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>
62 #define NR_STRIPES 256
63 #define STRIPE_SIZE PAGE_SIZE
64 #define STRIPE_SHIFT (PAGE_SHIFT - 9)
65 #define STRIPE_SECTORS (STRIPE_SIZE>>9)
66 #define IO_THRESHOLD 1
67 #define BYPASS_THRESHOLD 1
68 #define NR_HASH (PAGE_SIZE / sizeof(struct hlist_head))
69 #define HASH_MASK (NR_HASH - 1)
71 #define stripe_hash(conf, sect) (&((conf)->stripe_hashtbl[((sect) >> STRIPE_SHIFT) & HASH_MASK]))
73 /* bio's attached to a stripe+device for I/O are linked together in bi_sector
74 * order without overlap. There may be several bio's per stripe+device, and
75 * a bio could span several devices.
76 * When walking this list for a particular stripe+device, we must never proceed
77 * beyond a bio that extends past this device, as the next bio might no longer
79 * This macro is used to determine the 'next' bio in the list, given the sector
80 * of the current stripe+device
82 #define r5_next_bio(bio, sect) ( ( (bio)->bi_sector + ((bio)->bi_size>>9) < sect + STRIPE_SECTORS) ? (bio)->bi_next : NULL)
84 * The following can be used to debug the driver
86 #define RAID5_PARANOIA 1
87 #if RAID5_PARANOIA && defined(CONFIG_SMP)
88 # define CHECK_DEVLOCK() assert_spin_locked(&conf->device_lock)
90 # define CHECK_DEVLOCK()
98 #define printk_rl(args...) ((void) (printk_ratelimit() && printk(args)))
101 * We maintain a biased count of active stripes in the bottom 16 bits of
102 * bi_phys_segments, and a count of processed stripes in the upper 16 bits
104 static inline int raid5_bi_phys_segments(struct bio *bio)
106 return bio->bi_phys_segments & 0xffff;
109 static inline int raid5_bi_hw_segments(struct bio *bio)
111 return (bio->bi_phys_segments >> 16) & 0xffff;
114 static inline int raid5_dec_bi_phys_segments(struct bio *bio)
116 --bio->bi_phys_segments;
117 return raid5_bi_phys_segments(bio);
120 static inline int raid5_dec_bi_hw_segments(struct bio *bio)
122 unsigned short val = raid5_bi_hw_segments(bio);
125 bio->bi_phys_segments = (val << 16) | raid5_bi_phys_segments(bio);
129 static inline void raid5_set_bi_hw_segments(struct bio *bio, unsigned int cnt)
131 bio->bi_phys_segments = raid5_bi_phys_segments(bio) || (cnt << 16);
134 /* Find first data disk in a raid6 stripe */
135 static inline int raid6_d0(struct stripe_head *sh)
138 /* ddf always start from first device */
140 /* md starts just after Q block */
141 if (sh->qd_idx == sh->disks - 1)
144 return sh->qd_idx + 1;
146 static inline int raid6_next_disk(int disk, int raid_disks)
149 return (disk < raid_disks) ? disk : 0;
152 /* When walking through the disks in a raid5, starting at raid6_d0,
153 * We need to map each disk to a 'slot', where the data disks are slot
154 * 0 .. raid_disks-3, the parity disk is raid_disks-2 and the Q disk
155 * is raid_disks-1. This help does that mapping.
157 static int raid6_idx_to_slot(int idx, struct stripe_head *sh,
158 int *count, int syndrome_disks)
164 if (idx == sh->pd_idx)
165 return syndrome_disks;
166 if (idx == sh->qd_idx)
167 return syndrome_disks + 1;
173 static void return_io(struct bio *return_bi)
175 struct bio *bi = return_bi;
178 return_bi = bi->bi_next;
186 static void print_raid5_conf (raid5_conf_t *conf);
188 static int stripe_operations_active(struct stripe_head *sh)
190 return sh->check_state || sh->reconstruct_state ||
191 test_bit(STRIPE_BIOFILL_RUN, &sh->state) ||
192 test_bit(STRIPE_COMPUTE_RUN, &sh->state);
195 static void __release_stripe(raid5_conf_t *conf, struct stripe_head *sh)
197 if (atomic_dec_and_test(&sh->count)) {
198 BUG_ON(!list_empty(&sh->lru));
199 BUG_ON(atomic_read(&conf->active_stripes)==0);
200 if (test_bit(STRIPE_HANDLE, &sh->state)) {
201 if (test_bit(STRIPE_DELAYED, &sh->state)) {
202 list_add_tail(&sh->lru, &conf->delayed_list);
203 blk_plug_device(conf->mddev->queue);
204 } else if (test_bit(STRIPE_BIT_DELAY, &sh->state) &&
205 sh->bm_seq - conf->seq_write > 0) {
206 list_add_tail(&sh->lru, &conf->bitmap_list);
207 blk_plug_device(conf->mddev->queue);
209 clear_bit(STRIPE_BIT_DELAY, &sh->state);
210 list_add_tail(&sh->lru, &conf->handle_list);
212 md_wakeup_thread(conf->mddev->thread);
214 BUG_ON(stripe_operations_active(sh));
215 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
216 atomic_dec(&conf->preread_active_stripes);
217 if (atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD)
218 md_wakeup_thread(conf->mddev->thread);
220 atomic_dec(&conf->active_stripes);
221 if (!test_bit(STRIPE_EXPANDING, &sh->state)) {
222 list_add_tail(&sh->lru, &conf->inactive_list);
223 wake_up(&conf->wait_for_stripe);
224 if (conf->retry_read_aligned)
225 md_wakeup_thread(conf->mddev->thread);
231 static void release_stripe(struct stripe_head *sh)
233 raid5_conf_t *conf = sh->raid_conf;
236 spin_lock_irqsave(&conf->device_lock, flags);
237 __release_stripe(conf, sh);
238 spin_unlock_irqrestore(&conf->device_lock, flags);
241 static inline void remove_hash(struct stripe_head *sh)
243 pr_debug("remove_hash(), stripe %llu\n",
244 (unsigned long long)sh->sector);
246 hlist_del_init(&sh->hash);
249 static inline void insert_hash(raid5_conf_t *conf, struct stripe_head *sh)
251 struct hlist_head *hp = stripe_hash(conf, sh->sector);
253 pr_debug("insert_hash(), stripe %llu\n",
254 (unsigned long long)sh->sector);
257 hlist_add_head(&sh->hash, hp);
261 /* find an idle stripe, make sure it is unhashed, and return it. */
262 static struct stripe_head *get_free_stripe(raid5_conf_t *conf)
264 struct stripe_head *sh = NULL;
265 struct list_head *first;
268 if (list_empty(&conf->inactive_list))
270 first = conf->inactive_list.next;
271 sh = list_entry(first, struct stripe_head, lru);
272 list_del_init(first);
274 atomic_inc(&conf->active_stripes);
279 static void shrink_buffers(struct stripe_head *sh, int num)
284 for (i=0; i<num ; i++) {
288 sh->dev[i].page = NULL;
293 static int grow_buffers(struct stripe_head *sh, int num)
297 for (i=0; i<num; i++) {
300 if (!(page = alloc_page(GFP_KERNEL))) {
303 sh->dev[i].page = page;
308 static void raid5_build_block(struct stripe_head *sh, int i, int previous);
309 static void stripe_set_idx(sector_t stripe, raid5_conf_t *conf, int previous,
310 struct stripe_head *sh);
312 static void init_stripe(struct stripe_head *sh, sector_t sector, int previous)
314 raid5_conf_t *conf = sh->raid_conf;
317 BUG_ON(atomic_read(&sh->count) != 0);
318 BUG_ON(test_bit(STRIPE_HANDLE, &sh->state));
319 BUG_ON(stripe_operations_active(sh));
322 pr_debug("init_stripe called, stripe %llu\n",
323 (unsigned long long)sh->sector);
327 sh->generation = conf->generation - previous;
328 sh->disks = previous ? conf->previous_raid_disks : conf->raid_disks;
330 stripe_set_idx(sector, conf, previous, sh);
334 for (i = sh->disks; i--; ) {
335 struct r5dev *dev = &sh->dev[i];
337 if (dev->toread || dev->read || dev->towrite || dev->written ||
338 test_bit(R5_LOCKED, &dev->flags)) {
339 printk(KERN_ERR "sector=%llx i=%d %p %p %p %p %d\n",
340 (unsigned long long)sh->sector, i, dev->toread,
341 dev->read, dev->towrite, dev->written,
342 test_bit(R5_LOCKED, &dev->flags));
346 raid5_build_block(sh, i, previous);
348 insert_hash(conf, sh);
351 static struct stripe_head *__find_stripe(raid5_conf_t *conf, sector_t sector,
354 struct stripe_head *sh;
355 struct hlist_node *hn;
358 pr_debug("__find_stripe, sector %llu\n", (unsigned long long)sector);
359 hlist_for_each_entry(sh, hn, stripe_hash(conf, sector), hash)
360 if (sh->sector == sector && sh->generation == generation)
362 pr_debug("__stripe %llu not in cache\n", (unsigned long long)sector);
366 static void unplug_slaves(mddev_t *mddev);
367 static void raid5_unplug_device(struct request_queue *q);
369 static struct stripe_head *
370 get_active_stripe(raid5_conf_t *conf, sector_t sector,
371 int previous, int noblock, int noquiesce)
373 struct stripe_head *sh;
375 pr_debug("get_stripe, sector %llu\n", (unsigned long long)sector);
377 spin_lock_irq(&conf->device_lock);
380 wait_event_lock_irq(conf->wait_for_stripe,
381 conf->quiesce == 0 || noquiesce,
382 conf->device_lock, /* nothing */);
383 sh = __find_stripe(conf, sector, conf->generation - previous);
385 if (!conf->inactive_blocked)
386 sh = get_free_stripe(conf);
387 if (noblock && sh == NULL)
390 conf->inactive_blocked = 1;
391 wait_event_lock_irq(conf->wait_for_stripe,
392 !list_empty(&conf->inactive_list) &&
393 (atomic_read(&conf->active_stripes)
394 < (conf->max_nr_stripes *3/4)
395 || !conf->inactive_blocked),
397 raid5_unplug_device(conf->mddev->queue)
399 conf->inactive_blocked = 0;
401 init_stripe(sh, sector, previous);
403 if (atomic_read(&sh->count)) {
404 BUG_ON(!list_empty(&sh->lru)
405 && !test_bit(STRIPE_EXPANDING, &sh->state));
407 if (!test_bit(STRIPE_HANDLE, &sh->state))
408 atomic_inc(&conf->active_stripes);
409 if (list_empty(&sh->lru) &&
410 !test_bit(STRIPE_EXPANDING, &sh->state))
412 list_del_init(&sh->lru);
415 } while (sh == NULL);
418 atomic_inc(&sh->count);
420 spin_unlock_irq(&conf->device_lock);
425 raid5_end_read_request(struct bio *bi, int error);
427 raid5_end_write_request(struct bio *bi, int error);
429 static void ops_run_io(struct stripe_head *sh, struct stripe_head_state *s)
431 raid5_conf_t *conf = sh->raid_conf;
432 int i, disks = sh->disks;
436 for (i = disks; i--; ) {
440 if (test_and_clear_bit(R5_Wantwrite, &sh->dev[i].flags))
442 else if (test_and_clear_bit(R5_Wantread, &sh->dev[i].flags))
447 bi = &sh->dev[i].req;
451 bi->bi_end_io = raid5_end_write_request;
453 bi->bi_end_io = raid5_end_read_request;
456 rdev = rcu_dereference(conf->disks[i].rdev);
457 if (rdev && test_bit(Faulty, &rdev->flags))
460 atomic_inc(&rdev->nr_pending);
464 if (s->syncing || s->expanding || s->expanded)
465 md_sync_acct(rdev->bdev, STRIPE_SECTORS);
467 set_bit(STRIPE_IO_STARTED, &sh->state);
469 bi->bi_bdev = rdev->bdev;
470 pr_debug("%s: for %llu schedule op %ld on disc %d\n",
471 __func__, (unsigned long long)sh->sector,
473 atomic_inc(&sh->count);
474 bi->bi_sector = sh->sector + rdev->data_offset;
475 bi->bi_flags = 1 << BIO_UPTODATE;
479 bi->bi_io_vec = &sh->dev[i].vec;
480 bi->bi_io_vec[0].bv_len = STRIPE_SIZE;
481 bi->bi_io_vec[0].bv_offset = 0;
482 bi->bi_size = STRIPE_SIZE;
485 test_bit(R5_ReWrite, &sh->dev[i].flags))
486 atomic_add(STRIPE_SECTORS,
487 &rdev->corrected_errors);
488 generic_make_request(bi);
491 set_bit(STRIPE_DEGRADED, &sh->state);
492 pr_debug("skip op %ld on disc %d for sector %llu\n",
493 bi->bi_rw, i, (unsigned long long)sh->sector);
494 clear_bit(R5_LOCKED, &sh->dev[i].flags);
495 set_bit(STRIPE_HANDLE, &sh->state);
500 static struct dma_async_tx_descriptor *
501 async_copy_data(int frombio, struct bio *bio, struct page *page,
502 sector_t sector, struct dma_async_tx_descriptor *tx)
505 struct page *bio_page;
508 struct async_submit_ctl submit;
509 enum async_tx_flags flags = 0;
511 if (bio->bi_sector >= sector)
512 page_offset = (signed)(bio->bi_sector - sector) * 512;
514 page_offset = (signed)(sector - bio->bi_sector) * -512;
517 flags |= ASYNC_TX_FENCE;
518 init_async_submit(&submit, flags, tx, NULL, NULL, NULL);
520 bio_for_each_segment(bvl, bio, i) {
521 int len = bio_iovec_idx(bio, i)->bv_len;
525 if (page_offset < 0) {
526 b_offset = -page_offset;
527 page_offset += b_offset;
531 if (len > 0 && page_offset + len > STRIPE_SIZE)
532 clen = STRIPE_SIZE - page_offset;
537 b_offset += bio_iovec_idx(bio, i)->bv_offset;
538 bio_page = bio_iovec_idx(bio, i)->bv_page;
540 tx = async_memcpy(page, bio_page, page_offset,
541 b_offset, clen, &submit);
543 tx = async_memcpy(bio_page, page, b_offset,
544 page_offset, clen, &submit);
546 /* chain the operations */
547 submit.depend_tx = tx;
549 if (clen < len) /* hit end of page */
557 static void ops_complete_biofill(void *stripe_head_ref)
559 struct stripe_head *sh = stripe_head_ref;
560 struct bio *return_bi = NULL;
561 raid5_conf_t *conf = sh->raid_conf;
564 pr_debug("%s: stripe %llu\n", __func__,
565 (unsigned long long)sh->sector);
567 /* clear completed biofills */
568 spin_lock_irq(&conf->device_lock);
569 for (i = sh->disks; i--; ) {
570 struct r5dev *dev = &sh->dev[i];
572 /* acknowledge completion of a biofill operation */
573 /* and check if we need to reply to a read request,
574 * new R5_Wantfill requests are held off until
575 * !STRIPE_BIOFILL_RUN
577 if (test_and_clear_bit(R5_Wantfill, &dev->flags)) {
578 struct bio *rbi, *rbi2;
583 while (rbi && rbi->bi_sector <
584 dev->sector + STRIPE_SECTORS) {
585 rbi2 = r5_next_bio(rbi, dev->sector);
586 if (!raid5_dec_bi_phys_segments(rbi)) {
587 rbi->bi_next = return_bi;
594 spin_unlock_irq(&conf->device_lock);
595 clear_bit(STRIPE_BIOFILL_RUN, &sh->state);
597 return_io(return_bi);
599 set_bit(STRIPE_HANDLE, &sh->state);
603 static void ops_run_biofill(struct stripe_head *sh)
605 struct dma_async_tx_descriptor *tx = NULL;
606 raid5_conf_t *conf = sh->raid_conf;
607 struct async_submit_ctl submit;
610 pr_debug("%s: stripe %llu\n", __func__,
611 (unsigned long long)sh->sector);
613 for (i = sh->disks; i--; ) {
614 struct r5dev *dev = &sh->dev[i];
615 if (test_bit(R5_Wantfill, &dev->flags)) {
617 spin_lock_irq(&conf->device_lock);
618 dev->read = rbi = dev->toread;
620 spin_unlock_irq(&conf->device_lock);
621 while (rbi && rbi->bi_sector <
622 dev->sector + STRIPE_SECTORS) {
623 tx = async_copy_data(0, rbi, dev->page,
625 rbi = r5_next_bio(rbi, dev->sector);
630 atomic_inc(&sh->count);
631 init_async_submit(&submit, ASYNC_TX_ACK, tx, ops_complete_biofill, sh, NULL);
632 async_trigger_callback(&submit);
635 static void mark_target_uptodate(struct stripe_head *sh, int target)
642 tgt = &sh->dev[target];
643 set_bit(R5_UPTODATE, &tgt->flags);
644 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
645 clear_bit(R5_Wantcompute, &tgt->flags);
648 static void ops_complete_compute(void *stripe_head_ref)
650 struct stripe_head *sh = stripe_head_ref;
652 pr_debug("%s: stripe %llu\n", __func__,
653 (unsigned long long)sh->sector);
655 /* mark the computed target(s) as uptodate */
656 mark_target_uptodate(sh, sh->ops.target);
657 mark_target_uptodate(sh, sh->ops.target2);
659 clear_bit(STRIPE_COMPUTE_RUN, &sh->state);
660 if (sh->check_state == check_state_compute_run)
661 sh->check_state = check_state_compute_result;
662 set_bit(STRIPE_HANDLE, &sh->state);
666 /* return a pointer to the address conversion region of the scribble buffer */
667 static addr_conv_t *to_addr_conv(struct stripe_head *sh,
668 struct raid5_percpu *percpu)
670 return percpu->scribble + sizeof(struct page *) * (sh->disks + 2);
673 static struct dma_async_tx_descriptor *
674 ops_run_compute5(struct stripe_head *sh, struct raid5_percpu *percpu)
676 int disks = sh->disks;
677 struct page **xor_srcs = percpu->scribble;
678 int target = sh->ops.target;
679 struct r5dev *tgt = &sh->dev[target];
680 struct page *xor_dest = tgt->page;
682 struct dma_async_tx_descriptor *tx;
683 struct async_submit_ctl submit;
686 pr_debug("%s: stripe %llu block: %d\n",
687 __func__, (unsigned long long)sh->sector, target);
688 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
690 for (i = disks; i--; )
692 xor_srcs[count++] = sh->dev[i].page;
694 atomic_inc(&sh->count);
696 init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST, NULL,
697 ops_complete_compute, sh, to_addr_conv(sh, percpu));
698 if (unlikely(count == 1))
699 tx = async_memcpy(xor_dest, xor_srcs[0], 0, 0, STRIPE_SIZE, &submit);
701 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
706 /* set_syndrome_sources - populate source buffers for gen_syndrome
707 * @srcs - (struct page *) array of size sh->disks
708 * @sh - stripe_head to parse
710 * Populates srcs in proper layout order for the stripe and returns the
711 * 'count' of sources to be used in a call to async_gen_syndrome. The P
712 * destination buffer is recorded in srcs[count] and the Q destination
713 * is recorded in srcs[count+1]].
