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 NR_HASH (PAGE_SIZE / sizeof(struct hlist_head))
67 #define HASH_MASK (NR_HASH - 1)
69 #define stripe_hash(conf, sect) (&((conf)->stripe_hashtbl[((sect) >> STRIPE_SHIFT) & HASH_MASK]))
71 /* bio's attached to a stripe+device for I/O are linked together in bi_sector
72 * order without overlap. There may be several bio's per stripe+device, and
73 * a bio could span several devices.
74 * When walking this list for a particular stripe+device, we must never proceed
75 * beyond a bio that extends past this device, as the next bio might no longer
77 * This macro is used to determine the 'next' bio in the list, given the sector
78 * of the current stripe+device
80 #define r5_next_bio(bio, sect) ( ( (bio)->bi_sector + ((bio)->bi_size>>9) < sect + STRIPE_SECTORS) ? (bio)->bi_next : NULL)
82 * The following can be used to debug the driver
84 #define RAID5_PARANOIA 1
85 #if RAID5_PARANOIA && defined(CONFIG_SMP)
86 # define CHECK_DEVLOCK() assert_spin_locked(&conf->device_lock)
88 # define CHECK_DEVLOCK()
96 #if !RAID6_USE_EMPTY_ZERO_PAGE
97 /* In .bss so it's zeroed */
98 const char raid6_empty_zero_page[PAGE_SIZE] __attribute__((aligned(256)));
101 static inline int raid6_next_disk(int disk, int raid_disks)
104 return (disk < raid_disks) ? disk : 0;
107 static void return_io(struct bio *return_bi)
109 struct bio *bi = return_bi;
111 int bytes = bi->bi_size;
113 return_bi = bi->bi_next;
116 bi->bi_end_io(bi, bytes,
117 test_bit(BIO_UPTODATE, &bi->bi_flags)
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(request_queue_t *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 BUG_ON(sh->ops.count < 0);
386 raid5_end_read_request(struct bio *bi, unsigned int bytes_done, int error);
388 raid5_end_write_request (struct bio *bi, unsigned int bytes_done, int error);
390 static void ops_run_io(struct stripe_head *sh)
392 raid5_conf_t *conf = sh->raid_conf;
393 int i, disks = sh->disks;
397 for (i = disks; i--; ) {
401 if (test_and_clear_bit(R5_Wantwrite, &sh->dev[i].flags))
403 else if (test_and_clear_bit(R5_Wantread, &sh->dev[i].flags))
408 bi = &sh->dev[i].req;
412 bi->bi_end_io = raid5_end_write_request;
414 bi->bi_end_io = raid5_end_read_request;
417 rdev = rcu_dereference(conf->disks[i].rdev);
418 if (rdev && test_bit(Faulty, &rdev->flags))
421 atomic_inc(&rdev->nr_pending);
425 if (test_bit(STRIPE_SYNCING, &sh->state) ||
426 test_bit(STRIPE_EXPAND_SOURCE, &sh->state) ||
427 test_bit(STRIPE_EXPAND_READY, &sh->state))
428 md_sync_acct(rdev->bdev, STRIPE_SECTORS);
430 bi->bi_bdev = rdev->bdev;
431 pr_debug("%s: for %llu schedule op %ld on disc %d\n",
432 __FUNCTION__, (unsigned long long)sh->sector,
434 atomic_inc(&sh->count);
435 bi->bi_sector = sh->sector + rdev->data_offset;
436 bi->bi_flags = 1 << BIO_UPTODATE;
440 bi->bi_io_vec = &sh->dev[i].vec;
441 bi->bi_io_vec[0].bv_len = STRIPE_SIZE;
442 bi->bi_io_vec[0].bv_offset = 0;
443 bi->bi_size = STRIPE_SIZE;
446 test_bit(R5_ReWrite, &sh->dev[i].flags))
447 atomic_add(STRIPE_SECTORS,
448 &rdev->corrected_errors);
449 generic_make_request(bi);
452 set_bit(STRIPE_DEGRADED, &sh->state);
453 pr_debug("skip op %ld on disc %d for sector %llu\n",
454 bi->bi_rw, i, (unsigned long long)sh->sector);
455 clear_bit(R5_LOCKED, &sh->dev[i].flags);
456 set_bit(STRIPE_HANDLE, &sh->state);
461 static struct dma_async_tx_descriptor *
462 async_copy_data(int frombio, struct bio *bio, struct page *page,
463 sector_t sector, struct dma_async_tx_descriptor *tx)
466 struct page *bio_page;
470 if (bio->bi_sector >= sector)
471 page_offset = (signed)(bio->bi_sector - sector) * 512;
473 page_offset = (signed)(sector - bio->bi_sector) * -512;
474 bio_for_each_segment(bvl, bio, i) {
475 int len = bio_iovec_idx(bio, i)->bv_len;
479 if (page_offset < 0) {
480 b_offset = -page_offset;
481 page_offset += b_offset;
485 if (len > 0 && page_offset + len > STRIPE_SIZE)
486 clen = STRIPE_SIZE - page_offset;
491 b_offset += bio_iovec_idx(bio, i)->bv_offset;
492 bio_page = bio_iovec_idx(bio, i)->bv_page;
494 tx = async_memcpy(page, bio_page, page_offset,
496 ASYNC_TX_DEP_ACK | ASYNC_TX_KMAP_SRC,
499 tx = async_memcpy(bio_page, page, b_offset,
501 ASYNC_TX_DEP_ACK | ASYNC_TX_KMAP_DST,
504 if (clen < len) /* hit end of page */
512 static void ops_complete_biofill(void *stripe_head_ref)
514 struct stripe_head *sh = stripe_head_ref;
515 struct bio *return_bi = NULL;
516 raid5_conf_t *conf = sh->raid_conf;
517 int i, more_to_read = 0;
519 pr_debug("%s: stripe %llu\n", __FUNCTION__,
520 (unsigned long long)sh->sector);
522 /* clear completed biofills */
523 for (i = sh->disks; i--; ) {
524 struct r5dev *dev = &sh->dev[i];
525 /* check if this stripe has new incoming reads */
529 /* acknowledge completion of a biofill operation */
530 /* and check if we need to reply to a read request
532 if (test_bit(R5_Wantfill, &dev->flags) && !dev->toread) {
533 struct bio *rbi, *rbi2;
534 clear_bit(R5_Wantfill, &dev->flags);
536 /* The access to dev->read is outside of the
537 * spin_lock_irq(&conf->device_lock), but is protected
538 * by the STRIPE_OP_BIOFILL pending bit
543 while (rbi && rbi->bi_sector <
544 dev->sector + STRIPE_SECTORS) {
545 rbi2 = r5_next_bio(rbi, dev->sector);
546 spin_lock_irq(&conf->device_lock);
547 if (--rbi->bi_phys_segments == 0) {
548 rbi->bi_next = return_bi;
551 spin_unlock_irq(&conf->device_lock);
556 clear_bit(STRIPE_OP_BIOFILL, &sh->ops.ack);
557 clear_bit(STRIPE_OP_BIOFILL, &sh->ops.pending);
559 return_io(return_bi);
562 set_bit(STRIPE_HANDLE, &sh->state);
566 static void ops_run_biofill(struct stripe_head *sh)
568 struct dma_async_tx_descriptor *tx = NULL;
569 raid5_conf_t *conf = sh->raid_conf;
572 pr_debug("%s: stripe %llu\n", __FUNCTION__,
573 (unsigned long long)sh->sector);
575 for (i = sh->disks; i--; ) {
576 struct r5dev *dev = &sh->dev[i];
577 if (test_bit(R5_Wantfill, &dev->flags)) {
579 spin_lock_irq(&conf->device_lock);
580 dev->read = rbi = dev->toread;
582 spin_unlock_irq(&conf->device_lock);
583 while (rbi && rbi->bi_sector <
584 dev->sector + STRIPE_SECTORS) {
585 tx = async_copy_data(0, rbi, dev->page,
587 rbi = r5_next_bio(rbi, dev->sector);
592 atomic_inc(&sh->count);
593 async_trigger_callback(ASYNC_TX_DEP_ACK | ASYNC_TX_ACK, tx,
594 ops_complete_biofill, sh);
597 static void ops_complete_compute5(void *stripe_head_ref)
599 struct stripe_head *sh = stripe_head_ref;
600 int target = sh->ops.target;
601 struct r5dev *tgt = &sh->dev[target];
603 pr_debug("%s: stripe %llu\n", __FUNCTION__,
604 (unsigned long long)sh->sector);
606 set_bit(R5_UPTODATE, &tgt->flags);
607 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
608 clear_bit(R5_Wantcompute, &tgt->flags);
609 set_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.complete);
610 set_bit(STRIPE_HANDLE, &sh->state);
614 static struct dma_async_tx_descriptor *
615 ops_run_compute5(struct stripe_head *sh, unsigned long pending)
617 /* kernel stack size limits the total number of disks */
618 int disks = sh->disks;
619 struct page *xor_srcs[disks];
620 int target = sh->ops.target;
621 struct r5dev *tgt = &sh->dev[target];
622 struct page *xor_dest = tgt->page;
624 struct dma_async_tx_descriptor *tx;
627 pr_debug("%s: stripe %llu block: %d\n",
628 __FUNCTION__, (unsigned long long)sh->sector, target);
629 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
631 for (i = disks; i--; )
633 xor_srcs[count++] = sh->dev[i].page;
635 atomic_inc(&sh->count);
637 if (unlikely(count == 1))
638 tx = async_memcpy(xor_dest, xor_srcs[0], 0, 0, STRIPE_SIZE,
639 0, NULL, ops_complete_compute5, sh);
641 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE,
642 ASYNC_TX_XOR_ZERO_DST, NULL,
643 ops_complete_compute5, sh);
645 /* ack now if postxor is not set to be run */
646 if (tx && !test_bit(STRIPE_OP_POSTXOR, &pending))
652 static void ops_complete_prexor(void *stripe_head_ref)
654 struct stripe_head *sh = stripe_head_ref;
656 pr_debug("%s: stripe %llu\n", __FUNCTION__,
657 (unsigned long long)sh->sector);
659 set_bit(STRIPE_OP_PREXOR, &sh->ops.complete);
662 static struct dma_async_tx_descriptor *
663 ops_run_prexor(struct stripe_head *sh, struct dma_async_tx_descriptor *tx)
665 /* kernel stack size limits the total number of disks */
666 int disks = sh->disks;
667 struct page *xor_srcs[disks];
668 int count = 0, pd_idx = sh->pd_idx, i;
670 /* existing parity data subtracted */
671 struct page *xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
673 pr_debug("%s: stripe %llu\n", __FUNCTION__,
674 (unsigned long long)sh->sector);
676 for (i = disks; i--; ) {
677 struct r5dev *dev = &sh->dev[i];
678 /* Only process blocks that are known to be uptodate */
679 if (dev->towrite && test_bit(R5_Wantprexor, &dev->flags))
680 xor_srcs[count++] = dev->page;
683 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE,
684 ASYNC_TX_DEP_ACK | ASYNC_TX_XOR_DROP_DST, tx,
685 ops_complete_prexor, sh);
690 static struct dma_async_tx_descriptor *
691 ops_run_biodrain(struct stripe_head *sh, struct dma_async_tx_descriptor *tx)
693 int disks = sh->disks;
694 int pd_idx = sh->pd_idx, i;
696 /* check if prexor is active which means only process blocks
697 * that are part of a read-modify-write (Wantprexor)
699 int prexor = test_bit(STRIPE_OP_PREXOR, &sh->ops.pending);
701 pr_debug("%s: stripe %llu\n", __FUNCTION__,
702 (unsigned long long)sh->sector);
704 for (i = disks; i--; ) {
705 struct r5dev *dev = &sh->dev[i];
710 if (prexor) { /* rmw */
712 test_bit(R5_Wantprexor, &dev->flags))
715 if (i != pd_idx && dev->towrite &&
716 test_bit(R5_LOCKED, &dev->flags))
723 spin_lock(&sh->lock);
724 chosen = dev->towrite;
726 BUG_ON(dev->written);
727 wbi = dev->written = chosen;
728 spin_unlock(&sh->lock);
730 while (wbi && wbi->bi_sector <
731 dev->sector + STRIPE_SECTORS) {
732 tx = async_copy_data(1, wbi, dev->page,
734 wbi = r5_next_bio(wbi, dev->sector);
742 static void ops_complete_postxor(void *stripe_head_ref)
744 struct stripe_head *sh = stripe_head_ref;
746 pr_debug("%s: stripe %llu\n", __FUNCTION__,
747 (unsigned long long)sh->sector);
749 set_bit(STRIPE_OP_POSTXOR, &sh->ops.complete);
750 set_bit(STRIPE_HANDLE, &sh->state);
754 static void ops_complete_write(void *stripe_head_ref)
756 struct stripe_head *sh = stripe_head_ref;
757 int disks = sh->disks, i, pd_idx = sh->pd_idx;
759 pr_debug("%s: stripe %llu\n", __FUNCTION__,
760 (unsigned long long)sh->sector);
762 for (i = disks; i--; ) {
763 struct r5dev *dev = &sh->dev[i];
764 if (dev->written || i == pd_idx)
765 set_bit(R5_UPTODATE, &dev->flags);
768 set_bit(STRIPE_OP_BIODRAIN, &sh->ops.complete);
769 set_bit(STRIPE_OP_POSTXOR, &sh->ops.complete);
771 set_bit(STRIPE_HANDLE, &sh->state);
776 ops_run_postxor(struct stripe_head *sh, struct dma_async_tx_descriptor *tx)
778 /* kernel stack size limits the total number of disks */
779 int disks = sh->disks;
780 struct page *xor_srcs[disks];
782 int count = 0, pd_idx = sh->pd_idx, i;
783 struct page *xor_dest;
784 int prexor = test_bit(STRIPE_OP_PREXOR, &sh->ops.pending);
786 dma_async_tx_callback callback;
788 pr_debug("%s: stripe %llu\n", __FUNCTION__,
789 (unsigned long long)sh->sector);
791 /* check if prexor is active which means only process blocks
792 * that are part of a read-modify-write (written)
795 xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
796 for (i = disks; i--; ) {
797 struct r5dev *dev = &sh->dev[i];
799 xor_srcs[count++] = dev->page;
802 xor_dest = sh->dev[pd_idx].page;
803 for (i = disks; i--; ) {
804 struct r5dev *dev = &sh->dev[i];
806 xor_srcs[count++] = dev->page;
810 /* check whether this postxor is part of a write */
811 callback = test_bit(STRIPE_OP_BIODRAIN, &sh->ops.pending) ?
812 ops_complete_write : ops_complete_postxor;
814 /* 1/ if we prexor'd then the dest is reused as a source
815 * 2/ if we did not prexor then we are redoing the parity
816 * set ASYNC_TX_XOR_DROP_DST and ASYNC_TX_XOR_ZERO_DST
817 * for the synchronous xor case
819 flags = ASYNC_TX_DEP_ACK | ASYNC_TX_ACK |
820 (prexor ? ASYNC_TX_XOR_DROP_DST : ASYNC_TX_XOR_ZERO_DST);
822 atomic_inc(&sh->count);
824 if (unlikely(count == 1)) {
825 flags &= ~(ASYNC_TX_XOR_DROP_DST | ASYNC_TX_XOR_ZERO_DST);
826 tx = async_memcpy(xor_dest, xor_srcs[0], 0, 0, STRIPE_SIZE,
827 flags, tx, callback, sh);
829 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE,
830 flags, tx, callback, sh);
833 static void ops_complete_check(void *stripe_head_ref)
835 struct stripe_head *sh = stripe_head_ref;
836 int pd_idx = sh->pd_idx;
838 pr_debug("%s: stripe %llu\n", __FUNCTION__,
839 (unsigned long long)sh->sector);
841 if (test_and_clear_bit(STRIPE_OP_MOD_DMA_CHECK, &sh->ops.pending) &&
842 sh->ops.zero_sum_result == 0)
843 set_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
845 set_bit(STRIPE_OP_CHECK, &sh->ops.complete);
846 set_bit(STRIPE_HANDLE, &sh->state);
850 static void ops_run_check(struct stripe_head *sh)
852 /* kernel stack size limits the total number of disks */
853 int disks = sh->disks;
854 struct page *xor_srcs[disks];
855 struct dma_async_tx_descriptor *tx;
857 int count = 0, pd_idx = sh->pd_idx, i;
858 struct page *xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
860 pr_debug("%s: stripe %llu\n", __FUNCTION__,
861 (unsigned long long)sh->sector);
863 for (i = disks; i--; ) {
864 struct r5dev *dev = &sh->dev[i];
866 xor_srcs[count++] = dev->page;
869 tx = async_xor_zero_sum(xor_dest, xor_srcs, 0, count, STRIPE_SIZE,
870 &sh->ops.zero_sum_result, 0, NULL, NULL, NULL);
873 set_bit(STRIPE_OP_MOD_DMA_CHECK, &sh->ops.pending);
875 clear_bit(STRIPE_OP_MOD_DMA_CHECK, &sh->ops.pending);
877 atomic_inc(&sh->count);
878 tx = async_trigger_callback(ASYNC_TX_DEP_ACK | ASYNC_TX_ACK, tx,
879 ops_complete_check, sh);
882 static void raid5_run_ops(struct stripe_head *sh, unsigned long pending)
884 int overlap_clear = 0, i, disks = sh->disks;
885 struct dma_async_tx_descriptor *tx = NULL;
887 if (test_bit(STRIPE_OP_BIOFILL, &pending)) {
892 if (test_bit(STRIPE_OP_COMPUTE_BLK, &pending))
893 tx = ops_run_compute5(sh, pending);
895 if (test_bit(STRIPE_OP_PREXOR, &pending))
896 tx = ops_run_prexor(sh, tx);
898 if (test_bit(STRIPE_OP_BIODRAIN, &pending)) {
899 tx = ops_run_biodrain(sh, tx);
903 if (test_bit(STRIPE_OP_POSTXOR, &pending))
904 ops_run_postxor(sh, tx);
906 if (test_bit(STRIPE_OP_CHECK, &pending))
909 if (test_bit(STRIPE_OP_IO, &pending))
913 for (i = disks; i--; ) {
914 struct r5dev *dev = &sh->dev[i];
915 if (test_and_clear_bit(R5_Overlap, &dev->flags))
916 wake_up(&sh->raid_conf->wait_for_overlap);
920 static int grow_one_stripe(raid5_conf_t *conf)
922 struct stripe_head *sh;
923 sh = kmem_cache_alloc(conf->slab_cache, GFP_KERNEL);
926 memset(sh, 0, sizeof(*sh) + (conf->raid_disks-1)*sizeof(struct r5dev));
927 sh->raid_conf = conf;
928 spin_lock_init(&sh->lock);
930 if (grow_buffers(sh, conf->raid_disks)) {
931 shrink_buffers(sh, conf->raid_disks);
932 kmem_cache_free(conf->slab_cache, sh);
935 sh->disks = conf->raid_disks;
936 /* we just created an active stripe so... */
937 atomic_set(&sh->count, 1);
938 atomic_inc(&conf->active_stripes);
939 INIT_LIST_HEAD(&sh->lru);
944 static int grow_stripes(raid5_conf_t *conf, int num)
946 struct kmem_cache *sc;
947 int devs = conf->raid_disks;
949 sprintf(conf->cache_name[0], "raid5-%s", mdname(conf->mddev));
950 sprintf(conf->cache_name[1], "raid5-%s-alt", mdname(conf->mddev));
951 conf->active_name = 0;
952 sc = kmem_cache_create(conf->cache_name[conf->active_name],
953 sizeof(struct stripe_head)+(devs-1)*sizeof(struct r5dev),
957 conf->slab_cache = sc;
958 conf->pool_size = devs;
960 if (!grow_one_stripe(conf))
965 #ifdef CONFIG_MD_RAID5_RESHAPE
966 static int resize_stripes(raid5_conf_t *conf, int newsize)
968 /* Make all the stripes able to hold 'newsize' devices.
