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 ops_complete_compute5(void *stripe_head_ref)
628 struct stripe_head *sh = stripe_head_ref;
629 int target = sh->ops.target;
630 struct r5dev *tgt = &sh->dev[target];
632 pr_debug("%s: stripe %llu\n", __func__,
633 (unsigned long long)sh->sector);
635 set_bit(R5_UPTODATE, &tgt->flags);
636 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
637 clear_bit(R5_Wantcompute, &tgt->flags);
638 clear_bit(STRIPE_COMPUTE_RUN, &sh->state);
639 if (sh->check_state == check_state_compute_run)
640 sh->check_state = check_state_compute_result;
641 set_bit(STRIPE_HANDLE, &sh->state);
645 /* return a pointer to the address conversion region of the scribble buffer */
646 static addr_conv_t *to_addr_conv(struct stripe_head *sh,
647 struct raid5_percpu *percpu)
649 return percpu->scribble + sizeof(struct page *) * (sh->disks + 2);
652 static struct dma_async_tx_descriptor *
653 ops_run_compute5(struct stripe_head *sh, struct raid5_percpu *percpu)
655 int disks = sh->disks;
656 struct page **xor_srcs = percpu->scribble;
657 int target = sh->ops.target;
658 struct r5dev *tgt = &sh->dev[target];
659 struct page *xor_dest = tgt->page;
661 struct dma_async_tx_descriptor *tx;
662 struct async_submit_ctl submit;
665 pr_debug("%s: stripe %llu block: %d\n",
666 __func__, (unsigned long long)sh->sector, target);
667 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
669 for (i = disks; i--; )
671 xor_srcs[count++] = sh->dev[i].page;
673 atomic_inc(&sh->count);
675 init_async_submit(&submit, ASYNC_TX_XOR_ZERO_DST, NULL,
676 ops_complete_compute5, sh, to_addr_conv(sh, percpu));
677 if (unlikely(count == 1))
678 tx = async_memcpy(xor_dest, xor_srcs[0], 0, 0, STRIPE_SIZE, &submit);
680 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
685 static void ops_complete_prexor(void *stripe_head_ref)
687 struct stripe_head *sh = stripe_head_ref;
689 pr_debug("%s: stripe %llu\n", __func__,
690 (unsigned long long)sh->sector);
693 static struct dma_async_tx_descriptor *
694 ops_run_prexor(struct stripe_head *sh, struct raid5_percpu *percpu,
695 struct dma_async_tx_descriptor *tx)
697 int disks = sh->disks;
698 struct page **xor_srcs = percpu->scribble;
699 int count = 0, pd_idx = sh->pd_idx, i;
700 struct async_submit_ctl submit;
702 /* existing parity data subtracted */
703 struct page *xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
705 pr_debug("%s: stripe %llu\n", __func__,
706 (unsigned long long)sh->sector);
708 for (i = disks; i--; ) {
709 struct r5dev *dev = &sh->dev[i];
710 /* Only process blocks that are known to be uptodate */
711 if (test_bit(R5_Wantdrain, &dev->flags))
712 xor_srcs[count++] = dev->page;
715 init_async_submit(&submit, ASYNC_TX_XOR_DROP_DST, tx,
716 ops_complete_prexor, sh, to_addr_conv(sh, percpu));
717 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
722 static struct dma_async_tx_descriptor *
723 ops_run_biodrain(struct stripe_head *sh, struct dma_async_tx_descriptor *tx)
725 int disks = sh->disks;
728 pr_debug("%s: stripe %llu\n", __func__,
729 (unsigned long long)sh->sector);
731 for (i = disks; i--; ) {
732 struct r5dev *dev = &sh->dev[i];
735 if (test_and_clear_bit(R5_Wantdrain, &dev->flags)) {
738 spin_lock(&sh->lock);
739 chosen = dev->towrite;
741 BUG_ON(dev->written);
742 wbi = dev->written = chosen;
743 spin_unlock(&sh->lock);
745 while (wbi && wbi->bi_sector <
746 dev->sector + STRIPE_SECTORS) {
747 tx = async_copy_data(1, wbi, dev->page,
749 wbi = r5_next_bio(wbi, dev->sector);
757 static void ops_complete_postxor(void *stripe_head_ref)
759 struct stripe_head *sh = stripe_head_ref;
760 int disks = sh->disks, i, pd_idx = sh->pd_idx;
762 pr_debug("%s: stripe %llu\n", __func__,
763 (unsigned long long)sh->sector);
765 for (i = disks; i--; ) {
766 struct r5dev *dev = &sh->dev[i];
767 if (dev->written || i == pd_idx)
768 set_bit(R5_UPTODATE, &dev->flags);
771 if (sh->reconstruct_state == reconstruct_state_drain_run)
772 sh->reconstruct_state = reconstruct_state_drain_result;
773 else if (sh->reconstruct_state == reconstruct_state_prexor_drain_run)
774 sh->reconstruct_state = reconstruct_state_prexor_drain_result;
776 BUG_ON(sh->reconstruct_state != reconstruct_state_run);
777 sh->reconstruct_state = reconstruct_state_result;
780 set_bit(STRIPE_HANDLE, &sh->state);
785 ops_run_postxor(struct stripe_head *sh, struct raid5_percpu *percpu,
786 struct dma_async_tx_descriptor *tx)
788 int disks = sh->disks;
789 struct page **xor_srcs = percpu->scribble;
790 struct async_submit_ctl submit;
791 int count = 0, pd_idx = sh->pd_idx, i;
792 struct page *xor_dest;
796 pr_debug("%s: stripe %llu\n", __func__,
797 (unsigned long long)sh->sector);
799 /* check if prexor is active which means only process blocks
800 * that are part of a read-modify-write (written)
802 if (sh->reconstruct_state == reconstruct_state_prexor_drain_run) {
804 xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
805 for (i = disks; i--; ) {
806 struct r5dev *dev = &sh->dev[i];
808 xor_srcs[count++] = dev->page;
811 xor_dest = sh->dev[pd_idx].page;
812 for (i = disks; i--; ) {
813 struct r5dev *dev = &sh->dev[i];
815 xor_srcs[count++] = dev->page;
819 /* 1/ if we prexor'd then the dest is reused as a source
820 * 2/ if we did not prexor then we are redoing the parity
821 * set ASYNC_TX_XOR_DROP_DST and ASYNC_TX_XOR_ZERO_DST
822 * for the synchronous xor case
824 flags = ASYNC_TX_ACK |
825 (prexor ? ASYNC_TX_XOR_DROP_DST : ASYNC_TX_XOR_ZERO_DST);
827 atomic_inc(&sh->count);
829 init_async_submit(&submit, flags, tx, ops_complete_postxor, sh,
830 to_addr_conv(sh, percpu));
831 if (unlikely(count == 1))
832 tx = async_memcpy(xor_dest, xor_srcs[0], 0, 0, STRIPE_SIZE, &submit);
834 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
837 static void ops_complete_check(void *stripe_head_ref)
839 struct stripe_head *sh = stripe_head_ref;
841 pr_debug("%s: stripe %llu\n", __func__,
842 (unsigned long long)sh->sector);
844 sh->check_state = check_state_check_result;
845 set_bit(STRIPE_HANDLE, &sh->state);
849 static void ops_run_check(struct stripe_head *sh, struct raid5_percpu *percpu)
851 int disks = sh->disks;
852 struct page **xor_srcs = percpu->scribble;
853 struct dma_async_tx_descriptor *tx;
854 struct async_submit_ctl submit;
856 int count = 0, pd_idx = sh->pd_idx, i;
857 struct page *xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
859 pr_debug("%s: stripe %llu\n", __func__,
860 (unsigned long long)sh->sector);
862 for (i = disks; i--; ) {
863 struct r5dev *dev = &sh->dev[i];
865 xor_srcs[count++] = dev->page;
868 init_async_submit(&submit, 0, NULL, NULL, NULL,
869 to_addr_conv(sh, percpu));
870 tx = async_xor_val(xor_dest, xor_srcs, 0, count, STRIPE_SIZE,
871 &sh->ops.zero_sum_result, &submit);
873 atomic_inc(&sh->count);
874 init_async_submit(&submit, ASYNC_TX_ACK, tx, ops_complete_check, sh, NULL);
875 tx = async_trigger_callback(&submit);
878 static void raid5_run_ops(struct stripe_head *sh, unsigned long ops_request)
880 int overlap_clear = 0, i, disks = sh->disks;
881 struct dma_async_tx_descriptor *tx = NULL;
882 raid5_conf_t *conf = sh->raid_conf;
883 struct raid5_percpu *percpu;
887 percpu = per_cpu_ptr(conf->percpu, cpu);
888 if (test_bit(STRIPE_OP_BIOFILL, &ops_request)) {
893 if (test_bit(STRIPE_OP_COMPUTE_BLK, &ops_request)) {
894 tx = ops_run_compute5(sh, percpu);
895 /* terminate the chain if postxor is not set to be run */
896 if (tx && !test_bit(STRIPE_OP_POSTXOR, &ops_request))
900 if (test_bit(STRIPE_OP_PREXOR, &ops_request))
901 tx = ops_run_prexor(sh, percpu, tx);
903 if (test_bit(STRIPE_OP_BIODRAIN, &ops_request)) {
904 tx = ops_run_biodrain(sh, tx);
908 if (test_bit(STRIPE_OP_POSTXOR, &ops_request))
909 ops_run_postxor(sh, percpu, tx);
911 if (test_bit(STRIPE_OP_CHECK, &ops_request))
912 ops_run_check(sh, percpu);
915 for (i = disks; i--; ) {
916 struct r5dev *dev = &sh->dev[i];
917 if (test_and_clear_bit(R5_Overlap, &dev->flags))
918 wake_up(&sh->raid_conf->wait_for_overlap);
923 static int grow_one_stripe(raid5_conf_t *conf)
925 struct stripe_head *sh;
926 sh = kmem_cache_alloc(conf->slab_cache, GFP_KERNEL);
929 memset(sh, 0, sizeof(*sh) + (conf->raid_disks-1)*sizeof(struct r5dev));
930 sh->raid_conf = conf;
931 spin_lock_init(&sh->lock);
933 if (grow_buffers(sh, conf->raid_disks)) {
934 shrink_buffers(sh, conf->raid_disks);
935 kmem_cache_free(conf->slab_cache, sh);
938 sh->disks = conf->raid_disks;
939 /* we just created an active stripe so... */
940 atomic_set(&sh->count, 1);
941 atomic_inc(&conf->active_stripes);
942 INIT_LIST_HEAD(&sh->lru);
947 static int grow_stripes(raid5_conf_t *conf, int num)
949 struct kmem_cache *sc;
950 int devs = conf->raid_disks;
952 sprintf(conf->cache_name[0],
953 "raid%d-%s", conf->level, mdname(conf->mddev));
954 sprintf(conf->cache_name[1],
955 "raid%d-%s-alt", conf->level, mdname(conf->mddev));
956 conf->active_name = 0;
957 sc = kmem_cache_create(conf->cache_name[conf->active_name],
958 sizeof(struct stripe_head)+(devs-1)*sizeof(struct r5dev),
962 conf->slab_cache = sc;
963 conf->pool_size = devs;
965 if (!grow_one_stripe(conf))
971 * scribble_len - return the required size of the scribble region
972 * @num - total number of disks in the array
974 * The size must be enough to contain:
975 * 1/ a struct page pointer for each device in the array +2
976 * 2/ room to convert each entry in (1) to its corresponding dma
977 * (dma_map_page()) or page (page_address()) address.
979 * Note: the +2 is for the destination buffers of the ddf/raid6 case where we
980 * calculate over all devices (not just the data blocks), using zeros in place
981 * of the P and Q blocks.
983 static size_t scribble_len(int num)
987 len = sizeof(struct page *) * (num+2) + sizeof(addr_conv_t) * (num+2);
992 static int resize_stripes(raid5_conf_t *conf, int newsize)
994 /* Make all the stripes able to hold 'newsize' devices.
995 * New slots in each stripe get 'page' set to a new page.
997 * This happens in stages:
998 * 1/ create a new kmem_cache and allocate the required number of
1000 * 2/ gather all the old stripe_heads and tranfer the pages across
1001 * to the new stripe_heads. This will have the side effect of
1002 * freezing the array as once all stripe_heads have been collected,
1003 * no IO will be possible. Old stripe heads are freed once their
1004 * pages have been transferred over, and the old kmem_cache is
1005 * freed when all stripes are done.
1006 * 3/ reallocate conf->disks to be suitable bigger. If this fails,
1007 * we simple return a failre status - no need to clean anything up.
1008 * 4/ allocate new pages for the new slots in the new stripe_heads.
1009 * If this fails, we don't bother trying the shrink the
1010 * stripe_heads down again, we just leave them as they are.
1011 * As each stripe_head is processed the new one is released into
1014 * Once step2 is started, we cannot afford to wait for a write,
1015 * so we use GFP_NOIO allocations.
1017 struct stripe_head *osh, *nsh;
1018 LIST_HEAD(newstripes);
1019 struct disk_info *ndisks;
1022 struct kmem_cache *sc;
1025 if (newsize <= conf->pool_size)
1026 return 0; /* never bother to shrink */
1028 err = md_allow_write(conf->mddev);
1033 sc = kmem_cache_create(conf->cache_name[1-conf->active_name],
1034 sizeof(struct stripe_head)+(newsize-1)*sizeof(struct r5dev),
1039 for (i = conf->max_nr_stripes; i; i--) {
1040 nsh = kmem_cache_alloc(sc, GFP_KERNEL);
1044 memset(nsh, 0, sizeof(*nsh) + (newsize-1)*sizeof(struct r5dev));
1046 nsh->raid_conf = conf;
1047 spin_lock_init(&nsh->lock);
1049 list_add(&nsh->lru, &newstripes);
1052 /* didn't get enough, give up */
1053 while (!list_empty(&newstripes)) {
1054 nsh = list_entry(newstripes.next, struct stripe_head, lru);
1055 list_del(&nsh->lru);
1056 kmem_cache_free(sc, nsh);
1058 kmem_cache_destroy(sc);
1061 /* Step 2 - Must use GFP_NOIO now.
1062 * OK, we have enough stripes, start collecting inactive
1063 * stripes and copying them over
1065 list_for_each_entry(nsh, &newstripes, lru) {
1066 spin_lock_irq(&conf->device_lock);
1067 wait_event_lock_irq(conf->wait_for_stripe,
1068 !list_empty(&conf->inactive_list),
1070 unplug_slaves(conf->mddev)
1072 osh = get_free_stripe(conf);
1073 spin_unlock_irq(&conf->device_lock);
1074 atomic_set(&nsh->count, 1);
1075 for(i=0; i<conf->pool_size; i++)
1076 nsh->dev[i].page = osh->dev[i].page;
1077 for( ; i<newsize; i++)
1078 nsh->dev[i].page = NULL;
1079 kmem_cache_free(conf->slab_cache, osh);
1081 kmem_cache_destroy(conf->slab_cache);
1084 * At this point, we are holding all the stripes so the array
1085 * is completely stalled, so now is a good time to resize
1086 * conf->disks and the scribble region
1088 ndisks = kzalloc(newsize * sizeof(struct disk_info), GFP_NOIO);
1090 for (i=0; i<conf->raid_disks; i++)
1091 ndisks[i] = conf->disks[i];
1093 conf->disks = ndisks;
1098 conf->scribble_len = scribble_len(newsize);
1099 for_each_present_cpu(cpu) {
1100 struct raid5_percpu *percpu;
1103 percpu = per_cpu_ptr(conf->percpu, cpu);
1104 scribble = kmalloc(conf->scribble_len, GFP_NOIO);
1107 kfree(percpu->scribble);
1108 percpu->scribble = scribble;
1116 /* Step 4, return new stripes to service */
1117 while(!list_empty(&newstripes)) {
1118 nsh = list_entry(newstripes.next, struct stripe_head, lru);
1119 list_del_init(&nsh->lru);
1121 for (i=conf->raid_disks; i < newsize; i++)
1122 if (nsh->dev[i].page == NULL) {
1123 struct page *p = alloc_page(GFP_NOIO);
1124 nsh->dev[i].page = p;
1128 release_stripe(nsh);
1130 /* critical section pass, GFP_NOIO no longer needed */
1132 conf->slab_cache = sc;
1133 conf->active_name = 1-conf->active_name;
1134 conf->pool_size = newsize;
1138 static int drop_one_stripe(raid5_conf_t *conf)
1140 struct stripe_head *sh;
1142 spin_lock_irq(&conf->device_lock);
1143 sh = get_free_stripe(conf);
1144 spin_unlock_irq(&conf->device_lock);
1147 BUG_ON(atomic_read(&sh->count));
1148 shrink_buffers(sh, conf->pool_size);
1149 kmem_cache_free(conf->slab_cache, sh);
1150 atomic_dec(&conf->active_stripes);
1154 static void shrink_stripes(raid5_conf_t *conf)
1156 while (drop_one_stripe(conf))
1159 if (conf->slab_cache)
1160 kmem_cache_destroy(conf->slab_cache);
1161 conf->slab_cache = NULL;
1164 static void raid5_end_read_request(struct bio * bi, int error)
1166 struct stripe_head *sh = bi->bi_private;
1167 raid5_conf_t *conf = sh->raid_conf;
1168 int disks = sh->disks, i;
1169 int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
1170 char b[BDEVNAME_SIZE];
1174 for (i=0 ; i<disks; i++)
1175 if (bi == &sh->dev[i].req)
1178 pr_debug("end_read_request %llu/%d, count: %d, uptodate %d.\n",
1179 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
1187 set_bit(R5_UPTODATE, &sh->dev[i].flags);
1188 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
1189 rdev = conf->disks[i].rdev;
1190 printk_rl(KERN_INFO "raid5:%s: read error corrected"
1191 " (%lu sectors at %llu on %s)\n",
1192 mdname(conf->mddev), STRIPE_SECTORS,
1193 (unsigned long long)(sh->sector
1194 + rdev->data_offset),
1195 bdevname(rdev->bdev, b));
1196 clear_bit(R5_ReadError, &sh->dev[i].flags);
1197 clear_bit(R5_ReWrite, &sh->dev[i].flags);
1199 if (atomic_read(&conf->disks[i].rdev->read_errors))
1200 atomic_set(&conf->disks[i].rdev->read_errors, 0);
1202 const char *bdn = bdevname(conf->disks[i].rdev->bdev, b);
1204 rdev = conf->disks[i].rdev;
1206 clear_bit(R5_UPTODATE, &sh->dev[i].flags);
1207 atomic_inc(&rdev->read_errors);
1208 if (conf->mddev->degraded)
1209 printk_rl(KERN_WARNING
1210 "raid5:%s: read error not correctable "
1211 "(sector %llu on %s).\n",
1212 mdname(conf->mddev),
1213 (unsigned long long)(sh->sector
1214 + rdev->data_offset),
1216 else if (test_bit(R5_ReWrite, &sh->dev[i].flags))
1218 printk_rl(KERN_WARNING
1219 "raid5:%s: read error NOT corrected!! "
1220 "(sector %llu on %s).\n",
1221 mdname(conf->mddev),
1222 (unsigned long long)(sh->sector
1223 + rdev->data_offset),
1225 else if (atomic_read(&rdev->read_errors)
1226 > conf->max_nr_stripes)
1228 "raid5:%s: Too many read errors, failing device %s.\n",
1229 mdname(conf->mddev), bdn);
1233 set_bit(R5_ReadError, &sh->dev[i].flags);
1235 clear_bit(R5_ReadError, &sh->dev[i].flags);
1236 clear_bit(R5_ReWrite, &sh->dev[i].flags);
1237 md_error(conf->mddev, rdev);
1240 rdev_dec_pending(conf->disks[i].rdev, conf->mddev);
1241 clear_bit(R5_LOCKED, &sh->dev[i].flags);
1242 set_bit(STRIPE_HANDLE, &sh->state);
1246 static void raid5_end_write_request(struct bio *bi, int error)
1248 struct stripe_head *sh = bi->bi_private;
1249 raid5_conf_t *conf = sh->raid_conf;
1250 int disks = sh->disks, i;
1251 int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
1253 for (i=0 ; i<disks; i++)
1254 if (bi == &sh->dev[i].req)
1257 pr_debug("end_write_request %llu/%d, count %d, uptodate: %d.\n",
1258 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
1266 md_error(conf->mddev, conf->disks[i].rdev);
1268 rdev_dec_pending(conf->disks[i].rdev, conf->mddev);
1270 clear_bit(R5_LOCKED, &sh->dev[i].flags);
1271 set_bit(STRIPE_HANDLE, &sh->state);
1276 static sector_t compute_blocknr(struct stripe_head *sh, int i, int previous);
1278 static void raid5_build_block(struct stripe_head *sh, int i, int previous)
1280 struct r5dev *dev = &sh->dev[i];
1282 bio_init(&dev->req);
1283 dev->req.bi_io_vec = &dev->vec;
1285 dev->req.bi_max_vecs++;
1286 dev->vec.bv_page = dev->page;
1287 dev->vec.bv_len = STRIPE_SIZE;
1288 dev->vec.bv_offset = 0;
1290 dev->req.bi_sector = sh->sector;
1291 dev->req.bi_private = sh;
1294 dev->sector = compute_blocknr(sh, i, previous);
1297 static void error(mddev_t *mddev, mdk_rdev_t *rdev)
1299 char b[BDEVNAME_SIZE];
1300 raid5_conf_t *conf = (raid5_conf_t *) mddev->private;
1301 pr_debug("raid5: error called\n");
1303 if (!test_bit(Faulty, &rdev->flags)) {
1304 set_bit(MD_CHANGE_DEVS, &mddev->flags);
1305 if (test_and_clear_bit(In_sync, &rdev->flags)) {
1306 unsigned long flags;
1307 spin_lock_irqsave(&conf->device_lock, flags);
1309 spin_unlock_irqrestore(&conf->device_lock, flags);
1311 * if recovery was running, make sure it aborts.
