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");
1931 schedule_reconstruction(struct stripe_head *sh, struct stripe_head_state *s,
1932 int rcw, int expand)
1934 int i, pd_idx = sh->pd_idx, disks = sh->disks;
1935 raid5_conf_t *conf = sh->raid_conf;
1936 int level = conf->level;
1939 /* if we are not expanding this is a proper write request, and
1940 * there will be bios with new data to be drained into the
1944 sh->reconstruct_state = reconstruct_state_drain_run;
1945 set_bit(STRIPE_OP_BIODRAIN, &s->ops_request);
1947 sh->reconstruct_state = reconstruct_state_run;
1949 set_bit(STRIPE_OP_RECONSTRUCT, &s->ops_request);
1951 for (i = disks; i--; ) {
1952 struct r5dev *dev = &sh->dev[i];
1955 set_bit(R5_LOCKED, &dev->flags);
1956 set_bit(R5_Wantdrain, &dev->flags);
1958 clear_bit(R5_UPTODATE, &dev->flags);
1962 if (s->locked + conf->max_degraded == disks)
1963 if (!test_and_set_bit(STRIPE_FULL_WRITE, &sh->state))
1964 atomic_inc(&conf->pending_full_writes);
1967 BUG_ON(!(test_bit(R5_UPTODATE, &sh->dev[pd_idx].flags) ||
1968 test_bit(R5_Wantcompute, &sh->dev[pd_idx].flags)));
1970 sh->reconstruct_state = reconstruct_state_prexor_drain_run;
1971 set_bit(STRIPE_OP_PREXOR, &s->ops_request);
1972 set_bit(STRIPE_OP_BIODRAIN, &s->ops_request);
1973 set_bit(STRIPE_OP_RECONSTRUCT, &s->ops_request);
1975 for (i = disks; i--; ) {
1976 struct r5dev *dev = &sh->dev[i];
1981 (test_bit(R5_UPTODATE, &dev->flags) ||
1982 test_bit(R5_Wantcompute, &dev->flags))) {
1983 set_bit(R5_Wantdrain, &dev->flags);
1984 set_bit(R5_LOCKED, &dev->flags);
1985 clear_bit(R5_UPTODATE, &dev->flags);
1991 /* keep the parity disk(s) locked while asynchronous operations
1994 set_bit(R5_LOCKED, &sh->dev[pd_idx].flags);
1995 clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
1999 int qd_idx = sh->qd_idx;
2000 struct r5dev *dev = &sh->dev[qd_idx];
2002 set_bit(R5_LOCKED, &dev->flags);
2003 clear_bit(R5_UPTODATE, &dev->flags);
2007 pr_debug("%s: stripe %llu locked: %d ops_request: %lx\n",
2008 __func__, (unsigned long long)sh->sector,
2009 s->locked, s->ops_request);
2013 * Each stripe/dev can have one or more bion attached.
2014 * toread/towrite point to the first in a chain.
2015 * The bi_next chain must be in order.
2017 static int add_stripe_bio(struct stripe_head *sh, struct bio *bi, int dd_idx, int forwrite)
2020 raid5_conf_t *conf = sh->raid_conf;
2023 pr_debug("adding bh b#%llu to stripe s#%llu\n",
2024 (unsigned long long)bi->bi_sector,
2025 (unsigned long long)sh->sector);
2028 spin_lock(&sh->lock);
2029 spin_lock_irq(&conf->device_lock);
2031 bip = &sh->dev[dd_idx].towrite;
2032 if (*bip == NULL && sh->dev[dd_idx].written == NULL)
2035 bip = &sh->dev[dd_idx].toread;
2036 while (*bip && (*bip)->bi_sector < bi->bi_sector) {
2037 if ((*bip)->bi_sector + ((*bip)->bi_size >> 9) > bi->bi_sector)
2039 bip = & (*bip)->bi_next;
2041 if (*bip && (*bip)->bi_sector < bi->bi_sector + ((bi->bi_size)>>9))
2044 BUG_ON(*bip && bi->bi_next && (*bip) != bi->bi_next);
2048 bi->bi_phys_segments++;
2049 spin_unlock_irq(&conf->device_lock);
2050 spin_unlock(&sh->lock);
2052 pr_debug("added bi b#%llu to stripe s#%llu, disk %d.\n",
2053 (unsigned long long)bi->bi_sector,
2054 (unsigned long long)sh->sector, dd_idx);
2056 if (conf->mddev->bitmap && firstwrite) {
2057 bitmap_startwrite(conf->mddev->bitmap, sh->sector,
2059 sh->bm_seq = conf->seq_flush+1;
2060 set_bit(STRIPE_BIT_DELAY, &sh->state);
2064 /* check if page is covered */
2065 sector_t sector = sh->dev[dd_idx].sector;
2066 for (bi=sh->dev[dd_idx].towrite;
2067 sector < sh->dev[dd_idx].sector + STRIPE_SECTORS &&
2068 bi && bi->bi_sector <= sector;
2069 bi = r5_next_bio(bi, sh->dev[dd_idx].sector)) {
2070 if (bi->bi_sector + (bi->bi_size>>9) >= sector)
2071 sector = bi->bi_sector + (bi->bi_size>>9);
2073 if (sector >= sh->dev[dd_idx].sector + STRIPE_SECTORS)
2074 set_bit(R5_OVERWRITE, &sh->dev[dd_idx].flags);
2079 set_bit(R5_Overlap, &sh->dev[dd_idx].flags);
2080 spin_unlock_irq(&conf->device_lock);
2081 spin_unlock(&sh->lock);
2085 static void end_reshape(raid5_conf_t *conf);
2087 static void stripe_set_idx(sector_t stripe, raid5_conf_t *conf, int previous,
2088 struct stripe_head *sh)
2090 int sectors_per_chunk =
2091 previous ? (conf->prev_chunk >> 9)
2092 : (conf->chunk_size >> 9);
2094 int chunk_offset = sector_div(stripe, sectors_per_chunk);
2095 int disks = previous ? conf->previous_raid_disks : conf->raid_disks;
2097 raid5_compute_sector(conf,
2098 stripe * (disks - conf->max_degraded)
2099 *sectors_per_chunk + chunk_offset,
2105 handle_failed_stripe(raid5_conf_t *conf, struct stripe_head *sh,
2106 struct stripe_head_state *s, int disks,
2107 struct bio **return_bi)
2110 for (i = disks; i--; ) {
2114 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
2117 rdev = rcu_dereference(conf->disks[i].rdev);
2118 if (rdev && test_bit(In_sync, &rdev->flags))
2119 /* multiple read failures in one stripe */
2120 md_error(conf->mddev, rdev);
2123 spin_lock_irq(&conf->device_lock);
2124 /* fail all writes first */
2125 bi = sh->dev[i].towrite;
2126 sh->dev[i].towrite = NULL;
2132 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
2133 wake_up(&conf->wait_for_overlap);
2135 while (bi && bi->bi_sector <
2136 sh->dev[i].sector + STRIPE_SECTORS) {
2137 struct bio *nextbi = r5_next_bio(bi, sh->dev[i].sector);
2138 clear_bit(BIO_UPTODATE, &bi->bi_flags);
2139 if (!raid5_dec_bi_phys_segments(bi)) {
2140 md_write_end(conf->mddev);
2141 bi->bi_next = *return_bi;
2146 /* and fail all 'written' */
2147 bi = sh->dev[i].written;
2148 sh->dev[i].written = NULL;
2149 if (bi) bitmap_end = 1;
2150 while (bi && bi->bi_sector <
2151 sh->dev[i].sector + STRIPE_SECTORS) {
2152 struct bio *bi2 = r5_next_bio(bi, sh->dev[i].sector);
2153 clear_bit(BIO_UPTODATE, &bi->bi_flags);
2154 if (!raid5_dec_bi_phys_segments(bi)) {
2155 md_write_end(conf->mddev);
2156 bi->bi_next = *return_bi;
2162 /* fail any reads if this device is non-operational and
2163 * the data has not reached the cache yet.
2165 if (!test_bit(R5_Wantfill, &sh->dev[i].flags) &&
2166 (!test_bit(R5_Insync, &sh->dev[i].flags) ||
2167 test_bit(R5_ReadError, &sh->dev[i].flags))) {
2168 bi = sh->dev[i].toread;
2169 sh->dev[i].toread = NULL;
2170 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
2171 wake_up(&conf->wait_for_overlap);
2172 if (bi) s->to_read--;
2173 while (bi && bi->bi_sector <
2174 sh->dev[i].sector + STRIPE_SECTORS) {
2175 struct bio *nextbi =
2176 r5_next_bio(bi, sh->dev[i].sector);
2177 clear_bit(BIO_UPTODATE, &bi->bi_flags);
2178 if (!raid5_dec_bi_phys_segments(bi)) {
2179 bi->bi_next = *return_bi;
2185 spin_unlock_irq(&conf->device_lock);
2187 bitmap_endwrite(conf->mddev->bitmap, sh->sector,
2188 STRIPE_SECTORS, 0, 0);
2191 if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
2192 if (atomic_dec_and_test(&conf->pending_full_writes))
2193 md_wakeup_thread(conf->mddev->thread);
2196 /* fetch_block5 - checks the given member device to see if its data needs
2197 * to be read or computed to satisfy a request.
2199 * Returns 1 when no more member devices need to be checked, otherwise returns
2200 * 0 to tell the loop in handle_stripe_fill5 to continue
2202 static int fetch_block5(struct stripe_head *sh, struct stripe_head_state *s,
2203 int disk_idx, int disks)
2205 struct r5dev *dev = &sh->dev[disk_idx];
2206 struct r5dev *failed_dev = &sh->dev[s->failed_num];
2208 /* is the data in this block needed, and can we get it? */
2209 if (!test_bit(R5_LOCKED, &dev->flags) &&
2210 !test_bit(R5_UPTODATE, &dev->flags) &&
2212 (dev->towrite && !test_bit(R5_OVERWRITE, &dev->flags)) ||
2213 s->syncing || s->expanding ||
2215 (failed_dev->toread ||
2216 (failed_dev->towrite &&
2217 !test_bit(R5_OVERWRITE, &failed_dev->flags)))))) {
2218 /* We would like to get this block, possibly by computing it,
2219 * otherwise read it if the backing disk is insync
2221 if ((s->uptodate == disks - 1) &&
2222 (s->failed && disk_idx == s->failed_num)) {
2223 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
2224 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
2225 set_bit(R5_Wantcompute, &dev->flags);
2226 sh->ops.target = disk_idx;
2227 sh->ops.target2 = -1;
2229 /* Careful: from this point on 'uptodate' is in the eye
2230 * of raid_run_ops which services 'compute' operations
2231 * before writes. R5_Wantcompute flags a block that will
2232 * be R5_UPTODATE by the time it is needed for a
2233 * subsequent operation.
2236 return 1; /* uptodate + compute == disks */
2237 } else if (test_bit(R5_Insync, &dev->flags)) {
2238 set_bit(R5_LOCKED, &dev->flags);
2239 set_bit(R5_Wantread, &dev->flags);
2241 pr_debug("Reading block %d (sync=%d)\n", disk_idx,
2250 * handle_stripe_fill5 - read or compute data to satisfy pending requests.
2252 static void handle_stripe_fill5(struct stripe_head *sh,
2253 struct stripe_head_state *s, int disks)
2257 /* look for blocks to read/compute, skip this if a compute
2258 * is already in flight, or if the stripe contents are in the
2259 * midst of changing due to a write
2261 if (!test_bit(STRIPE_COMPUTE_RUN, &sh->state) && !sh->check_state &&
2262 !sh->reconstruct_state)
2263 for (i = disks; i--; )
2264 if (fetch_block5(sh, s, i, disks))
2266 set_bit(STRIPE_HANDLE, &sh->state);
2269 /* fetch_block6 - checks the given member device to see if its data needs
2270 * to be read or computed to satisfy a request.
2272 * Returns 1 when no more member devices need to be checked, otherwise returns
2273 * 0 to tell the loop in handle_stripe_fill6 to continue
2275 static int fetch_block6(struct stripe_head *sh, struct stripe_head_state *s,
2276 struct r6_state *r6s, int disk_idx, int disks)
2278 struct r5dev *dev = &sh->dev[disk_idx];
2279 struct r5dev *fdev[2] = { &sh->dev[r6s->failed_num[0]],
2280 &sh->dev[r6s->failed_num[1]] };
2282 if (!test_bit(R5_LOCKED, &dev->flags) &&
2283 !test_bit(R5_UPTODATE, &dev->flags) &&
2285 (dev->towrite && !test_bit(R5_OVERWRITE, &dev->flags)) ||
2286 s->syncing || s->expanding ||
2288 (fdev[0]->toread || s->to_write)) ||
2290 (fdev[1]->toread || s->to_write)))) {
2291 /* we would like to get this block, possibly by computing it,
2292 * otherwise read it if the backing disk is insync
2294 BUG_ON(test_bit(R5_Wantcompute, &dev->flags));
2295 BUG_ON(test_bit(R5_Wantread, &dev->flags));
2296 if ((s->uptodate == disks - 1) &&
2297 (s->failed && (disk_idx == r6s->failed_num[0] ||
2298 disk_idx == r6s->failed_num[1]))) {
2299 /* have disk failed, and we're requested to fetch it;
2302 pr_debug("Computing stripe %llu block %d\n",
2303 (unsigned long long)sh->sector, disk_idx);
2304 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
2305 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
2306 set_bit(R5_Wantcompute, &dev->flags);
2307 sh->ops.target = disk_idx;
2308 sh->ops.target2 = -1; /* no 2nd target */
2312 } else if (s->uptodate == disks-2 && s->failed >= 2) {
2313 /* Computing 2-failure is *very* expensive; only
2314 * do it if failed >= 2
2317 for (other = disks; other--; ) {
2318 if (other == disk_idx)
2320 if (!test_bit(R5_UPTODATE,
2321 &sh->dev[other].flags))
2325 pr_debug("Computing stripe %llu blocks %d,%d\n",
2326 (unsigned long long)sh->sector,
2328 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
2329 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
2330 set_bit(R5_Wantcompute, &sh->dev[disk_idx].flags);
2331 set_bit(R5_Wantcompute, &sh->dev[other].flags);
2332 sh->ops.target = disk_idx;
2333 sh->ops.target2 = other;
2337 } else if (test_bit(R5_Insync, &dev->flags)) {
2338 set_bit(R5_LOCKED, &dev->flags);
2339 set_bit(R5_Wantread, &dev->flags);
2341 pr_debug("Reading block %d (sync=%d)\n",
2342 disk_idx, s->syncing);
2350 * handle_stripe_fill6 - read or compute data to satisfy pending requests.
2352 static void handle_stripe_fill6(struct stripe_head *sh,
2353 struct stripe_head_state *s, struct r6_state *r6s,
2358 /* look for blocks to read/compute, skip this if a compute
2359 * is already in flight, or if the stripe contents are in the
2360 * midst of changing due to a write
2362 if (!test_bit(STRIPE_COMPUTE_RUN, &sh->state) && !sh->check_state &&
2363 !sh->reconstruct_state)
2364 for (i = disks; i--; )
2365 if (fetch_block6(sh, s, r6s, i, disks))
2367 set_bit(STRIPE_HANDLE, &sh->state);
2371 /* handle_stripe_clean_event
2372 * any written block on an uptodate or failed drive can be returned.
2373 * Note that if we 'wrote' to a failed drive, it will be UPTODATE, but
2374 * never LOCKED, so we don't need to test 'failed' directly.