715 static int set_syndrome_sources(struct page **srcs, struct stripe_head *sh)
717 int disks = sh->disks;
718 int syndrome_disks = sh->ddf_layout ? disks : (disks - 2);
719 int d0_idx = raid6_d0(sh);
723 for (i = 0; i < disks; i++)
729 int slot = raid6_idx_to_slot(i, sh, &count, syndrome_disks);
731 srcs[slot] = sh->dev[i].page;
732 i = raid6_next_disk(i, disks);
733 } while (i != d0_idx);
735 return syndrome_disks;
738 static struct dma_async_tx_descriptor *
739 ops_run_compute6_1(struct stripe_head *sh, struct raid5_percpu *percpu)
741 int disks = sh->disks;
742 struct page **blocks = percpu->scribble;
744 int qd_idx = sh->qd_idx;
745 struct dma_async_tx_descriptor *tx;
746 struct async_submit_ctl submit;
752 if (sh->ops.target < 0)
753 target = sh->ops.target2;
754 else if (sh->ops.target2 < 0)
755 target = sh->ops.target;
757 /* we should only have one valid target */
760 pr_debug("%s: stripe %llu block: %d\n",
761 __func__, (unsigned long long)sh->sector, target);
763 tgt = &sh->dev[target];
764 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
767 atomic_inc(&sh->count);
769 if (target == qd_idx) {
770 count = set_syndrome_sources(blocks, sh);
771 blocks[count] = NULL; /* regenerating p is not necessary */
772 BUG_ON(blocks[count+1] != dest); /* q should already be set */
773 init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
774 ops_complete_compute, sh,
775 to_addr_conv(sh, percpu));
776 tx = async_gen_syndrome(blocks, 0, count+2, STRIPE_SIZE, &submit);
778 /* Compute any data- or p-drive using XOR */
780 for (i = disks; i-- ; ) {
781 if (i == target || i == qd_idx)
783 blocks[count++] = sh->dev[i].page;
786 init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST,
787 NULL, ops_complete_compute, sh,
788 to_addr_conv(sh, percpu));
789 tx = async_xor(dest, blocks, 0, count, STRIPE_SIZE, &submit);
795 static struct dma_async_tx_descriptor *
796 ops_run_compute6_2(struct stripe_head *sh, struct raid5_percpu *percpu)
798 int i, count, disks = sh->disks;
799 int syndrome_disks = sh->ddf_layout ? disks : disks-2;
800 int d0_idx = raid6_d0(sh);
801 int faila = -1, failb = -1;
802 int target = sh->ops.target;
803 int target2 = sh->ops.target2;
804 struct r5dev *tgt = &sh->dev[target];
805 struct r5dev *tgt2 = &sh->dev[target2];
806 struct dma_async_tx_descriptor *tx;
807 struct page **blocks = percpu->scribble;
808 struct async_submit_ctl submit;
810 pr_debug("%s: stripe %llu block1: %d block2: %d\n",
811 __func__, (unsigned long long)sh->sector, target, target2);
812 BUG_ON(target < 0 || target2 < 0);
813 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
814 BUG_ON(!test_bit(R5_Wantcompute, &tgt2->flags));
816 /* we need to open-code set_syndrome_sources to handle the
817 * slot number conversion for 'faila' and 'failb'
819 for (i = 0; i < disks ; i++)
824 int slot = raid6_idx_to_slot(i, sh, &count, syndrome_disks);
826 blocks[slot] = sh->dev[i].page;
832 i = raid6_next_disk(i, disks);
833 } while (i != d0_idx);
835 BUG_ON(faila == failb);
838 pr_debug("%s: stripe: %llu faila: %d failb: %d\n",
839 __func__, (unsigned long long)sh->sector, faila, failb);
841 atomic_inc(&sh->count);
843 if (failb == syndrome_disks+1) {
844 /* Q disk is one of the missing disks */
845 if (faila == syndrome_disks) {
846 /* Missing P+Q, just recompute */
847 init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
848 ops_complete_compute, sh,
849 to_addr_conv(sh, percpu));
850 return async_gen_syndrome(blocks, 0, syndrome_disks+2,
851 STRIPE_SIZE, &submit);
855 int qd_idx = sh->qd_idx;
857 /* Missing D+Q: recompute D from P, then recompute Q */
858 if (target == qd_idx)
859 data_target = target2;
861 data_target = target;
864 for (i = disks; i-- ; ) {
865 if (i == data_target || i == qd_idx)
867 blocks[count++] = sh->dev[i].page;
869 dest = sh->dev[data_target].page;
870 init_async_submit(&submit,
871 ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST,
873 to_addr_conv(sh, percpu));
874 tx = async_xor(dest, blocks, 0, count, STRIPE_SIZE,
877 count = set_syndrome_sources(blocks, sh);
878 init_async_submit(&submit, ASYNC_TX_FENCE, tx,
879 ops_complete_compute, sh,
880 to_addr_conv(sh, percpu));
881 return async_gen_syndrome(blocks, 0, count+2,
882 STRIPE_SIZE, &submit);
885 init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
886 ops_complete_compute, sh,
887 to_addr_conv(sh, percpu));
888 if (failb == syndrome_disks) {
889 /* We're missing D+P. */
890 return async_raid6_datap_recov(syndrome_disks+2,
894 /* We're missing D+D. */
895 return async_raid6_2data_recov(syndrome_disks+2,
896 STRIPE_SIZE, faila, failb,
903 static void ops_complete_prexor(void *stripe_head_ref)
905 struct stripe_head *sh = stripe_head_ref;
907 pr_debug("%s: stripe %llu\n", __func__,
908 (unsigned long long)sh->sector);
911 static struct dma_async_tx_descriptor *
912 ops_run_prexor(struct stripe_head *sh, struct raid5_percpu *percpu,
913 struct dma_async_tx_descriptor *tx)
915 int disks = sh->disks;
916 struct page **xor_srcs = percpu->scribble;
917 int count = 0, pd_idx = sh->pd_idx, i;
918 struct async_submit_ctl submit;
920 /* existing parity data subtracted */
921 struct page *xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
923 pr_debug("%s: stripe %llu\n", __func__,
924 (unsigned long long)sh->sector);
926 for (i = disks; i--; ) {
927 struct r5dev *dev = &sh->dev[i];
928 /* Only process blocks that are known to be uptodate */
929 if (test_bit(R5_Wantdrain, &dev->flags))
930 xor_srcs[count++] = dev->page;
933 init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_DROP_DST, tx,
934 ops_complete_prexor, sh, to_addr_conv(sh, percpu));
935 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
940 static struct dma_async_tx_descriptor *
941 ops_run_biodrain(struct stripe_head *sh, struct dma_async_tx_descriptor *tx)
943 int disks = sh->disks;
946 pr_debug("%s: stripe %llu\n", __func__,
947 (unsigned long long)sh->sector);
949 for (i = disks; i--; ) {
950 struct r5dev *dev = &sh->dev[i];
953 if (test_and_clear_bit(R5_Wantdrain, &dev->flags)) {
956 spin_lock(&sh->lock);
957 chosen = dev->towrite;
959 BUG_ON(dev->written);
960 wbi = dev->written = chosen;
961 spin_unlock(&sh->lock);
963 while (wbi && wbi->bi_sector <
964 dev->sector + STRIPE_SECTORS) {
965 tx = async_copy_data(1, wbi, dev->page,
967 wbi = r5_next_bio(wbi, dev->sector);
975 static void ops_complete_reconstruct(void *stripe_head_ref)
977 struct stripe_head *sh = stripe_head_ref;
978 int disks = sh->disks;
979 int pd_idx = sh->pd_idx;
980 int qd_idx = sh->qd_idx;
983 pr_debug("%s: stripe %llu\n", __func__,
984 (unsigned long long)sh->sector);
986 for (i = disks; i--; ) {
987 struct r5dev *dev = &sh->dev[i];
989 if (dev->written || i == pd_idx || i == qd_idx)
990 set_bit(R5_UPTODATE, &dev->flags);
993 if (sh->reconstruct_state == reconstruct_state_drain_run)
994 sh->reconstruct_state = reconstruct_state_drain_result;
995 else if (sh->reconstruct_state == reconstruct_state_prexor_drain_run)
996 sh->reconstruct_state = reconstruct_state_prexor_drain_result;
998 BUG_ON(sh->reconstruct_state != reconstruct_state_run);
999 sh->reconstruct_state = reconstruct_state_result;
1002 set_bit(STRIPE_HANDLE, &sh->state);
1007 ops_run_reconstruct5(struct stripe_head *sh, struct raid5_percpu *percpu,
1008 struct dma_async_tx_descriptor *tx)
1010 int disks = sh->disks;
1011 struct page **xor_srcs = percpu->scribble;
1012 struct async_submit_ctl submit;
1013 int count = 0, pd_idx = sh->pd_idx, i;
1014 struct page *xor_dest;
1016 unsigned long flags;
1018 pr_debug("%s: stripe %llu\n", __func__,
1019 (unsigned long long)sh->sector);
1021 /* check if prexor is active which means only process blocks
1022 * that are part of a read-modify-write (written)
1024 if (sh->reconstruct_state == reconstruct_state_prexor_drain_run) {
1026 xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
1027 for (i = disks; i--; ) {
1028 struct r5dev *dev = &sh->dev[i];
1030 xor_srcs[count++] = dev->page;
1033 xor_dest = sh->dev[pd_idx].page;
1034 for (i = disks; i--; ) {
1035 struct r5dev *dev = &sh->dev[i];
1037 xor_srcs[count++] = dev->page;
1041 /* 1/ if we prexor'd then the dest is reused as a source
1042 * 2/ if we did not prexor then we are redoing the parity
1043 * set ASYNC_TX_XOR_DROP_DST and ASYNC_TX_XOR_ZERO_DST
1044 * for the synchronous xor case
1046 flags = ASYNC_TX_ACK |
1047 (prexor ? ASYNC_TX_XOR_DROP_DST : ASYNC_TX_XOR_ZERO_DST);
1049 atomic_inc(&sh->count);
1051 init_async_submit(&submit, flags, tx, ops_complete_reconstruct, sh,
1052 to_addr_conv(sh, percpu));
1053 if (unlikely(count == 1))
1054 tx = async_memcpy(xor_dest, xor_srcs[0], 0, 0, STRIPE_SIZE, &submit);
1056 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
1060 ops_run_reconstruct6(struct stripe_head *sh, struct raid5_percpu *percpu,
1061 struct dma_async_tx_descriptor *tx)
1063 struct async_submit_ctl submit;
1064 struct page **blocks = percpu->scribble;
1067 pr_debug("%s: stripe %llu\n", __func__, (unsigned long long)sh->sector);
1069 count = set_syndrome_sources(blocks, sh);
1071 atomic_inc(&sh->count);
1073 init_async_submit(&submit, ASYNC_TX_ACK, tx, ops_complete_reconstruct,
1074 sh, to_addr_conv(sh, percpu));
1075 async_gen_syndrome(blocks, 0, count+2, STRIPE_SIZE, &submit);
1078 static void ops_complete_check(void *stripe_head_ref)
1080 struct stripe_head *sh = stripe_head_ref;
1082 pr_debug("%s: stripe %llu\n", __func__,
1083 (unsigned long long)sh->sector);
1085 sh->check_state = check_state_check_result;
1086 set_bit(STRIPE_HANDLE, &sh->state);
1090 static void ops_run_check_p(struct stripe_head *sh, struct raid5_percpu *percpu)
1092 int disks = sh->disks;
1093 int pd_idx = sh->pd_idx;
1094 int qd_idx = sh->qd_idx;
1095 struct page *xor_dest;
1096 struct page **xor_srcs = percpu->scribble;
1097 struct dma_async_tx_descriptor *tx;
1098 struct async_submit_ctl submit;
1102 pr_debug("%s: stripe %llu\n", __func__,
1103 (unsigned long long)sh->sector);
1106 xor_dest = sh->dev[pd_idx].page;
1107 xor_srcs[count++] = xor_dest;
1108 for (i = disks; i--; ) {
1109 if (i == pd_idx || i == qd_idx)
1111 xor_srcs[count++] = sh->dev[i].page;
1114 init_async_submit(&submit, 0, NULL, NULL, NULL,
1115 to_addr_conv(sh, percpu));
1116 tx = async_xor_val(xor_dest, xor_srcs, 0, count, STRIPE_SIZE,
1117 &sh->ops.zero_sum_result, &submit);
1119 atomic_inc(&sh->count);
1120 init_async_submit(&submit, ASYNC_TX_ACK, tx, ops_complete_check, sh, NULL);
1121 tx = async_trigger_callback(&submit);
1124 static void ops_run_check_pq(struct stripe_head *sh, struct raid5_percpu *percpu, int checkp)
1126 struct page **srcs = percpu->scribble;
1127 struct async_submit_ctl submit;
1130 pr_debug("%s: stripe %llu checkp: %d\n", __func__,
1131 (unsigned long long)sh->sector, checkp);
1133 count = set_syndrome_sources(srcs, sh);
1137 atomic_inc(&sh->count);
1138 init_async_submit(&submit, ASYNC_TX_ACK, NULL, ops_complete_check,
1139 sh, to_addr_conv(sh, percpu));
1140 async_syndrome_val(srcs, 0, count+2, STRIPE_SIZE,
1141 &sh->ops.zero_sum_result, percpu->spare_page, &submit);
1144 static void __raid_run_ops(struct stripe_head *sh, unsigned long ops_request)
1146 int overlap_clear = 0, i, disks = sh->disks;
1147 struct dma_async_tx_descriptor *tx = NULL;
1148 raid5_conf_t *conf = sh->raid_conf;
1149 int level = conf->level;
1150 struct raid5_percpu *percpu;
1154 percpu = per_cpu_ptr(conf->percpu, cpu);
1155 if (test_bit(STRIPE_OP_BIOFILL, &ops_request)) {
1156 ops_run_biofill(sh);
1160 if (test_bit(STRIPE_OP_COMPUTE_BLK, &ops_request)) {
1162 tx = ops_run_compute5(sh, percpu);
1164 if (sh->ops.target2 < 0 || sh->ops.target < 0)
1165 tx = ops_run_compute6_1(sh, percpu);
1167 tx = ops_run_compute6_2(sh, percpu);
1169 /* terminate the chain if reconstruct is not set to be run */
1170 if (tx && !test_bit(STRIPE_OP_RECONSTRUCT, &ops_request))
1174 if (test_bit(STRIPE_OP_PREXOR, &ops_request))
1175 tx = ops_run_prexor(sh, percpu, tx);
1177 if (test_bit(STRIPE_OP_BIODRAIN, &ops_request)) {
1178 tx = ops_run_biodrain(sh, tx);
1182 if (test_bit(STRIPE_OP_RECONSTRUCT, &ops_request)) {
1184 ops_run_reconstruct5(sh, percpu, tx);
1186 ops_run_reconstruct6(sh, percpu, tx);
1189 if (test_bit(STRIPE_OP_CHECK, &ops_request)) {
1190 if (sh->check_state == check_state_run)
1191 ops_run_check_p(sh, percpu);
1192 else if (sh->check_state == check_state_run_q)
1193 ops_run_check_pq(sh, percpu, 0);
1194 else if (sh->check_state == check_state_run_pq)
1195 ops_run_check_pq(sh, percpu, 1);
1201 for (i = disks; i--; ) {
1202 struct r5dev *dev = &sh->dev[i];
1203 if (test_and_clear_bit(R5_Overlap, &dev->flags))
1204 wake_up(&sh->raid_conf->wait_for_overlap);
1209 #ifdef CONFIG_MULTICORE_RAID456
1210 static void async_run_ops(void *param, async_cookie_t cookie)
1212 struct stripe_head *sh = param;
1213 unsigned long ops_request = sh->ops.request;
1215 clear_bit_unlock(STRIPE_OPS_REQ_PENDING, &sh->state);
1216 wake_up(&sh->ops.wait_for_ops);
1218 __raid_run_ops(sh, ops_request);
1222 static void raid_run_ops(struct stripe_head *sh, unsigned long ops_request)
1224 /* since handle_stripe can be called outside of raid5d context
1225 * we need to ensure sh->ops.request is de-staged before another
1228 wait_event(sh->ops.wait_for_ops,
1229 !test_and_set_bit_lock(STRIPE_OPS_REQ_PENDING, &sh->state));
1230 sh->ops.request = ops_request;
1232 atomic_inc(&sh->count);
1233 async_schedule(async_run_ops, sh);
1236 #define raid_run_ops __raid_run_ops
1239 static int grow_one_stripe(raid5_conf_t *conf)
1241 struct stripe_head *sh;
1242 int disks = max(conf->raid_disks, conf->previous_raid_disks);
1243 sh = kmem_cache_alloc(conf->slab_cache, GFP_KERNEL);
1246 memset(sh, 0, sizeof(*sh) + (disks-1)*sizeof(struct r5dev));
1247 sh->raid_conf = conf;
1248 spin_lock_init(&sh->lock);
1249 #ifdef CONFIG_MULTICORE_RAID456
1250 init_waitqueue_head(&sh->ops.wait_for_ops);
1253 if (grow_buffers(sh, disks)) {
1254 shrink_buffers(sh, disks);
1255 kmem_cache_free(conf->slab_cache, sh);
1258 /* we just created an active stripe so... */
1259 atomic_set(&sh->count, 1);
1260 atomic_inc(&conf->active_stripes);
1261 INIT_LIST_HEAD(&sh->lru);
1266 static int grow_stripes(raid5_conf_t *conf, int num)
1268 struct kmem_cache *sc;
1269 int devs = max(conf->raid_disks, conf->previous_raid_disks);
1271 sprintf(conf->cache_name[0],
1272 "raid%d-%s", conf->level, mdname(conf->mddev));
1273 sprintf(conf->cache_name[1],
1274 "raid%d-%s-alt", conf->level, mdname(conf->mddev));
1275 conf->active_name = 0;
1276 sc = kmem_cache_create(conf->cache_name[conf->active_name],
1277 sizeof(struct stripe_head)+(devs-1)*sizeof(struct r5dev),
1281 conf->slab_cache = sc;
1282 conf->pool_size = devs;
1284 if (!grow_one_stripe(conf))
1290 * scribble_len - return the required size of the scribble region
1291 * @num - total number of disks in the array
1293 * The size must be enough to contain:
1294 * 1/ a struct page pointer for each device in the array +2
1295 * 2/ room to convert each entry in (1) to its corresponding dma
1296 * (dma_map_page()) or page (page_address()) address.
1298 * Note: the +2 is for the destination buffers of the ddf/raid6 case where we
1299 * calculate over all devices (not just the data blocks), using zeros in place
1300 * of the P and Q blocks.
1302 static size_t scribble_len(int num)
1306 len = sizeof(struct page *) * (num+2) + sizeof(addr_conv_t) * (num+2);
1311 static int resize_stripes(raid5_conf_t *conf, int newsize)
1313 /* Make all the stripes able to hold 'newsize' devices.
1314 * New slots in each stripe get 'page' set to a new page.
1316 * This happens in stages:
1317 * 1/ create a new kmem_cache and allocate the required number of
1319 * 2/ gather all the old stripe_heads and tranfer the pages across
1320 * to the new stripe_heads. This will have the side effect of
1321 * freezing the array as once all stripe_heads have been collected,
1322 * no IO will be possible. Old stripe heads are freed once their
1323 * pages have been transferred over, and the old kmem_cache is
1324 * freed when all stripes are done.
1325 * 3/ reallocate conf->disks to be suitable bigger. If this fails,
1326 * we simple return a failre status - no need to clean anything up.
1327 * 4/ allocate new pages for the new slots in the new stripe_heads.
1328 * If this fails, we don't bother trying the shrink the
1329 * stripe_heads down again, we just leave them as they are.
1330 * As each stripe_head is processed the new one is released into
1333 * Once step2 is started, we cannot afford to wait for a write,
1334 * so we use GFP_NOIO allocations.
1336 struct stripe_head *osh, *nsh;
1337 LIST_HEAD(newstripes);
1338 struct disk_info *ndisks;
1341 struct kmem_cache *sc;
1344 if (newsize <= conf->pool_size)
1345 return 0; /* never bother to shrink */
1347 err = md_allow_write(conf->mddev);
1352 sc = kmem_cache_create(conf->cache_name[1-conf->active_name],
1353 sizeof(struct stripe_head)+(newsize-1)*sizeof(struct r5dev),
1358 for (i = conf->max_nr_stripes; i; i--) {
1359 nsh = kmem_cache_alloc(sc, GFP_KERNEL);
1363 memset(nsh, 0, sizeof(*nsh) + (newsize-1)*sizeof(struct r5dev));
1365 nsh->raid_conf = conf;
1366 spin_lock_init(&nsh->lock);
1367 #ifdef CONFIG_MULTICORE_RAID456
1368 init_waitqueue_head(&nsh->ops.wait_for_ops);
1371 list_add(&nsh->lru, &newstripes);
1374 /* didn't get enough, give up */
1375 while (!list_empty(&newstripes)) {
1376 nsh = list_entry(newstripes.next, struct stripe_head, lru);
1377 list_del(&nsh->lru);
1378 kmem_cache_free(sc, nsh);
1380 kmem_cache_destroy(sc);
1383 /* Step 2 - Must use GFP_NOIO now.
1384 * OK, we have enough stripes, start collecting inactive
1385 * stripes and copying them over
1387 list_for_each_entry(nsh, &newstripes, lru) {
1388 spin_lock_irq(&conf->device_lock);
1389 wait_event_lock_irq(conf->wait_for_stripe,
1390 !list_empty(&conf->inactive_list),
1392 unplug_slaves(conf->mddev)
1394 osh = get_free_stripe(conf);
1395 spin_unlock_irq(&conf->device_lock);
1396 atomic_set(&nsh->count, 1);
1397 for(i=0; i<conf->pool_size; i++)
1398 nsh->dev[i].page = osh->dev[i].page;
1399 for( ; i<newsize; i++)
1400 nsh->dev[i].page = NULL;
1401 kmem_cache_free(conf->slab_cache, osh);
1403 kmem_cache_destroy(conf->slab_cache);
1406 * At this point, we are holding all the stripes so the array
1407 * is completely stalled, so now is a good time to resize
1408 * conf->disks and the scribble region
1410 ndisks = kzalloc(newsize * sizeof(struct disk_info), GFP_NOIO);
1412 for (i=0; i<conf->raid_disks; i++)
1413 ndisks[i] = conf->disks[i];
1415 conf->disks = ndisks;
1420 conf->scribble_len = scribble_len(newsize);
1421 for_each_present_cpu(cpu) {
1422 struct raid5_percpu *percpu;
1425 percpu = per_cpu_ptr(conf->percpu, cpu);
1426 scribble = kmalloc(conf->scribble_len, GFP_NOIO);
1429 kfree(percpu->scribble);
1430 percpu->scribble = scribble;
1438 /* Step 4, return new stripes to service */
1439 while(!list_empty(&newstripes)) {
1440 nsh = list_entry(newstripes.next, struct stripe_head, lru);
1441 list_del_init(&nsh->lru);
1443 for (i=conf->raid_disks; i < newsize; i++)
1444 if (nsh->dev[i].page == NULL) {
1445 struct page *p = alloc_page(GFP_NOIO);
1446 nsh->dev[i].page = p;
1450 release_stripe(nsh);
1452 /* critical section pass, GFP_NOIO no longer needed */
1454 conf->slab_cache = sc;
1455 conf->active_name = 1-conf->active_name;
1456 conf->pool_size = newsize;
1460 static int drop_one_stripe(raid5_conf_t *conf)
1462 struct stripe_head *sh;
1464 spin_lock_irq(&conf->device_lock);
1465 sh = get_free_stripe(conf);
1466 spin_unlock_irq(&conf->device_lock);
1469 BUG_ON(atomic_read(&sh->count));
1470 shrink_buffers(sh, conf->pool_size);
1471 kmem_cache_free(conf->slab_cache, sh);
1472 atomic_dec(&conf->active_stripes);
1476 static void shrink_stripes(raid5_conf_t *conf)
1478 while (drop_one_stripe(conf))
1481 if (conf->slab_cache)
1482 kmem_cache_destroy(conf->slab_cache);
1483 conf->slab_cache = NULL;
1486 static void raid5_end_read_request(struct bio * bi, int error)
1488 struct stripe_head *sh = bi->bi_private;
1489 raid5_conf_t *conf = sh->raid_conf;
1490 int disks = sh->disks, i;
1491 int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
1492 char b[BDEVNAME_SIZE];
1496 for (i=0 ; i<disks; i++)
1497 if (bi == &sh->dev[i].req)
1500 pr_debug("end_read_request %llu/%d, count: %d, uptodate %d.\n",
1501 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
1509 set_bit(R5_UPTODATE, &sh->dev[i].flags);
1510 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
1511 rdev = conf->disks[i].rdev;
1512 printk_rl(KERN_INFO "md/raid:%s: read error corrected"
1513 " (%lu sectors at %llu on %s)\n",
1514 mdname(conf->mddev), STRIPE_SECTORS,
1515 (unsigned long long)(sh->sector
1516 + rdev->data_offset),
1517 bdevname(rdev->bdev, b));
1518 clear_bit(R5_ReadError, &sh->dev[i].flags);
1519 clear_bit(R5_ReWrite, &sh->dev[i].flags);
1521 if (atomic_read(&conf->disks[i].rdev->read_errors))
1522 atomic_set(&conf->disks[i].rdev->read_errors, 0);
1524 const char *bdn = bdevname(conf->disks[i].rdev->bdev, b);
1526 rdev = conf->disks[i].rdev;
1528 clear_bit(R5_UPTODATE, &sh->dev[i].flags);
1529 atomic_inc(&rdev->read_errors);
1530 if (conf->mddev->degraded >= conf->max_degraded)
1531 printk_rl(KERN_WARNING
1532 "md/raid:%s: read error not correctable "
1533 "(sector %llu on %s).\n",
1534 mdname(conf->mddev),
1535 (unsigned long long)(sh->sector
1536 + rdev->data_offset),
1538 else if (test_bit(R5_ReWrite, &sh->dev[i].flags))
1540 printk_rl(KERN_WARNING
1541 "md/raid:%s: read error NOT corrected!! "
1542 "(sector %llu on %s).\n",
1543 mdname(conf->mddev),
1544 (unsigned long long)(sh->sector
1545 + rdev->data_offset),
1547 else if (atomic_read(&rdev->read_errors)
1548 > conf->max_nr_stripes)
1550 "md/raid:%s: Too many read errors, failing device %s.\n",
1551 mdname(conf->mddev), bdn);
1555 set_bit(R5_ReadError, &sh->dev[i].flags);
1557 clear_bit(R5_ReadError, &sh->dev[i].flags);
1558 clear_bit(R5_ReWrite, &sh->dev[i].flags);
1559 md_error(conf->mddev, rdev);
1562 rdev_dec_pending(conf->disks[i].rdev, conf->mddev);
1563 clear_bit(R5_LOCKED, &sh->dev[i].flags);
1564 set_bit(STRIPE_HANDLE, &sh->state);
1568 static void raid5_end_write_request(struct bio *bi, int error)
1570 struct stripe_head *sh = bi->bi_private;
1571 raid5_conf_t *conf = sh->raid_conf;
1572 int disks = sh->disks, i;
1573 int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
1575 for (i=0 ; i<disks; i++)
1576 if (bi == &sh->dev[i].req)
1579 pr_debug("end_write_request %llu/%d, count %d, uptodate: %d.\n",
1580 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
1588 md_error(conf->mddev, conf->disks[i].rdev);
1590 rdev_dec_pending(conf->disks[i].rdev, conf->mddev);
1592 clear_bit(R5_LOCKED, &sh->dev[i].flags);
1593 set_bit(STRIPE_HANDLE, &sh->state);
1598 static sector_t compute_blocknr(struct stripe_head *sh, int i, int previous);
1600 static void raid5_build_block(struct stripe_head *sh, int i, int previous)
1602 struct r5dev *dev = &sh->dev[i];
1604 bio_init(&dev->req);
1605 dev->req.bi_io_vec = &dev->vec;
1607 dev->req.bi_max_vecs++;
1608 dev->vec.bv_page = dev->page;
1609 dev->vec.bv_len = STRIPE_SIZE;
1610 dev->vec.bv_offset = 0;
1612 dev->req.bi_sector = sh->sector;
1613 dev->req.bi_private = sh;
1616 dev->sector = compute_blocknr(sh, i, previous);
1619 static void error(mddev_t *mddev, mdk_rdev_t *rdev)
1621 char b[BDEVNAME_SIZE];
1622 raid5_conf_t *conf = mddev->private;
1623 pr_debug("raid456: error called\n");
1625 if (!test_bit(Faulty, &rdev->flags)) {
1626 set_bit(MD_CHANGE_DEVS, &mddev->flags);
1627 if (test_and_clear_bit(In_sync, &rdev->flags)) {
1628 unsigned long flags;
1629 spin_lock_irqsave(&conf->device_lock, flags);
1631 spin_unlock_irqrestore(&conf->device_lock, flags);
1633 * if recovery was running, make sure it aborts.