1313 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1315 set_bit(Faulty, &rdev->flags);
1317 "raid5: Disk failure on %s, disabling device.\n"
1318 "raid5: Operation continuing on %d devices.\n",
1319 bdevname(rdev->bdev,b), conf->raid_disks - mddev->degraded);
1324 * Input: a 'big' sector number,
1325 * Output: index of the data and parity disk, and the sector # in them.
1327 static sector_t raid5_compute_sector(raid5_conf_t *conf, sector_t r_sector,
1328 int previous, int *dd_idx,
1329 struct stripe_head *sh)
1332 unsigned long chunk_number;
1333 unsigned int chunk_offset;
1336 sector_t new_sector;
1337 int algorithm = previous ? conf->prev_algo
1339 int sectors_per_chunk = previous ? (conf->prev_chunk >> 9)
1340 : (conf->chunk_size >> 9);
1341 int raid_disks = previous ? conf->previous_raid_disks
1343 int data_disks = raid_disks - conf->max_degraded;
1345 /* First compute the information on this sector */
1348 * Compute the chunk number and the sector offset inside the chunk
1350 chunk_offset = sector_div(r_sector, sectors_per_chunk);
1351 chunk_number = r_sector;
1352 BUG_ON(r_sector != chunk_number);
1355 * Compute the stripe number
1357 stripe = chunk_number / data_disks;
1360 * Compute the data disk and parity disk indexes inside the stripe
1362 *dd_idx = chunk_number % data_disks;
1365 * Select the parity disk based on the user selected algorithm.
1367 pd_idx = qd_idx = ~0;
1368 switch(conf->level) {
1370 pd_idx = data_disks;
1373 switch (algorithm) {
1374 case ALGORITHM_LEFT_ASYMMETRIC:
1375 pd_idx = data_disks - stripe % raid_disks;
1376 if (*dd_idx >= pd_idx)
1379 case ALGORITHM_RIGHT_ASYMMETRIC:
1380 pd_idx = stripe % raid_disks;
1381 if (*dd_idx >= pd_idx)
1384 case ALGORITHM_LEFT_SYMMETRIC:
1385 pd_idx = data_disks - stripe % raid_disks;
1386 *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
1388 case ALGORITHM_RIGHT_SYMMETRIC:
1389 pd_idx = stripe % raid_disks;
1390 *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
1392 case ALGORITHM_PARITY_0:
1396 case ALGORITHM_PARITY_N:
1397 pd_idx = data_disks;
1400 printk(KERN_ERR "raid5: unsupported algorithm %d\n",
1407 switch (algorithm) {
1408 case ALGORITHM_LEFT_ASYMMETRIC:
1409 pd_idx = raid_disks - 1 - (stripe % raid_disks);
1410 qd_idx = pd_idx + 1;
1411 if (pd_idx == raid_disks-1) {
1412 (*dd_idx)++; /* Q D D D P */
1414 } else if (*dd_idx >= pd_idx)
1415 (*dd_idx) += 2; /* D D P Q D */
1417 case ALGORITHM_RIGHT_ASYMMETRIC:
1418 pd_idx = stripe % raid_disks;
1419 qd_idx = pd_idx + 1;
1420 if (pd_idx == raid_disks-1) {
1421 (*dd_idx)++; /* Q D D D P */
1423 } else if (*dd_idx >= pd_idx)
1424 (*dd_idx) += 2; /* D D P Q D */
1426 case ALGORITHM_LEFT_SYMMETRIC:
1427 pd_idx = raid_disks - 1 - (stripe % raid_disks);
1428 qd_idx = (pd_idx + 1) % raid_disks;
1429 *dd_idx = (pd_idx + 2 + *dd_idx) % raid_disks;
1431 case ALGORITHM_RIGHT_SYMMETRIC:
1432 pd_idx = stripe % raid_disks;
1433 qd_idx = (pd_idx + 1) % raid_disks;
1434 *dd_idx = (pd_idx + 2 + *dd_idx) % raid_disks;
1437 case ALGORITHM_PARITY_0:
1442 case ALGORITHM_PARITY_N:
1443 pd_idx = data_disks;
1444 qd_idx = data_disks + 1;
1447 case ALGORITHM_ROTATING_ZERO_RESTART:
1448 /* Exactly the same as RIGHT_ASYMMETRIC, but or
1449 * of blocks for computing Q is different.
1451 pd_idx = stripe % raid_disks;
1452 qd_idx = pd_idx + 1;
1453 if (pd_idx == raid_disks-1) {
1454 (*dd_idx)++; /* Q D D D P */
1456 } else if (*dd_idx >= pd_idx)
1457 (*dd_idx) += 2; /* D D P Q D */
1461 case ALGORITHM_ROTATING_N_RESTART:
1462 /* Same a left_asymmetric, by first stripe is
1463 * D D D P Q rather than
1466 pd_idx = raid_disks - 1 - ((stripe + 1) % raid_disks);
1467 qd_idx = pd_idx + 1;
1468 if (pd_idx == raid_disks-1) {
1469 (*dd_idx)++; /* Q D D D P */
1471 } else if (*dd_idx >= pd_idx)
1472 (*dd_idx) += 2; /* D D P Q D */
1476 case ALGORITHM_ROTATING_N_CONTINUE:
1477 /* Same as left_symmetric but Q is before P */
1478 pd_idx = raid_disks - 1 - (stripe % raid_disks);
1479 qd_idx = (pd_idx + raid_disks - 1) % raid_disks;
1480 *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
1484 case ALGORITHM_LEFT_ASYMMETRIC_6:
1485 /* RAID5 left_asymmetric, with Q on last device */
1486 pd_idx = data_disks - stripe % (raid_disks-1);
1487 if (*dd_idx >= pd_idx)
1489 qd_idx = raid_disks - 1;
1492 case ALGORITHM_RIGHT_ASYMMETRIC_6:
1493 pd_idx = stripe % (raid_disks-1);
1494 if (*dd_idx >= pd_idx)
1496 qd_idx = raid_disks - 1;
1499 case ALGORITHM_LEFT_SYMMETRIC_6:
1500 pd_idx = data_disks - stripe % (raid_disks-1);
1501 *dd_idx = (pd_idx + 1 + *dd_idx) % (raid_disks-1);
1502 qd_idx = raid_disks - 1;
1505 case ALGORITHM_RIGHT_SYMMETRIC_6:
1506 pd_idx = stripe % (raid_disks-1);
1507 *dd_idx = (pd_idx + 1 + *dd_idx) % (raid_disks-1);
1508 qd_idx = raid_disks - 1;
1511 case ALGORITHM_PARITY_0_6:
1514 qd_idx = raid_disks - 1;
1519 printk(KERN_CRIT "raid6: unsupported algorithm %d\n",
1527 sh->pd_idx = pd_idx;
1528 sh->qd_idx = qd_idx;
1529 sh->ddf_layout = ddf_layout;
1532 * Finally, compute the new sector number
1534 new_sector = (sector_t)stripe * sectors_per_chunk + chunk_offset;
1539 static sector_t compute_blocknr(struct stripe_head *sh, int i, int previous)
1541 raid5_conf_t *conf = sh->raid_conf;
1542 int raid_disks = sh->disks;
1543 int data_disks = raid_disks - conf->max_degraded;
1544 sector_t new_sector = sh->sector, check;
1545 int sectors_per_chunk = previous ? (conf->prev_chunk >> 9)
1546 : (conf->chunk_size >> 9);
1547 int algorithm = previous ? conf->prev_algo
1551 int chunk_number, dummy1, dd_idx = i;
1553 struct stripe_head sh2;
1556 chunk_offset = sector_div(new_sector, sectors_per_chunk);
1557 stripe = new_sector;
1558 BUG_ON(new_sector != stripe);
1560 if (i == sh->pd_idx)
1562 switch(conf->level) {
1565 switch (algorithm) {
1566 case ALGORITHM_LEFT_ASYMMETRIC:
1567 case ALGORITHM_RIGHT_ASYMMETRIC:
1571 case ALGORITHM_LEFT_SYMMETRIC:
1572 case ALGORITHM_RIGHT_SYMMETRIC:
1575 i -= (sh->pd_idx + 1);
1577 case ALGORITHM_PARITY_0:
1580 case ALGORITHM_PARITY_N:
1583 printk(KERN_ERR "raid5: unsupported algorithm %d\n",
1589 if (i == sh->qd_idx)
1590 return 0; /* It is the Q disk */
1591 switch (algorithm) {
1592 case ALGORITHM_LEFT_ASYMMETRIC:
1593 case ALGORITHM_RIGHT_ASYMMETRIC:
1594 case ALGORITHM_ROTATING_ZERO_RESTART:
1595 case ALGORITHM_ROTATING_N_RESTART:
1596 if (sh->pd_idx == raid_disks-1)
1597 i--; /* Q D D D P */
1598 else if (i > sh->pd_idx)
1599 i -= 2; /* D D P Q D */
1601 case ALGORITHM_LEFT_SYMMETRIC:
1602 case ALGORITHM_RIGHT_SYMMETRIC:
1603 if (sh->pd_idx == raid_disks-1)
1604 i--; /* Q D D D P */
1609 i -= (sh->pd_idx + 2);
1612 case ALGORITHM_PARITY_0:
1615 case ALGORITHM_PARITY_N:
1617 case ALGORITHM_ROTATING_N_CONTINUE:
1618 if (sh->pd_idx == 0)
1619 i--; /* P D D D Q */
1620 else if (i > sh->pd_idx)
1621 i -= 2; /* D D Q P D */
1623 case ALGORITHM_LEFT_ASYMMETRIC_6:
1624 case ALGORITHM_RIGHT_ASYMMETRIC_6:
1628 case ALGORITHM_LEFT_SYMMETRIC_6:
1629 case ALGORITHM_RIGHT_SYMMETRIC_6:
1631 i += data_disks + 1;
1632 i -= (sh->pd_idx + 1);
1634 case ALGORITHM_PARITY_0_6:
1638 printk(KERN_CRIT "raid6: unsupported algorithm %d\n",
1645 chunk_number = stripe * data_disks + i;
1646 r_sector = (sector_t)chunk_number * sectors_per_chunk + chunk_offset;
1648 check = raid5_compute_sector(conf, r_sector,
1649 previous, &dummy1, &sh2);
1650 if (check != sh->sector || dummy1 != dd_idx || sh2.pd_idx != sh->pd_idx
1651 || sh2.qd_idx != sh->qd_idx) {
1652 printk(KERN_ERR "compute_blocknr: map not correct\n");
1661 * Copy data between a page in the stripe cache, and one or more bion
1662 * The page could align with the middle of the bio, or there could be
1663 * several bion, each with several bio_vecs, which cover part of the page
1664 * Multiple bion are linked together on bi_next. There may be extras
1665 * at the end of this list. We ignore them.
1667 static void copy_data(int frombio, struct bio *bio,
1671 char *pa = page_address(page);
1672 struct bio_vec *bvl;
1676 if (bio->bi_sector >= sector)
1677 page_offset = (signed)(bio->bi_sector - sector) * 512;
1679 page_offset = (signed)(sector - bio->bi_sector) * -512;
1680 bio_for_each_segment(bvl, bio, i) {
1681 int len = bio_iovec_idx(bio,i)->bv_len;
1685 if (page_offset < 0) {
1686 b_offset = -page_offset;
1687 page_offset += b_offset;
1691 if (len > 0 && page_offset + len > STRIPE_SIZE)
1692 clen = STRIPE_SIZE - page_offset;
1696 char *ba = __bio_kmap_atomic(bio, i, KM_USER0);
1698 memcpy(pa+page_offset, ba+b_offset, clen);
1700 memcpy(ba+b_offset, pa+page_offset, clen);
1701 __bio_kunmap_atomic(ba, KM_USER0);
1703 if (clen < len) /* hit end of page */
1709 #define check_xor() do { \
1710 if (count == MAX_XOR_BLOCKS) { \
1711 xor_blocks(count, STRIPE_SIZE, dest, ptr);\
1716 static void compute_parity6(struct stripe_head *sh, int method)
1718 raid5_conf_t *conf = sh->raid_conf;
1719 int i, pd_idx, qd_idx, d0_idx, disks = sh->disks, count;
1720 int syndrome_disks = sh->ddf_layout ? disks : (disks - 2);
1722 /**** FIX THIS: This could be very bad if disks is close to 256 ****/
1723 void *ptrs[syndrome_disks+2];
1725 pd_idx = sh->pd_idx;
1726 qd_idx = sh->qd_idx;
1727 d0_idx = raid6_d0(sh);
1729 pr_debug("compute_parity, stripe %llu, method %d\n",
1730 (unsigned long long)sh->sector, method);
1733 case READ_MODIFY_WRITE:
1734 BUG(); /* READ_MODIFY_WRITE N/A for RAID-6 */
1735 case RECONSTRUCT_WRITE:
1736 for (i= disks; i-- ;)
1737 if ( i != pd_idx && i != qd_idx && sh->dev[i].towrite ) {
1738 chosen = sh->dev[i].towrite;
1739 sh->dev[i].towrite = NULL;
1741 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
1742 wake_up(&conf->wait_for_overlap);
1744 BUG_ON(sh->dev[i].written);
1745 sh->dev[i].written = chosen;
1749 BUG(); /* Not implemented yet */
1752 for (i = disks; i--;)
1753 if (sh->dev[i].written) {
1754 sector_t sector = sh->dev[i].sector;
1755 struct bio *wbi = sh->dev[i].written;
1756 while (wbi && wbi->bi_sector < sector + STRIPE_SECTORS) {
1757 copy_data(1, wbi, sh->dev[i].page, sector);
1758 wbi = r5_next_bio(wbi, sector);
1761 set_bit(R5_LOCKED, &sh->dev[i].flags);
1762 set_bit(R5_UPTODATE, &sh->dev[i].flags);
1765 /* Note that unlike RAID-5, the ordering of the disks matters greatly.*/
1767 for (i = 0; i < disks; i++)
1768 ptrs[i] = (void *)raid6_empty_zero_page;
1773 int slot = raid6_idx_to_slot(i, sh, &count, syndrome_disks);
1775 ptrs[slot] = page_address(sh->dev[i].page);
1776 if (slot < syndrome_disks &&
1777 !test_bit(R5_UPTODATE, &sh->dev[i].flags)) {
1778 printk(KERN_ERR "block %d/%d not uptodate "
1779 "on parity calc\n", i, count);
1783 i = raid6_next_disk(i, disks);
1784 } while (i != d0_idx);
1785 BUG_ON(count != syndrome_disks);
1787 raid6_call.gen_syndrome(syndrome_disks+2, STRIPE_SIZE, ptrs);
1790 case RECONSTRUCT_WRITE:
1791 set_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
1792 set_bit(R5_UPTODATE, &sh->dev[qd_idx].flags);
1793 set_bit(R5_LOCKED, &sh->dev[pd_idx].flags);
1794 set_bit(R5_LOCKED, &sh->dev[qd_idx].flags);
1797 set_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
1798 set_bit(R5_UPTODATE, &sh->dev[qd_idx].flags);
1804 /* Compute one missing block */
1805 static void compute_block_1(struct stripe_head *sh, int dd_idx, int nozero)
1807 int i, count, disks = sh->disks;
1808 void *ptr[MAX_XOR_BLOCKS], *dest, *p;
1809 int qd_idx = sh->qd_idx;
1811 pr_debug("compute_block_1, stripe %llu, idx %d\n",
1812 (unsigned long long)sh->sector, dd_idx);
1814 if ( dd_idx == qd_idx ) {
1815 /* We're actually computing the Q drive */
1816 compute_parity6(sh, UPDATE_PARITY);
1818 dest = page_address(sh->dev[dd_idx].page);
1819 if (!nozero) memset(dest, 0, STRIPE_SIZE);
1821 for (i = disks ; i--; ) {
1822 if (i == dd_idx || i == qd_idx)
1824 p = page_address(sh->dev[i].page);
1825 if (test_bit(R5_UPTODATE, &sh->dev[i].flags))
1828 printk("compute_block() %d, stripe %llu, %d"
1829 " not present\n", dd_idx,
1830 (unsigned long long)sh->sector, i);
1835 xor_blocks(count, STRIPE_SIZE, dest, ptr);
1836 if (!nozero) set_bit(R5_UPTODATE, &sh->dev[dd_idx].flags);
1837 else clear_bit(R5_UPTODATE, &sh->dev[dd_idx].flags);
1841 /* Compute two missing blocks */
1842 static void compute_block_2(struct stripe_head *sh, int dd_idx1, int dd_idx2)
1844 int i, count, disks = sh->disks;
1845 int syndrome_disks = sh->ddf_layout ? disks : disks-2;
1846 int d0_idx = raid6_d0(sh);
1847 int faila = -1, failb = -1;
1848 /**** FIX THIS: This could be very bad if disks is close to 256 ****/
1849 void *ptrs[syndrome_disks+2];
1851 for (i = 0; i < disks ; i++)
1852 ptrs[i] = (void *)raid6_empty_zero_page;
1856 int slot = raid6_idx_to_slot(i, sh, &count, syndrome_disks);
1858 ptrs[slot] = page_address(sh->dev[i].page);
1864 i = raid6_next_disk(i, disks);
1865 } while (i != d0_idx);
1866 BUG_ON(count != syndrome_disks);
1868 BUG_ON(faila == failb);
1869 if ( failb < faila ) { int tmp = faila; faila = failb; failb = tmp; }
1871 pr_debug("compute_block_2, stripe %llu, idx %d,%d (%d,%d)\n",
1872 (unsigned long long)sh->sector, dd_idx1, dd_idx2,
1875 if (failb == syndrome_disks+1) {
1876 /* Q disk is one of the missing disks */
1877 if (faila == syndrome_disks) {
1878 /* Missing P+Q, just recompute */
1879 compute_parity6(sh, UPDATE_PARITY);
1882 /* We're missing D+Q; recompute D from P */
1883 compute_block_1(sh, ((dd_idx1 == sh->qd_idx) ?