2376 static void handle_stripe_clean_event(raid5_conf_t *conf,
2377 struct stripe_head *sh, int disks, struct bio **return_bi)
2382 for (i = disks; i--; )
2383 if (sh->dev[i].written) {
2385 if (!test_bit(R5_LOCKED, &dev->flags) &&
2386 test_bit(R5_UPTODATE, &dev->flags)) {
2387 /* We can return any write requests */
2388 struct bio *wbi, *wbi2;
2390 pr_debug("Return write for disc %d\n", i);
2391 spin_lock_irq(&conf->device_lock);
2393 dev->written = NULL;
2394 while (wbi && wbi->bi_sector <
2395 dev->sector + STRIPE_SECTORS) {
2396 wbi2 = r5_next_bio(wbi, dev->sector);
2397 if (!raid5_dec_bi_phys_segments(wbi)) {
2398 md_write_end(conf->mddev);
2399 wbi->bi_next = *return_bi;
2404 if (dev->towrite == NULL)
2406 spin_unlock_irq(&conf->device_lock);
2408 bitmap_endwrite(conf->mddev->bitmap,
2411 !test_bit(STRIPE_DEGRADED, &sh->state),
2416 if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
2417 if (atomic_dec_and_test(&conf->pending_full_writes))
2418 md_wakeup_thread(conf->mddev->thread);
2421 static void handle_stripe_dirtying5(raid5_conf_t *conf,
2422 struct stripe_head *sh, struct stripe_head_state *s, int disks)
2424 int rmw = 0, rcw = 0, i;
2425 for (i = disks; i--; ) {
2426 /* would I have to read this buffer for read_modify_write */
2427 struct r5dev *dev = &sh->dev[i];
2428 if ((dev->towrite || i == sh->pd_idx) &&
2429 !test_bit(R5_LOCKED, &dev->flags) &&
2430 !(test_bit(R5_UPTODATE, &dev->flags) ||
2431 test_bit(R5_Wantcompute, &dev->flags))) {
2432 if (test_bit(R5_Insync, &dev->flags))
2435 rmw += 2*disks; /* cannot read it */
2437 /* Would I have to read this buffer for reconstruct_write */
2438 if (!test_bit(R5_OVERWRITE, &dev->flags) && i != sh->pd_idx &&
2439 !test_bit(R5_LOCKED, &dev->flags) &&
2440 !(test_bit(R5_UPTODATE, &dev->flags) ||
2441 test_bit(R5_Wantcompute, &dev->flags))) {
2442 if (test_bit(R5_Insync, &dev->flags)) rcw++;
2447 pr_debug("for sector %llu, rmw=%d rcw=%d\n",
2448 (unsigned long long)sh->sector, rmw, rcw);
2449 set_bit(STRIPE_HANDLE, &sh->state);
2450 if (rmw < rcw && rmw > 0)
2451 /* prefer read-modify-write, but need to get some data */
2452 for (i = disks; i--; ) {
2453 struct r5dev *dev = &sh->dev[i];
2454 if ((dev->towrite || i == sh->pd_idx) &&
2455 !test_bit(R5_LOCKED, &dev->flags) &&
2456 !(test_bit(R5_UPTODATE, &dev->flags) ||
2457 test_bit(R5_Wantcompute, &dev->flags)) &&
2458 test_bit(R5_Insync, &dev->flags)) {
2460 test_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2461 pr_debug("Read_old block "
2462 "%d for r-m-w\n", i);
2463 set_bit(R5_LOCKED, &dev->flags);
2464 set_bit(R5_Wantread, &dev->flags);
2467 set_bit(STRIPE_DELAYED, &sh->state);
2468 set_bit(STRIPE_HANDLE, &sh->state);
2472 if (rcw <= rmw && rcw > 0)
2473 /* want reconstruct write, but need to get some data */
2474 for (i = disks; i--; ) {
2475 struct r5dev *dev = &sh->dev[i];
2476 if (!test_bit(R5_OVERWRITE, &dev->flags) &&
2478 !test_bit(R5_LOCKED, &dev->flags) &&
2479 !(test_bit(R5_UPTODATE, &dev->flags) ||
2480 test_bit(R5_Wantcompute, &dev->flags)) &&
2481 test_bit(R5_Insync, &dev->flags)) {
2483 test_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2484 pr_debug("Read_old block "
2485 "%d for Reconstruct\n", i);
2486 set_bit(R5_LOCKED, &dev->flags);
2487 set_bit(R5_Wantread, &dev->flags);
2490 set_bit(STRIPE_DELAYED, &sh->state);
2491 set_bit(STRIPE_HANDLE, &sh->state);
2495 /* now if nothing is locked, and if we have enough data,
2496 * we can start a write request
2498 /* since handle_stripe can be called at any time we need to handle the
2499 * case where a compute block operation has been submitted and then a
2500 * subsequent call wants to start a write request. raid_run_ops only
2501 * handles the case where compute block and reconstruct are requested
2502 * simultaneously. If this is not the case then new writes need to be
2503 * held off until the compute completes.
2505 if ((s->req_compute || !test_bit(STRIPE_COMPUTE_RUN, &sh->state)) &&
2506 (s->locked == 0 && (rcw == 0 || rmw == 0) &&
2507 !test_bit(STRIPE_BIT_DELAY, &sh->state)))
2508 schedule_reconstruction(sh, s, rcw == 0, 0);
2511 static void handle_stripe_dirtying6(raid5_conf_t *conf,
2512 struct stripe_head *sh, struct stripe_head_state *s,
2513 struct r6_state *r6s, int disks)
2515 int rcw = 0, pd_idx = sh->pd_idx, i;
2516 int qd_idx = sh->qd_idx;
2518 set_bit(STRIPE_HANDLE, &sh->state);
2519 for (i = disks; i--; ) {
2520 struct r5dev *dev = &sh->dev[i];
2521 /* check if we haven't enough data */
2522 if (!test_bit(R5_OVERWRITE, &dev->flags) &&
2523 i != pd_idx && i != qd_idx &&
2524 !test_bit(R5_LOCKED, &dev->flags) &&
2525 !(test_bit(R5_UPTODATE, &dev->flags) ||
2526 test_bit(R5_Wantcompute, &dev->flags))) {
2528 if (!test_bit(R5_Insync, &dev->flags))
2529 continue; /* it's a failed drive */
2532 test_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2533 pr_debug("Read_old stripe %llu "
2534 "block %d for Reconstruct\n",
2535 (unsigned long long)sh->sector, i);
2536 set_bit(R5_LOCKED, &dev->flags);
2537 set_bit(R5_Wantread, &dev->flags);
2540 pr_debug("Request delayed stripe %llu "
2541 "block %d for Reconstruct\n",
2542 (unsigned long long)sh->sector, i);
2543 set_bit(STRIPE_DELAYED, &sh->state);
2544 set_bit(STRIPE_HANDLE, &sh->state);
2548 /* now if nothing is locked, and if we have enough data, we can start a
2551 if ((s->req_compute || !test_bit(STRIPE_COMPUTE_RUN, &sh->state)) &&
2552 s->locked == 0 && rcw == 0 &&
2553 !test_bit(STRIPE_BIT_DELAY, &sh->state)) {
2554 schedule_reconstruction(sh, s, 1, 0);
2558 static void handle_parity_checks5(raid5_conf_t *conf, struct stripe_head *sh,
2559 struct stripe_head_state *s, int disks)
2561 struct r5dev *dev = NULL;
2563 set_bit(STRIPE_HANDLE, &sh->state);
2565 switch (sh->check_state) {
2566 case check_state_idle:
2567 /* start a new check operation if there are no failures */
2568 if (s->failed == 0) {
2569 BUG_ON(s->uptodate != disks);
2570 sh->check_state = check_state_run;
2571 set_bit(STRIPE_OP_CHECK, &s->ops_request);
2572 clear_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags);
2576 dev = &sh->dev[s->failed_num];
2578 case check_state_compute_result:
2579 sh->check_state = check_state_idle;
2581 dev = &sh->dev[sh->pd_idx];
2583 /* check that a write has not made the stripe insync */
2584 if (test_bit(STRIPE_INSYNC, &sh->state))
2587 /* either failed parity check, or recovery is happening */
2588 BUG_ON(!test_bit(R5_UPTODATE, &dev->flags));
2589 BUG_ON(s->uptodate != disks);
2591 set_bit(R5_LOCKED, &dev->flags);
2593 set_bit(R5_Wantwrite, &dev->flags);
2595 clear_bit(STRIPE_DEGRADED, &sh->state);
2596 set_bit(STRIPE_INSYNC, &sh->state);
2598 case check_state_run:
2599 break; /* we will be called again upon completion */
2600 case check_state_check_result:
2601 sh->check_state = check_state_idle;
2603 /* if a failure occurred during the check operation, leave
2604 * STRIPE_INSYNC not set and let the stripe be handled again
2609 /* handle a successful check operation, if parity is correct
2610 * we are done. Otherwise update the mismatch count and repair
2611 * parity if !MD_RECOVERY_CHECK
2613 if ((sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) == 0)
2614 /* parity is correct (on disc,
2615 * not in buffer any more)
2617 set_bit(STRIPE_INSYNC, &sh->state);
2619 conf->mddev->resync_mismatches += STRIPE_SECTORS;
2620 if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery))
2621 /* don't try to repair!! */
2622 set_bit(STRIPE_INSYNC, &sh->state);
2624 sh->check_state = check_state_compute_run;
2625 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
2626 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
2627 set_bit(R5_Wantcompute,
2628 &sh->dev[sh->pd_idx].flags);
2629 sh->ops.target = sh->pd_idx;
2630 sh->ops.target2 = -1;
2635 case check_state_compute_run:
2638 printk(KERN_ERR "%s: unknown check_state: %d sector: %llu\n",
2639 __func__, sh->check_state,
2640 (unsigned long long) sh->sector);
2646 static void handle_parity_checks6(raid5_conf_t *conf, struct stripe_head *sh,
2647 struct stripe_head_state *s,
2648 struct r6_state *r6s, int disks)
2650 int pd_idx = sh->pd_idx;
2651 int qd_idx = sh->qd_idx;
2654 set_bit(STRIPE_HANDLE, &sh->state);
2656 BUG_ON(s->failed > 2);
2658 /* Want to check and possibly repair P and Q.
2659 * However there could be one 'failed' device, in which
2660 * case we can only check one of them, possibly using the
2661 * other to generate missing data
2664 switch (sh->check_state) {
2665 case check_state_idle:
2666 /* start a new check operation if there are < 2 failures */
2667 if (s->failed == r6s->q_failed) {
2668 /* The only possible failed device holds Q, so it
2669 * makes sense to check P (If anything else were failed,
2670 * we would have used P to recreate it).
2672 sh->check_state = check_state_run;
2674 if (!r6s->q_failed && s->failed < 2) {
2675 /* Q is not failed, and we didn't use it to generate
2676 * anything, so it makes sense to check it
2678 if (sh->check_state == check_state_run)
2679 sh->check_state = check_state_run_pq;
2681 sh->check_state = check_state_run_q;
2684 /* discard potentially stale zero_sum_result */
2685 sh->ops.zero_sum_result = 0;
2687 if (sh->check_state == check_state_run) {
2688 /* async_xor_zero_sum destroys the contents of P */
2689 clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
2692 if (sh->check_state >= check_state_run &&
2693 sh->check_state <= check_state_run_pq) {
2694 /* async_syndrome_zero_sum preserves P and Q, so
2695 * no need to mark them !uptodate here
2697 set_bit(STRIPE_OP_CHECK, &s->ops_request);
2701 /* we have 2-disk failure */
2702 BUG_ON(s->failed != 2);
2704 case check_state_compute_result:
2705 sh->check_state = check_state_idle;
2707 /* check that a write has not made the stripe insync */
2708 if (test_bit(STRIPE_INSYNC, &sh->state))
2711 /* now write out any block on a failed drive,
2712 * or P or Q if they were recomputed
2714 BUG_ON(s->uptodate < disks - 1); /* We don't need Q to recover */
2715 if (s->failed == 2) {
2716 dev = &sh->dev[r6s->failed_num[1]];
2718 set_bit(R5_LOCKED, &dev->flags);
2719 set_bit(R5_Wantwrite, &dev->flags);
2721 if (s->failed >= 1) {
2722 dev = &sh->dev[r6s->failed_num[0]];
2724 set_bit(R5_LOCKED, &dev->flags);
2725 set_bit(R5_Wantwrite, &dev->flags);
2727 if (sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) {
2728 dev = &sh->dev[pd_idx];
2730 set_bit(R5_LOCKED, &dev->flags);
2731 set_bit(R5_Wantwrite, &dev->flags);
2733 if (sh->ops.zero_sum_result & SUM_CHECK_Q_RESULT) {
2734 dev = &sh->dev[qd_idx];
2736 set_bit(R5_LOCKED, &dev->flags);
2737 set_bit(R5_Wantwrite, &dev->flags);
2739 clear_bit(STRIPE_DEGRADED, &sh->state);
2741 set_bit(STRIPE_INSYNC, &sh->state);
2743 case check_state_run:
2744 case check_state_run_q:
2745 case check_state_run_pq:
2746 break; /* we will be called again upon completion */
2747 case check_state_check_result:
2748 sh->check_state = check_state_idle;
2750 /* handle a successful check operation, if parity is correct
2751 * we are done. Otherwise update the mismatch count and repair
2752 * parity if !MD_RECOVERY_CHECK
2754 if (sh->ops.zero_sum_result == 0) {
2755 /* both parities are correct */
2757 set_bit(STRIPE_INSYNC, &sh->state);
2759 /* in contrast to the raid5 case we can validate
2760 * parity, but still have a failure to write
2763 sh->check_state = check_state_compute_result;
2764 /* Returning at this point means that we may go
2765 * off and bring p and/or q uptodate again so
2766 * we make sure to check zero_sum_result again
2767 * to verify if p or q need writeback
2771 conf->mddev->resync_mismatches += STRIPE_SECTORS;
2772 if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery))
2773 /* don't try to repair!! */
2774 set_bit(STRIPE_INSYNC, &sh->state);
2776 int *target = &sh->ops.target;
2778 sh->ops.target = -1;
2779 sh->ops.target2 = -1;
2780 sh->check_state = check_state_compute_run;
2781 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
2782 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
2783 if (sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) {
2784 set_bit(R5_Wantcompute,
2785 &sh->dev[pd_idx].flags);
2787 target = &sh->ops.target2;
2790 if (sh->ops.zero_sum_result & SUM_CHECK_Q_RESULT) {
2791 set_bit(R5_Wantcompute,
2792 &sh->dev[qd_idx].flags);
2799 case check_state_compute_run:
2802 printk(KERN_ERR "%s: unknown check_state: %d sector: %llu\n",
2803 __func__, sh->check_state,
2804 (unsigned long long) sh->sector);
2809 static void handle_stripe_expansion(raid5_conf_t *conf, struct stripe_head *sh,
2810 struct r6_state *r6s)
2814 /* We have read all the blocks in this stripe and now we need to
2815 * copy some of them into a target stripe for expand.