1635 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1637 set_bit(Faulty, &rdev->flags);
1639 "md/raid:%s: Disk failure on %s, disabling device.\n"
1641 "md/raid:%s: Operation continuing on %d devices.\n",
1643 bdevname(rdev->bdev, b),
1645 conf->raid_disks - mddev->degraded);
1650 * Input: a 'big' sector number,
1651 * Output: index of the data and parity disk, and the sector # in them.
1653 static sector_t raid5_compute_sector(raid5_conf_t *conf, sector_t r_sector,
1654 int previous, int *dd_idx,
1655 struct stripe_head *sh)
1658 unsigned long chunk_number;
1659 unsigned int chunk_offset;
1662 sector_t new_sector;
1663 int algorithm = previous ? conf->prev_algo
1665 int sectors_per_chunk = previous ? conf->prev_chunk_sectors
1666 : conf->chunk_sectors;
1667 int raid_disks = previous ? conf->previous_raid_disks
1669 int data_disks = raid_disks - conf->max_degraded;
1671 /* First compute the information on this sector */
1674 * Compute the chunk number and the sector offset inside the chunk
1676 chunk_offset = sector_div(r_sector, sectors_per_chunk);
1677 chunk_number = r_sector;
1678 BUG_ON(r_sector != chunk_number);
1681 * Compute the stripe number
1683 stripe = chunk_number / data_disks;
1686 * Compute the data disk and parity disk indexes inside the stripe
1688 *dd_idx = chunk_number % data_disks;
1691 * Select the parity disk based on the user selected algorithm.
1693 pd_idx = qd_idx = ~0;
1694 switch(conf->level) {
1696 pd_idx = data_disks;
1699 switch (algorithm) {
1700 case ALGORITHM_LEFT_ASYMMETRIC:
1701 pd_idx = data_disks - stripe % raid_disks;
1702 if (*dd_idx >= pd_idx)
1705 case ALGORITHM_RIGHT_ASYMMETRIC:
1706 pd_idx = stripe % raid_disks;
1707 if (*dd_idx >= pd_idx)
1710 case ALGORITHM_LEFT_SYMMETRIC:
1711 pd_idx = data_disks - stripe % raid_disks;
1712 *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
1714 case ALGORITHM_RIGHT_SYMMETRIC:
1715 pd_idx = stripe % raid_disks;
1716 *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
1718 case ALGORITHM_PARITY_0:
1722 case ALGORITHM_PARITY_N:
1723 pd_idx = data_disks;
1731 switch (algorithm) {
1732 case ALGORITHM_LEFT_ASYMMETRIC:
1733 pd_idx = raid_disks - 1 - (stripe % raid_disks);
1734 qd_idx = pd_idx + 1;
1735 if (pd_idx == raid_disks-1) {
1736 (*dd_idx)++; /* Q D D D P */
1738 } else if (*dd_idx >= pd_idx)
1739 (*dd_idx) += 2; /* D D P Q D */
1741 case ALGORITHM_RIGHT_ASYMMETRIC:
1742 pd_idx = stripe % raid_disks;
1743 qd_idx = pd_idx + 1;
1744 if (pd_idx == raid_disks-1) {
1745 (*dd_idx)++; /* Q D D D P */
1747 } else if (*dd_idx >= pd_idx)
1748 (*dd_idx) += 2; /* D D P Q D */
1750 case ALGORITHM_LEFT_SYMMETRIC:
1751 pd_idx = raid_disks - 1 - (stripe % raid_disks);
1752 qd_idx = (pd_idx + 1) % raid_disks;
1753 *dd_idx = (pd_idx + 2 + *dd_idx) % raid_disks;
1755 case ALGORITHM_RIGHT_SYMMETRIC:
1756 pd_idx = stripe % raid_disks;
1757 qd_idx = (pd_idx + 1) % raid_disks;
1758 *dd_idx = (pd_idx + 2 + *dd_idx) % raid_disks;
1761 case ALGORITHM_PARITY_0:
1766 case ALGORITHM_PARITY_N:
1767 pd_idx = data_disks;
1768 qd_idx = data_disks + 1;
1771 case ALGORITHM_ROTATING_ZERO_RESTART:
1772 /* Exactly the same as RIGHT_ASYMMETRIC, but or
1773 * of blocks for computing Q is different.
1775 pd_idx = stripe % raid_disks;
1776 qd_idx = pd_idx + 1;
1777 if (pd_idx == raid_disks-1) {
1778 (*dd_idx)++; /* Q D D D P */
1780 } else if (*dd_idx >= pd_idx)
1781 (*dd_idx) += 2; /* D D P Q D */
1785 case ALGORITHM_ROTATING_N_RESTART:
1786 /* Same a left_asymmetric, by first stripe is
1787 * D D D P Q rather than
1790 pd_idx = raid_disks - 1 - ((stripe + 1) % raid_disks);
1791 qd_idx = pd_idx + 1;
1792 if (pd_idx == raid_disks-1) {
1793 (*dd_idx)++; /* Q D D D P */
1795 } else if (*dd_idx >= pd_idx)
1796 (*dd_idx) += 2; /* D D P Q D */
1800 case ALGORITHM_ROTATING_N_CONTINUE:
1801 /* Same as left_symmetric but Q is before P */
1802 pd_idx = raid_disks - 1 - (stripe % raid_disks);
1803 qd_idx = (pd_idx + raid_disks - 1) % raid_disks;
1804 *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
1808 case ALGORITHM_LEFT_ASYMMETRIC_6:
1809 /* RAID5 left_asymmetric, with Q on last device */
1810 pd_idx = data_disks - stripe % (raid_disks-1);
1811 if (*dd_idx >= pd_idx)
1813 qd_idx = raid_disks - 1;
1816 case ALGORITHM_RIGHT_ASYMMETRIC_6:
1817 pd_idx = stripe % (raid_disks-1);
1818 if (*dd_idx >= pd_idx)
1820 qd_idx = raid_disks - 1;
1823 case ALGORITHM_LEFT_SYMMETRIC_6:
1824 pd_idx = data_disks - stripe % (raid_disks-1);
1825 *dd_idx = (pd_idx + 1 + *dd_idx) % (raid_disks-1);
1826 qd_idx = raid_disks - 1;
1829 case ALGORITHM_RIGHT_SYMMETRIC_6:
1830 pd_idx = stripe % (raid_disks-1);
1831 *dd_idx = (pd_idx + 1 + *dd_idx) % (raid_disks-1);
1832 qd_idx = raid_disks - 1;
1835 case ALGORITHM_PARITY_0_6:
1838 qd_idx = raid_disks - 1;
1848 sh->pd_idx = pd_idx;
1849 sh->qd_idx = qd_idx;
1850 sh->ddf_layout = ddf_layout;
1853 * Finally, compute the new sector number
1855 new_sector = (sector_t)stripe * sectors_per_chunk + chunk_offset;
1860 static sector_t compute_blocknr(struct stripe_head *sh, int i, int previous)
1862 raid5_conf_t *conf = sh->raid_conf;
1863 int raid_disks = sh->disks;
1864 int data_disks = raid_disks - conf->max_degraded;
1865 sector_t new_sector = sh->sector, check;
1866 int sectors_per_chunk = previous ? conf->prev_chunk_sectors
1867 : conf->chunk_sectors;
1868 int algorithm = previous ? conf->prev_algo
1872 int chunk_number, dummy1, dd_idx = i;
1874 struct stripe_head sh2;
1877 chunk_offset = sector_div(new_sector, sectors_per_chunk);
1878 stripe = new_sector;
1879 BUG_ON(new_sector != stripe);
1881 if (i == sh->pd_idx)
1883 switch(conf->level) {
1886 switch (algorithm) {
1887 case ALGORITHM_LEFT_ASYMMETRIC:
1888 case ALGORITHM_RIGHT_ASYMMETRIC:
1892 case ALGORITHM_LEFT_SYMMETRIC:
1893 case ALGORITHM_RIGHT_SYMMETRIC:
1896 i -= (sh->pd_idx + 1);
1898 case ALGORITHM_PARITY_0:
1901 case ALGORITHM_PARITY_N:
1908 if (i == sh->qd_idx)
1909 return 0; /* It is the Q disk */
1910 switch (algorithm) {
1911 case ALGORITHM_LEFT_ASYMMETRIC:
1912 case ALGORITHM_RIGHT_ASYMMETRIC:
1913 case ALGORITHM_ROTATING_ZERO_RESTART:
1914 case ALGORITHM_ROTATING_N_RESTART:
1915 if (sh->pd_idx == raid_disks-1)
1916 i--; /* Q D D D P */
1917 else if (i > sh->pd_idx)
1918 i -= 2; /* D D P Q D */
1920 case ALGORITHM_LEFT_SYMMETRIC:
1921 case ALGORITHM_RIGHT_SYMMETRIC:
1922 if (sh->pd_idx == raid_disks-1)
1923 i--; /* Q D D D P */
1928 i -= (sh->pd_idx + 2);
1931 case ALGORITHM_PARITY_0:
1934 case ALGORITHM_PARITY_N:
1936 case ALGORITHM_ROTATING_N_CONTINUE:
1937 /* Like left_symmetric, but P is before Q */
1938 if (sh->pd_idx == 0)
1939 i--; /* P D D D Q */
1944 i -= (sh->pd_idx + 1);
1947 case ALGORITHM_LEFT_ASYMMETRIC_6:
1948 case ALGORITHM_RIGHT_ASYMMETRIC_6:
1952 case ALGORITHM_LEFT_SYMMETRIC_6:
1953 case ALGORITHM_RIGHT_SYMMETRIC_6:
1955 i += data_disks + 1;
1956 i -= (sh->pd_idx + 1);
1958 case ALGORITHM_PARITY_0_6:
1967 chunk_number = stripe * data_disks + i;
1968 r_sector = (sector_t)chunk_number * sectors_per_chunk + chunk_offset;
1970 check = raid5_compute_sector(conf, r_sector,
1971 previous, &dummy1, &sh2);
1972 if (check != sh->sector || dummy1 != dd_idx || sh2.pd_idx != sh->pd_idx
1973 || sh2.qd_idx != sh->qd_idx) {
1974 printk(KERN_ERR "md/raid:%s: compute_blocknr: map not correct\n",
1975 mdname(conf->mddev));
1983 schedule_reconstruction(struct stripe_head *sh, struct stripe_head_state *s,
1984 int rcw, int expand)
1986 int i, pd_idx = sh->pd_idx, disks = sh->disks;
1987 raid5_conf_t *conf = sh->raid_conf;
1988 int level = conf->level;
1991 /* if we are not expanding this is a proper write request, and
1992 * there will be bios with new data to be drained into the
1996 sh->reconstruct_state = reconstruct_state_drain_run;
1997 set_bit(STRIPE_OP_BIODRAIN, &s->ops_request);
1999 sh->reconstruct_state = reconstruct_state_run;
2001 set_bit(STRIPE_OP_RECONSTRUCT, &s->ops_request);
2003 for (i = disks; i--; ) {
2004 struct r5dev *dev = &sh->dev[i];
2007 set_bit(R5_LOCKED, &dev->flags);
2008 set_bit(R5_Wantdrain, &dev->flags);
2010 clear_bit(R5_UPTODATE, &dev->flags);
2014 if (s->locked + conf->max_degraded == disks)
2015 if (!test_and_set_bit(STRIPE_FULL_WRITE, &sh->state))
2016 atomic_inc(&conf->pending_full_writes);
2019 BUG_ON(!(test_bit(R5_UPTODATE, &sh->dev[pd_idx].flags) ||
2020 test_bit(R5_Wantcompute, &sh->dev[pd_idx].flags)));
2022 sh->reconstruct_state = reconstruct_state_prexor_drain_run;
2023 set_bit(STRIPE_OP_PREXOR, &s->ops_request);
2024 set_bit(STRIPE_OP_BIODRAIN, &s->ops_request);
2025 set_bit(STRIPE_OP_RECONSTRUCT, &s->ops_request);
2027 for (i = disks; i--; ) {
2028 struct r5dev *dev = &sh->dev[i];
2033 (test_bit(R5_UPTODATE, &dev->flags) ||
2034 test_bit(R5_Wantcompute, &dev->flags))) {
2035 set_bit(R5_Wantdrain, &dev->flags);
2036 set_bit(R5_LOCKED, &dev->flags);
2037 clear_bit(R5_UPTODATE, &dev->flags);
2043 /* keep the parity disk(s) locked while asynchronous operations
2046 set_bit(R5_LOCKED, &sh->dev[pd_idx].flags);
2047 clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
2051 int qd_idx = sh->qd_idx;
2052 struct r5dev *dev = &sh->dev[qd_idx];
2054 set_bit(R5_LOCKED, &dev->flags);
2055 clear_bit(R5_UPTODATE, &dev->flags);
2059 pr_debug("%s: stripe %llu locked: %d ops_request: %lx\n",
2060 __func__, (unsigned long long)sh->sector,
2061 s->locked, s->ops_request);
2065 * Each stripe/dev can have one or more bion attached.
2066 * toread/towrite point to the first in a chain.
2067 * The bi_next chain must be in order.
2069 static int add_stripe_bio(struct stripe_head *sh, struct bio *bi, int dd_idx, int forwrite)
2072 raid5_conf_t *conf = sh->raid_conf;
2075 pr_debug("adding bh b#%llu to stripe s#%llu\n",
2076 (unsigned long long)bi->bi_sector,
2077 (unsigned long long)sh->sector);
2080 spin_lock(&sh->lock);
2081 spin_lock_irq(&conf->device_lock);
2083 bip = &sh->dev[dd_idx].towrite;
2084 if (*bip == NULL && sh->dev[dd_idx].written == NULL)
2087 bip = &sh->dev[dd_idx].toread;
2088 while (*bip && (*bip)->bi_sector < bi->bi_sector) {
2089 if ((*bip)->bi_sector + ((*bip)->bi_size >> 9) > bi->bi_sector)
2091 bip = & (*bip)->bi_next;
2093 if (*bip && (*bip)->bi_sector < bi->bi_sector + ((bi->bi_size)>>9))
2096 BUG_ON(*bip && bi->bi_next && (*bip) != bi->bi_next);
2100 bi->bi_phys_segments++;
2101 spin_unlock_irq(&conf->device_lock);
2102 spin_unlock(&sh->lock);
2104 pr_debug("added bi b#%llu to stripe s#%llu, disk %d.\n",
2105 (unsigned long long)bi->bi_sector,
2106 (unsigned long long)sh->sector, dd_idx);
2108 if (conf->mddev->bitmap && firstwrite) {
2109 bitmap_startwrite(conf->mddev->bitmap, sh->sector,
2111 sh->bm_seq = conf->seq_flush+1;
2112 set_bit(STRIPE_BIT_DELAY, &sh->state);
2116 /* check if page is covered */
2117 sector_t sector = sh->dev[dd_idx].sector;
2118 for (bi=sh->dev[dd_idx].towrite;
2119 sector < sh->dev[dd_idx].sector + STRIPE_SECTORS &&
2120 bi && bi->bi_sector <= sector;
2121 bi = r5_next_bio(bi, sh->dev[dd_idx].sector)) {
2122 if (bi->bi_sector + (bi->bi_size>>9) >= sector)
2123 sector = bi->bi_sector + (bi->bi_size>>9);
2125 if (sector >= sh->dev[dd_idx].sector + STRIPE_SECTORS)
2126 set_bit(R5_OVERWRITE, &sh->dev[dd_idx].flags);
2131 set_bit(R5_Overlap, &sh->dev[dd_idx].flags);
2132 spin_unlock_irq(&conf->device_lock);
2133 spin_unlock(&sh->lock);
2137 static void end_reshape(raid5_conf_t *conf);
2139 static void stripe_set_idx(sector_t stripe, raid5_conf_t *conf, int previous,
2140 struct stripe_head *sh)
2142 int sectors_per_chunk =
2143 previous ? conf->prev_chunk_sectors : conf->chunk_sectors;
2145 int chunk_offset = sector_div(stripe, sectors_per_chunk);
2146 int disks = previous ? conf->previous_raid_disks : conf->raid_disks;
2148 raid5_compute_sector(conf,
2149 stripe * (disks - conf->max_degraded)
2150 *sectors_per_chunk + chunk_offset,
2156 handle_failed_stripe(raid5_conf_t *conf, struct stripe_head *sh,
2157 struct stripe_head_state *s, int disks,
2158 struct bio **return_bi)
2161 for (i = disks; i--; ) {
2165 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
2168 rdev = rcu_dereference(conf->disks[i].rdev);
2169 if (rdev && test_bit(In_sync, &rdev->flags))
2170 /* multiple read failures in one stripe */
2171 md_error(conf->mddev, rdev);
2174 spin_lock_irq(&conf->device_lock);
2175 /* fail all writes first */
2176 bi = sh->dev[i].towrite;
2177 sh->dev[i].towrite = NULL;
2183 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
2184 wake_up(&conf->wait_for_overlap);
2186 while (bi && bi->bi_sector <
2187 sh->dev[i].sector + STRIPE_SECTORS) {
2188 struct bio *nextbi = r5_next_bio(bi, sh->dev[i].sector);
2189 clear_bit(BIO_UPTODATE, &bi->bi_flags);
2190 if (!raid5_dec_bi_phys_segments(bi)) {
2191 md_write_end(conf->mddev);
2192 bi->bi_next = *return_bi;
2197 /* and fail all 'written' */
2198 bi = sh->dev[i].written;
2199 sh->dev[i].written = NULL;
2200 if (bi) bitmap_end = 1;
2201 while (bi && bi->bi_sector <
2202 sh->dev[i].sector + STRIPE_SECTORS) {
2203 struct bio *bi2 = r5_next_bio(bi, sh->dev[i].sector);
2204 clear_bit(BIO_UPTODATE, &bi->bi_flags);
2205 if (!raid5_dec_bi_phys_segments(bi)) {
2206 md_write_end(conf->mddev);
2207 bi->bi_next = *return_bi;
2213 /* fail any reads if this device is non-operational and
2214 * the data has not reached the cache yet.
2216 if (!test_bit(R5_Wantfill, &sh->dev[i].flags) &&
2217 (!test_bit(R5_Insync, &sh->dev[i].flags) ||
2218 test_bit(R5_ReadError, &sh->dev[i].flags))) {
2219 bi = sh->dev[i].toread;
2220 sh->dev[i].toread = NULL;
2221 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
2222 wake_up(&conf->wait_for_overlap);
2223 if (bi) s->to_read--;
2224 while (bi && bi->bi_sector <
2225 sh->dev[i].sector + STRIPE_SECTORS) {
2226 struct bio *nextbi =
2227 r5_next_bio(bi, sh->dev[i].sector);
2228 clear_bit(BIO_UPTODATE, &bi->bi_flags);
2229 if (!raid5_dec_bi_phys_segments(bi)) {
2230 bi->bi_next = *return_bi;
2236 spin_unlock_irq(&conf->device_lock);
2238 bitmap_endwrite(conf->mddev->bitmap, sh->sector,
2239 STRIPE_SECTORS, 0, 0);
2242 if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
2243 if (atomic_dec_and_test(&conf->pending_full_writes))
2244 md_wakeup_thread(conf->mddev->thread);
2247 /* fetch_block5 - checks the given member device to see if its data needs
2248 * to be read or computed to satisfy a request.
2250 * Returns 1 when no more member devices need to be checked, otherwise returns
2251 * 0 to tell the loop in handle_stripe_fill5 to continue
2253 static int fetch_block5(struct stripe_head *sh, struct stripe_head_state *s,
2254 int disk_idx, int disks)
2256 struct r5dev *dev = &sh->dev[disk_idx];
2257 struct r5dev *failed_dev = &sh->dev[s->failed_num];
2259 /* is the data in this block needed, and can we get it? */
2260 if (!test_bit(R5_LOCKED, &dev->flags) &&
2261 !test_bit(R5_UPTODATE, &dev->flags) &&
2263 (dev->towrite && !test_bit(R5_OVERWRITE, &dev->flags)) ||
2264 s->syncing || s->expanding ||
2266 (failed_dev->toread ||
2267 (failed_dev->towrite &&
2268 !test_bit(R5_OVERWRITE, &failed_dev->flags)))))) {
2269 /* We would like to get this block, possibly by computing it,
2270 * otherwise read it if the backing disk is insync
2272 if ((s->uptodate == disks - 1) &&
2273 (s->failed && disk_idx == s->failed_num)) {
2274 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
2275 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
2276 set_bit(R5_Wantcompute, &dev->flags);
2277 sh->ops.target = disk_idx;
2278 sh->ops.target2 = -1;
2280 /* Careful: from this point on 'uptodate' is in the eye
2281 * of raid_run_ops which services 'compute' operations
2282 * before writes. R5_Wantcompute flags a block that will
2283 * be R5_UPTODATE by the time it is needed for a
2284 * subsequent operation.
2287 return 1; /* uptodate + compute == disks */
2288 } else if (test_bit(R5_Insync, &dev->flags)) {
2289 set_bit(R5_LOCKED, &dev->flags);
2290 set_bit(R5_Wantread, &dev->flags);
2292 pr_debug("Reading block %d (sync=%d)\n", disk_idx,
2301 * handle_stripe_fill5 - read or compute data to satisfy pending requests.
2303 static void handle_stripe_fill5(struct stripe_head *sh,
2304 struct stripe_head_state *s, int disks)
2308 /* look for blocks to read/compute, skip this if a compute
2309 * is already in flight, or if the stripe contents are in the
2310 * midst of changing due to a write
2312 if (!test_bit(STRIPE_COMPUTE_RUN, &sh->state) && !sh->check_state &&
2313 !sh->reconstruct_state)
2314 for (i = disks; i--; )
2315 if (fetch_block5(sh, s, i, disks))
2317 set_bit(STRIPE_HANDLE, &sh->state);
2320 /* fetch_block6 - checks the given member device to see if its data needs
2321 * to be read or computed to satisfy a request.
