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/async.h>
51 #include <linux/seq_file.h>
52 #include <linux/cpu.h>
61 #define NR_STRIPES 256
62 #define STRIPE_SIZE PAGE_SIZE
63 #define STRIPE_SHIFT (PAGE_SHIFT - 9)
64 #define STRIPE_SECTORS (STRIPE_SIZE>>9)
65 #define IO_THRESHOLD 1
66 #define BYPASS_THRESHOLD 1
67 #define NR_HASH (PAGE_SIZE / sizeof(struct hlist_head))
68 #define HASH_MASK (NR_HASH - 1)
70 #define stripe_hash(conf, sect) (&((conf)->stripe_hashtbl[((sect) >> STRIPE_SHIFT) & HASH_MASK]))
72 /* bio's attached to a stripe+device for I/O are linked together in bi_sector
73 * order without overlap. There may be several bio's per stripe+device, and
74 * a bio could span several devices.
75 * When walking this list for a particular stripe+device, we must never proceed
76 * beyond a bio that extends past this device, as the next bio might no longer
78 * This macro is used to determine the 'next' bio in the list, given the sector
79 * of the current stripe+device
81 #define r5_next_bio(bio, sect) ( ( (bio)->bi_sector + ((bio)->bi_size>>9) < sect + STRIPE_SECTORS) ? (bio)->bi_next : NULL)
83 * The following can be used to debug the driver
85 #define RAID5_PARANOIA 1
86 #if RAID5_PARANOIA && defined(CONFIG_SMP)
87 # define CHECK_DEVLOCK() assert_spin_locked(&conf->device_lock)
89 # define CHECK_DEVLOCK()
97 #define printk_rl(args...) ((void) (printk_ratelimit() && printk(args)))
100 * We maintain a biased count of active stripes in the bottom 16 bits of
101 * bi_phys_segments, and a count of processed stripes in the upper 16 bits
103 static inline int raid5_bi_phys_segments(struct bio *bio)
105 return bio->bi_phys_segments & 0xffff;
108 static inline int raid5_bi_hw_segments(struct bio *bio)
110 return (bio->bi_phys_segments >> 16) & 0xffff;
113 static inline int raid5_dec_bi_phys_segments(struct bio *bio)
115 --bio->bi_phys_segments;
116 return raid5_bi_phys_segments(bio);
119 static inline int raid5_dec_bi_hw_segments(struct bio *bio)
121 unsigned short val = raid5_bi_hw_segments(bio);
124 bio->bi_phys_segments = (val << 16) | raid5_bi_phys_segments(bio);
128 static inline void raid5_set_bi_hw_segments(struct bio *bio, unsigned int cnt)
130 bio->bi_phys_segments = raid5_bi_phys_segments(bio) || (cnt << 16);
133 /* Find first data disk in a raid6 stripe */
134 static inline int raid6_d0(struct stripe_head *sh)
137 /* ddf always start from first device */
139 /* md starts just after Q block */
140 if (sh->qd_idx == sh->disks - 1)
143 return sh->qd_idx + 1;
145 static inline int raid6_next_disk(int disk, int raid_disks)
148 return (disk < raid_disks) ? disk : 0;
151 /* When walking through the disks in a raid5, starting at raid6_d0,
152 * We need to map each disk to a 'slot', where the data disks are slot
153 * 0 .. raid_disks-3, the parity disk is raid_disks-2 and the Q disk
154 * is raid_disks-1. This help does that mapping.
156 static int raid6_idx_to_slot(int idx, struct stripe_head *sh,
157 int *count, int syndrome_disks)
161 if (idx == sh->pd_idx)
162 return syndrome_disks;
163 if (idx == sh->qd_idx)
164 return syndrome_disks + 1;
169 static void return_io(struct bio *return_bi)
171 struct bio *bi = return_bi;
174 return_bi = bi->bi_next;
182 static void print_raid5_conf (raid5_conf_t *conf);
184 static int stripe_operations_active(struct stripe_head *sh)
186 return sh->check_state || sh->reconstruct_state ||
187 test_bit(STRIPE_BIOFILL_RUN, &sh->state) ||
188 test_bit(STRIPE_COMPUTE_RUN, &sh->state);
191 static void __release_stripe(raid5_conf_t *conf, struct stripe_head *sh)
193 if (atomic_dec_and_test(&sh->count)) {
194 BUG_ON(!list_empty(&sh->lru));
195 BUG_ON(atomic_read(&conf->active_stripes)==0);
196 if (test_bit(STRIPE_HANDLE, &sh->state)) {
197 if (test_bit(STRIPE_DELAYED, &sh->state)) {
198 list_add_tail(&sh->lru, &conf->delayed_list);
199 blk_plug_device(conf->mddev->queue);
200 } else if (test_bit(STRIPE_BIT_DELAY, &sh->state) &&
201 sh->bm_seq - conf->seq_write > 0) {
202 list_add_tail(&sh->lru, &conf->bitmap_list);
203 blk_plug_device(conf->mddev->queue);
205 clear_bit(STRIPE_BIT_DELAY, &sh->state);
206 list_add_tail(&sh->lru, &conf->handle_list);
208 md_wakeup_thread(conf->mddev->thread);
210 BUG_ON(stripe_operations_active(sh));
211 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
212 atomic_dec(&conf->preread_active_stripes);
213 if (atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD)
214 md_wakeup_thread(conf->mddev->thread);
216 atomic_dec(&conf->active_stripes);
217 if (!test_bit(STRIPE_EXPANDING, &sh->state)) {
218 list_add_tail(&sh->lru, &conf->inactive_list);
219 wake_up(&conf->wait_for_stripe);
220 if (conf->retry_read_aligned)
221 md_wakeup_thread(conf->mddev->thread);
227 static void release_stripe(struct stripe_head *sh)
229 raid5_conf_t *conf = sh->raid_conf;
232 spin_lock_irqsave(&conf->device_lock, flags);
233 __release_stripe(conf, sh);
234 spin_unlock_irqrestore(&conf->device_lock, flags);
237 static inline void remove_hash(struct stripe_head *sh)
239 pr_debug("remove_hash(), stripe %llu\n",
240 (unsigned long long)sh->sector);
242 hlist_del_init(&sh->hash);
245 static inline void insert_hash(raid5_conf_t *conf, struct stripe_head *sh)
247 struct hlist_head *hp = stripe_hash(conf, sh->sector);
249 pr_debug("insert_hash(), stripe %llu\n",
250 (unsigned long long)sh->sector);
253 hlist_add_head(&sh->hash, hp);
257 /* find an idle stripe, make sure it is unhashed, and return it. */
258 static struct stripe_head *get_free_stripe(raid5_conf_t *conf)
260 struct stripe_head *sh = NULL;
261 struct list_head *first;
264 if (list_empty(&conf->inactive_list))
266 first = conf->inactive_list.next;
267 sh = list_entry(first, struct stripe_head, lru);
268 list_del_init(first);
270 atomic_inc(&conf->active_stripes);
275 static void shrink_buffers(struct stripe_head *sh, int num)
280 for (i=0; i<num ; i++) {
284 sh->dev[i].page = NULL;
289 static int grow_buffers(struct stripe_head *sh, int num)
293 for (i=0; i<num; i++) {
296 if (!(page = alloc_page(GFP_KERNEL))) {
299 sh->dev[i].page = page;
304 static void raid5_build_block(struct stripe_head *sh, int i, int previous);
305 static void stripe_set_idx(sector_t stripe, raid5_conf_t *conf, int previous,
306 struct stripe_head *sh);
308 static void init_stripe(struct stripe_head *sh, sector_t sector, int previous)
310 raid5_conf_t *conf = sh->raid_conf;
313 BUG_ON(atomic_read(&sh->count) != 0);
314 BUG_ON(test_bit(STRIPE_HANDLE, &sh->state));
315 BUG_ON(stripe_operations_active(sh));
318 pr_debug("init_stripe called, stripe %llu\n",
319 (unsigned long long)sh->sector);
323 sh->generation = conf->generation - previous;
324 sh->disks = previous ? conf->previous_raid_disks : conf->raid_disks;
326 stripe_set_idx(sector, conf, previous, sh);
330 for (i = sh->disks; i--; ) {
331 struct r5dev *dev = &sh->dev[i];
333 if (dev->toread || dev->read || dev->towrite || dev->written ||
334 test_bit(R5_LOCKED, &dev->flags)) {
335 printk(KERN_ERR "sector=%llx i=%d %p %p %p %p %d\n",
336 (unsigned long long)sh->sector, i, dev->toread,
337 dev->read, dev->towrite, dev->written,
338 test_bit(R5_LOCKED, &dev->flags));
342 raid5_build_block(sh, i, previous);
344 insert_hash(conf, sh);
347 static struct stripe_head *__find_stripe(raid5_conf_t *conf, sector_t sector,
350 struct stripe_head *sh;
351 struct hlist_node *hn;
354 pr_debug("__find_stripe, sector %llu\n", (unsigned long long)sector);
355 hlist_for_each_entry(sh, hn, stripe_hash(conf, sector), hash)
356 if (sh->sector == sector && sh->generation == generation)
358 pr_debug("__stripe %llu not in cache\n", (unsigned long long)sector);
362 static void unplug_slaves(mddev_t *mddev);
363 static void raid5_unplug_device(struct request_queue *q);
365 static struct stripe_head *
366 get_active_stripe(raid5_conf_t *conf, sector_t sector,
367 int previous, int noblock, int noquiesce)
369 struct stripe_head *sh;
371 pr_debug("get_stripe, sector %llu\n", (unsigned long long)sector);
373 spin_lock_irq(&conf->device_lock);
376 wait_event_lock_irq(conf->wait_for_stripe,
377 conf->quiesce == 0 || noquiesce,
378 conf->device_lock, /* nothing */);
379 sh = __find_stripe(conf, sector, conf->generation - previous);
381 if (!conf->inactive_blocked)
382 sh = get_free_stripe(conf);
383 if (noblock && sh == NULL)
386 conf->inactive_blocked = 1;
387 wait_event_lock_irq(conf->wait_for_stripe,
388 !list_empty(&conf->inactive_list) &&
389 (atomic_read(&conf->active_stripes)
390 < (conf->max_nr_stripes *3/4)
391 || !conf->inactive_blocked),
393 raid5_unplug_device(conf->mddev->queue)
395 conf->inactive_blocked = 0;
397 init_stripe(sh, sector, previous);
399 if (atomic_read(&sh->count)) {
400 BUG_ON(!list_empty(&sh->lru)
401 && !test_bit(STRIPE_EXPANDING, &sh->state));
403 if (!test_bit(STRIPE_HANDLE, &sh->state))
404 atomic_inc(&conf->active_stripes);
405 if (list_empty(&sh->lru) &&
406 !test_bit(STRIPE_EXPANDING, &sh->state))
408 list_del_init(&sh->lru);
411 } while (sh == NULL);
414 atomic_inc(&sh->count);
416 spin_unlock_irq(&conf->device_lock);
421 raid5_end_read_request(struct bio *bi, int error);
423 raid5_end_write_request(struct bio *bi, int error);
425 static void ops_run_io(struct stripe_head *sh, struct stripe_head_state *s)
427 raid5_conf_t *conf = sh->raid_conf;
428 int i, disks = sh->disks;
432 for (i = disks; i--; ) {
436 if (test_and_clear_bit(R5_Wantwrite, &sh->dev[i].flags))
438 else if (test_and_clear_bit(R5_Wantread, &sh->dev[i].flags))
443 bi = &sh->dev[i].req;
447 bi->bi_end_io = raid5_end_write_request;
449 bi->bi_end_io = raid5_end_read_request;
452 rdev = rcu_dereference(conf->disks[i].rdev);
453 if (rdev && test_bit(Faulty, &rdev->flags))
456 atomic_inc(&rdev->nr_pending);
460 if (s->syncing || s->expanding || s->expanded)
461 md_sync_acct(rdev->bdev, STRIPE_SECTORS);
463 set_bit(STRIPE_IO_STARTED, &sh->state);
465 bi->bi_bdev = rdev->bdev;
466 pr_debug("%s: for %llu schedule op %ld on disc %d\n",
467 __func__, (unsigned long long)sh->sector,
469 atomic_inc(&sh->count);
470 bi->bi_sector = sh->sector + rdev->data_offset;
471 bi->bi_flags = 1 << BIO_UPTODATE;
475 bi->bi_io_vec = &sh->dev[i].vec;
476 bi->bi_io_vec[0].bv_len = STRIPE_SIZE;
477 bi->bi_io_vec[0].bv_offset = 0;
478 bi->bi_size = STRIPE_SIZE;
481 test_bit(R5_ReWrite, &sh->dev[i].flags))
482 atomic_add(STRIPE_SECTORS,
483 &rdev->corrected_errors);
484 generic_make_request(bi);
487 set_bit(STRIPE_DEGRADED, &sh->state);
488 pr_debug("skip op %ld on disc %d for sector %llu\n",
489 bi->bi_rw, i, (unsigned long long)sh->sector);
490 clear_bit(R5_LOCKED, &sh->dev[i].flags);
491 set_bit(STRIPE_HANDLE, &sh->state);
496 static struct dma_async_tx_descriptor *
497 async_copy_data(int frombio, struct bio *bio, struct page *page,
498 sector_t sector, struct dma_async_tx_descriptor *tx)
501 struct page *bio_page;
504 struct async_submit_ctl submit;
506 if (bio->bi_sector >= sector)
507 page_offset = (signed)(bio->bi_sector - sector) * 512;
509 page_offset = (signed)(sector - bio->bi_sector) * -512;
511 init_async_submit(&submit, 0, tx, NULL, NULL, NULL);
512 bio_for_each_segment(bvl, bio, i) {
513 int len = bio_iovec_idx(bio, i)->bv_len;
517 if (page_offset < 0) {
518 b_offset = -page_offset;
519 page_offset += b_offset;
523 if (len > 0 && page_offset + len > STRIPE_SIZE)
524 clen = STRIPE_SIZE - page_offset;
529 b_offset += bio_iovec_idx(bio, i)->bv_offset;
530 bio_page = bio_iovec_idx(bio, i)->bv_page;
532 tx = async_memcpy(page, bio_page, page_offset,
533 b_offset, clen, &submit);
535 tx = async_memcpy(bio_page, page, b_offset,
536 page_offset, clen, &submit);
538 /* chain the operations */
539 submit.depend_tx = tx;
541 if (clen < len) /* hit end of page */
549 static void ops_complete_biofill(void *stripe_head_ref)
551 struct stripe_head *sh = stripe_head_ref;
552 struct bio *return_bi = NULL;
553 raid5_conf_t *conf = sh->raid_conf;
556 pr_debug("%s: stripe %llu\n", __func__,
557 (unsigned long long)sh->sector);
559 /* clear completed biofills */
560 spin_lock_irq(&conf->device_lock);
561 for (i = sh->disks; i--; ) {
562 struct r5dev *dev = &sh->dev[i];
564 /* acknowledge completion of a biofill operation */
565 /* and check if we need to reply to a read request,
566 * new R5_Wantfill requests are held off until
567 * !STRIPE_BIOFILL_RUN
569 if (test_and_clear_bit(R5_Wantfill, &dev->flags)) {
570 struct bio *rbi, *rbi2;
575 while (rbi && rbi->bi_sector <
576 dev->sector + STRIPE_SECTORS) {
577 rbi2 = r5_next_bio(rbi, dev->sector);
578 if (!raid5_dec_bi_phys_segments(rbi)) {
579 rbi->bi_next = return_bi;
586 spin_unlock_irq(&conf->device_lock);
587 clear_bit(STRIPE_BIOFILL_RUN, &sh->state);
589 return_io(return_bi);
591 set_bit(STRIPE_HANDLE, &sh->state);
595 static void ops_run_biofill(struct stripe_head *sh)
597 struct dma_async_tx_descriptor *tx = NULL;
598 raid5_conf_t *conf = sh->raid_conf;
599 struct async_submit_ctl submit;
602 pr_debug("%s: stripe %llu\n", __func__,
603 (unsigned long long)sh->sector);
605 for (i = sh->disks; i--; ) {
606 struct r5dev *dev = &sh->dev[i];
607 if (test_bit(R5_Wantfill, &dev->flags)) {
609 spin_lock_irq(&conf->device_lock);
610 dev->read = rbi = dev->toread;
612 spin_unlock_irq(&conf->device_lock);
613 while (rbi && rbi->bi_sector <
614 dev->sector + STRIPE_SECTORS) {
615 tx = async_copy_data(0, rbi, dev->page,
617 rbi = r5_next_bio(rbi, dev->sector);
622 atomic_inc(&sh->count);
623 init_async_submit(&submit, ASYNC_TX_ACK, tx, ops_complete_biofill, sh, NULL);
624 async_trigger_callback(&submit);
627 static void mark_target_uptodate(struct stripe_head *sh, int target)
634 tgt = &sh->dev[target];
635 set_bit(R5_UPTODATE, &tgt->flags);
636 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
637 clear_bit(R5_Wantcompute, &tgt->flags);
640 static void ops_complete_compute(void *stripe_head_ref)
642 struct stripe_head *sh = stripe_head_ref;
644 pr_debug("%s: stripe %llu\n", __func__,
645 (unsigned long long)sh->sector);
647 /* mark the computed target(s) as uptodate */
648 mark_target_uptodate(sh, sh->ops.target);
649 mark_target_uptodate(sh, sh->ops.target2);
651 clear_bit(STRIPE_COMPUTE_RUN, &sh->state);
652 if (sh->check_state == check_state_compute_run)
653 sh->check_state = check_state_compute_result;
654 set_bit(STRIPE_HANDLE, &sh->state);
658 /* return a pointer to the address conversion region of the scribble buffer */
659 static addr_conv_t *to_addr_conv(struct stripe_head *sh,
660 struct raid5_percpu *percpu)
662 return percpu->scribble + sizeof(struct page *) * (sh->disks + 2);
665 static struct dma_async_tx_descriptor *
666 ops_run_compute5(struct stripe_head *sh, struct raid5_percpu *percpu)
668 int disks = sh->disks;
669 struct page **xor_srcs = percpu->scribble;
670 int target = sh->ops.target;
671 struct r5dev *tgt = &sh->dev[target];
672 struct page *xor_dest = tgt->page;
674 struct dma_async_tx_descriptor *tx;
675 struct async_submit_ctl submit;
678 pr_debug("%s: stripe %llu block: %d\n",
679 __func__, (unsigned long long)sh->sector, target);
680 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
682 for (i = disks; i--; )
684 xor_srcs[count++] = sh->dev[i].page;
686 atomic_inc(&sh->count);
688 init_async_submit(&submit, ASYNC_TX_XOR_ZERO_DST, NULL,
689 ops_complete_compute, sh, to_addr_conv(sh, percpu));
690 if (unlikely(count == 1))
691 tx = async_memcpy(xor_dest, xor_srcs[0], 0, 0, STRIPE_SIZE, &submit);
693 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
698 /* set_syndrome_sources - populate source buffers for gen_syndrome
699 * @srcs - (struct page *) array of size sh->disks
700 * @sh - stripe_head to parse
702 * Populates srcs in proper layout order for the stripe and returns the
703 * 'count' of sources to be used in a call to async_gen_syndrome. The P
704 * destination buffer is recorded in srcs[count] and the Q destination
705 * is recorded in srcs[count+1]].