969 * New slots in each stripe get 'page' set to a new page.
971 * This happens in stages:
972 * 1/ create a new kmem_cache and allocate the required number of
974 * 2/ gather all the old stripe_heads and tranfer the pages across
975 * to the new stripe_heads. This will have the side effect of
976 * freezing the array as once all stripe_heads have been collected,
977 * no IO will be possible. Old stripe heads are freed once their
978 * pages have been transferred over, and the old kmem_cache is
979 * freed when all stripes are done.
980 * 3/ reallocate conf->disks to be suitable bigger. If this fails,
981 * we simple return a failre status - no need to clean anything up.
982 * 4/ allocate new pages for the new slots in the new stripe_heads.
983 * If this fails, we don't bother trying the shrink the
984 * stripe_heads down again, we just leave them as they are.
985 * As each stripe_head is processed the new one is released into
988 * Once step2 is started, we cannot afford to wait for a write,
989 * so we use GFP_NOIO allocations.
991 struct stripe_head *osh, *nsh;
992 LIST_HEAD(newstripes);
993 struct disk_info *ndisks;
995 struct kmem_cache *sc;
998 if (newsize <= conf->pool_size)
999 return 0; /* never bother to shrink */
1001 md_allow_write(conf->mddev);
1004 sc = kmem_cache_create(conf->cache_name[1-conf->active_name],
1005 sizeof(struct stripe_head)+(newsize-1)*sizeof(struct r5dev),
1010 for (i = conf->max_nr_stripes; i; i--) {
1011 nsh = kmem_cache_alloc(sc, GFP_KERNEL);
1015 memset(nsh, 0, sizeof(*nsh) + (newsize-1)*sizeof(struct r5dev));
1017 nsh->raid_conf = conf;
1018 spin_lock_init(&nsh->lock);
1020 list_add(&nsh->lru, &newstripes);
1023 /* didn't get enough, give up */
1024 while (!list_empty(&newstripes)) {
1025 nsh = list_entry(newstripes.next, struct stripe_head, lru);
1026 list_del(&nsh->lru);
1027 kmem_cache_free(sc, nsh);
1029 kmem_cache_destroy(sc);
1032 /* Step 2 - Must use GFP_NOIO now.
1033 * OK, we have enough stripes, start collecting inactive
1034 * stripes and copying them over
1036 list_for_each_entry(nsh, &newstripes, lru) {
1037 spin_lock_irq(&conf->device_lock);
1038 wait_event_lock_irq(conf->wait_for_stripe,
1039 !list_empty(&conf->inactive_list),
1041 unplug_slaves(conf->mddev)
1043 osh = get_free_stripe(conf);
1044 spin_unlock_irq(&conf->device_lock);
1045 atomic_set(&nsh->count, 1);
1046 for(i=0; i<conf->pool_size; i++)
1047 nsh->dev[i].page = osh->dev[i].page;
1048 for( ; i<newsize; i++)
1049 nsh->dev[i].page = NULL;
1050 kmem_cache_free(conf->slab_cache, osh);
1052 kmem_cache_destroy(conf->slab_cache);
1055 * At this point, we are holding all the stripes so the array
1056 * is completely stalled, so now is a good time to resize
1059 ndisks = kzalloc(newsize * sizeof(struct disk_info), GFP_NOIO);
1061 for (i=0; i<conf->raid_disks; i++)
1062 ndisks[i] = conf->disks[i];
1064 conf->disks = ndisks;
1068 /* Step 4, return new stripes to service */
1069 while(!list_empty(&newstripes)) {
1070 nsh = list_entry(newstripes.next, struct stripe_head, lru);
1071 list_del_init(&nsh->lru);
1072 for (i=conf->raid_disks; i < newsize; i++)
1073 if (nsh->dev[i].page == NULL) {
1074 struct page *p = alloc_page(GFP_NOIO);
1075 nsh->dev[i].page = p;
1079 release_stripe(nsh);
1081 /* critical section pass, GFP_NOIO no longer needed */
1083 conf->slab_cache = sc;
1084 conf->active_name = 1-conf->active_name;
1085 conf->pool_size = newsize;
1090 static int drop_one_stripe(raid5_conf_t *conf)
1092 struct stripe_head *sh;
1094 spin_lock_irq(&conf->device_lock);
1095 sh = get_free_stripe(conf);
1096 spin_unlock_irq(&conf->device_lock);
1099 BUG_ON(atomic_read(&sh->count));
1100 shrink_buffers(sh, conf->pool_size);
1101 kmem_cache_free(conf->slab_cache, sh);
1102 atomic_dec(&conf->active_stripes);
1106 static void shrink_stripes(raid5_conf_t *conf)
1108 while (drop_one_stripe(conf))
1111 if (conf->slab_cache)
1112 kmem_cache_destroy(conf->slab_cache);
1113 conf->slab_cache = NULL;
1116 static int raid5_end_read_request(struct bio * bi, unsigned int bytes_done,
1119 struct stripe_head *sh = bi->bi_private;
1120 raid5_conf_t *conf = sh->raid_conf;
1121 int disks = sh->disks, i;
1122 int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
1123 char b[BDEVNAME_SIZE];
1129 for (i=0 ; i<disks; i++)
1130 if (bi == &sh->dev[i].req)
1133 pr_debug("end_read_request %llu/%d, count: %d, uptodate %d.\n",
1134 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
1142 set_bit(R5_UPTODATE, &sh->dev[i].flags);
1143 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
1144 rdev = conf->disks[i].rdev;
1145 printk(KERN_INFO "raid5:%s: read error corrected (%lu sectors at %llu on %s)\n",
1146 mdname(conf->mddev), STRIPE_SECTORS,
1147 (unsigned long long)sh->sector + rdev->data_offset,
1148 bdevname(rdev->bdev, b));
1149 clear_bit(R5_ReadError, &sh->dev[i].flags);
1150 clear_bit(R5_ReWrite, &sh->dev[i].flags);
1152 if (atomic_read(&conf->disks[i].rdev->read_errors))
1153 atomic_set(&conf->disks[i].rdev->read_errors, 0);
1155 const char *bdn = bdevname(conf->disks[i].rdev->bdev, b);
1157 rdev = conf->disks[i].rdev;
1159 clear_bit(R5_UPTODATE, &sh->dev[i].flags);
1160 atomic_inc(&rdev->read_errors);
1161 if (conf->mddev->degraded)
1162 printk(KERN_WARNING "raid5:%s: read error not correctable (sector %llu on %s).\n",
1163 mdname(conf->mddev),
1164 (unsigned long long)sh->sector + rdev->data_offset,
1166 else if (test_bit(R5_ReWrite, &sh->dev[i].flags))
1168 printk(KERN_WARNING "raid5:%s: read error NOT corrected!! (sector %llu on %s).\n",
1169 mdname(conf->mddev),
1170 (unsigned long long)sh->sector + rdev->data_offset,
1172 else if (atomic_read(&rdev->read_errors)
1173 > conf->max_nr_stripes)
1175 "raid5:%s: Too many read errors, failing device %s.\n",
1176 mdname(conf->mddev), bdn);
1180 set_bit(R5_ReadError, &sh->dev[i].flags);
1182 clear_bit(R5_ReadError, &sh->dev[i].flags);
1183 clear_bit(R5_ReWrite, &sh->dev[i].flags);
1184 md_error(conf->mddev, rdev);
1187 rdev_dec_pending(conf->disks[i].rdev, conf->mddev);
1188 clear_bit(R5_LOCKED, &sh->dev[i].flags);
1189 set_bit(STRIPE_HANDLE, &sh->state);
1194 static int raid5_end_write_request (struct bio *bi, unsigned int bytes_done,
1197 struct stripe_head *sh = bi->bi_private;
1198 raid5_conf_t *conf = sh->raid_conf;
1199 int disks = sh->disks, i;
1200 int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
1205 for (i=0 ; i<disks; i++)
1206 if (bi == &sh->dev[i].req)
1209 pr_debug("end_write_request %llu/%d, count %d, uptodate: %d.\n",
1210 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
1218 md_error(conf->mddev, conf->disks[i].rdev);
1220 rdev_dec_pending(conf->disks[i].rdev, conf->mddev);
1222 clear_bit(R5_LOCKED, &sh->dev[i].flags);
1223 set_bit(STRIPE_HANDLE, &sh->state);
1229 static sector_t compute_blocknr(struct stripe_head *sh, int i);
1231 static void raid5_build_block (struct stripe_head *sh, int i)
1233 struct r5dev *dev = &sh->dev[i];
1235 bio_init(&dev->req);
1236 dev->req.bi_io_vec = &dev->vec;
1238 dev->req.bi_max_vecs++;
1239 dev->vec.bv_page = dev->page;
1240 dev->vec.bv_len = STRIPE_SIZE;
1241 dev->vec.bv_offset = 0;
1243 dev->req.bi_sector = sh->sector;
1244 dev->req.bi_private = sh;
1247 dev->sector = compute_blocknr(sh, i);
1250 static void error(mddev_t *mddev, mdk_rdev_t *rdev)
1252 char b[BDEVNAME_SIZE];
1253 raid5_conf_t *conf = (raid5_conf_t *) mddev->private;
1254 pr_debug("raid5: error called\n");
1256 if (!test_bit(Faulty, &rdev->flags)) {
1257 set_bit(MD_CHANGE_DEVS, &mddev->flags);
1258 if (test_and_clear_bit(In_sync, &rdev->flags)) {
1259 unsigned long flags;
1260 spin_lock_irqsave(&conf->device_lock, flags);
1262 spin_unlock_irqrestore(&conf->device_lock, flags);
1264 * if recovery was running, make sure it aborts.
1266 set_bit(MD_RECOVERY_ERR, &mddev->recovery);
1268 set_bit(Faulty, &rdev->flags);
1270 "raid5: Disk failure on %s, disabling device."
1271 " Operation continuing on %d devices\n",
1272 bdevname(rdev->bdev,b), conf->raid_disks - mddev->degraded);
1277 * Input: a 'big' sector number,
1278 * Output: index of the data and parity disk, and the sector # in them.
1280 static sector_t raid5_compute_sector(sector_t r_sector, unsigned int raid_disks,
1281 unsigned int data_disks, unsigned int * dd_idx,
1282 unsigned int * pd_idx, raid5_conf_t *conf)
1285 unsigned long chunk_number;
1286 unsigned int chunk_offset;
1287 sector_t new_sector;
1288 int sectors_per_chunk = conf->chunk_size >> 9;
1290 /* First compute the information on this sector */
1293 * Compute the chunk number and the sector offset inside the chunk
1295 chunk_offset = sector_div(r_sector, sectors_per_chunk);
1296 chunk_number = r_sector;
1297 BUG_ON(r_sector != chunk_number);
1300 * Compute the stripe number
1302 stripe = chunk_number / data_disks;
1305 * Compute the data disk and parity disk indexes inside the stripe
1307 *dd_idx = chunk_number % data_disks;
1310 * Select the parity disk based on the user selected algorithm.
1312 switch(conf->level) {
1314 *pd_idx = data_disks;
1317 switch (conf->algorithm) {
1318 case ALGORITHM_LEFT_ASYMMETRIC:
1319 *pd_idx = data_disks - stripe % raid_disks;
1320 if (*dd_idx >= *pd_idx)
1323 case ALGORITHM_RIGHT_ASYMMETRIC:
1324 *pd_idx = stripe % raid_disks;
1325 if (*dd_idx >= *pd_idx)
1328 case ALGORITHM_LEFT_SYMMETRIC:
1329 *pd_idx = data_disks - stripe % raid_disks;
1330 *dd_idx = (*pd_idx + 1 + *dd_idx) % raid_disks;
1332 case ALGORITHM_RIGHT_SYMMETRIC:
1333 *pd_idx = stripe % raid_disks;
1334 *dd_idx = (*pd_idx + 1 + *dd_idx) % raid_disks;
1337 printk(KERN_ERR "raid5: unsupported algorithm %d\n",
1343 /**** FIX THIS ****/
1344 switch (conf->algorithm) {
1345 case ALGORITHM_LEFT_ASYMMETRIC:
1346 *pd_idx = raid_disks - 1 - (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_RIGHT_ASYMMETRIC:
1353 *pd_idx = stripe % raid_disks;
1354 if (*pd_idx == raid_disks-1)
1355 (*dd_idx)++; /* Q D D D P */
1356 else if (*dd_idx >= *pd_idx)
1357 (*dd_idx) += 2; /* D D P Q D */
1359 case ALGORITHM_LEFT_SYMMETRIC:
1360 *pd_idx = raid_disks - 1 - (stripe % raid_disks);
1361 *dd_idx = (*pd_idx + 2 + *dd_idx) % raid_disks;
1363 case ALGORITHM_RIGHT_SYMMETRIC:
1364 *pd_idx = stripe % raid_disks;
1365 *dd_idx = (*pd_idx + 2 + *dd_idx) % raid_disks;
1368 printk (KERN_CRIT "raid6: unsupported algorithm %d\n",
1375 * Finally, compute the new sector number
1377 new_sector = (sector_t)stripe * sectors_per_chunk + chunk_offset;
1382 static sector_t compute_blocknr(struct stripe_head *sh, int i)
1384 raid5_conf_t *conf = sh->raid_conf;
1385 int raid_disks = sh->disks;
1386 int data_disks = raid_disks - conf->max_degraded;
1387 sector_t new_sector = sh->sector, check;
1388 int sectors_per_chunk = conf->chunk_size >> 9;
1391 int chunk_number, dummy1, dummy2, dd_idx = i;
1395 chunk_offset = sector_div(new_sector, sectors_per_chunk);
1396 stripe = new_sector;
1397 BUG_ON(new_sector != stripe);
1399 if (i == sh->pd_idx)
1401 switch(conf->level) {
1404 switch (conf->algorithm) {
1405 case ALGORITHM_LEFT_ASYMMETRIC:
1406 case ALGORITHM_RIGHT_ASYMMETRIC:
1410 case ALGORITHM_LEFT_SYMMETRIC:
1411 case ALGORITHM_RIGHT_SYMMETRIC:
1414 i -= (sh->pd_idx + 1);
1417 printk(KERN_ERR "raid5: unsupported algorithm %d\n",
1422 if (i == raid6_next_disk(sh->pd_idx, raid_disks))
1423 return 0; /* It is the Q disk */
1424 switch (conf->algorithm) {
1425 case ALGORITHM_LEFT_ASYMMETRIC:
1426 case ALGORITHM_RIGHT_ASYMMETRIC:
1427 if (sh->pd_idx == raid_disks-1)
1428 i--; /* Q D D D P */
1429 else if (i > sh->pd_idx)
1430 i -= 2; /* D D P Q D */
1432 case ALGORITHM_LEFT_SYMMETRIC:
1433 case ALGORITHM_RIGHT_SYMMETRIC:
1434 if (sh->pd_idx == raid_disks-1)
1435 i--; /* Q D D D P */
1440 i -= (sh->pd_idx + 2);
1444 printk (KERN_CRIT "raid6: unsupported algorithm %d\n",
1450 chunk_number = stripe * data_disks + i;
1451 r_sector = (sector_t)chunk_number * sectors_per_chunk + chunk_offset;
1453 check = raid5_compute_sector (r_sector, raid_disks, data_disks, &dummy1, &dummy2, conf);
1454 if (check != sh->sector || dummy1 != dd_idx || dummy2 != sh->pd_idx) {
1455 printk(KERN_ERR "compute_blocknr: map not correct\n");
1464 * Copy data between a page in the stripe cache, and one or more bion
1465 * The page could align with the middle of the bio, or there could be
1466 * several bion, each with several bio_vecs, which cover part of the page
1467 * Multiple bion are linked together on bi_next. There may be extras
1468 * at the end of this list. We ignore them.