2323 * Returns 1 when no more member devices need to be checked, otherwise returns
2324 * 0 to tell the loop in handle_stripe_fill6 to continue
2326 static int fetch_block6(struct stripe_head *sh, struct stripe_head_state *s,
2327 struct r6_state *r6s, int disk_idx, int disks)
2329 struct r5dev *dev = &sh->dev[disk_idx];
2330 struct r5dev *fdev[2] = { &sh->dev[r6s->failed_num[0]],
2331 &sh->dev[r6s->failed_num[1]] };
2333 if (!test_bit(R5_LOCKED, &dev->flags) &&
2334 !test_bit(R5_UPTODATE, &dev->flags) &&
2336 (dev->towrite && !test_bit(R5_OVERWRITE, &dev->flags)) ||
2337 s->syncing || s->expanding ||
2339 (fdev[0]->toread || s->to_write)) ||
2341 (fdev[1]->toread || s->to_write)))) {
2342 /* we would like to get this block, possibly by computing it,
2343 * otherwise read it if the backing disk is insync
2345 BUG_ON(test_bit(R5_Wantcompute, &dev->flags));
2346 BUG_ON(test_bit(R5_Wantread, &dev->flags));
2347 if ((s->uptodate == disks - 1) &&
2348 (s->failed && (disk_idx == r6s->failed_num[0] ||
2349 disk_idx == r6s->failed_num[1]))) {
2350 /* have disk failed, and we're requested to fetch it;
2353 pr_debug("Computing stripe %llu block %d\n",
2354 (unsigned long long)sh->sector, disk_idx);
2355 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
2356 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
2357 set_bit(R5_Wantcompute, &dev->flags);
2358 sh->ops.target = disk_idx;
2359 sh->ops.target2 = -1; /* no 2nd target */
2363 } else if (s->uptodate == disks-2 && s->failed >= 2) {
2364 /* Computing 2-failure is *very* expensive; only
2365 * do it if failed >= 2
2368 for (other = disks; other--; ) {
2369 if (other == disk_idx)
2371 if (!test_bit(R5_UPTODATE,
2372 &sh->dev[other].flags))
2376 pr_debug("Computing stripe %llu blocks %d,%d\n",
2377 (unsigned long long)sh->sector,
2379 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
2380 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
2381 set_bit(R5_Wantcompute, &sh->dev[disk_idx].flags);
2382 set_bit(R5_Wantcompute, &sh->dev[other].flags);
2383 sh->ops.target = disk_idx;
2384 sh->ops.target2 = other;
2388 } else if (test_bit(R5_Insync, &dev->flags)) {
2389 set_bit(R5_LOCKED, &dev->flags);
2390 set_bit(R5_Wantread, &dev->flags);
2392 pr_debug("Reading block %d (sync=%d)\n",
2393 disk_idx, s->syncing);
2401 * handle_stripe_fill6 - read or compute data to satisfy pending requests.
2403 static void handle_stripe_fill6(struct stripe_head *sh,
2404 struct stripe_head_state *s, struct r6_state *r6s,
2409 /* look for blocks to read/compute, skip this if a compute
2410 * is already in flight, or if the stripe contents are in the
2411 * midst of changing due to a write
2413 if (!test_bit(STRIPE_COMPUTE_RUN, &sh->state) && !sh->check_state &&
2414 !sh->reconstruct_state)
2415 for (i = disks; i--; )
2416 if (fetch_block6(sh, s, r6s, i, disks))
2418 set_bit(STRIPE_HANDLE, &sh->state);
2422 /* handle_stripe_clean_event
2423 * any written block on an uptodate or failed drive can be returned.
2424 * Note that if we 'wrote' to a failed drive, it will be UPTODATE, but
2425 * never LOCKED, so we don't need to test 'failed' directly.
2427 static void handle_stripe_clean_event(raid5_conf_t *conf,
2428 struct stripe_head *sh, int disks, struct bio **return_bi)
2433 for (i = disks; i--; )
2434 if (sh->dev[i].written) {
2436 if (!test_bit(R5_LOCKED, &dev->flags) &&
2437 test_bit(R5_UPTODATE, &dev->flags)) {
2438 /* We can return any write requests */
2439 struct bio *wbi, *wbi2;
2441 pr_debug("Return write for disc %d\n", i);
2442 spin_lock_irq(&conf->device_lock);
2444 dev->written = NULL;
2445 while (wbi && wbi->bi_sector <
2446 dev->sector + STRIPE_SECTORS) {
2447 wbi2 = r5_next_bio(wbi, dev->sector);
2448 if (!raid5_dec_bi_phys_segments(wbi)) {
2449 md_write_end(conf->mddev);
2450 wbi->bi_next = *return_bi;
2455 if (dev->towrite == NULL)
2457 spin_unlock_irq(&conf->device_lock);
2459 bitmap_endwrite(conf->mddev->bitmap,
2462 !test_bit(STRIPE_DEGRADED, &sh->state),
2467 if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
2468 if (atomic_dec_and_test(&conf->pending_full_writes))
2469 md_wakeup_thread(conf->mddev->thread);
2472 static void handle_stripe_dirtying5(raid5_conf_t *conf,
2473 struct stripe_head *sh, struct stripe_head_state *s, int disks)
2475 int rmw = 0, rcw = 0, i;
2476 for (i = disks; i--; ) {
2477 /* would I have to read this buffer for read_modify_write */
2478 struct r5dev *dev = &sh->dev[i];
2479 if ((dev->towrite || i == sh->pd_idx) &&
2480 !test_bit(R5_LOCKED, &dev->flags) &&
2481 !(test_bit(R5_UPTODATE, &dev->flags) ||
2482 test_bit(R5_Wantcompute, &dev->flags))) {
2483 if (test_bit(R5_Insync, &dev->flags))
2486 rmw += 2*disks; /* cannot read it */
2488 /* Would I have to read this buffer for reconstruct_write */
2489 if (!test_bit(R5_OVERWRITE, &dev->flags) && i != sh->pd_idx &&
2490 !test_bit(R5_LOCKED, &dev->flags) &&
2491 !(test_bit(R5_UPTODATE, &dev->flags) ||
2492 test_bit(R5_Wantcompute, &dev->flags))) {
2493 if (test_bit(R5_Insync, &dev->flags)) rcw++;
2498 pr_debug("for sector %llu, rmw=%d rcw=%d\n",
2499 (unsigned long long)sh->sector, rmw, rcw);
2500 set_bit(STRIPE_HANDLE, &sh->state);
2501 if (rmw < rcw && rmw > 0)
2502 /* prefer read-modify-write, but need to get some data */
2503 for (i = disks; i--; ) {
2504 struct r5dev *dev = &sh->dev[i];
2505 if ((dev->towrite || i == sh->pd_idx) &&
2506 !test_bit(R5_LOCKED, &dev->flags) &&
2507 !(test_bit(R5_UPTODATE, &dev->flags) ||
2508 test_bit(R5_Wantcompute, &dev->flags)) &&
2509 test_bit(R5_Insync, &dev->flags)) {
2511 test_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2512 pr_debug("Read_old block "
2513 "%d for r-m-w\n", i);
2514 set_bit(R5_LOCKED, &dev->flags);
2515 set_bit(R5_Wantread, &dev->flags);
2518 set_bit(STRIPE_DELAYED, &sh->state);
2519 set_bit(STRIPE_HANDLE, &sh->state);
2523 if (rcw <= rmw && rcw > 0)
2524 /* want reconstruct write, but need to get some data */
2525 for (i = disks; i--; ) {
2526 struct r5dev *dev = &sh->dev[i];
2527 if (!test_bit(R5_OVERWRITE, &dev->flags) &&
2529 !test_bit(R5_LOCKED, &dev->flags) &&
2530 !(test_bit(R5_UPTODATE, &dev->flags) ||
2531 test_bit(R5_Wantcompute, &dev->flags)) &&
2532 test_bit(R5_Insync, &dev->flags)) {
2534 test_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2535 pr_debug("Read_old block "
2536 "%d for Reconstruct\n", i);
2537 set_bit(R5_LOCKED, &dev->flags);
2538 set_bit(R5_Wantread, &dev->flags);
2541 set_bit(STRIPE_DELAYED, &sh->state);
2542 set_bit(STRIPE_HANDLE, &sh->state);
2546 /* now if nothing is locked, and if we have enough data,
2547 * we can start a write request
2549 /* since handle_stripe can be called at any time we need to handle the
2550 * case where a compute block operation has been submitted and then a
2551 * subsequent call wants to start a write request. raid_run_ops only
2552 * handles the case where compute block and reconstruct are requested
2553 * simultaneously. If this is not the case then new writes need to be
2554 * held off until the compute completes.
2556 if ((s->req_compute || !test_bit(STRIPE_COMPUTE_RUN, &sh->state)) &&
2557 (s->locked == 0 && (rcw == 0 || rmw == 0) &&
2558 !test_bit(STRIPE_BIT_DELAY, &sh->state)))
2559 schedule_reconstruction(sh, s, rcw == 0, 0);
2562 static void handle_stripe_dirtying6(raid5_conf_t *conf,
2563 struct stripe_head *sh, struct stripe_head_state *s,
2564 struct r6_state *r6s, int disks)
2566 int rcw = 0, pd_idx = sh->pd_idx, i;
2567 int qd_idx = sh->qd_idx;
2569 set_bit(STRIPE_HANDLE, &sh->state);
2570 for (i = disks; i--; ) {
2571 struct r5dev *dev = &sh->dev[i];
2572 /* check if we haven't enough data */
2573 if (!test_bit(R5_OVERWRITE, &dev->flags) &&
2574 i != pd_idx && i != qd_idx &&
2575 !test_bit(R5_LOCKED, &dev->flags) &&
2576 !(test_bit(R5_UPTODATE, &dev->flags) ||
2577 test_bit(R5_Wantcompute, &dev->flags))) {
2579 if (!test_bit(R5_Insync, &dev->flags))
2580 continue; /* it's a failed drive */
2583 test_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2584 pr_debug("Read_old stripe %llu "
2585 "block %d for Reconstruct\n",
2586 (unsigned long long)sh->sector, i);
2587 set_bit(R5_LOCKED, &dev->flags);
2588 set_bit(R5_Wantread, &dev->flags);
2591 pr_debug("Request delayed stripe %llu "
2592 "block %d for Reconstruct\n",
2593 (unsigned long long)sh->sector, i);
2594 set_bit(STRIPE_DELAYED, &sh->state);
2595 set_bit(STRIPE_HANDLE, &sh->state);
2599 /* now if nothing is locked, and if we have enough data, we can start a
2602 if ((s->req_compute || !test_bit(STRIPE_COMPUTE_RUN, &sh->state)) &&
2603 s->locked == 0 && rcw == 0 &&
2604 !test_bit(STRIPE_BIT_DELAY, &sh->state)) {
2605 schedule_reconstruction(sh, s, 1, 0);
2609 static void handle_parity_checks5(raid5_conf_t *conf, struct stripe_head *sh,
2610 struct stripe_head_state *s, int disks)
2612 struct r5dev *dev = NULL;
2614 set_bit(STRIPE_HANDLE, &sh->state);
2616 switch (sh->check_state) {
2617 case check_state_idle:
2618 /* start a new check operation if there are no failures */
2619 if (s->failed == 0) {
2620 BUG_ON(s->uptodate != disks);
2621 sh->check_state = check_state_run;
2622 set_bit(STRIPE_OP_CHECK, &s->ops_request);
2623 clear_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags);
2627 dev = &sh->dev[s->failed_num];
2629 case check_state_compute_result:
2630 sh->check_state = check_state_idle;
2632 dev = &sh->dev[sh->pd_idx];
2634 /* check that a write has not made the stripe insync */
2635 if (test_bit(STRIPE_INSYNC, &sh->state))
2638 /* either failed parity check, or recovery is happening */
2639 BUG_ON(!test_bit(R5_UPTODATE, &dev->flags));
2640 BUG_ON(s->uptodate != disks);
2642 set_bit(R5_LOCKED, &dev->flags);
2644 set_bit(R5_Wantwrite, &dev->flags);
2646 clear_bit(STRIPE_DEGRADED, &sh->state);
2647 set_bit(STRIPE_INSYNC, &sh->state);
2649 case check_state_run:
2650 break; /* we will be called again upon completion */
2651 case check_state_check_result:
2652 sh->check_state = check_state_idle;
2654 /* if a failure occurred during the check operation, leave
2655 * STRIPE_INSYNC not set and let the stripe be handled again
2660 /* handle a successful check operation, if parity is correct
2661 * we are done. Otherwise update the mismatch count and repair
2662 * parity if !MD_RECOVERY_CHECK
2664 if ((sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) == 0)
2665 /* parity is correct (on disc,
2666 * not in buffer any more)
2668 set_bit(STRIPE_INSYNC, &sh->state);
2670 conf->mddev->resync_mismatches += STRIPE_SECTORS;
2671 if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery))
2672 /* don't try to repair!! */
2673 set_bit(STRIPE_INSYNC, &sh->state);
2675 sh->check_state = check_state_compute_run;
2676 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
2677 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
2678 set_bit(R5_Wantcompute,
2679 &sh->dev[sh->pd_idx].flags);
2680 sh->ops.target = sh->pd_idx;
2681 sh->ops.target2 = -1;
2686 case check_state_compute_run:
2689 printk(KERN_ERR "%s: unknown check_state: %d sector: %llu\n",
2690 __func__, sh->check_state,
2691 (unsigned long long) sh->sector);
2697 static void handle_parity_checks6(raid5_conf_t *conf, struct stripe_head *sh,
2698 struct stripe_head_state *s,
2699 struct r6_state *r6s, int disks)
2701 int pd_idx = sh->pd_idx;
2702 int qd_idx = sh->qd_idx;
2705 set_bit(STRIPE_HANDLE, &sh->state);
2707 BUG_ON(s->failed > 2);
2709 /* Want to check and possibly repair P and Q.
2710 * However there could be one 'failed' device, in which
2711 * case we can only check one of them, possibly using the
2712 * other to generate missing data
2715 switch (sh->check_state) {
2716 case check_state_idle:
2717 /* start a new check operation if there are < 2 failures */
2718 if (s->failed == r6s->q_failed) {
2719 /* The only possible failed device holds Q, so it
2720 * makes sense to check P (If anything else were failed,
2721 * we would have used P to recreate it).
2723 sh->check_state = check_state_run;
2725 if (!r6s->q_failed && s->failed < 2) {
2726 /* Q is not failed, and we didn't use it to generate
2727 * anything, so it makes sense to check it
2729 if (sh->check_state == check_state_run)
2730 sh->check_state = check_state_run_pq;
2732 sh->check_state = check_state_run_q;
2735 /* discard potentially stale zero_sum_result */
2736 sh->ops.zero_sum_result = 0;
2738 if (sh->check_state == check_state_run) {
2739 /* async_xor_zero_sum destroys the contents of P */
2740 clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
2743 if (sh->check_state >= check_state_run &&
2744 sh->check_state <= check_state_run_pq) {
2745 /* async_syndrome_zero_sum preserves P and Q, so
2746 * no need to mark them !uptodate here
2748 set_bit(STRIPE_OP_CHECK, &s->ops_request);
2752 /* we have 2-disk failure */
2753 BUG_ON(s->failed != 2);
2755 case check_state_compute_result:
2756 sh->check_state = check_state_idle;
2758 /* check that a write has not made the stripe insync */
2759 if (test_bit(STRIPE_INSYNC, &sh->state))
2762 /* now write out any block on a failed drive,
2763 * or P or Q if they were recomputed
2765 BUG_ON(s->uptodate < disks - 1); /* We don't need Q to recover */
2766 if (s->failed == 2) {
2767 dev = &sh->dev[r6s->failed_num[1]];
2769 set_bit(R5_LOCKED, &dev->flags);
2770 set_bit(R5_Wantwrite, &dev->flags);
2772 if (s->failed >= 1) {
2773 dev = &sh->dev[r6s->failed_num[0]];
2775 set_bit(R5_LOCKED, &dev->flags);
2776 set_bit(R5_Wantwrite, &dev->flags);
2778 if (sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) {
2779 dev = &sh->dev[pd_idx];
2781 set_bit(R5_LOCKED, &dev->flags);
2782 set_bit(R5_Wantwrite, &dev->flags);
2784 if (sh->ops.zero_sum_result & SUM_CHECK_Q_RESULT) {
2785 dev = &sh->dev[qd_idx];
2787 set_bit(R5_LOCKED, &dev->flags);
2788 set_bit(R5_Wantwrite, &dev->flags);
2790 clear_bit(STRIPE_DEGRADED, &sh->state);
2792 set_bit(STRIPE_INSYNC, &sh->state);
2794 case check_state_run:
2795 case check_state_run_q:
2796 case check_state_run_pq:
2797 break; /* we will be called again upon completion */
2798 case check_state_check_result:
2799 sh->check_state = check_state_idle;
2801 /* handle a successful check operation, if parity is correct
2802 * we are done. Otherwise update the mismatch count and repair
2803 * parity if !MD_RECOVERY_CHECK
2805 if (sh->ops.zero_sum_result == 0) {
2806 /* both parities are correct */
2808 set_bit(STRIPE_INSYNC, &sh->state);
2810 /* in contrast to the raid5 case we can validate
2811 * parity, but still have a failure to write
2814 sh->check_state = check_state_compute_result;
2815 /* Returning at this point means that we may go
2816 * off and bring p and/or q uptodate again so
2817 * we make sure to check zero_sum_result again
2818 * to verify if p or q need writeback
2822 conf->mddev->resync_mismatches += STRIPE_SECTORS;
2823 if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery))
2824 /* don't try to repair!! */
2825 set_bit(STRIPE_INSYNC, &sh->state);
2827 int *target = &sh->ops.target;
2829 sh->ops.target = -1;
2830 sh->ops.target2 = -1;
2831 sh->check_state = check_state_compute_run;
2832 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
2833 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
2834 if (sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) {
2835 set_bit(R5_Wantcompute,
2836 &sh->dev[pd_idx].flags);
2838 target = &sh->ops.target2;
2841 if (sh->ops.zero_sum_result & SUM_CHECK_Q_RESULT) {
2842 set_bit(R5_Wantcompute,
2843 &sh->dev[qd_idx].flags);
2850 case check_state_compute_run:
2853 printk(KERN_ERR "%s: unknown check_state: %d sector: %llu\n",
2854 __func__, sh->check_state,
2855 (unsigned long long) sh->sector);
2860 static void handle_stripe_expansion(raid5_conf_t *conf, struct stripe_head *sh,
2861 struct r6_state *r6s)
2865 /* We have read all the blocks in this stripe and now we need to
2866 * copy some of them into a target stripe for expand.
2868 struct dma_async_tx_descriptor *tx = NULL;
2869 clear_bit(STRIPE_EXPAND_SOURCE, &sh->state);
2870 for (i = 0; i < sh->disks; i++)
2871 if (i != sh->pd_idx && i != sh->qd_idx) {
2873 struct stripe_head *sh2;
2874 struct async_submit_ctl submit;
2876 sector_t bn = compute_blocknr(sh, i, 1);
2877 sector_t s = raid5_compute_sector(conf, bn, 0,
2879 sh2 = get_active_stripe(conf, s, 0, 1, 1);
2881 /* so far only the early blocks of this stripe
2882 * have been requested. When later blocks
2883 * get requested, we will try again
2886 if (!test_bit(STRIPE_EXPANDING, &sh2->state) ||
2887 test_bit(R5_Expanded, &sh2->dev[dd_idx].flags)) {
2888 /* must have already done this block */
2889 release_stripe(sh2);
2893 /* place all the copies on one channel */
2894 init_async_submit(&submit, 0, tx, NULL, NULL, NULL);
2895 tx = async_memcpy(sh2->dev[dd_idx].page,
2896 sh->dev[i].page, 0, 0, STRIPE_SIZE,
2899 set_bit(R5_Expanded, &sh2->dev[dd_idx].flags);
2900 set_bit(R5_UPTODATE, &sh2->dev[dd_idx].flags);
2901 for (j = 0; j < conf->raid_disks; j++)
2902 if (j != sh2->pd_idx &&
2903 (!r6s || j != sh2->qd_idx) &&
2904 !test_bit(R5_Expanded, &sh2->dev[j].flags))
2906 if (j == conf->raid_disks) {
2907 set_bit(STRIPE_EXPAND_READY, &sh2->state);
2908 set_bit(STRIPE_HANDLE, &sh2->state);
2910 release_stripe(sh2);
2913 /* done submitting copies, wait for them to complete */
2916 dma_wait_for_async_tx(tx);
2922 * handle_stripe - do things to a stripe.
2924 * We lock the stripe and then examine the state of various bits
2925 * to see what needs to be done.
2927 * return some read request which now have data
2928 * return some write requests which are safely on disc
2929 * schedule a read on some buffers
2930 * schedule a write of some buffers
2931 * return confirmation of parity correctness
2933 * buffers are taken off read_list or write_list, and bh_cache buffers
2934 * get BH_Lock set before the stripe lock is released.
2938 static void handle_stripe5(struct stripe_head *sh)
2940 raid5_conf_t *conf = sh->raid_conf;
2941 int disks = sh->disks, i;
2942 struct bio *return_bi = NULL;
2943 struct stripe_head_state s;
2945 mdk_rdev_t *blocked_rdev = NULL;
2947 int dec_preread_active = 0;
2949 memset(&s, 0, sizeof(s));
2950 pr_debug("handling stripe %llu, state=%#lx cnt=%d, pd_idx=%d check:%d "
2951 "reconstruct:%d\n", (unsigned long long)sh->sector, sh->state,
2952 atomic_read(&sh->count), sh->pd_idx, sh->check_state,
2953 sh->reconstruct_state);
2955 spin_lock(&sh->lock);
2956 clear_bit(STRIPE_HANDLE, &sh->state);
2957 clear_bit(STRIPE_DELAYED, &sh->state);
2959 s.syncing = test_bit(STRIPE_SYNCING, &sh->state);
2960 s.expanding = test_bit(STRIPE_EXPAND_SOURCE, &sh->state);
2961 s.expanded = test_bit(STRIPE_EXPAND_READY, &sh->state);
2963 /* Now to look around and see what can be done */
2965 for (i=disks; i--; ) {
2969 clear_bit(R5_Insync, &dev->flags);
2971 pr_debug("check %d: state 0x%lx toread %p read %p write %p "
2972 "written %p\n", i, dev->flags, dev->toread, dev->read,
2973 dev->towrite, dev->written);
2975 /* maybe we can request a biofill operation
2977 * new wantfill requests are only permitted while
2978 * ops_complete_biofill is guaranteed to be inactive
2980 if (test_bit(R5_UPTODATE, &dev->flags) && dev->toread &&
2981 !test_bit(STRIPE_BIOFILL_RUN, &sh->state))
2982 set_bit(R5_Wantfill, &dev->flags);
2984 /* now count some things */
2985 if (test_bit(R5_LOCKED, &dev->flags)) s.locked++;
2986 if (test_bit(R5_UPTODATE, &dev->flags)) s.uptodate++;
2987 if (test_bit(R5_Wantcompute, &dev->flags)) s.compute++;
2989 if (test_bit(R5_Wantfill, &dev->flags))
2991 else if (dev->toread)
2995 if (!test_bit(R5_OVERWRITE, &dev->flags))
3000 rdev = rcu_dereference(conf->disks[i].rdev);
3001 if (blocked_rdev == NULL &&
3002 rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
3003 blocked_rdev = rdev;
3004 atomic_inc(&rdev->nr_pending);
3006 if (!rdev || !test_bit(In_sync, &rdev->flags)) {
3007 /* The ReadError flag will just be confusing now */
3008 clear_bit(R5_ReadError, &dev->flags);
3009 clear_bit(R5_ReWrite, &dev->flags);
3011 if (!rdev || !test_bit(In_sync, &rdev->flags)
3012 || test_bit(R5_ReadError, &dev->flags)) {
3016 set_bit(R5_Insync, &dev->flags);
3020 if (unlikely(blocked_rdev)) {
3021 if (s.syncing || s.expanding || s.expanded ||
3022 s.to_write || s.written) {
3023 set_bit(STRIPE_HANDLE, &sh->state);
3026 /* There is nothing for the blocked_rdev to block */
3027 rdev_dec_pending(blocked_rdev, conf->mddev);
3028 blocked_rdev = NULL;
3031 if (s.to_fill && !test_bit(STRIPE_BIOFILL_RUN, &sh->state)) {
3032 set_bit(STRIPE_OP_BIOFILL, &s.ops_request);
3033 set_bit(STRIPE_BIOFILL_RUN, &sh->state);
3036 pr_debug("locked=%d uptodate=%d to_read=%d"
3037 " to_write=%d failed=%d failed_num=%d\n",
3038 s.locked, s.uptodate, s.to_read, s.to_write,
3039 s.failed, s.failed_num);
3040 /* check if the array has lost two devices and, if so, some requests might
3043 if (s.failed > 1 && s.to_read+s.to_write+s.written)
3044 handle_failed_stripe(conf, sh, &s, disks, &return_bi);
3045 if (s.failed > 1 && s.syncing) {
3046 md_done_sync(conf->mddev, STRIPE_SECTORS,0);
3047 clear_bit(STRIPE_SYNCING, &sh->state);
3051 /* might be able to return some write requests if the parity block
3052 * is safe, or on a failed drive
3054 dev = &sh->dev[sh->pd_idx];
3056 ((test_bit(R5_Insync, &dev->flags) &&
3057 !test_bit(R5_LOCKED, &dev->flags) &&
3058 test_bit(R5_UPTODATE, &dev->flags)) ||
3059 (s.failed == 1 && s.failed_num == sh->pd_idx)))
3060 handle_stripe_clean_event(conf, sh, disks, &return_bi);
3062 /* Now we might consider reading some blocks, either to check/generate
3063 * parity, or to satisfy requests
3064 * or to load a block that is being partially written.
