2 * raid5.c : Multiple Devices driver for Linux
3 * Copyright (C) 1996, 1997 Ingo Molnar, Miguel de Icaza, Gadi Oxman
4 * Copyright (C) 1999, 2000 Ingo Molnar
5 * Copyright (C) 2002, 2003 H. Peter Anvin
7 * RAID-4/5/6 management functions.
8 * Thanks to Penguin Computing for making the RAID-6 development possible
9 * by donating a test server!
11 * This program is free software; you can redistribute it and/or modify
12 * it under the terms of the GNU General Public License as published by
13 * the Free Software Foundation; either version 2, or (at your option)
16 * You should have received a copy of the GNU General Public License
17 * (for example /usr/src/linux/COPYING); if not, write to the Free
18 * Software Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
24 * The sequencing for updating the bitmap reliably is a little
25 * subtle (and I got it wrong the first time) so it deserves some
28 * We group bitmap updates into batches. Each batch has a number.
29 * We may write out several batches at once, but that isn't very important.
30 * conf->bm_write is the number of the last batch successfully written.
31 * conf->bm_flush is the number of the last batch that was closed to
33 * When we discover that we will need to write to any block in a stripe
34 * (in add_stripe_bio) we update the in-memory bitmap and record in sh->bm_seq
35 * the number of the batch it will be in. This is bm_flush+1.
36 * When we are ready to do a write, if that batch hasn't been written yet,
37 * we plug the array and queue the stripe for later.
38 * When an unplug happens, we increment bm_flush, thus closing the current
40 * When we notice that bm_flush > bm_write, we write out all pending updates
41 * to the bitmap, and advance bm_write to where bm_flush was.
42 * This may occasionally write a bit out twice, but is sure never to
46 #include <linux/module.h>
47 #include <linux/slab.h>
48 #include <linux/highmem.h>
49 #include <linux/bitops.h>
50 #include <linux/kthread.h>
51 #include <asm/atomic.h>
54 #include <linux/raid/bitmap.h>
55 #include <linux/async_tx.h>
61 #define NR_STRIPES 256
62 #define STRIPE_SIZE PAGE_SIZE
63 #define STRIPE_SHIFT (PAGE_SHIFT - 9)
64 #define STRIPE_SECTORS (STRIPE_SIZE>>9)
65 #define IO_THRESHOLD 1
66 #define BYPASS_THRESHOLD 1
67 #define NR_HASH (PAGE_SIZE / sizeof(struct hlist_head))
68 #define HASH_MASK (NR_HASH - 1)
70 #define stripe_hash(conf, sect) (&((conf)->stripe_hashtbl[((sect) >> STRIPE_SHIFT) & HASH_MASK]))
72 /* bio's attached to a stripe+device for I/O are linked together in bi_sector
73 * order without overlap. There may be several bio's per stripe+device, and
74 * a bio could span several devices.
75 * When walking this list for a particular stripe+device, we must never proceed
76 * beyond a bio that extends past this device, as the next bio might no longer
78 * This macro is used to determine the 'next' bio in the list, given the sector
79 * of the current stripe+device
81 #define r5_next_bio(bio, sect) ( ( (bio)->bi_sector + ((bio)->bi_size>>9) < sect + STRIPE_SECTORS) ? (bio)->bi_next : NULL)
83 * The following can be used to debug the driver
85 #define RAID5_PARANOIA 1
86 #if RAID5_PARANOIA && defined(CONFIG_SMP)
87 # define CHECK_DEVLOCK() assert_spin_locked(&conf->device_lock)
89 # define CHECK_DEVLOCK()
97 #define printk_rl(args...) ((void) (printk_ratelimit() && printk(args)))
99 #if !RAID6_USE_EMPTY_ZERO_PAGE
100 /* In .bss so it's zeroed */
101 const char raid6_empty_zero_page[PAGE_SIZE] __attribute__((aligned(256)));
104 static inline int raid6_next_disk(int disk, int raid_disks)
107 return (disk < raid_disks) ? disk : 0;
110 static void return_io(struct bio *return_bi)
112 struct bio *bi = return_bi;
115 return_bi = bi->bi_next;
123 static void print_raid5_conf (raid5_conf_t *conf);
125 static void __release_stripe(raid5_conf_t *conf, struct stripe_head *sh)
127 if (atomic_dec_and_test(&sh->count)) {
128 BUG_ON(!list_empty(&sh->lru));
129 BUG_ON(atomic_read(&conf->active_stripes)==0);
130 if (test_bit(STRIPE_HANDLE, &sh->state)) {
131 if (test_bit(STRIPE_DELAYED, &sh->state)) {
132 list_add_tail(&sh->lru, &conf->delayed_list);
133 blk_plug_device(conf->mddev->queue);
134 } else if (test_bit(STRIPE_BIT_DELAY, &sh->state) &&
135 sh->bm_seq - conf->seq_write > 0) {
136 list_add_tail(&sh->lru, &conf->bitmap_list);
137 blk_plug_device(conf->mddev->queue);
139 clear_bit(STRIPE_BIT_DELAY, &sh->state);
140 list_add_tail(&sh->lru, &conf->handle_list);
142 md_wakeup_thread(conf->mddev->thread);
144 BUG_ON(sh->ops.pending);
145 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
146 atomic_dec(&conf->preread_active_stripes);
147 if (atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD)
148 md_wakeup_thread(conf->mddev->thread);
150 atomic_dec(&conf->active_stripes);
151 if (!test_bit(STRIPE_EXPANDING, &sh->state)) {
152 list_add_tail(&sh->lru, &conf->inactive_list);
153 wake_up(&conf->wait_for_stripe);
154 if (conf->retry_read_aligned)
155 md_wakeup_thread(conf->mddev->thread);
160 static void release_stripe(struct stripe_head *sh)
162 raid5_conf_t *conf = sh->raid_conf;
165 spin_lock_irqsave(&conf->device_lock, flags);
166 __release_stripe(conf, sh);
167 spin_unlock_irqrestore(&conf->device_lock, flags);
170 static inline void remove_hash(struct stripe_head *sh)
172 pr_debug("remove_hash(), stripe %llu\n",
173 (unsigned long long)sh->sector);
175 hlist_del_init(&sh->hash);
178 static inline void insert_hash(raid5_conf_t *conf, struct stripe_head *sh)
180 struct hlist_head *hp = stripe_hash(conf, sh->sector);
182 pr_debug("insert_hash(), stripe %llu\n",
183 (unsigned long long)sh->sector);
186 hlist_add_head(&sh->hash, hp);
190 /* find an idle stripe, make sure it is unhashed, and return it. */
191 static struct stripe_head *get_free_stripe(raid5_conf_t *conf)
193 struct stripe_head *sh = NULL;
194 struct list_head *first;
197 if (list_empty(&conf->inactive_list))
199 first = conf->inactive_list.next;
200 sh = list_entry(first, struct stripe_head, lru);
201 list_del_init(first);
203 atomic_inc(&conf->active_stripes);
208 static void shrink_buffers(struct stripe_head *sh, int num)
213 for (i=0; i<num ; i++) {
217 sh->dev[i].page = NULL;
222 static int grow_buffers(struct stripe_head *sh, int num)
226 for (i=0; i<num; i++) {
229 if (!(page = alloc_page(GFP_KERNEL))) {
232 sh->dev[i].page = page;
237 static void raid5_build_block (struct stripe_head *sh, int i);
239 static void init_stripe(struct stripe_head *sh, sector_t sector, int pd_idx, int disks)
241 raid5_conf_t *conf = sh->raid_conf;
244 BUG_ON(atomic_read(&sh->count) != 0);
245 BUG_ON(test_bit(STRIPE_HANDLE, &sh->state));
246 BUG_ON(sh->ops.pending || sh->ops.ack || sh->ops.complete);
249 pr_debug("init_stripe called, stripe %llu\n",
250 (unsigned long long)sh->sector);
260 for (i = sh->disks; i--; ) {
261 struct r5dev *dev = &sh->dev[i];
263 if (dev->toread || dev->read || dev->towrite || dev->written ||
264 test_bit(R5_LOCKED, &dev->flags)) {
265 printk(KERN_ERR "sector=%llx i=%d %p %p %p %p %d\n",
266 (unsigned long long)sh->sector, i, dev->toread,
267 dev->read, dev->towrite, dev->written,
268 test_bit(R5_LOCKED, &dev->flags));
272 raid5_build_block(sh, i);
274 insert_hash(conf, sh);
277 static struct stripe_head *__find_stripe(raid5_conf_t *conf, sector_t sector, int disks)
279 struct stripe_head *sh;
280 struct hlist_node *hn;
283 pr_debug("__find_stripe, sector %llu\n", (unsigned long long)sector);
284 hlist_for_each_entry(sh, hn, stripe_hash(conf, sector), hash)
285 if (sh->sector == sector && sh->disks == disks)
287 pr_debug("__stripe %llu not in cache\n", (unsigned long long)sector);
291 static void unplug_slaves(mddev_t *mddev);
292 static void raid5_unplug_device(struct request_queue *q);
294 static struct stripe_head *get_active_stripe(raid5_conf_t *conf, sector_t sector, int disks,
295 int pd_idx, int noblock)
297 struct stripe_head *sh;
299 pr_debug("get_stripe, sector %llu\n", (unsigned long long)sector);
301 spin_lock_irq(&conf->device_lock);
304 wait_event_lock_irq(conf->wait_for_stripe,
306 conf->device_lock, /* nothing */);
307 sh = __find_stripe(conf, sector, disks);
309 if (!conf->inactive_blocked)
310 sh = get_free_stripe(conf);
311 if (noblock && sh == NULL)
314 conf->inactive_blocked = 1;
315 wait_event_lock_irq(conf->wait_for_stripe,
316 !list_empty(&conf->inactive_list) &&
317 (atomic_read(&conf->active_stripes)
318 < (conf->max_nr_stripes *3/4)
319 || !conf->inactive_blocked),
321 raid5_unplug_device(conf->mddev->queue)
323 conf->inactive_blocked = 0;
325 init_stripe(sh, sector, pd_idx, disks);
327 if (atomic_read(&sh->count)) {
328 BUG_ON(!list_empty(&sh->lru));
330 if (!test_bit(STRIPE_HANDLE, &sh->state))
331 atomic_inc(&conf->active_stripes);
332 if (list_empty(&sh->lru) &&
333 !test_bit(STRIPE_EXPANDING, &sh->state))
335 list_del_init(&sh->lru);
338 } while (sh == NULL);
341 atomic_inc(&sh->count);
343 spin_unlock_irq(&conf->device_lock);
347 /* test_and_ack_op() ensures that we only dequeue an operation once */
348 #define test_and_ack_op(op, pend) \
350 if (test_bit(op, &sh->ops.pending) && \
351 !test_bit(op, &sh->ops.complete)) { \
352 if (test_and_set_bit(op, &sh->ops.ack)) \
353 clear_bit(op, &pend); \
357 clear_bit(op, &pend); \
360 /* find new work to run, do not resubmit work that is already
363 static unsigned long get_stripe_work(struct stripe_head *sh)
365 unsigned long pending;
368 pending = sh->ops.pending;
370 test_and_ack_op(STRIPE_OP_BIOFILL, pending);
371 test_and_ack_op(STRIPE_OP_COMPUTE_BLK, pending);
372 test_and_ack_op(STRIPE_OP_PREXOR, pending);
373 test_and_ack_op(STRIPE_OP_BIODRAIN, pending);
374 test_and_ack_op(STRIPE_OP_POSTXOR, pending);
375 test_and_ack_op(STRIPE_OP_CHECK, pending);
377 sh->ops.count -= ack;
378 if (unlikely(sh->ops.count < 0)) {
379 printk(KERN_ERR "pending: %#lx ops.pending: %#lx ops.ack: %#lx "
380 "ops.complete: %#lx\n", pending, sh->ops.pending,
381 sh->ops.ack, sh->ops.complete);
389 raid5_end_read_request(struct bio *bi, int error);
391 raid5_end_write_request(struct bio *bi, int error);
393 static void ops_run_io(struct stripe_head *sh, struct stripe_head_state *s)
395 raid5_conf_t *conf = sh->raid_conf;
396 int i, disks = sh->disks;
400 for (i = disks; i--; ) {
404 if (test_and_clear_bit(R5_Wantwrite, &sh->dev[i].flags))
406 else if (test_and_clear_bit(R5_Wantread, &sh->dev[i].flags))
411 bi = &sh->dev[i].req;
415 bi->bi_end_io = raid5_end_write_request;
417 bi->bi_end_io = raid5_end_read_request;
420 rdev = rcu_dereference(conf->disks[i].rdev);
421 if (rdev && test_bit(Faulty, &rdev->flags))
424 atomic_inc(&rdev->nr_pending);
428 if (s->syncing || s->expanding || s->expanded)
429 md_sync_acct(rdev->bdev, STRIPE_SECTORS);
431 set_bit(STRIPE_IO_STARTED, &sh->state);
433 bi->bi_bdev = rdev->bdev;
434 pr_debug("%s: for %llu schedule op %ld on disc %d\n",
435 __func__, (unsigned long long)sh->sector,
437 atomic_inc(&sh->count);
438 bi->bi_sector = sh->sector + rdev->data_offset;
439 bi->bi_flags = 1 << BIO_UPTODATE;
443 bi->bi_io_vec = &sh->dev[i].vec;
444 bi->bi_io_vec[0].bv_len = STRIPE_SIZE;
445 bi->bi_io_vec[0].bv_offset = 0;
446 bi->bi_size = STRIPE_SIZE;
449 test_bit(R5_ReWrite, &sh->dev[i].flags))
450 atomic_add(STRIPE_SECTORS,
451 &rdev->corrected_errors);
452 generic_make_request(bi);
455 set_bit(STRIPE_DEGRADED, &sh->state);
456 pr_debug("skip op %ld on disc %d for sector %llu\n",
457 bi->bi_rw, i, (unsigned long long)sh->sector);
458 clear_bit(R5_LOCKED, &sh->dev[i].flags);
459 set_bit(STRIPE_HANDLE, &sh->state);
464 static struct dma_async_tx_descriptor *
465 async_copy_data(int frombio, struct bio *bio, struct page *page,
466 sector_t sector, struct dma_async_tx_descriptor *tx)
469 struct page *bio_page;
473 if (bio->bi_sector >= sector)
474 page_offset = (signed)(bio->bi_sector - sector) * 512;
476 page_offset = (signed)(sector - bio->bi_sector) * -512;
477 bio_for_each_segment(bvl, bio, i) {
478 int len = bio_iovec_idx(bio, i)->bv_len;
482 if (page_offset < 0) {
