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/module.h>
47 #include <linux/slab.h>
48 #include <linux/highmem.h>
49 #include <linux/bitops.h>
50 #include <linux/kthread.h>
51 #include <asm/atomic.h>
54 #include <linux/raid/bitmap.h>
55 #include <linux/async_tx.h>
61 #define NR_STRIPES 256
62 #define STRIPE_SIZE PAGE_SIZE
63 #define STRIPE_SHIFT (PAGE_SHIFT - 9)
64 #define STRIPE_SECTORS (STRIPE_SIZE>>9)
65 #define IO_THRESHOLD 1
66 #define BYPASS_THRESHOLD 1
67 #define NR_HASH (PAGE_SIZE / sizeof(struct hlist_head))
68 #define HASH_MASK (NR_HASH - 1)
70 #define stripe_hash(conf, sect) (&((conf)->stripe_hashtbl[((sect) >> STRIPE_SHIFT) & HASH_MASK]))
72 /* bio's attached to a stripe+device for I/O are linked together in bi_sector
73 * order without overlap. There may be several bio's per stripe+device, and
74 * a bio could span several devices.
75 * When walking this list for a particular stripe+device, we must never proceed
76 * beyond a bio that extends past this device, as the next bio might no longer
78 * This macro is used to determine the 'next' bio in the list, given the sector
79 * of the current stripe+device
81 #define r5_next_bio(bio, sect) ( ( (bio)->bi_sector + ((bio)->bi_size>>9) < sect + STRIPE_SECTORS) ? (bio)->bi_next : NULL)
83 * The following can be used to debug the driver
85 #define RAID5_PARANOIA 1
86 #if RAID5_PARANOIA && defined(CONFIG_SMP)
87 # define CHECK_DEVLOCK() assert_spin_locked(&conf->device_lock)
89 # define CHECK_DEVLOCK()
97 #define printk_rl(args...) ((void) (printk_ratelimit() && printk(args)))
99 #if !RAID6_USE_EMPTY_ZERO_PAGE
100 /* In .bss so it's zeroed */
101 const char raid6_empty_zero_page[PAGE_SIZE] __attribute__((aligned(256)));
104 static inline int raid6_next_disk(int disk, int raid_disks)
107 return (disk < raid_disks) ? disk : 0;
110 static void return_io(struct bio *return_bi)
112 struct bio *bi = return_bi;
115 return_bi = bi->bi_next;
123 static void print_raid5_conf (raid5_conf_t *conf);
125 static void __release_stripe(raid5_conf_t *conf, struct stripe_head *sh)
127 if (atomic_dec_and_test(&sh->count)) {
128 BUG_ON(!list_empty(&sh->lru));
129 BUG_ON(atomic_read(&conf->active_stripes)==0);
130 if (test_bit(STRIPE_HANDLE, &sh->state)) {
131 if (test_bit(STRIPE_DELAYED, &sh->state)) {
132 list_add_tail(&sh->lru, &conf->delayed_list);
133 blk_plug_device(conf->mddev->queue);
134 } else if (test_bit(STRIPE_BIT_DELAY, &sh->state) &&
135 sh->bm_seq - conf->seq_write > 0) {
136 list_add_tail(&sh->lru, &conf->bitmap_list);
137 blk_plug_device(conf->mddev->queue);
139 clear_bit(STRIPE_BIT_DELAY, &sh->state);
140 list_add_tail(&sh->lru, &conf->handle_list);
142 md_wakeup_thread(conf->mddev->thread);
144 BUG_ON(sh->ops.pending);
145 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
146 atomic_dec(&conf->preread_active_stripes);
147 if (atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD)
148 md_wakeup_thread(conf->mddev->thread);
150 atomic_dec(&conf->active_stripes);
151 if (!test_bit(STRIPE_EXPANDING, &sh->state)) {
152 list_add_tail(&sh->lru, &conf->inactive_list);
153 wake_up(&conf->wait_for_stripe);
154 if (conf->retry_read_aligned)
155 md_wakeup_thread(conf->mddev->thread);
160 static void release_stripe(struct stripe_head *sh)
162 raid5_conf_t *conf = sh->raid_conf;
165 spin_lock_irqsave(&conf->device_lock, flags);
166 __release_stripe(conf, sh);
167 spin_unlock_irqrestore(&conf->device_lock, flags);
170 static inline void remove_hash(struct stripe_head *sh)
172 pr_debug("remove_hash(), stripe %llu\n",
173 (unsigned long long)sh->sector);
175 hlist_del_init(&sh->hash);
178 static inline void insert_hash(raid5_conf_t *conf, struct stripe_head *sh)
180 struct hlist_head *hp = stripe_hash(conf, sh->sector);
182 pr_debug("insert_hash(), stripe %llu\n",
183 (unsigned long long)sh->sector);
186 hlist_add_head(&sh->hash, hp);
190 /* find an idle stripe, make sure it is unhashed, and return it. */
191 static struct stripe_head *get_free_stripe(raid5_conf_t *conf)
193 struct stripe_head *sh = NULL;
194 struct list_head *first;
197 if (list_empty(&conf->inactive_list))
199 first = conf->inactive_list.next;
200 sh = list_entry(first, struct stripe_head, lru);
201 list_del_init(first);
203 atomic_inc(&conf->active_stripes);
208 static void shrink_buffers(struct stripe_head *sh, int num)
213 for (i=0; i<num ; i++) {
217 sh->dev[i].page = NULL;
222 static int grow_buffers(struct stripe_head *sh, int num)
226 for (i=0; i<num; i++) {
229 if (!(page = alloc_page(GFP_KERNEL))) {
232 sh->dev[i].page = page;
237 static void raid5_build_block (struct stripe_head *sh, int i);
239 static void init_stripe(struct stripe_head *sh, sector_t sector, int pd_idx, int disks)
241 raid5_conf_t *conf = sh->raid_conf;
244 BUG_ON(atomic_read(&sh->count) != 0);
245 BUG_ON(test_bit(STRIPE_HANDLE, &sh->state));
246 BUG_ON(sh->ops.pending || sh->ops.ack || sh->ops.complete);
249 pr_debug("init_stripe called, stripe %llu\n",
250 (unsigned long long)sh->sector);
260 for (i = sh->disks; i--; ) {
261 struct r5dev *dev = &sh->dev[i];
263 if (dev->toread || dev->read || dev->towrite || dev->written ||
264 test_bit(R5_LOCKED, &dev->flags)) {
265 printk(KERN_ERR "sector=%llx i=%d %p %p %p %p %d\n",
266 (unsigned long long)sh->sector, i, dev->toread,
267 dev->read, dev->towrite, dev->written,
268 test_bit(R5_LOCKED, &dev->flags));
272 raid5_build_block(sh, i);
274 insert_hash(conf, sh);
277 static struct stripe_head *__find_stripe(raid5_conf_t *conf, sector_t sector, int disks)
279 struct stripe_head *sh;
280 struct hlist_node *hn;
283 pr_debug("__find_stripe, sector %llu\n", (unsigned long long)sector);
284 hlist_for_each_entry(sh, hn, stripe_hash(conf, sector), hash)
285 if (sh->sector == sector && sh->disks == disks)
287 pr_debug("__stripe %llu not in cache\n", (unsigned long long)sector);
291 static void unplug_slaves(mddev_t *mddev);
292 static void raid5_unplug_device(struct request_queue *q);
294 static struct stripe_head *get_active_stripe(raid5_conf_t *conf, sector_t sector, int disks,
295 int pd_idx, int noblock)
297 struct stripe_head *sh;
299 pr_debug("get_stripe, sector %llu\n", (unsigned long long)sector);
301 spin_lock_irq(&conf->device_lock);
304 wait_event_lock_irq(conf->wait_for_stripe,
306 conf->device_lock, /* nothing */);
307 sh = __find_stripe(conf, sector, disks);
309 if (!conf->inactive_blocked)
310 sh = get_free_stripe(conf);
311 if (noblock && sh == NULL)
314 conf->inactive_blocked = 1;
315 wait_event_lock_irq(conf->wait_for_stripe,
316 !list_empty(&conf->inactive_list) &&
317 (atomic_read(&conf->active_stripes)
318 < (conf->max_nr_stripes *3/4)
319 || !conf->inactive_blocked),
321 raid5_unplug_device(conf->mddev->queue)
323 conf->inactive_blocked = 0;
325 init_stripe(sh, sector, pd_idx, disks);
327 if (atomic_read(&sh->count)) {
328 BUG_ON(!list_empty(&sh->lru));
330 if (!test_bit(STRIPE_HANDLE, &sh->state))
331 atomic_inc(&conf->active_stripes);
332 if (list_empty(&sh->lru) &&
333 !test_bit(STRIPE_EXPANDING, &sh->state))
335 list_del_init(&sh->lru);
338 } while (sh == NULL);
341 atomic_inc(&sh->count);
343 spin_unlock_irq(&conf->device_lock);
347 /* test_and_ack_op() ensures that we only dequeue an operation once */
348 #define test_and_ack_op(op, pend) \
350 if (test_bit(op, &sh->ops.pending) && \
351 !test_bit(op, &sh->ops.complete)) { \
352 if (test_and_set_bit(op, &sh->ops.ack)) \
353 clear_bit(op, &pend); \
357 clear_bit(op, &pend); \
360 /* find new work to run, do not resubmit work that is already
363 static unsigned long get_stripe_work(struct stripe_head *sh)
365 unsigned long pending;
368 pending = sh->ops.pending;
370 test_and_ack_op(STRIPE_OP_BIOFILL, pending);
371 test_and_ack_op(STRIPE_OP_COMPUTE_BLK, pending);
372 test_and_ack_op(STRIPE_OP_PREXOR, pending);
373 test_and_ack_op(STRIPE_OP_BIODRAIN, pending);
374 test_and_ack_op(STRIPE_OP_POSTXOR, pending);
375 test_and_ack_op(STRIPE_OP_CHECK, pending);
377 sh->ops.count -= ack;
378 if (unlikely(sh->ops.count < 0)) {
379 printk(KERN_ERR "pending: %#lx ops.pending: %#lx ops.ack: %#lx "
380 "ops.complete: %#lx\n", pending, sh->ops.pending,
381 sh->ops.ack, sh->ops.complete);
389 raid5_end_read_request(struct bio *bi, int error);
391 raid5_end_write_request(struct bio *bi, int error);
393 static void ops_run_io(struct stripe_head *sh, struct stripe_head_state *s)
395 raid5_conf_t *conf = sh->raid_conf;
396 int i, disks = sh->disks;
400 for (i = disks; i--; ) {
404 if (test_and_clear_bit(R5_Wantwrite, &sh->dev[i].flags))
406 else if (test_and_clear_bit(R5_Wantread, &sh->dev[i].flags))
411 bi = &sh->dev[i].req;
415 bi->bi_end_io = raid5_end_write_request;
417 bi->bi_end_io = raid5_end_read_request;
420 rdev = rcu_dereference(conf->disks[i].rdev);
421 if (rdev && test_bit(Faulty, &rdev->flags))
424 atomic_inc(&rdev->nr_pending);
428 if (s->syncing || s->expanding || s->expanded)
429 md_sync_acct(rdev->bdev, STRIPE_SECTORS);
431 set_bit(STRIPE_IO_STARTED, &sh->state);
433 bi->bi_bdev = rdev->bdev;
434 pr_debug("%s: for %llu schedule op %ld on disc %d\n",
435 __func__, (unsigned long long)sh->sector,
437 atomic_inc(&sh->count);
438 bi->bi_sector = sh->sector + rdev->data_offset;
439 bi->bi_flags = 1 << BIO_UPTODATE;
443 bi->bi_io_vec = &sh->dev[i].vec;
444 bi->bi_io_vec[0].bv_len = STRIPE_SIZE;
445 bi->bi_io_vec[0].bv_offset = 0;
446 bi->bi_size = STRIPE_SIZE;
449 test_bit(R5_ReWrite, &sh->dev[i].flags))
450 atomic_add(STRIPE_SECTORS,
451 &rdev->corrected_errors);
452 generic_make_request(bi);
455 set_bit(STRIPE_DEGRADED, &sh->state);
456 pr_debug("skip op %ld on disc %d for sector %llu\n",
457 bi->bi_rw, i, (unsigned long long)sh->sector);
458 clear_bit(R5_LOCKED, &sh->dev[i].flags);
459 set_bit(STRIPE_HANDLE, &sh->state);
464 static struct dma_async_tx_descriptor *
465 async_copy_data(int frombio, struct bio *bio, struct page *page,
466 sector_t sector, struct dma_async_tx_descriptor *tx)
469 struct page *bio_page;
473 if (bio->bi_sector >= sector)
474 page_offset = (signed)(bio->bi_sector - sector) * 512;
476 page_offset = (signed)(sector - bio->bi_sector) * -512;
477 bio_for_each_segment(bvl, bio, i) {
478 int len = bio_iovec_idx(bio, i)->bv_len;
482 if (page_offset < 0) {
483 b_offset = -page_offset;
484 page_offset += b_offset;
488 if (len > 0 && page_offset + len > STRIPE_SIZE)
489 clen = STRIPE_SIZE - page_offset;
494 b_offset += bio_iovec_idx(bio, i)->bv_offset;
495 bio_page = bio_iovec_idx(bio, i)->bv_page;
497 tx = async_memcpy(page, bio_page, page_offset,
502 tx = async_memcpy(bio_page, page, b_offset,
507 if (clen < len) /* hit end of page */
515 static void ops_complete_biofill(void *stripe_head_ref)
517 struct stripe_head *sh = stripe_head_ref;
518 struct bio *return_bi = NULL;
519 raid5_conf_t *conf = sh->raid_conf;
522 pr_debug("%s: stripe %llu\n", __func__,
523 (unsigned long long)sh->sector);
525 /* clear completed biofills */
526 spin_lock_irq(&conf->device_lock);
527 for (i = sh->disks; i--; ) {
528 struct r5dev *dev = &sh->dev[i];
530 /* acknowledge completion of a biofill operation */
531 /* and check if we need to reply to a read request,
532 * new R5_Wantfill requests are held off until
533 * !STRIPE_BIOFILL_RUN
535 if (test_and_clear_bit(R5_Wantfill, &dev->flags)) {
536 struct bio *rbi, *rbi2;
541 while (rbi && rbi->bi_sector <
542 dev->sector + STRIPE_SECTORS) {
543 rbi2 = r5_next_bio(rbi, dev->sector);
544 if (--rbi->bi_phys_segments == 0) {
545 rbi->bi_next = return_bi;
552 spin_unlock_irq(&conf->device_lock);
553 clear_bit(STRIPE_BIOFILL_RUN, &sh->state);
555 return_io(return_bi);
557 set_bit(STRIPE_HANDLE, &sh->state);
561 static void ops_run_biofill(struct stripe_head *sh)
563 struct dma_async_tx_descriptor *tx = NULL;
564 raid5_conf_t *conf = sh->raid_conf;
567 pr_debug("%s: stripe %llu\n", __func__,
568 (unsigned long long)sh->sector);
570 for (i = sh->disks; i--; ) {
571 struct r5dev *dev = &sh->dev[i];
572 if (test_bit(R5_Wantfill, &dev->flags)) {
574 spin_lock_irq(&conf->device_lock);
575 dev->read = rbi = dev->toread;
577 spin_unlock_irq(&conf->device_lock);
578 while (rbi && rbi->bi_sector <
579 dev->sector + STRIPE_SECTORS) {
580 tx = async_copy_data(0, rbi, dev->page,
582 rbi = r5_next_bio(rbi, dev->sector);
587 atomic_inc(&sh->count);
588 async_trigger_callback(ASYNC_TX_DEP_ACK | ASYNC_TX_ACK, tx,
589 ops_complete_biofill, sh);
592 static void ops_complete_compute5(void *stripe_head_ref)
594 struct stripe_head *sh = stripe_head_ref;
595 int target = sh->ops.target;
596 struct r5dev *tgt = &sh->dev[target];
598 pr_debug("%s: stripe %llu\n", __func__,
599 (unsigned long long)sh->sector);
601 set_bit(R5_UPTODATE, &tgt->flags);
602 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
603 clear_bit(R5_Wantcompute, &tgt->flags);
604 clear_bit(STRIPE_COMPUTE_RUN, &sh->state);
605 if (sh->check_state == check_state_compute_run)
606 sh->check_state = check_state_compute_result;
607 set_bit(STRIPE_HANDLE, &sh->state);
611 static struct dma_async_tx_descriptor *
612 ops_run_compute5(struct stripe_head *sh, unsigned long pending)
614 /* kernel stack size limits the total number of disks */
615 int disks = sh->disks;
616 struct page *xor_srcs[disks];
617 int target = sh->ops.target;
618 struct r5dev *tgt = &sh->dev[target];
619 struct page *xor_dest = tgt->page;
621 struct dma_async_tx_descriptor *tx;
624 pr_debug("%s: stripe %llu block: %d\n",
625 __func__, (unsigned long long)sh->sector, target);
626 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
628 for (i = disks; i--; )
630 xor_srcs[count++] = sh->dev[i].page;
632 atomic_inc(&sh->count);
634 if (unlikely(count == 1))
635 tx = async_memcpy(xor_dest, xor_srcs[0], 0, 0, STRIPE_SIZE,
636 0, NULL, ops_complete_compute5, sh);
638 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE,
639 ASYNC_TX_XOR_ZERO_DST, NULL,
640 ops_complete_compute5, sh);
642 /* ack now if postxor is not set to be run */
643 if (tx && !test_bit(STRIPE_OP_POSTXOR, &pending))
649 static void ops_complete_prexor(void *stripe_head_ref)
651 struct stripe_head *sh = stripe_head_ref;
653 pr_debug("%s: stripe %llu\n", __func__,
654 (unsigned long long)sh->sector);
656 set_bit(STRIPE_OP_PREXOR, &sh->ops.complete);
659 static struct dma_async_tx_descriptor *
660 ops_run_prexor(struct stripe_head *sh, struct dma_async_tx_descriptor *tx)
662 /* kernel stack size limits the total number of disks */
663 int disks = sh->disks;
664 struct page *xor_srcs[disks];
665 int count = 0, pd_idx = sh->pd_idx, i;
667 /* existing parity data subtracted */
668 struct page *xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
670 pr_debug("%s: stripe %llu\n", __func__,
671 (unsigned long long)sh->sector);
673 for (i = disks; i--; ) {
674 struct r5dev *dev = &sh->dev[i];
675 /* Only process blocks that are known to be uptodate */
676 if (dev->towrite && test_bit(R5_Wantprexor, &dev->flags))
677 xor_srcs[count++] = dev->page;
680 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE,
681 ASYNC_TX_DEP_ACK | ASYNC_TX_XOR_DROP_DST, tx,
682 ops_complete_prexor, sh);
687 static struct dma_async_tx_descriptor *
688 ops_run_biodrain(struct stripe_head *sh, struct dma_async_tx_descriptor *tx,
689 unsigned long pending)
691 int disks = sh->disks;
692 int pd_idx = sh->pd_idx, i;
694 /* check if prexor is active which means only process blocks
695 * that are part of a read-modify-write (Wantprexor)
697 int prexor = test_bit(STRIPE_OP_PREXOR, &pending);
699 pr_debug("%s: stripe %llu\n", __func__,
700 (unsigned long long)sh->sector);
702 for (i = disks; i--; ) {
703 struct r5dev *dev = &sh->dev[i];
708 if (prexor) { /* rmw */
710 test_bit(R5_Wantprexor, &dev->flags))
713 if (i != pd_idx && dev->towrite &&
714 test_bit(R5_LOCKED, &dev->flags))
721 spin_lock(&sh->lock);
722 chosen = dev->towrite;
724 BUG_ON(dev->written);
725 wbi = dev->written = chosen;
726 spin_unlock(&sh->lock);
728 while (wbi && wbi->bi_sector <
729 dev->sector + STRIPE_SECTORS) {
730 tx = async_copy_data(1, wbi, dev->page,
732 wbi = r5_next_bio(wbi, dev->sector);
740 static void ops_complete_postxor(void *stripe_head_ref)
742 struct stripe_head *sh = stripe_head_ref;
744 pr_debug("%s: stripe %llu\n", __func__,
745 (unsigned long long)sh->sector);
747 set_bit(STRIPE_OP_POSTXOR, &sh->ops.complete);
748 set_bit(STRIPE_HANDLE, &sh->state);
752 static void ops_complete_write(void *stripe_head_ref)
754 struct stripe_head *sh = stripe_head_ref;
755 int disks = sh->disks, i, pd_idx = sh->pd_idx;
757 pr_debug("%s: stripe %llu\n", __func__,
758 (unsigned long long)sh->sector);
760 for (i = disks; i--; ) {
761 struct r5dev *dev = &sh->dev[i];
762 if (dev->written || i == pd_idx)
763 set_bit(R5_UPTODATE, &dev->flags);
766 set_bit(STRIPE_OP_BIODRAIN, &sh->ops.complete);
767 set_bit(STRIPE_OP_POSTXOR, &sh->ops.complete);
769 set_bit(STRIPE_HANDLE, &sh->state);
774 ops_run_postxor(struct stripe_head *sh, struct dma_async_tx_descriptor *tx,
775 unsigned long pending)
777 /* kernel stack size limits the total number of disks */
778 int disks = sh->disks;
779 struct page *xor_srcs[disks];
781 int count = 0, pd_idx = sh->pd_idx, i;
782 struct page *xor_dest;
783 int prexor = test_bit(STRIPE_OP_PREXOR, &pending);
785 dma_async_tx_callback callback;
787 pr_debug("%s: stripe %llu\n", __func__,
788 (unsigned long long)sh->sector);
790 /* check if prexor is active which means only process blocks
791 * that are part of a read-modify-write (written)
794 xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
795 for (i = disks; i--; ) {
796 struct r5dev *dev = &sh->dev[i];
798 xor_srcs[count++] = dev->page;
801 xor_dest = sh->dev[pd_idx].page;
802 for (i = disks; i--; ) {
803 struct r5dev *dev = &sh->dev[i];
805 xor_srcs[count++] = dev->page;
809 /* check whether this postxor is part of a write */
810 callback = test_bit(STRIPE_OP_BIODRAIN, &pending) ?