1886 compute_parity6(sh, UPDATE_PARITY); /* Is this necessary? */
1891 /* We're missing D+P or D+D; */
1892 if (failb == syndrome_disks) {
1893 /* We're missing D+P. */
1894 raid6_datap_recov(syndrome_disks+2, STRIPE_SIZE, faila, ptrs);
1896 /* We're missing D+D. */
1897 raid6_2data_recov(syndrome_disks+2, STRIPE_SIZE, faila, failb,
1901 /* Both the above update both missing blocks */
1902 set_bit(R5_UPTODATE, &sh->dev[dd_idx1].flags);
1903 set_bit(R5_UPTODATE, &sh->dev[dd_idx2].flags);
1907 schedule_reconstruction5(struct stripe_head *sh, struct stripe_head_state *s,
1908 int rcw, int expand)
1910 int i, pd_idx = sh->pd_idx, disks = sh->disks;
1913 /* if we are not expanding this is a proper write request, and
1914 * there will be bios with new data to be drained into the
1918 sh->reconstruct_state = reconstruct_state_drain_run;
1919 set_bit(STRIPE_OP_BIODRAIN, &s->ops_request);
1921 sh->reconstruct_state = reconstruct_state_run;
1923 set_bit(STRIPE_OP_POSTXOR, &s->ops_request);
1925 for (i = disks; i--; ) {
1926 struct r5dev *dev = &sh->dev[i];
1929 set_bit(R5_LOCKED, &dev->flags);
1930 set_bit(R5_Wantdrain, &dev->flags);
1932 clear_bit(R5_UPTODATE, &dev->flags);
1936 if (s->locked + 1 == disks)
1937 if (!test_and_set_bit(STRIPE_FULL_WRITE, &sh->state))
1938 atomic_inc(&sh->raid_conf->pending_full_writes);
1940 BUG_ON(!(test_bit(R5_UPTODATE, &sh->dev[pd_idx].flags) ||
1941 test_bit(R5_Wantcompute, &sh->dev[pd_idx].flags)));
1943 sh->reconstruct_state = reconstruct_state_prexor_drain_run;
1944 set_bit(STRIPE_OP_PREXOR, &s->ops_request);
1945 set_bit(STRIPE_OP_BIODRAIN, &s->ops_request);
1946 set_bit(STRIPE_OP_POSTXOR, &s->ops_request);
1948 for (i = disks; i--; ) {
1949 struct r5dev *dev = &sh->dev[i];
1954 (test_bit(R5_UPTODATE, &dev->flags) ||
1955 test_bit(R5_Wantcompute, &dev->flags))) {
1956 set_bit(R5_Wantdrain, &dev->flags);
1957 set_bit(R5_LOCKED, &dev->flags);
1958 clear_bit(R5_UPTODATE, &dev->flags);
1964 /* keep the parity disk locked while asynchronous operations
1967 set_bit(R5_LOCKED, &sh->dev[pd_idx].flags);
1968 clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
1971 pr_debug("%s: stripe %llu locked: %d ops_request: %lx\n",
1972 __func__, (unsigned long long)sh->sector,
1973 s->locked, s->ops_request);
1977 * Each stripe/dev can have one or more bion attached.
1978 * toread/towrite point to the first in a chain.
1979 * The bi_next chain must be in order.
1981 static int add_stripe_bio(struct stripe_head *sh, struct bio *bi, int dd_idx, int forwrite)
1984 raid5_conf_t *conf = sh->raid_conf;
1987 pr_debug("adding bh b#%llu to stripe s#%llu\n",
1988 (unsigned long long)bi->bi_sector,
1989 (unsigned long long)sh->sector);
1992 spin_lock(&sh->lock);
1993 spin_lock_irq(&conf->device_lock);
1995 bip = &sh->dev[dd_idx].towrite;
1996 if (*bip == NULL && sh->dev[dd_idx].written == NULL)
1999 bip = &sh->dev[dd_idx].toread;
2000 while (*bip && (*bip)->bi_sector < bi->bi_sector) {
2001 if ((*bip)->bi_sector + ((*bip)->bi_size >> 9) > bi->bi_sector)
2003 bip = & (*bip)->bi_next;
2005 if (*bip && (*bip)->bi_sector < bi->bi_sector + ((bi->bi_size)>>9))
2008 BUG_ON(*bip && bi->bi_next && (*bip) != bi->bi_next);
2012 bi->bi_phys_segments++;
2013 spin_unlock_irq(&conf->device_lock);
2014 spin_unlock(&sh->lock);
2016 pr_debug("added bi b#%llu to stripe s#%llu, disk %d.\n",
2017 (unsigned long long)bi->bi_sector,
2018 (unsigned long long)sh->sector, dd_idx);
2020 if (conf->mddev->bitmap && firstwrite) {
2021 bitmap_startwrite(conf->mddev->bitmap, sh->sector,
2023 sh->bm_seq = conf->seq_flush+1;
2024 set_bit(STRIPE_BIT_DELAY, &sh->state);
2028 /* check if page is covered */
2029 sector_t sector = sh->dev[dd_idx].sector;
2030 for (bi=sh->dev[dd_idx].towrite;
2031 sector < sh->dev[dd_idx].sector + STRIPE_SECTORS &&
2032 bi && bi->bi_sector <= sector;
2033 bi = r5_next_bio(bi, sh->dev[dd_idx].sector)) {
2034 if (bi->bi_sector + (bi->bi_size>>9) >= sector)
2035 sector = bi->bi_sector + (bi->bi_size>>9);
2037 if (sector >= sh->dev[dd_idx].sector + STRIPE_SECTORS)
2038 set_bit(R5_OVERWRITE, &sh->dev[dd_idx].flags);
2043 set_bit(R5_Overlap, &sh->dev[dd_idx].flags);
2044 spin_unlock_irq(&conf->device_lock);
2045 spin_unlock(&sh->lock);
2049 static void end_reshape(raid5_conf_t *conf);
2051 static int page_is_zero(struct page *p)
2053 char *a = page_address(p);
2054 return ((*(u32*)a) == 0 &&
2055 memcmp(a, a+4, STRIPE_SIZE-4)==0);
2058 static void stripe_set_idx(sector_t stripe, raid5_conf_t *conf, int previous,
2059 struct stripe_head *sh)
2061 int sectors_per_chunk =
2062 previous ? (conf->prev_chunk >> 9)
2063 : (conf->chunk_size >> 9);
2065 int chunk_offset = sector_div(stripe, sectors_per_chunk);
2066 int disks = previous ? conf->previous_raid_disks : conf->raid_disks;
2068 raid5_compute_sector(conf,
2069 stripe * (disks - conf->max_degraded)
2070 *sectors_per_chunk + chunk_offset,
2076 handle_failed_stripe(raid5_conf_t *conf, struct stripe_head *sh,
2077 struct stripe_head_state *s, int disks,
2078 struct bio **return_bi)
2081 for (i = disks; i--; ) {
2085 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
2088 rdev = rcu_dereference(conf->disks[i].rdev);
2089 if (rdev && test_bit(In_sync, &rdev->flags))
2090 /* multiple read failures in one stripe */
2091 md_error(conf->mddev, rdev);
2094 spin_lock_irq(&conf->device_lock);
2095 /* fail all writes first */
2096 bi = sh->dev[i].towrite;
2097 sh->dev[i].towrite = NULL;
2103 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
2104 wake_up(&conf->wait_for_overlap);
2106 while (bi && bi->bi_sector <
2107 sh->dev[i].sector + STRIPE_SECTORS) {
2108 struct bio *nextbi = r5_next_bio(bi, sh->dev[i].sector);
2109 clear_bit(BIO_UPTODATE, &bi->bi_flags);
2110 if (!raid5_dec_bi_phys_segments(bi)) {
2111 md_write_end(conf->mddev);
2112 bi->bi_next = *return_bi;
2117 /* and fail all 'written' */
2118 bi = sh->dev[i].written;
2119 sh->dev[i].written = NULL;
2120 if (bi) bitmap_end = 1;
2121 while (bi && bi->bi_sector <
2122 sh->dev[i].sector + STRIPE_SECTORS) {
2123 struct bio *bi2 = r5_next_bio(bi, sh->dev[i].sector);
2124 clear_bit(BIO_UPTODATE, &bi->bi_flags);
2125 if (!raid5_dec_bi_phys_segments(bi)) {
2126 md_write_end(conf->mddev);
2127 bi->bi_next = *return_bi;
2133 /* fail any reads if this device is non-operational and
2134 * the data has not reached the cache yet.
2136 if (!test_bit(R5_Wantfill, &sh->dev[i].flags) &&
2137 (!test_bit(R5_Insync, &sh->dev[i].flags) ||
2138 test_bit(R5_ReadError, &sh->dev[i].flags))) {
2139 bi = sh->dev[i].toread;
2140 sh->dev[i].toread = NULL;
2141 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
2142 wake_up(&conf->wait_for_overlap);
2143 if (bi) s->to_read--;
2144 while (bi && bi->bi_sector <
2145 sh->dev[i].sector + STRIPE_SECTORS) {
2146 struct bio *nextbi =
2147 r5_next_bio(bi, sh->dev[i].sector);
2148 clear_bit(BIO_UPTODATE, &bi->bi_flags);
2149 if (!raid5_dec_bi_phys_segments(bi)) {
2150 bi->bi_next = *return_bi;
2156 spin_unlock_irq(&conf->device_lock);
2158 bitmap_endwrite(conf->mddev->bitmap, sh->sector,
2159 STRIPE_SECTORS, 0, 0);
2162 if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
2163 if (atomic_dec_and_test(&conf->pending_full_writes))
2164 md_wakeup_thread(conf->mddev->thread);
2167 /* fetch_block5 - checks the given member device to see if its data needs
2168 * to be read or computed to satisfy a request.
2170 * Returns 1 when no more member devices need to be checked, otherwise returns
2171 * 0 to tell the loop in handle_stripe_fill5 to continue
2173 static int fetch_block5(struct stripe_head *sh, struct stripe_head_state *s,
2174 int disk_idx, int disks)
2176 struct r5dev *dev = &sh->dev[disk_idx];
2177 struct r5dev *failed_dev = &sh->dev[s->failed_num];
2179 /* is the data in this block needed, and can we get it? */
2180 if (!test_bit(R5_LOCKED, &dev->flags) &&
2181 !test_bit(R5_UPTODATE, &dev->flags) &&
2183 (dev->towrite && !test_bit(R5_OVERWRITE, &dev->flags)) ||
2184 s->syncing || s->expanding ||
2186 (failed_dev->toread ||
2187 (failed_dev->towrite &&
2188 !test_bit(R5_OVERWRITE, &failed_dev->flags)))))) {
2189 /* We would like to get this block, possibly by computing it,
2190 * otherwise read it if the backing disk is insync
2192 if ((s->uptodate == disks - 1) &&
2193 (s->failed && disk_idx == s->failed_num)) {
2194 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
2195 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
2196 set_bit(R5_Wantcompute, &dev->flags);
2197 sh->ops.target = disk_idx;
2199 /* Careful: from this point on 'uptodate' is in the eye
2200 * of raid5_run_ops which services 'compute' operations
2201 * before writes. R5_Wantcompute flags a block that will
2202 * be R5_UPTODATE by the time it is needed for a
2203 * subsequent operation.
2206 return 1; /* uptodate + compute == disks */
2207 } else if (test_bit(R5_Insync, &dev->flags)) {
2208 set_bit(R5_LOCKED, &dev->flags);
2209 set_bit(R5_Wantread, &dev->flags);
2211 pr_debug("Reading block %d (sync=%d)\n", disk_idx,
2220 * handle_stripe_fill5 - read or compute data to satisfy pending requests.
2222 static void handle_stripe_fill5(struct stripe_head *sh,
2223 struct stripe_head_state *s, int disks)
2227 /* look for blocks to read/compute, skip this if a compute
2228 * is already in flight, or if the stripe contents are in the
2229 * midst of changing due to a write
2231 if (!test_bit(STRIPE_COMPUTE_RUN, &sh->state) && !sh->check_state &&
2232 !sh->reconstruct_state)
2233 for (i = disks; i--; )
2234 if (fetch_block5(sh, s, i, disks))
2236 set_bit(STRIPE_HANDLE, &sh->state);
2239 static void handle_stripe_fill6(struct stripe_head *sh,
2240 struct stripe_head_state *s, struct r6_state *r6s,
2244 for (i = disks; i--; ) {
2245 struct r5dev *dev = &sh->dev[i];
2246 if (!test_bit(R5_LOCKED, &dev->flags) &&
2247 !test_bit(R5_UPTODATE, &dev->flags) &&
2248 (dev->toread || (dev->towrite &&
2249 !test_bit(R5_OVERWRITE, &dev->flags)) ||
2250 s->syncing || s->expanding ||
2252 (sh->dev[r6s->failed_num[0]].toread ||
2255 (sh->dev[r6s->failed_num[1]].toread ||
2257 /* we would like to get this block, possibly
2258 * by computing it, but we might not be able to
2260 if ((s->uptodate == disks - 1) &&
2261 (s->failed && (i == r6s->failed_num[0] ||
2262 i == r6s->failed_num[1]))) {
2263 pr_debug("Computing stripe %llu block %d\n",
2264 (unsigned long long)sh->sector, i);
2265 compute_block_1(sh, i, 0);
2267 } else if ( s->uptodate == disks-2 && s->failed >= 2 ) {
2268 /* Computing 2-failure is *very* expensive; only
2269 * do it if failed >= 2
2272 for (other = disks; other--; ) {
2275 if (!test_bit(R5_UPTODATE,
2276 &sh->dev[other].flags))
2280 pr_debug("Computing stripe %llu blocks %d,%d\n",
2281 (unsigned long long)sh->sector,
2283 compute_block_2(sh, i, other);
2285 } else if (test_bit(R5_Insync, &dev->flags)) {
2286 set_bit(R5_LOCKED, &dev->flags);
2287 set_bit(R5_Wantread, &dev->flags);
2289 pr_debug("Reading block %d (sync=%d)\n",
2294 set_bit(STRIPE_HANDLE, &sh->state);
2298 /* handle_stripe_clean_event
2299 * any written block on an uptodate or failed drive can be returned.
2300 * Note that if we 'wrote' to a failed drive, it will be UPTODATE, but
2301 * never LOCKED, so we don't need to test 'failed' directly.
2303 static void handle_stripe_clean_event(raid5_conf_t *conf,
2304 struct stripe_head *sh, int disks, struct bio **return_bi)
2309 for (i = disks; i--; )
2310 if (sh->dev[i].written) {
2312 if (!test_bit(R5_LOCKED, &dev->flags) &&
2313 test_bit(R5_UPTODATE, &dev->flags)) {
2314 /* We can return any write requests */
2315 struct bio *wbi, *wbi2;
2317 pr_debug("Return write for disc %d\n", i);
2318 spin_lock_irq(&conf->device_lock);
2320 dev->written = NULL;
2321 while (wbi && wbi->bi_sector <
2322 dev->sector + STRIPE_SECTORS) {
2323 wbi2 = r5_next_bio(wbi, dev->sector);
2324 if (!raid5_dec_bi_phys_segments(wbi)) {
2325 md_write_end(conf->mddev);
2326 wbi->bi_next = *return_bi;
2331 if (dev->towrite == NULL)
2333 spin_unlock_irq(&conf->device_lock);
2335 bitmap_endwrite(conf->mddev->bitmap,
2338 !test_bit(STRIPE_DEGRADED, &sh->state),
2343 if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
2344 if (atomic_dec_and_test(&conf->pending_full_writes))
2345 md_wakeup_thread(conf->mddev->thread);
2348 static void handle_stripe_dirtying5(raid5_conf_t *conf,
2349 struct stripe_head *sh, struct stripe_head_state *s, int disks)
2351 int rmw = 0, rcw = 0, i;
2352 for (i = disks; i--; ) {
2353 /* would I have to read this buffer for read_modify_write */
2354 struct r5dev *dev = &sh->dev[i];
2355 if ((dev->towrite || i == sh->pd_idx) &&
2356 !test_bit(R5_LOCKED, &dev->flags) &&
2357 !(test_bit(R5_UPTODATE, &dev->flags) ||
2358 test_bit(R5_Wantcompute, &dev->flags))) {
2359 if (test_bit(R5_Insync, &dev->flags))
2362 rmw += 2*disks; /* cannot read it */
2364 /* Would I have to read this buffer for reconstruct_write */
2365 if (!test_bit(R5_OVERWRITE, &dev->flags) && i != sh->pd_idx &&
2366 !test_bit(R5_LOCKED, &dev->flags) &&
2367 !(test_bit(R5_UPTODATE, &dev->flags) ||
2368 test_bit(R5_Wantcompute, &dev->flags))) {
2369 if (test_bit(R5_Insync, &dev->flags)) rcw++;
2374 pr_debug("for sector %llu, rmw=%d rcw=%d\n",
2375 (unsigned long long)sh->sector, rmw, rcw);
2376 set_bit(STRIPE_HANDLE, &sh->state);
2377 if (rmw < rcw && rmw > 0)
2378 /* prefer read-modify-write, but need to get some data */
2379 for (i = disks; i--; ) {
2380 struct r5dev *dev = &sh->dev[i];
2381 if ((dev->towrite || i == sh->pd_idx) &&
2382 !test_bit(R5_LOCKED, &dev->flags) &&
2383 !(test_bit(R5_UPTODATE, &dev->flags) ||
2384 test_bit(R5_Wantcompute, &dev->flags)) &&
2385 test_bit(R5_Insync, &dev->flags)) {
2387 test_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2388 pr_debug("Read_old block "
2389 "%d for r-m-w\n", i);
2390 set_bit(R5_LOCKED, &dev->flags);
2391 set_bit(R5_Wantread, &dev->flags);
2394 set_bit(STRIPE_DELAYED, &sh->state);
2395 set_bit(STRIPE_HANDLE, &sh->state);
2399 if (rcw <= rmw && rcw > 0)
2400 /* want reconstruct write, but need to get some data */
2401 for (i = disks; i--; ) {
2402 struct r5dev *dev = &sh->dev[i];
2403 if (!test_bit(R5_OVERWRITE, &dev->flags) &&
2405 !test_bit(R5_LOCKED, &dev->flags) &&
2406 !(test_bit(R5_UPTODATE, &dev->flags) ||
2407 test_bit(R5_Wantcompute, &dev->flags)) &&
2408 test_bit(R5_Insync, &dev->flags)) {
2410 test_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2411 pr_debug("Read_old block "
2412 "%d for Reconstruct\n", i);
2413 set_bit(R5_LOCKED, &dev->flags);
2414 set_bit(R5_Wantread, &dev->flags);
2417 set_bit(STRIPE_DELAYED, &sh->state);
2418 set_bit(STRIPE_HANDLE, &sh->state);
2422 /* now if nothing is locked, and if we have enough data,
2423 * we can start a write request
2425 /* since handle_stripe can be called at any time we need to handle the
2426 * case where a compute block operation has been submitted and then a
2427 * subsequent call wants to start a write request. raid5_run_ops only
2428 * handles the case where compute block and postxor are requested
2429 * simultaneously. If this is not the case then new writes need to be
2430 * held off until the compute completes.