483 b_offset = -page_offset;
484 page_offset += b_offset;
488 if (len > 0 && page_offset + len > STRIPE_SIZE)
489 clen = STRIPE_SIZE - page_offset;
494 b_offset += bio_iovec_idx(bio, i)->bv_offset;
495 bio_page = bio_iovec_idx(bio, i)->bv_page;
497 tx = async_memcpy(page, bio_page, page_offset,
502 tx = async_memcpy(bio_page, page, b_offset,
507 if (clen < len) /* hit end of page */
515 static void ops_complete_biofill(void *stripe_head_ref)
517 struct stripe_head *sh = stripe_head_ref;
518 struct bio *return_bi = NULL;
519 raid5_conf_t *conf = sh->raid_conf;
522 pr_debug("%s: stripe %llu\n", __func__,
523 (unsigned long long)sh->sector);
525 /* clear completed biofills */
526 for (i = sh->disks; i--; ) {
527 struct r5dev *dev = &sh->dev[i];
529 /* acknowledge completion of a biofill operation */
530 /* and check if we need to reply to a read request,
531 * new R5_Wantfill requests are held off until
532 * !test_bit(STRIPE_OP_BIOFILL, &sh->ops.pending)
534 if (test_and_clear_bit(R5_Wantfill, &dev->flags)) {
535 struct bio *rbi, *rbi2;
537 /* The access to dev->read is outside of the
538 * spin_lock_irq(&conf->device_lock), but is protected
539 * by the STRIPE_OP_BIOFILL pending bit
544 while (rbi && rbi->bi_sector <
545 dev->sector + STRIPE_SECTORS) {
546 rbi2 = r5_next_bio(rbi, dev->sector);
547 spin_lock_irq(&conf->device_lock);
548 if (--rbi->bi_phys_segments == 0) {
549 rbi->bi_next = return_bi;
552 spin_unlock_irq(&conf->device_lock);
557 set_bit(STRIPE_OP_BIOFILL, &sh->ops.complete);
559 return_io(return_bi);
561 set_bit(STRIPE_HANDLE, &sh->state);
565 static void ops_run_biofill(struct stripe_head *sh)
567 struct dma_async_tx_descriptor *tx = NULL;
568 raid5_conf_t *conf = sh->raid_conf;
571 pr_debug("%s: stripe %llu\n", __func__,
572 (unsigned long long)sh->sector);
574 for (i = sh->disks; i--; ) {
575 struct r5dev *dev = &sh->dev[i];
576 if (test_bit(R5_Wantfill, &dev->flags)) {
578 spin_lock_irq(&conf->device_lock);
579 dev->read = rbi = dev->toread;
581 spin_unlock_irq(&conf->device_lock);
582 while (rbi && rbi->bi_sector <
583 dev->sector + STRIPE_SECTORS) {
584 tx = async_copy_data(0, rbi, dev->page,
586 rbi = r5_next_bio(rbi, dev->sector);
591 atomic_inc(&sh->count);
592 async_trigger_callback(ASYNC_TX_DEP_ACK | ASYNC_TX_ACK, tx,
593 ops_complete_biofill, sh);
596 static void ops_complete_compute5(void *stripe_head_ref)
598 struct stripe_head *sh = stripe_head_ref;
599 int target = sh->ops.target;
600 struct r5dev *tgt = &sh->dev[target];
602 pr_debug("%s: stripe %llu\n", __func__,
603 (unsigned long long)sh->sector);
605 set_bit(R5_UPTODATE, &tgt->flags);
606 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
607 clear_bit(R5_Wantcompute, &tgt->flags);
608 set_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.complete);
609 set_bit(STRIPE_HANDLE, &sh->state);
613 static struct dma_async_tx_descriptor *
614 ops_run_compute5(struct stripe_head *sh, unsigned long pending)
616 /* kernel stack size limits the total number of disks */
617 int disks = sh->disks;
618 struct page *xor_srcs[disks];
619 int target = sh->ops.target;
620 struct r5dev *tgt = &sh->dev[target];
621 struct page *xor_dest = tgt->page;
623 struct dma_async_tx_descriptor *tx;
626 pr_debug("%s: stripe %llu block: %d\n",
627 __func__, (unsigned long long)sh->sector, target);
628 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
630 for (i = disks; i--; )
632 xor_srcs[count++] = sh->dev[i].page;
634 atomic_inc(&sh->count);
636 if (unlikely(count == 1))
637 tx = async_memcpy(xor_dest, xor_srcs[0], 0, 0, STRIPE_SIZE,
638 0, NULL, ops_complete_compute5, sh);
640 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE,
641 ASYNC_TX_XOR_ZERO_DST, NULL,
642 ops_complete_compute5, sh);
644 /* ack now if postxor is not set to be run */
645 if (tx && !test_bit(STRIPE_OP_POSTXOR, &pending))
651 static void ops_complete_prexor(void *stripe_head_ref)
653 struct stripe_head *sh = stripe_head_ref;
655 pr_debug("%s: stripe %llu\n", __func__,
656 (unsigned long long)sh->sector);
658 set_bit(STRIPE_OP_PREXOR, &sh->ops.complete);
661 static struct dma_async_tx_descriptor *
662 ops_run_prexor(struct stripe_head *sh, struct dma_async_tx_descriptor *tx)
664 /* kernel stack size limits the total number of disks */
665 int disks = sh->disks;
666 struct page *xor_srcs[disks];
667 int count = 0, pd_idx = sh->pd_idx, i;
669 /* existing parity data subtracted */
670 struct page *xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
672 pr_debug("%s: stripe %llu\n", __func__,
673 (unsigned long long)sh->sector);
675 for (i = disks; i--; ) {
676 struct r5dev *dev = &sh->dev[i];
677 /* Only process blocks that are known to be uptodate */
678 if (dev->towrite && test_bit(R5_Wantprexor, &dev->flags))
679 xor_srcs[count++] = dev->page;
682 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE,
683 ASYNC_TX_DEP_ACK | ASYNC_TX_XOR_DROP_DST, tx,
684 ops_complete_prexor, sh);
689 static struct dma_async_tx_descriptor *
690 ops_run_biodrain(struct stripe_head *sh, struct dma_async_tx_descriptor *tx,
691 unsigned long pending)
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, &pending);
701 pr_debug("%s: stripe %llu\n", __func__,
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", __func__,
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", __func__,
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,
777 unsigned long pending)
779 /* kernel stack size limits the total number of disks */
780 int disks = sh->disks;
781 struct page *xor_srcs[disks];
783 int count = 0, pd_idx = sh->pd_idx, i;
784 struct page *xor_dest;
785 int prexor = test_bit(STRIPE_OP_PREXOR, &pending);
787 dma_async_tx_callback callback;
789 pr_debug("%s: stripe %llu\n", __func__,
790 (unsigned long long)sh->sector);
792 /* check if prexor is active which means only process blocks
793 * that are part of a read-modify-write (written)
796 xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
797 for (i = disks; i--; ) {
798 struct r5dev *dev = &sh->dev[i];
800 xor_srcs[count++] = dev->page;
803 xor_dest = sh->dev[pd_idx].page;
804 for (i = disks; i--; ) {
805 struct r5dev *dev = &sh->dev[i];
807 xor_srcs[count++] = dev->page;
811 /* check whether this postxor is part of a write */
812 callback = test_bit(STRIPE_OP_BIODRAIN, &pending) ?
813 ops_complete_write : ops_complete_postxor;
815 /* 1/ if we prexor'd then the dest is reused as a source
816 * 2/ if we did not prexor then we are redoing the parity
817 * set ASYNC_TX_XOR_DROP_DST and ASYNC_TX_XOR_ZERO_DST
818 * for the synchronous xor case
820 flags = ASYNC_TX_DEP_ACK | ASYNC_TX_ACK |
821 (prexor ? ASYNC_TX_XOR_DROP_DST : ASYNC_TX_XOR_ZERO_DST);
823 atomic_inc(&sh->count);
825 if (unlikely(count == 1)) {
826 flags &= ~(ASYNC_TX_XOR_DROP_DST | ASYNC_TX_XOR_ZERO_DST);
827 tx = async_memcpy(xor_dest, xor_srcs[0], 0, 0, STRIPE_SIZE,
828 flags, tx, callback, sh);
830 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE,
831 flags, tx, callback, sh);
834 static void ops_complete_check(void *stripe_head_ref)
836 struct stripe_head *sh = stripe_head_ref;
838 pr_debug("%s: stripe %llu\n", __func__,
839 (unsigned long long)sh->sector);
841 set_bit(STRIPE_OP_CHECK, &sh->ops.complete);
842 set_bit(STRIPE_HANDLE, &sh->state);
846 static void ops_run_check(struct stripe_head *sh)
848 /* kernel stack size limits the total number of disks */
849 int disks = sh->disks;
850 struct page *xor_srcs[disks];
851 struct dma_async_tx_descriptor *tx;
853 int count = 0, pd_idx = sh->pd_idx, i;
854 struct page *xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
856 pr_debug("%s: stripe %llu\n", __func__,
857 (unsigned long long)sh->sector);
859 for (i = disks; i--; ) {
860 struct r5dev *dev = &sh->dev[i];
862 xor_srcs[count++] = dev->page;
865 tx = async_xor_zero_sum(xor_dest, xor_srcs, 0, count, STRIPE_SIZE,
866 &sh->ops.zero_sum_result, 0, NULL, NULL, NULL);
868 atomic_inc(&sh->count);
869 tx = async_trigger_callback(ASYNC_TX_DEP_ACK | ASYNC_TX_ACK, tx,
870 ops_complete_check, sh);
873 static void raid5_run_ops(struct stripe_head *sh, unsigned long pending)
875 int overlap_clear = 0, i, disks = sh->disks;
876 struct dma_async_tx_descriptor *tx = NULL;
878 if (test_bit(STRIPE_OP_BIOFILL, &pending)) {
883 if (test_bit(STRIPE_OP_COMPUTE_BLK, &pending))
884 tx = ops_run_compute5(sh, pending);
886 if (test_bit(STRIPE_OP_PREXOR, &pending))
887 tx = ops_run_prexor(sh, tx);
889 if (test_bit(STRIPE_OP_BIODRAIN, &pending)) {
890 tx = ops_run_biodrain(sh, tx, pending);
894 if (test_bit(STRIPE_OP_POSTXOR, &pending))
895 ops_run_postxor(sh, tx, pending);
897 if (test_bit(STRIPE_OP_CHECK, &pending))
901 for (i = disks; i--; ) {
902 struct r5dev *dev = &sh->dev[i];
903 if (test_and_clear_bit(R5_Overlap, &dev->flags))
904 wake_up(&sh->raid_conf->wait_for_overlap);
908 static int grow_one_stripe(raid5_conf_t *conf)
910 struct stripe_head *sh;
911 sh = kmem_cache_alloc(conf->slab_cache, GFP_KERNEL);
914 memset(sh, 0, sizeof(*sh) + (conf->raid_disks-1)*sizeof(struct r5dev));
915 sh->raid_conf = conf;
916 spin_lock_init(&sh->lock);
918 if (grow_buffers(sh, conf->raid_disks)) {
919 shrink_buffers(sh, conf->raid_disks);
920 kmem_cache_free(conf->slab_cache, sh);
923 sh->disks = conf->raid_disks;
924 /* we just created an active stripe so... */
925 atomic_set(&sh->count, 1);
926 atomic_inc(&conf->active_stripes);
927 INIT_LIST_HEAD(&sh->lru);
932 static int grow_stripes(raid5_conf_t *conf, int num)
934 struct kmem_cache *sc;
935 int devs = conf->raid_disks;
937 sprintf(conf->cache_name[0], "raid5-%s", mdname(conf->mddev));
938 sprintf(conf->cache_name[1], "raid5-%s-alt", mdname(conf->mddev));
939 conf->active_name = 0;
940 sc = kmem_cache_create(conf->cache_name[conf->active_name],
941 sizeof(struct stripe_head)+(devs-1)*sizeof(struct r5dev),
945 conf->slab_cache = sc;
946 conf->pool_size = devs;
948 if (!grow_one_stripe(conf))
953 #ifdef CONFIG_MD_RAID5_RESHAPE
954 static int resize_stripes(raid5_conf_t *conf, int newsize)
956 /* Make all the stripes able to hold 'newsize' devices.
957 * New slots in each stripe get 'page' set to a new page.
959 * This happens in stages:
960 * 1/ create a new kmem_cache and allocate the required number of
962 * 2/ gather all the old stripe_heads and tranfer the pages across
963 * to the new stripe_heads. This will have the side effect of
964 * freezing the array as once all stripe_heads have been collected,
965 * no IO will be possible. Old stripe heads are freed once their
966 * pages have been transferred over, and the old kmem_cache is
967 * freed when all stripes are done.
968 * 3/ reallocate conf->disks to be suitable bigger. If this fails,
969 * we simple return a failre status - no need to clean anything up.
970 * 4/ allocate new pages for the new slots in the new stripe_heads.
971 * If this fails, we don't bother trying the shrink the
972 * stripe_heads down again, we just leave them as they are.
973 * As each stripe_head is processed the new one is released into
976 * Once step2 is started, we cannot afford to wait for a write,
977 * so we use GFP_NOIO allocations.
979 struct stripe_head *osh, *nsh;
980 LIST_HEAD(newstripes);
981 struct disk_info *ndisks;
983 struct kmem_cache *sc;
986 if (newsize <= conf->pool_size)
987 return 0; /* never bother to shrink */
989 md_allow_write(conf->mddev);
992 sc = kmem_cache_create(conf->cache_name[1-conf->active_name],
993 sizeof(struct stripe_head)+(newsize-1)*sizeof(struct r5dev),
998 for (i = conf->max_nr_stripes; i; i--) {
999 nsh = kmem_cache_alloc(sc, GFP_KERNEL);
1003 memset(nsh, 0, sizeof(*nsh) + (newsize-1)*sizeof(struct r5dev));
1005 nsh->raid_conf = conf;
1006 spin_lock_init(&nsh->lock);
1008 list_add(&nsh->lru, &newstripes);
1011 /* didn't get enough, give up */
1012 while (!list_empty(&newstripes)) {
1013 nsh = list_entry(newstripes.next, struct stripe_head, lru);
1014 list_del(&nsh->lru);
1015 kmem_cache_free(sc, nsh);
1017 kmem_cache_destroy(sc);
1020 /* Step 2 - Must use GFP_NOIO now.
