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);
376 if (test_and_clear_bit(STRIPE_OP_IO, &sh->ops.pending))
379 sh->ops.count -= ack;
380 if (unlikely(sh->ops.count < 0)) {
381 printk(KERN_ERR "pending: %#lx ops.pending: %#lx ops.ack: %#lx "
382 "ops.complete: %#lx\n", pending, sh->ops.pending,
383 sh->ops.ack, sh->ops.complete);
391 raid5_end_read_request(struct bio *bi, int error);
393 raid5_end_write_request(struct bio *bi, int error);
395 static void ops_run_io(struct stripe_head *sh)
397 raid5_conf_t *conf = sh->raid_conf;
398 int i, disks = sh->disks;
402 set_bit(STRIPE_IO_STARTED, &sh->state);
403 for (i = disks; i--; ) {
407 if (test_and_clear_bit(R5_Wantwrite, &sh->dev[i].flags))
409 else if (test_and_clear_bit(R5_Wantread, &sh->dev[i].flags))
414 bi = &sh->dev[i].req;
418 bi->bi_end_io = raid5_end_write_request;
420 bi->bi_end_io = raid5_end_read_request;
423 rdev = rcu_dereference(conf->disks[i].rdev);
424 if (rdev && test_bit(Faulty, &rdev->flags))
427 atomic_inc(&rdev->nr_pending);
431 if (test_bit(STRIPE_SYNCING, &sh->state) ||
432 test_bit(STRIPE_EXPAND_SOURCE, &sh->state) ||
433 test_bit(STRIPE_EXPAND_READY, &sh->state))
434 md_sync_acct(rdev->bdev, STRIPE_SECTORS);
436 bi->bi_bdev = rdev->bdev;
437 pr_debug("%s: for %llu schedule op %ld on disc %d\n",
438 __func__, (unsigned long long)sh->sector,
440 atomic_inc(&sh->count);
441 bi->bi_sector = sh->sector + rdev->data_offset;
442 bi->bi_flags = 1 << BIO_UPTODATE;
446 bi->bi_io_vec = &sh->dev[i].vec;
447 bi->bi_io_vec[0].bv_len = STRIPE_SIZE;
448 bi->bi_io_vec[0].bv_offset = 0;
449 bi->bi_size = STRIPE_SIZE;
452 test_bit(R5_ReWrite, &sh->dev[i].flags))
453 atomic_add(STRIPE_SECTORS,
454 &rdev->corrected_errors);
455 generic_make_request(bi);
458 set_bit(STRIPE_DEGRADED, &sh->state);
459 pr_debug("skip op %ld on disc %d for sector %llu\n",
460 bi->bi_rw, i, (unsigned long long)sh->sector);
461 clear_bit(R5_LOCKED, &sh->dev[i].flags);
462 set_bit(STRIPE_HANDLE, &sh->state);
467 static struct dma_async_tx_descriptor *
468 async_copy_data(int frombio, struct bio *bio, struct page *page,
469 sector_t sector, struct dma_async_tx_descriptor *tx)
472 struct page *bio_page;
476 if (bio->bi_sector >= sector)
477 page_offset = (signed)(bio->bi_sector - sector) * 512;
479 page_offset = (signed)(sector - bio->bi_sector) * -512;
480 bio_for_each_segment(bvl, bio, i) {
481 int len = bio_iovec_idx(bio, i)->bv_len;
485 if (page_offset < 0) {
486 b_offset = -page_offset;
487 page_offset += b_offset;
491 if (len > 0 && page_offset + len > STRIPE_SIZE)
492 clen = STRIPE_SIZE - page_offset;
497 b_offset += bio_iovec_idx(bio, i)->bv_offset;
498 bio_page = bio_iovec_idx(bio, i)->bv_page;
500 tx = async_memcpy(page, bio_page, page_offset,
505 tx = async_memcpy(bio_page, page, b_offset,
510 if (clen < len) /* hit end of page */
518 static void ops_complete_biofill(void *stripe_head_ref)
520 struct stripe_head *sh = stripe_head_ref;
521 struct bio *return_bi = NULL;
522 raid5_conf_t *conf = sh->raid_conf;
525 pr_debug("%s: stripe %llu\n", __func__,
526 (unsigned long long)sh->sector);
528 /* clear completed biofills */
529 for (i = sh->disks; i--; ) {
530 struct r5dev *dev = &sh->dev[i];
532 /* acknowledge completion of a biofill operation */
533 /* and check if we need to reply to a read request,
534 * new R5_Wantfill requests are held off until
535 * !test_bit(STRIPE_OP_BIOFILL, &sh->ops.pending)
537 if (test_and_clear_bit(R5_Wantfill, &dev->flags)) {
538 struct bio *rbi, *rbi2;
540 /* The access to dev->read is outside of the
541 * spin_lock_irq(&conf->device_lock), but is protected
542 * by the STRIPE_OP_BIOFILL pending bit
547 while (rbi && rbi->bi_sector <
548 dev->sector + STRIPE_SECTORS) {
549 rbi2 = r5_next_bio(rbi, dev->sector);
550 spin_lock_irq(&conf->device_lock);
551 if (--rbi->bi_phys_segments == 0) {
552 rbi->bi_next = return_bi;
555 spin_unlock_irq(&conf->device_lock);
560 set_bit(STRIPE_OP_BIOFILL, &sh->ops.complete);
562 return_io(return_bi);
564 set_bit(STRIPE_HANDLE, &sh->state);
568 static void ops_run_biofill(struct stripe_head *sh)
570 struct dma_async_tx_descriptor *tx = NULL;
571 raid5_conf_t *conf = sh->raid_conf;
574 pr_debug("%s: stripe %llu\n", __func__,
575 (unsigned long long)sh->sector);
577 for (i = sh->disks; i--; ) {
578 struct r5dev *dev = &sh->dev[i];
579 if (test_bit(R5_Wantfill, &dev->flags)) {
581 spin_lock_irq(&conf->device_lock);
582 dev->read = rbi = dev->toread;
584 spin_unlock_irq(&conf->device_lock);
585 while (rbi && rbi->bi_sector <
586 dev->sector + STRIPE_SECTORS) {
587 tx = async_copy_data(0, rbi, dev->page,
589 rbi = r5_next_bio(rbi, dev->sector);
594 atomic_inc(&sh->count);
595 async_trigger_callback(ASYNC_TX_DEP_ACK | ASYNC_TX_ACK, tx,
596 ops_complete_biofill, sh);
599 static void ops_complete_compute5(void *stripe_head_ref)
601 struct stripe_head *sh = stripe_head_ref;
602 int target = sh->ops.target;
603 struct r5dev *tgt = &sh->dev[target];
605 pr_debug("%s: stripe %llu\n", __func__,
606 (unsigned long long)sh->sector);
608 set_bit(R5_UPTODATE, &tgt->flags);
609 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
610 clear_bit(R5_Wantcompute, &tgt->flags);
611 set_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.complete);
612 set_bit(STRIPE_HANDLE, &sh->state);
616 static struct dma_async_tx_descriptor *
617 ops_run_compute5(struct stripe_head *sh, unsigned long pending)
619 /* kernel stack size limits the total number of disks */
620 int disks = sh->disks;
621 struct page *xor_srcs[disks];
622 int target = sh->ops.target;
623 struct r5dev *tgt = &sh->dev[target];
624 struct page *xor_dest = tgt->page;
626 struct dma_async_tx_descriptor *tx;
629 pr_debug("%s: stripe %llu block: %d\n",
630 __func__, (unsigned long long)sh->sector, target);
631 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
633 for (i = disks; i--; )
635 xor_srcs[count++] = sh->dev[i].page;
637 atomic_inc(&sh->count);
639 if (unlikely(count == 1))
640 tx = async_memcpy(xor_dest, xor_srcs[0], 0, 0, STRIPE_SIZE,
641 0, NULL, ops_complete_compute5, sh);
643 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE,
644 ASYNC_TX_XOR_ZERO_DST, NULL,
645 ops_complete_compute5, sh);
647 /* ack now if postxor is not set to be run */
648 if (tx && !test_bit(STRIPE_OP_POSTXOR, &pending))
654 static void ops_complete_prexor(void *stripe_head_ref)
656 struct stripe_head *sh = stripe_head_ref;
658 pr_debug("%s: stripe %llu\n", __func__,
659 (unsigned long long)sh->sector);
661 set_bit(STRIPE_OP_PREXOR, &sh->ops.complete);
664 static struct dma_async_tx_descriptor *
665 ops_run_prexor(struct stripe_head *sh, struct dma_async_tx_descriptor *tx)
667 /* kernel stack size limits the total number of disks */
668 int disks = sh->disks;
669 struct page *xor_srcs[disks];
670 int count = 0, pd_idx = sh->pd_idx, i;
672 /* existing parity data subtracted */
673 struct page *xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
675 pr_debug("%s: stripe %llu\n", __func__,
676 (unsigned long long)sh->sector);
678 for (i = disks; i--; ) {
679 struct r5dev *dev = &sh->dev[i];
680 /* Only process blocks that are known to be uptodate */
681 if (dev->towrite && test_bit(R5_Wantprexor, &dev->flags))
682 xor_srcs[count++] = dev->page;
685 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE,
686 ASYNC_TX_DEP_ACK | ASYNC_TX_XOR_DROP_DST, tx,
687 ops_complete_prexor, sh);
692 static struct dma_async_tx_descriptor *
693 ops_run_biodrain(struct stripe_head *sh, struct dma_async_tx_descriptor *tx,
694 unsigned long pending)
696 int disks = sh->disks;
697 int pd_idx = sh->pd_idx, i;
699 /* check if prexor is active which means only process blocks
700 * that are part of a read-modify-write (Wantprexor)
702 int prexor = test_bit(STRIPE_OP_PREXOR, &pending);
704 pr_debug("%s: stripe %llu\n", __func__,
705 (unsigned long long)sh->sector);
707 for (i = disks; i--; ) {
708 struct r5dev *dev = &sh->dev[i];
713 if (prexor) { /* rmw */
715 test_bit(R5_Wantprexor, &dev->flags))
718 if (i != pd_idx && dev->towrite &&
719 test_bit(R5_LOCKED, &dev->flags))
726 spin_lock(&sh->lock);
727 chosen = dev->towrite;
729 BUG_ON(dev->written);
730 wbi = dev->written = chosen;
731 spin_unlock(&sh->lock);
733 while (wbi && wbi->bi_sector <
734 dev->sector + STRIPE_SECTORS) {
735 tx = async_copy_data(1, wbi, dev->page,
737 wbi = r5_next_bio(wbi, dev->sector);
745 static void ops_complete_postxor(void *stripe_head_ref)
747 struct stripe_head *sh = stripe_head_ref;
749 pr_debug("%s: stripe %llu\n", __func__,
750 (unsigned long long)sh->sector);
752 set_bit(STRIPE_OP_POSTXOR, &sh->ops.complete);
753 set_bit(STRIPE_HANDLE, &sh->state);
757 static void ops_complete_write(void *stripe_head_ref)
759 struct stripe_head *sh = stripe_head_ref;
760 int disks = sh->disks, i, pd_idx = sh->pd_idx;
762 pr_debug("%s: stripe %llu\n", __func__,
763 (unsigned long long)sh->sector);
765 for (i = disks; i--; ) {
766 struct r5dev *dev = &sh->dev[i];
767 if (dev->written || i == pd_idx)
768 set_bit(R5_UPTODATE, &dev->flags);
771 set_bit(STRIPE_OP_BIODRAIN, &sh->ops.complete);
772 set_bit(STRIPE_OP_POSTXOR, &sh->ops.complete);
774 set_bit(STRIPE_HANDLE, &sh->state);
779 ops_run_postxor(struct stripe_head *sh, struct dma_async_tx_descriptor *tx,
780 unsigned long pending)
782 /* kernel stack size limits the total number of disks */
783 int disks = sh->disks;
784 struct page *xor_srcs[disks];
786 int count = 0, pd_idx = sh->pd_idx, i;
787 struct page *xor_dest;
788 int prexor = test_bit(STRIPE_OP_PREXOR, &pending);
790 dma_async_tx_callback callback;
792 pr_debug("%s: stripe %llu\n", __func__,
793 (unsigned long long)sh->sector);
795 /* check if prexor is active which means only process blocks
796 * that are part of a read-modify-write (written)
799 xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
800 for (i = disks; i--; ) {
801 struct r5dev *dev = &sh->dev[i];
803 xor_srcs[count++] = dev->page;
806 xor_dest = sh->dev[pd_idx].page;
807 for (i = disks; i--; ) {
808 struct r5dev *dev = &sh->dev[i];
810 xor_srcs[count++] = dev->page;
814 /* check whether this postxor is part of a write */
815 callback = test_bit(STRIPE_OP_BIODRAIN, &pending) ?
816 ops_complete_write : ops_complete_postxor;
818 /* 1/ if we prexor'd then the dest is reused as a source
819 * 2/ if we did not prexor then we are redoing the parity
820 * set ASYNC_TX_XOR_DROP_DST and ASYNC_TX_XOR_ZERO_DST
821 * for the synchronous xor case
823 flags = ASYNC_TX_DEP_ACK | ASYNC_TX_ACK |
824 (prexor ? ASYNC_TX_XOR_DROP_DST : ASYNC_TX_XOR_ZERO_DST);
826 atomic_inc(&sh->count);
828 if (unlikely(count == 1)) {
829 flags &= ~(ASYNC_TX_XOR_DROP_DST | ASYNC_TX_XOR_ZERO_DST);
830 tx = async_memcpy(xor_dest, xor_srcs[0], 0, 0, STRIPE_SIZE,
831 flags, tx, callback, sh);
833 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE,
834 flags, tx, callback, sh);
837 static void ops_complete_check(void *stripe_head_ref)
839 struct stripe_head *sh = stripe_head_ref;
841 pr_debug("%s: stripe %llu\n", __func__,
842 (unsigned long long)sh->sector);
844 set_bit(STRIPE_OP_CHECK, &sh->ops.complete);
845 set_bit(STRIPE_HANDLE, &sh->state);
849 static void ops_run_check(struct stripe_head *sh)
851 /* kernel stack size limits the total number of disks */
852 int disks = sh->disks;
853 struct page *xor_srcs[disks];
854 struct dma_async_tx_descriptor *tx;
856 int count = 0, pd_idx = sh->pd_idx, i;
857 struct page *xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
859 pr_debug("%s: stripe %llu\n", __func__,
860 (unsigned long long)sh->sector);
862 for (i = disks; i--; ) {
863 struct r5dev *dev = &sh->dev[i];
865 xor_srcs[count++] = dev->page;
868 tx = async_xor_zero_sum(xor_dest, xor_srcs, 0, count, STRIPE_SIZE,
869 &sh->ops.zero_sum_result, 0, NULL, NULL, NULL);
871 atomic_inc(&sh->count);
872 tx = async_trigger_callback(ASYNC_TX_DEP_ACK | ASYNC_TX_ACK, tx,
873 ops_complete_check, sh);
876 static void raid5_run_ops(struct stripe_head *sh, unsigned long pending)
878 int overlap_clear = 0, i, disks = sh->disks;
879 struct dma_async_tx_descriptor *tx = NULL;
881 if (test_bit(STRIPE_OP_BIOFILL, &pending)) {
886 if (test_bit(STRIPE_OP_COMPUTE_BLK, &pending))
887 tx = ops_run_compute5(sh, pending);
889 if (test_bit(STRIPE_OP_PREXOR, &pending))
890 tx = ops_run_prexor(sh, tx);
892 if (test_bit(STRIPE_OP_BIODRAIN, &pending)) {
893 tx = ops_run_biodrain(sh, tx, pending);
897 if (test_bit(STRIPE_OP_POSTXOR, &pending))
898 ops_run_postxor(sh, tx, pending);
900 if (test_bit(STRIPE_OP_CHECK, &pending))
903 if (test_bit(STRIPE_OP_IO, &pending))
907 for (i = disks; i--; ) {
908 struct r5dev *dev = &sh->dev[i];
909 if (test_and_clear_bit(R5_Overlap, &dev->flags))
910 wake_up(&sh->raid_conf->wait_for_overlap);
914 static int grow_one_stripe(raid5_conf_t *conf)
916 struct stripe_head *sh;
917 sh = kmem_cache_alloc(conf->slab_cache, GFP_KERNEL);
920 memset(sh, 0, sizeof(*sh) + (conf->raid_disks-1)*sizeof(struct r5dev));
921 sh->raid_conf = conf;
922 spin_lock_init(&sh->lock);
924 if (grow_buffers(sh, conf->raid_disks)) {
925 shrink_buffers(sh, conf->raid_disks);
926 kmem_cache_free(conf->slab_cache, sh);
929 sh->disks = conf->raid_disks;
930 /* we just created an active stripe so... */
931 atomic_set(&sh->count, 1);
932 atomic_inc(&conf->active_stripes);
933 INIT_LIST_HEAD(&sh->lru);
938 static int grow_stripes(raid5_conf_t *conf, int num)
940 struct kmem_cache *sc;
941 int devs = conf->raid_disks;
943 sprintf(conf->cache_name[0], "raid5-%s", mdname(conf->mddev));
944 sprintf(conf->cache_name[1], "raid5-%s-alt", mdname(conf->mddev));
945 conf->active_name = 0;
946 sc = kmem_cache_create(conf->cache_name[conf->active_name],
947 sizeof(struct stripe_head)+(devs-1)*sizeof(struct r5dev),
951 conf->slab_cache = sc;
952 conf->pool_size = devs;
954 if (!grow_one_stripe(conf))
959 #ifdef CONFIG_MD_RAID5_RESHAPE
960 static int resize_stripes(raid5_conf_t *conf, int newsize)
962 /* Make all the stripes able to hold 'newsize' devices.
963 * New slots in each stripe get 'page' set to a new page.
965 * This happens in stages:
966 * 1/ create a new kmem_cache and allocate the required number of
968 * 2/ gather all the old stripe_heads and tranfer the pages across
969 * to the new stripe_heads. This will have the side effect of
970 * freezing the array as once all stripe_heads have been collected,
971 * no IO will be possible. Old stripe heads are freed once their
972 * pages have been transferred over, and the old kmem_cache is
973 * freed when all stripes are done.
974 * 3/ reallocate conf->disks to be suitable bigger. If this fails,
975 * we simple return a failre status - no need to clean anything up.
976 * 4/ allocate new pages for the new slots in the new stripe_heads.
977 * If this fails, we don't bother trying the shrink the
978 * stripe_heads down again, we just leave them as they are.
979 * As each stripe_head is processed the new one is released into
982 * Once step2 is started, we cannot afford to wait for a write,
983 * so we use GFP_NOIO allocations.
