_for_overlap);
2426 if (bi) s->to_read--;
2427 while (bi && bi->bi_sector <
2428 sh->dev[i].sector + STRIPE_SECTORS) {
2429 struct bio *nextbi =
2430 r5_next_bio(bi, sh->dev[i].sector);
2431 clear_bit(BIO_UPTODATE, &bi->bi_flags);
2432 if (!raid5_dec_bi_phys_segments(bi)) {
2433 bi->bi_next = *return_bi;
2439 spin_unlock_irq(&conf->device_lock);
2441 bitmap_endwrite(conf->mddev->bitmap, sh->sector,
2442 STRIPE_SECTORS, 0, 0);
2445 if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
2446 if (atomic_dec_and_test(&conf->pending_full_writes))
2447 md_wakeup_thread(conf->mddev->thread);
2450 /* fetch_block5 - checks the given member device to see if its data needs
2451 * to be read or computed to satisfy a request.
2453 * Returns 1 when no more member devices need to be checked, otherwise returns
2454 * 0 to tell the loop in handle_stripe_fill5 to continue
2456 static int fetch_block5(struct stripe_head *sh, struct stripe_head_state *s,
2457 int disk_idx, int disks)
2459 struct r5dev *dev = &sh->dev[disk_idx];
2460 struct r5dev *failed_dev = &sh->dev[s->failed_num];
2462 /* is the data in this block needed, and can we get it? */
2463 if (!test_bit(R5_LOCKED, &dev->flags) &&
2464 !test_bit(R5_UPTODATE, &dev->flags) &&
2466 (dev->towrite && !test_bit(R5_OVERWRITE, &dev->flags)) ||
2467 s->syncing || s->expanding ||
2469 (failed_dev->toread ||
2470 (failed_dev->towrite &&
2471 !test_bit(R5_OVERWRITE, &failed_dev->flags)))))) {
2472 /* We would like to get this block, possibly by computing it,
2473 * otherwise read it if the backing disk is insync
2475 if ((s->uptodate == disks - 1) &&
2476 (s->failed && disk_idx == s->failed_num)) {
2477 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
2478 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
2479 set_bit(R5_Wantcompute, &dev->flags);
2480 sh->ops.target = disk_idx;
2481 sh->ops.target2 = -1;
2483 /* Careful: from this point on 'uptodate' is in the eye
2484 * of raid_run_ops which services 'compute' operations
2485 * before writes. R5_Wantcompute flags a block that will
2486 * be R5_UPTODATE by the time it is needed for a
2487 * subsequent operation.
2490 return 1; /* uptodate + compute == disks */
2491 } else if (test_bit(R5_Insync, &dev->flags)) {
2492 set_bit(R5_LOCKED, &dev->flags);
2493 set_bit(R5_Wantread, &dev->flags);
2495 pr_debug("Reading block %d (sync=%d)\n", disk_idx,
2504 * handle_stripe_fill5 - read or compute data to satisfy pending requests.
2506 static void handle_stripe_fill5(struct stripe_head *sh,
2507 struct stripe_head_state *s, int disks)
2511 /* look for blocks to read/compute, skip this if a compute
2512 * is already in flight, or if the stripe contents are in the
2513 * midst of changing due to a write
2515 if (!test_bit(STRIPE_COMPUTE_RUN, &sh->state) && !sh->check_state &&
2516 !sh->reconstruct_state)
2517 for (i = disks; i--; )
2518 if (fetch_block5(sh, s, i, disks))
2520 set_bit(STRIPE_HANDLE, &sh->state);
2523 /* fetch_block6 - checks the given member device to see if its data needs
2524 * to be read or computed to satisfy a request.
2526 * Returns 1 when no more member devices need to be checked, otherwise returns
2527 * 0 to tell the loop in handle_stripe_fill6 to continue
2529 static int fetch_block6(struct stripe_head *sh, struct stripe_head_state *s,
2530 struct r6_state *r6s, int disk_idx, int disks)
2532 struct r5dev *dev = &sh->dev[disk_idx];
2533 struct r5dev *fdev[2] = { &sh->dev[r6s->failed_num[0]],
2534 &sh->dev[r6s->failed_num[1]] };
2536 if (!test_bit(R5_LOCKED, &dev->flags) &&
2537 !test_bit(R5_UPTODATE, &dev->flags) &&
2539 (dev->towrite && !test_bit(R5_OVERWRITE, &dev->flags)) ||
2540 s->syncing || s->expanding ||
2542 (fdev[0]->toread || s->to_write)) ||
2544 (fdev[1]->toread || s->to_write)))) {
2545 /* we would like to get this block, possibly by computing it,
2546 * otherwise read it if the backing disk is insync
2548 BUG_ON(test_bit(R5_Wantcompute, &dev->flags));
2549 BUG_ON(test_bit(R5_Wantread, &dev->flags));
2550 if ((s->uptodate == disks - 1) &&
2551 (s->failed && (disk_idx == r6s->failed_num[0] ||
2552 disk_idx == r6s->failed_num[1]))) {
2553 /* have disk failed, and we're requested to fetch it;
2556 pr_debug("Computing stripe %llu block %d\n",
2557 (unsigned long long)sh->sector, disk_idx);
2558 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
2559 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
2560 set_bit(R5_Wantcompute, &dev->flags);
2561 sh->ops.target = disk_idx;
2562 sh->ops.target2 = -1; /* no 2nd target */
2566 } else if (s->uptodate == disks-2 && s->failed >= 2) {
2567 /* Computing 2-failure is *very* expensive; only
2568 * do it if failed >= 2
2571 for (other = disks; other--; ) {
2572 if (other == disk_idx)
2574 if (!test_bit(R5_UPTODATE,
2575 &sh->dev[other].flags))
2579 pr_debug("Computing stripe %llu blocks %d,%d\n",
2580 (unsigned long long)sh->sector,
2582 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
2583 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
2584 set_bit(R5_Wantcompute, &sh->dev[disk_idx].flags);
2585 set_bit(R5_Wantcompute, &sh->dev[other].flags);
2586 sh->ops.target = disk_idx;
2587 sh->ops.target2 = other;
2591 } else if (test_bit(R5_Insync, &dev->flags)) {
2592 set_bit(R5_LOCKED, &dev->flags);
2593 set_bit(R5_Wantread, &dev->flags);
2595 pr_debug("Reading block %d (sync=%d)\n",
2596 disk_idx, s->syncing);
2604 * handle_stripe_fill6 - read or compute data to satisfy pending requests.
2606 static void handle_stripe_fill6(struct stripe_head *sh,
2607 struct stripe_head_state *s, struct r6_state *r6s,
2612 /* look for blocks to read/compute, skip this if a compute
2613 * is already in flight, or if the stripe contents are in the
2614 * midst of changing due to a write
2616 if (!test_bit(STRIPE_COMPUTE_RUN, &sh->state) && !sh->check_state &&
2617 !sh->reconstruct_state)
2618 for (i = disks; i--; )
2619 if (fetch_block6(sh, s, r6s, i, disks))
2621 set_bit(STRIPE_HANDLE, &sh->state);
2625 /* handle_stripe_clean_event
2626 * any written block on an uptodate or failed drive can be returned.
2627 * Note that if we 'wrote' to a failed drive, it will be UPTODATE, but
2628 * never LOCKED, so we don't need to test 'failed' directly.
2630 static void handle_stripe_clean_event(raid5_conf_t *conf,
2631 struct stripe_head *sh, int disks, struct bio **return_bi)
2636 for (i = disks; i--; )
2637 if (sh->dev[i].written) {
2639 if (!test_bit(R5_LOCKED, &dev->flags) &&
2640 test_bit(R5_UPTODATE, &dev->flags)) {
2641 /* We can return any write requests */
2642 struct bio *wbi, *wbi2;
2644 pr_debug("Return write for disc %d\n", i);
2645 spin_lock_irq(&conf->device_lock);
2647 dev->written = NULL;
2648 while (wbi && wbi->bi_sector <
2649 dev->sector + STRIPE_SECTORS) {
2650 wbi2 = r5_next_bio(wbi, dev->sector);
2651 if (!raid5_dec_bi_phys_segments(wbi)) {
2652 md_write_end(conf->mddev);
2653 wbi->bi_next = *return_bi;
2658 if (dev->towrite == NULL)
2660 spin_unlock_irq(&conf->device_lock);
2662 bitmap_endwrite(conf->mddev->bitmap,
2665 !test_bit(STRIPE_DEGRADED, &sh->state),
2670 if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
2671 if (atomic_dec_and_test(&conf->pending_full_writes))
2672 md_wakeup_thread(conf->mddev->thread);
2675 static void handle_stripe_dirtying5(raid5_conf_t *conf,
2676 struct stripe_head *sh, struct stripe_head_state *s, int disks)
2678 int rmw = 0, rcw = 0, i;
2679 for (i = disks; i--; ) {
2680 /* would I have to read this buffer for read_modify_write */
2681 struct r5dev *dev = &sh->dev[i];
2682 if ((dev->towrite || i == sh->pd_idx) &&
2683 !test_bit(R5_LOCKED, &dev->flags) &&
2684 !(test_bit(R5_UPTODATE, &dev->flags) ||
2685 test_bit(R5_Wantcompute, &dev->flags))) {
2686 if (test_bit(R5_Insync, &dev->flags))
2689 rmw += 2*disks; /* cannot read it */
2691 /* Would I have to read this buffer for reconstruct_write */
2692 if (!test_bit(R5_OVERWRITE, &dev->flags) && i != sh->pd_idx &&
2693 !test_bit(R5_LOCKED, &dev->flags) &&
2694 !(test_bit(R5_UPTODATE, &dev->flags) ||
2695 test_bit(R5_Wantcompute, &dev->flags))) {
2696 if (test_bit(R5_Insync, &dev->flags)) rcw++;
2701 pr_debug("for sector %llu, rmw=%d rcw=%d\n",
2702 (unsigned long long)sh->sector, rmw, rcw);
2703 set_bit(STRIPE_HANDLE, &sh->state);
2704 if (rmw < rcw && rmw > 0)
2705 /* prefer read-modify-write, but need to get some data */
2706 for (i = disks; i--; ) {
2707 struct r5dev *dev = &sh->dev[i];
2708 if ((dev->towrite || i == sh->pd_idx) &&
2709 !test_bit(R5_LOCKED, &dev->flags) &&
2710 !(test_bit(R5_UPTODATE, &dev->flags) ||
2711 test_bit(R5_Wantcompute, &dev->flags)) &&
2712 test_bit(R5_Insync, &dev->flags)) {
2714 test_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2715 pr_debug("Read_old block "
2716 "%d for r-m-w\n", i);
2717 set_bit(R5_LOCKED, &dev->flags);
2718 set_bit(R5_Wantread, &dev->flags);
2721 set_bit(STRIPE_DELAYED, &sh->state);
2722 set_bit(STRIPE_HANDLE, &sh->state);
2726 if (rcw <= rmw && rcw > 0)
2727 /* want reconstruct write, but need to get some data */
2728 for (i = disks; i--; ) {
2729 struct r5dev *dev = &sh->dev[i];
2730 if (!test_bit(R5_OVERWRITE, &dev->flags) &&
2732 !test_bit(R5_LOCKED, &dev->flags) &&
2733 !(test_bit(R5_UPTODATE, &dev->flags) ||
2734 test_bit(R5_Wantcompute, &dev->flags)) &&
2735 test_bit(R5_Insync, &dev->flags)) {
2737 test_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2738 pr_debug("Read_old block "
2739 "%d for Reconstruct\n", i);
2740 set_bit(R5_LOCKED, &dev->flags);
2741 set_bit(R5_Wantread, &dev->flags);
2744 set_bit(STRIPE_DELAYED, &sh->state);
2745 set_bit(STRIPE_HANDLE, &sh->state);
2749 /* now if nothing is locked, and if we have enough data,
2750 * we can start a write request
2752 /* since handle_stripe can be called at any time we need to handle the
2753 * case where a compute block operation has been submitted and then a
2754 * subsequent call wants to start a write request. raid_run_ops only
2755 * handles the case where compute block and reconstruct are requested
2756 * simultaneously. If this is not the case then new writes need to be
2757 * held off until the compute completes.