1470 static void copy_data(int frombio, struct bio *bio,
1474 char *pa = page_address(page);
1475 struct bio_vec *bvl;
1479 if (bio->bi_sector >= sector)
1480 page_offset = (signed)(bio->bi_sector - sector) * 512;
1482 page_offset = (signed)(sector - bio->bi_sector) * -512;
1483 bio_for_each_segment(bvl, bio, i) {
1484 int len = bio_iovec_idx(bio,i)->bv_len;
1488 if (page_offset < 0) {
1489 b_offset = -page_offset;
1490 page_offset += b_offset;
1494 if (len > 0 && page_offset + len > STRIPE_SIZE)
1495 clen = STRIPE_SIZE - page_offset;
1499 char *ba = __bio_kmap_atomic(bio, i, KM_USER0);
1501 memcpy(pa+page_offset, ba+b_offset, clen);
1503 memcpy(ba+b_offset, pa+page_offset, clen);
1504 __bio_kunmap_atomic(ba, KM_USER0);
1506 if (clen < len) /* hit end of page */
1512 #define check_xor() do { \
1513 if (count == MAX_XOR_BLOCKS) { \
1514 xor_blocks(count, STRIPE_SIZE, dest, ptr);\
1520 static void compute_block(struct stripe_head *sh, int dd_idx)
1522 int i, count, disks = sh->disks;
1523 void *ptr[MAX_XOR_BLOCKS], *dest, *p;
1525 pr_debug("compute_block, stripe %llu, idx %d\n",
1526 (unsigned long long)sh->sector, dd_idx);
1528 dest = page_address(sh->dev[dd_idx].page);
1529 memset(dest, 0, STRIPE_SIZE);
1531 for (i = disks ; i--; ) {
1534 p = page_address(sh->dev[i].page);
1535 if (test_bit(R5_UPTODATE, &sh->dev[i].flags))
1538 printk(KERN_ERR "compute_block() %d, stripe %llu, %d"
1539 " not present\n", dd_idx,
1540 (unsigned long long)sh->sector, i);
1545 xor_blocks(count, STRIPE_SIZE, dest, ptr);
1546 set_bit(R5_UPTODATE, &sh->dev[dd_idx].flags);
1549 static void compute_parity5(struct stripe_head *sh, int method)
1551 raid5_conf_t *conf = sh->raid_conf;
1552 int i, pd_idx = sh->pd_idx, disks = sh->disks, count;
1553 void *ptr[MAX_XOR_BLOCKS], *dest;
1556 pr_debug("compute_parity5, stripe %llu, method %d\n",
1557 (unsigned long long)sh->sector, method);
1560 dest = page_address(sh->dev[pd_idx].page);
1562 case READ_MODIFY_WRITE:
1563 BUG_ON(!test_bit(R5_UPTODATE, &sh->dev[pd_idx].flags));
1564 for (i=disks ; i-- ;) {
1567 if (sh->dev[i].towrite &&
1568 test_bit(R5_UPTODATE, &sh->dev[i].flags)) {
1569 ptr[count++] = page_address(sh->dev[i].page);
1570 chosen = sh->dev[i].towrite;
1571 sh->dev[i].towrite = NULL;
1573 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
1574 wake_up(&conf->wait_for_overlap);
1576 BUG_ON(sh->dev[i].written);
1577 sh->dev[i].written = chosen;
1582 case RECONSTRUCT_WRITE:
1583 memset(dest, 0, STRIPE_SIZE);
1584 for (i= disks; i-- ;)
1585 if (i!=pd_idx && sh->dev[i].towrite) {
1586 chosen = sh->dev[i].towrite;
1587 sh->dev[i].towrite = NULL;
1589 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
1590 wake_up(&conf->wait_for_overlap);
1592 BUG_ON(sh->dev[i].written);
1593 sh->dev[i].written = chosen;
1600 xor_blocks(count, STRIPE_SIZE, dest, ptr);
1604 for (i = disks; i--;)
1605 if (sh->dev[i].written) {
1606 sector_t sector = sh->dev[i].sector;
1607 struct bio *wbi = sh->dev[i].written;
1608 while (wbi && wbi->bi_sector < sector + STRIPE_SECTORS) {
1609 copy_data(1, wbi, sh->dev[i].page, sector);
1610 wbi = r5_next_bio(wbi, sector);
1613 set_bit(R5_LOCKED, &sh->dev[i].flags);
1614 set_bit(R5_UPTODATE, &sh->dev[i].flags);
1618 case RECONSTRUCT_WRITE:
1622 ptr[count++] = page_address(sh->dev[i].page);
1626 case READ_MODIFY_WRITE:
1627 for (i = disks; i--;)
1628 if (sh->dev[i].written) {
1629 ptr[count++] = page_address(sh->dev[i].page);
1634 xor_blocks(count, STRIPE_SIZE, dest, ptr);
1636 if (method != CHECK_PARITY) {
1637 set_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
1638 set_bit(R5_LOCKED, &sh->dev[pd_idx].flags);
1640 clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
1643 static void compute_parity6(struct stripe_head *sh, int method)
1645 raid6_conf_t *conf = sh->raid_conf;
1646 int i, pd_idx = sh->pd_idx, qd_idx, d0_idx, disks = sh->disks, count;
1648 /**** FIX THIS: This could be very bad if disks is close to 256 ****/
1651 qd_idx = raid6_next_disk(pd_idx, disks);
1652 d0_idx = raid6_next_disk(qd_idx, disks);
1654 pr_debug("compute_parity, stripe %llu, method %d\n",
1655 (unsigned long long)sh->sector, method);
1658 case READ_MODIFY_WRITE:
1659 BUG(); /* READ_MODIFY_WRITE N/A for RAID-6 */
1660 case RECONSTRUCT_WRITE:
1661 for (i= disks; i-- ;)
1662 if ( i != pd_idx && i != qd_idx && sh->dev[i].towrite ) {
1663 chosen = sh->dev[i].towrite;
1664 sh->dev[i].towrite = NULL;
1666 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
1667 wake_up(&conf->wait_for_overlap);
1669 BUG_ON(sh->dev[i].written);
1670 sh->dev[i].written = chosen;
1674 BUG(); /* Not implemented yet */
1677 for (i = disks; i--;)
1678 if (sh->dev[i].written) {
1679 sector_t sector = sh->dev[i].sector;
1680 struct bio *wbi = sh->dev[i].written;
1681 while (wbi && wbi->bi_sector < sector + STRIPE_SECTORS) {
1682 copy_data(1, wbi, sh->dev[i].page, sector);
1683 wbi = r5_next_bio(wbi, sector);
1686 set_bit(R5_LOCKED, &sh->dev[i].flags);
1687 set_bit(R5_UPTODATE, &sh->dev[i].flags);
1691 // case RECONSTRUCT_WRITE:
1692 // case CHECK_PARITY:
1693 // case UPDATE_PARITY:
1694 /* Note that unlike RAID-5, the ordering of the disks matters greatly. */
1695 /* FIX: Is this ordering of drives even remotely optimal? */
1699 ptrs[count++] = page_address(sh->dev[i].page);
1700 if (count <= disks-2 && !test_bit(R5_UPTODATE, &sh->dev[i].flags))
1701 printk("block %d/%d not uptodate on parity calc\n", i,count);
1702 i = raid6_next_disk(i, disks);
1703 } while ( i != d0_idx );
1707 raid6_call.gen_syndrome(disks, STRIPE_SIZE, ptrs);
1710 case RECONSTRUCT_WRITE:
1711 set_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
1712 set_bit(R5_UPTODATE, &sh->dev[qd_idx].flags);
1713 set_bit(R5_LOCKED, &sh->dev[pd_idx].flags);
1714 set_bit(R5_LOCKED, &sh->dev[qd_idx].flags);
1717 set_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
1718 set_bit(R5_UPTODATE, &sh->dev[qd_idx].flags);
1724 /* Compute one missing block */
1725 static void compute_block_1(struct stripe_head *sh, int dd_idx, int nozero)
1727 int i, count, disks = sh->disks;
1728 void *ptr[MAX_XOR_BLOCKS], *dest, *p;
1729 int pd_idx = sh->pd_idx;
1730 int qd_idx = raid6_next_disk(pd_idx, disks);
1732 pr_debug("compute_block_1, stripe %llu, idx %d\n",
1733 (unsigned long long)sh->sector, dd_idx);
1735 if ( dd_idx == qd_idx ) {
1736 /* We're actually computing the Q drive */
1737 compute_parity6(sh, UPDATE_PARITY);
1739 dest = page_address(sh->dev[dd_idx].page);
1740 if (!nozero) memset(dest, 0, STRIPE_SIZE);
1742 for (i = disks ; i--; ) {
1743 if (i == dd_idx || i == qd_idx)
1745 p = page_address(sh->dev[i].page);
1746 if (test_bit(R5_UPTODATE, &sh->dev[i].flags))
1749 printk("compute_block() %d, stripe %llu, %d"
1750 " not present\n", dd_idx,
1751 (unsigned long long)sh->sector, i);
1756 xor_blocks(count, STRIPE_SIZE, dest, ptr);
1757 if (!nozero) set_bit(R5_UPTODATE, &sh->dev[dd_idx].flags);
1758 else clear_bit(R5_UPTODATE, &sh->dev[dd_idx].flags);
1762 /* Compute two missing blocks */
1763 static void compute_block_2(struct stripe_head *sh, int dd_idx1, int dd_idx2)
1765 int i, count, disks = sh->disks;
1766 int pd_idx = sh->pd_idx;
1767 int qd_idx = raid6_next_disk(pd_idx, disks);
1768 int d0_idx = raid6_next_disk(qd_idx, disks);
1771 /* faila and failb are disk numbers relative to d0_idx */
1772 /* pd_idx become disks-2 and qd_idx become disks-1 */
1773 faila = (dd_idx1 < d0_idx) ? dd_idx1+(disks-d0_idx) : dd_idx1-d0_idx;
1774 failb = (dd_idx2 < d0_idx) ? dd_idx2+(disks-d0_idx) : dd_idx2-d0_idx;
1776 BUG_ON(faila == failb);
1777 if ( failb < faila ) { int tmp = faila; faila = failb; failb = tmp; }
1779 pr_debug("compute_block_2, stripe %llu, idx %d,%d (%d,%d)\n",
1780 (unsigned long long)sh->sector, dd_idx1, dd_idx2, faila, failb);
1782 if ( failb == disks-1 ) {
1783 /* Q disk is one of the missing disks */
1784 if ( faila == disks-2 ) {
1785 /* Missing P+Q, just recompute */
1786 compute_parity6(sh, UPDATE_PARITY);
1789 /* We're missing D+Q; recompute D from P */
1790 compute_block_1(sh, (dd_idx1 == qd_idx) ? dd_idx2 : dd_idx1, 0);
1791 compute_parity6(sh, UPDATE_PARITY); /* Is this necessary? */
1796 /* We're missing D+P or D+D; build pointer table */
1798 /**** FIX THIS: This could be very bad if disks is close to 256 ****/
1804 ptrs[count++] = page_address(sh->dev[i].page);
1805 i = raid6_next_disk(i, disks);
1806 if (i != dd_idx1 && i != dd_idx2 &&
1807 !test_bit(R5_UPTODATE, &sh->dev[i].flags))
1808 printk("compute_2 with missing block %d/%d\n", count, i);
1809 } while ( i != d0_idx );
1811 if ( failb == disks-2 ) {
1812 /* We're missing D+P. */
1813 raid6_datap_recov(disks, STRIPE_SIZE, faila, ptrs);
1815 /* We're missing D+D. */
1816 raid6_2data_recov(disks, STRIPE_SIZE, faila, failb, ptrs);
1819 /* Both the above update both missing blocks */
1820 set_bit(R5_UPTODATE, &sh->dev[dd_idx1].flags);
1821 set_bit(R5_UPTODATE, &sh->dev[dd_idx2].flags);
1826 handle_write_operations5(struct stripe_head *sh, int rcw, int expand)
1828 int i, pd_idx = sh->pd_idx, disks = sh->disks;
1832 /* if we are not expanding this is a proper write request, and
1833 * there will be bios with new data to be drained into the
1837 set_bit(STRIPE_OP_BIODRAIN, &sh->ops.pending);
1841 set_bit(STRIPE_OP_POSTXOR, &sh->ops.pending);
1844 for (i = disks; i--; ) {
1845 struct r5dev *dev = &sh->dev[i];
1848 set_bit(R5_LOCKED, &dev->flags);
1850 clear_bit(R5_UPTODATE, &dev->flags);
1855 BUG_ON(!(test_bit(R5_UPTODATE, &sh->dev[pd_idx].flags) ||
1856 test_bit(R5_Wantcompute, &sh->dev[pd_idx].flags)));
1858 set_bit(STRIPE_OP_PREXOR, &sh->ops.pending);
1859 set_bit(STRIPE_OP_BIODRAIN, &sh->ops.pending);
1860 set_bit(STRIPE_OP_POSTXOR, &sh->ops.pending);
1864 for (i = disks; i--; ) {
1865 struct r5dev *dev = &sh->dev[i];
1869 /* For a read-modify write there may be blocks that are
1870 * locked for reading while others are ready to be
1871 * written so we distinguish these blocks by the
1875 (test_bit(R5_UPTODATE, &dev->flags) ||
1876 test_bit(R5_Wantcompute, &dev->flags))) {
1877 set_bit(R5_Wantprexor, &dev->flags);
1878 set_bit(R5_LOCKED, &dev->flags);
1879 clear_bit(R5_UPTODATE, &dev->flags);
1885 /* keep the parity disk locked while asynchronous operations
1888 set_bit(R5_LOCKED, &sh->dev[pd_idx].flags);
1889 clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
1892 pr_debug("%s: stripe %llu locked: %d pending: %lx\n",
1893 __FUNCTION__, (unsigned long long)sh->sector,
1894 locked, sh->ops.pending);
1900 * Each stripe/dev can have one or more bion attached.
1901 * toread/towrite point to the first in a chain.
1902 * The bi_next chain must be in order.
1904 static int add_stripe_bio(struct stripe_head *sh, struct bio *bi, int dd_idx, int forwrite)
1907 raid5_conf_t *conf = sh->raid_conf;
1910 pr_debug("adding bh b#%llu to stripe s#%llu\n",
1911 (unsigned long long)bi->bi_sector,
1912 (unsigned long long)sh->sector);
1915 spin_lock(&sh->lock);
1916 spin_lock_irq(&conf->device_lock);
1918 bip = &sh->dev[dd_idx].towrite;
1919 if (*bip == NULL && sh->dev[dd_idx].written == NULL)
1922 bip = &sh->dev[dd_idx].toread;
1923 while (*bip && (*bip)->bi_sector < bi->bi_sector) {
1924 if ((*bip)->bi_sector + ((*bip)->bi_size >> 9) > bi->bi_sector)
1926 bip = & (*bip)->bi_next;
1928 if (*bip && (*bip)->bi_sector < bi->bi_sector + ((bi->bi_size)>>9))
1931 BUG_ON(*bip && bi->bi_next && (*bip) != bi->bi_next);
1935 bi->bi_phys_segments ++;
1936 spin_unlock_irq(&conf->device_lock);
1937 spin_unlock(&sh->lock);
1939 pr_debug("added bi b#%llu to stripe s#%llu, disk %d.\n",
1940 (unsigned long long)bi->bi_sector,
1941 (unsigned long long)sh->sector, dd_idx);
1943 if (conf->mddev->bitmap && firstwrite) {
1944 bitmap_startwrite(conf->mddev->bitmap, sh->sector,
1946 sh->bm_seq = conf->seq_flush+1;
1947 set_bit(STRIPE_BIT_DELAY, &sh->state);
1951 /* check if page is covered */
1952 sector_t sector = sh->dev[dd_idx].sector;
1953 for (bi=sh->dev[dd_idx].towrite;
1954 sector < sh->dev[dd_idx].sector + STRIPE_SECTORS &&
1955 bi && bi->bi_sector <= sector;
1956 bi = r5_next_bio(bi, sh->dev[dd_idx].sector)) {
1957 if (bi->bi_sector + (bi->bi_size>>9) >= sector)
1958 sector = bi->bi_sector + (bi->bi_size>>9);
1960 if (sector >= sh->dev[dd_idx].sector + STRIPE_SECTORS)
1961 set_bit(R5_OVERWRITE, &sh->dev[dd_idx].flags);
1966 set_bit(R5_Overlap, &sh->dev[dd_idx].flags);
1967 spin_unlock_irq(&conf->device_lock);
1968 spin_unlock(&sh->lock);
1972 static void end_reshape(raid5_conf_t *conf);
1974 static int page_is_zero(struct page *p)
1976 char *a = page_address(p);
1977 return ((*(u32*)a) == 0 &&
1978 memcmp(a, a+4, STRIPE_SIZE-4)==0);
1981 static int stripe_to_pdidx(sector_t stripe, raid5_conf_t *conf, int disks)
1983 int sectors_per_chunk = conf->chunk_size >> 9;
1985 int chunk_offset = sector_div(stripe, sectors_per_chunk);
1987 raid5_compute_sector(stripe * (disks - conf->max_degraded)
1988 *sectors_per_chunk + chunk_offset,
1989 disks, disks - conf->max_degraded,
1990 &dd_idx, &pd_idx, conf);
1995 handle_requests_to_failed_array(raid5_conf_t *conf, struct stripe_head *sh,
1996 struct stripe_head_state *s, int disks,
1997 struct bio **return_bi)
2000 for (i = disks; i--; ) {
2004 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
2007 rdev = rcu_dereference(conf->disks[i].rdev);
2008 if (rdev && test_bit(In_sync, &rdev->flags))
2009 /* multiple read failures in one stripe */
2010 md_error(conf->mddev, rdev);
2013 spin_lock_irq(&conf->device_lock);
2014 /* fail all writes first */
2015 bi = sh->dev[i].towrite;
2016 sh->dev[i].towrite = NULL;
2022 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
2023 wake_up(&conf->wait_for_overlap);
2025 while (bi && bi->bi_sector <
2026 sh->dev[i].sector + STRIPE_SECTORS) {
2027 struct bio *nextbi = r5_next_bio(bi, sh->dev[i].sector);
2028 clear_bit(BIO_UPTODATE, &bi->bi_flags);
2029 if (--bi->bi_phys_segments == 0) {
2030 md_write_end(conf->mddev);
2031 bi->bi_next = *return_bi;
2036 /* and fail all 'written' */
2037 bi = sh->dev[i].written;
2038 sh->dev[i].written = NULL;
2039 if (bi) bitmap_end = 1;
2040 while (bi && bi->bi_sector <
2041 sh->dev[i].sector + STRIPE_SECTORS) {
2042 struct bio *bi2 = r5_next_bio(bi, sh->dev[i].sector);
2043 clear_bit(BIO_UPTODATE, &bi->bi_flags);
2044 if (--bi->bi_phys_segments == 0) {
2045 md_write_end(conf->mddev);
2046 bi->bi_next = *return_bi;
2052 /* fail any reads if this device is non-operational */
2053 if (!test_bit(R5_Insync, &sh->dev[i].flags) ||
2054 test_bit(R5_ReadError, &sh->dev[i].flags)) {
2055 bi = sh->dev[i].toread;
2056 sh->dev[i].toread = NULL;
2057 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
2058 wake_up(&conf->wait_for_overlap);
2059 if (bi) s->to_read--;
2060 while (bi && bi->bi_sector <
2061 sh->dev[i].sector + STRIPE_SECTORS) {
2062 struct bio *nextbi =
2063 r5_next_bio(bi, sh->dev[i].sector);
2064 clear_bit(BIO_UPTODATE, &bi->bi_flags);
2065 if (--bi->bi_phys_segments == 0) {
2066 bi->bi_next = *return_bi;
2072 spin_unlock_irq(&conf->device_lock);
2074 bitmap_endwrite(conf->mddev->bitmap, sh->sector,
2075 STRIPE_SECTORS, 0, 0);
2080 /* __handle_issuing_new_read_requests5 - returns 0 if there are no more disks
2083 static int __handle_issuing_new_read_requests5(struct stripe_head *sh,
2084 struct stripe_head_state *s, int disk_idx, int disks)
2086 struct r5dev *dev = &sh->dev[disk_idx];
2087 struct r5dev *failed_dev = &sh->dev[s->failed_num];
2089 /* don't schedule compute operations or reads on the parity block while
2090 * a check is in flight
2092 if ((disk_idx == sh->pd_idx) &&
2093 test_bit(STRIPE_OP_CHECK, &sh->ops.pending))
2096 /* is the data in this block needed, and can we get it? */
2097 if (!test_bit(R5_LOCKED, &dev->flags) &&
2098 !test_bit(R5_UPTODATE, &dev->flags) && (dev->toread ||
2099 (dev->towrite && !test_bit(R5_OVERWRITE, &dev->flags)) ||
2100 s->syncing || s->expanding || (s->failed &&
2101 (failed_dev->toread || (failed_dev->towrite &&
2102 !test_bit(R5_OVERWRITE, &failed_dev->flags)
2104 /* 1/ We would like to get this block, possibly by computing it,
2105 * but we might not be able to.