2432 if ((s->req_compute || !test_bit(STRIPE_COMPUTE_RUN, &sh->state)) &&
2433 (s->locked == 0 && (rcw == 0 || rmw == 0) &&
2434 !test_bit(STRIPE_BIT_DELAY, &sh->state)))
2435 schedule_reconstruction5(sh, s, rcw == 0, 0);
2438 static void handle_stripe_dirtying6(raid5_conf_t *conf,
2439 struct stripe_head *sh, struct stripe_head_state *s,
2440 struct r6_state *r6s, int disks)
2442 int rcw = 0, must_compute = 0, pd_idx = sh->pd_idx, i;
2443 int qd_idx = sh->qd_idx;
2444 for (i = disks; i--; ) {
2445 struct r5dev *dev = &sh->dev[i];
2446 /* Would I have to read this buffer for reconstruct_write */
2447 if (!test_bit(R5_OVERWRITE, &dev->flags)
2448 && i != pd_idx && i != qd_idx
2449 && (!test_bit(R5_LOCKED, &dev->flags)
2451 !test_bit(R5_UPTODATE, &dev->flags)) {
2452 if (test_bit(R5_Insync, &dev->flags)) rcw++;
2454 pr_debug("raid6: must_compute: "
2455 "disk %d flags=%#lx\n", i, dev->flags);
2460 pr_debug("for sector %llu, rcw=%d, must_compute=%d\n",
2461 (unsigned long long)sh->sector, rcw, must_compute);
2462 set_bit(STRIPE_HANDLE, &sh->state);
2465 /* want reconstruct write, but need to get some data */
2466 for (i = disks; i--; ) {
2467 struct r5dev *dev = &sh->dev[i];
2468 if (!test_bit(R5_OVERWRITE, &dev->flags)
2469 && !(s->failed == 0 && (i == pd_idx || i == qd_idx))
2470 && !test_bit(R5_LOCKED, &dev->flags) &&
2471 !test_bit(R5_UPTODATE, &dev->flags) &&
2472 test_bit(R5_Insync, &dev->flags)) {
2474 test_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2475 pr_debug("Read_old stripe %llu "
2476 "block %d for Reconstruct\n",
2477 (unsigned long long)sh->sector, i);
2478 set_bit(R5_LOCKED, &dev->flags);
2479 set_bit(R5_Wantread, &dev->flags);
2482 pr_debug("Request delayed stripe %llu "
2483 "block %d for Reconstruct\n",
2484 (unsigned long long)sh->sector, i);
2485 set_bit(STRIPE_DELAYED, &sh->state);
2486 set_bit(STRIPE_HANDLE, &sh->state);
2490 /* now if nothing is locked, and if we have enough data, we can start a
2493 if (s->locked == 0 && rcw == 0 &&
2494 !test_bit(STRIPE_BIT_DELAY, &sh->state)) {
2495 if (must_compute > 0) {
2496 /* We have failed blocks and need to compute them */
2497 switch (s->failed) {
2501 compute_block_1(sh, r6s->failed_num[0], 0);
2504 compute_block_2(sh, r6s->failed_num[0],
2505 r6s->failed_num[1]);
2507 default: /* This request should have been failed? */
2512 pr_debug("Computing parity for stripe %llu\n",
2513 (unsigned long long)sh->sector);
2514 compute_parity6(sh, RECONSTRUCT_WRITE);
2515 /* now every locked buffer is ready to be written */
2516 for (i = disks; i--; )
2517 if (test_bit(R5_LOCKED, &sh->dev[i].flags)) {
2518 pr_debug("Writing stripe %llu block %d\n",
2519 (unsigned long long)sh->sector, i);
2521 set_bit(R5_Wantwrite, &sh->dev[i].flags);
2523 if (s->locked == disks)
2524 if (!test_and_set_bit(STRIPE_FULL_WRITE, &sh->state))
2525 atomic_inc(&conf->pending_full_writes);
2526 /* after a RECONSTRUCT_WRITE, the stripe MUST be in-sync */
2527 set_bit(STRIPE_INSYNC, &sh->state);
2529 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2530 atomic_dec(&conf->preread_active_stripes);
2531 if (atomic_read(&conf->preread_active_stripes) <
2533 md_wakeup_thread(conf->mddev->thread);
2538 static void handle_parity_checks5(raid5_conf_t *conf, struct stripe_head *sh,
2539 struct stripe_head_state *s, int disks)
2541 struct r5dev *dev = NULL;
2543 set_bit(STRIPE_HANDLE, &sh->state);
2545 switch (sh->check_state) {
2546 case check_state_idle:
2547 /* start a new check operation if there are no failures */
2548 if (s->failed == 0) {
2549 BUG_ON(s->uptodate != disks);
2550 sh->check_state = check_state_run;
2551 set_bit(STRIPE_OP_CHECK, &s->ops_request);
2552 clear_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags);
2556 dev = &sh->dev[s->failed_num];
2558 case check_state_compute_result:
2559 sh->check_state = check_state_idle;
2561 dev = &sh->dev[sh->pd_idx];
2563 /* check that a write has not made the stripe insync */
2564 if (test_bit(STRIPE_INSYNC, &sh->state))
2567 /* either failed parity check, or recovery is happening */
2568 BUG_ON(!test_bit(R5_UPTODATE, &dev->flags));
2569 BUG_ON(s->uptodate != disks);
2571 set_bit(R5_LOCKED, &dev->flags);
2573 set_bit(R5_Wantwrite, &dev->flags);
2575 clear_bit(STRIPE_DEGRADED, &sh->state);
2576 set_bit(STRIPE_INSYNC, &sh->state);
2578 case check_state_run:
2579 break; /* we will be called again upon completion */
2580 case check_state_check_result:
2581 sh->check_state = check_state_idle;
2583 /* if a failure occurred during the check operation, leave
2584 * STRIPE_INSYNC not set and let the stripe be handled again
2589 /* handle a successful check operation, if parity is correct
2590 * we are done. Otherwise update the mismatch count and repair
2591 * parity if !MD_RECOVERY_CHECK
2593 if (sh->ops.zero_sum_result == 0)
2594 /* parity is correct (on disc,
2595 * not in buffer any more)
2597 set_bit(STRIPE_INSYNC, &sh->state);
2599 conf->mddev->resync_mismatches += STRIPE_SECTORS;
2600 if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery))
2601 /* don't try to repair!! */
2602 set_bit(STRIPE_INSYNC, &sh->state);
2604 sh->check_state = check_state_compute_run;
2605 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
2606 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
2607 set_bit(R5_Wantcompute,
2608 &sh->dev[sh->pd_idx].flags);
2609 sh->ops.target = sh->pd_idx;
2614 case check_state_compute_run:
2617 printk(KERN_ERR "%s: unknown check_state: %d sector: %llu\n",
2618 __func__, sh->check_state,
2619 (unsigned long long) sh->sector);
2625 static void handle_parity_checks6(raid5_conf_t *conf, struct stripe_head *sh,
2626 struct stripe_head_state *s,
2627 struct r6_state *r6s, int disks)
2629 int update_p = 0, update_q = 0;
2631 int pd_idx = sh->pd_idx;
2632 int qd_idx = sh->qd_idx;
2634 struct page *tmp_page;
2636 set_bit(STRIPE_HANDLE, &sh->state);
2638 BUG_ON(s->failed > 2);
2639 BUG_ON(s->uptodate < disks);
2640 /* Want to check and possibly repair P and Q.
2641 * However there could be one 'failed' device, in which
2642 * case we can only check one of them, possibly using the
2643 * other to generate missing data
2646 tmp_page = per_cpu_ptr(conf->percpu, cpu)->spare_page;
2647 if (s->failed == r6s->q_failed) {
2648 /* The only possible failed device holds 'Q', so it
2649 * makes sense to check P (If anything else were failed,
2650 * we would have used P to recreate it).
2652 compute_block_1(sh, pd_idx, 1);
2653 if (!page_is_zero(sh->dev[pd_idx].page)) {
2654 compute_block_1(sh, pd_idx, 0);
2658 if (!r6s->q_failed && s->failed < 2) {
2659 /* q is not failed, and we didn't use it to generate
2660 * anything, so it makes sense to check it
2662 memcpy(page_address(tmp_page),
2663 page_address(sh->dev[qd_idx].page),
2665 compute_parity6(sh, UPDATE_PARITY);
2666 if (memcmp(page_address(tmp_page),
2667 page_address(sh->dev[qd_idx].page),
2668 STRIPE_SIZE) != 0) {
2669 clear_bit(STRIPE_INSYNC, &sh->state);
2675 if (update_p || update_q) {
2676 conf->mddev->resync_mismatches += STRIPE_SECTORS;
2677 if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery))
2678 /* don't try to repair!! */
2679 update_p = update_q = 0;
2682 /* now write out any block on a failed drive,
2683 * or P or Q if they need it
2686 if (s->failed == 2) {
2687 dev = &sh->dev[r6s->failed_num[1]];
2689 set_bit(R5_LOCKED, &dev->flags);
2690 set_bit(R5_Wantwrite, &dev->flags);
2692 if (s->failed >= 1) {
2693 dev = &sh->dev[r6s->failed_num[0]];
2695 set_bit(R5_LOCKED, &dev->flags);
2696 set_bit(R5_Wantwrite, &dev->flags);
2700 dev = &sh->dev[pd_idx];
2702 set_bit(R5_LOCKED, &dev->flags);
2703 set_bit(R5_Wantwrite, &dev->flags);
2706 dev = &sh->dev[qd_idx];
2708 set_bit(R5_LOCKED, &dev->flags);
2709 set_bit(R5_Wantwrite, &dev->flags);
2711 clear_bit(STRIPE_DEGRADED, &sh->state);
2713 set_bit(STRIPE_INSYNC, &sh->state);
2716 static void handle_stripe_expansion(raid5_conf_t *conf, struct stripe_head *sh,
2717 struct r6_state *r6s)
2721 /* We have read all the blocks in this stripe and now we need to
2722 * copy some of them into a target stripe for expand.
2724 struct dma_async_tx_descriptor *tx = NULL;
2725 clear_bit(STRIPE_EXPAND_SOURCE, &sh->state);
2726 for (i = 0; i < sh->disks; i++)
2727 if (i != sh->pd_idx && i != sh->qd_idx) {
2729 struct stripe_head *sh2;
2730 struct async_submit_ctl submit;
2732 sector_t bn = compute_blocknr(sh, i, 1);
2733 sector_t s = raid5_compute_sector(conf, bn, 0,
2735 sh2 = get_active_stripe(conf, s, 0, 1);
2737 /* so far only the early blocks of this stripe
2738 * have been requested. When later blocks
2739 * get requested, we will try again
2742 if (!test_bit(STRIPE_EXPANDING, &sh2->state) ||
2743 test_bit(R5_Expanded, &sh2->dev[dd_idx].flags)) {
2744 /* must have already done this block */
2745 release_stripe(sh2);
2749 /* place all the copies on one channel */
2750 init_async_submit(&submit, 0, tx, NULL, NULL, NULL);
2751 tx = async_memcpy(sh2->dev[dd_idx].page,
2752 sh->dev[i].page, 0, 0, STRIPE_SIZE,
2755 set_bit(R5_Expanded, &sh2->dev[dd_idx].flags);
2756 set_bit(R5_UPTODATE, &sh2->dev[dd_idx].flags);
2757 for (j = 0; j < conf->raid_disks; j++)
2758 if (j != sh2->pd_idx &&
2759 (!r6s || j != sh2->qd_idx) &&
2760 !test_bit(R5_Expanded, &sh2->dev[j].flags))
2762 if (j == conf->raid_disks) {
2763 set_bit(STRIPE_EXPAND_READY, &sh2->state);
2764 set_bit(STRIPE_HANDLE, &sh2->state);
2766 release_stripe(sh2);
2769 /* done submitting copies, wait for them to complete */
2772 dma_wait_for_async_tx(tx);
2778 * handle_stripe - do things to a stripe.
2780 * We lock the stripe and then examine the state of various bits
2781 * to see what needs to be done.
2783 * return some read request which now have data
2784 * return some write requests which are safely on disc
2785 * schedule a read on some buffers
2786 * schedule a write of some buffers
2787 * return confirmation of parity correctness
2789 * buffers are taken off read_list or write_list, and bh_cache buffers
2790 * get BH_Lock set before the stripe lock is released.
2794 static bool handle_stripe5(struct stripe_head *sh)
2796 raid5_conf_t *conf = sh->raid_conf;
2797 int disks = sh->disks, i;
2798 struct bio *return_bi = NULL;
2799 struct stripe_head_state s;
2801 mdk_rdev_t *blocked_rdev = NULL;
2804 memset(&s, 0, sizeof(s));
2805 pr_debug("handling stripe %llu, state=%#lx cnt=%d, pd_idx=%d check:%d "
2806 "reconstruct:%d\n", (unsigned long long)sh->sector, sh->state,
2807 atomic_read(&sh->count), sh->pd_idx, sh->check_state,
2808 sh->reconstruct_state);
2810 spin_lock(&sh->lock);
2811 clear_bit(STRIPE_HANDLE, &sh->state);
2812 clear_bit(STRIPE_DELAYED, &sh->state);
2814 s.syncing = test_bit(STRIPE_SYNCING, &sh->state);
2815 s.expanding = test_bit(STRIPE_EXPAND_SOURCE, &sh->state);
2816 s.expanded = test_bit(STRIPE_EXPAND_READY, &sh->state);
2818 /* Now to look around and see what can be done */
2820 for (i=disks; i--; ) {
2822 struct r5dev *dev = &sh->dev[i];
2823 clear_bit(R5_Insync, &dev->flags);
2825 pr_debug("check %d: state 0x%lx toread %p read %p write %p "
2826 "written %p\n", i, dev->flags, dev->toread, dev->read,
2827 dev->towrite, dev->written);
2829 /* maybe we can request a biofill operation
2831 * new wantfill requests are only permitted while
2832 * ops_complete_biofill is guaranteed to be inactive
2834 if (test_bit(R5_UPTODATE, &dev->flags) && dev->toread &&
2835 !test_bit(STRIPE_BIOFILL_RUN, &sh->state))
2836 set_bit(R5_Wantfill, &dev->flags);
2838 /* now count some things */
2839 if (test_bit(R5_LOCKED, &dev->flags)) s.locked++;
2840 if (test_bit(R5_UPTODATE, &dev->flags)) s.uptodate++;
2841 if (test_bit(R5_Wantcompute, &dev->flags)) s.compute++;
2843 if (test_bit(R5_Wantfill, &dev->flags))
2845 else if (dev->toread)
2849 if (!test_bit(R5_OVERWRITE, &dev->flags))
2854 rdev = rcu_dereference(conf->disks[i].rdev);
2855 if (blocked_rdev == NULL &&
2856 rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
2857 blocked_rdev = rdev;
2858 atomic_inc(&rdev->nr_pending);
2860 if (!rdev || !test_bit(In_sync, &rdev->flags)) {
2861 /* The ReadError flag will just be confusing now */
2862 clear_bit(R5_ReadError, &dev->flags);
2863 clear_bit(R5_ReWrite, &dev->flags);
2865 if (!rdev || !test_bit(In_sync, &rdev->flags)
2866 || test_bit(R5_ReadError, &dev->flags)) {
2870 set_bit(R5_Insync, &dev->flags);
2874 if (unlikely(blocked_rdev)) {
2875 if (s.syncing || s.expanding || s.expanded ||
2876 s.to_write || s.written) {
2877 set_bit(STRIPE_HANDLE, &sh->state);
2880 /* There is nothing for the blocked_rdev to block */
2881 rdev_dec_pending(blocked_rdev, conf->mddev);
2882 blocked_rdev = NULL;
2885 if (s.to_fill && !test_bit(STRIPE_BIOFILL_RUN, &sh->state)) {
2886 set_bit(STRIPE_OP_BIOFILL, &s.ops_request);
2887 set_bit(STRIPE_BIOFILL_RUN, &sh->state);
2890 pr_debug("locked=%d uptodate=%d to_read=%d"
2891 " to_write=%d failed=%d failed_num=%d\n",
2892 s.locked, s.uptodate, s.to_read, s.to_write,
2893 s.failed, s.failed_num);
2894 /* check if the array has lost two devices and, if so, some requests might
2897 if (s.failed > 1 && s.to_read+s.to_write+s.written)
2898 handle_failed_stripe(conf, sh, &s, disks, &return_bi);
2899 if (s.failed > 1 && s.syncing) {
2900 md_done_sync(conf->mddev, STRIPE_SECTORS,0);
2901 clear_bit(STRIPE_SYNCING, &sh->state);
2905 /* might be able to return some write requests if the parity block
2906 * is safe, or on a failed drive
2908 dev = &sh->dev[sh->pd_idx];
2910 ((test_bit(R5_Insync, &dev->flags) &&
2911 !test_bit(R5_LOCKED, &dev->flags) &&
2912 test_bit(R5_UPTODATE, &dev->flags)) ||
2913 (s.failed == 1 && s.failed_num == sh->pd_idx)))
2914 handle_stripe_clean_event(conf, sh, disks, &return_bi);
2916 /* Now we might consider reading some blocks, either to check/generate
2917 * parity, or to satisfy requests
2918 * or to load a block that is being partially written.
2920 if (s.to_read || s.non_overwrite ||
2921 (s.syncing && (s.uptodate + s.compute < disks)) || s.expanding)
2922 handle_stripe_fill5(sh, &s, disks);
2924 /* Now we check to see if any write operations have recently
2928 if (sh->reconstruct_state == reconstruct_state_prexor_drain_result)
2930 if (sh->reconstruct_state == reconstruct_state_drain_result ||
2931 sh->reconstruct_state == reconstruct_state_prexor_drain_result) {
2932 sh->reconstruct_state = reconstruct_state_idle;
2934 /* All the 'written' buffers and the parity block are ready to
2935 * be written back to disk
2937 BUG_ON(!test_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags));
2938 for (i = disks; i--; ) {
2940 if (test_bit(R5_LOCKED, &dev->flags) &&
2941 (i == sh->pd_idx || dev->written)) {
2942 pr_debug("Writing block %d\n", i);
2943 set_bit(R5_Wantwrite, &dev->flags);
2946 if (!test_bit(R5_Insync, &dev->flags) ||
2947 (i == sh->pd_idx && s.failed == 0))
2948 set_bit(STRIPE_INSYNC, &sh->state);
2951 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2952 atomic_dec(&conf->preread_active_stripes);
2953 if (atomic_read(&conf->preread_active_stripes) <
2955 md_wakeup_thread(conf->mddev->thread);
2959 /* Now to consider new write requests and what else, if anything
2960 * should be read. We do not handle new writes when:
2961 * 1/ A 'write' operation (copy+xor) is already in flight.
2962 * 2/ A 'check' operation is in flight, as it may clobber the parity
2965 if (s.to_write && !sh->reconstruct_state && !sh->check_state)
2966 handle_stripe_dirtying5(conf, sh, &s, disks);
2968 /* maybe we need to check and possibly fix the parity for this stripe
2969 * Any reads will already have been scheduled, so we just see if enough
2970 * data is available. The parity check is held off while parity
2971 * dependent operations are in flight.
2973 if (sh->check_state ||
2974 (s.syncing && s.locked == 0 &&
2975 !test_bit(STRIPE_COMPUTE_RUN, &sh->state) &&
2976 !test_bit(STRIPE_INSYNC, &sh->state)))
2977 handle_parity_checks5(conf, sh, &s, disks);
2979 if (s.syncing && s.locked == 0 && test_bit(STRIPE_INSYNC, &sh->state)) {
2980 md_done_sync(conf->mddev, STRIPE_SECTORS,1);
2981 clear_bit(STRIPE_SYNCING, &sh->state);
2984 /* If the failed drive is just a ReadError, then we might need to progress
2985 * the repair/check process
2987 if (s.failed == 1 && !conf->mddev->ro &&
2988 test_bit(R5_ReadError, &sh->dev[s.failed_num].flags)
2989 && !test_bit(R5_LOCKED, &sh->dev[s.failed_num].flags)
2990 && test_bit(R5_UPTODATE, &sh->dev[s.failed_num].flags)
2992 dev = &sh->dev[s.failed_num];
2993 if (!test_bit(R5_ReWrite, &dev->flags)) {
2994 set_bit(R5_Wantwrite, &dev->flags);
2995 set_bit(R5_ReWrite, &dev->flags);
2996 set_bit(R5_LOCKED, &dev->flags);
2999 /* let's read it back */
3000 set_bit(R5_Wantread, &dev->flags);
3001 set_bit(R5_LOCKED, &dev->flags);
3006 /* Finish reconstruct operations initiated by the expansion process */
3007 if (sh->reconstruct_state == reconstruct_state_result) {
3008 struct stripe_head *sh2
3009 = get_active_stripe(conf, sh->sector, 1, 1);
3010 if (sh2 && test_bit(STRIPE_EXPAND_SOURCE, &sh2->state)) {
3011 /* sh cannot be written until sh2 has been read.