2759 if ((s->req_compute || !test_bit(STRIPE_COMPUTE_RUN, &sh->state)) &&
2760 (s->locked == 0 && (rcw == 0 || rmw == 0) &&
2761 !test_bit(STRIPE_BIT_DELAY, &sh->state)))
2762 schedule_reconstruction(sh, s, rcw == 0, 0);
2765 static void handle_stripe_dirtying6(raid5_conf_t *conf,
2766 struct stripe_head *sh, struct stripe_head_state *s,
2767 struct r6_state *r6s, int disks)
2769 int rcw = 0, pd_idx = sh->pd_idx, i;
2770 int qd_idx = sh->qd_idx;
2772 set_bit(STRIPE_HANDLE, &sh->state);
2773 for (i = disks; i--; ) {
2774 struct r5dev *dev = &sh->dev[i];
2775 /* check if we haven't enough data */
2776 if (!test_bit(R5_OVERWRITE, &dev->flags) &&
2777 i != pd_idx && i != qd_idx &&
2778 !test_bit(R5_LOCKED, &dev->flags) &&
2779 !(test_bit(R5_UPTODATE, &dev->flags) ||
2780 test_bit(R5_Wantcompute, &dev->flags))) {
2782 if (!test_bit(R5_Insync, &dev->flags))
2783 continue; /* it's a failed drive */
2786 test_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2787 pr_debug("Read_old stripe %llu "
2788 "block %d for Reconstruct\n",
2789 (unsigned long long)sh->sector, i);
2790 set_bit(R5_LOCKED, &dev->flags);
2791 set_bit(R5_Wantread, &dev->flags);
2794 pr_debug("Request delayed stripe %llu "
2795 "block %d for Reconstruct\n",
2796 (unsigned long long)sh->sector, i);
2797 set_bit(STRIPE_DELAYED, &sh->state);
2798 set_bit(STRIPE_HANDLE, &sh->state);
2802 /* now if nothing is locked, and if we have enough data, we can start a
2805 if ((s->req_compute || !test_bit(STRIPE_COMPUTE_RUN, &sh->state)) &&
2806 s->locked == 0 && rcw == 0 &&
2807 !test_bit(STRIPE_BIT_DELAY, &sh->state)) {
2808 schedule_reconstruction(sh, s, 1, 0);
2812 static void handle_parity_checks5(raid5_conf_t *conf, struct stripe_head *sh,
2813 struct stripe_head_state *s, int disks)
2815 struct r5dev *dev = NULL;
2817 set_bit(STRIPE_HANDLE, &sh->state);
2819 switch (sh->check_state) {
2820 case check_state_idle:
2821 /* start a new check operation if there are no failures */
2822 if (s->failed == 0) {
2823 BUG_ON(s->uptodate != disks);
2824 sh->check_state = check_state_run;
2825 set_bit(STRIPE_OP_CHECK, &s->ops_request);
2826 clear_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags);
2830 dev = &sh->dev[s->failed_num];
2832 case check_state_compute_result:
2833 sh->check_state = check_state_idle;
2835 dev = &sh->dev[sh->pd_idx];
2837 /* check that a write has not made the stripe insync */
2838 if (test_bit(STRIPE_INSYNC, &sh->state))
2841 /* either failed parity check, or recovery is happening */
2842 BUG_ON(!test_bit(R5_UPTODATE, &dev->flags));
2843 BUG_ON(s->uptodate != disks);
2845 set_bit(R5_LOCKED, &dev->flags);
2847 set_bit(R5_Wantwrite, &dev->flags);
2849 clear_bit(STRIPE_DEGRADED, &sh->state);
2850 set_bit(STRIPE_INSYNC, &sh->state);
2852 case check_state_run:
2853 break; /* we will be called again upon completion */
2854 case check_state_check_result:
2855 sh->check_state = check_state_idle;
2857 /* if a failure occurred during the check operation, leave
2858 * STRIPE_INSYNC not set and let the stripe be handled again
2863 /* handle a successful check operation, if parity is correct
2864 * we are done. Otherwise update the mismatch count and repair
2865 * parity if !MD_RECOVERY_CHECK
2867 if ((sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) == 0)
2868 /* parity is correct (on disc,
2869 * not in buffer any more)
2871 set_bit(STRIPE_INSYNC, &sh->state);
2873 conf->mddev->resync_mismatches += STRIPE_SECTORS;
2874 if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery))
2875 /* don't try to repair!! */
2876 set_bit(STRIPE_INSYNC, &sh->state);
2878 sh->check_state = check_state_compute_run;
2879 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
2880 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
2881 set_bit(R5_Wantcompute,
2882 &sh->dev[sh->pd_idx].flags);
2883 sh->ops.target = sh->pd_idx;
2884 sh->ops.target2 = -1;
2889 case check_state_compute_run:
2892 printk(KERN_ERR "%s: unknown check_state: %d sector: %llu\n",
2893 __func__, sh->check_state,
2894 (unsigned long long) sh->sector);
2900 static void handle_parity_checks6(raid5_conf_t *conf, struct stripe_head *sh,
2901 struct stripe_head_state *s,
2902 struct r6_state *r6s, int disks)
2904 int pd_idx = sh->pd_idx;
2905 int qd_idx = sh->qd_idx;
2908 set_bit(STRIPE_HANDLE, &sh->state);
2910 BUG_ON(s->failed > 2);
2912 /* Want to check and possibly repair P and Q.
2913 * However there could be one 'failed' device, in which
2914 * case we can only check one of them, possibly using the
2915 * other to generate missing data
2918 switch (sh->check_state) {
2919 case check_state_idle:
2920 /* start a new check operation if there are < 2 failures */
2921 if (s->failed == r6s->q_failed) {
2922 /* The only possible failed device holds Q, so it
2923 * makes sense to check P (If anything else were failed,
2924 * we would have used P to recreate it).
2926 sh->check_state = check_state_run;
2928 if (!r6s->q_failed && s->failed < 2) {
2929 /* Q is not failed, and we didn't use it to generate
2930 * anything, so it makes sense to check it
2932 if (sh->check_state == check_state_run)
2933 sh->check_state = check_state_run_pq;
2935 sh->check_state = check_state_run_q;
2938 /* discard potentially stale zero_sum_result */
2939 sh->ops.zero_sum_result = 0;
2941 if (sh->check_state == check_state_run) {
2942 /* async_xor_zero_sum destroys the contents of P */
2943 clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
2946 if (sh->check_state >= check_state_run &&
2947 sh->check_state <= check_state_run_pq) {
2948 /* async_syndrome_zero_sum preserves P and Q, so
2949 * no need to mark them !uptodate here
2951 set_bit(STRIPE_OP_CHECK, &s->ops_request);
2955 /* we have 2-disk failure */
2956 BUG_ON(s->failed != 2);
2958 case check_state_compute_result:
2959 sh->check_state = check_state_idle;
2961 /* check that a write has not made the stripe insync */
2962 if (test_bit(STRIPE_INSYNC, &sh->state))
2965 /* now write out any block on a failed drive,
2966 * or P or Q if they were recomputed
2968 BUG_ON(s->uptodate < disks - 1); /* We don't need Q to recover */
2969 if (s->failed == 2) {
2970 dev = &sh->dev[r6s->failed_num[1]];
2972 set_bit(R5_LOCKED, &dev->flags);
2973 set_bit(R5_Wantwrite, &dev->flags);
2975 if (s->failed >= 1) {
2976 dev = &sh->dev[r6s->failed_num[0]];
2978 set_bit(R5_LOCKED, &dev->flags);
2979 set_bit(R5_Wantwrite, &dev->flags);
2981 if (sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) {
2982 dev = &sh->dev[pd_idx];
2984 set_bit(R5_LOCKED, &dev->flags);
2985 set_bit(R5_Wantwrite, &dev->flags);
2987 if (sh->ops.zero_sum_result & SUM_CHECK_Q_RESULT) {
2988 dev = &sh->dev[qd_idx];
2990 set_bit(R5_LOCKED, &dev->flags);
2991 set_bit(R5_Wantwrite, &dev->flags);
2993 clear_bit(STRIPE_DEGRADED, &sh->state);
2995 set_bit(STRIPE_INSYNC, &sh->state);
2997 case check_state_run:
2998 case check_state_run_q:
2999 case check_state_run_pq:
3000 break; /* we will be called again upon completion */
3001 case check_state_check_result:
3002 sh->check_state = check_state_idle;
3004 /* handle a successful check operation, if parity is correct
3005 * we are done. Otherwise update the mismatch count and repair
3006 * parity if !MD_RECOVERY_CHECK
3008 if (sh->ops.zero_sum_result == 0) {
3009 /* both parities are correct */
3011 set_bit(STRIPE_INSYNC, &sh->state);
3013 /* in contrast to the raid5 case we can validate
3014 * parity, but still have a failure to write
3017 sh->check_state = check_state_compute_result;
3018 /* Returning at this point means that we may go
3019 * off and bring p and/or q uptodate again so
3020 * we make sure to check zero_sum_result again
3021 * to verify if p or q need writeback
3025 conf->mddev->resync_mismatches += STRIPE_SECTORS;
3026 if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery))
3027 /* don't try to repair!! */
3028 set_bit(STRIPE_INSYNC, &sh->state);
3030 int *target = &sh->ops.target;
3032 sh->ops.target = -1;
3033 sh->ops.target2 = -1;
3034 sh->check_state = check_state_compute_run;
3035 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
3036 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
3037 if (sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) {
3038 set_bit(R5_Wantcompute,
3039 &sh->dev[pd_idx].flags);
3041 target = &sh->ops.target2;
3044 if (sh->ops.zero_sum_result & SUM_CHECK_Q_RESULT) {
3045 set_bit(R5_Wantcompute,
3046 &sh->dev[qd_idx].flags);
3053 case check_state_compute_run:
3056 printk(KERN_ERR "%s: unknown check_state: %d sector: %llu\n",
3057 __func__, sh->check_state,
3058 (unsigned long long) sh->sector);
3063 static void handle_stripe_expansion(raid5_conf_t *conf, struct stripe_head *sh,
3064 struct r6_state *r6s)
3068 /* We have read all the blocks in this stripe and now we need to
3069 * copy some of them into a target stripe for expand.
3071 struct dma_async_tx_descriptor *tx = NULL;
3072 clear_bit(STRIPE_EXPAND_SOURCE, &sh->state);
3073 for (i = 0; i < sh->disks; i++)
3074 if (i != sh->pd_idx && i != sh->qd_idx) {
3076 struct stripe_head *sh2;
3077 struct async_submit_ctl submit;
3079 sector_t bn = compute_blocknr(sh, i, 1);
3080 sector_t s = raid5_compute_sector(conf, bn, 0,
3082 sh2 = get_active_stripe(conf, s, 0, 1);
3084 /* so far only the early blocks of this stripe
3085 * have been requested. When later blocks
3086 * get requested, we will try again
3089 if (!test_bit(STRIPE_EXPANDING, &sh2->state) ||
3090 test_bit(R5_Expanded, &sh2->dev[dd_idx].flags)) {
3091 /* must have already done this block */
3092 release_stripe(sh2);
3096 /* place all the copies on one channel */
3097 init_async_submit(&submit, 0, tx, NULL, NULL, NULL);
3098 tx = async_memcpy(sh2->dev[dd_idx].page,
3099 sh->dev[i].page, 0, 0, STRIPE_SIZE,
3102 set_bit(R5_Expanded, &sh2->dev[dd_idx].flags);
3103 set_bit(R5_UPTODATE, &sh2->dev[dd_idx].flags);
3104 for (j = 0; j < conf->raid_disks; j++)
3105 if (j != sh2->pd_idx &&
3106 (!r6s || j != sh2->qd_idx) &&
3107 !test_bit(R5_Expanded, &sh2->dev[j].flags))
3109 if (j == conf->raid_disks) {
3110 set_bit(STRIPE_EXPAND_READY, &sh2->state);
3111 set_bit(STRIPE_HANDLE, &sh2->state);
3113 release_stripe(sh2);
3116 /* done submitting copies, wait for them to complete */
3119 dma_wait_for_async_tx(tx);
3125 * handle_stripe - do things to a stripe.
3127 * We lock the stripe and then examine the state of various bits
3128 * to see what needs to be done.
3130 * return some read request which now have data
3131 * return some write requests which are safely on disc
3132 * schedule a read on some buffers
3133 * schedule a write of some buffers
3134 * return confirmation of parity correctness
3136 * buffers are taken off read_list or write_list, and bh_cache buffers
3137 * get BH_Lock set before the stripe lock is released.
3141 static bool handle_stripe5(struct stripe_head *sh)
3143 raid5_conf_t *conf = sh->raid_conf;
3144 int disks = sh->disks, i;
3145 struct bio *return_bi = NULL;
3146 struct stripe_head_state s;
3148 mdk_rdev_t *blocked_rdev = NULL;
3151 memset(&s, 0, sizeof(s));
3152 pr_debug("handling stripe %llu, state=%#lx cnt=%d, pd_idx=%d check:%d "
3153 "reconstruct:%d\n", (unsigned long long)sh->sector, sh->state,
3154 atomic_read(&sh->count), sh->pd_idx, sh->check_state,
3155 sh->reconstruct_state);
3157 spin_lock(&sh->lock);
3158 clear_bit(STRIPE_HANDLE, &sh->state);
3159 clear_bit(STRIPE_DELAYED, &sh->state);
3161 s.syncing = test_bit(STRIPE_SYNCING, &sh->state);
3162 s.expanding = test_bit(STRIPE_EXPAND_SOURCE, &sh->state);
3163 s.expanded = test_bit(STRIPE_EXPAND_READY, &sh->state);
3165 /* Now to look around and see what can be done */
3167 for (i=disks; i--; ) {
3169 struct r5dev *dev = &sh->dev[i];
3170 clear_bit(R5_Insync, &dev->flags);
3172 pr_debug("check %d: state 0x%lx toread %p read %p write %p "
3173 "written %p\n", i, dev->flags, dev->toread, dev->read,
3174 dev->towrite, dev->written);
3176 /* maybe we can request a biofill operation
3178 * new wantfill requests are only permitted while
3179 * ops_complete_biofill is guaranteed to be inactive
3181 if (test_bit(R5_UPTODATE, &dev->flags) && dev->toread &&
3182 !test_bit(STRIPE_BIOFILL_RUN, &sh->state))
3183 set_bit(R5_Wantfill, &dev->flags);
3185 /* now count some things */
3186 if (test_bit(R5_LOCKED, &dev->flags)) s.locked++;
3187 if (test_bit(R5_UPTODATE, &dev->flags)) s.uptodate++;
3188 if (test_bit(R5_Wantcompute, &dev->flags)) s.compute++;
3190 if (test_bit(R5_Wantfill, &dev->flags))
3192 else if (dev->toread)
3196 if (!test_bit(R5_OVERWRITE, &dev->flags))
3201 rdev = rcu_dereference(conf->disks[i].rdev);
3202 if (blocked_rdev == NULL &&
3203 rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
3204 blocked_rdev = rdev;
3205 atomic_inc(&rdev->nr_pending);
3207 if (!rdev || !test_bit(In_sync, &rdev->flags)) {
3208 /* The ReadError flag will just be confusing now */
3209 clear_bit(R5_ReadError, &dev->flags);
3210 clear_bit(R5_ReWrite, &dev->flags);
3212 if (!rdev || !test_bit(In_sync, &rdev->flags)
3213 || test_bit(R5_ReadError, &dev->flags)) {
3217 set_bit(R5_Insync, &dev->flags);
3221 if (unlikely(blocked_rdev)) {
3222 if (s.syncing || s.expanding || s.expanded ||
3223 s.to_write || s.written) {
3224 set_bit(STRIPE_HANDLE, &sh->state);
3227 /* There is nothing for the blocked_rdev to block */
3228 rdev_dec_pending(blocked_rdev, conf->mddev);
3229 blocked_rdev = NULL;
3232 if (s.to_fill && !test_bit(STRIPE_BIOFILL_RUN, &sh->state)) {
3233 set_bit(STRIPE_OP_BIOFILL, &s.ops_request);
3234 set_bit(STRIPE_BIOFILL_RUN, &sh->state);
3237 pr_debug("locked=%d uptodate=%d to_read=%d"
3238 " to_write=%d failed=%d failed_num=%d\n",
3239 s.locked, s.uptodate, s.to_read, s.to_write,
3240 s.failed, s.failed_num);
3241 /* check if the array has lost two devices and, if so, some requests might
3244 if (s.failed > 1 && s.to_read+s.to_write+s.written)
3245 handle_failed_stripe(conf, sh, &s, disks, &return_bi);
3246 if (s.failed > 1 && s.syncing) {
3247 md_done_sync(conf->mddev, STRIPE_SECTORS,0);
3248 clear_bit(STRIPE_SYNCING, &sh->state);
3252 /* might be able to return some write requests if the parity block
3253 * is safe, or on a failed drive
3255 dev = &sh->dev[sh->pd_idx];
3257 ((test_bit(R5_Insync, &dev->flags) &&
3258 !test_bit(R5_LOCKED, &dev->flags) &&
3259 test_bit(R5_UPTODATE, &dev->flags)) ||
3260 (s.failed == 1 && s.failed_num == sh->pd_idx)))
3261 handle_stripe_clean_event(conf, sh, disks, &return_bi);
3263 /* Now we might consider reading some blocks, either to check/generate
3264 * parity, or to satisfy requests
3265 * or to load a block that is being partially written.