811 ops_complete_write : ops_complete_postxor;
813 /* 1/ if we prexor'd then the dest is reused as a source
814 * 2/ if we did not prexor then we are redoing the parity
815 * set ASYNC_TX_XOR_DROP_DST and ASYNC_TX_XOR_ZERO_DST
816 * for the synchronous xor case
818 flags = ASYNC_TX_DEP_ACK | ASYNC_TX_ACK |
819 (prexor ? ASYNC_TX_XOR_DROP_DST : ASYNC_TX_XOR_ZERO_DST);
821 atomic_inc(&sh->count);
823 if (unlikely(count == 1)) {
824 flags &= ~(ASYNC_TX_XOR_DROP_DST | ASYNC_TX_XOR_ZERO_DST);
825 tx = async_memcpy(xor_dest, xor_srcs[0], 0, 0, STRIPE_SIZE,
826 flags, tx, callback, sh);
828 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE,
829 flags, tx, callback, sh);
832 static void ops_complete_check(void *stripe_head_ref)
834 struct stripe_head *sh = stripe_head_ref;
836 pr_debug("%s: stripe %llu\n", __func__,
837 (unsigned long long)sh->sector);
839 sh->check_state = check_state_check_result;
840 set_bit(STRIPE_HANDLE, &sh->state);
844 static void ops_run_check(struct stripe_head *sh)
846 /* kernel stack size limits the total number of disks */
847 int disks = sh->disks;
848 struct page *xor_srcs[disks];
849 struct dma_async_tx_descriptor *tx;
851 int count = 0, pd_idx = sh->pd_idx, i;
852 struct page *xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
854 pr_debug("%s: stripe %llu\n", __func__,
855 (unsigned long long)sh->sector);
857 for (i = disks; i--; ) {
858 struct r5dev *dev = &sh->dev[i];
860 xor_srcs[count++] = dev->page;
863 tx = async_xor_zero_sum(xor_dest, xor_srcs, 0, count, STRIPE_SIZE,
864 &sh->ops.zero_sum_result, 0, NULL, NULL, NULL);
866 atomic_inc(&sh->count);
867 tx = async_trigger_callback(ASYNC_TX_DEP_ACK | ASYNC_TX_ACK, tx,
868 ops_complete_check, sh);
871 static void raid5_run_ops(struct stripe_head *sh, unsigned long pending,
872 unsigned long ops_request)
874 int overlap_clear = 0, i, disks = sh->disks;
875 struct dma_async_tx_descriptor *tx = NULL;
877 if (test_bit(STRIPE_OP_BIOFILL, &ops_request)) {
882 if (test_bit(STRIPE_OP_COMPUTE_BLK, &ops_request))
883 tx = ops_run_compute5(sh, pending);
885 if (test_bit(STRIPE_OP_PREXOR, &pending))
886 tx = ops_run_prexor(sh, tx);
888 if (test_bit(STRIPE_OP_BIODRAIN, &pending)) {
889 tx = ops_run_biodrain(sh, tx, pending);
893 if (test_bit(STRIPE_OP_POSTXOR, &pending))
894 ops_run_postxor(sh, tx, pending);
896 if (test_bit(STRIPE_OP_CHECK, &ops_request))
900 for (i = disks; i--; ) {
901 struct r5dev *dev = &sh->dev[i];
902 if (test_and_clear_bit(R5_Overlap, &dev->flags))
903 wake_up(&sh->raid_conf->wait_for_overlap);
907 static int grow_one_stripe(raid5_conf_t *conf)
909 struct stripe_head *sh;
910 sh = kmem_cache_alloc(conf->slab_cache, GFP_KERNEL);
913 memset(sh, 0, sizeof(*sh) + (conf->raid_disks-1)*sizeof(struct r5dev));
914 sh->raid_conf = conf;
915 spin_lock_init(&sh->lock);
917 if (grow_buffers(sh, conf->raid_disks)) {
918 shrink_buffers(sh, conf->raid_disks);
919 kmem_cache_free(conf->slab_cache, sh);
922 sh->disks = conf->raid_disks;
923 /* we just created an active stripe so... */
924 atomic_set(&sh->count, 1);
925 atomic_inc(&conf->active_stripes);
926 INIT_LIST_HEAD(&sh->lru);
931 static int grow_stripes(raid5_conf_t *conf, int num)
933 struct kmem_cache *sc;
934 int devs = conf->raid_disks;
936 sprintf(conf->cache_name[0], "raid5-%s", mdname(conf->mddev));
937 sprintf(conf->cache_name[1], "raid5-%s-alt", mdname(conf->mddev));
938 conf->active_name = 0;
939 sc = kmem_cache_create(conf->cache_name[conf->active_name],
940 sizeof(struct stripe_head)+(devs-1)*sizeof(struct r5dev),
944 conf->slab_cache = sc;
945 conf->pool_size = devs;
947 if (!grow_one_stripe(conf))
952 #ifdef CONFIG_MD_RAID5_RESHAPE
953 static int resize_stripes(raid5_conf_t *conf, int newsize)
955 /* Make all the stripes able to hold 'newsize' devices.
956 * New slots in each stripe get 'page' set to a new page.
958 * This happens in stages:
959 * 1/ create a new kmem_cache and allocate the required number of
961 * 2/ gather all the old stripe_heads and tranfer the pages across
962 * to the new stripe_heads. This will have the side effect of
963 * freezing the array as once all stripe_heads have been collected,
964 * no IO will be possible. Old stripe heads are freed once their
965 * pages have been transferred over, and the old kmem_cache is
966 * freed when all stripes are done.
967 * 3/ reallocate conf->disks to be suitable bigger. If this fails,
968 * we simple return a failre status - no need to clean anything up.
969 * 4/ allocate new pages for the new slots in the new stripe_heads.
970 * If this fails, we don't bother trying the shrink the
971 * stripe_heads down again, we just leave them as they are.
972 * As each stripe_head is processed the new one is released into
975 * Once step2 is started, we cannot afford to wait for a write,
976 * so we use GFP_NOIO allocations.
978 struct stripe_head *osh, *nsh;
979 LIST_HEAD(newstripes);
980 struct disk_info *ndisks;
982 struct kmem_cache *sc;
985 if (newsize <= conf->pool_size)
986 return 0; /* never bother to shrink */
988 md_allow_write(conf->mddev);
991 sc = kmem_cache_create(conf->cache_name[1-conf->active_name],
992 sizeof(struct stripe_head)+(newsize-1)*sizeof(struct r5dev),
997 for (i = conf->max_nr_stripes; i; i--) {
998 nsh = kmem_cache_alloc(sc, GFP_KERNEL);
1002 memset(nsh, 0, sizeof(*nsh) + (newsize-1)*sizeof(struct r5dev));
1004 nsh->raid_conf = conf;
1005 spin_lock_init(&nsh->lock);
1007 list_add(&nsh->lru, &newstripes);
1010 /* didn't get enough, give up */
1011 while (!list_empty(&newstripes)) {
1012 nsh = list_entry(newstripes.next, struct stripe_head, lru);
1013 list_del(&nsh->lru);
1014 kmem_cache_free(sc, nsh);
1016 kmem_cache_destroy(sc);
1019 /* Step 2 - Must use GFP_NOIO now.
1020 * OK, we have enough stripes, start collecting inactive
1021 * stripes and copying them over
1023 list_for_each_entry(nsh, &newstripes, lru) {
1024 spin_lock_irq(&conf->device_lock);
1025 wait_event_lock_irq(conf->wait_for_stripe,
1026 !list_empty(&conf->inactive_list),
1028 unplug_slaves(conf->mddev)
1030 osh = get_free_stripe(conf);
1031 spin_unlock_irq(&conf->device_lock);
1032 atomic_set(&nsh->count, 1);
1033 for(i=0; i<conf->pool_size; i++)
1034 nsh->dev[i].page = osh->dev[i].page;
1035 for( ; i<newsize; i++)
1036 nsh->dev[i].page = NULL;
1037 kmem_cache_free(conf->slab_cache, osh);
1039 kmem_cache_destroy(conf->slab_cache);
1042 * At this point, we are holding all the stripes so the array
1043 * is completely stalled, so now is a good time to resize
1046 ndisks = kzalloc(newsize * sizeof(struct disk_info), GFP_NOIO);
1048 for (i=0; i<conf->raid_disks; i++)
1049 ndisks[i] = conf->disks[i];
1051 conf->disks = ndisks;
1055 /* Step 4, return new stripes to service */
1056 while(!list_empty(&newstripes)) {
1057 nsh = list_entry(newstripes.next, struct stripe_head, lru);
1058 list_del_init(&nsh->lru);
1059 for (i=conf->raid_disks; i < newsize; i++)
1060 if (nsh->dev[i].page == NULL) {
1061 struct page *p = alloc_page(GFP_NOIO);
1062 nsh->dev[i].page = p;
1066 release_stripe(nsh);
1068 /* critical section pass, GFP_NOIO no longer needed */
1070 conf->slab_cache = sc;
1071 conf->active_name = 1-conf->active_name;
1072 conf->pool_size = newsize;
1077 static int drop_one_stripe(raid5_conf_t *conf)
1079 struct stripe_head *sh;
1081 spin_lock_irq(&conf->device_lock);
1082 sh = get_free_stripe(conf);
1083 spin_unlock_irq(&conf->device_lock);
1086 BUG_ON(atomic_read(&sh->count));
1087 shrink_buffers(sh, conf->pool_size);
1088 kmem_cache_free(conf->slab_cache, sh);
1089 atomic_dec(&conf->active_stripes);
1093 static void shrink_stripes(raid5_conf_t *conf)
1095 while (drop_one_stripe(conf))
1098 if (conf->slab_cache)
1099 kmem_cache_destroy(conf->slab_cache);
1100 conf->slab_cache = NULL;
1103 static void raid5_end_read_request(struct bio * bi, int error)
1105 struct stripe_head *sh = bi->bi_private;
1106 raid5_conf_t *conf = sh->raid_conf;
1107 int disks = sh->disks, i;
1108 int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
1109 char b[BDEVNAME_SIZE];
1113 for (i=0 ; i<disks; i++)
1114 if (bi == &sh->dev[i].req)
1117 pr_debug("end_read_request %llu/%d, count: %d, uptodate %d.\n",
1118 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
1126 set_bit(R5_UPTODATE, &sh->dev[i].flags);
1127 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
1128 rdev = conf->disks[i].rdev;
1129 printk_rl(KERN_INFO "raid5:%s: read error corrected"
1130 " (%lu sectors at %llu on %s)\n",
1131 mdname(conf->mddev), STRIPE_SECTORS,
1132 (unsigned long long)(sh->sector
1133 + rdev->data_offset),
1134 bdevname(rdev->bdev, b));
1135 clear_bit(R5_ReadError, &sh->dev[i].flags);
1136 clear_bit(R5_ReWrite, &sh->dev[i].flags);
1138 if (atomic_read(&conf->disks[i].rdev->read_errors))
1139 atomic_set(&conf->disks[i].rdev->read_errors, 0);
1141 const char *bdn = bdevname(conf->disks[i].rdev->bdev, b);
1143 rdev = conf->disks[i].rdev;
1145 clear_bit(R5_UPTODATE, &sh->dev[i].flags);
1146 atomic_inc(&rdev->read_errors);
1147 if (conf->mddev->degraded)
1148 printk_rl(KERN_WARNING
1149 "raid5:%s: read error not correctable "
1150 "(sector %llu on %s).\n",
1151 mdname(conf->mddev),
1152 (unsigned long long)(sh->sector
1153 + rdev->data_offset),
1155 else if (test_bit(R5_ReWrite, &sh->dev[i].flags))
1157 printk_rl(KERN_WARNING
1158 "raid5:%s: read error NOT corrected!! "
1159 "(sector %llu on %s).\n",
1160 mdname(conf->mddev),
1161 (unsigned long long)(sh->sector
1162 + rdev->data_offset),
1164 else if (atomic_read(&rdev->read_errors)
1165 > conf->max_nr_stripes)
1167 "raid5:%s: Too many read errors, failing device %s.\n",
1168 mdname(conf->mddev), bdn);
1172 set_bit(R5_ReadError, &sh->dev[i].flags);
1174 clear_bit(R5_ReadError, &sh->dev[i].flags);
1175 clear_bit(R5_ReWrite, &sh->dev[i].flags);
1176 md_error(conf->mddev, rdev);
1179 rdev_dec_pending(conf->disks[i].rdev, conf->mddev);
1180 clear_bit(R5_LOCKED, &sh->dev[i].flags);
1181 set_bit(STRIPE_HANDLE, &sh->state);
1185 static void raid5_end_write_request (struct bio *bi, int error)
1187 struct stripe_head *sh = bi->bi_private;
1188 raid5_conf_t *conf = sh->raid_conf;
1189 int disks = sh->disks, i;
1190 int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
1192 for (i=0 ; i<disks; i++)
1193 if (bi == &sh->dev[i].req)
1196 pr_debug("end_write_request %llu/%d, count %d, uptodate: %d.\n",
1197 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
1205 md_error(conf->mddev, conf->disks[i].rdev);
1207 rdev_dec_pending(conf->disks[i].rdev, conf->mddev);
1209 clear_bit(R5_LOCKED, &sh->dev[i].flags);
1210 set_bit(STRIPE_HANDLE, &sh->state);
1215 static sector_t compute_blocknr(struct stripe_head *sh, int i);
1217 static void raid5_build_block (struct stripe_head *sh, int i)
1219 struct r5dev *dev = &sh->dev[i];
1221 bio_init(&dev->req);
1222 dev->req.bi_io_vec = &dev->vec;
1224 dev->req.bi_max_vecs++;
1225 dev->vec.bv_page = dev->page;
1226 dev->vec.bv_len = STRIPE_SIZE;
1227 dev->vec.bv_offset = 0;
1229 dev->req.bi_sector = sh->sector;
1230 dev->req.bi_private = sh;
1233 dev->sector = compute_blocknr(sh, i);
1236 static void error(mddev_t *mddev, mdk_rdev_t *rdev)
1238 char b[BDEVNAME_SIZE];
1239 raid5_conf_t *conf = (raid5_conf_t *) mddev->private;
1240 pr_debug("raid5: error called\n");
1242 if (!test_bit(Faulty, &rdev->flags)) {
1243 set_bit(MD_CHANGE_DEVS, &mddev->flags);
1244 if (test_and_clear_bit(In_sync, &rdev->flags)) {
1245 unsigned long flags;
1246 spin_lock_irqsave(&conf->device_lock, flags);
1248 spin_unlock_irqrestore(&conf->device_lock, flags);
1250 * if recovery was running, make sure it aborts.
1252 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1254 set_bit(Faulty, &rdev->flags);
1256 "raid5: Disk failure on %s, disabling device.\n"
1257 "raid5: Operation continuing on %d devices.\n",
1258 bdevname(rdev->bdev,b), conf->raid_disks - mddev->degraded);
1263 * Input: a 'big' sector number,
1264 * Output: index of the data and parity disk, and the sector # in them.
1266 static sector_t raid5_compute_sector(sector_t r_sector, unsigned int raid_disks,
1267 unsigned int data_disks, unsigned int * dd_idx,
1268 unsigned int * pd_idx, raid5_conf_t *conf)
1271 unsigned long chunk_number;
1272 unsigned int chunk_offset;
1273 sector_t new_sector;
1274 int sectors_per_chunk = conf->chunk_size >> 9;
1276 /* First compute the information on this sector */
1279 * Compute the chunk number and the sector offset inside the chunk
1281 chunk_offset = sector_div(r_sector, sectors_per_chunk);
1282 chunk_number = r_sector;
1283 BUG_ON(r_sector != chunk_number);
1286 * Compute the stripe number
1288 stripe = chunk_number / data_disks;
1291 * Compute the data disk and parity disk indexes inside the stripe
1293 *dd_idx = chunk_number % data_disks;
1296 * Select the parity disk based on the user selected algorithm.
