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>
53 #include <linux/slab.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 "raid5:%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)
1531 printk_rl(KERN_WARNING
1532 "raid5:%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 "raid5:%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 "raid5:%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 = (raid5_conf_t *) mddev->private;
1623 pr_debug("raid5: 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 "raid5: Disk failure on %s, disabling device.\n"
1640 "raid5: Operation continuing on %d devices.\n",
1641 bdevname(rdev->bdev,b), conf->raid_disks - mddev->degraded);
1646 * Input: a 'big' sector number,
1647 * Output: index of the data and parity disk, and the sector # in them.
1649 static sector_t raid5_compute_sector(raid5_conf_t *conf, sector_t r_sector,
1650 int previous, int *dd_idx,
1651 struct stripe_head *sh)
1654 sector_t chunk_number;
1655 unsigned int chunk_offset;
1658 sector_t new_sector;
1659 int algorithm = previous ? conf->prev_algo
1661 int sectors_per_chunk = previous ? conf->prev_chunk_sectors
1662 : conf->chunk_sectors;
1663 int raid_disks = previous ? conf->previous_raid_disks
1665 int data_disks = raid_disks - conf->max_degraded;
1667 /* First compute the information on this sector */
1670 * Compute the chunk number and the sector offset inside the chunk
1672 chunk_offset = sector_div(r_sector, sectors_per_chunk);
1673 chunk_number = r_sector;
1676 * Compute the stripe number
1678 stripe = chunk_number;
1679 *dd_idx = sector_div(stripe, data_disks);
1682 * Select the parity disk based on the user selected algorithm.
1684 pd_idx = qd_idx = ~0;
1685 switch(conf->level) {
1687 pd_idx = data_disks;
1690 switch (algorithm) {
1691 case ALGORITHM_LEFT_ASYMMETRIC:
1692 pd_idx = data_disks - stripe % raid_disks;
1693 if (*dd_idx >= pd_idx)
1696 case ALGORITHM_RIGHT_ASYMMETRIC:
1697 pd_idx = stripe % raid_disks;
1698 if (*dd_idx >= pd_idx)
1701 case ALGORITHM_LEFT_SYMMETRIC:
1702 pd_idx = data_disks - stripe % raid_disks;
1703 *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
1705 case ALGORITHM_RIGHT_SYMMETRIC:
1706 pd_idx = stripe % raid_disks;
1707 *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
1709 case ALGORITHM_PARITY_0:
1713 case ALGORITHM_PARITY_N:
1714 pd_idx = data_disks;
1717 printk(KERN_ERR "raid5: unsupported algorithm %d\n",
1724 switch (algorithm) {
1725 case ALGORITHM_LEFT_ASYMMETRIC:
1726 pd_idx = raid_disks - 1 - (stripe % raid_disks);
1727 qd_idx = pd_idx + 1;
1728 if (pd_idx == raid_disks-1) {
1729 (*dd_idx)++; /* Q D D D P */
1731 } else if (*dd_idx >= pd_idx)
1732 (*dd_idx) += 2; /* D D P Q D */
1734 case ALGORITHM_RIGHT_ASYMMETRIC:
1735 pd_idx = stripe % raid_disks;
1736 qd_idx = pd_idx + 1;
1737 if (pd_idx == raid_disks-1) {
1738 (*dd_idx)++; /* Q D D D P */
1740 } else if (*dd_idx >= pd_idx)
1741 (*dd_idx) += 2; /* D D P Q D */
1743 case ALGORITHM_LEFT_SYMMETRIC:
1744 pd_idx = raid_disks - 1 - (stripe % raid_disks);
1745 qd_idx = (pd_idx + 1) % raid_disks;
1746 *dd_idx = (pd_idx + 2 + *dd_idx) % raid_disks;
1748 case ALGORITHM_RIGHT_SYMMETRIC:
1749 pd_idx = stripe % raid_disks;
1750 qd_idx = (pd_idx + 1) % raid_disks;
1751 *dd_idx = (pd_idx + 2 + *dd_idx) % raid_disks;
1754 case ALGORITHM_PARITY_0:
1759 case ALGORITHM_PARITY_N:
1760 pd_idx = data_disks;
1761 qd_idx = data_disks + 1;
1764 case ALGORITHM_ROTATING_ZERO_RESTART:
1765 /* Exactly the same as RIGHT_ASYMMETRIC, but or
1766 * of blocks for computing Q is different.
1768 pd_idx = stripe % raid_disks;
1769 qd_idx = pd_idx + 1;
1770 if (pd_idx == raid_disks-1) {
1771 (*dd_idx)++; /* Q D D D P */
1773 } else if (*dd_idx >= pd_idx)
1774 (*dd_idx) += 2; /* D D P Q D */
1778 case ALGORITHM_ROTATING_N_RESTART:
1779 /* Same a left_asymmetric, by first stripe is
1780 * D D D P Q rather than
1783 pd_idx = raid_disks - 1 - ((stripe + 1) % raid_disks);
1784 qd_idx = pd_idx + 1;
1785 if (pd_idx == raid_disks-1) {
1786 (*dd_idx)++; /* Q D D D P */
1788 } else if (*dd_idx >= pd_idx)
1789 (*dd_idx) += 2; /* D D P Q D */
1793 case ALGORITHM_ROTATING_N_CONTINUE:
1794 /* Same as left_symmetric but Q is before P */
1795 pd_idx = raid_disks - 1 - (stripe % raid_disks);
1796 qd_idx = (pd_idx + raid_disks - 1) % raid_disks;
1797 *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
1801 case ALGORITHM_LEFT_ASYMMETRIC_6:
1802 /* RAID5 left_asymmetric, with Q on last device */
1803 pd_idx = data_disks - stripe % (raid_disks-1);
1804 if (*dd_idx >= pd_idx)
1806 qd_idx = raid_disks - 1;
1809 case ALGORITHM_RIGHT_ASYMMETRIC_6:
1810 pd_idx = stripe % (raid_disks-1);
1811 if (*dd_idx >= pd_idx)
1813 qd_idx = raid_disks - 1;
1816 case ALGORITHM_LEFT_SYMMETRIC_6:
1817 pd_idx = data_disks - stripe % (raid_disks-1);
1818 *dd_idx = (pd_idx + 1 + *dd_idx) % (raid_disks-1);
1819 qd_idx = raid_disks - 1;
1822 case ALGORITHM_RIGHT_SYMMETRIC_6:
1823 pd_idx = stripe % (raid_disks-1);
1824 *dd_idx = (pd_idx + 1 + *dd_idx) % (raid_disks-1);
1825 qd_idx = raid_disks - 1;
1828 case ALGORITHM_PARITY_0_6:
1831 qd_idx = raid_disks - 1;
1836 printk(KERN_CRIT "raid6: unsupported algorithm %d\n",
1844 sh->pd_idx = pd_idx;
1845 sh->qd_idx = qd_idx;
1846 sh->ddf_layout = ddf_layout;
1849 * Finally, compute the new sector number
1851 new_sector = (sector_t)stripe * sectors_per_chunk + chunk_offset;
1856 static sector_t compute_blocknr(struct stripe_head *sh, int i, int previous)
1858 raid5_conf_t *conf = sh->raid_conf;
1859 int raid_disks = sh->disks;
1860 int data_disks = raid_disks - conf->max_degraded;
1861 sector_t new_sector = sh->sector, check;
1862 int sectors_per_chunk = previous ? conf->prev_chunk_sectors
1863 : conf->chunk_sectors;
1864 int algorithm = previous ? conf->prev_algo
1868 sector_t chunk_number;
1869 int dummy1, dd_idx = i;
1871 struct stripe_head sh2;
1874 chunk_offset = sector_div(new_sector, sectors_per_chunk);
1875 stripe = new_sector;
1877 if (i == sh->pd_idx)
1879 switch(conf->level) {
1882 switch (algorithm) {
1883 case ALGORITHM_LEFT_ASYMMETRIC:
1884 case ALGORITHM_RIGHT_ASYMMETRIC:
1888 case ALGORITHM_LEFT_SYMMETRIC:
1889 case ALGORITHM_RIGHT_SYMMETRIC:
1892 i -= (sh->pd_idx + 1);
1894 case ALGORITHM_PARITY_0:
1897 case ALGORITHM_PARITY_N:
1900 printk(KERN_ERR "raid5: unsupported algorithm %d\n",
1906 if (i == sh->qd_idx)
1907 return 0; /* It is the Q disk */
1908 switch (algorithm) {
1909 case ALGORITHM_LEFT_ASYMMETRIC:
1910 case ALGORITHM_RIGHT_ASYMMETRIC:
1911 case ALGORITHM_ROTATING_ZERO_RESTART:
1912 case ALGORITHM_ROTATING_N_RESTART:
1913 if (sh->pd_idx == raid_disks-1)
1914 i--; /* Q D D D P */
1915 else if (i > sh->pd_idx)
1916 i -= 2; /* D D P Q D */
1918 case ALGORITHM_LEFT_SYMMETRIC:
1919 case ALGORITHM_RIGHT_SYMMETRIC:
1920 if (sh->pd_idx == raid_disks-1)
1921 i--; /* Q D D D P */
1926 i -= (sh->pd_idx + 2);
1929 case ALGORITHM_PARITY_0:
1932 case ALGORITHM_PARITY_N:
1934 case ALGORITHM_ROTATING_N_CONTINUE:
1935 /* Like left_symmetric, but P is before Q */
1936 if (sh->pd_idx == 0)
1937 i--; /* P D D D Q */
1942 i -= (sh->pd_idx + 1);
1945 case ALGORITHM_LEFT_ASYMMETRIC_6:
1946 case ALGORITHM_RIGHT_ASYMMETRIC_6:
1950 case ALGORITHM_LEFT_SYMMETRIC_6:
1951 case ALGORITHM_RIGHT_SYMMETRIC_6:
1953 i += data_disks + 1;
1954 i -= (sh->pd_idx + 1);
1956 case ALGORITHM_PARITY_0_6:
1960 printk(KERN_CRIT "raid6: unsupported algorithm %d\n",
1967 chunk_number = stripe * data_disks + i;
1968 r_sector = 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 "compute_blocknr: map not correct\n");
1982 schedule_reconstruction(struct stripe_head *sh, struct stripe_head_state *s,
1983 int rcw, int expand)
1985 int i, pd_idx = sh->pd_idx, disks = sh->disks;
1986 raid5_conf_t *conf = sh->raid_conf;
1987 int level = conf->level;
1990 /* if we are not expanding this is a proper write request, and
1991 * there will be bios with new data to be drained into the
1995 sh->reconstruct_state = reconstruct_state_drain_run;
1996 set_bit(STRIPE_OP_BIODRAIN, &s->ops_request);
1998 sh->reconstruct_state = reconstruct_state_run;
2000 set_bit(STRIPE_OP_RECONSTRUCT, &s->ops_request);
2002 for (i = disks; i--; ) {
2003 struct r5dev *dev = &sh->dev[i];
2006 set_bit(R5_LOCKED, &dev->flags);
2007 set_bit(R5_Wantdrain, &dev->flags);
2009 clear_bit(R5_UPTODATE, &dev->flags);
2013 if (s->locked + conf->max_degraded == disks)
2014 if (!test_and_set_bit(STRIPE_FULL_WRITE, &sh->state))
2015 atomic_inc(&conf->pending_full_writes);
2018 BUG_ON(!(test_bit(R5_UPTODATE, &sh->dev[pd_idx].flags) ||
2019 test_bit(R5_Wantcompute, &sh->dev[pd_idx].flags)));
2021 sh->reconstruct_state = reconstruct_state_prexor_drain_run;
2022 set_bit(STRIPE_OP_PREXOR, &s->ops_request);
2023 set_bit(STRIPE_OP_BIODRAIN, &s->ops_request);
2024 set_bit(STRIPE_OP_RECONSTRUCT, &s->ops_request);
2026 for (i = disks; i--; ) {
2027 struct r5dev *dev = &sh->dev[i];
2032 (test_bit(R5_UPTODATE, &dev->flags) ||
2033 test_bit(R5_Wantcompute, &dev->flags))) {
2034 set_bit(R5_Wantdrain, &dev->flags);
2035 set_bit(R5_LOCKED, &dev->flags);
2036 clear_bit(R5_UPTODATE, &dev->flags);
2042 /* keep the parity disk(s) locked while asynchronous operations
2045 set_bit(R5_LOCKED, &sh->dev[pd_idx].flags);
2046 clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
2050 int qd_idx = sh->qd_idx;
2051 struct r5dev *dev = &sh->dev[qd_idx];
2053 set_bit(R5_LOCKED, &dev->flags);
2054 clear_bit(R5_UPTODATE, &dev->flags);
2058 pr_debug("%s: stripe %llu locked: %d ops_request: %lx\n",
2059 __func__, (unsigned long long)sh->sector,
2060 s->locked, s->ops_request);
2064 * Each stripe/dev can have one or more bion attached.
2065 * toread/towrite point to the first in a chain.
2066 * The bi_next chain must be in order.
2068 static int add_stripe_bio(struct stripe_head *sh, struct bio *bi, int dd_idx, int forwrite)
2071 raid5_conf_t *conf = sh->raid_conf;
2074 pr_debug("adding bh b#%llu to stripe s#%llu\n",
2075 (unsigned long long)bi->bi_sector,
2076 (unsigned long long)sh->sector);
2079 spin_lock(&sh->lock);
2080 spin_lock_irq(&conf->device_lock);
2082 bip = &sh->dev[dd_idx].towrite;
2083 if (*bip == NULL && sh->dev[dd_idx].written == NULL)
2086 bip = &sh->dev[dd_idx].toread;
2087 while (*bip && (*bip)->bi_sector < bi->bi_sector) {
2088 if ((*bip)->bi_sector + ((*bip)->bi_size >> 9) > bi->bi_sector)
2090 bip = & (*bip)->bi_next;
2092 if (*bip && (*bip)->bi_sector < bi->bi_sector + ((bi->bi_size)>>9))
2095 BUG_ON(*bip && bi->bi_next && (*bip) != bi->bi_next);
2099 bi->bi_phys_segments++;
2100 spin_unlock_irq(&conf->device_lock);
2101 spin_unlock(&sh->lock);
2103 pr_debug("added bi b#%llu to stripe s#%llu, disk %d.\n",
2104 (unsigned long long)bi->bi_sector,
2105 (unsigned long long)sh->sector, dd_idx);
2107 if (conf->mddev->bitmap && firstwrite) {
2108 bitmap_startwrite(conf->mddev->bitmap, sh->sector,
2110 sh->bm_seq = conf->seq_flush+1;
2111 set_bit(STRIPE_BIT_DELAY, &sh->state);
2115 /* check if page is covered */
2116 sector_t sector = sh->dev[dd_idx].sector;
2117 for (bi=sh->dev[dd_idx].towrite;
2118 sector < sh->dev[dd_idx].sector + STRIPE_SECTORS &&
2119 bi && bi->bi_sector <= sector;
2120 bi = r5_next_bio(bi, sh->dev[dd_idx].sector)) {
2121 if (bi->bi_sector + (bi->bi_size>>9) >= sector)
2122 sector = bi->bi_sector + (bi->bi_size>>9);
2124 if (sector >= sh->dev[dd_idx].sector + STRIPE_SECTORS)
2125 set_bit(R5_OVERWRITE, &sh->dev[dd_idx].flags);
2130 set_bit(R5_Overlap, &sh->dev[dd_idx].flags);
2131 spin_unlock_irq(&conf->device_lock);
2132 spin_unlock(&sh->lock);
2136 static void end_reshape(raid5_conf_t *conf);
2138 static void stripe_set_idx(sector_t stripe, raid5_conf_t *conf, int previous,
2139 struct stripe_head *sh)
2141 int sectors_per_chunk =
2142 previous ? conf->prev_chunk_sectors : conf->chunk_sectors;
2144 int chunk_offset = sector_div(stripe, sectors_per_chunk);
2145 int disks = previous ? conf->previous_raid_disks : conf->raid_disks;
2147 raid5_compute_sector(conf,
2148 stripe * (disks - conf->max_degraded)
2149 *sectors_per_chunk + chunk_offset,
2155 handle_failed_stripe(raid5_conf_t *conf, struct stripe_head *sh,
2156 struct stripe_head_state *s, int disks,
2157 struct bio **return_bi)
2160 for (i = disks; i--; ) {
2164 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
2167 rdev = rcu_dereference(conf->disks[i].rdev);
2168 if (rdev && test_bit(In_sync, &rdev->flags))
2169 /* multiple read failures in one stripe */
2170 md_error(conf->mddev, rdev);
2173 spin_lock_irq(&conf->device_lock);
2174 /* fail all writes first */
2175 bi = sh->dev[i].towrite;
2176 sh->dev[i].towrite = NULL;
2182 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
2183 wake_up(&conf->wait_for_overlap);
2185 while (bi && bi->bi_sector <
2186 sh->dev[i].sector + STRIPE_SECTORS) {
2187 struct bio *nextbi = r5_next_bio(bi, sh->dev[i].sector);
2188 clear_bit(BIO_UPTODATE, &bi->bi_flags);
2189 if (!raid5_dec_bi_phys_segments(bi)) {
2190 md_write_end(conf->mddev);
2191 bi->bi_next = *return_bi;
2196 /* and fail all 'written' */
2197 bi = sh->dev[i].written;
2198 sh->dev[i].written = NULL;
2199 if (bi) bitmap_end = 1;
2200 while (bi && bi->bi_sector <
2201 sh->dev[i].sector + STRIPE_SECTORS) {
2202 struct bio *bi2 = r5_next_bio(bi, sh->dev[i].sector);
2203 clear_bit(BIO_UPTODATE, &bi->bi_flags);
2204 if (!raid5_dec_bi_phys_segments(bi)) {
2205 md_write_end(conf->mddev);
2206 bi->bi_next = *return_bi;
2212 /* fail any reads if this device is non-operational and
2213 * the data has not reached the cache yet.