707 static int set_syndrome_sources(struct page **srcs, struct stripe_head *sh)
709 int disks = sh->disks;
710 int syndrome_disks = sh->ddf_layout ? disks : (disks - 2);
711 int d0_idx = raid6_d0(sh);
715 for (i = 0; i < disks; i++)
716 srcs[i] = (void *)raid6_empty_zero_page;
721 int slot = raid6_idx_to_slot(i, sh, &count, syndrome_disks);
723 srcs[slot] = sh->dev[i].page;
724 i = raid6_next_disk(i, disks);
725 } while (i != d0_idx);
726 BUG_ON(count != syndrome_disks);
731 static struct dma_async_tx_descriptor *
732 ops_run_compute6_1(struct stripe_head *sh, struct raid5_percpu *percpu)
734 int disks = sh->disks;
735 struct page **blocks = percpu->scribble;
737 int qd_idx = sh->qd_idx;
738 struct dma_async_tx_descriptor *tx;
739 struct async_submit_ctl submit;
745 if (sh->ops.target < 0)
746 target = sh->ops.target2;
747 else if (sh->ops.target2 < 0)
748 target = sh->ops.target;
750 /* we should only have one valid target */
753 pr_debug("%s: stripe %llu block: %d\n",
754 __func__, (unsigned long long)sh->sector, target);
756 tgt = &sh->dev[target];
757 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
760 atomic_inc(&sh->count);
762 if (target == qd_idx) {
763 count = set_syndrome_sources(blocks, sh);
764 blocks[count] = NULL; /* regenerating p is not necessary */
765 BUG_ON(blocks[count+1] != dest); /* q should already be set */
766 init_async_submit(&submit, 0, NULL, ops_complete_compute, sh,
767 to_addr_conv(sh, percpu));
768 tx = async_gen_syndrome(blocks, 0, count+2, STRIPE_SIZE, &submit);
770 /* Compute any data- or p-drive using XOR */
772 for (i = disks; i-- ; ) {
773 if (i == target || i == qd_idx)
775 blocks[count++] = sh->dev[i].page;
778 init_async_submit(&submit, ASYNC_TX_XOR_ZERO_DST, NULL,
779 ops_complete_compute, sh,
780 to_addr_conv(sh, percpu));
781 tx = async_xor(dest, blocks, 0, count, STRIPE_SIZE, &submit);
787 static struct dma_async_tx_descriptor *
788 ops_run_compute6_2(struct stripe_head *sh, struct raid5_percpu *percpu)
790 int i, count, disks = sh->disks;
791 int syndrome_disks = sh->ddf_layout ? disks : disks-2;
792 int d0_idx = raid6_d0(sh);
793 int faila = -1, failb = -1;
794 int target = sh->ops.target;
795 int target2 = sh->ops.target2;
796 struct r5dev *tgt = &sh->dev[target];
797 struct r5dev *tgt2 = &sh->dev[target2];
798 struct dma_async_tx_descriptor *tx;
799 struct page **blocks = percpu->scribble;
800 struct async_submit_ctl submit;
802 pr_debug("%s: stripe %llu block1: %d block2: %d\n",
803 __func__, (unsigned long long)sh->sector, target, target2);
804 BUG_ON(target < 0 || target2 < 0);
805 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
806 BUG_ON(!test_bit(R5_Wantcompute, &tgt2->flags));
808 /* we need to open-code set_syndrome_sources to handle to the
809 * slot number conversion for 'faila' and 'failb'
811 for (i = 0; i < disks ; i++)
812 blocks[i] = (void *)raid6_empty_zero_page;
816 int slot = raid6_idx_to_slot(i, sh, &count, syndrome_disks);
818 blocks[slot] = sh->dev[i].page;
824 i = raid6_next_disk(i, disks);
825 } while (i != d0_idx);
826 BUG_ON(count != syndrome_disks);
828 BUG_ON(faila == failb);
831 pr_debug("%s: stripe: %llu faila: %d failb: %d\n",
832 __func__, (unsigned long long)sh->sector, faila, failb);
834 atomic_inc(&sh->count);
836 if (failb == syndrome_disks+1) {
837 /* Q disk is one of the missing disks */
838 if (faila == syndrome_disks) {
839 /* Missing P+Q, just recompute */
840 init_async_submit(&submit, 0, NULL, ops_complete_compute,
841 sh, to_addr_conv(sh, percpu));
842 return async_gen_syndrome(blocks, 0, count+2,
843 STRIPE_SIZE, &submit);
847 int qd_idx = sh->qd_idx;
849 /* Missing D+Q: recompute D from P, then recompute Q */
850 if (target == qd_idx)
851 data_target = target2;
853 data_target = target;
856 for (i = disks; i-- ; ) {
857 if (i == data_target || i == qd_idx)
859 blocks[count++] = sh->dev[i].page;
861 dest = sh->dev[data_target].page;
862 init_async_submit(&submit, ASYNC_TX_XOR_ZERO_DST, NULL,
863 NULL, NULL, to_addr_conv(sh, percpu));
864 tx = async_xor(dest, blocks, 0, count, STRIPE_SIZE,
867 count = set_syndrome_sources(blocks, sh);
868 init_async_submit(&submit, 0, tx, ops_complete_compute,
869 sh, to_addr_conv(sh, percpu));
870 return async_gen_syndrome(blocks, 0, count+2,
871 STRIPE_SIZE, &submit);
875 init_async_submit(&submit, 0, NULL, ops_complete_compute, sh,
876 to_addr_conv(sh, percpu));
877 if (failb == syndrome_disks) {
878 /* We're missing D+P. */
879 return async_raid6_datap_recov(syndrome_disks+2, STRIPE_SIZE,
880 faila, blocks, &submit);
882 /* We're missing D+D. */
883 return async_raid6_2data_recov(syndrome_disks+2, STRIPE_SIZE,
884 faila, failb, blocks, &submit);
889 static void ops_complete_prexor(void *stripe_head_ref)
891 struct stripe_head *sh = stripe_head_ref;
893 pr_debug("%s: stripe %llu\n", __func__,
894 (unsigned long long)sh->sector);
897 static struct dma_async_tx_descriptor *
898 ops_run_prexor(struct stripe_head *sh, struct raid5_percpu *percpu,
899 struct dma_async_tx_descriptor *tx)
901 int disks = sh->disks;
902 struct page **xor_srcs = percpu->scribble;
903 int count = 0, pd_idx = sh->pd_idx, i;
904 struct async_submit_ctl submit;
906 /* existing parity data subtracted */
907 struct page *xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
909 pr_debug("%s: stripe %llu\n", __func__,
910 (unsigned long long)sh->sector);
912 for (i = disks; i--; ) {
913 struct r5dev *dev = &sh->dev[i];
914 /* Only process blocks that are known to be uptodate */
915 if (test_bit(R5_Wantdrain, &dev->flags))
916 xor_srcs[count++] = dev->page;
919 init_async_submit(&submit, ASYNC_TX_XOR_DROP_DST, tx,
920 ops_complete_prexor, sh, to_addr_conv(sh, percpu));
921 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
926 static struct dma_async_tx_descriptor *
927 ops_run_biodrain(struct stripe_head *sh, struct dma_async_tx_descriptor *tx)
929 int disks = sh->disks;
932 pr_debug("%s: stripe %llu\n", __func__,
933 (unsigned long long)sh->sector);
935 for (i = disks; i--; ) {
936 struct r5dev *dev = &sh->dev[i];
939 if (test_and_clear_bit(R5_Wantdrain, &dev->flags)) {
942 spin_lock(&sh->lock);
943 chosen = dev->towrite;
945 BUG_ON(dev->written);
946 wbi = dev->written = chosen;
947 spin_unlock(&sh->lock);
949 while (wbi && wbi->bi_sector <
950 dev->sector + STRIPE_SECTORS) {
951 tx = async_copy_data(1, wbi, dev->page,
953 wbi = r5_next_bio(wbi, dev->sector);
961 static void ops_complete_reconstruct(void *stripe_head_ref)
963 struct stripe_head *sh = stripe_head_ref;
964 int disks = sh->disks;
965 int pd_idx = sh->pd_idx;
966 int qd_idx = sh->qd_idx;
969 pr_debug("%s: stripe %llu\n", __func__,
970 (unsigned long long)sh->sector);
972 for (i = disks; i--; ) {
973 struct r5dev *dev = &sh->dev[i];
975 if (dev->written || i == pd_idx || i == qd_idx)
976 set_bit(R5_UPTODATE, &dev->flags);
979 if (sh->reconstruct_state == reconstruct_state_drain_run)
980 sh->reconstruct_state = reconstruct_state_drain_result;
981 else if (sh->reconstruct_state == reconstruct_state_prexor_drain_run)
982 sh->reconstruct_state = reconstruct_state_prexor_drain_result;
984 BUG_ON(sh->reconstruct_state != reconstruct_state_run);
985 sh->reconstruct_state = reconstruct_state_result;
988 set_bit(STRIPE_HANDLE, &sh->state);
993 ops_run_reconstruct5(struct stripe_head *sh, struct raid5_percpu *percpu,
994 struct dma_async_tx_descriptor *tx)
996 int disks = sh->disks;
997 struct page **xor_srcs = percpu->scribble;
998 struct async_submit_ctl submit;
999 int count = 0, pd_idx = sh->pd_idx, i;
1000 struct page *xor_dest;
1002 unsigned long flags;
1004 pr_debug("%s: stripe %llu\n", __func__,
1005 (unsigned long long)sh->sector);
1007 /* check if prexor is active which means only process blocks
1008 * that are part of a read-modify-write (written)
1010 if (sh->reconstruct_state == reconstruct_state_prexor_drain_run) {
1012 xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
1013 for (i = disks; i--; ) {
1014 struct r5dev *dev = &sh->dev[i];
1016 xor_srcs[count++] = dev->page;
1019 xor_dest = sh->dev[pd_idx].page;
1020 for (i = disks; i--; ) {
1021 struct r5dev *dev = &sh->dev[i];
1023 xor_srcs[count++] = dev->page;
1027 /* 1/ if we prexor'd then the dest is reused as a source
1028 * 2/ if we did not prexor then we are redoing the parity
1029 * set ASYNC_TX_XOR_DROP_DST and ASYNC_TX_XOR_ZERO_DST
1030 * for the synchronous xor case
1032 flags = ASYNC_TX_ACK |
1033 (prexor ? ASYNC_TX_XOR_DROP_DST : ASYNC_TX_XOR_ZERO_DST);
1035 atomic_inc(&sh->count);
1037 init_async_submit(&submit, flags, tx, ops_complete_reconstruct, sh,
1038 to_addr_conv(sh, percpu));
1039 if (unlikely(count == 1))
1040 tx = async_memcpy(xor_dest, xor_srcs[0], 0, 0, STRIPE_SIZE, &submit);
1042 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
1046 ops_run_reconstruct6(struct stripe_head *sh, struct raid5_percpu *percpu,
1047 struct dma_async_tx_descriptor *tx)
1049 struct async_submit_ctl submit;
1050 struct page **blocks = percpu->scribble;
1053 pr_debug("%s: stripe %llu\n", __func__, (unsigned long long)sh->sector);
1055 count = set_syndrome_sources(blocks, sh);
1057 atomic_inc(&sh->count);
1059 init_async_submit(&submit, ASYNC_TX_ACK, tx, ops_complete_reconstruct,
1060 sh, to_addr_conv(sh, percpu));
1061 async_gen_syndrome(blocks, 0, count+2, STRIPE_SIZE, &submit);
1064 static void ops_complete_check(void *stripe_head_ref)
1066 struct stripe_head *sh = stripe_head_ref;
1068 pr_debug("%s: stripe %llu\n", __func__,
1069 (unsigned long long)sh->sector);
1071 sh->check_state = check_state_check_result;
1072 set_bit(STRIPE_HANDLE, &sh->state);
1076 static void ops_run_check_p(struct stripe_head *sh, struct raid5_percpu *percpu)
1078 int disks = sh->disks;
1079 int pd_idx = sh->pd_idx;
1080 int qd_idx = sh->qd_idx;
1081 struct page *xor_dest;
1082 struct page **xor_srcs = percpu->scribble;
1083 struct dma_async_tx_descriptor *tx;
1084 struct async_submit_ctl submit;
1088 pr_debug("%s: stripe %llu\n", __func__,
1089 (unsigned long long)sh->sector);
1092 xor_dest = sh->dev[pd_idx].page;
1093 xor_srcs[count++] = xor_dest;
1094 for (i = disks; i--; ) {
1095 if (i == pd_idx || i == qd_idx)
1097 xor_srcs[count++] = sh->dev[i].page;
1100 init_async_submit(&submit, 0, NULL, NULL, NULL,
1101 to_addr_conv(sh, percpu));
1102 tx = async_xor_val(xor_dest, xor_srcs, 0, count, STRIPE_SIZE,
1103 &sh->ops.zero_sum_result, &submit);
1105 atomic_inc(&sh->count);
1106 init_async_submit(&submit, ASYNC_TX_ACK, tx, ops_complete_check, sh, NULL);
1107 tx = async_trigger_callback(&submit);
1110 static void ops_run_check_pq(struct stripe_head *sh, struct raid5_percpu *percpu, int checkp)
1112 struct page **srcs = percpu->scribble;
1113 struct async_submit_ctl submit;
1116 pr_debug("%s: stripe %llu checkp: %d\n", __func__,
1117 (unsigned long long)sh->sector, checkp);
1119 count = set_syndrome_sources(srcs, sh);
1123 atomic_inc(&sh->count);
1124 init_async_submit(&submit, ASYNC_TX_ACK, NULL, ops_complete_check,
1125 sh, to_addr_conv(sh, percpu));
1126 async_syndrome_val(srcs, 0, count+2, STRIPE_SIZE,
1127 &sh->ops.zero_sum_result, percpu->spare_page, &submit);
1130 static void raid_run_ops(struct stripe_head *sh, unsigned long ops_request)
1132 int overlap_clear = 0, i, disks = sh->disks;
1133 struct dma_async_tx_descriptor *tx = NULL;
1134 raid5_conf_t *conf = sh->raid_conf;
1135 int level = conf->level;
1136 struct raid5_percpu *percpu;
1140 percpu = per_cpu_ptr(conf->percpu, cpu);
1141 if (test_bit(STRIPE_OP_BIOFILL, &ops_request)) {
1142 ops_run_biofill(sh);
1146 if (test_bit(STRIPE_OP_COMPUTE_BLK, &ops_request)) {
1148 tx = ops_run_compute5(sh, percpu);
1150 if (sh->ops.target2 < 0 || sh->ops.target < 0)
1151 tx = ops_run_compute6_1(sh, percpu);
1153 tx = ops_run_compute6_2(sh, percpu);
1155 /* terminate the chain if reconstruct is not set to be run */
1156 if (tx && !test_bit(STRIPE_OP_RECONSTRUCT, &ops_request))
1160 if (test_bit(STRIPE_OP_PREXOR, &ops_request))
1161 tx = ops_run_prexor(sh, percpu, tx);
1163 if (test_bit(STRIPE_OP_BIODRAIN, &ops_request)) {
1164 tx = ops_run_biodrain(sh, tx);
1168 if (test_bit(STRIPE_OP_RECONSTRUCT, &ops_request)) {
1170 ops_run_reconstruct5(sh, percpu, tx);
1172 ops_run_reconstruct6(sh, percpu, tx);
1175 if (test_bit(STRIPE_OP_CHECK, &ops_request)) {
1176 if (sh->check_state == check_state_run)
1177 ops_run_check_p(sh, percpu);
1178 else if (sh->check_state == check_state_run_q)
1179 ops_run_check_pq(sh, percpu, 0);
1180 else if (sh->check_state == check_state_run_pq)
1181 ops_run_check_pq(sh, percpu, 1);
1187 for (i = disks; i--; ) {
1188 struct r5dev *dev = &sh->dev[i];
1189 if (test_and_clear_bit(R5_Overlap, &dev->flags))
1190 wake_up(&sh->raid_conf->wait_for_overlap);
1195 static int grow_one_stripe(raid5_conf_t *conf)
1197 struct stripe_head *sh;
1198 sh = kmem_cache_alloc(conf->slab_cache, GFP_KERNEL);
1201 memset(sh, 0, sizeof(*sh) + (conf->raid_disks-1)*sizeof(struct r5dev));
1202 sh->raid_conf = conf;
1203 spin_lock_init(&sh->lock);
1205 if (grow_buffers(sh, conf->raid_disks)) {
1206 shrink_buffers(sh, conf->raid_disks);
1207 kmem_cache_free(conf->slab_cache, sh);
1210 sh->disks = conf->raid_disks;
1211 /* we just created an active stripe so... */
1212 atomic_set(&sh->count, 1);
1213 atomic_inc(&conf->active_stripes);
1214 INIT_LIST_HEAD(&sh->lru);
1219 static int grow_stripes(raid5_conf_t *conf, int num)
1221 struct kmem_cache *sc;
1222 int devs = conf->raid_disks;
1224 sprintf(conf->cache_name[0],
1225 "raid%d-%s", conf->level, mdname(conf->mddev));
1226 sprintf(conf->cache_name[1],
1227 "raid%d-%s-alt", conf->level, mdname(conf->mddev));
1228 conf->active_name = 0;
1229 sc = kmem_cache_create(conf->cache_name[conf->active_name],
1230 sizeof(struct stripe_head)+(devs-1)*sizeof(struct r5dev),
1234 conf->slab_cache = sc;
1235 conf->pool_size = devs;
1237 if (!grow_one_stripe(conf))
1243 * scribble_len - return the required size of the scribble region
1244 * @num - total number of disks in the array
1246 * The size must be enough to contain:
1247 * 1/ a struct page pointer for each device in the array +2
1248 * 2/ room to convert each entry in (1) to its corresponding dma
1249 * (dma_map_page()) or page (page_address()) address.
1251 * Note: the +2 is for the destination buffers of the ddf/raid6 case where we
1252 * calculate over all devices (not just the data blocks), using zeros in place
1253 * of the P and Q blocks.
1255 static size_t scribble_len(int num)
1259 len = sizeof(struct page *) * (num+2) + sizeof(addr_conv_t) * (num+2);
1264 static int resize_stripes(raid5_conf_t *conf, int newsize)
1266 /* Make all the stripes able to hold 'newsize' devices.
1267 * New slots in each stripe get 'page' set to a new page.
1269 * This happens in stages:
1270 * 1/ create a new kmem_cache and allocate the required number of
1272 * 2/ gather all the old stripe_heads and tranfer the pages across
1273 * to the new stripe_heads. This will have the side effect of
1274 * freezing the array as once all stripe_heads have been collected,
1275 * no IO will be possible. Old stripe heads are freed once their
1276 * pages have been transferred over, and the old kmem_cache is
1277 * freed when all stripes are done.
1278 * 3/ reallocate conf->disks to be suitable bigger. If this fails,
1279 * we simple return a failre status - no need to clean anything up.
1280 * 4/ allocate new pages for the new slots in the new stripe_heads.
1281 * If this fails, we don't bother trying the shrink the
1282 * stripe_heads down again, we just leave them as they are.
1283 * As each stripe_head is processed the new one is released into
1286 * Once step2 is started, we cannot afford to wait for a write,
1287 * so we use GFP_NOIO allocations.
1289 struct stripe_head *osh, *nsh;
1290 LIST_HEAD(newstripes);
1291 struct disk_info *ndisks;
1294 struct kmem_cache *sc;
1297 if (newsize <= conf->pool_size)
1298 return 0; /* never bother to shrink */
1300 err = md_allow_write(conf->mddev);
1305 sc = kmem_cache_create(conf->cache_name[1-conf->active_name],
1306 sizeof(struct stripe_head)+(newsize-1)*sizeof(struct r5dev),
1311 for (i = conf->max_nr_stripes; i; i--) {
1312 nsh = kmem_cache_alloc(sc, GFP_KERNEL);
1316 memset(nsh, 0, sizeof(*nsh) + (newsize-1)*sizeof(struct r5dev));
1318 nsh->raid_conf = conf;
1319 spin_lock_init(&nsh->lock);
1321 list_add(&nsh->lru, &newstripes);
1324 /* didn't get enough, give up */
1325 while (!list_empty(&newstripes)) {
1326 nsh = list_entry(newstripes.next, struct stripe_head, lru);
1327 list_del(&nsh->lru);
1328 kmem_cache_free(sc, nsh);
1330 kmem_cache_destroy(sc);
1333 /* Step 2 - Must use GFP_NOIO now.
1334 * OK, we have enough stripes, start collecting inactive
1335 * stripes and copying them over
1337 list_for_each_entry(nsh, &newstripes, lru) {
1338 spin_lock_irq(&conf->device_lock);
1339 wait_event_lock_irq(conf->wait_for_stripe,
1340 !list_empty(&conf->inactive_list),
1342 unplug_slaves(conf->mddev)
1344 osh = get_free_stripe(conf);
1345 spin_unlock_irq(&conf->device_lock);
1346 atomic_set(&nsh->count, 1);
1347 for(i=0; i<conf->pool_size; i++)
1348 nsh->dev[i].page = osh->dev[i].page;
1349 for( ; i<newsize; i++)
1350 nsh->dev[i].page = NULL;
1351 kmem_cache_free(conf->slab_cache, osh);
1353 kmem_cache_destroy(conf->slab_cache);
1356 * At this point, we are holding all the stripes so the array
1357 * is completely stalled, so now is a good time to resize
1358 * conf->disks and the scribble region
1360 ndisks = kzalloc(newsize * sizeof(struct disk_info), GFP_NOIO);
1362 for (i=0; i<conf->raid_disks; i++)
1363 ndisks[i] = conf->disks[i];
1365 conf->disks = ndisks;
1370 conf->scribble_len = scribble_len(newsize);
1371 for_each_present_cpu(cpu) {
1372 struct raid5_percpu *percpu;
1375 percpu = per_cpu_ptr(conf->percpu, cpu);
1376 scribble = kmalloc(conf->scribble_len, GFP_NOIO);
1379 kfree(percpu->scribble);
1380 percpu->scribble = scribble;
1388 /* Step 4, return new stripes to service */
1389 while(!list_empty(&newstripes)) {
1390 nsh = list_entry(newstripes.next, struct stripe_head, lru);
1391 list_del_init(&nsh->lru);
1393 for (i=conf->raid_disks; i < newsize; i++)
1394 if (nsh->dev[i].page == NULL) {
1395 struct page *p = alloc_page(GFP_NOIO);
1396 nsh->dev[i].page = p;
1400 release_stripe(nsh);
1402 /* critical section pass, GFP_NOIO no longer needed */
1404 conf->slab_cache = sc;
1405 conf->active_name = 1-conf->active_name;
1406 conf->pool_size = newsize;
1410 static int drop_one_stripe(raid5_conf_t *conf)
1412 struct stripe_head *sh;
1414 spin_lock_irq(&conf->device_lock);
1415 sh = get_free_stripe(conf);
1416 spin_unlock_irq(&conf->device_lock);
1419 BUG_ON(atomic_read(&sh->count));
1420 shrink_buffers(sh, conf->pool_size);
1421 kmem_cache_free(conf->slab_cache, sh);
1422 atomic_dec(&conf->active_stripes);
1426 static void shrink_stripes(raid5_conf_t *conf)
1428 while (drop_one_stripe(conf))
1431 if (conf->slab_cache)
1432 kmem_cache_destroy(conf->slab_cache);
1433 conf->slab_cache = NULL;
1436 static void raid5_end_read_request(struct bio * bi, int error)
1438 struct stripe_head *sh = bi->bi_private;
1439 raid5_conf_t *conf = sh->raid_conf;
1440 int disks = sh->disks, i;
1441 int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
1442 char b[BDEVNAME_SIZE];
1446 for (i=0 ; i<disks; i++)
1447 if (bi == &sh->dev[i].req)
1450 pr_debug("end_read_request %llu/%d, count: %d, uptodate %d.\n",
1451 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
1459 set_bit(R5_UPTODATE, &sh->dev[i].flags);
1460 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
1461 rdev = conf->disks[i].rdev;
1462 printk_rl(KERN_INFO "raid5:%s: read error corrected"
1463 " (%lu sectors at %llu on %s)\n",
1464 mdname(conf->mddev), STRIPE_SECTORS,
1465 (unsigned long long)(sh->sector
1466 + rdev->data_offset),
1467 bdevname(rdev->bdev, b));
1468 clear_bit(R5_ReadError, &sh->dev[i].flags);
1469 clear_bit(R5_ReWrite, &sh->dev[i].flags);
1471 if (atomic_read(&conf->disks[i].rdev->read_errors))
1472 atomic_set(&conf->disks[i].rdev->read_errors, 0);
1474 const char *bdn = bdevname(conf->disks[i].rdev->bdev, b);
1476 rdev = conf->disks[i].rdev;
1478 clear_bit(R5_UPTODATE, &sh->dev[i].flags);
1479 atomic_inc(&rdev->read_errors);
1480 if (conf->mddev->degraded)
1481 printk_rl(KERN_WARNING
1482 "raid5:%s: read error not correctable "
1483 "(sector %llu on %s).\n",
1484 mdname(conf->mddev),
1485 (unsigned long long)(sh->sector
1486 + rdev->data_offset),
1488 else if (test_bit(R5_ReWrite, &sh->dev[i].flags))
1490 printk_rl(KERN_WARNING
1491 "raid5:%s: read error NOT corrected!! "
1492 "(sector %llu on %s).\n",
1493 mdname(conf->mddev),
1494 (unsigned long long)(sh->sector
1495 + rdev->data_offset),
1497 else if (atomic_read(&rdev->read_errors)
1498 > conf->max_nr_stripes)
1500 "raid5:%s: Too many read errors, failing device %s.\n",
1501 mdname(conf->mddev), bdn);
1505 set_bit(R5_ReadError, &sh->dev[i].flags);
1507 clear_bit(R5_ReadError, &sh->dev[i].flags);
1508 clear_bit(R5_ReWrite, &sh->dev[i].flags);
1509 md_error(conf->mddev, rdev);
1512 rdev_dec_pending(conf->disks[i].rdev, conf->mddev);
1513 clear_bit(R5_LOCKED, &sh->dev[i].flags);
1514 set_bit(STRIPE_HANDLE, &sh->state);
1518 static void raid5_end_write_request(struct bio *bi, int error)
1520 struct stripe_head *sh = bi->bi_private;
1521 raid5_conf_t *conf = sh->raid_conf;
1522 int disks = sh->disks, i;
1523 int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
1525 for (i=0 ; i<disks; i++)
1526 if (bi == &sh->dev[i].req)
1529 pr_debug("end_write_request %llu/%d, count %d, uptodate: %d.\n",
1530 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
1538 md_error(conf->mddev, conf->disks[i].rdev);
1540 rdev_dec_pending(conf->disks[i].rdev, conf->mddev);
1542 clear_bit(R5_LOCKED, &sh->dev[i].flags);
1543 set_bit(STRIPE_HANDLE, &sh->state);
1548 static sector_t compute_blocknr(struct stripe_head *sh, int i, int previous);
1550 static void raid5_build_block(struct stripe_head *sh, int i, int previous)
1552 struct r5dev *dev = &sh->dev[i];
1554 bio_init(&dev->req);
1555 dev->req.bi_io_vec = &dev->vec;
1557 dev->req.bi_max_vecs++;
1558 dev->vec.bv_page = dev->page;
1559 dev->vec.bv_len = STRIPE_SIZE;
1560 dev->vec.bv_offset = 0;
1562 dev->req.bi_sector = sh->sector;
1563 dev->req.bi_private = sh;
1566 dev->sector = compute_blocknr(sh, i, previous);
1569 static void error(mddev_t *mddev, mdk_rdev_t *rdev)
1571 char b[BDEVNAME_SIZE];
1572 raid5_conf_t *conf = (raid5_conf_t *) mddev->private;
1573 pr_debug("raid5: error called\n");
1575 if (!test_bit(Faulty, &rdev->flags)) {
1576 set_bit(MD_CHANGE_DEVS, &mddev->flags);
1577 if (test_and_clear_bit(In_sync, &rdev->flags)) {
1578 unsigned long flags;
1579 spin_lock_irqsave(&conf->device_lock, flags);
1581 spin_unlock_irqrestore(&conf->device_lock, flags);
1583 * if recovery was running, make sure it aborts.