3066 if (s.to_read || s.non_overwrite ||
3067 (s.syncing && (s.uptodate + s.compute < disks)) || s.expanding)
3068 handle_stripe_fill5(sh, &s, disks);
3070 /* Now we check to see if any write operations have recently
3074 if (sh->reconstruct_state == reconstruct_state_prexor_drain_result)
3076 if (sh->reconstruct_state == reconstruct_state_drain_result ||
3077 sh->reconstruct_state == reconstruct_state_prexor_drain_result) {
3078 sh->reconstruct_state = reconstruct_state_idle;
3080 /* All the 'written' buffers and the parity block are ready to
3081 * be written back to disk
3083 BUG_ON(!test_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags));
3084 for (i = disks; i--; ) {
3086 if (test_bit(R5_LOCKED, &dev->flags) &&
3087 (i == sh->pd_idx || dev->written)) {
3088 pr_debug("Writing block %d\n", i);
3089 set_bit(R5_Wantwrite, &dev->flags);
3092 if (!test_bit(R5_Insync, &dev->flags) ||
3093 (i == sh->pd_idx && s.failed == 0))
3094 set_bit(STRIPE_INSYNC, &sh->state);
3097 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
3098 dec_preread_active = 1;
3101 /* Now to consider new write requests and what else, if anything
3102 * should be read. We do not handle new writes when:
3103 * 1/ A 'write' operation (copy+xor) is already in flight.
3104 * 2/ A 'check' operation is in flight, as it may clobber the parity
3107 if (s.to_write && !sh->reconstruct_state && !sh->check_state)
3108 handle_stripe_dirtying5(conf, sh, &s, disks);
3110 /* maybe we need to check and possibly fix the parity for this stripe
3111 * Any reads will already have been scheduled, so we just see if enough
3112 * data is available. The parity check is held off while parity
3113 * dependent operations are in flight.
3115 if (sh->check_state ||
3116 (s.syncing && s.locked == 0 &&
3117 !test_bit(STRIPE_COMPUTE_RUN, &sh->state) &&
3118 !test_bit(STRIPE_INSYNC, &sh->state)))
3119 handle_parity_checks5(conf, sh, &s, disks);
3121 if (s.syncing && s.locked == 0 && test_bit(STRIPE_INSYNC, &sh->state)) {
3122 md_done_sync(conf->mddev, STRIPE_SECTORS,1);
3123 clear_bit(STRIPE_SYNCING, &sh->state);
3126 /* If the failed drive is just a ReadError, then we might need to progress
3127 * the repair/check process
3129 if (s.failed == 1 && !conf->mddev->ro &&
3130 test_bit(R5_ReadError, &sh->dev[s.failed_num].flags)
3131 && !test_bit(R5_LOCKED, &sh->dev[s.failed_num].flags)
3132 && test_bit(R5_UPTODATE, &sh->dev[s.failed_num].flags)
3134 dev = &sh->dev[s.failed_num];
3135 if (!test_bit(R5_ReWrite, &dev->flags)) {
3136 set_bit(R5_Wantwrite, &dev->flags);
3137 set_bit(R5_ReWrite, &dev->flags);
3138 set_bit(R5_LOCKED, &dev->flags);
3141 /* let's read it back */
3142 set_bit(R5_Wantread, &dev->flags);
3143 set_bit(R5_LOCKED, &dev->flags);
3148 /* Finish reconstruct operations initiated by the expansion process */
3149 if (sh->reconstruct_state == reconstruct_state_result) {
3150 struct stripe_head *sh2
3151 = get_active_stripe(conf, sh->sector, 1, 1, 1);
3152 if (sh2 && test_bit(STRIPE_EXPAND_SOURCE, &sh2->state)) {
3153 /* sh cannot be written until sh2 has been read.
3154 * so arrange for sh to be delayed a little
3156 set_bit(STRIPE_DELAYED, &sh->state);
3157 set_bit(STRIPE_HANDLE, &sh->state);
3158 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE,
3160 atomic_inc(&conf->preread_active_stripes);
3161 release_stripe(sh2);
3165 release_stripe(sh2);
3167 sh->reconstruct_state = reconstruct_state_idle;
3168 clear_bit(STRIPE_EXPANDING, &sh->state);
3169 for (i = conf->raid_disks; i--; ) {
3170 set_bit(R5_Wantwrite, &sh->dev[i].flags);
3171 set_bit(R5_LOCKED, &sh->dev[i].flags);
3176 if (s.expanded && test_bit(STRIPE_EXPANDING, &sh->state) &&
3177 !sh->reconstruct_state) {
3178 /* Need to write out all blocks after computing parity */
3179 sh->disks = conf->raid_disks;
3180 stripe_set_idx(sh->sector, conf, 0, sh);
3181 schedule_reconstruction(sh, &s, 1, 1);
3182 } else if (s.expanded && !sh->reconstruct_state && s.locked == 0) {
3183 clear_bit(STRIPE_EXPAND_READY, &sh->state);
3184 atomic_dec(&conf->reshape_stripes);
3185 wake_up(&conf->wait_for_overlap);
3186 md_done_sync(conf->mddev, STRIPE_SECTORS, 1);
3189 if (s.expanding && s.locked == 0 &&
3190 !test_bit(STRIPE_COMPUTE_RUN, &sh->state))
3191 handle_stripe_expansion(conf, sh, NULL);
3194 spin_unlock(&sh->lock);
3196 /* wait for this device to become unblocked */
3197 if (unlikely(blocked_rdev))
3198 md_wait_for_blocked_rdev(blocked_rdev, conf->mddev);
3201 raid_run_ops(sh, s.ops_request);
3205 if (dec_preread_active) {
3206 /* We delay this until after ops_run_io so that if make_request
3207 * is waiting on a barrier, it won't continue until the writes
3208 * have actually been submitted.
3210 atomic_dec(&conf->preread_active_stripes);
3211 if (atomic_read(&conf->preread_active_stripes) <
3213 md_wakeup_thread(conf->mddev->thread);
3215 return_io(return_bi);
3218 static void handle_stripe6(struct stripe_head *sh)
3220 raid5_conf_t *conf = sh->raid_conf;
3221 int disks = sh->disks;
3222 struct bio *return_bi = NULL;
3223 int i, pd_idx = sh->pd_idx, qd_idx = sh->qd_idx;
3224 struct stripe_head_state s;
3225 struct r6_state r6s;
3226 struct r5dev *dev, *pdev, *qdev;
3227 mdk_rdev_t *blocked_rdev = NULL;
3228 int dec_preread_active = 0;
3230 pr_debug("handling stripe %llu, state=%#lx cnt=%d, "
3231 "pd_idx=%d, qd_idx=%d\n, check:%d, reconstruct:%d\n",
3232 (unsigned long long)sh->sector, sh->state,
3233 atomic_read(&sh->count), pd_idx, qd_idx,
3234 sh->check_state, sh->reconstruct_state);
3235 memset(&s, 0, sizeof(s));
3237 spin_lock(&sh->lock);
3238 clear_bit(STRIPE_HANDLE, &sh->state);
3239 clear_bit(STRIPE_DELAYED, &sh->state);
3241 s.syncing = test_bit(STRIPE_SYNCING, &sh->state);
3242 s.expanding = test_bit(STRIPE_EXPAND_SOURCE, &sh->state);
3243 s.expanded = test_bit(STRIPE_EXPAND_READY, &sh->state);
3244 /* Now to look around and see what can be done */
3247 for (i=disks; i--; ) {
3250 clear_bit(R5_Insync, &dev->flags);
3252 pr_debug("check %d: state 0x%lx read %p write %p written %p\n",
3253 i, dev->flags, dev->toread, dev->towrite, dev->written);
3254 /* maybe we can reply to a read
3256 * new wantfill requests are only permitted while
3257 * ops_complete_biofill is guaranteed to be inactive
3259 if (test_bit(R5_UPTODATE, &dev->flags) && dev->toread &&
3260 !test_bit(STRIPE_BIOFILL_RUN, &sh->state))
3261 set_bit(R5_Wantfill, &dev->flags);
3263 /* now count some things */
3264 if (test_bit(R5_LOCKED, &dev->flags)) s.locked++;
3265 if (test_bit(R5_UPTODATE, &dev->flags)) s.uptodate++;
3266 if (test_bit(R5_Wantcompute, &dev->flags)) {
3268 BUG_ON(s.compute > 2);
3271 if (test_bit(R5_Wantfill, &dev->flags)) {
3273 } else if (dev->toread)
3277 if (!test_bit(R5_OVERWRITE, &dev->flags))
3282 rdev = rcu_dereference(conf->disks[i].rdev);
3283 if (blocked_rdev == NULL &&
3284 rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
3285 blocked_rdev = rdev;
3286 atomic_inc(&rdev->nr_pending);
3288 if (!rdev || !test_bit(In_sync, &rdev->flags)) {
3289 /* The ReadError flag will just be confusing now */
3290 clear_bit(R5_ReadError, &dev->flags);
3291 clear_bit(R5_ReWrite, &dev->flags);
3293 if (!rdev || !test_bit(In_sync, &rdev->flags)
3294 || test_bit(R5_ReadError, &dev->flags)) {
3296 r6s.failed_num[s.failed] = i;
3299 set_bit(R5_Insync, &dev->flags);
3303 if (unlikely(blocked_rdev)) {
3304 if (s.syncing || s.expanding || s.expanded ||
3305 s.to_write || s.written) {
3306 set_bit(STRIPE_HANDLE, &sh->state);
3309 /* There is nothing for the blocked_rdev to block */
3310 rdev_dec_pending(blocked_rdev, conf->mddev);
3311 blocked_rdev = NULL;
3314 if (s.to_fill && !test_bit(STRIPE_BIOFILL_RUN, &sh->state)) {
3315 set_bit(STRIPE_OP_BIOFILL, &s.ops_request);
3316 set_bit(STRIPE_BIOFILL_RUN, &sh->state);
3319 pr_debug("locked=%d uptodate=%d to_read=%d"
3320 " to_write=%d failed=%d failed_num=%d,%d\n",
3321 s.locked, s.uptodate, s.to_read, s.to_write, s.failed,
3322 r6s.failed_num[0], r6s.failed_num[1]);
3323 /* check if the array has lost >2 devices and, if so, some requests
3324 * might need to be failed
3326 if (s.failed > 2 && s.to_read+s.to_write+s.written)
3327 handle_failed_stripe(conf, sh, &s, disks, &return_bi);
3328 if (s.failed > 2 && s.syncing) {
3329 md_done_sync(conf->mddev, STRIPE_SECTORS,0);
3330 clear_bit(STRIPE_SYNCING, &sh->state);
3335 * might be able to return some write requests if the parity blocks
3336 * are safe, or on a failed drive
3338 pdev = &sh->dev[pd_idx];
3339 r6s.p_failed = (s.failed >= 1 && r6s.failed_num[0] == pd_idx)
3340 || (s.failed >= 2 && r6s.failed_num[1] == pd_idx);
3341 qdev = &sh->dev[qd_idx];
3342 r6s.q_failed = (s.failed >= 1 && r6s.failed_num[0] == qd_idx)
3343 || (s.failed >= 2 && r6s.failed_num[1] == qd_idx);
3346 ( r6s.p_failed || ((test_bit(R5_Insync, &pdev->flags)
3347 && !test_bit(R5_LOCKED, &pdev->flags)
3348 && test_bit(R5_UPTODATE, &pdev->flags)))) &&
3349 ( r6s.q_failed || ((test_bit(R5_Insync, &qdev->flags)
3350 && !test_bit(R5_LOCKED, &qdev->flags)
3351 && test_bit(R5_UPTODATE, &qdev->flags)))))
3352 handle_stripe_clean_event(conf, sh, disks, &return_bi);
3354 /* Now we might consider reading some blocks, either to check/generate
3355 * parity, or to satisfy requests
3356 * or to load a block that is being partially written.
3358 if (s.to_read || s.non_overwrite || (s.to_write && s.failed) ||
3359 (s.syncing && (s.uptodate + s.compute < disks)) || s.expanding)
3360 handle_stripe_fill6(sh, &s, &r6s, disks);
3362 /* Now we check to see if any write operations have recently
3365 if (sh->reconstruct_state == reconstruct_state_drain_result) {
3367 sh->reconstruct_state = reconstruct_state_idle;
3368 /* All the 'written' buffers and the parity blocks are ready to
3369 * be written back to disk
3371 BUG_ON(!test_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags));
3372 BUG_ON(!test_bit(R5_UPTODATE, &sh->dev[qd_idx].flags));
3373 for (i = disks; i--; ) {
3375 if (test_bit(R5_LOCKED, &dev->flags) &&
3376 (i == sh->pd_idx || i == qd_idx ||
3378 pr_debug("Writing block %d\n", i);
3379 BUG_ON(!test_bit(R5_UPTODATE, &dev->flags));
3380 set_bit(R5_Wantwrite, &dev->flags);
3381 if (!test_bit(R5_Insync, &dev->flags) ||
3382 ((i == sh->pd_idx || i == qd_idx) &&
3384 set_bit(STRIPE_INSYNC, &sh->state);
3387 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
3388 dec_preread_active = 1;
3391 /* Now to consider new write requests and what else, if anything
3392 * should be read. We do not handle new writes when:
3393 * 1/ A 'write' operation (copy+gen_syndrome) is already in flight.
3394 * 2/ A 'check' operation is in flight, as it may clobber the parity
3397 if (s.to_write && !sh->reconstruct_state && !sh->check_state)
3398 handle_stripe_dirtying6(conf, sh, &s, &r6s, disks);
3400 /* maybe we need to check and possibly fix the parity for this stripe
3401 * Any reads will already have been scheduled, so we just see if enough
3402 * data is available. The parity check is held off while parity
3403 * dependent operations are in flight.
3405 if (sh->check_state ||
3406 (s.syncing && s.locked == 0 &&
3407 !test_bit(STRIPE_COMPUTE_RUN, &sh->state) &&
3408 !test_bit(STRIPE_INSYNC, &sh->state)))
3409 handle_parity_checks6(conf, sh, &s, &r6s, disks);
3411 if (s.syncing && s.locked == 0 && test_bit(STRIPE_INSYNC, &sh->state)) {
3412 md_done_sync(conf->mddev, STRIPE_SECTORS,1);
3413 clear_bit(STRIPE_SYNCING, &sh->state);
3416 /* If the failed drives are just a ReadError, then we might need
3417 * to progress the repair/check process
3419 if (s.failed <= 2 && !conf->mddev->ro)
3420 for (i = 0; i < s.failed; i++) {
3421 dev = &sh->dev[r6s.failed_num[i]];
3422 if (test_bit(R5_ReadError, &dev->flags)
3423 && !test_bit(R5_LOCKED, &dev->flags)
3424 && test_bit(R5_UPTODATE, &dev->flags)
3426 if (!test_bit(R5_ReWrite, &dev->flags)) {
3427 set_bit(R5_Wantwrite, &dev->flags);
3428 set_bit(R5_ReWrite, &dev->flags);
3429 set_bit(R5_LOCKED, &dev->flags);
3432 /* let's read it back */
3433 set_bit(R5_Wantread, &dev->flags);
3434 set_bit(R5_LOCKED, &dev->flags);
3440 /* Finish reconstruct operations initiated by the expansion process */
3441 if (sh->reconstruct_state == reconstruct_state_result) {
3442 sh->reconstruct_state = reconstruct_state_idle;
3443 clear_bit(STRIPE_EXPANDING, &sh->state);
3444 for (i = conf->raid_disks; i--; ) {
3445 set_bit(R5_Wantwrite, &sh->dev[i].flags);
3446 set_bit(R5_LOCKED, &sh->dev[i].flags);
3451 if (s.expanded && test_bit(STRIPE_EXPANDING, &sh->state) &&
3452 !sh->reconstruct_state) {
3453 struct stripe_head *sh2
3454 = get_active_stripe(conf, sh->sector, 1, 1, 1);
3455 if (sh2 && test_bit(STRIPE_EXPAND_SOURCE, &sh2->state)) {
3456 /* sh cannot be written until sh2 has been read.
3457 * so arrange for sh to be delayed a little
3459 set_bit(STRIPE_DELAYED, &sh->state);
3460 set_bit(STRIPE_HANDLE, &sh->state);
3461 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE,
3463 atomic_inc(&conf->preread_active_stripes);
3464 release_stripe(sh2);
3468 release_stripe(sh2);
3470 /* Need to write out all blocks after computing P&Q */
3471 sh->disks = conf->raid_disks;
3472 stripe_set_idx(sh->sector, conf, 0, sh);
3473 schedule_reconstruction(sh, &s, 1, 1);
3474 } else if (s.expanded && !sh->reconstruct_state && s.locked == 0) {
3475 clear_bit(STRIPE_EXPAND_READY, &sh->state);
3476 atomic_dec(&conf->reshape_stripes);
3477 wake_up(&conf->wait_for_overlap);
3478 md_done_sync(conf->mddev, STRIPE_SECTORS, 1);
3481 if (s.expanding && s.locked == 0 &&
3482 !test_bit(STRIPE_COMPUTE_RUN, &sh->state))
3483 handle_stripe_expansion(conf, sh, &r6s);
3486 spin_unlock(&sh->lock);
3488 /* wait for this device to become unblocked */
3489 if (unlikely(blocked_rdev))
3490 md_wait_for_blocked_rdev(blocked_rdev, conf->mddev);
3493 raid_run_ops(sh, s.ops_request);
3498 if (dec_preread_active) {
3499 /* We delay this until after ops_run_io so that if make_request
3500 * is waiting on a barrier, it won't continue until the writes
3501 * have actually been submitted.
3503 atomic_dec(&conf->preread_active_stripes);
3504 if (atomic_read(&conf->preread_active_stripes) <
3506 md_wakeup_thread(conf->mddev->thread);
3509 return_io(return_bi);
3512 static void handle_stripe(struct stripe_head *sh)
3514 if (sh->raid_conf->level == 6)
3520 static void raid5_activate_delayed(raid5_conf_t *conf)
3522 if (atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD) {
3523 while (!list_empty(&conf->delayed_list)) {
3524 struct list_head *l = conf->delayed_list.next;
3525 struct stripe_head *sh;
3526 sh = list_entry(l, struct stripe_head, lru);
3528 clear_bit(STRIPE_DELAYED, &sh->state);
3529 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
3530 atomic_inc(&conf->preread_active_stripes);
3531 list_add_tail(&sh->lru, &conf->hold_list);
3534 blk_plug_device(conf->mddev->queue);
3537 static void activate_bit_delay(raid5_conf_t *conf)
3539 /* device_lock is held */
3540 struct list_head head;
3541 list_add(&head, &conf->bitmap_list);
3542 list_del_init(&conf->bitmap_list);
3543 while (!list_empty(&head)) {
3544 struct stripe_head *sh = list_entry(head.next, struct stripe_head, lru);
3545 list_del_init(&sh->lru);
3546 atomic_inc(&sh->count);
3547 __release_stripe(conf, sh);
3551 static void unplug_slaves(mddev_t *mddev)
3553 raid5_conf_t *conf = mddev->private;
3555 int devs = max(conf->raid_disks, conf->previous_raid_disks);
3558 for (i = 0; i < devs; i++) {
3559 mdk_rdev_t *rdev = rcu_dereference(conf->disks[i].rdev);
3560 if (rdev && !test_bit(Faulty, &rdev->flags) && atomic_read(&rdev->nr_pending)) {
3561 struct request_queue *r_queue = bdev_get_queue(rdev->bdev);
3563 atomic_inc(&rdev->nr_pending);
3566 blk_unplug(r_queue);
3568 rdev_dec_pending(rdev, mddev);
3575 static void raid5_unplug_device(struct request_queue *q)
3577 mddev_t *mddev = q->queuedata;
3578 raid5_conf_t *conf = mddev->private;
3579 unsigned long flags;
3581 spin_lock_irqsave(&conf->device_lock, flags);
3583 if (blk_remove_plug(q)) {
3585 raid5_activate_delayed(conf);
3587 md_wakeup_thread(mddev->thread);
3589 spin_unlock_irqrestore(&conf->device_lock, flags);
3591 unplug_slaves(mddev);
3594 static int raid5_congested(void *data, int bits)
3596 mddev_t *mddev = data;
3597 raid5_conf_t *conf = mddev->private;
3599 /* No difference between reads and writes. Just check
3600 * how busy the stripe_cache is
3603 if (mddev_congested(mddev, bits))
3605 if (conf->inactive_blocked)
3609 if (list_empty_careful(&conf->inactive_list))
3615 /* We want read requests to align with chunks where possible,
3616 * but write requests don't need to.
3618 static int raid5_mergeable_bvec(struct request_queue *q,
3619 struct bvec_merge_data *bvm,
3620 struct bio_vec *biovec)
3622 mddev_t *mddev = q->queuedata;
3623 sector_t sector = bvm->bi_sector + get_start_sect(bvm->bi_bdev);
3625 unsigned int chunk_sectors = mddev->chunk_sectors;
3626 unsigned int bio_sectors = bvm->bi_size >> 9;
3628 if ((bvm->bi_rw & 1) == WRITE)
3629 return biovec->bv_len; /* always allow writes to be mergeable */
3631 if (mddev->new_chunk_sectors < mddev->chunk_sectors)
3632 chunk_sectors = mddev->new_chunk_sectors;
3633 max = (chunk_sectors - ((sector & (chunk_sectors - 1)) + bio_sectors)) << 9;
3634 if (max < 0) max = 0;
3635 if (max <= biovec->bv_len && bio_sectors == 0)
3636 return biovec->bv_len;
3642 static int in_chunk_boundary(mddev_t *mddev, struct bio *bio)
3644 sector_t sector = bio->bi_sector + get_start_sect(bio->bi_bdev);
3645 unsigned int chunk_sectors = mddev->chunk_sectors;
3646 unsigned int bio_sectors = bio->bi_size >> 9;
3648 if (mddev->new_chunk_sectors < mddev->chunk_sectors)
3649 chunk_sectors = mddev->new_chunk_sectors;
3650 return chunk_sectors >=
3651 ((sector & (chunk_sectors - 1)) + bio_sectors);
3655 * add bio to the retry LIFO ( in O(1) ... we are in interrupt )
3656 * later sampled by raid5d.