2215 if (!test_bit(R5_Wantfill, &sh->dev[i].flags) &&
2216 (!test_bit(R5_Insync, &sh->dev[i].flags) ||
2217 test_bit(R5_ReadError, &sh->dev[i].flags))) {
2218 bi = sh->dev[i].toread;
2219 sh->dev[i].toread = NULL;
2220 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
2221 wake_up(&conf->wait_for_overlap);
2222 if (bi) s->to_read--;
2223 while (bi && bi->bi_sector <
2224 sh->dev[i].sector + STRIPE_SECTORS) {
2225 struct bio *nextbi =
2226 r5_next_bio(bi, sh->dev[i].sector);
2227 clear_bit(BIO_UPTODATE, &bi->bi_flags);
2228 if (!raid5_dec_bi_phys_segments(bi)) {
2229 bi->bi_next = *return_bi;
2235 spin_unlock_irq(&conf->device_lock);
2237 bitmap_endwrite(conf->mddev->bitmap, sh->sector,
2238 STRIPE_SECTORS, 0, 0);
2241 if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
2242 if (atomic_dec_and_test(&conf->pending_full_writes))
2243 md_wakeup_thread(conf->mddev->thread);
2246 /* fetch_block5 - checks the given member device to see if its data needs
2247 * to be read or computed to satisfy a request.
2249 * Returns 1 when no more member devices need to be checked, otherwise returns
2250 * 0 to tell the loop in handle_stripe_fill5 to continue
2252 static int fetch_block5(struct stripe_head *sh, struct stripe_head_state *s,
2253 int disk_idx, int disks)
2255 struct r5dev *dev = &sh->dev[disk_idx];
2256 struct r5dev *failed_dev = &sh->dev[s->failed_num];
2258 /* is the data in this block needed, and can we get it? */
2259 if (!test_bit(R5_LOCKED, &dev->flags) &&
2260 !test_bit(R5_UPTODATE, &dev->flags) &&
2262 (dev->towrite && !test_bit(R5_OVERWRITE, &dev->flags)) ||
2263 s->syncing || s->expanding ||
2265 (failed_dev->toread ||
2266 (failed_dev->towrite &&
2267 !test_bit(R5_OVERWRITE, &failed_dev->flags)))))) {
2268 /* We would like to get this block, possibly by computing it,
2269 * otherwise read it if the backing disk is insync
2271 if ((s->uptodate == disks - 1) &&
2272 (s->failed && disk_idx == s->failed_num)) {
2273 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
2274 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
2275 set_bit(R5_Wantcompute, &dev->flags);
2276 sh->ops.target = disk_idx;
2277 sh->ops.target2 = -1;
2279 /* Careful: from this point on 'uptodate' is in the eye
2280 * of raid_run_ops which services 'compute' operations
2281 * before writes. R5_Wantcompute flags a block that will
2282 * be R5_UPTODATE by the time it is needed for a
2283 * subsequent operation.
2286 return 1; /* uptodate + compute == disks */
2287 } else if (test_bit(R5_Insync, &dev->flags)) {
2288 set_bit(R5_LOCKED, &dev->flags);
2289 set_bit(R5_Wantread, &dev->flags);
2291 pr_debug("Reading block %d (sync=%d)\n", disk_idx,
2300 * handle_stripe_fill5 - read or compute data to satisfy pending requests.
2302 static void handle_stripe_fill5(struct stripe_head *sh,
2303 struct stripe_head_state *s, int disks)
2307 /* look for blocks to read/compute, skip this if a compute
2308 * is already in flight, or if the stripe contents are in the
2309 * midst of changing due to a write
2311 if (!test_bit(STRIPE_COMPUTE_RUN, &sh->state) && !sh->check_state &&
2312 !sh->reconstruct_state)
2313 for (i = disks; i--; )
2314 if (fetch_block5(sh, s, i, disks))
2316 set_bit(STRIPE_HANDLE, &sh->state);
2319 /* fetch_block6 - checks the given member device to see if its data needs
2320 * to be read or computed to satisfy a request.
2322 * Returns 1 when no more member devices need to be checked, otherwise returns
2323 * 0 to tell the loop in handle_stripe_fill6 to continue
2325 static int fetch_block6(struct stripe_head *sh, struct stripe_head_state *s,
2326 struct r6_state *r6s, int disk_idx, int disks)
2328 struct r5dev *dev = &sh->dev[disk_idx];
2329 struct r5dev *fdev[2] = { &sh->dev[r6s->failed_num[0]],
2330 &sh->dev[r6s->failed_num[1]] };
2332 if (!test_bit(R5_LOCKED, &dev->flags) &&
2333 !test_bit(R5_UPTODATE, &dev->flags) &&
2335 (dev->towrite && !test_bit(R5_OVERWRITE, &dev->flags)) ||
2336 s->syncing || s->expanding ||
2338 (fdev[0]->toread || s->to_write)) ||
2340 (fdev[1]->toread || s->to_write)))) {
2341 /* we would like to get this block, possibly by computing it,
2342 * otherwise read it if the backing disk is insync
2344 BUG_ON(test_bit(R5_Wantcompute, &dev->flags));
2345 BUG_ON(test_bit(R5_Wantread, &dev->flags));
2346 if ((s->uptodate == disks - 1) &&
2347 (s->failed && (disk_idx == r6s->failed_num[0] ||
2348 disk_idx == r6s->failed_num[1]))) {
2349 /* have disk failed, and we're requested to fetch it;
2352 pr_debug("Computing stripe %llu block %d\n",
2353 (unsigned long long)sh->sector, disk_idx);
2354 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
2355 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
2356 set_bit(R5_Wantcompute, &dev->flags);
2357 sh->ops.target = disk_idx;
2358 sh->ops.target2 = -1; /* no 2nd target */
2362 } else if (s->uptodate == disks-2 && s->failed >= 2) {
2363 /* Computing 2-failure is *very* expensive; only
2364 * do it if failed >= 2
2367 for (other = disks; other--; ) {
2368 if (other == disk_idx)
2370 if (!test_bit(R5_UPTODATE,
2371 &sh->dev[other].flags))
2375 pr_debug("Computing stripe %llu blocks %d,%d\n",
2376 (unsigned long long)sh->sector,
2378 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
2379 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
2380 set_bit(R5_Wantcompute, &sh->dev[disk_idx].flags);
2381 set_bit(R5_Wantcompute, &sh->dev[other].flags);
2382 sh->ops.target = disk_idx;
2383 sh->ops.target2 = other;
2387 } else if (test_bit(R5_Insync, &dev->flags)) {
2388 set_bit(R5_LOCKED, &dev->flags);
2389 set_bit(R5_Wantread, &dev->flags);
2391 pr_debug("Reading block %d (sync=%d)\n",
2392 disk_idx, s->syncing);
2400 * handle_stripe_fill6 - read or compute data to satisfy pending requests.
2402 static void handle_stripe_fill6(struct stripe_head *sh,
2403 struct stripe_head_state *s, struct r6_state *r6s,
2408 /* look for blocks to read/compute, skip this if a compute
2409 * is already in flight, or if the stripe contents are in the
2410 * midst of changing due to a write
2412 if (!test_bit(STRIPE_COMPUTE_RUN, &sh->state) && !sh->check_state &&
2413 !sh->reconstruct_state)
2414 for (i = disks; i--; )
2415 if (fetch_block6(sh, s, r6s, i, disks))
2417 set_bit(STRIPE_HANDLE, &sh->state);
2421 /* handle_stripe_clean_event
2422 * any written block on an uptodate or failed drive can be returned.
2423 * Note that if we 'wrote' to a failed drive, it will be UPTODATE, but
2424 * never LOCKED, so we don't need to test 'failed' directly.
2426 static void handle_stripe_clean_event(raid5_conf_t *conf,
2427 struct stripe_head *sh, int disks, struct bio **return_bi)
2432 for (i = disks; i--; )
2433 if (sh->dev[i].written) {
2435 if (!test_bit(R5_LOCKED, &dev->flags) &&
2436 test_bit(R5_UPTODATE, &dev->flags)) {
2437 /* We can return any write requests */
2438 struct bio *wbi, *wbi2;
2440 pr_debug("Return write for disc %d\n", i);
2441 spin_lock_irq(&conf->device_lock);
2443 dev->written = NULL;
2444 while (wbi && wbi->bi_sector <
2445 dev->sector + STRIPE_SECTORS) {
2446 wbi2 = r5_next_bio(wbi, dev->sector);
2447 if (!raid5_dec_bi_phys_segments(wbi)) {
2448 md_write_end(conf->mddev);
2449 wbi->bi_next = *return_bi;
2454 if (dev->towrite == NULL)
2456 spin_unlock_irq(&conf->device_lock);
2458 bitmap_endwrite(conf->mddev->bitmap,
2461 !test_bit(STRIPE_DEGRADED, &sh->state),
2466 if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
2467 if (atomic_dec_and_test(&conf->pending_full_writes))
2468 md_wakeup_thread(conf->mddev->thread);
2471 static void handle_stripe_dirtying5(raid5_conf_t *conf,
2472 struct stripe_head *sh, struct stripe_head_state *s, int disks)
2474 int rmw = 0, rcw = 0, i;
2475 for (i = disks; i--; ) {
2476 /* would I have to read this buffer for read_modify_write */
2477 struct r5dev *dev = &sh->dev[i];
2478 if ((dev->towrite || i == sh->pd_idx) &&
2479 !test_bit(R5_LOCKED, &dev->flags) &&
2480 !(test_bit(R5_UPTODATE, &dev->flags) ||
2481 test_bit(R5_Wantcompute, &dev->flags))) {
2482 if (test_bit(R5_Insync, &dev->flags))
2485 rmw += 2*disks; /* cannot read it */
2487 /* Would I have to read this buffer for reconstruct_write */
2488 if (!test_bit(R5_OVERWRITE, &dev->flags) && i != sh->pd_idx &&
2489 !test_bit(R5_LOCKED, &dev->flags) &&
2490 !(test_bit(R5_UPTODATE, &dev->flags) ||
2491 test_bit(R5_Wantcompute, &dev->flags))) {
2492 if (test_bit(R5_Insync, &dev->flags)) rcw++;
2497 pr_debug("for sector %llu, rmw=%d rcw=%d\n",
2498 (unsigned long long)sh->sector, rmw, rcw);
2499 set_bit(STRIPE_HANDLE, &sh->state);
2500 if (rmw < rcw && rmw > 0)
2501 /* prefer read-modify-write, but need to get some data */
2502 for (i = disks; i--; ) {
2503 struct r5dev *dev = &sh->dev[i];
2504 if ((dev->towrite || i == sh->pd_idx) &&
2505 !test_bit(R5_LOCKED, &dev->flags) &&
2506 !(test_bit(R5_UPTODATE, &dev->flags) ||
2507 test_bit(R5_Wantcompute, &dev->flags)) &&
2508 test_bit(R5_Insync, &dev->flags)) {
2510 test_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2511 pr_debug("Read_old block "
2512 "%d for r-m-w\n", i);
2513 set_bit(R5_LOCKED, &dev->flags);
2514 set_bit(R5_Wantread, &dev->flags);
2517 set_bit(STRIPE_DELAYED, &sh->state);
2518 set_bit(STRIPE_HANDLE, &sh->state);
2522 if (rcw <= rmw && rcw > 0)
2523 /* want reconstruct write, but need to get some data */
2524 for (i = disks; i--; ) {
2525 struct r5dev *dev = &sh->dev[i];
2526 if (!test_bit(R5_OVERWRITE, &dev->flags) &&
2528 !test_bit(R5_LOCKED, &dev->flags) &&
2529 !(test_bit(R5_UPTODATE, &dev->flags) ||
2530 test_bit(R5_Wantcompute, &dev->flags)) &&
2531 test_bit(R5_Insync, &dev->flags)) {
2533 test_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2534 pr_debug("Read_old block "
2535 "%d for Reconstruct\n", i);
2536 set_bit(R5_LOCKED, &dev->flags);
2537 set_bit(R5_Wantread, &dev->flags);
2540 set_bit(STRIPE_DELAYED, &sh->state);
2541 set_bit(STRIPE_HANDLE, &sh->state);
2545 /* now if nothing is locked, and if we have enough data,
2546 * we can start a write request
2548 /* since handle_stripe can be called at any time we need to handle the
2549 * case where a compute block operation has been submitted and then a
2550 * subsequent call wants to start a write request. raid_run_ops only
2551 * handles the case where compute block and reconstruct are requested
2552 * simultaneously. If this is not the case then new writes need to be
2553 * held off until the compute completes.
2555 if ((s->req_compute || !test_bit(STRIPE_COMPUTE_RUN, &sh->state)) &&
2556 (s->locked == 0 && (rcw == 0 || rmw == 0) &&
2557 !test_bit(STRIPE_BIT_DELAY, &sh->state)))
2558 schedule_reconstruction(sh, s, rcw == 0, 0);
2561 static void handle_stripe_dirtying6(raid5_conf_t *conf,
2562 struct stripe_head *sh, struct stripe_head_state *s,
2563 struct r6_state *r6s, int disks)
2565 int rcw = 0, pd_idx = sh->pd_idx, i;
2566 int qd_idx = sh->qd_idx;
2568 set_bit(STRIPE_HANDLE, &sh->state);
2569 for (i = disks; i--; ) {
2570 struct r5dev *dev = &sh->dev[i];
2571 /* check if we haven't enough data */
2572 if (!test_bit(R5_OVERWRITE, &dev->flags) &&
2573 i != pd_idx && i != qd_idx &&
2574 !test_bit(R5_LOCKED, &dev->flags) &&
2575 !(test_bit(R5_UPTODATE, &dev->flags) ||
2576 test_bit(R5_Wantcompute, &dev->flags))) {
2578 if (!test_bit(R5_Insync, &dev->flags))
2579 continue; /* it's a failed drive */
2582 test_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2583 pr_debug("Read_old stripe %llu "
2584 "block %d for Reconstruct\n",
2585 (unsigned long long)sh->sector, i);
2586 set_bit(R5_LOCKED, &dev->flags);
2587 set_bit(R5_Wantread, &dev->flags);
2590 pr_debug("Request delayed stripe %llu "
2591 "block %d for Reconstruct\n",
2592 (unsigned long long)sh->sector, i);
2593 set_bit(STRIPE_DELAYED, &sh->state);
2594 set_bit(STRIPE_HANDLE, &sh->state);
2598 /* now if nothing is locked, and if we have enough data, we can start a
2601 if ((s->req_compute || !test_bit(STRIPE_COMPUTE_RUN, &sh->state)) &&
2602 s->locked == 0 && rcw == 0 &&
2603 !test_bit(STRIPE_BIT_DELAY, &sh->state)) {
2604 schedule_reconstruction(sh, s, 1, 0);
2608 static void handle_parity_checks5(raid5_conf_t *conf, struct stripe_head *sh,
2609 struct stripe_head_state *s, int disks)
2611 struct r5dev *dev = NULL;
2613 set_bit(STRIPE_HANDLE, &sh->state);
2615 switch (sh->check_state) {
2616 case check_state_idle:
2617 /* start a new check operation if there are no failures */
2618 if (s->failed == 0) {
2619 BUG_ON(s->uptodate != disks);
2620 sh->check_state = check_state_run;
2621 set_bit(STRIPE_OP_CHECK, &s->ops_request);
2622 clear_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags);
2626 dev = &sh->dev[s->failed_num];
2628 case check_state_compute_result:
2629 sh->check_state = check_state_idle;
2631 dev = &sh->dev[sh->pd_idx];
2633 /* check that a write has not made the stripe insync */
2634 if (test_bit(STRIPE_INSYNC, &sh->state))
2637 /* either failed parity check, or recovery is happening */
2638 BUG_ON(!test_bit(R5_UPTODATE, &dev->flags));
2639 BUG_ON(s->uptodate != disks);
2641 set_bit(R5_LOCKED, &dev->flags);
2643 set_bit(R5_Wantwrite, &dev->flags);
2645 clear_bit(STRIPE_DEGRADED, &sh->state);
2646 set_bit(STRIPE_INSYNC, &sh->state);
2648 case check_state_run:
2649 break; /* we will be called again upon completion */
2650 case check_state_check_result:
2651 sh->check_state = check_state_idle;
2653 /* if a failure occurred during the check operation, leave
2654 * STRIPE_INSYNC not set and let the stripe be handled again
2659 /* handle a successful check operation, if parity is correct
2660 * we are done. Otherwise update the mismatch count and repair
2661 * parity if !MD_RECOVERY_CHECK
2663 if ((sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) == 0)
2664 /* parity is correct (on disc,
2665 * not in buffer any more)
2667 set_bit(STRIPE_INSYNC, &sh->state);
2669 conf->mddev->resync_mismatches += STRIPE_SECTORS;
2670 if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery))
2671 /* don't try to repair!! */
2672 set_bit(STRIPE_INSYNC, &sh->state);
2674 sh->check_state = check_state_compute_run;
2675 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
2676 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
2677 set_bit(R5_Wantcompute,
2678 &sh->dev[sh->pd_idx].flags);
2679 sh->ops.target = sh->pd_idx;
2680 sh->ops.target2 = -1;
2685 case check_state_compute_run:
2688 printk(KERN_ERR "%s: unknown check_state: %d sector: %llu\n",
2689 __func__, sh->check_state,
2690 (unsigned long long) sh->sector);
2696 static void handle_parity_checks6(raid5_conf_t *conf, struct stripe_head *sh,
2697 struct stripe_head_state *s,
2698 struct r6_state *r6s, int disks)
2700 int pd_idx = sh->pd_idx;
2701 int qd_idx = sh->qd_idx;
2704 set_bit(STRIPE_HANDLE, &sh->state);
2706 BUG_ON(s->failed > 2);
2708 /* Want to check and possibly repair P and Q.