3012 * so arrange for sh to be delayed a little
3014 set_bit(STRIPE_DELAYED, &sh->state);
3015 set_bit(STRIPE_HANDLE, &sh->state);
3016 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE,
3018 atomic_inc(&conf->preread_active_stripes);
3019 release_stripe(sh2);
3023 release_stripe(sh2);
3025 sh->reconstruct_state = reconstruct_state_idle;
3026 clear_bit(STRIPE_EXPANDING, &sh->state);
3027 for (i = conf->raid_disks; i--; ) {
3028 set_bit(R5_Wantwrite, &sh->dev[i].flags);
3029 set_bit(R5_LOCKED, &sh->dev[i].flags);
3034 if (s.expanded && test_bit(STRIPE_EXPANDING, &sh->state) &&
3035 !sh->reconstruct_state) {
3036 /* Need to write out all blocks after computing parity */
3037 sh->disks = conf->raid_disks;
3038 stripe_set_idx(sh->sector, conf, 0, sh);
3039 schedule_reconstruction5(sh, &s, 1, 1);
3040 } else if (s.expanded && !sh->reconstruct_state && s.locked == 0) {
3041 clear_bit(STRIPE_EXPAND_READY, &sh->state);
3042 atomic_dec(&conf->reshape_stripes);
3043 wake_up(&conf->wait_for_overlap);
3044 md_done_sync(conf->mddev, STRIPE_SECTORS, 1);
3047 if (s.expanding && s.locked == 0 &&
3048 !test_bit(STRIPE_COMPUTE_RUN, &sh->state))
3049 handle_stripe_expansion(conf, sh, NULL);
3052 spin_unlock(&sh->lock);
3054 /* wait for this device to become unblocked */
3055 if (unlikely(blocked_rdev))
3056 md_wait_for_blocked_rdev(blocked_rdev, conf->mddev);
3059 raid5_run_ops(sh, s.ops_request);
3063 return_io(return_bi);
3065 return blocked_rdev == NULL;
3068 static bool handle_stripe6(struct stripe_head *sh)
3070 raid5_conf_t *conf = sh->raid_conf;
3071 int disks = sh->disks;
3072 struct bio *return_bi = NULL;
3073 int i, pd_idx = sh->pd_idx, qd_idx = sh->qd_idx;
3074 struct stripe_head_state s;
3075 struct r6_state r6s;
3076 struct r5dev *dev, *pdev, *qdev;
3077 mdk_rdev_t *blocked_rdev = NULL;
3079 pr_debug("handling stripe %llu, state=%#lx cnt=%d, "
3080 "pd_idx=%d, qd_idx=%d\n",
3081 (unsigned long long)sh->sector, sh->state,
3082 atomic_read(&sh->count), pd_idx, qd_idx);
3083 memset(&s, 0, sizeof(s));
3085 spin_lock(&sh->lock);
3086 clear_bit(STRIPE_HANDLE, &sh->state);
3087 clear_bit(STRIPE_DELAYED, &sh->state);
3089 s.syncing = test_bit(STRIPE_SYNCING, &sh->state);
3090 s.expanding = test_bit(STRIPE_EXPAND_SOURCE, &sh->state);
3091 s.expanded = test_bit(STRIPE_EXPAND_READY, &sh->state);
3092 /* Now to look around and see what can be done */
3095 for (i=disks; i--; ) {
3098 clear_bit(R5_Insync, &dev->flags);
3100 pr_debug("check %d: state 0x%lx read %p write %p written %p\n",
3101 i, dev->flags, dev->toread, dev->towrite, dev->written);
3102 /* maybe we can reply to a read */
3103 if (test_bit(R5_UPTODATE, &dev->flags) && dev->toread) {
3104 struct bio *rbi, *rbi2;
3105 pr_debug("Return read for disc %d\n", i);
3106 spin_lock_irq(&conf->device_lock);
3109 if (test_and_clear_bit(R5_Overlap, &dev->flags))
3110 wake_up(&conf->wait_for_overlap);
3111 spin_unlock_irq(&conf->device_lock);
3112 while (rbi && rbi->bi_sector < dev->sector + STRIPE_SECTORS) {
3113 copy_data(0, rbi, dev->page, dev->sector);
3114 rbi2 = r5_next_bio(rbi, dev->sector);
3115 spin_lock_irq(&conf->device_lock);
3116 if (!raid5_dec_bi_phys_segments(rbi)) {
3117 rbi->bi_next = return_bi;
3120 spin_unlock_irq(&conf->device_lock);
3125 /* now count some things */
3126 if (test_bit(R5_LOCKED, &dev->flags)) s.locked++;
3127 if (test_bit(R5_UPTODATE, &dev->flags)) s.uptodate++;
3134 if (!test_bit(R5_OVERWRITE, &dev->flags))
3139 rdev = rcu_dereference(conf->disks[i].rdev);
3140 if (blocked_rdev == NULL &&
3141 rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
3142 blocked_rdev = rdev;
3143 atomic_inc(&rdev->nr_pending);
3145 if (!rdev || !test_bit(In_sync, &rdev->flags)) {
3146 /* The ReadError flag will just be confusing now */
3147 clear_bit(R5_ReadError, &dev->flags);
3148 clear_bit(R5_ReWrite, &dev->flags);
3150 if (!rdev || !test_bit(In_sync, &rdev->flags)
3151 || test_bit(R5_ReadError, &dev->flags)) {
3153 r6s.failed_num[s.failed] = i;
3156 set_bit(R5_Insync, &dev->flags);
3160 if (unlikely(blocked_rdev)) {
3161 if (s.syncing || s.expanding || s.expanded ||
3162 s.to_write || s.written) {
3163 set_bit(STRIPE_HANDLE, &sh->state);
3166 /* There is nothing for the blocked_rdev to block */
3167 rdev_dec_pending(blocked_rdev, conf->mddev);
3168 blocked_rdev = NULL;
3171 pr_debug("locked=%d uptodate=%d to_read=%d"
3172 " to_write=%d failed=%d failed_num=%d,%d\n",
3173 s.locked, s.uptodate, s.to_read, s.to_write, s.failed,
3174 r6s.failed_num[0], r6s.failed_num[1]);
3175 /* check if the array has lost >2 devices and, if so, some requests
3176 * might need to be failed
3178 if (s.failed > 2 && s.to_read+s.to_write+s.written)
3179 handle_failed_stripe(conf, sh, &s, disks, &return_bi);
3180 if (s.failed > 2 && s.syncing) {
3181 md_done_sync(conf->mddev, STRIPE_SECTORS,0);
3182 clear_bit(STRIPE_SYNCING, &sh->state);
3187 * might be able to return some write requests if the parity blocks
3188 * are safe, or on a failed drive
3190 pdev = &sh->dev[pd_idx];
3191 r6s.p_failed = (s.failed >= 1 && r6s.failed_num[0] == pd_idx)
3192 || (s.failed >= 2 && r6s.failed_num[1] == pd_idx);
3193 qdev = &sh->dev[qd_idx];
3194 r6s.q_failed = (s.failed >= 1 && r6s.failed_num[0] == qd_idx)
3195 || (s.failed >= 2 && r6s.failed_num[1] == qd_idx);
3198 ( r6s.p_failed || ((test_bit(R5_Insync, &pdev->flags)
3199 && !test_bit(R5_LOCKED, &pdev->flags)
3200 && test_bit(R5_UPTODATE, &pdev->flags)))) &&
3201 ( r6s.q_failed || ((test_bit(R5_Insync, &qdev->flags)
3202 && !test_bit(R5_LOCKED, &qdev->flags)
3203 && test_bit(R5_UPTODATE, &qdev->flags)))))
3204 handle_stripe_clean_event(conf, sh, disks, &return_bi);
3206 /* Now we might consider reading some blocks, either to check/generate
3207 * parity, or to satisfy requests
3208 * or to load a block that is being partially written.
3210 if (s.to_read || s.non_overwrite || (s.to_write && s.failed) ||
3211 (s.syncing && (s.uptodate < disks)) || s.expanding)
3212 handle_stripe_fill6(sh, &s, &r6s, disks);
3214 /* now to consider writing and what else, if anything should be read */
3216 handle_stripe_dirtying6(conf, sh, &s, &r6s, disks);
3218 /* maybe we need to check and possibly fix the parity for this stripe
3219 * Any reads will already have been scheduled, so we just see if enough
3222 if (s.syncing && s.locked == 0 && !test_bit(STRIPE_INSYNC, &sh->state))
3223 handle_parity_checks6(conf, sh, &s, &r6s, disks);
3225 if (s.syncing && s.locked == 0 && test_bit(STRIPE_INSYNC, &sh->state)) {
3226 md_done_sync(conf->mddev, STRIPE_SECTORS,1);
3227 clear_bit(STRIPE_SYNCING, &sh->state);
3230 /* If the failed drives are just a ReadError, then we might need
3231 * to progress the repair/check process
3233 if (s.failed <= 2 && !conf->mddev->ro)
3234 for (i = 0; i < s.failed; i++) {
3235 dev = &sh->dev[r6s.failed_num[i]];
3236 if (test_bit(R5_ReadError, &dev->flags)
3237 && !test_bit(R5_LOCKED, &dev->flags)
3238 && test_bit(R5_UPTODATE, &dev->flags)
3240 if (!test_bit(R5_ReWrite, &dev->flags)) {
3241 set_bit(R5_Wantwrite, &dev->flags);
3242 set_bit(R5_ReWrite, &dev->flags);
3243 set_bit(R5_LOCKED, &dev->flags);
3245 /* let's read it back */
3246 set_bit(R5_Wantread, &dev->flags);
3247 set_bit(R5_LOCKED, &dev->flags);
3252 if (s.expanded && test_bit(STRIPE_EXPANDING, &sh->state)) {
3253 struct stripe_head *sh2
3254 = get_active_stripe(conf, sh->sector, 1, 1);
3255 if (sh2 && test_bit(STRIPE_EXPAND_SOURCE, &sh2->state)) {
3256 /* sh cannot be written until sh2 has been read.
3257 * so arrange for sh to be delayed a little
3259 set_bit(STRIPE_DELAYED, &sh->state);
3260 set_bit(STRIPE_HANDLE, &sh->state);
3261 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE,
3263 atomic_inc(&conf->preread_active_stripes);
3264 release_stripe(sh2);
3268 release_stripe(sh2);
3270 /* Need to write out all blocks after computing P&Q */
3271 sh->disks = conf->raid_disks;
3272 stripe_set_idx(sh->sector, conf, 0, sh);
3273 compute_parity6(sh, RECONSTRUCT_WRITE);
3274 for (i = conf->raid_disks ; i-- ; ) {
3275 set_bit(R5_LOCKED, &sh->dev[i].flags);
3277 set_bit(R5_Wantwrite, &sh->dev[i].flags);
3279 clear_bit(STRIPE_EXPANDING, &sh->state);
3280 } else if (s.expanded) {
3281 clear_bit(STRIPE_EXPAND_READY, &sh->state);
3282 atomic_dec(&conf->reshape_stripes);
3283 wake_up(&conf->wait_for_overlap);
3284 md_done_sync(conf->mddev, STRIPE_SECTORS, 1);
3287 if (s.expanding && s.locked == 0 &&
3288 !test_bit(STRIPE_COMPUTE_RUN, &sh->state))
3289 handle_stripe_expansion(conf, sh, &r6s);
3292 spin_unlock(&sh->lock);
3294 /* wait for this device to become unblocked */
3295 if (unlikely(blocked_rdev))
3296 md_wait_for_blocked_rdev(blocked_rdev, conf->mddev);
3300 return_io(return_bi);
3302 return blocked_rdev == NULL;
3305 /* returns true if the stripe was handled */
3306 static bool handle_stripe(struct stripe_head *sh)
3308 if (sh->raid_conf->level == 6)
3309 return handle_stripe6(sh);
3311 return handle_stripe5(sh);
3314 static void raid5_activate_delayed(raid5_conf_t *conf)
3316 if (atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD) {
3317 while (!list_empty(&conf->delayed_list)) {
3318 struct list_head *l = conf->delayed_list.next;
3319 struct stripe_head *sh;
3320 sh = list_entry(l, struct stripe_head, lru);
3322 clear_bit(STRIPE_DELAYED, &sh->state);
3323 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
3324 atomic_inc(&conf->preread_active_stripes);
3325 list_add_tail(&sh->lru, &conf->hold_list);
3328 blk_plug_device(conf->mddev->queue);
3331 static void activate_bit_delay(raid5_conf_t *conf)
3333 /* device_lock is held */
3334 struct list_head head;
3335 list_add(&head, &conf->bitmap_list);
3336 list_del_init(&conf->bitmap_list);
3337 while (!list_empty(&head)) {
3338 struct stripe_head *sh = list_entry(head.next, struct stripe_head, lru);
3339 list_del_init(&sh->lru);
3340 atomic_inc(&sh->count);
3341 __release_stripe(conf, sh);
3345 static void unplug_slaves(mddev_t *mddev)
3347 raid5_conf_t *conf = mddev_to_conf(mddev);
3351 for (i=0; i<mddev->raid_disks; i++) {
3352 mdk_rdev_t *rdev = rcu_dereference(conf->disks[i].rdev);
3353 if (rdev && !test_bit(Faulty, &rdev->flags) && atomic_read(&rdev->nr_pending)) {
3354 struct request_queue *r_queue = bdev_get_queue(rdev->bdev);
3356 atomic_inc(&rdev->nr_pending);
3359 blk_unplug(r_queue);
3361 rdev_dec_pending(rdev, mddev);
3368 static void raid5_unplug_device(struct request_queue *q)
3370 mddev_t *mddev = q->queuedata;
3371 raid5_conf_t *conf = mddev_to_conf(mddev);
3372 unsigned long flags;
3374 spin_lock_irqsave(&conf->device_lock, flags);
3376 if (blk_remove_plug(q)) {
3378 raid5_activate_delayed(conf);
3380 md_wakeup_thread(mddev->thread);
3382 spin_unlock_irqrestore(&conf->device_lock, flags);
3384 unplug_slaves(mddev);
3387 static int raid5_congested(void *data, int bits)
3389 mddev_t *mddev = data;
3390 raid5_conf_t *conf = mddev_to_conf(mddev);
3392 /* No difference between reads and writes. Just check
3393 * how busy the stripe_cache is
3395 if (conf->inactive_blocked)
3399 if (list_empty_careful(&conf->inactive_list))
3405 /* We want read requests to align with chunks where possible,
3406 * but write requests don't need to.
3408 static int raid5_mergeable_bvec(struct request_queue *q,
3409 struct bvec_merge_data *bvm,
3410 struct bio_vec *biovec)
3412 mddev_t *mddev = q->queuedata;
3413 sector_t sector = bvm->bi_sector + get_start_sect(bvm->bi_bdev);
3415 unsigned int chunk_sectors = mddev->chunk_size >> 9;
3416 unsigned int bio_sectors = bvm->bi_size >> 9;
3418 if ((bvm->bi_rw & 1) == WRITE)
3419 return biovec->bv_len; /* always allow writes to be mergeable */
3421 if (mddev->new_chunk < mddev->chunk_size)
3422 chunk_sectors = mddev->new_chunk >> 9;
3423 max = (chunk_sectors - ((sector & (chunk_sectors - 1)) + bio_sectors)) << 9;
3424 if (max < 0) max = 0;
3425 if (max <= biovec->bv_len && bio_sectors == 0)
3426 return biovec->bv_len;
3432 static int in_chunk_boundary(mddev_t *mddev, struct bio *bio)
3434 sector_t sector = bio->bi_sector + get_start_sect(bio->bi_bdev);
3435 unsigned int chunk_sectors = mddev->chunk_size >> 9;
3436 unsigned int bio_sectors = bio->bi_size >> 9;
3438 if (mddev->new_chunk < mddev->chunk_size)
3439 chunk_sectors = mddev->new_chunk >> 9;
3440 return chunk_sectors >=
3441 ((sector & (chunk_sectors - 1)) + bio_sectors);
3445 * add bio to the retry LIFO ( in O(1) ... we are in interrupt )
3446 * later sampled by raid5d.
3448 static void add_bio_to_retry(struct bio *bi,raid5_conf_t *conf)
3450 unsigned long flags;
3452 spin_lock_irqsave(&conf->device_lock, flags);
3454 bi->bi_next = conf->retry_read_aligned_list;
3455 conf->retry_read_aligned_list = bi;
3457 spin_unlock_irqrestore(&conf->device_lock, flags);
3458 md_wakeup_thread(conf->mddev->thread);
3462 static struct bio *remove_bio_from_retry(raid5_conf_t *conf)
3466 bi = conf->retry_read_aligned;
3468 conf->retry_read_aligned = NULL;
3471 bi = conf->retry_read_aligned_list;
3473 conf->retry_read_aligned_list = bi->bi_next;
3476 * this sets the active strip count to 1 and the processed
3477 * strip count to zero (upper 8 bits)
3479 bi->bi_phys_segments = 1; /* biased count of active stripes */
3487 * The "raid5_align_endio" should check if the read succeeded and if it
3488 * did, call bio_endio on the original bio (having bio_put the new bio
3490 * If the read failed..