3658 static void add_bio_to_retry(struct bio *bi,raid5_conf_t *conf)
3660 unsigned long flags;
3662 spin_lock_irqsave(&conf->device_lock, flags);
3664 bi->bi_next = conf->retry_read_aligned_list;
3665 conf->retry_read_aligned_list = bi;
3667 spin_unlock_irqrestore(&conf->device_lock, flags);
3668 md_wakeup_thread(conf->mddev->thread);
3672 static struct bio *remove_bio_from_retry(raid5_conf_t *conf)
3676 bi = conf->retry_read_aligned;
3678 conf->retry_read_aligned = NULL;
3681 bi = conf->retry_read_aligned_list;
3683 conf->retry_read_aligned_list = bi->bi_next;
3686 * this sets the active strip count to 1 and the processed
3687 * strip count to zero (upper 8 bits)
3689 bi->bi_phys_segments = 1; /* biased count of active stripes */
3697 * The "raid5_align_endio" should check if the read succeeded and if it
3698 * did, call bio_endio on the original bio (having bio_put the new bio
3700 * If the read failed..
3702 static void raid5_align_endio(struct bio *bi, int error)
3704 struct bio* raid_bi = bi->bi_private;
3707 int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
3712 rdev = (void*)raid_bi->bi_next;
3713 raid_bi->bi_next = NULL;
3714 mddev = rdev->mddev;
3715 conf = mddev->private;
3717 rdev_dec_pending(rdev, conf->mddev);
3719 if (!error && uptodate) {
3720 bio_endio(raid_bi, 0);
3721 if (atomic_dec_and_test(&conf->active_aligned_reads))
3722 wake_up(&conf->wait_for_stripe);
3727 pr_debug("raid5_align_endio : io error...handing IO for a retry\n");
3729 add_bio_to_retry(raid_bi, conf);
3732 static int bio_fits_rdev(struct bio *bi)
3734 struct request_queue *q = bdev_get_queue(bi->bi_bdev);
3736 if ((bi->bi_size>>9) > queue_max_sectors(q))
3738 blk_recount_segments(q, bi);
3739 if (bi->bi_phys_segments > queue_max_segments(q))
3742 if (q->merge_bvec_fn)
3743 /* it's too hard to apply the merge_bvec_fn at this stage,
3752 static int chunk_aligned_read(mddev_t *mddev, struct bio * raid_bio)
3754 raid5_conf_t *conf = mddev->private;
3756 struct bio* align_bi;
3759 if (!in_chunk_boundary(mddev, raid_bio)) {
3760 pr_debug("chunk_aligned_read : non aligned\n");
3764 * use bio_clone to make a copy of the bio
3766 align_bi = bio_clone(raid_bio, GFP_NOIO);
3770 * set bi_end_io to a new function, and set bi_private to the
3773 align_bi->bi_end_io = raid5_align_endio;
3774 align_bi->bi_private = raid_bio;
3778 align_bi->bi_sector = raid5_compute_sector(conf, raid_bio->bi_sector,
3783 rdev = rcu_dereference(conf->disks[dd_idx].rdev);
3784 if (rdev && test_bit(In_sync, &rdev->flags)) {
3785 atomic_inc(&rdev->nr_pending);
3787 raid_bio->bi_next = (void*)rdev;
3788 align_bi->bi_bdev = rdev->bdev;
3789 align_bi->bi_flags &= ~(1 << BIO_SEG_VALID);
3790 align_bi->bi_sector += rdev->data_offset;
3792 if (!bio_fits_rdev(align_bi)) {
3793 /* too big in some way */
3795 rdev_dec_pending(rdev, mddev);
3799 spin_lock_irq(&conf->device_lock);
3800 wait_event_lock_irq(conf->wait_for_stripe,
3802 conf->device_lock, /* nothing */);
3803 atomic_inc(&conf->active_aligned_reads);
3804 spin_unlock_irq(&conf->device_lock);
3806 generic_make_request(align_bi);
3815 /* __get_priority_stripe - get the next stripe to process
3817 * Full stripe writes are allowed to pass preread active stripes up until
3818 * the bypass_threshold is exceeded. In general the bypass_count
3819 * increments when the handle_list is handled before the hold_list; however, it
3820 * will not be incremented when STRIPE_IO_STARTED is sampled set signifying a
3821 * stripe with in flight i/o. The bypass_count will be reset when the
3822 * head of the hold_list has changed, i.e. the head was promoted to the
3825 static struct stripe_head *__get_priority_stripe(raid5_conf_t *conf)
3827 struct stripe_head *sh;
3829 pr_debug("%s: handle: %s hold: %s full_writes: %d bypass_count: %d\n",
3831 list_empty(&conf->handle_list) ? "empty" : "busy",
3832 list_empty(&conf->hold_list) ? "empty" : "busy",
3833 atomic_read(&conf->pending_full_writes), conf->bypass_count);
3835 if (!list_empty(&conf->handle_list)) {
3836 sh = list_entry(conf->handle_list.next, typeof(*sh), lru);
3838 if (list_empty(&conf->hold_list))
3839 conf->bypass_count = 0;
3840 else if (!test_bit(STRIPE_IO_STARTED, &sh->state)) {
3841 if (conf->hold_list.next == conf->last_hold)
3842 conf->bypass_count++;
3844 conf->last_hold = conf->hold_list.next;
3845 conf->bypass_count -= conf->bypass_threshold;
3846 if (conf->bypass_count < 0)
3847 conf->bypass_count = 0;
3850 } else if (!list_empty(&conf->hold_list) &&
3851 ((conf->bypass_threshold &&
3852 conf->bypass_count > conf->bypass_threshold) ||
3853 atomic_read(&conf->pending_full_writes) == 0)) {
3854 sh = list_entry(conf->hold_list.next,
3856 conf->bypass_count -= conf->bypass_threshold;
3857 if (conf->bypass_count < 0)
3858 conf->bypass_count = 0;
3862 list_del_init(&sh->lru);
3863 atomic_inc(&sh->count);
3864 BUG_ON(atomic_read(&sh->count) != 1);
3868 static int make_request(mddev_t *mddev, struct bio * bi)
3870 raid5_conf_t *conf = mddev->private;
3872 sector_t new_sector;
3873 sector_t logical_sector, last_sector;
3874 struct stripe_head *sh;
3875 const int rw = bio_data_dir(bi);
3878 if (unlikely(bio_rw_flagged(bi, BIO_RW_BARRIER))) {
3879 /* Drain all pending writes. We only really need
3880 * to ensure they have been submitted, but this is
3883 mddev->pers->quiesce(mddev, 1);
3884 mddev->pers->quiesce(mddev, 0);
3885 md_barrier_request(mddev, bi);
3889 md_write_start(mddev, bi);
3892 mddev->reshape_position == MaxSector &&
3893 chunk_aligned_read(mddev,bi))
3896 logical_sector = bi->bi_sector & ~((sector_t)STRIPE_SECTORS-1);
3897 last_sector = bi->bi_sector + (bi->bi_size>>9);
3899 bi->bi_phys_segments = 1; /* over-loaded to count active stripes */
3901 for (;logical_sector < last_sector; logical_sector += STRIPE_SECTORS) {
3903 int disks, data_disks;
3908 disks = conf->raid_disks;
3909 prepare_to_wait(&conf->wait_for_overlap, &w, TASK_UNINTERRUPTIBLE);
3910 if (unlikely(conf->reshape_progress != MaxSector)) {
3911 /* spinlock is needed as reshape_progress may be
3912 * 64bit on a 32bit platform, and so it might be
3913 * possible to see a half-updated value
3914 * Ofcourse reshape_progress could change after
3915 * the lock is dropped, so once we get a reference
3916 * to the stripe that we think it is, we will have
3919 spin_lock_irq(&conf->device_lock);
3920 if (mddev->delta_disks < 0
3921 ? logical_sector < conf->reshape_progress
3922 : logical_sector >= conf->reshape_progress) {
3923 disks = conf->previous_raid_disks;
3926 if (mddev->delta_disks < 0
3927 ? logical_sector < conf->reshape_safe
3928 : logical_sector >= conf->reshape_safe) {
3929 spin_unlock_irq(&conf->device_lock);
3934 spin_unlock_irq(&conf->device_lock);
3936 data_disks = disks - conf->max_degraded;
3938 new_sector = raid5_compute_sector(conf, logical_sector,
3941 pr_debug("raid456: make_request, sector %llu logical %llu\n",
3942 (unsigned long long)new_sector,
3943 (unsigned long long)logical_sector);
3945 sh = get_active_stripe(conf, new_sector, previous,
3946 (bi->bi_rw&RWA_MASK), 0);
3948 if (unlikely(previous)) {
3949 /* expansion might have moved on while waiting for a
3950 * stripe, so we must do the range check again.
3951 * Expansion could still move past after this
3952 * test, but as we are holding a reference to
3953 * 'sh', we know that if that happens,
3954 * STRIPE_EXPANDING will get set and the expansion
3955 * won't proceed until we finish with the stripe.
3958 spin_lock_irq(&conf->device_lock);
3959 if (mddev->delta_disks < 0
3960 ? logical_sector >= conf->reshape_progress
3961 : logical_sector < conf->reshape_progress)
3962 /* mismatch, need to try again */
3964 spin_unlock_irq(&conf->device_lock);
3972 if (bio_data_dir(bi) == WRITE &&
3973 logical_sector >= mddev->suspend_lo &&
3974 logical_sector < mddev->suspend_hi) {
3976 /* As the suspend_* range is controlled by
3977 * userspace, we want an interruptible
3980 flush_signals(current);
3981 prepare_to_wait(&conf->wait_for_overlap,
3982 &w, TASK_INTERRUPTIBLE);
3983 if (logical_sector >= mddev->suspend_lo &&
3984 logical_sector < mddev->suspend_hi)
3989 if (test_bit(STRIPE_EXPANDING, &sh->state) ||
3990 !add_stripe_bio(sh, bi, dd_idx, (bi->bi_rw&RW_MASK))) {
3991 /* Stripe is busy expanding or
3992 * add failed due to overlap. Flush everything
3995 raid5_unplug_device(mddev->queue);
4000 finish_wait(&conf->wait_for_overlap, &w);
4001 set_bit(STRIPE_HANDLE, &sh->state);
4002 clear_bit(STRIPE_DELAYED, &sh->state);
4003 if (mddev->barrier &&
4004 !test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
4005 atomic_inc(&conf->preread_active_stripes);
4008 /* cannot get stripe for read-ahead, just give-up */
4009 clear_bit(BIO_UPTODATE, &bi->bi_flags);
4010 finish_wait(&conf->wait_for_overlap, &w);
4015 spin_lock_irq(&conf->device_lock);
4016 remaining = raid5_dec_bi_phys_segments(bi);
4017 spin_unlock_irq(&conf->device_lock);
4018 if (remaining == 0) {
4021 md_write_end(mddev);
4026 if (mddev->barrier) {
4027 /* We need to wait for the stripes to all be handled.
4028 * So: wait for preread_active_stripes to drop to 0.
4030 wait_event(mddev->thread->wqueue,
4031 atomic_read(&conf->preread_active_stripes) == 0);
4036 static sector_t raid5_size(mddev_t *mddev, sector_t sectors, int raid_disks);
4038 static sector_t reshape_request(mddev_t *mddev, sector_t sector_nr, int *skipped)
4040 /* reshaping is quite different to recovery/resync so it is
4041 * handled quite separately ... here.
4043 * On each call to sync_request, we gather one chunk worth of
4044 * destination stripes and flag them as expanding.
4045 * Then we find all the source stripes and request reads.
4046 * As the reads complete, handle_stripe will copy the data
4047 * into the destination stripe and release that stripe.
4049 raid5_conf_t *conf = mddev->private;
4050 struct stripe_head *sh;
4051 sector_t first_sector, last_sector;
4052 int raid_disks = conf->previous_raid_disks;
4053 int data_disks = raid_disks - conf->max_degraded;
4054 int new_data_disks = conf->raid_disks - conf->max_degraded;
4057 sector_t writepos, readpos, safepos;
4058 sector_t stripe_addr;
4059 int reshape_sectors;
4060 struct list_head stripes;
4062 if (sector_nr == 0) {
4063 /* If restarting in the middle, skip the initial sectors */
4064 if (mddev->delta_disks < 0 &&
4065 conf->reshape_progress < raid5_size(mddev, 0, 0)) {
4066 sector_nr = raid5_size(mddev, 0, 0)
4067 - conf->reshape_progress;
4068 } else if (mddev->delta_disks >= 0 &&
4069 conf->reshape_progress > 0)
4070 sector_nr = conf->reshape_progress;
4071 sector_div(sector_nr, new_data_disks);
4073 mddev->curr_resync_completed = sector_nr;
4074 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
4080 /* We need to process a full chunk at a time.
4081 * If old and new chunk sizes differ, we need to process the
4084 if (mddev->new_chunk_sectors > mddev->chunk_sectors)
4085 reshape_sectors = mddev->new_chunk_sectors;
4087 reshape_sectors = mddev->chunk_sectors;
4089 /* we update the metadata when there is more than 3Meg
4090 * in the block range (that is rather arbitrary, should
4091 * probably be time based) or when the data about to be
4092 * copied would over-write the source of the data at
4093 * the front of the range.
4094 * i.e. one new_stripe along from reshape_progress new_maps
4095 * to after where reshape_safe old_maps to
4097 writepos = conf->reshape_progress;
4098 sector_div(writepos, new_data_disks);
4099 readpos = conf->reshape_progress;
4100 sector_div(readpos, data_disks);
4101 safepos = conf->reshape_safe;
4102 sector_div(safepos, data_disks);
4103 if (mddev->delta_disks < 0) {
4104 writepos -= min_t(sector_t, reshape_sectors, writepos);
4105 readpos += reshape_sectors;
4106 safepos += reshape_sectors;
4108 writepos += reshape_sectors;
4109 readpos -= min_t(sector_t, reshape_sectors, readpos);
4110 safepos -= min_t(sector_t, reshape_sectors, safepos);
4113 /* 'writepos' is the most advanced device address we might write.
4114 * 'readpos' is the least advanced device address we might read.
4115 * 'safepos' is the least address recorded in the metadata as having
4117 * If 'readpos' is behind 'writepos', then there is no way that we can
4118 * ensure safety in the face of a crash - that must be done by userspace
4119 * making a backup of the data. So in that case there is no particular
4120 * rush to update metadata.
4121 * Otherwise if 'safepos' is behind 'writepos', then we really need to
4122 * update the metadata to advance 'safepos' to match 'readpos' so that
4123 * we can be safe in the event of a crash.
4124 * So we insist on updating metadata if safepos is behind writepos and
4125 * readpos is beyond writepos.
4126 * In any case, update the metadata every 10 seconds.
4127 * Maybe that number should be configurable, but I'm not sure it is
4128 * worth it.... maybe it could be a multiple of safemode_delay???
4130 if ((mddev->delta_disks < 0
4131 ? (safepos > writepos && readpos < writepos)
4132 : (safepos < writepos && readpos > writepos)) ||
4133 time_after(jiffies, conf->reshape_checkpoint + 10*HZ)) {
4134 /* Cannot proceed until we've updated the superblock... */
4135 wait_event(conf->wait_for_overlap,
4136 atomic_read(&conf->reshape_stripes)==0);
4137 mddev->reshape_position = conf->reshape_progress;
4138 mddev->curr_resync_completed = mddev->curr_resync;
4139 conf->reshape_checkpoint = jiffies;
4140 set_bit(MD_CHANGE_DEVS, &mddev->flags);
4141 md_wakeup_thread(mddev->thread);
4142 wait_event(mddev->sb_wait, mddev->flags == 0 ||
4143 kthread_should_stop());
4144 spin_lock_irq(&conf->device_lock);
4145 conf->reshape_safe = mddev->reshape_position;
4146 spin_unlock_irq(&conf->device_lock);
4147 wake_up(&conf->wait_for_overlap);
4148 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
4151 if (mddev->delta_disks < 0) {
4152 BUG_ON(conf->reshape_progress == 0);
4153 stripe_addr = writepos;
4154 BUG_ON((mddev->dev_sectors &
4155 ~((sector_t)reshape_sectors - 1))
4156 - reshape_sectors - stripe_addr
4159 BUG_ON(writepos != sector_nr + reshape_sectors);
4160 stripe_addr = sector_nr;
4162 INIT_LIST_HEAD(&stripes);
4163 for (i = 0; i < reshape_sectors; i += STRIPE_SECTORS) {
4165 int skipped_disk = 0;
4166 sh = get_active_stripe(conf, stripe_addr+i, 0, 0, 1);
4167 set_bit(STRIPE_EXPANDING, &sh->state);
4168 atomic_inc(&conf->reshape_stripes);
4169 /* If any of this stripe is beyond the end of the old
4170 * array, then we need to zero those blocks
4172 for (j=sh->disks; j--;) {
4174 if (j == sh->pd_idx)
4176 if (conf->level == 6 &&
4179 s = compute_blocknr(sh, j, 0);
4180 if (s < raid5_size(mddev, 0, 0)) {
4184 memset(page_address(sh->dev[j].page), 0, STRIPE_SIZE);
4185 set_bit(R5_Expanded, &sh->dev[j].flags);
4186 set_bit(R5_UPTODATE, &sh->dev[j].flags);
4188 if (!skipped_disk) {
4189 set_bit(STRIPE_EXPAND_READY, &sh->state);
4190 set_bit(STRIPE_HANDLE, &sh->state);
4192 list_add(&sh->lru, &stripes);
4194 spin_lock_irq(&conf->device_lock);
4195 if (mddev->delta_disks < 0)
4196 conf->reshape_progress -= reshape_sectors * new_data_disks;
4198 conf->reshape_progress += reshape_sectors * new_data_disks;
4199 spin_unlock_irq(&conf->device_lock);
4200 /* Ok, those stripe are ready. We can start scheduling
4201 * reads on the source stripes.
4202 * The source stripes are determined by mapping the first and last
4203 * block on the destination stripes.
4206 raid5_compute_sector(conf, stripe_addr*(new_data_disks),
4209 raid5_compute_sector(conf, ((stripe_addr+reshape_sectors)
4210 * new_data_disks - 1),
4212 if (last_sector >= mddev->dev_sectors)
4213 last_sector = mddev->dev_sectors - 1;
4214 while (first_sector <= last_sector) {
4215 sh = get_active_stripe(conf, first_sector, 1, 0, 1);
4216 set_bit(STRIPE_EXPAND_SOURCE, &sh->state);
4217 set_bit(STRIPE_HANDLE, &sh->state);
4219 first_sector += STRIPE_SECTORS;
4221 /* Now that the sources are clearly marked, we can release
4222 * the destination stripes
4224 while (!list_empty(&stripes)) {
4225 sh = list_entry(stripes.next, struct stripe_head, lru);
4226 list_del_init(&sh->lru);
4229 /* If this takes us to the resync_max point where we have to pause,
4230 * then we need to write out the superblock.
4232 sector_nr += reshape_sectors;
4233 if ((sector_nr - mddev->curr_resync_completed) * 2
4234 >= mddev->resync_max - mddev->curr_resync_completed) {
4235 /* Cannot proceed until we've updated the superblock... */
4236 wait_event(conf->wait_for_overlap,
4237 atomic_read(&conf->reshape_stripes) == 0);
4238 mddev->reshape_position = conf->reshape_progress;
4239 mddev->curr_resync_completed = mddev->curr_resync + reshape_sectors;
4240 conf->reshape_checkpoint = jiffies;
4241 set_bit(MD_CHANGE_DEVS, &mddev->flags);
4242 md_wakeup_thread(mddev->thread);
4243 wait_event(mddev->sb_wait,
4244 !test_bit(MD_CHANGE_DEVS, &mddev->flags)
4245 || kthread_should_stop());
4246 spin_lock_irq(&conf->device_lock);
4247 conf->reshape_safe = mddev->reshape_position;
4248 spin_unlock_irq(&conf->device_lock);
4249 wake_up(&conf->wait_for_overlap);
4250 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
4252 return reshape_sectors;
4255 /* FIXME go_faster isn't used */
4256 static inline sector_t sync_request(mddev_t *mddev, sector_t sector_nr, int *skipped, int go_faster)
4258 raid5_conf_t *conf = mddev->private;
4259 struct stripe_head *sh;
4260 sector_t max_sector = mddev->dev_sectors;
4262 int still_degraded = 0;
4265 if (sector_nr >= max_sector) {
4266 /* just being told to finish up .. nothing much to do */
4267 unplug_slaves(mddev);
4269 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery)) {
4274 if (mddev->curr_resync < max_sector) /* aborted */
4275 bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
4277 else /* completed sync */
4279 bitmap_close_sync(mddev->bitmap);
4284 /* Allow raid5_quiesce to complete */
4285 wait_event(conf->wait_for_overlap, conf->quiesce != 2);
4287 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
4288 return reshape_request(mddev, sector_nr, skipped);
4290 /* No need to check resync_max as we never do more than one
4291 * stripe, and as resync_max will always be on a chunk boundary,
4292 * if the check in md_do_sync didn't fire, there is no chance
4293 * of overstepping resync_max here
4296 /* if there is too many failed drives and we are trying
4297 * to resync, then assert that we are finished, because there is
4298 * nothing we can do.
4300 if (mddev->degraded >= conf->max_degraded &&
4301 test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
4302 sector_t rv = mddev->dev_sectors - sector_nr;
4306 if (!bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
4307 !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
4308 !conf->fullsync && sync_blocks >= STRIPE_SECTORS) {
4309 /* we can skip this block, and probably more */
4310 sync_blocks /= STRIPE_SECTORS;
4312 return sync_blocks * STRIPE_SECTORS; /* keep things rounded to whole stripes */
4316 bitmap_cond_end_sync(mddev->bitmap, sector_nr);
4318 sh = get_active_stripe(conf, sector_nr, 0, 1, 0);
4320 sh = get_active_stripe(conf, sector_nr, 0, 0, 0);
4321 /* make sure we don't swamp the stripe cache if someone else
4322 * is trying to get access
4324 schedule_timeout_uninterruptible(1);
4326 /* Need to check if array will still be degraded after recovery/resync
4327 * We don't need to check the 'failed' flag as when that gets set,
4330 for (i = 0; i < conf->raid_disks; i++)
4331 if (conf->disks[i].rdev == NULL)
4334 bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, still_degraded);
4336 spin_lock(&sh->lock);
4337 set_bit(STRIPE_SYNCING, &sh->state);
4338 clear_bit(STRIPE_INSYNC, &sh->state);
4339 spin_unlock(&sh->lock);
4344 return STRIPE_SECTORS;
4347 static int retry_aligned_read(raid5_conf_t *conf, struct bio *raid_bio)
4349 /* We may not be able to submit a whole bio at once as there
4350 * may not be enough stripe_heads available.
4351 * We cannot pre-allocate enough stripe_heads as we may need
4352 * more than exist in the cache (if we allow ever large chunks).
4353 * So we do one stripe head at a time and record in
4354 * ->bi_hw_segments how many have been done.
4356 * We *know* that this entire raid_bio is in one chunk, so
4357 * it will be only one 'dd_idx' and only need one call to raid5_compute_sector.
4359 struct stripe_head *sh;
4361 sector_t sector, logical_sector, last_sector;
4366 logical_sector = raid_bio->bi_sector & ~((sector_t)STRIPE_SECTORS-1);
4367 sector = raid5_compute_sector(conf, logical_sector,
4369 last_sector = raid_bio->bi_sector + (raid_bio->bi_size>>9);
4371 for (; logical_sector < last_sector;
4372 logical_sector += STRIPE_SECTORS,
4373 sector += STRIPE_SECTORS,
4376 if (scnt < raid5_bi_hw_segments(raid_bio))
4377 /* already done this stripe */
4380 sh = get_active_stripe(conf, sector, 0, 1, 0);
4383 /* failed to get a stripe - must wait */
4384 raid5_set_bi_hw_segments(raid_bio, scnt);
4385 conf->retry_read_aligned = raid_bio;
4389 set_bit(R5_ReadError, &sh->dev[dd_idx].flags);
4390 if (!add_stripe_bio(sh, raid_bio, dd_idx, 0)) {
4392 raid5_set_bi_hw_segments(raid_bio, scnt);
4393 conf->retry_read_aligned = raid_bio;
4401 spin_lock_irq(&conf->device_lock);
4402 remaining = raid5_dec_bi_phys_segments(raid_bio);
4403 spin_unlock_irq(&conf->device_lock);
4405 bio_endio(raid_bio, 0);
4406 if (atomic_dec_and_test(&conf->active_aligned_reads))
4407 wake_up(&conf->wait_for_stripe);
4413 * This is our raid5 kernel thread.