2107 * 2/ Since parity check operations potentially make the parity
2108 * block !uptodate it will need to be refreshed before any
2109 * compute operations on data disks are scheduled.
2111 * 3/ We hold off parity block re-reads until check operations
2114 if ((s->uptodate == disks - 1) &&
2115 !test_bit(STRIPE_OP_CHECK, &sh->ops.pending)) {
2116 set_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.pending);
2117 set_bit(R5_Wantcompute, &dev->flags);
2118 sh->ops.target = disk_idx;
2121 /* Careful: from this point on 'uptodate' is in the eye
2122 * of raid5_run_ops which services 'compute' operations
2123 * before writes. R5_Wantcompute flags a block that will
2124 * be R5_UPTODATE by the time it is needed for a
2125 * subsequent operation.
2128 return 0; /* uptodate + compute == disks */
2129 } else if ((s->uptodate < disks - 1) &&
2130 test_bit(R5_Insync, &dev->flags)) {
2131 /* Note: we hold off compute operations while checks are
2132 * in flight, but we still prefer 'compute' over 'read'
2133 * hence we only read if (uptodate < * disks-1)
2135 set_bit(R5_LOCKED, &dev->flags);
2136 set_bit(R5_Wantread, &dev->flags);
2137 if (!test_and_set_bit(STRIPE_OP_IO, &sh->ops.pending))
2140 pr_debug("Reading block %d (sync=%d)\n", disk_idx,
2148 static void handle_issuing_new_read_requests5(struct stripe_head *sh,
2149 struct stripe_head_state *s, int disks)
2153 /* Clear completed compute operations. Parity recovery
2154 * (STRIPE_OP_MOD_REPAIR_PD) implies a write-back which is handled
2155 * later on in this routine
2157 if (test_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.complete) &&
2158 !test_bit(STRIPE_OP_MOD_REPAIR_PD, &sh->ops.pending)) {
2159 clear_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.complete);
2160 clear_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.ack);
2161 clear_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.pending);
2164 /* look for blocks to read/compute, skip this if a compute
2165 * is already in flight, or if the stripe contents are in the
2166 * midst of changing due to a write
2168 if (!test_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.pending) &&
2169 !test_bit(STRIPE_OP_PREXOR, &sh->ops.pending) &&
2170 !test_bit(STRIPE_OP_POSTXOR, &sh->ops.pending)) {
2171 for (i = disks; i--; )
2172 if (__handle_issuing_new_read_requests5(
2173 sh, s, i, disks) == 0)
2176 set_bit(STRIPE_HANDLE, &sh->state);
2179 static void handle_issuing_new_read_requests6(struct stripe_head *sh,
2180 struct stripe_head_state *s, struct r6_state *r6s,
2184 for (i = disks; i--; ) {
2185 struct r5dev *dev = &sh->dev[i];
2186 if (!test_bit(R5_LOCKED, &dev->flags) &&
2187 !test_bit(R5_UPTODATE, &dev->flags) &&
2188 (dev->toread || (dev->towrite &&
2189 !test_bit(R5_OVERWRITE, &dev->flags)) ||
2190 s->syncing || s->expanding ||
2192 (sh->dev[r6s->failed_num[0]].toread ||
2195 (sh->dev[r6s->failed_num[1]].toread ||
2197 /* we would like to get this block, possibly
2198 * by computing it, but we might not be able to
2200 if (s->uptodate == disks-1) {
2201 pr_debug("Computing stripe %llu block %d\n",
2202 (unsigned long long)sh->sector, i);
2203 compute_block_1(sh, i, 0);
2205 } else if ( s->uptodate == disks-2 && s->failed >= 2 ) {
2206 /* Computing 2-failure is *very* expensive; only
2207 * do it if failed >= 2
2210 for (other = disks; other--; ) {
2213 if (!test_bit(R5_UPTODATE,
2214 &sh->dev[other].flags))
2218 pr_debug("Computing stripe %llu blocks %d,%d\n",
2219 (unsigned long long)sh->sector,
2221 compute_block_2(sh, i, other);
2223 } else if (test_bit(R5_Insync, &dev->flags)) {
2224 set_bit(R5_LOCKED, &dev->flags);
2225 set_bit(R5_Wantread, &dev->flags);
2227 pr_debug("Reading block %d (sync=%d)\n",
2232 set_bit(STRIPE_HANDLE, &sh->state);
2236 /* handle_completed_write_requests
2237 * any written block on an uptodate or failed drive can be returned.
2238 * Note that if we 'wrote' to a failed drive, it will be UPTODATE, but
2239 * never LOCKED, so we don't need to test 'failed' directly.
2241 static void handle_completed_write_requests(raid5_conf_t *conf,
2242 struct stripe_head *sh, int disks, struct bio **return_bi)
2247 for (i = disks; i--; )
2248 if (sh->dev[i].written) {
2250 if (!test_bit(R5_LOCKED, &dev->flags) &&
2251 test_bit(R5_UPTODATE, &dev->flags)) {
2252 /* We can return any write requests */
2253 struct bio *wbi, *wbi2;
2255 pr_debug("Return write for disc %d\n", i);
2256 spin_lock_irq(&conf->device_lock);
2258 dev->written = NULL;
2259 while (wbi && wbi->bi_sector <
2260 dev->sector + STRIPE_SECTORS) {
2261 wbi2 = r5_next_bio(wbi, dev->sector);
2262 if (--wbi->bi_phys_segments == 0) {
2263 md_write_end(conf->mddev);
2264 wbi->bi_next = *return_bi;
2269 if (dev->towrite == NULL)
2271 spin_unlock_irq(&conf->device_lock);
2273 bitmap_endwrite(conf->mddev->bitmap,
2276 !test_bit(STRIPE_DEGRADED, &sh->state),
2282 static void handle_issuing_new_write_requests5(raid5_conf_t *conf,
2283 struct stripe_head *sh, struct stripe_head_state *s, int disks)
2285 int rmw = 0, rcw = 0, i;
2286 for (i = disks; i--; ) {
2287 /* would I have to read this buffer for read_modify_write */
2288 struct r5dev *dev = &sh->dev[i];
2289 if ((dev->towrite || i == sh->pd_idx) &&
2290 !test_bit(R5_LOCKED, &dev->flags) &&
2291 !(test_bit(R5_UPTODATE, &dev->flags) ||
2292 test_bit(R5_Wantcompute, &dev->flags))) {
2293 if (test_bit(R5_Insync, &dev->flags))
2296 rmw += 2*disks; /* cannot read it */
2298 /* Would I have to read this buffer for reconstruct_write */
2299 if (!test_bit(R5_OVERWRITE, &dev->flags) && i != sh->pd_idx &&
2300 !test_bit(R5_LOCKED, &dev->flags) &&
2301 !(test_bit(R5_UPTODATE, &dev->flags) ||
2302 test_bit(R5_Wantcompute, &dev->flags))) {
2303 if (test_bit(R5_Insync, &dev->flags)) rcw++;
2308 pr_debug("for sector %llu, rmw=%d rcw=%d\n",
2309 (unsigned long long)sh->sector, rmw, rcw);
2310 set_bit(STRIPE_HANDLE, &sh->state);
2311 if (rmw < rcw && rmw > 0)
2312 /* prefer read-modify-write, but need to get some data */
2313 for (i = disks; i--; ) {
2314 struct r5dev *dev = &sh->dev[i];
2315 if ((dev->towrite || i == sh->pd_idx) &&
2316 !test_bit(R5_LOCKED, &dev->flags) &&
2317 !(test_bit(R5_UPTODATE, &dev->flags) ||
2318 test_bit(R5_Wantcompute, &dev->flags)) &&
2319 test_bit(R5_Insync, &dev->flags)) {
2321 test_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2322 pr_debug("Read_old block "
2323 "%d for r-m-w\n", i);
2324 set_bit(R5_LOCKED, &dev->flags);
2325 set_bit(R5_Wantread, &dev->flags);
2328 set_bit(STRIPE_DELAYED, &sh->state);
2329 set_bit(STRIPE_HANDLE, &sh->state);
2333 if (rcw <= rmw && rcw > 0)
2334 /* want reconstruct write, but need to get some data */
2335 for (i = disks; i--; ) {
2336 struct r5dev *dev = &sh->dev[i];
2337 if (!test_bit(R5_OVERWRITE, &dev->flags) &&
2339 !test_bit(R5_LOCKED, &dev->flags) &&
2340 !(test_bit(R5_UPTODATE, &dev->flags) ||
2341 test_bit(R5_Wantcompute, &dev->flags)) &&
2342 test_bit(R5_Insync, &dev->flags)) {
2344 test_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2345 pr_debug("Read_old block "
2346 "%d for Reconstruct\n", i);
2347 set_bit(R5_LOCKED, &dev->flags);
2348 set_bit(R5_Wantread, &dev->flags);
2351 set_bit(STRIPE_DELAYED, &sh->state);
2352 set_bit(STRIPE_HANDLE, &sh->state);
2356 /* now if nothing is locked, and if we have enough data,
2357 * we can start a write request
2359 /* since handle_stripe can be called at any time we need to handle the
2360 * case where a compute block operation has been submitted and then a
2361 * subsequent call wants to start a write request. raid5_run_ops only
2362 * handles the case where compute block and postxor are requested
2363 * simultaneously. If this is not the case then new writes need to be
2364 * held off until the compute completes.
2366 if ((s->req_compute ||
2367 !test_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.pending)) &&
2368 (s->locked == 0 && (rcw == 0 || rmw == 0) &&
2369 !test_bit(STRIPE_BIT_DELAY, &sh->state)))
2370 s->locked += handle_write_operations5(sh, rcw == 0, 0);
2373 static void handle_issuing_new_write_requests6(raid5_conf_t *conf,
2374 struct stripe_head *sh, struct stripe_head_state *s,
2375 struct r6_state *r6s, int disks)
2377 int rcw = 0, must_compute = 0, pd_idx = sh->pd_idx, i;
2378 int qd_idx = r6s->qd_idx;
2379 for (i = disks; i--; ) {
2380 struct r5dev *dev = &sh->dev[i];
2381 /* Would I have to read this buffer for reconstruct_write */
2382 if (!test_bit(R5_OVERWRITE, &dev->flags)
2383 && i != pd_idx && i != qd_idx
2384 && (!test_bit(R5_LOCKED, &dev->flags)
2386 !test_bit(R5_UPTODATE, &dev->flags)) {
2387 if (test_bit(R5_Insync, &dev->flags)) rcw++;
2389 pr_debug("raid6: must_compute: "
2390 "disk %d flags=%#lx\n", i, dev->flags);
2395 pr_debug("for sector %llu, rcw=%d, must_compute=%d\n",
2396 (unsigned long long)sh->sector, rcw, must_compute);
2397 set_bit(STRIPE_HANDLE, &sh->state);
2400 /* want reconstruct write, but need to get some data */
2401 for (i = disks; i--; ) {
2402 struct r5dev *dev = &sh->dev[i];
2403 if (!test_bit(R5_OVERWRITE, &dev->flags)
2404 && !(s->failed == 0 && (i == pd_idx || i == qd_idx))
2405 && !test_bit(R5_LOCKED, &dev->flags) &&
2406 !test_bit(R5_UPTODATE, &dev->flags) &&
2407 test_bit(R5_Insync, &dev->flags)) {
2409 test_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2410 pr_debug("Read_old stripe %llu "
2411 "block %d for Reconstruct\n",
2412 (unsigned long long)sh->sector, i);
2413 set_bit(R5_LOCKED, &dev->flags);
2414 set_bit(R5_Wantread, &dev->flags);
2417 pr_debug("Request delayed stripe %llu "
2418 "block %d for Reconstruct\n",
2419 (unsigned long long)sh->sector, i);
2420 set_bit(STRIPE_DELAYED, &sh->state);
2421 set_bit(STRIPE_HANDLE, &sh->state);
2425 /* now if nothing is locked, and if we have enough data, we can start a
2428 if (s->locked == 0 && rcw == 0 &&
2429 !test_bit(STRIPE_BIT_DELAY, &sh->state)) {
2430 if (must_compute > 0) {
2431 /* We have failed blocks and need to compute them */
2432 switch (s->failed) {
2436 compute_block_1(sh, r6s->failed_num[0], 0);
2439 compute_block_2(sh, r6s->failed_num[0],
2440 r6s->failed_num[1]);
2442 default: /* This request should have been failed? */
2447 pr_debug("Computing parity for stripe %llu\n",
2448 (unsigned long long)sh->sector);
2449 compute_parity6(sh, RECONSTRUCT_WRITE);
2450 /* now every locked buffer is ready to be written */
2451 for (i = disks; i--; )
2452 if (test_bit(R5_LOCKED, &sh->dev[i].flags)) {
2453 pr_debug("Writing stripe %llu block %d\n",
2454 (unsigned long long)sh->sector, i);
2456 set_bit(R5_Wantwrite, &sh->dev[i].flags);
2458 /* after a RECONSTRUCT_WRITE, the stripe MUST be in-sync */
2459 set_bit(STRIPE_INSYNC, &sh->state);
2461 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2462 atomic_dec(&conf->preread_active_stripes);
2463 if (atomic_read(&conf->preread_active_stripes) <
2465 md_wakeup_thread(conf->mddev->thread);
2470 static void handle_parity_checks5(raid5_conf_t *conf, struct stripe_head *sh,
2471 struct stripe_head_state *s, int disks)
2473 set_bit(STRIPE_HANDLE, &sh->state);
2474 if (s->failed == 0) {
2475 BUG_ON(s->uptodate != disks);
2476 compute_parity5(sh, CHECK_PARITY);
2478 if (page_is_zero(sh->dev[sh->pd_idx].page)) {
2479 /* parity is correct (on disc, not in buffer any more)
2481 set_bit(STRIPE_INSYNC, &sh->state);
2483 conf->mddev->resync_mismatches += STRIPE_SECTORS;
2484 if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery))
2485 /* don't try to repair!! */
2486 set_bit(STRIPE_INSYNC, &sh->state);
2488 compute_block(sh, sh->pd_idx);
2493 if (!test_bit(STRIPE_INSYNC, &sh->state)) {
2495 /* either failed parity check, or recovery is happening */
2497 s->failed_num = sh->pd_idx;
2498 dev = &sh->dev[s->failed_num];
2499 BUG_ON(!test_bit(R5_UPTODATE, &dev->flags));
2500 BUG_ON(s->uptodate != disks);
2502 set_bit(R5_LOCKED, &dev->flags);
2503 set_bit(R5_Wantwrite, &dev->flags);
2504 clear_bit(STRIPE_DEGRADED, &sh->state);
2506 set_bit(STRIPE_INSYNC, &sh->state);
2511 static void handle_parity_checks6(raid5_conf_t *conf, struct stripe_head *sh,
2512 struct stripe_head_state *s,
2513 struct r6_state *r6s, struct page *tmp_page,
2516 int update_p = 0, update_q = 0;
2518 int pd_idx = sh->pd_idx;
2519 int qd_idx = r6s->qd_idx;
2521 set_bit(STRIPE_HANDLE, &sh->state);
2523 BUG_ON(s->failed > 2);
2524 BUG_ON(s->uptodate < disks);
2525 /* Want to check and possibly repair P and Q.
2526 * However there could be one 'failed' device, in which
2527 * case we can only check one of them, possibly using the
2528 * other to generate missing data
2531 /* If !tmp_page, we cannot do the calculations,
2532 * but as we have set STRIPE_HANDLE, we will soon be called
2533 * by stripe_handle with a tmp_page - just wait until then.
2536 if (s->failed == r6s->q_failed) {
2537 /* The only possible failed device holds 'Q', so it
2538 * makes sense to check P (If anything else were failed,
2539 * we would have used P to recreate it).
2541 compute_block_1(sh, pd_idx, 1);
2542 if (!page_is_zero(sh->dev[pd_idx].page)) {
2543 compute_block_1(sh, pd_idx, 0);
2547 if (!r6s->q_failed && s->failed < 2) {
2548 /* q is not failed, and we didn't use it to generate
2549 * anything, so it makes sense to check it
2551 memcpy(page_address(tmp_page),
2552 page_address(sh->dev[qd_idx].page),
2554 compute_parity6(sh, UPDATE_PARITY);
2555 if (memcmp(page_address(tmp_page),
2556 page_address(sh->dev[qd_idx].page),
2557 STRIPE_SIZE) != 0) {
2558 clear_bit(STRIPE_INSYNC, &sh->state);
2562 if (update_p || update_q) {
2563 conf->mddev->resync_mismatches += STRIPE_SECTORS;
2564 if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery))
2565 /* don't try to repair!! */
2566 update_p = update_q = 0;
2569 /* now write out any block on a failed drive,
2570 * or P or Q if they need it
2573 if (s->failed == 2) {
2574 dev = &sh->dev[r6s->failed_num[1]];
2576 set_bit(R5_LOCKED, &dev->flags);
2577 set_bit(R5_Wantwrite, &dev->flags);
2579 if (s->failed >= 1) {
2580 dev = &sh->dev[r6s->failed_num[0]];
2582 set_bit(R5_LOCKED, &dev->flags);
2583 set_bit(R5_Wantwrite, &dev->flags);
2587 dev = &sh->dev[pd_idx];
2589 set_bit(R5_LOCKED, &dev->flags);
2590 set_bit(R5_Wantwrite, &dev->flags);
2593 dev = &sh->dev[qd_idx];
2595 set_bit(R5_LOCKED, &dev->flags);
2596 set_bit(R5_Wantwrite, &dev->flags);
2598 clear_bit(STRIPE_DEGRADED, &sh->state);
2600 set_bit(STRIPE_INSYNC, &sh->state);
2604 static void handle_stripe_expansion(raid5_conf_t *conf, struct stripe_head *sh,
2605 struct r6_state *r6s)
2609 /* We have read all the blocks in this stripe and now we need to
2610 * copy some of them into a target stripe for expand.