1585 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1587 set_bit(Faulty, &rdev->flags);
1589 "raid5: Disk failure on %s, disabling device.\n"
1590 "raid5: Operation continuing on %d devices.\n",
1591 bdevname(rdev->bdev,b), conf->raid_disks - mddev->degraded);
1596 * Input: a 'big' sector number,
1597 * Output: index of the data and parity disk, and the sector # in them.
1599 static sector_t raid5_compute_sector(raid5_conf_t *conf, sector_t r_sector,
1600 int previous, int *dd_idx,
1601 struct stripe_head *sh)
1604 unsigned long chunk_number;
1605 unsigned int chunk_offset;
1608 sector_t new_sector;
1609 int algorithm = previous ? conf->prev_algo
1611 int sectors_per_chunk = previous ? (conf->prev_chunk >> 9)
1612 : (conf->chunk_size >> 9);
1613 int raid_disks = previous ? conf->previous_raid_disks
1615 int data_disks = raid_disks - conf->max_degraded;
1617 /* First compute the information on this sector */
1620 * Compute the chunk number and the sector offset inside the chunk
1622 chunk_offset = sector_div(r_sector, sectors_per_chunk);
1623 chunk_number = r_sector;
1624 BUG_ON(r_sector != chunk_number);
1627 * Compute the stripe number
1629 stripe = chunk_number / data_disks;
1632 * Compute the data disk and parity disk indexes inside the stripe
1634 *dd_idx = chunk_number % data_disks;
1637 * Select the parity disk based on the user selected algorithm.
1639 pd_idx = qd_idx = ~0;
1640 switch(conf->level) {
1642 pd_idx = data_disks;
1645 switch (algorithm) {
1646 case ALGORITHM_LEFT_ASYMMETRIC:
1647 pd_idx = data_disks - stripe % raid_disks;
1648 if (*dd_idx >= pd_idx)
1651 case ALGORITHM_RIGHT_ASYMMETRIC:
1652 pd_idx = stripe % raid_disks;
1653 if (*dd_idx >= pd_idx)
1656 case ALGORITHM_LEFT_SYMMETRIC:
1657 pd_idx = data_disks - stripe % raid_disks;
1658 *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
1660 case ALGORITHM_RIGHT_SYMMETRIC:
1661 pd_idx = stripe % raid_disks;
1662 *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
1664 case ALGORITHM_PARITY_0:
1668 case ALGORITHM_PARITY_N:
1669 pd_idx = data_disks;
1672 printk(KERN_ERR "raid5: unsupported algorithm %d\n",
1679 switch (algorithm) {
1680 case ALGORITHM_LEFT_ASYMMETRIC:
1681 pd_idx = raid_disks - 1 - (stripe % raid_disks);
1682 qd_idx = pd_idx + 1;
1683 if (pd_idx == raid_disks-1) {
1684 (*dd_idx)++; /* Q D D D P */
1686 } else if (*dd_idx >= pd_idx)
1687 (*dd_idx) += 2; /* D D P Q D */
1689 case ALGORITHM_RIGHT_ASYMMETRIC:
1690 pd_idx = stripe % raid_disks;
1691 qd_idx = pd_idx + 1;
1692 if (pd_idx == raid_disks-1) {
1693 (*dd_idx)++; /* Q D D D P */
1695 } else if (*dd_idx >= pd_idx)
1696 (*dd_idx) += 2; /* D D P Q D */
1698 case ALGORITHM_LEFT_SYMMETRIC:
1699 pd_idx = raid_disks - 1 - (stripe % raid_disks);
1700 qd_idx = (pd_idx + 1) % raid_disks;
1701 *dd_idx = (pd_idx + 2 + *dd_idx) % raid_disks;
1703 case ALGORITHM_RIGHT_SYMMETRIC:
1704 pd_idx = stripe % raid_disks;
1705 qd_idx = (pd_idx + 1) % raid_disks;
1706 *dd_idx = (pd_idx + 2 + *dd_idx) % raid_disks;
1709 case ALGORITHM_PARITY_0:
1714 case ALGORITHM_PARITY_N:
1715 pd_idx = data_disks;
1716 qd_idx = data_disks + 1;
1719 case ALGORITHM_ROTATING_ZERO_RESTART:
1720 /* Exactly the same as RIGHT_ASYMMETRIC, but or
1721 * of blocks for computing Q is different.
1723 pd_idx = stripe % raid_disks;
1724 qd_idx = pd_idx + 1;
1725 if (pd_idx == raid_disks-1) {
1726 (*dd_idx)++; /* Q D D D P */
1728 } else if (*dd_idx >= pd_idx)
1729 (*dd_idx) += 2; /* D D P Q D */
1733 case ALGORITHM_ROTATING_N_RESTART:
1734 /* Same a left_asymmetric, by first stripe is
1735 * D D D P Q rather than
1738 pd_idx = raid_disks - 1 - ((stripe + 1) % raid_disks);
1739 qd_idx = pd_idx + 1;
1740 if (pd_idx == raid_disks-1) {
1741 (*dd_idx)++; /* Q D D D P */
1743 } else if (*dd_idx >= pd_idx)
1744 (*dd_idx) += 2; /* D D P Q D */
1748 case ALGORITHM_ROTATING_N_CONTINUE:
1749 /* Same as left_symmetric but Q is before P */
1750 pd_idx = raid_disks - 1 - (stripe % raid_disks);
1751 qd_idx = (pd_idx + raid_disks - 1) % raid_disks;
1752 *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
1756 case ALGORITHM_LEFT_ASYMMETRIC_6:
1757 /* RAID5 left_asymmetric, with Q on last device */
1758 pd_idx = data_disks - stripe % (raid_disks-1);
1759 if (*dd_idx >= pd_idx)
1761 qd_idx = raid_disks - 1;
1764 case ALGORITHM_RIGHT_ASYMMETRIC_6:
1765 pd_idx = stripe % (raid_disks-1);
1766 if (*dd_idx >= pd_idx)
1768 qd_idx = raid_disks - 1;
1771 case ALGORITHM_LEFT_SYMMETRIC_6:
1772 pd_idx = data_disks - stripe % (raid_disks-1);
1773 *dd_idx = (pd_idx + 1 + *dd_idx) % (raid_disks-1);
1774 qd_idx = raid_disks - 1;
1777 case ALGORITHM_RIGHT_SYMMETRIC_6:
1778 pd_idx = stripe % (raid_disks-1);
1779 *dd_idx = (pd_idx + 1 + *dd_idx) % (raid_disks-1);
1780 qd_idx = raid_disks - 1;
1783 case ALGORITHM_PARITY_0_6:
1786 qd_idx = raid_disks - 1;
1791 printk(KERN_CRIT "raid6: unsupported algorithm %d\n",
1799 sh->pd_idx = pd_idx;
1800 sh->qd_idx = qd_idx;
1801 sh->ddf_layout = ddf_layout;
1804 * Finally, compute the new sector number
1806 new_sector = (sector_t)stripe * sectors_per_chunk + chunk_offset;
1811 static sector_t compute_blocknr(struct stripe_head *sh, int i, int previous)
1813 raid5_conf_t *conf = sh->raid_conf;
1814 int raid_disks = sh->disks;
1815 int data_disks = raid_disks - conf->max_degraded;
1816 sector_t new_sector = sh->sector, check;
1817 int sectors_per_chunk = previous ? (conf->prev_chunk >> 9)
1818 : (conf->chunk_size >> 9);
1819 int algorithm = previous ? conf->prev_algo
1823 int chunk_number, dummy1, dd_idx = i;
1825 struct stripe_head sh2;
1828 chunk_offset = sector_div(new_sector, sectors_per_chunk);
1829 stripe = new_sector;
1830 BUG_ON(new_sector != stripe);
1832 if (i == sh->pd_idx)
1834 switch(conf->level) {
1837 switch (algorithm) {
1838 case ALGORITHM_LEFT_ASYMMETRIC:
1839 case ALGORITHM_RIGHT_ASYMMETRIC:
1843 case ALGORITHM_LEFT_SYMMETRIC:
1844 case ALGORITHM_RIGHT_SYMMETRIC:
1847 i -= (sh->pd_idx + 1);
1849 case ALGORITHM_PARITY_0:
1852 case ALGORITHM_PARITY_N:
1855 printk(KERN_ERR "raid5: unsupported algorithm %d\n",
1861 if (i == sh->qd_idx)
1862 return 0; /* It is the Q disk */
1863 switch (algorithm) {
1864 case ALGORITHM_LEFT_ASYMMETRIC:
1865 case ALGORITHM_RIGHT_ASYMMETRIC:
1866 case ALGORITHM_ROTATING_ZERO_RESTART:
1867 case ALGORITHM_ROTATING_N_RESTART:
1868 if (sh->pd_idx == raid_disks-1)
1869 i--; /* Q D D D P */
1870 else if (i > sh->pd_idx)
1871 i -= 2; /* D D P Q D */
1873 case ALGORITHM_LEFT_SYMMETRIC:
1874 case ALGORITHM_RIGHT_SYMMETRIC:
1875 if (sh->pd_idx == raid_disks-1)
1876 i--; /* Q D D D P */
1881 i -= (sh->pd_idx + 2);
1884 case ALGORITHM_PARITY_0:
1887 case ALGORITHM_PARITY_N:
1889 case ALGORITHM_ROTATING_N_CONTINUE:
1890 if (sh->pd_idx == 0)
1891 i--; /* P D D D Q */
1892 else if (i > sh->pd_idx)
1893 i -= 2; /* D D Q P D */
1895 case ALGORITHM_LEFT_ASYMMETRIC_6:
1896 case ALGORITHM_RIGHT_ASYMMETRIC_6:
1900 case ALGORITHM_LEFT_SYMMETRIC_6:
1901 case ALGORITHM_RIGHT_SYMMETRIC_6:
1903 i += data_disks + 1;
1904 i -= (sh->pd_idx + 1);
1906 case ALGORITHM_PARITY_0_6:
1910 printk(KERN_CRIT "raid6: unsupported algorithm %d\n",
1917 chunk_number = stripe * data_disks + i;
1918 r_sector = (sector_t)chunk_number * sectors_per_chunk + chunk_offset;
1920 check = raid5_compute_sector(conf, r_sector,
1921 previous, &dummy1, &sh2);
1922 if (check != sh->sector || dummy1 != dd_idx || sh2.pd_idx != sh->pd_idx
1923 || sh2.qd_idx != sh->qd_idx) {
1924 printk(KERN_ERR "compute_blocknr: map not correct\n");
1932 schedule_reconstruction(struct stripe_head *sh, struct stripe_head_state *s,
1933 int rcw, int expand)
1935 int i, pd_idx = sh->pd_idx, disks = sh->disks;
1936 raid5_conf_t *conf = sh->raid_conf;
1937 int level = conf->level;
1940 /* if we are not expanding this is a proper write request, and
1941 * there will be bios with new data to be drained into the
1945 sh->reconstruct_state = reconstruct_state_drain_run;
1946 set_bit(STRIPE_OP_BIODRAIN, &s->ops_request);
1948 sh->reconstruct_state = reconstruct_state_run;
1950 set_bit(STRIPE_OP_RECONSTRUCT, &s->ops_request);
1952 for (i = disks; i--; ) {
1953 struct r5dev *dev = &sh->dev[i];
1956 set_bit(R5_LOCKED, &dev->flags);
1957 set_bit(R5_Wantdrain, &dev->flags);
1959 clear_bit(R5_UPTODATE, &dev->flags);
1963 if (s->locked + conf->max_degraded == disks)
1964 if (!test_and_set_bit(STRIPE_FULL_WRITE, &sh->state))
1965 atomic_inc(&conf->pending_full_writes);
1968 BUG_ON(!(test_bit(R5_UPTODATE, &sh->dev[pd_idx].flags) ||
1969 test_bit(R5_Wantcompute, &sh->dev[pd_idx].flags)));
1971 sh->reconstruct_state = reconstruct_state_prexor_drain_run;
1972 set_bit(STRIPE_OP_PREXOR, &s->ops_request);
1973 set_bit(STRIPE_OP_BIODRAIN, &s->ops_request);
1974 set_bit(STRIPE_OP_RECONSTRUCT, &s->ops_request);
1976 for (i = disks; i--; ) {
1977 struct r5dev *dev = &sh->dev[i];
1982 (test_bit(R5_UPTODATE, &dev->flags) ||
1983 test_bit(R5_Wantcompute, &dev->flags))) {
1984 set_bit(R5_Wantdrain, &dev->flags);
1985 set_bit(R5_LOCKED, &dev->flags);
1986 clear_bit(R5_UPTODATE, &dev->flags);
1992 /* keep the parity disk(s) locked while asynchronous operations
1995 set_bit(R5_LOCKED, &sh->dev[pd_idx].flags);
1996 clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
2000 int qd_idx = sh->qd_idx;
2001 struct r5dev *dev = &sh->dev[qd_idx];
2003 set_bit(R5_LOCKED, &dev->flags);
2004 clear_bit(R5_UPTODATE, &dev->flags);
2008 pr_debug("%s: stripe %llu locked: %d ops_request: %lx\n",
2009 __func__, (unsigned long long)sh->sector,
2010 s->locked, s->ops_request);
2014 * Each stripe/dev can have one or more bion attached.
2015 * toread/towrite point to the first in a chain.
2016 * The bi_next chain must be in order.
2018 static int add_stripe_bio(struct stripe_head *sh, struct bio *bi, int dd_idx, int forwrite)
2021 raid5_conf_t *conf = sh->raid_conf;
2024 pr_debug("adding bh b#%llu to stripe s#%llu\n",
2025 (unsigned long long)bi->bi_sector,
2026 (unsigned long long)sh->sector);
2029 spin_lock(&sh->lock);
2030 spin_lock_irq(&conf->device_lock);
2032 bip = &sh->dev[dd_idx].towrite;
2033 if (*bip == NULL && sh->dev[dd_idx].written == NULL)
2036 bip = &sh->dev[dd_idx].toread;
2037 while (*bip && (*bip)->bi_sector < bi->bi_sector) {
2038 if ((*bip)->bi_sector + ((*bip)->bi_size >> 9) > bi->bi_sector)
2040 bip = & (*bip)->bi_next;
2042 if (*bip && (*bip)->bi_sector < bi->bi_sector + ((bi->bi_size)>>9))
2045 BUG_ON(*bip && bi->bi_next && (*bip) != bi->bi_next);
2049 bi->bi_phys_segments++;
2050 spin_unlock_irq(&conf->device_lock);
2051 spin_unlock(&sh->lock);
2053 pr_debug("added bi b#%llu to stripe s#%llu, disk %d.\n",
2054 (unsigned long long)bi->bi_sector,
2055 (unsigned long long)sh->sector, dd_idx);
2057 if (conf->mddev->bitmap && firstwrite) {
2058 bitmap_startwrite(conf->mddev->bitmap, sh->sector,
2060 sh->bm_seq = conf->seq_flush+1;
2061 set_bit(STRIPE_BIT_DELAY, &sh->state);
2065 /* check if page is covered */
2066 sector_t sector = sh->dev[dd_idx].sector;
2067 for (bi=sh->dev[dd_idx].towrite;
2068 sector < sh->dev[dd_idx].sector + STRIPE_SECTORS &&
2069 bi && bi->bi_sector <= sector;
2070 bi = r5_next_bio(bi, sh->dev[dd_idx].sector)) {
2071 if (bi->bi_sector + (bi->bi_size>>9) >= sector)
2072 sector = bi->bi_sector + (bi->bi_size>>9);
2074 if (sector >= sh->dev[dd_idx].sector + STRIPE_SECTORS)
2075 set_bit(R5_OVERWRITE, &sh->dev[dd_idx].flags);
2080 set_bit(R5_Overlap, &sh->dev[dd_idx].flags);
2081 spin_unlock_irq(&conf->device_lock);
2082 spin_unlock(&sh->lock);
2086 static void end_reshape(raid5_conf_t *conf);
2088 static void stripe_set_idx(sector_t stripe, raid5_conf_t *conf, int previous,
2089 struct stripe_head *sh)
2091 int sectors_per_chunk =
2092 previous ? (conf->prev_chunk >> 9)
2093 : (conf->chunk_size >> 9);
2095 int chunk_offset = sector_div(stripe, sectors_per_chunk);
2096 int disks = previous ? conf->previous_raid_disks : conf->raid_disks;
2098 raid5_compute_sector(conf,
2099 stripe * (disks - conf->max_degraded)
2100 *sectors_per_chunk + chunk_offset,
2106 handle_failed_stripe(raid5_conf_t *conf, struct stripe_head *sh,
2107 struct stripe_head_state *s, int disks,
2108 struct bio **return_bi)
2111 for (i = disks; i--; ) {
2115 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
2118 rdev = rcu_dereference(conf->disks[i].rdev);
2119 if (rdev && test_bit(In_sync, &rdev->flags))
2120 /* multiple read failures in one stripe */
2121 md_error(conf->mddev, rdev);
2124 spin_lock_irq(&conf->device_lock);
2125 /* fail all writes first */
2126 bi = sh->dev[i].towrite;
2127 sh->dev[i].towrite = NULL;
2133 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
2134 wake_up(&conf->wait_for_overlap);
2136 while (bi && bi->bi_sector <
2137 sh->dev[i].sector + STRIPE_SECTORS) {
2138 struct bio *nextbi = r5_next_bio(bi, sh->dev[i].sector);
2139 clear_bit(BIO_UPTODATE, &bi->bi_flags);
2140 if (!raid5_dec_bi_phys_segments(bi)) {
2141 md_write_end(conf->mddev);
2142 bi->bi_next = *return_bi;
2147 /* and fail all 'written' */
2148 bi = sh->dev[i].written;
2149 sh->dev[i].written = NULL;
2150 if (bi) bitmap_end = 1;
2151 while (bi && bi->bi_sector <
2152 sh->dev[i].sector + STRIPE_SECTORS) {
2153 struct bio *bi2 = r5_next_bio(bi, sh->dev[i].sector);
2154 clear_bit(BIO_UPTODATE, &bi->bi_flags);
2155 if (!raid5_dec_bi_phys_segments(bi)) {
2156 md_write_end(conf->mddev);
2157 bi->bi_next = *return_bi;
2163 /* fail any reads if this device is non-operational and
2164 * the data has not reached the cache yet.
2166 if (!test_bit(R5_Wantfill, &sh->dev[i].flags) &&
2167 (!test_bit(R5_Insync, &sh->dev[i].flags) ||
2168 test_bit(R5_ReadError, &sh->dev[i].flags))) {
2169 bi = sh->dev[i].toread;
2170 sh->dev[i].toread = NULL;
2171 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
2172 wake_up(&conf->wait_for_overlap);
2173 if (bi) s->to_read--;
2174 while (bi && bi->bi_sector <
2175 sh->dev[i].sector + STRIPE_SECTORS) {
2176 struct bio *nextbi =
2177 r5_next_bio(bi, sh->dev[i].sector);
2178 clear_bit(BIO_UPTODATE, &bi->bi_flags);
2179 if (!raid5_dec_bi_phys_segments(bi)) {
2180 bi->bi_next = *return_bi;
2186 spin_unlock_irq(&conf->device_lock);
2188 bitmap_endwrite(conf->mddev->bitmap, sh->sector,
2189 STRIPE_SECTORS, 0, 0);
2192 if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
2193 if (atomic_dec_and_test(&conf->pending_full_writes))
2194 md_wakeup_thread(conf->mddev->thread);
2197 /* fetch_block5 - checks the given member device to see if its data needs
2198 * to be read or computed to satisfy a request.
2200 * Returns 1 when no more member devices need to be checked, otherwise returns
2201 * 0 to tell the loop in handle_stripe_fill5 to continue
2203 static int fetch_block5(struct stripe_head *sh, struct stripe_head_state *s,
2204 int disk_idx, int disks)
2206 struct r5dev *dev = &sh->dev[disk_idx];
2207 struct r5dev *failed_dev = &sh->dev[s->failed_num];
2209 /* is the data in this block needed, and can we get it? */
2210 if (!test_bit(R5_LOCKED, &dev->flags) &&
2211 !test_bit(R5_UPTODATE, &dev->flags) &&
2213 (dev->towrite && !test_bit(R5_OVERWRITE, &dev->flags)) ||
2214 s->syncing || s->expanding ||
2216 (failed_dev->toread ||
2217 (failed_dev->towrite &&
2218 !test_bit(R5_OVERWRITE, &failed_dev->flags)))))) {
2219 /* We would like to get this block, possibly by computing it,
2220 * otherwise read it if the backing disk is insync
2222 if ((s->uptodate == disks - 1) &&
2223 (s->failed && disk_idx == s->failed_num)) {
2224 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
2225 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
2226 set_bit(R5_Wantcompute, &dev->flags);
2227 sh->ops.target = disk_idx;
2228 sh->ops.target2 = -1;
2230 /* Careful: from this point on 'uptodate' is in the eye
2231 * of raid_run_ops which services 'compute' operations
2232 * before writes. R5_Wantcompute flags a block that will
2233 * be R5_UPTODATE by the time it is needed for a
2234 * subsequent operation.
2237 return 1; /* uptodate + compute == disks */
2238 } else if (test_bit(R5_Insync, &dev->flags)) {
2239 set_bit(R5_LOCKED, &dev->flags);
2240 set_bit(R5_Wantread, &dev->flags);
2242 pr_debug("Reading block %d (sync=%d)\n", disk_idx,
2251 * handle_stripe_fill5 - read or compute data to satisfy pending requests.