1298 switch(conf->level) {
1300 *pd_idx = data_disks;
1303 switch (conf->algorithm) {
1304 case ALGORITHM_LEFT_ASYMMETRIC:
1305 *pd_idx = data_disks - stripe % raid_disks;
1306 if (*dd_idx >= *pd_idx)
1309 case ALGORITHM_RIGHT_ASYMMETRIC:
1310 *pd_idx = stripe % raid_disks;
1311 if (*dd_idx >= *pd_idx)
1314 case ALGORITHM_LEFT_SYMMETRIC:
1315 *pd_idx = data_disks - stripe % raid_disks;
1316 *dd_idx = (*pd_idx + 1 + *dd_idx) % raid_disks;
1318 case ALGORITHM_RIGHT_SYMMETRIC:
1319 *pd_idx = stripe % raid_disks;
1320 *dd_idx = (*pd_idx + 1 + *dd_idx) % raid_disks;
1323 printk(KERN_ERR "raid5: unsupported algorithm %d\n",
1329 /**** FIX THIS ****/
1330 switch (conf->algorithm) {
1331 case ALGORITHM_LEFT_ASYMMETRIC:
1332 *pd_idx = raid_disks - 1 - (stripe % raid_disks);
1333 if (*pd_idx == raid_disks-1)
1334 (*dd_idx)++; /* Q D D D P */
1335 else if (*dd_idx >= *pd_idx)
1336 (*dd_idx) += 2; /* D D P Q D */
1338 case ALGORITHM_RIGHT_ASYMMETRIC:
1339 *pd_idx = stripe % raid_disks;
1340 if (*pd_idx == raid_disks-1)
1341 (*dd_idx)++; /* Q D D D P */
1342 else if (*dd_idx >= *pd_idx)
1343 (*dd_idx) += 2; /* D D P Q D */
1345 case ALGORITHM_LEFT_SYMMETRIC:
1346 *pd_idx = raid_disks - 1 - (stripe % raid_disks);
1347 *dd_idx = (*pd_idx + 2 + *dd_idx) % raid_disks;
1349 case ALGORITHM_RIGHT_SYMMETRIC:
1350 *pd_idx = stripe % raid_disks;
1351 *dd_idx = (*pd_idx + 2 + *dd_idx) % raid_disks;
1354 printk (KERN_CRIT "raid6: unsupported algorithm %d\n",
1361 * Finally, compute the new sector number
1363 new_sector = (sector_t)stripe * sectors_per_chunk + chunk_offset;
1368 static sector_t compute_blocknr(struct stripe_head *sh, int i)
1370 raid5_conf_t *conf = sh->raid_conf;
1371 int raid_disks = sh->disks;
1372 int data_disks = raid_disks - conf->max_degraded;
1373 sector_t new_sector = sh->sector, check;
1374 int sectors_per_chunk = conf->chunk_size >> 9;
1377 int chunk_number, dummy1, dummy2, dd_idx = i;
1381 chunk_offset = sector_div(new_sector, sectors_per_chunk);
1382 stripe = new_sector;
1383 BUG_ON(new_sector != stripe);
1385 if (i == sh->pd_idx)
1387 switch(conf->level) {
1390 switch (conf->algorithm) {
1391 case ALGORITHM_LEFT_ASYMMETRIC:
1392 case ALGORITHM_RIGHT_ASYMMETRIC:
1396 case ALGORITHM_LEFT_SYMMETRIC:
1397 case ALGORITHM_RIGHT_SYMMETRIC:
1400 i -= (sh->pd_idx + 1);
1403 printk(KERN_ERR "raid5: unsupported algorithm %d\n",
1408 if (i == raid6_next_disk(sh->pd_idx, raid_disks))
1409 return 0; /* It is the Q disk */
1410 switch (conf->algorithm) {
1411 case ALGORITHM_LEFT_ASYMMETRIC:
1412 case ALGORITHM_RIGHT_ASYMMETRIC:
1413 if (sh->pd_idx == raid_disks-1)
1414 i--; /* Q D D D P */
1415 else if (i > sh->pd_idx)
1416 i -= 2; /* D D P Q D */
1418 case ALGORITHM_LEFT_SYMMETRIC:
1419 case ALGORITHM_RIGHT_SYMMETRIC:
1420 if (sh->pd_idx == raid_disks-1)
1421 i--; /* Q D D D P */
1426 i -= (sh->pd_idx + 2);
1430 printk (KERN_CRIT "raid6: unsupported algorithm %d\n",
1436 chunk_number = stripe * data_disks + i;
1437 r_sector = (sector_t)chunk_number * sectors_per_chunk + chunk_offset;
1439 check = raid5_compute_sector (r_sector, raid_disks, data_disks, &dummy1, &dummy2, conf);
1440 if (check != sh->sector || dummy1 != dd_idx || dummy2 != sh->pd_idx) {
1441 printk(KERN_ERR "compute_blocknr: map not correct\n");
1450 * Copy data between a page in the stripe cache, and one or more bion
1451 * The page could align with the middle of the bio, or there could be
1452 * several bion, each with several bio_vecs, which cover part of the page
1453 * Multiple bion are linked together on bi_next. There may be extras
1454 * at the end of this list. We ignore them.
1456 static void copy_data(int frombio, struct bio *bio,
1460 char *pa = page_address(page);
1461 struct bio_vec *bvl;
1465 if (bio->bi_sector >= sector)
1466 page_offset = (signed)(bio->bi_sector - sector) * 512;
1468 page_offset = (signed)(sector - bio->bi_sector) * -512;
1469 bio_for_each_segment(bvl, bio, i) {
1470 int len = bio_iovec_idx(bio,i)->bv_len;
1474 if (page_offset < 0) {
1475 b_offset = -page_offset;
1476 page_offset += b_offset;
1480 if (len > 0 && page_offset + len > STRIPE_SIZE)
1481 clen = STRIPE_SIZE - page_offset;
1485 char *ba = __bio_kmap_atomic(bio, i, KM_USER0);
1487 memcpy(pa+page_offset, ba+b_offset, clen);
1489 memcpy(ba+b_offset, pa+page_offset, clen);
1490 __bio_kunmap_atomic(ba, KM_USER0);
1492 if (clen < len) /* hit end of page */
1498 #define check_xor() do { \
1499 if (count == MAX_XOR_BLOCKS) { \
1500 xor_blocks(count, STRIPE_SIZE, dest, ptr);\
1505 static void compute_parity6(struct stripe_head *sh, int method)
1507 raid6_conf_t *conf = sh->raid_conf;
1508 int i, pd_idx = sh->pd_idx, qd_idx, d0_idx, disks = sh->disks, count;
1510 /**** FIX THIS: This could be very bad if disks is close to 256 ****/
1513 qd_idx = raid6_next_disk(pd_idx, disks);
1514 d0_idx = raid6_next_disk(qd_idx, disks);
1516 pr_debug("compute_parity, stripe %llu, method %d\n",
1517 (unsigned long long)sh->sector, method);
1520 case READ_MODIFY_WRITE:
1521 BUG(); /* READ_MODIFY_WRITE N/A for RAID-6 */
1522 case RECONSTRUCT_WRITE:
1523 for (i= disks; i-- ;)
1524 if ( i != pd_idx && i != qd_idx && sh->dev[i].towrite ) {
1525 chosen = sh->dev[i].towrite;
1526 sh->dev[i].towrite = NULL;
1528 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
1529 wake_up(&conf->wait_for_overlap);
1531 BUG_ON(sh->dev[i].written);
1532 sh->dev[i].written = chosen;
1536 BUG(); /* Not implemented yet */
1539 for (i = disks; i--;)
1540 if (sh->dev[i].written) {
1541 sector_t sector = sh->dev[i].sector;
1542 struct bio *wbi = sh->dev[i].written;
1543 while (wbi && wbi->bi_sector < sector + STRIPE_SECTORS) {
1544 copy_data(1, wbi, sh->dev[i].page, sector);
1545 wbi = r5_next_bio(wbi, sector);
1548 set_bit(R5_LOCKED, &sh->dev[i].flags);
1549 set_bit(R5_UPTODATE, &sh->dev[i].flags);
1553 // case RECONSTRUCT_WRITE:
1554 // case CHECK_PARITY:
1555 // case UPDATE_PARITY:
1556 /* Note that unlike RAID-5, the ordering of the disks matters greatly. */
1557 /* FIX: Is this ordering of drives even remotely optimal? */
1561 ptrs[count++] = page_address(sh->dev[i].page);
1562 if (count <= disks-2 && !test_bit(R5_UPTODATE, &sh->dev[i].flags))
1563 printk("block %d/%d not uptodate on parity calc\n", i,count);
1564 i = raid6_next_disk(i, disks);
1565 } while ( i != d0_idx );
1569 raid6_call.gen_syndrome(disks, STRIPE_SIZE, ptrs);
1572 case RECONSTRUCT_WRITE:
1573 set_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
1574 set_bit(R5_UPTODATE, &sh->dev[qd_idx].flags);
1575 set_bit(R5_LOCKED, &sh->dev[pd_idx].flags);
1576 set_bit(R5_LOCKED, &sh->dev[qd_idx].flags);
1579 set_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
1580 set_bit(R5_UPTODATE, &sh->dev[qd_idx].flags);
1586 /* Compute one missing block */
1587 static void compute_block_1(struct stripe_head *sh, int dd_idx, int nozero)
1589 int i, count, disks = sh->disks;
1590 void *ptr[MAX_XOR_BLOCKS], *dest, *p;
1591 int pd_idx = sh->pd_idx;
1592 int qd_idx = raid6_next_disk(pd_idx, disks);
1594 pr_debug("compute_block_1, stripe %llu, idx %d\n",
1595 (unsigned long long)sh->sector, dd_idx);
1597 if ( dd_idx == qd_idx ) {
1598 /* We're actually computing the Q drive */
1599 compute_parity6(sh, UPDATE_PARITY);
1601 dest = page_address(sh->dev[dd_idx].page);
1602 if (!nozero) memset(dest, 0, STRIPE_SIZE);
1604 for (i = disks ; i--; ) {
1605 if (i == dd_idx || i == qd_idx)
1607 p = page_address(sh->dev[i].page);
1608 if (test_bit(R5_UPTODATE, &sh->dev[i].flags))
1611 printk("compute_block() %d, stripe %llu, %d"
1612 " not present\n", dd_idx,
1613 (unsigned long long)sh->sector, i);
1618 xor_blocks(count, STRIPE_SIZE, dest, ptr);
1619 if (!nozero) set_bit(R5_UPTODATE, &sh->dev[dd_idx].flags);
1620 else clear_bit(R5_UPTODATE, &sh->dev[dd_idx].flags);
1624 /* Compute two missing blocks */
1625 static void compute_block_2(struct stripe_head *sh, int dd_idx1, int dd_idx2)
1627 int i, count, disks = sh->disks;
1628 int pd_idx = sh->pd_idx;
1629 int qd_idx = raid6_next_disk(pd_idx, disks);
1630 int d0_idx = raid6_next_disk(qd_idx, disks);
1633 /* faila and failb are disk numbers relative to d0_idx */
1634 /* pd_idx become disks-2 and qd_idx become disks-1 */
1635 faila = (dd_idx1 < d0_idx) ? dd_idx1+(disks-d0_idx) : dd_idx1-d0_idx;
1636 failb = (dd_idx2 < d0_idx) ? dd_idx2+(disks-d0_idx) : dd_idx2-d0_idx;
1638 BUG_ON(faila == failb);
1639 if ( failb < faila ) { int tmp = faila; faila = failb; failb = tmp; }
1641 pr_debug("compute_block_2, stripe %llu, idx %d,%d (%d,%d)\n",
1642 (unsigned long long)sh->sector, dd_idx1, dd_idx2, faila, failb);
1644 if ( failb == disks-1 ) {
1645 /* Q disk is one of the missing disks */
1646 if ( faila == disks-2 ) {
1647 /* Missing P+Q, just recompute */
1648 compute_parity6(sh, UPDATE_PARITY);
1651 /* We're missing D+Q; recompute D from P */
1652 compute_block_1(sh, (dd_idx1 == qd_idx) ? dd_idx2 : dd_idx1, 0);
1653 compute_parity6(sh, UPDATE_PARITY); /* Is this necessary? */
1658 /* We're missing D+P or D+D; build pointer table */
1660 /**** FIX THIS: This could be very bad if disks is close to 256 ****/
1666 ptrs[count++] = page_address(sh->dev[i].page);
1667 i = raid6_next_disk(i, disks);
1668 if (i != dd_idx1 && i != dd_idx2 &&
1669 !test_bit(R5_UPTODATE, &sh->dev[i].flags))
1670 printk("compute_2 with missing block %d/%d\n", count, i);
1671 } while ( i != d0_idx );
1673 if ( failb == disks-2 ) {
1674 /* We're missing D+P. */
1675 raid6_datap_recov(disks, STRIPE_SIZE, faila, ptrs);
1677 /* We're missing D+D. */
1678 raid6_2data_recov(disks, STRIPE_SIZE, faila, failb, ptrs);
1681 /* Both the above update both missing blocks */
1682 set_bit(R5_UPTODATE, &sh->dev[dd_idx1].flags);
1683 set_bit(R5_UPTODATE, &sh->dev[dd_idx2].flags);
1688 handle_write_operations5(struct stripe_head *sh, int rcw, int expand)
1690 int i, pd_idx = sh->pd_idx, disks = sh->disks;
1694 /* if we are not expanding this is a proper write request, and
1695 * there will be bios with new data to be drained into the
1699 set_bit(STRIPE_OP_BIODRAIN, &sh->ops.pending);
1703 set_bit(STRIPE_OP_POSTXOR, &sh->ops.pending);
1706 for (i = disks; i--; ) {
1707 struct r5dev *dev = &sh->dev[i];
1710 set_bit(R5_LOCKED, &dev->flags);
1712 clear_bit(R5_UPTODATE, &dev->flags);
1716 if (locked + 1 == disks)
1717 if (!test_and_set_bit(STRIPE_FULL_WRITE, &sh->state))
1718 atomic_inc(&sh->raid_conf->pending_full_writes);
1720 BUG_ON(!(test_bit(R5_UPTODATE, &sh->dev[pd_idx].flags) ||
1721 test_bit(R5_Wantcompute, &sh->dev[pd_idx].flags)));
1723 set_bit(STRIPE_OP_PREXOR, &sh->ops.pending);
1724 set_bit(STRIPE_OP_BIODRAIN, &sh->ops.pending);
1725 set_bit(STRIPE_OP_POSTXOR, &sh->ops.pending);
1729 for (i = disks; i--; ) {
1730 struct r5dev *dev = &sh->dev[i];
1734 /* For a read-modify write there may be blocks that are
1735 * locked for reading while others are ready to be
1736 * written so we distinguish these blocks by the
1740 (test_bit(R5_UPTODATE, &dev->flags) ||
1741 test_bit(R5_Wantcompute, &dev->flags))) {
1742 set_bit(R5_Wantprexor, &dev->flags);
1743 set_bit(R5_LOCKED, &dev->flags);
1744 clear_bit(R5_UPTODATE, &dev->flags);
1750 /* keep the parity disk locked while asynchronous operations
1753 set_bit(R5_LOCKED, &sh->dev[pd_idx].flags);
1754 clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
1757 pr_debug("%s: stripe %llu locked: %d pending: %lx\n",
1758 __func__, (unsigned long long)sh->sector,
1759 locked, sh->ops.pending);
1765 * Each stripe/dev can have one or more bion attached.
1766 * toread/towrite point to the first in a chain.
1767 * The bi_next chain must be in order.
1769 static int add_stripe_bio(struct stripe_head *sh, struct bio *bi, int dd_idx, int forwrite)
1772 raid5_conf_t *conf = sh->raid_conf;
1775 pr_debug("adding bh b#%llu to stripe s#%llu\n",
1776 (unsigned long long)bi->bi_sector,
1777 (unsigned long long)sh->sector);
1780 spin_lock(&sh->lock);
1781 spin_lock_irq(&conf->device_lock);
1783 bip = &sh->dev[dd_idx].towrite;
1784 if (*bip == NULL && sh->dev[dd_idx].written == NULL)
1787 bip = &sh->dev[dd_idx].toread;
1788 while (*bip && (*bip)->bi_sector < bi->bi_sector) {
1789 if ((*bip)->bi_sector + ((*bip)->bi_size >> 9) > bi->bi_sector)
1791 bip = & (*bip)->bi_next;
1793 if (*bip && (*bip)->bi_sector < bi->bi_sector + ((bi->bi_size)>>9))
1796 BUG_ON(*bip && bi->bi_next && (*bip) != bi->bi_next);
1800 bi->bi_phys_segments ++;
1801 spin_unlock_irq(&conf->device_lock);
1802 spin_unlock(&sh->lock);
1804 pr_debug("added bi b#%llu to stripe s#%llu, disk %d.\n",
1805 (unsigned long long)bi->bi_sector,
1806 (unsigned long long)sh->sector, dd_idx);
1808 if (conf->mddev->bitmap && firstwrite) {
1809 bitmap_startwrite(conf->mddev->bitmap, sh->sector,
1811 sh->bm_seq = conf->seq_flush+1;
1812 set_bit(STRIPE_BIT_DELAY, &sh->state);
1816 /* check if page is covered */
1817 sector_t sector = sh->dev[dd_idx].sector;
1818 for (bi=sh->dev[dd_idx].towrite;
1819 sector < sh->dev[dd_idx].sector + STRIPE_SECTORS &&
1820 bi && bi->bi_sector <= sector;
1821 bi = r5_next_bio(bi, sh->dev[dd_idx].sector)) {
1822 if (bi->bi_sector + (bi->bi_size>>9) >= sector)
1823 sector = bi->bi_sector + (bi->bi_size>>9);
1825 if (sector >= sh->dev[dd_idx].sector + STRIPE_SECTORS)
1826 set_bit(R5_OVERWRITE, &sh->dev[dd_idx].flags);
1831 set_bit(R5_Overlap, &sh->dev[dd_idx].flags);
1832 spin_unlock_irq(&conf->device_lock);
1833 spin_unlock(&sh->lock);
1837 static void end_reshape(raid5_conf_t *conf);
1839 static int page_is_zero(struct page *p)
1841 char *a = page_address(p);
1842 return ((*(u32*)a) == 0 &&
1843 memcmp(a, a+4, STRIPE_SIZE-4)==0);
1846 static int stripe_to_pdidx(sector_t stripe, raid5_conf_t *conf, int disks)
1848 int sectors_per_chunk = conf->chunk_size >> 9;
1850 int chunk_offset = sector_div(stripe, sectors_per_chunk);
1852 raid5_compute_sector(stripe * (disks - conf->max_degraded)
1853 *sectors_per_chunk + chunk_offset,
1854 disks, disks - conf->max_degraded,
1855 &dd_idx, &pd_idx, conf);
1860 handle_requests_to_failed_array(raid5_conf_t *conf, struct stripe_head *sh,
1861 struct stripe_head_state *s, int disks,
1862 struct bio **return_bi)
1865 for (i = disks; i--; ) {
1869 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
1872 rdev = rcu_dereference(conf->disks[i].rdev);
1873 if (rdev && test_bit(In_sync, &rdev->flags))
1874 /* multiple read failures in one stripe */
1875 md_error(conf->mddev, rdev);
1878 spin_lock_irq(&conf->device_lock);
1879 /* fail all writes first */
1880 bi = sh->dev[i].towrite;
1881 sh->dev[i].towrite = NULL;
1887 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
1888 wake_up(&conf->wait_for_overlap);
1890 while (bi && bi->bi_sector <
1891 sh->dev[i].sector + STRIPE_SECTORS) {
1892 struct bio *nextbi = r5_next_bio(bi, sh->dev[i].sector);
1893 clear_bit(BIO_UPTODATE, &bi->bi_flags);
1894 if (--bi->bi_phys_segments == 0) {
1895 md_write_end(conf->mddev);
1896 bi->bi_next = *return_bi;
1901 /* and fail all 'written' */
1902 bi = sh->dev[i].written;
1903 sh->dev[i].written = NULL;
1904 if (bi) bitmap_end = 1;
1905 while (bi && bi->bi_sector <
1906 sh->dev[i].sector + STRIPE_SECTORS) {
1907 struct bio *bi2 = r5_next_bio(bi, sh->dev[i].sector);
1908 clear_bit(BIO_UPTODATE, &bi->bi_flags);
1909 if (--bi->bi_phys_segments == 0) {
1910 md_write_end(conf->mddev);
1911 bi->bi_next = *return_bi;
1917 /* fail any reads if this device is non-operational and
1918 * the data has not reached the cache yet.