2817 struct dma_async_tx_descriptor *tx = NULL;
2818 clear_bit(STRIPE_EXPAND_SOURCE, &sh->state);
2819 for (i = 0; i < sh->disks; i++)
2820 if (i != sh->pd_idx && i != sh->qd_idx) {
2822 struct stripe_head *sh2;
2823 struct async_submit_ctl submit;
2825 sector_t bn = compute_blocknr(sh, i, 1);
2826 sector_t s = raid5_compute_sector(conf, bn, 0,
2828 sh2 = get_active_stripe(conf, s, 0, 1);
2830 /* so far only the early blocks of this stripe
2831 * have been requested. When later blocks
2832 * get requested, we will try again
2835 if (!test_bit(STRIPE_EXPANDING, &sh2->state) ||
2836 test_bit(R5_Expanded, &sh2->dev[dd_idx].flags)) {
2837 /* must have already done this block */
2838 release_stripe(sh2);
2842 /* place all the copies on one channel */
2843 init_async_submit(&submit, 0, tx, NULL, NULL, NULL);
2844 tx = async_memcpy(sh2->dev[dd_idx].page,
2845 sh->dev[i].page, 0, 0, STRIPE_SIZE,
2848 set_bit(R5_Expanded, &sh2->dev[dd_idx].flags);
2849 set_bit(R5_UPTODATE, &sh2->dev[dd_idx].flags);
2850 for (j = 0; j < conf->raid_disks; j++)
2851 if (j != sh2->pd_idx &&
2852 (!r6s || j != sh2->qd_idx) &&
2853 !test_bit(R5_Expanded, &sh2->dev[j].flags))
2855 if (j == conf->raid_disks) {
2856 set_bit(STRIPE_EXPAND_READY, &sh2->state);
2857 set_bit(STRIPE_HANDLE, &sh2->state);
2859 release_stripe(sh2);
2862 /* done submitting copies, wait for them to complete */
2865 dma_wait_for_async_tx(tx);
2871 * handle_stripe - do things to a stripe.
2873 * We lock the stripe and then examine the state of various bits
2874 * to see what needs to be done.
2876 * return some read request which now have data
2877 * return some write requests which are safely on disc
2878 * schedule a read on some buffers
2879 * schedule a write of some buffers
2880 * return confirmation of parity correctness
2882 * buffers are taken off read_list or write_list, and bh_cache buffers
2883 * get BH_Lock set before the stripe lock is released.
2887 static bool handle_stripe5(struct stripe_head *sh)
2889 raid5_conf_t *conf = sh->raid_conf;
2890 int disks = sh->disks, i;
2891 struct bio *return_bi = NULL;
2892 struct stripe_head_state s;
2894 mdk_rdev_t *blocked_rdev = NULL;
2897 memset(&s, 0, sizeof(s));
2898 pr_debug("handling stripe %llu, state=%#lx cnt=%d, pd_idx=%d check:%d "
2899 "reconstruct:%d\n", (unsigned long long)sh->sector, sh->state,
2900 atomic_read(&sh->count), sh->pd_idx, sh->check_state,
2901 sh->reconstruct_state);
2903 spin_lock(&sh->lock);
2904 clear_bit(STRIPE_HANDLE, &sh->state);
2905 clear_bit(STRIPE_DELAYED, &sh->state);
2907 s.syncing = test_bit(STRIPE_SYNCING, &sh->state);
2908 s.expanding = test_bit(STRIPE_EXPAND_SOURCE, &sh->state);
2909 s.expanded = test_bit(STRIPE_EXPAND_READY, &sh->state);
2911 /* Now to look around and see what can be done */
2913 for (i=disks; i--; ) {
2915 struct r5dev *dev = &sh->dev[i];
2916 clear_bit(R5_Insync, &dev->flags);
2918 pr_debug("check %d: state 0x%lx toread %p read %p write %p "
2919 "written %p\n", i, dev->flags, dev->toread, dev->read,
2920 dev->towrite, dev->written);
2922 /* maybe we can request a biofill operation
2924 * new wantfill requests are only permitted while
2925 * ops_complete_biofill is guaranteed to be inactive
2927 if (test_bit(R5_UPTODATE, &dev->flags) && dev->toread &&
2928 !test_bit(STRIPE_BIOFILL_RUN, &sh->state))
2929 set_bit(R5_Wantfill, &dev->flags);
2931 /* now count some things */
2932 if (test_bit(R5_LOCKED, &dev->flags)) s.locked++;
2933 if (test_bit(R5_UPTODATE, &dev->flags)) s.uptodate++;
2934 if (test_bit(R5_Wantcompute, &dev->flags)) s.compute++;
2936 if (test_bit(R5_Wantfill, &dev->flags))
2938 else if (dev->toread)
2942 if (!test_bit(R5_OVERWRITE, &dev->flags))
2947 rdev = rcu_dereference(conf->disks[i].rdev);
2948 if (blocked_rdev == NULL &&
2949 rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
2950 blocked_rdev = rdev;
2951 atomic_inc(&rdev->nr_pending);
2953 if (!rdev || !test_bit(In_sync, &rdev->flags)) {
2954 /* The ReadError flag will just be confusing now */
2955 clear_bit(R5_ReadError, &dev->flags);
2956 clear_bit(R5_ReWrite, &dev->flags);
2958 if (!rdev || !test_bit(In_sync, &rdev->flags)
2959 || test_bit(R5_ReadError, &dev->flags)) {
2963 set_bit(R5_Insync, &dev->flags);
2967 if (unlikely(blocked_rdev)) {
2968 if (s.syncing || s.expanding || s.expanded ||
2969 s.to_write || s.written) {
2970 set_bit(STRIPE_HANDLE, &sh->state);
2973 /* There is nothing for the blocked_rdev to block */
2974 rdev_dec_pending(blocked_rdev, conf->mddev);
2975 blocked_rdev = NULL;
2978 if (s.to_fill && !test_bit(STRIPE_BIOFILL_RUN, &sh->state)) {
2979 set_bit(STRIPE_OP_BIOFILL, &s.ops_request);
2980 set_bit(STRIPE_BIOFILL_RUN, &sh->state);
2983 pr_debug("locked=%d uptodate=%d to_read=%d"
2984 " to_write=%d failed=%d failed_num=%d\n",
2985 s.locked, s.uptodate, s.to_read, s.to_write,
2986 s.failed, s.failed_num);
2987 /* check if the array has lost two devices and, if so, some requests might
2990 if (s.failed > 1 && s.to_read+s.to_write+s.written)
2991 handle_failed_stripe(conf, sh, &s, disks, &return_bi);
2992 if (s.failed > 1 && s.syncing) {
2993 md_done_sync(conf->mddev, STRIPE_SECTORS,0);
2994 clear_bit(STRIPE_SYNCING, &sh->state);
2998 /* might be able to return some write requests if the parity block
2999 * is safe, or on a failed drive
3001 dev = &sh->dev[sh->pd_idx];
3003 ((test_bit(R5_Insync, &dev->flags) &&
3004 !test_bit(R5_LOCKED, &dev->flags) &&
3005 test_bit(R5_UPTODATE, &dev->flags)) ||
3006 (s.failed == 1 && s.failed_num == sh->pd_idx)))
3007 handle_stripe_clean_event(conf, sh, disks, &return_bi);
3009 /* Now we might consider reading some blocks, either to check/generate
3010 * parity, or to satisfy requests
3011 * or to load a block that is being partially written.
3013 if (s.to_read || s.non_overwrite ||
3014 (s.syncing && (s.uptodate + s.compute < disks)) || s.expanding)
3015 handle_stripe_fill5(sh, &s, disks);
3017 /* Now we check to see if any write operations have recently
3021 if (sh->reconstruct_state == reconstruct_state_prexor_drain_result)
3023 if (sh->reconstruct_state == reconstruct_state_drain_result ||
3024 sh->reconstruct_state == reconstruct_state_prexor_drain_result) {
3025 sh->reconstruct_state = reconstruct_state_idle;
3027 /* All the 'written' buffers and the parity block are ready to
3028 * be written back to disk
3030 BUG_ON(!test_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags));
3031 for (i = disks; i--; ) {
3033 if (test_bit(R5_LOCKED, &dev->flags) &&
3034 (i == sh->pd_idx || dev->written)) {
3035 pr_debug("Writing block %d\n", i);
3036 set_bit(R5_Wantwrite, &dev->flags);
3039 if (!test_bit(R5_Insync, &dev->flags) ||
3040 (i == sh->pd_idx && s.failed == 0))
3041 set_bit(STRIPE_INSYNC, &sh->state);
3044 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
3045 atomic_dec(&conf->preread_active_stripes);
3046 if (atomic_read(&conf->preread_active_stripes) <
3048 md_wakeup_thread(conf->mddev->thread);
3052 /* Now to consider new write requests and what else, if anything
3053 * should be read. We do not handle new writes when:
3054 * 1/ A 'write' operation (copy+xor) is already in flight.
3055 * 2/ A 'check' operation is in flight, as it may clobber the parity
3058 if (s.to_write && !sh->reconstruct_state && !sh->check_state)
3059 handle_stripe_dirtying5(conf, sh, &s, disks);
3061 /* maybe we need to check and possibly fix the parity for this stripe
3062 * Any reads will already have been scheduled, so we just see if enough
3063 * data is available. The parity check is held off while parity
3064 * dependent operations are in flight.
3066 if (sh->check_state ||
3067 (s.syncing && s.locked == 0 &&
3068 !test_bit(STRIPE_COMPUTE_RUN, &sh->state) &&
3069 !test_bit(STRIPE_INSYNC, &sh->state)))
3070 handle_parity_checks5(conf, sh, &s, disks);
3072 if (s.syncing && s.locked == 0 && test_bit(STRIPE_INSYNC, &sh->state)) {
3073 md_done_sync(conf->mddev, STRIPE_SECTORS,1);
3074 clear_bit(STRIPE_SYNCING, &sh->state);
3077 /* If the failed drive is just a ReadError, then we might need to progress
3078 * the repair/check process
3080 if (s.failed == 1 && !conf->mddev->ro &&
3081 test_bit(R5_ReadError, &sh->dev[s.failed_num].flags)
3082 && !test_bit(R5_LOCKED, &sh->dev[s.failed_num].flags)
3083 && test_bit(R5_UPTODATE, &sh->dev[s.failed_num].flags)
3085 dev = &sh->dev[s.failed_num];
3086 if (!test_bit(R5_ReWrite, &dev->flags)) {
3087 set_bit(R5_Wantwrite, &dev->flags);
3088 set_bit(R5_ReWrite, &dev->flags);
3089 set_bit(R5_LOCKED, &dev->flags);
3092 /* let's read it back */
3093 set_bit(R5_Wantread, &dev->flags);
3094 set_bit(R5_LOCKED, &dev->flags);
3099 /* Finish reconstruct operations initiated by the expansion process */
3100 if (sh->reconstruct_state == reconstruct_state_result) {
3101 struct stripe_head *sh2
3102 = get_active_stripe(conf, sh->sector, 1, 1);
3103 if (sh2 && test_bit(STRIPE_EXPAND_SOURCE, &sh2->state)) {
3104 /* sh cannot be written until sh2 has been read.
3105 * so arrange for sh to be delayed a little
3107 set_bit(STRIPE_DELAYED, &sh->state);
3108 set_bit(STRIPE_HANDLE, &sh->state);
3109 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE,
3111 atomic_inc(&conf->preread_active_stripes);
3112 release_stripe(sh2);
3116 release_stripe(sh2);
3118 sh->reconstruct_state = reconstruct_state_idle;
3119 clear_bit(STRIPE_EXPANDING, &sh->state);
3120 for (i = conf->raid_disks; i--; ) {
3121 set_bit(R5_Wantwrite, &sh->dev[i].flags);
3122 set_bit(R5_LOCKED, &sh->dev[i].flags);
3127 if (s.expanded && test_bit(STRIPE_EXPANDING, &sh->state) &&
3128 !sh->reconstruct_state) {
3129 /* Need to write out all blocks after computing parity */
3130 sh->disks = conf->raid_disks;
3131 stripe_set_idx(sh->sector, conf, 0, sh);
3132 schedule_reconstruction(sh, &s, 1, 1);
3133 } else if (s.expanded && !sh->reconstruct_state && s.locked == 0) {
3134 clear_bit(STRIPE_EXPAND_READY, &sh->state);
3135 atomic_dec(&conf->reshape_stripes);
3136 wake_up(&conf->wait_for_overlap);
3137 md_done_sync(conf->mddev, STRIPE_SECTORS, 1);
3140 if (s.expanding && s.locked == 0 &&
3141 !test_bit(STRIPE_COMPUTE_RUN, &sh->state))
3142 handle_stripe_expansion(conf, sh, NULL);
3145 spin_unlock(&sh->lock);
3147 /* wait for this device to become unblocked */
3148 if (unlikely(blocked_rdev))
3149 md_wait_for_blocked_rdev(blocked_rdev, conf->mddev);
3152 raid_run_ops(sh, s.ops_request);
3156 return_io(return_bi);
3158 return blocked_rdev == NULL;
3161 static bool handle_stripe6(struct stripe_head *sh)
3163 raid5_conf_t *conf = sh->raid_conf;
3164 int disks = sh->disks;
3165 struct bio *return_bi = NULL;
3166 int i, pd_idx = sh->pd_idx, qd_idx = sh->qd_idx;
3167 struct stripe_head_state s;
3168 struct r6_state r6s;
3169 struct r5dev *dev, *pdev, *qdev;
3170 mdk_rdev_t *blocked_rdev = NULL;
3172 pr_debug("handling stripe %llu, state=%#lx cnt=%d, "
3173 "pd_idx=%d, qd_idx=%d\n, check:%d, reconstruct:%d\n",
3174 (unsigned long long)sh->sector, sh->state,
3175 atomic_read(&sh->count), pd_idx, qd_idx,
3176 sh->check_state, sh->reconstruct_state);
3177 memset(&s, 0, sizeof(s));
3179 spin_lock(&sh->lock);
3180 clear_bit(STRIPE_HANDLE, &sh->state);
3181 clear_bit(STRIPE_DELAYED, &sh->state);
3183 s.syncing = test_bit(STRIPE_SYNCING, &sh->state);
3184 s.expanding = test_bit(STRIPE_EXPAND_SOURCE, &sh->state);
3185 s.expanded = test_bit(STRIPE_EXPAND_READY, &sh->state);
3186 /* Now to look around and see what can be done */
3189 for (i=disks; i--; ) {
3192 clear_bit(R5_Insync, &dev->flags);
3194 pr_debug("check %d: state 0x%lx read %p write %p written %p\n",
3195 i, dev->flags, dev->toread, dev->towrite, dev->written);
3196 /* maybe we can reply to a read
3198 * new wantfill requests are only permitted while
3199 * ops_complete_biofill is guaranteed to be inactive
3201 if (test_bit(R5_UPTODATE, &dev->flags) && dev->toread &&
3202 !test_bit(STRIPE_BIOFILL_RUN, &sh->state))
3203 set_bit(R5_Wantfill, &dev->flags);
3205 /* now count some things */
3206 if (test_bit(R5_LOCKED, &dev->flags)) s.locked++;
3207 if (test_bit(R5_UPTODATE, &dev->flags)) s.uptodate++;
3208 if (test_bit(R5_Wantcompute, &dev->flags))
3209 BUG_ON(++s.compute > 2);
3211 if (test_bit(R5_Wantfill, &dev->flags)) {
3213 } else if (dev->toread)
3217 if (!test_bit(R5_OVERWRITE, &dev->flags))
3222 rdev = rcu_dereference(conf->disks[i].rdev);
3223 if (blocked_rdev == NULL &&
3224 rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
3225 blocked_rdev = rdev;
3226 atomic_inc(&rdev->nr_pending);
3228 if (!rdev || !test_bit(In_sync, &rdev->flags)) {
3229 /* The ReadError flag will just be confusing now */
3230 clear_bit(R5_ReadError, &dev->flags);
3231 clear_bit(R5_ReWrite, &dev->flags);
3233 if (!rdev || !test_bit(In_sync, &rdev->flags)
3234 || test_bit(R5_ReadError, &dev->flags)) {
3236 r6s.failed_num[s.failed] = i;
3239 set_bit(R5_Insync, &dev->flags);
3243 if (unlikely(blocked_rdev)) {
3244 if (s.syncing || s.expanding || s.expanded ||
3245 s.to_write || s.written) {
3246 set_bit(STRIPE_HANDLE, &sh->state);
3249 /* There is nothing for the blocked_rdev to block */
3250 rdev_dec_pending(blocked_rdev, conf->mddev);
3251 blocked_rdev = NULL;
3254 if (s.to_fill && !test_bit(STRIPE_BIOFILL_RUN, &sh->state)) {
3255 set_bit(STRIPE_OP_BIOFILL, &s.ops_request);
3256 set_bit(STRIPE_BIOFILL_RUN, &sh->state);
3259 pr_debug("locked=%d uptodate=%d to_read=%d"
3260 " to_write=%d failed=%d failed_num=%d,%d\n",
3261 s.locked, s.uptodate, s.to_read, s.to_write, s.failed,
3262 r6s.failed_num[0], r6s.failed_num[1]);
3263 /* check if the array has lost >2 devices and, if so, some requests
3264 * might need to be failed
3266 if (s.failed > 2 && s.to_read+s.to_write+s.written)
3267 handle_failed_stripe(conf, sh, &s, disks, &return_bi);
3268 if (s.failed > 2 && s.syncing) {
3269 md_done_sync(conf->mddev, STRIPE_SECTORS,0);
3270 clear_bit(STRIPE_SYNCING, &sh->state);
3275 * might be able to return some write requests if the parity blocks
3276 * are safe, or on a failed drive
3278 pdev = &sh->dev[pd_idx];
3279 r6s.p_failed = (s.failed >= 1 && r6s.failed_num[0] == pd_idx)
3280 || (s.failed >= 2 && r6s.failed_num[1] == pd_idx);
3281 qdev = &sh->dev[qd_idx];
3282 r6s.q_failed = (s.failed >= 1 && r6s.failed_num[0] == qd_idx)
3283 || (s.failed >= 2 && r6s.failed_num[1] == qd_idx);
3286 ( r6s.p_failed || ((test_bit(R5_Insync, &pdev->flags)
3287 && !test_bit(R5_LOCKED, &pdev->flags)
3288 && test_bit(R5_UPTODATE, &pdev->flags)))) &&
3289 ( r6s.q_failed || ((test_bit(R5_Insync, &qdev->flags)
3290 && !test_bit(R5_LOCKED, &qdev->flags)
3291 && test_bit(R5_UPTODATE, &qdev->flags)))))
3292 handle_stripe_clean_event(conf, sh, disks, &return_bi);
3294 /* Now we might consider reading some blocks, either to check/generate
3295 * parity, or to satisfy requests
3296 * or to load a block that is being partially written.