985 struct stripe_head *osh, *nsh;
986 LIST_HEAD(newstripes);
987 struct disk_info *ndisks;
989 struct kmem_cache *sc;
992 if (newsize <= conf->pool_size)
993 return 0; /* never bother to shrink */
995 md_allow_write(conf->mddev);
998 sc = kmem_cache_create(conf->cache_name[1-conf->active_name],
999 sizeof(struct stripe_head)+(newsize-1)*sizeof(struct r5dev),
1004 for (i = conf->max_nr_stripes; i; i--) {
1005 nsh = kmem_cache_alloc(sc, GFP_KERNEL);
1009 memset(nsh, 0, sizeof(*nsh) + (newsize-1)*sizeof(struct r5dev));
1011 nsh->raid_conf = conf;
1012 spin_lock_init(&nsh->lock);
1014 list_add(&nsh->lru, &newstripes);
1017 /* didn't get enough, give up */
1018 while (!list_empty(&newstripes)) {
1019 nsh = list_entry(newstripes.next, struct stripe_head, lru);
1020 list_del(&nsh->lru);
1021 kmem_cache_free(sc, nsh);
1023 kmem_cache_destroy(sc);
1026 /* Step 2 - Must use GFP_NOIO now.
1027 * OK, we have enough stripes, start collecting inactive
1028 * stripes and copying them over
1030 list_for_each_entry(nsh, &newstripes, lru) {
1031 spin_lock_irq(&conf->device_lock);
1032 wait_event_lock_irq(conf->wait_for_stripe,
1033 !list_empty(&conf->inactive_list),
1035 unplug_slaves(conf->mddev)
1037 osh = get_free_stripe(conf);
1038 spin_unlock_irq(&conf->device_lock);
1039 atomic_set(&nsh->count, 1);
1040 for(i=0; i<conf->pool_size; i++)
1041 nsh->dev[i].page = osh->dev[i].page;
1042 for( ; i<newsize; i++)
1043 nsh->dev[i].page = NULL;
1044 kmem_cache_free(conf->slab_cache, osh);
1046 kmem_cache_destroy(conf->slab_cache);
1049 * At this point, we are holding all the stripes so the array
1050 * is completely stalled, so now is a good time to resize
1053 ndisks = kzalloc(newsize * sizeof(struct disk_info), GFP_NOIO);
1055 for (i=0; i<conf->raid_disks; i++)
1056 ndisks[i] = conf->disks[i];
1058 conf->disks = ndisks;
1062 /* Step 4, return new stripes to service */
1063 while(!list_empty(&newstripes)) {
1064 nsh = list_entry(newstripes.next, struct stripe_head, lru);
1065 list_del_init(&nsh->lru);
1066 for (i=conf->raid_disks; i < newsize; i++)
1067 if (nsh->dev[i].page == NULL) {
1068 struct page *p = alloc_page(GFP_NOIO);
1069 nsh->dev[i].page = p;
1073 release_stripe(nsh);
1075 /* critical section pass, GFP_NOIO no longer needed */
1077 conf->slab_cache = sc;
1078 conf->active_name = 1-conf->active_name;
1079 conf->pool_size = newsize;
1084 static int drop_one_stripe(raid5_conf_t *conf)
1086 struct stripe_head *sh;
1088 spin_lock_irq(&conf->device_lock);
1089 sh = get_free_stripe(conf);
1090 spin_unlock_irq(&conf->device_lock);
1093 BUG_ON(atomic_read(&sh->count));
1094 shrink_buffers(sh, conf->pool_size);
1095 kmem_cache_free(conf->slab_cache, sh);
1096 atomic_dec(&conf->active_stripes);
1100 static void shrink_stripes(raid5_conf_t *conf)
1102 while (drop_one_stripe(conf))
1105 if (conf->slab_cache)
1106 kmem_cache_destroy(conf->slab_cache);
1107 conf->slab_cache = NULL;
1110 static void raid5_end_read_request(struct bio * bi, int error)
1112 struct stripe_head *sh = bi->bi_private;
1113 raid5_conf_t *conf = sh->raid_conf;
1114 int disks = sh->disks, i;
1115 int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
1116 char b[BDEVNAME_SIZE];
1120 for (i=0 ; i<disks; i++)
1121 if (bi == &sh->dev[i].req)
1124 pr_debug("end_read_request %llu/%d, count: %d, uptodate %d.\n",
1125 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
1133 set_bit(R5_UPTODATE, &sh->dev[i].flags);
1134 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
1135 rdev = conf->disks[i].rdev;
1136 printk_rl(KERN_INFO "raid5:%s: read error corrected"
1137 " (%lu sectors at %llu on %s)\n",
1138 mdname(conf->mddev), STRIPE_SECTORS,
1139 (unsigned long long)(sh->sector
1140 + rdev->data_offset),
1141 bdevname(rdev->bdev, b));
1142 clear_bit(R5_ReadError, &sh->dev[i].flags);
1143 clear_bit(R5_ReWrite, &sh->dev[i].flags);
1145 if (atomic_read(&conf->disks[i].rdev->read_errors))
1146 atomic_set(&conf->disks[i].rdev->read_errors, 0);
1148 const char *bdn = bdevname(conf->disks[i].rdev->bdev, b);
1150 rdev = conf->disks[i].rdev;
1152 clear_bit(R5_UPTODATE, &sh->dev[i].flags);
1153 atomic_inc(&rdev->read_errors);
1154 if (conf->mddev->degraded)
1155 printk_rl(KERN_WARNING
1156 "raid5:%s: read error not correctable "
1157 "(sector %llu on %s).\n",
1158 mdname(conf->mddev),
1159 (unsigned long long)(sh->sector
1160 + rdev->data_offset),
1162 else if (test_bit(R5_ReWrite, &sh->dev[i].flags))
1164 printk_rl(KERN_WARNING
1165 "raid5:%s: read error NOT corrected!! "
1166 "(sector %llu on %s).\n",
1167 mdname(conf->mddev),
1168 (unsigned long long)(sh->sector
1169 + rdev->data_offset),
1171 else if (atomic_read(&rdev->read_errors)
1172 > conf->max_nr_stripes)
1174 "raid5:%s: Too many read errors, failing device %s.\n",
1175 mdname(conf->mddev), bdn);
1179 set_bit(R5_ReadError, &sh->dev[i].flags);
1181 clear_bit(R5_ReadError, &sh->dev[i].flags);
1182 clear_bit(R5_ReWrite, &sh->dev[i].flags);
1183 md_error(conf->mddev, rdev);
1186 rdev_dec_pending(conf->disks[i].rdev, conf->mddev);
1187 clear_bit(R5_LOCKED, &sh->dev[i].flags);
1188 set_bit(STRIPE_HANDLE, &sh->state);
1192 static void raid5_end_write_request (struct bio *bi, int error)
1194 struct stripe_head *sh = bi->bi_private;
1195 raid5_conf_t *conf = sh->raid_conf;
1196 int disks = sh->disks, i;
1197 int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
1199 for (i=0 ; i<disks; i++)
1200 if (bi == &sh->dev[i].req)
1203 pr_debug("end_write_request %llu/%d, count %d, uptodate: %d.\n",
1204 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
1212 md_error(conf->mddev, conf->disks[i].rdev);
1214 rdev_dec_pending(conf->disks[i].rdev, conf->mddev);
1216 clear_bit(R5_LOCKED, &sh->dev[i].flags);
1217 set_bit(STRIPE_HANDLE, &sh->state);
1222 static sector_t compute_blocknr(struct stripe_head *sh, int i);
1224 static void raid5_build_block (struct stripe_head *sh, int i)
1226 struct r5dev *dev = &sh->dev[i];
1228 bio_init(&dev->req);
1229 dev->req.bi_io_vec = &dev->vec;
1231 dev->req.bi_max_vecs++;
1232 dev->vec.bv_page = dev->page;
1233 dev->vec.bv_len = STRIPE_SIZE;
1234 dev->vec.bv_offset = 0;
1236 dev->req.bi_sector = sh->sector;
1237 dev->req.bi_private = sh;
1240 dev->sector = compute_blocknr(sh, i);
1243 static void error(mddev_t *mddev, mdk_rdev_t *rdev)
1245 char b[BDEVNAME_SIZE];
1246 raid5_conf_t *conf = (raid5_conf_t *) mddev->private;
1247 pr_debug("raid5: error called\n");
1249 if (!test_bit(Faulty, &rdev->flags)) {
1250 set_bit(MD_CHANGE_DEVS, &mddev->flags);
1251 if (test_and_clear_bit(In_sync, &rdev->flags)) {
1252 unsigned long flags;
1253 spin_lock_irqsave(&conf->device_lock, flags);
1255 spin_unlock_irqrestore(&conf->device_lock, flags);
1257 * if recovery was running, make sure it aborts.
1259 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1261 set_bit(Faulty, &rdev->flags);
1263 "raid5: Disk failure on %s, disabling device.\n"
1264 "raid5: Operation continuing on %d devices.\n",
1265 bdevname(rdev->bdev,b), conf->raid_disks - mddev->degraded);
1270 * Input: a 'big' sector number,
1271 * Output: index of the data and parity disk, and the sector # in them.
1273 static sector_t raid5_compute_sector(sector_t r_sector, unsigned int raid_disks,
1274 unsigned int data_disks, unsigned int * dd_idx,
1275 unsigned int * pd_idx, raid5_conf_t *conf)
1278 unsigned long chunk_number;
1279 unsigned int chunk_offset;
1280 sector_t new_sector;
1281 int sectors_per_chunk = conf->chunk_size >> 9;
1283 /* First compute the information on this sector */
1286 * Compute the chunk number and the sector offset inside the chunk
1288 chunk_offset = sector_div(r_sector, sectors_per_chunk);
1289 chunk_number = r_sector;
1290 BUG_ON(r_sector != chunk_number);
1293 * Compute the stripe number
1295 stripe = chunk_number / data_disks;
1298 * Compute the data disk and parity disk indexes inside the stripe
1300 *dd_idx = chunk_number % data_disks;
1303 * Select the parity disk based on the user selected algorithm.
1305 switch(conf->level) {
1307 *pd_idx = data_disks;
1310 switch (conf->algorithm) {
1311 case ALGORITHM_LEFT_ASYMMETRIC:
1312 *pd_idx = data_disks - stripe % raid_disks;
1313 if (*dd_idx >= *pd_idx)
1316 case ALGORITHM_RIGHT_ASYMMETRIC:
1317 *pd_idx = stripe % raid_disks;
1318 if (*dd_idx >= *pd_idx)
1321 case ALGORITHM_LEFT_SYMMETRIC:
1322 *pd_idx = data_disks - stripe % raid_disks;
1323 *dd_idx = (*pd_idx + 1 + *dd_idx) % raid_disks;
1325 case ALGORITHM_RIGHT_SYMMETRIC:
1326 *pd_idx = stripe % raid_disks;
1327 *dd_idx = (*pd_idx + 1 + *dd_idx) % raid_disks;
1330 printk(KERN_ERR "raid5: unsupported algorithm %d\n",
1336 /**** FIX THIS ****/
1337 switch (conf->algorithm) {
1338 case ALGORITHM_LEFT_ASYMMETRIC:
1339 *pd_idx = raid_disks - 1 - (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_RIGHT_ASYMMETRIC:
1346 *pd_idx = stripe % raid_disks;
1347 if (*pd_idx == raid_disks-1)
1348 (*dd_idx)++; /* Q D D D P */
1349 else if (*dd_idx >= *pd_idx)
1350 (*dd_idx) += 2; /* D D P Q D */
1352 case ALGORITHM_LEFT_SYMMETRIC:
1353 *pd_idx = raid_disks - 1 - (stripe % raid_disks);
1354 *dd_idx = (*pd_idx + 2 + *dd_idx) % raid_disks;
1356 case ALGORITHM_RIGHT_SYMMETRIC:
1357 *pd_idx = stripe % raid_disks;
1358 *dd_idx = (*pd_idx + 2 + *dd_idx) % raid_disks;
1361 printk (KERN_CRIT "raid6: unsupported algorithm %d\n",
1368 * Finally, compute the new sector number
1370 new_sector = (sector_t)stripe * sectors_per_chunk + chunk_offset;
1375 static sector_t compute_blocknr(struct stripe_head *sh, int i)
1377 raid5_conf_t *conf = sh->raid_conf;
1378 int raid_disks = sh->disks;
1379 int data_disks = raid_disks - conf->max_degraded;
1380 sector_t new_sector = sh->sector, check;
1381 int sectors_per_chunk = conf->chunk_size >> 9;
1384 int chunk_number, dummy1, dummy2, dd_idx = i;
1388 chunk_offset = sector_div(new_sector, sectors_per_chunk);
1389 stripe = new_sector;
1390 BUG_ON(new_sector != stripe);
1392 if (i == sh->pd_idx)
1394 switch(conf->level) {
1397 switch (conf->algorithm) {
1398 case ALGORITHM_LEFT_ASYMMETRIC:
1399 case ALGORITHM_RIGHT_ASYMMETRIC:
1403 case ALGORITHM_LEFT_SYMMETRIC:
1404 case ALGORITHM_RIGHT_SYMMETRIC:
1407 i -= (sh->pd_idx + 1);
1410 printk(KERN_ERR "raid5: unsupported algorithm %d\n",
1415 if (i == raid6_next_disk(sh->pd_idx, raid_disks))
1416 return 0; /* It is the Q disk */
1417 switch (conf->algorithm) {
1418 case ALGORITHM_LEFT_ASYMMETRIC:
1419 case ALGORITHM_RIGHT_ASYMMETRIC:
1420 if (sh->pd_idx == raid_disks-1)
1421 i--; /* Q D D D P */
1422 else if (i > sh->pd_idx)
1423 i -= 2; /* D D P Q D */
1425 case ALGORITHM_LEFT_SYMMETRIC:
1426 case ALGORITHM_RIGHT_SYMMETRIC:
1427 if (sh->pd_idx == raid_disks-1)
1428 i--; /* Q D D D P */
1433 i -= (sh->pd_idx + 2);
1437 printk (KERN_CRIT "raid6: unsupported algorithm %d\n",
1443 chunk_number = stripe * data_disks + i;
1444 r_sector = (sector_t)chunk_number * sectors_per_chunk + chunk_offset;
1446 check = raid5_compute_sector (r_sector, raid_disks, data_disks, &dummy1, &dummy2, conf);
1447 if (check != sh->sector || dummy1 != dd_idx || dummy2 != sh->pd_idx) {
1448 printk(KERN_ERR "compute_blocknr: map not correct\n");
1457 * Copy data between a page in the stripe cache, and one or more bion
1458 * The page could align with the middle of the bio, or there could be
1459 * several bion, each with several bio_vecs, which cover part of the page
1460 * Multiple bion are linked together on bi_next. There may be extras
1461 * at the end of this list. We ignore them.