4415 * We scan the hash table for stripes which can be handled now.
4416 * During the scan, completed stripes are saved for us by the interrupt
4417 * handler, so that they will not have to wait for our next wakeup.
4419 static void raid5d(mddev_t *mddev)
4421 struct stripe_head *sh;
4422 raid5_conf_t *conf = mddev->private;
4425 pr_debug("+++ raid5d active\n");
4427 md_check_recovery(mddev);
4430 spin_lock_irq(&conf->device_lock);
4434 if (conf->seq_flush != conf->seq_write) {
4435 int seq = conf->seq_flush;
4436 spin_unlock_irq(&conf->device_lock);
4437 bitmap_unplug(mddev->bitmap);
4438 spin_lock_irq(&conf->device_lock);
4439 conf->seq_write = seq;
4440 activate_bit_delay(conf);
4443 while ((bio = remove_bio_from_retry(conf))) {
4445 spin_unlock_irq(&conf->device_lock);
4446 ok = retry_aligned_read(conf, bio);
4447 spin_lock_irq(&conf->device_lock);
4453 sh = __get_priority_stripe(conf);
4457 spin_unlock_irq(&conf->device_lock);
4464 spin_lock_irq(&conf->device_lock);
4466 pr_debug("%d stripes handled\n", handled);
4468 spin_unlock_irq(&conf->device_lock);
4470 async_tx_issue_pending_all();
4471 unplug_slaves(mddev);
4473 pr_debug("--- raid5d inactive\n");
4477 raid5_show_stripe_cache_size(mddev_t *mddev, char *page)
4479 raid5_conf_t *conf = mddev->private;
4481 return sprintf(page, "%d\n", conf->max_nr_stripes);
4487 raid5_store_stripe_cache_size(mddev_t *mddev, const char *page, size_t len)
4489 raid5_conf_t *conf = mddev->private;
4493 if (len >= PAGE_SIZE)
4498 if (strict_strtoul(page, 10, &new))
4500 if (new <= 16 || new > 32768)
4502 while (new < conf->max_nr_stripes) {
4503 if (drop_one_stripe(conf))
4504 conf->max_nr_stripes--;
4508 err = md_allow_write(mddev);
4511 while (new > conf->max_nr_stripes) {
4512 if (grow_one_stripe(conf))
4513 conf->max_nr_stripes++;
4519 static struct md_sysfs_entry
4520 raid5_stripecache_size = __ATTR(stripe_cache_size, S_IRUGO | S_IWUSR,
4521 raid5_show_stripe_cache_size,
4522 raid5_store_stripe_cache_size);
4525 raid5_show_preread_threshold(mddev_t *mddev, char *page)
4527 raid5_conf_t *conf = mddev->private;
4529 return sprintf(page, "%d\n", conf->bypass_threshold);
4535 raid5_store_preread_threshold(mddev_t *mddev, const char *page, size_t len)
4537 raid5_conf_t *conf = mddev->private;
4539 if (len >= PAGE_SIZE)
4544 if (strict_strtoul(page, 10, &new))
4546 if (new > conf->max_nr_stripes)
4548 conf->bypass_threshold = new;
4552 static struct md_sysfs_entry
4553 raid5_preread_bypass_threshold = __ATTR(preread_bypass_threshold,
4555 raid5_show_preread_threshold,
4556 raid5_store_preread_threshold);
4559 stripe_cache_active_show(mddev_t *mddev, char *page)
4561 raid5_conf_t *conf = mddev->private;
4563 return sprintf(page, "%d\n", atomic_read(&conf->active_stripes));
4568 static struct md_sysfs_entry
4569 raid5_stripecache_active = __ATTR_RO(stripe_cache_active);
4571 static struct attribute *raid5_attrs[] = {
4572 &raid5_stripecache_size.attr,
4573 &raid5_stripecache_active.attr,
4574 &raid5_preread_bypass_threshold.attr,
4577 static struct attribute_group raid5_attrs_group = {
4579 .attrs = raid5_attrs,
4583 raid5_size(mddev_t *mddev, sector_t sectors, int raid_disks)
4585 raid5_conf_t *conf = mddev->private;
4588 sectors = mddev->dev_sectors;
4590 /* size is defined by the smallest of previous and new size */
4591 raid_disks = min(conf->raid_disks, conf->previous_raid_disks);
4593 sectors &= ~((sector_t)mddev->chunk_sectors - 1);
4594 sectors &= ~((sector_t)mddev->new_chunk_sectors - 1);
4595 return sectors * (raid_disks - conf->max_degraded);
4598 static void raid5_free_percpu(raid5_conf_t *conf)
4600 struct raid5_percpu *percpu;
4607 for_each_possible_cpu(cpu) {
4608 percpu = per_cpu_ptr(conf->percpu, cpu);
4609 safe_put_page(percpu->spare_page);
4610 kfree(percpu->scribble);
4612 #ifdef CONFIG_HOTPLUG_CPU
4613 unregister_cpu_notifier(&conf->cpu_notify);
4617 free_percpu(conf->percpu);
4620 static void free_conf(raid5_conf_t *conf)
4622 shrink_stripes(conf);
4623 raid5_free_percpu(conf);
4625 kfree(conf->stripe_hashtbl);
4629 #ifdef CONFIG_HOTPLUG_CPU
4630 static int raid456_cpu_notify(struct notifier_block *nfb, unsigned long action,
4633 raid5_conf_t *conf = container_of(nfb, raid5_conf_t, cpu_notify);
4634 long cpu = (long)hcpu;
4635 struct raid5_percpu *percpu = per_cpu_ptr(conf->percpu, cpu);
4638 case CPU_UP_PREPARE:
4639 case CPU_UP_PREPARE_FROZEN:
4640 if (conf->level == 6 && !percpu->spare_page)
4641 percpu->spare_page = alloc_page(GFP_KERNEL);
4642 if (!percpu->scribble)
4643 percpu->scribble = kmalloc(conf->scribble_len, GFP_KERNEL);
4645 if (!percpu->scribble ||
4646 (conf->level == 6 && !percpu->spare_page)) {
4647 safe_put_page(percpu->spare_page);
4648 kfree(percpu->scribble);
4649 pr_err("%s: failed memory allocation for cpu%ld\n",
4655 case CPU_DEAD_FROZEN:
4656 safe_put_page(percpu->spare_page);
4657 kfree(percpu->scribble);
4658 percpu->spare_page = NULL;
4659 percpu->scribble = NULL;
4668 static int raid5_alloc_percpu(raid5_conf_t *conf)
4671 struct page *spare_page;
4672 struct raid5_percpu __percpu *allcpus;
4676 allcpus = alloc_percpu(struct raid5_percpu);
4679 conf->percpu = allcpus;
4683 for_each_present_cpu(cpu) {
4684 if (conf->level == 6) {
4685 spare_page = alloc_page(GFP_KERNEL);
4690 per_cpu_ptr(conf->percpu, cpu)->spare_page = spare_page;
4692 scribble = kmalloc(conf->scribble_len, GFP_KERNEL);
4697 per_cpu_ptr(conf->percpu, cpu)->scribble = scribble;
4699 #ifdef CONFIG_HOTPLUG_CPU
4700 conf->cpu_notify.notifier_call = raid456_cpu_notify;
4701 conf->cpu_notify.priority = 0;
4703 err = register_cpu_notifier(&conf->cpu_notify);
4710 static raid5_conf_t *setup_conf(mddev_t *mddev)
4713 int raid_disk, memory, max_disks;
4715 struct disk_info *disk;
4717 if (mddev->new_level != 5
4718 && mddev->new_level != 4
4719 && mddev->new_level != 6) {
4720 printk(KERN_ERR "md/raid:%s: raid level not set to 4/5/6 (%d)\n",
4721 mdname(mddev), mddev->new_level);
4722 return ERR_PTR(-EIO);
4724 if ((mddev->new_level == 5
4725 && !algorithm_valid_raid5(mddev->new_layout)) ||
4726 (mddev->new_level == 6
4727 && !algorithm_valid_raid6(mddev->new_layout))) {
4728 printk(KERN_ERR "md/raid:%s: layout %d not supported\n",
4729 mdname(mddev), mddev->new_layout);
4730 return ERR_PTR(-EIO);
4732 if (mddev->new_level == 6 && mddev->raid_disks < 4) {
4733 printk(KERN_ERR "md/raid:%s: not enough configured devices (%d, minimum 4)\n",
4734 mdname(mddev), mddev->raid_disks);
4735 return ERR_PTR(-EINVAL);
4738 if (!mddev->new_chunk_sectors ||
4739 (mddev->new_chunk_sectors << 9) % PAGE_SIZE ||
4740 !is_power_of_2(mddev->new_chunk_sectors)) {
4741 printk(KERN_ERR "md/raid:%s: invalid chunk size %d\n",
4742 mdname(mddev), mddev->new_chunk_sectors << 9);
4743 return ERR_PTR(-EINVAL);
4746 conf = kzalloc(sizeof(raid5_conf_t), GFP_KERNEL);
4749 spin_lock_init(&conf->device_lock);
4750 init_waitqueue_head(&conf->wait_for_stripe);
4751 init_waitqueue_head(&conf->wait_for_overlap);
4752 INIT_LIST_HEAD(&conf->handle_list);
4753 INIT_LIST_HEAD(&conf->hold_list);
4754 INIT_LIST_HEAD(&conf->delayed_list);
4755 INIT_LIST_HEAD(&conf->bitmap_list);
4756 INIT_LIST_HEAD(&conf->inactive_list);
4757 atomic_set(&conf->active_stripes, 0);
4758 atomic_set(&conf->preread_active_stripes, 0);
4759 atomic_set(&conf->active_aligned_reads, 0);
4760 conf->bypass_threshold = BYPASS_THRESHOLD;
4762 conf->raid_disks = mddev->raid_disks;
4763 if (mddev->reshape_position == MaxSector)
4764 conf->previous_raid_disks = mddev->raid_disks;
4766 conf->previous_raid_disks = mddev->raid_disks - mddev->delta_disks;
4767 max_disks = max(conf->raid_disks, conf->previous_raid_disks);
4768 conf->scribble_len = scribble_len(max_disks);
4770 conf->disks = kzalloc(max_disks * sizeof(struct disk_info),
4775 conf->mddev = mddev;
4777 if ((conf->stripe_hashtbl = kzalloc(PAGE_SIZE, GFP_KERNEL)) == NULL)
4780 conf->level = mddev->new_level;
4781 if (raid5_alloc_percpu(conf) != 0)
4784 pr_debug("raid456: run(%s) called.\n", mdname(mddev));
4786 list_for_each_entry(rdev, &mddev->disks, same_set) {
4787 raid_disk = rdev->raid_disk;
4788 if (raid_disk >= max_disks
4791 disk = conf->disks + raid_disk;
4795 if (test_bit(In_sync, &rdev->flags)) {
4796 char b[BDEVNAME_SIZE];
4797 printk(KERN_INFO "md/raid:%s: device %s operational as raid"
4799 mdname(mddev), bdevname(rdev->bdev, b), raid_disk);
4801 /* Cannot rely on bitmap to complete recovery */
4805 conf->chunk_sectors = mddev->new_chunk_sectors;
4806 conf->level = mddev->new_level;
4807 if (conf->level == 6)
4808 conf->max_degraded = 2;
4810 conf->max_degraded = 1;
4811 conf->algorithm = mddev->new_layout;
4812 conf->max_nr_stripes = NR_STRIPES;
4813 conf->reshape_progress = mddev->reshape_position;
4814 if (conf->reshape_progress != MaxSector) {
4815 conf->prev_chunk_sectors = mddev->chunk_sectors;
4816 conf->prev_algo = mddev->layout;
4819 memory = conf->max_nr_stripes * (sizeof(struct stripe_head) +
4820 max_disks * ((sizeof(struct bio) + PAGE_SIZE))) / 1024;
4821 if (grow_stripes(conf, conf->max_nr_stripes)) {
4823 "md/raid:%s: couldn't allocate %dkB for buffers\n",
4824 mdname(mddev), memory);
4827 printk(KERN_INFO "md/raid:%s: allocated %dkB\n",
4828 mdname(mddev), memory);
4830 conf->thread = md_register_thread(raid5d, mddev, NULL);
4831 if (!conf->thread) {
4833 "md/raid:%s: couldn't allocate thread.\n",
4843 return ERR_PTR(-EIO);
4845 return ERR_PTR(-ENOMEM);
4849 static int only_parity(int raid_disk, int algo, int raid_disks, int max_degraded)
4852 case ALGORITHM_PARITY_0:
4853 if (raid_disk < max_degraded)
4856 case ALGORITHM_PARITY_N:
4857 if (raid_disk >= raid_disks - max_degraded)
4860 case ALGORITHM_PARITY_0_6:
4861 if (raid_disk == 0 ||
4862 raid_disk == raid_disks - 1)
4865 case ALGORITHM_LEFT_ASYMMETRIC_6:
4866 case ALGORITHM_RIGHT_ASYMMETRIC_6:
4867 case ALGORITHM_LEFT_SYMMETRIC_6:
4868 case ALGORITHM_RIGHT_SYMMETRIC_6:
4869 if (raid_disk == raid_disks - 1)
4875 static int run(mddev_t *mddev)
4878 int working_disks = 0, chunk_size;
4879 int dirty_parity_disks = 0;
4881 sector_t reshape_offset = 0;
4883 if (mddev->recovery_cp != MaxSector)
4884 printk(KERN_NOTICE "md/raid:%s: not clean"
4885 " -- starting background reconstruction\n",
4887 if (mddev->reshape_position != MaxSector) {
4888 /* Check that we can continue the reshape.
4889 * Currently only disks can change, it must
4890 * increase, and we must be past the point where
4891 * a stripe over-writes itself
4893 sector_t here_new, here_old;
4895 int max_degraded = (mddev->level == 6 ? 2 : 1);
4897 if (mddev->new_level != mddev->level) {
4898 printk(KERN_ERR "md/raid:%s: unsupported reshape "
4899 "required - aborting.\n",
4903 old_disks = mddev->raid_disks - mddev->delta_disks;
4904 /* reshape_position must be on a new-stripe boundary, and one
4905 * further up in new geometry must map after here in old
4908 here_new = mddev->reshape_position;
4909 if (sector_div(here_new, mddev->new_chunk_sectors *
4910 (mddev->raid_disks - max_degraded))) {
4911 printk(KERN_ERR "md/raid:%s: reshape_position not "
4912 "on a stripe boundary\n", mdname(mddev));
4915 reshape_offset = here_new * mddev->new_chunk_sectors;
4916 /* here_new is the stripe we will write to */
4917 here_old = mddev->reshape_position;
4918 sector_div(here_old, mddev->chunk_sectors *
4919 (old_disks-max_degraded));
4920 /* here_old is the first stripe that we might need to read
4922 if (mddev->delta_disks == 0) {
4923 /* We cannot be sure it is safe to start an in-place
4924 * reshape. It is only safe if user-space if monitoring
4925 * and taking constant backups.
4926 * mdadm always starts a situation like this in
4927 * readonly mode so it can take control before
4928 * allowing any writes. So just check for that.
4930 if ((here_new * mddev->new_chunk_sectors !=
4931 here_old * mddev->chunk_sectors) ||
4933 printk(KERN_ERR "md/raid:%s: in-place reshape must be started"
4934 " in read-only mode - aborting\n",
4938 } else if (mddev->delta_disks < 0
4939 ? (here_new * mddev->new_chunk_sectors <=
4940 here_old * mddev->chunk_sectors)
4941 : (here_new * mddev->new_chunk_sectors >=
4942 here_old * mddev->chunk_sectors)) {
4943 /* Reading from the same stripe as writing to - bad */
4944 printk(KERN_ERR "md/raid:%s: reshape_position too early for "
4945 "auto-recovery - aborting.\n",
4949 printk(KERN_INFO "md/raid:%s: reshape will continue\n",
4951 /* OK, we should be able to continue; */
4953 BUG_ON(mddev->level != mddev->new_level);
4954 BUG_ON(mddev->layout != mddev->new_layout);
4955 BUG_ON(mddev->chunk_sectors != mddev->new_chunk_sectors);
4956 BUG_ON(mddev->delta_disks != 0);
4959 if (mddev->private == NULL)
4960 conf = setup_conf(mddev);
4962 conf = mddev->private;
4965 return PTR_ERR(conf);
4967 mddev->thread = conf->thread;
4968 conf->thread = NULL;
4969 mddev->private = conf;
4972 * 0 for a fully functional array, 1 or 2 for a degraded array.
4974 list_for_each_entry(rdev, &mddev->disks, same_set) {
4975 if (rdev->raid_disk < 0)
4977 if (test_bit(In_sync, &rdev->flags))
4979 /* This disc is not fully in-sync. However if it
4980 * just stored parity (beyond the recovery_offset),
4981 * when we don't need to be concerned about the
4982 * array being dirty.
4983 * When reshape goes 'backwards', we never have
4984 * partially completed devices, so we only need
4985 * to worry about reshape going forwards.
4987 /* Hack because v0.91 doesn't store recovery_offset properly. */
4988 if (mddev->major_version == 0 &&
4989 mddev->minor_version > 90)
4990 rdev->recovery_offset = reshape_offset;
4992 if (rdev->recovery_offset < reshape_offset) {
4993 /* We need to check old and new layout */
4994 if (!only_parity(rdev->raid_disk,
4997 conf->max_degraded))
5000 if (!only_parity(rdev->raid_disk,
5002 conf->previous_raid_disks,
5003 conf->max_degraded))
5005 dirty_parity_disks++;
5008 mddev->degraded = (max(conf->raid_disks, conf->previous_raid_disks)
5011 if (mddev->degraded > conf->max_degraded) {
5012 printk(KERN_ERR "md/raid:%s: not enough operational devices"
5013 " (%d/%d failed)\n",
5014 mdname(mddev), mddev->degraded, conf->raid_disks);
5018 /* device size must be a multiple of chunk size */
5019 mddev->dev_sectors &= ~(mddev->chunk_sectors - 1);
5020 mddev->resync_max_sectors = mddev->dev_sectors;
5022 if (mddev->degraded > dirty_parity_disks &&
5023 mddev->recovery_cp != MaxSector) {
5024 if (mddev->ok_start_degraded)
5026 "md/raid:%s: starting dirty degraded array"
5027 " - data corruption possible.\n",
5031 "md/raid:%s: cannot start dirty degraded array.\n",
5037 if (mddev->degraded == 0)
5038 printk(KERN_INFO "md/raid:%s: raid level %d active with %d out of %d"
5039 " devices, algorithm %d\n", mdname(mddev), conf->level,
5040 mddev->raid_disks-mddev->degraded, mddev->raid_disks,
5043 printk(KERN_ALERT "md/raid:%s: raid level %d active with %d"
5044 " out of %d devices, algorithm %d\n",
5045 mdname(mddev), conf->level,
5046 mddev->raid_disks - mddev->degraded,
5047 mddev->raid_disks, mddev->new_layout);
5049 print_raid5_conf(conf);
5051 if (conf->reshape_progress != MaxSector) {
5052 conf->reshape_safe = conf->reshape_progress;
5053 atomic_set(&conf->reshape_stripes, 0);
5054 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
5055 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
5056 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
5057 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
5058 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
5062 /* read-ahead size must cover two whole stripes, which is
5063 * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
5066 int data_disks = conf->previous_raid_disks - conf->max_degraded;
5067 int stripe = data_disks *
5068 ((mddev->chunk_sectors << 9) / PAGE_SIZE);
5069 if (mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
5070 mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
5073 /* Ok, everything is just fine now */
5074 if (mddev->to_remove == &raid5_attrs_group)
5075 mddev->to_remove = NULL;
5076 else if (sysfs_create_group(&mddev->kobj, &raid5_attrs_group))
5078 "md/raid:%s: failed to create sysfs attributes.\n",
5081 mddev->queue->queue_lock = &conf->device_lock;
5083 mddev->queue->unplug_fn = raid5_unplug_device;
5084 mddev->queue->backing_dev_info.congested_data = mddev;
5085 mddev->queue->backing_dev_info.congested_fn = raid5_congested;
5087 md_set_array_sectors(mddev, raid5_size(mddev, 0, 0));
5089 blk_queue_merge_bvec(mddev->queue, raid5_mergeable_bvec);
5090 chunk_size = mddev->chunk_sectors << 9;
5091 blk_queue_io_min(mddev->queue, chunk_size);
5092 blk_queue_io_opt(mddev->queue, chunk_size *
5093 (conf->raid_disks - conf->max_degraded));
5095 list_for_each_entry(rdev, &mddev->disks, same_set)
5096 disk_stack_limits(mddev->gendisk, rdev->bdev,
5097 rdev->data_offset << 9);
5101 md_unregister_thread(mddev->thread);
5102 mddev->thread = NULL;
5104 print_raid5_conf(conf);
5107 mddev->private = NULL;
5108 printk(KERN_ALERT "md/raid:%s: failed to run raid set.\n", mdname(mddev));
5112 static int stop(mddev_t *mddev)
5114 raid5_conf_t *conf = mddev->private;
5116 md_unregister_thread(mddev->thread);
5117 mddev->thread = NULL;
5118 mddev->queue->backing_dev_info.congested_fn = NULL;
5119 blk_sync_queue(mddev->queue); /* the unplug fn references 'conf'*/
5121 mddev->private = NULL;
5122 mddev->to_remove = &raid5_attrs_group;
5127 static void print_sh(struct seq_file *seq, struct stripe_head *sh)
5131 seq_printf(seq, "sh %llu, pd_idx %d, state %ld.\n",
5132 (unsigned long long)sh->sector, sh->pd_idx, sh->state);
5133 seq_printf(seq, "sh %llu, count %d.\n",
5134 (unsigned long long)sh->sector, atomic_read(&sh->count));
5135 seq_printf(seq, "sh %llu, ", (unsigned long long)sh->sector);
5136 for (i = 0; i < sh->disks; i++) {
5137 seq_printf(seq, "(cache%d: %p %ld) ",
5138 i, sh->dev[i].page, sh->dev[i].flags);
5140 seq_printf(seq, "\n");
5143 static void printall(struct seq_file *seq, raid5_conf_t *conf)
5145 struct stripe_head *sh;
5146 struct hlist_node *hn;
5149 spin_lock_irq(&conf->device_lock);
5150 for (i = 0; i < NR_HASH; i++) {
5151 hlist_for_each_entry(sh, hn, &conf->stripe_hashtbl[i], hash) {
5152 if (sh->raid_conf != conf)
5157 spin_unlock_irq(&conf->device_lock);
5161 static void status(struct seq_file *seq, mddev_t *mddev)
5163 raid5_conf_t *conf = mddev->private;
5166 seq_printf(seq, " level %d, %dk chunk, algorithm %d", mddev->level,
5167 mddev->chunk_sectors / 2, mddev->layout);
5168 seq_printf (seq, " [%d/%d] [", conf->raid_disks, conf->raid_disks - mddev->degraded);
5169 for (i = 0; i < conf->raid_disks; i++)
5170 seq_printf (seq, "%s",
5171 conf->disks[i].rdev &&
5172 test_bit(In_sync, &conf->disks[i].rdev->flags) ? "U" : "_");
5173 seq_printf (seq, "]");
5175 seq_printf (seq, "\n");
5176 printall(seq, conf);
5180 static void print_raid5_conf (raid5_conf_t *conf)
5183 struct disk_info *tmp;
5185 printk(KERN_DEBUG "RAID conf printout:\n");
5187 printk("(conf==NULL)\n");
5190 printk(KERN_DEBUG " --- level:%d rd:%d wd:%d\n", conf->level,
5192 conf->raid_disks - conf->mddev->degraded);
5194 for (i = 0; i < conf->raid_disks; i++) {
5195 char b[BDEVNAME_SIZE];
5196 tmp = conf->disks + i;
5198 printk(KERN_DEBUG " disk %d, o:%d, dev:%s\n",
5199 i, !test_bit(Faulty, &tmp->rdev->flags),
5200 bdevname(tmp->rdev->bdev, b));
5204 static int raid5_spare_active(mddev_t *mddev)
5207 raid5_conf_t *conf = mddev->private;
5208 struct disk_info *tmp;
5210 for (i = 0; i < conf->raid_disks; i++) {
5211 tmp = conf->disks + i;
5213 && !test_bit(Faulty, &tmp->rdev->flags)
5214 && !test_and_set_bit(In_sync, &tmp->rdev->flags)) {
5215 unsigned long flags;
5216 spin_lock_irqsave(&conf->device_lock, flags);
5218 spin_unlock_irqrestore(&conf->device_lock, flags);
5221 print_raid5_conf(conf);
5225 static int raid5_remove_disk(mddev_t *mddev, int number)
5227 raid5_conf_t *conf = mddev->private;
5230 struct disk_info *p = conf->disks + number;
5232 print_raid5_conf(conf);
5235 if (number >= conf->raid_disks &&
5236 conf->reshape_progress == MaxSector)
5237 clear_bit(In_sync, &rdev->flags);
5239 if (test_bit(In_sync, &rdev->flags) ||
5240 atomic_read(&rdev->nr_pending)) {
5244 /* Only remove non-faulty devices if recovery
5247 if (!test_bit(Faulty, &rdev->flags) &&
5248 mddev->degraded <= conf->max_degraded &&
5249 number < conf->raid_disks) {
5255 if (atomic_read(&rdev->nr_pending)) {
5256 /* lost the race, try later */
5263 print_raid5_conf(conf);
5267 static int raid5_add_disk(mddev_t *mddev, mdk_rdev_t *rdev)
5269 raid5_conf_t *conf = mddev->private;
5272 struct disk_info *p;
5274 int last = conf->raid_disks - 1;
5276 if (mddev->degraded > conf->max_degraded)
5277 /* no point adding a device */
5280 if (rdev->raid_disk >= 0)
5281 first = last = rdev->raid_disk;
5284 * find the disk ... but prefer rdev->saved_raid_disk
5287 if (rdev->saved_raid_disk >= 0 &&
5288 rdev->saved_raid_disk >= first &&
5289 conf->disks[rdev->saved_raid_disk].rdev == NULL)
5290 disk = rdev->saved_raid_disk;
5293 for ( ; disk <= last ; disk++)
5294 if ((p=conf->disks + disk)->rdev == NULL) {
5295 clear_bit(In_sync, &rdev->flags);
5296 rdev->raid_disk = disk;
5298 if (rdev->saved_raid_disk != disk)
5300 rcu_assign_pointer(p->rdev, rdev);
5303 print_raid5_conf(conf);
5307 static int raid5_resize(mddev_t *mddev, sector_t sectors)
5309 /* no resync is happening, and there is enough space
5310 * on all devices, so we can resize.