2612 clear_bit(STRIPE_EXPAND_SOURCE, &sh->state);
2613 for (i = 0; i < sh->disks; i++)
2614 if (i != sh->pd_idx && (r6s && i != r6s->qd_idx)) {
2615 int dd_idx, pd_idx, j;
2616 struct stripe_head *sh2;
2618 sector_t bn = compute_blocknr(sh, i);
2619 sector_t s = raid5_compute_sector(bn, conf->raid_disks,
2621 conf->max_degraded, &dd_idx,
2623 sh2 = get_active_stripe(conf, s, conf->raid_disks,
2626 /* so far only the early blocks of this stripe
2627 * have been requested. When later blocks
2628 * get requested, we will try again
2631 if (!test_bit(STRIPE_EXPANDING, &sh2->state) ||
2632 test_bit(R5_Expanded, &sh2->dev[dd_idx].flags)) {
2633 /* must have already done this block */
2634 release_stripe(sh2);
2637 memcpy(page_address(sh2->dev[dd_idx].page),
2638 page_address(sh->dev[i].page),
2640 set_bit(R5_Expanded, &sh2->dev[dd_idx].flags);
2641 set_bit(R5_UPTODATE, &sh2->dev[dd_idx].flags);
2642 for (j = 0; j < conf->raid_disks; j++)
2643 if (j != sh2->pd_idx &&
2644 (r6s && j != r6s->qd_idx) &&
2645 !test_bit(R5_Expanded, &sh2->dev[j].flags))
2647 if (j == conf->raid_disks) {
2648 set_bit(STRIPE_EXPAND_READY, &sh2->state);
2649 set_bit(STRIPE_HANDLE, &sh2->state);
2651 release_stripe(sh2);
2656 * handle_stripe - do things to a stripe.
2658 * We lock the stripe and then examine the state of various bits
2659 * to see what needs to be done.
2661 * return some read request which now have data
2662 * return some write requests which are safely on disc
2663 * schedule a read on some buffers
2664 * schedule a write of some buffers
2665 * return confirmation of parity correctness
2667 * buffers are taken off read_list or write_list, and bh_cache buffers
2668 * get BH_Lock set before the stripe lock is released.
2672 static void handle_stripe5(struct stripe_head *sh)
2674 raid5_conf_t *conf = sh->raid_conf;
2675 int disks = sh->disks, i;
2676 struct bio *return_bi = NULL;
2677 struct stripe_head_state s;
2679 unsigned long pending = 0;
2681 memset(&s, 0, sizeof(s));
2682 pr_debug("handling stripe %llu, state=%#lx cnt=%d, pd_idx=%d "
2683 "ops=%lx:%lx:%lx\n", (unsigned long long)sh->sector, sh->state,
2684 atomic_read(&sh->count), sh->pd_idx,
2685 sh->ops.pending, sh->ops.ack, sh->ops.complete);
2687 spin_lock(&sh->lock);
2688 clear_bit(STRIPE_HANDLE, &sh->state);
2689 clear_bit(STRIPE_DELAYED, &sh->state);
2691 s.syncing = test_bit(STRIPE_SYNCING, &sh->state);
2692 s.expanding = test_bit(STRIPE_EXPAND_SOURCE, &sh->state);
2693 s.expanded = test_bit(STRIPE_EXPAND_READY, &sh->state);
2694 /* Now to look around and see what can be done */
2697 for (i=disks; i--; ) {
2699 struct r5dev *dev = &sh->dev[i];
2700 clear_bit(R5_Insync, &dev->flags);
2702 pr_debug("check %d: state 0x%lx read %p write %p written %p\n",
2703 i, dev->flags, dev->toread, dev->towrite, dev->written);
2704 /* maybe we can reply to a read */
2705 if (test_bit(R5_UPTODATE, &dev->flags) && dev->toread) {
2706 struct bio *rbi, *rbi2;
2707 pr_debug("Return read for disc %d\n", i);
2708 spin_lock_irq(&conf->device_lock);
2711 if (test_and_clear_bit(R5_Overlap, &dev->flags))
2712 wake_up(&conf->wait_for_overlap);
2713 spin_unlock_irq(&conf->device_lock);
2714 while (rbi && rbi->bi_sector < dev->sector + STRIPE_SECTORS) {
2715 copy_data(0, rbi, dev->page, dev->sector);
2716 rbi2 = r5_next_bio(rbi, dev->sector);
2717 spin_lock_irq(&conf->device_lock);
2718 if (--rbi->bi_phys_segments == 0) {
2719 rbi->bi_next = return_bi;
2722 spin_unlock_irq(&conf->device_lock);
2727 /* now count some things */
2728 if (test_bit(R5_LOCKED, &dev->flags)) s.locked++;
2729 if (test_bit(R5_UPTODATE, &dev->flags)) s.uptodate++;
2730 if (test_bit(R5_Wantcompute, &dev->flags)) s.compute++;
2736 if (!test_bit(R5_OVERWRITE, &dev->flags))
2741 rdev = rcu_dereference(conf->disks[i].rdev);
2742 if (!rdev || !test_bit(In_sync, &rdev->flags)) {
2743 /* The ReadError flag will just be confusing now */
2744 clear_bit(R5_ReadError, &dev->flags);
2745 clear_bit(R5_ReWrite, &dev->flags);
2747 if (!rdev || !test_bit(In_sync, &rdev->flags)
2748 || test_bit(R5_ReadError, &dev->flags)) {
2752 set_bit(R5_Insync, &dev->flags);
2755 pr_debug("locked=%d uptodate=%d to_read=%d"
2756 " to_write=%d failed=%d failed_num=%d\n",
2757 s.locked, s.uptodate, s.to_read, s.to_write,
2758 s.failed, s.failed_num);
2759 /* check if the array has lost two devices and, if so, some requests might
2762 if (s.failed > 1 && s.to_read+s.to_write+s.written)
2763 handle_requests_to_failed_array(conf, sh, &s, disks,
2765 if (s.failed > 1 && s.syncing) {
2766 md_done_sync(conf->mddev, STRIPE_SECTORS,0);
2767 clear_bit(STRIPE_SYNCING, &sh->state);
2771 /* might be able to return some write requests if the parity block
2772 * is safe, or on a failed drive
2774 dev = &sh->dev[sh->pd_idx];
2776 ((test_bit(R5_Insync, &dev->flags) &&
2777 !test_bit(R5_LOCKED, &dev->flags) &&
2778 test_bit(R5_UPTODATE, &dev->flags)) ||
2779 (s.failed == 1 && s.failed_num == sh->pd_idx)))
2780 handle_completed_write_requests(conf, sh, disks, &return_bi);
2782 /* Now we might consider reading some blocks, either to check/generate
2783 * parity, or to satisfy requests
2784 * or to load a block that is being partially written.
2786 if (s.to_read || s.non_overwrite ||
2787 (s.syncing && (s.uptodate + s.compute < disks)) || s.expanding ||
2788 test_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.pending))
2789 handle_issuing_new_read_requests5(sh, &s, disks);
2791 /* Now we check to see if any write operations have recently
2795 /* leave prexor set until postxor is done, allows us to distinguish
2796 * a rmw from a rcw during biodrain
2798 if (test_bit(STRIPE_OP_PREXOR, &sh->ops.complete) &&
2799 test_bit(STRIPE_OP_POSTXOR, &sh->ops.complete)) {
2801 clear_bit(STRIPE_OP_PREXOR, &sh->ops.complete);
2802 clear_bit(STRIPE_OP_PREXOR, &sh->ops.ack);
2803 clear_bit(STRIPE_OP_PREXOR, &sh->ops.pending);
2805 for (i = disks; i--; )
2806 clear_bit(R5_Wantprexor, &sh->dev[i].flags);
2809 /* if only POSTXOR is set then this is an 'expand' postxor */
2810 if (test_bit(STRIPE_OP_BIODRAIN, &sh->ops.complete) &&
2811 test_bit(STRIPE_OP_POSTXOR, &sh->ops.complete)) {
2813 clear_bit(STRIPE_OP_BIODRAIN, &sh->ops.complete);
2814 clear_bit(STRIPE_OP_BIODRAIN, &sh->ops.ack);
2815 clear_bit(STRIPE_OP_BIODRAIN, &sh->ops.pending);
2817 clear_bit(STRIPE_OP_POSTXOR, &sh->ops.complete);
2818 clear_bit(STRIPE_OP_POSTXOR, &sh->ops.ack);
2819 clear_bit(STRIPE_OP_POSTXOR, &sh->ops.pending);
2821 /* All the 'written' buffers and the parity block are ready to
2822 * be written back to disk
2824 BUG_ON(!test_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags));
2825 for (i = disks; i--; ) {
2827 if (test_bit(R5_LOCKED, &dev->flags) &&
2828 (i == sh->pd_idx || dev->written)) {
2829 pr_debug("Writing block %d\n", i);
2830 set_bit(R5_Wantwrite, &dev->flags);
2831 if (!test_and_set_bit(
2832 STRIPE_OP_IO, &sh->ops.pending))
2834 if (!test_bit(R5_Insync, &dev->flags) ||
2835 (i == sh->pd_idx && s.failed == 0))
2836 set_bit(STRIPE_INSYNC, &sh->state);
2839 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2840 atomic_dec(&conf->preread_active_stripes);
2841 if (atomic_read(&conf->preread_active_stripes) <
2843 md_wakeup_thread(conf->mddev->thread);
2847 /* Now to consider new write requests and what else, if anything
2848 * should be read. We do not handle new writes when:
2849 * 1/ A 'write' operation (copy+xor) is already in flight.
2850 * 2/ A 'check' operation is in flight, as it may clobber the parity
2853 if (s.to_write && !test_bit(STRIPE_OP_POSTXOR, &sh->ops.pending) &&
2854 !test_bit(STRIPE_OP_CHECK, &sh->ops.pending))
2855 handle_issuing_new_write_requests5(conf, sh, &s, disks);
2857 /* maybe we need to check and possibly fix the parity for this stripe
2858 * Any reads will already have been scheduled, so we just see if enough data
2861 if (s.syncing && s.locked == 0 &&
2862 !test_bit(STRIPE_INSYNC, &sh->state))
2863 handle_parity_checks5(conf, sh, &s, disks);
2864 if (s.syncing && s.locked == 0 && test_bit(STRIPE_INSYNC, &sh->state)) {
2865 md_done_sync(conf->mddev, STRIPE_SECTORS,1);
2866 clear_bit(STRIPE_SYNCING, &sh->state);
2869 /* If the failed drive is just a ReadError, then we might need to progress
2870 * the repair/check process
2872 if (s.failed == 1 && !conf->mddev->ro &&
2873 test_bit(R5_ReadError, &sh->dev[s.failed_num].flags)
2874 && !test_bit(R5_LOCKED, &sh->dev[s.failed_num].flags)
2875 && test_bit(R5_UPTODATE, &sh->dev[s.failed_num].flags)
2877 dev = &sh->dev[s.failed_num];
2878 if (!test_bit(R5_ReWrite, &dev->flags)) {
2879 set_bit(R5_Wantwrite, &dev->flags);
2880 set_bit(R5_ReWrite, &dev->flags);
2881 set_bit(R5_LOCKED, &dev->flags);
2884 /* let's read it back */
2885 set_bit(R5_Wantread, &dev->flags);
2886 set_bit(R5_LOCKED, &dev->flags);
2891 if (s.expanded && test_bit(STRIPE_EXPANDING, &sh->state)) {
2892 /* Need to write out all blocks after computing parity */
2893 sh->disks = conf->raid_disks;
2894 sh->pd_idx = stripe_to_pdidx(sh->sector, conf, conf->raid_disks);
2895 compute_parity5(sh, RECONSTRUCT_WRITE);
2896 for (i = conf->raid_disks; i--; ) {
2897 set_bit(R5_LOCKED, &sh->dev[i].flags);
2899 set_bit(R5_Wantwrite, &sh->dev[i].flags);
2901 clear_bit(STRIPE_EXPANDING, &sh->state);
2902 } else if (s.expanded) {
2903 clear_bit(STRIPE_EXPAND_READY, &sh->state);
2904 atomic_dec(&conf->reshape_stripes);
2905 wake_up(&conf->wait_for_overlap);
2906 md_done_sync(conf->mddev, STRIPE_SECTORS, 1);
2909 if (s.expanding && s.locked == 0)
2910 handle_stripe_expansion(conf, sh, NULL);
2913 pending = get_stripe_work(sh);
2915 spin_unlock(&sh->lock);
2918 raid5_run_ops(sh, pending);
2920 return_io(return_bi);
2922 for (i=disks; i-- ;) {
2926 if (test_and_clear_bit(R5_Wantwrite, &sh->dev[i].flags))
2928 else if (test_and_clear_bit(R5_Wantread, &sh->dev[i].flags))
2933 bi = &sh->dev[i].req;
2937 bi->bi_end_io = raid5_end_write_request;
2939 bi->bi_end_io = raid5_end_read_request;
2942 rdev = rcu_dereference(conf->disks[i].rdev);
2943 if (rdev && test_bit(Faulty, &rdev->flags))
2946 atomic_inc(&rdev->nr_pending);
2950 if (s.syncing || s.expanding || s.expanded)
2951 md_sync_acct(rdev->bdev, STRIPE_SECTORS);
2953 bi->bi_bdev = rdev->bdev;
2954 pr_debug("for %llu schedule op %ld on disc %d\n",
2955 (unsigned long long)sh->sector, bi->bi_rw, i);
2956 atomic_inc(&sh->count);
2957 bi->bi_sector = sh->sector + rdev->data_offset;
2958 bi->bi_flags = 1 << BIO_UPTODATE;
2960 bi->bi_max_vecs = 1;
2962 bi->bi_io_vec = &sh->dev[i].vec;
2963 bi->bi_io_vec[0].bv_len = STRIPE_SIZE;
2964 bi->bi_io_vec[0].bv_offset = 0;
2965 bi->bi_size = STRIPE_SIZE;
2968 test_bit(R5_ReWrite, &sh->dev[i].flags))
2969 atomic_add(STRIPE_SECTORS, &rdev->corrected_errors);
2970 generic_make_request(bi);
2973 set_bit(STRIPE_DEGRADED, &sh->state);
2974 pr_debug("skip op %ld on disc %d for sector %llu\n",
2975 bi->bi_rw, i, (unsigned long long)sh->sector);
2976 clear_bit(R5_LOCKED, &sh->dev[i].flags);
2977 set_bit(STRIPE_HANDLE, &sh->state);
2982 static void handle_stripe6(struct stripe_head *sh, struct page *tmp_page)
2984 raid6_conf_t *conf = sh->raid_conf;
2985 int disks = sh->disks;
2986 struct bio *return_bi = NULL;
2987 int i, pd_idx = sh->pd_idx;
2988 struct stripe_head_state s;
2989 struct r6_state r6s;
2990 struct r5dev *dev, *pdev, *qdev;
2992 r6s.qd_idx = raid6_next_disk(pd_idx, disks);
2993 pr_debug("handling stripe %llu, state=%#lx cnt=%d, "
2994 "pd_idx=%d, qd_idx=%d\n",
2995 (unsigned long long)sh->sector, sh->state,
2996 atomic_read(&sh->count), pd_idx, r6s.qd_idx);
2997 memset(&s, 0, sizeof(s));
2999 spin_lock(&sh->lock);
3000 clear_bit(STRIPE_HANDLE, &sh->state);
3001 clear_bit(STRIPE_DELAYED, &sh->state);
3003 s.syncing = test_bit(STRIPE_SYNCING, &sh->state);
3004 s.expanding = test_bit(STRIPE_EXPAND_SOURCE, &sh->state);
3005 s.expanded = test_bit(STRIPE_EXPAND_READY, &sh->state);
3006 /* Now to look around and see what can be done */
3009 for (i=disks; i--; ) {
3012 clear_bit(R5_Insync, &dev->flags);
3014 pr_debug("check %d: state 0x%lx read %p write %p written %p\n",
3015 i, dev->flags, dev->toread, dev->towrite, dev->written);
3016 /* maybe we can reply to a read */
3017 if (test_bit(R5_UPTODATE, &dev->flags) && dev->toread) {
3018 struct bio *rbi, *rbi2;
3019 pr_debug("Return read for disc %d\n", i);
3020 spin_lock_irq(&conf->device_lock);
3023 if (test_and_clear_bit(R5_Overlap, &dev->flags))
3024 wake_up(&conf->wait_for_overlap);
3025 spin_unlock_irq(&conf->device_lock);
3026 while (rbi && rbi->bi_sector < dev->sector + STRIPE_SECTORS) {
3027 copy_data(0, rbi, dev->page, dev->sector);
3028 rbi2 = r5_next_bio(rbi, dev->sector);
3029 spin_lock_irq(&conf->device_lock);
3030 if (--rbi->bi_phys_segments == 0) {
3031 rbi->bi_next = return_bi;
3034 spin_unlock_irq(&conf->device_lock);
3039 /* now count some things */
3040 if (test_bit(R5_LOCKED, &dev->flags)) s.locked++;
3041 if (test_bit(R5_UPTODATE, &dev->flags)) s.uptodate++;
3048 if (!test_bit(R5_OVERWRITE, &dev->flags))
3053 rdev = rcu_dereference(conf->disks[i].rdev);
3054 if (!rdev || !test_bit(In_sync, &rdev->flags)) {
3055 /* The ReadError flag will just be confusing now */
3056 clear_bit(R5_ReadError, &dev->flags);
3057 clear_bit(R5_ReWrite, &dev->flags);
3059 if (!rdev || !test_bit(In_sync, &rdev->flags)
3060 || test_bit(R5_ReadError, &dev->flags)) {
3062 r6s.failed_num[s.failed] = i;
3065 set_bit(R5_Insync, &dev->flags);
3068 pr_debug("locked=%d uptodate=%d to_read=%d"
3069 " to_write=%d failed=%d failed_num=%d,%d\n",
3070 s.locked, s.uptodate, s.to_read, s.to_write, s.failed,
3071 r6s.failed_num[0], r6s.failed_num[1]);
3072 /* check if the array has lost >2 devices and, if so, some requests
3073 * might need to be failed
3075 if (s.failed > 2 && s.to_read+s.to_write+s.written)
3076 handle_requests_to_failed_array(conf, sh, &s, disks,
3078 if (s.failed > 2 && s.syncing) {
3079 md_done_sync(conf->mddev, STRIPE_SECTORS,0);
3080 clear_bit(STRIPE_SYNCING, &sh->state);
3085 * might be able to return some write requests if the parity blocks
3086 * are safe, or on a failed drive
3088 pdev = &sh->dev[pd_idx];
3089 r6s.p_failed = (s.failed >= 1 && r6s.failed_num[0] == pd_idx)
3090 || (s.failed >= 2 && r6s.failed_num[1] == pd_idx);
3091 qdev = &sh->dev[r6s.qd_idx];
3092 r6s.q_failed = (s.failed >= 1 && r6s.failed_num[0] == r6s.qd_idx)
3093 || (s.failed >= 2 && r6s.failed_num[1] == r6s.qd_idx);
3096 ( r6s.p_failed || ((test_bit(R5_Insync, &pdev->flags)
3097 && !test_bit(R5_LOCKED, &pdev->flags)
3098 && test_bit(R5_UPTODATE, &pdev->flags)))) &&
3099 ( r6s.q_failed || ((test_bit(R5_Insync, &qdev->flags)
3100 && !test_bit(R5_LOCKED, &qdev->flags)
3101 && test_bit(R5_UPTODATE, &qdev->flags)))))
3102 handle_completed_write_requests(conf, sh, disks, &return_bi);
3104 /* Now we might consider reading some blocks, either to check/generate
3105 * parity, or to satisfy requests
3106 * or to load a block that is being partially written.