1463 static void copy_data(int frombio, struct bio *bio,
1467 char *pa = page_address(page);
1468 struct bio_vec *bvl;
1472 if (bio->bi_sector >= sector)
1473 page_offset = (signed)(bio->bi_sector - sector) * 512;
1475 page_offset = (signed)(sector - bio->bi_sector) * -512;
1476 bio_for_each_segment(bvl, bio, i) {
1477 int len = bio_iovec_idx(bio,i)->bv_len;
1481 if (page_offset < 0) {
1482 b_offset = -page_offset;
1483 page_offset += b_offset;
1487 if (len > 0 && page_offset + len > STRIPE_SIZE)
1488 clen = STRIPE_SIZE - page_offset;
1492 char *ba = __bio_kmap_atomic(bio, i, KM_USER0);
1494 memcpy(pa+page_offset, ba+b_offset, clen);
1496 memcpy(ba+b_offset, pa+page_offset, clen);
1497 __bio_kunmap_atomic(ba, KM_USER0);
1499 if (clen < len) /* hit end of page */
1505 #define check_xor() do { \
1506 if (count == MAX_XOR_BLOCKS) { \
1507 xor_blocks(count, STRIPE_SIZE, dest, ptr);\
1512 static void compute_parity6(struct stripe_head *sh, int method)
1514 raid6_conf_t *conf = sh->raid_conf;
1515 int i, pd_idx = sh->pd_idx, qd_idx, d0_idx, disks = sh->disks, count;
1517 /**** FIX THIS: This could be very bad if disks is close to 256 ****/
1520 qd_idx = raid6_next_disk(pd_idx, disks);
1521 d0_idx = raid6_next_disk(qd_idx, disks);
1523 pr_debug("compute_parity, stripe %llu, method %d\n",
1524 (unsigned long long)sh->sector, method);
1527 case READ_MODIFY_WRITE:
1528 BUG(); /* READ_MODIFY_WRITE N/A for RAID-6 */
1529 case RECONSTRUCT_WRITE:
1530 for (i= disks; i-- ;)
1531 if ( i != pd_idx && i != qd_idx && sh->dev[i].towrite ) {
1532 chosen = sh->dev[i].towrite;
1533 sh->dev[i].towrite = NULL;
1535 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
1536 wake_up(&conf->wait_for_overlap);
1538 BUG_ON(sh->dev[i].written);
1539 sh->dev[i].written = chosen;
1543 BUG(); /* Not implemented yet */
1546 for (i = disks; i--;)
1547 if (sh->dev[i].written) {
1548 sector_t sector = sh->dev[i].sector;
1549 struct bio *wbi = sh->dev[i].written;
1550 while (wbi && wbi->bi_sector < sector + STRIPE_SECTORS) {
1551 copy_data(1, wbi, sh->dev[i].page, sector);
1552 wbi = r5_next_bio(wbi, sector);
1555 set_bit(R5_LOCKED, &sh->dev[i].flags);
1556 set_bit(R5_UPTODATE, &sh->dev[i].flags);
1560 // case RECONSTRUCT_WRITE:
1561 // case CHECK_PARITY:
1562 // case UPDATE_PARITY:
1563 /* Note that unlike RAID-5, the ordering of the disks matters greatly. */
1564 /* FIX: Is this ordering of drives even remotely optimal? */
1568 ptrs[count++] = page_address(sh->dev[i].page);
1569 if (count <= disks-2 && !test_bit(R5_UPTODATE, &sh->dev[i].flags))
1570 printk("block %d/%d not uptodate on parity calc\n", i,count);
1571 i = raid6_next_disk(i, disks);
1572 } while ( i != d0_idx );
1576 raid6_call.gen_syndrome(disks, STRIPE_SIZE, ptrs);
1579 case RECONSTRUCT_WRITE:
1580 set_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
1581 set_bit(R5_UPTODATE, &sh->dev[qd_idx].flags);
1582 set_bit(R5_LOCKED, &sh->dev[pd_idx].flags);
1583 set_bit(R5_LOCKED, &sh->dev[qd_idx].flags);
1586 set_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
1587 set_bit(R5_UPTODATE, &sh->dev[qd_idx].flags);
1593 /* Compute one missing block */
1594 static void compute_block_1(struct stripe_head *sh, int dd_idx, int nozero)
1596 int i, count, disks = sh->disks;
1597 void *ptr[MAX_XOR_BLOCKS], *dest, *p;
1598 int pd_idx = sh->pd_idx;
1599 int qd_idx = raid6_next_disk(pd_idx, disks);
1601 pr_debug("compute_block_1, stripe %llu, idx %d\n",
1602 (unsigned long long)sh->sector, dd_idx);
1604 if ( dd_idx == qd_idx ) {
1605 /* We're actually computing the Q drive */
1606 compute_parity6(sh, UPDATE_PARITY);
1608 dest = page_address(sh->dev[dd_idx].page);
1609 if (!nozero) memset(dest, 0, STRIPE_SIZE);
1611 for (i = disks ; i--; ) {
1612 if (i == dd_idx || i == qd_idx)
1614 p = page_address(sh->dev[i].page);
1615 if (test_bit(R5_UPTODATE, &sh->dev[i].flags))
1618 printk("compute_block() %d, stripe %llu, %d"
1619 " not present\n", dd_idx,
1620 (unsigned long long)sh->sector, i);
1625 xor_blocks(count, STRIPE_SIZE, dest, ptr);
1626 if (!nozero) set_bit(R5_UPTODATE, &sh->dev[dd_idx].flags);
1627 else clear_bit(R5_UPTODATE, &sh->dev[dd_idx].flags);
1631 /* Compute two missing blocks */
1632 static void compute_block_2(struct stripe_head *sh, int dd_idx1, int dd_idx2)
1634 int i, count, disks = sh->disks;
1635 int pd_idx = sh->pd_idx;
1636 int qd_idx = raid6_next_disk(pd_idx, disks);
1637 int d0_idx = raid6_next_disk(qd_idx, disks);
1640 /* faila and failb are disk numbers relative to d0_idx */
1641 /* pd_idx become disks-2 and qd_idx become disks-1 */
1642 faila = (dd_idx1 < d0_idx) ? dd_idx1+(disks-d0_idx) : dd_idx1-d0_idx;
1643 failb = (dd_idx2 < d0_idx) ? dd_idx2+(disks-d0_idx) : dd_idx2-d0_idx;
1645 BUG_ON(faila == failb);
1646 if ( failb < faila ) { int tmp = faila; faila = failb; failb = tmp; }
1648 pr_debug("compute_block_2, stripe %llu, idx %d,%d (%d,%d)\n",
1649 (unsigned long long)sh->sector, dd_idx1, dd_idx2, faila, failb);
1651 if ( failb == disks-1 ) {
1652 /* Q disk is one of the missing disks */
1653 if ( faila == disks-2 ) {
1654 /* Missing P+Q, just recompute */
1655 compute_parity6(sh, UPDATE_PARITY);
1658 /* We're missing D+Q; recompute D from P */
1659 compute_block_1(sh, (dd_idx1 == qd_idx) ? dd_idx2 : dd_idx1, 0);
1660 compute_parity6(sh, UPDATE_PARITY); /* Is this necessary? */
1665 /* We're missing D+P or D+D; build pointer table */
1667 /**** FIX THIS: This could be very bad if disks is close to 256 ****/
1673 ptrs[count++] = page_address(sh->dev[i].page);
1674 i = raid6_next_disk(i, disks);
1675 if (i != dd_idx1 && i != dd_idx2 &&
1676 !test_bit(R5_UPTODATE, &sh->dev[i].flags))
1677 printk("compute_2 with missing block %d/%d\n", count, i);
1678 } while ( i != d0_idx );
1680 if ( failb == disks-2 ) {
1681 /* We're missing D+P. */
1682 raid6_datap_recov(disks, STRIPE_SIZE, faila, ptrs);
1684 /* We're missing D+D. */
1685 raid6_2data_recov(disks, STRIPE_SIZE, faila, failb, ptrs);
1688 /* Both the above update both missing blocks */
1689 set_bit(R5_UPTODATE, &sh->dev[dd_idx1].flags);
1690 set_bit(R5_UPTODATE, &sh->dev[dd_idx2].flags);
1695 handle_write_operations5(struct stripe_head *sh, int rcw, int expand)
1697 int i, pd_idx = sh->pd_idx, disks = sh->disks;
1701 /* if we are not expanding this is a proper write request, and
1702 * there will be bios with new data to be drained into the
1706 set_bit(STRIPE_OP_BIODRAIN, &sh->ops.pending);
1710 set_bit(STRIPE_OP_POSTXOR, &sh->ops.pending);
1713 for (i = disks; i--; ) {
1714 struct r5dev *dev = &sh->dev[i];
1717 set_bit(R5_LOCKED, &dev->flags);
1719 clear_bit(R5_UPTODATE, &dev->flags);
1723 if (locked + 1 == disks)
1724 if (!test_and_set_bit(STRIPE_FULL_WRITE, &sh->state))
1725 atomic_inc(&sh->raid_conf->pending_full_writes);
1727 BUG_ON(!(test_bit(R5_UPTODATE, &sh->dev[pd_idx].flags) ||
1728 test_bit(R5_Wantcompute, &sh->dev[pd_idx].flags)));
1730 set_bit(STRIPE_OP_PREXOR, &sh->ops.pending);
1731 set_bit(STRIPE_OP_BIODRAIN, &sh->ops.pending);
1732 set_bit(STRIPE_OP_POSTXOR, &sh->ops.pending);
1736 for (i = disks; i--; ) {
1737 struct r5dev *dev = &sh->dev[i];
1741 /* For a read-modify write there may be blocks that are
1742 * locked for reading while others are ready to be
1743 * written so we distinguish these blocks by the
1747 (test_bit(R5_UPTODATE, &dev->flags) ||
1748 test_bit(R5_Wantcompute, &dev->flags))) {
1749 set_bit(R5_Wantprexor, &dev->flags);
1750 set_bit(R5_LOCKED, &dev->flags);
1751 clear_bit(R5_UPTODATE, &dev->flags);
1757 /* keep the parity disk locked while asynchronous operations
1760 set_bit(R5_LOCKED, &sh->dev[pd_idx].flags);
1761 clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
1764 pr_debug("%s: stripe %llu locked: %d pending: %lx\n",
1765 __func__, (unsigned long long)sh->sector,
1766 locked, sh->ops.pending);
1772 * Each stripe/dev can have one or more bion attached.
1773 * toread/towrite point to the first in a chain.
1774 * The bi_next chain must be in order.
1776 static int add_stripe_bio(struct stripe_head *sh, struct bio *bi, int dd_idx, int forwrite)
1779 raid5_conf_t *conf = sh->raid_conf;
1782 pr_debug("adding bh b#%llu to stripe s#%llu\n",
1783 (unsigned long long)bi->bi_sector,
1784 (unsigned long long)sh->sector);
1787 spin_lock(&sh->lock);
1788 spin_lock_irq(&conf->device_lock);
1790 bip = &sh->dev[dd_idx].towrite;
1791 if (*bip == NULL && sh->dev[dd_idx].written == NULL)
1794 bip = &sh->dev[dd_idx].toread;
1795 while (*bip && (*bip)->bi_sector < bi->bi_sector) {
1796 if ((*bip)->bi_sector + ((*bip)->bi_size >> 9) > bi->bi_sector)
1798 bip = & (*bip)->bi_next;
1800 if (*bip && (*bip)->bi_sector < bi->bi_sector + ((bi->bi_size)>>9))
1803 BUG_ON(*bip && bi->bi_next && (*bip) != bi->bi_next);
1807 bi->bi_phys_segments ++;
1808 spin_unlock_irq(&conf->device_lock);
1809 spin_unlock(&sh->lock);
1811 pr_debug("added bi b#%llu to stripe s#%llu, disk %d.\n",
1812 (unsigned long long)bi->bi_sector,
1813 (unsigned long long)sh->sector, dd_idx);
1815 if (conf->mddev->bitmap && firstwrite) {
1816 bitmap_startwrite(conf->mddev->bitmap, sh->sector,
1818 sh->bm_seq = conf->seq_flush+1;
1819 set_bit(STRIPE_BIT_DELAY, &sh->state);
1823 /* check if page is covered */
1824 sector_t sector = sh->dev[dd_idx].sector;
1825 for (bi=sh->dev[dd_idx].towrite;
1826 sector < sh->dev[dd_idx].sector + STRIPE_SECTORS &&
1827 bi && bi->bi_sector <= sector;
1828 bi = r5_next_bio(bi, sh->dev[dd_idx].sector)) {
1829 if (bi->bi_sector + (bi->bi_size>>9) >= sector)
1830 sector = bi->bi_sector + (bi->bi_size>>9);
1832 if (sector >= sh->dev[dd_idx].sector + STRIPE_SECTORS)
1833 set_bit(R5_OVERWRITE, &sh->dev[dd_idx].flags);
1838 set_bit(R5_Overlap, &sh->dev[dd_idx].flags);
1839 spin_unlock_irq(&conf->device_lock);
1840 spin_unlock(&sh->lock);
1844 static void end_reshape(raid5_conf_t *conf);
1846 static int page_is_zero(struct page *p)
1848 char *a = page_address(p);
1849 return ((*(u32*)a) == 0 &&
1850 memcmp(a, a+4, STRIPE_SIZE-4)==0);
1853 static int stripe_to_pdidx(sector_t stripe, raid5_conf_t *conf, int disks)
1855 int sectors_per_chunk = conf->chunk_size >> 9;
1857 int chunk_offset = sector_div(stripe, sectors_per_chunk);
1859 raid5_compute_sector(stripe * (disks - conf->max_degraded)
1860 *sectors_per_chunk + chunk_offset,
1861 disks, disks - conf->max_degraded,
1862 &dd_idx, &pd_idx, conf);
1867 handle_requests_to_failed_array(raid5_conf_t *conf, struct stripe_head *sh,
1868 struct stripe_head_state *s, int disks,
1869 struct bio **return_bi)
1872 for (i = disks; i--; ) {
1876 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
1879 rdev = rcu_dereference(conf->disks[i].rdev);
1880 if (rdev && test_bit(In_sync, &rdev->flags))
1881 /* multiple read failures in one stripe */
1882 md_error(conf->mddev, rdev);
1885 spin_lock_irq(&conf->device_lock);
1886 /* fail all writes first */
1887 bi = sh->dev[i].towrite;
1888 sh->dev[i].towrite = NULL;
1894 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
1895 wake_up(&conf->wait_for_overlap);
1897 while (bi && bi->bi_sector <
1898 sh->dev[i].sector + STRIPE_SECTORS) {
1899 struct bio *nextbi = r5_next_bio(bi, sh->dev[i].sector);
1900 clear_bit(BIO_UPTODATE, &bi->bi_flags);
1901 if (--bi->bi_phys_segments == 0) {
1902 md_write_end(conf->mddev);
1903 bi->bi_next = *return_bi;
1908 /* and fail all 'written' */
1909 bi = sh->dev[i].written;
1910 sh->dev[i].written = NULL;
1911 if (bi) bitmap_end = 1;
1912 while (bi && bi->bi_sector <
1913 sh->dev[i].sector + STRIPE_SECTORS) {
1914 struct bio *bi2 = r5_next_bio(bi, sh->dev[i].sector);
1915 clear_bit(BIO_UPTODATE, &bi->bi_flags);
1916 if (--bi->bi_phys_segments == 0) {
1917 md_write_end(conf->mddev);
1918 bi->bi_next = *return_bi;
1924 /* fail any reads if this device is non-operational and
1925 * the data has not reached the cache yet.
1927 if (!test_bit(R5_Wantfill, &sh->dev[i].flags) &&
1928 (!test_bit(R5_Insync, &sh->dev[i].flags) ||
1929 test_bit(R5_ReadError, &sh->dev[i].flags))) {
1930 bi = sh->dev[i].toread;
1931 sh->dev[i].toread = NULL;
1932 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
1933 wake_up(&conf->wait_for_overlap);
1934 if (bi) s->to_read--;
1935 while (bi && bi->bi_sector <
1936 sh->dev[i].sector + STRIPE_SECTORS) {
1937 struct bio *nextbi =
1938 r5_next_bio(bi, sh->dev[i].sector);
1939 clear_bit(BIO_UPTODATE, &bi->bi_flags);
1940 if (--bi->bi_phys_segments == 0) {
1941 bi->bi_next = *return_bi;
1947 spin_unlock_irq(&conf->device_lock);
1949 bitmap_endwrite(conf->mddev->bitmap, sh->sector,
1950 STRIPE_SECTORS, 0, 0);
1953 if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
1954 if (atomic_dec_and_test(&conf->pending_full_writes))
1955 md_wakeup_thread(conf->mddev->thread);
1958 /* __handle_issuing_new_read_requests5 - returns 0 if there are no more disks
1961 static int __handle_issuing_new_read_requests5(struct stripe_head *sh,
1962 struct stripe_head_state *s, int disk_idx, int disks)
1964 struct r5dev *dev = &sh->dev[disk_idx];
1965 struct r5dev *failed_dev = &sh->dev[s->failed_num];
1967 /* don't schedule compute operations or reads on the parity block while
1968 * a check is in flight
1970 if ((disk_idx == sh->pd_idx) &&
1971 test_bit(STRIPE_OP_CHECK, &sh->ops.pending))
1974 /* is the data in this block needed, and can we get it? */
1975 if (!test_bit(R5_LOCKED, &dev->flags) &&
1976 !test_bit(R5_UPTODATE, &dev->flags) && (dev->toread ||
1977 (dev->towrite && !test_bit(R5_OVERWRITE, &dev->flags)) ||
1978 s->syncing || s->expanding || (s->failed &&
1979 (failed_dev->toread || (failed_dev->towrite &&
1980 !test_bit(R5_OVERWRITE, &failed_dev->flags)
1982 /* 1/ We would like to get this block, possibly by computing it,
1983 * but we might not be able to.
1985 * 2/ Since parity check operations potentially make the parity
1986 * block !uptodate it will need to be refreshed before any
1987 * compute operations on data disks are scheduled.
1989 * 3/ We hold off parity block re-reads until check operations
1992 if ((s->uptodate == disks - 1) &&
1993 (s->failed && disk_idx == s->failed_num) &&
1994 !test_bit(STRIPE_OP_CHECK, &sh->ops.pending)) {
1995 set_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.pending);
1996 set_bit(R5_Wantcompute, &dev->flags);
1997 sh->ops.target = disk_idx;
2000 /* Careful: from this point on 'uptodate' is in the eye
2001 * of raid5_run_ops which services 'compute' operations
2002 * before writes. R5_Wantcompute flags a block that will
2003 * be R5_UPTODATE by the time it is needed for a
2004 * subsequent operation.
2007 return 0; /* uptodate + compute == disks */
2008 } else if ((s->uptodate < disks - 1) &&
2009 test_bit(R5_Insync, &dev->flags)) {
2010 /* Note: we hold off compute operations while checks are
2011 * in flight, but we still prefer 'compute' over 'read'
2012 * hence we only read if (uptodate < * disks-1)
2014 set_bit(R5_LOCKED, &dev->flags);
2015 set_bit(R5_Wantread, &dev->flags);
2016 if (!test_and_set_bit(STRIPE_OP_IO, &sh->ops.pending))
2019 pr_debug("Reading block %d (sync=%d)\n", disk_idx,
2027 static void handle_issuing_new_read_requests5(struct stripe_head *sh,
2028 struct stripe_head_state *s, int disks)
2032 /* Clear completed compute operations. Parity recovery
2033 * (STRIPE_OP_MOD_REPAIR_PD) implies a write-back which is handled
2034 * later on in this routine
2036 if (test_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.complete) &&
2037 !test_bit(STRIPE_OP_MOD_REPAIR_PD, &sh->ops.pending)) {
2038 clear_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.complete);
2039 clear_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.ack);
2040 clear_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.pending);
2043 /* look for blocks to read/compute, skip this if a compute
2044 * is already in flight, or if the stripe contents are in the
2045 * midst of changing due to a write
2047 if (!test_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.pending) &&
2048 !test_bit(STRIPE_OP_PREXOR, &sh->ops.pending) &&
2049 !test_bit(STRIPE_OP_POSTXOR, &sh->ops.pending)) {
2050 for (i = disks; i--; )
2051 if (__handle_issuing_new_read_requests5(
2052 sh, s, i, disks) == 0)
2055 set_bit(STRIPE_HANDLE, &sh->state);
2058 static void handle_issuing_new_read_requests6(struct stripe_head *sh,
2059 struct stripe_head_state *s, struct r6_state *r6s,
2063 for (i = disks; i--; ) {
2064 struct r5dev *dev = &sh->dev[i];
2065 if (!test_bit(R5_LOCKED, &dev->flags) &&
2066 !test_bit(R5_UPTODATE, &dev->flags) &&
2067 (dev->toread || (dev->towrite &&
2068 !test_bit(R5_OVERWRITE, &dev->flags)) ||
2069 s->syncing || s->expanding ||
2071 (sh->dev[r6s->failed_num[0]].toread ||
2074 (sh->dev[r6s->failed_num[1]].toread ||
2076 /* we would like to get this block, possibly
2077 * by computing it, but we might not be able to
2079 if ((s->uptodate == disks - 1) &&
2080 (s->failed && (i == r6s->failed_num[0] ||
2081 i == r6s->failed_num[1]))) {
2082 pr_debug("Computing stripe %llu block %d\n",
2083 (unsigned long long)sh->sector, i);
2084 compute_block_1(sh, i, 0);
2086 } else if ( s->uptodate == disks-2 && s->failed >= 2 ) {
2087 /* Computing 2-failure is *very* expensive; only
2088 * do it if failed >= 2
2091 for (other = disks; other--; ) {
2094 if (!test_bit(R5_UPTODATE,
2095 &sh->dev[other].flags))
2099 pr_debug("Computing stripe %llu blocks %d,%d\n",
2100 (unsigned long long)sh->sector,
2102 compute_block_2(sh, i, other);
2104 } else if (test_bit(R5_Insync, &dev->flags)) {
2105 set_bit(R5_LOCKED, &dev->flags);
2106 set_bit(R5_Wantread, &dev->flags);
2108 pr_debug("Reading block %d (sync=%d)\n",
2113 set_bit(STRIPE_HANDLE, &sh->state);
2117 /* handle_completed_write_requests
2118 * any written block on an uptodate or failed drive can be returned.
2119 * Note that if we 'wrote' to a failed drive, it will be UPTODATE, but
2120 * never LOCKED, so we don't need to test 'failed' directly.