3298 if (s.to_read || s.non_overwrite || (s.to_write && s.failed) ||
3299 (s.syncing && (s.uptodate + s.compute < disks)) || s.expanding)
3300 handle_stripe_fill6(sh, &s, &r6s, disks);
3302 /* Now we check to see if any write operations have recently
3305 if (sh->reconstruct_state == reconstruct_state_drain_result) {
3306 int qd_idx = sh->qd_idx;
3308 sh->reconstruct_state = reconstruct_state_idle;
3309 /* All the 'written' buffers and the parity blocks are ready to
3310 * be written back to disk
3312 BUG_ON(!test_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags));
3313 BUG_ON(!test_bit(R5_UPTODATE, &sh->dev[qd_idx].flags));
3314 for (i = disks; i--; ) {
3316 if (test_bit(R5_LOCKED, &dev->flags) &&
3317 (i == sh->pd_idx || i == qd_idx ||
3319 pr_debug("Writing block %d\n", i);
3320 BUG_ON(!test_bit(R5_UPTODATE, &dev->flags));
3321 set_bit(R5_Wantwrite, &dev->flags);
3322 if (!test_bit(R5_Insync, &dev->flags) ||
3323 ((i == sh->pd_idx || i == qd_idx) &&
3325 set_bit(STRIPE_INSYNC, &sh->state);
3328 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
3329 atomic_dec(&conf->preread_active_stripes);
3330 if (atomic_read(&conf->preread_active_stripes) <
3332 md_wakeup_thread(conf->mddev->thread);
3336 /* Now to consider new write requests and what else, if anything
3337 * should be read. We do not handle new writes when:
3338 * 1/ A 'write' operation (copy+gen_syndrome) is already in flight.
3339 * 2/ A 'check' operation is in flight, as it may clobber the parity
3342 if (s.to_write && !sh->reconstruct_state && !sh->check_state)
3343 handle_stripe_dirtying6(conf, sh, &s, &r6s, disks);
3345 /* maybe we need to check and possibly fix the parity for this stripe
3346 * Any reads will already have been scheduled, so we just see if enough
3347 * data is available. The parity check is held off while parity
3348 * dependent operations are in flight.
3350 if (sh->check_state ||
3351 (s.syncing && s.locked == 0 &&
3352 !test_bit(STRIPE_COMPUTE_RUN, &sh->state) &&
3353 !test_bit(STRIPE_INSYNC, &sh->state)))
3354 handle_parity_checks6(conf, sh, &s, &r6s, disks);
3356 if (s.syncing && s.locked == 0 && test_bit(STRIPE_INSYNC, &sh->state)) {
3357 md_done_sync(conf->mddev, STRIPE_SECTORS,1);
3358 clear_bit(STRIPE_SYNCING, &sh->state);
3361 /* If the failed drives are just a ReadError, then we might need
3362 * to progress the repair/check process
3364 if (s.failed <= 2 && !conf->mddev->ro)
3365 for (i = 0; i < s.failed; i++) {
3366 dev = &sh->dev[r6s.failed_num[i]];
3367 if (test_bit(R5_ReadError, &dev->flags)
3368 && !test_bit(R5_LOCKED, &dev->flags)
3369 && test_bit(R5_UPTODATE, &dev->flags)
3371 if (!test_bit(R5_ReWrite, &dev->flags)) {
3372 set_bit(R5_Wantwrite, &dev->flags);
3373 set_bit(R5_ReWrite, &dev->flags);
3374 set_bit(R5_LOCKED, &dev->flags);
3377 /* let's read it back */
3378 set_bit(R5_Wantread, &dev->flags);
3379 set_bit(R5_LOCKED, &dev->flags);
3385 /* Finish reconstruct operations initiated by the expansion process */
3386 if (sh->reconstruct_state == reconstruct_state_result) {
3387 sh->reconstruct_state = reconstruct_state_idle;
3388 clear_bit(STRIPE_EXPANDING, &sh->state);
3389 for (i = conf->raid_disks; i--; ) {
3390 set_bit(R5_Wantwrite, &sh->dev[i].flags);
3391 set_bit(R5_LOCKED, &sh->dev[i].flags);
3396 if (s.expanded && test_bit(STRIPE_EXPANDING, &sh->state) &&
3397 !sh->reconstruct_state) {
3398 struct stripe_head *sh2
3399 = get_active_stripe(conf, sh->sector, 1, 1);
3400 if (sh2 && test_bit(STRIPE_EXPAND_SOURCE, &sh2->state)) {
3401 /* sh cannot be written until sh2 has been read.
3402 * so arrange for sh to be delayed a little
3404 set_bit(STRIPE_DELAYED, &sh->state);
3405 set_bit(STRIPE_HANDLE, &sh->state);
3406 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE,
3408 atomic_inc(&conf->preread_active_stripes);
3409 release_stripe(sh2);
3413 release_stripe(sh2);
3415 /* Need to write out all blocks after computing P&Q */
3416 sh->disks = conf->raid_disks;
3417 stripe_set_idx(sh->sector, conf, 0, sh);
3418 schedule_reconstruction(sh, &s, 1, 1);
3419 } else if (s.expanded && !sh->reconstruct_state && s.locked == 0) {
3420 clear_bit(STRIPE_EXPAND_READY, &sh->state);
3421 atomic_dec(&conf->reshape_stripes);
3422 wake_up(&conf->wait_for_overlap);
3423 md_done_sync(conf->mddev, STRIPE_SECTORS, 1);
3426 if (s.expanding && s.locked == 0 &&
3427 !test_bit(STRIPE_COMPUTE_RUN, &sh->state))
3428 handle_stripe_expansion(conf, sh, &r6s);
3431 spin_unlock(&sh->lock);
3433 /* wait for this device to become unblocked */
3434 if (unlikely(blocked_rdev))
3435 md_wait_for_blocked_rdev(blocked_rdev, conf->mddev);
3438 raid_run_ops(sh, s.ops_request);
3442 return_io(return_bi);
3444 return blocked_rdev == NULL;
3447 /* returns true if the stripe was handled */
3448 static bool handle_stripe(struct stripe_head *sh)
3450 if (sh->raid_conf->level == 6)
3451 return handle_stripe6(sh);
3453 return handle_stripe5(sh);
3456 static void raid5_activate_delayed(raid5_conf_t *conf)
3458 if (atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD) {
3459 while (!list_empty(&conf->delayed_list)) {
3460 struct list_head *l = conf->delayed_list.next;
3461 struct stripe_head *sh;
3462 sh = list_entry(l, struct stripe_head, lru);
3464 clear_bit(STRIPE_DELAYED, &sh->state);
3465 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
3466 atomic_inc(&conf->preread_active_stripes);
3467 list_add_tail(&sh->lru, &conf->hold_list);
3470 blk_plug_device(conf->mddev->queue);
3473 static void activate_bit_delay(raid5_conf_t *conf)
3475 /* device_lock is held */
3476 struct list_head head;
3477 list_add(&head, &conf->bitmap_list);
3478 list_del_init(&conf->bitmap_list);
3479 while (!list_empty(&head)) {
3480 struct stripe_head *sh = list_entry(head.next, struct stripe_head, lru);
3481 list_del_init(&sh->lru);
3482 atomic_inc(&sh->count);
3483 __release_stripe(conf, sh);
3487 static void unplug_slaves(mddev_t *mddev)
3489 raid5_conf_t *conf = mddev_to_conf(mddev);
3493 for (i=0; i<mddev->raid_disks; i++) {
3494 mdk_rdev_t *rdev = rcu_dereference(conf->disks[i].rdev);
3495 if (rdev && !test_bit(Faulty, &rdev->flags) && atomic_read(&rdev->nr_pending)) {
3496 struct request_queue *r_queue = bdev_get_queue(rdev->bdev);
3498 atomic_inc(&rdev->nr_pending);
3501 blk_unplug(r_queue);
3503 rdev_dec_pending(rdev, mddev);
3510 static void raid5_unplug_device(struct request_queue *q)
3512 mddev_t *mddev = q->queuedata;
3513 raid5_conf_t *conf = mddev_to_conf(mddev);
3514 unsigned long flags;
3516 spin_lock_irqsave(&conf->device_lock, flags);
3518 if (blk_remove_plug(q)) {
3520 raid5_activate_delayed(conf);
3522 md_wakeup_thread(mddev->thread);
3524 spin_unlock_irqrestore(&conf->device_lock, flags);
3526 unplug_slaves(mddev);
3529 static int raid5_congested(void *data, int bits)
3531 mddev_t *mddev = data;
3532 raid5_conf_t *conf = mddev_to_conf(mddev);
3534 /* No difference between reads and writes. Just check
3535 * how busy the stripe_cache is
3537 if (conf->inactive_blocked)
3541 if (list_empty_careful(&conf->inactive_list))
3547 /* We want read requests to align with chunks where possible,
3548 * but write requests don't need to.
3550 static int raid5_mergeable_bvec(struct request_queue *q,
3551 struct bvec_merge_data *bvm,
3552 struct bio_vec *biovec)
3554 mddev_t *mddev = q->queuedata;
3555 sector_t sector = bvm->bi_sector + get_start_sect(bvm->bi_bdev);
3557 unsigned int chunk_sectors = mddev->chunk_size >> 9;
3558 unsigned int bio_sectors = bvm->bi_size >> 9;
3560 if ((bvm->bi_rw & 1) == WRITE)
3561 return biovec->bv_len; /* always allow writes to be mergeable */
3563 if (mddev->new_chunk < mddev->chunk_size)
3564 chunk_sectors = mddev->new_chunk >> 9;
3565 max = (chunk_sectors - ((sector & (chunk_sectors - 1)) + bio_sectors)) << 9;
3566 if (max < 0) max = 0;
3567 if (max <= biovec->bv_len && bio_sectors == 0)
3568 return biovec->bv_len;
3574 static int in_chunk_boundary(mddev_t *mddev, struct bio *bio)
3576 sector_t sector = bio->bi_sector + get_start_sect(bio->bi_bdev);
3577 unsigned int chunk_sectors = mddev->chunk_size >> 9;
3578 unsigned int bio_sectors = bio->bi_size >> 9;
3580 if (mddev->new_chunk < mddev->chunk_size)
3581 chunk_sectors = mddev->new_chunk >> 9;
3582 return chunk_sectors >=
3583 ((sector & (chunk_sectors - 1)) + bio_sectors);
3587 * add bio to the retry LIFO ( in O(1) ... we are in interrupt )
3588 * later sampled by raid5d.
3590 static void add_bio_to_retry(struct bio *bi,raid5_conf_t *conf)
3592 unsigned long flags;
3594 spin_lock_irqsave(&conf->device_lock, flags);
3596 bi->bi_next = conf->retry_read_aligned_list;
3597 conf->retry_read_aligned_list = bi;
3599 spin_unlock_irqrestore(&conf->device_lock, flags);
3600 md_wakeup_thread(conf->mddev->thread);
3604 static struct bio *remove_bio_from_retry(raid5_conf_t *conf)
3608 bi = conf->retry_read_aligned;
3610 conf->retry_read_aligned = NULL;
3613 bi = conf->retry_read_aligned_list;
3615 conf->retry_read_aligned_list = bi->bi_next;
3618 * this sets the active strip count to 1 and the processed
3619 * strip count to zero (upper 8 bits)
3621 bi->bi_phys_segments = 1; /* biased count of active stripes */
3629 * The "raid5_align_endio" should check if the read succeeded and if it
3630 * did, call bio_endio on the original bio (having bio_put the new bio
3632 * If the read failed..