3267 if (s.to_read || s.non_overwrite ||
3268 (s.syncing && (s.uptodate + s.compute < disks)) || s.expanding)
3269 handle_stripe_fill5(sh, &s, disks);
3271 /* Now we check to see if any write operations have recently
3275 if (sh->reconstruct_state == reconstruct_state_prexor_drain_result)
3277 if (sh->reconstruct_state == reconstruct_state_drain_result ||
3278 sh->reconstruct_state == reconstruct_state_prexor_drain_result) {
3279 sh->reconstruct_state = reconstruct_state_idle;
3281 /* All the 'written' buffers and the parity block are ready to
3282 * be written back to disk
3284 BUG_ON(!test_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags));
3285 for (i = disks; i--; ) {
3287 if (test_bit(R5_LOCKED, &dev->flags) &&
3288 (i == sh->pd_idx || dev->written)) {
3289 pr_debug("Writing block %d\n", i);
3290 set_bit(R5_Wantwrite, &dev->flags);
3293 if (!test_bit(R5_Insync, &dev->flags) ||
3294 (i == sh->pd_idx && s.failed == 0))
3295 set_bit(STRIPE_INSYNC, &sh->state);
3298 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
3299 atomic_dec(&conf->preread_active_stripes);
3300 if (atomic_read(&conf->preread_active_stripes) <
3302 md_wakeup_thread(conf->mddev->thread);
3306 /* Now to consider new write requests and what else, if anything
3307 * should be read. We do not handle new writes when:
3308 * 1/ A 'write' operation (copy+xor) is already in flight.
3309 * 2/ A 'check' operation is in flight, as it may clobber the parity
3312 if (s.to_write && !sh->reconstruct_state && !sh->check_state)
3313 handle_stripe_dirtying5(conf, sh, &s, disks);
3315 /* maybe we need to check and possibly fix the parity for this stripe
3316 * Any reads will already have been scheduled, so we just see if enough
3317 * data is available. The parity check is held off while parity
3318 * dependent operations are in flight.
3320 if (sh->check_state ||
3321 (s.syncing && s.locked == 0 &&
3322 !test_bit(STRIPE_COMPUTE_RUN, &sh->state) &&
3323 !test_bit(STRIPE_INSYNC, &sh->state)))
3324 handle_parity_checks5(conf, sh, &s, disks);
3326 if (s.syncing && s.locked == 0 && test_bit(STRIPE_INSYNC, &sh->state)) {
3327 md_done_sync(conf->mddev, STRIPE_SECTORS,1);
3328 clear_bit(STRIPE_SYNCING, &sh->state);
3331 /* If the failed drive is just a ReadError, then we might need to progress
3332 * the repair/check process
3334 if (s.failed == 1 && !conf->mddev->ro &&
3335 test_bit(R5_ReadError, &sh->dev[s.failed_num].flags)
3336 && !test_bit(R5_LOCKED, &sh->dev[s.failed_num].flags)
3337 && test_bit(R5_UPTODATE, &sh->dev[s.failed_num].flags)
3339 dev = &sh->dev[s.failed_num];
3340 if (!test_bit(R5_ReWrite, &dev->flags)) {
3341 set_bit(R5_Wantwrite, &dev->flags);
3342 set_bit(R5_ReWrite, &dev->flags);
3343 set_bit(R5_LOCKED, &dev->flags);
3346 /* let's read it back */
3347 set_bit(R5_Wantread, &dev->flags);
3348 set_bit(R5_LOCKED, &dev->flags);
3353 /* Finish reconstruct operations initiated by the expansion process */
3354 if (sh->reconstruct_state == reconstruct_state_result) {
3355 struct stripe_head *sh2
3356 = get_active_stripe(conf, sh->sector, 1, 1);
3357 if (sh2 && test_bit(STRIPE_EXPAND_SOURCE, &sh2->state)) {
3358 /* sh cannot be written until sh2 has been read.
3359 * so arrange for sh to be delayed a little
3361 set_bit(STRIPE_DELAYED, &sh->state);
3362 set_bit(STRIPE_HANDLE, &sh->state);
3363 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE,
3365 atomic_inc(&conf->preread_active_stripes);
3366 release_stripe(sh2);
3370 release_stripe(sh2);
3372 sh->reconstruct_state = reconstruct_state_idle;
3373 clear_bit(STRIPE_EXPANDING, &sh->state);
3374 for (i = conf->raid_disks; i--; ) {
3375 set_bit(R5_Wantwrite, &sh->dev[i].flags);
3376 set_bit(R5_LOCKED, &sh->dev[i].flags);
3381 if (s.expanded && test_bit(STRIPE_EXPANDING, &sh->state) &&
3382 !sh->reconstruct_state) {
3383 /* Need to write out all blocks after computing parity */
3384 sh->disks = conf->raid_disks;
3385 stripe_set_idx(sh->sector, conf, 0, sh);
3386 schedule_reconstruction(sh, &s, 1, 1);
3387 } else if (s.expanded && !sh->reconstruct_state && s.locked == 0) {
3388 clear_bit(STRIPE_EXPAND_READY, &sh->state);
3389 atomic_dec(&conf->reshape_stripes);
3390 wake_up(&conf->wait_for_overlap);
3391 md_done_sync(conf->mddev, STRIPE_SECTORS, 1);
3394 if (s.expanding && s.locked == 0 &&
3395 !test_bit(STRIPE_COMPUTE_RUN, &sh->state))
3396 handle_stripe_expansion(conf, sh, NULL);
3399 spin_unlock(&sh->lock);
3401 /* wait for this device to become unblocked */
3402 if (unlikely(blocked_rdev))
3403 md_wait_for_blocked_rdev(blocked_rdev, conf->mddev);
3406 raid_run_ops(sh, s.ops_request);
3410 return_io(return_bi);
3412 return blocked_rdev == NULL;
3415 static bool handle_stripe6(struct stripe_head *sh)
3417 raid5_conf_t *conf = sh->raid_conf;
3418 int disks = sh->disks;
3419 struct bio *return_bi = NULL;
3420 int i, pd_idx = sh->pd_idx, qd_idx = sh->qd_idx;
3421 struct stripe_head_state s;
3422 struct r6_state r6s;
3423 struct r5dev *dev, *pdev, *qdev;
3424 mdk_rdev_t *blocked_rdev = NULL;
3426 pr_debug("handling stripe %llu, state=%#lx cnt=%d, "
3427 "pd_idx=%d, qd_idx=%d\n, check:%d, reconstruct:%d\n",
3428 (unsigned long long)sh->sector, sh->state,
3429 atomic_read(&sh->count), pd_idx, qd_idx,
3430 sh->check_state, sh->reconstruct_state);
3431 memset(&s, 0, sizeof(s));
3433 spin_lock(&sh->lock);
3434 clear_bit(STRIPE_HANDLE, &sh->state);
3435 clear_bit(STRIPE_DELAYED, &sh->state);
3437 s.syncing = test_bit(STRIPE_SYNCING, &sh->state);
3438 s.expanding = test_bit(STRIPE_EXPAND_SOURCE, &sh->state);
3439 s.expanded = test_bit(STRIPE_EXPAND_READY, &sh->state);
3440 /* Now to look around and see what can be done */
3443 for (i=disks; i--; ) {
3446 clear_bit(R5_Insync, &dev->flags);
3448 pr_debug("check %d: state 0x%lx read %p write %p written %p\n",
3449 i, dev->flags, dev->toread, dev->towrite, dev->written);
3450 /* maybe we can reply to a read
3452 * new wantfill requests are only permitted while
3453 * ops_complete_biofill is guaranteed to be inactive
3455 if (test_bit(R5_UPTODATE, &dev->flags) && dev->toread &&
3456 !test_bit(STRIPE_BIOFILL_RUN, &sh->state))
3457 set_bit(R5_Wantfill, &dev->flags);
3459 /* now count some things */
3460 if (test_bit(R5_LOCKED, &dev->flags)) s.locked++;
3461 if (test_bit(R5_UPTODATE, &dev->flags)) s.uptodate++;
3462 if (test_bit(R5_Wantcompute, &dev->flags))
3463 BUG_ON(++s.compute > 2);
3465 if (test_bit(R5_Wantfill, &dev->flags)) {
3467 } else if (dev->toread)
3471 if (!test_bit(R5_OVERWRITE, &dev->flags))
3476 rdev = rcu_dereference(conf->disks[i].rdev);
3477 if (blocked_rdev == NULL &&
3478 rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
3479 blocked_rdev = rdev;
3480 atomic_inc(&rdev->nr_pending);
3482 if (!rdev || !test_bit(In_sync, &rdev->flags)) {
3483 /* The ReadError flag will just be confusing now */
3484 clear_bit(R5_ReadError, &dev->flags);
3485 clear_bit(R5_ReWrite, &dev->flags);
3487 if (!rdev || !test_bit(In_sync, &rdev->flags)
3488 || test_bit(R5_ReadError, &dev->flags)) {
3490 r6s.failed_num[s.failed] = i;
3493 set_bit(R5_Insync, &dev->flags);
3497 if (unlikely(blocked_rdev)) {
3498 if (s.syncing || s.expanding || s.expanded ||
3499 s.to_write || s.written) {
3500 set_bit(STRIPE_HANDLE, &sh->state);
3503 /* There is nothing for the blocked_rdev to block */
3504 rdev_dec_pending(blocked_rdev, conf->mddev);
3505 blocked_rdev = NULL;
3508 if (s.to_fill && !test_bit(STRIPE_BIOFILL_RUN, &sh->state)) {
3509 set_bit(STRIPE_OP_BIOFILL, &s.ops_request);
3510 set_bit(STRIPE_BIOFILL_RUN, &sh->state);
3513 pr_debug("locked=%d uptodate=%d to_read=%d"
3514 " to_write=%d failed=%d failed_num=%d,%d\n",
3515 s.locked, s.uptodate, s.to_read, s.to_write, s.failed,
3516 r6s.failed_num[0], r6s.failed_num[1]);
3517 /* check if the array has lost >2 devices and, if so, some requests
3518 * might need to be failed
3520 if (s.failed > 2 && s.to_read+s.to_write+s.written)
3521 handle_failed_stripe(conf, sh, &s, disks, &return_bi);
3522 if (s.failed > 2 && s.syncing) {
3523 md_done_sync(conf->mddev, STRIPE_SECTORS,0);
3524 clear_bit(STRIPE_SYNCING, &sh->state);
3529 * might be able to return some write requests if the parity blocks
3530 * are safe, or on a failed drive
3532 pdev = &sh->dev[pd_idx];
3533 r6s.p_failed = (s.failed >= 1 && r6s.failed_num[0] == pd_idx)
3534 || (s.failed >= 2 && r6s.failed_num[1] == pd_idx);
3535 qdev = &sh->dev[qd_idx];
3536 r6s.q_failed = (s.failed >= 1 && r6s.failed_num[0] == qd_idx)
3537 || (s.failed >= 2 && r6s.failed_num[1] == qd_idx);
3540 ( r6s.p_failed || ((test_bit(R5_Insync, &pdev->flags)
3541 && !test_bit(R5_LOCKED, &pdev->flags)
3542 && test_bit(R5_UPTODATE, &pdev->flags)))) &&
3543 ( r6s.q_failed || ((test_bit(R5_Insync, &qdev->flags)
3544 && !test_bit(R5_LOCKED, &qdev->flags)
3545 && test_bit(R5_UPTODATE, &qdev->flags)))))
3546 handle_stripe_clean_event(conf, sh, disks, &return_bi);
3548 /* Now we might consider reading some blocks, either to check/generate
3549 * parity, or to satisfy requests
3550 * or to load a block that is being partially written.
3552 if (s.to_read || s.non_overwrite || (s.to_write && s.failed) ||
3553 (s.syncing && (s.uptodate + s.compute < disks)) || s.expanding)
3554 handle_stripe_fill6(sh, &s, &r6s, disks);
3556 /* Now we check to see if any write operations have recently
3559 if (sh->reconstruct_state == reconstruct_state_drain_result) {
3560 int qd_idx = sh->qd_idx;
3562 sh->reconstruct_state = reconstruct_state_idle;
3563 /* All the 'written' buffers and the parity blocks are ready to
3564 * be written back to disk
3566 BUG_ON(!test_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags));
3567 BUG_ON(!test_bit(R5_UPTODATE, &sh->dev[qd_idx].flags));
3568 for (i = disks; i--; ) {
3570 if (test_bit(R5_LOCKED, &dev->flags) &&
3571 (i == sh->pd_idx || i == qd_idx ||
3573 pr_debug("Writing block %d\n", i);
3574 BUG_ON(!test_bit(R5_UPTODATE, &dev->flags));
3575 set_bit(R5_Wantwrite, &dev->flags);
3576 if (!test_bit(R5_Insync, &dev->flags) ||
3577 ((i == sh->pd_idx || i == qd_idx) &&
3579 set_bit(STRIPE_INSYNC, &sh->state);
3582 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
3583 atomic_dec(&conf->preread_active_stripes);
3584 if (atomic_read(&conf->preread_active_stripes) <
3586 md_wakeup_thread(conf->mddev->thread);
3590 /* Now to consider new write requests and what else, if anything
3591 * should be read. We do not handle new writes when:
3592 * 1/ A 'write' operation (copy+gen_syndrome) is already in flight.
3593 * 2/ A 'check' operation is in flight, as it may clobber the parity
3596 if (s.to_write && !sh->reconstruct_state && !sh->check_state)
3597 handle_stripe_dirtying6(conf, sh, &s, &r6s, disks);
3599 /* maybe we need to check and possibly fix the parity for this stripe
3600 * Any reads will already have been scheduled, so we just see if enough
3601 * data is available. The parity check is held off while parity
3602 * dependent operations are in flight.