2253 static void handle_stripe_fill5(struct stripe_head *sh,
2254 struct stripe_head_state *s, int disks)
2258 /* look for blocks to read/compute, skip this if a compute
2259 * is already in flight, or if the stripe contents are in the
2260 * midst of changing due to a write
2262 if (!test_bit(STRIPE_COMPUTE_RUN, &sh->state) && !sh->check_state &&
2263 !sh->reconstruct_state)
2264 for (i = disks; i--; )
2265 if (fetch_block5(sh, s, i, disks))
2267 set_bit(STRIPE_HANDLE, &sh->state);
2270 /* fetch_block6 - checks the given member device to see if its data needs
2271 * to be read or computed to satisfy a request.
2273 * Returns 1 when no more member devices need to be checked, otherwise returns
2274 * 0 to tell the loop in handle_stripe_fill6 to continue
2276 static int fetch_block6(struct stripe_head *sh, struct stripe_head_state *s,
2277 struct r6_state *r6s, int disk_idx, int disks)
2279 struct r5dev *dev = &sh->dev[disk_idx];
2280 struct r5dev *fdev[2] = { &sh->dev[r6s->failed_num[0]],
2281 &sh->dev[r6s->failed_num[1]] };
2283 if (!test_bit(R5_LOCKED, &dev->flags) &&
2284 !test_bit(R5_UPTODATE, &dev->flags) &&
2286 (dev->towrite && !test_bit(R5_OVERWRITE, &dev->flags)) ||
2287 s->syncing || s->expanding ||
2289 (fdev[0]->toread || s->to_write)) ||
2291 (fdev[1]->toread || s->to_write)))) {
2292 /* we would like to get this block, possibly by computing it,
2293 * otherwise read it if the backing disk is insync
2295 BUG_ON(test_bit(R5_Wantcompute, &dev->flags));
2296 BUG_ON(test_bit(R5_Wantread, &dev->flags));
2297 if ((s->uptodate == disks - 1) &&
2298 (s->failed && (disk_idx == r6s->failed_num[0] ||
2299 disk_idx == r6s->failed_num[1]))) {
2300 /* have disk failed, and we're requested to fetch it;
2303 pr_debug("Computing stripe %llu block %d\n",
2304 (unsigned long long)sh->sector, disk_idx);
2305 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
2306 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
2307 set_bit(R5_Wantcompute, &dev->flags);
2308 sh->ops.target = disk_idx;
2309 sh->ops.target2 = -1; /* no 2nd target */
2313 } else if (s->uptodate == disks-2 && s->failed >= 2) {
2314 /* Computing 2-failure is *very* expensive; only
2315 * do it if failed >= 2
2318 for (other = disks; other--; ) {
2319 if (other == disk_idx)
2321 if (!test_bit(R5_UPTODATE,
2322 &sh->dev[other].flags))
2326 pr_debug("Computing stripe %llu blocks %d,%d\n",
2327 (unsigned long long)sh->sector,
2329 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
2330 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
2331 set_bit(R5_Wantcompute, &sh->dev[disk_idx].flags);
2332 set_bit(R5_Wantcompute, &sh->dev[other].flags);
2333 sh->ops.target = disk_idx;
2334 sh->ops.target2 = other;
2338 } else if (test_bit(R5_Insync, &dev->flags)) {
2339 set_bit(R5_LOCKED, &dev->flags);
2340 set_bit(R5_Wantread, &dev->flags);
2342 pr_debug("Reading block %d (sync=%d)\n",
2343 disk_idx, s->syncing);
2351 * handle_stripe_fill6 - read or compute data to satisfy pending requests.
2353 static void handle_stripe_fill6(struct stripe_head *sh,
2354 struct stripe_head_state *s, struct r6_state *r6s,
2359 /* look for blocks to read/compute, skip this if a compute
2360 * is already in flight, or if the stripe contents are in the
2361 * midst of changing due to a write
2363 if (!test_bit(STRIPE_COMPUTE_RUN, &sh->state) && !sh->check_state &&
2364 !sh->reconstruct_state)
2365 for (i = disks; i--; )
2366 if (fetch_block6(sh, s, r6s, i, disks))
2368 set_bit(STRIPE_HANDLE, &sh->state);
2372 /* handle_stripe_clean_event
2373 * any written block on an uptodate or failed drive can be returned.
2374 * Note that if we 'wrote' to a failed drive, it will be UPTODATE, but
2375 * never LOCKED, so we don't need to test 'failed' directly.
2377 static void handle_stripe_clean_event(raid5_conf_t *conf,
2378 struct stripe_head *sh, int disks, struct bio **return_bi)
2383 for (i = disks; i--; )
2384 if (sh->dev[i].written) {
2386 if (!test_bit(R5_LOCKED, &dev->flags) &&
2387 test_bit(R5_UPTODATE, &dev->flags)) {
2388 /* We can return any write requests */
2389 struct bio *wbi, *wbi2;
2391 pr_debug("Return write for disc %d\n", i);
2392 spin_lock_irq(&conf->device_lock);
2394 dev->written = NULL;
2395 while (wbi && wbi->bi_sector <
2396 dev->sector + STRIPE_SECTORS) {
2397 wbi2 = r5_next_bio(wbi, dev->sector);
2398 if (!raid5_dec_bi_phys_segments(wbi)) {
2399 md_write_end(conf->mddev);
2400 wbi->bi_next = *return_bi;
2405 if (dev->towrite == NULL)
2407 spin_unlock_irq(&conf->device_lock);
2409 bitmap_endwrite(conf->mddev->bitmap,
2412 !test_bit(STRIPE_DEGRADED, &sh->state),
2417 if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
2418 if (atomic_dec_and_test(&conf->pending_full_writes))
2419 md_wakeup_thread(conf->mddev->thread);
2422 static void handle_stripe_dirtying5(raid5_conf_t *conf,
2423 struct stripe_head *sh, struct stripe_head_state *s, int disks)
2425 int rmw = 0, rcw = 0, i;
2426 for (i = disks; i--; ) {
2427 /* would I have to read this buffer for read_modify_write */
2428 struct r5dev *dev = &sh->dev[i];
2429 if ((dev->towrite || i == sh->pd_idx) &&
2430 !test_bit(R5_LOCKED, &dev->flags) &&
2431 !(test_bit(R5_UPTODATE, &dev->flags) ||
2432 test_bit(R5_Wantcompute, &dev->flags))) {
2433 if (test_bit(R5_Insync, &dev->flags))
2436 rmw += 2*disks; /* cannot read it */
2438 /* Would I have to read this buffer for reconstruct_write */
2439 if (!test_bit(R5_OVERWRITE, &dev->flags) && i != sh->pd_idx &&
2440 !test_bit(R5_LOCKED, &dev->flags) &&
2441 !(test_bit(R5_UPTODATE, &dev->flags) ||
2442 test_bit(R5_Wantcompute, &dev->flags))) {
2443 if (test_bit(R5_Insync, &dev->flags)) rcw++;
2448 pr_debug("for sector %llu, rmw=%d rcw=%d\n",
2449 (unsigned long long)sh->sector, rmw, rcw);
2450 set_bit(STRIPE_HANDLE, &sh->state);
2451 if (rmw < rcw && rmw > 0)
2452 /* prefer read-modify-write, but need to get some data */
2453 for (i = disks; i--; ) {
2454 struct r5dev *dev = &sh->dev[i];
2455 if ((dev->towrite || i == sh->pd_idx) &&
2456 !test_bit(R5_LOCKED, &dev->flags) &&
2457 !(test_bit(R5_UPTODATE, &dev->flags) ||
2458 test_bit(R5_Wantcompute, &dev->flags)) &&
2459 test_bit(R5_Insync, &dev->flags)) {
2461 test_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2462 pr_debug("Read_old block "
2463 "%d for r-m-w\n", i);
2464 set_bit(R5_LOCKED, &dev->flags);
2465 set_bit(R5_Wantread, &dev->flags);
2468 set_bit(STRIPE_DELAYED, &sh->state);
2469 set_bit(STRIPE_HANDLE, &sh->state);
2473 if (rcw <= rmw && rcw > 0)
2474 /* want reconstruct write, but need to get some data */
2475 for (i = disks; i--; ) {
2476 struct r5dev *dev = &sh->dev[i];
2477 if (!test_bit(R5_OVERWRITE, &dev->flags) &&
2479 !test_bit(R5_LOCKED, &dev->flags) &&
2480 !(test_bit(R5_UPTODATE, &dev->flags) ||
2481 test_bit(R5_Wantcompute, &dev->flags)) &&
2482 test_bit(R5_Insync, &dev->flags)) {
2484 test_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2485 pr_debug("Read_old block "
2486 "%d for Reconstruct\n", i);
2487 set_bit(R5_LOCKED, &dev->flags);
2488 set_bit(R5_Wantread, &dev->flags);
2491 set_bit(STRIPE_DELAYED, &sh->state);
2492 set_bit(STRIPE_HANDLE, &sh->state);
2496 /* now if nothing is locked, and if we have enough data,
2497 * we can start a write request
2499 /* since handle_stripe can be called at any time we need to handle the
2500 * case where a compute block operation has been submitted and then a
2501 * subsequent call wants to start a write request. raid_run_ops only
2502 * handles the case where compute block and reconstruct are requested
2503 * simultaneously. If this is not the case then new writes need to be
2504 * held off until the compute completes.
2506 if ((s->req_compute || !test_bit(STRIPE_COMPUTE_RUN, &sh->state)) &&
2507 (s->locked == 0 && (rcw == 0 || rmw == 0) &&
2508 !test_bit(STRIPE_BIT_DELAY, &sh->state)))
2509 schedule_reconstruction(sh, s, rcw == 0, 0);
2512 static void handle_stripe_dirtying6(raid5_conf_t *conf,
2513 struct stripe_head *sh, struct stripe_head_state *s,
2514 struct r6_state *r6s, int disks)
2516 int rcw = 0, pd_idx = sh->pd_idx, i;
2517 int qd_idx = sh->qd_idx;
2519 set_bit(STRIPE_HANDLE, &sh->state);
2520 for (i = disks; i--; ) {
2521 struct r5dev *dev = &sh->dev[i];
2522 /* check if we haven't enough data */
2523 if (!test_bit(R5_OVERWRITE, &dev->flags) &&
2524 i != pd_idx && i != qd_idx &&
2525 !test_bit(R5_LOCKED, &dev->flags) &&
2526 !(test_bit(R5_UPTODATE, &dev->flags) ||
2527 test_bit(R5_Wantcompute, &dev->flags))) {
2529 if (!test_bit(R5_Insync, &dev->flags))
2530 continue; /* it's a failed drive */
2533 test_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2534 pr_debug("Read_old stripe %llu "
2535 "block %d for Reconstruct\n",
2536 (unsigned long long)sh->sector, i);
2537 set_bit(R5_LOCKED, &dev->flags);
2538 set_bit(R5_Wantread, &dev->flags);
2541 pr_debug("Request delayed stripe %llu "
2542 "block %d for Reconstruct\n",
2543 (unsigned long long)sh->sector, i);
2544 set_bit(STRIPE_DELAYED, &sh->state);
2545 set_bit(STRIPE_HANDLE, &sh->state);
2549 /* now if nothing is locked, and if we have enough data, we can start a
2552 if ((s->req_compute || !test_bit(STRIPE_COMPUTE_RUN, &sh->state)) &&
2553 s->locked == 0 && rcw == 0 &&
2554 !test_bit(STRIPE_BIT_DELAY, &sh->state)) {
2555 schedule_reconstruction(sh, s, 1, 0);
2559 static void handle_parity_checks5(raid5_conf_t *conf, struct stripe_head *sh,
2560 struct stripe_head_state *s, int disks)
2562 struct r5dev *dev = NULL;
2564 set_bit(STRIPE_HANDLE, &sh->state);
2566 switch (sh->check_state) {
2567 case check_state_idle:
2568 /* start a new check operation if there are no failures */
2569 if (s->failed == 0) {
2570 BUG_ON(s->uptodate != disks);
2571 sh->check_state = check_state_run;
2572 set_bit(STRIPE_OP_CHECK, &s->ops_request);
2573 clear_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags);
2577 dev = &sh->dev[s->failed_num];
2579 case check_state_compute_result:
2580 sh->check_state = check_state_idle;
2582 dev = &sh->dev[sh->pd_idx];
2584 /* check that a write has not made the stripe insync */
2585 if (test_bit(STRIPE_INSYNC, &sh->state))
2588 /* either failed parity check, or recovery is happening */
2589 BUG_ON(!test_bit(R5_UPTODATE, &dev->flags));
2590 BUG_ON(s->uptodate != disks);
2592 set_bit(R5_LOCKED, &dev->flags);
2594 set_bit(R5_Wantwrite, &dev->flags);
2596 clear_bit(STRIPE_DEGRADED, &sh->state);
2597 set_bit(STRIPE_INSYNC, &sh->state);
2599 case check_state_run:
2600 break; /* we will be called again upon completion */
2601 case check_state_check_result:
2602 sh->check_state = check_state_idle;
2604 /* if a failure occurred during the check operation, leave
2605 * STRIPE_INSYNC not set and let the stripe be handled again
2610 /* handle a successful check operation, if parity is correct
2611 * we are done. Otherwise update the mismatch count and repair
2612 * parity if !MD_RECOVERY_CHECK
2614 if ((sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) == 0)
2615 /* parity is correct (on disc,
2616 * not in buffer any more)
2618 set_bit(STRIPE_INSYNC, &sh->state);
2620 conf->mddev->resync_mismatches += STRIPE_SECTORS;
2621 if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery))
2622 /* don't try to repair!! */
2623 set_bit(STRIPE_INSYNC, &sh->state);
2625 sh->check_state = check_state_compute_run;
2626 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
2627 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
2628 set_bit(R5_Wantcompute,
2629 &sh->dev[sh->pd_idx].flags);
2630 sh->ops.target = sh->pd_idx;
2631 sh->ops.target2 = -1;
2636 case check_state_compute_run:
2639 printk(KERN_ERR "%s: unknown check_state: %d sector: %llu\n",
2640 __func__, sh->check_state,
2641 (unsigned long long) sh->sector);
2647 static void handle_parity_checks6(raid5_conf_t *conf, struct stripe_head *sh,
2648 struct stripe_head_state *s,
2649 struct r6_state *r6s, int disks)
2651 int pd_idx = sh->pd_idx;
2652 int qd_idx = sh->qd_idx;
2655 set_bit(STRIPE_HANDLE, &sh->state);
2657 BUG_ON(s->failed > 2);
2659 /* Want to check and possibly repair P and Q.
2660 * However there could be one 'failed' device, in which
2661 * case we can only check one of them, possibly using the
2662 * other to generate missing data
2665 switch (sh->check_state) {
2666 case check_state_idle:
2667 /* start a new check operation if there are < 2 failures */
2668 if (s->failed == r6s->q_failed) {
2669 /* The only possible failed device holds Q, so it
2670 * makes sense to check P (If anything else were failed,
2671 * we would have used P to recreate it).
2673 sh->check_state = check_state_run;
2675 if (!r6s->q_failed && s->failed < 2) {
2676 /* Q is not failed, and we didn't use it to generate
2677 * anything, so it makes sense to check it
2679 if (sh->check_state == check_state_run)
2680 sh->check_state = check_state_run_pq;
2682 sh->check_state = check_state_run_q;
2685 /* discard potentially stale zero_sum_result */
2686 sh->ops.zero_sum_result = 0;
2688 if (sh->check_state == check_state_run) {
2689 /* async_xor_zero_sum destroys the contents of P */
2690 clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
2693 if (sh->check_state >= check_state_run &&
2694 sh->check_state <= check_state_run_pq) {
2695 /* async_syndrome_zero_sum preserves P and Q, so
2696 * no need to mark them !uptodate here
2698 set_bit(STRIPE_OP_CHECK, &s->ops_request);
2702 /* we have 2-disk failure */
2703 BUG_ON(s->failed != 2);
2705 case check_state_compute_result:
2706 sh->check_state = check_state_idle;
2708 /* check that a write has not made the stripe insync */
2709 if (test_bit(STRIPE_INSYNC, &sh->state))
2712 /* now write out any block on a failed drive,
2713 * or P or Q if they were recomputed
2715 BUG_ON(s->uptodate < disks - 1); /* We don't need Q to recover */
2716 if (s->failed == 2) {
2717 dev = &sh->dev[r6s->failed_num[1]];
2719 set_bit(R5_LOCKED, &dev->flags);
2720 set_bit(R5_Wantwrite, &dev->flags);
2722 if (s->failed >= 1) {
2723 dev = &sh->dev[r6s->failed_num[0]];
2725 set_bit(R5_LOCKED, &dev->flags);
2726 set_bit(R5_Wantwrite, &dev->flags);
2728 if (sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) {
2729 dev = &sh->dev[pd_idx];
2731 set_bit(R5_LOCKED, &dev->flags);
2732 set_bit(R5_Wantwrite, &dev->flags);
2734 if (sh->ops.zero_sum_result & SUM_CHECK_Q_RESULT) {
2735 dev = &sh->dev[qd_idx];
2737 set_bit(R5_LOCKED, &dev->flags);
2738 set_bit(R5_Wantwrite, &dev->flags);
2740 clear_bit(STRIPE_DEGRADED, &sh->state);
2742 set_bit(STRIPE_INSYNC, &sh->state);
2744 case check_state_run:
2745 case check_state_run_q:
2746 case check_state_run_pq:
2747 break; /* we will be called again upon completion */
2748 case check_state_check_result:
2749 sh->check_state = check_state_idle;
2751 /* handle a successful check operation, if parity is correct
2752 * we are done. Otherwise update the mismatch count and repair
2753 * parity if !MD_RECOVERY_CHECK
2755 if (sh->ops.zero_sum_result == 0) {
2756 /* both parities are correct */
2758 set_bit(STRIPE_INSYNC, &sh->state);
2760 /* in contrast to the raid5 case we can validate
2761 * parity, but still have a failure to write
2764 sh->check_state = check_state_compute_result;
2765 /* Returning at this point means that we may go
2766 * off and bring p and/or q uptodate again so
2767 * we make sure to check zero_sum_result again
2768 * to verify if p or q need writeback
2772 conf->mddev->resync_mismatches += STRIPE_SECTORS;
2773 if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery))
2774 /* don't try to repair!! */
2775 set_bit(STRIPE_INSYNC, &sh->state);
2777 int *target = &sh->ops.target;
2779 sh->ops.target = -1;
2780 sh->ops.target2 = -1;
2781 sh->check_state = check_state_compute_run;
2782 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
2783 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
2784 if (sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) {
2785 set_bit(R5_Wantcompute,
2786 &sh->dev[pd_idx].flags);
2788 target = &sh->ops.target2;
2791 if (sh->ops.zero_sum_result & SUM_CHECK_Q_RESULT) {
2792 set_bit(R5_Wantcompute,
2793 &sh->dev[qd_idx].flags);
2800 case check_state_compute_run:
2803 printk(KERN_ERR "%s: unknown check_state: %d sector: %llu\n",
2804 __func__, sh->check_state,
2805 (unsigned long long) sh->sector);
2810 static void handle_stripe_expansion(raid5_conf_t *conf, struct stripe_head *sh,
2811 struct r6_state *r6s)
2815 /* We have read all the blocks in this stripe and now we need to
2816 * copy some of them into a target stripe for expand.
2818 struct dma_async_tx_descriptor *tx = NULL;
2819 clear_bit(STRIPE_EXPAND_SOURCE, &sh->state);
2820 for (i = 0; i < sh->disks; i++)
2821 if (i != sh->pd_idx && i != sh->qd_idx) {
2823 struct stripe_head *sh2;
2824 struct async_submit_ctl submit;
2826 sector_t bn = compute_blocknr(sh, i, 1);
2827 sector_t s = raid5_compute_sector(conf, bn, 0,
2829 sh2 = get_active_stripe(conf, s, 0, 1, 1);
2831 /* so far only the early blocks of this stripe
2832 * have been requested. When later blocks
2833 * get requested, we will try again
2836 if (!test_bit(STRIPE_EXPANDING, &sh2->state) ||
2837 test_bit(R5_Expanded, &sh2->dev[dd_idx].flags)) {
2838 /* must have already done this block */
2839 release_stripe(sh2);
2843 /* place all the copies on one channel */
2844 init_async_submit(&submit, 0, tx, NULL, NULL, NULL);
2845 tx = async_memcpy(sh2->dev[dd_idx].page,
2846 sh->dev[i].page, 0, 0, STRIPE_SIZE,
2849 set_bit(R5_Expanded, &sh2->dev[dd_idx].flags);
2850 set_bit(R5_UPTODATE, &sh2->dev[dd_idx].flags);
2851 for (j = 0; j < conf->raid_disks; j++)
2852 if (j != sh2->pd_idx &&
2853 (!r6s || j != sh2->qd_idx) &&
2854 !test_bit(R5_Expanded, &sh2->dev[j].flags))
2856 if (j == conf->raid_disks) {
2857 set_bit(STRIPE_EXPAND_READY, &sh2->state);
2858 set_bit(STRIPE_HANDLE, &sh2->state);
2860 release_stripe(sh2);
2863 /* done submitting copies, wait for them to complete */
2866 dma_wait_for_async_tx(tx);
2872 * handle_stripe - do things to a stripe.
2874 * We lock the stripe and then examine the state of various bits
2875 * to see what needs to be done.
2877 * return some read request which now have data
2878 * return some write requests which are safely on disc
2879 * schedule a read on some buffers
2880 * schedule a write of some buffers
2881 * return confirmation of parity correctness
2883 * buffers are taken off read_list or write_list, and bh_cache buffers
2884 * get BH_Lock set before the stripe lock is released.