3634 static void raid5_align_endio(struct bio *bi, int error)
3636 struct bio* raid_bi = bi->bi_private;
3639 int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
3644 mddev = raid_bi->bi_bdev->bd_disk->queue->queuedata;
3645 conf = mddev_to_conf(mddev);
3646 rdev = (void*)raid_bi->bi_next;
3647 raid_bi->bi_next = NULL;
3649 rdev_dec_pending(rdev, conf->mddev);
3651 if (!error && uptodate) {
3652 bio_endio(raid_bi, 0);
3653 if (atomic_dec_and_test(&conf->active_aligned_reads))
3654 wake_up(&conf->wait_for_stripe);
3659 pr_debug("raid5_align_endio : io error...handing IO for a retry\n");
3661 add_bio_to_retry(raid_bi, conf);
3664 static int bio_fits_rdev(struct bio *bi)
3666 struct request_queue *q = bdev_get_queue(bi->bi_bdev);
3668 if ((bi->bi_size>>9) > q->max_sectors)
3670 blk_recount_segments(q, bi);
3671 if (bi->bi_phys_segments > q->max_phys_segments)
3674 if (q->merge_bvec_fn)
3675 /* it's too hard to apply the merge_bvec_fn at this stage,
3684 static int chunk_aligned_read(struct request_queue *q, struct bio * raid_bio)
3686 mddev_t *mddev = q->queuedata;
3687 raid5_conf_t *conf = mddev_to_conf(mddev);
3688 unsigned int dd_idx;
3689 struct bio* align_bi;
3692 if (!in_chunk_boundary(mddev, raid_bio)) {
3693 pr_debug("chunk_aligned_read : non aligned\n");
3697 * use bio_clone to make a copy of the bio
3699 align_bi = bio_clone(raid_bio, GFP_NOIO);
3703 * set bi_end_io to a new function, and set bi_private to the
3706 align_bi->bi_end_io = raid5_align_endio;
3707 align_bi->bi_private = raid_bio;
3711 align_bi->bi_sector = raid5_compute_sector(conf, raid_bio->bi_sector,
3716 rdev = rcu_dereference(conf->disks[dd_idx].rdev);
3717 if (rdev && test_bit(In_sync, &rdev->flags)) {
3718 atomic_inc(&rdev->nr_pending);
3720 raid_bio->bi_next = (void*)rdev;
3721 align_bi->bi_bdev = rdev->bdev;
3722 align_bi->bi_flags &= ~(1 << BIO_SEG_VALID);
3723 align_bi->bi_sector += rdev->data_offset;
3725 if (!bio_fits_rdev(align_bi)) {
3726 /* too big in some way */
3728 rdev_dec_pending(rdev, mddev);
3732 spin_lock_irq(&conf->device_lock);
3733 wait_event_lock_irq(conf->wait_for_stripe,
3735 conf->device_lock, /* nothing */);
3736 atomic_inc(&conf->active_aligned_reads);
3737 spin_unlock_irq(&conf->device_lock);
3739 generic_make_request(align_bi);
3748 /* __get_priority_stripe - get the next stripe to process
3750 * Full stripe writes are allowed to pass preread active stripes up until
3751 * the bypass_threshold is exceeded. In general the bypass_count
3752 * increments when the handle_list is handled before the hold_list; however, it
3753 * will not be incremented when STRIPE_IO_STARTED is sampled set signifying a
3754 * stripe with in flight i/o. The bypass_count will be reset when the
3755 * head of the hold_list has changed, i.e. the head was promoted to the
3758 static struct stripe_head *__get_priority_stripe(raid5_conf_t *conf)
3760 struct stripe_head *sh;
3762 pr_debug("%s: handle: %s hold: %s full_writes: %d bypass_count: %d\n",
3764 list_empty(&conf->handle_list) ? "empty" : "busy",
3765 list_empty(&conf->hold_list) ? "empty" : "busy",
3766 atomic_read(&conf->pending_full_writes), conf->bypass_count);
3768 if (!list_empty(&conf->handle_list)) {
3769 sh = list_entry(conf->handle_list.next, typeof(*sh), lru);
3771 if (list_empty(&conf->hold_list))
3772 conf->bypass_count = 0;
3773 else if (!test_bit(STRIPE_IO_STARTED, &sh->state)) {
3774 if (conf->hold_list.next == conf->last_hold)
3775 conf->bypass_count++;
3777 conf->last_hold = conf->hold_list.next;
3778 conf->bypass_count -= conf->bypass_threshold;
3779 if (conf->bypass_count < 0)
3780 conf->bypass_count = 0;
3783 } else if (!list_empty(&conf->hold_list) &&
3784 ((conf->bypass_threshold &&
3785 conf->bypass_count > conf->bypass_threshold) ||
3786 atomic_read(&conf->pending_full_writes) == 0)) {
3787 sh = list_entry(conf->hold_list.next,
3789 conf->bypass_count -= conf->bypass_threshold;
3790 if (conf->bypass_count < 0)
3791 conf->bypass_count = 0;
3795 list_del_init(&sh->lru);
3796 atomic_inc(&sh->count);
3797 BUG_ON(atomic_read(&sh->count) != 1);
3801 static int make_request(struct request_queue *q, struct bio * bi)
3803 mddev_t *mddev = q->queuedata;
3804 raid5_conf_t *conf = mddev_to_conf(mddev);
3806 sector_t new_sector;
3807 sector_t logical_sector, last_sector;
3808 struct stripe_head *sh;
3809 const int rw = bio_data_dir(bi);
3812 if (unlikely(bio_barrier(bi))) {
3813 bio_endio(bi, -EOPNOTSUPP);
3817 md_write_start(mddev, bi);
3819 cpu = part_stat_lock();
3820 part_stat_inc(cpu, &mddev->gendisk->part0, ios[rw]);
3821 part_stat_add(cpu, &mddev->gendisk->part0, sectors[rw],
3826 mddev->reshape_position == MaxSector &&
3827 chunk_aligned_read(q,bi))
3830 logical_sector = bi->bi_sector & ~((sector_t)STRIPE_SECTORS-1);
3831 last_sector = bi->bi_sector + (bi->bi_size>>9);
3833 bi->bi_phys_segments = 1; /* over-loaded to count active stripes */
3835 for (;logical_sector < last_sector; logical_sector += STRIPE_SECTORS) {
3837 int disks, data_disks;
3842 disks = conf->raid_disks;
3843 prepare_to_wait(&conf->wait_for_overlap, &w, TASK_UNINTERRUPTIBLE);
3844 if (unlikely(conf->reshape_progress != MaxSector)) {
3845 /* spinlock is needed as reshape_progress may be
3846 * 64bit on a 32bit platform, and so it might be
3847 * possible to see a half-updated value
3848 * Ofcourse reshape_progress could change after
3849 * the lock is dropped, so once we get a reference
3850 * to the stripe that we think it is, we will have
3853 spin_lock_irq(&conf->device_lock);
3854 if (mddev->delta_disks < 0
3855 ? logical_sector < conf->reshape_progress
3856 : logical_sector >= conf->reshape_progress) {
3857 disks = conf->previous_raid_disks;
3860 if (mddev->delta_disks < 0
3861 ? logical_sector < conf->reshape_safe
3862 : logical_sector >= conf->reshape_safe) {
3863 spin_unlock_irq(&conf->device_lock);
3868 spin_unlock_irq(&conf->device_lock);
3870 data_disks = disks - conf->max_degraded;
3872 new_sector = raid5_compute_sector(conf, logical_sector,
3875 pr_debug("raid5: make_request, sector %llu logical %llu\n",
3876 (unsigned long long)new_sector,
3877 (unsigned long long)logical_sector);
3879 sh = get_active_stripe(conf, new_sector, previous,
3880 (bi->bi_rw&RWA_MASK));
3882 if (unlikely(previous)) {
3883 /* expansion might have moved on while waiting for a
3884 * stripe, so we must do the range check again.
3885 * Expansion could still move past after this
3886 * test, but as we are holding a reference to
3887 * 'sh', we know that if that happens,
3888 * STRIPE_EXPANDING will get set and the expansion
3889 * won't proceed until we finish with the stripe.
3892 spin_lock_irq(&conf->device_lock);
3893 if (mddev->delta_disks < 0
3894 ? logical_sector >= conf->reshape_progress
3895 : logical_sector < conf->reshape_progress)
3896 /* mismatch, need to try again */
3898 spin_unlock_irq(&conf->device_lock);
3904 /* FIXME what if we get a false positive because these
3905 * are being updated.
3907 if (logical_sector >= mddev->suspend_lo &&
3908 logical_sector < mddev->suspend_hi) {
3914 if (test_bit(STRIPE_EXPANDING, &sh->state) ||
3915 !add_stripe_bio(sh, bi, dd_idx, (bi->bi_rw&RW_MASK))) {
3916 /* Stripe is busy expanding or
3917 * add failed due to overlap. Flush everything
3920 raid5_unplug_device(mddev->queue);
3925 finish_wait(&conf->wait_for_overlap, &w);
3926 set_bit(STRIPE_HANDLE, &sh->state);
3927 clear_bit(STRIPE_DELAYED, &sh->state);
3930 /* cannot get stripe for read-ahead, just give-up */
3931 clear_bit(BIO_UPTODATE, &bi->bi_flags);
3932 finish_wait(&conf->wait_for_overlap, &w);
3937 spin_lock_irq(&conf->device_lock);
3938 remaining = raid5_dec_bi_phys_segments(bi);
3939 spin_unlock_irq(&conf->device_lock);
3940 if (remaining == 0) {
3943 md_write_end(mddev);
3950 static sector_t raid5_size(mddev_t *mddev, sector_t sectors, int raid_disks);
3952 static sector_t reshape_request(mddev_t *mddev, sector_t sector_nr, int *skipped)
3954 /* reshaping is quite different to recovery/resync so it is
3955 * handled quite separately ... here.
3957 * On each call to sync_request, we gather one chunk worth of
3958 * destination stripes and flag them as expanding.
3959 * Then we find all the source stripes and request reads.
3960 * As the reads complete, handle_stripe will copy the data
3961 * into the destination stripe and release that stripe.
3963 raid5_conf_t *conf = (raid5_conf_t *) mddev->private;
3964 struct stripe_head *sh;
3965 sector_t first_sector, last_sector;
3966 int raid_disks = conf->previous_raid_disks;
3967 int data_disks = raid_disks - conf->max_degraded;
3968 int new_data_disks = conf->raid_disks - conf->max_degraded;
3971 sector_t writepos, readpos, safepos;
3972 sector_t stripe_addr;
3973 int reshape_sectors;
3974 struct list_head stripes;
3976 if (sector_nr == 0) {
3977 /* If restarting in the middle, skip the initial sectors */
3978 if (mddev->delta_disks < 0 &&
3979 conf->reshape_progress < raid5_size(mddev, 0, 0)) {
3980 sector_nr = raid5_size(mddev, 0, 0)
3981 - conf->reshape_progress;
3982 } else if (mddev->delta_disks > 0 &&
3983 conf->reshape_progress > 0)
3984 sector_nr = conf->reshape_progress;
3985 sector_div(sector_nr, new_data_disks);
3992 /* We need to process a full chunk at a time.
3993 * If old and new chunk sizes differ, we need to process the
3996 if (mddev->new_chunk > mddev->chunk_size)
3997 reshape_sectors = mddev->new_chunk / 512;
3999 reshape_sectors = mddev->chunk_size / 512;
4001 /* we update the metadata when there is more than 3Meg
4002 * in the block range (that is rather arbitrary, should
4003 * probably be time based) or when the data about to be
4004 * copied would over-write the source of the data at
4005 * the front of the range.
4006 * i.e. one new_stripe along from reshape_progress new_maps
4007 * to after where reshape_safe old_maps to
4009 writepos = conf->reshape_progress;
4010 sector_div(writepos, new_data_disks);
4011 readpos = conf->reshape_progress;
4012 sector_div(readpos, data_disks);
4013 safepos = conf->reshape_safe;
4014 sector_div(safepos, data_disks);
4015 if (mddev->delta_disks < 0) {
4016 writepos -= reshape_sectors;
4017 readpos += reshape_sectors;
4018 safepos += reshape_sectors;
4020 writepos += reshape_sectors;
4021 readpos -= reshape_sectors;
4022 safepos -= reshape_sectors;
4025 /* 'writepos' is the most advanced device address we might write.
4026 * 'readpos' is the least advanced device address we might read.
4027 * 'safepos' is the least address recorded in the metadata as having
4029 * If 'readpos' is behind 'writepos', then there is no way that we can
4030 * ensure safety in the face of a crash - that must be done by userspace
4031 * making a backup of the data. So in that case there is no particular
4032 * rush to update metadata.
4033 * Otherwise if 'safepos' is behind 'writepos', then we really need to
4034 * update the metadata to advance 'safepos' to match 'readpos' so that
4035 * we can be safe in the event of a crash.
4036 * So we insist on updating metadata if safepos is behind writepos and
4037 * readpos is beyond writepos.
4038 * In any case, update the metadata every 10 seconds.
4039 * Maybe that number should be configurable, but I'm not sure it is
4040 * worth it.... maybe it could be a multiple of safemode_delay???
4042 if ((mddev->delta_disks < 0
4043 ? (safepos > writepos && readpos < writepos)
4044 : (safepos < writepos && readpos > writepos)) ||
4045 time_after(jiffies, conf->reshape_checkpoint + 10*HZ)) {
4046 /* Cannot proceed until we've updated the superblock... */
4047 wait_event(conf->wait_for_overlap,
4048 atomic_read(&conf->reshape_stripes)==0);
4049 mddev->reshape_position = conf->reshape_progress;
4050 conf->reshape_checkpoint = jiffies;
4051 set_bit(MD_CHANGE_DEVS, &mddev->flags);
4052 md_wakeup_thread(mddev->thread);
4053 wait_event(mddev->sb_wait, mddev->flags == 0 ||
4054 kthread_should_stop());
4055 spin_lock_irq(&conf->device_lock);
4056 conf->reshape_safe = mddev->reshape_position;
4057 spin_unlock_irq(&conf->device_lock);
4058 wake_up(&conf->wait_for_overlap);
4061 if (mddev->delta_disks < 0) {
4062 BUG_ON(conf->reshape_progress == 0);
4063 stripe_addr = writepos;
4064 BUG_ON((mddev->dev_sectors &
4065 ~((sector_t)reshape_sectors - 1))
4066 - reshape_sectors - stripe_addr
4069 BUG_ON(writepos != sector_nr + reshape_sectors);
4070 stripe_addr = sector_nr;
4072 INIT_LIST_HEAD(&stripes);
4073 for (i = 0; i < reshape_sectors; i += STRIPE_SECTORS) {
4076 sh = get_active_stripe(conf, stripe_addr+i, 0, 0);
4077 set_bit(STRIPE_EXPANDING, &sh->state);
4078 atomic_inc(&conf->reshape_stripes);
4079 /* If any of this stripe is beyond the end of the old
4080 * array, then we need to zero those blocks
4082 for (j=sh->disks; j--;) {
4084 if (j == sh->pd_idx)
4086 if (conf->level == 6 &&
4089 s = compute_blocknr(sh, j, 0);
4090 if (s < raid5_size(mddev, 0, 0)) {
4094 memset(page_address(sh->dev[j].page), 0, STRIPE_SIZE);
4095 set_bit(R5_Expanded, &sh->dev[j].flags);
4096 set_bit(R5_UPTODATE, &sh->dev[j].flags);
4099 set_bit(STRIPE_EXPAND_READY, &sh->state);
4100 set_bit(STRIPE_HANDLE, &sh->state);
4102 list_add(&sh->lru, &stripes);
4104 spin_lock_irq(&conf->device_lock);
4105 if (mddev->delta_disks < 0)
4106 conf->reshape_progress -= reshape_sectors * new_data_disks;
4108 conf->reshape_progress += reshape_sectors * new_data_disks;
4109 spin_unlock_irq(&conf->device_lock);
4110 /* Ok, those stripe are ready. We can start scheduling
4111 * reads on the source stripes.
4112 * The source stripes are determined by mapping the first and last
4113 * block on the destination stripes.
4116 raid5_compute_sector(conf, stripe_addr*(new_data_disks),
4119 raid5_compute_sector(conf, ((stripe_addr+conf->chunk_size/512)
4120 *(new_data_disks) - 1),
4122 if (last_sector >= mddev->dev_sectors)
4123 last_sector = mddev->dev_sectors - 1;
4124 while (first_sector <= last_sector) {
4125 sh = get_active_stripe(conf, first_sector, 1, 0);
4126 set_bit(STRIPE_EXPAND_SOURCE, &sh->state);
4127 set_bit(STRIPE_HANDLE, &sh->state);
4129 first_sector += STRIPE_SECTORS;
4131 /* Now that the sources are clearly marked, we can release
4132 * the destination stripes
4134 while (!list_empty(&stripes)) {
4135 sh = list_entry(stripes.next, struct stripe_head, lru);
4136 list_del_init(&sh->lru);
4139 /* If this takes us to the resync_max point where we have to pause,
4140 * then we need to write out the superblock.