1021 * OK, we have enough stripes, start collecting inactive
1022 * stripes and copying them over
1024 list_for_each_entry(nsh, &newstripes, lru) {
1025 spin_lock_irq(&conf->device_lock);
1026 wait_event_lock_irq(conf->wait_for_stripe,
1027 !list_empty(&conf->inactive_list),
1029 unplug_slaves(conf->mddev)
1031 osh = get_free_stripe(conf);
1032 spin_unlock_irq(&conf->device_lock);
1033 atomic_set(&nsh->count, 1);
1034 for(i=0; i<conf->pool_size; i++)
1035 nsh->dev[i].page = osh->dev[i].page;
1036 for( ; i<newsize; i++)
1037 nsh->dev[i].page = NULL;
1038 kmem_cache_free(conf->slab_cache, osh);
1040 kmem_cache_destroy(conf->slab_cache);
1043 * At this point, we are holding all the stripes so the array
1044 * is completely stalled, so now is a good time to resize
1047 ndisks = kzalloc(newsize * sizeof(struct disk_info), GFP_NOIO);
1049 for (i=0; i<conf->raid_disks; i++)
1050 ndisks[i] = conf->disks[i];
1052 conf->disks = ndisks;
1056 /* Step 4, return new stripes to service */
1057 while(!list_empty(&newstripes)) {
1058 nsh = list_entry(newstripes.next, struct stripe_head, lru);
1059 list_del_init(&nsh->lru);
1060 for (i=conf->raid_disks; i < newsize; i++)
1061 if (nsh->dev[i].page == NULL) {
1062 struct page *p = alloc_page(GFP_NOIO);
1063 nsh->dev[i].page = p;
1067 release_stripe(nsh);
1069 /* critical section pass, GFP_NOIO no longer needed */
1071 conf->slab_cache = sc;
1072 conf->active_name = 1-conf->active_name;
1073 conf->pool_size = newsize;
1078 static int drop_one_stripe(raid5_conf_t *conf)
1080 struct stripe_head *sh;
1082 spin_lock_irq(&conf->device_lock);
1083 sh = get_free_stripe(conf);
1084 spin_unlock_irq(&conf->device_lock);
1087 BUG_ON(atomic_read(&sh->count));
1088 shrink_buffers(sh, conf->pool_size);
1089 kmem_cache_free(conf->slab_cache, sh);
1090 atomic_dec(&conf->active_stripes);
1094 static void shrink_stripes(raid5_conf_t *conf)
1096 while (drop_one_stripe(conf))
1099 if (conf->slab_cache)
1100 kmem_cache_destroy(conf->slab_cache);
1101 conf->slab_cache = NULL;
1104 static void raid5_end_read_request(struct bio * bi, int error)
1106 struct stripe_head *sh = bi->bi_private;
1107 raid5_conf_t *conf = sh->raid_conf;
1108 int disks = sh->disks, i;
1109 int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
1110 char b[BDEVNAME_SIZE];
1114 for (i=0 ; i<disks; i++)
1115 if (bi == &sh->dev[i].req)
1118 pr_debug("end_read_request %llu/%d, count: %d, uptodate %d.\n",
1119 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
1127 set_bit(R5_UPTODATE, &sh->dev[i].flags);
1128 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
1129 rdev = conf->disks[i].rdev;
1130 printk_rl(KERN_INFO "raid5:%s: read error corrected"
1131 " (%lu sectors at %llu on %s)\n",
1132 mdname(conf->mddev), STRIPE_SECTORS,
1133 (unsigned long long)(sh->sector
1134 + rdev->data_offset),
1135 bdevname(rdev->bdev, b));
1136 clear_bit(R5_ReadError, &sh->dev[i].flags);
1137 clear_bit(R5_ReWrite, &sh->dev[i].flags);
1139 if (atomic_read(&conf->disks[i].rdev->read_errors))
1140 atomic_set(&conf->disks[i].rdev->read_errors, 0);
1142 const char *bdn = bdevname(conf->disks[i].rdev->bdev, b);
1144 rdev = conf->disks[i].rdev;
1146 clear_bit(R5_UPTODATE, &sh->dev[i].flags);
1147 atomic_inc(&rdev->read_errors);
1148 if (conf->mddev->degraded)
1149 printk_rl(KERN_WARNING
1150 "raid5:%s: read error not correctable "
1151 "(sector %llu on %s).\n",
1152 mdname(conf->mddev),
1153 (unsigned long long)(sh->sector
1154 + rdev->data_offset),
1156 else if (test_bit(R5_ReWrite, &sh->dev[i].flags))
1158 printk_rl(KERN_WARNING
1159 "raid5:%s: read error NOT corrected!! "
1160 "(sector %llu on %s).\n",
1161 mdname(conf->mddev),
1162 (unsigned long long)(sh->sector
1163 + rdev->data_offset),
1165 else if (atomic_read(&rdev->read_errors)
1166 > conf->max_nr_stripes)
1168 "raid5:%s: Too many read errors, failing device %s.\n",
1169 mdname(conf->mddev), bdn);
1173 set_bit(R5_ReadError, &sh->dev[i].flags);
1175 clear_bit(R5_ReadError, &sh->dev[i].flags);
1176 clear_bit(R5_ReWrite, &sh->dev[i].flags);
1177 md_error(conf->mddev, rdev);
1180 rdev_dec_pending(conf->disks[i].rdev, conf->mddev);
1181 clear_bit(R5_LOCKED, &sh->dev[i].flags);
1182 set_bit(STRIPE_HANDLE, &sh->state);
1186 static void raid5_end_write_request (struct bio *bi, int error)
1188 struct stripe_head *sh = bi->bi_private;
1189 raid5_conf_t *conf = sh->raid_conf;
1190 int disks = sh->disks, i;
1191 int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
1193 for (i=0 ; i<disks; i++)
1194 if (bi == &sh->dev[i].req)
1197 pr_debug("end_write_request %llu/%d, count %d, uptodate: %d.\n",
1198 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
1206 md_error(conf->mddev, conf->disks[i].rdev);
1208 rdev_dec_pending(conf->disks[i].rdev, conf->mddev);
1210 clear_bit(R5_LOCKED, &sh->dev[i].flags);
1211 set_bit(STRIPE_HANDLE, &sh->state);
1216 static sector_t compute_blocknr(struct stripe_head *sh, int i);
1218 static void raid5_build_block (struct stripe_head *sh, int i)
1220 struct r5dev *dev = &sh->dev[i];
1222 bio_init(&dev->req);
1223 dev->req.bi_io_vec = &dev->vec;
1225 dev->req.bi_max_vecs++;
1226 dev->vec.bv_page = dev->page;
1227 dev->vec.bv_len = STRIPE_SIZE;
1228 dev->vec.bv_offset = 0;
1230 dev->req.bi_sector = sh->sector;
1231 dev->req.bi_private = sh;
1234 dev->sector = compute_blocknr(sh, i);
1237 static void error(mddev_t *mddev, mdk_rdev_t *rdev)
1239 char b[BDEVNAME_SIZE];
1240 raid5_conf_t *conf = (raid5_conf_t *) mddev->private;
1241 pr_debug("raid5: error called\n");
1243 if (!test_bit(Faulty, &rdev->flags)) {
1244 set_bit(MD_CHANGE_DEVS, &mddev->flags);
1245 if (test_and_clear_bit(In_sync, &rdev->flags)) {
1246 unsigned long flags;
1247 spin_lock_irqsave(&conf->device_lock, flags);
1249 spin_unlock_irqrestore(&conf->device_lock, flags);
1251 * if recovery was running, make sure it aborts.
1253 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1255 set_bit(Faulty, &rdev->flags);
1257 "raid5: Disk failure on %s, disabling device.\n"
1258 "raid5: Operation continuing on %d devices.\n",
1259 bdevname(rdev->bdev,b), conf->raid_disks - mddev->degraded);
1264 * Input: a 'big' sector number,
1265 * Output: index of the data and parity disk, and the sector # in them.
1267 static sector_t raid5_compute_sector(sector_t r_sector, unsigned int raid_disks,
1268 unsigned int data_disks, unsigned int * dd_idx,
1269 unsigned int * pd_idx, raid5_conf_t *conf)
1272 unsigned long chunk_number;
1273 unsigned int chunk_offset;
1274 sector_t new_sector;
1275 int sectors_per_chunk = conf->chunk_size >> 9;
1277 /* First compute the information on this sector */
1280 * Compute the chunk number and the sector offset inside the chunk
1282 chunk_offset = sector_div(r_sector, sectors_per_chunk);
1283 chunk_number = r_sector;
1284 BUG_ON(r_sector != chunk_number);
1287 * Compute the stripe number
1289 stripe = chunk_number / data_disks;
1292 * Compute the data disk and parity disk indexes inside the stripe
1294 *dd_idx = chunk_number % data_disks;
1297 * Select the parity disk based on the user selected algorithm.
1299 switch(conf->level) {
1301 *pd_idx = data_disks;
1304 switch (conf->algorithm) {
1305 case ALGORITHM_LEFT_ASYMMETRIC:
1306 *pd_idx = data_disks - stripe % raid_disks;
1307 if (*dd_idx >= *pd_idx)
1310 case ALGORITHM_RIGHT_ASYMMETRIC:
1311 *pd_idx = stripe % raid_disks;
1312 if (*dd_idx >= *pd_idx)
1315 case ALGORITHM_LEFT_SYMMETRIC:
1316 *pd_idx = data_disks - stripe % raid_disks;
1317 *dd_idx = (*pd_idx + 1 + *dd_idx) % raid_disks;
1319 case ALGORITHM_RIGHT_SYMMETRIC:
1320 *pd_idx = stripe % raid_disks;
1321 *dd_idx = (*pd_idx + 1 + *dd_idx) % raid_disks;
1324 printk(KERN_ERR "raid5: unsupported algorithm %d\n",
1330 /**** FIX THIS ****/
1331 switch (conf->algorithm) {
1332 case ALGORITHM_LEFT_ASYMMETRIC:
1333 *pd_idx = raid_disks - 1 - (stripe % raid_disks);
1334 if (*pd_idx == raid_disks-1)
1335 (*dd_idx)++; /* Q D D D P */
1336 else if (*dd_idx >= *pd_idx)
1337 (*dd_idx) += 2; /* D D P Q D */
1339 case ALGORITHM_RIGHT_ASYMMETRIC:
1340 *pd_idx = stripe % raid_disks;
1341 if (*pd_idx == raid_disks-1)
1342 (*dd_idx)++; /* Q D D D P */
1343 else if (*dd_idx >= *pd_idx)
1344 (*dd_idx) += 2; /* D D P Q D */
1346 case ALGORITHM_LEFT_SYMMETRIC:
1347 *pd_idx = raid_disks - 1 - (stripe % raid_disks);
1348 *dd_idx = (*pd_idx + 2 + *dd_idx) % raid_disks;
1350 case ALGORITHM_RIGHT_SYMMETRIC:
1351 *pd_idx = stripe % raid_disks;
1352 *dd_idx = (*pd_idx + 2 + *dd_idx) % raid_disks;
1355 printk (KERN_CRIT "raid6: unsupported algorithm %d\n",
1362 * Finally, compute the new sector number
1364 new_sector = (sector_t)stripe * sectors_per_chunk + chunk_offset;
1369 static sector_t compute_blocknr(struct stripe_head *sh, int i)
1371 raid5_conf_t *conf = sh->raid_conf;
1372 int raid_disks = sh->disks;
1373 int data_disks = raid_disks - conf->max_degraded;
1374 sector_t new_sector = sh->sector, check;
1375 int sectors_per_chunk = conf->chunk_size >> 9;
1378 int chunk_number, dummy1, dummy2, dd_idx = i;
1382 chunk_offset = sector_div(new_sector, sectors_per_chunk);
1383 stripe = new_sector;
1384 BUG_ON(new_sector != stripe);
1386 if (i == sh->pd_idx)
1388 switch(conf->level) {
1391 switch (conf->algorithm) {
1392 case ALGORITHM_LEFT_ASYMMETRIC:
1393 case ALGORITHM_RIGHT_ASYMMETRIC:
1397 case ALGORITHM_LEFT_SYMMETRIC:
1398 case ALGORITHM_RIGHT_SYMMETRIC:
1401 i -= (sh->pd_idx + 1);
1404 printk(KERN_ERR "raid5: unsupported algorithm %d\n",
1409 if (i == raid6_next_disk(sh->pd_idx, raid_disks))
1410 return 0; /* It is the Q disk */
1411 switch (conf->algorithm) {
1412 case ALGORITHM_LEFT_ASYMMETRIC:
1413 case ALGORITHM_RIGHT_ASYMMETRIC:
1414 if (sh->pd_idx == raid_disks-1)
1415 i--; /* Q D D D P */
1416 else if (i > sh->pd_idx)
1417 i -= 2; /* D D P Q D */
1419 case ALGORITHM_LEFT_SYMMETRIC:
1420 case ALGORITHM_RIGHT_SYMMETRIC:
1421 if (sh->pd_idx == raid_disks-1)
1422 i--; /* Q D D D P */
1427 i -= (sh->pd_idx + 2);
1431 printk (KERN_CRIT "raid6: unsupported algorithm %d\n",
1437 chunk_number = stripe * data_disks + i;
1438 r_sector = (sector_t)chunk_number * sectors_per_chunk + chunk_offset;
1440 check = raid5_compute_sector (r_sector, raid_disks, data_disks, &dummy1, &dummy2, conf);
1441 if (check != sh->sector || dummy1 != dd_idx || dummy2 != sh->pd_idx) {
1442 printk(KERN_ERR "compute_blocknr: map not correct\n");
1451 * Copy data between a page in the stripe cache, and one or more bion
1452 * The page could align with the middle of the bio, or there could be
1453 * several bion, each with several bio_vecs, which cover part of the page
1454 * Multiple bion are linked together on bi_next. There may be extras
1455 * at the end of this list. We ignore them.