1920 if (!test_bit(R5_Wantfill, &sh->dev[i].flags) &&
1921 (!test_bit(R5_Insync, &sh->dev[i].flags) ||
1922 test_bit(R5_ReadError, &sh->dev[i].flags))) {
1923 bi = sh->dev[i].toread;
1924 sh->dev[i].toread = NULL;
1925 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
1926 wake_up(&conf->wait_for_overlap);
1927 if (bi) s->to_read--;
1928 while (bi && bi->bi_sector <
1929 sh->dev[i].sector + STRIPE_SECTORS) {
1930 struct bio *nextbi =
1931 r5_next_bio(bi, sh->dev[i].sector);
1932 clear_bit(BIO_UPTODATE, &bi->bi_flags);
1933 if (--bi->bi_phys_segments == 0) {
1934 bi->bi_next = *return_bi;
1940 spin_unlock_irq(&conf->device_lock);
1942 bitmap_endwrite(conf->mddev->bitmap, sh->sector,
1943 STRIPE_SECTORS, 0, 0);
1946 if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
1947 if (atomic_dec_and_test(&conf->pending_full_writes))
1948 md_wakeup_thread(conf->mddev->thread);
1951 /* __handle_issuing_new_read_requests5 - returns 0 if there are no more disks
1954 static int __handle_issuing_new_read_requests5(struct stripe_head *sh,
1955 struct stripe_head_state *s, int disk_idx, int disks)
1957 struct r5dev *dev = &sh->dev[disk_idx];
1958 struct r5dev *failed_dev = &sh->dev[s->failed_num];
1960 /* is the data in this block needed, and can we get it? */
1961 if (!test_bit(R5_LOCKED, &dev->flags) &&
1962 !test_bit(R5_UPTODATE, &dev->flags) && (dev->toread ||
1963 (dev->towrite && !test_bit(R5_OVERWRITE, &dev->flags)) ||
1964 s->syncing || s->expanding || (s->failed &&
1965 (failed_dev->toread || (failed_dev->towrite &&
1966 !test_bit(R5_OVERWRITE, &failed_dev->flags)
1968 /* We would like to get this block, possibly by computing it,
1969 * otherwise read it if the backing disk is insync
1971 if ((s->uptodate == disks - 1) &&
1972 (s->failed && disk_idx == s->failed_num)) {
1973 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
1974 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
1975 set_bit(R5_Wantcompute, &dev->flags);
1976 sh->ops.target = disk_idx;
1978 /* Careful: from this point on 'uptodate' is in the eye
1979 * of raid5_run_ops which services 'compute' operations
1980 * before writes. R5_Wantcompute flags a block that will
1981 * be R5_UPTODATE by the time it is needed for a
1982 * subsequent operation.
1985 return 0; /* uptodate + compute == disks */
1986 } else if (test_bit(R5_Insync, &dev->flags)) {
1987 set_bit(R5_LOCKED, &dev->flags);
1988 set_bit(R5_Wantread, &dev->flags);
1990 pr_debug("Reading block %d (sync=%d)\n", disk_idx,
1998 static void handle_issuing_new_read_requests5(struct stripe_head *sh,
1999 struct stripe_head_state *s, int disks)
2003 /* look for blocks to read/compute, skip this if a compute
2004 * is already in flight, or if the stripe contents are in the
2005 * midst of changing due to a write
2007 if (!test_bit(STRIPE_COMPUTE_RUN, &sh->state) && !sh->check_state &&
2008 !test_bit(STRIPE_OP_PREXOR, &sh->ops.pending) &&
2009 !test_bit(STRIPE_OP_POSTXOR, &sh->ops.pending)) {
2010 for (i = disks; i--; )
2011 if (__handle_issuing_new_read_requests5(
2012 sh, s, i, disks) == 0)
2015 set_bit(STRIPE_HANDLE, &sh->state);
2018 static void handle_issuing_new_read_requests6(struct stripe_head *sh,
2019 struct stripe_head_state *s, struct r6_state *r6s,
2023 for (i = disks; i--; ) {
2024 struct r5dev *dev = &sh->dev[i];
2025 if (!test_bit(R5_LOCKED, &dev->flags) &&
2026 !test_bit(R5_UPTODATE, &dev->flags) &&
2027 (dev->toread || (dev->towrite &&
2028 !test_bit(R5_OVERWRITE, &dev->flags)) ||
2029 s->syncing || s->expanding ||
2031 (sh->dev[r6s->failed_num[0]].toread ||
2034 (sh->dev[r6s->failed_num[1]].toread ||
2036 /* we would like to get this block, possibly
2037 * by computing it, but we might not be able to
2039 if ((s->uptodate == disks - 1) &&
2040 (s->failed && (i == r6s->failed_num[0] ||
2041 i == r6s->failed_num[1]))) {
2042 pr_debug("Computing stripe %llu block %d\n",
2043 (unsigned long long)sh->sector, i);
2044 compute_block_1(sh, i, 0);
2046 } else if ( s->uptodate == disks-2 && s->failed >= 2 ) {
2047 /* Computing 2-failure is *very* expensive; only
2048 * do it if failed >= 2
2051 for (other = disks; other--; ) {
2054 if (!test_bit(R5_UPTODATE,
2055 &sh->dev[other].flags))
2059 pr_debug("Computing stripe %llu blocks %d,%d\n",
2060 (unsigned long long)sh->sector,
2062 compute_block_2(sh, i, other);
2064 } else if (test_bit(R5_Insync, &dev->flags)) {
2065 set_bit(R5_LOCKED, &dev->flags);
2066 set_bit(R5_Wantread, &dev->flags);
2068 pr_debug("Reading block %d (sync=%d)\n",
2073 set_bit(STRIPE_HANDLE, &sh->state);
2077 /* handle_completed_write_requests
2078 * any written block on an uptodate or failed drive can be returned.
2079 * Note that if we 'wrote' to a failed drive, it will be UPTODATE, but
2080 * never LOCKED, so we don't need to test 'failed' directly.
2082 static void handle_completed_write_requests(raid5_conf_t *conf,
2083 struct stripe_head *sh, int disks, struct bio **return_bi)
2088 for (i = disks; i--; )
2089 if (sh->dev[i].written) {
2091 if (!test_bit(R5_LOCKED, &dev->flags) &&
2092 test_bit(R5_UPTODATE, &dev->flags)) {
2093 /* We can return any write requests */
2094 struct bio *wbi, *wbi2;
2096 pr_debug("Return write for disc %d\n", i);
2097 spin_lock_irq(&conf->device_lock);
2099 dev->written = NULL;
2100 while (wbi && wbi->bi_sector <
2101 dev->sector + STRIPE_SECTORS) {
2102 wbi2 = r5_next_bio(wbi, dev->sector);
2103 if (--wbi->bi_phys_segments == 0) {
2104 md_write_end(conf->mddev);
2105 wbi->bi_next = *return_bi;
2110 if (dev->towrite == NULL)
2112 spin_unlock_irq(&conf->device_lock);
2114 bitmap_endwrite(conf->mddev->bitmap,
2117 !test_bit(STRIPE_DEGRADED, &sh->state),
2122 if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
2123 if (atomic_dec_and_test(&conf->pending_full_writes))
2124 md_wakeup_thread(conf->mddev->thread);
2127 static void handle_issuing_new_write_requests5(raid5_conf_t *conf,
2128 struct stripe_head *sh, struct stripe_head_state *s, int disks)
2130 int rmw = 0, rcw = 0, i;
2131 for (i = disks; i--; ) {
2132 /* would I have to read this buffer for read_modify_write */
2133 struct r5dev *dev = &sh->dev[i];
2134 if ((dev->towrite || i == sh->pd_idx) &&
2135 !test_bit(R5_LOCKED, &dev->flags) &&
2136 !(test_bit(R5_UPTODATE, &dev->flags) ||
2137 test_bit(R5_Wantcompute, &dev->flags))) {
2138 if (test_bit(R5_Insync, &dev->flags))
2141 rmw += 2*disks; /* cannot read it */
2143 /* Would I have to read this buffer for reconstruct_write */
2144 if (!test_bit(R5_OVERWRITE, &dev->flags) && i != sh->pd_idx &&
2145 !test_bit(R5_LOCKED, &dev->flags) &&
2146 !(test_bit(R5_UPTODATE, &dev->flags) ||
2147 test_bit(R5_Wantcompute, &dev->flags))) {
2148 if (test_bit(R5_Insync, &dev->flags)) rcw++;
2153 pr_debug("for sector %llu, rmw=%d rcw=%d\n",
2154 (unsigned long long)sh->sector, rmw, rcw);
2155 set_bit(STRIPE_HANDLE, &sh->state);
2156 if (rmw < rcw && rmw > 0)
2157 /* prefer read-modify-write, but need to get some data */
2158 for (i = disks; i--; ) {
2159 struct r5dev *dev = &sh->dev[i];
2160 if ((dev->towrite || i == sh->pd_idx) &&
2161 !test_bit(R5_LOCKED, &dev->flags) &&
2162 !(test_bit(R5_UPTODATE, &dev->flags) ||
2163 test_bit(R5_Wantcompute, &dev->flags)) &&
2164 test_bit(R5_Insync, &dev->flags)) {
2166 test_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2167 pr_debug("Read_old block "
2168 "%d for r-m-w\n", i);
2169 set_bit(R5_LOCKED, &dev->flags);
2170 set_bit(R5_Wantread, &dev->flags);
2173 set_bit(STRIPE_DELAYED, &sh->state);
2174 set_bit(STRIPE_HANDLE, &sh->state);
2178 if (rcw <= rmw && rcw > 0)
2179 /* want reconstruct write, but need to get some data */
2180 for (i = disks; i--; ) {
2181 struct r5dev *dev = &sh->dev[i];
2182 if (!test_bit(R5_OVERWRITE, &dev->flags) &&
2184 !test_bit(R5_LOCKED, &dev->flags) &&
2185 !(test_bit(R5_UPTODATE, &dev->flags) ||
2186 test_bit(R5_Wantcompute, &dev->flags)) &&
2187 test_bit(R5_Insync, &dev->flags)) {
2189 test_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2190 pr_debug("Read_old block "
2191 "%d for Reconstruct\n", i);
2192 set_bit(R5_LOCKED, &dev->flags);
2193 set_bit(R5_Wantread, &dev->flags);
2196 set_bit(STRIPE_DELAYED, &sh->state);
2197 set_bit(STRIPE_HANDLE, &sh->state);
2201 /* now if nothing is locked, and if we have enough data,
2202 * we can start a write request
2204 /* since handle_stripe can be called at any time we need to handle the
2205 * case where a compute block operation has been submitted and then a
2206 * subsequent call wants to start a write request. raid5_run_ops only
2207 * handles the case where compute block and postxor are requested
2208 * simultaneously. If this is not the case then new writes need to be
2209 * held off until the compute completes.
2211 if ((s->req_compute || !test_bit(STRIPE_COMPUTE_RUN, &sh->state)) &&
2212 (s->locked == 0 && (rcw == 0 || rmw == 0) &&
2213 !test_bit(STRIPE_BIT_DELAY, &sh->state)))
2214 s->locked += handle_write_operations5(sh, rcw == 0, 0);
2217 static void handle_issuing_new_write_requests6(raid5_conf_t *conf,
2218 struct stripe_head *sh, struct stripe_head_state *s,
2219 struct r6_state *r6s, int disks)
2221 int rcw = 0, must_compute = 0, pd_idx = sh->pd_idx, i;
2222 int qd_idx = r6s->qd_idx;
2223 for (i = disks; i--; ) {
2224 struct r5dev *dev = &sh->dev[i];
2225 /* Would I have to read this buffer for reconstruct_write */
2226 if (!test_bit(R5_OVERWRITE, &dev->flags)
2227 && i != pd_idx && i != qd_idx
2228 && (!test_bit(R5_LOCKED, &dev->flags)
2230 !test_bit(R5_UPTODATE, &dev->flags)) {
2231 if (test_bit(R5_Insync, &dev->flags)) rcw++;
2233 pr_debug("raid6: must_compute: "
2234 "disk %d flags=%#lx\n", i, dev->flags);
2239 pr_debug("for sector %llu, rcw=%d, must_compute=%d\n",
2240 (unsigned long long)sh->sector, rcw, must_compute);
2241 set_bit(STRIPE_HANDLE, &sh->state);
2244 /* want reconstruct write, but need to get some data */
2245 for (i = disks; i--; ) {
2246 struct r5dev *dev = &sh->dev[i];
2247 if (!test_bit(R5_OVERWRITE, &dev->flags)
2248 && !(s->failed == 0 && (i == pd_idx || i == qd_idx))
2249 && !test_bit(R5_LOCKED, &dev->flags) &&
2250 !test_bit(R5_UPTODATE, &dev->flags) &&
2251 test_bit(R5_Insync, &dev->flags)) {
2253 test_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2254 pr_debug("Read_old stripe %llu "
2255 "block %d for Reconstruct\n",
2256 (unsigned long long)sh->sector, i);
2257 set_bit(R5_LOCKED, &dev->flags);
2258 set_bit(R5_Wantread, &dev->flags);
2261 pr_debug("Request delayed stripe %llu "
2262 "block %d for Reconstruct\n",
2263 (unsigned long long)sh->sector, i);
2264 set_bit(STRIPE_DELAYED, &sh->state);
2265 set_bit(STRIPE_HANDLE, &sh->state);
2269 /* now if nothing is locked, and if we have enough data, we can start a
2272 if (s->locked == 0 && rcw == 0 &&
2273 !test_bit(STRIPE_BIT_DELAY, &sh->state)) {
2274 if (must_compute > 0) {
2275 /* We have failed blocks and need to compute them */
2276 switch (s->failed) {
2280 compute_block_1(sh, r6s->failed_num[0], 0);
2283 compute_block_2(sh, r6s->failed_num[0],
2284 r6s->failed_num[1]);
2286 default: /* This request should have been failed? */
2291 pr_debug("Computing parity for stripe %llu\n",
2292 (unsigned long long)sh->sector);
2293 compute_parity6(sh, RECONSTRUCT_WRITE);
2294 /* now every locked buffer is ready to be written */
2295 for (i = disks; i--; )
2296 if (test_bit(R5_LOCKED, &sh->dev[i].flags)) {
2297 pr_debug("Writing stripe %llu block %d\n",
2298 (unsigned long long)sh->sector, i);
2300 set_bit(R5_Wantwrite, &sh->dev[i].flags);
2302 if (s->locked == disks)
2303 if (!test_and_set_bit(STRIPE_FULL_WRITE, &sh->state))
2304 atomic_inc(&conf->pending_full_writes);
2305 /* after a RECONSTRUCT_WRITE, the stripe MUST be in-sync */
2306 set_bit(STRIPE_INSYNC, &sh->state);
2308 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2309 atomic_dec(&conf->preread_active_stripes);
2310 if (atomic_read(&conf->preread_active_stripes) <
2312 md_wakeup_thread(conf->mddev->thread);
2317 static void handle_parity_checks5(raid5_conf_t *conf, struct stripe_head *sh,
2318 struct stripe_head_state *s, int disks)
2320 struct r5dev *dev = NULL;
2322 set_bit(STRIPE_HANDLE, &sh->state);
2324 switch (sh->check_state) {
2325 case check_state_idle:
2326 /* start a new check operation if there are no failures */
2327 if (s->failed == 0) {
2328 BUG_ON(s->uptodate != disks);
2329 sh->check_state = check_state_run;
2330 set_bit(STRIPE_OP_CHECK, &s->ops_request);
2331 clear_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags);
2335 dev = &sh->dev[s->failed_num];
2337 case check_state_compute_result:
2338 sh->check_state = check_state_idle;
2340 dev = &sh->dev[sh->pd_idx];
2342 /* check that a write has not made the stripe insync */
2343 if (test_bit(STRIPE_INSYNC, &sh->state))
2346 /* either failed parity check, or recovery is happening */
2347 BUG_ON(!test_bit(R5_UPTODATE, &dev->flags));
2348 BUG_ON(s->uptodate != disks);
2350 set_bit(R5_LOCKED, &dev->flags);
2352 set_bit(R5_Wantwrite, &dev->flags);
2354 clear_bit(STRIPE_DEGRADED, &sh->state);
2355 set_bit(STRIPE_INSYNC, &sh->state);
2357 case check_state_run:
2358 break; /* we will be called again upon completion */
2359 case check_state_check_result:
2360 sh->check_state = check_state_idle;
2362 /* if a failure occurred during the check operation, leave
2363 * STRIPE_INSYNC not set and let the stripe be handled again
2368 /* handle a successful check operation, if parity is correct
2369 * we are done. Otherwise update the mismatch count and repair
2370 * parity if !MD_RECOVERY_CHECK
2372 if (sh->ops.zero_sum_result == 0)
2373 /* parity is correct (on disc,
2374 * not in buffer any more)
2376 set_bit(STRIPE_INSYNC, &sh->state);
2378 conf->mddev->resync_mismatches += STRIPE_SECTORS;
2379 if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery))
2380 /* don't try to repair!! */
2381 set_bit(STRIPE_INSYNC, &sh->state);
2383 sh->check_state = check_state_compute_run;
2384 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
2385 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
2386 set_bit(R5_Wantcompute,
2387 &sh->dev[sh->pd_idx].flags);
2388 sh->ops.target = sh->pd_idx;
2393 case check_state_compute_run:
2396 printk(KERN_ERR "%s: unknown check_state: %d sector: %llu\n",
2397 __func__, sh->check_state,
2398 (unsigned long long) sh->sector);
2404 static void handle_parity_checks6(raid5_conf_t *conf, struct stripe_head *sh,
2405 struct stripe_head_state *s,
2406 struct r6_state *r6s, struct page *tmp_page,
2409 int update_p = 0, update_q = 0;
2411 int pd_idx = sh->pd_idx;
2412 int qd_idx = r6s->qd_idx;
2414 set_bit(STRIPE_HANDLE, &sh->state);
2416 BUG_ON(s->failed > 2);
2417 BUG_ON(s->uptodate < disks);
2418 /* Want to check and possibly repair P and Q.
2419 * However there could be one 'failed' device, in which
2420 * case we can only check one of them, possibly using the
2421 * other to generate missing data
2424 /* If !tmp_page, we cannot do the calculations,
2425 * but as we have set STRIPE_HANDLE, we will soon be called
2426 * by stripe_handle with a tmp_page - just wait until then.
2429 if (s->failed == r6s->q_failed) {
2430 /* The only possible failed device holds 'Q', so it
2431 * makes sense to check P (If anything else were failed,
2432 * we would have used P to recreate it).