2122 static void handle_completed_write_requests(raid5_conf_t *conf,
2123 struct stripe_head *sh, int disks, struct bio **return_bi)
2128 for (i = disks; i--; )
2129 if (sh->dev[i].written) {
2131 if (!test_bit(R5_LOCKED, &dev->flags) &&
2132 test_bit(R5_UPTODATE, &dev->flags)) {
2133 /* We can return any write requests */
2134 struct bio *wbi, *wbi2;
2136 pr_debug("Return write for disc %d\n", i);
2137 spin_lock_irq(&conf->device_lock);
2139 dev->written = NULL;
2140 while (wbi && wbi->bi_sector <
2141 dev->sector + STRIPE_SECTORS) {
2142 wbi2 = r5_next_bio(wbi, dev->sector);
2143 if (--wbi->bi_phys_segments == 0) {
2144 md_write_end(conf->mddev);
2145 wbi->bi_next = *return_bi;
2150 if (dev->towrite == NULL)
2152 spin_unlock_irq(&conf->device_lock);
2154 bitmap_endwrite(conf->mddev->bitmap,
2157 !test_bit(STRIPE_DEGRADED, &sh->state),
2162 if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
2163 if (atomic_dec_and_test(&conf->pending_full_writes))
2164 md_wakeup_thread(conf->mddev->thread);
2167 static void handle_issuing_new_write_requests5(raid5_conf_t *conf,
2168 struct stripe_head *sh, struct stripe_head_state *s, int disks)
2170 int rmw = 0, rcw = 0, i;
2171 for (i = disks; i--; ) {
2172 /* would I have to read this buffer for read_modify_write */
2173 struct r5dev *dev = &sh->dev[i];
2174 if ((dev->towrite || i == sh->pd_idx) &&
2175 !test_bit(R5_LOCKED, &dev->flags) &&
2176 !(test_bit(R5_UPTODATE, &dev->flags) ||
2177 test_bit(R5_Wantcompute, &dev->flags))) {
2178 if (test_bit(R5_Insync, &dev->flags))
2181 rmw += 2*disks; /* cannot read it */
2183 /* Would I have to read this buffer for reconstruct_write */
2184 if (!test_bit(R5_OVERWRITE, &dev->flags) && i != sh->pd_idx &&
2185 !test_bit(R5_LOCKED, &dev->flags) &&
2186 !(test_bit(R5_UPTODATE, &dev->flags) ||
2187 test_bit(R5_Wantcompute, &dev->flags))) {
2188 if (test_bit(R5_Insync, &dev->flags)) rcw++;
2193 pr_debug("for sector %llu, rmw=%d rcw=%d\n",
2194 (unsigned long long)sh->sector, rmw, rcw);
2195 set_bit(STRIPE_HANDLE, &sh->state);
2196 if (rmw < rcw && rmw > 0)
2197 /* prefer read-modify-write, but need to get some data */
2198 for (i = disks; i--; ) {
2199 struct r5dev *dev = &sh->dev[i];
2200 if ((dev->towrite || i == sh->pd_idx) &&
2201 !test_bit(R5_LOCKED, &dev->flags) &&
2202 !(test_bit(R5_UPTODATE, &dev->flags) ||
2203 test_bit(R5_Wantcompute, &dev->flags)) &&
2204 test_bit(R5_Insync, &dev->flags)) {
2206 test_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2207 pr_debug("Read_old block "
2208 "%d for r-m-w\n", i);
2209 set_bit(R5_LOCKED, &dev->flags);
2210 set_bit(R5_Wantread, &dev->flags);
2211 if (!test_and_set_bit(
2212 STRIPE_OP_IO, &sh->ops.pending))
2216 set_bit(STRIPE_DELAYED, &sh->state);
2217 set_bit(STRIPE_HANDLE, &sh->state);
2221 if (rcw <= rmw && rcw > 0)
2222 /* want reconstruct write, but need to get some data */
2223 for (i = disks; i--; ) {
2224 struct r5dev *dev = &sh->dev[i];
2225 if (!test_bit(R5_OVERWRITE, &dev->flags) &&
2227 !test_bit(R5_LOCKED, &dev->flags) &&
2228 !(test_bit(R5_UPTODATE, &dev->flags) ||
2229 test_bit(R5_Wantcompute, &dev->flags)) &&
2230 test_bit(R5_Insync, &dev->flags)) {
2232 test_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2233 pr_debug("Read_old block "
2234 "%d for Reconstruct\n", i);
2235 set_bit(R5_LOCKED, &dev->flags);
2236 set_bit(R5_Wantread, &dev->flags);
2237 if (!test_and_set_bit(
2238 STRIPE_OP_IO, &sh->ops.pending))
2242 set_bit(STRIPE_DELAYED, &sh->state);
2243 set_bit(STRIPE_HANDLE, &sh->state);
2247 /* now if nothing is locked, and if we have enough data,
2248 * we can start a write request
2250 /* since handle_stripe can be called at any time we need to handle the
2251 * case where a compute block operation has been submitted and then a
2252 * subsequent call wants to start a write request. raid5_run_ops only
2253 * handles the case where compute block and postxor are requested
2254 * simultaneously. If this is not the case then new writes need to be
2255 * held off until the compute completes.
2257 if ((s->req_compute ||
2258 !test_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.pending)) &&
2259 (s->locked == 0 && (rcw == 0 || rmw == 0) &&
2260 !test_bit(STRIPE_BIT_DELAY, &sh->state)))
2261 s->locked += handle_write_operations5(sh, rcw == 0, 0);
2264 static void handle_issuing_new_write_requests6(raid5_conf_t *conf,
2265 struct stripe_head *sh, struct stripe_head_state *s,
2266 struct r6_state *r6s, int disks)
2268 int rcw = 0, must_compute = 0, pd_idx = sh->pd_idx, i;
2269 int qd_idx = r6s->qd_idx;
2270 for (i = disks; i--; ) {
2271 struct r5dev *dev = &sh->dev[i];
2272 /* Would I have to read this buffer for reconstruct_write */
2273 if (!test_bit(R5_OVERWRITE, &dev->flags)
2274 && i != pd_idx && i != qd_idx
2275 && (!test_bit(R5_LOCKED, &dev->flags)
2277 !test_bit(R5_UPTODATE, &dev->flags)) {
2278 if (test_bit(R5_Insync, &dev->flags)) rcw++;
2280 pr_debug("raid6: must_compute: "
2281 "disk %d flags=%#lx\n", i, dev->flags);
2286 pr_debug("for sector %llu, rcw=%d, must_compute=%d\n",
2287 (unsigned long long)sh->sector, rcw, must_compute);
2288 set_bit(STRIPE_HANDLE, &sh->state);
2291 /* want reconstruct write, but need to get some data */
2292 for (i = disks; i--; ) {
2293 struct r5dev *dev = &sh->dev[i];
2294 if (!test_bit(R5_OVERWRITE, &dev->flags)
2295 && !(s->failed == 0 && (i == pd_idx || i == qd_idx))
2296 && !test_bit(R5_LOCKED, &dev->flags) &&
2297 !test_bit(R5_UPTODATE, &dev->flags) &&
2298 test_bit(R5_Insync, &dev->flags)) {
2300 test_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2301 pr_debug("Read_old stripe %llu "
2302 "block %d for Reconstruct\n",
2303 (unsigned long long)sh->sector, i);
2304 set_bit(R5_LOCKED, &dev->flags);
2305 set_bit(R5_Wantread, &dev->flags);
2308 pr_debug("Request delayed stripe %llu "
2309 "block %d for Reconstruct\n",
2310 (unsigned long long)sh->sector, i);
2311 set_bit(STRIPE_DELAYED, &sh->state);
2312 set_bit(STRIPE_HANDLE, &sh->state);
2316 /* now if nothing is locked, and if we have enough data, we can start a
2319 if (s->locked == 0 && rcw == 0 &&
2320 !test_bit(STRIPE_BIT_DELAY, &sh->state)) {
2321 if (must_compute > 0) {
2322 /* We have failed blocks and need to compute them */
2323 switch (s->failed) {
2327 compute_block_1(sh, r6s->failed_num[0], 0);
2330 compute_block_2(sh, r6s->failed_num[0],
2331 r6s->failed_num[1]);
2333 default: /* This request should have been failed? */
2338 pr_debug("Computing parity for stripe %llu\n",
2339 (unsigned long long)sh->sector);
2340 compute_parity6(sh, RECONSTRUCT_WRITE);
2341 /* now every locked buffer is ready to be written */
2342 for (i = disks; i--; )
2343 if (test_bit(R5_LOCKED, &sh->dev[i].flags)) {
2344 pr_debug("Writing stripe %llu block %d\n",
2345 (unsigned long long)sh->sector, i);
2347 set_bit(R5_Wantwrite, &sh->dev[i].flags);
2349 if (s->locked == disks)
2350 if (!test_and_set_bit(STRIPE_FULL_WRITE, &sh->state))
2351 atomic_inc(&conf->pending_full_writes);
2352 /* after a RECONSTRUCT_WRITE, the stripe MUST be in-sync */
2353 set_bit(STRIPE_INSYNC, &sh->state);
2355 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2356 atomic_dec(&conf->preread_active_stripes);
2357 if (atomic_read(&conf->preread_active_stripes) <
2359 md_wakeup_thread(conf->mddev->thread);
2364 static void handle_parity_checks5(raid5_conf_t *conf, struct stripe_head *sh,
2365 struct stripe_head_state *s, int disks)
2367 int canceled_check = 0;
2369 set_bit(STRIPE_HANDLE, &sh->state);
2371 /* complete a check operation */
2372 if (test_and_clear_bit(STRIPE_OP_CHECK, &sh->ops.complete)) {
2373 clear_bit(STRIPE_OP_CHECK, &sh->ops.ack);
2374 clear_bit(STRIPE_OP_CHECK, &sh->ops.pending);
2375 if (s->failed == 0) {
2376 if (sh->ops.zero_sum_result == 0)
2377 /* parity is correct (on disc,
2378 * not in buffer any more)
2380 set_bit(STRIPE_INSYNC, &sh->state);
2382 conf->mddev->resync_mismatches +=
2385 MD_RECOVERY_CHECK, &conf->mddev->recovery))
2386 /* don't try to repair!! */
2387 set_bit(STRIPE_INSYNC, &sh->state);
2389 set_bit(STRIPE_OP_COMPUTE_BLK,
2391 set_bit(STRIPE_OP_MOD_REPAIR_PD,
2393 set_bit(R5_Wantcompute,
2394 &sh->dev[sh->pd_idx].flags);
2395 sh->ops.target = sh->pd_idx;
2401 canceled_check = 1; /* STRIPE_INSYNC is not set */
2404 /* start a new check operation if there are no failures, the stripe is
2405 * not insync, and a repair is not in flight
2407 if (s->failed == 0 &&
2408 !test_bit(STRIPE_INSYNC, &sh->state) &&
2409 !test_bit(STRIPE_OP_MOD_REPAIR_PD, &sh->ops.pending)) {
2410 if (!test_and_set_bit(STRIPE_OP_CHECK, &sh->ops.pending)) {
2411 BUG_ON(s->uptodate != disks);
2412 clear_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags);
2418 /* check if we can clear a parity disk reconstruct */
2419 if (test_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.complete) &&
2420 test_bit(STRIPE_OP_MOD_REPAIR_PD, &sh->ops.pending)) {
2422 clear_bit(STRIPE_OP_MOD_REPAIR_PD, &sh->ops.pending);
2423 clear_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.complete);
2424 clear_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.ack);
2425 clear_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.pending);
2429 /* Wait for check parity and compute block operations to complete
2430 * before write-back. If a failure occurred while the check operation
2431 * was in flight we need to cycle this stripe through handle_stripe
2432 * since the parity block may not be uptodate
2434 if (!canceled_check && !test_bit(STRIPE_INSYNC, &sh->state) &&
2435 !test_bit(STRIPE_OP_CHECK, &sh->ops.pending) &&
2436 !test_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.pending)) {
2438 /* either failed parity check, or recovery is happening */
2440 s->failed_num = sh->pd_idx;
2441 dev = &sh->dev[s->failed_num];
2442 BUG_ON(!test_bit(R5_UPTODATE, &dev->flags));
2443 BUG_ON(s->uptodate != disks);
2445 set_bit(R5_LOCKED, &dev->flags);
2446 set_bit(R5_Wantwrite, &dev->flags);
2447 if (!test_and_set_bit(STRIPE_OP_IO, &sh->ops.pending))
2450 clear_bit(STRIPE_DEGRADED, &sh->state);
2452 set_bit(STRIPE_INSYNC, &sh->state);
2457 static void handle_parity_checks6(raid5_conf_t *conf, struct stripe_head *sh,
2458 struct stripe_head_state *s,
2459 struct r6_state *r6s, struct page *tmp_page,
2462 int update_p = 0, update_q = 0;
2464 int pd_idx = sh->pd_idx;
2465 int qd_idx = r6s->qd_idx;
2467 set_bit(STRIPE_HANDLE, &sh->state);
2469 BUG_ON(s->failed > 2);
2470 BUG_ON(s->uptodate < disks);
2471 /* Want to check and possibly repair P and Q.
2472 * However there could be one 'failed' device, in which
2473 * case we can only check one of them, possibly using the
2474 * other to generate missing data
2477 /* If !tmp_page, we cannot do the calculations,
2478 * but as we have set STRIPE_HANDLE, we will soon be called
2479 * by stripe_handle with a tmp_page - just wait until then.
2482 if (s->failed == r6s->q_failed) {
2483 /* The only possible failed device holds 'Q', so it
2484 * makes sense to check P (If anything else were failed,
2485 * we would have used P to recreate it).
2487 compute_block_1(sh, pd_idx, 1);
2488 if (!page_is_zero(sh->dev[pd_idx].page)) {
2489 compute_block_1(sh, pd_idx, 0);
2493 if (!r6s->q_failed && s->failed < 2) {
2494 /* q is not failed, and we didn't use it to generate
2495 * anything, so it makes sense to check it
2497 memcpy(page_address(tmp_page),
2498 page_address(sh->dev[qd_idx].page),
2500 compute_parity6(sh, UPDATE_PARITY);
2501 if (memcmp(page_address(tmp_page),
2502 page_address(sh->dev[qd_idx].page),
2503 STRIPE_SIZE) != 0) {
2504 clear_bit(STRIPE_INSYNC, &sh->state);
2508 if (update_p || update_q) {
2509 conf->mddev->resync_mismatches += STRIPE_SECTORS;
2510 if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery))
2511 /* don't try to repair!! */
2512 update_p = update_q = 0;
2515 /* now write out any block on a failed drive,
2516 * or P or Q if they need it
2519 if (s->failed == 2) {
2520 dev = &sh->dev[r6s->failed_num[1]];
2522 set_bit(R5_LOCKED, &dev->flags);
2523 set_bit(R5_Wantwrite, &dev->flags);
2525 if (s->failed >= 1) {
2526 dev = &sh->dev[r6s->failed_num[0]];
2528 set_bit(R5_LOCKED, &dev->flags);
2529 set_bit(R5_Wantwrite, &dev->flags);
2533 dev = &sh->dev[pd_idx];
2535 set_bit(R5_LOCKED, &dev->flags);
2536 set_bit(R5_Wantwrite, &dev->flags);
2539 dev = &sh->dev[qd_idx];
2541 set_bit(R5_LOCKED, &dev->flags);
2542 set_bit(R5_Wantwrite, &dev->flags);
2544 clear_bit(STRIPE_DEGRADED, &sh->state);
2546 set_bit(STRIPE_INSYNC, &sh->state);
2550 static void handle_stripe_expansion(raid5_conf_t *conf, struct stripe_head *sh,
2551 struct r6_state *r6s)
2555 /* We have read all the blocks in this stripe and now we need to
2556 * copy some of them into a target stripe for expand.
2558 struct dma_async_tx_descriptor *tx = NULL;
2559 clear_bit(STRIPE_EXPAND_SOURCE, &sh->state);
2560 for (i = 0; i < sh->disks; i++)
2561 if (i != sh->pd_idx && (!r6s || i != r6s->qd_idx)) {
2562 int dd_idx, pd_idx, j;
2563 struct stripe_head *sh2;
2565 sector_t bn = compute_blocknr(sh, i);
2566 sector_t s = raid5_compute_sector(bn, conf->raid_disks,
2568 conf->max_degraded, &dd_idx,
2570 sh2 = get_active_stripe(conf, s, conf->raid_disks,
2573 /* so far only the early blocks of this stripe
2574 * have been requested. When later blocks
2575 * get requested, we will try again
2578 if (!test_bit(STRIPE_EXPANDING, &sh2->state) ||
2579 test_bit(R5_Expanded, &sh2->dev[dd_idx].flags)) {
2580 /* must have already done this block */
2581 release_stripe(sh2);
2585 /* place all the copies on one channel */
2586 tx = async_memcpy(sh2->dev[dd_idx].page,
2587 sh->dev[i].page, 0, 0, STRIPE_SIZE,
2588 ASYNC_TX_DEP_ACK, tx, NULL, NULL);
2590 set_bit(R5_Expanded, &sh2->dev[dd_idx].flags);
2591 set_bit(R5_UPTODATE, &sh2->dev[dd_idx].flags);
2592 for (j = 0; j < conf->raid_disks; j++)
2593 if (j != sh2->pd_idx &&
2594 (!r6s || j != raid6_next_disk(sh2->pd_idx,
2596 !test_bit(R5_Expanded, &sh2->dev[j].flags))
2598 if (j == conf->raid_disks) {
2599 set_bit(STRIPE_EXPAND_READY, &sh2->state);
2600 set_bit(STRIPE_HANDLE, &sh2->state);
2602 release_stripe(sh2);
2605 /* done submitting copies, wait for them to complete */
2608 dma_wait_for_async_tx(tx);
2614 * handle_stripe - do things to a stripe.
2616 * We lock the stripe and then examine the state of various bits
2617 * to see what needs to be done.
2619 * return some read request which now have data
2620 * return some write requests which are safely on disc
2621 * schedule a read on some buffers
2622 * schedule a write of some buffers
2623 * return confirmation of parity correctness
2625 * buffers are taken off read_list or write_list, and bh_cache buffers
2626 * get BH_Lock set before the stripe lock is released.