1457 static void copy_data(int frombio, struct bio *bio,
1461 char *pa = page_address(page);
1462 struct bio_vec *bvl;
1466 if (bio->bi_sector >= sector)
1467 page_offset = (signed)(bio->bi_sector - sector) * 512;
1469 page_offset = (signed)(sector - bio->bi_sector) * -512;
1470 bio_for_each_segment(bvl, bio, i) {
1471 int len = bio_iovec_idx(bio,i)->bv_len;
1475 if (page_offset < 0) {
1476 b_offset = -page_offset;
1477 page_offset += b_offset;
1481 if (len > 0 && page_offset + len > STRIPE_SIZE)
1482 clen = STRIPE_SIZE - page_offset;
1486 char *ba = __bio_kmap_atomic(bio, i, KM_USER0);
1488 memcpy(pa+page_offset, ba+b_offset, clen);
1490 memcpy(ba+b_offset, pa+page_offset, clen);
1491 __bio_kunmap_atomic(ba, KM_USER0);
1493 if (clen < len) /* hit end of page */
1499 #define check_xor() do { \
1500 if (count == MAX_XOR_BLOCKS) { \
1501 xor_blocks(count, STRIPE_SIZE, dest, ptr);\
1506 static void compute_parity6(struct stripe_head *sh, int method)
1508 raid6_conf_t *conf = sh->raid_conf;
1509 int i, pd_idx = sh->pd_idx, qd_idx, d0_idx, disks = sh->disks, count;
1511 /**** FIX THIS: This could be very bad if disks is close to 256 ****/
1514 qd_idx = raid6_next_disk(pd_idx, disks);
1515 d0_idx = raid6_next_disk(qd_idx, disks);
1517 pr_debug("compute_parity, stripe %llu, method %d\n",
1518 (unsigned long long)sh->sector, method);
1521 case READ_MODIFY_WRITE:
1522 BUG(); /* READ_MODIFY_WRITE N/A for RAID-6 */
1523 case RECONSTRUCT_WRITE:
1524 for (i= disks; i-- ;)
1525 if ( i != pd_idx && i != qd_idx && sh->dev[i].towrite ) {
1526 chosen = sh->dev[i].towrite;
1527 sh->dev[i].towrite = NULL;
1529 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
1530 wake_up(&conf->wait_for_overlap);
1532 BUG_ON(sh->dev[i].written);
1533 sh->dev[i].written = chosen;
1537 BUG(); /* Not implemented yet */
1540 for (i = disks; i--;)
1541 if (sh->dev[i].written) {
1542 sector_t sector = sh->dev[i].sector;
1543 struct bio *wbi = sh->dev[i].written;
1544 while (wbi && wbi->bi_sector < sector + STRIPE_SECTORS) {
1545 copy_data(1, wbi, sh->dev[i].page, sector);
1546 wbi = r5_next_bio(wbi, sector);
1549 set_bit(R5_LOCKED, &sh->dev[i].flags);
1550 set_bit(R5_UPTODATE, &sh->dev[i].flags);
1554 // case RECONSTRUCT_WRITE:
1555 // case CHECK_PARITY:
1556 // case UPDATE_PARITY:
1557 /* Note that unlike RAID-5, the ordering of the disks matters greatly. */
1558 /* FIX: Is this ordering of drives even remotely optimal? */
1562 ptrs[count++] = page_address(sh->dev[i].page);
1563 if (count <= disks-2 && !test_bit(R5_UPTODATE, &sh->dev[i].flags))
1564 printk("block %d/%d not uptodate on parity calc\n", i,count);
1565 i = raid6_next_disk(i, disks);
1566 } while ( i != d0_idx );
1570 raid6_call.gen_syndrome(disks, STRIPE_SIZE, ptrs);
1573 case RECONSTRUCT_WRITE:
1574 set_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
1575 set_bit(R5_UPTODATE, &sh->dev[qd_idx].flags);
1576 set_bit(R5_LOCKED, &sh->dev[pd_idx].flags);
1577 set_bit(R5_LOCKED, &sh->dev[qd_idx].flags);
1580 set_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
1581 set_bit(R5_UPTODATE, &sh->dev[qd_idx].flags);
1587 /* Compute one missing block */
1588 static void compute_block_1(struct stripe_head *sh, int dd_idx, int nozero)
1590 int i, count, disks = sh->disks;
1591 void *ptr[MAX_XOR_BLOCKS], *dest, *p;
1592 int pd_idx = sh->pd_idx;
1593 int qd_idx = raid6_next_disk(pd_idx, disks);
1595 pr_debug("compute_block_1, stripe %llu, idx %d\n",
1596 (unsigned long long)sh->sector, dd_idx);
1598 if ( dd_idx == qd_idx ) {
1599 /* We're actually computing the Q drive */
1600 compute_parity6(sh, UPDATE_PARITY);
1602 dest = page_address(sh->dev[dd_idx].page);
1603 if (!nozero) memset(dest, 0, STRIPE_SIZE);
1605 for (i = disks ; i--; ) {
1606 if (i == dd_idx || i == qd_idx)
1608 p = page_address(sh->dev[i].page);
1609 if (test_bit(R5_UPTODATE, &sh->dev[i].flags))
1612 printk("compute_block() %d, stripe %llu, %d"
1613 " not present\n", dd_idx,
1614 (unsigned long long)sh->sector, i);
1619 xor_blocks(count, STRIPE_SIZE, dest, ptr);
1620 if (!nozero) set_bit(R5_UPTODATE, &sh->dev[dd_idx].flags);
1621 else clear_bit(R5_UPTODATE, &sh->dev[dd_idx].flags);
1625 /* Compute two missing blocks */
1626 static void compute_block_2(struct stripe_head *sh, int dd_idx1, int dd_idx2)
1628 int i, count, disks = sh->disks;
1629 int pd_idx = sh->pd_idx;
1630 int qd_idx = raid6_next_disk(pd_idx, disks);
1631 int d0_idx = raid6_next_disk(qd_idx, disks);
1634 /* faila and failb are disk numbers relative to d0_idx */
1635 /* pd_idx become disks-2 and qd_idx become disks-1 */
1636 faila = (dd_idx1 < d0_idx) ? dd_idx1+(disks-d0_idx) : dd_idx1-d0_idx;
1637 failb = (dd_idx2 < d0_idx) ? dd_idx2+(disks-d0_idx) : dd_idx2-d0_idx;
1639 BUG_ON(faila == failb);
1640 if ( failb < faila ) { int tmp = faila; faila = failb; failb = tmp; }
1642 pr_debug("compute_block_2, stripe %llu, idx %d,%d (%d,%d)\n",
1643 (unsigned long long)sh->sector, dd_idx1, dd_idx2, faila, failb);
1645 if ( failb == disks-1 ) {
1646 /* Q disk is one of the missing disks */
1647 if ( faila == disks-2 ) {
1648 /* Missing P+Q, just recompute */
1649 compute_parity6(sh, UPDATE_PARITY);
1652 /* We're missing D+Q; recompute D from P */
1653 compute_block_1(sh, (dd_idx1 == qd_idx) ? dd_idx2 : dd_idx1, 0);
1654 compute_parity6(sh, UPDATE_PARITY); /* Is this necessary? */
1659 /* We're missing D+P or D+D; build pointer table */
1661 /**** FIX THIS: This could be very bad if disks is close to 256 ****/
1667 ptrs[count++] = page_address(sh->dev[i].page);
1668 i = raid6_next_disk(i, disks);
1669 if (i != dd_idx1 && i != dd_idx2 &&
1670 !test_bit(R5_UPTODATE, &sh->dev[i].flags))
1671 printk("compute_2 with missing block %d/%d\n", count, i);
1672 } while ( i != d0_idx );
1674 if ( failb == disks-2 ) {
1675 /* We're missing D+P. */
1676 raid6_datap_recov(disks, STRIPE_SIZE, faila, ptrs);
1678 /* We're missing D+D. */
1679 raid6_2data_recov(disks, STRIPE_SIZE, faila, failb, ptrs);
1682 /* Both the above update both missing blocks */
1683 set_bit(R5_UPTODATE, &sh->dev[dd_idx1].flags);
1684 set_bit(R5_UPTODATE, &sh->dev[dd_idx2].flags);
1689 handle_write_operations5(struct stripe_head *sh, int rcw, int expand)
1691 int i, pd_idx = sh->pd_idx, disks = sh->disks;
1695 /* if we are not expanding this is a proper write request, and
1696 * there will be bios with new data to be drained into the
1700 set_bit(STRIPE_OP_BIODRAIN, &sh->ops.pending);
1704 set_bit(STRIPE_OP_POSTXOR, &sh->ops.pending);
1707 for (i = disks; i--; ) {
1708 struct r5dev *dev = &sh->dev[i];
1711 set_bit(R5_LOCKED, &dev->flags);
1713 clear_bit(R5_UPTODATE, &dev->flags);
1717 if (locked + 1 == disks)
1718 if (!test_and_set_bit(STRIPE_FULL_WRITE, &sh->state))
1719 atomic_inc(&sh->raid_conf->pending_full_writes);
1721 BUG_ON(!(test_bit(R5_UPTODATE, &sh->dev[pd_idx].flags) ||
1722 test_bit(R5_Wantcompute, &sh->dev[pd_idx].flags)));
1724 set_bit(STRIPE_OP_PREXOR, &sh->ops.pending);
1725 set_bit(STRIPE_OP_BIODRAIN, &sh->ops.pending);
1726 set_bit(STRIPE_OP_POSTXOR, &sh->ops.pending);
1730 for (i = disks; i--; ) {
1731 struct r5dev *dev = &sh->dev[i];
1735 /* For a read-modify write there may be blocks that are
1736 * locked for reading while others are ready to be
1737 * written so we distinguish these blocks by the
1741 (test_bit(R5_UPTODATE, &dev->flags) ||
1742 test_bit(R5_Wantcompute, &dev->flags))) {
1743 set_bit(R5_Wantprexor, &dev->flags);
1744 set_bit(R5_LOCKED, &dev->flags);
1745 clear_bit(R5_UPTODATE, &dev->flags);
1751 /* keep the parity disk locked while asynchronous operations
1754 set_bit(R5_LOCKED, &sh->dev[pd_idx].flags);
1755 clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
1758 pr_debug("%s: stripe %llu locked: %d pending: %lx\n",
1759 __func__, (unsigned long long)sh->sector,
1760 locked, sh->ops.pending);
1766 * Each stripe/dev can have one or more bion attached.
1767 * toread/towrite point to the first in a chain.
1768 * The bi_next chain must be in order.
1770 static int add_stripe_bio(struct stripe_head *sh, struct bio *bi, int dd_idx, int forwrite)
1773 raid5_conf_t *conf = sh->raid_conf;
1776 pr_debug("adding bh b#%llu to stripe s#%llu\n",
1777 (unsigned long long)bi->bi_sector,
1778 (unsigned long long)sh->sector);
1781 spin_lock(&sh->lock);
1782 spin_lock_irq(&conf->device_lock);
1784 bip = &sh->dev[dd_idx].towrite;
1785 if (*bip == NULL && sh->dev[dd_idx].written == NULL)
1788 bip = &sh->dev[dd_idx].toread;
1789 while (*bip && (*bip)->bi_sector < bi->bi_sector) {
1790 if ((*bip)->bi_sector + ((*bip)->bi_size >> 9) > bi->bi_sector)
1792 bip = & (*bip)->bi_next;
1794 if (*bip && (*bip)->bi_sector < bi->bi_sector + ((bi->bi_size)>>9))
1797 BUG_ON(*bip && bi->bi_next && (*bip) != bi->bi_next);
1801 bi->bi_phys_segments ++;
1802 spin_unlock_irq(&conf->device_lock);
1803 spin_unlock(&sh->lock);
1805 pr_debug("added bi b#%llu to stripe s#%llu, disk %d.\n",
1806 (unsigned long long)bi->bi_sector,
1807 (unsigned long long)sh->sector, dd_idx);
1809 if (conf->mddev->bitmap && firstwrite) {
1810 bitmap_startwrite(conf->mddev->bitmap, sh->sector,
1812 sh->bm_seq = conf->seq_flush+1;
1813 set_bit(STRIPE_BIT_DELAY, &sh->state);
1817 /* check if page is covered */
1818 sector_t sector = sh->dev[dd_idx].sector;
1819 for (bi=sh->dev[dd_idx].towrite;
1820 sector < sh->dev[dd_idx].sector + STRIPE_SECTORS &&
1821 bi && bi->bi_sector <= sector;
1822 bi = r5_next_bio(bi, sh->dev[dd_idx].sector)) {
1823 if (bi->bi_sector + (bi->bi_size>>9) >= sector)
1824 sector = bi->bi_sector + (bi->bi_size>>9);
1826 if (sector >= sh->dev[dd_idx].sector + STRIPE_SECTORS)
1827 set_bit(R5_OVERWRITE, &sh->dev[dd_idx].flags);
1832 set_bit(R5_Overlap, &sh->dev[dd_idx].flags);
1833 spin_unlock_irq(&conf->device_lock);
1834 spin_unlock(&sh->lock);
1838 static void end_reshape(raid5_conf_t *conf);
1840 static int page_is_zero(struct page *p)
1842 char *a = page_address(p);
1843 return ((*(u32*)a) == 0 &&
1844 memcmp(a, a+4, STRIPE_SIZE-4)==0);
1847 static int stripe_to_pdidx(sector_t stripe, raid5_conf_t *conf, int disks)
1849 int sectors_per_chunk = conf->chunk_size >> 9;
1851 int chunk_offset = sector_div(stripe, sectors_per_chunk);
1853 raid5_compute_sector(stripe * (disks - conf->max_degraded)
1854 *sectors_per_chunk + chunk_offset,
1855 disks, disks - conf->max_degraded,
1856 &dd_idx, &pd_idx, conf);
1861 handle_requests_to_failed_array(raid5_conf_t *conf, struct stripe_head *sh,
1862 struct stripe_head_state *s, int disks,
1863 struct bio **return_bi)
1866 for (i = disks; i--; ) {
1870 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
1873 rdev = rcu_dereference(conf->disks[i].rdev);
1874 if (rdev && test_bit(In_sync, &rdev->flags))
1875 /* multiple read failures in one stripe */
1876 md_error(conf->mddev, rdev);
1879 spin_lock_irq(&conf->device_lock);
1880 /* fail all writes first */
1881 bi = sh->dev[i].towrite;
1882 sh->dev[i].towrite = NULL;
1888 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
1889 wake_up(&conf->wait_for_overlap);
1891 while (bi && bi->bi_sector <
1892 sh->dev[i].sector + STRIPE_SECTORS) {
1893 struct bio *nextbi = r5_next_bio(bi, sh->dev[i].sector);
1894 clear_bit(BIO_UPTODATE, &bi->bi_flags);
1895 if (--bi->bi_phys_segments == 0) {
1896 md_write_end(conf->mddev);
1897 bi->bi_next = *return_bi;
1902 /* and fail all 'written' */
1903 bi = sh->dev[i].written;
1904 sh->dev[i].written = NULL;
1905 if (bi) bitmap_end = 1;
1906 while (bi && bi->bi_sector <
1907 sh->dev[i].sector + STRIPE_SECTORS) {
1908 struct bio *bi2 = r5_next_bio(bi, sh->dev[i].sector);
1909 clear_bit(BIO_UPTODATE, &bi->bi_flags);
1910 if (--bi->bi_phys_segments == 0) {
1911 md_write_end(conf->mddev);
1912 bi->bi_next = *return_bi;
1918 /* fail any reads if this device is non-operational and
1919 * the data has not reached the cache yet.
1921 if (!test_bit(R5_Wantfill, &sh->dev[i].flags) &&
1922 (!test_bit(R5_Insync, &sh->dev[i].flags) ||
1923 test_bit(R5_ReadError, &sh->dev[i].flags))) {
1924 bi = sh->dev[i].toread;
1925 sh->dev[i].toread = NULL;
1926 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
1927 wake_up(&conf->wait_for_overlap);
1928 if (bi) s->to_read--;
1929 while (bi && bi->bi_sector <
1930 sh->dev[i].sector + STRIPE_SECTORS) {
1931 struct bio *nextbi =
1932 r5_next_bio(bi, sh->dev[i].sector);
1933 clear_bit(BIO_UPTODATE, &bi->bi_flags);
1934 if (--bi->bi_phys_segments == 0) {
1935 bi->bi_next = *return_bi;
1941 spin_unlock_irq(&conf->device_lock);
1943 bitmap_endwrite(conf->mddev->bitmap, sh->sector,
1944 STRIPE_SECTORS, 0, 0);
1947 if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
1948 if (atomic_dec_and_test(&conf->pending_full_writes))
1949 md_wakeup_thread(conf->mddev->thread);
1952 /* __handle_issuing_new_read_requests5 - returns 0 if there are no more disks
1955 static int __handle_issuing_new_read_requests5(struct stripe_head *sh,
1956 struct stripe_head_state *s, int disk_idx, int disks)
1958 struct r5dev *dev = &sh->dev[disk_idx];
1959 struct r5dev *failed_dev = &sh->dev[s->failed_num];
1961 /* don't schedule compute operations or reads on the parity block while
1962 * a check is in flight
1964 if ((disk_idx == sh->pd_idx) &&
1965 test_bit(STRIPE_OP_CHECK, &sh->ops.pending))
1968 /* is the data in this block needed, and can we get it? */
1969 if (!test_bit(R5_LOCKED, &dev->flags) &&
1970 !test_bit(R5_UPTODATE, &dev->flags) && (dev->toread ||
1971 (dev->towrite && !test_bit(R5_OVERWRITE, &dev->flags)) ||
1972 s->syncing || s->expanding || (s->failed &&
1973 (failed_dev->toread || (failed_dev->towrite &&
1974 !test_bit(R5_OVERWRITE, &failed_dev->flags)
1976 /* 1/ We would like to get this block, possibly by computing it,
1977 * but we might not be able to.
1979 * 2/ Since parity check operations potentially make the parity
1980 * block !uptodate it will need to be refreshed before any
1981 * compute operations on data disks are scheduled.