2709 * However there could be one 'failed' device, in which
2710 * case we can only check one of them, possibly using the
2711 * other to generate missing data
2714 switch (sh->check_state) {
2715 case check_state_idle:
2716 /* start a new check operation if there are < 2 failures */
2717 if (s->failed == r6s->q_failed) {
2718 /* The only possible failed device holds Q, so it
2719 * makes sense to check P (If anything else were failed,
2720 * we would have used P to recreate it).
2722 sh->check_state = check_state_run;
2724 if (!r6s->q_failed && s->failed < 2) {
2725 /* Q is not failed, and we didn't use it to generate
2726 * anything, so it makes sense to check it
2728 if (sh->check_state == check_state_run)
2729 sh->check_state = check_state_run_pq;
2731 sh->check_state = check_state_run_q;
2734 /* discard potentially stale zero_sum_result */
2735 sh->ops.zero_sum_result = 0;
2737 if (sh->check_state == check_state_run) {
2738 /* async_xor_zero_sum destroys the contents of P */
2739 clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
2742 if (sh->check_state >= check_state_run &&
2743 sh->check_state <= check_state_run_pq) {
2744 /* async_syndrome_zero_sum preserves P and Q, so
2745 * no need to mark them !uptodate here
2747 set_bit(STRIPE_OP_CHECK, &s->ops_request);
2751 /* we have 2-disk failure */
2752 BUG_ON(s->failed != 2);
2754 case check_state_compute_result:
2755 sh->check_state = check_state_idle;
2757 /* check that a write has not made the stripe insync */
2758 if (test_bit(STRIPE_INSYNC, &sh->state))
2761 /* now write out any block on a failed drive,
2762 * or P or Q if they were recomputed
2764 BUG_ON(s->uptodate < disks - 1); /* We don't need Q to recover */
2765 if (s->failed == 2) {
2766 dev = &sh->dev[r6s->failed_num[1]];
2768 set_bit(R5_LOCKED, &dev->flags);
2769 set_bit(R5_Wantwrite, &dev->flags);
2771 if (s->failed >= 1) {
2772 dev = &sh->dev[r6s->failed_num[0]];
2774 set_bit(R5_LOCKED, &dev->flags);
2775 set_bit(R5_Wantwrite, &dev->flags);
2777 if (sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) {
2778 dev = &sh->dev[pd_idx];
2780 set_bit(R5_LOCKED, &dev->flags);
2781 set_bit(R5_Wantwrite, &dev->flags);
2783 if (sh->ops.zero_sum_result & SUM_CHECK_Q_RESULT) {
2784 dev = &sh->dev[qd_idx];
2786 set_bit(R5_LOCKED, &dev->flags);
2787 set_bit(R5_Wantwrite, &dev->flags);
2789 clear_bit(STRIPE_DEGRADED, &sh->state);
2791 set_bit(STRIPE_INSYNC, &sh->state);
2793 case check_state_run:
2794 case check_state_run_q:
2795 case check_state_run_pq:
2796 break; /* we will be called again upon completion */
2797 case check_state_check_result:
2798 sh->check_state = check_state_idle;
2800 /* handle a successful check operation, if parity is correct
2801 * we are done. Otherwise update the mismatch count and repair
2802 * parity if !MD_RECOVERY_CHECK
2804 if (sh->ops.zero_sum_result == 0) {
2805 /* both parities are correct */
2807 set_bit(STRIPE_INSYNC, &sh->state);
2809 /* in contrast to the raid5 case we can validate
2810 * parity, but still have a failure to write
2813 sh->check_state = check_state_compute_result;
2814 /* Returning at this point means that we may go
2815 * off and bring p and/or q uptodate again so
2816 * we make sure to check zero_sum_result again
2817 * to verify if p or q need writeback
2821 conf->mddev->resync_mismatches += STRIPE_SECTORS;
2822 if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery))
2823 /* don't try to repair!! */
2824 set_bit(STRIPE_INSYNC, &sh->state);
2826 int *target = &sh->ops.target;
2828 sh->ops.target = -1;
2829 sh->ops.target2 = -1;
2830 sh->check_state = check_state_compute_run;
2831 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
2832 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
2833 if (sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) {
2834 set_bit(R5_Wantcompute,
2835 &sh->dev[pd_idx].flags);
2837 target = &sh->ops.target2;
2840 if (sh->ops.zero_sum_result & SUM_CHECK_Q_RESULT) {
2841 set_bit(R5_Wantcompute,
2842 &sh->dev[qd_idx].flags);
2849 case check_state_compute_run:
2852 printk(KERN_ERR "%s: unknown check_state: %d sector: %llu\n",
2853 __func__, sh->check_state,
2854 (unsigned long long) sh->sector);
2859 static void handle_stripe_expansion(raid5_conf_t *conf, struct stripe_head *sh,
2860 struct r6_state *r6s)
2864 /* We have read all the blocks in this stripe and now we need to
2865 * copy some of them into a target stripe for expand.
2867 struct dma_async_tx_descriptor *tx = NULL;
2868 clear_bit(STRIPE_EXPAND_SOURCE, &sh->state);
2869 for (i = 0; i < sh->disks; i++)
2870 if (i != sh->pd_idx && i != sh->qd_idx) {
2872 struct stripe_head *sh2;
2873 struct async_submit_ctl submit;
2875 sector_t bn = compute_blocknr(sh, i, 1);
2876 sector_t s = raid5_compute_sector(conf, bn, 0,
2878 sh2 = get_active_stripe(conf, s, 0, 1, 1);
2880 /* so far only the early blocks of this stripe
2881 * have been requested. When later blocks
2882 * get requested, we will try again
2885 if (!test_bit(STRIPE_EXPANDING, &sh2->state) ||
2886 test_bit(R5_Expanded, &sh2->dev[dd_idx].flags)) {
2887 /* must have already done this block */
2888 release_stripe(sh2);
2892 /* place all the copies on one channel */
2893 init_async_submit(&submit, 0, tx, NULL, NULL, NULL);
2894 tx = async_memcpy(sh2->dev[dd_idx].page,
2895 sh->dev[i].page, 0, 0, STRIPE_SIZE,
2898 set_bit(R5_Expanded, &sh2->dev[dd_idx].flags);
2899 set_bit(R5_UPTODATE, &sh2->dev[dd_idx].flags);
2900 for (j = 0; j < conf->raid_disks; j++)
2901 if (j != sh2->pd_idx &&
2902 (!r6s || j != sh2->qd_idx) &&
2903 !test_bit(R5_Expanded, &sh2->dev[j].flags))
2905 if (j == conf->raid_disks) {
2906 set_bit(STRIPE_EXPAND_READY, &sh2->state);
2907 set_bit(STRIPE_HANDLE, &sh2->state);
2909 release_stripe(sh2);
2912 /* done submitting copies, wait for them to complete */
2915 dma_wait_for_async_tx(tx);
2921 * handle_stripe - do things to a stripe.
2923 * We lock the stripe and then examine the state of various bits
2924 * to see what needs to be done.
2926 * return some read request which now have data
2927 * return some write requests which are safely on disc
2928 * schedule a read on some buffers
2929 * schedule a write of some buffers
2930 * return confirmation of parity correctness
2932 * buffers are taken off read_list or write_list, and bh_cache buffers
2933 * get BH_Lock set before the stripe lock is released.
2937 static void handle_stripe5(struct stripe_head *sh)
2939 raid5_conf_t *conf = sh->raid_conf;
2940 int disks = sh->disks, i;
2941 struct bio *return_bi = NULL;
2942 struct stripe_head_state s;
2944 mdk_rdev_t *blocked_rdev = NULL;
2946 int dec_preread_active = 0;
2948 memset(&s, 0, sizeof(s));
2949 pr_debug("handling stripe %llu, state=%#lx cnt=%d, pd_idx=%d check:%d "
2950 "reconstruct:%d\n", (unsigned long long)sh->sector, sh->state,
2951 atomic_read(&sh->count), sh->pd_idx, sh->check_state,
2952 sh->reconstruct_state);
2954 spin_lock(&sh->lock);
2955 clear_bit(STRIPE_HANDLE, &sh->state);
2956 clear_bit(STRIPE_DELAYED, &sh->state);
2958 s.syncing = test_bit(STRIPE_SYNCING, &sh->state);
2959 s.expanding = test_bit(STRIPE_EXPAND_SOURCE, &sh->state);
2960 s.expanded = test_bit(STRIPE_EXPAND_READY, &sh->state);
2962 /* Now to look around and see what can be done */
2964 for (i=disks; i--; ) {
2968 clear_bit(R5_Insync, &dev->flags);
2970 pr_debug("check %d: state 0x%lx toread %p read %p write %p "
2971 "written %p\n", i, dev->flags, dev->toread, dev->read,
2972 dev->towrite, dev->written);
2974 /* maybe we can request a biofill operation
2976 * new wantfill requests are only permitted while
2977 * ops_complete_biofill is guaranteed to be inactive
2979 if (test_bit(R5_UPTODATE, &dev->flags) && dev->toread &&
2980 !test_bit(STRIPE_BIOFILL_RUN, &sh->state))
2981 set_bit(R5_Wantfill, &dev->flags);
2983 /* now count some things */
2984 if (test_bit(R5_LOCKED, &dev->flags)) s.locked++;
2985 if (test_bit(R5_UPTODATE, &dev->flags)) s.uptodate++;
2986 if (test_bit(R5_Wantcompute, &dev->flags)) s.compute++;
2988 if (test_bit(R5_Wantfill, &dev->flags))
2990 else if (dev->toread)
2994 if (!test_bit(R5_OVERWRITE, &dev->flags))
2999 rdev = rcu_dereference(conf->disks[i].rdev);
3000 if (blocked_rdev == NULL &&
3001 rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
3002 blocked_rdev = rdev;
3003 atomic_inc(&rdev->nr_pending);
3005 if (!rdev || !test_bit(In_sync, &rdev->flags)) {
3006 /* The ReadError flag will just be confusing now */
3007 clear_bit(R5_ReadError, &dev->flags);
3008 clear_bit(R5_ReWrite, &dev->flags);
3010 if (!rdev || !test_bit(In_sync, &rdev->flags)
3011 || test_bit(R5_ReadError, &dev->flags)) {
3015 set_bit(R5_Insync, &dev->flags);
3019 if (unlikely(blocked_rdev)) {
3020 if (s.syncing || s.expanding || s.expanded ||
3021 s.to_write || s.written) {
3022 set_bit(STRIPE_HANDLE, &sh->state);
3025 /* There is nothing for the blocked_rdev to block */
3026 rdev_dec_pending(blocked_rdev, conf->mddev);
3027 blocked_rdev = NULL;
3030 if (s.to_fill && !test_bit(STRIPE_BIOFILL_RUN, &sh->state)) {
3031 set_bit(STRIPE_OP_BIOFILL, &s.ops_request);
3032 set_bit(STRIPE_BIOFILL_RUN, &sh->state);
3035 pr_debug("locked=%d uptodate=%d to_read=%d"
3036 " to_write=%d failed=%d failed_num=%d\n",
3037 s.locked, s.uptodate, s.to_read, s.to_write,
3038 s.failed, s.failed_num);
3039 /* check if the array has lost two devices and, if so, some requests might
3042 if (s.failed > 1 && s.to_read+s.to_write+s.written)
3043 handle_failed_stripe(conf, sh, &s, disks, &return_bi);
3044 if (s.failed > 1 && s.syncing) {
3045 md_done_sync(conf->mddev, STRIPE_SECTORS,0);
3046 clear_bit(STRIPE_SYNCING, &sh->state);
3050 /* might be able to return some write requests if the parity block
3051 * is safe, or on a failed drive
3053 dev = &sh->dev[sh->pd_idx];
3055 ((test_bit(R5_Insync, &dev->flags) &&
3056 !test_bit(R5_LOCKED, &dev->flags) &&
3057 test_bit(R5_UPTODATE, &dev->flags)) ||
3058 (s.failed == 1 && s.failed_num == sh->pd_idx)))
3059 handle_stripe_clean_event(conf, sh, disks, &return_bi);
3061 /* Now we might consider reading some blocks, either to check/generate
3062 * parity, or to satisfy requests
3063 * or to load a block that is being partially written.
3065 if (s.to_read || s.non_overwrite ||
3066 (s.syncing && (s.uptodate + s.compute < disks)) || s.expanding)
3067 handle_stripe_fill5(sh, &s, disks);
3069 /* Now we check to see if any write operations have recently
3073 if (sh->reconstruct_state == reconstruct_state_prexor_drain_result)
3075 if (sh->reconstruct_state == reconstruct_state_drain_result ||
3076 sh->reconstruct_state == reconstruct_state_prexor_drain_result) {
3077 sh->reconstruct_state = reconstruct_state_idle;
3079 /* All the 'written' buffers and the parity block are ready to
3080 * be written back to disk
3082 BUG_ON(!test_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags));
3083 for (i = disks; i--; ) {
3085 if (test_bit(R5_LOCKED, &dev->flags) &&
3086 (i == sh->pd_idx || dev->written)) {
3087 pr_debug("Writing block %d\n", i);
3088 set_bit(R5_Wantwrite, &dev->flags);
3091 if (!test_bit(R5_Insync, &dev->flags) ||
3092 (i == sh->pd_idx && s.failed == 0))
3093 set_bit(STRIPE_INSYNC, &sh->state);
3096 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
3097 dec_preread_active = 1;
3100 /* Now to consider new write requests and what else, if anything
3101 * should be read. We do not handle new writes when:
3102 * 1/ A 'write' operation (copy+xor) is already in flight.
3103 * 2/ A 'check' operation is in flight, as it may clobber the parity
3106 if (s.to_write && !sh->reconstruct_state && !sh->check_state)
3107 handle_stripe_dirtying5(conf, sh, &s, disks);
3109 /* maybe we need to check and possibly fix the parity for this stripe
3110 * Any reads will already have been scheduled, so we just see if enough
3111 * data is available. The parity check is held off while parity
3112 * dependent operations are in flight.
3114 if (sh->check_state ||
3115 (s.syncing && s.locked == 0 &&
3116 !test_bit(STRIPE_COMPUTE_RUN, &sh->state) &&
3117 !test_bit(STRIPE_INSYNC, &sh->state)))
3118 handle_parity_checks5(conf, sh, &s, disks);
3120 if (s.syncing && s.locked == 0 && test_bit(STRIPE_INSYNC, &sh->state)) {
3121 md_done_sync(conf->mddev, STRIPE_SECTORS,1);
3122 clear_bit(STRIPE_SYNCING, &sh->state);
3125 /* If the failed drive is just a ReadError, then we might need to progress
3126 * the repair/check process
3128 if (s.failed == 1 && !conf->mddev->ro &&
3129 test_bit(R5_ReadError, &sh->dev[s.failed_num].flags)
3130 && !test_bit(R5_LOCKED, &sh->dev[s.failed_num].flags)
3131 && test_bit(R5_UPTODATE, &sh->dev[s.failed_num].flags)
3133 dev = &sh->dev[s.failed_num];
3134 if (!test_bit(R5_ReWrite, &dev->flags)) {
3135 set_bit(R5_Wantwrite, &dev->flags);
3136 set_bit(R5_ReWrite, &dev->flags);
3137 set_bit(R5_LOCKED, &dev->flags);
3140 /* let's read it back */
3141 set_bit(R5_Wantread, &dev->flags);
3142 set_bit(R5_LOCKED, &dev->flags);
3147 /* Finish reconstruct operations initiated by the expansion process */
3148 if (sh->reconstruct_state == reconstruct_state_result) {
3149 struct stripe_head *sh2
3150 = get_active_stripe(conf, sh->sector, 1, 1, 1);
3151 if (sh2 && test_bit(STRIPE_EXPAND_SOURCE, &sh2->state)) {
3152 /* sh cannot be written until sh2 has been read.