3108 if (s.to_read || s.non_overwrite || (s.to_write && s.failed) ||
3109 (s.syncing && (s.uptodate < disks)) || s.expanding)
3110 handle_issuing_new_read_requests6(sh, &s, &r6s, disks);
3112 /* now to consider writing and what else, if anything should be read */
3114 handle_issuing_new_write_requests6(conf, sh, &s, &r6s, disks);
3116 /* maybe we need to check and possibly fix the parity for this stripe
3117 * Any reads will already have been scheduled, so we just see if enough
3120 if (s.syncing && s.locked == 0 && !test_bit(STRIPE_INSYNC, &sh->state))
3121 handle_parity_checks6(conf, sh, &s, &r6s, tmp_page, disks);
3123 if (s.syncing && s.locked == 0 && test_bit(STRIPE_INSYNC, &sh->state)) {
3124 md_done_sync(conf->mddev, STRIPE_SECTORS,1);
3125 clear_bit(STRIPE_SYNCING, &sh->state);
3128 /* If the failed drives are just a ReadError, then we might need
3129 * to progress the repair/check process
3131 if (s.failed <= 2 && !conf->mddev->ro)
3132 for (i = 0; i < s.failed; i++) {
3133 dev = &sh->dev[r6s.failed_num[i]];
3134 if (test_bit(R5_ReadError, &dev->flags)
3135 && !test_bit(R5_LOCKED, &dev->flags)
3136 && test_bit(R5_UPTODATE, &dev->flags)
3138 if (!test_bit(R5_ReWrite, &dev->flags)) {
3139 set_bit(R5_Wantwrite, &dev->flags);
3140 set_bit(R5_ReWrite, &dev->flags);
3141 set_bit(R5_LOCKED, &dev->flags);
3143 /* let's read it back */
3144 set_bit(R5_Wantread, &dev->flags);
3145 set_bit(R5_LOCKED, &dev->flags);
3150 if (s.expanded && test_bit(STRIPE_EXPANDING, &sh->state)) {
3151 /* Need to write out all blocks after computing P&Q */
3152 sh->disks = conf->raid_disks;
3153 sh->pd_idx = stripe_to_pdidx(sh->sector, conf,
3155 compute_parity6(sh, RECONSTRUCT_WRITE);
3156 for (i = conf->raid_disks ; i-- ; ) {
3157 set_bit(R5_LOCKED, &sh->dev[i].flags);
3159 set_bit(R5_Wantwrite, &sh->dev[i].flags);
3161 clear_bit(STRIPE_EXPANDING, &sh->state);
3162 } else if (s.expanded) {
3163 clear_bit(STRIPE_EXPAND_READY, &sh->state);
3164 atomic_dec(&conf->reshape_stripes);
3165 wake_up(&conf->wait_for_overlap);
3166 md_done_sync(conf->mddev, STRIPE_SECTORS, 1);
3169 if (s.expanding && s.locked == 0)
3170 handle_stripe_expansion(conf, sh, &r6s);
3172 spin_unlock(&sh->lock);
3174 return_io(return_bi);
3176 for (i=disks; i-- ;) {
3180 if (test_and_clear_bit(R5_Wantwrite, &sh->dev[i].flags))
3182 else if (test_and_clear_bit(R5_Wantread, &sh->dev[i].flags))
3187 bi = &sh->dev[i].req;
3191 bi->bi_end_io = raid5_end_write_request;
3193 bi->bi_end_io = raid5_end_read_request;
3196 rdev = rcu_dereference(conf->disks[i].rdev);
3197 if (rdev && test_bit(Faulty, &rdev->flags))
3200 atomic_inc(&rdev->nr_pending);
3204 if (s.syncing || s.expanding || s.expanded)
3205 md_sync_acct(rdev->bdev, STRIPE_SECTORS);
3207 bi->bi_bdev = rdev->bdev;
3208 pr_debug("for %llu schedule op %ld on disc %d\n",
3209 (unsigned long long)sh->sector, bi->bi_rw, i);
3210 atomic_inc(&sh->count);
3211 bi->bi_sector = sh->sector + rdev->data_offset;
3212 bi->bi_flags = 1 << BIO_UPTODATE;
3214 bi->bi_max_vecs = 1;
3216 bi->bi_io_vec = &sh->dev[i].vec;
3217 bi->bi_io_vec[0].bv_len = STRIPE_SIZE;
3218 bi->bi_io_vec[0].bv_offset = 0;
3219 bi->bi_size = STRIPE_SIZE;
3222 test_bit(R5_ReWrite, &sh->dev[i].flags))
3223 atomic_add(STRIPE_SECTORS, &rdev->corrected_errors);
3224 generic_make_request(bi);
3227 set_bit(STRIPE_DEGRADED, &sh->state);
3228 pr_debug("skip op %ld on disc %d for sector %llu\n",
3229 bi->bi_rw, i, (unsigned long long)sh->sector);
3230 clear_bit(R5_LOCKED, &sh->dev[i].flags);
3231 set_bit(STRIPE_HANDLE, &sh->state);
3236 static void handle_stripe(struct stripe_head *sh, struct page *tmp_page)
3238 if (sh->raid_conf->level == 6)
3239 handle_stripe6(sh, tmp_page);
3246 static void raid5_activate_delayed(raid5_conf_t *conf)
3248 if (atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD) {
3249 while (!list_empty(&conf->delayed_list)) {
3250 struct list_head *l = conf->delayed_list.next;
3251 struct stripe_head *sh;
3252 sh = list_entry(l, struct stripe_head, lru);
3254 clear_bit(STRIPE_DELAYED, &sh->state);
3255 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
3256 atomic_inc(&conf->preread_active_stripes);
3257 list_add_tail(&sh->lru, &conf->handle_list);
3262 static void activate_bit_delay(raid5_conf_t *conf)
3264 /* device_lock is held */
3265 struct list_head head;
3266 list_add(&head, &conf->bitmap_list);
3267 list_del_init(&conf->bitmap_list);
3268 while (!list_empty(&head)) {
3269 struct stripe_head *sh = list_entry(head.next, struct stripe_head, lru);
3270 list_del_init(&sh->lru);
3271 atomic_inc(&sh->count);
3272 __release_stripe(conf, sh);
3276 static void unplug_slaves(mddev_t *mddev)
3278 raid5_conf_t *conf = mddev_to_conf(mddev);
3282 for (i=0; i<mddev->raid_disks; i++) {
3283 mdk_rdev_t *rdev = rcu_dereference(conf->disks[i].rdev);
3284 if (rdev && !test_bit(Faulty, &rdev->flags) && atomic_read(&rdev->nr_pending)) {
3285 request_queue_t *r_queue = bdev_get_queue(rdev->bdev);
3287 atomic_inc(&rdev->nr_pending);
3290 if (r_queue->unplug_fn)
3291 r_queue->unplug_fn(r_queue);
3293 rdev_dec_pending(rdev, mddev);
3300 static void raid5_unplug_device(request_queue_t *q)
3302 mddev_t *mddev = q->queuedata;
3303 raid5_conf_t *conf = mddev_to_conf(mddev);
3304 unsigned long flags;
3306 spin_lock_irqsave(&conf->device_lock, flags);
3308 if (blk_remove_plug(q)) {
3310 raid5_activate_delayed(conf);
3312 md_wakeup_thread(mddev->thread);
3314 spin_unlock_irqrestore(&conf->device_lock, flags);
3316 unplug_slaves(mddev);
3319 static int raid5_issue_flush(request_queue_t *q, struct gendisk *disk,
3320 sector_t *error_sector)
3322 mddev_t *mddev = q->queuedata;
3323 raid5_conf_t *conf = mddev_to_conf(mddev);
3327 for (i=0; i<mddev->raid_disks && ret == 0; i++) {
3328 mdk_rdev_t *rdev = rcu_dereference(conf->disks[i].rdev);
3329 if (rdev && !test_bit(Faulty, &rdev->flags)) {
3330 struct block_device *bdev = rdev->bdev;
3331 request_queue_t *r_queue = bdev_get_queue(bdev);
3333 if (!r_queue->issue_flush_fn)
3336 atomic_inc(&rdev->nr_pending);
3338 ret = r_queue->issue_flush_fn(r_queue, bdev->bd_disk,
3340 rdev_dec_pending(rdev, mddev);
3349 static int raid5_congested(void *data, int bits)
3351 mddev_t *mddev = data;
3352 raid5_conf_t *conf = mddev_to_conf(mddev);
3354 /* No difference between reads and writes. Just check
3355 * how busy the stripe_cache is
3357 if (conf->inactive_blocked)
3361 if (list_empty_careful(&conf->inactive_list))
3367 /* We want read requests to align with chunks where possible,
3368 * but write requests don't need to.
3370 static int raid5_mergeable_bvec(request_queue_t *q, struct bio *bio, struct bio_vec *biovec)
3372 mddev_t *mddev = q->queuedata;
3373 sector_t sector = bio->bi_sector + get_start_sect(bio->bi_bdev);
3375 unsigned int chunk_sectors = mddev->chunk_size >> 9;
3376 unsigned int bio_sectors = bio->bi_size >> 9;
3378 if (bio_data_dir(bio) == WRITE)
3379 return biovec->bv_len; /* always allow writes to be mergeable */
3381 max = (chunk_sectors - ((sector & (chunk_sectors - 1)) + bio_sectors)) << 9;
3382 if (max < 0) max = 0;
3383 if (max <= biovec->bv_len && bio_sectors == 0)
3384 return biovec->bv_len;
3390 static int in_chunk_boundary(mddev_t *mddev, struct bio *bio)
3392 sector_t sector = bio->bi_sector + get_start_sect(bio->bi_bdev);
3393 unsigned int chunk_sectors = mddev->chunk_size >> 9;
3394 unsigned int bio_sectors = bio->bi_size >> 9;
3396 return chunk_sectors >=
3397 ((sector & (chunk_sectors - 1)) + bio_sectors);
3401 * add bio to the retry LIFO ( in O(1) ... we are in interrupt )
3402 * later sampled by raid5d.
3404 static void add_bio_to_retry(struct bio *bi,raid5_conf_t *conf)
3406 unsigned long flags;
3408 spin_lock_irqsave(&conf->device_lock, flags);
3410 bi->bi_next = conf->retry_read_aligned_list;
3411 conf->retry_read_aligned_list = bi;
3413 spin_unlock_irqrestore(&conf->device_lock, flags);
3414 md_wakeup_thread(conf->mddev->thread);
3418 static struct bio *remove_bio_from_retry(raid5_conf_t *conf)
3422 bi = conf->retry_read_aligned;
3424 conf->retry_read_aligned = NULL;
3427 bi = conf->retry_read_aligned_list;
3429 conf->retry_read_aligned_list = bi->bi_next;
3431 bi->bi_phys_segments = 1; /* biased count of active stripes */
3432 bi->bi_hw_segments = 0; /* count of processed stripes */
3440 * The "raid5_align_endio" should check if the read succeeded and if it
3441 * did, call bio_endio on the original bio (having bio_put the new bio
3443 * If the read failed..
3445 static int raid5_align_endio(struct bio *bi, unsigned int bytes, int error)
3447 struct bio* raid_bi = bi->bi_private;
3450 int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
3457 mddev = raid_bi->bi_bdev->bd_disk->queue->queuedata;
3458 conf = mddev_to_conf(mddev);
3459 rdev = (void*)raid_bi->bi_next;
3460 raid_bi->bi_next = NULL;
3462 rdev_dec_pending(rdev, conf->mddev);
3464 if (!error && uptodate) {
3465 bio_endio(raid_bi, bytes, 0);
3466 if (atomic_dec_and_test(&conf->active_aligned_reads))
3467 wake_up(&conf->wait_for_stripe);
3472 pr_debug("raid5_align_endio : io error...handing IO for a retry\n");
3474 add_bio_to_retry(raid_bi, conf);
3478 static int bio_fits_rdev(struct bio *bi)
3480 request_queue_t *q = bdev_get_queue(bi->bi_bdev);
3482 if ((bi->bi_size>>9) > q->max_sectors)
3484 blk_recount_segments(q, bi);
3485 if (bi->bi_phys_segments > q->max_phys_segments ||
3486 bi->bi_hw_segments > q->max_hw_segments)
3489 if (q->merge_bvec_fn)
3490 /* it's too hard to apply the merge_bvec_fn at this stage,
3499 static int chunk_aligned_read(request_queue_t *q, struct bio * raid_bio)
3501 mddev_t *mddev = q->queuedata;
3502 raid5_conf_t *conf = mddev_to_conf(mddev);
3503 const unsigned int raid_disks = conf->raid_disks;
3504 const unsigned int data_disks = raid_disks - conf->max_degraded;
3505 unsigned int dd_idx, pd_idx;
3506 struct bio* align_bi;
3509 if (!in_chunk_boundary(mddev, raid_bio)) {
3510 pr_debug("chunk_aligned_read : non aligned\n");
3514 * use bio_clone to make a copy of the bio
3516 align_bi = bio_clone(raid_bio, GFP_NOIO);
3520 * set bi_end_io to a new function, and set bi_private to the
3523 align_bi->bi_end_io = raid5_align_endio;
3524 align_bi->bi_private = raid_bio;
3528 align_bi->bi_sector = raid5_compute_sector(raid_bio->bi_sector,
3536 rdev = rcu_dereference(conf->disks[dd_idx].rdev);
3537 if (rdev && test_bit(In_sync, &rdev->flags)) {
3538 atomic_inc(&rdev->nr_pending);
3540 raid_bio->bi_next = (void*)rdev;
3541 align_bi->bi_bdev = rdev->bdev;
3542 align_bi->bi_flags &= ~(1 << BIO_SEG_VALID);
3543 align_bi->bi_sector += rdev->data_offset;
3545 if (!bio_fits_rdev(align_bi)) {
3546 /* too big in some way */
3548 rdev_dec_pending(rdev, mddev);
3552 spin_lock_irq(&conf->device_lock);
3553 wait_event_lock_irq(conf->wait_for_stripe,
3555 conf->device_lock, /* nothing */);
3556 atomic_inc(&conf->active_aligned_reads);
3557 spin_unlock_irq(&conf->device_lock);
3559 generic_make_request(align_bi);
3569 static int make_request(request_queue_t *q, struct bio * bi)
3571 mddev_t *mddev = q->queuedata;
3572 raid5_conf_t *conf = mddev_to_conf(mddev);
3573 unsigned int dd_idx, pd_idx;
3574 sector_t new_sector;
3575 sector_t logical_sector, last_sector;
3576 struct stripe_head *sh;
3577 const int rw = bio_data_dir(bi);
3580 if (unlikely(bio_barrier(bi))) {
3581 bio_endio(bi, bi->bi_size, -EOPNOTSUPP);
3585 md_write_start(mddev, bi);
3587 disk_stat_inc(mddev->gendisk, ios[rw]);
3588 disk_stat_add(mddev->gendisk, sectors[rw], bio_sectors(bi));
3591 mddev->reshape_position == MaxSector &&
3592 chunk_aligned_read(q,bi))
3595 logical_sector = bi->bi_sector & ~((sector_t)STRIPE_SECTORS-1);
3596 last_sector = bi->bi_sector + (bi->bi_size>>9);
3598 bi->bi_phys_segments = 1; /* over-loaded to count active stripes */
3600 for (;logical_sector < last_sector; logical_sector += STRIPE_SECTORS) {
3602 int disks, data_disks;
3605 prepare_to_wait(&conf->wait_for_overlap, &w, TASK_UNINTERRUPTIBLE);
3606 if (likely(conf->expand_progress == MaxSector))
3607 disks = conf->raid_disks;
3609 /* spinlock is needed as expand_progress may be
3610 * 64bit on a 32bit platform, and so it might be
3611 * possible to see a half-updated value
3612 * Ofcourse expand_progress could change after
3613 * the lock is dropped, so once we get a reference
3614 * to the stripe that we think it is, we will have
3617 spin_lock_irq(&conf->device_lock);
3618 disks = conf->raid_disks;
3619 if (logical_sector >= conf->expand_progress)
3620 disks = conf->previous_raid_disks;
3622 if (logical_sector >= conf->expand_lo) {
3623 spin_unlock_irq(&conf->device_lock);
3628 spin_unlock_irq(&conf->device_lock);
3630 data_disks = disks - conf->max_degraded;
3632 new_sector = raid5_compute_sector(logical_sector, disks, data_disks,
3633 &dd_idx, &pd_idx, conf);
3634 pr_debug("raid5: make_request, sector %llu logical %llu\n",
3635 (unsigned long long)new_sector,
3636 (unsigned long long)logical_sector);
3638 sh = get_active_stripe(conf, new_sector, disks, pd_idx, (bi->bi_rw&RWA_MASK));
3640 if (unlikely(conf->expand_progress != MaxSector)) {
3641 /* expansion might have moved on while waiting for a
3642 * stripe, so we must do the range check again.
3643 * Expansion could still move past after this
3644 * test, but as we are holding a reference to
3645 * 'sh', we know that if that happens,
3646 * STRIPE_EXPANDING will get set and the expansion
3647 * won't proceed until we finish with the stripe.
3650 spin_lock_irq(&conf->device_lock);
3651 if (logical_sector < conf->expand_progress &&
3652 disks == conf->previous_raid_disks)
3653 /* mismatch, need to try again */
3655 spin_unlock_irq(&conf->device_lock);
3661 /* FIXME what if we get a false positive because these
3662 * are being updated.
3664 if (logical_sector >= mddev->suspend_lo &&
3665 logical_sector < mddev->suspend_hi) {
3671 if (test_bit(STRIPE_EXPANDING, &sh->state) ||
3672 !add_stripe_bio(sh, bi, dd_idx, (bi->bi_rw&RW_MASK))) {
3673 /* Stripe is busy expanding or
3674 * add failed due to overlap. Flush everything
3677 raid5_unplug_device(mddev->queue);
3682 finish_wait(&conf->wait_for_overlap, &w);
3683 handle_stripe(sh, NULL);
3686 /* cannot get stripe for read-ahead, just give-up */
3687 clear_bit(BIO_UPTODATE, &bi->bi_flags);
3688 finish_wait(&conf->wait_for_overlap, &w);
3693 spin_lock_irq(&conf->device_lock);
3694 remaining = --bi->bi_phys_segments;
3695 spin_unlock_irq(&conf->device_lock);
3696 if (remaining == 0) {
3697 int bytes = bi->bi_size;
3700 md_write_end(mddev);
3702 bi->bi_end_io(bi, bytes,
3703 test_bit(BIO_UPTODATE, &bi->bi_flags)
3709 static sector_t reshape_request(mddev_t *mddev, sector_t sector_nr, int *skipped)
3711 /* reshaping is quite different to recovery/resync so it is
3712 * handled quite separately ... here.