3492 static void raid5_align_endio(struct bio *bi, int error)
3494 struct bio* raid_bi = bi->bi_private;
3497 int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
3502 mddev = raid_bi->bi_bdev->bd_disk->queue->queuedata;
3503 conf = mddev_to_conf(mddev);
3504 rdev = (void*)raid_bi->bi_next;
3505 raid_bi->bi_next = NULL;
3507 rdev_dec_pending(rdev, conf->mddev);
3509 if (!error && uptodate) {
3510 bio_endio(raid_bi, 0);
3511 if (atomic_dec_and_test(&conf->active_aligned_reads))
3512 wake_up(&conf->wait_for_stripe);
3517 pr_debug("raid5_align_endio : io error...handing IO for a retry\n");
3519 add_bio_to_retry(raid_bi, conf);
3522 static int bio_fits_rdev(struct bio *bi)
3524 struct request_queue *q = bdev_get_queue(bi->bi_bdev);
3526 if ((bi->bi_size>>9) > q->max_sectors)
3528 blk_recount_segments(q, bi);
3529 if (bi->bi_phys_segments > q->max_phys_segments)
3532 if (q->merge_bvec_fn)
3533 /* it's too hard to apply the merge_bvec_fn at this stage,
3542 static int chunk_aligned_read(struct request_queue *q, struct bio * raid_bio)
3544 mddev_t *mddev = q->queuedata;
3545 raid5_conf_t *conf = mddev_to_conf(mddev);
3546 unsigned int dd_idx;
3547 struct bio* align_bi;
3550 if (!in_chunk_boundary(mddev, raid_bio)) {
3551 pr_debug("chunk_aligned_read : non aligned\n");
3555 * use bio_clone to make a copy of the bio
3557 align_bi = bio_clone(raid_bio, GFP_NOIO);
3561 * set bi_end_io to a new function, and set bi_private to the
3564 align_bi->bi_end_io = raid5_align_endio;
3565 align_bi->bi_private = raid_bio;
3569 align_bi->bi_sector = raid5_compute_sector(conf, raid_bio->bi_sector,
3574 rdev = rcu_dereference(conf->disks[dd_idx].rdev);
3575 if (rdev && test_bit(In_sync, &rdev->flags)) {
3576 atomic_inc(&rdev->nr_pending);
3578 raid_bio->bi_next = (void*)rdev;
3579 align_bi->bi_bdev = rdev->bdev;
3580 align_bi->bi_flags &= ~(1 << BIO_SEG_VALID);
3581 align_bi->bi_sector += rdev->data_offset;
3583 if (!bio_fits_rdev(align_bi)) {
3584 /* too big in some way */
3586 rdev_dec_pending(rdev, mddev);
3590 spin_lock_irq(&conf->device_lock);
3591 wait_event_lock_irq(conf->wait_for_stripe,
3593 conf->device_lock, /* nothing */);
3594 atomic_inc(&conf->active_aligned_reads);
3595 spin_unlock_irq(&conf->device_lock);
3597 generic_make_request(align_bi);
3606 /* __get_priority_stripe - get the next stripe to process
3608 * Full stripe writes are allowed to pass preread active stripes up until
3609 * the bypass_threshold is exceeded. In general the bypass_count
3610 * increments when the handle_list is handled before the hold_list; however, it
3611 * will not be incremented when STRIPE_IO_STARTED is sampled set signifying a
3612 * stripe with in flight i/o. The bypass_count will be reset when the
3613 * head of the hold_list has changed, i.e. the head was promoted to the
3616 static struct stripe_head *__get_priority_stripe(raid5_conf_t *conf)
3618 struct stripe_head *sh;
3620 pr_debug("%s: handle: %s hold: %s full_writes: %d bypass_count: %d\n",
3622 list_empty(&conf->handle_list) ? "empty" : "busy",
3623 list_empty(&conf->hold_list) ? "empty" : "busy",
3624 atomic_read(&conf->pending_full_writes), conf->bypass_count);
3626 if (!list_empty(&conf->handle_list)) {
3627 sh = list_entry(conf->handle_list.next, typeof(*sh), lru);
3629 if (list_empty(&conf->hold_list))
3630 conf->bypass_count = 0;
3631 else if (!test_bit(STRIPE_IO_STARTED, &sh->state)) {
3632 if (conf->hold_list.next == conf->last_hold)
3633 conf->bypass_count++;
3635 conf->last_hold = conf->hold_list.next;
3636 conf->bypass_count -= conf->bypass_threshold;
3637 if (conf->bypass_count < 0)
3638 conf->bypass_count = 0;
3641 } else if (!list_empty(&conf->hold_list) &&
3642 ((conf->bypass_threshold &&
3643 conf->bypass_count > conf->bypass_threshold) ||
3644 atomic_read(&conf->pending_full_writes) == 0)) {
3645 sh = list_entry(conf->hold_list.next,
3647 conf->bypass_count -= conf->bypass_threshold;
3648 if (conf->bypass_count < 0)
3649 conf->bypass_count = 0;
3653 list_del_init(&sh->lru);
3654 atomic_inc(&sh->count);
3655 BUG_ON(atomic_read(&sh->count) != 1);
3659 static int make_request(struct request_queue *q, struct bio * bi)
3661 mddev_t *mddev = q->queuedata;
3662 raid5_conf_t *conf = mddev_to_conf(mddev);
3664 sector_t new_sector;
3665 sector_t logical_sector, last_sector;
3666 struct stripe_head *sh;
3667 const int rw = bio_data_dir(bi);
3670 if (unlikely(bio_barrier(bi))) {
3671 bio_endio(bi, -EOPNOTSUPP);
3675 md_write_start(mddev, bi);
3677 cpu = part_stat_lock();
3678 part_stat_inc(cpu, &mddev->gendisk->part0, ios[rw]);
3679 part_stat_add(cpu, &mddev->gendisk->part0, sectors[rw],
3684 mddev->reshape_position == MaxSector &&
3685 chunk_aligned_read(q,bi))
3688 logical_sector = bi->bi_sector & ~((sector_t)STRIPE_SECTORS-1);
3689 last_sector = bi->bi_sector + (bi->bi_size>>9);
3691 bi->bi_phys_segments = 1; /* over-loaded to count active stripes */
3693 for (;logical_sector < last_sector; logical_sector += STRIPE_SECTORS) {
3695 int disks, data_disks;
3700 disks = conf->raid_disks;
3701 prepare_to_wait(&conf->wait_for_overlap, &w, TASK_UNINTERRUPTIBLE);
3702 if (unlikely(conf->reshape_progress != MaxSector)) {
3703 /* spinlock is needed as reshape_progress may be
3704 * 64bit on a 32bit platform, and so it might be
3705 * possible to see a half-updated value
3706 * Ofcourse reshape_progress could change after
3707 * the lock is dropped, so once we get a reference
3708 * to the stripe that we think it is, we will have
3711 spin_lock_irq(&conf->device_lock);
3712 if (mddev->delta_disks < 0
3713 ? logical_sector < conf->reshape_progress
3714 : logical_sector >= conf->reshape_progress) {
3715 disks = conf->previous_raid_disks;
3718 if (mddev->delta_disks < 0
3719 ? logical_sector < conf->reshape_safe
3720 : logical_sector >= conf->reshape_safe) {
3721 spin_unlock_irq(&conf->device_lock);
3726 spin_unlock_irq(&conf->device_lock);
3728 data_disks = disks - conf->max_degraded;
3730 new_sector = raid5_compute_sector(conf, logical_sector,
3733 pr_debug("raid5: make_request, sector %llu logical %llu\n",
3734 (unsigned long long)new_sector,
3735 (unsigned long long)logical_sector);
3737 sh = get_active_stripe(conf, new_sector, previous,
3738 (bi->bi_rw&RWA_MASK));
3740 if (unlikely(previous)) {
3741 /* expansion might have moved on while waiting for a
3742 * stripe, so we must do the range check again.
3743 * Expansion could still move past after this
3744 * test, but as we are holding a reference to
3745 * 'sh', we know that if that happens,
3746 * STRIPE_EXPANDING will get set and the expansion
3747 * won't proceed until we finish with the stripe.
3750 spin_lock_irq(&conf->device_lock);
3751 if (mddev->delta_disks < 0
3752 ? logical_sector >= conf->reshape_progress
3753 : logical_sector < conf->reshape_progress)
3754 /* mismatch, need to try again */
3756 spin_unlock_irq(&conf->device_lock);
3762 /* FIXME what if we get a false positive because these
3763 * are being updated.
3765 if (logical_sector >= mddev->suspend_lo &&
3766 logical_sector < mddev->suspend_hi) {
3772 if (test_bit(STRIPE_EXPANDING, &sh->state) ||
3773 !add_stripe_bio(sh, bi, dd_idx, (bi->bi_rw&RW_MASK))) {
3774 /* Stripe is busy expanding or
3775 * add failed due to overlap. Flush everything
3778 raid5_unplug_device(mddev->queue);
3783 finish_wait(&conf->wait_for_overlap, &w);
3784 set_bit(STRIPE_HANDLE, &sh->state);
3785 clear_bit(STRIPE_DELAYED, &sh->state);
3788 /* cannot get stripe for read-ahead, just give-up */
3789 clear_bit(BIO_UPTODATE, &bi->bi_flags);
3790 finish_wait(&conf->wait_for_overlap, &w);
3795 spin_lock_irq(&conf->device_lock);
3796 remaining = raid5_dec_bi_phys_segments(bi);
3797 spin_unlock_irq(&conf->device_lock);
3798 if (remaining == 0) {
3801 md_write_end(mddev);
3808 static sector_t raid5_size(mddev_t *mddev, sector_t sectors, int raid_disks);
3810 static sector_t reshape_request(mddev_t *mddev, sector_t sector_nr, int *skipped)
3812 /* reshaping is quite different to recovery/resync so it is
3813 * handled quite separately ... here.
3815 * On each call to sync_request, we gather one chunk worth of
3816 * destination stripes and flag them as expanding.
3817 * Then we find all the source stripes and request reads.
3818 * As the reads complete, handle_stripe will copy the data
3819 * into the destination stripe and release that stripe.
3821 raid5_conf_t *conf = (raid5_conf_t *) mddev->private;
3822 struct stripe_head *sh;
3823 sector_t first_sector, last_sector;
3824 int raid_disks = conf->previous_raid_disks;
3825 int data_disks = raid_disks - conf->max_degraded;
3826 int new_data_disks = conf->raid_disks - conf->max_degraded;
3829 sector_t writepos, readpos, safepos;
3830 sector_t stripe_addr;
3831 int reshape_sectors;
3832 struct list_head stripes;
3834 if (sector_nr == 0) {
3835 /* If restarting in the middle, skip the initial sectors */
3836 if (mddev->delta_disks < 0 &&
3837 conf->reshape_progress < raid5_size(mddev, 0, 0)) {
3838 sector_nr = raid5_size(mddev, 0, 0)
3839 - conf->reshape_progress;
3840 } else if (mddev->delta_disks > 0 &&
3841 conf->reshape_progress > 0)
3842 sector_nr = conf->reshape_progress;
3843 sector_div(sector_nr, new_data_disks);
3850 /* We need to process a full chunk at a time.
3851 * If old and new chunk sizes differ, we need to process the
3854 if (mddev->new_chunk > mddev->chunk_size)
3855 reshape_sectors = mddev->new_chunk / 512;
3857 reshape_sectors = mddev->chunk_size / 512;
3859 /* we update the metadata when there is more than 3Meg
3860 * in the block range (that is rather arbitrary, should
3861 * probably be time based) or when the data about to be
3862 * copied would over-write the source of the data at
3863 * the front of the range.
3864 * i.e. one new_stripe along from reshape_progress new_maps
3865 * to after where reshape_safe old_maps to
3867 writepos = conf->reshape_progress;
3868 sector_div(writepos, new_data_disks);
3869 readpos = conf->reshape_progress;
3870 sector_div(readpos, data_disks);
3871 safepos = conf->reshape_safe;
3872 sector_div(safepos, data_disks);
3873 if (mddev->delta_disks < 0) {
3874 writepos -= reshape_sectors;
3875 readpos += reshape_sectors;
3876 safepos += reshape_sectors;
3878 writepos += reshape_sectors;
3879 readpos -= reshape_sectors;
3880 safepos -= reshape_sectors;
3883 /* 'writepos' is the most advanced device address we might write.
3884 * 'readpos' is the least advanced device address we might read.
3885 * 'safepos' is the least address recorded in the metadata as having
3887 * If 'readpos' is behind 'writepos', then there is no way that we can
3888 * ensure safety in the face of a crash - that must be done by userspace
3889 * making a backup of the data. So in that case there is no particular
3890 * rush to update metadata.
3891 * Otherwise if 'safepos' is behind 'writepos', then we really need to
3892 * update the metadata to advance 'safepos' to match 'readpos' so that
3893 * we can be safe in the event of a crash.
3894 * So we insist on updating metadata if safepos is behind writepos and
3895 * readpos is beyond writepos.
3896 * In any case, update the metadata every 10 seconds.
3897 * Maybe that number should be configurable, but I'm not sure it is
3898 * worth it.... maybe it could be a multiple of safemode_delay???
3900 if ((mddev->delta_disks < 0
3901 ? (safepos > writepos && readpos < writepos)
3902 : (safepos < writepos && readpos > writepos)) ||
3903 time_after(jiffies, conf->reshape_checkpoint + 10*HZ)) {
3904 /* Cannot proceed until we've updated the superblock... */
3905 wait_event(conf->wait_for_overlap,
3906 atomic_read(&conf->reshape_stripes)==0);
3907 mddev->reshape_position = conf->reshape_progress;
3908 conf->reshape_checkpoint = jiffies;
3909 set_bit(MD_CHANGE_DEVS, &mddev->flags);
3910 md_wakeup_thread(mddev->thread);
3911 wait_event(mddev->sb_wait, mddev->flags == 0 ||
3912 kthread_should_stop());
3913 spin_lock_irq(&conf->device_lock);
3914 conf->reshape_safe = mddev->reshape_position;
3915 spin_unlock_irq(&conf->device_lock);
3916 wake_up(&conf->wait_for_overlap);
3919 if (mddev->delta_disks < 0) {
3920 BUG_ON(conf->reshape_progress == 0);
3921 stripe_addr = writepos;
3922 BUG_ON((mddev->dev_sectors &
3923 ~((sector_t)reshape_sectors - 1))
3924 - reshape_sectors - stripe_addr
3927 BUG_ON(writepos != sector_nr + reshape_sectors);
3928 stripe_addr = sector_nr;
3930 INIT_LIST_HEAD(&stripes);
3931 for (i = 0; i < reshape_sectors; i += STRIPE_SECTORS) {
3934 sh = get_active_stripe(conf, stripe_addr+i, 0, 0);
3935 set_bit(STRIPE_EXPANDING, &sh->state);
3936 atomic_inc(&conf->reshape_stripes);
3937 /* If any of this stripe is beyond the end of the old
3938 * array, then we need to zero those blocks
3940 for (j=sh->disks; j--;) {
3942 if (j == sh->pd_idx)
3944 if (conf->level == 6 &&
3947 s = compute_blocknr(sh, j, 0);
3948 if (s < raid5_size(mddev, 0, 0)) {
3952 memset(page_address(sh->dev[j].page), 0, STRIPE_SIZE);
3953 set_bit(R5_Expanded, &sh->dev[j].flags);
3954 set_bit(R5_UPTODATE, &sh->dev[j].flags);
3957 set_bit(STRIPE_EXPAND_READY, &sh->state);
3958 set_bit(STRIPE_HANDLE, &sh->state);
3960 list_add(&sh->lru, &stripes);
3962 spin_lock_irq(&conf->device_lock);
3963 if (mddev->delta_disks < 0)
3964 conf->reshape_progress -= reshape_sectors * new_data_disks;
3966 conf->reshape_progress += reshape_sectors * new_data_disks;
3967 spin_unlock_irq(&conf->device_lock);
3968 /* Ok, those stripe are ready. We can start scheduling
3969 * reads on the source stripes.
3970 * The source stripes are determined by mapping the first and last
3971 * block on the destination stripes.
3974 raid5_compute_sector(conf, stripe_addr*(new_data_disks),
3977 raid5_compute_sector(conf, ((stripe_addr+conf->chunk_size/512)
3978 *(new_data_disks) - 1),
3980 if (last_sector >= mddev->dev_sectors)
3981 last_sector = mddev->dev_sectors - 1;
3982 while (first_sector <= last_sector) {
3983 sh = get_active_stripe(conf, first_sector, 1, 0);
3984 set_bit(STRIPE_EXPAND_SOURCE, &sh->state);
3985 set_bit(STRIPE_HANDLE, &sh->state);
3987 first_sector += STRIPE_SECTORS;
3989 /* Now that the sources are clearly marked, we can release
3990 * the destination stripes
3992 while (!list_empty(&stripes)) {
3993 sh = list_entry(stripes.next, struct stripe_head, lru);
3994 list_del_init(&sh->lru);
3997 /* If this takes us to the resync_max point where we have to pause,
3998 * then we need to write out the superblock.
4000 sector_nr += reshape_sectors;
4001 if (sector_nr >= mddev->resync_max) {
4002 /* Cannot proceed until we've updated the superblock... */
4003 wait_event(conf->wait_for_overlap,
4004 atomic_read(&conf->reshape_stripes) == 0);
4005 mddev->reshape_position = conf->reshape_progress;
4006 conf->reshape_checkpoint = jiffies;
4007 set_bit(MD_CHANGE_DEVS, &mddev->flags);
4008 md_wakeup_thread(mddev->thread);
4009 wait_event(mddev->sb_wait,
4010 !test_bit(MD_CHANGE_DEVS, &mddev->flags)
4011 || kthread_should_stop());
4012 spin_lock_irq(&conf->device_lock);
4013 conf->reshape_safe = mddev->reshape_position;
4014 spin_unlock_irq(&conf->device_lock);
4015 wake_up(&conf->wait_for_overlap);
4017 return reshape_sectors;
4020 /* FIXME go_faster isn't used */
4021 static inline sector_t sync_request(mddev_t *mddev, sector_t sector_nr, int *skipped, int go_faster)
4023 raid5_conf_t *conf = (raid5_conf_t *) mddev->private;
4024 struct stripe_head *sh;
4025 sector_t max_sector = mddev->dev_sectors;
4027 int still_degraded = 0;
4030 if (sector_nr >= max_sector) {
4031 /* just being told to finish up .. nothing much to do */
4032 unplug_slaves(mddev);
4034 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery)) {
4039 if (mddev->curr_resync < max_sector) /* aborted */
4040 bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
4042 else /* completed sync */
4044 bitmap_close_sync(mddev->bitmap);
4049 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
4050 return reshape_request(mddev, sector_nr, skipped);
4052 /* No need to check resync_max as we never do more than one
4053 * stripe, and as resync_max will always be on a chunk boundary,
4054 * if the check in md_do_sync didn't fire, there is no chance
4055 * of overstepping resync_max here
4058 /* if there is too many failed drives and we are trying
4059 * to resync, then assert that we are finished, because there is
4060 * nothing we can do.
4062 if (mddev->degraded >= conf->max_degraded &&
4063 test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
4064 sector_t rv = mddev->dev_sectors - sector_nr;
4068 if (!bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
4069 !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
4070 !conf->fullsync && sync_blocks >= STRIPE_SECTORS) {
4071 /* we can skip this block, and probably more */
4072 sync_blocks /= STRIPE_SECTORS;
4074 return sync_blocks * STRIPE_SECTORS; /* keep things rounded to whole stripes */
4078 bitmap_cond_end_sync(mddev->bitmap, sector_nr);
4080 sh = get_active_stripe(conf, sector_nr, 0, 1);
4082 sh = get_active_stripe(conf, sector_nr, 0, 0);
4083 /* make sure we don't swamp the stripe cache if someone else
4084 * is trying to get access
4086 schedule_timeout_uninterruptible(1);
4088 /* Need to check if array will still be degraded after recovery/resync
4089 * We don't need to check the 'failed' flag as when that gets set,
4092 for (i=0; i<mddev->raid_disks; i++)
4093 if (conf->disks[i].rdev == NULL)
4096 bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, still_degraded);
4098 spin_lock(&sh->lock);
4099 set_bit(STRIPE_SYNCING, &sh->state);
4100 clear_bit(STRIPE_INSYNC, &sh->state);
4101 spin_unlock(&sh->lock);
4103 /* wait for any blocked device to be handled */
4104 while (unlikely(!handle_stripe(sh)))
4108 return STRIPE_SECTORS;
4111 static int retry_aligned_read(raid5_conf_t *conf, struct bio *raid_bio)
4113 /* We may not be able to submit a whole bio at once as there
4114 * may not be enough stripe_heads available.
4115 * We cannot pre-allocate enough stripe_heads as we may need
4116 * more than exist in the cache (if we allow ever large chunks).
4117 * So we do one stripe head at a time and record in
4118 * ->bi_hw_segments how many have been done.
4120 * We *know* that this entire raid_bio is in one chunk, so
4121 * it will be only one 'dd_idx' and only need one call to raid5_compute_sector.
4123 struct stripe_head *sh;
4125 sector_t sector, logical_sector, last_sector;
4130 logical_sector = raid_bio->bi_sector & ~((sector_t)STRIPE_SECTORS-1);
4131 sector = raid5_compute_sector(conf, logical_sector,
4133 last_sector = raid_bio->bi_sector + (raid_bio->bi_size>>9);
4135 for (; logical_sector < last_sector;
4136 logical_sector += STRIPE_SECTORS,
4137 sector += STRIPE_SECTORS,
4140 if (scnt < raid5_bi_hw_segments(raid_bio))
4141 /* already done this stripe */
4144 sh = get_active_stripe(conf, sector, 0, 1);
4147 /* failed to get a stripe - must wait */
4148 raid5_set_bi_hw_segments(raid_bio, scnt);
4149 conf->retry_read_aligned = raid_bio;
4153 set_bit(R5_ReadError, &sh->dev[dd_idx].flags);
4154 if (!add_stripe_bio(sh, raid_bio, dd_idx, 0)) {
4156 raid5_set_bi_hw_segments(raid_bio, scnt);
4157 conf->retry_read_aligned = raid_bio;
4165 spin_lock_irq(&conf->device_lock);
4166 remaining = raid5_dec_bi_phys_segments(raid_bio);
4167 spin_unlock_irq(&conf->device_lock);
4169 bio_endio(raid_bio, 0);
4170 if (atomic_dec_and_test(&conf->active_aligned_reads))
4171 wake_up(&conf->wait_for_stripe);
4178 * This is our raid5 kernel thread.
4180 * We scan the hash table for stripes which can be handled now.
4181 * During the scan, completed stripes are saved for us by the interrupt
4182 * handler, so that they will not have to wait for our next wakeup.