1983 * 3/ We hold off parity block re-reads until check operations
1986 if ((s->uptodate == disks - 1) &&
1987 (s->failed && disk_idx == s->failed_num) &&
1988 !test_bit(STRIPE_OP_CHECK, &sh->ops.pending)) {
1989 set_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.pending);
1990 set_bit(R5_Wantcompute, &dev->flags);
1991 sh->ops.target = disk_idx;
1994 /* Careful: from this point on 'uptodate' is in the eye
1995 * of raid5_run_ops which services 'compute' operations
1996 * before writes. R5_Wantcompute flags a block that will
1997 * be R5_UPTODATE by the time it is needed for a
1998 * subsequent operation.
2001 return 0; /* uptodate + compute == disks */
2002 } else if ((s->uptodate < disks - 1) &&
2003 test_bit(R5_Insync, &dev->flags)) {
2004 /* Note: we hold off compute operations while checks are
2005 * in flight, but we still prefer 'compute' over 'read'
2006 * hence we only read if (uptodate < * disks-1)
2008 set_bit(R5_LOCKED, &dev->flags);
2009 set_bit(R5_Wantread, &dev->flags);
2011 pr_debug("Reading block %d (sync=%d)\n", disk_idx,
2019 static void handle_issuing_new_read_requests5(struct stripe_head *sh,
2020 struct stripe_head_state *s, int disks)
2024 /* Clear completed compute operations. Parity recovery
2025 * (STRIPE_OP_MOD_REPAIR_PD) implies a write-back which is handled
2026 * later on in this routine
2028 if (test_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.complete) &&
2029 !test_bit(STRIPE_OP_MOD_REPAIR_PD, &sh->ops.pending)) {
2030 clear_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.complete);
2031 clear_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.ack);
2032 clear_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.pending);
2035 /* look for blocks to read/compute, skip this if a compute
2036 * is already in flight, or if the stripe contents are in the
2037 * midst of changing due to a write
2039 if (!test_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.pending) &&
2040 !test_bit(STRIPE_OP_PREXOR, &sh->ops.pending) &&
2041 !test_bit(STRIPE_OP_POSTXOR, &sh->ops.pending)) {
2042 for (i = disks; i--; )
2043 if (__handle_issuing_new_read_requests5(
2044 sh, s, i, disks) == 0)
2047 set_bit(STRIPE_HANDLE, &sh->state);
2050 static void handle_issuing_new_read_requests6(struct stripe_head *sh,
2051 struct stripe_head_state *s, struct r6_state *r6s,
2055 for (i = disks; i--; ) {
2056 struct r5dev *dev = &sh->dev[i];
2057 if (!test_bit(R5_LOCKED, &dev->flags) &&
2058 !test_bit(R5_UPTODATE, &dev->flags) &&
2059 (dev->toread || (dev->towrite &&
2060 !test_bit(R5_OVERWRITE, &dev->flags)) ||
2061 s->syncing || s->expanding ||
2063 (sh->dev[r6s->failed_num[0]].toread ||
2066 (sh->dev[r6s->failed_num[1]].toread ||
2068 /* we would like to get this block, possibly
2069 * by computing it, but we might not be able to
2071 if ((s->uptodate == disks - 1) &&
2072 (s->failed && (i == r6s->failed_num[0] ||
2073 i == r6s->failed_num[1]))) {
2074 pr_debug("Computing stripe %llu block %d\n",
2075 (unsigned long long)sh->sector, i);
2076 compute_block_1(sh, i, 0);
2078 } else if ( s->uptodate == disks-2 && s->failed >= 2 ) {
2079 /* Computing 2-failure is *very* expensive; only
2080 * do it if failed >= 2
2083 for (other = disks; other--; ) {
2086 if (!test_bit(R5_UPTODATE,
2087 &sh->dev[other].flags))
2091 pr_debug("Computing stripe %llu blocks %d,%d\n",
2092 (unsigned long long)sh->sector,
2094 compute_block_2(sh, i, other);
2096 } else if (test_bit(R5_Insync, &dev->flags)) {
2097 set_bit(R5_LOCKED, &dev->flags);
2098 set_bit(R5_Wantread, &dev->flags);
2100 pr_debug("Reading block %d (sync=%d)\n",
2105 set_bit(STRIPE_HANDLE, &sh->state);
2109 /* handle_completed_write_requests
2110 * any written block on an uptodate or failed drive can be returned.
2111 * Note that if we 'wrote' to a failed drive, it will be UPTODATE, but
2112 * never LOCKED, so we don't need to test 'failed' directly.
2114 static void handle_completed_write_requests(raid5_conf_t *conf,
2115 struct stripe_head *sh, int disks, struct bio **return_bi)
2120 for (i = disks; i--; )
2121 if (sh->dev[i].written) {
2123 if (!test_bit(R5_LOCKED, &dev->flags) &&
2124 test_bit(R5_UPTODATE, &dev->flags)) {
2125 /* We can return any write requests */
2126 struct bio *wbi, *wbi2;
2128 pr_debug("Return write for disc %d\n", i);
2129 spin_lock_irq(&conf->device_lock);
2131 dev->written = NULL;
2132 while (wbi && wbi->bi_sector <
2133 dev->sector + STRIPE_SECTORS) {
2134 wbi2 = r5_next_bio(wbi, dev->sector);
2135 if (--wbi->bi_phys_segments == 0) {
2136 md_write_end(conf->mddev);
2137 wbi->bi_next = *return_bi;
2142 if (dev->towrite == NULL)
2144 spin_unlock_irq(&conf->device_lock);
2146 bitmap_endwrite(conf->mddev->bitmap,
2149 !test_bit(STRIPE_DEGRADED, &sh->state),
2154 if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
2155 if (atomic_dec_and_test(&conf->pending_full_writes))
2156 md_wakeup_thread(conf->mddev->thread);
2159 static void handle_issuing_new_write_requests5(raid5_conf_t *conf,
2160 struct stripe_head *sh, struct stripe_head_state *s, int disks)
2162 int rmw = 0, rcw = 0, i;
2163 for (i = disks; i--; ) {
2164 /* would I have to read this buffer for read_modify_write */
2165 struct r5dev *dev = &sh->dev[i];
2166 if ((dev->towrite || i == sh->pd_idx) &&
2167 !test_bit(R5_LOCKED, &dev->flags) &&
2168 !(test_bit(R5_UPTODATE, &dev->flags) ||
2169 test_bit(R5_Wantcompute, &dev->flags))) {
2170 if (test_bit(R5_Insync, &dev->flags))
2173 rmw += 2*disks; /* cannot read it */
2175 /* Would I have to read this buffer for reconstruct_write */
2176 if (!test_bit(R5_OVERWRITE, &dev->flags) && i != sh->pd_idx &&
2177 !test_bit(R5_LOCKED, &dev->flags) &&
2178 !(test_bit(R5_UPTODATE, &dev->flags) ||
2179 test_bit(R5_Wantcompute, &dev->flags))) {
2180 if (test_bit(R5_Insync, &dev->flags)) rcw++;
2185 pr_debug("for sector %llu, rmw=%d rcw=%d\n",
2186 (unsigned long long)sh->sector, rmw, rcw);
2187 set_bit(STRIPE_HANDLE, &sh->state);
2188 if (rmw < rcw && rmw > 0)
2189 /* prefer read-modify-write, but need to get some data */
2190 for (i = disks; i--; ) {
2191 struct r5dev *dev = &sh->dev[i];
2192 if ((dev->towrite || i == sh->pd_idx) &&
2193 !test_bit(R5_LOCKED, &dev->flags) &&
2194 !(test_bit(R5_UPTODATE, &dev->flags) ||
2195 test_bit(R5_Wantcompute, &dev->flags)) &&
2196 test_bit(R5_Insync, &dev->flags)) {
2198 test_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2199 pr_debug("Read_old block "
2200 "%d for r-m-w\n", i);
2201 set_bit(R5_LOCKED, &dev->flags);
2202 set_bit(R5_Wantread, &dev->flags);
2205 set_bit(STRIPE_DELAYED, &sh->state);
2206 set_bit(STRIPE_HANDLE, &sh->state);
2210 if (rcw <= rmw && rcw > 0)
2211 /* want reconstruct write, but need to get some data */
2212 for (i = disks; i--; ) {
2213 struct r5dev *dev = &sh->dev[i];
2214 if (!test_bit(R5_OVERWRITE, &dev->flags) &&
2216 !test_bit(R5_LOCKED, &dev->flags) &&
2217 !(test_bit(R5_UPTODATE, &dev->flags) ||
2218 test_bit(R5_Wantcompute, &dev->flags)) &&
2219 test_bit(R5_Insync, &dev->flags)) {
2221 test_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2222 pr_debug("Read_old block "
2223 "%d for Reconstruct\n", i);
2224 set_bit(R5_LOCKED, &dev->flags);
2225 set_bit(R5_Wantread, &dev->flags);
2228 set_bit(STRIPE_DELAYED, &sh->state);
2229 set_bit(STRIPE_HANDLE, &sh->state);
2233 /* now if nothing is locked, and if we have enough data,
2234 * we can start a write request
2236 /* since handle_stripe can be called at any time we need to handle the
2237 * case where a compute block operation has been submitted and then a
2238 * subsequent call wants to start a write request. raid5_run_ops only
2239 * handles the case where compute block and postxor are requested
2240 * simultaneously. If this is not the case then new writes need to be
2241 * held off until the compute completes.
2243 if ((s->req_compute ||
2244 !test_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.pending)) &&
2245 (s->locked == 0 && (rcw == 0 || rmw == 0) &&
2246 !test_bit(STRIPE_BIT_DELAY, &sh->state)))
2247 s->locked += handle_write_operations5(sh, rcw == 0, 0);
2250 static void handle_issuing_new_write_requests6(raid5_conf_t *conf,
2251 struct stripe_head *sh, struct stripe_head_state *s,
2252 struct r6_state *r6s, int disks)
2254 int rcw = 0, must_compute = 0, pd_idx = sh->pd_idx, i;
2255 int qd_idx = r6s->qd_idx;
2256 for (i = disks; i--; ) {
2257 struct r5dev *dev = &sh->dev[i];
2258 /* Would I have to read this buffer for reconstruct_write */
2259 if (!test_bit(R5_OVERWRITE, &dev->flags)
2260 && i != pd_idx && i != qd_idx
2261 && (!test_bit(R5_LOCKED, &dev->flags)
2263 !test_bit(R5_UPTODATE, &dev->flags)) {
2264 if (test_bit(R5_Insync, &dev->flags)) rcw++;
2266 pr_debug("raid6: must_compute: "
2267 "disk %d flags=%#lx\n", i, dev->flags);
2272 pr_debug("for sector %llu, rcw=%d, must_compute=%d\n",
2273 (unsigned long long)sh->sector, rcw, must_compute);
2274 set_bit(STRIPE_HANDLE, &sh->state);
2277 /* want reconstruct write, but need to get some data */
2278 for (i = disks; i--; ) {
2279 struct r5dev *dev = &sh->dev[i];
2280 if (!test_bit(R5_OVERWRITE, &dev->flags)
2281 && !(s->failed == 0 && (i == pd_idx || i == qd_idx))
2282 && !test_bit(R5_LOCKED, &dev->flags) &&
2283 !test_bit(R5_UPTODATE, &dev->flags) &&
2284 test_bit(R5_Insync, &dev->flags)) {
2286 test_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2287 pr_debug("Read_old stripe %llu "
2288 "block %d for Reconstruct\n",
2289 (unsigned long long)sh->sector, i);
2290 set_bit(R5_LOCKED, &dev->flags);
2291 set_bit(R5_Wantread, &dev->flags);
2294 pr_debug("Request delayed stripe %llu "
2295 "block %d for Reconstruct\n",
2296 (unsigned long long)sh->sector, i);
2297 set_bit(STRIPE_DELAYED, &sh->state);
2298 set_bit(STRIPE_HANDLE, &sh->state);
2302 /* now if nothing is locked, and if we have enough data, we can start a
2305 if (s->locked == 0 && rcw == 0 &&
2306 !test_bit(STRIPE_BIT_DELAY, &sh->state)) {
2307 if (must_compute > 0) {
2308 /* We have failed blocks and need to compute them */
2309 switch (s->failed) {
2313 compute_block_1(sh, r6s->failed_num[0], 0);
2316 compute_block_2(sh, r6s->failed_num[0],
2317 r6s->failed_num[1]);
2319 default: /* This request should have been failed? */
2324 pr_debug("Computing parity for stripe %llu\n",
2325 (unsigned long long)sh->sector);
2326 compute_parity6(sh, RECONSTRUCT_WRITE);
2327 /* now every locked buffer is ready to be written */
2328 for (i = disks; i--; )
2329 if (test_bit(R5_LOCKED, &sh->dev[i].flags)) {
2330 pr_debug("Writing stripe %llu block %d\n",
2331 (unsigned long long)sh->sector, i);
2333 set_bit(R5_Wantwrite, &sh->dev[i].flags);
2335 if (s->locked == disks)
2336 if (!test_and_set_bit(STRIPE_FULL_WRITE, &sh->state))
2337 atomic_inc(&conf->pending_full_writes);
2338 /* after a RECONSTRUCT_WRITE, the stripe MUST be in-sync */
2339 set_bit(STRIPE_INSYNC, &sh->state);
2341 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2342 atomic_dec(&conf->preread_active_stripes);
2343 if (atomic_read(&conf->preread_active_stripes) <
2345 md_wakeup_thread(conf->mddev->thread);
2350 static void handle_parity_checks5(raid5_conf_t *conf, struct stripe_head *sh,
2351 struct stripe_head_state *s, int disks)
2353 int canceled_check = 0;
2355 set_bit(STRIPE_HANDLE, &sh->state);
2357 /* complete a check operation */
2358 if (test_and_clear_bit(STRIPE_OP_CHECK, &sh->ops.complete)) {
2359 clear_bit(STRIPE_OP_CHECK, &sh->ops.ack);
2360 clear_bit(STRIPE_OP_CHECK, &sh->ops.pending);
2361 if (s->failed == 0) {
2362 if (sh->ops.zero_sum_result == 0)
2363 /* parity is correct (on disc,
2364 * not in buffer any more)
2366 set_bit(STRIPE_INSYNC, &sh->state);
2368 conf->mddev->resync_mismatches +=
2371 MD_RECOVERY_CHECK, &conf->mddev->recovery))
2372 /* don't try to repair!! */
2373 set_bit(STRIPE_INSYNC, &sh->state);
2375 set_bit(STRIPE_OP_COMPUTE_BLK,
2377 set_bit(STRIPE_OP_MOD_REPAIR_PD,
2379 set_bit(R5_Wantcompute,
2380 &sh->dev[sh->pd_idx].flags);
2381 sh->ops.target = sh->pd_idx;
2387 canceled_check = 1; /* STRIPE_INSYNC is not set */
2390 /* start a new check operation if there are no failures, the stripe is
2391 * not insync, and a repair is not in flight
2393 if (s->failed == 0 &&
2394 !test_bit(STRIPE_INSYNC, &sh->state) &&
2395 !test_bit(STRIPE_OP_MOD_REPAIR_PD, &sh->ops.pending)) {
2396 if (!test_and_set_bit(STRIPE_OP_CHECK, &sh->ops.pending)) {
2397 BUG_ON(s->uptodate != disks);
2398 clear_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags);
2404 /* check if we can clear a parity disk reconstruct */
2405 if (test_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.complete) &&
2406 test_bit(STRIPE_OP_MOD_REPAIR_PD, &sh->ops.pending)) {
2408 clear_bit(STRIPE_OP_MOD_REPAIR_PD, &sh->ops.pending);
2409 clear_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.complete);
2410 clear_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.ack);
2411 clear_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.pending);
2415 /* Wait for check parity and compute block operations to complete
2416 * before write-back. If a failure occurred while the check operation
2417 * was in flight we need to cycle this stripe through handle_stripe
2418 * since the parity block may not be uptodate
2420 if (!canceled_check && !test_bit(STRIPE_INSYNC, &sh->state) &&
2421 !test_bit(STRIPE_OP_CHECK, &sh->ops.pending) &&
2422 !test_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.pending)) {
2424 /* either failed parity check, or recovery is happening */
2426 s->failed_num = sh->pd_idx;
2427 dev = &sh->dev[s->failed_num];
2428 BUG_ON(!test_bit(R5_UPTODATE, &dev->flags));
2429 BUG_ON(s->uptodate != disks);
2431 set_bit(R5_LOCKED, &dev->flags);
2432 set_bit(R5_Wantwrite, &dev->flags);
2434 clear_bit(STRIPE_DEGRADED, &sh->state);
2436 set_bit(STRIPE_INSYNC, &sh->state);
2441 static void handle_parity_checks6(raid5_conf_t *conf, struct stripe_head *sh,
2442 struct stripe_head_state *s,
2443 struct r6_state *r6s, struct page *tmp_page,
2446 int update_p = 0, update_q = 0;
2448 int pd_idx = sh->pd_idx;
2449 int qd_idx = r6s->qd_idx;
2451 set_bit(STRIPE_HANDLE, &sh->state);
2453 BUG_ON(s->failed > 2);
2454 BUG_ON(s->uptodate < disks);
2455 /* Want to check and possibly repair P and Q.