5311 * We need to make sure resync covers any new space.
5312 * If the array is shrinking we should possibly wait until
5313 * any io in the removed space completes, but it hardly seems
5316 sectors &= ~((sector_t)mddev->chunk_sectors - 1);
5317 md_set_array_sectors(mddev, raid5_size(mddev, sectors,
5318 mddev->raid_disks));
5319 if (mddev->array_sectors >
5320 raid5_size(mddev, sectors, mddev->raid_disks))
5322 set_capacity(mddev->gendisk, mddev->array_sectors);
5323 revalidate_disk(mddev->gendisk);
5324 if (sectors > mddev->dev_sectors && mddev->recovery_cp == MaxSector) {
5325 mddev->recovery_cp = mddev->dev_sectors;
5326 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
5328 mddev->dev_sectors = sectors;
5329 mddev->resync_max_sectors = sectors;
5333 static int check_stripe_cache(mddev_t *mddev)
5335 /* Can only proceed if there are plenty of stripe_heads.
5336 * We need a minimum of one full stripe,, and for sensible progress
5337 * it is best to have about 4 times that.
5338 * If we require 4 times, then the default 256 4K stripe_heads will
5339 * allow for chunk sizes up to 256K, which is probably OK.
5340 * If the chunk size is greater, user-space should request more
5341 * stripe_heads first.
5343 raid5_conf_t *conf = mddev->private;
5344 if (((mddev->chunk_sectors << 9) / STRIPE_SIZE) * 4
5345 > conf->max_nr_stripes ||
5346 ((mddev->new_chunk_sectors << 9) / STRIPE_SIZE) * 4
5347 > conf->max_nr_stripes) {
5348 printk(KERN_WARNING "md/raid:%s: reshape: not enough stripes. Needed %lu\n",
5350 ((max(mddev->chunk_sectors, mddev->new_chunk_sectors) << 9)
5357 static int check_reshape(mddev_t *mddev)
5359 raid5_conf_t *conf = mddev->private;
5361 if (mddev->delta_disks == 0 &&
5362 mddev->new_layout == mddev->layout &&
5363 mddev->new_chunk_sectors == mddev->chunk_sectors)
5364 return 0; /* nothing to do */
5366 /* Cannot grow a bitmap yet */
5368 if (mddev->degraded > conf->max_degraded)
5370 if (mddev->delta_disks < 0) {
5371 /* We might be able to shrink, but the devices must
5372 * be made bigger first.
5373 * For raid6, 4 is the minimum size.
5374 * Otherwise 2 is the minimum
5377 if (mddev->level == 6)
5379 if (mddev->raid_disks + mddev->delta_disks < min)
5383 if (!check_stripe_cache(mddev))
5386 return resize_stripes(conf, conf->raid_disks + mddev->delta_disks);
5389 static int raid5_start_reshape(mddev_t *mddev)
5391 raid5_conf_t *conf = mddev->private;
5394 int added_devices = 0;
5395 unsigned long flags;
5397 if (test_bit(MD_RECOVERY_RUNNING, &mddev->recovery))
5400 if (!check_stripe_cache(mddev))
5403 list_for_each_entry(rdev, &mddev->disks, same_set)
5404 if (rdev->raid_disk < 0 &&
5405 !test_bit(Faulty, &rdev->flags))
5408 if (spares - mddev->degraded < mddev->delta_disks - conf->max_degraded)
5409 /* Not enough devices even to make a degraded array
5414 /* Refuse to reduce size of the array. Any reductions in
5415 * array size must be through explicit setting of array_size
5418 if (raid5_size(mddev, 0, conf->raid_disks + mddev->delta_disks)
5419 < mddev->array_sectors) {
5420 printk(KERN_ERR "md/raid:%s: array size must be reduced "
5421 "before number of disks\n", mdname(mddev));
5425 atomic_set(&conf->reshape_stripes, 0);
5426 spin_lock_irq(&conf->device_lock);
5427 conf->previous_raid_disks = conf->raid_disks;
5428 conf->raid_disks += mddev->delta_disks;
5429 conf->prev_chunk_sectors = conf->chunk_sectors;
5430 conf->chunk_sectors = mddev->new_chunk_sectors;
5431 conf->prev_algo = conf->algorithm;
5432 conf->algorithm = mddev->new_layout;
5433 if (mddev->delta_disks < 0)
5434 conf->reshape_progress = raid5_size(mddev, 0, 0);
5436 conf->reshape_progress = 0;
5437 conf->reshape_safe = conf->reshape_progress;
5439 spin_unlock_irq(&conf->device_lock);
5441 /* Add some new drives, as many as will fit.
5442 * We know there are enough to make the newly sized array work.
5444 list_for_each_entry(rdev, &mddev->disks, same_set)
5445 if (rdev->raid_disk < 0 &&
5446 !test_bit(Faulty, &rdev->flags)) {
5447 if (raid5_add_disk(mddev, rdev) == 0) {
5449 if (rdev->raid_disk >= conf->previous_raid_disks) {
5450 set_bit(In_sync, &rdev->flags);
5453 rdev->recovery_offset = 0;
5454 sprintf(nm, "rd%d", rdev->raid_disk);
5455 if (sysfs_create_link(&mddev->kobj,
5458 "md/raid:%s: failed to create "
5465 /* When a reshape changes the number of devices, ->degraded
5466 * is measured against the large of the pre and post number of
5468 if (mddev->delta_disks > 0) {
5469 spin_lock_irqsave(&conf->device_lock, flags);
5470 mddev->degraded += (conf->raid_disks - conf->previous_raid_disks)
5472 spin_unlock_irqrestore(&conf->device_lock, flags);
5474 mddev->raid_disks = conf->raid_disks;
5475 mddev->reshape_position = conf->reshape_progress;
5476 set_bit(MD_CHANGE_DEVS, &mddev->flags);
5478 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
5479 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
5480 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
5481 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
5482 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
5484 if (!mddev->sync_thread) {
5485 mddev->recovery = 0;
5486 spin_lock_irq(&conf->device_lock);
5487 mddev->raid_disks = conf->raid_disks = conf->previous_raid_disks;
5488 conf->reshape_progress = MaxSector;
5489 spin_unlock_irq(&conf->device_lock);
5492 conf->reshape_checkpoint = jiffies;
5493 md_wakeup_thread(mddev->sync_thread);
5494 md_new_event(mddev);
5498 /* This is called from the reshape thread and should make any
5499 * changes needed in 'conf'
5501 static void end_reshape(raid5_conf_t *conf)
5504 if (!test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery)) {
5506 spin_lock_irq(&conf->device_lock);
5507 conf->previous_raid_disks = conf->raid_disks;
5508 conf->reshape_progress = MaxSector;
5509 spin_unlock_irq(&conf->device_lock);
5510 wake_up(&conf->wait_for_overlap);
5512 /* read-ahead size must cover two whole stripes, which is
5513 * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
5516 int data_disks = conf->raid_disks - conf->max_degraded;
5517 int stripe = data_disks * ((conf->chunk_sectors << 9)
5519 if (conf->mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
5520 conf->mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
5525 /* This is called from the raid5d thread with mddev_lock held.
5526 * It makes config changes to the device.
5528 static void raid5_finish_reshape(mddev_t *mddev)
5530 raid5_conf_t *conf = mddev->private;
5532 if (!test_bit(MD_RECOVERY_INTR, &mddev->recovery)) {
5534 if (mddev->delta_disks > 0) {
5535 md_set_array_sectors(mddev, raid5_size(mddev, 0, 0));
5536 set_capacity(mddev->gendisk, mddev->array_sectors);
5537 revalidate_disk(mddev->gendisk);
5540 mddev->degraded = conf->raid_disks;
5541 for (d = 0; d < conf->raid_disks ; d++)
5542 if (conf->disks[d].rdev &&
5544 &conf->disks[d].rdev->flags))
5546 for (d = conf->raid_disks ;
5547 d < conf->raid_disks - mddev->delta_disks;
5549 mdk_rdev_t *rdev = conf->disks[d].rdev;
5550 if (rdev && raid5_remove_disk(mddev, d) == 0) {
5552 sprintf(nm, "rd%d", rdev->raid_disk);
5553 sysfs_remove_link(&mddev->kobj, nm);
5554 rdev->raid_disk = -1;
5558 mddev->layout = conf->algorithm;
5559 mddev->chunk_sectors = conf->chunk_sectors;
5560 mddev->reshape_position = MaxSector;
5561 mddev->delta_disks = 0;
5565 static void raid5_quiesce(mddev_t *mddev, int state)
5567 raid5_conf_t *conf = mddev->private;
5570 case 2: /* resume for a suspend */
5571 wake_up(&conf->wait_for_overlap);
5574 case 1: /* stop all writes */
5575 spin_lock_irq(&conf->device_lock);
5576 /* '2' tells resync/reshape to pause so that all
5577 * active stripes can drain
5580 wait_event_lock_irq(conf->wait_for_stripe,
5581 atomic_read(&conf->active_stripes) == 0 &&
5582 atomic_read(&conf->active_aligned_reads) == 0,
5583 conf->device_lock, /* nothing */);
5585 spin_unlock_irq(&conf->device_lock);
5586 /* allow reshape to continue */
5587 wake_up(&conf->wait_for_overlap);
5590 case 0: /* re-enable writes */
5591 spin_lock_irq(&conf->device_lock);
5593 wake_up(&conf->wait_for_stripe);
5594 wake_up(&conf->wait_for_overlap);
5595 spin_unlock_irq(&conf->device_lock);
5601 static void *raid45_takeover_raid0(mddev_t *mddev, int level)
5603 struct raid0_private_data *raid0_priv = mddev->private;
5605 /* for raid0 takeover only one zone is supported */
5606 if (raid0_priv->nr_strip_zones > 1) {
5607 printk(KERN_ERR "md/raid:%s: cannot takeover raid0 with more than one zone.\n",
5609 return ERR_PTR(-EINVAL);
5612 mddev->new_level = level;
5613 mddev->new_layout = ALGORITHM_PARITY_N;
5614 mddev->new_chunk_sectors = mddev->chunk_sectors;
5615 mddev->raid_disks += 1;
5616 mddev->delta_disks = 1;
5617 /* make sure it will be not marked as dirty */
5618 mddev->recovery_cp = MaxSector;
5620 return setup_conf(mddev);
5624 static void *raid5_takeover_raid1(mddev_t *mddev)
5628 if (mddev->raid_disks != 2 ||
5629 mddev->degraded > 1)
5630 return ERR_PTR(-EINVAL);
5632 /* Should check if there are write-behind devices? */
5634 chunksect = 64*2; /* 64K by default */
5636 /* The array must be an exact multiple of chunksize */
5637 while (chunksect && (mddev->array_sectors & (chunksect-1)))
5640 if ((chunksect<<9) < STRIPE_SIZE)
5641 /* array size does not allow a suitable chunk size */
5642 return ERR_PTR(-EINVAL);
5644 mddev->new_level = 5;
5645 mddev->new_layout = ALGORITHM_LEFT_SYMMETRIC;
5646 mddev->new_chunk_sectors = chunksect;
5648 return setup_conf(mddev);
5651 static void *raid5_takeover_raid6(mddev_t *mddev)
5655 switch (mddev->layout) {
5656 case ALGORITHM_LEFT_ASYMMETRIC_6:
5657 new_layout = ALGORITHM_LEFT_ASYMMETRIC;
5659 case ALGORITHM_RIGHT_ASYMMETRIC_6:
5660 new_layout = ALGORITHM_RIGHT_ASYMMETRIC;
5662 case ALGORITHM_LEFT_SYMMETRIC_6:
5663 new_layout = ALGORITHM_LEFT_SYMMETRIC;
5665 case ALGORITHM_RIGHT_SYMMETRIC_6:
5666 new_layout = ALGORITHM_RIGHT_SYMMETRIC;
5668 case ALGORITHM_PARITY_0_6:
5669 new_layout = ALGORITHM_PARITY_0;
5671 case ALGORITHM_PARITY_N:
5672 new_layout = ALGORITHM_PARITY_N;
5675 return ERR_PTR(-EINVAL);
5677 mddev->new_level = 5;
5678 mddev->new_layout = new_layout;
5679 mddev->delta_disks = -1;
5680 mddev->raid_disks -= 1;
5681 return setup_conf(mddev);
5685 static int raid5_check_reshape(mddev_t *mddev)
5687 /* For a 2-drive array, the layout and chunk size can be changed
5688 * immediately as not restriping is needed.
5689 * For larger arrays we record the new value - after validation
5690 * to be used by a reshape pass.
5692 raid5_conf_t *conf = mddev->private;
5693 int new_chunk = mddev->new_chunk_sectors;
5695 if (mddev->new_layout >= 0 && !algorithm_valid_raid5(mddev->new_layout))
5697 if (new_chunk > 0) {
5698 if (!is_power_of_2(new_chunk))
5700 if (new_chunk < (PAGE_SIZE>>9))
5702 if (mddev->array_sectors & (new_chunk-1))
5703 /* not factor of array size */
5707 /* They look valid */
5709 if (mddev->raid_disks == 2) {
5710 /* can make the change immediately */
5711 if (mddev->new_layout >= 0) {
5712 conf->algorithm = mddev->new_layout;
5713 mddev->layout = mddev->new_layout;
5715 if (new_chunk > 0) {
5716 conf->chunk_sectors = new_chunk ;
5717 mddev->chunk_sectors = new_chunk;
5719 set_bit(MD_CHANGE_DEVS, &mddev->flags);
5720 md_wakeup_thread(mddev->thread);
5722 return check_reshape(mddev);
5725 static int raid6_check_reshape(mddev_t *mddev)
5727 int new_chunk = mddev->new_chunk_sectors;
5729 if (mddev->new_layout >= 0 && !algorithm_valid_raid6(mddev->new_layout))
5731 if (new_chunk > 0) {
5732 if (!is_power_of_2(new_chunk))
5734 if (new_chunk < (PAGE_SIZE >> 9))
5736 if (mddev->array_sectors & (new_chunk-1))
5737 /* not factor of array size */
5741 /* They look valid */
5742 return check_reshape(mddev);
5745 static void *raid5_takeover(mddev_t *mddev)
5747 /* raid5 can take over:
5748 * raid0 - if there is only one strip zone - make it a raid4 layout
5749 * raid1 - if there are two drives. We need to know the chunk size
5750 * raid4 - trivial - just use a raid4 layout.
5751 * raid6 - Providing it is a *_6 layout
5753 if (mddev->level == 0)
5754 return raid45_takeover_raid0(mddev, 5);
5755 if (mddev->level == 1)
5756 return raid5_takeover_raid1(mddev);
5757 if (mddev->level == 4) {
5758 mddev->new_layout = ALGORITHM_PARITY_N;
5759 mddev->new_level = 5;
5760 return setup_conf(mddev);
5762 if (mddev->level == 6)
5763 return raid5_takeover_raid6(mddev);
5765 return ERR_PTR(-EINVAL);
5768 static void *raid4_takeover(mddev_t *mddev)
5770 /* raid4 can take over:
5771 * raid0 - if there is only one strip zone
5772 * raid5 - if layout is right
5774 if (mddev->level == 0)
5775 return raid45_takeover_raid0(mddev, 4);
5776 if (mddev->level == 5 &&
5777 mddev->layout == ALGORITHM_PARITY_N) {
5778 mddev->new_layout = 0;
5779 mddev->new_level = 4;
5780 return setup_conf(mddev);
5782 return ERR_PTR(-EINVAL);
5785 static struct mdk_personality raid5_personality;
5787 static void *raid6_takeover(mddev_t *mddev)
5789 /* Currently can only take over a raid5. We map the
5790 * personality to an equivalent raid6 personality
5791 * with the Q block at the end.
5795 if (mddev->pers != &raid5_personality)
5796 return ERR_PTR(-EINVAL);
5797 if (mddev->degraded > 1)
5798 return ERR_PTR(-EINVAL);
5799 if (mddev->raid_disks > 253)
5800 return ERR_PTR(-EINVAL);
5801 if (mddev->raid_disks < 3)
5802 return ERR_PTR(-EINVAL);
5804 switch (mddev->layout) {
5805 case ALGORITHM_LEFT_ASYMMETRIC:
5806 new_layout = ALGORITHM_LEFT_ASYMMETRIC_6;
5808 case ALGORITHM_RIGHT_ASYMMETRIC:
5809 new_layout = ALGORITHM_RIGHT_ASYMMETRIC_6;
5811 case ALGORITHM_LEFT_SYMMETRIC:
5812 new_layout = ALGORITHM_LEFT_SYMMETRIC_6;
5814 case ALGORITHM_RIGHT_SYMMETRIC:
5815 new_layout = ALGORITHM_RIGHT_SYMMETRIC_6;
5817 case ALGORITHM_PARITY_0:
5818 new_layout = ALGORITHM_PARITY_0_6;
5820 case ALGORITHM_PARITY_N:
5821 new_layout = ALGORITHM_PARITY_N;
5824 return ERR_PTR(-EINVAL);
5826 mddev->new_level = 6;
5827 mddev->new_layout = new_layout;
5828 mddev->delta_disks = 1;
5829 mddev->raid_disks += 1;
5830 return setup_conf(mddev);
5834 static struct mdk_personality raid6_personality =
5838 .owner = THIS_MODULE,
5839 .make_request = make_request,
5843 .error_handler = error,
5844 .hot_add_disk = raid5_add_disk,
5845 .hot_remove_disk= raid5_remove_disk,
5846 .spare_active = raid5_spare_active,
5847 .sync_request = sync_request,
5848 .resize = raid5_resize,
5850 .check_reshape = raid6_check_reshape,
5851 .start_reshape = raid5_start_reshape,
5852 .finish_reshape = raid5_finish_reshape,
5853 .quiesce = raid5_quiesce,
5854 .takeover = raid6_takeover,
5856 static struct mdk_personality raid5_personality =
5860 .owner = THIS_MODULE,
5861 .make_request = make_request,
5865 .error_handler = error,
5866 .hot_add_disk = raid5_add_disk,
5867 .hot_remove_disk= raid5_remove_disk,
5868 .spare_active = raid5_spare_active,
5869 .sync_request = sync_request,
5870 .resize = raid5_resize,
5872 .check_reshape = raid5_check_reshape,
5873 .start_reshape = raid5_start_reshape,
5874 .finish_reshape = raid5_finish_reshape,
5875 .quiesce = raid5_quiesce,
5876 .takeover = raid5_takeover,
5879 static struct mdk_personality raid4_personality =
5883 .owner = THIS_MODULE,
5884 .make_request = make_request,
5888 .error_handler = error,
5889 .hot_add_disk = raid5_add_disk,
5890 .hot_remove_disk= raid5_remove_disk,
5891 .spare_active = raid5_spare_active,
5892 .sync_request = sync_request,
5893 .resize = raid5_resize,
5895 .check_reshape = raid5_check_reshape,
5896 .start_reshape = raid5_start_reshape,
5897 .finish_reshape = raid5_finish_reshape,
5898 .quiesce = raid5_quiesce,
5899 .takeover = raid4_takeover,
5902 static int __init raid5_init(void)
5904 register_md_personality(&raid6_personality);
5905 register_md_personality(&raid5_personality);
5906 register_md_personality(&raid4_personality);
5910 static void raid5_exit(void)
5912 unregister_md_personality(&raid6_personality);
5913 unregister_md_personality(&raid5_personality);
5914 unregister_md_personality(&raid4_personality);
5917 module_init(raid5_init);
5918 module_exit(raid5_exit);
5919 MODULE_LICENSE("GPL");
5920 MODULE_DESCRIPTION("RAID4/5/6 (striping with parity) personality for MD");
5921 MODULE_ALIAS("md-personality-4"); /* RAID5 */
5922 MODULE_ALIAS("md-raid5");
5923 MODULE_ALIAS("md-raid4");
5924 MODULE_ALIAS("md-level-5");
5925 MODULE_ALIAS("md-level-4");
5926 MODULE_ALIAS("md-personality-8"); /* RAID6 */
5927 MODULE_ALIAS("md-raid6");
5928 MODULE_ALIAS("md-level-6");
5930 /* This used to be two separate modules, they were: */
5931 MODULE_ALIAS("raid5");
5932 MODULE_ALIAS("raid6");