2888 static bool handle_stripe5(struct stripe_head *sh)
2890 raid5_conf_t *conf = sh->raid_conf;
2891 int disks = sh->disks, i;
2892 struct bio *return_bi = NULL;
2893 struct stripe_head_state s;
2895 mdk_rdev_t *blocked_rdev = NULL;
2898 memset(&s, 0, sizeof(s));
2899 pr_debug("handling stripe %llu, state=%#lx cnt=%d, pd_idx=%d check:%d "
2900 "reconstruct:%d\n", (unsigned long long)sh->sector, sh->state,
2901 atomic_read(&sh->count), sh->pd_idx, sh->check_state,
2902 sh->reconstruct_state);
2904 spin_lock(&sh->lock);
2905 clear_bit(STRIPE_HANDLE, &sh->state);
2906 clear_bit(STRIPE_DELAYED, &sh->state);
2908 s.syncing = test_bit(STRIPE_SYNCING, &sh->state);
2909 s.expanding = test_bit(STRIPE_EXPAND_SOURCE, &sh->state);
2910 s.expanded = test_bit(STRIPE_EXPAND_READY, &sh->state);
2912 /* Now to look around and see what can be done */
2914 for (i=disks; i--; ) {
2916 struct r5dev *dev = &sh->dev[i];
2917 clear_bit(R5_Insync, &dev->flags);
2919 pr_debug("check %d: state 0x%lx toread %p read %p write %p "
2920 "written %p\n", i, dev->flags, dev->toread, dev->read,
2921 dev->towrite, dev->written);
2923 /* maybe we can request a biofill operation
2925 * new wantfill requests are only permitted while
2926 * ops_complete_biofill is guaranteed to be inactive
2928 if (test_bit(R5_UPTODATE, &dev->flags) && dev->toread &&
2929 !test_bit(STRIPE_BIOFILL_RUN, &sh->state))
2930 set_bit(R5_Wantfill, &dev->flags);
2932 /* now count some things */
2933 if (test_bit(R5_LOCKED, &dev->flags)) s.locked++;
2934 if (test_bit(R5_UPTODATE, &dev->flags)) s.uptodate++;
2935 if (test_bit(R5_Wantcompute, &dev->flags)) s.compute++;
2937 if (test_bit(R5_Wantfill, &dev->flags))
2939 else if (dev->toread)
2943 if (!test_bit(R5_OVERWRITE, &dev->flags))
2948 rdev = rcu_dereference(conf->disks[i].rdev);
2949 if (blocked_rdev == NULL &&
2950 rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
2951 blocked_rdev = rdev;
2952 atomic_inc(&rdev->nr_pending);
2954 if (!rdev || !test_bit(In_sync, &rdev->flags)) {
2955 /* The ReadError flag will just be confusing now */
2956 clear_bit(R5_ReadError, &dev->flags);
2957 clear_bit(R5_ReWrite, &dev->flags);
2959 if (!rdev || !test_bit(In_sync, &rdev->flags)
2960 || test_bit(R5_ReadError, &dev->flags)) {
2964 set_bit(R5_Insync, &dev->flags);
2968 if (unlikely(blocked_rdev)) {
2969 if (s.syncing || s.expanding || s.expanded ||
2970 s.to_write || s.written) {
2971 set_bit(STRIPE_HANDLE, &sh->state);
2974 /* There is nothing for the blocked_rdev to block */
2975 rdev_dec_pending(blocked_rdev, conf->mddev);
2976 blocked_rdev = NULL;
2979 if (s.to_fill && !test_bit(STRIPE_BIOFILL_RUN, &sh->state)) {
2980 set_bit(STRIPE_OP_BIOFILL, &s.ops_request);
2981 set_bit(STRIPE_BIOFILL_RUN, &sh->state);
2984 pr_debug("locked=%d uptodate=%d to_read=%d"
2985 " to_write=%d failed=%d failed_num=%d\n",
2986 s.locked, s.uptodate, s.to_read, s.to_write,
2987 s.failed, s.failed_num);
2988 /* check if the array has lost two devices and, if so, some requests might
2991 if (s.failed > 1 && s.to_read+s.to_write+s.written)
2992 handle_failed_stripe(conf, sh, &s, disks, &return_bi);
2993 if (s.failed > 1 && s.syncing) {
2994 md_done_sync(conf->mddev, STRIPE_SECTORS,0);
2995 clear_bit(STRIPE_SYNCING, &sh->state);
2999 /* might be able to return some write requests if the parity block
3000 * is safe, or on a failed drive
3002 dev = &sh->dev[sh->pd_idx];
3004 ((test_bit(R5_Insync, &dev->flags) &&
3005 !test_bit(R5_LOCKED, &dev->flags) &&
3006 test_bit(R5_UPTODATE, &dev->flags)) ||
3007 (s.failed == 1 && s.failed_num == sh->pd_idx)))
3008 handle_stripe_clean_event(conf, sh, disks, &return_bi);
3010 /* Now we might consider reading some blocks, either to check/generate
3011 * parity, or to satisfy requests
3012 * or to load a block that is being partially written.
3014 if (s.to_read || s.non_overwrite ||
3015 (s.syncing && (s.uptodate + s.compute < disks)) || s.expanding)
3016 handle_stripe_fill5(sh, &s, disks);
3018 /* Now we check to see if any write operations have recently
3022 if (sh->reconstruct_state == reconstruct_state_prexor_drain_result)
3024 if (sh->reconstruct_state == reconstruct_state_drain_result ||
3025 sh->reconstruct_state == reconstruct_state_prexor_drain_result) {
3026 sh->reconstruct_state = reconstruct_state_idle;
3028 /* All the 'written' buffers and the parity block are ready to
3029 * be written back to disk
3031 BUG_ON(!test_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags));
3032 for (i = disks; i--; ) {
3034 if (test_bit(R5_LOCKED, &dev->flags) &&
3035 (i == sh->pd_idx || dev->written)) {
3036 pr_debug("Writing block %d\n", i);
3037 set_bit(R5_Wantwrite, &dev->flags);
3040 if (!test_bit(R5_Insync, &dev->flags) ||
3041 (i == sh->pd_idx && s.failed == 0))
3042 set_bit(STRIPE_INSYNC, &sh->state);
3045 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
3046 atomic_dec(&conf->preread_active_stripes);
3047 if (atomic_read(&conf->preread_active_stripes) <
3049 md_wakeup_thread(conf->mddev->thread);
3053 /* Now to consider new write requests and what else, if anything
3054 * should be read. We do not handle new writes when:
3055 * 1/ A 'write' operation (copy+xor) is already in flight.
3056 * 2/ A 'check' operation is in flight, as it may clobber the parity
3059 if (s.to_write && !sh->reconstruct_state && !sh->check_state)
3060 handle_stripe_dirtying5(conf, sh, &s, disks);
3062 /* maybe we need to check and possibly fix the parity for this stripe
3063 * Any reads will already have been scheduled, so we just see if enough
3064 * data is available. The parity check is held off while parity
3065 * dependent operations are in flight.
3067 if (sh->check_state ||
3068 (s.syncing && s.locked == 0 &&
3069 !test_bit(STRIPE_COMPUTE_RUN, &sh->state) &&
3070 !test_bit(STRIPE_INSYNC, &sh->state)))
3071 handle_parity_checks5(conf, sh, &s, disks);
3073 if (s.syncing && s.locked == 0 && test_bit(STRIPE_INSYNC, &sh->state)) {
3074 md_done_sync(conf->mddev, STRIPE_SECTORS,1);
3075 clear_bit(STRIPE_SYNCING, &sh->state);
3078 /* If the failed drive is just a ReadError, then we might need to progress
3079 * the repair/check process
3081 if (s.failed == 1 && !conf->mddev->ro &&
3082 test_bit(R5_ReadError, &sh->dev[s.failed_num].flags)
3083 && !test_bit(R5_LOCKED, &sh->dev[s.failed_num].flags)
3084 && test_bit(R5_UPTODATE, &sh->dev[s.failed_num].flags)
3086 dev = &sh->dev[s.failed_num];
3087 if (!test_bit(R5_ReWrite, &dev->flags)) {
3088 set_bit(R5_Wantwrite, &dev->flags);
3089 set_bit(R5_ReWrite, &dev->flags);
3090 set_bit(R5_LOCKED, &dev->flags);
3093 /* let's read it back */
3094 set_bit(R5_Wantread, &dev->flags);
3095 set_bit(R5_LOCKED, &dev->flags);
3100 /* Finish reconstruct operations initiated by the expansion process */
3101 if (sh->reconstruct_state == reconstruct_state_result) {
3102 struct stripe_head *sh2
3103 = get_active_stripe(conf, sh->sector, 1, 1, 1);
3104 if (sh2 && test_bit(STRIPE_EXPAND_SOURCE, &sh2->state)) {
3105 /* sh cannot be written until sh2 has been read.
3106 * so arrange for sh to be delayed a little
3108 set_bit(STRIPE_DELAYED, &sh->state);
3109 set_bit(STRIPE_HANDLE, &sh->state);
3110 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE,
3112 atomic_inc(&conf->preread_active_stripes);
3113 release_stripe(sh2);
3117 release_stripe(sh2);
3119 sh->reconstruct_state = reconstruct_state_idle;
3120 clear_bit(STRIPE_EXPANDING, &sh->state);
3121 for (i = conf->raid_disks; i--; ) {
3122 set_bit(R5_Wantwrite, &sh->dev[i].flags);
3123 set_bit(R5_LOCKED, &sh->dev[i].flags);
3128 if (s.expanded && test_bit(STRIPE_EXPANDING, &sh->state) &&
3129 !sh->reconstruct_state) {
3130 /* Need to write out all blocks after computing parity */
3131 sh->disks = conf->raid_disks;
3132 stripe_set_idx(sh->sector, conf, 0, sh);
3133 schedule_reconstruction(sh, &s, 1, 1);
3134 } else if (s.expanded && !sh->reconstruct_state && s.locked == 0) {
3135 clear_bit(STRIPE_EXPAND_READY, &sh->state);
3136 atomic_dec(&conf->reshape_stripes);
3137 wake_up(&conf->wait_for_overlap);
3138 md_done_sync(conf->mddev, STRIPE_SECTORS, 1);
3141 if (s.expanding && s.locked == 0 &&
3142 !test_bit(STRIPE_COMPUTE_RUN, &sh->state))
3143 handle_stripe_expansion(conf, sh, NULL);
3146 spin_unlock(&sh->lock);
3148 /* wait for this device to become unblocked */
3149 if (unlikely(blocked_rdev))
3150 md_wait_for_blocked_rdev(blocked_rdev, conf->mddev);
3153 raid_run_ops(sh, s.ops_request);
3157 return_io(return_bi);
3159 return blocked_rdev == NULL;
3162 static bool handle_stripe6(struct stripe_head *sh)
3164 raid5_conf_t *conf = sh->raid_conf;
3165 int disks = sh->disks;
3166 struct bio *return_bi = NULL;
3167 int i, pd_idx = sh->pd_idx, qd_idx = sh->qd_idx;
3168 struct stripe_head_state s;
3169 struct r6_state r6s;
3170 struct r5dev *dev, *pdev, *qdev;
3171 mdk_rdev_t *blocked_rdev = NULL;
3173 pr_debug("handling stripe %llu, state=%#lx cnt=%d, "
3174 "pd_idx=%d, qd_idx=%d\n, check:%d, reconstruct:%d\n",
3175 (unsigned long long)sh->sector, sh->state,
3176 atomic_read(&sh->count), pd_idx, qd_idx,
3177 sh->check_state, sh->reconstruct_state);
3178 memset(&s, 0, sizeof(s));
3180 spin_lock(&sh->lock);
3181 clear_bit(STRIPE_HANDLE, &sh->state);
3182 clear_bit(STRIPE_DELAYED, &sh->state);
3184 s.syncing = test_bit(STRIPE_SYNCING, &sh->state);
3185 s.expanding = test_bit(STRIPE_EXPAND_SOURCE, &sh->state);
3186 s.expanded = test_bit(STRIPE_EXPAND_READY, &sh->state);
3187 /* Now to look around and see what can be done */
3190 for (i=disks; i--; ) {
3193 clear_bit(R5_Insync, &dev->flags);
3195 pr_debug("check %d: state 0x%lx read %p write %p written %p\n",
3196 i, dev->flags, dev->toread, dev->towrite, dev->written);
3197 /* maybe we can reply to a read
3199 * new wantfill requests are only permitted while
3200 * ops_complete_biofill is guaranteed to be inactive
3202 if (test_bit(R5_UPTODATE, &dev->flags) && dev->toread &&
3203 !test_bit(STRIPE_BIOFILL_RUN, &sh->state))
3204 set_bit(R5_Wantfill, &dev->flags);
3206 /* now count some things */
3207 if (test_bit(R5_LOCKED, &dev->flags)) s.locked++;
3208 if (test_bit(R5_UPTODATE, &dev->flags)) s.uptodate++;
3209 if (test_bit(R5_Wantcompute, &dev->flags))
3210 BUG_ON(++s.compute > 2);
3212 if (test_bit(R5_Wantfill, &dev->flags)) {
3214 } else if (dev->toread)
3218 if (!test_bit(R5_OVERWRITE, &dev->flags))
3223 rdev = rcu_dereference(conf->disks[i].rdev);
3224 if (blocked_rdev == NULL &&
3225 rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
3226 blocked_rdev = rdev;
3227 atomic_inc(&rdev->nr_pending);
3229 if (!rdev || !test_bit(In_sync, &rdev->flags)) {
3230 /* The ReadError flag will just be confusing now */
3231 clear_bit(R5_ReadError, &dev->flags);
3232 clear_bit(R5_ReWrite, &dev->flags);
3234 if (!rdev || !test_bit(In_sync, &rdev->flags)
3235 || test_bit(R5_ReadError, &dev->flags)) {
3237 r6s.failed_num[s.failed] = i;
3240 set_bit(R5_Insync, &dev->flags);
3244 if (unlikely(blocked_rdev)) {
3245 if (s.syncing || s.expanding || s.expanded ||
3246 s.to_write || s.written) {
3247 set_bit(STRIPE_HANDLE, &sh->state);
3250 /* There is nothing for the blocked_rdev to block */
3251 rdev_dec_pending(blocked_rdev, conf->mddev);
3252 blocked_rdev = NULL;
3255 if (s.to_fill && !test_bit(STRIPE_BIOFILL_RUN, &sh->state)) {
3256 set_bit(STRIPE_OP_BIOFILL, &s.ops_request);
3257 set_bit(STRIPE_BIOFILL_RUN, &sh->state);
3260 pr_debug("locked=%d uptodate=%d to_read=%d"
3261 " to_write=%d failed=%d failed_num=%d,%d\n",
3262 s.locked, s.uptodate, s.to_read, s.to_write, s.failed,
3263 r6s.failed_num[0], r6s.failed_num[1]);
3264 /* check if the array has lost >2 devices and, if so, some requests
3265 * might need to be failed
3267 if (s.failed > 2 && s.to_read+s.to_write+s.written)
3268 handle_failed_stripe(conf, sh, &s, disks, &return_bi);
3269 if (s.failed > 2 && s.syncing) {
3270 md_done_sync(conf->mddev, STRIPE_SECTORS,0);
3271 clear_bit(STRIPE_SYNCING, &sh->state);
3276 * might be able to return some write requests if the parity blocks
3277 * are safe, or on a failed drive
3279 pdev = &sh->dev[pd_idx];
3280 r6s.p_failed = (s.failed >= 1 && r6s.failed_num[0] == pd_idx)
3281 || (s.failed >= 2 && r6s.failed_num[1] == pd_idx);
3282 qdev = &sh->dev[qd_idx];
3283 r6s.q_failed = (s.failed >= 1 && r6s.failed_num[0] == qd_idx)
3284 || (s.failed >= 2 && r6s.failed_num[1] == qd_idx);
3287 ( r6s.p_failed || ((test_bit(R5_Insync, &pdev->flags)
3288 && !test_bit(R5_LOCKED, &pdev->flags)
3289 && test_bit(R5_UPTODATE, &pdev->flags)))) &&
3290 ( r6s.q_failed || ((test_bit(R5_Insync, &qdev->flags)
3291 && !test_bit(R5_LOCKED, &qdev->flags)
3292 && test_bit(R5_UPTODATE, &qdev->flags)))))
3293 handle_stripe_clean_event(conf, sh, disks, &return_bi);
3295 /* Now we might consider reading some blocks, either to check/generate
3296 * parity, or to satisfy requests
3297 * or to load a block that is being partially written.
3299 if (s.to_read || s.non_overwrite || (s.to_write && s.failed) ||
3300 (s.syncing && (s.uptodate + s.compute < disks)) || s.expanding)
3301 handle_stripe_fill6(sh, &s, &r6s, disks);
3303 /* Now we check to see if any write operations have recently
3306 if (sh->reconstruct_state == reconstruct_state_drain_result) {
3307 int qd_idx = sh->qd_idx;
3309 sh->reconstruct_state = reconstruct_state_idle;
3310 /* All the 'written' buffers and the parity blocks are ready to
3311 * be written back to disk
3313 BUG_ON(!test_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags));
3314 BUG_ON(!test_bit(R5_UPTODATE, &sh->dev[qd_idx].flags));
3315 for (i = disks; i--; ) {
3317 if (test_bit(R5_LOCKED, &dev->flags) &&
3318 (i == sh->pd_idx || i == qd_idx ||
3320 pr_debug("Writing block %d\n", i);
3321 BUG_ON(!test_bit(R5_UPTODATE, &dev->flags));
3322 set_bit(R5_Wantwrite, &dev->flags);
3323 if (!test_bit(R5_Insync, &dev->flags) ||
3324 ((i == sh->pd_idx || i == qd_idx) &&
3326 set_bit(STRIPE_INSYNC, &sh->state);
3329 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
3330 atomic_dec(&conf->preread_active_stripes);
3331 if (atomic_read(&conf->preread_active_stripes) <
3333 md_wakeup_thread(conf->mddev->thread);
3337 /* Now to consider new write requests and what else, if anything
3338 * should be read. We do not handle new writes when:
3339 * 1/ A 'write' operation (copy+gen_syndrome) is already in flight.
3340 * 2/ A 'check' operation is in flight, as it may clobber the parity
3343 if (s.to_write && !sh->reconstruct_state && !sh->check_state)
3344 handle_stripe_dirtying6(conf, sh, &s, &r6s, disks);
3346 /* maybe we need to check and possibly fix the parity for this stripe
3347 * Any reads will already have been scheduled, so we just see if enough
3348 * data is available. The parity check is held off while parity
3349 * dependent operations are in flight.
3351 if (sh->check_state ||
3352 (s.syncing && s.locked == 0 &&
3353 !test_bit(STRIPE_COMPUTE_RUN, &sh->state) &&
3354 !test_bit(STRIPE_INSYNC, &sh->state)))
3355 handle_parity_checks6(conf, sh, &s, &r6s, disks);
3357 if (s.syncing && s.locked == 0 && test_bit(STRIPE_INSYNC, &sh->state)) {
3358 md_done_sync(conf->mddev, STRIPE_SECTORS,1);
3359 clear_bit(STRIPE_SYNCING, &sh->state);
3362 /* If the failed drives are just a ReadError, then we might need
3363 * to progress the repair/check process
3365 if (s.failed <= 2 && !conf->mddev->ro)
3366 for (i = 0; i < s.failed; i++) {
3367 dev = &sh->dev[r6s.failed_num[i]];
3368 if (test_bit(R5_ReadError, &dev->flags)
3369 && !test_bit(R5_LOCKED, &dev->flags)
3370 && test_bit(R5_UPTODATE, &dev->flags)
3372 if (!test_bit(R5_ReWrite, &dev->flags)) {
3373 set_bit(R5_Wantwrite, &dev->flags);
3374 set_bit(R5_ReWrite, &dev->flags);
3375 set_bit(R5_LOCKED, &dev->flags);
3378 /* let's read it back */
3379 set_bit(R5_Wantread, &dev->flags);
3380 set_bit(R5_LOCKED, &dev->flags);
3386 /* Finish reconstruct operations initiated by the expansion process */
3387 if (sh->reconstruct_state == reconstruct_state_result) {
3388 sh->reconstruct_state = reconstruct_state_idle;
3389 clear_bit(STRIPE_EXPANDING, &sh->state);
3390 for (i = conf->raid_disks; i--; ) {
3391 set_bit(R5_Wantwrite, &sh->dev[i].flags);
3392 set_bit(R5_LOCKED, &sh->dev[i].flags);
3397 if (s.expanded && test_bit(STRIPE_EXPANDING, &sh->state) &&
3398 !sh->reconstruct_state) {
3399 struct stripe_head *sh2
3400 = get_active_stripe(conf, sh->sector, 1, 1, 1);
3401 if (sh2 && test_bit(STRIPE_EXPAND_SOURCE, &sh2->state)) {
3402 /* sh cannot be written until sh2 has been read.
3403 * so arrange for sh to be delayed a little
3405 set_bit(STRIPE_DELAYED, &sh->state);
3406 set_bit(STRIPE_HANDLE, &sh->state);
3407 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE,
3409 atomic_inc(&conf->preread_active_stripes);
3410 release_stripe(sh2);
3414 release_stripe(sh2);
3416 /* Need to write out all blocks after computing P&Q */
3417 sh->disks = conf->raid_disks;
3418 stripe_set_idx(sh->sector, conf, 0, sh);
3419 schedule_reconstruction(sh, &s, 1, 1);
3420 } else if (s.expanded && !sh->reconstruct_state && s.locked == 0) {
3421 clear_bit(STRIPE_EXPAND_READY, &sh->state);
3422 atomic_dec(&conf->reshape_stripes);
3423 wake_up(&conf->wait_for_overlap);
3424 md_done_sync(conf->mddev, STRIPE_SECTORS, 1);
3427 if (s.expanding && s.locked == 0 &&
3428 !test_bit(STRIPE_COMPUTE_RUN, &sh->state))
3429 handle_stripe_expansion(conf, sh, &r6s);
3432 spin_unlock(&sh->lock);
3434 /* wait for this device to become unblocked */
3435 if (unlikely(blocked_rdev))
3436 md_wait_for_blocked_rdev(blocked_rdev, conf->mddev);
3439 raid_run_ops(sh, s.ops_request);
3443 return_io(return_bi);
3445 return blocked_rdev == NULL;
3448 /* returns true if the stripe was handled */
3449 static bool handle_stripe(struct stripe_head *sh)
3451 if (sh->raid_conf->level == 6)
3452 return handle_stripe6(sh);
3454 return handle_stripe5(sh);
3457 static void raid5_activate_delayed(raid5_conf_t *conf)
3459 if (atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD) {
3460 while (!list_empty(&conf->delayed_list)) {
3461 struct list_head *l = conf->delayed_list.next;
3462 struct stripe_head *sh;
3463 sh = list_entry(l, struct stripe_head, lru);
3465 clear_bit(STRIPE_DELAYED, &sh->state);
3466 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
3467 atomic_inc(&conf->preread_active_stripes);
3468 list_add_tail(&sh->lru, &conf->hold_list);
3471 blk_plug_device(conf->mddev->queue);
3474 static void activate_bit_delay(raid5_conf_t *conf)
3476 /* device_lock is held */
3477 struct list_head head;
3478 list_add(&head, &conf->bitmap_list);
3479 list_del_init(&conf->bitmap_list);
3480 while (!list_empty(&head)) {
3481 struct stripe_head *sh = list_entry(head.next, struct stripe_head, lru);
3482 list_del_init(&sh->lru);
3483 atomic_inc(&sh->count);
3484 __release_stripe(conf, sh);
3488 static void unplug_slaves(mddev_t *mddev)
3490 raid5_conf_t *conf = mddev_to_conf(mddev);
3494 for (i = 0; i < conf->raid_disks; i++) {
3495 mdk_rdev_t *rdev = rcu_dereference(conf->disks[i].rdev);
3496 if (rdev && !test_bit(Faulty, &rdev->flags) && atomic_read(&rdev->nr_pending)) {
3497 struct request_queue *r_queue = bdev_get_queue(rdev->bdev);
3499 atomic_inc(&rdev->nr_pending);
3502 blk_unplug(r_queue);
3504 rdev_dec_pending(rdev, mddev);
3511 static void raid5_unplug_device(struct request_queue *q)
3513 mddev_t *mddev = q->queuedata;
3514 raid5_conf_t *conf = mddev_to_conf(mddev);
3515 unsigned long flags;
3517 spin_lock_irqsave(&conf->device_lock, flags);
3519 if (blk_remove_plug(q)) {
3521 raid5_activate_delayed(conf);
3523 md_wakeup_thread(mddev->thread);
3525 spin_unlock_irqrestore(&conf->device_lock, flags);
3527 unplug_slaves(mddev);
3530 static int raid5_congested(void *data, int bits)
3532 mddev_t *mddev = data;
3533 raid5_conf_t *conf = mddev_to_conf(mddev);
3535 /* No difference between reads and writes. Just check
3536 * how busy the stripe_cache is
3538 if (conf->inactive_blocked)
3542 if (list_empty_careful(&conf->inactive_list))
3548 /* We want read requests to align with chunks where possible,
3549 * but write requests don't need to.