2630 static void handle_stripe5(struct stripe_head *sh)
2632 raid5_conf_t *conf = sh->raid_conf;
2633 int disks = sh->disks, i;
2634 struct bio *return_bi = NULL;
2635 struct stripe_head_state s;
2637 unsigned long pending = 0;
2638 mdk_rdev_t *blocked_rdev = NULL;
2641 memset(&s, 0, sizeof(s));
2642 pr_debug("handling stripe %llu, state=%#lx cnt=%d, pd_idx=%d "
2643 "ops=%lx:%lx:%lx\n", (unsigned long long)sh->sector, sh->state,
2644 atomic_read(&sh->count), sh->pd_idx,
2645 sh->ops.pending, sh->ops.ack, sh->ops.complete);
2647 spin_lock(&sh->lock);
2648 clear_bit(STRIPE_HANDLE, &sh->state);
2649 clear_bit(STRIPE_DELAYED, &sh->state);
2651 s.syncing = test_bit(STRIPE_SYNCING, &sh->state);
2652 s.expanding = test_bit(STRIPE_EXPAND_SOURCE, &sh->state);
2653 s.expanded = test_bit(STRIPE_EXPAND_READY, &sh->state);
2654 /* Now to look around and see what can be done */
2656 /* clean-up completed biofill operations */
2657 if (test_bit(STRIPE_OP_BIOFILL, &sh->ops.complete)) {
2658 clear_bit(STRIPE_OP_BIOFILL, &sh->ops.pending);
2659 clear_bit(STRIPE_OP_BIOFILL, &sh->ops.ack);
2660 clear_bit(STRIPE_OP_BIOFILL, &sh->ops.complete);
2664 for (i=disks; i--; ) {
2666 struct r5dev *dev = &sh->dev[i];
2667 clear_bit(R5_Insync, &dev->flags);
2669 pr_debug("check %d: state 0x%lx toread %p read %p write %p "
2670 "written %p\n", i, dev->flags, dev->toread, dev->read,
2671 dev->towrite, dev->written);
2673 /* maybe we can request a biofill operation
2675 * new wantfill requests are only permitted while
2676 * STRIPE_OP_BIOFILL is clear
2678 if (test_bit(R5_UPTODATE, &dev->flags) && dev->toread &&
2679 !test_bit(STRIPE_OP_BIOFILL, &sh->ops.pending))
2680 set_bit(R5_Wantfill, &dev->flags);
2682 /* now count some things */
2683 if (test_bit(R5_LOCKED, &dev->flags)) s.locked++;
2684 if (test_bit(R5_UPTODATE, &dev->flags)) s.uptodate++;
2685 if (test_bit(R5_Wantcompute, &dev->flags)) s.compute++;
2687 if (test_bit(R5_Wantfill, &dev->flags))
2689 else if (dev->toread)
2693 if (!test_bit(R5_OVERWRITE, &dev->flags))
2698 rdev = rcu_dereference(conf->disks[i].rdev);
2699 if (rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
2700 blocked_rdev = rdev;
2701 atomic_inc(&rdev->nr_pending);
2704 if (!rdev || !test_bit(In_sync, &rdev->flags)) {
2705 /* The ReadError flag will just be confusing now */
2706 clear_bit(R5_ReadError, &dev->flags);
2707 clear_bit(R5_ReWrite, &dev->flags);
2709 if (!rdev || !test_bit(In_sync, &rdev->flags)
2710 || test_bit(R5_ReadError, &dev->flags)) {
2714 set_bit(R5_Insync, &dev->flags);
2718 if (unlikely(blocked_rdev)) {
2719 set_bit(STRIPE_HANDLE, &sh->state);
2723 if (s.to_fill && !test_and_set_bit(STRIPE_OP_BIOFILL, &sh->ops.pending))
2726 pr_debug("locked=%d uptodate=%d to_read=%d"
2727 " to_write=%d failed=%d failed_num=%d\n",
2728 s.locked, s.uptodate, s.to_read, s.to_write,
2729 s.failed, s.failed_num);
2730 /* check if the array has lost two devices and, if so, some requests might
2733 if (s.failed > 1 && s.to_read+s.to_write+s.written)
2734 handle_requests_to_failed_array(conf, sh, &s, disks,
2736 if (s.failed > 1 && s.syncing) {
2737 md_done_sync(conf->mddev, STRIPE_SECTORS,0);
2738 clear_bit(STRIPE_SYNCING, &sh->state);
2742 /* might be able to return some write requests if the parity block
2743 * is safe, or on a failed drive
2745 dev = &sh->dev[sh->pd_idx];
2747 ((test_bit(R5_Insync, &dev->flags) &&
2748 !test_bit(R5_LOCKED, &dev->flags) &&
2749 test_bit(R5_UPTODATE, &dev->flags)) ||
2750 (s.failed == 1 && s.failed_num == sh->pd_idx)))
2751 handle_completed_write_requests(conf, sh, disks, &return_bi);
2753 /* Now we might consider reading some blocks, either to check/generate
2754 * parity, or to satisfy requests
2755 * or to load a block that is being partially written.
2757 if (s.to_read || s.non_overwrite ||
2758 (s.syncing && (s.uptodate + s.compute < disks)) || s.expanding ||
2759 test_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.pending))
2760 handle_issuing_new_read_requests5(sh, &s, disks);
2762 /* Now we check to see if any write operations have recently
2766 /* leave prexor set until postxor is done, allows us to distinguish
2767 * a rmw from a rcw during biodrain
2770 if (test_bit(STRIPE_OP_PREXOR, &sh->ops.complete) &&
2771 test_bit(STRIPE_OP_POSTXOR, &sh->ops.complete)) {
2774 clear_bit(STRIPE_OP_PREXOR, &sh->ops.complete);
2775 clear_bit(STRIPE_OP_PREXOR, &sh->ops.ack);
2776 clear_bit(STRIPE_OP_PREXOR, &sh->ops.pending);
2778 for (i = disks; i--; )
2779 clear_bit(R5_Wantprexor, &sh->dev[i].flags);
2782 /* if only POSTXOR is set then this is an 'expand' postxor */
2783 if (test_bit(STRIPE_OP_BIODRAIN, &sh->ops.complete) &&
2784 test_bit(STRIPE_OP_POSTXOR, &sh->ops.complete)) {
2786 clear_bit(STRIPE_OP_BIODRAIN, &sh->ops.complete);
2787 clear_bit(STRIPE_OP_BIODRAIN, &sh->ops.ack);
2788 clear_bit(STRIPE_OP_BIODRAIN, &sh->ops.pending);
2790 clear_bit(STRIPE_OP_POSTXOR, &sh->ops.complete);
2791 clear_bit(STRIPE_OP_POSTXOR, &sh->ops.ack);
2792 clear_bit(STRIPE_OP_POSTXOR, &sh->ops.pending);
2794 /* All the 'written' buffers and the parity block are ready to
2795 * be written back to disk
2797 BUG_ON(!test_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags));
2798 for (i = disks; i--; ) {
2800 if (test_bit(R5_LOCKED, &dev->flags) &&
2801 (i == sh->pd_idx || dev->written)) {
2802 pr_debug("Writing block %d\n", i);
2803 set_bit(R5_Wantwrite, &dev->flags);
2804 if (!test_and_set_bit(
2805 STRIPE_OP_IO, &sh->ops.pending))
2809 if (!test_bit(R5_Insync, &dev->flags) ||
2810 (i == sh->pd_idx && s.failed == 0))
2811 set_bit(STRIPE_INSYNC, &sh->state);
2814 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2815 atomic_dec(&conf->preread_active_stripes);
2816 if (atomic_read(&conf->preread_active_stripes) <
2818 md_wakeup_thread(conf->mddev->thread);
2822 /* Now to consider new write requests and what else, if anything
2823 * should be read. We do not handle new writes when:
2824 * 1/ A 'write' operation (copy+xor) is already in flight.
2825 * 2/ A 'check' operation is in flight, as it may clobber the parity
2828 if (s.to_write && !test_bit(STRIPE_OP_POSTXOR, &sh->ops.pending) &&
2829 !test_bit(STRIPE_OP_CHECK, &sh->ops.pending))
2830 handle_issuing_new_write_requests5(conf, sh, &s, disks);
2832 /* maybe we need to check and possibly fix the parity for this stripe
2833 * Any reads will already have been scheduled, so we just see if enough
2834 * data is available. The parity check is held off while parity
2835 * dependent operations are in flight.
2837 if ((s.syncing && s.locked == 0 &&
2838 !test_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.pending) &&
2839 !test_bit(STRIPE_INSYNC, &sh->state)) ||
2840 test_bit(STRIPE_OP_CHECK, &sh->ops.pending) ||
2841 test_bit(STRIPE_OP_MOD_REPAIR_PD, &sh->ops.pending))
2842 handle_parity_checks5(conf, sh, &s, disks);
2844 if (s.syncing && s.locked == 0 && test_bit(STRIPE_INSYNC, &sh->state)) {
2845 md_done_sync(conf->mddev, STRIPE_SECTORS,1);
2846 clear_bit(STRIPE_SYNCING, &sh->state);
2849 /* If the failed drive is just a ReadError, then we might need to progress
2850 * the repair/check process
2852 if (s.failed == 1 && !conf->mddev->ro &&
2853 test_bit(R5_ReadError, &sh->dev[s.failed_num].flags)
2854 && !test_bit(R5_LOCKED, &sh->dev[s.failed_num].flags)
2855 && test_bit(R5_UPTODATE, &sh->dev[s.failed_num].flags)
2857 dev = &sh->dev[s.failed_num];
2858 if (!test_bit(R5_ReWrite, &dev->flags)) {
2859 set_bit(R5_Wantwrite, &dev->flags);
2860 if (!test_and_set_bit(STRIPE_OP_IO, &sh->ops.pending))
2862 set_bit(R5_ReWrite, &dev->flags);
2863 set_bit(R5_LOCKED, &dev->flags);
2866 /* let's read it back */
2867 set_bit(R5_Wantread, &dev->flags);
2868 if (!test_and_set_bit(STRIPE_OP_IO, &sh->ops.pending))
2870 set_bit(R5_LOCKED, &dev->flags);
2875 /* Finish postxor operations initiated by the expansion
2878 if (test_bit(STRIPE_OP_POSTXOR, &sh->ops.complete) &&
2879 !test_bit(STRIPE_OP_BIODRAIN, &sh->ops.pending)) {
2881 clear_bit(STRIPE_EXPANDING, &sh->state);
2883 clear_bit(STRIPE_OP_POSTXOR, &sh->ops.pending);
2884 clear_bit(STRIPE_OP_POSTXOR, &sh->ops.ack);
2885 clear_bit(STRIPE_OP_POSTXOR, &sh->ops.complete);
2887 for (i = conf->raid_disks; i--; ) {
2888 set_bit(R5_Wantwrite, &sh->dev[i].flags);
2889 set_bit(R5_LOCKED, &dev->flags);
2891 if (!test_and_set_bit(STRIPE_OP_IO, &sh->ops.pending))
2896 if (s.expanded && test_bit(STRIPE_EXPANDING, &sh->state) &&
2897 !test_bit(STRIPE_OP_POSTXOR, &sh->ops.pending)) {
2898 /* Need to write out all blocks after computing parity */
2899 sh->disks = conf->raid_disks;
2900 sh->pd_idx = stripe_to_pdidx(sh->sector, conf,
2902 s.locked += handle_write_operations5(sh, 1, 1);
2903 } else if (s.expanded &&
2905 !test_bit(STRIPE_OP_POSTXOR, &sh->ops.pending)) {
2906 clear_bit(STRIPE_EXPAND_READY, &sh->state);
2907 atomic_dec(&conf->reshape_stripes);
2908 wake_up(&conf->wait_for_overlap);
2909 md_done_sync(conf->mddev, STRIPE_SECTORS, 1);
2912 if (s.expanding && s.locked == 0 &&
2913 !test_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.pending))
2914 handle_stripe_expansion(conf, sh, NULL);
2917 pending = get_stripe_work(sh);
2920 spin_unlock(&sh->lock);
2922 /* wait for this device to become unblocked */
2923 if (unlikely(blocked_rdev))
2924 md_wait_for_blocked_rdev(blocked_rdev, conf->mddev);
2927 raid5_run_ops(sh, pending);
2929 return_io(return_bi);
2933 static void handle_stripe6(struct stripe_head *sh, struct page *tmp_page)
2935 raid6_conf_t *conf = sh->raid_conf;
2936 int disks = sh->disks;
2937 struct bio *return_bi = NULL;
2938 int i, pd_idx = sh->pd_idx;
2939 struct stripe_head_state s;
2940 struct r6_state r6s;
2941 struct r5dev *dev, *pdev, *qdev;
2942 mdk_rdev_t *blocked_rdev = NULL;
2944 r6s.qd_idx = raid6_next_disk(pd_idx, disks);
2945 pr_debug("handling stripe %llu, state=%#lx cnt=%d, "
2946 "pd_idx=%d, qd_idx=%d\n",
2947 (unsigned long long)sh->sector, sh->state,
2948 atomic_read(&sh->count), pd_idx, r6s.qd_idx);
2949 memset(&s, 0, sizeof(s));
2951 spin_lock(&sh->lock);
2952 clear_bit(STRIPE_HANDLE, &sh->state);
2953 clear_bit(STRIPE_DELAYED, &sh->state);
2955 s.syncing = test_bit(STRIPE_SYNCING, &sh->state);
2956 s.expanding = test_bit(STRIPE_EXPAND_SOURCE, &sh->state);
2957 s.expanded = test_bit(STRIPE_EXPAND_READY, &sh->state);
2958 /* Now to look around and see what can be done */
2961 for (i=disks; i--; ) {
2964 clear_bit(R5_Insync, &dev->flags);
2966 pr_debug("check %d: state 0x%lx read %p write %p written %p\n",
2967 i, dev->flags, dev->toread, dev->towrite, dev->written);
2968 /* maybe we can reply to a read */
2969 if (test_bit(R5_UPTODATE, &dev->flags) && dev->toread) {
2970 struct bio *rbi, *rbi2;
2971 pr_debug("Return read for disc %d\n", i);
2972 spin_lock_irq(&conf->device_lock);
2975 if (test_and_clear_bit(R5_Overlap, &dev->flags))
2976 wake_up(&conf->wait_for_overlap);
2977 spin_unlock_irq(&conf->device_lock);
2978 while (rbi && rbi->bi_sector < dev->sector + STRIPE_SECTORS) {
2979 copy_data(0, rbi, dev->page, dev->sector);
2980 rbi2 = r5_next_bio(rbi, dev->sector);
2981 spin_lock_irq(&conf->device_lock);
2982 if (--rbi->bi_phys_segments == 0) {
2983 rbi->bi_next = return_bi;
2986 spin_unlock_irq(&conf->device_lock);
2991 /* now count some things */
2992 if (test_bit(R5_LOCKED, &dev->flags)) s.locked++;
2993 if (test_bit(R5_UPTODATE, &dev->flags)) s.uptodate++;
3000 if (!test_bit(R5_OVERWRITE, &dev->flags))
3005 rdev = rcu_dereference(conf->disks[i].rdev);
3006 if (rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
3007 blocked_rdev = rdev;
3008 atomic_inc(&rdev->nr_pending);
3011 if (!rdev || !test_bit(In_sync, &rdev->flags)) {
3012 /* The ReadError flag will just be confusing now */
3013 clear_bit(R5_ReadError, &dev->flags);
3014 clear_bit(R5_ReWrite, &dev->flags);
3016 if (!rdev || !test_bit(In_sync, &rdev->flags)
3017 || test_bit(R5_ReadError, &dev->flags)) {
3019 r6s.failed_num[s.failed] = i;
3022 set_bit(R5_Insync, &dev->flags);
3026 if (unlikely(blocked_rdev)) {
3027 set_bit(STRIPE_HANDLE, &sh->state);
3030 pr_debug("locked=%d uptodate=%d to_read=%d"
3031 " to_write=%d failed=%d failed_num=%d,%d\n",
3032 s.locked, s.uptodate, s.to_read, s.to_write, s.failed,
3033 r6s.failed_num[0], r6s.failed_num[1]);
3034 /* check if the array has lost >2 devices and, if so, some requests
3035 * might need to be failed
3037 if (s.failed > 2 && s.to_read+s.to_write+s.written)
3038 handle_requests_to_failed_array(conf, sh, &s, disks,
3040 if (s.failed > 2 && s.syncing) {
3041 md_done_sync(conf->mddev, STRIPE_SECTORS,0);
3042 clear_bit(STRIPE_SYNCING, &sh->state);
3047 * might be able to return some write requests if the parity blocks
3048 * are safe, or on a failed drive
3050 pdev = &sh->dev[pd_idx];
3051 r6s.p_failed = (s.failed >= 1 && r6s.failed_num[0] == pd_idx)
3052 || (s.failed >= 2 && r6s.failed_num[1] == pd_idx);
3053 qdev = &sh->dev[r6s.qd_idx];
3054 r6s.q_failed = (s.failed >= 1 && r6s.failed_num[0] == r6s.qd_idx)
3055 || (s.failed >= 2 && r6s.failed_num[1] == r6s.qd_idx);
3058 ( r6s.p_failed || ((test_bit(R5_Insync, &pdev->flags)
3059 && !test_bit(R5_LOCKED, &pdev->flags)
3060 && test_bit(R5_UPTODATE, &pdev->flags)))) &&
3061 ( r6s.q_failed || ((test_bit(R5_Insync, &qdev->flags)
3062 && !test_bit(R5_LOCKED, &qdev->flags)
3063 && test_bit(R5_UPTODATE, &qdev->flags)))))
3064 handle_completed_write_requests(conf, sh, disks, &return_bi);
3066 /* Now we might consider reading some blocks, either to check/generate
3067 * parity, or to satisfy requests
3068 * or to load a block that is being partially written.