4184 static void raid5d(mddev_t *mddev)
4186 struct stripe_head *sh;
4187 raid5_conf_t *conf = mddev_to_conf(mddev);
4190 pr_debug("+++ raid5d active\n");
4192 md_check_recovery(mddev);
4195 spin_lock_irq(&conf->device_lock);
4199 if (conf->seq_flush != conf->seq_write) {
4200 int seq = conf->seq_flush;
4201 spin_unlock_irq(&conf->device_lock);
4202 bitmap_unplug(mddev->bitmap);
4203 spin_lock_irq(&conf->device_lock);
4204 conf->seq_write = seq;
4205 activate_bit_delay(conf);
4208 while ((bio = remove_bio_from_retry(conf))) {
4210 spin_unlock_irq(&conf->device_lock);
4211 ok = retry_aligned_read(conf, bio);
4212 spin_lock_irq(&conf->device_lock);
4218 sh = __get_priority_stripe(conf);
4222 spin_unlock_irq(&conf->device_lock);
4228 spin_lock_irq(&conf->device_lock);
4230 pr_debug("%d stripes handled\n", handled);
4232 spin_unlock_irq(&conf->device_lock);
4234 async_tx_issue_pending_all();
4235 unplug_slaves(mddev);
4237 pr_debug("--- raid5d inactive\n");
4241 raid5_show_stripe_cache_size(mddev_t *mddev, char *page)
4243 raid5_conf_t *conf = mddev_to_conf(mddev);
4245 return sprintf(page, "%d\n", conf->max_nr_stripes);
4251 raid5_store_stripe_cache_size(mddev_t *mddev, const char *page, size_t len)
4253 raid5_conf_t *conf = mddev_to_conf(mddev);
4257 if (len >= PAGE_SIZE)
4262 if (strict_strtoul(page, 10, &new))
4264 if (new <= 16 || new > 32768)
4266 while (new < conf->max_nr_stripes) {
4267 if (drop_one_stripe(conf))
4268 conf->max_nr_stripes--;
4272 err = md_allow_write(mddev);
4275 while (new > conf->max_nr_stripes) {
4276 if (grow_one_stripe(conf))
4277 conf->max_nr_stripes++;
4283 static struct md_sysfs_entry
4284 raid5_stripecache_size = __ATTR(stripe_cache_size, S_IRUGO | S_IWUSR,
4285 raid5_show_stripe_cache_size,
4286 raid5_store_stripe_cache_size);
4289 raid5_show_preread_threshold(mddev_t *mddev, char *page)
4291 raid5_conf_t *conf = mddev_to_conf(mddev);
4293 return sprintf(page, "%d\n", conf->bypass_threshold);
4299 raid5_store_preread_threshold(mddev_t *mddev, const char *page, size_t len)
4301 raid5_conf_t *conf = mddev_to_conf(mddev);
4303 if (len >= PAGE_SIZE)
4308 if (strict_strtoul(page, 10, &new))
4310 if (new > conf->max_nr_stripes)
4312 conf->bypass_threshold = new;
4316 static struct md_sysfs_entry
4317 raid5_preread_bypass_threshold = __ATTR(preread_bypass_threshold,
4319 raid5_show_preread_threshold,
4320 raid5_store_preread_threshold);
4323 stripe_cache_active_show(mddev_t *mddev, char *page)
4325 raid5_conf_t *conf = mddev_to_conf(mddev);
4327 return sprintf(page, "%d\n", atomic_read(&conf->active_stripes));
4332 static struct md_sysfs_entry
4333 raid5_stripecache_active = __ATTR_RO(stripe_cache_active);
4335 static struct attribute *raid5_attrs[] = {
4336 &raid5_stripecache_size.attr,
4337 &raid5_stripecache_active.attr,
4338 &raid5_preread_bypass_threshold.attr,
4341 static struct attribute_group raid5_attrs_group = {
4343 .attrs = raid5_attrs,
4347 raid5_size(mddev_t *mddev, sector_t sectors, int raid_disks)
4349 raid5_conf_t *conf = mddev_to_conf(mddev);
4352 sectors = mddev->dev_sectors;
4354 /* size is defined by the smallest of previous and new size */
4355 if (conf->raid_disks < conf->previous_raid_disks)
4356 raid_disks = conf->raid_disks;
4358 raid_disks = conf->previous_raid_disks;
4361 sectors &= ~((sector_t)mddev->chunk_size/512 - 1);
4362 sectors &= ~((sector_t)mddev->new_chunk/512 - 1);
4363 return sectors * (raid_disks - conf->max_degraded);
4366 static void raid5_free_percpu(raid5_conf_t *conf)
4368 struct raid5_percpu *percpu;
4375 for_each_possible_cpu(cpu) {
4376 percpu = per_cpu_ptr(conf->percpu, cpu);
4377 safe_put_page(percpu->spare_page);
4378 kfree(percpu->scribble);
4380 #ifdef CONFIG_HOTPLUG_CPU
4381 unregister_cpu_notifier(&conf->cpu_notify);
4385 free_percpu(conf->percpu);
4388 static void free_conf(raid5_conf_t *conf)
4390 shrink_stripes(conf);
4391 raid5_free_percpu(conf);
4393 kfree(conf->stripe_hashtbl);
4397 #ifdef CONFIG_HOTPLUG_CPU
4398 static int raid456_cpu_notify(struct notifier_block *nfb, unsigned long action,
4401 raid5_conf_t *conf = container_of(nfb, raid5_conf_t, cpu_notify);
4402 long cpu = (long)hcpu;
4403 struct raid5_percpu *percpu = per_cpu_ptr(conf->percpu, cpu);
4406 case CPU_UP_PREPARE:
4407 case CPU_UP_PREPARE_FROZEN:
4408 if (conf->level == 6 && !percpu->spare_page)
4409 percpu->spare_page = alloc_page(GFP_KERNEL);
4410 if (!percpu->scribble)
4411 percpu->scribble = kmalloc(conf->scribble_len, GFP_KERNEL);
4413 if (!percpu->scribble ||
4414 (conf->level == 6 && !percpu->spare_page)) {
4415 safe_put_page(percpu->spare_page);
4416 kfree(percpu->scribble);
4417 pr_err("%s: failed memory allocation for cpu%ld\n",
4423 case CPU_DEAD_FROZEN:
4424 safe_put_page(percpu->spare_page);
4425 kfree(percpu->scribble);
4426 percpu->spare_page = NULL;
4427 percpu->scribble = NULL;
4436 static int raid5_alloc_percpu(raid5_conf_t *conf)
4439 struct page *spare_page;
4440 struct raid5_percpu *allcpus;
4444 allcpus = alloc_percpu(struct raid5_percpu);
4447 conf->percpu = allcpus;
4451 for_each_present_cpu(cpu) {
4452 if (conf->level == 6) {
4453 spare_page = alloc_page(GFP_KERNEL);
4458 per_cpu_ptr(conf->percpu, cpu)->spare_page = spare_page;
4460 scribble = kmalloc(scribble_len(conf->raid_disks), GFP_KERNEL);
4465 per_cpu_ptr(conf->percpu, cpu)->scribble = scribble;
4467 #ifdef CONFIG_HOTPLUG_CPU
4468 conf->cpu_notify.notifier_call = raid456_cpu_notify;
4469 conf->cpu_notify.priority = 0;
4471 err = register_cpu_notifier(&conf->cpu_notify);
4478 static raid5_conf_t *setup_conf(mddev_t *mddev)
4481 int raid_disk, memory;
4483 struct disk_info *disk;
4485 if (mddev->new_level != 5
4486 && mddev->new_level != 4
4487 && mddev->new_level != 6) {
4488 printk(KERN_ERR "raid5: %s: raid level not set to 4/5/6 (%d)\n",
4489 mdname(mddev), mddev->new_level);
4490 return ERR_PTR(-EIO);
4492 if ((mddev->new_level == 5
4493 && !algorithm_valid_raid5(mddev->new_layout)) ||
4494 (mddev->new_level == 6
4495 && !algorithm_valid_raid6(mddev->new_layout))) {
4496 printk(KERN_ERR "raid5: %s: layout %d not supported\n",
4497 mdname(mddev), mddev->new_layout);
4498 return ERR_PTR(-EIO);
4500 if (mddev->new_level == 6 && mddev->raid_disks < 4) {
4501 printk(KERN_ERR "raid6: not enough configured devices for %s (%d, minimum 4)\n",
4502 mdname(mddev), mddev->raid_disks);
4503 return ERR_PTR(-EINVAL);
4506 if (!mddev->new_chunk || mddev->new_chunk % PAGE_SIZE) {
4507 printk(KERN_ERR "raid5: invalid chunk size %d for %s\n",
4508 mddev->new_chunk, mdname(mddev));
4509 return ERR_PTR(-EINVAL);
4512 conf = kzalloc(sizeof(raid5_conf_t), GFP_KERNEL);
4516 conf->raid_disks = mddev->raid_disks;
4517 conf->scribble_len = scribble_len(conf->raid_disks);
4518 if (mddev->reshape_position == MaxSector)
4519 conf->previous_raid_disks = mddev->raid_disks;
4521 conf->previous_raid_disks = mddev->raid_disks - mddev->delta_disks;
4523 conf->disks = kzalloc(conf->raid_disks * sizeof(struct disk_info),
4528 conf->mddev = mddev;
4530 if ((conf->stripe_hashtbl = kzalloc(PAGE_SIZE, GFP_KERNEL)) == NULL)
4533 conf->level = mddev->new_level;
4534 if (raid5_alloc_percpu(conf) != 0)
4537 spin_lock_init(&conf->device_lock);
4538 init_waitqueue_head(&conf->wait_for_stripe);
4539 init_waitqueue_head(&conf->wait_for_overlap);
4540 INIT_LIST_HEAD(&conf->handle_list);
4541 INIT_LIST_HEAD(&conf->hold_list);
4542 INIT_LIST_HEAD(&conf->delayed_list);
4543 INIT_LIST_HEAD(&conf->bitmap_list);
4544 INIT_LIST_HEAD(&conf->inactive_list);
4545 atomic_set(&conf->active_stripes, 0);
4546 atomic_set(&conf->preread_active_stripes, 0);
4547 atomic_set(&conf->active_aligned_reads, 0);
4548 conf->bypass_threshold = BYPASS_THRESHOLD;
4550 pr_debug("raid5: run(%s) called.\n", mdname(mddev));
4552 list_for_each_entry(rdev, &mddev->disks, same_set) {
4553 raid_disk = rdev->raid_disk;
4554 if (raid_disk >= conf->raid_disks
4557 disk = conf->disks + raid_disk;
4561 if (test_bit(In_sync, &rdev->flags)) {
4562 char b[BDEVNAME_SIZE];
4563 printk(KERN_INFO "raid5: device %s operational as raid"
4564 " disk %d\n", bdevname(rdev->bdev,b),
4567 /* Cannot rely on bitmap to complete recovery */
4571 conf->chunk_size = mddev->new_chunk;
4572 if (conf->level == 6)
4573 conf->max_degraded = 2;
4575 conf->max_degraded = 1;
4576 conf->algorithm = mddev->new_layout;
4577 conf->max_nr_stripes = NR_STRIPES;
4578 conf->reshape_progress = mddev->reshape_position;
4579 if (conf->reshape_progress != MaxSector) {
4580 conf->prev_chunk = mddev->chunk_size;
4581 conf->prev_algo = mddev->layout;
4584 memory = conf->max_nr_stripes * (sizeof(struct stripe_head) +
4585 conf->raid_disks * ((sizeof(struct bio) + PAGE_SIZE))) / 1024;
4586 if (grow_stripes(conf, conf->max_nr_stripes)) {
4588 "raid5: couldn't allocate %dkB for buffers\n", memory);
4591 printk(KERN_INFO "raid5: allocated %dkB for %s\n",
4592 memory, mdname(mddev));
4594 conf->thread = md_register_thread(raid5d, mddev, "%s_raid5");
4595 if (!conf->thread) {
4597 "raid5: couldn't allocate thread for %s\n",
4607 return ERR_PTR(-EIO);
4609 return ERR_PTR(-ENOMEM);
4612 static int run(mddev_t *mddev)
4615 int working_disks = 0;
4618 if (mddev->reshape_position != MaxSector) {
4619 /* Check that we can continue the reshape.
4620 * Currently only disks can change, it must
4621 * increase, and we must be past the point where
4622 * a stripe over-writes itself
4624 sector_t here_new, here_old;
4626 int max_degraded = (mddev->level == 6 ? 2 : 1);
4628 if (mddev->new_level != mddev->level) {
4629 printk(KERN_ERR "raid5: %s: unsupported reshape "
4630 "required - aborting.\n",
4634 old_disks = mddev->raid_disks - mddev->delta_disks;
4635 /* reshape_position must be on a new-stripe boundary, and one
4636 * further up in new geometry must map after here in old
4639 here_new = mddev->reshape_position;
4640 if (sector_div(here_new, (mddev->new_chunk>>9)*
4641 (mddev->raid_disks - max_degraded))) {
4642 printk(KERN_ERR "raid5: reshape_position not "
4643 "on a stripe boundary\n");
4646 /* here_new is the stripe we will write to */
4647 here_old = mddev->reshape_position;
4648 sector_div(here_old, (mddev->chunk_size>>9)*
4649 (old_disks-max_degraded));
4650 /* here_old is the first stripe that we might need to read
4652 if (here_new >= here_old) {
4653 /* Reading from the same stripe as writing to - bad */
4654 printk(KERN_ERR "raid5: reshape_position too early for "
4655 "auto-recovery - aborting.\n");
4658 printk(KERN_INFO "raid5: reshape will continue\n");
4659 /* OK, we should be able to continue; */
4661 BUG_ON(mddev->level != mddev->new_level);
4662 BUG_ON(mddev->layout != mddev->new_layout);
4663 BUG_ON(mddev->chunk_size != mddev->new_chunk);
4664 BUG_ON(mddev->delta_disks != 0);
4667 if (mddev->private == NULL)
4668 conf = setup_conf(mddev);
4670 conf = mddev->private;
4673 return PTR_ERR(conf);
4675 mddev->thread = conf->thread;
4676 conf->thread = NULL;
4677 mddev->private = conf;
4680 * 0 for a fully functional array, 1 or 2 for a degraded array.
4682 list_for_each_entry(rdev, &mddev->disks, same_set)
4683 if (rdev->raid_disk >= 0 &&
4684 test_bit(In_sync, &rdev->flags))
4687 mddev->degraded = conf->raid_disks - working_disks;
4689 if (mddev->degraded > conf->max_degraded) {
4690 printk(KERN_ERR "raid5: not enough operational devices for %s"
4691 " (%d/%d failed)\n",
4692 mdname(mddev), mddev->degraded, conf->raid_disks);
4696 /* device size must be a multiple of chunk size */
4697 mddev->dev_sectors &= ~(mddev->chunk_size / 512 - 1);
4698 mddev->resync_max_sectors = mddev->dev_sectors;
4700 if (mddev->degraded > 0 &&
4701 mddev->recovery_cp != MaxSector) {
4702 if (mddev->ok_start_degraded)
4704 "raid5: starting dirty degraded array: %s"
4705 "- data corruption possible.\n",
4709 "raid5: cannot start dirty degraded array for %s\n",
4715 if (mddev->degraded == 0)
4716 printk("raid5: raid level %d set %s active with %d out of %d"
4717 " devices, algorithm %d\n", conf->level, mdname(mddev),
4718 mddev->raid_disks-mddev->degraded, mddev->raid_disks,
4721 printk(KERN_ALERT "raid5: raid level %d set %s active with %d"
4722 " out of %d devices, algorithm %d\n", conf->level,
4723 mdname(mddev), mddev->raid_disks - mddev->degraded,
4724 mddev->raid_disks, mddev->new_layout);
4726 print_raid5_conf(conf);
4728 if (conf->reshape_progress != MaxSector) {
4729 printk("...ok start reshape thread\n");
4730 conf->reshape_safe = conf->reshape_progress;
4731 atomic_set(&conf->reshape_stripes, 0);
4732 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
4733 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
4734 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
4735 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
4736 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
4740 /* read-ahead size must cover two whole stripes, which is
4741 * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
4744 int data_disks = conf->previous_raid_disks - conf->max_degraded;
4745 int stripe = data_disks *
4746 (mddev->chunk_size / PAGE_SIZE);
4747 if (mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
4748 mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
4751 /* Ok, everything is just fine now */
4752 if (sysfs_create_group(&mddev->kobj, &raid5_attrs_group))
4754 "raid5: failed to create sysfs attributes for %s\n",
4757 mddev->queue->queue_lock = &conf->device_lock;
4759 mddev->queue->unplug_fn = raid5_unplug_device;
4760 mddev->queue->backing_dev_info.congested_data = mddev;
4761 mddev->queue->backing_dev_info.congested_fn = raid5_congested;
4763 md_set_array_sectors(mddev, raid5_size(mddev, 0, 0));
4765 blk_queue_merge_bvec(mddev->queue, raid5_mergeable_bvec);
4769 md_unregister_thread(mddev->thread);
4770 mddev->thread = NULL;
4772 print_raid5_conf(conf);
4775 mddev->private = NULL;
4776 printk(KERN_ALERT "raid5: failed to run raid set %s\n", mdname(mddev));
4782 static int stop(mddev_t *mddev)
4784 raid5_conf_t *conf = (raid5_conf_t *) mddev->private;
4786 md_unregister_thread(mddev->thread);
4787 mddev->thread = NULL;
4788 mddev->queue->backing_dev_info.congested_fn = NULL;
4789 blk_sync_queue(mddev->queue); /* the unplug fn references 'conf'*/
4790 sysfs_remove_group(&mddev->kobj, &raid5_attrs_group);
4792 mddev->private = NULL;
4797 static void print_sh(struct seq_file *seq, struct stripe_head *sh)
4801 seq_printf(seq, "sh %llu, pd_idx %d, state %ld.\n",
4802 (unsigned long long)sh->sector, sh->pd_idx, sh->state);
4803 seq_printf(seq, "sh %llu, count %d.\n",
4804 (unsigned long long)sh->sector, atomic_read(&sh->count));
4805 seq_printf(seq, "sh %llu, ", (unsigned long long)sh->sector);
4806 for (i = 0; i < sh->disks; i++) {
4807 seq_printf(seq, "(cache%d: %p %ld) ",
4808 i, sh->dev[i].page, sh->dev[i].flags);
4810 seq_printf(seq, "\n");
4813 static void printall(struct seq_file *seq, raid5_conf_t *conf)
4815 struct stripe_head *sh;
4816 struct hlist_node *hn;
4819 spin_lock_irq(&conf->device_lock);
4820 for (i = 0; i < NR_HASH; i++) {
4821 hlist_for_each_entry(sh, hn, &conf->stripe_hashtbl[i], hash) {
4822 if (sh->raid_conf != conf)
4827 spin_unlock_irq(&conf->device_lock);
4831 static void status(struct seq_file *seq, mddev_t *mddev)
4833 raid5_conf_t *conf = (raid5_conf_t *) mddev->private;
4836 seq_printf (seq, " level %d, %dk chunk, algorithm %d", mddev->level, mddev->chunk_size >> 10, mddev->layout);
4837 seq_printf (seq, " [%d/%d] [", conf->raid_disks, conf->raid_disks - mddev->degraded);
4838 for (i = 0; i < conf->raid_disks; i++)
4839 seq_printf (seq, "%s",
4840 conf->disks[i].rdev &&
4841 test_bit(In_sync, &conf->disks[i].rdev->flags) ? "U" : "_");
4842 seq_printf (seq, "]");
4844 seq_printf (seq, "\n");
4845 printall(seq, conf);
4849 static void print_raid5_conf (raid5_conf_t *conf)
4852 struct disk_info *tmp;
4854 printk("RAID5 conf printout:\n");
4856 printk("(conf==NULL)\n");
4859 printk(" --- rd:%d wd:%d\n", conf->raid_disks,
4860 conf->raid_disks - conf->mddev->degraded);
4862 for (i = 0; i < conf->raid_disks; i++) {
4863 char b[BDEVNAME_SIZE];
4864 tmp = conf->disks + i;
4866 printk(" disk %d, o:%d, dev:%s\n",
4867 i, !test_bit(Faulty, &tmp->rdev->flags),
4868 bdevname(tmp->rdev->bdev,b));
4872 static int raid5_spare_active(mddev_t *mddev)
4875 raid5_conf_t *conf = mddev->private;
4876 struct disk_info *tmp;
4878 for (i = 0; i < conf->raid_disks; i++) {
4879 tmp = conf->disks + i;
4881 && !test_bit(Faulty, &tmp->rdev->flags)
4882 && !test_and_set_bit(In_sync, &tmp->rdev->flags)) {
4883 unsigned long flags;
4884 spin_lock_irqsave(&conf->device_lock, flags);
4886 spin_unlock_irqrestore(&conf->device_lock, flags);
4889 print_raid5_conf(conf);
4893 static int raid5_remove_disk(mddev_t *mddev, int number)
4895 raid5_conf_t *conf = mddev->private;
4898 struct disk_info *p = conf->disks + number;
4900 print_raid5_conf(conf);
4903 if (number >= conf->raid_disks &&
4904 conf->reshape_progress == MaxSector)
4905 clear_bit(In_sync, &rdev->flags);
4907 if (test_bit(In_sync, &rdev->flags) ||
4908 atomic_read(&rdev->nr_pending)) {
4912 /* Only remove non-faulty devices if recovery
4915 if (!test_bit(Faulty, &rdev->flags) &&
4916 mddev->degraded <= conf->max_degraded &&
4917 number < conf->raid_disks) {
4923 if (atomic_read(&rdev->nr_pending)) {
4924 /* lost the race, try later */
4931 print_raid5_conf(conf);
4935 static int raid5_add_disk(mddev_t *mddev, mdk_rdev_t *rdev)
4937 raid5_conf_t *conf = mddev->private;
4940 struct disk_info *p;
4942 int last = conf->raid_disks - 1;
4944 if (mddev->degraded > conf->max_degraded)
4945 /* no point adding a device */
4948 if (rdev->raid_disk >= 0)
4949 first = last = rdev->raid_disk;
4952 * find the disk ... but prefer rdev->saved_raid_disk
4955 if (rdev->saved_raid_disk >= 0 &&
4956 rdev->saved_raid_disk >= first &&
4957 conf->disks[rdev->saved_raid_disk].rdev == NULL)
4958 disk = rdev->saved_raid_disk;
4961 for ( ; disk <= last ; disk++)
4962 if ((p=conf->disks + disk)->rdev == NULL) {
4963 clear_bit(In_sync, &rdev->flags);
4964 rdev->raid_disk = disk;
4966 if (rdev->saved_raid_disk != disk)
4968 rcu_assign_pointer(p->rdev, rdev);
4971 print_raid5_conf(conf);
4975 static int raid5_resize(mddev_t *mddev, sector_t sectors)
4977 /* no resync is happening, and there is enough space
4978 * on all devices, so we can resize.