4142 sector_nr += reshape_sectors;
4143 if (sector_nr >= mddev->resync_max) {
4144 /* Cannot proceed until we've updated the superblock... */
4145 wait_event(conf->wait_for_overlap,
4146 atomic_read(&conf->reshape_stripes) == 0);
4147 mddev->reshape_position = conf->reshape_progress;
4148 conf->reshape_checkpoint = jiffies;
4149 set_bit(MD_CHANGE_DEVS, &mddev->flags);
4150 md_wakeup_thread(mddev->thread);
4151 wait_event(mddev->sb_wait,
4152 !test_bit(MD_CHANGE_DEVS, &mddev->flags)
4153 || kthread_should_stop());
4154 spin_lock_irq(&conf->device_lock);
4155 conf->reshape_safe = mddev->reshape_position;
4156 spin_unlock_irq(&conf->device_lock);
4157 wake_up(&conf->wait_for_overlap);
4159 return reshape_sectors;
4162 /* FIXME go_faster isn't used */
4163 static inline sector_t sync_request(mddev_t *mddev, sector_t sector_nr, int *skipped, int go_faster)
4165 raid5_conf_t *conf = (raid5_conf_t *) mddev->private;
4166 struct stripe_head *sh;
4167 sector_t max_sector = mddev->dev_sectors;
4169 int still_degraded = 0;
4172 if (sector_nr >= max_sector) {
4173 /* just being told to finish up .. nothing much to do */
4174 unplug_slaves(mddev);
4176 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery)) {
4181 if (mddev->curr_resync < max_sector) /* aborted */
4182 bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
4184 else /* completed sync */
4186 bitmap_close_sync(mddev->bitmap);
4191 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
4192 return reshape_request(mddev, sector_nr, skipped);
4194 /* No need to check resync_max as we never do more than one
4195 * stripe, and as resync_max will always be on a chunk boundary,
4196 * if the check in md_do_sync didn't fire, there is no chance
4197 * of overstepping resync_max here
4200 /* if there is too many failed drives and we are trying
4201 * to resync, then assert that we are finished, because there is
4202 * nothing we can do.
4204 if (mddev->degraded >= conf->max_degraded &&
4205 test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
4206 sector_t rv = mddev->dev_sectors - sector_nr;
4210 if (!bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
4211 !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
4212 !conf->fullsync && sync_blocks >= STRIPE_SECTORS) {
4213 /* we can skip this block, and probably more */
4214 sync_blocks /= STRIPE_SECTORS;
4216 return sync_blocks * STRIPE_SECTORS; /* keep things rounded to whole stripes */
4220 bitmap_cond_end_sync(mddev->bitmap, sector_nr);
4222 sh = get_active_stripe(conf, sector_nr, 0, 1);
4224 sh = get_active_stripe(conf, sector_nr, 0, 0);
4225 /* make sure we don't swamp the stripe cache if someone else
4226 * is trying to get access
4228 schedule_timeout_uninterruptible(1);
4230 /* Need to check if array will still be degraded after recovery/resync
4231 * We don't need to check the 'failed' flag as when that gets set,
4234 for (i=0; i<mddev->raid_disks; i++)
4235 if (conf->disks[i].rdev == NULL)
4238 bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, still_degraded);
4240 spin_lock(&sh->lock);
4241 set_bit(STRIPE_SYNCING, &sh->state);
4242 clear_bit(STRIPE_INSYNC, &sh->state);
4243 spin_unlock(&sh->lock);
4245 /* wait for any blocked device to be handled */
4246 while (unlikely(!handle_stripe(sh)))
4250 return STRIPE_SECTORS;
4253 static int retry_aligned_read(raid5_conf_t *conf, struct bio *raid_bio)
4255 /* We may not be able to submit a whole bio at once as there
4256 * may not be enough stripe_heads available.
4257 * We cannot pre-allocate enough stripe_heads as we may need
4258 * more than exist in the cache (if we allow ever large chunks).
4259 * So we do one stripe head at a time and record in
4260 * ->bi_hw_segments how many have been done.
4262 * We *know* that this entire raid_bio is in one chunk, so
4263 * it will be only one 'dd_idx' and only need one call to raid5_compute_sector.
4265 struct stripe_head *sh;
4267 sector_t sector, logical_sector, last_sector;
4272 logical_sector = raid_bio->bi_sector & ~((sector_t)STRIPE_SECTORS-1);
4273 sector = raid5_compute_sector(conf, logical_sector,
4275 last_sector = raid_bio->bi_sector + (raid_bio->bi_size>>9);
4277 for (; logical_sector < last_sector;
4278 logical_sector += STRIPE_SECTORS,
4279 sector += STRIPE_SECTORS,
4282 if (scnt < raid5_bi_hw_segments(raid_bio))
4283 /* already done this stripe */
4286 sh = get_active_stripe(conf, sector, 0, 1);
4289 /* failed to get a stripe - must wait */
4290 raid5_set_bi_hw_segments(raid_bio, scnt);
4291 conf->retry_read_aligned = raid_bio;
4295 set_bit(R5_ReadError, &sh->dev[dd_idx].flags);
4296 if (!add_stripe_bio(sh, raid_bio, dd_idx, 0)) {
4298 raid5_set_bi_hw_segments(raid_bio, scnt);
4299 conf->retry_read_aligned = raid_bio;
4307 spin_lock_irq(&conf->device_lock);
4308 remaining = raid5_dec_bi_phys_segments(raid_bio);
4309 spin_unlock_irq(&conf->device_lock);
4311 bio_endio(raid_bio, 0);
4312 if (atomic_dec_and_test(&conf->active_aligned_reads))
4313 wake_up(&conf->wait_for_stripe);
4320 * This is our raid5 kernel thread.
4322 * We scan the hash table for stripes which can be handled now.
4323 * During the scan, completed stripes are saved for us by the interrupt
4324 * handler, so that they will not have to wait for our next wakeup.
4326 static void raid5d(mddev_t *mddev)
4328 struct stripe_head *sh;
4329 raid5_conf_t *conf = mddev_to_conf(mddev);
4332 pr_debug("+++ raid5d active\n");
4334 md_check_recovery(mddev);
4337 spin_lock_irq(&conf->device_lock);
4341 if (conf->seq_flush != conf->seq_write) {
4342 int seq = conf->seq_flush;
4343 spin_unlock_irq(&conf->device_lock);
4344 bitmap_unplug(mddev->bitmap);
4345 spin_lock_irq(&conf->device_lock);
4346 conf->seq_write = seq;
4347 activate_bit_delay(conf);
4350 while ((bio = remove_bio_from_retry(conf))) {
4352 spin_unlock_irq(&conf->device_lock);
4353 ok = retry_aligned_read(conf, bio);
4354 spin_lock_irq(&conf->device_lock);
4360 sh = __get_priority_stripe(conf);
4364 spin_unlock_irq(&conf->device_lock);
4370 spin_lock_irq(&conf->device_lock);
4372 pr_debug("%d stripes handled\n", handled);
4374 spin_unlock_irq(&conf->device_lock);
4376 async_tx_issue_pending_all();
4377 unplug_slaves(mddev);
4379 pr_debug("--- raid5d inactive\n");
4383 raid5_show_stripe_cache_size(mddev_t *mddev, char *page)
4385 raid5_conf_t *conf = mddev_to_conf(mddev);
4387 return sprintf(page, "%d\n", conf->max_nr_stripes);
4393 raid5_store_stripe_cache_size(mddev_t *mddev, const char *page, size_t len)
4395 raid5_conf_t *conf = mddev_to_conf(mddev);
4399 if (len >= PAGE_SIZE)
4404 if (strict_strtoul(page, 10, &new))
4406 if (new <= 16 || new > 32768)
4408 while (new < conf->max_nr_stripes) {
4409 if (drop_one_stripe(conf))
4410 conf->max_nr_stripes--;
4414 err = md_allow_write(mddev);
4417 while (new > conf->max_nr_stripes) {
4418 if (grow_one_stripe(conf))
4419 conf->max_nr_stripes++;
4425 static struct md_sysfs_entry
4426 raid5_stripecache_size = __ATTR(stripe_cache_size, S_IRUGO | S_IWUSR,
4427 raid5_show_stripe_cache_size,
4428 raid5_store_stripe_cache_size);
4431 raid5_show_preread_threshold(mddev_t *mddev, char *page)
4433 raid5_conf_t *conf = mddev_to_conf(mddev);
4435 return sprintf(page, "%d\n", conf->bypass_threshold);
4441 raid5_store_preread_threshold(mddev_t *mddev, const char *page, size_t len)
4443 raid5_conf_t *conf = mddev_to_conf(mddev);
4445 if (len >= PAGE_SIZE)
4450 if (strict_strtoul(page, 10, &new))
4452 if (new > conf->max_nr_stripes)
4454 conf->bypass_threshold = new;
4458 static struct md_sysfs_entry
4459 raid5_preread_bypass_threshold = __ATTR(preread_bypass_threshold,
4461 raid5_show_preread_threshold,
4462 raid5_store_preread_threshold);
4465 stripe_cache_active_show(mddev_t *mddev, char *page)
4467 raid5_conf_t *conf = mddev_to_conf(mddev);
4469 return sprintf(page, "%d\n", atomic_read(&conf->active_stripes));
4474 static struct md_sysfs_entry
4475 raid5_stripecache_active = __ATTR_RO(stripe_cache_active);
4477 static struct attribute *raid5_attrs[] = {
4478 &raid5_stripecache_size.attr,
4479 &raid5_stripecache_active.attr,
4480 &raid5_preread_bypass_threshold.attr,
4483 static struct attribute_group raid5_attrs_group = {
4485 .attrs = raid5_attrs,
4489 raid5_size(mddev_t *mddev, sector_t sectors, int raid_disks)
4491 raid5_conf_t *conf = mddev_to_conf(mddev);
4494 sectors = mddev->dev_sectors;
4496 /* size is defined by the smallest of previous and new size */
4497 if (conf->raid_disks < conf->previous_raid_disks)
4498 raid_disks = conf->raid_disks;
4500 raid_disks = conf->previous_raid_disks;
4503 sectors &= ~((sector_t)mddev->chunk_size/512 - 1);
4504 sectors &= ~((sector_t)mddev->new_chunk/512 - 1);
4505 return sectors * (raid_disks - conf->max_degraded);
4508 static void raid5_free_percpu(raid5_conf_t *conf)
4510 struct raid5_percpu *percpu;
4517 for_each_possible_cpu(cpu) {
4518 percpu = per_cpu_ptr(conf->percpu, cpu);
4519 safe_put_page(percpu->spare_page);
4520 kfree(percpu->scribble);
4522 #ifdef CONFIG_HOTPLUG_CPU
4523 unregister_cpu_notifier(&conf->cpu_notify);
4527 free_percpu(conf->percpu);
4530 static void free_conf(raid5_conf_t *conf)
4532 shrink_stripes(conf);
4533 raid5_free_percpu(conf);
4535 kfree(conf->stripe_hashtbl);
4539 #ifdef CONFIG_HOTPLUG_CPU
4540 static int raid456_cpu_notify(struct notifier_block *nfb, unsigned long action,
4543 raid5_conf_t *conf = container_of(nfb, raid5_conf_t, cpu_notify);
4544 long cpu = (long)hcpu;
4545 struct raid5_percpu *percpu = per_cpu_ptr(conf->percpu, cpu);
4548 case CPU_UP_PREPARE:
4549 case CPU_UP_PREPARE_FROZEN:
4550 if (conf->level == 6 && !percpu->spare_page)
4551 percpu->spare_page = alloc_page(GFP_KERNEL);
4552 if (!percpu->scribble)
4553 percpu->scribble = kmalloc(conf->scribble_len, GFP_KERNEL);
4555 if (!percpu->scribble ||
4556 (conf->level == 6 && !percpu->spare_page)) {
4557 safe_put_page(percpu->spare_page);
4558 kfree(percpu->scribble);
4559 pr_err("%s: failed memory allocation for cpu%ld\n",
4565 case CPU_DEAD_FROZEN:
4566 safe_put_page(percpu->spare_page);
4567 kfree(percpu->scribble);
4568 percpu->spare_page = NULL;
4569 percpu->scribble = NULL;
4578 static int raid5_alloc_percpu(raid5_conf_t *conf)
4581 struct page *spare_page;
4582 struct raid5_percpu *allcpus;
4586 allcpus = alloc_percpu(struct raid5_percpu);
4589 conf->percpu = allcpus;
4593 for_each_present_cpu(cpu) {
4594 if (conf->level == 6) {
4595 spare_page = alloc_page(GFP_KERNEL);
4600 per_cpu_ptr(conf->percpu, cpu)->spare_page = spare_page;
4602 scribble = kmalloc(scribble_len(conf->raid_disks), GFP_KERNEL);
4607 per_cpu_ptr(conf->percpu, cpu)->scribble = scribble;
4609 #ifdef CONFIG_HOTPLUG_CPU
4610 conf->cpu_notify.notifier_call = raid456_cpu_notify;
4611 conf->cpu_notify.priority = 0;
4613 err = register_cpu_notifier(&conf->cpu_notify);
4620 static raid5_conf_t *setup_conf(mddev_t *mddev)
4623 int raid_disk, memory;
4625 struct disk_info *disk;
4627 if (mddev->new_level != 5
4628 && mddev->new_level != 4
4629 && mddev->new_level != 6) {
4630 printk(KERN_ERR "raid5: %s: raid level not set to 4/5/6 (%d)\n",
4631 mdname(mddev), mddev->new_level);
4632 return ERR_PTR(-EIO);
4634 if ((mddev->new_level == 5
4635 && !algorithm_valid_raid5(mddev->new_layout)) ||
4636 (mddev->new_level == 6
4637 && !algorithm_valid_raid6(mddev->new_layout))) {
4638 printk(KERN_ERR "raid5: %s: layout %d not supported\n",
4639 mdname(mddev), mddev->new_layout);
4640 return ERR_PTR(-EIO);
4642 if (mddev->new_level == 6 && mddev->raid_disks < 4) {
4643 printk(KERN_ERR "raid6: not enough configured devices for %s (%d, minimum 4)\n",
4644 mdname(mddev), mddev->raid_disks);
4645 return ERR_PTR(-EINVAL);
4648 if (!mddev->new_chunk || mddev->new_chunk % PAGE_SIZE) {
4649 printk(KERN_ERR "raid5: invalid chunk size %d for %s\n",
4650 mddev->new_chunk, mdname(mddev));
4651 return ERR_PTR(-EINVAL);
4654 conf = kzalloc(sizeof(raid5_conf_t), GFP_KERNEL);
4658 conf->raid_disks = mddev->raid_disks;
4659 conf->scribble_len = scribble_len(conf->raid_disks);
4660 if (mddev->reshape_position == MaxSector)
4661 conf->previous_raid_disks = mddev->raid_disks;
4663 conf->previous_raid_disks = mddev->raid_disks - mddev->delta_disks;
4665 conf->disks = kzalloc(conf->raid_disks * sizeof(struct disk_info),
4670 conf->mddev = mddev;
4672 if ((conf->stripe_hashtbl = kzalloc(PAGE_SIZE, GFP_KERNEL)) == NULL)
4675 conf->level = mddev->new_level;
4676 if (raid5_alloc_percpu(conf) != 0)
4679 spin_lock_init(&conf->device_lock);
4680 init_waitqueue_head(&conf->wait_for_stripe);
4681 init_waitqueue_head(&conf->wait_for_overlap);
4682 INIT_LIST_HEAD(&conf->handle_list);
4683 INIT_LIST_HEAD(&conf->hold_list);
4684 INIT_LIST_HEAD(&conf->delayed_list);
4685 INIT_LIST_HEAD(&conf->bitmap_list);
4686 INIT_LIST_HEAD(&conf->inactive_list);
4687 atomic_set(&conf->active_stripes, 0);
4688 atomic_set(&conf->preread_active_stripes, 0);
4689 atomic_set(&conf->active_aligned_reads, 0);
4690 conf->bypass_threshold = BYPASS_THRESHOLD;
4692 pr_debug("raid5: run(%s) called.\n", mdname(mddev));
4694 list_for_each_entry(rdev, &mddev->disks, same_set) {
4695 raid_disk = rdev->raid_disk;
4696 if (raid_disk >= conf->raid_disks
4699 disk = conf->disks + raid_disk;
4703 if (test_bit(In_sync, &rdev->flags)) {
4704 char b[BDEVNAME_SIZE];
4705 printk(KERN_INFO "raid5: device %s operational as raid"
4706 " disk %d\n", bdevname(rdev->bdev,b),
4709 /* Cannot rely on bitmap to complete recovery */
4713 conf->chunk_size = mddev->new_chunk;
4714 if (conf->level == 6)
4715 conf->max_degraded = 2;
4717 conf->max_degraded = 1;
4718 conf->algorithm = mddev->new_layout;
4719 conf->max_nr_stripes = NR_STRIPES;
4720 conf->reshape_progress = mddev->reshape_position;
4721 if (conf->reshape_progress != MaxSector) {
4722 conf->prev_chunk = mddev->chunk_size;
4723 conf->prev_algo = mddev->layout;
4726 memory = conf->max_nr_stripes * (sizeof(struct stripe_head) +
4727 conf->raid_disks * ((sizeof(struct bio) + PAGE_SIZE))) / 1024;
4728 if (grow_stripes(conf, conf->max_nr_stripes)) {
4730 "raid5: couldn't allocate %dkB for buffers\n", memory);
4733 printk(KERN_INFO "raid5: allocated %dkB for %s\n",
4734 memory, mdname(mddev));
4736 conf->thread = md_register_thread(raid5d, mddev, "%s_raid5");
4737 if (!conf->thread) {
4739 "raid5: couldn't allocate thread for %s\n",
4749 return ERR_PTR(-EIO);
4751 return ERR_PTR(-ENOMEM);
4754 static int run(mddev_t *mddev)
4757 int working_disks = 0;
4760 if (mddev->reshape_position != MaxSector) {
4761 /* Check that we can continue the reshape.