3551 static int raid5_mergeable_bvec(struct request_queue *q,
3552 struct bvec_merge_data *bvm,
3553 struct bio_vec *biovec)
3555 mddev_t *mddev = q->queuedata;
3556 sector_t sector = bvm->bi_sector + get_start_sect(bvm->bi_bdev);
3558 unsigned int chunk_sectors = mddev->chunk_size >> 9;
3559 unsigned int bio_sectors = bvm->bi_size >> 9;
3561 if ((bvm->bi_rw & 1) == WRITE)
3562 return biovec->bv_len; /* always allow writes to be mergeable */
3564 if (mddev->new_chunk < mddev->chunk_size)
3565 chunk_sectors = mddev->new_chunk >> 9;
3566 max = (chunk_sectors - ((sector & (chunk_sectors - 1)) + bio_sectors)) << 9;
3567 if (max < 0) max = 0;
3568 if (max <= biovec->bv_len && bio_sectors == 0)
3569 return biovec->bv_len;
3575 static int in_chunk_boundary(mddev_t *mddev, struct bio *bio)
3577 sector_t sector = bio->bi_sector + get_start_sect(bio->bi_bdev);
3578 unsigned int chunk_sectors = mddev->chunk_size >> 9;
3579 unsigned int bio_sectors = bio->bi_size >> 9;
3581 if (mddev->new_chunk < mddev->chunk_size)
3582 chunk_sectors = mddev->new_chunk >> 9;
3583 return chunk_sectors >=
3584 ((sector & (chunk_sectors - 1)) + bio_sectors);
3588 * add bio to the retry LIFO ( in O(1) ... we are in interrupt )
3589 * later sampled by raid5d.
3591 static void add_bio_to_retry(struct bio *bi,raid5_conf_t *conf)
3593 unsigned long flags;
3595 spin_lock_irqsave(&conf->device_lock, flags);
3597 bi->bi_next = conf->retry_read_aligned_list;
3598 conf->retry_read_aligned_list = bi;
3600 spin_unlock_irqrestore(&conf->device_lock, flags);
3601 md_wakeup_thread(conf->mddev->thread);
3605 static struct bio *remove_bio_from_retry(raid5_conf_t *conf)
3609 bi = conf->retry_read_aligned;
3611 conf->retry_read_aligned = NULL;
3614 bi = conf->retry_read_aligned_list;
3616 conf->retry_read_aligned_list = bi->bi_next;
3619 * this sets the active strip count to 1 and the processed
3620 * strip count to zero (upper 8 bits)
3622 bi->bi_phys_segments = 1; /* biased count of active stripes */
3630 * The "raid5_align_endio" should check if the read succeeded and if it
3631 * did, call bio_endio on the original bio (having bio_put the new bio
3633 * If the read failed..
3635 static void raid5_align_endio(struct bio *bi, int error)
3637 struct bio* raid_bi = bi->bi_private;
3640 int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
3645 mddev = raid_bi->bi_bdev->bd_disk->queue->queuedata;
3646 conf = mddev_to_conf(mddev);
3647 rdev = (void*)raid_bi->bi_next;
3648 raid_bi->bi_next = NULL;
3650 rdev_dec_pending(rdev, conf->mddev);
3652 if (!error && uptodate) {
3653 bio_endio(raid_bi, 0);
3654 if (atomic_dec_and_test(&conf->active_aligned_reads))
3655 wake_up(&conf->wait_for_stripe);
3660 pr_debug("raid5_align_endio : io error...handing IO for a retry\n");
3662 add_bio_to_retry(raid_bi, conf);
3665 static int bio_fits_rdev(struct bio *bi)
3667 struct request_queue *q = bdev_get_queue(bi->bi_bdev);
3669 if ((bi->bi_size>>9) > q->max_sectors)
3671 blk_recount_segments(q, bi);
3672 if (bi->bi_phys_segments > q->max_phys_segments)
3675 if (q->merge_bvec_fn)
3676 /* it's too hard to apply the merge_bvec_fn at this stage,
3685 static int chunk_aligned_read(struct request_queue *q, struct bio * raid_bio)
3687 mddev_t *mddev = q->queuedata;
3688 raid5_conf_t *conf = mddev_to_conf(mddev);
3689 unsigned int dd_idx;
3690 struct bio* align_bi;
3693 if (!in_chunk_boundary(mddev, raid_bio)) {
3694 pr_debug("chunk_aligned_read : non aligned\n");
3698 * use bio_clone to make a copy of the bio
3700 align_bi = bio_clone(raid_bio, GFP_NOIO);
3704 * set bi_end_io to a new function, and set bi_private to the
3707 align_bi->bi_end_io = raid5_align_endio;
3708 align_bi->bi_private = raid_bio;
3712 align_bi->bi_sector = raid5_compute_sector(conf, raid_bio->bi_sector,
3717 rdev = rcu_dereference(conf->disks[dd_idx].rdev);
3718 if (rdev && test_bit(In_sync, &rdev->flags)) {
3719 atomic_inc(&rdev->nr_pending);
3721 raid_bio->bi_next = (void*)rdev;
3722 align_bi->bi_bdev = rdev->bdev;
3723 align_bi->bi_flags &= ~(1 << BIO_SEG_VALID);
3724 align_bi->bi_sector += rdev->data_offset;
3726 if (!bio_fits_rdev(align_bi)) {
3727 /* too big in some way */
3729 rdev_dec_pending(rdev, mddev);
3733 spin_lock_irq(&conf->device_lock);
3734 wait_event_lock_irq(conf->wait_for_stripe,
3736 conf->device_lock, /* nothing */);
3737 atomic_inc(&conf->active_aligned_reads);
3738 spin_unlock_irq(&conf->device_lock);
3740 generic_make_request(align_bi);
3749 /* __get_priority_stripe - get the next stripe to process
3751 * Full stripe writes are allowed to pass preread active stripes up until
3752 * the bypass_threshold is exceeded. In general the bypass_count
3753 * increments when the handle_list is handled before the hold_list; however, it
3754 * will not be incremented when STRIPE_IO_STARTED is sampled set signifying a
3755 * stripe with in flight i/o. The bypass_count will be reset when the
3756 * head of the hold_list has changed, i.e. the head was promoted to the
3759 static struct stripe_head *__get_priority_stripe(raid5_conf_t *conf)
3761 struct stripe_head *sh;
3763 pr_debug("%s: handle: %s hold: %s full_writes: %d bypass_count: %d\n",
3765 list_empty(&conf->handle_list) ? "empty" : "busy",
3766 list_empty(&conf->hold_list) ? "empty" : "busy",
3767 atomic_read(&conf->pending_full_writes), conf->bypass_count);
3769 if (!list_empty(&conf->handle_list)) {
3770 sh = list_entry(conf->handle_list.next, typeof(*sh), lru);
3772 if (list_empty(&conf->hold_list))
3773 conf->bypass_count = 0;
3774 else if (!test_bit(STRIPE_IO_STARTED, &sh->state)) {
3775 if (conf->hold_list.next == conf->last_hold)
3776 conf->bypass_count++;
3778 conf->last_hold = conf->hold_list.next;
3779 conf->bypass_count -= conf->bypass_threshold;
3780 if (conf->bypass_count < 0)
3781 conf->bypass_count = 0;
3784 } else if (!list_empty(&conf->hold_list) &&
3785 ((conf->bypass_threshold &&
3786 conf->bypass_count > conf->bypass_threshold) ||
3787 atomic_read(&conf->pending_full_writes) == 0)) {
3788 sh = list_entry(conf->hold_list.next,
3790 conf->bypass_count -= conf->bypass_threshold;
3791 if (conf->bypass_count < 0)
3792 conf->bypass_count = 0;
3796 list_del_init(&sh->lru);
3797 atomic_inc(&sh->count);
3798 BUG_ON(atomic_read(&sh->count) != 1);
3802 static int make_request(struct request_queue *q, struct bio * bi)
3804 mddev_t *mddev = q->queuedata;
3805 raid5_conf_t *conf = mddev_to_conf(mddev);
3807 sector_t new_sector;
3808 sector_t logical_sector, last_sector;
3809 struct stripe_head *sh;
3810 const int rw = bio_data_dir(bi);
3813 if (unlikely(bio_barrier(bi))) {
3814 bio_endio(bi, -EOPNOTSUPP);
3818 md_write_start(mddev, bi);
3820 cpu = part_stat_lock();
3821 part_stat_inc(cpu, &mddev->gendisk->part0, ios[rw]);
3822 part_stat_add(cpu, &mddev->gendisk->part0, sectors[rw],
3827 mddev->reshape_position == MaxSector &&
3828 chunk_aligned_read(q,bi))
3831 logical_sector = bi->bi_sector & ~((sector_t)STRIPE_SECTORS-1);
3832 last_sector = bi->bi_sector + (bi->bi_size>>9);
3834 bi->bi_phys_segments = 1; /* over-loaded to count active stripes */
3836 for (;logical_sector < last_sector; logical_sector += STRIPE_SECTORS) {
3838 int disks, data_disks;
3843 disks = conf->raid_disks;
3844 prepare_to_wait(&conf->wait_for_overlap, &w, TASK_UNINTERRUPTIBLE);
3845 if (unlikely(conf->reshape_progress != MaxSector)) {
3846 /* spinlock is needed as reshape_progress may be
3847 * 64bit on a 32bit platform, and so it might be
3848 * possible to see a half-updated value
3849 * Ofcourse reshape_progress could change after
3850 * the lock is dropped, so once we get a reference
3851 * to the stripe that we think it is, we will have
3854 spin_lock_irq(&conf->device_lock);
3855 if (mddev->delta_disks < 0
3856 ? logical_sector < conf->reshape_progress
3857 : logical_sector >= conf->reshape_progress) {
3858 disks = conf->previous_raid_disks;
3861 if (mddev->delta_disks < 0
3862 ? logical_sector < conf->reshape_safe
3863 : logical_sector >= conf->reshape_safe) {
3864 spin_unlock_irq(&conf->device_lock);
3869 spin_unlock_irq(&conf->device_lock);
3871 data_disks = disks - conf->max_degraded;
3873 new_sector = raid5_compute_sector(conf, logical_sector,
3876 pr_debug("raid5: make_request, sector %llu logical %llu\n",
3877 (unsigned long long)new_sector,
3878 (unsigned long long)logical_sector);
3880 sh = get_active_stripe(conf, new_sector, previous,
3881 (bi->bi_rw&RWA_MASK), 0);
3883 if (unlikely(previous)) {
3884 /* expansion might have moved on while waiting for a
3885 * stripe, so we must do the range check again.
3886 * Expansion could still move past after this
3887 * test, but as we are holding a reference to
3888 * 'sh', we know that if that happens,
3889 * STRIPE_EXPANDING will get set and the expansion
3890 * won't proceed until we finish with the stripe.
3893 spin_lock_irq(&conf->device_lock);
3894 if (mddev->delta_disks < 0
3895 ? logical_sector >= conf->reshape_progress
3896 : logical_sector < conf->reshape_progress)
3897 /* mismatch, need to try again */
3899 spin_unlock_irq(&conf->device_lock);
3905 /* FIXME what if we get a false positive because these
3906 * are being updated.
3908 if (logical_sector >= mddev->suspend_lo &&
3909 logical_sector < mddev->suspend_hi) {
3915 if (test_bit(STRIPE_EXPANDING, &sh->state) ||
3916 !add_stripe_bio(sh, bi, dd_idx, (bi->bi_rw&RW_MASK))) {
3917 /* Stripe is busy expanding or
3918 * add failed due to overlap. Flush everything
3921 raid5_unplug_device(mddev->queue);
3926 finish_wait(&conf->wait_for_overlap, &w);
3927 set_bit(STRIPE_HANDLE, &sh->state);
3928 clear_bit(STRIPE_DELAYED, &sh->state);
3931 /* cannot get stripe for read-ahead, just give-up */
3932 clear_bit(BIO_UPTODATE, &bi->bi_flags);
3933 finish_wait(&conf->wait_for_overlap, &w);
3938 spin_lock_irq(&conf->device_lock);
3939 remaining = raid5_dec_bi_phys_segments(bi);
3940 spin_unlock_irq(&conf->device_lock);
3941 if (remaining == 0) {
3944 md_write_end(mddev);
3951 static sector_t raid5_size(mddev_t *mddev, sector_t sectors, int raid_disks);
3953 static sector_t reshape_request(mddev_t *mddev, sector_t sector_nr, int *skipped)
3955 /* reshaping is quite different to recovery/resync so it is
3956 * handled quite separately ... here.
3958 * On each call to sync_request, we gather one chunk worth of
3959 * destination stripes and flag them as expanding.
3960 * Then we find all the source stripes and request reads.
3961 * As the reads complete, handle_stripe will copy the data
3962 * into the destination stripe and release that stripe.
3964 raid5_conf_t *conf = (raid5_conf_t *) mddev->private;
3965 struct stripe_head *sh;
3966 sector_t first_sector, last_sector;
3967 int raid_disks = conf->previous_raid_disks;
3968 int data_disks = raid_disks - conf->max_degraded;
3969 int new_data_disks = conf->raid_disks - conf->max_degraded;
3972 sector_t writepos, readpos, safepos;
3973 sector_t stripe_addr;
3974 int reshape_sectors;
3975 struct list_head stripes;
3977 if (sector_nr == 0) {
3978 /* If restarting in the middle, skip the initial sectors */
3979 if (mddev->delta_disks < 0 &&
3980 conf->reshape_progress < raid5_size(mddev, 0, 0)) {
3981 sector_nr = raid5_size(mddev, 0, 0)
3982 - conf->reshape_progress;
3983 } else if (mddev->delta_disks > 0 &&
3984 conf->reshape_progress > 0)
3985 sector_nr = conf->reshape_progress;
3986 sector_div(sector_nr, new_data_disks);
3993 /* We need to process a full chunk at a time.
3994 * If old and new chunk sizes differ, we need to process the
3997 if (mddev->new_chunk > mddev->chunk_size)
3998 reshape_sectors = mddev->new_chunk / 512;
4000 reshape_sectors = mddev->chunk_size / 512;
4002 /* we update the metadata when there is more than 3Meg
4003 * in the block range (that is rather arbitrary, should
4004 * probably be time based) or when the data about to be
4005 * copied would over-write the source of the data at
4006 * the front of the range.
4007 * i.e. one new_stripe along from reshape_progress new_maps
4008 * to after where reshape_safe old_maps to
4010 writepos = conf->reshape_progress;
4011 sector_div(writepos, new_data_disks);
4012 readpos = conf->reshape_progress;
4013 sector_div(readpos, data_disks);
4014 safepos = conf->reshape_safe;
4015 sector_div(safepos, data_disks);
4016 if (mddev->delta_disks < 0) {
4017 writepos -= min_t(sector_t, reshape_sectors, writepos);
4018 readpos += reshape_sectors;
4019 safepos += reshape_sectors;
4021 writepos += reshape_sectors;
4022 readpos -= min_t(sector_t, reshape_sectors, readpos);
4023 safepos -= min_t(sector_t, reshape_sectors, safepos);
4026 /* 'writepos' is the most advanced device address we might write.
4027 * 'readpos' is the least advanced device address we might read.
4028 * 'safepos' is the least address recorded in the metadata as having
4030 * If 'readpos' is behind 'writepos', then there is no way that we can
4031 * ensure safety in the face of a crash - that must be done by userspace
4032 * making a backup of the data. So in that case there is no particular
4033 * rush to update metadata.
4034 * Otherwise if 'safepos' is behind 'writepos', then we really need to
4035 * update the metadata to advance 'safepos' to match 'readpos' so that
4036 * we can be safe in the event of a crash.
4037 * So we insist on updating metadata if safepos is behind writepos and
4038 * readpos is beyond writepos.
4039 * In any case, update the metadata every 10 seconds.
4040 * Maybe that number should be configurable, but I'm not sure it is
4041 * worth it.... maybe it could be a multiple of safemode_delay???
4043 if ((mddev->delta_disks < 0
4044 ? (safepos > writepos && readpos < writepos)
4045 : (safepos < writepos && readpos > writepos)) ||
4046 time_after(jiffies, conf->reshape_checkpoint + 10*HZ)) {
4047 /* Cannot proceed until we've updated the superblock... */
4048 wait_event(conf->wait_for_overlap,
4049 atomic_read(&conf->reshape_stripes)==0);
4050 mddev->reshape_position = conf->reshape_progress;
4051 mddev->curr_resync_completed = mddev->curr_resync;
4052 conf->reshape_checkpoint = jiffies;
4053 set_bit(MD_CHANGE_DEVS, &mddev->flags);
4054 md_wakeup_thread(mddev->thread);
4055 wait_event(mddev->sb_wait, mddev->flags == 0 ||
4056 kthread_should_stop());
4057 spin_lock_irq(&conf->device_lock);
4058 conf->reshape_safe = mddev->reshape_position;
4059 spin_unlock_irq(&conf->device_lock);
4060 wake_up(&conf->wait_for_overlap);
4061 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
4064 if (mddev->delta_disks < 0) {
4065 BUG_ON(conf->reshape_progress == 0);
4066 stripe_addr = writepos;
4067 BUG_ON((mddev->dev_sectors &
4068 ~((sector_t)reshape_sectors - 1))
4069 - reshape_sectors - stripe_addr
4072 BUG_ON(writepos != sector_nr + reshape_sectors);
4073 stripe_addr = sector_nr;
4075 INIT_LIST_HEAD(&stripes);
4076 for (i = 0; i < reshape_sectors; i += STRIPE_SECTORS) {
4079 sh = get_active_stripe(conf, stripe_addr+i, 0, 0, 1);
4080 set_bit(STRIPE_EXPANDING, &sh->state);
4081 atomic_inc(&conf->reshape_stripes);
4082 /* If any of this stripe is beyond the end of the old
4083 * array, then we need to zero those blocks
4085 for (j=sh->disks; j--;) {
4087 if (j == sh->pd_idx)
4089 if (conf->level == 6 &&
4092 s = compute_blocknr(sh, j, 0);
4093 if (s < raid5_size(mddev, 0, 0)) {
4097 memset(page_address(sh->dev[j].page), 0, STRIPE_SIZE);
4098 set_bit(R5_Expanded, &sh->dev[j].flags);
4099 set_bit(R5_UPTODATE, &sh->dev[j].flags);
4102 set_bit(STRIPE_EXPAND_READY, &sh->state);
4103 set_bit(STRIPE_HANDLE, &sh->state);
4105 list_add(&sh->lru, &stripes);
4107 spin_lock_irq(&conf->device_lock);
4108 if (mddev->delta_disks < 0)
4109 conf->reshape_progress -= reshape_sectors * new_data_disks;
4111 conf->reshape_progress += reshape_sectors * new_data_disks;
4112 spin_unlock_irq(&conf->device_lock);
4113 /* Ok, those stripe are ready. We can start scheduling
4114 * reads on the source stripes.
4115 * The source stripes are determined by mapping the first and last
4116 * block on the destination stripes.
4119 raid5_compute_sector(conf, stripe_addr*(new_data_disks),
4122 raid5_compute_sector(conf, ((stripe_addr+reshape_sectors)
4123 *(new_data_disks) - 1),
4125 if (last_sector >= mddev->dev_sectors)
4126 last_sector = mddev->dev_sectors - 1;
4127 while (first_sector <= last_sector) {
4128 sh = get_active_stripe(conf, first_sector, 1, 0, 1);
4129 set_bit(STRIPE_EXPAND_SOURCE, &sh->state);
4130 set_bit(STRIPE_HANDLE, &sh->state);
4132 first_sector += STRIPE_SECTORS;
4134 /* Now that the sources are clearly marked, we can release
4135 * the destination stripes
4137 while (!list_empty(&stripes)) {
4138 sh = list_entry(stripes.next, struct stripe_head, lru);
4139 list_del_init(&sh->lru);
4142 /* If this takes us to the resync_max point where we have to pause,
4143 * then we need to write out the superblock.
4145 sector_nr += reshape_sectors;
4146 if ((sector_nr - mddev->curr_resync_completed) * 2
4147 >= mddev->resync_max - mddev->curr_resync_completed) {
4148 /* Cannot proceed until we've updated the superblock... */
4149 wait_event(conf->wait_for_overlap,
4150 atomic_read(&conf->reshape_stripes) == 0);
4151 mddev->reshape_position = conf->reshape_progress;
4152 mddev->curr_resync_completed = mddev->curr_resync;
4153 conf->reshape_checkpoint = jiffies;
4154 set_bit(MD_CHANGE_DEVS, &mddev->flags);
4155 md_wakeup_thread(mddev->thread);
4156 wait_event(mddev->sb_wait,
4157 !test_bit(MD_CHANGE_DEVS, &mddev->flags)
4158 || kthread_should_stop());
4159 spin_lock_irq(&conf->device_lock);
4160 conf->reshape_safe = mddev->reshape_position;
4161 spin_unlock_irq(&conf->device_lock);
4162 wake_up(&conf->wait_for_overlap);
4163 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
4165 return reshape_sectors;
4168 /* FIXME go_faster isn't used */
4169 static inline sector_t sync_request(mddev_t *mddev, sector_t sector_nr, int *skipped, int go_faster)
4171 raid5_conf_t *conf = (raid5_conf_t *) mddev->private;
4172 struct stripe_head *sh;
4173 sector_t max_sector = mddev->dev_sectors;
4175 int still_degraded = 0;
4178 if (sector_nr >= max_sector) {
4179 /* just being told to finish up .. nothing much to do */
4180 unplug_slaves(mddev);
4182 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery)) {
4187 if (mddev->curr_resync < max_sector) /* aborted */
4188 bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
4190 else /* completed sync */
4192 bitmap_close_sync(mddev->bitmap);
4197 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
4198 return reshape_request(mddev, sector_nr, skipped);
4200 /* No need to check resync_max as we never do more than one
4201 * stripe, and as resync_max will always be on a chunk boundary,
4202 * if the check in md_do_sync didn't fire, there is no chance
4203 * of overstepping resync_max here
4206 /* if there is too many failed drives and we are trying
4207 * to resync, then assert that we are finished, because there is
4208 * nothing we can do.