3070 if (s.to_read || s.non_overwrite || (s.to_write && s.failed) ||
3071 (s.syncing && (s.uptodate < disks)) || s.expanding)
3072 handle_issuing_new_read_requests6(sh, &s, &r6s, disks);
3074 /* now to consider writing and what else, if anything should be read */
3076 handle_issuing_new_write_requests6(conf, sh, &s, &r6s, disks);
3078 /* maybe we need to check and possibly fix the parity for this stripe
3079 * Any reads will already have been scheduled, so we just see if enough
3082 if (s.syncing && s.locked == 0 && !test_bit(STRIPE_INSYNC, &sh->state))
3083 handle_parity_checks6(conf, sh, &s, &r6s, tmp_page, disks);
3085 if (s.syncing && s.locked == 0 && test_bit(STRIPE_INSYNC, &sh->state)) {
3086 md_done_sync(conf->mddev, STRIPE_SECTORS,1);
3087 clear_bit(STRIPE_SYNCING, &sh->state);
3090 /* If the failed drives are just a ReadError, then we might need
3091 * to progress the repair/check process
3093 if (s.failed <= 2 && !conf->mddev->ro)
3094 for (i = 0; i < s.failed; i++) {
3095 dev = &sh->dev[r6s.failed_num[i]];
3096 if (test_bit(R5_ReadError, &dev->flags)
3097 && !test_bit(R5_LOCKED, &dev->flags)
3098 && test_bit(R5_UPTODATE, &dev->flags)
3100 if (!test_bit(R5_ReWrite, &dev->flags)) {
3101 set_bit(R5_Wantwrite, &dev->flags);
3102 set_bit(R5_ReWrite, &dev->flags);
3103 set_bit(R5_LOCKED, &dev->flags);
3105 /* let's read it back */
3106 set_bit(R5_Wantread, &dev->flags);
3107 set_bit(R5_LOCKED, &dev->flags);
3112 if (s.expanded && test_bit(STRIPE_EXPANDING, &sh->state)) {
3113 /* Need to write out all blocks after computing P&Q */
3114 sh->disks = conf->raid_disks;
3115 sh->pd_idx = stripe_to_pdidx(sh->sector, conf,
3117 compute_parity6(sh, RECONSTRUCT_WRITE);
3118 for (i = conf->raid_disks ; i-- ; ) {
3119 set_bit(R5_LOCKED, &sh->dev[i].flags);
3121 set_bit(R5_Wantwrite, &sh->dev[i].flags);
3123 clear_bit(STRIPE_EXPANDING, &sh->state);
3124 } else if (s.expanded) {
3125 clear_bit(STRIPE_EXPAND_READY, &sh->state);
3126 atomic_dec(&conf->reshape_stripes);
3127 wake_up(&conf->wait_for_overlap);
3128 md_done_sync(conf->mddev, STRIPE_SECTORS, 1);
3131 if (s.expanding && s.locked == 0 &&
3132 !test_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.pending))
3133 handle_stripe_expansion(conf, sh, &r6s);
3136 spin_unlock(&sh->lock);
3138 /* wait for this device to become unblocked */
3139 if (unlikely(blocked_rdev))
3140 md_wait_for_blocked_rdev(blocked_rdev, conf->mddev);
3142 return_io(return_bi);
3144 for (i=disks; i-- ;) {
3148 if (test_and_clear_bit(R5_Wantwrite, &sh->dev[i].flags))
3150 else if (test_and_clear_bit(R5_Wantread, &sh->dev[i].flags))
3155 set_bit(STRIPE_IO_STARTED, &sh->state);
3157 bi = &sh->dev[i].req;
3161 bi->bi_end_io = raid5_end_write_request;
3163 bi->bi_end_io = raid5_end_read_request;
3166 rdev = rcu_dereference(conf->disks[i].rdev);
3167 if (rdev && test_bit(Faulty, &rdev->flags))
3170 atomic_inc(&rdev->nr_pending);
3174 if (s.syncing || s.expanding || s.expanded)
3175 md_sync_acct(rdev->bdev, STRIPE_SECTORS);
3177 bi->bi_bdev = rdev->bdev;
3178 pr_debug("for %llu schedule op %ld on disc %d\n",
3179 (unsigned long long)sh->sector, bi->bi_rw, i);
3180 atomic_inc(&sh->count);
3181 bi->bi_sector = sh->sector + rdev->data_offset;
3182 bi->bi_flags = 1 << BIO_UPTODATE;
3184 bi->bi_max_vecs = 1;
3186 bi->bi_io_vec = &sh->dev[i].vec;
3187 bi->bi_io_vec[0].bv_len = STRIPE_SIZE;
3188 bi->bi_io_vec[0].bv_offset = 0;
3189 bi->bi_size = STRIPE_SIZE;
3192 test_bit(R5_ReWrite, &sh->dev[i].flags))
3193 atomic_add(STRIPE_SECTORS, &rdev->corrected_errors);
3194 generic_make_request(bi);
3197 set_bit(STRIPE_DEGRADED, &sh->state);
3198 pr_debug("skip op %ld on disc %d for sector %llu\n",
3199 bi->bi_rw, i, (unsigned long long)sh->sector);
3200 clear_bit(R5_LOCKED, &sh->dev[i].flags);
3201 set_bit(STRIPE_HANDLE, &sh->state);
3206 static void handle_stripe(struct stripe_head *sh, struct page *tmp_page)
3208 if (sh->raid_conf->level == 6)
3209 handle_stripe6(sh, tmp_page);
3216 static void raid5_activate_delayed(raid5_conf_t *conf)
3218 if (atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD) {
3219 while (!list_empty(&conf->delayed_list)) {
3220 struct list_head *l = conf->delayed_list.next;
3221 struct stripe_head *sh;
3222 sh = list_entry(l, struct stripe_head, lru);
3224 clear_bit(STRIPE_DELAYED, &sh->state);
3225 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
3226 atomic_inc(&conf->preread_active_stripes);
3227 list_add_tail(&sh->lru, &conf->hold_list);
3230 blk_plug_device(conf->mddev->queue);
3233 static void activate_bit_delay(raid5_conf_t *conf)
3235 /* device_lock is held */
3236 struct list_head head;
3237 list_add(&head, &conf->bitmap_list);
3238 list_del_init(&conf->bitmap_list);
3239 while (!list_empty(&head)) {
3240 struct stripe_head *sh = list_entry(head.next, struct stripe_head, lru);
3241 list_del_init(&sh->lru);
3242 atomic_inc(&sh->count);
3243 __release_stripe(conf, sh);
3247 static void unplug_slaves(mddev_t *mddev)
3249 raid5_conf_t *conf = mddev_to_conf(mddev);
3253 for (i=0; i<mddev->raid_disks; i++) {
3254 mdk_rdev_t *rdev = rcu_dereference(conf->disks[i].rdev);
3255 if (rdev && !test_bit(Faulty, &rdev->flags) && atomic_read(&rdev->nr_pending)) {
3256 struct request_queue *r_queue = bdev_get_queue(rdev->bdev);
3258 atomic_inc(&rdev->nr_pending);
3261 blk_unplug(r_queue);
3263 rdev_dec_pending(rdev, mddev);
3270 static void raid5_unplug_device(struct request_queue *q)
3272 mddev_t *mddev = q->queuedata;
3273 raid5_conf_t *conf = mddev_to_conf(mddev);
3274 unsigned long flags;
3276 spin_lock_irqsave(&conf->device_lock, flags);
3278 if (blk_remove_plug(q)) {
3280 raid5_activate_delayed(conf);
3282 md_wakeup_thread(mddev->thread);
3284 spin_unlock_irqrestore(&conf->device_lock, flags);
3286 unplug_slaves(mddev);
3289 static int raid5_congested(void *data, int bits)
3291 mddev_t *mddev = data;
3292 raid5_conf_t *conf = mddev_to_conf(mddev);
3294 /* No difference between reads and writes. Just check
3295 * how busy the stripe_cache is
3297 if (conf->inactive_blocked)
3301 if (list_empty_careful(&conf->inactive_list))
3307 /* We want read requests to align with chunks where possible,
3308 * but write requests don't need to.
3310 static int raid5_mergeable_bvec(struct request_queue *q, struct bio *bio, struct bio_vec *biovec)
3312 mddev_t *mddev = q->queuedata;
3313 sector_t sector = bio->bi_sector + get_start_sect(bio->bi_bdev);
3315 unsigned int chunk_sectors = mddev->chunk_size >> 9;
3316 unsigned int bio_sectors = bio->bi_size >> 9;
3318 if (bio_data_dir(bio) == WRITE)
3319 return biovec->bv_len; /* always allow writes to be mergeable */
3321 max = (chunk_sectors - ((sector & (chunk_sectors - 1)) + bio_sectors)) << 9;
3322 if (max < 0) max = 0;
3323 if (max <= biovec->bv_len && bio_sectors == 0)
3324 return biovec->bv_len;
3330 static int in_chunk_boundary(mddev_t *mddev, struct bio *bio)
3332 sector_t sector = bio->bi_sector + get_start_sect(bio->bi_bdev);
3333 unsigned int chunk_sectors = mddev->chunk_size >> 9;
3334 unsigned int bio_sectors = bio->bi_size >> 9;
3336 return chunk_sectors >=
3337 ((sector & (chunk_sectors - 1)) + bio_sectors);
3341 * add bio to the retry LIFO ( in O(1) ... we are in interrupt )
3342 * later sampled by raid5d.
3344 static void add_bio_to_retry(struct bio *bi,raid5_conf_t *conf)
3346 unsigned long flags;
3348 spin_lock_irqsave(&conf->device_lock, flags);
3350 bi->bi_next = conf->retry_read_aligned_list;
3351 conf->retry_read_aligned_list = bi;
3353 spin_unlock_irqrestore(&conf->device_lock, flags);
3354 md_wakeup_thread(conf->mddev->thread);
3358 static struct bio *remove_bio_from_retry(raid5_conf_t *conf)
3362 bi = conf->retry_read_aligned;
3364 conf->retry_read_aligned = NULL;
3367 bi = conf->retry_read_aligned_list;
3369 conf->retry_read_aligned_list = bi->bi_next;
3371 bi->bi_phys_segments = 1; /* biased count of active stripes */
3372 bi->bi_hw_segments = 0; /* count of processed stripes */
3380 * The "raid5_align_endio" should check if the read succeeded and if it
3381 * did, call bio_endio on the original bio (having bio_put the new bio
3383 * If the read failed..
3385 static void raid5_align_endio(struct bio *bi, int error)
3387 struct bio* raid_bi = bi->bi_private;
3390 int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
3395 mddev = raid_bi->bi_bdev->bd_disk->queue->queuedata;
3396 conf = mddev_to_conf(mddev);
3397 rdev = (void*)raid_bi->bi_next;
3398 raid_bi->bi_next = NULL;
3400 rdev_dec_pending(rdev, conf->mddev);
3402 if (!error && uptodate) {
3403 bio_endio(raid_bi, 0);
3404 if (atomic_dec_and_test(&conf->active_aligned_reads))
3405 wake_up(&conf->wait_for_stripe);
3410 pr_debug("raid5_align_endio : io error...handing IO for a retry\n");
3412 add_bio_to_retry(raid_bi, conf);
3415 static int bio_fits_rdev(struct bio *bi)
3417 struct request_queue *q = bdev_get_queue(bi->bi_bdev);
3419 if ((bi->bi_size>>9) > q->max_sectors)
3421 blk_recount_segments(q, bi);
3422 if (bi->bi_phys_segments > q->max_phys_segments ||
3423 bi->bi_hw_segments > q->max_hw_segments)
3426 if (q->merge_bvec_fn)
3427 /* it's too hard to apply the merge_bvec_fn at this stage,
3436 static int chunk_aligned_read(struct request_queue *q, struct bio * raid_bio)
3438 mddev_t *mddev = q->queuedata;
3439 raid5_conf_t *conf = mddev_to_conf(mddev);
3440 const unsigned int raid_disks = conf->raid_disks;
3441 const unsigned int data_disks = raid_disks - conf->max_degraded;
3442 unsigned int dd_idx, pd_idx;
3443 struct bio* align_bi;
3446 if (!in_chunk_boundary(mddev, raid_bio)) {
3447 pr_debug("chunk_aligned_read : non aligned\n");
3451 * use bio_clone to make a copy of the bio
3453 align_bi = bio_clone(raid_bio, GFP_NOIO);
3457 * set bi_end_io to a new function, and set bi_private to the
3460 align_bi->bi_end_io = raid5_align_endio;
3461 align_bi->bi_private = raid_bio;
3465 align_bi->bi_sector = raid5_compute_sector(raid_bio->bi_sector,
3473 rdev = rcu_dereference(conf->disks[dd_idx].rdev);
3474 if (rdev && test_bit(In_sync, &rdev->flags)) {
3475 atomic_inc(&rdev->nr_pending);
3477 raid_bio->bi_next = (void*)rdev;
3478 align_bi->bi_bdev = rdev->bdev;
3479 align_bi->bi_flags &= ~(1 << BIO_SEG_VALID);
3480 align_bi->bi_sector += rdev->data_offset;
3482 if (!bio_fits_rdev(align_bi)) {
3483 /* too big in some way */
3485 rdev_dec_pending(rdev, mddev);
3489 spin_lock_irq(&conf->device_lock);
3490 wait_event_lock_irq(conf->wait_for_stripe,
3492 conf->device_lock, /* nothing */);
3493 atomic_inc(&conf->active_aligned_reads);
3494 spin_unlock_irq(&conf->device_lock);
3496 generic_make_request(align_bi);
3505 /* __get_priority_stripe - get the next stripe to process
3507 * Full stripe writes are allowed to pass preread active stripes up until
3508 * the bypass_threshold is exceeded. In general the bypass_count
3509 * increments when the handle_list is handled before the hold_list; however, it
3510 * will not be incremented when STRIPE_IO_STARTED is sampled set signifying a
3511 * stripe with in flight i/o. The bypass_count will be reset when the
3512 * head of the hold_list has changed, i.e. the head was promoted to the
3515 static struct stripe_head *__get_priority_stripe(raid5_conf_t *conf)
3517 struct stripe_head *sh;
3519 pr_debug("%s: handle: %s hold: %s full_writes: %d bypass_count: %d\n",
3521 list_empty(&conf->handle_list) ? "empty" : "busy",
3522 list_empty(&conf->hold_list) ? "empty" : "busy",
3523 atomic_read(&conf->pending_full_writes), conf->bypass_count);
3525 if (!list_empty(&conf->handle_list)) {
3526 sh = list_entry(conf->handle_list.next, typeof(*sh), lru);
3528 if (list_empty(&conf->hold_list))
3529 conf->bypass_count = 0;
3530 else if (!test_bit(STRIPE_IO_STARTED, &sh->state)) {
3531 if (conf->hold_list.next == conf->last_hold)
3532 conf->bypass_count++;
3534 conf->last_hold = conf->hold_list.next;
3535 conf->bypass_count -= conf->bypass_threshold;
3536 if (conf->bypass_count < 0)
3537 conf->bypass_count = 0;
3540 } else if (!list_empty(&conf->hold_list) &&
3541 ((conf->bypass_threshold &&
3542 conf->bypass_count > conf->bypass_threshold) ||
3543 atomic_read(&conf->pending_full_writes) == 0)) {
3544 sh = list_entry(conf->hold_list.next,
3546 conf->bypass_count -= conf->bypass_threshold;
3547 if (conf->bypass_count < 0)
3548 conf->bypass_count = 0;
3552 list_del_init(&sh->lru);
3553 atomic_inc(&sh->count);
3554 BUG_ON(atomic_read(&sh->count) != 1);
3558 static int make_request(struct request_queue *q, struct bio * bi)
3560 mddev_t *mddev = q->queuedata;
3561 raid5_conf_t *conf = mddev_to_conf(mddev);
3562 unsigned int dd_idx, pd_idx;
3563 sector_t new_sector;
3564 sector_t logical_sector, last_sector;
3565 struct stripe_head *sh;
3566 const int rw = bio_data_dir(bi);
3569 if (unlikely(bio_barrier(bi))) {
3570 bio_endio(bi, -EOPNOTSUPP);
3574 md_write_start(mddev, bi);
3576 disk_stat_inc(mddev->gendisk, ios[rw]);
3577 disk_stat_add(mddev->gendisk, sectors[rw], bio_sectors(bi));
3580 mddev->reshape_position == MaxSector &&
3581 chunk_aligned_read(q,bi))
3584 logical_sector = bi->bi_sector & ~((sector_t)STRIPE_SECTORS-1);
3585 last_sector = bi->bi_sector + (bi->bi_size>>9);
3587 bi->bi_phys_segments = 1; /* over-loaded to count active stripes */
3589 for (;logical_sector < last_sector; logical_sector += STRIPE_SECTORS) {
3591 int disks, data_disks;
3594 prepare_to_wait(&conf->wait_for_overlap, &w, TASK_UNINTERRUPTIBLE);
3595 if (likely(conf->expand_progress == MaxSector))
3596 disks = conf->raid_disks;
3598 /* spinlock is needed as expand_progress may be
3599 * 64bit on a 32bit platform, and so it might be
3600 * possible to see a half-updated value
3601 * Ofcourse expand_progress could change after
3602 * the lock is dropped, so once we get a reference
3603 * to the stripe that we think it is, we will have
3606 spin_lock_irq(&conf->device_lock);
3607 disks = conf->raid_disks;
3608 if (logical_sector >= conf->expand_progress)
3609 disks = conf->previous_raid_disks;
3611 if (logical_sector >= conf->expand_lo) {
3612 spin_unlock_irq(&conf->device_lock);
3617 spin_unlock_irq(&conf->device_lock);
3619 data_disks = disks - conf->max_degraded;
3621 new_sector = raid5_compute_sector(logical_sector, disks, data_disks,
3622 &dd_idx, &pd_idx, conf);
3623 pr_debug("raid5: make_request, sector %llu logical %llu\n",
3624 (unsigned long long)new_sector,
3625 (unsigned long long)logical_sector);
3627 sh = get_active_stripe(conf, new_sector, disks, pd_idx, (bi->bi_rw&RWA_MASK));
3629 if (unlikely(conf->expand_progress != MaxSector)) {
3630 /* expansion might have moved on while waiting for a
3631 * stripe, so we must do the range check again.
3632 * Expansion could still move past after this
3633 * test, but as we are holding a reference to
3634 * 'sh', we know that if that happens,
3635 * STRIPE_EXPANDING will get set and the expansion
3636 * won't proceed until we finish with the stripe.
3639 spin_lock_irq(&conf->device_lock);
3640 if (logical_sector < conf->expand_progress &&
3641 disks == conf->previous_raid_disks)
3642 /* mismatch, need to try again */
3644 spin_unlock_irq(&conf->device_lock);
3650 /* FIXME what if we get a false positive because these
3651 * are being updated.
3653 if (logical_sector >= mddev->suspend_lo &&
3654 logical_sector < mddev->suspend_hi) {
3660 if (test_bit(STRIPE_EXPANDING, &sh->state) ||
3661 !add_stripe_bio(sh, bi, dd_idx, (bi->bi_rw&RW_MASK))) {
3662 /* Stripe is busy expanding or
3663 * add failed due to overlap. Flush everything
3666 raid5_unplug_device(mddev->queue);
3671 finish_wait(&conf->wait_for_overlap, &w);
3672 set_bit(STRIPE_HANDLE, &sh->state);
3673 clear_bit(STRIPE_DELAYED, &sh->state);
3676 /* cannot get stripe for read-ahead, just give-up */
3677 clear_bit(BIO_UPTODATE, &bi->bi_flags);
3678 finish_wait(&conf->wait_for_overlap, &w);
3683 spin_lock_irq(&conf->device_lock);
3684 remaining = --bi->bi_phys_segments;
3685 spin_unlock_irq(&conf->device_lock);
3686 if (remaining == 0) {
3689 md_write_end(mddev);
3696 static sector_t reshape_request(mddev_t *mddev, sector_t sector_nr, int *skipped)
3698 /* reshaping is quite different to recovery/resync so it is
3699 * handled quite separately ... here.
3701 * On each call to sync_request, we gather one chunk worth of
3702 * destination stripes and flag them as expanding.
3703 * Then we find all the source stripes and request reads.
3704 * As the reads complete, handle_stripe will copy the data
3705 * into the destination stripe and release that stripe.
3707 raid5_conf_t *conf = (raid5_conf_t *) mddev->private;
3708 struct stripe_head *sh;
3710 sector_t first_sector, last_sector;
3711 int raid_disks = conf->previous_raid_disks;
3712 int data_disks = raid_disks - conf->max_degraded;
3713 int new_data_disks = conf->raid_disks - conf->max_degraded;
3716 sector_t writepos, safepos, gap;
3718 if (sector_nr == 0 &&
3719 conf->expand_progress != 0) {
3720 /* restarting in the middle, skip the initial sectors */
3721 sector_nr = conf->expand_progress;
3722 sector_div(sector_nr, new_data_disks);
3727 /* we update the metadata when there is more than 3Meg
3728 * in the block range (that is rather arbitrary, should
3729 * probably be time based) or when the data about to be
3730 * copied would over-write the source of the data at
3731 * the front of the range.