2434 compute_block_1(sh, pd_idx, 1);
2435 if (!page_is_zero(sh->dev[pd_idx].page)) {
2436 compute_block_1(sh, pd_idx, 0);
2440 if (!r6s->q_failed && s->failed < 2) {
2441 /* q is not failed, and we didn't use it to generate
2442 * anything, so it makes sense to check it
2444 memcpy(page_address(tmp_page),
2445 page_address(sh->dev[qd_idx].page),
2447 compute_parity6(sh, UPDATE_PARITY);
2448 if (memcmp(page_address(tmp_page),
2449 page_address(sh->dev[qd_idx].page),
2450 STRIPE_SIZE) != 0) {
2451 clear_bit(STRIPE_INSYNC, &sh->state);
2455 if (update_p || update_q) {
2456 conf->mddev->resync_mismatches += STRIPE_SECTORS;
2457 if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery))
2458 /* don't try to repair!! */
2459 update_p = update_q = 0;
2462 /* now write out any block on a failed drive,
2463 * or P or Q if they need it
2466 if (s->failed == 2) {
2467 dev = &sh->dev[r6s->failed_num[1]];
2469 set_bit(R5_LOCKED, &dev->flags);
2470 set_bit(R5_Wantwrite, &dev->flags);
2472 if (s->failed >= 1) {
2473 dev = &sh->dev[r6s->failed_num[0]];
2475 set_bit(R5_LOCKED, &dev->flags);
2476 set_bit(R5_Wantwrite, &dev->flags);
2480 dev = &sh->dev[pd_idx];
2482 set_bit(R5_LOCKED, &dev->flags);
2483 set_bit(R5_Wantwrite, &dev->flags);
2486 dev = &sh->dev[qd_idx];
2488 set_bit(R5_LOCKED, &dev->flags);
2489 set_bit(R5_Wantwrite, &dev->flags);
2491 clear_bit(STRIPE_DEGRADED, &sh->state);
2493 set_bit(STRIPE_INSYNC, &sh->state);
2497 static void handle_stripe_expansion(raid5_conf_t *conf, struct stripe_head *sh,
2498 struct r6_state *r6s)
2502 /* We have read all the blocks in this stripe and now we need to
2503 * copy some of them into a target stripe for expand.
2505 struct dma_async_tx_descriptor *tx = NULL;
2506 clear_bit(STRIPE_EXPAND_SOURCE, &sh->state);
2507 for (i = 0; i < sh->disks; i++)
2508 if (i != sh->pd_idx && (!r6s || i != r6s->qd_idx)) {
2509 int dd_idx, pd_idx, j;
2510 struct stripe_head *sh2;
2512 sector_t bn = compute_blocknr(sh, i);
2513 sector_t s = raid5_compute_sector(bn, conf->raid_disks,
2515 conf->max_degraded, &dd_idx,
2517 sh2 = get_active_stripe(conf, s, conf->raid_disks,
2520 /* so far only the early blocks of this stripe
2521 * have been requested. When later blocks
2522 * get requested, we will try again
2525 if (!test_bit(STRIPE_EXPANDING, &sh2->state) ||
2526 test_bit(R5_Expanded, &sh2->dev[dd_idx].flags)) {
2527 /* must have already done this block */
2528 release_stripe(sh2);
2532 /* place all the copies on one channel */
2533 tx = async_memcpy(sh2->dev[dd_idx].page,
2534 sh->dev[i].page, 0, 0, STRIPE_SIZE,
2535 ASYNC_TX_DEP_ACK, tx, NULL, NULL);
2537 set_bit(R5_Expanded, &sh2->dev[dd_idx].flags);
2538 set_bit(R5_UPTODATE, &sh2->dev[dd_idx].flags);
2539 for (j = 0; j < conf->raid_disks; j++)
2540 if (j != sh2->pd_idx &&
2541 (!r6s || j != raid6_next_disk(sh2->pd_idx,
2543 !test_bit(R5_Expanded, &sh2->dev[j].flags))
2545 if (j == conf->raid_disks) {
2546 set_bit(STRIPE_EXPAND_READY, &sh2->state);
2547 set_bit(STRIPE_HANDLE, &sh2->state);
2549 release_stripe(sh2);
2552 /* done submitting copies, wait for them to complete */
2555 dma_wait_for_async_tx(tx);
2561 * handle_stripe - do things to a stripe.
2563 * We lock the stripe and then examine the state of various bits
2564 * to see what needs to be done.
2566 * return some read request which now have data
2567 * return some write requests which are safely on disc
2568 * schedule a read on some buffers
2569 * schedule a write of some buffers
2570 * return confirmation of parity correctness
2572 * buffers are taken off read_list or write_list, and bh_cache buffers
2573 * get BH_Lock set before the stripe lock is released.
2577 static void handle_stripe5(struct stripe_head *sh)
2579 raid5_conf_t *conf = sh->raid_conf;
2580 int disks = sh->disks, i;
2581 struct bio *return_bi = NULL;
2582 struct stripe_head_state s;
2584 unsigned long pending = 0;
2585 mdk_rdev_t *blocked_rdev = NULL;
2588 memset(&s, 0, sizeof(s));
2589 pr_debug("handling stripe %llu, state=%#lx cnt=%d, pd_idx=%d "
2590 "ops=%lx:%lx:%lx\n", (unsigned long long)sh->sector, sh->state,
2591 atomic_read(&sh->count), sh->pd_idx,
2592 sh->ops.pending, sh->ops.ack, sh->ops.complete);
2594 spin_lock(&sh->lock);
2595 clear_bit(STRIPE_HANDLE, &sh->state);
2596 clear_bit(STRIPE_DELAYED, &sh->state);
2598 s.syncing = test_bit(STRIPE_SYNCING, &sh->state);
2599 s.expanding = test_bit(STRIPE_EXPAND_SOURCE, &sh->state);
2600 s.expanded = test_bit(STRIPE_EXPAND_READY, &sh->state);
2602 /* Now to look around and see what can be done */
2604 for (i=disks; i--; ) {
2606 struct r5dev *dev = &sh->dev[i];
2607 clear_bit(R5_Insync, &dev->flags);
2609 pr_debug("check %d: state 0x%lx toread %p read %p write %p "
2610 "written %p\n", i, dev->flags, dev->toread, dev->read,
2611 dev->towrite, dev->written);
2613 /* maybe we can request a biofill operation
2615 * new wantfill requests are only permitted while
2616 * ops_complete_biofill is guaranteed to be inactive
2618 if (test_bit(R5_UPTODATE, &dev->flags) && dev->toread &&
2619 !test_bit(STRIPE_BIOFILL_RUN, &sh->state))
2620 set_bit(R5_Wantfill, &dev->flags);
2622 /* now count some things */
2623 if (test_bit(R5_LOCKED, &dev->flags)) s.locked++;
2624 if (test_bit(R5_UPTODATE, &dev->flags)) s.uptodate++;
2625 if (test_bit(R5_Wantcompute, &dev->flags)) s.compute++;
2627 if (test_bit(R5_Wantfill, &dev->flags))
2629 else if (dev->toread)
2633 if (!test_bit(R5_OVERWRITE, &dev->flags))
2638 rdev = rcu_dereference(conf->disks[i].rdev);
2639 if (rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
2640 blocked_rdev = rdev;
2641 atomic_inc(&rdev->nr_pending);
2644 if (!rdev || !test_bit(In_sync, &rdev->flags)) {
2645 /* The ReadError flag will just be confusing now */
2646 clear_bit(R5_ReadError, &dev->flags);
2647 clear_bit(R5_ReWrite, &dev->flags);
2649 if (!rdev || !test_bit(In_sync, &rdev->flags)
2650 || test_bit(R5_ReadError, &dev->flags)) {
2654 set_bit(R5_Insync, &dev->flags);
2658 if (unlikely(blocked_rdev)) {
2659 set_bit(STRIPE_HANDLE, &sh->state);
2663 if (s.to_fill && !test_bit(STRIPE_BIOFILL_RUN, &sh->state)) {
2664 set_bit(STRIPE_OP_BIOFILL, &s.ops_request);
2665 set_bit(STRIPE_BIOFILL_RUN, &sh->state);
2668 pr_debug("locked=%d uptodate=%d to_read=%d"
2669 " to_write=%d failed=%d failed_num=%d\n",
2670 s.locked, s.uptodate, s.to_read, s.to_write,
2671 s.failed, s.failed_num);
2672 /* check if the array has lost two devices and, if so, some requests might
2675 if (s.failed > 1 && s.to_read+s.to_write+s.written)
2676 handle_requests_to_failed_array(conf, sh, &s, disks,
2678 if (s.failed > 1 && s.syncing) {
2679 md_done_sync(conf->mddev, STRIPE_SECTORS,0);
2680 clear_bit(STRIPE_SYNCING, &sh->state);
2684 /* might be able to return some write requests if the parity block
2685 * is safe, or on a failed drive
2687 dev = &sh->dev[sh->pd_idx];
2689 ((test_bit(R5_Insync, &dev->flags) &&
2690 !test_bit(R5_LOCKED, &dev->flags) &&
2691 test_bit(R5_UPTODATE, &dev->flags)) ||
2692 (s.failed == 1 && s.failed_num == sh->pd_idx)))
2693 handle_completed_write_requests(conf, sh, disks, &return_bi);
2695 /* Now we might consider reading some blocks, either to check/generate
2696 * parity, or to satisfy requests
2697 * or to load a block that is being partially written.
2699 if (s.to_read || s.non_overwrite ||
2700 (s.syncing && (s.uptodate + s.compute < disks)) || s.expanding)
2701 handle_issuing_new_read_requests5(sh, &s, disks);
2703 /* Now we check to see if any write operations have recently
2707 /* leave prexor set until postxor is done, allows us to distinguish
2708 * a rmw from a rcw during biodrain
2711 if (test_bit(STRIPE_OP_PREXOR, &sh->ops.complete) &&
2712 test_bit(STRIPE_OP_POSTXOR, &sh->ops.complete)) {
2715 clear_bit(STRIPE_OP_PREXOR, &sh->ops.complete);
2716 clear_bit(STRIPE_OP_PREXOR, &sh->ops.ack);
2717 clear_bit(STRIPE_OP_PREXOR, &sh->ops.pending);
2719 for (i = disks; i--; )
2720 clear_bit(R5_Wantprexor, &sh->dev[i].flags);
2723 /* if only POSTXOR is set then this is an 'expand' postxor */
2724 if (test_bit(STRIPE_OP_BIODRAIN, &sh->ops.complete) &&
2725 test_bit(STRIPE_OP_POSTXOR, &sh->ops.complete)) {
2727 clear_bit(STRIPE_OP_BIODRAIN, &sh->ops.complete);
2728 clear_bit(STRIPE_OP_BIODRAIN, &sh->ops.ack);
2729 clear_bit(STRIPE_OP_BIODRAIN, &sh->ops.pending);
2731 clear_bit(STRIPE_OP_POSTXOR, &sh->ops.complete);
2732 clear_bit(STRIPE_OP_POSTXOR, &sh->ops.ack);
2733 clear_bit(STRIPE_OP_POSTXOR, &sh->ops.pending);
2735 /* All the 'written' buffers and the parity block are ready to
2736 * be written back to disk
2738 BUG_ON(!test_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags));
2739 for (i = disks; i--; ) {
2741 if (test_bit(R5_LOCKED, &dev->flags) &&
2742 (i == sh->pd_idx || dev->written)) {
2743 pr_debug("Writing block %d\n", i);
2744 set_bit(R5_Wantwrite, &dev->flags);
2747 if (!test_bit(R5_Insync, &dev->flags) ||
2748 (i == sh->pd_idx && s.failed == 0))
2749 set_bit(STRIPE_INSYNC, &sh->state);
2752 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2753 atomic_dec(&conf->preread_active_stripes);
2754 if (atomic_read(&conf->preread_active_stripes) <
2756 md_wakeup_thread(conf->mddev->thread);
2760 /* Now to consider new write requests and what else, if anything
2761 * should be read. We do not handle new writes when:
2762 * 1/ A 'write' operation (copy+xor) is already in flight.
2763 * 2/ A 'check' operation is in flight, as it may clobber the parity
2766 if (s.to_write && !test_bit(STRIPE_OP_POSTXOR, &sh->ops.pending) &&
2768 handle_issuing_new_write_requests5(conf, sh, &s, disks);
2770 /* maybe we need to check and possibly fix the parity for this stripe
2771 * Any reads will already have been scheduled, so we just see if enough
2772 * data is available. The parity check is held off while parity
2773 * dependent operations are in flight.
2775 if (sh->check_state ||
2776 (s.syncing && s.locked == 0 &&
2777 !test_bit(STRIPE_COMPUTE_RUN, &sh->state) &&
2778 !test_bit(STRIPE_INSYNC, &sh->state)))
2779 handle_parity_checks5(conf, sh, &s, disks);
2781 if (s.syncing && s.locked == 0 && test_bit(STRIPE_INSYNC, &sh->state)) {
2782 md_done_sync(conf->mddev, STRIPE_SECTORS,1);
2783 clear_bit(STRIPE_SYNCING, &sh->state);
2786 /* If the failed drive is just a ReadError, then we might need to progress
2787 * the repair/check process
2789 if (s.failed == 1 && !conf->mddev->ro &&
2790 test_bit(R5_ReadError, &sh->dev[s.failed_num].flags)
2791 && !test_bit(R5_LOCKED, &sh->dev[s.failed_num].flags)
2792 && test_bit(R5_UPTODATE, &sh->dev[s.failed_num].flags)
2794 dev = &sh->dev[s.failed_num];
2795 if (!test_bit(R5_ReWrite, &dev->flags)) {
2796 set_bit(R5_Wantwrite, &dev->flags);
2797 set_bit(R5_ReWrite, &dev->flags);
2798 set_bit(R5_LOCKED, &dev->flags);
2801 /* let's read it back */
2802 set_bit(R5_Wantread, &dev->flags);
2803 set_bit(R5_LOCKED, &dev->flags);
2808 /* Finish postxor operations initiated by the expansion
2811 if (test_bit(STRIPE_OP_POSTXOR, &sh->ops.complete) &&
2812 !test_bit(STRIPE_OP_BIODRAIN, &sh->ops.pending)) {
2814 clear_bit(STRIPE_EXPANDING, &sh->state);
2816 clear_bit(STRIPE_OP_POSTXOR, &sh->ops.pending);
2817 clear_bit(STRIPE_OP_POSTXOR, &sh->ops.ack);
2818 clear_bit(STRIPE_OP_POSTXOR, &sh->ops.complete);
2820 for (i = conf->raid_disks; i--; )
2821 set_bit(R5_Wantwrite, &sh->dev[i].flags);
2822 set_bit(R5_LOCKED, &dev->flags);
2826 if (s.expanded && test_bit(STRIPE_EXPANDING, &sh->state) &&
2827 !test_bit(STRIPE_OP_POSTXOR, &sh->ops.pending)) {
2828 /* Need to write out all blocks after computing parity */
2829 sh->disks = conf->raid_disks;
2830 sh->pd_idx = stripe_to_pdidx(sh->sector, conf,
2832 s.locked += handle_write_operations5(sh, 1, 1);
2833 } else if (s.expanded &&
2835 !test_bit(STRIPE_OP_POSTXOR, &sh->ops.pending)) {
2836 clear_bit(STRIPE_EXPAND_READY, &sh->state);
2837 atomic_dec(&conf->reshape_stripes);
2838 wake_up(&conf->wait_for_overlap);
2839 md_done_sync(conf->mddev, STRIPE_SECTORS, 1);
2842 if (s.expanding && s.locked == 0 &&
2843 !test_bit(STRIPE_COMPUTE_RUN, &sh->state))
2844 handle_stripe_expansion(conf, sh, NULL);
2847 pending = get_stripe_work(sh);
2850 spin_unlock(&sh->lock);
2852 /* wait for this device to become unblocked */
2853 if (unlikely(blocked_rdev))
2854 md_wait_for_blocked_rdev(blocked_rdev, conf->mddev);
2856 if (pending || s.ops_request)
2857 raid5_run_ops(sh, pending, s.ops_request);
2861 return_io(return_bi);
2864 static void handle_stripe6(struct stripe_head *sh, struct page *tmp_page)
2866 raid6_conf_t *conf = sh->raid_conf;
2867 int disks = sh->disks;
2868 struct bio *return_bi = NULL;
2869 int i, pd_idx = sh->pd_idx;
2870 struct stripe_head_state s;
2871 struct r6_state r6s;
2872 struct r5dev *dev, *pdev, *qdev;
2873 mdk_rdev_t *blocked_rdev = NULL;
2875 r6s.qd_idx = raid6_next_disk(pd_idx, disks);
2876 pr_debug("handling stripe %llu, state=%#lx cnt=%d, "
2877 "pd_idx=%d, qd_idx=%d\n",
2878 (unsigned long long)sh->sector, sh->state,
2879 atomic_read(&sh->count), pd_idx, r6s.qd_idx);
2880 memset(&s, 0, sizeof(s));
2882 spin_lock(&sh->lock);
2883 clear_bit(STRIPE_HANDLE, &sh->state);
2884 clear_bit(STRIPE_DELAYED, &sh->state);
2886 s.syncing = test_bit(STRIPE_SYNCING, &sh->state);
2887 s.expanding = test_bit(STRIPE_EXPAND_SOURCE, &sh->state);
2888 s.expanded = test_bit(STRIPE_EXPAND_READY, &sh->state);
2889 /* Now to look around and see what can be done */
2892 for (i=disks; i--; ) {
2895 clear_bit(R5_Insync, &dev->flags);
2897 pr_debug("check %d: state 0x%lx read %p write %p written %p\n",
2898 i, dev->flags, dev->toread, dev->towrite, dev->written);
2899 /* maybe we can reply to a read */
2900 if (test_bit(R5_UPTODATE, &dev->flags) && dev->toread) {
2901 struct bio *rbi, *rbi2;
2902 pr_debug("Return read for disc %d\n", i);
2903 spin_lock_irq(&conf->device_lock);
2906 if (test_and_clear_bit(R5_Overlap, &dev->flags))
2907 wake_up(&conf->wait_for_overlap);
2908 spin_unlock_irq(&conf->device_lock);
2909 while (rbi && rbi->bi_sector < dev->sector + STRIPE_SECTORS) {
2910 copy_data(0, rbi, dev->page, dev->sector);
2911 rbi2 = r5_next_bio(rbi, dev->sector);
2912 spin_lock_irq(&conf->device_lock);
2913 if (--rbi->bi_phys_segments == 0) {
2914 rbi->bi_next = return_bi;
2917 spin_unlock_irq(&conf->device_lock);
2922 /* now count some things */
2923 if (test_bit(R5_LOCKED, &dev->flags)) s.locked++;
2924 if (test_bit(R5_UPTODATE, &dev->flags)) s.uptodate++;
2931 if (!test_bit(R5_OVERWRITE, &dev->flags))
2936 rdev = rcu_dereference(conf->disks[i].rdev);
2937 if (rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
2938 blocked_rdev = rdev;
2939 atomic_inc(&rdev->nr_pending);
2942 if (!rdev || !test_bit(In_sync, &rdev->flags)) {
2943 /* The ReadError flag will just be confusing now */
2944 clear_bit(R5_ReadError, &dev->flags);
2945 clear_bit(R5_ReWrite, &dev->flags);
2947 if (!rdev || !test_bit(In_sync, &rdev->flags)
2948 || test_bit(R5_ReadError, &dev->flags)) {
2950 r6s.failed_num[s.failed] = i;
2953 set_bit(R5_Insync, &dev->flags);
2957 if (unlikely(blocked_rdev)) {
2958 set_bit(STRIPE_HANDLE, &sh->state);
2961 pr_debug("locked=%d uptodate=%d to_read=%d"
2962 " to_write=%d failed=%d failed_num=%d,%d\n",
2963 s.locked, s.uptodate, s.to_read, s.to_write, s.failed,
2964 r6s.failed_num[0], r6s.failed_num[1]);
2965 /* check if the array has lost >2 devices and, if so, some requests
2966 * might need to be failed
2968 if (s.failed > 2 && s.to_read+s.to_write+s.written)
2969 handle_requests_to_failed_array(conf, sh, &s, disks,
2971 if (s.failed > 2 && s.syncing) {
2972 md_done_sync(conf->mddev, STRIPE_SECTORS,0);
2973 clear_bit(STRIPE_SYNCING, &sh->state);
2978 * might be able to return some write requests if the parity blocks
2979 * are safe, or on a failed drive
2981 pdev = &sh->dev[pd_idx];
2982 r6s.p_failed = (s.failed >= 1 && r6s.failed_num[0] == pd_idx)
2983 || (s.failed >= 2 && r6s.failed_num[1] == pd_idx);
2984 qdev = &sh->dev[r6s.qd_idx];
2985 r6s.q_failed = (s.failed >= 1 && r6s.failed_num[0] == r6s.qd_idx)
2986 || (s.failed >= 2 && r6s.failed_num[1] == r6s.qd_idx);
2989 ( r6s.p_failed || ((test_bit(R5_Insync, &pdev->flags)
2990 && !test_bit(R5_LOCKED, &pdev->flags)
2991 && test_bit(R5_UPTODATE, &pdev->flags)))) &&
2992 ( r6s.q_failed || ((test_bit(R5_Insync, &qdev->flags)
2993 && !test_bit(R5_LOCKED, &qdev->flags)
2994 && test_bit(R5_UPTODATE, &qdev->flags)))))
2995 handle_completed_write_requests(conf, sh, disks, &return_bi);
2997 /* Now we might consider reading some blocks, either to check/generate
2998 * parity, or to satisfy requests
2999 * or to load a block that is being partially written.