3714 * On each call to sync_request, we gather one chunk worth of
3715 * destination stripes and flag them as expanding.
3716 * Then we find all the source stripes and request reads.
3717 * As the reads complete, handle_stripe will copy the data
3718 * into the destination stripe and release that stripe.
3720 raid5_conf_t *conf = (raid5_conf_t *) mddev->private;
3721 struct stripe_head *sh;
3723 sector_t first_sector, last_sector;
3724 int raid_disks = conf->previous_raid_disks;
3725 int data_disks = raid_disks - conf->max_degraded;
3726 int new_data_disks = conf->raid_disks - conf->max_degraded;
3729 sector_t writepos, safepos, gap;
3731 if (sector_nr == 0 &&
3732 conf->expand_progress != 0) {
3733 /* restarting in the middle, skip the initial sectors */
3734 sector_nr = conf->expand_progress;
3735 sector_div(sector_nr, new_data_disks);
3740 /* we update the metadata when there is more than 3Meg
3741 * in the block range (that is rather arbitrary, should
3742 * probably be time based) or when the data about to be
3743 * copied would over-write the source of the data at
3744 * the front of the range.
3745 * i.e. one new_stripe forward from expand_progress new_maps
3746 * to after where expand_lo old_maps to
3748 writepos = conf->expand_progress +
3749 conf->chunk_size/512*(new_data_disks);
3750 sector_div(writepos, new_data_disks);
3751 safepos = conf->expand_lo;
3752 sector_div(safepos, data_disks);
3753 gap = conf->expand_progress - conf->expand_lo;
3755 if (writepos >= safepos ||
3756 gap > (new_data_disks)*3000*2 /*3Meg*/) {
3757 /* Cannot proceed until we've updated the superblock... */
3758 wait_event(conf->wait_for_overlap,
3759 atomic_read(&conf->reshape_stripes)==0);
3760 mddev->reshape_position = conf->expand_progress;
3761 set_bit(MD_CHANGE_DEVS, &mddev->flags);
3762 md_wakeup_thread(mddev->thread);
3763 wait_event(mddev->sb_wait, mddev->flags == 0 ||
3764 kthread_should_stop());
3765 spin_lock_irq(&conf->device_lock);
3766 conf->expand_lo = mddev->reshape_position;
3767 spin_unlock_irq(&conf->device_lock);
3768 wake_up(&conf->wait_for_overlap);
3771 for (i=0; i < conf->chunk_size/512; i+= STRIPE_SECTORS) {
3774 pd_idx = stripe_to_pdidx(sector_nr+i, conf, conf->raid_disks);
3775 sh = get_active_stripe(conf, sector_nr+i,
3776 conf->raid_disks, pd_idx, 0);
3777 set_bit(STRIPE_EXPANDING, &sh->state);
3778 atomic_inc(&conf->reshape_stripes);
3779 /* If any of this stripe is beyond the end of the old
3780 * array, then we need to zero those blocks
3782 for (j=sh->disks; j--;) {
3784 if (j == sh->pd_idx)
3786 if (conf->level == 6 &&
3787 j == raid6_next_disk(sh->pd_idx, sh->disks))
3789 s = compute_blocknr(sh, j);
3790 if (s < (mddev->array_size<<1)) {
3794 memset(page_address(sh->dev[j].page), 0, STRIPE_SIZE);
3795 set_bit(R5_Expanded, &sh->dev[j].flags);
3796 set_bit(R5_UPTODATE, &sh->dev[j].flags);
3799 set_bit(STRIPE_EXPAND_READY, &sh->state);
3800 set_bit(STRIPE_HANDLE, &sh->state);
3804 spin_lock_irq(&conf->device_lock);
3805 conf->expand_progress = (sector_nr + i) * new_data_disks;
3806 spin_unlock_irq(&conf->device_lock);
3807 /* Ok, those stripe are ready. We can start scheduling
3808 * reads on the source stripes.
3809 * The source stripes are determined by mapping the first and last
3810 * block on the destination stripes.
3813 raid5_compute_sector(sector_nr*(new_data_disks),
3814 raid_disks, data_disks,
3815 &dd_idx, &pd_idx, conf);
3817 raid5_compute_sector((sector_nr+conf->chunk_size/512)
3818 *(new_data_disks) -1,
3819 raid_disks, data_disks,
3820 &dd_idx, &pd_idx, conf);
3821 if (last_sector >= (mddev->size<<1))
3822 last_sector = (mddev->size<<1)-1;
3823 while (first_sector <= last_sector) {
3824 pd_idx = stripe_to_pdidx(first_sector, conf,
3825 conf->previous_raid_disks);
3826 sh = get_active_stripe(conf, first_sector,
3827 conf->previous_raid_disks, pd_idx, 0);
3828 set_bit(STRIPE_EXPAND_SOURCE, &sh->state);
3829 set_bit(STRIPE_HANDLE, &sh->state);
3831 first_sector += STRIPE_SECTORS;
3833 return conf->chunk_size>>9;
3836 /* FIXME go_faster isn't used */
3837 static inline sector_t sync_request(mddev_t *mddev, sector_t sector_nr, int *skipped, int go_faster)
3839 raid5_conf_t *conf = (raid5_conf_t *) mddev->private;
3840 struct stripe_head *sh;
3842 int raid_disks = conf->raid_disks;
3843 sector_t max_sector = mddev->size << 1;
3845 int still_degraded = 0;
3848 if (sector_nr >= max_sector) {
3849 /* just being told to finish up .. nothing much to do */
3850 unplug_slaves(mddev);
3851 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery)) {
3856 if (mddev->curr_resync < max_sector) /* aborted */
3857 bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
3859 else /* completed sync */
3861 bitmap_close_sync(mddev->bitmap);
3866 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
3867 return reshape_request(mddev, sector_nr, skipped);
3869 /* if there is too many failed drives and we are trying
3870 * to resync, then assert that we are finished, because there is
3871 * nothing we can do.
3873 if (mddev->degraded >= conf->max_degraded &&
3874 test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
3875 sector_t rv = (mddev->size << 1) - sector_nr;
3879 if (!bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
3880 !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
3881 !conf->fullsync && sync_blocks >= STRIPE_SECTORS) {
3882 /* we can skip this block, and probably more */
3883 sync_blocks /= STRIPE_SECTORS;
3885 return sync_blocks * STRIPE_SECTORS; /* keep things rounded to whole stripes */
3888 pd_idx = stripe_to_pdidx(sector_nr, conf, raid_disks);
3889 sh = get_active_stripe(conf, sector_nr, raid_disks, pd_idx, 1);
3891 sh = get_active_stripe(conf, sector_nr, raid_disks, pd_idx, 0);
3892 /* make sure we don't swamp the stripe cache if someone else
3893 * is trying to get access
3895 schedule_timeout_uninterruptible(1);
3897 /* Need to check if array will still be degraded after recovery/resync
3898 * We don't need to check the 'failed' flag as when that gets set,
3901 for (i=0; i<mddev->raid_disks; i++)
3902 if (conf->disks[i].rdev == NULL)
3905 bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, still_degraded);
3907 spin_lock(&sh->lock);
3908 set_bit(STRIPE_SYNCING, &sh->state);
3909 clear_bit(STRIPE_INSYNC, &sh->state);
3910 spin_unlock(&sh->lock);
3912 handle_stripe(sh, NULL);
3915 return STRIPE_SECTORS;
3918 static int retry_aligned_read(raid5_conf_t *conf, struct bio *raid_bio)
3920 /* We may not be able to submit a whole bio at once as there
3921 * may not be enough stripe_heads available.
3922 * We cannot pre-allocate enough stripe_heads as we may need
3923 * more than exist in the cache (if we allow ever large chunks).
3924 * So we do one stripe head at a time and record in
3925 * ->bi_hw_segments how many have been done.
3927 * We *know* that this entire raid_bio is in one chunk, so
3928 * it will be only one 'dd_idx' and only need one call to raid5_compute_sector.
3930 struct stripe_head *sh;
3932 sector_t sector, logical_sector, last_sector;
3937 logical_sector = raid_bio->bi_sector & ~((sector_t)STRIPE_SECTORS-1);
3938 sector = raid5_compute_sector( logical_sector,
3940 conf->raid_disks - conf->max_degraded,
3944 last_sector = raid_bio->bi_sector + (raid_bio->bi_size>>9);
3946 for (; logical_sector < last_sector;
3947 logical_sector += STRIPE_SECTORS,
3948 sector += STRIPE_SECTORS,
3951 if (scnt < raid_bio->bi_hw_segments)
3952 /* already done this stripe */
3955 sh = get_active_stripe(conf, sector, conf->raid_disks, pd_idx, 1);
3958 /* failed to get a stripe - must wait */
3959 raid_bio->bi_hw_segments = scnt;
3960 conf->retry_read_aligned = raid_bio;
3964 set_bit(R5_ReadError, &sh->dev[dd_idx].flags);
3965 if (!add_stripe_bio(sh, raid_bio, dd_idx, 0)) {
3967 raid_bio->bi_hw_segments = scnt;
3968 conf->retry_read_aligned = raid_bio;
3972 handle_stripe(sh, NULL);
3976 spin_lock_irq(&conf->device_lock);
3977 remaining = --raid_bio->bi_phys_segments;
3978 spin_unlock_irq(&conf->device_lock);
3979 if (remaining == 0) {
3980 int bytes = raid_bio->bi_size;
3982 raid_bio->bi_size = 0;
3983 raid_bio->bi_end_io(raid_bio, bytes,
3984 test_bit(BIO_UPTODATE, &raid_bio->bi_flags)
3987 if (atomic_dec_and_test(&conf->active_aligned_reads))
3988 wake_up(&conf->wait_for_stripe);
3995 * This is our raid5 kernel thread.
3997 * We scan the hash table for stripes which can be handled now.
3998 * During the scan, completed stripes are saved for us by the interrupt
3999 * handler, so that they will not have to wait for our next wakeup.
4001 static void raid5d (mddev_t *mddev)
4003 struct stripe_head *sh;
4004 raid5_conf_t *conf = mddev_to_conf(mddev);
4007 pr_debug("+++ raid5d active\n");
4009 md_check_recovery(mddev);
4012 spin_lock_irq(&conf->device_lock);
4014 struct list_head *first;
4017 if (conf->seq_flush != conf->seq_write) {
4018 int seq = conf->seq_flush;
4019 spin_unlock_irq(&conf->device_lock);
4020 bitmap_unplug(mddev->bitmap);
4021 spin_lock_irq(&conf->device_lock);
4022 conf->seq_write = seq;
4023 activate_bit_delay(conf);
4026 if (list_empty(&conf->handle_list) &&
4027 atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD &&
4028 !blk_queue_plugged(mddev->queue) &&
4029 !list_empty(&conf->delayed_list))
4030 raid5_activate_delayed(conf);
4032 while ((bio = remove_bio_from_retry(conf))) {
4034 spin_unlock_irq(&conf->device_lock);
4035 ok = retry_aligned_read(conf, bio);
4036 spin_lock_irq(&conf->device_lock);
4042 if (list_empty(&conf->handle_list)) {
4043 async_tx_issue_pending_all();
4047 first = conf->handle_list.next;
4048 sh = list_entry(first, struct stripe_head, lru);
4050 list_del_init(first);
4051 atomic_inc(&sh->count);
4052 BUG_ON(atomic_read(&sh->count)!= 1);
4053 spin_unlock_irq(&conf->device_lock);
4056 handle_stripe(sh, conf->spare_page);
4059 spin_lock_irq(&conf->device_lock);
4061 pr_debug("%d stripes handled\n", handled);
4063 spin_unlock_irq(&conf->device_lock);
4065 unplug_slaves(mddev);
4067 pr_debug("--- raid5d inactive\n");
4071 raid5_show_stripe_cache_size(mddev_t *mddev, char *page)
4073 raid5_conf_t *conf = mddev_to_conf(mddev);
4075 return sprintf(page, "%d\n", conf->max_nr_stripes);
4081 raid5_store_stripe_cache_size(mddev_t *mddev, const char *page, size_t len)
4083 raid5_conf_t *conf = mddev_to_conf(mddev);
4086 if (len >= PAGE_SIZE)
4091 new = simple_strtoul(page, &end, 10);
4092 if (!*page || (*end && *end != '\n') )
4094 if (new <= 16 || new > 32768)
4096 while (new < conf->max_nr_stripes) {
4097 if (drop_one_stripe(conf))
4098 conf->max_nr_stripes--;
4102 md_allow_write(mddev);
4103 while (new > conf->max_nr_stripes) {
4104 if (grow_one_stripe(conf))
4105 conf->max_nr_stripes++;
4111 static struct md_sysfs_entry
4112 raid5_stripecache_size = __ATTR(stripe_cache_size, S_IRUGO | S_IWUSR,
4113 raid5_show_stripe_cache_size,
4114 raid5_store_stripe_cache_size);
4117 stripe_cache_active_show(mddev_t *mddev, char *page)
4119 raid5_conf_t *conf = mddev_to_conf(mddev);
4121 return sprintf(page, "%d\n", atomic_read(&conf->active_stripes));
4126 static struct md_sysfs_entry
4127 raid5_stripecache_active = __ATTR_RO(stripe_cache_active);
4129 static struct attribute *raid5_attrs[] = {
4130 &raid5_stripecache_size.attr,
4131 &raid5_stripecache_active.attr,
4134 static struct attribute_group raid5_attrs_group = {
4136 .attrs = raid5_attrs,
4139 static int run(mddev_t *mddev)
4142 int raid_disk, memory;
4144 struct disk_info *disk;
4145 struct list_head *tmp;
4146 int working_disks = 0;
4148 if (mddev->level != 5 && mddev->level != 4 && mddev->level != 6) {
4149 printk(KERN_ERR "raid5: %s: raid level not set to 4/5/6 (%d)\n",
4150 mdname(mddev), mddev->level);
4154 if (mddev->reshape_position != MaxSector) {
4155 /* Check that we can continue the reshape.
4156 * Currently only disks can change, it must
4157 * increase, and we must be past the point where
4158 * a stripe over-writes itself
4160 sector_t here_new, here_old;
4162 int max_degraded = (mddev->level == 5 ? 1 : 2);
4164 if (mddev->new_level != mddev->level ||
4165 mddev->new_layout != mddev->layout ||
4166 mddev->new_chunk != mddev->chunk_size) {
4167 printk(KERN_ERR "raid5: %s: unsupported reshape "
4168 "required - aborting.\n",
4172 if (mddev->delta_disks <= 0) {
4173 printk(KERN_ERR "raid5: %s: unsupported reshape "
4174 "(reduce disks) required - aborting.\n",
4178 old_disks = mddev->raid_disks - mddev->delta_disks;
4179 /* reshape_position must be on a new-stripe boundary, and one
4180 * further up in new geometry must map after here in old
4183 here_new = mddev->reshape_position;
4184 if (sector_div(here_new, (mddev->chunk_size>>9)*
4185 (mddev->raid_disks - max_degraded))) {
4186 printk(KERN_ERR "raid5: reshape_position not "
4187 "on a stripe boundary\n");
4190 /* here_new is the stripe we will write to */
4191 here_old = mddev->reshape_position;
4192 sector_div(here_old, (mddev->chunk_size>>9)*
4193 (old_disks-max_degraded));
4194 /* here_old is the first stripe that we might need to read
4196 if (here_new >= here_old) {
4197 /* Reading from the same stripe as writing to - bad */
4198 printk(KERN_ERR "raid5: reshape_position too early for "
4199 "auto-recovery - aborting.\n");
4202 printk(KERN_INFO "raid5: reshape will continue\n");
4203 /* OK, we should be able to continue; */
4207 mddev->private = kzalloc(sizeof (raid5_conf_t), GFP_KERNEL);
4208 if ((conf = mddev->private) == NULL)
4210 if (mddev->reshape_position == MaxSector) {
4211 conf->previous_raid_disks = conf->raid_disks = mddev->raid_disks;
4213 conf->raid_disks = mddev->raid_disks;
4214 conf->previous_raid_disks = mddev->raid_disks - mddev->delta_disks;
4217 conf->disks = kzalloc(conf->raid_disks * sizeof(struct disk_info),
4222 conf->mddev = mddev;
4224 if ((conf->stripe_hashtbl = kzalloc(PAGE_SIZE, GFP_KERNEL)) == NULL)
4227 if (mddev->level == 6) {
4228 conf->spare_page = alloc_page(GFP_KERNEL);
4229 if (!conf->spare_page)
4232 spin_lock_init(&conf->device_lock);
4233 init_waitqueue_head(&conf->wait_for_stripe);
4234 init_waitqueue_head(&conf->wait_for_overlap);
4235 INIT_LIST_HEAD(&conf->handle_list);
4236 INIT_LIST_HEAD(&conf->delayed_list);
4237 INIT_LIST_HEAD(&conf->bitmap_list);
4238 INIT_LIST_HEAD(&conf->inactive_list);
4239 atomic_set(&conf->active_stripes, 0);
4240 atomic_set(&conf->preread_active_stripes, 0);
4241 atomic_set(&conf->active_aligned_reads, 0);
4243 pr_debug("raid5: run(%s) called.\n", mdname(mddev));
4245 ITERATE_RDEV(mddev,rdev,tmp) {
4246 raid_disk = rdev->raid_disk;
4247 if (raid_disk >= conf->raid_disks
4250 disk = conf->disks + raid_disk;
4254 if (test_bit(In_sync, &rdev->flags)) {
4255 char b[BDEVNAME_SIZE];
4256 printk(KERN_INFO "raid5: device %s operational as raid"
4257 " disk %d\n", bdevname(rdev->bdev,b),
4264 * 0 for a fully functional array, 1 or 2 for a degraded array.