4210 if (mddev->degraded >= conf->max_degraded &&
4211 test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
4212 sector_t rv = mddev->dev_sectors - sector_nr;
4216 if (!bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
4217 !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
4218 !conf->fullsync && sync_blocks >= STRIPE_SECTORS) {
4219 /* we can skip this block, and probably more */
4220 sync_blocks /= STRIPE_SECTORS;
4222 return sync_blocks * STRIPE_SECTORS; /* keep things rounded to whole stripes */
4226 bitmap_cond_end_sync(mddev->bitmap, sector_nr);
4228 sh = get_active_stripe(conf, sector_nr, 0, 1, 0);
4230 sh = get_active_stripe(conf, sector_nr, 0, 0, 0);
4231 /* make sure we don't swamp the stripe cache if someone else
4232 * is trying to get access
4234 schedule_timeout_uninterruptible(1);
4236 /* Need to check if array will still be degraded after recovery/resync
4237 * We don't need to check the 'failed' flag as when that gets set,
4240 for (i = 0; i < conf->raid_disks; i++)
4241 if (conf->disks[i].rdev == NULL)
4244 bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, still_degraded);
4246 spin_lock(&sh->lock);
4247 set_bit(STRIPE_SYNCING, &sh->state);
4248 clear_bit(STRIPE_INSYNC, &sh->state);
4249 spin_unlock(&sh->lock);
4251 /* wait for any blocked device to be handled */
4252 while (unlikely(!handle_stripe(sh)))
4256 return STRIPE_SECTORS;
4259 static int retry_aligned_read(raid5_conf_t *conf, struct bio *raid_bio)
4261 /* We may not be able to submit a whole bio at once as there
4262 * may not be enough stripe_heads available.
4263 * We cannot pre-allocate enough stripe_heads as we may need
4264 * more than exist in the cache (if we allow ever large chunks).
4265 * So we do one stripe head at a time and record in
4266 * ->bi_hw_segments how many have been done.
4268 * We *know* that this entire raid_bio is in one chunk, so
4269 * it will be only one 'dd_idx' and only need one call to raid5_compute_sector.
4271 struct stripe_head *sh;
4273 sector_t sector, logical_sector, last_sector;
4278 logical_sector = raid_bio->bi_sector & ~((sector_t)STRIPE_SECTORS-1);
4279 sector = raid5_compute_sector(conf, logical_sector,
4281 last_sector = raid_bio->bi_sector + (raid_bio->bi_size>>9);
4283 for (; logical_sector < last_sector;
4284 logical_sector += STRIPE_SECTORS,
4285 sector += STRIPE_SECTORS,
4288 if (scnt < raid5_bi_hw_segments(raid_bio))
4289 /* already done this stripe */
4292 sh = get_active_stripe(conf, sector, 0, 1, 0);
4295 /* failed to get a stripe - must wait */
4296 raid5_set_bi_hw_segments(raid_bio, scnt);
4297 conf->retry_read_aligned = raid_bio;
4301 set_bit(R5_ReadError, &sh->dev[dd_idx].flags);
4302 if (!add_stripe_bio(sh, raid_bio, dd_idx, 0)) {
4304 raid5_set_bi_hw_segments(raid_bio, scnt);
4305 conf->retry_read_aligned = raid_bio;
4313 spin_lock_irq(&conf->device_lock);
4314 remaining = raid5_dec_bi_phys_segments(raid_bio);
4315 spin_unlock_irq(&conf->device_lock);
4317 bio_endio(raid_bio, 0);
4318 if (atomic_dec_and_test(&conf->active_aligned_reads))
4319 wake_up(&conf->wait_for_stripe);
4323 #ifdef CONFIG_MULTICORE_RAID456
4324 static void __process_stripe(void *param, async_cookie_t cookie)
4326 struct stripe_head *sh = param;
4332 static void process_stripe(struct stripe_head *sh, struct list_head *domain)
4334 async_schedule_domain(__process_stripe, sh, domain);
4337 static void synchronize_stripe_processing(struct list_head *domain)
4339 async_synchronize_full_domain(domain);
4342 static void process_stripe(struct stripe_head *sh, struct list_head *domain)
4349 static void synchronize_stripe_processing(struct list_head *domain)
4356 * This is our raid5 kernel thread.
4358 * We scan the hash table for stripes which can be handled now.
4359 * During the scan, completed stripes are saved for us by the interrupt
4360 * handler, so that they will not have to wait for our next wakeup.
4362 static void raid5d(mddev_t *mddev)
4364 struct stripe_head *sh;
4365 raid5_conf_t *conf = mddev_to_conf(mddev);
4367 LIST_HEAD(raid_domain);
4369 pr_debug("+++ raid5d active\n");
4371 md_check_recovery(mddev);
4374 spin_lock_irq(&conf->device_lock);
4378 if (conf->seq_flush != conf->seq_write) {
4379 int seq = conf->seq_flush;
4380 spin_unlock_irq(&conf->device_lock);
4381 bitmap_unplug(mddev->bitmap);
4382 spin_lock_irq(&conf->device_lock);
4383 conf->seq_write = seq;
4384 activate_bit_delay(conf);
4387 while ((bio = remove_bio_from_retry(conf))) {
4389 spin_unlock_irq(&conf->device_lock);
4390 ok = retry_aligned_read(conf, bio);
4391 spin_lock_irq(&conf->device_lock);
4397 sh = __get_priority_stripe(conf);
4401 spin_unlock_irq(&conf->device_lock);
4404 process_stripe(sh, &raid_domain);
4406 spin_lock_irq(&conf->device_lock);
4408 pr_debug("%d stripes handled\n", handled);
4410 spin_unlock_irq(&conf->device_lock);
4412 synchronize_stripe_processing(&raid_domain);
4413 async_tx_issue_pending_all();
4414 unplug_slaves(mddev);
4416 pr_debug("--- raid5d inactive\n");
4420 raid5_show_stripe_cache_size(mddev_t *mddev, char *page)
4422 raid5_conf_t *conf = mddev_to_conf(mddev);
4424 return sprintf(page, "%d\n", conf->max_nr_stripes);
4430 raid5_store_stripe_cache_size(mddev_t *mddev, const char *page, size_t len)
4432 raid5_conf_t *conf = mddev_to_conf(mddev);
4436 if (len >= PAGE_SIZE)
4441 if (strict_strtoul(page, 10, &new))
4443 if (new <= 16 || new > 32768)
4445 while (new < conf->max_nr_stripes) {
4446 if (drop_one_stripe(conf))
4447 conf->max_nr_stripes--;
4451 err = md_allow_write(mddev);
4454 while (new > conf->max_nr_stripes) {
4455 if (grow_one_stripe(conf))
4456 conf->max_nr_stripes++;
4462 static struct md_sysfs_entry
4463 raid5_stripecache_size = __ATTR(stripe_cache_size, S_IRUGO | S_IWUSR,
4464 raid5_show_stripe_cache_size,
4465 raid5_store_stripe_cache_size);
4468 raid5_show_preread_threshold(mddev_t *mddev, char *page)
4470 raid5_conf_t *conf = mddev_to_conf(mddev);
4472 return sprintf(page, "%d\n", conf->bypass_threshold);
4478 raid5_store_preread_threshold(mddev_t *mddev, const char *page, size_t len)
4480 raid5_conf_t *conf = mddev_to_conf(mddev);
4482 if (len >= PAGE_SIZE)
4487 if (strict_strtoul(page, 10, &new))
4489 if (new > conf->max_nr_stripes)
4491 conf->bypass_threshold = new;
4495 static struct md_sysfs_entry
4496 raid5_preread_bypass_threshold = __ATTR(preread_bypass_threshold,
4498 raid5_show_preread_threshold,
4499 raid5_store_preread_threshold);
4502 stripe_cache_active_show(mddev_t *mddev, char *page)
4504 raid5_conf_t *conf = mddev_to_conf(mddev);
4506 return sprintf(page, "%d\n", atomic_read(&conf->active_stripes));
4511 static struct md_sysfs_entry
4512 raid5_stripecache_active = __ATTR_RO(stripe_cache_active);
4514 static struct attribute *raid5_attrs[] = {
4515 &raid5_stripecache_size.attr,
4516 &raid5_stripecache_active.attr,
4517 &raid5_preread_bypass_threshold.attr,
4520 static struct attribute_group raid5_attrs_group = {
4522 .attrs = raid5_attrs,
4526 raid5_size(mddev_t *mddev, sector_t sectors, int raid_disks)
4528 raid5_conf_t *conf = mddev_to_conf(mddev);
4531 sectors = mddev->dev_sectors;
4533 /* size is defined by the smallest of previous and new size */
4534 if (conf->raid_disks < conf->previous_raid_disks)
4535 raid_disks = conf->raid_disks;
4537 raid_disks = conf->previous_raid_disks;
4540 sectors &= ~((sector_t)mddev->chunk_size/512 - 1);
4541 sectors &= ~((sector_t)mddev->new_chunk/512 - 1);
4542 return sectors * (raid_disks - conf->max_degraded);
4545 static void raid5_free_percpu(raid5_conf_t *conf)
4547 struct raid5_percpu *percpu;
4554 for_each_possible_cpu(cpu) {
4555 percpu = per_cpu_ptr(conf->percpu, cpu);
4556 safe_put_page(percpu->spare_page);
4557 kfree(percpu->scribble);
4559 #ifdef CONFIG_HOTPLUG_CPU
4560 unregister_cpu_notifier(&conf->cpu_notify);
4564 free_percpu(conf->percpu);
4567 static void free_conf(raid5_conf_t *conf)
4569 shrink_stripes(conf);
4570 raid5_free_percpu(conf);
4572 kfree(conf->stripe_hashtbl);
4576 #ifdef CONFIG_HOTPLUG_CPU
4577 static int raid456_cpu_notify(struct notifier_block *nfb, unsigned long action,
4580 raid5_conf_t *conf = container_of(nfb, raid5_conf_t, cpu_notify);
4581 long cpu = (long)hcpu;
4582 struct raid5_percpu *percpu = per_cpu_ptr(conf->percpu, cpu);
4585 case CPU_UP_PREPARE:
4586 case CPU_UP_PREPARE_FROZEN:
4587 if (conf->level == 6 && !percpu->spare_page)
4588 percpu->spare_page = alloc_page(GFP_KERNEL);
4589 if (!percpu->scribble)
4590 percpu->scribble = kmalloc(conf->scribble_len, GFP_KERNEL);
4592 if (!percpu->scribble ||
4593 (conf->level == 6 && !percpu->spare_page)) {
4594 safe_put_page(percpu->spare_page);
4595 kfree(percpu->scribble);
4596 pr_err("%s: failed memory allocation for cpu%ld\n",
4602 case CPU_DEAD_FROZEN:
4603 safe_put_page(percpu->spare_page);
4604 kfree(percpu->scribble);
4605 percpu->spare_page = NULL;
4606 percpu->scribble = NULL;
4615 static int raid5_alloc_percpu(raid5_conf_t *conf)
4618 struct page *spare_page;
4619 struct raid5_percpu *allcpus;
4623 allcpus = alloc_percpu(struct raid5_percpu);
4626 conf->percpu = allcpus;
4630 for_each_present_cpu(cpu) {
4631 if (conf->level == 6) {
4632 spare_page = alloc_page(GFP_KERNEL);
4637 per_cpu_ptr(conf->percpu, cpu)->spare_page = spare_page;
4639 scribble = kmalloc(scribble_len(conf->raid_disks), GFP_KERNEL);
4644 per_cpu_ptr(conf->percpu, cpu)->scribble = scribble;
4646 #ifdef CONFIG_HOTPLUG_CPU
4647 conf->cpu_notify.notifier_call = raid456_cpu_notify;
4648 conf->cpu_notify.priority = 0;
4650 err = register_cpu_notifier(&conf->cpu_notify);
4657 static raid5_conf_t *setup_conf(mddev_t *mddev)
4660 int raid_disk, memory;
4662 struct disk_info *disk;
4664 if (mddev->new_level != 5
4665 && mddev->new_level != 4
4666 && mddev->new_level != 6) {
4667 printk(KERN_ERR "raid5: %s: raid level not set to 4/5/6 (%d)\n",
4668 mdname(mddev), mddev->new_level);
4669 return ERR_PTR(-EIO);
4671 if ((mddev->new_level == 5
4672 && !algorithm_valid_raid5(mddev->new_layout)) ||
4673 (mddev->new_level == 6
4674 && !algorithm_valid_raid6(mddev->new_layout))) {
4675 printk(KERN_ERR "raid5: %s: layout %d not supported\n",
4676 mdname(mddev), mddev->new_layout);
4677 return ERR_PTR(-EIO);
4679 if (mddev->new_level == 6 && mddev->raid_disks < 4) {
4680 printk(KERN_ERR "raid6: not enough configured devices for %s (%d, minimum 4)\n",
4681 mdname(mddev), mddev->raid_disks);
4682 return ERR_PTR(-EINVAL);
4685 if (!mddev->new_chunk || mddev->new_chunk % PAGE_SIZE) {
4686 printk(KERN_ERR "raid5: invalid chunk size %d for %s\n",
4687 mddev->new_chunk, mdname(mddev));
4688 return ERR_PTR(-EINVAL);
4691 conf = kzalloc(sizeof(raid5_conf_t), GFP_KERNEL);
4695 conf->raid_disks = mddev->raid_disks;
4696 conf->scribble_len = scribble_len(conf->raid_disks);
4697 if (mddev->reshape_position == MaxSector)
4698 conf->previous_raid_disks = mddev->raid_disks;
4700 conf->previous_raid_disks = mddev->raid_disks - mddev->delta_disks;
4702 conf->disks = kzalloc(conf->raid_disks * sizeof(struct disk_info),
4707 conf->mddev = mddev;
4709 if ((conf->stripe_hashtbl = kzalloc(PAGE_SIZE, GFP_KERNEL)) == NULL)
4712 conf->level = mddev->new_level;
4713 if (raid5_alloc_percpu(conf) != 0)
4716 spin_lock_init(&conf->device_lock);
4717 init_waitqueue_head(&conf->wait_for_stripe);
4718 init_waitqueue_head(&conf->wait_for_overlap);
4719 INIT_LIST_HEAD(&conf->handle_list);
4720 INIT_LIST_HEAD(&conf->hold_list);
4721 INIT_LIST_HEAD(&conf->delayed_list);
4722 INIT_LIST_HEAD(&conf->bitmap_list);
4723 INIT_LIST_HEAD(&conf->inactive_list);
4724 atomic_set(&conf->active_stripes, 0);
4725 atomic_set(&conf->preread_active_stripes, 0);
4726 atomic_set(&conf->active_aligned_reads, 0);
4727 conf->bypass_threshold = BYPASS_THRESHOLD;
4729 pr_debug("raid5: run(%s) called.\n", mdname(mddev));
4731 list_for_each_entry(rdev, &mddev->disks, same_set) {
4732 raid_disk = rdev->raid_disk;
4733 if (raid_disk >= conf->raid_disks
4736 disk = conf->disks + raid_disk;
4740 if (test_bit(In_sync, &rdev->flags)) {
4741 char b[BDEVNAME_SIZE];
4742 printk(KERN_INFO "raid5: device %s operational as raid"
4743 " disk %d\n", bdevname(rdev->bdev,b),
4746 /* Cannot rely on bitmap to complete recovery */
4750 conf->chunk_size = mddev->new_chunk;
4751 if (conf->level == 6)
4752 conf->max_degraded = 2;
4754 conf->max_degraded = 1;
4755 conf->algorithm = mddev->new_layout;
4756 conf->max_nr_stripes = NR_STRIPES;
4757 conf->reshape_progress = mddev->reshape_position;
4758 if (conf->reshape_progress != MaxSector) {
4759 conf->prev_chunk = mddev->chunk_size;
4760 conf->prev_algo = mddev->layout;
4763 memory = conf->max_nr_stripes * (sizeof(struct stripe_head) +
4764 conf->raid_disks * ((sizeof(struct bio) + PAGE_SIZE))) / 1024;
4765 if (grow_stripes(conf, conf->max_nr_stripes)) {
4767 "raid5: couldn't allocate %dkB for buffers\n", memory);
4770 printk(KERN_INFO "raid5: allocated %dkB for %s\n",
4771 memory, mdname(mddev));
4773 conf->thread = md_register_thread(raid5d, mddev, "%s_raid5");
4774 if (!conf->thread) {
4776 "raid5: couldn't allocate thread for %s\n",
4786 return ERR_PTR(-EIO);
4788 return ERR_PTR(-ENOMEM);
4791 static int run(mddev_t *mddev)
4794 int working_disks = 0;
4797 if (mddev->reshape_position != MaxSector) {
4798 /* Check that we can continue the reshape.
4799 * Currently only disks can change, it must
4800 * increase, and we must be past the point where
4801 * a stripe over-writes itself
4803 sector_t here_new, here_old;
4805 int max_degraded = (mddev->level == 6 ? 2 : 1);
4807 if (mddev->new_level != mddev->level) {
4808 printk(KERN_ERR "raid5: %s: unsupported reshape "
4809 "required - aborting.\n",
4813 old_disks = mddev->raid_disks - mddev->delta_disks;
4814 /* reshape_position must be on a new-stripe boundary, and one
4815 * further up in new geometry must map after here in old
4818 here_new = mddev->reshape_position;
4819 if (sector_div(here_new, (mddev->new_chunk>>9)*
4820 (mddev->raid_disks - max_degraded))) {
4821 printk(KERN_ERR "raid5: reshape_position not "
4822 "on a stripe boundary\n");
4825 /* here_new is the stripe we will write to */
4826 here_old = mddev->reshape_position;
4827 sector_div(here_old, (mddev->chunk_size>>9)*
4828 (old_disks-max_degraded));
4829 /* here_old is the first stripe that we might need to read
4831 if (here_new >= here_old) {
4832 /* Reading from the same stripe as writing to - bad */
4833 printk(KERN_ERR "raid5: reshape_position too early for "
4834 "auto-recovery - aborting.\n");
4837 printk(KERN_INFO "raid5: reshape will continue\n");
4838 /* OK, we should be able to continue; */
4840 BUG_ON(mddev->level != mddev->new_level);
4841 BUG_ON(mddev->layout != mddev->new_layout);
4842 BUG_ON(mddev->chunk_size != mddev->new_chunk);
4843 BUG_ON(mddev->delta_disks != 0);
4846 if (mddev->private == NULL)
4847 conf = setup_conf(mddev);
4849 conf = mddev->private;
4852 return PTR_ERR(conf);
4854 mddev->thread = conf->thread;
4855 conf->thread = NULL;
4856 mddev->private = conf;
4859 * 0 for a fully functional array, 1 or 2 for a degraded array.