3604 if (sh->check_state ||
3605 (s.syncing && s.locked == 0 &&
3606 !test_bit(STRIPE_COMPUTE_RUN, &sh->state) &&
3607 !test_bit(STRIPE_INSYNC, &sh->state)))
3608 handle_parity_checks6(conf, sh, &s, &r6s, disks);
3610 if (s.syncing && s.locked == 0 && test_bit(STRIPE_INSYNC, &sh->state)) {
3611 md_done_sync(conf->mddev, STRIPE_SECTORS,1);
3612 clear_bit(STRIPE_SYNCING, &sh->state);
3615 /* If the failed drives are just a ReadError, then we might need
3616 * to progress the repair/check process
3618 if (s.failed <= 2 && !conf->mddev->ro)
3619 for (i = 0; i < s.failed; i++) {
3620 dev = &sh->dev[r6s.failed_num[i]];
3621 if (test_bit(R5_ReadError, &dev->flags)
3622 && !test_bit(R5_LOCKED, &dev->flags)
3623 && test_bit(R5_UPTODATE, &dev->flags)
3625 if (!test_bit(R5_ReWrite, &dev->flags)) {
3626 set_bit(R5_Wantwrite, &dev->flags);
3627 set_bit(R5_ReWrite, &dev->flags);
3628 set_bit(R5_LOCKED, &dev->flags);
3631 /* let's read it back */
3632 set_bit(R5_Wantread, &dev->flags);
3633 set_bit(R5_LOCKED, &dev->flags);
3639 /* Finish reconstruct operations initiated by the expansion process */
3640 if (sh->reconstruct_state == reconstruct_state_result) {
3641 sh->reconstruct_state = reconstruct_state_idle;
3642 clear_bit(STRIPE_EXPANDING, &sh->state);
3643 for (i = conf->raid_disks; i--; ) {
3644 set_bit(R5_Wantwrite, &sh->dev[i].flags);
3645 set_bit(R5_LOCKED, &sh->dev[i].flags);
3650 if (s.expanded && test_bit(STRIPE_EXPANDING, &sh->state) &&
3651 !sh->reconstruct_state) {
3652 struct stripe_head *sh2
3653 = get_active_stripe(conf, sh->sector, 1, 1);
3654 if (sh2 && test_bit(STRIPE_EXPAND_SOURCE, &sh2->state)) {
3655 /* sh cannot be written until sh2 has been read.
3656 * so arrange for sh to be delayed a little
3658 set_bit(STRIPE_DELAYED, &sh->state);
3659 set_bit(STRIPE_HANDLE, &sh->state);
3660 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE,
3662 atomic_inc(&conf->preread_active_stripes);
3663 release_stripe(sh2);
3667 release_stripe(sh2);
3669 /* Need to write out all blocks after computing P&Q */
3670 sh->disks = conf->raid_disks;
3671 stripe_set_idx(sh->sector, conf, 0, sh);
3672 schedule_reconstruction(sh, &s, 1, 1);
3673 } else if (s.expanded && !sh->reconstruct_state && s.locked == 0) {
3674 clear_bit(STRIPE_EXPAND_READY, &sh->state);
3675 atomic_dec(&conf->reshape_stripes);
3676 wake_up(&conf->wait_for_overlap);
3677 md_done_sync(conf->mddev, STRIPE_SECTORS, 1);
3680 if (s.expanding && s.locked == 0 &&
3681 !test_bit(STRIPE_COMPUTE_RUN, &sh->state))
3682 handle_stripe_expansion(conf, sh, &r6s);
3685 spin_unlock(&sh->lock);
3687 /* wait for this device to become unblocked */
3688 if (unlikely(blocked_rdev))
3689 md_wait_for_blocked_rdev(blocked_rdev, conf->mddev);
3692 raid_run_ops(sh, s.ops_request);
3696 return_io(return_bi);
3698 return blocked_rdev == NULL;
3701 /* returns true if the stripe was handled */
3702 static bool handle_stripe(struct stripe_head *sh)
3704 if (sh->raid_conf->level == 6)
3705 return handle_stripe6(sh);
3707 return handle_stripe5(sh);
3710 static void raid5_activate_delayed(raid5_conf_t *conf)
3712 if (atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD) {
3713 while (!list_empty(&conf->delayed_list)) {
3714 struct list_head *l = conf->delayed_list.next;
3715 struct stripe_head *sh;
3716 sh = list_entry(l, struct stripe_head, lru);
3718 clear_bit(STRIPE_DELAYED, &sh->state);
3719 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
3720 atomic_inc(&conf->preread_active_stripes);
3721 list_add_tail(&sh->lru, &conf->hold_list);
3724 blk_plug_device(conf->mddev->queue);
3727 static void activate_bit_delay(raid5_conf_t *conf)
3729 /* device_lock is held */
3730 struct list_head head;
3731 list_add(&head, &conf->bitmap_list);
3732 list_del_init(&conf->bitmap_list);
3733 while (!list_empty(&head)) {
3734 struct stripe_head *sh = list_entry(head.next, struct stripe_head, lru);
3735 list_del_init(&sh->lru);
3736 atomic_inc(&sh->count);
3737 __release_stripe(conf, sh);
3741 static void unplug_slaves(mddev_t *mddev)
3743 raid5_conf_t *conf = mddev_to_conf(mddev);
3747 for (i=0; i<mddev->raid_disks; i++) {
3748 mdk_rdev_t *rdev = rcu_dereference(conf->disks[i].rdev);
3749 if (rdev && !test_bit(Faulty, &rdev->flags) && atomic_read(&rdev->nr_pending)) {
3750 struct request_queue *r_queue = bdev_get_queue(rdev->bdev);
3752 atomic_inc(&rdev->nr_pending);
3755 blk_unplug(r_queue);
3757 rdev_dec_pending(rdev, mddev);
3764 static void raid5_unplug_device(struct request_queue *q)
3766 mddev_t *mddev = q->queuedata;
3767 raid5_conf_t *conf = mddev_to_conf(mddev);
3768 unsigned long flags;
3770 spin_lock_irqsave(&conf->device_lock, flags);
3772 if (blk_remove_plug(q)) {
3774 raid5_activate_delayed(conf);
3776 md_wakeup_thread(mddev->thread);
3778 spin_unlock_irqrestore(&conf->device_lock, flags);
3780 unplug_slaves(mddev);
3783 static int raid5_congested(void *data, int bits)
3785 mddev_t *mddev = data;
3786 raid5_conf_t *conf = mddev_to_conf(mddev);
3788 /* No difference between reads and writes. Just check
3789 * how busy the stripe_cache is
3791 if (conf->inactive_blocked)
3795 if (list_empty_careful(&conf->inactive_list))
3801 /* We want read requests to align with chunks where possible,
3802 * but write requests don't need to.
3804 static int raid5_mergeable_bvec(struct request_queue *q,
3805 struct bvec_merge_data *bvm,
3806 struct bio_vec *biovec)
3808 mddev_t *mddev = q->queuedata;
3809 sector_t sector = bvm->bi_sector + get_start_sect(bvm->bi_bdev);
3811 unsigned int chunk_sectors = mddev->chunk_size >> 9;
3812 unsigned int bio_sectors = bvm->bi_size >> 9;
3814 if ((bvm->bi_rw & 1) == WRITE)
3815 return biovec->bv_len; /* always allow writes to be mergeable */
3817 if (mddev->new_chunk < mddev->chunk_size)
3818 chunk_sectors = mddev->new_chunk >> 9;
3819 max = (chunk_sectors - ((sector & (chunk_sectors - 1)) + bio_sectors)) << 9;
3820 if (max < 0) max = 0;
3821 if (max <= biovec->bv_len && bio_sectors == 0)
3822 return biovec->bv_len;
3828 static int in_chunk_boundary(mddev_t *mddev, struct bio *bio)
3830 sector_t sector = bio->bi_sector + get_start_sect(bio->bi_bdev);
3831 unsigned int chunk_sectors = mddev->chunk_size >> 9;
3832 unsigned int bio_sectors = bio->bi_size >> 9;
3834 if (mddev->new_chunk < mddev->chunk_size)
3835 chunk_sectors = mddev->new_chunk >> 9;
3836 return chunk_sectors >=
3837 ((sector & (chunk_sectors - 1)) + bio_sectors);
3841 * add bio to the retry LIFO ( in O(1) ... we are in interrupt )
3842 * later sampled by raid5d.
3844 static void add_bio_to_retry(struct bio *bi,raid5_conf_t *conf)
3846 unsigned long flags;
3848 spin_lock_irqsave(&conf->device_lock, flags);
3850 bi->bi_next = conf->retry_read_aligned_list;
3851 conf->retry_read_aligned_list = bi;
3853 spin_unlock_irqrestore(&conf->device_lock, flags);
3854 md_wakeup_thread(conf->mddev->thread);
3858 static struct bio *remove_bio_from_retry(raid5_conf_t *conf)
3862 bi = conf->retry_read_aligned;
3864 conf->retry_read_aligned = NULL;
3867 bi = conf->retry_read_aligned_list;
3869 conf->retry_read_aligned_list = bi->bi_next;
3872 * this sets the active strip count to 1 and the processed
3873 * strip count to zero (upper 8 bits)
3875 bi->bi_phys_segments = 1; /* biased count of active stripes */
3883 * The "raid5_align_endio" should check if the read succeeded and if it
3884 * did, call bio_endio on the original bio (having bio_put the new bio
3886 * If the read failed..
3888 static void raid5_align_endio(struct bio *bi, int error)
3890 struct bio* raid_bi = bi->bi_private;
3893 int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
3898 mddev = raid_bi->bi_bdev->bd_disk->queue->queuedata;
3899 conf = mddev_to_conf(mddev);
3900 rdev = (void*)raid_bi->bi_next;
3901 raid_bi->bi_next = NULL;
3903 rdev_dec_pending(rdev, conf->mddev);
3905 if (!error && uptodate) {
3906 bio_endio(raid_bi, 0);
3907 if (atomic_dec_and_test(&conf->active_aligned_reads))
3908 wake_up(&conf->wait_for_stripe);
3913 pr_debug("raid5_align_endio : io error...handing IO for a retry\n");
3915 add_bio_to_retry(raid_bi, conf);
3918 static int bio_fits_rdev(struct bio *bi)
3920 struct request_queue *q = bdev_get_queue(bi->bi_bdev);
3922 if ((bi->bi_size>>9) > q->max_sectors)
3924 blk_recount_segments(q, bi);
3925 if (bi->bi_phys_segments > q->max_phys_segments)
3928 if (q->merge_bvec_fn)
3929 /* it's too hard to apply the merge_bvec_fn at this stage,
3938 static int chunk_aligned_read(struct request_queue *q, struct bio * raid_bio)
3940 mddev_t *mddev = q->queuedata;
3941 raid5_conf_t *conf = mddev_to_conf(mddev);
3942 unsigned int dd_idx;
3943 struct bio* align_bi;
3946 if (!in_chunk_boundary(mddev, raid_bio)) {
3947 pr_debug("chunk_aligned_read : non aligned\n");
3951 * use bio_clone to make a copy of the bio
3953 align_bi = bio_clone(raid_bio, GFP_NOIO);
3957 * set bi_end_io to a new function, and set bi_private to the
3960 align_bi->bi_end_io = raid5_align_endio;
3961 align_bi->bi_private = raid_bio;
3965 align_bi->bi_sector = raid5_compute_sector(conf, raid_bio->bi_sector,
3970 rdev = rcu_dereference(conf->disks[dd_idx].rdev);
3971 if (rdev && test_bit(In_sync, &rdev->flags)) {
3972 atomic_inc(&rdev->nr_pending);
3974 raid_bio->bi_next = (void*)rdev;
3975 align_bi->bi_bdev = rdev->bdev;
3976 align_bi->bi_flags &= ~(1 << BIO_SEG_VALID);
3977 align_bi->bi_sector += rdev->data_offset;
3979 if (!bio_fits_rdev(align_bi)) {
3980 /* too big in some way */
3982 rdev_dec_pending(rdev, mddev);
3986 spin_lock_irq(&conf->device_lock);
3987 wait_event_lock_irq(conf->wait_for_stripe,
3989 conf->device_lock, /* nothing */);
3990 atomic_inc(&conf->active_aligned_reads);
3991 spin_unlock_irq(&conf->device_lock);
3993 generic_make_request(align_bi);
4002 /* __get_priority_stripe - get the next stripe to process
4004 * Full stripe writes are allowed to pass preread active stripes up until
4005 * the bypass_threshold is exceeded. In general the bypass_count
4006 * increments when the handle_list is handled before the hold_list; however, it
4007 * will not be incremented when STRIPE_IO_STARTED is sampled set signifying a
4008 * stripe with in flight i/o. The bypass_count will be reset when the
4009 * head of the hold_list has changed, i.e. the head was promoted to the
4012 static struct stripe_head *__get_priority_stripe(raid5_conf_t *conf)
4014 struct stripe_head *sh;
4016 pr_debug("%s: handle: %s hold: %s full_writes: %d bypass_count: %d\n",
4018 list_empty(&conf->handle_list) ? "empty" : "busy",
4019 list_empty(&conf->hold_list) ? "empty" : "busy",
4020 atomic_read(&conf->pending_full_writes), conf->bypass_count);
4022 if (!list_empty(&conf->handle_list)) {
4023 sh = list_entry(conf->handle_list.next, typeof(*sh), lru);
4025 if (list_empty(&conf->hold_list))
4026 conf->bypass_count = 0;
4027 else if (!test_bit(STRIPE_IO_STARTED, &sh->state)) {
4028 if (conf->hold_list.next == conf->last_hold)
4029 conf->bypass_count++;
4031 conf->last_hold = conf->hold_list.next;
4032 conf->bypass_count -= conf->bypass_threshold;
4033 if (conf->bypass_count < 0)
4034 conf->bypass_count = 0;
4037 } else if (!list_empty(&conf->hold_list) &&
4038 ((conf->bypass_threshold &&
4039 conf->bypass_count > conf->bypass_threshold) ||
4040 atomic_read(&conf->pending_full_writes) == 0)) {
4041 sh = list_entry(conf->hold_list.next,
4043 conf->bypass_count -= conf->bypass_threshold;
4044 if (conf->bypass_count < 0)
4045 conf->bypass_count = 0;
4049 list_del_init(&sh->lru);
4050 atomic_inc(&sh->count);
4051 BUG_ON(atomic_read(&sh->count) != 1);
4055 static int make_request(struct request_queue *q, struct bio * bi)
4057 mddev_t *mddev = q->queuedata;
4058 raid5_conf_t *conf = mddev_to_conf(mddev);
4060 sector_t new_sector;
4061 sector_t logical_sector, last_sector;
4062 struct stripe_head *sh;
4063 const int rw = bio_data_dir(bi);
4066 if (unlikely(bio_barrier(bi))) {
4067 bio_endio(bi, -EOPNOTSUPP);
4071 md_write_start(mddev, bi);
4073 cpu = part_stat_lock();
4074 part_stat_inc(cpu, &mddev->gendisk->part0, ios[rw]);
4075 part_stat_add(cpu, &mddev->gendisk->part0, sectors[rw],
4080 mddev->reshape_position == MaxSector &&
4081 chunk_aligned_read(q,bi))
4084 logical_sector = bi->bi_sector & ~((sector_t)STRIPE_SECTORS-1);
4085 last_sector = bi->bi_sector + (bi->bi_size>>9);
4087 bi->bi_phys_segments = 1; /* over-loaded to count active stripes */
4089 for (;logical_sector < last_sector; logical_sector += STRIPE_SECTORS) {
4091 int disks, data_disks;
4096 disks = conf->raid_disks;
4097 prepare_to_wait(&conf->wait_for_overlap, &w, TASK_UNINTERRUPTIBLE);
4098 if (unlikely(conf->reshape_progress != MaxSector)) {
4099 /* spinlock is needed as reshape_progress may be
4100 * 64bit on a 32bit platform, and so it might be
4101 * possible to see a half-updated value
4102 * Ofcourse reshape_progress could change after
4103 * the lock is dropped, so once we get a reference
4104 * to the stripe that we think it is, we will have
4107 spin_lock_irq(&conf->device_lock);
4108 if (mddev->delta_disks < 0
4109 ? logical_sector < conf->reshape_progress
4110 : logical_sector >= conf->reshape_progress) {
4111 disks = conf->previous_raid_disks;
4114 if (mddev->delta_disks < 0
4115 ? logical_sector < conf->reshape_safe
4116 : logical_sector >= conf->reshape_safe) {
4117 spin_unlock_irq(&conf->device_lock);
4122 spin_unlock_irq(&conf->device_lock);
4124 data_disks = disks - conf->max_degraded;
4126 new_sector = raid5_compute_sector(conf, logical_sector,
4129 pr_debug("raid5: make_request, sector %llu logical %llu\n",
4130 (unsigned long long)new_sector,
4131 (unsigned long long)logical_sector);
4133 sh = get_active_stripe(conf, new_sector, previous,
4134 (bi->bi_rw&RWA_MASK));
4136 if (unlikely(previous)) {
4137 /* expansion might have moved on while waiting for a
4138 * stripe, so we must do the range check again.