3153 * so arrange for sh to be delayed a little
3155 set_bit(STRIPE_DELAYED, &sh->state);
3156 set_bit(STRIPE_HANDLE, &sh->state);
3157 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE,
3159 atomic_inc(&conf->preread_active_stripes);
3160 release_stripe(sh2);
3164 release_stripe(sh2);
3166 sh->reconstruct_state = reconstruct_state_idle;
3167 clear_bit(STRIPE_EXPANDING, &sh->state);
3168 for (i = conf->raid_disks; i--; ) {
3169 set_bit(R5_Wantwrite, &sh->dev[i].flags);
3170 set_bit(R5_LOCKED, &sh->dev[i].flags);
3175 if (s.expanded && test_bit(STRIPE_EXPANDING, &sh->state) &&
3176 !sh->reconstruct_state) {
3177 /* Need to write out all blocks after computing parity */
3178 sh->disks = conf->raid_disks;
3179 stripe_set_idx(sh->sector, conf, 0, sh);
3180 schedule_reconstruction(sh, &s, 1, 1);
3181 } else if (s.expanded && !sh->reconstruct_state && s.locked == 0) {
3182 clear_bit(STRIPE_EXPAND_READY, &sh->state);
3183 atomic_dec(&conf->reshape_stripes);
3184 wake_up(&conf->wait_for_overlap);
3185 md_done_sync(conf->mddev, STRIPE_SECTORS, 1);
3188 if (s.expanding && s.locked == 0 &&
3189 !test_bit(STRIPE_COMPUTE_RUN, &sh->state))
3190 handle_stripe_expansion(conf, sh, NULL);
3193 spin_unlock(&sh->lock);
3195 /* wait for this device to become unblocked */
3196 if (unlikely(blocked_rdev))
3197 md_wait_for_blocked_rdev(blocked_rdev, conf->mddev);
3200 raid_run_ops(sh, s.ops_request);
3204 if (dec_preread_active) {
3205 /* We delay this until after ops_run_io so that if make_request
3206 * is waiting on a barrier, it won't continue until the writes
3207 * have actually been submitted.
3209 atomic_dec(&conf->preread_active_stripes);
3210 if (atomic_read(&conf->preread_active_stripes) <
3212 md_wakeup_thread(conf->mddev->thread);
3214 return_io(return_bi);
3217 static void handle_stripe6(struct stripe_head *sh)
3219 raid5_conf_t *conf = sh->raid_conf;
3220 int disks = sh->disks;
3221 struct bio *return_bi = NULL;
3222 int i, pd_idx = sh->pd_idx, qd_idx = sh->qd_idx;
3223 struct stripe_head_state s;
3224 struct r6_state r6s;
3225 struct r5dev *dev, *pdev, *qdev;
3226 mdk_rdev_t *blocked_rdev = NULL;
3227 int dec_preread_active = 0;
3229 pr_debug("handling stripe %llu, state=%#lx cnt=%d, "
3230 "pd_idx=%d, qd_idx=%d\n, check:%d, reconstruct:%d\n",
3231 (unsigned long long)sh->sector, sh->state,
3232 atomic_read(&sh->count), pd_idx, qd_idx,
3233 sh->check_state, sh->reconstruct_state);
3234 memset(&s, 0, sizeof(s));
3236 spin_lock(&sh->lock);
3237 clear_bit(STRIPE_HANDLE, &sh->state);
3238 clear_bit(STRIPE_DELAYED, &sh->state);
3240 s.syncing = test_bit(STRIPE_SYNCING, &sh->state);
3241 s.expanding = test_bit(STRIPE_EXPAND_SOURCE, &sh->state);
3242 s.expanded = test_bit(STRIPE_EXPAND_READY, &sh->state);
3243 /* Now to look around and see what can be done */
3246 for (i=disks; i--; ) {
3249 clear_bit(R5_Insync, &dev->flags);
3251 pr_debug("check %d: state 0x%lx read %p write %p written %p\n",
3252 i, dev->flags, dev->toread, dev->towrite, dev->written);
3253 /* maybe we can reply to a read
3255 * new wantfill requests are only permitted while
3256 * ops_complete_biofill is guaranteed to be inactive
3258 if (test_bit(R5_UPTODATE, &dev->flags) && dev->toread &&
3259 !test_bit(STRIPE_BIOFILL_RUN, &sh->state))
3260 set_bit(R5_Wantfill, &dev->flags);
3262 /* now count some things */
3263 if (test_bit(R5_LOCKED, &dev->flags)) s.locked++;
3264 if (test_bit(R5_UPTODATE, &dev->flags)) s.uptodate++;
3265 if (test_bit(R5_Wantcompute, &dev->flags)) {
3267 BUG_ON(s.compute > 2);
3270 if (test_bit(R5_Wantfill, &dev->flags)) {
3272 } else if (dev->toread)
3276 if (!test_bit(R5_OVERWRITE, &dev->flags))
3281 rdev = rcu_dereference(conf->disks[i].rdev);
3282 if (blocked_rdev == NULL &&
3283 rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
3284 blocked_rdev = rdev;
3285 atomic_inc(&rdev->nr_pending);
3287 if (!rdev || !test_bit(In_sync, &rdev->flags)) {
3288 /* The ReadError flag will just be confusing now */
3289 clear_bit(R5_ReadError, &dev->flags);
3290 clear_bit(R5_ReWrite, &dev->flags);
3292 if (!rdev || !test_bit(In_sync, &rdev->flags)
3293 || test_bit(R5_ReadError, &dev->flags)) {
3295 r6s.failed_num[s.failed] = i;
3298 set_bit(R5_Insync, &dev->flags);
3302 if (unlikely(blocked_rdev)) {
3303 if (s.syncing || s.expanding || s.expanded ||
3304 s.to_write || s.written) {
3305 set_bit(STRIPE_HANDLE, &sh->state);
3308 /* There is nothing for the blocked_rdev to block */
3309 rdev_dec_pending(blocked_rdev, conf->mddev);
3310 blocked_rdev = NULL;
3313 if (s.to_fill && !test_bit(STRIPE_BIOFILL_RUN, &sh->state)) {
3314 set_bit(STRIPE_OP_BIOFILL, &s.ops_request);
3315 set_bit(STRIPE_BIOFILL_RUN, &sh->state);
3318 pr_debug("locked=%d uptodate=%d to_read=%d"
3319 " to_write=%d failed=%d failed_num=%d,%d\n",
3320 s.locked, s.uptodate, s.to_read, s.to_write, s.failed,
3321 r6s.failed_num[0], r6s.failed_num[1]);
3322 /* check if the array has lost >2 devices and, if so, some requests
3323 * might need to be failed
3325 if (s.failed > 2 && s.to_read+s.to_write+s.written)
3326 handle_failed_stripe(conf, sh, &s, disks, &return_bi);
3327 if (s.failed > 2 && s.syncing) {
3328 md_done_sync(conf->mddev, STRIPE_SECTORS,0);
3329 clear_bit(STRIPE_SYNCING, &sh->state);
3334 * might be able to return some write requests if the parity blocks
3335 * are safe, or on a failed drive
3337 pdev = &sh->dev[pd_idx];
3338 r6s.p_failed = (s.failed >= 1 && r6s.failed_num[0] == pd_idx)
3339 || (s.failed >= 2 && r6s.failed_num[1] == pd_idx);
3340 qdev = &sh->dev[qd_idx];
3341 r6s.q_failed = (s.failed >= 1 && r6s.failed_num[0] == qd_idx)
3342 || (s.failed >= 2 && r6s.failed_num[1] == qd_idx);
3345 ( r6s.p_failed || ((test_bit(R5_Insync, &pdev->flags)
3346 && !test_bit(R5_LOCKED, &pdev->flags)
3347 && test_bit(R5_UPTODATE, &pdev->flags)))) &&
3348 ( r6s.q_failed || ((test_bit(R5_Insync, &qdev->flags)
3349 && !test_bit(R5_LOCKED, &qdev->flags)
3350 && test_bit(R5_UPTODATE, &qdev->flags)))))
3351 handle_stripe_clean_event(conf, sh, disks, &return_bi);
3353 /* Now we might consider reading some blocks, either to check/generate
3354 * parity, or to satisfy requests
3355 * or to load a block that is being partially written.
3357 if (s.to_read || s.non_overwrite || (s.to_write && s.failed) ||
3358 (s.syncing && (s.uptodate + s.compute < disks)) || s.expanding)
3359 handle_stripe_fill6(sh, &s, &r6s, disks);
3361 /* Now we check to see if any write operations have recently
3364 if (sh->reconstruct_state == reconstruct_state_drain_result) {
3366 sh->reconstruct_state = reconstruct_state_idle;
3367 /* All the 'written' buffers and the parity blocks are ready to
3368 * be written back to disk
3370 BUG_ON(!test_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags));
3371 BUG_ON(!test_bit(R5_UPTODATE, &sh->dev[qd_idx].flags));
3372 for (i = disks; i--; ) {
3374 if (test_bit(R5_LOCKED, &dev->flags) &&
3375 (i == sh->pd_idx || i == qd_idx ||
3377 pr_debug("Writing block %d\n", i);
3378 BUG_ON(!test_bit(R5_UPTODATE, &dev->flags));
3379 set_bit(R5_Wantwrite, &dev->flags);
3380 if (!test_bit(R5_Insync, &dev->flags) ||
3381 ((i == sh->pd_idx || i == qd_idx) &&
3383 set_bit(STRIPE_INSYNC, &sh->state);
3386 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
3387 dec_preread_active = 1;
3390 /* Now to consider new write requests and what else, if anything
3391 * should be read. We do not handle new writes when:
3392 * 1/ A 'write' operation (copy+gen_syndrome) is already in flight.
3393 * 2/ A 'check' operation is in flight, as it may clobber the parity
3396 if (s.to_write && !sh->reconstruct_state && !sh->check_state)
3397 handle_stripe_dirtying6(conf, sh, &s, &r6s, disks);
3399 /* maybe we need to check and possibly fix the parity for this stripe
3400 * Any reads will already have been scheduled, so we just see if enough
3401 * data is available. The parity check is held off while parity
3402 * dependent operations are in flight.
3404 if (sh->check_state ||
3405 (s.syncing && s.locked == 0 &&
3406 !test_bit(STRIPE_COMPUTE_RUN, &sh->state) &&
3407 !test_bit(STRIPE_INSYNC, &sh->state)))
3408 handle_parity_checks6(conf, sh, &s, &r6s, disks);
3410 if (s.syncing && s.locked == 0 && test_bit(STRIPE_INSYNC, &sh->state)) {
3411 md_done_sync(conf->mddev, STRIPE_SECTORS,1);
3412 clear_bit(STRIPE_SYNCING, &sh->state);
3415 /* If the failed drives are just a ReadError, then we might need
3416 * to progress the repair/check process
3418 if (s.failed <= 2 && !conf->mddev->ro)
3419 for (i = 0; i < s.failed; i++) {
3420 dev = &sh->dev[r6s.failed_num[i]];
3421 if (test_bit(R5_ReadError, &dev->flags)
3422 && !test_bit(R5_LOCKED, &dev->flags)
3423 && test_bit(R5_UPTODATE, &dev->flags)
3425 if (!test_bit(R5_ReWrite, &dev->flags)) {
3426 set_bit(R5_Wantwrite, &dev->flags);
3427 set_bit(R5_ReWrite, &dev->flags);
3428 set_bit(R5_LOCKED, &dev->flags);
3431 /* let's read it back */
3432 set_bit(R5_Wantread, &dev->flags);
3433 set_bit(R5_LOCKED, &dev->flags);
3439 /* Finish reconstruct operations initiated by the expansion process */
3440 if (sh->reconstruct_state == reconstruct_state_result) {
3441 sh->reconstruct_state = reconstruct_state_idle;
3442 clear_bit(STRIPE_EXPANDING, &sh->state);
3443 for (i = conf->raid_disks; i--; ) {
3444 set_bit(R5_Wantwrite, &sh->dev[i].flags);
3445 set_bit(R5_LOCKED, &sh->dev[i].flags);
3450 if (s.expanded && test_bit(STRIPE_EXPANDING, &sh->state) &&
3451 !sh->reconstruct_state) {
3452 struct stripe_head *sh2
3453 = get_active_stripe(conf, sh->sector, 1, 1, 1);
3454 if (sh2 && test_bit(STRIPE_EXPAND_SOURCE, &sh2->state)) {
3455 /* sh cannot be written until sh2 has been read.
3456 * so arrange for sh to be delayed a little
3458 set_bit(STRIPE_DELAYED, &sh->state);
3459 set_bit(STRIPE_HANDLE, &sh->state);
3460 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE,
3462 atomic_inc(&conf->preread_active_stripes);
3463 release_stripe(sh2);
3467 release_stripe(sh2);
3469 /* Need to write out all blocks after computing P&Q */
3470 sh->disks = conf->raid_disks;
3471 stripe_set_idx(sh->sector, conf, 0, sh);
3472 schedule_reconstruction(sh, &s, 1, 1);
3473 } else if (s.expanded && !sh->reconstruct_state && s.locked == 0) {
3474 clear_bit(STRIPE_EXPAND_READY, &sh->state);
3475 atomic_dec(&conf->reshape_stripes);
3476 wake_up(&conf->wait_for_overlap);
3477 md_done_sync(conf->mddev, STRIPE_SECTORS, 1);
3480 if (s.expanding && s.locked == 0 &&
3481 !test_bit(STRIPE_COMPUTE_RUN, &sh->state))
3482 handle_stripe_expansion(conf, sh, &r6s);
3485 spin_unlock(&sh->lock);
3487 /* wait for this device to become unblocked */
3488 if (unlikely(blocked_rdev))
3489 md_wait_for_blocked_rdev(blocked_rdev, conf->mddev);
3492 raid_run_ops(sh, s.ops_request);
3497 if (dec_preread_active) {
3498 /* We delay this until after ops_run_io so that if make_request
3499 * is waiting on a barrier, it won't continue until the writes
3500 * have actually been submitted.
3502 atomic_dec(&conf->preread_active_stripes);
3503 if (atomic_read(&conf->preread_active_stripes) <
3505 md_wakeup_thread(conf->mddev->thread);
3508 return_io(return_bi);
3511 static void handle_stripe(struct stripe_head *sh)
3513 if (sh->raid_conf->level == 6)
3519 static void raid5_activate_delayed(raid5_conf_t *conf)
3521 if (atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD) {
3522 while (!list_empty(&conf->delayed_list)) {
3523 struct list_head *l = conf->delayed_list.next;
3524 struct stripe_head *sh;
3525 sh = list_entry(l, struct stripe_head, lru);
3527 clear_bit(STRIPE_DELAYED, &sh->state);
3528 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
3529 atomic_inc(&conf->preread_active_stripes);
3530 list_add_tail(&sh->lru, &conf->hold_list);
3533 blk_plug_device(conf->mddev->queue);
3536 static void activate_bit_delay(raid5_conf_t *conf)
3538 /* device_lock is held */
3539 struct list_head head;
3540 list_add(&head, &conf->bitmap_list);
3541 list_del_init(&conf->bitmap_list);
3542 while (!list_empty(&head)) {
3543 struct stripe_head *sh = list_entry(head.next, struct stripe_head, lru);
3544 list_del_init(&sh->lru);
3545 atomic_inc(&sh->count);
3546 __release_stripe(conf, sh);
3550 static void unplug_slaves(mddev_t *mddev)
3552 raid5_conf_t *conf = mddev->private;
3554 int devs = max(conf->raid_disks, conf->previous_raid_disks);
3557 for (i = 0; i < devs; i++) {
3558 mdk_rdev_t *rdev = rcu_dereference(conf->disks[i].rdev);
3559 if (rdev && !test_bit(Faulty, &rdev->flags) && atomic_read(&rdev->nr_pending)) {
3560 struct request_queue *r_queue = bdev_get_queue(rdev->bdev);
3562 atomic_inc(&rdev->nr_pending);
3565 blk_unplug(r_queue);
3567 rdev_dec_pending(rdev, mddev);
3574 static void raid5_unplug_device(struct request_queue *q)
3576 mddev_t *mddev = q->queuedata;
3577 raid5_conf_t *conf = mddev->private;
3578 unsigned long flags;
3580 spin_lock_irqsave(&conf->device_lock, flags);
3582 if (blk_remove_plug(q)) {
3584 raid5_activate_delayed(conf);
3586 md_wakeup_thread(mddev->thread);
3588 spin_unlock_irqrestore(&conf->device_lock, flags);
3590 unplug_slaves(mddev);
3593 static int raid5_congested(void *data, int bits)
3595 mddev_t *mddev = data;
3596 raid5_conf_t *conf = mddev->private;
3598 /* No difference between reads and writes. Just check
3599 * how busy the stripe_cache is
3602 if (mddev_congested(mddev, bits))
3604 if (conf->inactive_blocked)
3608 if (list_empty_careful(&conf->inactive_list))
3614 /* We want read requests to align with chunks where possible,
3615 * but write requests don't need to.