4266 mddev->degraded = conf->raid_disks - working_disks;
4267 conf->mddev = mddev;
4268 conf->chunk_size = mddev->chunk_size;
4269 conf->level = mddev->level;
4270 if (conf->level == 6)
4271 conf->max_degraded = 2;
4273 conf->max_degraded = 1;
4274 conf->algorithm = mddev->layout;
4275 conf->max_nr_stripes = NR_STRIPES;
4276 conf->expand_progress = mddev->reshape_position;
4278 /* device size must be a multiple of chunk size */
4279 mddev->size &= ~(mddev->chunk_size/1024 -1);
4280 mddev->resync_max_sectors = mddev->size << 1;
4282 if (conf->level == 6 && conf->raid_disks < 4) {
4283 printk(KERN_ERR "raid6: not enough configured devices for %s (%d, minimum 4)\n",
4284 mdname(mddev), conf->raid_disks);
4287 if (!conf->chunk_size || conf->chunk_size % 4) {
4288 printk(KERN_ERR "raid5: invalid chunk size %d for %s\n",
4289 conf->chunk_size, mdname(mddev));
4292 if (conf->algorithm > ALGORITHM_RIGHT_SYMMETRIC) {
4294 "raid5: unsupported parity algorithm %d for %s\n",
4295 conf->algorithm, mdname(mddev));
4298 if (mddev->degraded > conf->max_degraded) {
4299 printk(KERN_ERR "raid5: not enough operational devices for %s"
4300 " (%d/%d failed)\n",
4301 mdname(mddev), mddev->degraded, conf->raid_disks);
4305 if (mddev->degraded > 0 &&
4306 mddev->recovery_cp != MaxSector) {
4307 if (mddev->ok_start_degraded)
4309 "raid5: starting dirty degraded array: %s"
4310 "- data corruption possible.\n",
4314 "raid5: cannot start dirty degraded array for %s\n",
4321 mddev->thread = md_register_thread(raid5d, mddev, "%s_raid5");
4322 if (!mddev->thread) {
4324 "raid5: couldn't allocate thread for %s\n",
4329 memory = conf->max_nr_stripes * (sizeof(struct stripe_head) +
4330 conf->raid_disks * ((sizeof(struct bio) + PAGE_SIZE))) / 1024;
4331 if (grow_stripes(conf, conf->max_nr_stripes)) {
4333 "raid5: couldn't allocate %dkB for buffers\n", memory);
4334 shrink_stripes(conf);
4335 md_unregister_thread(mddev->thread);
4338 printk(KERN_INFO "raid5: allocated %dkB for %s\n",
4339 memory, mdname(mddev));
4341 if (mddev->degraded == 0)
4342 printk("raid5: raid level %d set %s active with %d out of %d"
4343 " devices, algorithm %d\n", conf->level, mdname(mddev),
4344 mddev->raid_disks-mddev->degraded, mddev->raid_disks,
4347 printk(KERN_ALERT "raid5: raid level %d set %s active with %d"
4348 " out of %d devices, algorithm %d\n", conf->level,
4349 mdname(mddev), mddev->raid_disks - mddev->degraded,
4350 mddev->raid_disks, conf->algorithm);
4352 print_raid5_conf(conf);
4354 if (conf->expand_progress != MaxSector) {
4355 printk("...ok start reshape thread\n");
4356 conf->expand_lo = conf->expand_progress;
4357 atomic_set(&conf->reshape_stripes, 0);
4358 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
4359 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
4360 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
4361 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
4362 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
4366 /* read-ahead size must cover two whole stripes, which is
4367 * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
4370 int data_disks = conf->previous_raid_disks - conf->max_degraded;
4371 int stripe = data_disks *
4372 (mddev->chunk_size / PAGE_SIZE);
4373 if (mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
4374 mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
4377 /* Ok, everything is just fine now */
4378 if (sysfs_create_group(&mddev->kobj, &raid5_attrs_group))
4380 "raid5: failed to create sysfs attributes for %s\n",
4383 mddev->queue->unplug_fn = raid5_unplug_device;
4384 mddev->queue->issue_flush_fn = raid5_issue_flush;
4385 mddev->queue->backing_dev_info.congested_data = mddev;
4386 mddev->queue->backing_dev_info.congested_fn = raid5_congested;
4388 mddev->array_size = mddev->size * (conf->previous_raid_disks -
4389 conf->max_degraded);
4391 blk_queue_merge_bvec(mddev->queue, raid5_mergeable_bvec);
4396 print_raid5_conf(conf);
4397 safe_put_page(conf->spare_page);
4399 kfree(conf->stripe_hashtbl);
4402 mddev->private = NULL;
4403 printk(KERN_ALERT "raid5: failed to run raid set %s\n", mdname(mddev));
4409 static int stop(mddev_t *mddev)
4411 raid5_conf_t *conf = (raid5_conf_t *) mddev->private;
4413 md_unregister_thread(mddev->thread);
4414 mddev->thread = NULL;
4415 shrink_stripes(conf);
4416 kfree(conf->stripe_hashtbl);
4417 mddev->queue->backing_dev_info.congested_fn = NULL;
4418 blk_sync_queue(mddev->queue); /* the unplug fn references 'conf'*/
4419 sysfs_remove_group(&mddev->kobj, &raid5_attrs_group);
4422 mddev->private = NULL;
4427 static void print_sh (struct seq_file *seq, struct stripe_head *sh)
4431 seq_printf(seq, "sh %llu, pd_idx %d, state %ld.\n",
4432 (unsigned long long)sh->sector, sh->pd_idx, sh->state);
4433 seq_printf(seq, "sh %llu, count %d.\n",
4434 (unsigned long long)sh->sector, atomic_read(&sh->count));
4435 seq_printf(seq, "sh %llu, ", (unsigned long long)sh->sector);
4436 for (i = 0; i < sh->disks; i++) {
4437 seq_printf(seq, "(cache%d: %p %ld) ",
4438 i, sh->dev[i].page, sh->dev[i].flags);
4440 seq_printf(seq, "\n");
4443 static void printall (struct seq_file *seq, raid5_conf_t *conf)
4445 struct stripe_head *sh;
4446 struct hlist_node *hn;
4449 spin_lock_irq(&conf->device_lock);
4450 for (i = 0; i < NR_HASH; i++) {
4451 hlist_for_each_entry(sh, hn, &conf->stripe_hashtbl[i], hash) {
4452 if (sh->raid_conf != conf)
4457 spin_unlock_irq(&conf->device_lock);
4461 static void status (struct seq_file *seq, mddev_t *mddev)
4463 raid5_conf_t *conf = (raid5_conf_t *) mddev->private;
4466 seq_printf (seq, " level %d, %dk chunk, algorithm %d", mddev->level, mddev->chunk_size >> 10, mddev->layout);
4467 seq_printf (seq, " [%d/%d] [", conf->raid_disks, conf->raid_disks - mddev->degraded);
4468 for (i = 0; i < conf->raid_disks; i++)
4469 seq_printf (seq, "%s",
4470 conf->disks[i].rdev &&
4471 test_bit(In_sync, &conf->disks[i].rdev->flags) ? "U" : "_");
4472 seq_printf (seq, "]");
4474 seq_printf (seq, "\n");
4475 printall(seq, conf);
4479 static void print_raid5_conf (raid5_conf_t *conf)
4482 struct disk_info *tmp;
4484 printk("RAID5 conf printout:\n");
4486 printk("(conf==NULL)\n");
4489 printk(" --- rd:%d wd:%d\n", conf->raid_disks,
4490 conf->raid_disks - conf->mddev->degraded);
4492 for (i = 0; i < conf->raid_disks; i++) {
4493 char b[BDEVNAME_SIZE];
4494 tmp = conf->disks + i;
4496 printk(" disk %d, o:%d, dev:%s\n",
4497 i, !test_bit(Faulty, &tmp->rdev->flags),
4498 bdevname(tmp->rdev->bdev,b));
4502 static int raid5_spare_active(mddev_t *mddev)
4505 raid5_conf_t *conf = mddev->private;
4506 struct disk_info *tmp;
4508 for (i = 0; i < conf->raid_disks; i++) {
4509 tmp = conf->disks + i;
4511 && !test_bit(Faulty, &tmp->rdev->flags)
4512 && !test_and_set_bit(In_sync, &tmp->rdev->flags)) {
4513 unsigned long flags;
4514 spin_lock_irqsave(&conf->device_lock, flags);
4516 spin_unlock_irqrestore(&conf->device_lock, flags);
4519 print_raid5_conf(conf);
4523 static int raid5_remove_disk(mddev_t *mddev, int number)
4525 raid5_conf_t *conf = mddev->private;
4528 struct disk_info *p = conf->disks + number;
4530 print_raid5_conf(conf);
4533 if (test_bit(In_sync, &rdev->flags) ||
4534 atomic_read(&rdev->nr_pending)) {
4540 if (atomic_read(&rdev->nr_pending)) {
4541 /* lost the race, try later */
4548 print_raid5_conf(conf);
4552 static int raid5_add_disk(mddev_t *mddev, mdk_rdev_t *rdev)
4554 raid5_conf_t *conf = mddev->private;
4557 struct disk_info *p;
4559 if (mddev->degraded > conf->max_degraded)
4560 /* no point adding a device */
4564 * find the disk ... but prefer rdev->saved_raid_disk
4567 if (rdev->saved_raid_disk >= 0 &&
4568 conf->disks[rdev->saved_raid_disk].rdev == NULL)
4569 disk = rdev->saved_raid_disk;
4572 for ( ; disk < conf->raid_disks; disk++)
4573 if ((p=conf->disks + disk)->rdev == NULL) {
4574 clear_bit(In_sync, &rdev->flags);
4575 rdev->raid_disk = disk;
4577 if (rdev->saved_raid_disk != disk)
4579 rcu_assign_pointer(p->rdev, rdev);
4582 print_raid5_conf(conf);
4586 static int raid5_resize(mddev_t *mddev, sector_t sectors)
4588 /* no resync is happening, and there is enough space
4589 * on all devices, so we can resize.
4590 * We need to make sure resync covers any new space.
4591 * If the array is shrinking we should possibly wait until
4592 * any io in the removed space completes, but it hardly seems
4595 raid5_conf_t *conf = mddev_to_conf(mddev);
4597 sectors &= ~((sector_t)mddev->chunk_size/512 - 1);
4598 mddev->array_size = (sectors * (mddev->raid_disks-conf->max_degraded))>>1;
4599 set_capacity(mddev->gendisk, mddev->array_size << 1);
4601 if (sectors/2 > mddev->size && mddev->recovery_cp == MaxSector) {
4602 mddev->recovery_cp = mddev->size << 1;
4603 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
4605 mddev->size = sectors /2;
4606 mddev->resync_max_sectors = sectors;
4610 #ifdef CONFIG_MD_RAID5_RESHAPE
4611 static int raid5_check_reshape(mddev_t *mddev)
4613 raid5_conf_t *conf = mddev_to_conf(mddev);
4616 if (mddev->delta_disks < 0 ||
4617 mddev->new_level != mddev->level)
4618 return -EINVAL; /* Cannot shrink array or change level yet */
4619 if (mddev->delta_disks == 0)
4620 return 0; /* nothing to do */
4622 /* Can only proceed if there are plenty of stripe_heads.
4623 * We need a minimum of one full stripe,, and for sensible progress
4624 * it is best to have about 4 times that.
4625 * If we require 4 times, then the default 256 4K stripe_heads will
4626 * allow for chunk sizes up to 256K, which is probably OK.
4627 * If the chunk size is greater, user-space should request more
4628 * stripe_heads first.
4630 if ((mddev->chunk_size / STRIPE_SIZE) * 4 > conf->max_nr_stripes ||
4631 (mddev->new_chunk / STRIPE_SIZE) * 4 > conf->max_nr_stripes) {
4632 printk(KERN_WARNING "raid5: reshape: not enough stripes. Needed %lu\n",
4633 (mddev->chunk_size / STRIPE_SIZE)*4);
4637 err = resize_stripes(conf, conf->raid_disks + mddev->delta_disks);
4641 if (mddev->degraded > conf->max_degraded)
4643 /* looks like we might be able to manage this */
4647 static int raid5_start_reshape(mddev_t *mddev)
4649 raid5_conf_t *conf = mddev_to_conf(mddev);
4651 struct list_head *rtmp;
4653 int added_devices = 0;
4654 unsigned long flags;
4656 if (test_bit(MD_RECOVERY_RUNNING, &mddev->recovery))
4659 ITERATE_RDEV(mddev, rdev, rtmp)
4660 if (rdev->raid_disk < 0 &&
4661 !test_bit(Faulty, &rdev->flags))
4664 if (spares - mddev->degraded < mddev->delta_disks - conf->max_degraded)
4665 /* Not enough devices even to make a degraded array
4670 atomic_set(&conf->reshape_stripes, 0);
4671 spin_lock_irq(&conf->device_lock);
4672 conf->previous_raid_disks = conf->raid_disks;
4673 conf->raid_disks += mddev->delta_disks;
4674 conf->expand_progress = 0;
4675 conf->expand_lo = 0;
4676 spin_unlock_irq(&conf->device_lock);
4678 /* Add some new drives, as many as will fit.
4679 * We know there are enough to make the newly sized array work.
4681 ITERATE_RDEV(mddev, rdev, rtmp)
4682 if (rdev->raid_disk < 0 &&
4683 !test_bit(Faulty, &rdev->flags)) {
4684 if (raid5_add_disk(mddev, rdev)) {
4686 set_bit(In_sync, &rdev->flags);
4688 rdev->recovery_offset = 0;
4689 sprintf(nm, "rd%d", rdev->raid_disk);
4690 if (sysfs_create_link(&mddev->kobj,
4693 "raid5: failed to create "
4694 " link %s for %s\n",
4700 spin_lock_irqsave(&conf->device_lock, flags);
4701 mddev->degraded = (conf->raid_disks - conf->previous_raid_disks) - added_devices;
4702 spin_unlock_irqrestore(&conf->device_lock, flags);
4703 mddev->raid_disks = conf->raid_disks;
4704 mddev->reshape_position = 0;
4705 set_bit(MD_CHANGE_DEVS, &mddev->flags);
4707 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
4708 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
4709 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
4710 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
4711 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
4713 if (!mddev->sync_thread) {
4714 mddev->recovery = 0;
4715 spin_lock_irq(&conf->device_lock);
4716 mddev->raid_disks = conf->raid_disks = conf->previous_raid_disks;
4717 conf->expand_progress = MaxSector;
4718 spin_unlock_irq(&conf->device_lock);
4721 md_wakeup_thread(mddev->sync_thread);
4722 md_new_event(mddev);
4727 static void end_reshape(raid5_conf_t *conf)
4729 struct block_device *bdev;
4731 if (!test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery)) {
4732 conf->mddev->array_size = conf->mddev->size *
4733 (conf->raid_disks - conf->max_degraded);
4734 set_capacity(conf->mddev->gendisk, conf->mddev->array_size << 1);
4735 conf->mddev->changed = 1;
4737 bdev = bdget_disk(conf->mddev->gendisk, 0);
4739 mutex_lock(&bdev->bd_inode->i_mutex);
4740 i_size_write(bdev->bd_inode, (loff_t)conf->mddev->array_size << 10);
4741 mutex_unlock(&bdev->bd_inode->i_mutex);
4744 spin_lock_irq(&conf->device_lock);
4745 conf->expand_progress = MaxSector;
4746 spin_unlock_irq(&conf->device_lock);
4747 conf->mddev->reshape_position = MaxSector;
4749 /* read-ahead size must cover two whole stripes, which is
4750 * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
4753 int data_disks = conf->previous_raid_disks - conf->max_degraded;
4754 int stripe = data_disks *
4755 (conf->mddev->chunk_size / PAGE_SIZE);
4756 if (conf->mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
4757 conf->mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
4762 static void raid5_quiesce(mddev_t *mddev, int state)
4764 raid5_conf_t *conf = mddev_to_conf(mddev);
4767 case 2: /* resume for a suspend */
4768 wake_up(&conf->wait_for_overlap);
4771 case 1: /* stop all writes */
4772 spin_lock_irq(&conf->device_lock);
4774 wait_event_lock_irq(conf->wait_for_stripe,
4775 atomic_read(&conf->active_stripes) == 0 &&
4776 atomic_read(&conf->active_aligned_reads) == 0,
4777 conf->device_lock, /* nothing */);
4778 spin_unlock_irq(&conf->device_lock);
4781 case 0: /* re-enable writes */
4782 spin_lock_irq(&conf->device_lock);
4784 wake_up(&conf->wait_for_stripe);
4785 wake_up(&conf->wait_for_overlap);
4786 spin_unlock_irq(&conf->device_lock);
4791 static struct mdk_personality raid6_personality =
4795 .owner = THIS_MODULE,
4796 .make_request = make_request,
4800 .error_handler = error,
4801 .hot_add_disk = raid5_add_disk,
4802 .hot_remove_disk= raid5_remove_disk,
4803 .spare_active = raid5_spare_active,
4804 .sync_request = sync_request,
4805 .resize = raid5_resize,
4806 #ifdef CONFIG_MD_RAID5_RESHAPE
4807 .check_reshape = raid5_check_reshape,
4808 .start_reshape = raid5_start_reshape,
4810 .quiesce = raid5_quiesce,
4812 static struct mdk_personality raid5_personality =
4816 .owner = THIS_MODULE,
4817 .make_request = make_request,
4821 .error_handler = error,
4822 .hot_add_disk = raid5_add_disk,
4823 .hot_remove_disk= raid5_remove_disk,
4824 .spare_active = raid5_spare_active,
4825 .sync_request = sync_request,
4826 .resize = raid5_resize,
4827 #ifdef CONFIG_MD_RAID5_RESHAPE
4828 .check_reshape = raid5_check_reshape,
4829 .start_reshape = raid5_start_reshape,
4831 .quiesce = raid5_quiesce,
4834 static struct mdk_personality raid4_personality =
4838 .owner = THIS_MODULE,
4839 .make_request = make_request,
4843 .error_handler = error,
4844 .hot_add_disk = raid5_add_disk,
4845 .hot_remove_disk= raid5_remove_disk,
4846 .spare_active = raid5_spare_active,
4847 .sync_request = sync_request,
4848 .resize = raid5_resize,
4849 #ifdef CONFIG_MD_RAID5_RESHAPE
4850 .check_reshape = raid5_check_reshape,
4851 .start_reshape = raid5_start_reshape,
4853 .quiesce = raid5_quiesce,
4856 static int __init raid5_init(void)
4860 e = raid6_select_algo();
4863 register_md_personality(&raid6_personality);
4864 register_md_personality(&raid5_personality);
4865 register_md_personality(&raid4_personality);
4869 static void raid5_exit(void)
4871 unregister_md_personality(&raid6_personality);
4872 unregister_md_personality(&raid5_personality);
4873 unregister_md_personality(&raid4_personality);
4876 module_init(raid5_init);
4877 module_exit(raid5_exit);
4878 MODULE_LICENSE("GPL");
4879 MODULE_ALIAS("md-personality-4"); /* RAID5 */
4880 MODULE_ALIAS("md-raid5");
4881 MODULE_ALIAS("md-raid4");
4882 MODULE_ALIAS("md-level-5");
4883 MODULE_ALIAS("md-level-4");
4884 MODULE_ALIAS("md-personality-8"); /* RAID6 */
4885 MODULE_ALIAS("md-raid6");
4886 MODULE_ALIAS("md-level-6");
4888 /* This used to be two separate modules, they were: */
4889 MODULE_ALIAS("raid5");
4890 MODULE_ALIAS("raid6");