4139 * Expansion could still move past after this
4140 * test, but as we are holding a reference to
4141 * 'sh', we know that if that happens,
4142 * STRIPE_EXPANDING will get set and the expansion
4143 * won't proceed until we finish with the stripe.
4146 spin_lock_irq(&conf->device_lock);
4147 if (mddev->delta_disks < 0
4148 ? logical_sector >= conf->reshape_progress
4149 : logical_sector < conf->reshape_progress)
4150 /* mismatch, need to try again */
4152 spin_unlock_irq(&conf->device_lock);
4158 /* FIXME what if we get a false positive because these
4159 * are being updated.
4161 if (logical_sector >= mddev->suspend_lo &&
4162 logical_sector < mddev->suspend_hi) {
4168 if (test_bit(STRIPE_EXPANDING, &sh->state) ||
4169 !add_stripe_bio(sh, bi, dd_idx, (bi->bi_rw&RW_MASK))) {
4170 /* Stripe is busy expanding or
4171 * add failed due to overlap. Flush everything
4174 raid5_unplug_device(mddev->queue);
4179 finish_wait(&conf->wait_for_overlap, &w);
4180 set_bit(STRIPE_HANDLE, &sh->state);
4181 clear_bit(STRIPE_DELAYED, &sh->state);
4184 /* cannot get stripe for read-ahead, just give-up */
4185 clear_bit(BIO_UPTODATE, &bi->bi_flags);
4186 finish_wait(&conf->wait_for_overlap, &w);
4191 spin_lock_irq(&conf->device_lock);
4192 remaining = raid5_dec_bi_phys_segments(bi);
4193 spin_unlock_irq(&conf->device_lock);
4194 if (remaining == 0) {
4197 md_write_end(mddev);
4204 static sector_t raid5_size(mddev_t *mddev, sector_t sectors, int raid_disks);
4206 static sector_t reshape_request(mddev_t *mddev, sector_t sector_nr, int *skipped)
4208 /* reshaping is quite different to recovery/resync so it is
4209 * handled quite separately ... here.
4211 * On each call to sync_request, we gather one chunk worth of
4212 * destination stripes and flag them as expanding.
4213 * Then we find all the source stripes and request reads.
4214 * As the reads complete, handle_stripe will copy the data
4215 * into the destination stripe and release that stripe.
4217 raid5_conf_t *conf = (raid5_conf_t *) mddev->private;
4218 struct stripe_head *sh;
4219 sector_t first_sector, last_sector;
4220 int raid_disks = conf->previous_raid_disks;
4221 int data_disks = raid_disks - conf->max_degraded;
4222 int new_data_disks = conf->raid_disks - conf->max_degraded;
4225 sector_t writepos, readpos, safepos;
4226 sector_t stripe_addr;
4227 int reshape_sectors;
4228 struct list_head stripes;
4230 if (sector_nr == 0) {
4231 /* If restarting in the middle, skip the initial sectors */
4232 if (mddev->delta_disks < 0 &&
4233 conf->reshape_progress < raid5_size(mddev, 0, 0)) {
4234 sector_nr = raid5_size(mddev, 0, 0)
4235 - conf->reshape_progress;
4236 } else if (mddev->delta_disks > 0 &&
4237 conf->reshape_progress > 0)
4238 sector_nr = conf->reshape_progress;
4239 sector_div(sector_nr, new_data_disks);
4246 /* We need to process a full chunk at a time.
4247 * If old and new chunk sizes differ, we need to process the
4250 if (mddev->new_chunk > mddev->chunk_size)
4251 reshape_sectors = mddev->new_chunk / 512;
4253 reshape_sectors = mddev->chunk_size / 512;
4255 /* we update the metadata when there is more than 3Meg
4256 * in the block range (that is rather arbitrary, should
4257 * probably be time based) or when the data about to be
4258 * copied would over-write the source of the data at
4259 * the front of the range.
4260 * i.e. one new_stripe along from reshape_progress new_maps
4261 * to after where reshape_safe old_maps to
4263 writepos = conf->reshape_progress;
4264 sector_div(writepos, new_data_disks);
4265 readpos = conf->reshape_progress;
4266 sector_div(readpos, data_disks);
4267 safepos = conf->reshape_safe;
4268 sector_div(safepos, data_disks);
4269 if (mddev->delta_disks < 0) {
4270 writepos -= reshape_sectors;
4271 readpos += reshape_sectors;
4272 safepos += reshape_sectors;
4274 writepos += reshape_sectors;
4275 readpos -= reshape_sectors;
4276 safepos -= reshape_sectors;
4279 /* 'writepos' is the most advanced device address we might write.
4280 * 'readpos' is the least advanced device address we might read.
4281 * 'safepos' is the least address recorded in the metadata as having
4283 * If 'readpos' is behind 'writepos', then there is no way that we can
4284 * ensure safety in the face of a crash - that must be done by userspace
4285 * making a backup of the data. So in that case there is no particular
4286 * rush to update metadata.
4287 * Otherwise if 'safepos' is behind 'writepos', then we really need to
4288 * update the metadata to advance 'safepos' to match 'readpos' so that
4289 * we can be safe in the event of a crash.
4290 * So we insist on updating metadata if safepos is behind writepos and
4291 * readpos is beyond writepos.
4292 * In any case, update the metadata every 10 seconds.
4293 * Maybe that number should be configurable, but I'm not sure it is
4294 * worth it.... maybe it could be a multiple of safemode_delay???
4296 if ((mddev->delta_disks < 0
4297 ? (safepos > writepos && readpos < writepos)
4298 : (safepos < writepos && readpos > writepos)) ||
4299 time_after(jiffies, conf->reshape_checkpoint + 10*HZ)) {
4300 /* Cannot proceed until we've updated the superblock... */
4301 wait_event(conf->wait_for_overlap,
4302 atomic_read(&conf->reshape_stripes)==0);
4303 mddev->reshape_position = conf->reshape_progress;
4304 conf->reshape_checkpoint = jiffies;
4305 set_bit(MD_CHANGE_DEVS, &mddev->flags);
4306 md_wakeup_thread(mddev->thread);
4307 wait_event(mddev->sb_wait, mddev->flags == 0 ||
4308 kthread_should_stop());
4309 spin_lock_irq(&conf->device_lock);
4310 conf->reshape_safe = mddev->reshape_position;
4311 spin_unlock_irq(&conf->device_lock);
4312 wake_up(&conf->wait_for_overlap);
4315 if (mddev->delta_disks < 0) {
4316 BUG_ON(conf->reshape_progress == 0);
4317 stripe_addr = writepos;
4318 BUG_ON((mddev->dev_sectors &
4319 ~((sector_t)reshape_sectors - 1))
4320 - reshape_sectors - stripe_addr
4323 BUG_ON(writepos != sector_nr + reshape_sectors);
4324 stripe_addr = sector_nr;
4326 INIT_LIST_HEAD(&stripes);
4327 for (i = 0; i < reshape_sectors; i += STRIPE_SECTORS) {
4330 sh = get_active_stripe(conf, stripe_addr+i, 0, 0);
4331 set_bit(STRIPE_EXPANDING, &sh->state);
4332 atomic_inc(&conf->reshape_stripes);
4333 /* If any of this stripe is beyond the end of the old
4334 * array, then we need to zero those blocks
4336 for (j=sh->disks; j--;) {
4338 if (j == sh->pd_idx)
4340 if (conf->level == 6 &&
4343 s = compute_blocknr(sh, j, 0);
4344 if (s < raid5_size(mddev, 0, 0)) {
4348 memset(page_address(sh->dev[j].page), 0, STRIPE_SIZE);
4349 set_bit(R5_Expanded, &sh->dev[j].flags);
4350 set_bit(R5_UPTODATE, &sh->dev[j].flags);
4353 set_bit(STRIPE_EXPAND_READY, &sh->state);
4354 set_bit(STRIPE_HANDLE, &sh->state);
4356 list_add(&sh->lru, &stripes);
4358 spin_lock_irq(&conf->device_lock);
4359 if (mddev->delta_disks < 0)
4360 conf->reshape_progress -= reshape_sectors * new_data_disks;
4362 conf->reshape_progress += reshape_sectors * new_data_disks;
4363 spin_unlock_irq(&conf->device_lock);
4364 /* Ok, those stripe are ready. We can start scheduling
4365 * reads on the source stripes.
4366 * The source stripes are determined by mapping the first and last
4367 * block on the destination stripes.
4370 raid5_compute_sector(conf, stripe_addr*(new_data_disks),
4373 raid5_compute_sector(conf, ((stripe_addr+conf->chunk_size/512)
4374 *(new_data_disks) - 1),
4376 if (last_sector >= mddev->dev_sectors)
4377 last_sector = mddev->dev_sectors - 1;
4378 while (first_sector <= last_sector) {
4379 sh = get_active_stripe(conf, first_sector, 1, 0);
4380 set_bit(STRIPE_EXPAND_SOURCE, &sh->state);
4381 set_bit(STRIPE_HANDLE, &sh->state);
4383 first_sector += STRIPE_SECTORS;
4385 /* Now that the sources are clearly marked, we can release
4386 * the destination stripes
4388 while (!list_empty(&stripes)) {
4389 sh = list_entry(stripes.next, struct stripe_head, lru);
4390 list_del_init(&sh->lru);
4393 /* If this takes us to the resync_max point where we have to pause,
4394 * then we need to write out the superblock.
4396 sector_nr += reshape_sectors;
4397 if (sector_nr >= mddev->resync_max) {
4398 /* Cannot proceed until we've updated the superblock... */
4399 wait_event(conf->wait_for_overlap,
4400 atomic_read(&conf->reshape_stripes) == 0);
4401 mddev->reshape_position = conf->reshape_progress;
4402 conf->reshape_checkpoint = jiffies;
4403 set_bit(MD_CHANGE_DEVS, &mddev->flags);
4404 md_wakeup_thread(mddev->thread);
4405 wait_event(mddev->sb_wait,
4406 !test_bit(MD_CHANGE_DEVS, &mddev->flags)
4407 || kthread_should_stop());
4408 spin_lock_irq(&conf->device_lock);
4409 conf->reshape_safe = mddev->reshape_position;
4410 spin_unlock_irq(&conf->device_lock);
4411 wake_up(&conf->wait_for_overlap);
4413 return reshape_sectors;
4416 /* FIXME go_faster isn't used */
4417 static inline sector_t sync_request(mddev_t *mddev, sector_t sector_nr, int *skipped, int go_faster)
4419 raid5_conf_t *conf = (raid5_conf_t *) mddev->private;
4420 struct stripe_head *sh;
4421 sector_t max_sector = mddev->dev_sectors;
4423 int still_degraded = 0;
4426 if (sector_nr >= max_sector) {
4427 /* just being told to finish up .. nothing much to do */
4428 unplug_slaves(mddev);
4430 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery)) {
4435 if (mddev->curr_resync < max_sector) /* aborted */
4436 bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
4438 else /* completed sync */
4440 bitmap_close_sync(mddev->bitmap);
4445 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
4446 return reshape_request(mddev, sector_nr, skipped);
4448 /* No need to check resync_max as we never do more than one
4449 * stripe, and as resync_max will always be on a chunk boundary,
4450 * if the check in md_do_sync didn't fire, there is no chance
4451 * of overstepping resync_max here
4454 /* if there is too many failed drives and we are trying
4455 * to resync, then assert that we are finished, because there is
4456 * nothing we can do.
4458 if (mddev->degraded >= conf->max_degraded &&
4459 test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
4460 sector_t rv = mddev->dev_sectors - sector_nr;
4464 if (!bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
4465 !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
4466 !conf->fullsync && sync_blocks >= STRIPE_SECTORS) {
4467 /* we can skip this block, and probably more */
4468 sync_blocks /= STRIPE_SECTORS;
4470 return sync_blocks * STRIPE_SECTORS; /* keep things rounded to whole stripes */
4474 bitmap_cond_end_sync(mddev->bitmap, sector_nr);
4476 sh = get_active_stripe(conf, sector_nr, 0, 1);
4478 sh = get_active_stripe(conf, sector_nr, 0, 0);
4479 /* make sure we don't swamp the stripe cache if someone else
4480 * is trying to get access
4482 schedule_timeout_uninterruptible(1);
4484 /* Need to check if array will still be degraded after recovery/resync
4485 * We don't need to check the 'failed' flag as when that gets set,
4488 for (i=0; i<mddev->raid_disks; i++)
4489 if (conf->disks[i].rdev == NULL)
4492 bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, still_degraded);
4494 spin_lock(&sh->lock);
4495 set_bit(STRIPE_SYNCING, &sh->state);
4496 clear_bit(STRIPE_INSYNC, &sh->state);
4497 spin_unlock(&sh->lock);
4499 /* wait for any blocked device to be handled */
4500 while (unlikely(!handle_stripe(sh)))
4504 return STRIPE_SECTORS;
4507 static int retry_aligned_read(raid5_conf_t *conf, struct bio *raid_bio)
4509 /* We may not be able to submit a whole bio at once as there
4510 * may not be enough stripe_heads available.