3617 static int raid5_mergeable_bvec(struct request_queue *q,
3618 struct bvec_merge_data *bvm,
3619 struct bio_vec *biovec)
3621 mddev_t *mddev = q->queuedata;
3622 sector_t sector = bvm->bi_sector + get_start_sect(bvm->bi_bdev);
3624 unsigned int chunk_sectors = mddev->chunk_sectors;
3625 unsigned int bio_sectors = bvm->bi_size >> 9;
3627 if ((bvm->bi_rw & 1) == WRITE)
3628 return biovec->bv_len; /* always allow writes to be mergeable */
3630 if (mddev->new_chunk_sectors < mddev->chunk_sectors)
3631 chunk_sectors = mddev->new_chunk_sectors;
3632 max = (chunk_sectors - ((sector & (chunk_sectors - 1)) + bio_sectors)) << 9;
3633 if (max < 0) max = 0;
3634 if (max <= biovec->bv_len && bio_sectors == 0)
3635 return biovec->bv_len;
3641 static int in_chunk_boundary(mddev_t *mddev, struct bio *bio)
3643 sector_t sector = bio->bi_sector + get_start_sect(bio->bi_bdev);
3644 unsigned int chunk_sectors = mddev->chunk_sectors;
3645 unsigned int bio_sectors = bio->bi_size >> 9;
3647 if (mddev->new_chunk_sectors < mddev->chunk_sectors)
3648 chunk_sectors = mddev->new_chunk_sectors;
3649 return chunk_sectors >=
3650 ((sector & (chunk_sectors - 1)) + bio_sectors);
3654 * add bio to the retry LIFO ( in O(1) ... we are in interrupt )
3655 * later sampled by raid5d.
3657 static void add_bio_to_retry(struct bio *bi,raid5_conf_t *conf)
3659 unsigned long flags;
3661 spin_lock_irqsave(&conf->device_lock, flags);
3663 bi->bi_next = conf->retry_read_aligned_list;
3664 conf->retry_read_aligned_list = bi;
3666 spin_unlock_irqrestore(&conf->device_lock, flags);
3667 md_wakeup_thread(conf->mddev->thread);
3671 static struct bio *remove_bio_from_retry(raid5_conf_t *conf)
3675 bi = conf->retry_read_aligned;
3677 conf->retry_read_aligned = NULL;
3680 bi = conf->retry_read_aligned_list;
3682 conf->retry_read_aligned_list = bi->bi_next;
3685 * this sets the active strip count to 1 and the processed
3686 * strip count to zero (upper 8 bits)
3688 bi->bi_phys_segments = 1; /* biased count of active stripes */
3696 * The "raid5_align_endio" should check if the read succeeded and if it
3697 * did, call bio_endio on the original bio (having bio_put the new bio
3699 * If the read failed..
3701 static void raid5_align_endio(struct bio *bi, int error)
3703 struct bio* raid_bi = bi->bi_private;
3706 int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
3711 mddev = raid_bi->bi_bdev->bd_disk->queue->queuedata;
3712 conf = mddev->private;
3713 rdev = (void*)raid_bi->bi_next;
3714 raid_bi->bi_next = NULL;
3716 rdev_dec_pending(rdev, conf->mddev);
3718 if (!error && uptodate) {
3719 bio_endio(raid_bi, 0);
3720 if (atomic_dec_and_test(&conf->active_aligned_reads))
3721 wake_up(&conf->wait_for_stripe);
3726 pr_debug("raid5_align_endio : io error...handing IO for a retry\n");
3728 add_bio_to_retry(raid_bi, conf);
3731 static int bio_fits_rdev(struct bio *bi)
3733 struct request_queue *q = bdev_get_queue(bi->bi_bdev);
3735 if ((bi->bi_size>>9) > queue_max_sectors(q))
3737 blk_recount_segments(q, bi);
3738 if (bi->bi_phys_segments > queue_max_segments(q))
3741 if (q->merge_bvec_fn)
3742 /* it's too hard to apply the merge_bvec_fn at this stage,
3751 static int chunk_aligned_read(struct request_queue *q, struct bio * raid_bio)
3753 mddev_t *mddev = q->queuedata;
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(struct request_queue *q, struct bio * bi)
3870 mddev_t *mddev = q->queuedata;
3871 raid5_conf_t *conf = mddev->private;
3873 sector_t new_sector;
3874 sector_t logical_sector, last_sector;
3875 struct stripe_head *sh;
3876 const int rw = bio_data_dir(bi);
3879 if (unlikely(bio_rw_flagged(bi, BIO_RW_BARRIER))) {
3880 /* Drain all pending writes. We only really need
3881 * to ensure they have been submitted, but this is
3884 mddev->pers->quiesce(mddev, 1);
3885 mddev->pers->quiesce(mddev, 0);
3886 md_barrier_request(mddev, bi);
3890 md_write_start(mddev, bi);
3892 cpu = part_stat_lock();
3893 part_stat_inc(cpu, &mddev->gendisk->part0, ios[rw]);
3894 part_stat_add(cpu, &mddev->gendisk->part0, sectors[rw],
3899 mddev->reshape_position == MaxSector &&
3900 chunk_aligned_read(q,bi))
3903 logical_sector = bi->bi_sector & ~((sector_t)STRIPE_SECTORS-1);
3904 last_sector = bi->bi_sector + (bi->bi_size>>9);
3906 bi->bi_phys_segments = 1; /* over-loaded to count active stripes */
3908 for (;logical_sector < last_sector; logical_sector += STRIPE_SECTORS) {
3910 int disks, data_disks;
3915 disks = conf->raid_disks;
3916 prepare_to_wait(&conf->wait_for_overlap, &w, TASK_UNINTERRUPTIBLE);
3917 if (unlikely(conf->reshape_progress != MaxSector)) {
3918 /* spinlock is needed as reshape_progress may be
3919 * 64bit on a 32bit platform, and so it might be
3920 * possible to see a half-updated value
3921 * Ofcourse reshape_progress could change after
3922 * the lock is dropped, so once we get a reference
3923 * to the stripe that we think it is, we will have
3926 spin_lock_irq(&conf->device_lock);
3927 if (mddev->delta_disks < 0
3928 ? logical_sector < conf->reshape_progress
3929 : logical_sector >= conf->reshape_progress) {
3930 disks = conf->previous_raid_disks;
3933 if (mddev->delta_disks < 0
3934 ? logical_sector < conf->reshape_safe
3935 : logical_sector >= conf->reshape_safe) {
3936 spin_unlock_irq(&conf->device_lock);
3941 spin_unlock_irq(&conf->device_lock);
3943 data_disks = disks - conf->max_degraded;
3945 new_sector = raid5_compute_sector(conf, logical_sector,
3948 pr_debug("raid5: make_request, sector %llu logical %llu\n",
3949 (unsigned long long)new_sector,
3950 (unsigned long long)logical_sector);
3952 sh = get_active_stripe(conf, new_sector, previous,
3953 (bi->bi_rw&RWA_MASK), 0);
3955 if (unlikely(previous)) {
3956 /* expansion might have moved on while waiting for a
3957 * stripe, so we must do the range check again.
3958 * Expansion could still move past after this
3959 * test, but as we are holding a reference to
3960 * 'sh', we know that if that happens,
3961 * STRIPE_EXPANDING will get set and the expansion
3962 * won't proceed until we finish with the stripe.
3965 spin_lock_irq(&conf->device_lock);
3966 if (mddev->delta_disks < 0
3967 ? logical_sector >= conf->reshape_progress
3968 : logical_sector < conf->reshape_progress)
3969 /* mismatch, need to try again */
3971 spin_unlock_irq(&conf->device_lock);
3979 if (bio_data_dir(bi) == WRITE &&
3980 logical_sector >= mddev->suspend_lo &&
3981 logical_sector < mddev->suspend_hi) {
3983 /* As the suspend_* range is controlled by
3984 * userspace, we want an interruptible
3987 flush_signals(current);
3988 prepare_to_wait(&conf->wait_for_overlap,
3989 &w, TASK_INTERRUPTIBLE);
3990 if (logical_sector >= mddev->suspend_lo &&
3991 logical_sector < mddev->suspend_hi)
3996 if (test_bit(STRIPE_EXPANDING, &sh->state) ||
3997 !add_stripe_bio(sh, bi, dd_idx, (bi->bi_rw&RW_MASK))) {
3998 /* Stripe is busy expanding or
3999 * add failed due to overlap. Flush everything
4002 raid5_unplug_device(mddev->queue);
4007 finish_wait(&conf->wait_for_overlap, &w);
4008 set_bit(STRIPE_HANDLE, &sh->state);
4009 clear_bit(STRIPE_DELAYED, &sh->state);
4010 if (mddev->barrier &&
4011 !test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
4012 atomic_inc(&conf->preread_active_stripes);
4015 /* cannot get stripe for read-ahead, just give-up */
4016 clear_bit(BIO_UPTODATE, &bi->bi_flags);
4017 finish_wait(&conf->wait_for_overlap, &w);
4022 spin_lock_irq(&conf->device_lock);
4023 remaining = raid5_dec_bi_phys_segments(bi);
4024 spin_unlock_irq(&conf->device_lock);
4025 if (remaining == 0) {
4028 md_write_end(mddev);
4033 if (mddev->barrier) {
4034 /* We need to wait for the stripes to all be handled.
4035 * So: wait for preread_active_stripes to drop to 0.
4037 wait_event(mddev->thread->wqueue,
4038 atomic_read(&conf->preread_active_stripes) == 0);
4043 static sector_t raid5_size(mddev_t *mddev, sector_t sectors, int raid_disks);
4045 static sector_t reshape_request(mddev_t *mddev, sector_t sector_nr, int *skipped)
4047 /* reshaping is quite different to recovery/resync so it is
4048 * handled quite separately ... here.
4050 * On each call to sync_request, we gather one chunk worth of
4051 * destination stripes and flag them as expanding.
4052 * Then we find all the source stripes and request reads.
4053 * As the reads complete, handle_stripe will copy the data
4054 * into the destination stripe and release that stripe.
4056 raid5_conf_t *conf = (raid5_conf_t *) mddev->private;
4057 struct stripe_head *sh;
4058 sector_t first_sector, last_sector;
4059 int raid_disks = conf->previous_raid_disks;
4060 int data_disks = raid_disks - conf->max_degraded;
4061 int new_data_disks = conf->raid_disks - conf->max_degraded;
4064 sector_t writepos, readpos, safepos;
4065 sector_t stripe_addr;
4066 int reshape_sectors;
4067 struct list_head stripes;
4069 if (sector_nr == 0) {
4070 /* If restarting in the middle, skip the initial sectors */
4071 if (mddev->delta_disks < 0 &&
4072 conf->reshape_progress < raid5_size(mddev, 0, 0)) {
4073 sector_nr = raid5_size(mddev, 0, 0)
4074 - conf->reshape_progress;
4075 } else if (mddev->delta_disks >= 0 &&
4076 conf->reshape_progress > 0)
4077 sector_nr = conf->reshape_progress;
4078 sector_div(sector_nr, new_data_disks);
4080 mddev->curr_resync_completed = sector_nr;
4081 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
4087 /* We need to process a full chunk at a time.
4088 * If old and new chunk sizes differ, we need to process the
4091 if (mddev->new_chunk_sectors > mddev->chunk_sectors)
4092 reshape_sectors = mddev->new_chunk_sectors;
4094 reshape_sectors = mddev->chunk_sectors;
4096 /* we update the metadata when there is more than 3Meg
4097 * in the block range (that is rather arbitrary, should
4098 * probably be time based) or when the data about to be
4099 * copied would over-write the source of the data at
4100 * the front of the range.
4101 * i.e. one new_stripe along from reshape_progress new_maps
4102 * to after where reshape_safe old_maps to
4104 writepos = conf->reshape_progress;
4105 sector_div(writepos, new_data_disks);
4106 readpos = conf->reshape_progress;
4107 sector_div(readpos, data_disks);
4108 safepos = conf->reshape_safe;
4109 sector_div(safepos, data_disks);
4110 if (mddev->delta_disks < 0) {
4111 writepos -= min_t(sector_t, reshape_sectors, writepos);
4112 readpos += reshape_sectors;
4113 safepos += reshape_sectors;
4115 writepos += reshape_sectors;
4116 readpos -= min_t(sector_t, reshape_sectors, readpos);
4117 safepos -= min_t(sector_t, reshape_sectors, safepos);
4120 /* 'writepos' is the most advanced device address we might write.
4121 * 'readpos' is the least advanced device address we might read.
4122 * 'safepos' is the least address recorded in the metadata as having
4124 * If 'readpos' is behind 'writepos', then there is no way that we can
4125 * ensure safety in the face of a crash - that must be done by userspace
4126 * making a backup of the data. So in that case there is no particular
4127 * rush to update metadata.
4128 * Otherwise if 'safepos' is behind 'writepos', then we really need to
4129 * update the metadata to advance 'safepos' to match 'readpos' so that
4130 * we can be safe in the event of a crash.
4131 * So we insist on updating metadata if safepos is behind writepos and
4132 * readpos is beyond writepos.
4133 * In any case, update the metadata every 10 seconds.
4134 * Maybe that number should be configurable, but I'm not sure it is
4135 * worth it.... maybe it could be a multiple of safemode_delay???
4137 if ((mddev->delta_disks < 0
4138 ? (safepos > writepos && readpos < writepos)
4139 : (safepos < writepos && readpos > writepos)) ||
4140 time_after(jiffies, conf->reshape_checkpoint + 10*HZ)) {
4141 /* Cannot proceed until we've updated the superblock... */
4142 wait_event(conf->wait_for_overlap,
4143 atomic_read(&conf->reshape_stripes)==0);
4144 mddev->reshape_position = conf->reshape_progress;
4145 mddev->curr_resync_completed = mddev->curr_resync;
4146 conf->reshape_checkpoint = jiffies;
4147 set_bit(MD_CHANGE_DEVS, &mddev->flags);
4148 md_wakeup_thread(mddev->thread);
4149 wait_event(mddev->sb_wait, mddev->flags == 0 ||
4150 kthread_should_stop());
4151 spin_lock_irq(&conf->device_lock);
4152 conf->reshape_safe = mddev->reshape_position;
4153 spin_unlock_irq(&conf->device_lock);
4154 wake_up(&conf->wait_for_overlap);
4155 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
4158 if (mddev->delta_disks < 0) {
4159 BUG_ON(conf->reshape_progress == 0);
4160 stripe_addr = writepos;
4161 BUG_ON((mddev->dev_sectors &
4162 ~((sector_t)reshape_sectors - 1))
4163 - reshape_sectors - stripe_addr
4166 BUG_ON(writepos != sector_nr + reshape_sectors);
4167 stripe_addr = sector_nr;
4169 INIT_LIST_HEAD(&stripes);
4170 for (i = 0; i < reshape_sectors; i += STRIPE_SECTORS) {
4172 int skipped_disk = 0;
4173 sh = get_active_stripe(conf, stripe_addr+i, 0, 0, 1);
4174 set_bit(STRIPE_EXPANDING, &sh->state);
4175 atomic_inc(&conf->reshape_stripes);
4176 /* If any of this stripe is beyond the end of the old
4177 * array, then we need to zero those blocks
4179 for (j=sh->disks; j--;) {
4181 if (j == sh->pd_idx)
4183 if (conf->level == 6 &&
4186 s = compute_blocknr(sh, j, 0);
4187 if (s < raid5_size(mddev, 0, 0)) {
4191 memset(page_address(sh->dev[j].page), 0, STRIPE_SIZE);
4192 set_bit(R5_Expanded, &sh->dev[j].flags);
4193 set_bit(R5_UPTODATE, &sh->dev[j].flags);
4195 if (!skipped_disk) {
4196 set_bit(STRIPE_EXPAND_READY, &sh->state);
4197 set_bit(STRIPE_HANDLE, &sh->state);
4199 list_add(&sh->lru, &stripes);
4201 spin_lock_irq(&conf->device_lock);
4202 if (mddev->delta_disks < 0)
4203 conf->reshape_progress -= reshape_sectors * new_data_disks;
4205 conf->reshape_progress += reshape_sectors * new_data_disks;
4206 spin_unlock_irq(&conf->device_lock);
4207 /* Ok, those stripe are ready. We can start scheduling
4208 * reads on the source stripes.
4209 * The source stripes are determined by mapping the first and last
4210 * block on the destination stripes.
4213 raid5_compute_sector(conf, stripe_addr*(new_data_disks),
4216 raid5_compute_sector(conf, ((stripe_addr+reshape_sectors)
4217 * new_data_disks - 1),
4219 if (last_sector >= mddev->dev_sectors)
4220 last_sector = mddev->dev_sectors - 1;
4221 while (first_sector <= last_sector) {
4222 sh = get_active_stripe(conf, first_sector, 1, 0, 1);
4223 set_bit(STRIPE_EXPAND_SOURCE, &sh->state);
4224 set_bit(STRIPE_HANDLE, &sh->state);
4226 first_sector += STRIPE_SECTORS;
4228 /* Now that the sources are clearly marked, we can release
4229 * the destination stripes
4231 while (!list_empty(&stripes)) {
4232 sh = list_entry(stripes.next, struct stripe_head, lru);
4233 list_del_init(&sh->lru);
4236 /* If this takes us to the resync_max point where we have to pause,
4237 * then we need to write out the superblock.