3001 if (s.to_read || s.non_overwrite || (s.to_write && s.failed) ||
3002 (s.syncing && (s.uptodate < disks)) || s.expanding)
3003 handle_issuing_new_read_requests6(sh, &s, &r6s, disks);
3005 /* now to consider writing and what else, if anything should be read */
3007 handle_issuing_new_write_requests6(conf, sh, &s, &r6s, disks);
3009 /* maybe we need to check and possibly fix the parity for this stripe
3010 * Any reads will already have been scheduled, so we just see if enough
3013 if (s.syncing && s.locked == 0 && !test_bit(STRIPE_INSYNC, &sh->state))
3014 handle_parity_checks6(conf, sh, &s, &r6s, tmp_page, disks);
3016 if (s.syncing && s.locked == 0 && test_bit(STRIPE_INSYNC, &sh->state)) {
3017 md_done_sync(conf->mddev, STRIPE_SECTORS,1);
3018 clear_bit(STRIPE_SYNCING, &sh->state);
3021 /* If the failed drives are just a ReadError, then we might need
3022 * to progress the repair/check process
3024 if (s.failed <= 2 && !conf->mddev->ro)
3025 for (i = 0; i < s.failed; i++) {
3026 dev = &sh->dev[r6s.failed_num[i]];
3027 if (test_bit(R5_ReadError, &dev->flags)
3028 && !test_bit(R5_LOCKED, &dev->flags)
3029 && test_bit(R5_UPTODATE, &dev->flags)
3031 if (!test_bit(R5_ReWrite, &dev->flags)) {
3032 set_bit(R5_Wantwrite, &dev->flags);
3033 set_bit(R5_ReWrite, &dev->flags);
3034 set_bit(R5_LOCKED, &dev->flags);
3036 /* let's read it back */
3037 set_bit(R5_Wantread, &dev->flags);
3038 set_bit(R5_LOCKED, &dev->flags);
3043 if (s.expanded && test_bit(STRIPE_EXPANDING, &sh->state)) {
3044 /* Need to write out all blocks after computing P&Q */
3045 sh->disks = conf->raid_disks;
3046 sh->pd_idx = stripe_to_pdidx(sh->sector, conf,
3048 compute_parity6(sh, RECONSTRUCT_WRITE);
3049 for (i = conf->raid_disks ; i-- ; ) {
3050 set_bit(R5_LOCKED, &sh->dev[i].flags);
3052 set_bit(R5_Wantwrite, &sh->dev[i].flags);
3054 clear_bit(STRIPE_EXPANDING, &sh->state);
3055 } else if (s.expanded) {
3056 clear_bit(STRIPE_EXPAND_READY, &sh->state);
3057 atomic_dec(&conf->reshape_stripes);
3058 wake_up(&conf->wait_for_overlap);
3059 md_done_sync(conf->mddev, STRIPE_SECTORS, 1);
3062 if (s.expanding && s.locked == 0 &&
3063 !test_bit(STRIPE_COMPUTE_RUN, &sh->state))
3064 handle_stripe_expansion(conf, sh, &r6s);
3067 spin_unlock(&sh->lock);
3069 /* wait for this device to become unblocked */
3070 if (unlikely(blocked_rdev))
3071 md_wait_for_blocked_rdev(blocked_rdev, conf->mddev);
3075 return_io(return_bi);
3078 static void handle_stripe(struct stripe_head *sh, struct page *tmp_page)
3080 if (sh->raid_conf->level == 6)
3081 handle_stripe6(sh, tmp_page);
3088 static void raid5_activate_delayed(raid5_conf_t *conf)
3090 if (atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD) {
3091 while (!list_empty(&conf->delayed_list)) {
3092 struct list_head *l = conf->delayed_list.next;
3093 struct stripe_head *sh;
3094 sh = list_entry(l, struct stripe_head, lru);
3096 clear_bit(STRIPE_DELAYED, &sh->state);
3097 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
3098 atomic_inc(&conf->preread_active_stripes);
3099 list_add_tail(&sh->lru, &conf->hold_list);
3102 blk_plug_device(conf->mddev->queue);
3105 static void activate_bit_delay(raid5_conf_t *conf)
3107 /* device_lock is held */
3108 struct list_head head;
3109 list_add(&head, &conf->bitmap_list);
3110 list_del_init(&conf->bitmap_list);
3111 while (!list_empty(&head)) {
3112 struct stripe_head *sh = list_entry(head.next, struct stripe_head, lru);
3113 list_del_init(&sh->lru);
3114 atomic_inc(&sh->count);
3115 __release_stripe(conf, sh);
3119 static void unplug_slaves(mddev_t *mddev)
3121 raid5_conf_t *conf = mddev_to_conf(mddev);
3125 for (i=0; i<mddev->raid_disks; i++) {
3126 mdk_rdev_t *rdev = rcu_dereference(conf->disks[i].rdev);
3127 if (rdev && !test_bit(Faulty, &rdev->flags) && atomic_read(&rdev->nr_pending)) {
3128 struct request_queue *r_queue = bdev_get_queue(rdev->bdev);
3130 atomic_inc(&rdev->nr_pending);
3133 blk_unplug(r_queue);
3135 rdev_dec_pending(rdev, mddev);
3142 static void raid5_unplug_device(struct request_queue *q)
3144 mddev_t *mddev = q->queuedata;
3145 raid5_conf_t *conf = mddev_to_conf(mddev);
3146 unsigned long flags;
3148 spin_lock_irqsave(&conf->device_lock, flags);
3150 if (blk_remove_plug(q)) {
3152 raid5_activate_delayed(conf);
3154 md_wakeup_thread(mddev->thread);
3156 spin_unlock_irqrestore(&conf->device_lock, flags);
3158 unplug_slaves(mddev);
3161 static int raid5_congested(void *data, int bits)
3163 mddev_t *mddev = data;
3164 raid5_conf_t *conf = mddev_to_conf(mddev);
3166 /* No difference between reads and writes. Just check
3167 * how busy the stripe_cache is
3169 if (conf->inactive_blocked)
3173 if (list_empty_careful(&conf->inactive_list))
3179 /* We want read requests to align with chunks where possible,
3180 * but write requests don't need to.
3182 static int raid5_mergeable_bvec(struct request_queue *q, struct bio *bio, struct bio_vec *biovec)
3184 mddev_t *mddev = q->queuedata;
3185 sector_t sector = bio->bi_sector + get_start_sect(bio->bi_bdev);
3187 unsigned int chunk_sectors = mddev->chunk_size >> 9;
3188 unsigned int bio_sectors = bio->bi_size >> 9;
3190 if (bio_data_dir(bio) == WRITE)
3191 return biovec->bv_len; /* always allow writes to be mergeable */
3193 max = (chunk_sectors - ((sector & (chunk_sectors - 1)) + bio_sectors)) << 9;
3194 if (max < 0) max = 0;
3195 if (max <= biovec->bv_len && bio_sectors == 0)
3196 return biovec->bv_len;
3202 static int in_chunk_boundary(mddev_t *mddev, struct bio *bio)
3204 sector_t sector = bio->bi_sector + get_start_sect(bio->bi_bdev);
3205 unsigned int chunk_sectors = mddev->chunk_size >> 9;
3206 unsigned int bio_sectors = bio->bi_size >> 9;
3208 return chunk_sectors >=
3209 ((sector & (chunk_sectors - 1)) + bio_sectors);
3213 * add bio to the retry LIFO ( in O(1) ... we are in interrupt )
3214 * later sampled by raid5d.
3216 static void add_bio_to_retry(struct bio *bi,raid5_conf_t *conf)
3218 unsigned long flags;
3220 spin_lock_irqsave(&conf->device_lock, flags);
3222 bi->bi_next = conf->retry_read_aligned_list;
3223 conf->retry_read_aligned_list = bi;
3225 spin_unlock_irqrestore(&conf->device_lock, flags);
3226 md_wakeup_thread(conf->mddev->thread);
3230 static struct bio *remove_bio_from_retry(raid5_conf_t *conf)
3234 bi = conf->retry_read_aligned;
3236 conf->retry_read_aligned = NULL;
3239 bi = conf->retry_read_aligned_list;
3241 conf->retry_read_aligned_list = bi->bi_next;
3243 bi->bi_phys_segments = 1; /* biased count of active stripes */
3244 bi->bi_hw_segments = 0; /* count of processed stripes */
3252 * The "raid5_align_endio" should check if the read succeeded and if it
3253 * did, call bio_endio on the original bio (having bio_put the new bio
3255 * If the read failed..
3257 static void raid5_align_endio(struct bio *bi, int error)
3259 struct bio* raid_bi = bi->bi_private;
3262 int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
3267 mddev = raid_bi->bi_bdev->bd_disk->queue->queuedata;
3268 conf = mddev_to_conf(mddev);
3269 rdev = (void*)raid_bi->bi_next;
3270 raid_bi->bi_next = NULL;
3272 rdev_dec_pending(rdev, conf->mddev);
3274 if (!error && uptodate) {
3275 bio_endio(raid_bi, 0);
3276 if (atomic_dec_and_test(&conf->active_aligned_reads))
3277 wake_up(&conf->wait_for_stripe);
3282 pr_debug("raid5_align_endio : io error...handing IO for a retry\n");
3284 add_bio_to_retry(raid_bi, conf);
3287 static int bio_fits_rdev(struct bio *bi)
3289 struct request_queue *q = bdev_get_queue(bi->bi_bdev);
3291 if ((bi->bi_size>>9) > q->max_sectors)
3293 blk_recount_segments(q, bi);
3294 if (bi->bi_phys_segments > q->max_phys_segments ||
3295 bi->bi_hw_segments > q->max_hw_segments)
3298 if (q->merge_bvec_fn)
3299 /* it's too hard to apply the merge_bvec_fn at this stage,
3308 static int chunk_aligned_read(struct request_queue *q, struct bio * raid_bio)
3310 mddev_t *mddev = q->queuedata;
3311 raid5_conf_t *conf = mddev_to_conf(mddev);
3312 const unsigned int raid_disks = conf->raid_disks;
3313 const unsigned int data_disks = raid_disks - conf->max_degraded;
3314 unsigned int dd_idx, pd_idx;
3315 struct bio* align_bi;
3318 if (!in_chunk_boundary(mddev, raid_bio)) {
3319 pr_debug("chunk_aligned_read : non aligned\n");
3323 * use bio_clone to make a copy of the bio
3325 align_bi = bio_clone(raid_bio, GFP_NOIO);
3329 * set bi_end_io to a new function, and set bi_private to the
3332 align_bi->bi_end_io = raid5_align_endio;
3333 align_bi->bi_private = raid_bio;
3337 align_bi->bi_sector = raid5_compute_sector(raid_bio->bi_sector,
3345 rdev = rcu_dereference(conf->disks[dd_idx].rdev);
3346 if (rdev && test_bit(In_sync, &rdev->flags)) {
3347 atomic_inc(&rdev->nr_pending);
3349 raid_bio->bi_next = (void*)rdev;
3350 align_bi->bi_bdev = rdev->bdev;
3351 align_bi->bi_flags &= ~(1 << BIO_SEG_VALID);
3352 align_bi->bi_sector += rdev->data_offset;
3354 if (!bio_fits_rdev(align_bi)) {
3355 /* too big in some way */
3357 rdev_dec_pending(rdev, mddev);
3361 spin_lock_irq(&conf->device_lock);
3362 wait_event_lock_irq(conf->wait_for_stripe,
3364 conf->device_lock, /* nothing */);
3365 atomic_inc(&conf->active_aligned_reads);
3366 spin_unlock_irq(&conf->device_lock);
3368 generic_make_request(align_bi);
3377 /* __get_priority_stripe - get the next stripe to process
3379 * Full stripe writes are allowed to pass preread active stripes up until
3380 * the bypass_threshold is exceeded. In general the bypass_count
3381 * increments when the handle_list is handled before the hold_list; however, it
3382 * will not be incremented when STRIPE_IO_STARTED is sampled set signifying a
3383 * stripe with in flight i/o. The bypass_count will be reset when the
3384 * head of the hold_list has changed, i.e. the head was promoted to the
3387 static struct stripe_head *__get_priority_stripe(raid5_conf_t *conf)
3389 struct stripe_head *sh;
3391 pr_debug("%s: handle: %s hold: %s full_writes: %d bypass_count: %d\n",
3393 list_empty(&conf->handle_list) ? "empty" : "busy",
3394 list_empty(&conf->hold_list) ? "empty" : "busy",
3395 atomic_read(&conf->pending_full_writes), conf->bypass_count);
3397 if (!list_empty(&conf->handle_list)) {
3398 sh = list_entry(conf->handle_list.next, typeof(*sh), lru);
3400 if (list_empty(&conf->hold_list))
3401 conf->bypass_count = 0;
3402 else if (!test_bit(STRIPE_IO_STARTED, &sh->state)) {
3403 if (conf->hold_list.next == conf->last_hold)
3404 conf->bypass_count++;
3406 conf->last_hold = conf->hold_list.next;
3407 conf->bypass_count -= conf->bypass_threshold;
3408 if (conf->bypass_count < 0)
3409 conf->bypass_count = 0;
3412 } else if (!list_empty(&conf->hold_list) &&
3413 ((conf->bypass_threshold &&
3414 conf->bypass_count > conf->bypass_threshold) ||
3415 atomic_read(&conf->pending_full_writes) == 0)) {
3416 sh = list_entry(conf->hold_list.next,
3418 conf->bypass_count -= conf->bypass_threshold;
3419 if (conf->bypass_count < 0)
3420 conf->bypass_count = 0;
3424 list_del_init(&sh->lru);
3425 atomic_inc(&sh->count);
3426 BUG_ON(atomic_read(&sh->count) != 1);
3430 static int make_request(struct request_queue *q, struct bio * bi)
3432 mddev_t *mddev = q->queuedata;
3433 raid5_conf_t *conf = mddev_to_conf(mddev);
3434 unsigned int dd_idx, pd_idx;
3435 sector_t new_sector;
3436 sector_t logical_sector, last_sector;
3437 struct stripe_head *sh;
3438 const int rw = bio_data_dir(bi);
3441 if (unlikely(bio_barrier(bi))) {
3442 bio_endio(bi, -EOPNOTSUPP);
3446 md_write_start(mddev, bi);
3448 disk_stat_inc(mddev->gendisk, ios[rw]);
3449 disk_stat_add(mddev->gendisk, sectors[rw], bio_sectors(bi));
3452 mddev->reshape_position == MaxSector &&
3453 chunk_aligned_read(q,bi))
3456 logical_sector = bi->bi_sector & ~((sector_t)STRIPE_SECTORS-1);
3457 last_sector = bi->bi_sector + (bi->bi_size>>9);
3459 bi->bi_phys_segments = 1; /* over-loaded to count active stripes */
3461 for (;logical_sector < last_sector; logical_sector += STRIPE_SECTORS) {
3463 int disks, data_disks;
3466 prepare_to_wait(&conf->wait_for_overlap, &w, TASK_UNINTERRUPTIBLE);
3467 if (likely(conf->expand_progress == MaxSector))
3468 disks = conf->raid_disks;
3470 /* spinlock is needed as expand_progress may be
3471 * 64bit on a 32bit platform, and so it might be
3472 * possible to see a half-updated value
3473 * Ofcourse expand_progress could change after
3474 * the lock is dropped, so once we get a reference
3475 * to the stripe that we think it is, we will have
3478 spin_lock_irq(&conf->device_lock);
3479 disks = conf->raid_disks;
3480 if (logical_sector >= conf->expand_progress)
3481 disks = conf->previous_raid_disks;
3483 if (logical_sector >= conf->expand_lo) {
3484 spin_unlock_irq(&conf->device_lock);
3489 spin_unlock_irq(&conf->device_lock);
3491 data_disks = disks - conf->max_degraded;
3493 new_sector = raid5_compute_sector(logical_sector, disks, data_disks,
3494 &dd_idx, &pd_idx, conf);
3495 pr_debug("raid5: make_request, sector %llu logical %llu\n",
3496 (unsigned long long)new_sector,
3497 (unsigned long long)logical_sector);
3499 sh = get_active_stripe(conf, new_sector, disks, pd_idx, (bi->bi_rw&RWA_MASK));
3501 if (unlikely(conf->expand_progress != MaxSector)) {
3502 /* expansion might have moved on while waiting for a
3503 * stripe, so we must do the range check again.
3504 * Expansion could still move past after this
3505 * test, but as we are holding a reference to
3506 * 'sh', we know that if that happens,
3507 * STRIPE_EXPANDING will get set and the expansion
3508 * won't proceed until we finish with the stripe.
3511 spin_lock_irq(&conf->device_lock);
3512 if (logical_sector < conf->expand_progress &&
3513 disks == conf->previous_raid_disks)
3514 /* mismatch, need to try again */
3516 spin_unlock_irq(&conf->device_lock);
3522 /* FIXME what if we get a false positive because these
3523 * are being updated.
3525 if (logical_sector >= mddev->suspend_lo &&
3526 logical_sector < mddev->suspend_hi) {
3532 if (test_bit(STRIPE_EXPANDING, &sh->state) ||
3533 !add_stripe_bio(sh, bi, dd_idx, (bi->bi_rw&RW_MASK))) {
3534 /* Stripe is busy expanding or
3535 * add failed due to overlap. Flush everything
3538 raid5_unplug_device(mddev->queue);
3543 finish_wait(&conf->wait_for_overlap, &w);
3544 set_bit(STRIPE_HANDLE, &sh->state);
3545 clear_bit(STRIPE_DELAYED, &sh->state);
3548 /* cannot get stripe for read-ahead, just give-up */
3549 clear_bit(BIO_UPTODATE, &bi->bi_flags);
3550 finish_wait(&conf->wait_for_overlap, &w);
3555 spin_lock_irq(&conf->device_lock);
3556 remaining = --bi->bi_phys_segments;
3557 spin_unlock_irq(&conf->device_lock);
3558 if (remaining == 0) {
3561 md_write_end(mddev);
3568 static sector_t reshape_request(mddev_t *mddev, sector_t sector_nr, int *skipped)
3570 /* reshaping is quite different to recovery/resync so it is
3571 * handled quite separately ... here.