4511 * We cannot pre-allocate enough stripe_heads as we may need
4512 * more than exist in the cache (if we allow ever large chunks).
4513 * So we do one stripe head at a time and record in
4514 * ->bi_hw_segments how many have been done.
4516 * We *know* that this entire raid_bio is in one chunk, so
4517 * it will be only one 'dd_idx' and only need one call to raid5_compute_sector.
4519 struct stripe_head *sh;
4521 sector_t sector, logical_sector, last_sector;
4526 logical_sector = raid_bio->bi_sector & ~((sector_t)STRIPE_SECTORS-1);
4527 sector = raid5_compute_sector(conf, logical_sector,
4529 last_sector = raid_bio->bi_sector + (raid_bio->bi_size>>9);
4531 for (; logical_sector < last_sector;
4532 logical_sector += STRIPE_SECTORS,
4533 sector += STRIPE_SECTORS,
4536 if (scnt < raid5_bi_hw_segments(raid_bio))
4537 /* already done this stripe */
4540 sh = get_active_stripe(conf, sector, 0, 1);
4543 /* failed to get a stripe - must wait */
4544 raid5_set_bi_hw_segments(raid_bio, scnt);
4545 conf->retry_read_aligned = raid_bio;
4549 set_bit(R5_ReadError, &sh->dev[dd_idx].flags);
4550 if (!add_stripe_bio(sh, raid_bio, dd_idx, 0)) {
4552 raid5_set_bi_hw_segments(raid_bio, scnt);
4553 conf->retry_read_aligned = raid_bio;
4561 spin_lock_irq(&conf->device_lock);
4562 remaining = raid5_dec_bi_phys_segments(raid_bio);
4563 spin_unlock_irq(&conf->device_lock);
4565 bio_endio(raid_bio, 0);
4566 if (atomic_dec_and_test(&conf->active_aligned_reads))
4567 wake_up(&conf->wait_for_stripe);
4574 * This is our raid5 kernel thread.
4576 * We scan the hash table for stripes which can be handled now.
4577 * During the scan, completed stripes are saved for us by the interrupt
4578 * handler, so that they will not have to wait for our next wakeup.
4580 static void raid5d(mddev_t *mddev)
4582 struct stripe_head *sh;
4583 raid5_conf_t *conf = mddev_to_conf(mddev);
4586 pr_debug("+++ raid5d active\n");
4588 md_check_recovery(mddev);
4591 spin_lock_irq(&conf->device_lock);
4595 if (conf->seq_flush != conf->seq_write) {
4596 int seq = conf->seq_flush;
4597 spin_unlock_irq(&conf->device_lock);
4598 bitmap_unplug(mddev->bitmap);
4599 spin_lock_irq(&conf->device_lock);
4600 conf->seq_write = seq;
4601 activate_bit_delay(conf);
4604 while ((bio = remove_bio_from_retry(conf))) {
4606 spin_unlock_irq(&conf->device_lock);
4607 ok = retry_aligned_read(conf, bio);
4608 spin_lock_irq(&conf->device_lock);
4614 sh = __get_priority_stripe(conf);
4618 spin_unlock_irq(&conf->device_lock);
4624 spin_lock_irq(&conf->device_lock);
4626 pr_debug("%d stripes handled\n", handled);
4628 spin_unlock_irq(&conf->device_lock);
4630 async_tx_issue_pending_all();
4631 unplug_slaves(mddev);
4633 pr_debug("--- raid5d inactive\n");
4637 raid5_show_stripe_cache_size(mddev_t *mddev, char *page)
4639 raid5_conf_t *conf = mddev_to_conf(mddev);
4641 return sprintf(page, "%d\n", conf->max_nr_stripes);
4647 raid5_store_stripe_cache_size(mddev_t *mddev, const char *page, size_t len)
4649 raid5_conf_t *conf = mddev_to_conf(mddev);
4653 if (len >= PAGE_SIZE)
4658 if (strict_strtoul(page, 10, &new))
4660 if (new <= 16 || new > 32768)
4662 while (new < conf->max_nr_stripes) {
4663 if (drop_one_stripe(conf))
4664 conf->max_nr_stripes--;
4668 err = md_allow_write(mddev);
4671 while (new > conf->max_nr_stripes) {
4672 if (grow_one_stripe(conf))
4673 conf->max_nr_stripes++;
4679 static struct md_sysfs_entry
4680 raid5_stripecache_size = __ATTR(stripe_cache_size, S_IRUGO | S_IWUSR,
4681 raid5_show_stripe_cache_size,
4682 raid5_store_stripe_cache_size);
4685 raid5_show_preread_threshold(mddev_t *mddev, char *page)
4687 raid5_conf_t *conf = mddev_to_conf(mddev);
4689 return sprintf(page, "%d\n", conf->bypass_threshold);
4695 raid5_store_preread_threshold(mddev_t *mddev, const char *page, size_t len)
4697 raid5_conf_t *conf = mddev_to_conf(mddev);
4699 if (len >= PAGE_SIZE)
4704 if (strict_strtoul(page, 10, &new))
4706 if (new > conf->max_nr_stripes)
4708 conf->bypass_threshold = new;
4712 static struct md_sysfs_entry
4713 raid5_preread_bypass_threshold = __ATTR(preread_bypass_threshold,
4715 raid5_show_preread_threshold,
4716 raid5_store_preread_threshold);
4719 stripe_cache_active_show(mddev_t *mddev, char *page)
4721 raid5_conf_t *conf = mddev_to_conf(mddev);
4723 return sprintf(page, "%d\n", atomic_read(&conf->active_stripes));
4728 static struct md_sysfs_entry
4729 raid5_stripecache_active = __ATTR_RO(stripe_cache_active);
4731 static struct attribute *raid5_attrs[] = {
4732 &raid5_stripecache_size.attr,
4733 &raid5_stripecache_active.attr,
4734 &raid5_preread_bypass_threshold.attr,
4737 static struct attribute_group raid5_attrs_group = {
4739 .attrs = raid5_attrs,
4743 raid5_size(mddev_t *mddev, sector_t sectors, int raid_disks)
4745 raid5_conf_t *conf = mddev_to_conf(mddev);
4748 sectors = mddev->dev_sectors;
4750 /* size is defined by the smallest of previous and new size */
4751 if (conf->raid_disks < conf->previous_raid_disks)
4752 raid_disks = conf->raid_disks;
4754 raid_disks = conf->previous_raid_disks;
4757 sectors &= ~((sector_t)mddev->chunk_size/512 - 1);
4758 sectors &= ~((sector_t)mddev->new_chunk/512 - 1);
4759 return sectors * (raid_disks - conf->max_degraded);
4762 static void raid5_free_percpu(raid5_conf_t *conf)
4764 struct raid5_percpu *percpu;
4771 for_each_possible_cpu(cpu) {
4772 percpu = per_cpu_ptr(conf->percpu, cpu);
4773 safe_put_page(percpu->spare_page);
4774 kfree(percpu->scribble);
4776 #ifdef CONFIG_HOTPLUG_CPU
4777 unregister_cpu_notifier(&conf->cpu_notify);
4781 free_percpu(conf->percpu);
4784 static void free_conf(raid5_conf_t *conf)
4786 shrink_stripes(conf);
4787 raid5_free_percpu(conf);
4789 kfree(conf->stripe_hashtbl);
4793 #ifdef CONFIG_HOTPLUG_CPU
4794 static int raid456_cpu_notify(struct notifier_block *nfb, unsigned long action,
4797 raid5_conf_t *conf = container_of(nfb, raid5_conf_t, cpu_notify);
4798 long cpu = (long)hcpu;
4799 struct raid5_percpu *percpu = per_cpu_ptr(conf->percpu, cpu);
4802 case CPU_UP_PREPARE:
4803 case CPU_UP_PREPARE_FROZEN:
4804 if (conf->level == 6 && !percpu->spare_page)
4805 percpu->spare_page = alloc_page(GFP_KERNEL);
4806 if (!percpu->scribble)
4807 percpu->scribble = kmalloc(conf->scribble_len, GFP_KERNEL);
4809 if (!percpu->scribble ||
4810 (conf->level == 6 && !percpu->spare_page)) {
4811 safe_put_page(percpu->spare_page);
4812 kfree(percpu->scribble);
4813 pr_err("%s: failed memory allocation for cpu%ld\n",
4819 case CPU_DEAD_FROZEN:
4820 safe_put_page(percpu->spare_page);
4821 kfree(percpu->scribble);
4822 percpu->spare_page = NULL;
4823 percpu->scribble = NULL;
4832 static int raid5_alloc_percpu(raid5_conf_t *conf)
4835 struct page *spare_page;
4836 struct raid5_percpu *allcpus;
4840 allcpus = alloc_percpu(struct raid5_percpu);
4843 conf->percpu = allcpus;
4847 for_each_present_cpu(cpu) {
4848 if (conf->level == 6) {
4849 spare_page = alloc_page(GFP_KERNEL);
4854 per_cpu_ptr(conf->percpu, cpu)->spare_page = spare_page;
4856 scribble = kmalloc(scribble_len(conf->raid_disks), GFP_KERNEL);
4861 per_cpu_ptr(conf->percpu, cpu)->scribble = scribble;
4863 #ifdef CONFIG_HOTPLUG_CPU
4864 conf->cpu_notify.notifier_call = raid456_cpu_notify;
4865 conf->cpu_notify.priority = 0;
4867 err = register_cpu_notifier(&conf->cpu_notify);
4874 static raid5_conf_t *setup_conf(mddev_t *mddev)
4877 int raid_disk, memory;
4879 struct disk_info *disk;
4881 if (mddev->new_level != 5
4882 && mddev->new_level != 4
4883 && mddev->new_level != 6) {
4884 printk(KERN_ERR "raid5: %s: raid level not set to 4/5/6 (%d)\n",
4885 mdname(mddev), mddev->new_level);
4886 return ERR_PTR(-EIO);
4888 if ((mddev->new_level == 5
4889 && !algorithm_valid_raid5(mddev->new_layout)) ||
4890 (mddev->new_level == 6
4891 && !algorithm_valid_raid6(mddev->new_layout))) {
4892 printk(KERN_ERR "raid5: %s: layout %d not supported\n",
4893 mdname(mddev), mddev->new_layout);
4894 return ERR_PTR(-EIO);
4896 if (mddev->new_level == 6 && mddev->raid_disks < 4) {
4897 printk(KERN_ERR "raid6: not enough configured devices for %s (%d, minimum 4)\n",
4898 mdname(mddev), mddev->raid_disks);
4899 return ERR_PTR(-EINVAL);
4902 if (!mddev->new_chunk || mddev->new_chunk % PAGE_SIZE) {
4903 printk(KERN_ERR "raid5: invalid chunk size %d for %s\n",
4904 mddev->new_chunk, mdname(mddev));
4905 return ERR_PTR(-EINVAL);
4908 conf = kzalloc(sizeof(raid5_conf_t), GFP_KERNEL);
4912 conf->raid_disks = mddev->raid_disks;
4913 conf->scribble_len = scribble_len(conf->raid_disks);
4914 if (mddev->reshape_position == MaxSector)
4915 conf->previous_raid_disks = mddev->raid_disks;
4917 conf->previous_raid_disks = mddev->raid_disks - mddev->delta_disks;
4919 conf->disks = kzalloc(conf->raid_disks * sizeof(struct disk_info),
4924 conf->mddev = mddev;
4926 if ((conf->stripe_hashtbl = kzalloc(PAGE_SIZE, GFP_KERNEL)) == NULL)
4929 conf->level = mddev->new_level;
4930 if (raid5_alloc_percpu(conf) != 0)
4933 spin_lock_init(&conf->device_lock);
4934 init_waitqueue_head(&conf->wait_for_stripe);
4935 init_waitqueue_head(&conf->wait_for_overlap);
4936 INIT_LIST_HEAD(&conf->handle_list);
4937 INIT_LIST_HEAD(&conf->hold_list);
4938 INIT_LIST_HEAD(&conf->delayed_list);
4939 INIT_LIST_HEAD(&conf->bitmap_list);
4940 INIT_LIST_HEAD(&conf->inactive_list);
4941 atomic_set(&conf->active_stripes, 0);
4942 atomic_set(&conf->preread_active_stripes, 0);
4943 atomic_set(&conf->active_aligned_reads, 0);
4944 conf->bypass_threshold = BYPASS_THRESHOLD;
4946 pr_debug("raid5: run(%s) called.\n", mdname(mddev));
4948 list_for_each_entry(rdev, &mddev->disks, same_set) {
4949 raid_disk = rdev->raid_disk;
4950 if (raid_disk >= conf->raid_disks
4953 disk = conf->disks + raid_disk;
4957 if (test_bit(In_sync, &rdev->flags)) {
4958 char b[BDEVNAME_SIZE];
4959 printk(KERN_INFO "raid5: device %s operational as raid"
4960 " disk %d\n", bdevname(rdev->bdev,b),
4963 /* Cannot rely on bitmap to complete recovery */
4967 conf->chunk_size = mddev->new_chunk;
4968 if (conf->level == 6)
4969 conf->max_degraded = 2;
4971 conf->max_degraded = 1;
4972 conf->algorithm = mddev->new_layout;
4973 conf->max_nr_stripes = NR_STRIPES;
4974 conf->reshape_progress = mddev->reshape_position;
4975 if (conf->reshape_progress != MaxSector) {
4976 conf->prev_chunk = mddev->chunk_size;
4977 conf->prev_algo = mddev->layout;
4980 memory = conf->max_nr_stripes * (sizeof(struct stripe_head) +
4981 conf->raid_disks * ((sizeof(struct bio) + PAGE_SIZE))) / 1024;
4982 if (grow_stripes(conf, conf->max_nr_stripes)) {
4984 "raid5: couldn't allocate %dkB for buffers\n", memory);
4987 printk(KERN_INFO "raid5: allocated %dkB for %s\n",
4988 memory, mdname(mddev));
4990 conf->thread = md_register_thread(raid5d, mddev, "%s_raid5");
4991 if (!conf->thread) {
4993 "raid5: couldn't allocate thread for %s\n",
5003 return ERR_PTR(-EIO);
5005 return ERR_PTR(-ENOMEM);
5008 static int run(mddev_t *mddev)
5011 int working_disks = 0;
5014 if (mddev->reshape_position != MaxSector) {
5015 /* Check that we can continue the reshape.