4239 sector_nr += reshape_sectors;
4240 if ((sector_nr - mddev->curr_resync_completed) * 2
4241 >= mddev->resync_max - mddev->curr_resync_completed) {
4242 /* Cannot proceed until we've updated the superblock... */
4243 wait_event(conf->wait_for_overlap,
4244 atomic_read(&conf->reshape_stripes) == 0);
4245 mddev->reshape_position = conf->reshape_progress;
4246 mddev->curr_resync_completed = mddev->curr_resync + reshape_sectors;
4247 conf->reshape_checkpoint = jiffies;
4248 set_bit(MD_CHANGE_DEVS, &mddev->flags);
4249 md_wakeup_thread(mddev->thread);
4250 wait_event(mddev->sb_wait,
4251 !test_bit(MD_CHANGE_DEVS, &mddev->flags)
4252 || kthread_should_stop());
4253 spin_lock_irq(&conf->device_lock);
4254 conf->reshape_safe = mddev->reshape_position;
4255 spin_unlock_irq(&conf->device_lock);
4256 wake_up(&conf->wait_for_overlap);
4257 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
4259 return reshape_sectors;
4262 /* FIXME go_faster isn't used */
4263 static inline sector_t sync_request(mddev_t *mddev, sector_t sector_nr, int *skipped, int go_faster)
4265 raid5_conf_t *conf = (raid5_conf_t *) mddev->private;
4266 struct stripe_head *sh;
4267 sector_t max_sector = mddev->dev_sectors;
4269 int still_degraded = 0;
4272 if (sector_nr >= max_sector) {
4273 /* just being told to finish up .. nothing much to do */
4274 unplug_slaves(mddev);
4276 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery)) {
4281 if (mddev->curr_resync < max_sector) /* aborted */
4282 bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
4284 else /* completed sync */
4286 bitmap_close_sync(mddev->bitmap);
4291 /* Allow raid5_quiesce to complete */
4292 wait_event(conf->wait_for_overlap, conf->quiesce != 2);
4294 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
4295 return reshape_request(mddev, sector_nr, skipped);
4297 /* No need to check resync_max as we never do more than one
4298 * stripe, and as resync_max will always be on a chunk boundary,
4299 * if the check in md_do_sync didn't fire, there is no chance
4300 * of overstepping resync_max here
4303 /* if there is too many failed drives and we are trying
4304 * to resync, then assert that we are finished, because there is
4305 * nothing we can do.
4307 if (mddev->degraded >= conf->max_degraded &&
4308 test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
4309 sector_t rv = mddev->dev_sectors - sector_nr;
4313 if (!bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
4314 !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
4315 !conf->fullsync && sync_blocks >= STRIPE_SECTORS) {
4316 /* we can skip this block, and probably more */
4317 sync_blocks /= STRIPE_SECTORS;
4319 return sync_blocks * STRIPE_SECTORS; /* keep things rounded to whole stripes */
4323 bitmap_cond_end_sync(mddev->bitmap, sector_nr);
4325 sh = get_active_stripe(conf, sector_nr, 0, 1, 0);
4327 sh = get_active_stripe(conf, sector_nr, 0, 0, 0);
4328 /* make sure we don't swamp the stripe cache if someone else
4329 * is trying to get access
4331 schedule_timeout_uninterruptible(1);
4333 /* Need to check if array will still be degraded after recovery/resync
4334 * We don't need to check the 'failed' flag as when that gets set,
4337 for (i = 0; i < conf->raid_disks; i++)
4338 if (conf->disks[i].rdev == NULL)
4341 bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, still_degraded);
4343 spin_lock(&sh->lock);
4344 set_bit(STRIPE_SYNCING, &sh->state);
4345 clear_bit(STRIPE_INSYNC, &sh->state);
4346 spin_unlock(&sh->lock);
4351 return STRIPE_SECTORS;
4354 static int retry_aligned_read(raid5_conf_t *conf, struct bio *raid_bio)
4356 /* We may not be able to submit a whole bio at once as there
4357 * may not be enough stripe_heads available.
4358 * We cannot pre-allocate enough stripe_heads as we may need
4359 * more than exist in the cache (if we allow ever large chunks).
4360 * So we do one stripe head at a time and record in
4361 * ->bi_hw_segments how many have been done.
4363 * We *know* that this entire raid_bio is in one chunk, so
4364 * it will be only one 'dd_idx' and only need one call to raid5_compute_sector.
4366 struct stripe_head *sh;
4368 sector_t sector, logical_sector, last_sector;
4373 logical_sector = raid_bio->bi_sector & ~((sector_t)STRIPE_SECTORS-1);
4374 sector = raid5_compute_sector(conf, logical_sector,
4376 last_sector = raid_bio->bi_sector + (raid_bio->bi_size>>9);
4378 for (; logical_sector < last_sector;
4379 logical_sector += STRIPE_SECTORS,
4380 sector += STRIPE_SECTORS,
4383 if (scnt < raid5_bi_hw_segments(raid_bio))
4384 /* already done this stripe */
4387 sh = get_active_stripe(conf, sector, 0, 1, 0);
4390 /* failed to get a stripe - must wait */
4391 raid5_set_bi_hw_segments(raid_bio, scnt);
4392 conf->retry_read_aligned = raid_bio;
4396 set_bit(R5_ReadError, &sh->dev[dd_idx].flags);
4397 if (!add_stripe_bio(sh, raid_bio, dd_idx, 0)) {
4399 raid5_set_bi_hw_segments(raid_bio, scnt);
4400 conf->retry_read_aligned = raid_bio;
4408 spin_lock_irq(&conf->device_lock);
4409 remaining = raid5_dec_bi_phys_segments(raid_bio);
4410 spin_unlock_irq(&conf->device_lock);
4412 bio_endio(raid_bio, 0);
4413 if (atomic_dec_and_test(&conf->active_aligned_reads))
4414 wake_up(&conf->wait_for_stripe);
4420 * This is our raid5 kernel thread.
4422 * We scan the hash table for stripes which can be handled now.
4423 * During the scan, completed stripes are saved for us by the interrupt
4424 * handler, so that they will not have to wait for our next wakeup.
4426 static void raid5d(mddev_t *mddev)
4428 struct stripe_head *sh;
4429 raid5_conf_t *conf = mddev->private;
4432 pr_debug("+++ raid5d active\n");
4434 md_check_recovery(mddev);
4437 spin_lock_irq(&conf->device_lock);
4441 if (conf->seq_flush != conf->seq_write) {
4442 int seq = conf->seq_flush;
4443 spin_unlock_irq(&conf->device_lock);
4444 bitmap_unplug(mddev->bitmap);
4445 spin_lock_irq(&conf->device_lock);
4446 conf->seq_write = seq;
4447 activate_bit_delay(conf);
4450 while ((bio = remove_bio_from_retry(conf))) {
4452 spin_unlock_irq(&conf->device_lock);
4453 ok = retry_aligned_read(conf, bio);
4454 spin_lock_irq(&conf->device_lock);
4460 sh = __get_priority_stripe(conf);
4464 spin_unlock_irq(&conf->device_lock);
4471 spin_lock_irq(&conf->device_lock);
4473 pr_debug("%d stripes handled\n", handled);
4475 spin_unlock_irq(&conf->device_lock);
4477 async_tx_issue_pending_all();
4478 unplug_slaves(mddev);
4480 pr_debug("--- raid5d inactive\n");
4484 raid5_show_stripe_cache_size(mddev_t *mddev, char *page)
4486 raid5_conf_t *conf = mddev->private;
4488 return sprintf(page, "%d\n", conf->max_nr_stripes);
4494 raid5_store_stripe_cache_size(mddev_t *mddev, const char *page, size_t len)
4496 raid5_conf_t *conf = mddev->private;
4500 if (len >= PAGE_SIZE)
4505 if (strict_strtoul(page, 10, &new))
4507 if (new <= 16 || new > 32768)
4509 while (new < conf->max_nr_stripes) {
4510 if (drop_one_stripe(conf))
4511 conf->max_nr_stripes--;
4515 err = md_allow_write(mddev);
4518 while (new > conf->max_nr_stripes) {
4519 if (grow_one_stripe(conf))
4520 conf->max_nr_stripes++;
4526 static struct md_sysfs_entry
4527 raid5_stripecache_size = __ATTR(stripe_cache_size, S_IRUGO | S_IWUSR,
4528 raid5_show_stripe_cache_size,
4529 raid5_store_stripe_cache_size);
4532 raid5_show_preread_threshold(mddev_t *mddev, char *page)
4534 raid5_conf_t *conf = mddev->private;
4536 return sprintf(page, "%d\n", conf->bypass_threshold);
4542 raid5_store_preread_threshold(mddev_t *mddev, const char *page, size_t len)
4544 raid5_conf_t *conf = mddev->private;
4546 if (len >= PAGE_SIZE)
4551 if (strict_strtoul(page, 10, &new))
4553 if (new > conf->max_nr_stripes)
4555 conf->bypass_threshold = new;
4559 static struct md_sysfs_entry
4560 raid5_preread_bypass_threshold = __ATTR(preread_bypass_threshold,
4562 raid5_show_preread_threshold,
4563 raid5_store_preread_threshold);
4566 stripe_cache_active_show(mddev_t *mddev, char *page)
4568 raid5_conf_t *conf = mddev->private;
4570 return sprintf(page, "%d\n", atomic_read(&conf->active_stripes));
4575 static struct md_sysfs_entry
4576 raid5_stripecache_active = __ATTR_RO(stripe_cache_active);
4578 static struct attribute *raid5_attrs[] = {
4579 &raid5_stripecache_size.attr,
4580 &raid5_stripecache_active.attr,
4581 &raid5_preread_bypass_threshold.attr,
4584 static struct attribute_group raid5_attrs_group = {
4586 .attrs = raid5_attrs,
4590 raid5_size(mddev_t *mddev, sector_t sectors, int raid_disks)
4592 raid5_conf_t *conf = mddev->private;
4595 sectors = mddev->dev_sectors;
4597 /* size is defined by the smallest of previous and new size */
4598 raid_disks = min(conf->raid_disks, conf->previous_raid_disks);
4600 sectors &= ~((sector_t)mddev->chunk_sectors - 1);
4601 sectors &= ~((sector_t)mddev->new_chunk_sectors - 1);
4602 return sectors * (raid_disks - conf->max_degraded);
4605 static void raid5_free_percpu(raid5_conf_t *conf)
4607 struct raid5_percpu *percpu;
4614 for_each_possible_cpu(cpu) {
4615 percpu = per_cpu_ptr(conf->percpu, cpu);
4616 safe_put_page(percpu->spare_page);
4617 kfree(percpu->scribble);
4619 #ifdef CONFIG_HOTPLUG_CPU
4620 unregister_cpu_notifier(&conf->cpu_notify);
4624 free_percpu(conf->percpu);
4627 static void free_conf(raid5_conf_t *conf)
4629 shrink_stripes(conf);
4630 raid5_free_percpu(conf);
4632 kfree(conf->stripe_hashtbl);
4636 #ifdef CONFIG_HOTPLUG_CPU
4637 static int raid456_cpu_notify(struct notifier_block *nfb, unsigned long action,
4640 raid5_conf_t *conf = container_of(nfb, raid5_conf_t, cpu_notify);
4641 long cpu = (long)hcpu;
4642 struct raid5_percpu *percpu = per_cpu_ptr(conf->percpu, cpu);
4645 case CPU_UP_PREPARE:
4646 case CPU_UP_PREPARE_FROZEN:
4647 if (conf->level == 6 && !percpu->spare_page)
4648 percpu->spare_page = alloc_page(GFP_KERNEL);
4649 if (!percpu->scribble)
4650 percpu->scribble = kmalloc(conf->scribble_len, GFP_KERNEL);
4652 if (!percpu->scribble ||
4653 (conf->level == 6 && !percpu->spare_page)) {
4654 safe_put_page(percpu->spare_page);
4655 kfree(percpu->scribble);
4656 pr_err("%s: failed memory allocation for cpu%ld\n",
4662 case CPU_DEAD_FROZEN:
4663 safe_put_page(percpu->spare_page);
4664 kfree(percpu->scribble);
4665 percpu->spare_page = NULL;
4666 percpu->scribble = NULL;
4675 static int raid5_alloc_percpu(raid5_conf_t *conf)
4678 struct page *spare_page;
4679 struct raid5_percpu __percpu *allcpus;
4683 allcpus = alloc_percpu(struct raid5_percpu);
4686 conf->percpu = allcpus;
4690 for_each_present_cpu(cpu) {
4691 if (conf->level == 6) {
4692 spare_page = alloc_page(GFP_KERNEL);
4697 per_cpu_ptr(conf->percpu, cpu)->spare_page = spare_page;
4699 scribble = kmalloc(conf->scribble_len, GFP_KERNEL);
4704 per_cpu_ptr(conf->percpu, cpu)->scribble = scribble;
4706 #ifdef CONFIG_HOTPLUG_CPU
4707 conf->cpu_notify.notifier_call = raid456_cpu_notify;
4708 conf->cpu_notify.priority = 0;
4710 err = register_cpu_notifier(&conf->cpu_notify);
4717 static raid5_conf_t *setup_conf(mddev_t *mddev)
4720 int raid_disk, memory, max_disks;
4722 struct disk_info *disk;
4724 if (mddev->new_level != 5
4725 && mddev->new_level != 4
4726 && mddev->new_level != 6) {
4727 printk(KERN_ERR "raid5: %s: raid level not set to 4/5/6 (%d)\n",
4728 mdname(mddev), mddev->new_level);
4729 return ERR_PTR(-EIO);
4731 if ((mddev->new_level == 5
4732 && !algorithm_valid_raid5(mddev->new_layout)) ||
4733 (mddev->new_level == 6
4734 && !algorithm_valid_raid6(mddev->new_layout))) {
4735 printk(KERN_ERR "raid5: %s: layout %d not supported\n",
4736 mdname(mddev), mddev->new_layout);
4737 return ERR_PTR(-EIO);
4739 if (mddev->new_level == 6 && mddev->raid_disks < 4) {
4740 printk(KERN_ERR "raid6: not enough configured devices for %s (%d, minimum 4)\n",
4741 mdname(mddev), mddev->raid_disks);
4742 return ERR_PTR(-EINVAL);
4745 if (!mddev->new_chunk_sectors ||
4746 (mddev->new_chunk_sectors << 9) % PAGE_SIZE ||
4747 !is_power_of_2(mddev->new_chunk_sectors)) {
4748 printk(KERN_ERR "raid5: invalid chunk size %d for %s\n",
4749 mddev->new_chunk_sectors << 9, mdname(mddev));
4750 return ERR_PTR(-EINVAL);
4753 conf = kzalloc(sizeof(raid5_conf_t), GFP_KERNEL);
4756 spin_lock_init(&conf->device_lock);
4757 init_waitqueue_head(&conf->wait_for_stripe);
4758 init_waitqueue_head(&conf->wait_for_overlap);
4759 INIT_LIST_HEAD(&conf->handle_list);
4760 INIT_LIST_HEAD(&conf->hold_list);
4761 INIT_LIST_HEAD(&conf->delayed_list);
4762 INIT_LIST_HEAD(&conf->bitmap_list);
4763 INIT_LIST_HEAD(&conf->inactive_list);
4764 atomic_set(&conf->active_stripes, 0);
4765 atomic_set(&conf->preread_active_stripes, 0);
4766 atomic_set(&conf->active_aligned_reads, 0);
4767 conf->bypass_threshold = BYPASS_THRESHOLD;
4769 conf->raid_disks = mddev->raid_disks;
4770 if (mddev->reshape_position == MaxSector)
4771 conf->previous_raid_disks = mddev->raid_disks;
4773 conf->previous_raid_disks = mddev->raid_disks - mddev->delta_disks;
4774 max_disks = max(conf->raid_disks, conf->previous_raid_disks);
4775 conf->scribble_len = scribble_len(max_disks);
4777 conf->disks = kzalloc(max_disks * sizeof(struct disk_info),
4782 conf->mddev = mddev;
4784 if ((conf->stripe_hashtbl = kzalloc(PAGE_SIZE, GFP_KERNEL)) == NULL)
4787 conf->level = mddev->new_level;
4788 if (raid5_alloc_percpu(conf) != 0)
4791 pr_debug("raid5: run(%s) called.\n", mdname(mddev));
4793 list_for_each_entry(rdev, &mddev->disks, same_set) {
4794 raid_disk = rdev->raid_disk;
4795 if (raid_disk >= max_disks
4798 disk = conf->disks + raid_disk;
4802 if (test_bit(In_sync, &rdev->flags)) {
4803 char b[BDEVNAME_SIZE];
4804 printk(KERN_INFO "raid5: device %s operational as raid"
4805 " disk %d\n", bdevname(rdev->bdev,b),
4808 /* Cannot rely on bitmap to complete recovery */
4812 conf->chunk_sectors = mddev->new_chunk_sectors;
4813 conf->level = mddev->new_level;
4814 if (conf->level == 6)
4815 conf->max_degraded = 2;
4817 conf->max_degraded = 1;
4818 conf->algorithm = mddev->new_layout;
4819 conf->max_nr_stripes = NR_STRIPES;
4820 conf->reshape_progress = mddev->reshape_position;
4821 if (conf->reshape_progress != MaxSector) {
4822 conf->prev_chunk_sectors = mddev->chunk_sectors;
4823 conf->prev_algo = mddev->layout;
4826 memory = conf->max_nr_stripes * (sizeof(struct stripe_head) +
4827 max_disks * ((sizeof(struct bio) + PAGE_SIZE))) / 1024;
4828 if (grow_stripes(conf, conf->max_nr_stripes)) {
4830 "raid5: couldn't allocate %dkB for buffers\n", memory);
4833 printk(KERN_INFO "raid5: allocated %dkB for %s\n",
4834 memory, mdname(mddev));
4836 conf->thread = md_register_thread(raid5d, mddev, NULL);
4837 if (!conf->thread) {
4839 "raid5: couldn't allocate thread for %s\n",
4849 return ERR_PTR(-EIO);
4851 return ERR_PTR(-ENOMEM);
4855 static int only_parity(int raid_disk, int algo, int raid_disks, int max_degraded)
4858 case ALGORITHM_PARITY_0:
4859 if (raid_disk < max_degraded)
4862 case ALGORITHM_PARITY_N:
4863 if (raid_disk >= raid_disks - max_degraded)
4866 case ALGORITHM_PARITY_0_6:
4867 if (raid_disk == 0 ||
4868 raid_disk == raid_disks - 1)
4871 case ALGORITHM_LEFT_ASYMMETRIC_6:
4872 case ALGORITHM_RIGHT_ASYMMETRIC_6:
4873 case ALGORITHM_LEFT_SYMMETRIC_6:
4874 case ALGORITHM_RIGHT_SYMMETRIC_6:
4875 if (raid_disk == raid_disks - 1)
4881 static int run(mddev_t *mddev)
4884 int working_disks = 0, chunk_size;
4885 int dirty_parity_disks = 0;
4887 sector_t reshape_offset = 0;
4889 if (mddev->recovery_cp != MaxSector)
4890 printk(KERN_NOTICE "raid5: %s is not clean"
4891 " -- starting background reconstruction\n",
4893 if (mddev->reshape_position != MaxSector) {
4894 /* Check that we can continue the reshape.