2456 * However there could be one 'failed' device, in which
2457 * case we can only check one of them, possibly using the
2458 * other to generate missing data
2461 /* If !tmp_page, we cannot do the calculations,
2462 * but as we have set STRIPE_HANDLE, we will soon be called
2463 * by stripe_handle with a tmp_page - just wait until then.
2466 if (s->failed == r6s->q_failed) {
2467 /* The only possible failed device holds 'Q', so it
2468 * makes sense to check P (If anything else were failed,
2469 * we would have used P to recreate it).
2471 compute_block_1(sh, pd_idx, 1);
2472 if (!page_is_zero(sh->dev[pd_idx].page)) {
2473 compute_block_1(sh, pd_idx, 0);
2477 if (!r6s->q_failed && s->failed < 2) {
2478 /* q is not failed, and we didn't use it to generate
2479 * anything, so it makes sense to check it
2481 memcpy(page_address(tmp_page),
2482 page_address(sh->dev[qd_idx].page),
2484 compute_parity6(sh, UPDATE_PARITY);
2485 if (memcmp(page_address(tmp_page),
2486 page_address(sh->dev[qd_idx].page),
2487 STRIPE_SIZE) != 0) {
2488 clear_bit(STRIPE_INSYNC, &sh->state);
2492 if (update_p || update_q) {
2493 conf->mddev->resync_mismatches += STRIPE_SECTORS;
2494 if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery))
2495 /* don't try to repair!! */
2496 update_p = update_q = 0;
2499 /* now write out any block on a failed drive,
2500 * or P or Q if they need it
2503 if (s->failed == 2) {
2504 dev = &sh->dev[r6s->failed_num[1]];
2506 set_bit(R5_LOCKED, &dev->flags);
2507 set_bit(R5_Wantwrite, &dev->flags);
2509 if (s->failed >= 1) {
2510 dev = &sh->dev[r6s->failed_num[0]];
2512 set_bit(R5_LOCKED, &dev->flags);
2513 set_bit(R5_Wantwrite, &dev->flags);
2517 dev = &sh->dev[pd_idx];
2519 set_bit(R5_LOCKED, &dev->flags);
2520 set_bit(R5_Wantwrite, &dev->flags);
2523 dev = &sh->dev[qd_idx];
2525 set_bit(R5_LOCKED, &dev->flags);
2526 set_bit(R5_Wantwrite, &dev->flags);
2528 clear_bit(STRIPE_DEGRADED, &sh->state);
2530 set_bit(STRIPE_INSYNC, &sh->state);
2534 static void handle_stripe_expansion(raid5_conf_t *conf, struct stripe_head *sh,
2535 struct r6_state *r6s)
2539 /* We have read all the blocks in this stripe and now we need to
2540 * copy some of them into a target stripe for expand.
2542 struct dma_async_tx_descriptor *tx = NULL;
2543 clear_bit(STRIPE_EXPAND_SOURCE, &sh->state);
2544 for (i = 0; i < sh->disks; i++)
2545 if (i != sh->pd_idx && (!r6s || i != r6s->qd_idx)) {
2546 int dd_idx, pd_idx, j;
2547 struct stripe_head *sh2;
2549 sector_t bn = compute_blocknr(sh, i);
2550 sector_t s = raid5_compute_sector(bn, conf->raid_disks,
2552 conf->max_degraded, &dd_idx,
2554 sh2 = get_active_stripe(conf, s, conf->raid_disks,
2557 /* so far only the early blocks of this stripe
2558 * have been requested. When later blocks
2559 * get requested, we will try again
2562 if (!test_bit(STRIPE_EXPANDING, &sh2->state) ||
2563 test_bit(R5_Expanded, &sh2->dev[dd_idx].flags)) {
2564 /* must have already done this block */
2565 release_stripe(sh2);
2569 /* place all the copies on one channel */
2570 tx = async_memcpy(sh2->dev[dd_idx].page,
2571 sh->dev[i].page, 0, 0, STRIPE_SIZE,
2572 ASYNC_TX_DEP_ACK, tx, NULL, NULL);
2574 set_bit(R5_Expanded, &sh2->dev[dd_idx].flags);
2575 set_bit(R5_UPTODATE, &sh2->dev[dd_idx].flags);
2576 for (j = 0; j < conf->raid_disks; j++)
2577 if (j != sh2->pd_idx &&
2578 (!r6s || j != raid6_next_disk(sh2->pd_idx,
2580 !test_bit(R5_Expanded, &sh2->dev[j].flags))
2582 if (j == conf->raid_disks) {
2583 set_bit(STRIPE_EXPAND_READY, &sh2->state);
2584 set_bit(STRIPE_HANDLE, &sh2->state);
2586 release_stripe(sh2);
2589 /* done submitting copies, wait for them to complete */
2592 dma_wait_for_async_tx(tx);
2598 * handle_stripe - do things to a stripe.
2600 * We lock the stripe and then examine the state of various bits
2601 * to see what needs to be done.
2603 * return some read request which now have data
2604 * return some write requests which are safely on disc
2605 * schedule a read on some buffers
2606 * schedule a write of some buffers
2607 * return confirmation of parity correctness
2609 * buffers are taken off read_list or write_list, and bh_cache buffers
2610 * get BH_Lock set before the stripe lock is released.
2614 static void handle_stripe5(struct stripe_head *sh)
2616 raid5_conf_t *conf = sh->raid_conf;
2617 int disks = sh->disks, i;
2618 struct bio *return_bi = NULL;
2619 struct stripe_head_state s;
2621 unsigned long pending = 0;
2622 mdk_rdev_t *blocked_rdev = NULL;
2625 memset(&s, 0, sizeof(s));
2626 pr_debug("handling stripe %llu, state=%#lx cnt=%d, pd_idx=%d "
2627 "ops=%lx:%lx:%lx\n", (unsigned long long)sh->sector, sh->state,
2628 atomic_read(&sh->count), sh->pd_idx,
2629 sh->ops.pending, sh->ops.ack, sh->ops.complete);
2631 spin_lock(&sh->lock);
2632 clear_bit(STRIPE_HANDLE, &sh->state);
2633 clear_bit(STRIPE_DELAYED, &sh->state);
2635 s.syncing = test_bit(STRIPE_SYNCING, &sh->state);
2636 s.expanding = test_bit(STRIPE_EXPAND_SOURCE, &sh->state);
2637 s.expanded = test_bit(STRIPE_EXPAND_READY, &sh->state);
2638 /* Now to look around and see what can be done */
2640 /* clean-up completed biofill operations */
2641 if (test_bit(STRIPE_OP_BIOFILL, &sh->ops.complete)) {
2642 clear_bit(STRIPE_OP_BIOFILL, &sh->ops.pending);
2643 clear_bit(STRIPE_OP_BIOFILL, &sh->ops.ack);
2644 clear_bit(STRIPE_OP_BIOFILL, &sh->ops.complete);
2648 for (i=disks; i--; ) {
2650 struct r5dev *dev = &sh->dev[i];
2651 clear_bit(R5_Insync, &dev->flags);
2653 pr_debug("check %d: state 0x%lx toread %p read %p write %p "
2654 "written %p\n", i, dev->flags, dev->toread, dev->read,
2655 dev->towrite, dev->written);
2657 /* maybe we can request a biofill operation
2659 * new wantfill requests are only permitted while
2660 * STRIPE_OP_BIOFILL is clear
2662 if (test_bit(R5_UPTODATE, &dev->flags) && dev->toread &&
2663 !test_bit(STRIPE_OP_BIOFILL, &sh->ops.pending))
2664 set_bit(R5_Wantfill, &dev->flags);
2666 /* now count some things */
2667 if (test_bit(R5_LOCKED, &dev->flags)) s.locked++;
2668 if (test_bit(R5_UPTODATE, &dev->flags)) s.uptodate++;
2669 if (test_bit(R5_Wantcompute, &dev->flags)) s.compute++;
2671 if (test_bit(R5_Wantfill, &dev->flags))
2673 else if (dev->toread)
2677 if (!test_bit(R5_OVERWRITE, &dev->flags))
2682 rdev = rcu_dereference(conf->disks[i].rdev);
2683 if (rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
2684 blocked_rdev = rdev;
2685 atomic_inc(&rdev->nr_pending);
2688 if (!rdev || !test_bit(In_sync, &rdev->flags)) {
2689 /* The ReadError flag will just be confusing now */
2690 clear_bit(R5_ReadError, &dev->flags);
2691 clear_bit(R5_ReWrite, &dev->flags);
2693 if (!rdev || !test_bit(In_sync, &rdev->flags)
2694 || test_bit(R5_ReadError, &dev->flags)) {
2698 set_bit(R5_Insync, &dev->flags);
2702 if (unlikely(blocked_rdev)) {
2703 set_bit(STRIPE_HANDLE, &sh->state);
2707 if (s.to_fill && !test_and_set_bit(STRIPE_OP_BIOFILL, &sh->ops.pending))
2710 pr_debug("locked=%d uptodate=%d to_read=%d"
2711 " to_write=%d failed=%d failed_num=%d\n",
2712 s.locked, s.uptodate, s.to_read, s.to_write,
2713 s.failed, s.failed_num);
2714 /* check if the array has lost two devices and, if so, some requests might
2717 if (s.failed > 1 && s.to_read+s.to_write+s.written)
2718 handle_requests_to_failed_array(conf, sh, &s, disks,
2720 if (s.failed > 1 && s.syncing) {
2721 md_done_sync(conf->mddev, STRIPE_SECTORS,0);
2722 clear_bit(STRIPE_SYNCING, &sh->state);
2726 /* might be able to return some write requests if the parity block
2727 * is safe, or on a failed drive
2729 dev = &sh->dev[sh->pd_idx];
2731 ((test_bit(R5_Insync, &dev->flags) &&
2732 !test_bit(R5_LOCKED, &dev->flags) &&
2733 test_bit(R5_UPTODATE, &dev->flags)) ||
2734 (s.failed == 1 && s.failed_num == sh->pd_idx)))
2735 handle_completed_write_requests(conf, sh, disks, &return_bi);
2737 /* Now we might consider reading some blocks, either to check/generate
2738 * parity, or to satisfy requests
2739 * or to load a block that is being partially written.
2741 if (s.to_read || s.non_overwrite ||
2742 (s.syncing && (s.uptodate + s.compute < disks)) || s.expanding ||
2743 test_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.pending))
2744 handle_issuing_new_read_requests5(sh, &s, disks);
2746 /* Now we check to see if any write operations have recently
2750 /* leave prexor set until postxor is done, allows us to distinguish
2751 * a rmw from a rcw during biodrain
2754 if (test_bit(STRIPE_OP_PREXOR, &sh->ops.complete) &&
2755 test_bit(STRIPE_OP_POSTXOR, &sh->ops.complete)) {
2758 clear_bit(STRIPE_OP_PREXOR, &sh->ops.complete);
2759 clear_bit(STRIPE_OP_PREXOR, &sh->ops.ack);
2760 clear_bit(STRIPE_OP_PREXOR, &sh->ops.pending);
2762 for (i = disks; i--; )
2763 clear_bit(R5_Wantprexor, &sh->dev[i].flags);
2766 /* if only POSTXOR is set then this is an 'expand' postxor */
2767 if (test_bit(STRIPE_OP_BIODRAIN, &sh->ops.complete) &&
2768 test_bit(STRIPE_OP_POSTXOR, &sh->ops.complete)) {
2770 clear_bit(STRIPE_OP_BIODRAIN, &sh->ops.complete);
2771 clear_bit(STRIPE_OP_BIODRAIN, &sh->ops.ack);
2772 clear_bit(STRIPE_OP_BIODRAIN, &sh->ops.pending);
2774 clear_bit(STRIPE_OP_POSTXOR, &sh->ops.complete);
2775 clear_bit(STRIPE_OP_POSTXOR, &sh->ops.ack);
2776 clear_bit(STRIPE_OP_POSTXOR, &sh->ops.pending);
2778 /* All the 'written' buffers and the parity block are ready to
2779 * be written back to disk
2781 BUG_ON(!test_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags));
2782 for (i = disks; i--; ) {
2784 if (test_bit(R5_LOCKED, &dev->flags) &&
2785 (i == sh->pd_idx || dev->written)) {
2786 pr_debug("Writing block %d\n", i);
2787 set_bit(R5_Wantwrite, &dev->flags);
2790 if (!test_bit(R5_Insync, &dev->flags) ||
2791 (i == sh->pd_idx && s.failed == 0))
2792 set_bit(STRIPE_INSYNC, &sh->state);
2795 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2796 atomic_dec(&conf->preread_active_stripes);
2797 if (atomic_read(&conf->preread_active_stripes) <
2799 md_wakeup_thread(conf->mddev->thread);
2803 /* Now to consider new write requests and what else, if anything
2804 * should be read. We do not handle new writes when:
2805 * 1/ A 'write' operation (copy+xor) is already in flight.