4979 * We need to make sure resync covers any new space.
4980 * If the array is shrinking we should possibly wait until
4981 * any io in the removed space completes, but it hardly seems
4984 sectors &= ~((sector_t)mddev->chunk_size/512 - 1);
4985 md_set_array_sectors(mddev, raid5_size(mddev, sectors,
4986 mddev->raid_disks));
4987 if (mddev->array_sectors >
4988 raid5_size(mddev, sectors, mddev->raid_disks))
4990 set_capacity(mddev->gendisk, mddev->array_sectors);
4992 if (sectors > mddev->dev_sectors && mddev->recovery_cp == MaxSector) {
4993 mddev->recovery_cp = mddev->dev_sectors;
4994 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
4996 mddev->dev_sectors = sectors;
4997 mddev->resync_max_sectors = sectors;
5001 static int raid5_check_reshape(mddev_t *mddev)
5003 raid5_conf_t *conf = mddev_to_conf(mddev);
5005 if (mddev->delta_disks == 0 &&
5006 mddev->new_layout == mddev->layout &&
5007 mddev->new_chunk == mddev->chunk_size)
5008 return -EINVAL; /* nothing to do */
5010 /* Cannot grow a bitmap yet */
5012 if (mddev->degraded > conf->max_degraded)
5014 if (mddev->delta_disks < 0) {
5015 /* We might be able to shrink, but the devices must
5016 * be made bigger first.
5017 * For raid6, 4 is the minimum size.
5018 * Otherwise 2 is the minimum
5021 if (mddev->level == 6)
5023 if (mddev->raid_disks + mddev->delta_disks < min)
5027 /* Can only proceed if there are plenty of stripe_heads.
5028 * We need a minimum of one full stripe,, and for sensible progress
5029 * it is best to have about 4 times that.
5030 * If we require 4 times, then the default 256 4K stripe_heads will
5031 * allow for chunk sizes up to 256K, which is probably OK.
5032 * If the chunk size is greater, user-space should request more
5033 * stripe_heads first.
5035 if ((mddev->chunk_size / STRIPE_SIZE) * 4 > conf->max_nr_stripes ||
5036 (mddev->new_chunk / STRIPE_SIZE) * 4 > conf->max_nr_stripes) {
5037 printk(KERN_WARNING "raid5: reshape: not enough stripes. Needed %lu\n",
5038 (max(mddev->chunk_size, mddev->new_chunk)
5043 return resize_stripes(conf, conf->raid_disks + mddev->delta_disks);
5046 static int raid5_start_reshape(mddev_t *mddev)
5048 raid5_conf_t *conf = mddev_to_conf(mddev);
5051 int added_devices = 0;
5052 unsigned long flags;
5054 if (test_bit(MD_RECOVERY_RUNNING, &mddev->recovery))
5057 list_for_each_entry(rdev, &mddev->disks, same_set)
5058 if (rdev->raid_disk < 0 &&
5059 !test_bit(Faulty, &rdev->flags))
5062 if (spares - mddev->degraded < mddev->delta_disks - conf->max_degraded)
5063 /* Not enough devices even to make a degraded array
5068 /* Refuse to reduce size of the array. Any reductions in
5069 * array size must be through explicit setting of array_size
5072 if (raid5_size(mddev, 0, conf->raid_disks + mddev->delta_disks)
5073 < mddev->array_sectors) {
5074 printk(KERN_ERR "md: %s: array size must be reduced "
5075 "before number of disks\n", mdname(mddev));
5079 atomic_set(&conf->reshape_stripes, 0);
5080 spin_lock_irq(&conf->device_lock);
5081 conf->previous_raid_disks = conf->raid_disks;
5082 conf->raid_disks += mddev->delta_disks;
5083 conf->prev_chunk = conf->chunk_size;
5084 conf->chunk_size = mddev->new_chunk;
5085 conf->prev_algo = conf->algorithm;
5086 conf->algorithm = mddev->new_layout;
5087 if (mddev->delta_disks < 0)
5088 conf->reshape_progress = raid5_size(mddev, 0, 0);
5090 conf->reshape_progress = 0;
5091 conf->reshape_safe = conf->reshape_progress;
5093 spin_unlock_irq(&conf->device_lock);
5095 /* Add some new drives, as many as will fit.
5096 * We know there are enough to make the newly sized array work.
5098 list_for_each_entry(rdev, &mddev->disks, same_set)
5099 if (rdev->raid_disk < 0 &&
5100 !test_bit(Faulty, &rdev->flags)) {
5101 if (raid5_add_disk(mddev, rdev) == 0) {
5103 set_bit(In_sync, &rdev->flags);
5105 rdev->recovery_offset = 0;
5106 sprintf(nm, "rd%d", rdev->raid_disk);
5107 if (sysfs_create_link(&mddev->kobj,
5110 "raid5: failed to create "
5111 " link %s for %s\n",
5117 if (mddev->delta_disks > 0) {
5118 spin_lock_irqsave(&conf->device_lock, flags);
5119 mddev->degraded = (conf->raid_disks - conf->previous_raid_disks)
5121 spin_unlock_irqrestore(&conf->device_lock, flags);
5123 mddev->raid_disks = conf->raid_disks;
5124 mddev->reshape_position = 0;
5125 set_bit(MD_CHANGE_DEVS, &mddev->flags);
5127 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
5128 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
5129 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
5130 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
5131 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
5133 if (!mddev->sync_thread) {
5134 mddev->recovery = 0;
5135 spin_lock_irq(&conf->device_lock);
5136 mddev->raid_disks = conf->raid_disks = conf->previous_raid_disks;
5137 conf->reshape_progress = MaxSector;
5138 spin_unlock_irq(&conf->device_lock);
5141 conf->reshape_checkpoint = jiffies;
5142 md_wakeup_thread(mddev->sync_thread);
5143 md_new_event(mddev);
5147 /* This is called from the reshape thread and should make any
5148 * changes needed in 'conf'
5150 static void end_reshape(raid5_conf_t *conf)
5153 if (!test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery)) {
5155 spin_lock_irq(&conf->device_lock);
5156 conf->previous_raid_disks = conf->raid_disks;
5157 conf->reshape_progress = MaxSector;
5158 spin_unlock_irq(&conf->device_lock);
5159 wake_up(&conf->wait_for_overlap);
5161 /* read-ahead size must cover two whole stripes, which is
5162 * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
5165 int data_disks = conf->raid_disks - conf->max_degraded;
5166 int stripe = data_disks * (conf->chunk_size
5168 if (conf->mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
5169 conf->mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
5174 /* This is called from the raid5d thread with mddev_lock held.
5175 * It makes config changes to the device.
5177 static void raid5_finish_reshape(mddev_t *mddev)
5179 struct block_device *bdev;
5180 raid5_conf_t *conf = mddev_to_conf(mddev);
5182 if (!test_bit(MD_RECOVERY_INTR, &mddev->recovery)) {
5184 if (mddev->delta_disks > 0) {
5185 md_set_array_sectors(mddev, raid5_size(mddev, 0, 0));
5186 set_capacity(mddev->gendisk, mddev->array_sectors);
5189 bdev = bdget_disk(mddev->gendisk, 0);
5191 mutex_lock(&bdev->bd_inode->i_mutex);
5192 i_size_write(bdev->bd_inode,
5193 (loff_t)mddev->array_sectors << 9);
5194 mutex_unlock(&bdev->bd_inode->i_mutex);
5199 mddev->degraded = conf->raid_disks;
5200 for (d = 0; d < conf->raid_disks ; d++)
5201 if (conf->disks[d].rdev &&
5203 &conf->disks[d].rdev->flags))
5205 for (d = conf->raid_disks ;
5206 d < conf->raid_disks - mddev->delta_disks;
5208 raid5_remove_disk(mddev, d);
5210 mddev->layout = conf->algorithm;
5211 mddev->chunk_size = conf->chunk_size;
5212 mddev->reshape_position = MaxSector;
5213 mddev->delta_disks = 0;
5217 static void raid5_quiesce(mddev_t *mddev, int state)
5219 raid5_conf_t *conf = mddev_to_conf(mddev);
5222 case 2: /* resume for a suspend */
5223 wake_up(&conf->wait_for_overlap);
5226 case 1: /* stop all writes */
5227 spin_lock_irq(&conf->device_lock);
5229 wait_event_lock_irq(conf->wait_for_stripe,
5230 atomic_read(&conf->active_stripes) == 0 &&
5231 atomic_read(&conf->active_aligned_reads) == 0,
5232 conf->device_lock, /* nothing */);
5233 spin_unlock_irq(&conf->device_lock);
5236 case 0: /* re-enable writes */
5237 spin_lock_irq(&conf->device_lock);
5239 wake_up(&conf->wait_for_stripe);
5240 wake_up(&conf->wait_for_overlap);
5241 spin_unlock_irq(&conf->device_lock);
5247 static void *raid5_takeover_raid1(mddev_t *mddev)
5251 if (mddev->raid_disks != 2 ||
5252 mddev->degraded > 1)
5253 return ERR_PTR(-EINVAL);
5255 /* Should check if there are write-behind devices? */
5257 chunksect = 64*2; /* 64K by default */
5259 /* The array must be an exact multiple of chunksize */
5260 while (chunksect && (mddev->array_sectors & (chunksect-1)))
5263 if ((chunksect<<9) < STRIPE_SIZE)
5264 /* array size does not allow a suitable chunk size */
5265 return ERR_PTR(-EINVAL);
5267 mddev->new_level = 5;
5268 mddev->new_layout = ALGORITHM_LEFT_SYMMETRIC;
5269 mddev->new_chunk = chunksect << 9;
5271 return setup_conf(mddev);
5274 static void *raid5_takeover_raid6(mddev_t *mddev)
5278 switch (mddev->layout) {
5279 case ALGORITHM_LEFT_ASYMMETRIC_6:
5280 new_layout = ALGORITHM_LEFT_ASYMMETRIC;
5282 case ALGORITHM_RIGHT_ASYMMETRIC_6:
5283 new_layout = ALGORITHM_RIGHT_ASYMMETRIC;
5285 case ALGORITHM_LEFT_SYMMETRIC_6:
5286 new_layout = ALGORITHM_LEFT_SYMMETRIC;
5288 case ALGORITHM_RIGHT_SYMMETRIC_6:
5289 new_layout = ALGORITHM_RIGHT_SYMMETRIC;
5291 case ALGORITHM_PARITY_0_6:
5292 new_layout = ALGORITHM_PARITY_0;
5294 case ALGORITHM_PARITY_N:
5295 new_layout = ALGORITHM_PARITY_N;
5298 return ERR_PTR(-EINVAL);
5300 mddev->new_level = 5;
5301 mddev->new_layout = new_layout;
5302 mddev->delta_disks = -1;
5303 mddev->raid_disks -= 1;
5304 return setup_conf(mddev);
5308 static int raid5_reconfig(mddev_t *mddev, int new_layout, int new_chunk)
5310 /* For a 2-drive array, the layout and chunk size can be changed
5311 * immediately as not restriping is needed.
5312 * For larger arrays we record the new value - after validation
5313 * to be used by a reshape pass.
5315 raid5_conf_t *conf = mddev_to_conf(mddev);
5317 if (new_layout >= 0 && !algorithm_valid_raid5(new_layout))
5319 if (new_chunk > 0) {
5320 if (new_chunk & (new_chunk-1))
5321 /* not a power of 2 */
5323 if (new_chunk < PAGE_SIZE)
5325 if (mddev->array_sectors & ((new_chunk>>9)-1))
5326 /* not factor of array size */
5330 /* They look valid */
5332 if (mddev->raid_disks == 2) {
5334 if (new_layout >= 0) {
5335 conf->algorithm = new_layout;
5336 mddev->layout = mddev->new_layout = new_layout;
5338 if (new_chunk > 0) {
5339 conf->chunk_size = new_chunk;
5340 mddev->chunk_size = mddev->new_chunk = new_chunk;
5342 set_bit(MD_CHANGE_DEVS, &mddev->flags);
5343 md_wakeup_thread(mddev->thread);
5345 if (new_layout >= 0)
5346 mddev->new_layout = new_layout;
5348 mddev->new_chunk = new_chunk;
5353 static int raid6_reconfig(mddev_t *mddev, int new_layout, int new_chunk)
5355 if (new_layout >= 0 && !algorithm_valid_raid6(new_layout))
5357 if (new_chunk > 0) {
5358 if (new_chunk & (new_chunk-1))
5359 /* not a power of 2 */
5361 if (new_chunk < PAGE_SIZE)
5363 if (mddev->array_sectors & ((new_chunk>>9)-1))
5364 /* not factor of array size */
5368 /* They look valid */
5370 if (new_layout >= 0)
5371 mddev->new_layout = new_layout;
5373 mddev->new_chunk = new_chunk;
5378 static void *raid5_takeover(mddev_t *mddev)
5380 /* raid5 can take over:
5381 * raid0 - if all devices are the same - make it a raid4 layout
5382 * raid1 - if there are two drives. We need to know the chunk size
5383 * raid4 - trivial - just use a raid4 layout.
5384 * raid6 - Providing it is a *_6 layout
5386 * For now, just do raid1
5389 if (mddev->level == 1)
5390 return raid5_takeover_raid1(mddev);
5391 if (mddev->level == 4) {
5392 mddev->new_layout = ALGORITHM_PARITY_N;
5393 mddev->new_level = 5;
5394 return setup_conf(mddev);
5396 if (mddev->level == 6)
5397 return raid5_takeover_raid6(mddev);
5399 return ERR_PTR(-EINVAL);
5403 static struct mdk_personality raid5_personality;
5405 static void *raid6_takeover(mddev_t *mddev)
5407 /* Currently can only take over a raid5. We map the
5408 * personality to an equivalent raid6 personality
5409 * with the Q block at the end.
5413 if (mddev->pers != &raid5_personality)
5414 return ERR_PTR(-EINVAL);
5415 if (mddev->degraded > 1)
5416 return ERR_PTR(-EINVAL);
5417 if (mddev->raid_disks > 253)
5418 return ERR_PTR(-EINVAL);
5419 if (mddev->raid_disks < 3)
5420 return ERR_PTR(-EINVAL);
5422 switch (mddev->layout) {
5423 case ALGORITHM_LEFT_ASYMMETRIC:
5424 new_layout = ALGORITHM_LEFT_ASYMMETRIC_6;
5426 case ALGORITHM_RIGHT_ASYMMETRIC:
5427 new_layout = ALGORITHM_RIGHT_ASYMMETRIC_6;
5429 case ALGORITHM_LEFT_SYMMETRIC:
5430 new_layout = ALGORITHM_LEFT_SYMMETRIC_6;
5432 case ALGORITHM_RIGHT_SYMMETRIC:
5433 new_layout = ALGORITHM_RIGHT_SYMMETRIC_6;
5435 case ALGORITHM_PARITY_0:
5436 new_layout = ALGORITHM_PARITY_0_6;
5438 case ALGORITHM_PARITY_N:
5439 new_layout = ALGORITHM_PARITY_N;
5442 return ERR_PTR(-EINVAL);
5444 mddev->new_level = 6;
5445 mddev->new_layout = new_layout;
5446 mddev->delta_disks = 1;
5447 mddev->raid_disks += 1;
5448 return setup_conf(mddev);
5452 static struct mdk_personality raid6_personality =
5456 .owner = THIS_MODULE,
5457 .make_request = make_request,
5461 .error_handler = error,
5462 .hot_add_disk = raid5_add_disk,
5463 .hot_remove_disk= raid5_remove_disk,
5464 .spare_active = raid5_spare_active,
5465 .sync_request = sync_request,
5466 .resize = raid5_resize,
5468 .check_reshape = raid5_check_reshape,
5469 .start_reshape = raid5_start_reshape,
5470 .finish_reshape = raid5_finish_reshape,
5471 .quiesce = raid5_quiesce,
5472 .takeover = raid6_takeover,
5473 .reconfig = raid6_reconfig,
5475 static struct mdk_personality raid5_personality =
5479 .owner = THIS_MODULE,
5480 .make_request = make_request,
5484 .error_handler = error,
5485 .hot_add_disk = raid5_add_disk,
5486 .hot_remove_disk= raid5_remove_disk,
5487 .spare_active = raid5_spare_active,
5488 .sync_request = sync_request,
5489 .resize = raid5_resize,
5491 .check_reshape = raid5_check_reshape,
5492 .start_reshape = raid5_start_reshape,
5493 .finish_reshape = raid5_finish_reshape,
5494 .quiesce = raid5_quiesce,
5495 .takeover = raid5_takeover,
5496 .reconfig = raid5_reconfig,
5499 static struct mdk_personality raid4_personality =
5503 .owner = THIS_MODULE,
5504 .make_request = make_request,
5508 .error_handler = error,
5509 .hot_add_disk = raid5_add_disk,
5510 .hot_remove_disk= raid5_remove_disk,
5511 .spare_active = raid5_spare_active,
5512 .sync_request = sync_request,
5513 .resize = raid5_resize,
5515 .check_reshape = raid5_check_reshape,
5516 .start_reshape = raid5_start_reshape,
5517 .finish_reshape = raid5_finish_reshape,
5518 .quiesce = raid5_quiesce,
5521 static int __init raid5_init(void)
5523 register_md_personality(&raid6_personality);
5524 register_md_personality(&raid5_personality);
5525 register_md_personality(&raid4_personality);
5529 static void raid5_exit(void)
5531 unregister_md_personality(&raid6_personality);
5532 unregister_md_personality(&raid5_personality);
5533 unregister_md_personality(&raid4_personality);
5536 module_init(raid5_init);
5537 module_exit(raid5_exit);
5538 MODULE_LICENSE("GPL");
5539 MODULE_ALIAS("md-personality-4"); /* RAID5 */
5540 MODULE_ALIAS("md-raid5");
5541 MODULE_ALIAS("md-raid4");
5542 MODULE_ALIAS("md-level-5");
5543 MODULE_ALIAS("md-level-4");
5544 MODULE_ALIAS("md-personality-8"); /* RAID6 */
5545 MODULE_ALIAS("md-raid6");
5546 MODULE_ALIAS("md-level-6");
5548 /* This used to be two separate modules, they were: */
5549 MODULE_ALIAS("raid5");
5550 MODULE_ALIAS("raid6");