4895 * Currently only disks can change, it must
4896 * increase, and we must be past the point where
4897 * a stripe over-writes itself
4899 sector_t here_new, here_old;
4901 int max_degraded = (mddev->level == 6 ? 2 : 1);
4903 if (mddev->new_level != mddev->level) {
4904 printk(KERN_ERR "raid5: %s: unsupported reshape "
4905 "required - aborting.\n",
4909 old_disks = mddev->raid_disks - mddev->delta_disks;
4910 /* reshape_position must be on a new-stripe boundary, and one
4911 * further up in new geometry must map after here in old
4914 here_new = mddev->reshape_position;
4915 if (sector_div(here_new, mddev->new_chunk_sectors *
4916 (mddev->raid_disks - max_degraded))) {
4917 printk(KERN_ERR "raid5: reshape_position not "
4918 "on a stripe boundary\n");
4921 reshape_offset = here_new * mddev->new_chunk_sectors;
4922 /* here_new is the stripe we will write to */
4923 here_old = mddev->reshape_position;
4924 sector_div(here_old, mddev->chunk_sectors *
4925 (old_disks-max_degraded));
4926 /* here_old is the first stripe that we might need to read
4928 if (mddev->delta_disks == 0) {
4929 /* We cannot be sure it is safe to start an in-place
4930 * reshape. It is only safe if user-space if monitoring
4931 * and taking constant backups.
4932 * mdadm always starts a situation like this in
4933 * readonly mode so it can take control before
4934 * allowing any writes. So just check for that.
4936 if ((here_new * mddev->new_chunk_sectors !=
4937 here_old * mddev->chunk_sectors) ||
4939 printk(KERN_ERR "raid5: in-place reshape must be started"
4940 " in read-only mode - aborting\n");
4943 } else if (mddev->delta_disks < 0
4944 ? (here_new * mddev->new_chunk_sectors <=
4945 here_old * mddev->chunk_sectors)
4946 : (here_new * mddev->new_chunk_sectors >=
4947 here_old * mddev->chunk_sectors)) {
4948 /* Reading from the same stripe as writing to - bad */
4949 printk(KERN_ERR "raid5: reshape_position too early for "
4950 "auto-recovery - aborting.\n");
4953 printk(KERN_INFO "raid5: reshape will continue\n");
4954 /* OK, we should be able to continue; */
4956 BUG_ON(mddev->level != mddev->new_level);
4957 BUG_ON(mddev->layout != mddev->new_layout);
4958 BUG_ON(mddev->chunk_sectors != mddev->new_chunk_sectors);
4959 BUG_ON(mddev->delta_disks != 0);
4962 if (mddev->private == NULL)
4963 conf = setup_conf(mddev);
4965 conf = mddev->private;
4968 return PTR_ERR(conf);
4970 mddev->thread = conf->thread;
4971 conf->thread = NULL;
4972 mddev->private = conf;
4975 * 0 for a fully functional array, 1 or 2 for a degraded array.
4977 list_for_each_entry(rdev, &mddev->disks, same_set) {
4978 if (rdev->raid_disk < 0)
4980 if (test_bit(In_sync, &rdev->flags))
4982 /* This disc is not fully in-sync. However if it
4983 * just stored parity (beyond the recovery_offset),
4984 * when we don't need to be concerned about the
4985 * array being dirty.
4986 * When reshape goes 'backwards', we never have
4987 * partially completed devices, so we only need
4988 * to worry about reshape going forwards.
4990 /* Hack because v0.91 doesn't store recovery_offset properly. */
4991 if (mddev->major_version == 0 &&
4992 mddev->minor_version > 90)
4993 rdev->recovery_offset = reshape_offset;
4995 printk("%d: w=%d pa=%d pr=%d m=%d a=%d r=%d op1=%d op2=%d\n",
4996 rdev->raid_disk, working_disks, conf->prev_algo,
4997 conf->previous_raid_disks, conf->max_degraded,
4998 conf->algorithm, conf->raid_disks,
4999 only_parity(rdev->raid_disk,
5001 conf->previous_raid_disks,
5002 conf->max_degraded),
5003 only_parity(rdev->raid_disk,
5006 conf->max_degraded));
5007 if (rdev->recovery_offset < reshape_offset) {
5008 /* We need to check old and new layout */
5009 if (!only_parity(rdev->raid_disk,
5012 conf->max_degraded))
5015 if (!only_parity(rdev->raid_disk,
5017 conf->previous_raid_disks,
5018 conf->max_degraded))
5020 dirty_parity_disks++;
5023 mddev->degraded = (max(conf->raid_disks, conf->previous_raid_disks)
5026 if (mddev->degraded > conf->max_degraded) {
5027 printk(KERN_ERR "raid5: not enough operational devices for %s"
5028 " (%d/%d failed)\n",
5029 mdname(mddev), mddev->degraded, conf->raid_disks);
5033 /* device size must be a multiple of chunk size */
5034 mddev->dev_sectors &= ~(mddev->chunk_sectors - 1);
5035 mddev->resync_max_sectors = mddev->dev_sectors;
5037 if (mddev->degraded > dirty_parity_disks &&
5038 mddev->recovery_cp != MaxSector) {
5039 if (mddev->ok_start_degraded)
5041 "raid5: starting dirty degraded array: %s"
5042 "- data corruption possible.\n",
5046 "raid5: cannot start dirty degraded array for %s\n",
5052 if (mddev->degraded == 0)
5053 printk("raid5: raid level %d set %s active with %d out of %d"
5054 " devices, algorithm %d\n", conf->level, mdname(mddev),
5055 mddev->raid_disks-mddev->degraded, mddev->raid_disks,
5058 printk(KERN_ALERT "raid5: raid level %d set %s active with %d"
5059 " out of %d devices, algorithm %d\n", conf->level,
5060 mdname(mddev), mddev->raid_disks - mddev->degraded,
5061 mddev->raid_disks, mddev->new_layout);
5063 print_raid5_conf(conf);
5065 if (conf->reshape_progress != MaxSector) {
5066 printk("...ok start reshape thread\n");
5067 conf->reshape_safe = conf->reshape_progress;
5068 atomic_set(&conf->reshape_stripes, 0);
5069 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
5070 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
5071 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
5072 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
5073 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
5077 /* read-ahead size must cover two whole stripes, which is
5078 * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
5081 int data_disks = conf->previous_raid_disks - conf->max_degraded;
5082 int stripe = data_disks *
5083 ((mddev->chunk_sectors << 9) / PAGE_SIZE);
5084 if (mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
5085 mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
5088 /* Ok, everything is just fine now */
5089 if (sysfs_create_group(&mddev->kobj, &raid5_attrs_group))
5091 "raid5: failed to create sysfs attributes for %s\n",
5094 mddev->queue->queue_lock = &conf->device_lock;
5096 mddev->queue->unplug_fn = raid5_unplug_device;
5097 mddev->queue->backing_dev_info.congested_data = mddev;
5098 mddev->queue->backing_dev_info.congested_fn = raid5_congested;
5100 md_set_array_sectors(mddev, raid5_size(mddev, 0, 0));
5102 blk_queue_merge_bvec(mddev->queue, raid5_mergeable_bvec);
5103 chunk_size = mddev->chunk_sectors << 9;
5104 blk_queue_io_min(mddev->queue, chunk_size);
5105 blk_queue_io_opt(mddev->queue, chunk_size *
5106 (conf->raid_disks - conf->max_degraded));
5108 list_for_each_entry(rdev, &mddev->disks, same_set)
5109 disk_stack_limits(mddev->gendisk, rdev->bdev,
5110 rdev->data_offset << 9);
5114 md_unregister_thread(mddev->thread);
5115 mddev->thread = NULL;
5117 print_raid5_conf(conf);
5120 mddev->private = NULL;
5121 printk(KERN_ALERT "raid5: failed to run raid set %s\n", mdname(mddev));
5127 static int stop(mddev_t *mddev)
5129 raid5_conf_t *conf = (raid5_conf_t *) mddev->private;
5131 md_unregister_thread(mddev->thread);
5132 mddev->thread = NULL;
5133 mddev->queue->backing_dev_info.congested_fn = NULL;
5134 blk_sync_queue(mddev->queue); /* the unplug fn references 'conf'*/
5136 mddev->private = &raid5_attrs_group;
5141 static void print_sh(struct seq_file *seq, struct stripe_head *sh)
5145 seq_printf(seq, "sh %llu, pd_idx %d, state %ld.\n",
5146 (unsigned long long)sh->sector, sh->pd_idx, sh->state);
5147 seq_printf(seq, "sh %llu, count %d.\n",
5148 (unsigned long long)sh->sector, atomic_read(&sh->count));
5149 seq_printf(seq, "sh %llu, ", (unsigned long long)sh->sector);
5150 for (i = 0; i < sh->disks; i++) {
5151 seq_printf(seq, "(cache%d: %p %ld) ",
5152 i, sh->dev[i].page, sh->dev[i].flags);
5154 seq_printf(seq, "\n");
5157 static void printall(struct seq_file *seq, raid5_conf_t *conf)
5159 struct stripe_head *sh;
5160 struct hlist_node *hn;
5163 spin_lock_irq(&conf->device_lock);
5164 for (i = 0; i < NR_HASH; i++) {
5165 hlist_for_each_entry(sh, hn, &conf->stripe_hashtbl[i], hash) {
5166 if (sh->raid_conf != conf)
5171 spin_unlock_irq(&conf->device_lock);
5175 static void status(struct seq_file *seq, mddev_t *mddev)
5177 raid5_conf_t *conf = (raid5_conf_t *) mddev->private;
5180 seq_printf(seq, " level %d, %dk chunk, algorithm %d", mddev->level,
5181 mddev->chunk_sectors / 2, mddev->layout);
5182 seq_printf (seq, " [%d/%d] [", conf->raid_disks, conf->raid_disks - mddev->degraded);
5183 for (i = 0; i < conf->raid_disks; i++)
5184 seq_printf (seq, "%s",
5185 conf->disks[i].rdev &&
5186 test_bit(In_sync, &conf->disks[i].rdev->flags) ? "U" : "_");
5187 seq_printf (seq, "]");
5189 seq_printf (seq, "\n");
5190 printall(seq, conf);
5194 static void print_raid5_conf (raid5_conf_t *conf)
5197 struct disk_info *tmp;
5199 printk("RAID5 conf printout:\n");
5201 printk("(conf==NULL)\n");
5204 printk(" --- rd:%d wd:%d\n", conf->raid_disks,
5205 conf->raid_disks - conf->mddev->degraded);
5207 for (i = 0; i < conf->raid_disks; i++) {
5208 char b[BDEVNAME_SIZE];
5209 tmp = conf->disks + i;
5211 printk(" disk %d, o:%d, dev:%s\n",
5212 i, !test_bit(Faulty, &tmp->rdev->flags),
5213 bdevname(tmp->rdev->bdev,b));
5217 static int raid5_spare_active(mddev_t *mddev)
5220 raid5_conf_t *conf = mddev->private;
5221 struct disk_info *tmp;
5223 for (i = 0; i < conf->raid_disks; i++) {
5224 tmp = conf->disks + i;
5226 && !test_bit(Faulty, &tmp->rdev->flags)
5227 && !test_and_set_bit(In_sync, &tmp->rdev->flags)) {
5228 unsigned long flags;
5229 spin_lock_irqsave(&conf->device_lock, flags);
5231 spin_unlock_irqrestore(&conf->device_lock, flags);
5234 print_raid5_conf(conf);
5238 static int raid5_remove_disk(mddev_t *mddev, int number)
5240 raid5_conf_t *conf = mddev->private;
5243 struct disk_info *p = conf->disks + number;
5245 print_raid5_conf(conf);
5248 if (number >= conf->raid_disks &&
5249 conf->reshape_progress == MaxSector)
5250 clear_bit(In_sync, &rdev->flags);
5252 if (test_bit(In_sync, &rdev->flags) ||
5253 atomic_read(&rdev->nr_pending)) {
5257 /* Only remove non-faulty devices if recovery
5260 if (!test_bit(Faulty, &rdev->flags) &&
5261 mddev->degraded <= conf->max_degraded &&
5262 number < conf->raid_disks) {
5268 if (atomic_read(&rdev->nr_pending)) {
5269 /* lost the race, try later */
5276 print_raid5_conf(conf);
5280 static int raid5_add_disk(mddev_t *mddev, mdk_rdev_t *rdev)
5282 raid5_conf_t *conf = mddev->private;
5285 struct disk_info *p;
5287 int last = conf->raid_disks - 1;
5289 if (mddev->degraded > conf->max_degraded)
5290 /* no point adding a device */
5293 if (rdev->raid_disk >= 0)
5294 first = last = rdev->raid_disk;
5297 * find the disk ... but prefer rdev->saved_raid_disk
5300 if (rdev->saved_raid_disk >= 0 &&
5301 rdev->saved_raid_disk >= first &&
5302 conf->disks[rdev->saved_raid_disk].rdev == NULL)
5303 disk = rdev->saved_raid_disk;
5306 for ( ; disk <= last ; disk++)
5307 if ((p=conf->disks + disk)->rdev == NULL) {
5308 clear_bit(In_sync, &rdev->flags);
5309 rdev->raid_disk = disk;
5311 if (rdev->saved_raid_disk != disk)
5313 rcu_assign_pointer(p->rdev, rdev);
5316 print_raid5_conf(conf);
5320 static int raid5_resize(mddev_t *mddev, sector_t sectors)
5322 /* no resync is happening, and there is enough space
5323 * on all devices, so we can resize.
5324 * We need to make sure resync covers any new space.
5325 * If the array is shrinking we should possibly wait until
5326 * any io in the removed space completes, but it hardly seems
5329 sectors &= ~((sector_t)mddev->chunk_sectors - 1);
5330 md_set_array_sectors(mddev, raid5_size(mddev, sectors,
5331 mddev->raid_disks));
5332 if (mddev->array_sectors >
5333 raid5_size(mddev, sectors, mddev->raid_disks))
5335 set_capacity(mddev->gendisk, mddev->array_sectors);
5337 revalidate_disk(mddev->gendisk);
5338 if (sectors > mddev->dev_sectors && mddev->recovery_cp == MaxSector) {
5339 mddev->recovery_cp = mddev->dev_sectors;
5340 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
5342 mddev->dev_sectors = sectors;
5343 mddev->resync_max_sectors = sectors;
5347 static int check_stripe_cache(mddev_t *mddev)
5349 /* Can only proceed if there are plenty of stripe_heads.
5350 * We need a minimum of one full stripe,, and for sensible progress
5351 * it is best to have about 4 times that.
5352 * If we require 4 times, then the default 256 4K stripe_heads will
5353 * allow for chunk sizes up to 256K, which is probably OK.
5354 * If the chunk size is greater, user-space should request more
5355 * stripe_heads first.