2806 * 2/ A 'check' operation is in flight, as it may clobber the parity
2809 if (s.to_write && !test_bit(STRIPE_OP_POSTXOR, &sh->ops.pending) &&
2810 !test_bit(STRIPE_OP_CHECK, &sh->ops.pending))
2811 handle_issuing_new_write_requests5(conf, sh, &s, disks);
2813 /* maybe we need to check and possibly fix the parity for this stripe
2814 * Any reads will already have been scheduled, so we just see if enough
2815 * data is available. The parity check is held off while parity
2816 * dependent operations are in flight.
2818 if ((s.syncing && s.locked == 0 &&
2819 !test_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.pending) &&
2820 !test_bit(STRIPE_INSYNC, &sh->state)) ||
2821 test_bit(STRIPE_OP_CHECK, &sh->ops.pending) ||
2822 test_bit(STRIPE_OP_MOD_REPAIR_PD, &sh->ops.pending))
2823 handle_parity_checks5(conf, sh, &s, disks);
2825 if (s.syncing && s.locked == 0 && test_bit(STRIPE_INSYNC, &sh->state)) {
2826 md_done_sync(conf->mddev, STRIPE_SECTORS,1);
2827 clear_bit(STRIPE_SYNCING, &sh->state);
2830 /* If the failed drive is just a ReadError, then we might need to progress
2831 * the repair/check process
2833 if (s.failed == 1 && !conf->mddev->ro &&
2834 test_bit(R5_ReadError, &sh->dev[s.failed_num].flags)
2835 && !test_bit(R5_LOCKED, &sh->dev[s.failed_num].flags)
2836 && test_bit(R5_UPTODATE, &sh->dev[s.failed_num].flags)
2838 dev = &sh->dev[s.failed_num];
2839 if (!test_bit(R5_ReWrite, &dev->flags)) {
2840 set_bit(R5_Wantwrite, &dev->flags);
2841 set_bit(R5_ReWrite, &dev->flags);
2842 set_bit(R5_LOCKED, &dev->flags);
2845 /* let's read it back */
2846 set_bit(R5_Wantread, &dev->flags);
2847 set_bit(R5_LOCKED, &dev->flags);
2852 /* Finish postxor operations initiated by the expansion
2855 if (test_bit(STRIPE_OP_POSTXOR, &sh->ops.complete) &&
2856 !test_bit(STRIPE_OP_BIODRAIN, &sh->ops.pending)) {
2858 clear_bit(STRIPE_EXPANDING, &sh->state);
2860 clear_bit(STRIPE_OP_POSTXOR, &sh->ops.pending);
2861 clear_bit(STRIPE_OP_POSTXOR, &sh->ops.ack);
2862 clear_bit(STRIPE_OP_POSTXOR, &sh->ops.complete);
2864 for (i = conf->raid_disks; i--; )
2865 set_bit(R5_Wantwrite, &sh->dev[i].flags);
2866 set_bit(R5_LOCKED, &dev->flags);
2870 if (s.expanded && test_bit(STRIPE_EXPANDING, &sh->state) &&
2871 !test_bit(STRIPE_OP_POSTXOR, &sh->ops.pending)) {
2872 /* Need to write out all blocks after computing parity */
2873 sh->disks = conf->raid_disks;
2874 sh->pd_idx = stripe_to_pdidx(sh->sector, conf,
2876 s.locked += handle_write_operations5(sh, 1, 1);
2877 } else if (s.expanded &&
2879 !test_bit(STRIPE_OP_POSTXOR, &sh->ops.pending)) {
2880 clear_bit(STRIPE_EXPAND_READY, &sh->state);
2881 atomic_dec(&conf->reshape_stripes);
2882 wake_up(&conf->wait_for_overlap);
2883 md_done_sync(conf->mddev, STRIPE_SECTORS, 1);
2886 if (s.expanding && s.locked == 0 &&
2887 !test_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.pending))
2888 handle_stripe_expansion(conf, sh, NULL);
2891 pending = get_stripe_work(sh);
2894 spin_unlock(&sh->lock);
2896 /* wait for this device to become unblocked */
2897 if (unlikely(blocked_rdev))
2898 md_wait_for_blocked_rdev(blocked_rdev, conf->mddev);
2901 raid5_run_ops(sh, pending);
2905 return_io(return_bi);
2908 static void handle_stripe6(struct stripe_head *sh, struct page *tmp_page)
2910 raid6_conf_t *conf = sh->raid_conf;
2911 int disks = sh->disks;
2912 struct bio *return_bi = NULL;
2913 int i, pd_idx = sh->pd_idx;
2914 struct stripe_head_state s;
2915 struct r6_state r6s;
2916 struct r5dev *dev, *pdev, *qdev;
2917 mdk_rdev_t *blocked_rdev = NULL;
2919 r6s.qd_idx = raid6_next_disk(pd_idx, disks);
2920 pr_debug("handling stripe %llu, state=%#lx cnt=%d, "
2921 "pd_idx=%d, qd_idx=%d\n",
2922 (unsigned long long)sh->sector, sh->state,
2923 atomic_read(&sh->count), pd_idx, r6s.qd_idx);
2924 memset(&s, 0, sizeof(s));
2926 spin_lock(&sh->lock);
2927 clear_bit(STRIPE_HANDLE, &sh->state);
2928 clear_bit(STRIPE_DELAYED, &sh->state);
2930 s.syncing = test_bit(STRIPE_SYNCING, &sh->state);
2931 s.expanding = test_bit(STRIPE_EXPAND_SOURCE, &sh->state);
2932 s.expanded = test_bit(STRIPE_EXPAND_READY, &sh->state);
2933 /* Now to look around and see what can be done */
2936 for (i=disks; i--; ) {
2939 clear_bit(R5_Insync, &dev->flags);
2941 pr_debug("check %d: state 0x%lx read %p write %p written %p\n",
2942 i, dev->flags, dev->toread, dev->towrite, dev->written);
2943 /* maybe we can reply to a read */
2944 if (test_bit(R5_UPTODATE, &dev->flags) && dev->toread) {
2945 struct bio *rbi, *rbi2;
2946 pr_debug("Return read for disc %d\n", i);
2947 spin_lock_irq(&conf->device_lock);
2950 if (test_and_clear_bit(R5_Overlap, &dev->flags))
2951 wake_up(&conf->wait_for_overlap);
2952 spin_unlock_irq(&conf->device_lock);
2953 while (rbi && rbi->bi_sector < dev->sector + STRIPE_SECTORS) {
2954 copy_data(0, rbi, dev->page, dev->sector);
2955 rbi2 = r5_next_bio(rbi, dev->sector);
2956 spin_lock_irq(&conf->device_lock);
2957 if (--rbi->bi_phys_segments == 0) {
2958 rbi->bi_next = return_bi;
2961 spin_unlock_irq(&conf->device_lock);
2966 /* now count some things */
2967 if (test_bit(R5_LOCKED, &dev->flags)) s.locked++;
2968 if (test_bit(R5_UPTODATE, &dev->flags)) s.uptodate++;
2975 if (!test_bit(R5_OVERWRITE, &dev->flags))
2980 rdev = rcu_dereference(conf->disks[i].rdev);
2981 if (rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
2982 blocked_rdev = rdev;
2983 atomic_inc(&rdev->nr_pending);
2986 if (!rdev || !test_bit(In_sync, &rdev->flags)) {
2987 /* The ReadError flag will just be confusing now */
2988 clear_bit(R5_ReadError, &dev->flags);
2989 clear_bit(R5_ReWrite, &dev->flags);
2991 if (!rdev || !test_bit(In_sync, &rdev->flags)
2992 || test_bit(R5_ReadError, &dev->flags)) {
2994 r6s.failed_num[s.failed] = i;
2997 set_bit(R5_Insync, &dev->flags);
3001 if (unlikely(blocked_rdev)) {
3002 set_bit(STRIPE_HANDLE, &sh->state);
3005 pr_debug("locked=%d uptodate=%d to_read=%d"
3006 " to_write=%d failed=%d failed_num=%d,%d\n",
3007 s.locked, s.uptodate, s.to_read, s.to_write, s.failed,
3008 r6s.failed_num[0], r6s.failed_num[1]);
3009 /* check if the array has lost >2 devices and, if so, some requests
3010 * might need to be failed
3012 if (s.failed > 2 && s.to_read+s.to_write+s.written)
3013 handle_requests_to_failed_array(conf, sh, &s, disks,
3015 if (s.failed > 2 && s.syncing) {
3016 md_done_sync(conf->mddev, STRIPE_SECTORS,0);
3017 clear_bit(STRIPE_SYNCING, &sh->state);
3022 * might be able to return some write requests if the parity blocks
3023 * are safe, or on a failed drive
3025 pdev = &sh->dev[pd_idx];
3026 r6s.p_failed = (s.failed >= 1 && r6s.failed_num[0] == pd_idx)
3027 || (s.failed >= 2 && r6s.failed_num[1] == pd_idx);
3028 qdev = &sh->dev[r6s.qd_idx];
3029 r6s.q_failed = (s.failed >= 1 && r6s.failed_num[0] == r6s.qd_idx)
3030 || (s.failed >= 2 && r6s.failed_num[1] == r6s.qd_idx);
3033 ( r6s.p_failed || ((test_bit(R5_Insync, &pdev->flags)
3034 && !test_bit(R5_LOCKED, &pdev->flags)
3035 && test_bit(R5_UPTODATE, &pdev->flags)))) &&
3036 ( r6s.q_failed || ((test_bit(R5_Insync, &qdev->flags)
3037 && !test_bit(R5_LOCKED, &qdev->flags)
3038 && test_bit(R5_UPTODATE, &qdev->flags)))))
3039 handle_completed_write_requests(conf, sh, disks, &return_bi);
3041 /* Now we might consider reading some blocks, either to check/generate
3042 * parity, or to satisfy requests
3043 * or to load a block that is being partially written.
3045 if (s.to_read || s.non_overwrite || (s.to_write && s.failed) ||
3046 (s.syncing && (s.uptodate < disks)) || s.expanding)
3047 handle_issuing_new_read_requests6(sh, &s, &r6s, disks);
3049 /* now to consider writing and what else, if anything should be read */
3051 handle_issuing_new_write_requests6(conf, sh, &s, &r6s, disks);
3053 /* maybe we need to check and possibly fix the parity for this stripe
3054 * Any reads will already have been scheduled, so we just see if enough
3057 if (s.syncing && s.locked == 0 && !test_bit(STRIPE_INSYNC, &sh->state))
3058 handle_parity_checks6(conf, sh, &s, &r6s, tmp_page, disks);
3060 if (s.syncing && s.locked == 0 && test_bit(STRIPE_INSYNC, &sh->state)) {
3061 md_done_sync(conf->mddev, STRIPE_SECTORS,1);
3062 clear_bit(STRIPE_SYNCING, &sh->state);
3065 /* If the failed drives are just a ReadError, then we might need
3066 * to progress the repair/check process
3068 if (s.failed <= 2 && !conf->mddev->ro)
3069 for (i = 0; i < s.failed; i++) {
3070 dev = &sh->dev[r6s.failed_num[i]];
3071 if (test_bit(R5_ReadError, &dev->flags)
3072 && !test_bit(R5_LOCKED, &dev->flags)
3073 && test_bit(R5_UPTODATE, &dev->flags)
3075 if (!test_bit(R5_ReWrite, &dev->flags)) {
3076 set_bit(R5_Wantwrite, &dev->flags);
3077 set_bit(R5_ReWrite, &dev->flags);
3078 set_bit(R5_LOCKED, &dev->flags);
3080 /* let's read it back */
3081 set_bit(R5_Wantread, &dev->flags);
3082 set_bit(R5_LOCKED, &dev->flags);
3087 if (s.expanded && test_bit(STRIPE_EXPANDING, &sh->state)) {
3088 /* Need to write out all blocks after computing P&Q */
3089 sh->disks = conf->raid_disks;
3090 sh->pd_idx = stripe_to_pdidx(sh->sector, conf,
3092 compute_parity6(sh, RECONSTRUCT_WRITE);
3093 for (i = conf->raid_disks ; i-- ; ) {
3094 set_bit(R5_LOCKED, &sh->dev[i].flags);
3096 set_bit(R5_Wantwrite, &sh->dev[i].flags);
3098 clear_bit(STRIPE_EXPANDING, &sh->state);
3099 } else if (s.expanded) {
3100 clear_bit(STRIPE_EXPAND_READY, &sh->state);
3101 atomic_dec(&conf->reshape_stripes);
3102 wake_up(&conf->wait_for_overlap);
3103 md_done_sync(conf->mddev, STRIPE_SECTORS, 1);
3106 if (s.expanding && s.locked == 0 &&
3107 !test_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.pending))
3108 handle_stripe_expansion(conf, sh, &r6s);
3111 spin_unlock(&sh->lock);
3113 /* wait for this device to become unblocked */
3114 if (unlikely(blocked_rdev))
3115 md_wait_for_blocked_rdev(blocked_rdev, conf->mddev);
3119 return_io(return_bi);
3122 static void handle_stripe(struct stripe_head *sh, struct page *tmp_page)
3124 if (sh->raid_conf->level == 6)
3125 handle_stripe6(sh, tmp_page);
3132 static void raid5_activate_delayed(raid5_conf_t *conf)
3134 if (atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD) {
3135 while (!list_empty(&conf->delayed_list)) {
3136 struct list_head *l = conf->delayed_list.next;
3137 struct stripe_head *sh;
3138 sh = list_entry(l, struct stripe_head, lru);
3140 clear_bit(STRIPE_DELAYED, &sh->state);
3141 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
3142 atomic_inc(&conf->preread_active_stripes);
3143 list_add_tail(&sh->lru, &conf->hold_list);
3146 blk_plug_device(conf->mddev->queue);
3149 static void activate_bit_delay(raid5_conf_t *conf)
3151 /* device_lock is held */
3152 struct list_head head;
3153 list_add(&head, &conf->bitmap_list);
3154 list_del_init(&conf->bitmap_list);
3155 while (!list_empty(&head)) {
3156 struct stripe_head *sh = list_entry(head.next, struct stripe_head, lru);
3157 list_del_init(&sh->lru);
3158 atomic_inc(&sh->count);
3159 __release_stripe(conf, sh);
3163 static void unplug_slaves(mddev_t *mddev)
3165 raid5_conf_t *conf = mddev_to_conf(mddev);
3169 for (i=0; i<mddev->raid_disks; i++) {
3170 mdk_rdev_t *rdev = rcu_dereference(conf->disks[i].rdev);
3171 if (rdev && !test_bit(Faulty, &rdev->flags) && atomic_read(&rdev->nr_pending)) {
3172 struct request_queue *r_queue = bdev_get_queue(rdev->bdev);
3174 atomic_inc(&rdev->nr_pending);
3177 blk_unplug(r_queue);
3179 rdev_dec_pending(rdev, mddev);
3186 static void raid5_unplug_device(struct request_queue *q)
3188 mddev_t *mddev = q->queuedata;
3189 raid5_conf_t *conf = mddev_to_conf(mddev);
3190 unsigned long flags;
3192 spin_lock_irqsave(&conf->device_lock, flags);
3194 if (blk_remove_plug(q)) {
3196 raid5_activate_delayed(conf);
3198 md_wakeup_thread(mddev->thread);
3200 spin_unlock_irqrestore(&conf->device_lock, flags);
3202 unplug_slaves(mddev);
3205 static int raid5_congested(void *data, int bits)
3207 mddev_t *mddev = data;
3208 raid5_conf_t *conf = mddev_to_conf(mddev);
3210 /* No difference between reads and writes. Just check
3211 * how busy the stripe_cache is
3213 if (conf->inactive_blocked)
3217 if (list_empty_careful(&conf->inactive_list))
3223 /* We want read requests to align with chunks where possible,
3224 * but write requests don't need to.