3573 * On each call to sync_request, we gather one chunk worth of
3574 * destination stripes and flag them as expanding.
3575 * Then we find all the source stripes and request reads.
3576 * As the reads complete, handle_stripe will copy the data
3577 * into the destination stripe and release that stripe.
3579 raid5_conf_t *conf = (raid5_conf_t *) mddev->private;
3580 struct stripe_head *sh;
3582 sector_t first_sector, last_sector;
3583 int raid_disks = conf->previous_raid_disks;
3584 int data_disks = raid_disks - conf->max_degraded;
3585 int new_data_disks = conf->raid_disks - conf->max_degraded;
3588 sector_t writepos, safepos, gap;
3590 if (sector_nr == 0 &&
3591 conf->expand_progress != 0) {
3592 /* restarting in the middle, skip the initial sectors */
3593 sector_nr = conf->expand_progress;
3594 sector_div(sector_nr, new_data_disks);
3599 /* we update the metadata when there is more than 3Meg
3600 * in the block range (that is rather arbitrary, should
3601 * probably be time based) or when the data about to be
3602 * copied would over-write the source of the data at
3603 * the front of the range.
3604 * i.e. one new_stripe forward from expand_progress new_maps
3605 * to after where expand_lo old_maps to
3607 writepos = conf->expand_progress +
3608 conf->chunk_size/512*(new_data_disks);
3609 sector_div(writepos, new_data_disks);
3610 safepos = conf->expand_lo;
3611 sector_div(safepos, data_disks);
3612 gap = conf->expand_progress - conf->expand_lo;
3614 if (writepos >= safepos ||
3615 gap > (new_data_disks)*3000*2 /*3Meg*/) {
3616 /* Cannot proceed until we've updated the superblock... */
3617 wait_event(conf->wait_for_overlap,
3618 atomic_read(&conf->reshape_stripes)==0);
3619 mddev->reshape_position = conf->expand_progress;
3620 set_bit(MD_CHANGE_DEVS, &mddev->flags);
3621 md_wakeup_thread(mddev->thread);
3622 wait_event(mddev->sb_wait, mddev->flags == 0 ||
3623 kthread_should_stop());
3624 spin_lock_irq(&conf->device_lock);
3625 conf->expand_lo = mddev->reshape_position;
3626 spin_unlock_irq(&conf->device_lock);
3627 wake_up(&conf->wait_for_overlap);
3630 for (i=0; i < conf->chunk_size/512; i+= STRIPE_SECTORS) {
3633 pd_idx = stripe_to_pdidx(sector_nr+i, conf, conf->raid_disks);
3634 sh = get_active_stripe(conf, sector_nr+i,
3635 conf->raid_disks, pd_idx, 0);
3636 set_bit(STRIPE_EXPANDING, &sh->state);
3637 atomic_inc(&conf->reshape_stripes);
3638 /* If any of this stripe is beyond the end of the old
3639 * array, then we need to zero those blocks
3641 for (j=sh->disks; j--;) {
3643 if (j == sh->pd_idx)
3645 if (conf->level == 6 &&
3646 j == raid6_next_disk(sh->pd_idx, sh->disks))
3648 s = compute_blocknr(sh, j);
3649 if (s < (mddev->array_size<<1)) {
3653 memset(page_address(sh->dev[j].page), 0, STRIPE_SIZE);
3654 set_bit(R5_Expanded, &sh->dev[j].flags);
3655 set_bit(R5_UPTODATE, &sh->dev[j].flags);
3658 set_bit(STRIPE_EXPAND_READY, &sh->state);
3659 set_bit(STRIPE_HANDLE, &sh->state);
3663 spin_lock_irq(&conf->device_lock);
3664 conf->expand_progress = (sector_nr + i) * new_data_disks;
3665 spin_unlock_irq(&conf->device_lock);
3666 /* Ok, those stripe are ready. We can start scheduling
3667 * reads on the source stripes.
3668 * The source stripes are determined by mapping the first and last
3669 * block on the destination stripes.
3672 raid5_compute_sector(sector_nr*(new_data_disks),
3673 raid_disks, data_disks,
3674 &dd_idx, &pd_idx, conf);
3676 raid5_compute_sector((sector_nr+conf->chunk_size/512)
3677 *(new_data_disks) -1,
3678 raid_disks, data_disks,
3679 &dd_idx, &pd_idx, conf);
3680 if (last_sector >= (mddev->size<<1))
3681 last_sector = (mddev->size<<1)-1;
3682 while (first_sector <= last_sector) {
3683 pd_idx = stripe_to_pdidx(first_sector, conf,
3684 conf->previous_raid_disks);
3685 sh = get_active_stripe(conf, first_sector,
3686 conf->previous_raid_disks, pd_idx, 0);
3687 set_bit(STRIPE_EXPAND_SOURCE, &sh->state);
3688 set_bit(STRIPE_HANDLE, &sh->state);
3690 first_sector += STRIPE_SECTORS;
3692 /* If this takes us to the resync_max point where we have to pause,
3693 * then we need to write out the superblock.
3695 sector_nr += conf->chunk_size>>9;
3696 if (sector_nr >= mddev->resync_max) {
3697 /* Cannot proceed until we've updated the superblock... */
3698 wait_event(conf->wait_for_overlap,
3699 atomic_read(&conf->reshape_stripes) == 0);
3700 mddev->reshape_position = conf->expand_progress;
3701 set_bit(MD_CHANGE_DEVS, &mddev->flags);
3702 md_wakeup_thread(mddev->thread);
3703 wait_event(mddev->sb_wait,
3704 !test_bit(MD_CHANGE_DEVS, &mddev->flags)
3705 || kthread_should_stop());
3706 spin_lock_irq(&conf->device_lock);
3707 conf->expand_lo = mddev->reshape_position;
3708 spin_unlock_irq(&conf->device_lock);
3709 wake_up(&conf->wait_for_overlap);
3711 return conf->chunk_size>>9;
3714 /* FIXME go_faster isn't used */
3715 static inline sector_t sync_request(mddev_t *mddev, sector_t sector_nr, int *skipped, int go_faster)
3717 raid5_conf_t *conf = (raid5_conf_t *) mddev->private;
3718 struct stripe_head *sh;
3720 int raid_disks = conf->raid_disks;
3721 sector_t max_sector = mddev->size << 1;
3723 int still_degraded = 0;
3726 if (sector_nr >= max_sector) {
3727 /* just being told to finish up .. nothing much to do */
3728 unplug_slaves(mddev);
3729 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery)) {
3734 if (mddev->curr_resync < max_sector) /* aborted */
3735 bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
3737 else /* completed sync */
3739 bitmap_close_sync(mddev->bitmap);
3744 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
3745 return reshape_request(mddev, sector_nr, skipped);
3747 /* No need to check resync_max as we never do more than one
3748 * stripe, and as resync_max will always be on a chunk boundary,
3749 * if the check in md_do_sync didn't fire, there is no chance
3750 * of overstepping resync_max here
3753 /* if there is too many failed drives and we are trying
3754 * to resync, then assert that we are finished, because there is
3755 * nothing we can do.
3757 if (mddev->degraded >= conf->max_degraded &&
3758 test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
3759 sector_t rv = (mddev->size << 1) - sector_nr;
3763 if (!bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
3764 !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
3765 !conf->fullsync && sync_blocks >= STRIPE_SECTORS) {
3766 /* we can skip this block, and probably more */
3767 sync_blocks /= STRIPE_SECTORS;
3769 return sync_blocks * STRIPE_SECTORS; /* keep things rounded to whole stripes */
3773 bitmap_cond_end_sync(mddev->bitmap, sector_nr);
3775 pd_idx = stripe_to_pdidx(sector_nr, conf, raid_disks);
3776 sh = get_active_stripe(conf, sector_nr, raid_disks, pd_idx, 1);
3778 sh = get_active_stripe(conf, sector_nr, raid_disks, pd_idx, 0);
3779 /* make sure we don't swamp the stripe cache if someone else
3780 * is trying to get access
3782 schedule_timeout_uninterruptible(1);
3784 /* Need to check if array will still be degraded after recovery/resync
3785 * We don't need to check the 'failed' flag as when that gets set,
3788 for (i=0; i<mddev->raid_disks; i++)
3789 if (conf->disks[i].rdev == NULL)
3792 bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, still_degraded);
3794 spin_lock(&sh->lock);
3795 set_bit(STRIPE_SYNCING, &sh->state);
3796 clear_bit(STRIPE_INSYNC, &sh->state);
3797 spin_unlock(&sh->lock);
3799 handle_stripe(sh, NULL);
3802 return STRIPE_SECTORS;
3805 static int retry_aligned_read(raid5_conf_t *conf, struct bio *raid_bio)
3807 /* We may not be able to submit a whole bio at once as there
3808 * may not be enough stripe_heads available.
3809 * We cannot pre-allocate enough stripe_heads as we may need
3810 * more than exist in the cache (if we allow ever large chunks).
3811 * So we do one stripe head at a time and record in
3812 * ->bi_hw_segments how many have been done.
3814 * We *know* that this entire raid_bio is in one chunk, so
3815 * it will be only one 'dd_idx' and only need one call to raid5_compute_sector.
3817 struct stripe_head *sh;
3819 sector_t sector, logical_sector, last_sector;
3824 logical_sector = raid_bio->bi_sector & ~((sector_t)STRIPE_SECTORS-1);
3825 sector = raid5_compute_sector( logical_sector,
3827 conf->raid_disks - conf->max_degraded,
3831 last_sector = raid_bio->bi_sector + (raid_bio->bi_size>>9);
3833 for (; logical_sector < last_sector;
3834 logical_sector += STRIPE_SECTORS,
3835 sector += STRIPE_SECTORS,
3838 if (scnt < raid_bio->bi_hw_segments)
3839 /* already done this stripe */
3842 sh = get_active_stripe(conf, sector, conf->raid_disks, pd_idx, 1);
3845 /* failed to get a stripe - must wait */
3846 raid_bio->bi_hw_segments = scnt;
3847 conf->retry_read_aligned = raid_bio;
3851 set_bit(R5_ReadError, &sh->dev[dd_idx].flags);
3852 if (!add_stripe_bio(sh, raid_bio, dd_idx, 0)) {
3854 raid_bio->bi_hw_segments = scnt;
3855 conf->retry_read_aligned = raid_bio;
3859 handle_stripe(sh, NULL);
3863 spin_lock_irq(&conf->device_lock);
3864 remaining = --raid_bio->bi_phys_segments;
3865 spin_unlock_irq(&conf->device_lock);
3867 bio_endio(raid_bio, 0);
3868 if (atomic_dec_and_test(&conf->active_aligned_reads))
3869 wake_up(&conf->wait_for_stripe);
3876 * This is our raid5 kernel thread.
3878 * We scan the hash table for stripes which can be handled now.
3879 * During the scan, completed stripes are saved for us by the interrupt
3880 * handler, so that they will not have to wait for our next wakeup.
3882 static void raid5d(mddev_t *mddev)
3884 struct stripe_head *sh;
3885 raid5_conf_t *conf = mddev_to_conf(mddev);
3888 pr_debug("+++ raid5d active\n");
3890 md_check_recovery(mddev);
3893 spin_lock_irq(&conf->device_lock);
3897 if (conf->seq_flush != conf->seq_write) {
3898 int seq = conf->seq_flush;
3899 spin_unlock_irq(&conf->device_lock);
3900 bitmap_unplug(mddev->bitmap);
3901 spin_lock_irq(&conf->device_lock);
3902 conf->seq_write = seq;
3903 activate_bit_delay(conf);
3906 while ((bio = remove_bio_from_retry(conf))) {
3908 spin_unlock_irq(&conf->device_lock);
3909 ok = retry_aligned_read(conf, bio);
3910 spin_lock_irq(&conf->device_lock);
3916 sh = __get_priority_stripe(conf);
3919 async_tx_issue_pending_all();
3922 spin_unlock_irq(&conf->device_lock);
3925 handle_stripe(sh, conf->spare_page);
3928 spin_lock_irq(&conf->device_lock);
3930 pr_debug("%d stripes handled\n", handled);
3932 spin_unlock_irq(&conf->device_lock);
3934 unplug_slaves(mddev);
3936 pr_debug("--- raid5d inactive\n");
3940 raid5_show_stripe_cache_size(mddev_t *mddev, char *page)
3942 raid5_conf_t *conf = mddev_to_conf(mddev);
3944 return sprintf(page, "%d\n", conf->max_nr_stripes);
3950 raid5_store_stripe_cache_size(mddev_t *mddev, const char *page, size_t len)
3952 raid5_conf_t *conf = mddev_to_conf(mddev);
3954 if (len >= PAGE_SIZE)
3959 if (strict_strtoul(page, 10, &new))
3961 if (new <= 16 || new > 32768)
3963 while (new < conf->max_nr_stripes) {
3964 if (drop_one_stripe(conf))
3965 conf->max_nr_stripes--;
3969 md_allow_write(mddev);
3970 while (new > conf->max_nr_stripes) {
3971 if (grow_one_stripe(conf))
3972 conf->max_nr_stripes++;
3978 static struct md_sysfs_entry
3979 raid5_stripecache_size = __ATTR(stripe_cache_size, S_IRUGO | S_IWUSR,
3980 raid5_show_stripe_cache_size,
3981 raid5_store_stripe_cache_size);
3984 raid5_show_preread_threshold(mddev_t *mddev, char *page)
3986 raid5_conf_t *conf = mddev_to_conf(mddev);
3988 return sprintf(page, "%d\n", conf->bypass_threshold);
3994 raid5_store_preread_threshold(mddev_t *mddev, const char *page, size_t len)
3996 raid5_conf_t *conf = mddev_to_conf(mddev);
3998 if (len >= PAGE_SIZE)
4003 if (strict_strtoul(page, 10, &new))
4005 if (new > conf->max_nr_stripes)
4007 conf->bypass_threshold = new;
4011 static struct md_sysfs_entry
4012 raid5_preread_bypass_threshold = __ATTR(preread_bypass_threshold,
4014 raid5_show_preread_threshold,
4015 raid5_store_preread_threshold);
4018 stripe_cache_active_show(mddev_t *mddev, char *page)
4020 raid5_conf_t *conf = mddev_to_conf(mddev);
4022 return sprintf(page, "%d\n", atomic_read(&conf->active_stripes));
4027 static struct md_sysfs_entry
4028 raid5_stripecache_active = __ATTR_RO(stripe_cache_active);
4030 static struct attribute *raid5_attrs[] = {
4031 &raid5_stripecache_size.attr,
4032 &raid5_stripecache_active.attr,
4033 &raid5_preread_bypass_threshold.attr,
4036 static struct attribute_group raid5_attrs_group = {
4038 .attrs = raid5_attrs,
4041 static int run(mddev_t *mddev)
4044 int raid_disk, memory;
4046 struct disk_info *disk;
4047 struct list_head *tmp;
4048 int working_disks = 0;
4050 if (mddev->level != 5 && mddev->level != 4 && mddev->level != 6) {
4051 printk(KERN_ERR "raid5: %s: raid level not set to 4/5/6 (%d)\n",
4052 mdname(mddev), mddev->level);
4056 if (mddev->reshape_position != MaxSector) {
4057 /* Check that we can continue the reshape.
4058 * Currently only disks can change, it must
4059 * increase, and we must be past the point where
4060 * a stripe over-writes itself
4062 sector_t here_new, here_old;
4064 int max_degraded = (mddev->level == 5 ? 1 : 2);
4066 if (mddev->new_level != mddev->level ||
4067 mddev->new_layout != mddev->layout ||
4068 mddev->new_chunk != mddev->chunk_size) {
4069 printk(KERN_ERR "raid5: %s: unsupported reshape "
4070 "required - aborting.\n",
4074 if (mddev->delta_disks <= 0) {
4075 printk(KERN_ERR "raid5: %s: unsupported reshape "
4076 "(reduce disks) required - aborting.\n",
4080 old_disks = mddev->raid_disks - mddev->delta_disks;
4081 /* reshape_position must be on a new-stripe boundary, and one
4082 * further up in new geometry must map after here in old
4085 here_new = mddev->reshape_position;
4086 if (sector_div(here_new, (mddev->chunk_size>>9)*
4087 (mddev->raid_disks - max_degraded))) {
4088 printk(KERN_ERR "raid5: reshape_position not "
4089 "on a stripe boundary\n");
4092 /* here_new is the stripe we will write to */
4093 here_old = mddev->reshape_position;
4094 sector_div(here_old, (mddev->chunk_size>>9)*
4095 (old_disks-max_degraded));
4096 /* here_old is the first stripe that we might need to read
4098 if (here_new >= here_old) {
4099 /* Reading from the same stripe as writing to - bad */
4100 printk(KERN_ERR "raid5: reshape_position too early for "
4101 "auto-recovery - aborting.\n");
4104 printk(KERN_INFO "raid5: reshape will continue\n");
4105 /* OK, we should be able to continue; */
4109 mddev->private = kzalloc(sizeof (raid5_conf_t), GFP_KERNEL);
4110 if ((conf = mddev->private) == NULL)
4112 if (mddev->reshape_position == MaxSector) {
4113 conf->previous_raid_disks = conf->raid_disks = mddev->raid_disks;
4115 conf->raid_disks = mddev->raid_disks;
4116 conf->previous_raid_disks = mddev->raid_disks - mddev->delta_disks;
4119 conf->disks = kzalloc(conf->raid_disks * sizeof(struct disk_info),
4124 conf->mddev = mddev;
4126 if ((conf->stripe_hashtbl = kzalloc(PAGE_SIZE, GFP_KERNEL)) == NULL)
4129 if (mddev->level == 6) {
4130 conf->spare_page = alloc_page(GFP_KERNEL);
4131 if (!conf->spare_page)
4134 spin_lock_init(&conf->device_lock);
4135 mddev->queue->queue_lock = &conf->device_lock;
4136 init_waitqueue_head(&conf->wait_for_stripe);
4137 init_waitqueue_head(&conf->wait_for_overlap);
4138 INIT_LIST_HEAD(&conf->handle_list);
4139 INIT_LIST_HEAD(&conf->hold_list);
4140 INIT_LIST_HEAD(&conf->delayed_list);
4141 INIT_LIST_HEAD(&conf->bitmap_list);
4142 INIT_LIST_HEAD(&conf->inactive_list);
4143 atomic_set(&conf->active_stripes, 0);
4144 atomic_set(&conf->preread_active_stripes, 0);
4145 atomic_set(&conf->active_aligned_reads, 0);
4146 conf->bypass_threshold = BYPASS_THRESHOLD;
4148 pr_debug("raid5: run(%s) called.\n", mdname(mddev));
4150 rdev_for_each(rdev, tmp, mddev) {
4151 raid_disk = rdev->raid_disk;
4152 if (raid_disk >= conf->raid_disks
4155 disk = conf->disks + raid_disk;
4159 if (test_bit(In_sync, &rdev->flags)) {
4160 char b[BDEVNAME_SIZE];
4161 printk(KERN_INFO "raid5: device %s operational as raid"
4162 " disk %d\n", bdevname(rdev->bdev,b),
4166 /* Cannot rely on bitmap to complete recovery */
4171 * 0 for a fully functional array, 1 or 2 for a degraded array.