5016 * Currently only disks can change, it must
5017 * increase, and we must be past the point where
5018 * a stripe over-writes itself
5020 sector_t here_new, here_old;
5022 int max_degraded = (mddev->level == 6 ? 2 : 1);
5024 if (mddev->new_level != mddev->level) {
5025 printk(KERN_ERR "raid5: %s: unsupported reshape "
5026 "required - aborting.\n",
5030 old_disks = mddev->raid_disks - mddev->delta_disks;
5031 /* reshape_position must be on a new-stripe boundary, and one
5032 * further up in new geometry must map after here in old
5035 here_new = mddev->reshape_position;
5036 if (sector_div(here_new, (mddev->new_chunk>>9)*
5037 (mddev->raid_disks - max_degraded))) {
5038 printk(KERN_ERR "raid5: reshape_position not "
5039 "on a stripe boundary\n");
5042 /* here_new is the stripe we will write to */
5043 here_old = mddev->reshape_position;
5044 sector_div(here_old, (mddev->chunk_size>>9)*
5045 (old_disks-max_degraded));
5046 /* here_old is the first stripe that we might need to read
5048 if (here_new >= here_old) {
5049 /* Reading from the same stripe as writing to - bad */
5050 printk(KERN_ERR "raid5: reshape_position too early for "
5051 "auto-recovery - aborting.\n");
5054 printk(KERN_INFO "raid5: reshape will continue\n");
5055 /* OK, we should be able to continue; */
5057 BUG_ON(mddev->level != mddev->new_level);
5058 BUG_ON(mddev->layout != mddev->new_layout);
5059 BUG_ON(mddev->chunk_size != mddev->new_chunk);
5060 BUG_ON(mddev->delta_disks != 0);
5063 if (mddev->private == NULL)
5064 conf = setup_conf(mddev);
5066 conf = mddev->private;
5069 return PTR_ERR(conf);
5071 mddev->thread = conf->thread;
5072 conf->thread = NULL;
5073 mddev->private = conf;
5076 * 0 for a fully functional array, 1 or 2 for a degraded array.
5078 list_for_each_entry(rdev, &mddev->disks, same_set)
5079 if (rdev->raid_disk >= 0 &&
5080 test_bit(In_sync, &rdev->flags))
5083 mddev->degraded = conf->raid_disks - working_disks;
5085 if (mddev->degraded > conf->max_degraded) {
5086 printk(KERN_ERR "raid5: not enough operational devices for %s"
5087 " (%d/%d failed)\n",
5088 mdname(mddev), mddev->degraded, conf->raid_disks);
5092 /* device size must be a multiple of chunk size */
5093 mddev->dev_sectors &= ~(mddev->chunk_size / 512 - 1);
5094 mddev->resync_max_sectors = mddev->dev_sectors;
5096 if (mddev->degraded > 0 &&
5097 mddev->recovery_cp != MaxSector) {
5098 if (mddev->ok_start_degraded)
5100 "raid5: starting dirty degraded array: %s"
5101 "- data corruption possible.\n",
5105 "raid5: cannot start dirty degraded array for %s\n",
5111 if (mddev->degraded == 0)
5112 printk("raid5: raid level %d set %s active with %d out of %d"
5113 " devices, algorithm %d\n", conf->level, mdname(mddev),
5114 mddev->raid_disks-mddev->degraded, mddev->raid_disks,
5117 printk(KERN_ALERT "raid5: raid level %d set %s active with %d"
5118 " out of %d devices, algorithm %d\n", conf->level,
5119 mdname(mddev), mddev->raid_disks - mddev->degraded,
5120 mddev->raid_disks, mddev->new_layout);
5122 print_raid5_conf(conf);
5124 if (conf->reshape_progress != MaxSector) {
5125 printk("...ok start reshape thread\n");
5126 conf->reshape_safe = conf->reshape_progress;
5127 atomic_set(&conf->reshape_stripes, 0);
5128 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
5129 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
5130 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
5131 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
5132 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
5136 /* read-ahead size must cover two whole stripes, which is
5137 * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
5140 int data_disks = conf->previous_raid_disks - conf->max_degraded;
5141 int stripe = data_disks *
5142 (mddev->chunk_size / PAGE_SIZE);
5143 if (mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
5144 mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
5147 /* Ok, everything is just fine now */
5148 if (sysfs_create_group(&mddev->kobj, &raid5_attrs_group))
5150 "raid5: failed to create sysfs attributes for %s\n",
5153 mddev->queue->queue_lock = &conf->device_lock;
5155 mddev->queue->unplug_fn = raid5_unplug_device;
5156 mddev->queue->backing_dev_info.congested_data = mddev;
5157 mddev->queue->backing_dev_info.congested_fn = raid5_congested;
5159 md_set_array_sectors(mddev, raid5_size(mddev, 0, 0));
5161 blk_queue_merge_bvec(mddev->queue, raid5_mergeable_bvec);
5165 md_unregister_thread(mddev->thread);
5166 mddev->thread = NULL;
5168 print_raid5_conf(conf);
5171 mddev->private = NULL;
5172 printk(KERN_ALERT "raid5: failed to run raid set %s\n", mdname(mddev));
5178 static int stop(mddev_t *mddev)
5180 raid5_conf_t *conf = (raid5_conf_t *) mddev->private;
5182 md_unregister_thread(mddev->thread);
5183 mddev->thread = NULL;
5184 mddev->queue->backing_dev_info.congested_fn = NULL;
5185 blk_sync_queue(mddev->queue); /* the unplug fn references 'conf'*/
5186 sysfs_remove_group(&mddev->kobj, &raid5_attrs_group);
5188 mddev->private = NULL;
5193 static void print_sh(struct seq_file *seq, struct stripe_head *sh)
5197 seq_printf(seq, "sh %llu, pd_idx %d, state %ld.\n",
5198 (unsigned long long)sh->sector, sh->pd_idx, sh->state);
5199 seq_printf(seq, "sh %llu, count %d.\n",
5200 (unsigned long long)sh->sector, atomic_read(&sh->count));
5201 seq_printf(seq, "sh %llu, ", (unsigned long long)sh->sector);
5202 for (i = 0; i < sh->disks; i++) {
5203 seq_printf(seq, "(cache%d: %p %ld) ",
5204 i, sh->dev[i].page, sh->dev[i].flags);
5206 seq_printf(seq, "\n");
5209 static void printall(struct seq_file *seq, raid5_conf_t *conf)
5211 struct stripe_head *sh;
5212 struct hlist_node *hn;
5215 spin_lock_irq(&conf->device_lock);
5216 for (i = 0; i < NR_HASH; i++) {
5217 hlist_for_each_entry(sh, hn, &conf->stripe_hashtbl[i], hash) {
5218 if (sh->raid_conf != conf)
5223 spin_unlock_irq(&conf->device_lock);
5227 static void status(struct seq_file *seq, mddev_t *mddev)
5229 raid5_conf_t *conf = (raid5_conf_t *) mddev->private;
5232 seq_printf (seq, " level %d, %dk chunk, algorithm %d", mddev->level, mddev->chunk_size >> 10, mddev->layout);
5233 seq_printf (seq, " [%d/%d] [", conf->raid_disks, conf->raid_disks - mddev->degraded);
5234 for (i = 0; i < conf->raid_disks; i++)
5235 seq_printf (seq, "%s",
5236 conf->disks[i].rdev &&
5237 test_bit(In_sync, &conf->disks[i].rdev->flags) ? "U" : "_");
5238 seq_printf (seq, "]");
5240 seq_printf (seq, "\n");
5241 printall(seq, conf);
5245 static void print_raid5_conf (raid5_conf_t *conf)
5248 struct disk_info *tmp;
5250 printk("RAID5 conf printout:\n");
5252 printk("(conf==NULL)\n");
5255 printk(" --- rd:%d wd:%d\n", conf->raid_disks,
5256 conf->raid_disks - conf->mddev->degraded);
5258 for (i = 0; i < conf->raid_disks; i++) {
5259 char b[BDEVNAME_SIZE];
5260 tmp = conf->disks + i;
5262 printk(" disk %d, o:%d, dev:%s\n",
5263 i, !test_bit(Faulty, &tmp->rdev->flags),
5264 bdevname(tmp->rdev->bdev,b));
5268 static int raid5_spare_active(mddev_t *mddev)
5271 raid5_conf_t *conf = mddev->private;
5272 struct disk_info *tmp;
5274 for (i = 0; i < conf->raid_disks; i++) {
5275 tmp = conf->disks + i;
5277 && !test_bit(Faulty, &tmp->rdev->flags)
5278 && !test_and_set_bit(In_sync, &tmp->rdev->flags)) {
5279 unsigned long flags;
5280 spin_lock_irqsave(&conf->device_lock, flags);
5282 spin_unlock_irqrestore(&conf->device_lock, flags);
5285 print_raid5_conf(conf);
5289 static int raid5_remove_disk(mddev_t *mddev, int number)
5291 raid5_conf_t *conf = mddev->private;
5294 struct disk_info *p = conf->disks + number;
5296 print_raid5_conf(conf);
5299 if (number >= conf->raid_disks &&
5300 conf->reshape_progress == MaxSector)
5301 clear_bit(In_sync, &rdev->flags);
5303 if (test_bit(In_sync, &rdev->flags) ||
5304 atomic_read(&rdev->nr_pending)) {
5308 /* Only remove non-faulty devices if recovery
5311 if (!test_bit(Faulty, &rdev->flags) &&
5312 mddev->degraded <= conf->max_degraded &&
5313 number < conf->raid_disks) {
5319 if (atomic_read(&rdev->nr_pending)) {
5320 /* lost the race, try later */
5327 print_raid5_conf(conf);
5331 static int raid5_add_disk(mddev_t *mddev, mdk_rdev_t *rdev)
5333 raid5_conf_t *conf = mddev->private;
5336 struct disk_info *p;
5338 int last = conf->raid_disks - 1;
5340 if (mddev->degraded > conf->max_degraded)
5341 /* no point adding a device */
5344 if (rdev->raid_disk >= 0)
5345 first = last = rdev->raid_disk;
5348 * find the disk ... but prefer rdev->saved_raid_disk
5351 if (rdev->saved_raid_disk >= 0 &&
5352 rdev->saved_raid_disk >= first &&
5353 conf->disks[rdev->saved_raid_disk].rdev == NULL)
5354 disk = rdev->saved_raid_disk;
5357 for ( ; disk <= last ; disk++)
5358 if ((p=conf->disks + disk)->rdev == NULL) {
5359 clear_bit(In_sync, &rdev->flags);
5360 rdev->raid_disk = disk;
5362 if (rdev->saved_raid_disk != disk)
5364 rcu_assign_pointer(p->rdev, rdev);
5367 print_raid5_conf(conf);
5371 static int raid5_resize(mddev_t *mddev, sector_t sectors)
5373 /* no resync is happening, and there is enough space
5374 * on all devices, so we can resize.
5375 * We need to make sure resync covers any new space.
5376 * If the array is shrinking we should possibly wait until
5377 * any io in the removed space completes, but it hardly seems
5380 sectors &= ~((sector_t)mddev->chunk_size/512 - 1);
5381 md_set_array_sectors(mddev, raid5_size(mddev, sectors,
5382 mddev->raid_disks));
5383 if (mddev->array_sectors >
5384 raid5_size(mddev, sectors, mddev->raid_disks))
5386 set_capacity(mddev->gendisk, mddev->array_sectors);
5388 if (sectors > mddev->dev_sectors && mddev->recovery_cp == MaxSector) {
5389 mddev->recovery_cp = mddev->dev_sectors;
5390 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
5392 mddev->dev_sectors = sectors;
5393 mddev->resync_max_sectors = sectors;
5397 static int raid5_check_reshape(mddev_t *mddev)
5399 raid5_conf_t *conf = mddev_to_conf(mddev);
5401 if (mddev->delta_disks == 0 &&
5402 mddev->new_layout == mddev->layout &&
5403 mddev->new_chunk == mddev->chunk_size)
5404 return -EINVAL; /* nothing to do */
5406 /* Cannot grow a bitmap yet */
5408 if (mddev->degraded > conf->max_degraded)
5410 if (mddev->delta_disks < 0) {
5411 /* We might be able to shrink, but the devices must
5412 * be made bigger first.
5413 * For raid6, 4 is the minimum size.
5414 * Otherwise 2 is the minimum
5417 if (mddev->level == 6)
5419 if (mddev->raid_disks + mddev->delta_disks < min)
5423 /* Can only proceed if there are plenty of stripe_heads.
5424 * We need a minimum of one full stripe,, and for sensible progress
5425 * it is best to have about 4 times that.
5426 * If we require 4 times, then the default 256 4K stripe_heads will
5427 * allow for chunk sizes up to 256K, which is probably OK.
5428 * If the chunk size is greater, user-space should request more
5429 * stripe_heads first.
5431 if ((mddev->chunk_size / STRIPE_SIZE) * 4 > conf->max_nr_stripes ||
5432 (mddev->new_chunk / STRIPE_SIZE) * 4 > conf->max_nr_stripes) {
5433 printk(KERN_WARNING "raid5: reshape: not enough stripes. Needed %lu\n",
5434 (max(mddev->chunk_size, mddev->new_chunk)
5439 return resize_stripes(conf, conf->raid_disks + mddev->delta_disks);
5442 static int raid5_start_reshape(mddev_t *mddev)
5444 raid5_conf_t *conf = mddev_to_conf(mddev);
5447 int added_devices = 0;
5448 unsigned long flags;
5450 if (test_bit(MD_RECOVERY_RUNNING, &mddev->recovery))
5453 list_for_each_entry(rdev, &mddev->disks, same_set)
5454 if (rdev->raid_disk < 0 &&
5455 !test_bit(Faulty, &rdev->flags))
5458 if (spares - mddev->degraded < mddev->delta_disks - conf->max_degraded)
5459 /* Not enough devices even to make a degraded array
5464 /* Refuse to reduce size of the array. Any reductions in
5465 * array size must be through explicit setting of array_size
5468 if (raid5_size(mddev, 0, conf->raid_disks + mddev->delta_disks)
5469 < mddev->array_sectors) {
5470 printk(KERN_ERR "md: %s: array size must be reduced "
5471 "before number of disks\n", mdname(mddev));
5475 atomic_set(&conf->reshape_stripes, 0);
5476 spin_lock_irq(&conf->device_lock);
5477 conf->previous_raid_disks = conf->raid_disks;
5478 conf->raid_disks += mddev->delta_disks;
5479 conf->prev_chunk = conf->chunk_size;
5480 conf->chunk_size = mddev->new_chunk;
5481 conf->prev_algo = conf->algorithm;
5482 conf->algorithm = mddev->new_layout;
5483 if (mddev->delta_disks < 0)
5484 conf->reshape_progress = raid5_size(mddev, 0, 0);
5486 conf->reshape_progress = 0;
5487 conf->reshape_safe = conf->reshape_progress;
5489 spin_unlock_irq(&conf->device_lock);
5491 /* Add some new drives, as many as will fit.