4762 * Currently only disks can change, it must
4763 * increase, and we must be past the point where
4764 * a stripe over-writes itself
4766 sector_t here_new, here_old;
4768 int max_degraded = (mddev->level == 6 ? 2 : 1);
4770 if (mddev->new_level != mddev->level) {
4771 printk(KERN_ERR "raid5: %s: unsupported reshape "
4772 "required - aborting.\n",
4776 old_disks = mddev->raid_disks - mddev->delta_disks;
4777 /* reshape_position must be on a new-stripe boundary, and one
4778 * further up in new geometry must map after here in old
4781 here_new = mddev->reshape_position;
4782 if (sector_div(here_new, (mddev->new_chunk>>9)*
4783 (mddev->raid_disks - max_degraded))) {
4784 printk(KERN_ERR "raid5: reshape_position not "
4785 "on a stripe boundary\n");
4788 /* here_new is the stripe we will write to */
4789 here_old = mddev->reshape_position;
4790 sector_div(here_old, (mddev->chunk_size>>9)*
4791 (old_disks-max_degraded));
4792 /* here_old is the first stripe that we might need to read
4794 if (here_new >= here_old) {
4795 /* Reading from the same stripe as writing to - bad */
4796 printk(KERN_ERR "raid5: reshape_position too early for "
4797 "auto-recovery - aborting.\n");
4800 printk(KERN_INFO "raid5: reshape will continue\n");
4801 /* OK, we should be able to continue; */
4803 BUG_ON(mddev->level != mddev->new_level);
4804 BUG_ON(mddev->layout != mddev->new_layout);
4805 BUG_ON(mddev->chunk_size != mddev->new_chunk);
4806 BUG_ON(mddev->delta_disks != 0);
4809 if (mddev->private == NULL)
4810 conf = setup_conf(mddev);
4812 conf = mddev->private;
4815 return PTR_ERR(conf);
4817 mddev->thread = conf->thread;
4818 conf->thread = NULL;
4819 mddev->private = conf;
4822 * 0 for a fully functional array, 1 or 2 for a degraded array.
4824 list_for_each_entry(rdev, &mddev->disks, same_set)
4825 if (rdev->raid_disk >= 0 &&
4826 test_bit(In_sync, &rdev->flags))
4829 mddev->degraded = conf->raid_disks - working_disks;
4831 if (mddev->degraded > conf->max_degraded) {
4832 printk(KERN_ERR "raid5: not enough operational devices for %s"
4833 " (%d/%d failed)\n",
4834 mdname(mddev), mddev->degraded, conf->raid_disks);
4838 /* device size must be a multiple of chunk size */
4839 mddev->dev_sectors &= ~(mddev->chunk_size / 512 - 1);
4840 mddev->resync_max_sectors = mddev->dev_sectors;
4842 if (mddev->degraded > 0 &&
4843 mddev->recovery_cp != MaxSector) {
4844 if (mddev->ok_start_degraded)
4846 "raid5: starting dirty degraded array: %s"
4847 "- data corruption possible.\n",
4851 "raid5: cannot start dirty degraded array for %s\n",
4857 if (mddev->degraded == 0)
4858 printk("raid5: raid level %d set %s active with %d out of %d"
4859 " devices, algorithm %d\n", conf->level, mdname(mddev),
4860 mddev->raid_disks-mddev->degraded, mddev->raid_disks,
4863 printk(KERN_ALERT "raid5: raid level %d set %s active with %d"
4864 " out of %d devices, algorithm %d\n", conf->level,
4865 mdname(mddev), mddev->raid_disks - mddev->degraded,
4866 mddev->raid_disks, mddev->new_layout);
4868 print_raid5_conf(conf);
4870 if (conf->reshape_progress != MaxSector) {
4871 printk("...ok start reshape thread\n");
4872 conf->reshape_safe = conf->reshape_progress;
4873 atomic_set(&conf->reshape_stripes, 0);
4874 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
4875 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
4876 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
4877 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
4878 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
4882 /* read-ahead size must cover two whole stripes, which is
4883 * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
4886 int data_disks = conf->previous_raid_disks - conf->max_degraded;
4887 int stripe = data_disks *
4888 (mddev->chunk_size / PAGE_SIZE);
4889 if (mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
4890 mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
4893 /* Ok, everything is just fine now */
4894 if (sysfs_create_group(&mddev->kobj, &raid5_attrs_group))
4896 "raid5: failed to create sysfs attributes for %s\n",
4899 mddev->queue->queue_lock = &conf->device_lock;
4901 mddev->queue->unplug_fn = raid5_unplug_device;
4902 mddev->queue->backing_dev_info.congested_data = mddev;
4903 mddev->queue->backing_dev_info.congested_fn = raid5_congested;
4905 md_set_array_sectors(mddev, raid5_size(mddev, 0, 0));
4907 blk_queue_merge_bvec(mddev->queue, raid5_mergeable_bvec);
4911 md_unregister_thread(mddev->thread);
4912 mddev->thread = NULL;
4914 print_raid5_conf(conf);
4917 mddev->private = NULL;
4918 printk(KERN_ALERT "raid5: failed to run raid set %s\n", mdname(mddev));
4924 static int stop(mddev_t *mddev)
4926 raid5_conf_t *conf = (raid5_conf_t *) mddev->private;
4928 md_unregister_thread(mddev->thread);
4929 mddev->thread = NULL;
4930 mddev->queue->backing_dev_info.congested_fn = NULL;
4931 blk_sync_queue(mddev->queue); /* the unplug fn references 'conf'*/
4932 sysfs_remove_group(&mddev->kobj, &raid5_attrs_group);
4934 mddev->private = NULL;
4939 static void print_sh(struct seq_file *seq, struct stripe_head *sh)
4943 seq_printf(seq, "sh %llu, pd_idx %d, state %ld.\n",
4944 (unsigned long long)sh->sector, sh->pd_idx, sh->state);
4945 seq_printf(seq, "sh %llu, count %d.\n",
4946 (unsigned long long)sh->sector, atomic_read(&sh->count));
4947 seq_printf(seq, "sh %llu, ", (unsigned long long)sh->sector);
4948 for (i = 0; i < sh->disks; i++) {
4949 seq_printf(seq, "(cache%d: %p %ld) ",
4950 i, sh->dev[i].page, sh->dev[i].flags);
4952 seq_printf(seq, "\n");
4955 static void printall(struct seq_file *seq, raid5_conf_t *conf)
4957 struct stripe_head *sh;
4958 struct hlist_node *hn;
4961 spin_lock_irq(&conf->device_lock);
4962 for (i = 0; i < NR_HASH; i++) {
4963 hlist_for_each_entry(sh, hn, &conf->stripe_hashtbl[i], hash) {
4964 if (sh->raid_conf != conf)
4969 spin_unlock_irq(&conf->device_lock);
4973 static void status(struct seq_file *seq, mddev_t *mddev)
4975 raid5_conf_t *conf = (raid5_conf_t *) mddev->private;
4978 seq_printf (seq, " level %d, %dk chunk, algorithm %d", mddev->level, mddev->chunk_size >> 10, mddev->layout);
4979 seq_printf (seq, " [%d/%d] [", conf->raid_disks, conf->raid_disks - mddev->degraded);
4980 for (i = 0; i < conf->raid_disks; i++)
4981 seq_printf (seq, "%s",
4982 conf->disks[i].rdev &&
4983 test_bit(In_sync, &conf->disks[i].rdev->flags) ? "U" : "_");
4984 seq_printf (seq, "]");
4986 seq_printf (seq, "\n");
4987 printall(seq, conf);
4991 static void print_raid5_conf (raid5_conf_t *conf)
4994 struct disk_info *tmp;
4996 printk("RAID5 conf printout:\n");
4998 printk("(conf==NULL)\n");
5001 printk(" --- rd:%d wd:%d\n", conf->raid_disks,
5002 conf->raid_disks - conf->mddev->degraded);
5004 for (i = 0; i < conf->raid_disks; i++) {
5005 char b[BDEVNAME_SIZE];
5006 tmp = conf->disks + i;
5008 printk(" disk %d, o:%d, dev:%s\n",
5009 i, !test_bit(Faulty, &tmp->rdev->flags),
5010 bdevname(tmp->rdev->bdev,b));
5014 static int raid5_spare_active(mddev_t *mddev)
5017 raid5_conf_t *conf = mddev->private;
5018 struct disk_info *tmp;
5020 for (i = 0; i < conf->raid_disks; i++) {
5021 tmp = conf->disks + i;
5023 && !test_bit(Faulty, &tmp->rdev->flags)
5024 && !test_and_set_bit(In_sync, &tmp->rdev->flags)) {
5025 unsigned long flags;
5026 spin_lock_irqsave(&conf->device_lock, flags);
5028 spin_unlock_irqrestore(&conf->device_lock, flags);
5031 print_raid5_conf(conf);
5035 static int raid5_remove_disk(mddev_t *mddev, int number)
5037 raid5_conf_t *conf = mddev->private;
5040 struct disk_info *p = conf->disks + number;
5042 print_raid5_conf(conf);
5045 if (number >= conf->raid_disks &&
5046 conf->reshape_progress == MaxSector)
5047 clear_bit(In_sync, &rdev->flags);
5049 if (test_bit(In_sync, &rdev->flags) ||
5050 atomic_read(&rdev->nr_pending)) {
5054 /* Only remove non-faulty devices if recovery
5057 if (!test_bit(Faulty, &rdev->flags) &&
5058 mddev->degraded <= conf->max_degraded &&
5059 number < conf->raid_disks) {
5065 if (atomic_read(&rdev->nr_pending)) {
5066 /* lost the race, try later */
5073 print_raid5_conf(conf);
5077 static int raid5_add_disk(mddev_t *mddev, mdk_rdev_t *rdev)
5079 raid5_conf_t *conf = mddev->private;
5082 struct disk_info *p;
5084 int last = conf->raid_disks - 1;
5086 if (mddev->degraded > conf->max_degraded)
5087 /* no point adding a device */
5090 if (rdev->raid_disk >= 0)
5091 first = last = rdev->raid_disk;
5094 * find the disk ... but prefer rdev->saved_raid_disk
5097 if (rdev->saved_raid_disk >= 0 &&
5098 rdev->saved_raid_disk >= first &&
5099 conf->disks[rdev->saved_raid_disk].rdev == NULL)
5100 disk = rdev->saved_raid_disk;
5103 for ( ; disk <= last ; disk++)
5104 if ((p=conf->disks + disk)->rdev == NULL) {
5105 clear_bit(In_sync, &rdev->flags);
5106 rdev->raid_disk = disk;
5108 if (rdev->saved_raid_disk != disk)
5110 rcu_assign_pointer(p->rdev, rdev);
5113 print_raid5_conf(conf);
5117 static int raid5_resize(mddev_t *mddev, sector_t sectors)
5119 /* no resync is happening, and there is enough space
5120 * on all devices, so we can resize.
5121 * We need to make sure resync covers any new space.
5122 * If the array is shrinking we should possibly wait until
5123 * any io in the removed space completes, but it hardly seems
5126 sectors &= ~((sector_t)mddev->chunk_size/512 - 1);
5127 md_set_array_sectors(mddev, raid5_size(mddev, sectors,
5128 mddev->raid_disks));
5129 if (mddev->array_sectors >
5130 raid5_size(mddev, sectors, mddev->raid_disks))
5132 set_capacity(mddev->gendisk, mddev->array_sectors);
5134 if (sectors > mddev->dev_sectors && mddev->recovery_cp == MaxSector) {
5135 mddev->recovery_cp = mddev->dev_sectors;
5136 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
5138 mddev->dev_sectors = sectors;
5139 mddev->resync_max_sectors = sectors;
5143 static int raid5_check_reshape(mddev_t *mddev)
5145 raid5_conf_t *conf = mddev_to_conf(mddev);
5147 if (mddev->delta_disks == 0 &&
5148 mddev->new_layout == mddev->layout &&
5149 mddev->new_chunk == mddev->chunk_size)
5150 return -EINVAL; /* nothing to do */
5152 /* Cannot grow a bitmap yet */
5154 if (mddev->degraded > conf->max_degraded)
5156 if (mddev->delta_disks < 0) {
5157 /* We might be able to shrink, but the devices must
5158 * be made bigger first.
5159 * For raid6, 4 is the minimum size.
5160 * Otherwise 2 is the minimum
5163 if (mddev->level == 6)
5165 if (mddev->raid_disks + mddev->delta_disks < min)
5169 /* Can only proceed if there are plenty of stripe_heads.
5170 * We need a minimum of one full stripe,, and for sensible progress
5171 * it is best to have about 4 times that.
5172 * If we require 4 times, then the default 256 4K stripe_heads will
5173 * allow for chunk sizes up to 256K, which is probably OK.
5174 * If the chunk size is greater, user-space should request more
5175 * stripe_heads first.