3732 * i.e. one new_stripe forward from expand_progress new_maps
3733 * to after where expand_lo old_maps to
3735 writepos = conf->expand_progress +
3736 conf->chunk_size/512*(new_data_disks);
3737 sector_div(writepos, new_data_disks);
3738 safepos = conf->expand_lo;
3739 sector_div(safepos, data_disks);
3740 gap = conf->expand_progress - conf->expand_lo;
3742 if (writepos >= safepos ||
3743 gap > (new_data_disks)*3000*2 /*3Meg*/) {
3744 /* Cannot proceed until we've updated the superblock... */
3745 wait_event(conf->wait_for_overlap,
3746 atomic_read(&conf->reshape_stripes)==0);
3747 mddev->reshape_position = conf->expand_progress;
3748 set_bit(MD_CHANGE_DEVS, &mddev->flags);
3749 md_wakeup_thread(mddev->thread);
3750 wait_event(mddev->sb_wait, mddev->flags == 0 ||
3751 kthread_should_stop());
3752 spin_lock_irq(&conf->device_lock);
3753 conf->expand_lo = mddev->reshape_position;
3754 spin_unlock_irq(&conf->device_lock);
3755 wake_up(&conf->wait_for_overlap);
3758 for (i=0; i < conf->chunk_size/512; i+= STRIPE_SECTORS) {
3761 pd_idx = stripe_to_pdidx(sector_nr+i, conf, conf->raid_disks);
3762 sh = get_active_stripe(conf, sector_nr+i,
3763 conf->raid_disks, pd_idx, 0);
3764 set_bit(STRIPE_EXPANDING, &sh->state);
3765 atomic_inc(&conf->reshape_stripes);
3766 /* If any of this stripe is beyond the end of the old
3767 * array, then we need to zero those blocks
3769 for (j=sh->disks; j--;) {
3771 if (j == sh->pd_idx)
3773 if (conf->level == 6 &&
3774 j == raid6_next_disk(sh->pd_idx, sh->disks))
3776 s = compute_blocknr(sh, j);
3777 if (s < (mddev->array_size<<1)) {
3781 memset(page_address(sh->dev[j].page), 0, STRIPE_SIZE);
3782 set_bit(R5_Expanded, &sh->dev[j].flags);
3783 set_bit(R5_UPTODATE, &sh->dev[j].flags);
3786 set_bit(STRIPE_EXPAND_READY, &sh->state);
3787 set_bit(STRIPE_HANDLE, &sh->state);
3791 spin_lock_irq(&conf->device_lock);
3792 conf->expand_progress = (sector_nr + i) * new_data_disks;
3793 spin_unlock_irq(&conf->device_lock);
3794 /* Ok, those stripe are ready. We can start scheduling
3795 * reads on the source stripes.
3796 * The source stripes are determined by mapping the first and last
3797 * block on the destination stripes.
3800 raid5_compute_sector(sector_nr*(new_data_disks),
3801 raid_disks, data_disks,
3802 &dd_idx, &pd_idx, conf);
3804 raid5_compute_sector((sector_nr+conf->chunk_size/512)
3805 *(new_data_disks) -1,
3806 raid_disks, data_disks,
3807 &dd_idx, &pd_idx, conf);
3808 if (last_sector >= (mddev->size<<1))
3809 last_sector = (mddev->size<<1)-1;
3810 while (first_sector <= last_sector) {
3811 pd_idx = stripe_to_pdidx(first_sector, conf,
3812 conf->previous_raid_disks);
3813 sh = get_active_stripe(conf, first_sector,
3814 conf->previous_raid_disks, pd_idx, 0);
3815 set_bit(STRIPE_EXPAND_SOURCE, &sh->state);
3816 set_bit(STRIPE_HANDLE, &sh->state);
3818 first_sector += STRIPE_SECTORS;
3820 /* If this takes us to the resync_max point where we have to pause,
3821 * then we need to write out the superblock.
3823 sector_nr += conf->chunk_size>>9;
3824 if (sector_nr >= mddev->resync_max) {
3825 /* Cannot proceed until we've updated the superblock... */
3826 wait_event(conf->wait_for_overlap,
3827 atomic_read(&conf->reshape_stripes) == 0);
3828 mddev->reshape_position = conf->expand_progress;
3829 set_bit(MD_CHANGE_DEVS, &mddev->flags);
3830 md_wakeup_thread(mddev->thread);
3831 wait_event(mddev->sb_wait,
3832 !test_bit(MD_CHANGE_DEVS, &mddev->flags)
3833 || kthread_should_stop());
3834 spin_lock_irq(&conf->device_lock);
3835 conf->expand_lo = mddev->reshape_position;
3836 spin_unlock_irq(&conf->device_lock);
3837 wake_up(&conf->wait_for_overlap);
3839 return conf->chunk_size>>9;
3842 /* FIXME go_faster isn't used */
3843 static inline sector_t sync_request(mddev_t *mddev, sector_t sector_nr, int *skipped, int go_faster)
3845 raid5_conf_t *conf = (raid5_conf_t *) mddev->private;
3846 struct stripe_head *sh;
3848 int raid_disks = conf->raid_disks;
3849 sector_t max_sector = mddev->size << 1;
3851 int still_degraded = 0;
3854 if (sector_nr >= max_sector) {
3855 /* just being told to finish up .. nothing much to do */
3856 unplug_slaves(mddev);
3857 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery)) {
3862 if (mddev->curr_resync < max_sector) /* aborted */
3863 bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
3865 else /* completed sync */
3867 bitmap_close_sync(mddev->bitmap);
3872 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
3873 return reshape_request(mddev, sector_nr, skipped);
3875 /* No need to check resync_max as we never do more than one
3876 * stripe, and as resync_max will always be on a chunk boundary,
3877 * if the check in md_do_sync didn't fire, there is no chance
3878 * of overstepping resync_max here
3881 /* if there is too many failed drives and we are trying
3882 * to resync, then assert that we are finished, because there is
3883 * nothing we can do.
3885 if (mddev->degraded >= conf->max_degraded &&
3886 test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
3887 sector_t rv = (mddev->size << 1) - sector_nr;
3891 if (!bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
3892 !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
3893 !conf->fullsync && sync_blocks >= STRIPE_SECTORS) {
3894 /* we can skip this block, and probably more */
3895 sync_blocks /= STRIPE_SECTORS;
3897 return sync_blocks * STRIPE_SECTORS; /* keep things rounded to whole stripes */
3901 bitmap_cond_end_sync(mddev->bitmap, sector_nr);
3903 pd_idx = stripe_to_pdidx(sector_nr, conf, raid_disks);
3904 sh = get_active_stripe(conf, sector_nr, raid_disks, pd_idx, 1);
3906 sh = get_active_stripe(conf, sector_nr, raid_disks, pd_idx, 0);
3907 /* make sure we don't swamp the stripe cache if someone else
3908 * is trying to get access
3910 schedule_timeout_uninterruptible(1);
3912 /* Need to check if array will still be degraded after recovery/resync
3913 * We don't need to check the 'failed' flag as when that gets set,
3916 for (i=0; i<mddev->raid_disks; i++)
3917 if (conf->disks[i].rdev == NULL)
3920 bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, still_degraded);
3922 spin_lock(&sh->lock);
3923 set_bit(STRIPE_SYNCING, &sh->state);
3924 clear_bit(STRIPE_INSYNC, &sh->state);
3925 spin_unlock(&sh->lock);
3927 handle_stripe(sh, NULL);
3930 return STRIPE_SECTORS;
3933 static int retry_aligned_read(raid5_conf_t *conf, struct bio *raid_bio)
3935 /* We may not be able to submit a whole bio at once as there
3936 * may not be enough stripe_heads available.
3937 * We cannot pre-allocate enough stripe_heads as we may need
3938 * more than exist in the cache (if we allow ever large chunks).
3939 * So we do one stripe head at a time and record in
3940 * ->bi_hw_segments how many have been done.
3942 * We *know* that this entire raid_bio is in one chunk, so
3943 * it will be only one 'dd_idx' and only need one call to raid5_compute_sector.
3945 struct stripe_head *sh;
3947 sector_t sector, logical_sector, last_sector;
3952 logical_sector = raid_bio->bi_sector & ~((sector_t)STRIPE_SECTORS-1);
3953 sector = raid5_compute_sector( logical_sector,
3955 conf->raid_disks - conf->max_degraded,
3959 last_sector = raid_bio->bi_sector + (raid_bio->bi_size>>9);
3961 for (; logical_sector < last_sector;
3962 logical_sector += STRIPE_SECTORS,
3963 sector += STRIPE_SECTORS,
3966 if (scnt < raid_bio->bi_hw_segments)
3967 /* already done this stripe */
3970 sh = get_active_stripe(conf, sector, conf->raid_disks, pd_idx, 1);
3973 /* failed to get a stripe - must wait */
3974 raid_bio->bi_hw_segments = scnt;
3975 conf->retry_read_aligned = raid_bio;
3979 set_bit(R5_ReadError, &sh->dev[dd_idx].flags);
3980 if (!add_stripe_bio(sh, raid_bio, dd_idx, 0)) {
3982 raid_bio->bi_hw_segments = scnt;
3983 conf->retry_read_aligned = raid_bio;
3987 handle_stripe(sh, NULL);
3991 spin_lock_irq(&conf->device_lock);
3992 remaining = --raid_bio->bi_phys_segments;
3993 spin_unlock_irq(&conf->device_lock);
3995 bio_endio(raid_bio, 0);
3996 if (atomic_dec_and_test(&conf->active_aligned_reads))
3997 wake_up(&conf->wait_for_stripe);
4004 * This is our raid5 kernel thread.
4006 * We scan the hash table for stripes which can be handled now.
4007 * During the scan, completed stripes are saved for us by the interrupt
4008 * handler, so that they will not have to wait for our next wakeup.
4010 static void raid5d(mddev_t *mddev)
4012 struct stripe_head *sh;
4013 raid5_conf_t *conf = mddev_to_conf(mddev);
4016 pr_debug("+++ raid5d active\n");
4018 md_check_recovery(mddev);
4021 spin_lock_irq(&conf->device_lock);
4025 if (conf->seq_flush != conf->seq_write) {
4026 int seq = conf->seq_flush;
4027 spin_unlock_irq(&conf->device_lock);
4028 bitmap_unplug(mddev->bitmap);
4029 spin_lock_irq(&conf->device_lock);
4030 conf->seq_write = seq;
4031 activate_bit_delay(conf);
4034 while ((bio = remove_bio_from_retry(conf))) {
4036 spin_unlock_irq(&conf->device_lock);
4037 ok = retry_aligned_read(conf, bio);
4038 spin_lock_irq(&conf->device_lock);
4044 sh = __get_priority_stripe(conf);
4047 async_tx_issue_pending_all();
4050 spin_unlock_irq(&conf->device_lock);
4053 handle_stripe(sh, conf->spare_page);
4056 spin_lock_irq(&conf->device_lock);
4058 pr_debug("%d stripes handled\n", handled);
4060 spin_unlock_irq(&conf->device_lock);
4062 unplug_slaves(mddev);
4064 pr_debug("--- raid5d inactive\n");
4068 raid5_show_stripe_cache_size(mddev_t *mddev, char *page)
4070 raid5_conf_t *conf = mddev_to_conf(mddev);
4072 return sprintf(page, "%d\n", conf->max_nr_stripes);
4078 raid5_store_stripe_cache_size(mddev_t *mddev, const char *page, size_t len)
4080 raid5_conf_t *conf = mddev_to_conf(mddev);
4082 if (len >= PAGE_SIZE)
4087 if (strict_strtoul(page, 10, &new))
4089 if (new <= 16 || new > 32768)
4091 while (new < conf->max_nr_stripes) {
4092 if (drop_one_stripe(conf))
4093 conf->max_nr_stripes--;
4097 md_allow_write(mddev);
4098 while (new > conf->max_nr_stripes) {
4099 if (grow_one_stripe(conf))
4100 conf->max_nr_stripes++;
4106 static struct md_sysfs_entry
4107 raid5_stripecache_size = __ATTR(stripe_cache_size, S_IRUGO | S_IWUSR,
4108 raid5_show_stripe_cache_size,
4109 raid5_store_stripe_cache_size);
4112 raid5_show_preread_threshold(mddev_t *mddev, char *page)
4114 raid5_conf_t *conf = mddev_to_conf(mddev);
4116 return sprintf(page, "%d\n", conf->bypass_threshold);
4122 raid5_store_preread_threshold(mddev_t *mddev, const char *page, size_t len)
4124 raid5_conf_t *conf = mddev_to_conf(mddev);
4126 if (len >= PAGE_SIZE)
4131 if (strict_strtoul(page, 10, &new))
4133 if (new > conf->max_nr_stripes)
4135 conf->bypass_threshold = new;
4139 static struct md_sysfs_entry
4140 raid5_preread_bypass_threshold = __ATTR(preread_bypass_threshold,
4142 raid5_show_preread_threshold,
4143 raid5_store_preread_threshold);
4146 stripe_cache_active_show(mddev_t *mddev, char *page)
4148 raid5_conf_t *conf = mddev_to_conf(mddev);
4150 return sprintf(page, "%d\n", atomic_read(&conf->active_stripes));
4155 static struct md_sysfs_entry
4156 raid5_stripecache_active = __ATTR_RO(stripe_cache_active);
4158 static struct attribute *raid5_attrs[] = {
4159 &raid5_stripecache_size.attr,
4160 &raid5_stripecache_active.attr,
4161 &raid5_preread_bypass_threshold.attr,
4164 static struct attribute_group raid5_attrs_group = {
4166 .attrs = raid5_attrs,
4169 static int run(mddev_t *mddev)
4172 int raid_disk, memory;
4174 struct disk_info *disk;
4175 struct list_head *tmp;
4176 int working_disks = 0;
4178 if (mddev->level != 5 && mddev->level != 4 && mddev->level != 6) {
4179 printk(KERN_ERR "raid5: %s: raid level not set to 4/5/6 (%d)\n",
4180 mdname(mddev), mddev->level);
4184 if (mddev->reshape_position != MaxSector) {
4185 /* Check that we can continue the reshape.
4186 * Currently only disks can change, it must
4187 * increase, and we must be past the point where
4188 * a stripe over-writes itself
4190 sector_t here_new, here_old;
4192 int max_degraded = (mddev->level == 5 ? 1 : 2);
4194 if (mddev->new_level != mddev->level ||
4195 mddev->new_layout != mddev->layout ||
4196 mddev->new_chunk != mddev->chunk_size) {
4197 printk(KERN_ERR "raid5: %s: unsupported reshape "
4198 "required - aborting.\n",
4202 if (mddev->delta_disks <= 0) {
4203 printk(KERN_ERR "raid5: %s: unsupported reshape "
4204 "(reduce disks) required - aborting.\n",
4208 old_disks = mddev->raid_disks - mddev->delta_disks;
4209 /* reshape_position must be on a new-stripe boundary, and one
4210 * further up in new geometry must map after here in old
4213 here_new = mddev->reshape_position;
4214 if (sector_div(here_new, (mddev->chunk_size>>9)*
4215 (mddev->raid_disks - max_degraded))) {
4216 printk(KERN_ERR "raid5: reshape_position not "
4217 "on a stripe boundary\n");
4220 /* here_new is the stripe we will write to */
4221 here_old = mddev->reshape_position;
4222 sector_div(here_old, (mddev->chunk_size>>9)*
4223 (old_disks-max_degraded));
4224 /* here_old is the first stripe that we might need to read
4226 if (here_new >= here_old) {
4227 /* Reading from the same stripe as writing to - bad */
4228 printk(KERN_ERR "raid5: reshape_position too early for "
4229 "auto-recovery - aborting.\n");
4232 printk(KERN_INFO "raid5: reshape will continue\n");
4233 /* OK, we should be able to continue; */
4237 mddev->private = kzalloc(sizeof (raid5_conf_t), GFP_KERNEL);
4238 if ((conf = mddev->private) == NULL)
4240 if (mddev->reshape_position == MaxSector) {
4241 conf->previous_raid_disks = conf->raid_disks = mddev->raid_disks;
4243 conf->raid_disks = mddev->raid_disks;
4244 conf->previous_raid_disks = mddev->raid_disks - mddev->delta_disks;
4247 conf->disks = kzalloc(conf->raid_disks * sizeof(struct disk_info),
4252 conf->mddev = mddev;
4254 if ((conf->stripe_hashtbl = kzalloc(PAGE_SIZE, GFP_KERNEL)) == NULL)
4257 if (mddev->level == 6) {
4258 conf->spare_page = alloc_page(GFP_KERNEL);
4259 if (!conf->spare_page)
4262 spin_lock_init(&conf->device_lock);
4263 mddev->queue->queue_lock = &conf->device_lock;
4264 init_waitqueue_head(&conf->wait_for_stripe);
4265 init_waitqueue_head(&conf->wait_for_overlap);
4266 INIT_LIST_HEAD(&conf->handle_list);
4267 INIT_LIST_HEAD(&conf->hold_list);
4268 INIT_LIST_HEAD(&conf->delayed_list);
4269 INIT_LIST_HEAD(&conf->bitmap_list);
4270 INIT_LIST_HEAD(&conf->inactive_list);
4271 atomic_set(&conf->active_stripes, 0);
4272 atomic_set(&conf->preread_active_stripes, 0);
4273 atomic_set(&conf->active_aligned_reads, 0);
4274 conf->bypass_threshold = BYPASS_THRESHOLD;
4276 pr_debug("raid5: run(%s) called.\n", mdname(mddev));
4278 rdev_for_each(rdev, tmp, mddev) {
4279 raid_disk = rdev->raid_disk;
4280 if (raid_disk >= conf->raid_disks
4283 disk = conf->disks + raid_disk;
4287 if (test_bit(In_sync, &rdev->flags)) {
4288 char b[BDEVNAME_SIZE];
4289 printk(KERN_INFO "raid5: device %s operational as raid"
4290 " disk %d\n", bdevname(rdev->bdev,b),
4294 /* Cannot rely on bitmap to complete recovery */
4299 * 0 for a fully functional array, 1 or 2 for a degraded array.