5492 * We know there are enough to make the newly sized array work.
5494 list_for_each_entry(rdev, &mddev->disks, same_set)
5495 if (rdev->raid_disk < 0 &&
5496 !test_bit(Faulty, &rdev->flags)) {
5497 if (raid5_add_disk(mddev, rdev) == 0) {
5499 set_bit(In_sync, &rdev->flags);
5501 rdev->recovery_offset = 0;
5502 sprintf(nm, "rd%d", rdev->raid_disk);
5503 if (sysfs_create_link(&mddev->kobj,
5506 "raid5: failed to create "
5507 " link %s for %s\n",
5513 if (mddev->delta_disks > 0) {
5514 spin_lock_irqsave(&conf->device_lock, flags);
5515 mddev->degraded = (conf->raid_disks - conf->previous_raid_disks)
5517 spin_unlock_irqrestore(&conf->device_lock, flags);
5519 mddev->raid_disks = conf->raid_disks;
5520 mddev->reshape_position = 0;
5521 set_bit(MD_CHANGE_DEVS, &mddev->flags);
5523 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
5524 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
5525 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
5526 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
5527 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
5529 if (!mddev->sync_thread) {
5530 mddev->recovery = 0;
5531 spin_lock_irq(&conf->device_lock);
5532 mddev->raid_disks = conf->raid_disks = conf->previous_raid_disks;
5533 conf->reshape_progress = MaxSector;
5534 spin_unlock_irq(&conf->device_lock);
5537 conf->reshape_checkpoint = jiffies;
5538 md_wakeup_thread(mddev->sync_thread);
5539 md_new_event(mddev);
5543 /* This is called from the reshape thread and should make any
5544 * changes needed in 'conf'
5546 static void end_reshape(raid5_conf_t *conf)
5549 if (!test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery)) {
5551 spin_lock_irq(&conf->device_lock);
5552 conf->previous_raid_disks = conf->raid_disks;
5553 conf->reshape_progress = MaxSector;
5554 spin_unlock_irq(&conf->device_lock);
5555 wake_up(&conf->wait_for_overlap);
5557 /* read-ahead size must cover two whole stripes, which is
5558 * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
5561 int data_disks = conf->raid_disks - conf->max_degraded;
5562 int stripe = data_disks * (conf->chunk_size
5564 if (conf->mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
5565 conf->mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
5570 /* This is called from the raid5d thread with mddev_lock held.
5571 * It makes config changes to the device.
5573 static void raid5_finish_reshape(mddev_t *mddev)
5575 struct block_device *bdev;
5576 raid5_conf_t *conf = mddev_to_conf(mddev);
5578 if (!test_bit(MD_RECOVERY_INTR, &mddev->recovery)) {
5580 if (mddev->delta_disks > 0) {
5581 md_set_array_sectors(mddev, raid5_size(mddev, 0, 0));
5582 set_capacity(mddev->gendisk, mddev->array_sectors);
5585 bdev = bdget_disk(mddev->gendisk, 0);
5587 mutex_lock(&bdev->bd_inode->i_mutex);
5588 i_size_write(bdev->bd_inode,
5589 (loff_t)mddev->array_sectors << 9);
5590 mutex_unlock(&bdev->bd_inode->i_mutex);
5595 mddev->degraded = conf->raid_disks;
5596 for (d = 0; d < conf->raid_disks ; d++)
5597 if (conf->disks[d].rdev &&
5599 &conf->disks[d].rdev->flags))
5601 for (d = conf->raid_disks ;
5602 d < conf->raid_disks - mddev->delta_disks;
5604 raid5_remove_disk(mddev, d);
5606 mddev->layout = conf->algorithm;
5607 mddev->chunk_size = conf->chunk_size;
5608 mddev->reshape_position = MaxSector;
5609 mddev->delta_disks = 0;
5613 static void raid5_quiesce(mddev_t *mddev, int state)
5615 raid5_conf_t *conf = mddev_to_conf(mddev);
5618 case 2: /* resume for a suspend */
5619 wake_up(&conf->wait_for_overlap);
5622 case 1: /* stop all writes */
5623 spin_lock_irq(&conf->device_lock);
5625 wait_event_lock_irq(conf->wait_for_stripe,
5626 atomic_read(&conf->active_stripes) == 0 &&
5627 atomic_read(&conf->active_aligned_reads) == 0,
5628 conf->device_lock, /* nothing */);
5629 spin_unlock_irq(&conf->device_lock);
5632 case 0: /* re-enable writes */
5633 spin_lock_irq(&conf->device_lock);
5635 wake_up(&conf->wait_for_stripe);
5636 wake_up(&conf->wait_for_overlap);
5637 spin_unlock_irq(&conf->device_lock);
5643 static void *raid5_takeover_raid1(mddev_t *mddev)
5647 if (mddev->raid_disks != 2 ||
5648 mddev->degraded > 1)
5649 return ERR_PTR(-EINVAL);
5651 /* Should check if there are write-behind devices? */
5653 chunksect = 64*2; /* 64K by default */
5655 /* The array must be an exact multiple of chunksize */
5656 while (chunksect && (mddev->array_sectors & (chunksect-1)))
5659 if ((chunksect<<9) < STRIPE_SIZE)
5660 /* array size does not allow a suitable chunk size */
5661 return ERR_PTR(-EINVAL);
5663 mddev->new_level = 5;
5664 mddev->new_layout = ALGORITHM_LEFT_SYMMETRIC;
5665 mddev->new_chunk = chunksect << 9;
5667 return setup_conf(mddev);
5670 static void *raid5_takeover_raid6(mddev_t *mddev)
5674 switch (mddev->layout) {
5675 case ALGORITHM_LEFT_ASYMMETRIC_6:
5676 new_layout = ALGORITHM_LEFT_ASYMMETRIC;
5678 case ALGORITHM_RIGHT_ASYMMETRIC_6:
5679 new_layout = ALGORITHM_RIGHT_ASYMMETRIC;
5681 case ALGORITHM_LEFT_SYMMETRIC_6:
5682 new_layout = ALGORITHM_LEFT_SYMMETRIC;
5684 case ALGORITHM_RIGHT_SYMMETRIC_6:
5685 new_layout = ALGORITHM_RIGHT_SYMMETRIC;
5687 case ALGORITHM_PARITY_0_6:
5688 new_layout = ALGORITHM_PARITY_0;
5690 case ALGORITHM_PARITY_N:
5691 new_layout = ALGORITHM_PARITY_N;
5694 return ERR_PTR(-EINVAL);
5696 mddev->new_level = 5;
5697 mddev->new_layout = new_layout;
5698 mddev->delta_disks = -1;
5699 mddev->raid_disks -= 1;
5700 return setup_conf(mddev);
5704 static int raid5_reconfig(mddev_t *mddev, int new_layout, int new_chunk)
5706 /* For a 2-drive array, the layout and chunk size can be changed
5707 * immediately as not restriping is needed.
5708 * For larger arrays we record the new value - after validation
5709 * to be used by a reshape pass.
5711 raid5_conf_t *conf = mddev_to_conf(mddev);
5713 if (new_layout >= 0 && !algorithm_valid_raid5(new_layout))
5715 if (new_chunk > 0) {
5716 if (new_chunk & (new_chunk-1))
5717 /* not a power of 2 */
5719 if (new_chunk < PAGE_SIZE)
5721 if (mddev->array_sectors & ((new_chunk>>9)-1))
5722 /* not factor of array size */
5726 /* They look valid */
5728 if (mddev->raid_disks == 2) {
5730 if (new_layout >= 0) {
5731 conf->algorithm = new_layout;
5732 mddev->layout = mddev->new_layout = new_layout;
5734 if (new_chunk > 0) {
5735 conf->chunk_size = new_chunk;
5736 mddev->chunk_size = mddev->new_chunk = new_chunk;
5738 set_bit(MD_CHANGE_DEVS, &mddev->flags);
5739 md_wakeup_thread(mddev->thread);
5741 if (new_layout >= 0)
5742 mddev->new_layout = new_layout;
5744 mddev->new_chunk = new_chunk;
5749 static int raid6_reconfig(mddev_t *mddev, int new_layout, int new_chunk)
5751 if (new_layout >= 0 && !algorithm_valid_raid6(new_layout))
5753 if (new_chunk > 0) {
5754 if (new_chunk & (new_chunk-1))
5755 /* not a power of 2 */
5757 if (new_chunk < PAGE_SIZE)
5759 if (mddev->array_sectors & ((new_chunk>>9)-1))
5760 /* not factor of array size */
5764 /* They look valid */
5766 if (new_layout >= 0)
5767 mddev->new_layout = new_layout;
5769 mddev->new_chunk = new_chunk;
5774 static void *raid5_takeover(mddev_t *mddev)
5776 /* raid5 can take over:
5777 * raid0 - if all devices are the same - make it a raid4 layout
5778 * raid1 - if there are two drives. We need to know the chunk size
5779 * raid4 - trivial - just use a raid4 layout.
5780 * raid6 - Providing it is a *_6 layout
5782 * For now, just do raid1
5785 if (mddev->level == 1)
5786 return raid5_takeover_raid1(mddev);
5787 if (mddev->level == 4) {
5788 mddev->new_layout = ALGORITHM_PARITY_N;
5789 mddev->new_level = 5;
5790 return setup_conf(mddev);
5792 if (mddev->level == 6)
5793 return raid5_takeover_raid6(mddev);
5795 return ERR_PTR(-EINVAL);
5799 static struct mdk_personality raid5_personality;
5801 static void *raid6_takeover(mddev_t *mddev)
5803 /* Currently can only take over a raid5. We map the
5804 * personality to an equivalent raid6 personality
5805 * with the Q block at the end.
5809 if (mddev->pers != &raid5_personality)
5810 return ERR_PTR(-EINVAL);
5811 if (mddev->degraded > 1)
5812 return ERR_PTR(-EINVAL);
5813 if (mddev->raid_disks > 253)
5814 return ERR_PTR(-EINVAL);
5815 if (mddev->raid_disks < 3)
5816 return ERR_PTR(-EINVAL);
5818 switch (mddev->layout) {
5819 case ALGORITHM_LEFT_ASYMMETRIC:
5820 new_layout = ALGORITHM_LEFT_ASYMMETRIC_6;
5822 case ALGORITHM_RIGHT_ASYMMETRIC:
5823 new_layout = ALGORITHM_RIGHT_ASYMMETRIC_6;
5825 case ALGORITHM_LEFT_SYMMETRIC:
5826 new_layout = ALGORITHM_LEFT_SYMMETRIC_6;
5828 case ALGORITHM_RIGHT_SYMMETRIC:
5829 new_layout = ALGORITHM_RIGHT_SYMMETRIC_6;
5831 case ALGORITHM_PARITY_0:
5832 new_layout = ALGORITHM_PARITY_0_6;
5834 case ALGORITHM_PARITY_N:
5835 new_layout = ALGORITHM_PARITY_N;
5838 return ERR_PTR(-EINVAL);
5840 mddev->new_level = 6;
5841 mddev->new_layout = new_layout;
5842 mddev->delta_disks = 1;
5843 mddev->raid_disks += 1;
5844 return setup_conf(mddev);
5848 static struct mdk_personality raid6_personality =
5852 .owner = THIS_MODULE,
5853 .make_request = make_request,
5857 .error_handler = error,
5858 .hot_add_disk = raid5_add_disk,
5859 .hot_remove_disk= raid5_remove_disk,
5860 .spare_active = raid5_spare_active,
5861 .sync_request = sync_request,
5862 .resize = raid5_resize,
5864 .check_reshape = raid5_check_reshape,
5865 .start_reshape = raid5_start_reshape,
5866 .finish_reshape = raid5_finish_reshape,
5867 .quiesce = raid5_quiesce,
5868 .takeover = raid6_takeover,
5869 .reconfig = raid6_reconfig,
5871 static struct mdk_personality raid5_personality =
5875 .owner = THIS_MODULE,
5876 .make_request = make_request,
5880 .error_handler = error,
5881 .hot_add_disk = raid5_add_disk,
5882 .hot_remove_disk= raid5_remove_disk,
5883 .spare_active = raid5_spare_active,
5884 .sync_request = sync_request,
5885 .resize = raid5_resize,
5887 .check_reshape = raid5_check_reshape,
5888 .start_reshape = raid5_start_reshape,
5889 .finish_reshape = raid5_finish_reshape,
5890 .quiesce = raid5_quiesce,
5891 .takeover = raid5_takeover,
5892 .reconfig = raid5_reconfig,
5895 static struct mdk_personality raid4_personality =
5899 .owner = THIS_MODULE,
5900 .make_request = make_request,
5904 .error_handler = error,
5905 .hot_add_disk = raid5_add_disk,
5906 .hot_remove_disk= raid5_remove_disk,
5907 .spare_active = raid5_spare_active,
5908 .sync_request = sync_request,
5909 .resize = raid5_resize,
5911 .check_reshape = raid5_check_reshape,
5912 .start_reshape = raid5_start_reshape,
5913 .finish_reshape = raid5_finish_reshape,
5914 .quiesce = raid5_quiesce,
5917 static int __init raid5_init(void)
5919 register_md_personality(&raid6_personality);
5920 register_md_personality(&raid5_personality);
5921 register_md_personality(&raid4_personality);
5925 static void raid5_exit(void)
5927 unregister_md_personality(&raid6_personality);
5928 unregister_md_personality(&raid5_personality);
5929 unregister_md_personality(&raid4_personality);
5932 module_init(raid5_init);
5933 module_exit(raid5_exit);
5934 MODULE_LICENSE("GPL");
5935 MODULE_ALIAS("md-personality-4"); /* RAID5 */
5936 MODULE_ALIAS("md-raid5");
5937 MODULE_ALIAS("md-raid4");
5938 MODULE_ALIAS("md-level-5");
5939 MODULE_ALIAS("md-level-4");
5940 MODULE_ALIAS("md-personality-8"); /* RAID6 */
5941 MODULE_ALIAS("md-raid6");
5942 MODULE_ALIAS("md-level-6");
5944 /* This used to be two separate modules, they were: */
5945 MODULE_ALIAS("raid5");
5946 MODULE_ALIAS("raid6");