5177 if ((mddev->chunk_size / STRIPE_SIZE) * 4 > conf->max_nr_stripes ||
5178 (mddev->new_chunk / STRIPE_SIZE) * 4 > conf->max_nr_stripes) {
5179 printk(KERN_WARNING "raid5: reshape: not enough stripes. Needed %lu\n",
5180 (max(mddev->chunk_size, mddev->new_chunk)
5185 return resize_stripes(conf, conf->raid_disks + mddev->delta_disks);
5188 static int raid5_start_reshape(mddev_t *mddev)
5190 raid5_conf_t *conf = mddev_to_conf(mddev);
5193 int added_devices = 0;
5194 unsigned long flags;
5196 if (test_bit(MD_RECOVERY_RUNNING, &mddev->recovery))
5199 list_for_each_entry(rdev, &mddev->disks, same_set)
5200 if (rdev->raid_disk < 0 &&
5201 !test_bit(Faulty, &rdev->flags))
5204 if (spares - mddev->degraded < mddev->delta_disks - conf->max_degraded)
5205 /* Not enough devices even to make a degraded array
5210 /* Refuse to reduce size of the array. Any reductions in
5211 * array size must be through explicit setting of array_size
5214 if (raid5_size(mddev, 0, conf->raid_disks + mddev->delta_disks)
5215 < mddev->array_sectors) {
5216 printk(KERN_ERR "md: %s: array size must be reduced "
5217 "before number of disks\n", mdname(mddev));
5221 atomic_set(&conf->reshape_stripes, 0);
5222 spin_lock_irq(&conf->device_lock);
5223 conf->previous_raid_disks = conf->raid_disks;
5224 conf->raid_disks += mddev->delta_disks;
5225 conf->prev_chunk = conf->chunk_size;
5226 conf->chunk_size = mddev->new_chunk;
5227 conf->prev_algo = conf->algorithm;
5228 conf->algorithm = mddev->new_layout;
5229 if (mddev->delta_disks < 0)
5230 conf->reshape_progress = raid5_size(mddev, 0, 0);
5232 conf->reshape_progress = 0;
5233 conf->reshape_safe = conf->reshape_progress;
5235 spin_unlock_irq(&conf->device_lock);
5237 /* Add some new drives, as many as will fit.
5238 * We know there are enough to make the newly sized array work.
5240 list_for_each_entry(rdev, &mddev->disks, same_set)
5241 if (rdev->raid_disk < 0 &&
5242 !test_bit(Faulty, &rdev->flags)) {
5243 if (raid5_add_disk(mddev, rdev) == 0) {
5245 set_bit(In_sync, &rdev->flags);
5247 rdev->recovery_offset = 0;
5248 sprintf(nm, "rd%d", rdev->raid_disk);
5249 if (sysfs_create_link(&mddev->kobj,
5252 "raid5: failed to create "
5253 " link %s for %s\n",
5259 if (mddev->delta_disks > 0) {
5260 spin_lock_irqsave(&conf->device_lock, flags);
5261 mddev->degraded = (conf->raid_disks - conf->previous_raid_disks)
5263 spin_unlock_irqrestore(&conf->device_lock, flags);
5265 mddev->raid_disks = conf->raid_disks;
5266 mddev->reshape_position = 0;
5267 set_bit(MD_CHANGE_DEVS, &mddev->flags);
5269 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
5270 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
5271 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
5272 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
5273 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
5275 if (!mddev->sync_thread) {
5276 mddev->recovery = 0;
5277 spin_lock_irq(&conf->device_lock);
5278 mddev->raid_disks = conf->raid_disks = conf->previous_raid_disks;
5279 conf->reshape_progress = MaxSector;
5280 spin_unlock_irq(&conf->device_lock);
5283 conf->reshape_checkpoint = jiffies;
5284 md_wakeup_thread(mddev->sync_thread);
5285 md_new_event(mddev);
5289 /* This is called from the reshape thread and should make any
5290 * changes needed in 'conf'
5292 static void end_reshape(raid5_conf_t *conf)
5295 if (!test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery)) {
5297 spin_lock_irq(&conf->device_lock);
5298 conf->previous_raid_disks = conf->raid_disks;
5299 conf->reshape_progress = MaxSector;
5300 spin_unlock_irq(&conf->device_lock);
5301 wake_up(&conf->wait_for_overlap);
5303 /* read-ahead size must cover two whole stripes, which is
5304 * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
5307 int data_disks = conf->raid_disks - conf->max_degraded;
5308 int stripe = data_disks * (conf->chunk_size
5310 if (conf->mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
5311 conf->mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
5316 /* This is called from the raid5d thread with mddev_lock held.
5317 * It makes config changes to the device.
5319 static void raid5_finish_reshape(mddev_t *mddev)
5321 struct block_device *bdev;
5322 raid5_conf_t *conf = mddev_to_conf(mddev);
5324 if (!test_bit(MD_RECOVERY_INTR, &mddev->recovery)) {
5326 if (mddev->delta_disks > 0) {
5327 md_set_array_sectors(mddev, raid5_size(mddev, 0, 0));
5328 set_capacity(mddev->gendisk, mddev->array_sectors);
5331 bdev = bdget_disk(mddev->gendisk, 0);
5333 mutex_lock(&bdev->bd_inode->i_mutex);
5334 i_size_write(bdev->bd_inode,
5335 (loff_t)mddev->array_sectors << 9);
5336 mutex_unlock(&bdev->bd_inode->i_mutex);
5341 mddev->degraded = conf->raid_disks;
5342 for (d = 0; d < conf->raid_disks ; d++)
5343 if (conf->disks[d].rdev &&
5345 &conf->disks[d].rdev->flags))
5347 for (d = conf->raid_disks ;
5348 d < conf->raid_disks - mddev->delta_disks;
5350 raid5_remove_disk(mddev, d);
5352 mddev->layout = conf->algorithm;
5353 mddev->chunk_size = conf->chunk_size;
5354 mddev->reshape_position = MaxSector;
5355 mddev->delta_disks = 0;
5359 static void raid5_quiesce(mddev_t *mddev, int state)
5361 raid5_conf_t *conf = mddev_to_conf(mddev);
5364 case 2: /* resume for a suspend */
5365 wake_up(&conf->wait_for_overlap);
5368 case 1: /* stop all writes */
5369 spin_lock_irq(&conf->device_lock);
5371 wait_event_lock_irq(conf->wait_for_stripe,
5372 atomic_read(&conf->active_stripes) == 0 &&
5373 atomic_read(&conf->active_aligned_reads) == 0,
5374 conf->device_lock, /* nothing */);
5375 spin_unlock_irq(&conf->device_lock);
5378 case 0: /* re-enable writes */
5379 spin_lock_irq(&conf->device_lock);
5381 wake_up(&conf->wait_for_stripe);
5382 wake_up(&conf->wait_for_overlap);
5383 spin_unlock_irq(&conf->device_lock);
5389 static void *raid5_takeover_raid1(mddev_t *mddev)
5393 if (mddev->raid_disks != 2 ||
5394 mddev->degraded > 1)
5395 return ERR_PTR(-EINVAL);
5397 /* Should check if there are write-behind devices? */
5399 chunksect = 64*2; /* 64K by default */
5401 /* The array must be an exact multiple of chunksize */
5402 while (chunksect && (mddev->array_sectors & (chunksect-1)))
5405 if ((chunksect<<9) < STRIPE_SIZE)
5406 /* array size does not allow a suitable chunk size */
5407 return ERR_PTR(-EINVAL);
5409 mddev->new_level = 5;
5410 mddev->new_layout = ALGORITHM_LEFT_SYMMETRIC;
5411 mddev->new_chunk = chunksect << 9;
5413 return setup_conf(mddev);
5416 static void *raid5_takeover_raid6(mddev_t *mddev)
5420 switch (mddev->layout) {
5421 case ALGORITHM_LEFT_ASYMMETRIC_6:
5422 new_layout = ALGORITHM_LEFT_ASYMMETRIC;
5424 case ALGORITHM_RIGHT_ASYMMETRIC_6:
5425 new_layout = ALGORITHM_RIGHT_ASYMMETRIC;
5427 case ALGORITHM_LEFT_SYMMETRIC_6:
5428 new_layout = ALGORITHM_LEFT_SYMMETRIC;
5430 case ALGORITHM_RIGHT_SYMMETRIC_6:
5431 new_layout = ALGORITHM_RIGHT_SYMMETRIC;
5433 case ALGORITHM_PARITY_0_6:
5434 new_layout = ALGORITHM_PARITY_0;
5436 case ALGORITHM_PARITY_N:
5437 new_layout = ALGORITHM_PARITY_N;
5440 return ERR_PTR(-EINVAL);
5442 mddev->new_level = 5;
5443 mddev->new_layout = new_layout;
5444 mddev->delta_disks = -1;
5445 mddev->raid_disks -= 1;
5446 return setup_conf(mddev);
5450 static int raid5_reconfig(mddev_t *mddev, int new_layout, int new_chunk)
5452 /* For a 2-drive array, the layout and chunk size can be changed
5453 * immediately as not restriping is needed.
5454 * For larger arrays we record the new value - after validation
5455 * to be used by a reshape pass.
5457 raid5_conf_t *conf = mddev_to_conf(mddev);
5459 if (new_layout >= 0 && !algorithm_valid_raid5(new_layout))
5461 if (new_chunk > 0) {
5462 if (new_chunk & (new_chunk-1))
5463 /* not a power of 2 */
5465 if (new_chunk < PAGE_SIZE)
5467 if (mddev->array_sectors & ((new_chunk>>9)-1))
5468 /* not factor of array size */
5472 /* They look valid */
5474 if (mddev->raid_disks == 2) {
5476 if (new_layout >= 0) {
5477 conf->algorithm = new_layout;
5478 mddev->layout = mddev->new_layout = new_layout;
5480 if (new_chunk > 0) {
5481 conf->chunk_size = new_chunk;
5482 mddev->chunk_size = mddev->new_chunk = new_chunk;
5484 set_bit(MD_CHANGE_DEVS, &mddev->flags);
5485 md_wakeup_thread(mddev->thread);
5487 if (new_layout >= 0)
5488 mddev->new_layout = new_layout;
5490 mddev->new_chunk = new_chunk;
5495 static int raid6_reconfig(mddev_t *mddev, int new_layout, int new_chunk)
5497 if (new_layout >= 0 && !algorithm_valid_raid6(new_layout))
5499 if (new_chunk > 0) {
5500 if (new_chunk & (new_chunk-1))
5501 /* not a power of 2 */
5503 if (new_chunk < PAGE_SIZE)
5505 if (mddev->array_sectors & ((new_chunk>>9)-1))
5506 /* not factor of array size */
5510 /* They look valid */
5512 if (new_layout >= 0)
5513 mddev->new_layout = new_layout;
5515 mddev->new_chunk = new_chunk;
5520 static void *raid5_takeover(mddev_t *mddev)
5522 /* raid5 can take over:
5523 * raid0 - if all devices are the same - make it a raid4 layout
5524 * raid1 - if there are two drives. We need to know the chunk size
5525 * raid4 - trivial - just use a raid4 layout.
5526 * raid6 - Providing it is a *_6 layout
5528 * For now, just do raid1
5531 if (mddev->level == 1)
5532 return raid5_takeover_raid1(mddev);
5533 if (mddev->level == 4) {
5534 mddev->new_layout = ALGORITHM_PARITY_N;
5535 mddev->new_level = 5;
5536 return setup_conf(mddev);
5538 if (mddev->level == 6)
5539 return raid5_takeover_raid6(mddev);
5541 return ERR_PTR(-EINVAL);
5545 static struct mdk_personality raid5_personality;
5547 static void *raid6_takeover(mddev_t *mddev)
5549 /* Currently can only take over a raid5. We map the
5550 * personality to an equivalent raid6 personality
5551 * with the Q block at the end.
5555 if (mddev->pers != &raid5_personality)
5556 return ERR_PTR(-EINVAL);
5557 if (mddev->degraded > 1)
5558 return ERR_PTR(-EINVAL);
5559 if (mddev->raid_disks > 253)
5560 return ERR_PTR(-EINVAL);
5561 if (mddev->raid_disks < 3)
5562 return ERR_PTR(-EINVAL);
5564 switch (mddev->layout) {
5565 case ALGORITHM_LEFT_ASYMMETRIC:
5566 new_layout = ALGORITHM_LEFT_ASYMMETRIC_6;
5568 case ALGORITHM_RIGHT_ASYMMETRIC:
5569 new_layout = ALGORITHM_RIGHT_ASYMMETRIC_6;
5571 case ALGORITHM_LEFT_SYMMETRIC:
5572 new_layout = ALGORITHM_LEFT_SYMMETRIC_6;
5574 case ALGORITHM_RIGHT_SYMMETRIC:
5575 new_layout = ALGORITHM_RIGHT_SYMMETRIC_6;
5577 case ALGORITHM_PARITY_0:
5578 new_layout = ALGORITHM_PARITY_0_6;
5580 case ALGORITHM_PARITY_N:
5581 new_layout = ALGORITHM_PARITY_N;
5584 return ERR_PTR(-EINVAL);
5586 mddev->new_level = 6;
5587 mddev->new_layout = new_layout;
5588 mddev->delta_disks = 1;
5589 mddev->raid_disks += 1;
5590 return setup_conf(mddev);
5594 static struct mdk_personality raid6_personality =
5598 .owner = THIS_MODULE,
5599 .make_request = make_request,
5603 .error_handler = error,
5604 .hot_add_disk = raid5_add_disk,
5605 .hot_remove_disk= raid5_remove_disk,
5606 .spare_active = raid5_spare_active,
5607 .sync_request = sync_request,
5608 .resize = raid5_resize,
5610 .check_reshape = raid5_check_reshape,
5611 .start_reshape = raid5_start_reshape,
5612 .finish_reshape = raid5_finish_reshape,
5613 .quiesce = raid5_quiesce,
5614 .takeover = raid6_takeover,
5615 .reconfig = raid6_reconfig,
5617 static struct mdk_personality raid5_personality =
5621 .owner = THIS_MODULE,
5622 .make_request = make_request,
5626 .error_handler = error,
5627 .hot_add_disk = raid5_add_disk,
5628 .hot_remove_disk= raid5_remove_disk,
5629 .spare_active = raid5_spare_active,
5630 .sync_request = sync_request,
5631 .resize = raid5_resize,
5633 .check_reshape = raid5_check_reshape,
5634 .start_reshape = raid5_start_reshape,
5635 .finish_reshape = raid5_finish_reshape,
5636 .quiesce = raid5_quiesce,
5637 .takeover = raid5_takeover,
5638 .reconfig = raid5_reconfig,
5641 static struct mdk_personality raid4_personality =
5645 .owner = THIS_MODULE,
5646 .make_request = make_request,
5650 .error_handler = error,
5651 .hot_add_disk = raid5_add_disk,
5652 .hot_remove_disk= raid5_remove_disk,
5653 .spare_active = raid5_spare_active,
5654 .sync_request = sync_request,
5655 .resize = raid5_resize,
5657 .check_reshape = raid5_check_reshape,
5658 .start_reshape = raid5_start_reshape,
5659 .finish_reshape = raid5_finish_reshape,
5660 .quiesce = raid5_quiesce,
5663 static int __init raid5_init(void)
5665 register_md_personality(&raid6_personality);
5666 register_md_personality(&raid5_personality);
5667 register_md_personality(&raid4_personality);
5671 static void raid5_exit(void)
5673 unregister_md_personality(&raid6_personality);
5674 unregister_md_personality(&raid5_personality);
5675 unregister_md_personality(&raid4_personality);
5678 module_init(raid5_init);
5679 module_exit(raid5_exit);
5680 MODULE_LICENSE("GPL");
5681 MODULE_ALIAS("md-personality-4"); /* RAID5 */
5682 MODULE_ALIAS("md-raid5");
5683 MODULE_ALIAS("md-raid4");
5684 MODULE_ALIAS("md-level-5");
5685 MODULE_ALIAS("md-level-4");
5686 MODULE_ALIAS("md-personality-8"); /* RAID6 */
5687 MODULE_ALIAS("md-raid6");
5688 MODULE_ALIAS("md-level-6");
5690 /* This used to be two separate modules, they were: */
5691 MODULE_ALIAS("raid5");
5692 MODULE_ALIAS("raid6");