4861 list_for_each_entry(rdev, &mddev->disks, same_set)
4862 if (rdev->raid_disk >= 0 &&
4863 test_bit(In_sync, &rdev->flags))
4866 mddev->degraded = conf->raid_disks - working_disks;
4868 if (mddev->degraded > conf->max_degraded) {
4869 printk(KERN_ERR "raid5: not enough operational devices for %s"
4870 " (%d/%d failed)\n",
4871 mdname(mddev), mddev->degraded, conf->raid_disks);
4875 /* device size must be a multiple of chunk size */
4876 mddev->dev_sectors &= ~(mddev->chunk_size / 512 - 1);
4877 mddev->resync_max_sectors = mddev->dev_sectors;
4879 if (mddev->degraded > 0 &&
4880 mddev->recovery_cp != MaxSector) {
4881 if (mddev->ok_start_degraded)
4883 "raid5: starting dirty degraded array: %s"
4884 "- data corruption possible.\n",
4888 "raid5: cannot start dirty degraded array for %s\n",
4894 if (mddev->degraded == 0)
4895 printk("raid5: raid level %d set %s active with %d out of %d"
4896 " devices, algorithm %d\n", conf->level, mdname(mddev),
4897 mddev->raid_disks-mddev->degraded, mddev->raid_disks,
4900 printk(KERN_ALERT "raid5: raid level %d set %s active with %d"
4901 " out of %d devices, algorithm %d\n", conf->level,
4902 mdname(mddev), mddev->raid_disks - mddev->degraded,
4903 mddev->raid_disks, mddev->new_layout);
4905 print_raid5_conf(conf);
4907 if (conf->reshape_progress != MaxSector) {
4908 printk("...ok start reshape thread\n");
4909 conf->reshape_safe = conf->reshape_progress;
4910 atomic_set(&conf->reshape_stripes, 0);
4911 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
4912 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
4913 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
4914 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
4915 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
4919 /* read-ahead size must cover two whole stripes, which is
4920 * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
4923 int data_disks = conf->previous_raid_disks - conf->max_degraded;
4924 int stripe = data_disks *
4925 (mddev->chunk_size / PAGE_SIZE);
4926 if (mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
4927 mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
4930 /* Ok, everything is just fine now */
4931 if (sysfs_create_group(&mddev->kobj, &raid5_attrs_group))
4933 "raid5: failed to create sysfs attributes for %s\n",
4936 mddev->queue->queue_lock = &conf->device_lock;
4938 mddev->queue->unplug_fn = raid5_unplug_device;
4939 mddev->queue->backing_dev_info.congested_data = mddev;
4940 mddev->queue->backing_dev_info.congested_fn = raid5_congested;
4942 md_set_array_sectors(mddev, raid5_size(mddev, 0, 0));
4944 blk_queue_merge_bvec(mddev->queue, raid5_mergeable_bvec);
4948 md_unregister_thread(mddev->thread);
4949 mddev->thread = NULL;
4951 print_raid5_conf(conf);
4954 mddev->private = NULL;
4955 printk(KERN_ALERT "raid5: failed to run raid set %s\n", mdname(mddev));
4961 static int stop(mddev_t *mddev)
4963 raid5_conf_t *conf = (raid5_conf_t *) mddev->private;
4965 md_unregister_thread(mddev->thread);
4966 mddev->thread = NULL;
4967 mddev->queue->backing_dev_info.congested_fn = NULL;
4968 blk_sync_queue(mddev->queue); /* the unplug fn references 'conf'*/
4969 sysfs_remove_group(&mddev->kobj, &raid5_attrs_group);
4971 mddev->private = NULL;
4976 static void print_sh(struct seq_file *seq, struct stripe_head *sh)
4980 seq_printf(seq, "sh %llu, pd_idx %d, state %ld.\n",
4981 (unsigned long long)sh->sector, sh->pd_idx, sh->state);
4982 seq_printf(seq, "sh %llu, count %d.\n",
4983 (unsigned long long)sh->sector, atomic_read(&sh->count));
4984 seq_printf(seq, "sh %llu, ", (unsigned long long)sh->sector);
4985 for (i = 0; i < sh->disks; i++) {
4986 seq_printf(seq, "(cache%d: %p %ld) ",
4987 i, sh->dev[i].page, sh->dev[i].flags);
4989 seq_printf(seq, "\n");
4992 static void printall(struct seq_file *seq, raid5_conf_t *conf)
4994 struct stripe_head *sh;
4995 struct hlist_node *hn;
4998 spin_lock_irq(&conf->device_lock);
4999 for (i = 0; i < NR_HASH; i++) {
5000 hlist_for_each_entry(sh, hn, &conf->stripe_hashtbl[i], hash) {
5001 if (sh->raid_conf != conf)
5006 spin_unlock_irq(&conf->device_lock);
5010 static void status(struct seq_file *seq, mddev_t *mddev)
5012 raid5_conf_t *conf = (raid5_conf_t *) mddev->private;
5015 seq_printf (seq, " level %d, %dk chunk, algorithm %d", mddev->level, mddev->chunk_size >> 10, mddev->layout);
5016 seq_printf (seq, " [%d/%d] [", conf->raid_disks, conf->raid_disks - mddev->degraded);
5017 for (i = 0; i < conf->raid_disks; i++)
5018 seq_printf (seq, "%s",
5019 conf->disks[i].rdev &&
5020 test_bit(In_sync, &conf->disks[i].rdev->flags) ? "U" : "_");
5021 seq_printf (seq, "]");
5023 seq_printf (seq, "\n");
5024 printall(seq, conf);
5028 static void print_raid5_conf (raid5_conf_t *conf)
5031 struct disk_info *tmp;
5033 printk("RAID5 conf printout:\n");
5035 printk("(conf==NULL)\n");
5038 printk(" --- rd:%d wd:%d\n", conf->raid_disks,
5039 conf->raid_disks - conf->mddev->degraded);
5041 for (i = 0; i < conf->raid_disks; i++) {
5042 char b[BDEVNAME_SIZE];
5043 tmp = conf->disks + i;
5045 printk(" disk %d, o:%d, dev:%s\n",
5046 i, !test_bit(Faulty, &tmp->rdev->flags),
5047 bdevname(tmp->rdev->bdev,b));
5051 static int raid5_spare_active(mddev_t *mddev)
5054 raid5_conf_t *conf = mddev->private;
5055 struct disk_info *tmp;
5057 for (i = 0; i < conf->raid_disks; i++) {
5058 tmp = conf->disks + i;
5060 && !test_bit(Faulty, &tmp->rdev->flags)
5061 && !test_and_set_bit(In_sync, &tmp->rdev->flags)) {
5062 unsigned long flags;
5063 spin_lock_irqsave(&conf->device_lock, flags);
5065 spin_unlock_irqrestore(&conf->device_lock, flags);
5068 print_raid5_conf(conf);
5072 static int raid5_remove_disk(mddev_t *mddev, int number)
5074 raid5_conf_t *conf = mddev->private;
5077 struct disk_info *p = conf->disks + number;
5079 print_raid5_conf(conf);
5082 if (number >= conf->raid_disks &&
5083 conf->reshape_progress == MaxSector)
5084 clear_bit(In_sync, &rdev->flags);
5086 if (test_bit(In_sync, &rdev->flags) ||
5087 atomic_read(&rdev->nr_pending)) {
5091 /* Only remove non-faulty devices if recovery
5094 if (!test_bit(Faulty, &rdev->flags) &&
5095 mddev->degraded <= conf->max_degraded &&
5096 number < conf->raid_disks) {
5102 if (atomic_read(&rdev->nr_pending)) {
5103 /* lost the race, try later */
5110 print_raid5_conf(conf);
5114 static int raid5_add_disk(mddev_t *mddev, mdk_rdev_t *rdev)
5116 raid5_conf_t *conf = mddev->private;
5119 struct disk_info *p;
5121 int last = conf->raid_disks - 1;
5123 if (mddev->degraded > conf->max_degraded)
5124 /* no point adding a device */
5127 if (rdev->raid_disk >= 0)
5128 first = last = rdev->raid_disk;
5131 * find the disk ... but prefer rdev->saved_raid_disk
5134 if (rdev->saved_raid_disk >= 0 &&
5135 rdev->saved_raid_disk >= first &&
5136 conf->disks[rdev->saved_raid_disk].rdev == NULL)
5137 disk = rdev->saved_raid_disk;
5140 for ( ; disk <= last ; disk++)
5141 if ((p=conf->disks + disk)->rdev == NULL) {
5142 clear_bit(In_sync, &rdev->flags);
5143 rdev->raid_disk = disk;
5145 if (rdev->saved_raid_disk != disk)
5147 rcu_assign_pointer(p->rdev, rdev);
5150 print_raid5_conf(conf);
5154 static int raid5_resize(mddev_t *mddev, sector_t sectors)
5156 /* no resync is happening, and there is enough space
5157 * on all devices, so we can resize.
5158 * We need to make sure resync covers any new space.
5159 * If the array is shrinking we should possibly wait until
5160 * any io in the removed space completes, but it hardly seems
5163 sectors &= ~((sector_t)mddev->chunk_size/512 - 1);
5164 md_set_array_sectors(mddev, raid5_size(mddev, sectors,
5165 mddev->raid_disks));
5166 if (mddev->array_sectors >
5167 raid5_size(mddev, sectors, mddev->raid_disks))
5169 set_capacity(mddev->gendisk, mddev->array_sectors);
5171 if (sectors > mddev->dev_sectors && mddev->recovery_cp == MaxSector) {
5172 mddev->recovery_cp = mddev->dev_sectors;
5173 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
5175 mddev->dev_sectors = sectors;
5176 mddev->resync_max_sectors = sectors;
5180 static int raid5_check_reshape(mddev_t *mddev)
5182 raid5_conf_t *conf = mddev_to_conf(mddev);
5184 if (mddev->delta_disks == 0 &&
5185 mddev->new_layout == mddev->layout &&
5186 mddev->new_chunk == mddev->chunk_size)
5187 return -EINVAL; /* nothing to do */
5189 /* Cannot grow a bitmap yet */
5191 if (mddev->degraded > conf->max_degraded)
5193 if (mddev->delta_disks < 0) {
5194 /* We might be able to shrink, but the devices must
5195 * be made bigger first.
5196 * For raid6, 4 is the minimum size.
5197 * Otherwise 2 is the minimum
5200 if (mddev->level == 6)
5202 if (mddev->raid_disks + mddev->delta_disks < min)
5206 /* Can only proceed if there are plenty of stripe_heads.
5207 * We need a minimum of one full stripe,, and for sensible progress
5208 * it is best to have about 4 times that.
5209 * If we require 4 times, then the default 256 4K stripe_heads will
5210 * allow for chunk sizes up to 256K, which is probably OK.
5211 * If the chunk size is greater, user-space should request more
5212 * stripe_heads first.
5214 if ((mddev->chunk_size / STRIPE_SIZE) * 4 > conf->max_nr_stripes ||
5215 (mddev->new_chunk / STRIPE_SIZE) * 4 > conf->max_nr_stripes) {
5216 printk(KERN_WARNING "raid5: reshape: not enough stripes. Needed %lu\n",
5217 (max(mddev->chunk_size, mddev->new_chunk)
5222 return resize_stripes(conf, conf->raid_disks + mddev->delta_disks);
5225 static int raid5_start_reshape(mddev_t *mddev)
5227 raid5_conf_t *conf = mddev_to_conf(mddev);
5230 int added_devices = 0;
5231 unsigned long flags;
5233 if (test_bit(MD_RECOVERY_RUNNING, &mddev->recovery))
5236 list_for_each_entry(rdev, &mddev->disks, same_set)
5237 if (rdev->raid_disk < 0 &&
5238 !test_bit(Faulty, &rdev->flags))
5241 if (spares - mddev->degraded < mddev->delta_disks - conf->max_degraded)
5242 /* Not enough devices even to make a degraded array
5247 /* Refuse to reduce size of the array. Any reductions in
5248 * array size must be through explicit setting of array_size
5251 if (raid5_size(mddev, 0, conf->raid_disks + mddev->delta_disks)
5252 < mddev->array_sectors) {
5253 printk(KERN_ERR "md: %s: array size must be reduced "
5254 "before number of disks\n", mdname(mddev));
5258 atomic_set(&conf->reshape_stripes, 0);
5259 spin_lock_irq(&conf->device_lock);
5260 conf->previous_raid_disks = conf->raid_disks;
5261 conf->raid_disks += mddev->delta_disks;
5262 conf->prev_chunk = conf->chunk_size;
5263 conf->chunk_size = mddev->new_chunk;
5264 conf->prev_algo = conf->algorithm;
5265 conf->algorithm = mddev->new_layout;
5266 if (mddev->delta_disks < 0)
5267 conf->reshape_progress = raid5_size(mddev, 0, 0);
5269 conf->reshape_progress = 0;
5270 conf->reshape_safe = conf->reshape_progress;
5272 spin_unlock_irq(&conf->device_lock);
5274 /* Add some new drives, as many as will fit.
5275 * We know there are enough to make the newly sized array work.
5277 list_for_each_entry(rdev, &mddev->disks, same_set)
5278 if (rdev->raid_disk < 0 &&
5279 !test_bit(Faulty, &rdev->flags)) {
5280 if (raid5_add_disk(mddev, rdev) == 0) {
5282 set_bit(In_sync, &rdev->flags);
5284 rdev->recovery_offset = 0;
5285 sprintf(nm, "rd%d", rdev->raid_disk);
5286 if (sysfs_create_link(&mddev->kobj,
5289 "raid5: failed to create "
5290 " link %s for %s\n",
5296 if (mddev->delta_disks > 0) {
5297 spin_lock_irqsave(&conf->device_lock, flags);
5298 mddev->degraded = (conf->raid_disks - conf->previous_raid_disks)
5300 spin_unlock_irqrestore(&conf->device_lock, flags);
5302 mddev->raid_disks = conf->raid_disks;
5303 mddev->reshape_position = 0;
5304 set_bit(MD_CHANGE_DEVS, &mddev->flags);
5306 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
5307 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
5308 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
5309 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
5310 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
5312 if (!mddev->sync_thread) {
5313 mddev->recovery = 0;
5314 spin_lock_irq(&conf->device_lock);
5315 mddev->raid_disks = conf->raid_disks = conf->previous_raid_disks;
5316 conf->reshape_progress = MaxSector;
5317 spin_unlock_irq(&conf->device_lock);
5320 conf->reshape_checkpoint = jiffies;
5321 md_wakeup_thread(mddev->sync_thread);
5322 md_new_event(mddev);
5326 /* This is called from the reshape thread and should make any
5327 * changes needed in 'conf'
5329 static void end_reshape(raid5_conf_t *conf)
5332 if (!test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery)) {
5334 spin_lock_irq(&conf->device_lock);
5335 conf->previous_raid_disks = conf->raid_disks;
5336 conf->reshape_progress = MaxSector;
5337 spin_unlock_irq(&conf->device_lock);
5338 wake_up(&conf->wait_for_overlap);
5340 /* read-ahead size must cover two whole stripes, which is
5341 * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
5344 int data_disks = conf->raid_disks - conf->max_degraded;
5345 int stripe = data_disks * (conf->chunk_size
5347 if (conf->mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
5348 conf->mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
5353 /* This is called from the raid5d thread with mddev_lock held.
5354 * It makes config changes to the device.
5356 static void raid5_finish_reshape(mddev_t *mddev)
5358 struct block_device *bdev;
5359 raid5_conf_t *conf = mddev_to_conf(mddev);
5361 if (!test_bit(MD_RECOVERY_INTR, &mddev->recovery)) {
5363 if (mddev->delta_disks > 0) {
5364 md_set_array_sectors(mddev, raid5_size(mddev, 0, 0));
5365 set_capacity(mddev->gendisk, mddev->array_sectors);
5368 bdev = bdget_disk(mddev->gendisk, 0);
5370 mutex_lock(&bdev->bd_inode->i_mutex);
5371 i_size_write(bdev->bd_inode,
5372 (loff_t)mddev->array_sectors << 9);
5373 mutex_unlock(&bdev->bd_inode->i_mutex);
5378 mddev->degraded = conf->raid_disks;
5379 for (d = 0; d < conf->raid_disks ; d++)
5380 if (conf->disks[d].rdev &&
5382 &conf->disks[d].rdev->flags))
5384 for (d = conf->raid_disks ;
5385 d < conf->raid_disks - mddev->delta_disks;
5387 raid5_remove_disk(mddev, d);
5389 mddev->layout = conf->algorithm;
5390 mddev->chunk_size = conf->chunk_size;
5391 mddev->reshape_position = MaxSector;
5392 mddev->delta_disks = 0;
5396 static void raid5_quiesce(mddev_t *mddev, int state)
5398 raid5_conf_t *conf = mddev_to_conf(mddev);
5401 case 2: /* resume for a suspend */
5402 wake_up(&conf->wait_for_overlap);
5405 case 1: /* stop all writes */
5406 spin_lock_irq(&conf->device_lock);
5408 wait_event_lock_irq(conf->wait_for_stripe,
5409 atomic_read(&conf->active_stripes) == 0 &&
5410 atomic_read(&conf->active_aligned_reads) == 0,
5411 conf->device_lock, /* nothing */);
5412 spin_unlock_irq(&conf->device_lock);
5415 case 0: /* re-enable writes */
5416 spin_lock_irq(&conf->device_lock);
5418 wake_up(&conf->wait_for_stripe);
5419 wake_up(&conf->wait_for_overlap);
5420 spin_unlock_irq(&conf->device_lock);
5426 static void *raid5_takeover_raid1(mddev_t *mddev)
5430 if (mddev->raid_disks != 2 ||
5431 mddev->degraded > 1)
5432 return ERR_PTR(-EINVAL);
5434 /* Should check if there are write-behind devices? */
5436 chunksect = 64*2; /* 64K by default */
5438 /* The array must be an exact multiple of chunksize */
5439 while (chunksect && (mddev->array_sectors & (chunksect-1)))
5442 if ((chunksect<<9) < STRIPE_SIZE)
5443 /* array size does not allow a suitable chunk size */
5444 return ERR_PTR(-EINVAL);
5446 mddev->new_level = 5;
5447 mddev->new_layout = ALGORITHM_LEFT_SYMMETRIC;
5448 mddev->new_chunk = chunksect << 9;
5450 return setup_conf(mddev);
5453 static void *raid5_takeover_raid6(mddev_t *mddev)
5457 switch (mddev->layout) {
5458 case ALGORITHM_LEFT_ASYMMETRIC_6:
5459 new_layout = ALGORITHM_LEFT_ASYMMETRIC;
5461 case ALGORITHM_RIGHT_ASYMMETRIC_6:
5462 new_layout = ALGORITHM_RIGHT_ASYMMETRIC;
5464 case ALGORITHM_LEFT_SYMMETRIC_6:
5465 new_layout = ALGORITHM_LEFT_SYMMETRIC;
5467 case ALGORITHM_RIGHT_SYMMETRIC_6:
5468 new_layout = ALGORITHM_RIGHT_SYMMETRIC;
5470 case ALGORITHM_PARITY_0_6:
5471 new_layout = ALGORITHM_PARITY_0;
5473 case ALGORITHM_PARITY_N:
5474 new_layout = ALGORITHM_PARITY_N;
5477 return ERR_PTR(-EINVAL);
5479 mddev->new_level = 5;
5480 mddev->new_layout = new_layout;
5481 mddev->delta_disks = -1;
5482 mddev->raid_disks -= 1;
5483 return setup_conf(mddev);
5487 static int raid5_reconfig(mddev_t *mddev, int new_layout, int new_chunk)
5489 /* For a 2-drive array, the layout and chunk size can be changed
5490 * immediately as not restriping is needed.
5491 * For larger arrays we record the new value - after validation
5492 * to be used by a reshape pass.
5494 raid5_conf_t *conf = mddev_to_conf(mddev);
5496 if (new_layout >= 0 && !algorithm_valid_raid5(new_layout))
5498 if (new_chunk > 0) {
5499 if (new_chunk & (new_chunk-1))
5500 /* not a power of 2 */
5502 if (new_chunk < PAGE_SIZE)
5504 if (mddev->array_sectors & ((new_chunk>>9)-1))
5505 /* not factor of array size */
5509 /* They look valid */
5511 if (mddev->raid_disks == 2) {
5513 if (new_layout >= 0) {
5514 conf->algorithm = new_layout;
5515 mddev->layout = mddev->new_layout = new_layout;
5517 if (new_chunk > 0) {
5518 conf->chunk_size = new_chunk;
5519 mddev->chunk_size = mddev->new_chunk = new_chunk;
5521 set_bit(MD_CHANGE_DEVS, &mddev->flags);
5522 md_wakeup_thread(mddev->thread);
5524 if (new_layout >= 0)
5525 mddev->new_layout = new_layout;
5527 mddev->new_chunk = new_chunk;
5532 static int raid6_reconfig(mddev_t *mddev, int new_layout, int new_chunk)
5534 if (new_layout >= 0 && !algorithm_valid_raid6(new_layout))
5536 if (new_chunk > 0) {
5537 if (new_chunk & (new_chunk-1))
5538 /* not a power of 2 */
5540 if (new_chunk < PAGE_SIZE)
5542 if (mddev->array_sectors & ((new_chunk>>9)-1))
5543 /* not factor of array size */
5547 /* They look valid */
5549 if (new_layout >= 0)
5550 mddev->new_layout = new_layout;
5552 mddev->new_chunk = new_chunk;
5557 static void *raid5_takeover(mddev_t *mddev)
5559 /* raid5 can take over:
5560 * raid0 - if all devices are the same - make it a raid4 layout
5561 * raid1 - if there are two drives. We need to know the chunk size
5562 * raid4 - trivial - just use a raid4 layout.
5563 * raid6 - Providing it is a *_6 layout
5565 * For now, just do raid1
5568 if (mddev->level == 1)
5569 return raid5_takeover_raid1(mddev);
5570 if (mddev->level == 4) {
5571 mddev->new_layout = ALGORITHM_PARITY_N;
5572 mddev->new_level = 5;
5573 return setup_conf(mddev);
5575 if (mddev->level == 6)
5576 return raid5_takeover_raid6(mddev);
5578 return ERR_PTR(-EINVAL);
5582 static struct mdk_personality raid5_personality;
5584 static void *raid6_takeover(mddev_t *mddev)
5586 /* Currently can only take over a raid5. We map the
5587 * personality to an equivalent raid6 personality
5588 * with the Q block at the end.
5592 if (mddev->pers != &raid5_personality)
5593 return ERR_PTR(-EINVAL);
5594 if (mddev->degraded > 1)
5595 return ERR_PTR(-EINVAL);
5596 if (mddev->raid_disks > 253)
5597 return ERR_PTR(-EINVAL);
5598 if (mddev->raid_disks < 3)
5599 return ERR_PTR(-EINVAL);
5601 switch (mddev->layout) {
5602 case ALGORITHM_LEFT_ASYMMETRIC:
5603 new_layout = ALGORITHM_LEFT_ASYMMETRIC_6;
5605 case ALGORITHM_RIGHT_ASYMMETRIC:
5606 new_layout = ALGORITHM_RIGHT_ASYMMETRIC_6;
5608 case ALGORITHM_LEFT_SYMMETRIC:
5609 new_layout = ALGORITHM_LEFT_SYMMETRIC_6;
5611 case ALGORITHM_RIGHT_SYMMETRIC:
5612 new_layout = ALGORITHM_RIGHT_SYMMETRIC_6;
5614 case ALGORITHM_PARITY_0:
5615 new_layout = ALGORITHM_PARITY_0_6;
5617 case ALGORITHM_PARITY_N:
5618 new_layout = ALGORITHM_PARITY_N;
5621 return ERR_PTR(-EINVAL);
5623 mddev->new_level = 6;
5624 mddev->new_layout = new_layout;
5625 mddev->delta_disks = 1;
5626 mddev->raid_disks += 1;
5627 return setup_conf(mddev);
5631 static struct mdk_personality raid6_personality =
5635 .owner = THIS_MODULE,
5636 .make_request = make_request,
5640 .error_handler = error,
5641 .hot_add_disk = raid5_add_disk,
5642 .hot_remove_disk= raid5_remove_disk,
5643 .spare_active = raid5_spare_active,
5644 .sync_request = sync_request,
5645 .resize = raid5_resize,
5647 .check_reshape = raid5_check_reshape,
5648 .start_reshape = raid5_start_reshape,
5649 .finish_reshape = raid5_finish_reshape,
5650 .quiesce = raid5_quiesce,
5651 .takeover = raid6_takeover,
5652 .reconfig = raid6_reconfig,
5654 static struct mdk_personality raid5_personality =
5658 .owner = THIS_MODULE,
5659 .make_request = make_request,
5663 .error_handler = error,
5664 .hot_add_disk = raid5_add_disk,
5665 .hot_remove_disk= raid5_remove_disk,
5666 .spare_active = raid5_spare_active,
5667 .sync_request = sync_request,
5668 .resize = raid5_resize,
5670 .check_reshape = raid5_check_reshape,
5671 .start_reshape = raid5_start_reshape,
5672 .finish_reshape = raid5_finish_reshape,
5673 .quiesce = raid5_quiesce,
5674 .takeover = raid5_takeover,
5675 .reconfig = raid5_reconfig,
5678 static struct mdk_personality raid4_personality =
5682 .owner = THIS_MODULE,
5683 .make_request = make_request,
5687 .error_handler = error,
5688 .hot_add_disk = raid5_add_disk,
5689 .hot_remove_disk= raid5_remove_disk,
5690 .spare_active = raid5_spare_active,
5691 .sync_request = sync_request,
5692 .resize = raid5_resize,
5694 .check_reshape = raid5_check_reshape,
5695 .start_reshape = raid5_start_reshape,
5696 .finish_reshape = raid5_finish_reshape,
5697 .quiesce = raid5_quiesce,
5700 static int __init raid5_init(void)
5702 register_md_personality(&raid6_personality);
5703 register_md_personality(&raid5_personality);
5704 register_md_personality(&raid4_personality);
5708 static void raid5_exit(void)
5710 unregister_md_personality(&raid6_personality);
5711 unregister_md_personality(&raid5_personality);
5712 unregister_md_personality(&raid4_personality);
5715 module_init(raid5_init);
5716 module_exit(raid5_exit);
5717 MODULE_LICENSE("GPL");
5718 MODULE_ALIAS("md-personality-4"); /* RAID5 */
5719 MODULE_ALIAS("md-raid5");
5720 MODULE_ALIAS("md-raid4");
5721 MODULE_ALIAS("md-level-5");
5722 MODULE_ALIAS("md-level-4");
5723 MODULE_ALIAS("md-personality-8"); /* RAID6 */
5724 MODULE_ALIAS("md-raid6");
5725 MODULE_ALIAS("md-level-6");
5727 /* This used to be two separate modules, they were: */
5728 MODULE_ALIAS("raid5");
5729 MODULE_ALIAS("raid6");