4173 mddev->degraded = conf->raid_disks - working_disks;
4174 conf->mddev = mddev;
4175 conf->chunk_size = mddev->chunk_size;
4176 conf->level = mddev->level;
4177 if (conf->level == 6)
4178 conf->max_degraded = 2;
4180 conf->max_degraded = 1;
4181 conf->algorithm = mddev->layout;
4182 conf->max_nr_stripes = NR_STRIPES;
4183 conf->expand_progress = mddev->reshape_position;
4185 /* device size must be a multiple of chunk size */
4186 mddev->size &= ~(mddev->chunk_size/1024 -1);
4187 mddev->resync_max_sectors = mddev->size << 1;
4189 if (conf->level == 6 && conf->raid_disks < 4) {
4190 printk(KERN_ERR "raid6: not enough configured devices for %s (%d, minimum 4)\n",
4191 mdname(mddev), conf->raid_disks);
4194 if (!conf->chunk_size || conf->chunk_size % 4) {
4195 printk(KERN_ERR "raid5: invalid chunk size %d for %s\n",
4196 conf->chunk_size, mdname(mddev));
4199 if (conf->algorithm > ALGORITHM_RIGHT_SYMMETRIC) {
4201 "raid5: unsupported parity algorithm %d for %s\n",
4202 conf->algorithm, mdname(mddev));
4205 if (mddev->degraded > conf->max_degraded) {
4206 printk(KERN_ERR "raid5: not enough operational devices for %s"
4207 " (%d/%d failed)\n",
4208 mdname(mddev), mddev->degraded, conf->raid_disks);
4212 if (mddev->degraded > 0 &&
4213 mddev->recovery_cp != MaxSector) {
4214 if (mddev->ok_start_degraded)
4216 "raid5: starting dirty degraded array: %s"
4217 "- data corruption possible.\n",
4221 "raid5: cannot start dirty degraded array for %s\n",
4228 mddev->thread = md_register_thread(raid5d, mddev, "%s_raid5");
4229 if (!mddev->thread) {
4231 "raid5: couldn't allocate thread for %s\n",
4236 memory = conf->max_nr_stripes * (sizeof(struct stripe_head) +
4237 conf->raid_disks * ((sizeof(struct bio) + PAGE_SIZE))) / 1024;
4238 if (grow_stripes(conf, conf->max_nr_stripes)) {
4240 "raid5: couldn't allocate %dkB for buffers\n", memory);
4241 shrink_stripes(conf);
4242 md_unregister_thread(mddev->thread);
4245 printk(KERN_INFO "raid5: allocated %dkB for %s\n",
4246 memory, mdname(mddev));
4248 if (mddev->degraded == 0)
4249 printk("raid5: raid level %d set %s active with %d out of %d"
4250 " devices, algorithm %d\n", conf->level, mdname(mddev),
4251 mddev->raid_disks-mddev->degraded, mddev->raid_disks,
4254 printk(KERN_ALERT "raid5: raid level %d set %s active with %d"
4255 " out of %d devices, algorithm %d\n", conf->level,
4256 mdname(mddev), mddev->raid_disks - mddev->degraded,
4257 mddev->raid_disks, conf->algorithm);
4259 print_raid5_conf(conf);
4261 if (conf->expand_progress != MaxSector) {
4262 printk("...ok start reshape thread\n");
4263 conf->expand_lo = conf->expand_progress;
4264 atomic_set(&conf->reshape_stripes, 0);
4265 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
4266 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
4267 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
4268 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
4269 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
4273 /* read-ahead size must cover two whole stripes, which is
4274 * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
4277 int data_disks = conf->previous_raid_disks - conf->max_degraded;
4278 int stripe = data_disks *
4279 (mddev->chunk_size / PAGE_SIZE);
4280 if (mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
4281 mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
4284 /* Ok, everything is just fine now */
4285 if (sysfs_create_group(&mddev->kobj, &raid5_attrs_group))
4287 "raid5: failed to create sysfs attributes for %s\n",
4290 mddev->queue->unplug_fn = raid5_unplug_device;
4291 mddev->queue->backing_dev_info.congested_data = mddev;
4292 mddev->queue->backing_dev_info.congested_fn = raid5_congested;
4294 mddev->array_size = mddev->size * (conf->previous_raid_disks -
4295 conf->max_degraded);
4297 blk_queue_merge_bvec(mddev->queue, raid5_mergeable_bvec);
4302 print_raid5_conf(conf);
4303 safe_put_page(conf->spare_page);
4305 kfree(conf->stripe_hashtbl);
4308 mddev->private = NULL;
4309 printk(KERN_ALERT "raid5: failed to run raid set %s\n", mdname(mddev));
4315 static int stop(mddev_t *mddev)
4317 raid5_conf_t *conf = (raid5_conf_t *) mddev->private;
4319 md_unregister_thread(mddev->thread);
4320 mddev->thread = NULL;
4321 shrink_stripes(conf);
4322 kfree(conf->stripe_hashtbl);
4323 mddev->queue->backing_dev_info.congested_fn = NULL;
4324 blk_sync_queue(mddev->queue); /* the unplug fn references 'conf'*/
4325 sysfs_remove_group(&mddev->kobj, &raid5_attrs_group);
4328 mddev->private = NULL;
4333 static void print_sh (struct seq_file *seq, struct stripe_head *sh)
4337 seq_printf(seq, "sh %llu, pd_idx %d, state %ld.\n",
4338 (unsigned long long)sh->sector, sh->pd_idx, sh->state);
4339 seq_printf(seq, "sh %llu, count %d.\n",
4340 (unsigned long long)sh->sector, atomic_read(&sh->count));
4341 seq_printf(seq, "sh %llu, ", (unsigned long long)sh->sector);
4342 for (i = 0; i < sh->disks; i++) {
4343 seq_printf(seq, "(cache%d: %p %ld) ",
4344 i, sh->dev[i].page, sh->dev[i].flags);
4346 seq_printf(seq, "\n");
4349 static void printall (struct seq_file *seq, raid5_conf_t *conf)
4351 struct stripe_head *sh;
4352 struct hlist_node *hn;
4355 spin_lock_irq(&conf->device_lock);
4356 for (i = 0; i < NR_HASH; i++) {
4357 hlist_for_each_entry(sh, hn, &conf->stripe_hashtbl[i], hash) {
4358 if (sh->raid_conf != conf)
4363 spin_unlock_irq(&conf->device_lock);
4367 static void status (struct seq_file *seq, mddev_t *mddev)
4369 raid5_conf_t *conf = (raid5_conf_t *) mddev->private;
4372 seq_printf (seq, " level %d, %dk chunk, algorithm %d", mddev->level, mddev->chunk_size >> 10, mddev->layout);
4373 seq_printf (seq, " [%d/%d] [", conf->raid_disks, conf->raid_disks - mddev->degraded);
4374 for (i = 0; i < conf->raid_disks; i++)
4375 seq_printf (seq, "%s",
4376 conf->disks[i].rdev &&
4377 test_bit(In_sync, &conf->disks[i].rdev->flags) ? "U" : "_");
4378 seq_printf (seq, "]");
4380 seq_printf (seq, "\n");
4381 printall(seq, conf);
4385 static void print_raid5_conf (raid5_conf_t *conf)
4388 struct disk_info *tmp;
4390 printk("RAID5 conf printout:\n");
4392 printk("(conf==NULL)\n");
4395 printk(" --- rd:%d wd:%d\n", conf->raid_disks,
4396 conf->raid_disks - conf->mddev->degraded);
4398 for (i = 0; i < conf->raid_disks; i++) {
4399 char b[BDEVNAME_SIZE];
4400 tmp = conf->disks + i;
4402 printk(" disk %d, o:%d, dev:%s\n",
4403 i, !test_bit(Faulty, &tmp->rdev->flags),
4404 bdevname(tmp->rdev->bdev,b));
4408 static int raid5_spare_active(mddev_t *mddev)
4411 raid5_conf_t *conf = mddev->private;
4412 struct disk_info *tmp;
4414 for (i = 0; i < conf->raid_disks; i++) {
4415 tmp = conf->disks + i;
4417 && !test_bit(Faulty, &tmp->rdev->flags)
4418 && !test_and_set_bit(In_sync, &tmp->rdev->flags)) {
4419 unsigned long flags;
4420 spin_lock_irqsave(&conf->device_lock, flags);
4422 spin_unlock_irqrestore(&conf->device_lock, flags);
4425 print_raid5_conf(conf);
4429 static int raid5_remove_disk(mddev_t *mddev, int number)
4431 raid5_conf_t *conf = mddev->private;
4434 struct disk_info *p = conf->disks + number;
4436 print_raid5_conf(conf);
4439 if (test_bit(In_sync, &rdev->flags) ||
4440 atomic_read(&rdev->nr_pending)) {
4444 /* Only remove non-faulty devices if recovery
4447 if (!test_bit(Faulty, &rdev->flags) &&
4448 mddev->degraded <= conf->max_degraded) {
4454 if (atomic_read(&rdev->nr_pending)) {
4455 /* lost the race, try later */
4462 print_raid5_conf(conf);
4466 static int raid5_add_disk(mddev_t *mddev, mdk_rdev_t *rdev)
4468 raid5_conf_t *conf = mddev->private;
4471 struct disk_info *p;
4473 int last = conf->raid_disks - 1;
4475 if (mddev->degraded > conf->max_degraded)
4476 /* no point adding a device */
4479 if (rdev->raid_disk >= 0)
4480 first = last = rdev->raid_disk;
4483 * find the disk ... but prefer rdev->saved_raid_disk
4486 if (rdev->saved_raid_disk >= 0 &&
4487 rdev->saved_raid_disk >= first &&
4488 conf->disks[rdev->saved_raid_disk].rdev == NULL)
4489 disk = rdev->saved_raid_disk;
4492 for ( ; disk <= last ; disk++)
4493 if ((p=conf->disks + disk)->rdev == NULL) {
4494 clear_bit(In_sync, &rdev->flags);
4495 rdev->raid_disk = disk;
4497 if (rdev->saved_raid_disk != disk)
4499 rcu_assign_pointer(p->rdev, rdev);
4502 print_raid5_conf(conf);
4506 static int raid5_resize(mddev_t *mddev, sector_t sectors)
4508 /* no resync is happening, and there is enough space
4509 * on all devices, so we can resize.
4510 * We need to make sure resync covers any new space.
4511 * If the array is shrinking we should possibly wait until
4512 * any io in the removed space completes, but it hardly seems
4515 raid5_conf_t *conf = mddev_to_conf(mddev);
4517 sectors &= ~((sector_t)mddev->chunk_size/512 - 1);
4518 mddev->array_size = (sectors * (mddev->raid_disks-conf->max_degraded))>>1;
4519 set_capacity(mddev->gendisk, mddev->array_size << 1);
4521 if (sectors/2 > mddev->size && mddev->recovery_cp == MaxSector) {
4522 mddev->recovery_cp = mddev->size << 1;
4523 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
4525 mddev->size = sectors /2;
4526 mddev->resync_max_sectors = sectors;
4530 #ifdef CONFIG_MD_RAID5_RESHAPE
4531 static int raid5_check_reshape(mddev_t *mddev)
4533 raid5_conf_t *conf = mddev_to_conf(mddev);
4536 if (mddev->delta_disks < 0 ||
4537 mddev->new_level != mddev->level)
4538 return -EINVAL; /* Cannot shrink array or change level yet */
4539 if (mddev->delta_disks == 0)
4540 return 0; /* nothing to do */
4542 /* Can only proceed if there are plenty of stripe_heads.
4543 * We need a minimum of one full stripe,, and for sensible progress
4544 * it is best to have about 4 times that.
4545 * If we require 4 times, then the default 256 4K stripe_heads will
4546 * allow for chunk sizes up to 256K, which is probably OK.
4547 * If the chunk size is greater, user-space should request more
4548 * stripe_heads first.
4550 if ((mddev->chunk_size / STRIPE_SIZE) * 4 > conf->max_nr_stripes ||
4551 (mddev->new_chunk / STRIPE_SIZE) * 4 > conf->max_nr_stripes) {
4552 printk(KERN_WARNING "raid5: reshape: not enough stripes. Needed %lu\n",
4553 (mddev->chunk_size / STRIPE_SIZE)*4);
4557 err = resize_stripes(conf, conf->raid_disks + mddev->delta_disks);
4561 if (mddev->degraded > conf->max_degraded)
4563 /* looks like we might be able to manage this */
4567 static int raid5_start_reshape(mddev_t *mddev)
4569 raid5_conf_t *conf = mddev_to_conf(mddev);
4571 struct list_head *rtmp;
4573 int added_devices = 0;
4574 unsigned long flags;
4576 if (test_bit(MD_RECOVERY_RUNNING, &mddev->recovery))
4579 rdev_for_each(rdev, rtmp, mddev)
4580 if (rdev->raid_disk < 0 &&
4581 !test_bit(Faulty, &rdev->flags))
4584 if (spares - mddev->degraded < mddev->delta_disks - conf->max_degraded)
4585 /* Not enough devices even to make a degraded array
4590 atomic_set(&conf->reshape_stripes, 0);
4591 spin_lock_irq(&conf->device_lock);
4592 conf->previous_raid_disks = conf->raid_disks;
4593 conf->raid_disks += mddev->delta_disks;
4594 conf->expand_progress = 0;
4595 conf->expand_lo = 0;
4596 spin_unlock_irq(&conf->device_lock);
4598 /* Add some new drives, as many as will fit.
4599 * We know there are enough to make the newly sized array work.
4601 rdev_for_each(rdev, rtmp, mddev)
4602 if (rdev->raid_disk < 0 &&
4603 !test_bit(Faulty, &rdev->flags)) {
4604 if (raid5_add_disk(mddev, rdev) == 0) {
4606 set_bit(In_sync, &rdev->flags);
4608 rdev->recovery_offset = 0;
4609 sprintf(nm, "rd%d", rdev->raid_disk);
4610 if (sysfs_create_link(&mddev->kobj,
4613 "raid5: failed to create "
4614 " link %s for %s\n",
4620 spin_lock_irqsave(&conf->device_lock, flags);
4621 mddev->degraded = (conf->raid_disks - conf->previous_raid_disks) - added_devices;
4622 spin_unlock_irqrestore(&conf->device_lock, flags);
4623 mddev->raid_disks = conf->raid_disks;
4624 mddev->reshape_position = 0;
4625 set_bit(MD_CHANGE_DEVS, &mddev->flags);
4627 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
4628 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
4629 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
4630 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
4631 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
4633 if (!mddev->sync_thread) {
4634 mddev->recovery = 0;
4635 spin_lock_irq(&conf->device_lock);
4636 mddev->raid_disks = conf->raid_disks = conf->previous_raid_disks;
4637 conf->expand_progress = MaxSector;
4638 spin_unlock_irq(&conf->device_lock);
4641 md_wakeup_thread(mddev->sync_thread);
4642 md_new_event(mddev);
4647 static void end_reshape(raid5_conf_t *conf)
4649 struct block_device *bdev;
4651 if (!test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery)) {
4652 conf->mddev->array_size = conf->mddev->size *
4653 (conf->raid_disks - conf->max_degraded);
4654 set_capacity(conf->mddev->gendisk, conf->mddev->array_size << 1);
4655 conf->mddev->changed = 1;
4657 bdev = bdget_disk(conf->mddev->gendisk, 0);
4659 mutex_lock(&bdev->bd_inode->i_mutex);
4660 i_size_write(bdev->bd_inode, (loff_t)conf->mddev->array_size << 10);
4661 mutex_unlock(&bdev->bd_inode->i_mutex);
4664 spin_lock_irq(&conf->device_lock);
4665 conf->expand_progress = MaxSector;
4666 spin_unlock_irq(&conf->device_lock);
4667 conf->mddev->reshape_position = MaxSector;
4669 /* read-ahead size must cover two whole stripes, which is
4670 * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
4673 int data_disks = conf->previous_raid_disks - conf->max_degraded;
4674 int stripe = data_disks *
4675 (conf->mddev->chunk_size / PAGE_SIZE);
4676 if (conf->mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
4677 conf->mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
4682 static void raid5_quiesce(mddev_t *mddev, int state)
4684 raid5_conf_t *conf = mddev_to_conf(mddev);
4687 case 2: /* resume for a suspend */
4688 wake_up(&conf->wait_for_overlap);
4691 case 1: /* stop all writes */
4692 spin_lock_irq(&conf->device_lock);
4694 wait_event_lock_irq(conf->wait_for_stripe,
4695 atomic_read(&conf->active_stripes) == 0 &&
4696 atomic_read(&conf->active_aligned_reads) == 0,
4697 conf->device_lock, /* nothing */);
4698 spin_unlock_irq(&conf->device_lock);
4701 case 0: /* re-enable writes */
4702 spin_lock_irq(&conf->device_lock);
4704 wake_up(&conf->wait_for_stripe);
4705 wake_up(&conf->wait_for_overlap);
4706 spin_unlock_irq(&conf->device_lock);
4711 static struct mdk_personality raid6_personality =
4715 .owner = THIS_MODULE,
4716 .make_request = make_request,
4720 .error_handler = error,
4721 .hot_add_disk = raid5_add_disk,
4722 .hot_remove_disk= raid5_remove_disk,
4723 .spare_active = raid5_spare_active,
4724 .sync_request = sync_request,
4725 .resize = raid5_resize,
4726 #ifdef CONFIG_MD_RAID5_RESHAPE
4727 .check_reshape = raid5_check_reshape,
4728 .start_reshape = raid5_start_reshape,
4730 .quiesce = raid5_quiesce,
4732 static struct mdk_personality raid5_personality =
4736 .owner = THIS_MODULE,
4737 .make_request = make_request,
4741 .error_handler = error,
4742 .hot_add_disk = raid5_add_disk,
4743 .hot_remove_disk= raid5_remove_disk,
4744 .spare_active = raid5_spare_active,
4745 .sync_request = sync_request,
4746 .resize = raid5_resize,
4747 #ifdef CONFIG_MD_RAID5_RESHAPE
4748 .check_reshape = raid5_check_reshape,
4749 .start_reshape = raid5_start_reshape,
4751 .quiesce = raid5_quiesce,
4754 static struct mdk_personality raid4_personality =
4758 .owner = THIS_MODULE,
4759 .make_request = make_request,
4763 .error_handler = error,
4764 .hot_add_disk = raid5_add_disk,
4765 .hot_remove_disk= raid5_remove_disk,
4766 .spare_active = raid5_spare_active,
4767 .sync_request = sync_request,
4768 .resize = raid5_resize,
4769 #ifdef CONFIG_MD_RAID5_RESHAPE
4770 .check_reshape = raid5_check_reshape,
4771 .start_reshape = raid5_start_reshape,
4773 .quiesce = raid5_quiesce,
4776 static int __init raid5_init(void)
4780 e = raid6_select_algo();
4783 register_md_personality(&raid6_personality);
4784 register_md_personality(&raid5_personality);
4785 register_md_personality(&raid4_personality);
4789 static void raid5_exit(void)
4791 unregister_md_personality(&raid6_personality);
4792 unregister_md_personality(&raid5_personality);
4793 unregister_md_personality(&raid4_personality);
4796 module_init(raid5_init);
4797 module_exit(raid5_exit);
4798 MODULE_LICENSE("GPL");
4799 MODULE_ALIAS("md-personality-4"); /* RAID5 */
4800 MODULE_ALIAS("md-raid5");
4801 MODULE_ALIAS("md-raid4");
4802 MODULE_ALIAS("md-level-5");
4803 MODULE_ALIAS("md-level-4");
4804 MODULE_ALIAS("md-personality-8"); /* RAID6 */
4805 MODULE_ALIAS("md-raid6");
4806 MODULE_ALIAS("md-level-6");
4808 /* This used to be two separate modules, they were: */
4809 MODULE_ALIAS("raid5");
4810 MODULE_ALIAS("raid6");