5357 raid5_conf_t *conf = mddev->private;
5358 if (((mddev->chunk_sectors << 9) / STRIPE_SIZE) * 4
5359 > conf->max_nr_stripes ||
5360 ((mddev->new_chunk_sectors << 9) / STRIPE_SIZE) * 4
5361 > conf->max_nr_stripes) {
5362 printk(KERN_WARNING "raid5: reshape: not enough stripes. Needed %lu\n",
5363 ((max(mddev->chunk_sectors, mddev->new_chunk_sectors) << 9)
5370 static int check_reshape(mddev_t *mddev)
5372 raid5_conf_t *conf = mddev->private;
5374 if (mddev->delta_disks == 0 &&
5375 mddev->new_layout == mddev->layout &&
5376 mddev->new_chunk_sectors == mddev->chunk_sectors)
5377 return 0; /* nothing to do */
5379 /* Cannot grow a bitmap yet */
5381 if (mddev->degraded > conf->max_degraded)
5383 if (mddev->delta_disks < 0) {
5384 /* We might be able to shrink, but the devices must
5385 * be made bigger first.
5386 * For raid6, 4 is the minimum size.
5387 * Otherwise 2 is the minimum
5390 if (mddev->level == 6)
5392 if (mddev->raid_disks + mddev->delta_disks < min)
5396 if (!check_stripe_cache(mddev))
5399 return resize_stripes(conf, conf->raid_disks + mddev->delta_disks);
5402 static int raid5_start_reshape(mddev_t *mddev)
5404 raid5_conf_t *conf = mddev->private;
5407 int added_devices = 0;
5408 unsigned long flags;
5410 if (test_bit(MD_RECOVERY_RUNNING, &mddev->recovery))
5413 if (!check_stripe_cache(mddev))
5416 list_for_each_entry(rdev, &mddev->disks, same_set)
5417 if (rdev->raid_disk < 0 &&
5418 !test_bit(Faulty, &rdev->flags))
5421 if (spares - mddev->degraded < mddev->delta_disks - conf->max_degraded)
5422 /* Not enough devices even to make a degraded array
5427 /* Refuse to reduce size of the array. Any reductions in
5428 * array size must be through explicit setting of array_size
5431 if (raid5_size(mddev, 0, conf->raid_disks + mddev->delta_disks)
5432 < mddev->array_sectors) {
5433 printk(KERN_ERR "md: %s: array size must be reduced "
5434 "before number of disks\n", mdname(mddev));
5438 atomic_set(&conf->reshape_stripes, 0);
5439 spin_lock_irq(&conf->device_lock);
5440 conf->previous_raid_disks = conf->raid_disks;
5441 conf->raid_disks += mddev->delta_disks;
5442 conf->prev_chunk_sectors = conf->chunk_sectors;
5443 conf->chunk_sectors = mddev->new_chunk_sectors;
5444 conf->prev_algo = conf->algorithm;
5445 conf->algorithm = mddev->new_layout;
5446 if (mddev->delta_disks < 0)
5447 conf->reshape_progress = raid5_size(mddev, 0, 0);
5449 conf->reshape_progress = 0;
5450 conf->reshape_safe = conf->reshape_progress;
5452 spin_unlock_irq(&conf->device_lock);
5454 /* Add some new drives, as many as will fit.
5455 * We know there are enough to make the newly sized array work.
5457 list_for_each_entry(rdev, &mddev->disks, same_set)
5458 if (rdev->raid_disk < 0 &&
5459 !test_bit(Faulty, &rdev->flags)) {
5460 if (raid5_add_disk(mddev, rdev) == 0) {
5462 if (rdev->raid_disk >= conf->previous_raid_disks) {
5463 set_bit(In_sync, &rdev->flags);
5466 rdev->recovery_offset = 0;
5467 sprintf(nm, "rd%d", rdev->raid_disk);
5468 if (sysfs_create_link(&mddev->kobj,
5471 "raid5: failed to create "
5472 " link %s for %s\n",
5478 /* When a reshape changes the number of devices, ->degraded
5479 * is measured against the large of the pre and post number of
5481 if (mddev->delta_disks > 0) {
5482 spin_lock_irqsave(&conf->device_lock, flags);
5483 mddev->degraded += (conf->raid_disks - conf->previous_raid_disks)
5485 spin_unlock_irqrestore(&conf->device_lock, flags);
5487 mddev->raid_disks = conf->raid_disks;
5488 mddev->reshape_position = conf->reshape_progress;
5489 set_bit(MD_CHANGE_DEVS, &mddev->flags);
5491 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
5492 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
5493 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
5494 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
5495 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
5497 if (!mddev->sync_thread) {
5498 mddev->recovery = 0;
5499 spin_lock_irq(&conf->device_lock);
5500 mddev->raid_disks = conf->raid_disks = conf->previous_raid_disks;
5501 conf->reshape_progress = MaxSector;
5502 spin_unlock_irq(&conf->device_lock);
5505 conf->reshape_checkpoint = jiffies;
5506 md_wakeup_thread(mddev->sync_thread);
5507 md_new_event(mddev);
5511 /* This is called from the reshape thread and should make any
5512 * changes needed in 'conf'
5514 static void end_reshape(raid5_conf_t *conf)
5517 if (!test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery)) {
5519 spin_lock_irq(&conf->device_lock);
5520 conf->previous_raid_disks = conf->raid_disks;
5521 conf->reshape_progress = MaxSector;
5522 spin_unlock_irq(&conf->device_lock);
5523 wake_up(&conf->wait_for_overlap);
5525 /* read-ahead size must cover two whole stripes, which is
5526 * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
5529 int data_disks = conf->raid_disks - conf->max_degraded;
5530 int stripe = data_disks * ((conf->chunk_sectors << 9)
5532 if (conf->mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
5533 conf->mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
5538 /* This is called from the raid5d thread with mddev_lock held.
5539 * It makes config changes to the device.
5541 static void raid5_finish_reshape(mddev_t *mddev)
5543 raid5_conf_t *conf = mddev->private;
5545 if (!test_bit(MD_RECOVERY_INTR, &mddev->recovery)) {
5547 if (mddev->delta_disks > 0) {
5548 md_set_array_sectors(mddev, raid5_size(mddev, 0, 0));
5549 set_capacity(mddev->gendisk, mddev->array_sectors);
5551 revalidate_disk(mddev->gendisk);
5554 mddev->degraded = conf->raid_disks;
5555 for (d = 0; d < conf->raid_disks ; d++)
5556 if (conf->disks[d].rdev &&
5558 &conf->disks[d].rdev->flags))
5560 for (d = conf->raid_disks ;
5561 d < conf->raid_disks - mddev->delta_disks;
5563 mdk_rdev_t *rdev = conf->disks[d].rdev;
5564 if (rdev && raid5_remove_disk(mddev, d) == 0) {
5566 sprintf(nm, "rd%d", rdev->raid_disk);
5567 sysfs_remove_link(&mddev->kobj, nm);
5568 rdev->raid_disk = -1;
5572 mddev->layout = conf->algorithm;
5573 mddev->chunk_sectors = conf->chunk_sectors;
5574 mddev->reshape_position = MaxSector;
5575 mddev->delta_disks = 0;
5579 static void raid5_quiesce(mddev_t *mddev, int state)
5581 raid5_conf_t *conf = mddev->private;
5584 case 2: /* resume for a suspend */
5585 wake_up(&conf->wait_for_overlap);
5588 case 1: /* stop all writes */
5589 spin_lock_irq(&conf->device_lock);
5590 /* '2' tells resync/reshape to pause so that all
5591 * active stripes can drain
5594 wait_event_lock_irq(conf->wait_for_stripe,
5595 atomic_read(&conf->active_stripes) == 0 &&
5596 atomic_read(&conf->active_aligned_reads) == 0,
5597 conf->device_lock, /* nothing */);
5599 spin_unlock_irq(&conf->device_lock);
5600 /* allow reshape to continue */
5601 wake_up(&conf->wait_for_overlap);
5604 case 0: /* re-enable writes */
5605 spin_lock_irq(&conf->device_lock);
5607 wake_up(&conf->wait_for_stripe);
5608 wake_up(&conf->wait_for_overlap);
5609 spin_unlock_irq(&conf->device_lock);
5615 static void *raid5_takeover_raid1(mddev_t *mddev)
5619 if (mddev->raid_disks != 2 ||
5620 mddev->degraded > 1)
5621 return ERR_PTR(-EINVAL);
5623 /* Should check if there are write-behind devices? */
5625 chunksect = 64*2; /* 64K by default */
5627 /* The array must be an exact multiple of chunksize */
5628 while (chunksect && (mddev->array_sectors & (chunksect-1)))
5631 if ((chunksect<<9) < STRIPE_SIZE)
5632 /* array size does not allow a suitable chunk size */
5633 return ERR_PTR(-EINVAL);
5635 mddev->new_level = 5;
5636 mddev->new_layout = ALGORITHM_LEFT_SYMMETRIC;
5637 mddev->new_chunk_sectors = chunksect;
5639 return setup_conf(mddev);
5642 static void *raid5_takeover_raid6(mddev_t *mddev)
5646 switch (mddev->layout) {
5647 case ALGORITHM_LEFT_ASYMMETRIC_6:
5648 new_layout = ALGORITHM_LEFT_ASYMMETRIC;
5650 case ALGORITHM_RIGHT_ASYMMETRIC_6:
5651 new_layout = ALGORITHM_RIGHT_ASYMMETRIC;
5653 case ALGORITHM_LEFT_SYMMETRIC_6:
5654 new_layout = ALGORITHM_LEFT_SYMMETRIC;
5656 case ALGORITHM_RIGHT_SYMMETRIC_6:
5657 new_layout = ALGORITHM_RIGHT_SYMMETRIC;
5659 case ALGORITHM_PARITY_0_6:
5660 new_layout = ALGORITHM_PARITY_0;
5662 case ALGORITHM_PARITY_N:
5663 new_layout = ALGORITHM_PARITY_N;
5666 return ERR_PTR(-EINVAL);
5668 mddev->new_level = 5;
5669 mddev->new_layout = new_layout;
5670 mddev->delta_disks = -1;
5671 mddev->raid_disks -= 1;
5672 return setup_conf(mddev);
5676 static int raid5_check_reshape(mddev_t *mddev)
5678 /* For a 2-drive array, the layout and chunk size can be changed
5679 * immediately as not restriping is needed.
5680 * For larger arrays we record the new value - after validation
5681 * to be used by a reshape pass.
5683 raid5_conf_t *conf = mddev->private;
5684 int new_chunk = mddev->new_chunk_sectors;
5686 if (mddev->new_layout >= 0 && !algorithm_valid_raid5(mddev->new_layout))
5688 if (new_chunk > 0) {
5689 if (!is_power_of_2(new_chunk))
5691 if (new_chunk < (PAGE_SIZE>>9))
5693 if (mddev->array_sectors & (new_chunk-1))
5694 /* not factor of array size */
5698 /* They look valid */
5700 if (mddev->raid_disks == 2) {
5701 /* can make the change immediately */
5702 if (mddev->new_layout >= 0) {
5703 conf->algorithm = mddev->new_layout;
5704 mddev->layout = mddev->new_layout;
5706 if (new_chunk > 0) {
5707 conf->chunk_sectors = new_chunk ;
5708 mddev->chunk_sectors = new_chunk;
5710 set_bit(MD_CHANGE_DEVS, &mddev->flags);
5711 md_wakeup_thread(mddev->thread);
5713 return check_reshape(mddev);
5716 static int raid6_check_reshape(mddev_t *mddev)
5718 int new_chunk = mddev->new_chunk_sectors;
5720 if (mddev->new_layout >= 0 && !algorithm_valid_raid6(mddev->new_layout))
5722 if (new_chunk > 0) {
5723 if (!is_power_of_2(new_chunk))
5725 if (new_chunk < (PAGE_SIZE >> 9))
5727 if (mddev->array_sectors & (new_chunk-1))
5728 /* not factor of array size */
5732 /* They look valid */
5733 return check_reshape(mddev);
5736 static void *raid5_takeover(mddev_t *mddev)
5738 /* raid5 can take over:
5739 * raid0 - if all devices are the same - make it a raid4 layout
5740 * raid1 - if there are two drives. We need to know the chunk size
5741 * raid4 - trivial - just use a raid4 layout.
5742 * raid6 - Providing it is a *_6 layout
5745 if (mddev->level == 1)
5746 return raid5_takeover_raid1(mddev);
5747 if (mddev->level == 4) {
5748 mddev->new_layout = ALGORITHM_PARITY_N;
5749 mddev->new_level = 5;
5750 return setup_conf(mddev);
5752 if (mddev->level == 6)
5753 return raid5_takeover_raid6(mddev);
5755 return ERR_PTR(-EINVAL);
5759 static struct mdk_personality raid5_personality;
5761 static void *raid6_takeover(mddev_t *mddev)
5763 /* Currently can only take over a raid5. We map the
5764 * personality to an equivalent raid6 personality
5765 * with the Q block at the end.
5769 if (mddev->pers != &raid5_personality)
5770 return ERR_PTR(-EINVAL);
5771 if (mddev->degraded > 1)
5772 return ERR_PTR(-EINVAL);
5773 if (mddev->raid_disks > 253)
5774 return ERR_PTR(-EINVAL);
5775 if (mddev->raid_disks < 3)
5776 return ERR_PTR(-EINVAL);
5778 switch (mddev->layout) {
5779 case ALGORITHM_LEFT_ASYMMETRIC:
5780 new_layout = ALGORITHM_LEFT_ASYMMETRIC_6;
5782 case ALGORITHM_RIGHT_ASYMMETRIC:
5783 new_layout = ALGORITHM_RIGHT_ASYMMETRIC_6;
5785 case ALGORITHM_LEFT_SYMMETRIC:
5786 new_layout = ALGORITHM_LEFT_SYMMETRIC_6;
5788 case ALGORITHM_RIGHT_SYMMETRIC:
5789 new_layout = ALGORITHM_RIGHT_SYMMETRIC_6;
5791 case ALGORITHM_PARITY_0:
5792 new_layout = ALGORITHM_PARITY_0_6;
5794 case ALGORITHM_PARITY_N:
5795 new_layout = ALGORITHM_PARITY_N;
5798 return ERR_PTR(-EINVAL);
5800 mddev->new_level = 6;
5801 mddev->new_layout = new_layout;
5802 mddev->delta_disks = 1;
5803 mddev->raid_disks += 1;
5804 return setup_conf(mddev);
5808 static struct mdk_personality raid6_personality =
5812 .owner = THIS_MODULE,
5813 .make_request = make_request,
5817 .error_handler = error,
5818 .hot_add_disk = raid5_add_disk,
5819 .hot_remove_disk= raid5_remove_disk,
5820 .spare_active = raid5_spare_active,
5821 .sync_request = sync_request,
5822 .resize = raid5_resize,
5824 .check_reshape = raid6_check_reshape,
5825 .start_reshape = raid5_start_reshape,
5826 .finish_reshape = raid5_finish_reshape,
5827 .quiesce = raid5_quiesce,
5828 .takeover = raid6_takeover,
5830 static struct mdk_personality raid5_personality =
5834 .owner = THIS_MODULE,
5835 .make_request = make_request,
5839 .error_handler = error,
5840 .hot_add_disk = raid5_add_disk,
5841 .hot_remove_disk= raid5_remove_disk,
5842 .spare_active = raid5_spare_active,
5843 .sync_request = sync_request,
5844 .resize = raid5_resize,
5846 .check_reshape = raid5_check_reshape,
5847 .start_reshape = raid5_start_reshape,
5848 .finish_reshape = raid5_finish_reshape,
5849 .quiesce = raid5_quiesce,
5850 .takeover = raid5_takeover,
5853 static struct mdk_personality raid4_personality =
5857 .owner = THIS_MODULE,
5858 .make_request = make_request,
5862 .error_handler = error,
5863 .hot_add_disk = raid5_add_disk,
5864 .hot_remove_disk= raid5_remove_disk,
5865 .spare_active = raid5_spare_active,
5866 .sync_request = sync_request,
5867 .resize = raid5_resize,
5869 .check_reshape = raid5_check_reshape,
5870 .start_reshape = raid5_start_reshape,
5871 .finish_reshape = raid5_finish_reshape,
5872 .quiesce = raid5_quiesce,
5875 static int __init raid5_init(void)
5877 register_md_personality(&raid6_personality);
5878 register_md_personality(&raid5_personality);
5879 register_md_personality(&raid4_personality);
5883 static void raid5_exit(void)
5885 unregister_md_personality(&raid6_personality);
5886 unregister_md_personality(&raid5_personality);
5887 unregister_md_personality(&raid4_personality);
5890 module_init(raid5_init);
5891 module_exit(raid5_exit);
5892 MODULE_LICENSE("GPL");
5893 MODULE_DESCRIPTION("RAID4/5/6 (striping with parity) personality for MD");
5894 MODULE_ALIAS("md-personality-4"); /* RAID5 */
5895 MODULE_ALIAS("md-raid5");
5896 MODULE_ALIAS("md-raid4");
5897 MODULE_ALIAS("md-level-5");
5898 MODULE_ALIAS("md-level-4");
5899 MODULE_ALIAS("md-personality-8"); /* RAID6 */
5900 MODULE_ALIAS("md-raid6");
5901 MODULE_ALIAS("md-level-6");
5903 /* This used to be two separate modules, they were: */
5904 MODULE_ALIAS("raid5");
5905 MODULE_ALIAS("raid6");