4301 mddev->degraded = conf->raid_disks - working_disks;
4302 conf->mddev = mddev;
4303 conf->chunk_size = mddev->chunk_size;
4304 conf->level = mddev->level;
4305 if (conf->level == 6)
4306 conf->max_degraded = 2;
4308 conf->max_degraded = 1;
4309 conf->algorithm = mddev->layout;
4310 conf->max_nr_stripes = NR_STRIPES;
4311 conf->expand_progress = mddev->reshape_position;
4313 /* device size must be a multiple of chunk size */
4314 mddev->size &= ~(mddev->chunk_size/1024 -1);
4315 mddev->resync_max_sectors = mddev->size << 1;
4317 if (conf->level == 6 && conf->raid_disks < 4) {
4318 printk(KERN_ERR "raid6: not enough configured devices for %s (%d, minimum 4)\n",
4319 mdname(mddev), conf->raid_disks);
4322 if (!conf->chunk_size || conf->chunk_size % 4) {
4323 printk(KERN_ERR "raid5: invalid chunk size %d for %s\n",
4324 conf->chunk_size, mdname(mddev));
4327 if (conf->algorithm > ALGORITHM_RIGHT_SYMMETRIC) {
4329 "raid5: unsupported parity algorithm %d for %s\n",
4330 conf->algorithm, mdname(mddev));
4333 if (mddev->degraded > conf->max_degraded) {
4334 printk(KERN_ERR "raid5: not enough operational devices for %s"
4335 " (%d/%d failed)\n",
4336 mdname(mddev), mddev->degraded, conf->raid_disks);
4340 if (mddev->degraded > 0 &&
4341 mddev->recovery_cp != MaxSector) {
4342 if (mddev->ok_start_degraded)
4344 "raid5: starting dirty degraded array: %s"
4345 "- data corruption possible.\n",
4349 "raid5: cannot start dirty degraded array for %s\n",
4356 mddev->thread = md_register_thread(raid5d, mddev, "%s_raid5");
4357 if (!mddev->thread) {
4359 "raid5: couldn't allocate thread for %s\n",
4364 memory = conf->max_nr_stripes * (sizeof(struct stripe_head) +
4365 conf->raid_disks * ((sizeof(struct bio) + PAGE_SIZE))) / 1024;
4366 if (grow_stripes(conf, conf->max_nr_stripes)) {
4368 "raid5: couldn't allocate %dkB for buffers\n", memory);
4369 shrink_stripes(conf);
4370 md_unregister_thread(mddev->thread);
4373 printk(KERN_INFO "raid5: allocated %dkB for %s\n",
4374 memory, mdname(mddev));
4376 if (mddev->degraded == 0)
4377 printk("raid5: raid level %d set %s active with %d out of %d"
4378 " devices, algorithm %d\n", conf->level, mdname(mddev),
4379 mddev->raid_disks-mddev->degraded, mddev->raid_disks,
4382 printk(KERN_ALERT "raid5: raid level %d set %s active with %d"
4383 " out of %d devices, algorithm %d\n", conf->level,
4384 mdname(mddev), mddev->raid_disks - mddev->degraded,
4385 mddev->raid_disks, conf->algorithm);
4387 print_raid5_conf(conf);
4389 if (conf->expand_progress != MaxSector) {
4390 printk("...ok start reshape thread\n");
4391 conf->expand_lo = conf->expand_progress;
4392 atomic_set(&conf->reshape_stripes, 0);
4393 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
4394 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
4395 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
4396 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
4397 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
4401 /* read-ahead size must cover two whole stripes, which is
4402 * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
4405 int data_disks = conf->previous_raid_disks - conf->max_degraded;
4406 int stripe = data_disks *
4407 (mddev->chunk_size / PAGE_SIZE);
4408 if (mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
4409 mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
4412 /* Ok, everything is just fine now */
4413 if (sysfs_create_group(&mddev->kobj, &raid5_attrs_group))
4415 "raid5: failed to create sysfs attributes for %s\n",
4418 mddev->queue->unplug_fn = raid5_unplug_device;
4419 mddev->queue->backing_dev_info.congested_data = mddev;
4420 mddev->queue->backing_dev_info.congested_fn = raid5_congested;
4422 mddev->array_size = mddev->size * (conf->previous_raid_disks -
4423 conf->max_degraded);
4425 blk_queue_merge_bvec(mddev->queue, raid5_mergeable_bvec);
4430 print_raid5_conf(conf);
4431 safe_put_page(conf->spare_page);
4433 kfree(conf->stripe_hashtbl);
4436 mddev->private = NULL;
4437 printk(KERN_ALERT "raid5: failed to run raid set %s\n", mdname(mddev));
4443 static int stop(mddev_t *mddev)
4445 raid5_conf_t *conf = (raid5_conf_t *) mddev->private;
4447 md_unregister_thread(mddev->thread);
4448 mddev->thread = NULL;
4449 shrink_stripes(conf);
4450 kfree(conf->stripe_hashtbl);
4451 mddev->queue->backing_dev_info.congested_fn = NULL;
4452 blk_sync_queue(mddev->queue); /* the unplug fn references 'conf'*/
4453 sysfs_remove_group(&mddev->kobj, &raid5_attrs_group);
4456 mddev->private = NULL;
4461 static void print_sh (struct seq_file *seq, struct stripe_head *sh)
4465 seq_printf(seq, "sh %llu, pd_idx %d, state %ld.\n",
4466 (unsigned long long)sh->sector, sh->pd_idx, sh->state);
4467 seq_printf(seq, "sh %llu, count %d.\n",
4468 (unsigned long long)sh->sector, atomic_read(&sh->count));
4469 seq_printf(seq, "sh %llu, ", (unsigned long long)sh->sector);
4470 for (i = 0; i < sh->disks; i++) {
4471 seq_printf(seq, "(cache%d: %p %ld) ",
4472 i, sh->dev[i].page, sh->dev[i].flags);
4474 seq_printf(seq, "\n");
4477 static void printall (struct seq_file *seq, raid5_conf_t *conf)
4479 struct stripe_head *sh;
4480 struct hlist_node *hn;
4483 spin_lock_irq(&conf->device_lock);
4484 for (i = 0; i < NR_HASH; i++) {
4485 hlist_for_each_entry(sh, hn, &conf->stripe_hashtbl[i], hash) {
4486 if (sh->raid_conf != conf)
4491 spin_unlock_irq(&conf->device_lock);
4495 static void status (struct seq_file *seq, mddev_t *mddev)
4497 raid5_conf_t *conf = (raid5_conf_t *) mddev->private;
4500 seq_printf (seq, " level %d, %dk chunk, algorithm %d", mddev->level, mddev->chunk_size >> 10, mddev->layout);
4501 seq_printf (seq, " [%d/%d] [", conf->raid_disks, conf->raid_disks - mddev->degraded);
4502 for (i = 0; i < conf->raid_disks; i++)
4503 seq_printf (seq, "%s",
4504 conf->disks[i].rdev &&
4505 test_bit(In_sync, &conf->disks[i].rdev->flags) ? "U" : "_");
4506 seq_printf (seq, "]");
4508 seq_printf (seq, "\n");
4509 printall(seq, conf);
4513 static void print_raid5_conf (raid5_conf_t *conf)
4516 struct disk_info *tmp;
4518 printk("RAID5 conf printout:\n");
4520 printk("(conf==NULL)\n");
4523 printk(" --- rd:%d wd:%d\n", conf->raid_disks,
4524 conf->raid_disks - conf->mddev->degraded);
4526 for (i = 0; i < conf->raid_disks; i++) {
4527 char b[BDEVNAME_SIZE];
4528 tmp = conf->disks + i;
4530 printk(" disk %d, o:%d, dev:%s\n",
4531 i, !test_bit(Faulty, &tmp->rdev->flags),
4532 bdevname(tmp->rdev->bdev,b));
4536 static int raid5_spare_active(mddev_t *mddev)
4539 raid5_conf_t *conf = mddev->private;
4540 struct disk_info *tmp;
4542 for (i = 0; i < conf->raid_disks; i++) {
4543 tmp = conf->disks + i;
4545 && !test_bit(Faulty, &tmp->rdev->flags)
4546 && !test_and_set_bit(In_sync, &tmp->rdev->flags)) {
4547 unsigned long flags;
4548 spin_lock_irqsave(&conf->device_lock, flags);
4550 spin_unlock_irqrestore(&conf->device_lock, flags);
4553 print_raid5_conf(conf);
4557 static int raid5_remove_disk(mddev_t *mddev, int number)
4559 raid5_conf_t *conf = mddev->private;
4562 struct disk_info *p = conf->disks + number;
4564 print_raid5_conf(conf);
4567 if (test_bit(In_sync, &rdev->flags) ||
4568 atomic_read(&rdev->nr_pending)) {
4572 /* Only remove non-faulty devices if recovery
4575 if (!test_bit(Faulty, &rdev->flags) &&
4576 mddev->degraded <= conf->max_degraded) {
4582 if (atomic_read(&rdev->nr_pending)) {
4583 /* lost the race, try later */
4590 print_raid5_conf(conf);
4594 static int raid5_add_disk(mddev_t *mddev, mdk_rdev_t *rdev)
4596 raid5_conf_t *conf = mddev->private;
4599 struct disk_info *p;
4601 int last = conf->raid_disks - 1;
4603 if (mddev->degraded > conf->max_degraded)
4604 /* no point adding a device */
4607 if (rdev->raid_disk >= 0)
4608 first = last = rdev->raid_disk;
4611 * find the disk ... but prefer rdev->saved_raid_disk
4614 if (rdev->saved_raid_disk >= 0 &&
4615 rdev->saved_raid_disk >= first &&
4616 conf->disks[rdev->saved_raid_disk].rdev == NULL)
4617 disk = rdev->saved_raid_disk;
4620 for ( ; disk <= last ; disk++)
4621 if ((p=conf->disks + disk)->rdev == NULL) {
4622 clear_bit(In_sync, &rdev->flags);
4623 rdev->raid_disk = disk;
4625 if (rdev->saved_raid_disk != disk)
4627 rcu_assign_pointer(p->rdev, rdev);
4630 print_raid5_conf(conf);
4634 static int raid5_resize(mddev_t *mddev, sector_t sectors)
4636 /* no resync is happening, and there is enough space
4637 * on all devices, so we can resize.
4638 * We need to make sure resync covers any new space.
4639 * If the array is shrinking we should possibly wait until
4640 * any io in the removed space completes, but it hardly seems
4643 raid5_conf_t *conf = mddev_to_conf(mddev);
4645 sectors &= ~((sector_t)mddev->chunk_size/512 - 1);
4646 mddev->array_size = (sectors * (mddev->raid_disks-conf->max_degraded))>>1;
4647 set_capacity(mddev->gendisk, mddev->array_size << 1);
4649 if (sectors/2 > mddev->size && mddev->recovery_cp == MaxSector) {
4650 mddev->recovery_cp = mddev->size << 1;
4651 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
4653 mddev->size = sectors /2;
4654 mddev->resync_max_sectors = sectors;
4658 #ifdef CONFIG_MD_RAID5_RESHAPE
4659 static int raid5_check_reshape(mddev_t *mddev)
4661 raid5_conf_t *conf = mddev_to_conf(mddev);
4664 if (mddev->delta_disks < 0 ||
4665 mddev->new_level != mddev->level)
4666 return -EINVAL; /* Cannot shrink array or change level yet */
4667 if (mddev->delta_disks == 0)
4668 return 0; /* nothing to do */
4670 /* Can only proceed if there are plenty of stripe_heads.
4671 * We need a minimum of one full stripe,, and for sensible progress
4672 * it is best to have about 4 times that.
4673 * If we require 4 times, then the default 256 4K stripe_heads will
4674 * allow for chunk sizes up to 256K, which is probably OK.
4675 * If the chunk size is greater, user-space should request more
4676 * stripe_heads first.
4678 if ((mddev->chunk_size / STRIPE_SIZE) * 4 > conf->max_nr_stripes ||
4679 (mddev->new_chunk / STRIPE_SIZE) * 4 > conf->max_nr_stripes) {
4680 printk(KERN_WARNING "raid5: reshape: not enough stripes. Needed %lu\n",
4681 (mddev->chunk_size / STRIPE_SIZE)*4);
4685 err = resize_stripes(conf, conf->raid_disks + mddev->delta_disks);
4689 if (mddev->degraded > conf->max_degraded)
4691 /* looks like we might be able to manage this */
4695 static int raid5_start_reshape(mddev_t *mddev)
4697 raid5_conf_t *conf = mddev_to_conf(mddev);
4699 struct list_head *rtmp;
4701 int added_devices = 0;
4702 unsigned long flags;
4704 if (test_bit(MD_RECOVERY_RUNNING, &mddev->recovery))
4707 rdev_for_each(rdev, rtmp, mddev)
4708 if (rdev->raid_disk < 0 &&
4709 !test_bit(Faulty, &rdev->flags))
4712 if (spares - mddev->degraded < mddev->delta_disks - conf->max_degraded)
4713 /* Not enough devices even to make a degraded array
4718 atomic_set(&conf->reshape_stripes, 0);
4719 spin_lock_irq(&conf->device_lock);
4720 conf->previous_raid_disks = conf->raid_disks;
4721 conf->raid_disks += mddev->delta_disks;
4722 conf->expand_progress = 0;
4723 conf->expand_lo = 0;
4724 spin_unlock_irq(&conf->device_lock);
4726 /* Add some new drives, as many as will fit.
4727 * We know there are enough to make the newly sized array work.
4729 rdev_for_each(rdev, rtmp, mddev)
4730 if (rdev->raid_disk < 0 &&
4731 !test_bit(Faulty, &rdev->flags)) {
4732 if (raid5_add_disk(mddev, rdev) == 0) {
4734 set_bit(In_sync, &rdev->flags);
4736 rdev->recovery_offset = 0;
4737 sprintf(nm, "rd%d", rdev->raid_disk);
4738 if (sysfs_create_link(&mddev->kobj,
4741 "raid5: failed to create "
4742 " link %s for %s\n",
4748 spin_lock_irqsave(&conf->device_lock, flags);
4749 mddev->degraded = (conf->raid_disks - conf->previous_raid_disks) - added_devices;
4750 spin_unlock_irqrestore(&conf->device_lock, flags);
4751 mddev->raid_disks = conf->raid_disks;
4752 mddev->reshape_position = 0;
4753 set_bit(MD_CHANGE_DEVS, &mddev->flags);
4755 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
4756 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
4757 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
4758 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
4759 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
4761 if (!mddev->sync_thread) {
4762 mddev->recovery = 0;
4763 spin_lock_irq(&conf->device_lock);
4764 mddev->raid_disks = conf->raid_disks = conf->previous_raid_disks;
4765 conf->expand_progress = MaxSector;
4766 spin_unlock_irq(&conf->device_lock);
4769 md_wakeup_thread(mddev->sync_thread);
4770 md_new_event(mddev);
4775 static void end_reshape(raid5_conf_t *conf)
4777 struct block_device *bdev;
4779 if (!test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery)) {
4780 conf->mddev->array_size = conf->mddev->size *
4781 (conf->raid_disks - conf->max_degraded);
4782 set_capacity(conf->mddev->gendisk, conf->mddev->array_size << 1);
4783 conf->mddev->changed = 1;
4785 bdev = bdget_disk(conf->mddev->gendisk, 0);
4787 mutex_lock(&bdev->bd_inode->i_mutex);
4788 i_size_write(bdev->bd_inode, (loff_t)conf->mddev->array_size << 10);
4789 mutex_unlock(&bdev->bd_inode->i_mutex);
4792 spin_lock_irq(&conf->device_lock);
4793 conf->expand_progress = MaxSector;
4794 spin_unlock_irq(&conf->device_lock);
4795 conf->mddev->reshape_position = MaxSector;
4797 /* read-ahead size must cover two whole stripes, which is
4798 * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
4801 int data_disks = conf->previous_raid_disks - conf->max_degraded;
4802 int stripe = data_disks *
4803 (conf->mddev->chunk_size / PAGE_SIZE);
4804 if (conf->mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
4805 conf->mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
4810 static void raid5_quiesce(mddev_t *mddev, int state)
4812 raid5_conf_t *conf = mddev_to_conf(mddev);
4815 case 2: /* resume for a suspend */
4816 wake_up(&conf->wait_for_overlap);
4819 case 1: /* stop all writes */
4820 spin_lock_irq(&conf->device_lock);
4822 wait_event_lock_irq(conf->wait_for_stripe,
4823 atomic_read(&conf->active_stripes) == 0 &&
4824 atomic_read(&conf->active_aligned_reads) == 0,
4825 conf->device_lock, /* nothing */);
4826 spin_unlock_irq(&conf->device_lock);
4829 case 0: /* re-enable writes */
4830 spin_lock_irq(&conf->device_lock);
4832 wake_up(&conf->wait_for_stripe);
4833 wake_up(&conf->wait_for_overlap);
4834 spin_unlock_irq(&conf->device_lock);
4839 static struct mdk_personality raid6_personality =
4843 .owner = THIS_MODULE,
4844 .make_request = make_request,
4848 .error_handler = error,
4849 .hot_add_disk = raid5_add_disk,
4850 .hot_remove_disk= raid5_remove_disk,
4851 .spare_active = raid5_spare_active,
4852 .sync_request = sync_request,
4853 .resize = raid5_resize,
4854 #ifdef CONFIG_MD_RAID5_RESHAPE
4855 .check_reshape = raid5_check_reshape,
4856 .start_reshape = raid5_start_reshape,
4858 .quiesce = raid5_quiesce,
4860 static struct mdk_personality raid5_personality =
4864 .owner = THIS_MODULE,
4865 .make_request = make_request,
4869 .error_handler = error,
4870 .hot_add_disk = raid5_add_disk,
4871 .hot_remove_disk= raid5_remove_disk,
4872 .spare_active = raid5_spare_active,
4873 .sync_request = sync_request,
4874 .resize = raid5_resize,
4875 #ifdef CONFIG_MD_RAID5_RESHAPE
4876 .check_reshape = raid5_check_reshape,
4877 .start_reshape = raid5_start_reshape,
4879 .quiesce = raid5_quiesce,
4882 static struct mdk_personality raid4_personality =
4886 .owner = THIS_MODULE,
4887 .make_request = make_request,
4891 .error_handler = error,
4892 .hot_add_disk = raid5_add_disk,
4893 .hot_remove_disk= raid5_remove_disk,
4894 .spare_active = raid5_spare_active,
4895 .sync_request = sync_request,
4896 .resize = raid5_resize,
4897 #ifdef CONFIG_MD_RAID5_RESHAPE
4898 .check_reshape = raid5_check_reshape,
4899 .start_reshape = raid5_start_reshape,
4901 .quiesce = raid5_quiesce,
4904 static int __init raid5_init(void)
4908 e = raid6_select_algo();
4911 register_md_personality(&raid6_personality);
4912 register_md_personality(&raid5_personality);
4913 register_md_personality(&raid4_personality);
4917 static void raid5_exit(void)
4919 unregister_md_personality(&raid6_personality);
4920 unregister_md_personality(&raid5_personality);
4921 unregister_md_personality(&raid4_personality);
4924 module_init(raid5_init);
4925 module_exit(raid5_exit);
4926 MODULE_LICENSE("GPL");
4927 MODULE_ALIAS("md-personality-4"); /* RAID5 */
4928 MODULE_ALIAS("md-raid5");
4929 MODULE_ALIAS("md-raid4");
4930 MODULE_ALIAS("md-level-5");
4931 MODULE_ALIAS("md-level-4");
4932 MODULE_ALIAS("md-personality-8"); /* RAID6 */
4933 MODULE_ALIAS("md-raid6");
4934 MODULE_ALIAS("md-level-6");
4936 /* This used to be two separate modules, they were: */
4937 MODULE_ALIAS("raid5");
4938 MODULE_ALIAS("raid6");