3226 static int raid5_mergeable_bvec(struct request_queue *q, struct bio *bio, struct bio_vec *biovec)
3228 mddev_t *mddev = q->queuedata;
3229 sector_t sector = bio->bi_sector + get_start_sect(bio->bi_bdev);
3231 unsigned int chunk_sectors = mddev->chunk_size >> 9;
3232 unsigned int bio_sectors = bio->bi_size >> 9;
3234 if (bio_data_dir(bio) == WRITE)
3235 return biovec->bv_len; /* always allow writes to be mergeable */
3237 max = (chunk_sectors - ((sector & (chunk_sectors - 1)) + bio_sectors)) << 9;
3238 if (max < 0) max = 0;
3239 if (max <= biovec->bv_len && bio_sectors == 0)
3240 return biovec->bv_len;
3246 static int in_chunk_boundary(mddev_t *mddev, struct bio *bio)
3248 sector_t sector = bio->bi_sector + get_start_sect(bio->bi_bdev);
3249 unsigned int chunk_sectors = mddev->chunk_size >> 9;
3250 unsigned int bio_sectors = bio->bi_size >> 9;
3252 return chunk_sectors >=
3253 ((sector & (chunk_sectors - 1)) + bio_sectors);
3257 * add bio to the retry LIFO ( in O(1) ... we are in interrupt )
3258 * later sampled by raid5d.
3260 static void add_bio_to_retry(struct bio *bi,raid5_conf_t *conf)
3262 unsigned long flags;
3264 spin_lock_irqsave(&conf->device_lock, flags);
3266 bi->bi_next = conf->retry_read_aligned_list;
3267 conf->retry_read_aligned_list = bi;
3269 spin_unlock_irqrestore(&conf->device_lock, flags);
3270 md_wakeup_thread(conf->mddev->thread);
3274 static struct bio *remove_bio_from_retry(raid5_conf_t *conf)
3278 bi = conf->retry_read_aligned;
3280 conf->retry_read_aligned = NULL;
3283 bi = conf->retry_read_aligned_list;
3285 conf->retry_read_aligned_list = bi->bi_next;
3287 bi->bi_phys_segments = 1; /* biased count of active stripes */
3288 bi->bi_hw_segments = 0; /* count of processed stripes */
3296 * The "raid5_align_endio" should check if the read succeeded and if it
3297 * did, call bio_endio on the original bio (having bio_put the new bio
3299 * If the read failed..
3301 static void raid5_align_endio(struct bio *bi, int error)
3303 struct bio* raid_bi = bi->bi_private;
3306 int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
3311 mddev = raid_bi->bi_bdev->bd_disk->queue->queuedata;
3312 conf = mddev_to_conf(mddev);
3313 rdev = (void*)raid_bi->bi_next;
3314 raid_bi->bi_next = NULL;
3316 rdev_dec_pending(rdev, conf->mddev);
3318 if (!error && uptodate) {
3319 bio_endio(raid_bi, 0);
3320 if (atomic_dec_and_test(&conf->active_aligned_reads))
3321 wake_up(&conf->wait_for_stripe);
3326 pr_debug("raid5_align_endio : io error...handing IO for a retry\n");
3328 add_bio_to_retry(raid_bi, conf);
3331 static int bio_fits_rdev(struct bio *bi)
3333 struct request_queue *q = bdev_get_queue(bi->bi_bdev);
3335 if ((bi->bi_size>>9) > q->max_sectors)
3337 blk_recount_segments(q, bi);
3338 if (bi->bi_phys_segments > q->max_phys_segments ||
3339 bi->bi_hw_segments > q->max_hw_segments)
3342 if (q->merge_bvec_fn)
3343 /* it's too hard to apply the merge_bvec_fn at this stage,
3352 static int chunk_aligned_read(struct request_queue *q, struct bio * raid_bio)
3354 mddev_t *mddev = q->queuedata;
3355 raid5_conf_t *conf = mddev_to_conf(mddev);
3356 const unsigned int raid_disks = conf->raid_disks;
3357 const unsigned int data_disks = raid_disks - conf->max_degraded;
3358 unsigned int dd_idx, pd_idx;
3359 struct bio* align_bi;
3362 if (!in_chunk_boundary(mddev, raid_bio)) {
3363 pr_debug("chunk_aligned_read : non aligned\n");
3367 * use bio_clone to make a copy of the bio
3369 align_bi = bio_clone(raid_bio, GFP_NOIO);
3373 * set bi_end_io to a new function, and set bi_private to the
3376 align_bi->bi_end_io = raid5_align_endio;
3377 align_bi->bi_private = raid_bio;
3381 align_bi->bi_sector = raid5_compute_sector(raid_bio->bi_sector,
3389 rdev = rcu_dereference(conf->disks[dd_idx].rdev);
3390 if (rdev && test_bit(In_sync, &rdev->flags)) {
3391 atomic_inc(&rdev->nr_pending);
3393 raid_bio->bi_next = (void*)rdev;
3394 align_bi->bi_bdev = rdev->bdev;
3395 align_bi->bi_flags &= ~(1 << BIO_SEG_VALID);
3396 align_bi->bi_sector += rdev->data_offset;
3398 if (!bio_fits_rdev(align_bi)) {
3399 /* too big in some way */
3401 rdev_dec_pending(rdev, mddev);
3405 spin_lock_irq(&conf->device_lock);
3406 wait_event_lock_irq(conf->wait_for_stripe,
3408 conf->device_lock, /* nothing */);
3409 atomic_inc(&conf->active_aligned_reads);
3410 spin_unlock_irq(&conf->device_lock);
3412 generic_make_request(align_bi);
3421 /* __get_priority_stripe - get the next stripe to process
3423 * Full stripe writes are allowed to pass preread active stripes up until
3424 * the bypass_threshold is exceeded. In general the bypass_count
3425 * increments when the handle_list is handled before the hold_list; however, it
3426 * will not be incremented when STRIPE_IO_STARTED is sampled set signifying a
3427 * stripe with in flight i/o. The bypass_count will be reset when the
3428 * head of the hold_list has changed, i.e. the head was promoted to the
3431 static struct stripe_head *__get_priority_stripe(raid5_conf_t *conf)
3433 struct stripe_head *sh;
3435 pr_debug("%s: handle: %s hold: %s full_writes: %d bypass_count: %d\n",
3437 list_empty(&conf->handle_list) ? "empty" : "busy",
3438 list_empty(&conf->hold_list) ? "empty" : "busy",
3439 atomic_read(&conf->pending_full_writes), conf->bypass_count);
3441 if (!list_empty(&conf->handle_list)) {
3442 sh = list_entry(conf->handle_list.next, typeof(*sh), lru);
3444 if (list_empty(&conf->hold_list))
3445 conf->bypass_count = 0;
3446 else if (!test_bit(STRIPE_IO_STARTED, &sh->state)) {
3447 if (conf->hold_list.next == conf->last_hold)
3448 conf->bypass_count++;
3450 conf->last_hold = conf->hold_list.next;
3451 conf->bypass_count -= conf->bypass_threshold;
3452 if (conf->bypass_count < 0)
3453 conf->bypass_count = 0;
3456 } else if (!list_empty(&conf->hold_list) &&
3457 ((conf->bypass_threshold &&
3458 conf->bypass_count > conf->bypass_threshold) ||
3459 atomic_read(&conf->pending_full_writes) == 0)) {
3460 sh = list_entry(conf->hold_list.next,
3462 conf->bypass_count -= conf->bypass_threshold;
3463 if (conf->bypass_count < 0)
3464 conf->bypass_count = 0;
3468 list_del_init(&sh->lru);
3469 atomic_inc(&sh->count);
3470 BUG_ON(atomic_read(&sh->count) != 1);
3474 static int make_request(struct request_queue *q, struct bio * bi)
3476 mddev_t *mddev = q->queuedata;
3477 raid5_conf_t *conf = mddev_to_conf(mddev);
3478 unsigned int dd_idx, pd_idx;
3479 sector_t new_sector;
3480 sector_t logical_sector, last_sector;
3481 struct stripe_head *sh;
3482 const int rw = bio_data_dir(bi);
3485 if (unlikely(bio_barrier(bi))) {
3486 bio_endio(bi, -EOPNOTSUPP);
3490 md_write_start(mddev, bi);
3492 disk_stat_inc(mddev->gendisk, ios[rw]);
3493 disk_stat_add(mddev->gendisk, sectors[rw], bio_sectors(bi));
3496 mddev->reshape_position == MaxSector &&
3497 chunk_aligned_read(q,bi))
3500 logical_sector = bi->bi_sector & ~((sector_t)STRIPE_SECTORS-1);
3501 last_sector = bi->bi_sector + (bi->bi_size>>9);
3503 bi->bi_phys_segments = 1; /* over-loaded to count active stripes */
3505 for (;logical_sector < last_sector; logical_sector += STRIPE_SECTORS) {
3507 int disks, data_disks;
3510 prepare_to_wait(&conf->wait_for_overlap, &w, TASK_UNINTERRUPTIBLE);
3511 if (likely(conf->expand_progress == MaxSector))
3512 disks = conf->raid_disks;
3514 /* spinlock is needed as expand_progress may be
3515 * 64bit on a 32bit platform, and so it might be
3516 * possible to see a half-updated value
3517 * Ofcourse expand_progress could change after
3518 * the lock is dropped, so once we get a reference
3519 * to the stripe that we think it is, we will have
3522 spin_lock_irq(&conf->device_lock);
3523 disks = conf->raid_disks;
3524 if (logical_sector >= conf->expand_progress)
3525 disks = conf->previous_raid_disks;
3527 if (logical_sector >= conf->expand_lo) {
3528 spin_unlock_irq(&conf->device_lock);
3533 spin_unlock_irq(&conf->device_lock);
3535 data_disks = disks - conf->max_degraded;
3537 new_sector = raid5_compute_sector(logical_sector, disks, data_disks,
3538 &dd_idx, &pd_idx, conf);
3539 pr_debug("raid5: make_request, sector %llu logical %llu\n",
3540 (unsigned long long)new_sector,
3541 (unsigned long long)logical_sector);
3543 sh = get_active_stripe(conf, new_sector, disks, pd_idx, (bi->bi_rw&RWA_MASK));
3545 if (unlikely(conf->expand_progress != MaxSector)) {
3546 /* expansion might have moved on while waiting for a
3547 * stripe, so we must do the range check again.
3548 * Expansion could still move past after this
3549 * test, but as we are holding a reference to
3550 * 'sh', we know that if that happens,
3551 * STRIPE_EXPANDING will get set and the expansion
3552 * won't proceed until we finish with the stripe.
3555 spin_lock_irq(&conf->device_lock);
3556 if (logical_sector < conf->expand_progress &&
3557 disks == conf->previous_raid_disks)
3558 /* mismatch, need to try again */
3560 spin_unlock_irq(&conf->device_lock);
3566 /* FIXME what if we get a false positive because these
3567 * are being updated.
3569 if (logical_sector >= mddev->suspend_lo &&
3570 logical_sector < mddev->suspend_hi) {
3576 if (test_bit(STRIPE_EXPANDING, &sh->state) ||
3577 !add_stripe_bio(sh, bi, dd_idx, (bi->bi_rw&RW_MASK))) {
3578 /* Stripe is busy expanding or
3579 * add failed due to overlap. Flush everything
3582 raid5_unplug_device(mddev->queue);
3587 finish_wait(&conf->wait_for_overlap, &w);
3588 set_bit(STRIPE_HANDLE, &sh->state);
3589 clear_bit(STRIPE_DELAYED, &sh->state);
3592 /* cannot get stripe for read-ahead, just give-up */
3593 clear_bit(BIO_UPTODATE, &bi->bi_flags);
3594 finish_wait(&conf->wait_for_overlap, &w);
3599 spin_lock_irq(&conf->device_lock);
3600 remaining = --bi->bi_phys_segments;
3601 spin_unlock_irq(&conf->device_lock);
3602 if (remaining == 0) {
3605 md_write_end(mddev);
3612 static sector_t reshape_request(mddev_t *mddev, sector_t sector_nr, int *skipped)
3614 /* reshaping is quite different to recovery/resync so it is
3615 * handled quite separately ... here.
3617 * On each call to sync_request, we gather one chunk worth of
3618 * destination stripes and flag them as expanding.
3619 * Then we find all the source stripes and request reads.
3620 * As the reads complete, handle_stripe will copy the data
3621 * into the destination stripe and release that stripe.
3623 raid5_conf_t *conf = (raid5_conf_t *) mddev->private;
3624 struct stripe_head *sh;
3626 sector_t first_sector, last_sector;
3627 int raid_disks = conf->previous_raid_disks;
3628 int data_disks = raid_disks - conf->max_degraded;
3629 int new_data_disks = conf->raid_disks - conf->max_degraded;
3632 sector_t writepos, safepos, gap;
3634 if (sector_nr == 0 &&
3635 conf->expand_progress != 0) {
3636 /* restarting in the middle, skip the initial sectors */
3637 sector_nr = conf->expand_progress;
3638 sector_div(sector_nr, new_data_disks);
3643 /* we update the metadata when there is more than 3Meg
3644 * in the block range (that is rather arbitrary, should
3645 * probably be time based) or when the data about to be
3646 * copied would over-write the source of the data at
3647 * the front of the range.
3648 * i.e. one new_stripe forward from expand_progress new_maps
3649 * to after where expand_lo old_maps to
3651 writepos = conf->expand_progress +
3652 conf->chunk_size/512*(new_data_disks);
3653 sector_div(writepos, new_data_disks);
3654 safepos = conf->expand_lo;
3655 sector_div(safepos, data_disks);
3656 gap = conf->expand_progress - conf->expand_lo;
3658 if (writepos >= safepos ||
3659 gap > (new_data_disks)*3000*2 /*3Meg*/) {
3660 /* Cannot proceed until we've updated the superblock... */
3661 wait_event(conf->wait_for_overlap,
3662 atomic_read(&conf->reshape_stripes)==0);
3663 mddev->reshape_position = conf->expand_progress;
3664 set_bit(MD_CHANGE_DEVS, &mddev->flags);
3665 md_wakeup_thread(mddev->thread);
3666 wait_event(mddev->sb_wait, mddev->flags == 0 ||
3667 kthread_should_stop());
3668 spin_lock_irq(&conf->device_lock);
3669 conf->expand_lo = mddev->reshape_position;
3670 spin_unlock_irq(&conf->device_lock);
3671 wake_up(&conf->wait_for_overlap);
3674 for (i=0; i < conf->chunk_size/512; i+= STRIPE_SECTORS) {
3677 pd_idx = stripe_to_pdidx(sector_nr+i, conf, conf->raid_disks);
3678 sh = get_active_stripe(conf, sector_nr+i,
3679 conf->raid_disks, pd_idx, 0);
3680 set_bit(STRIPE_EXPANDING, &sh->state);
3681 atomic_inc(&conf->reshape_stripes);
3682 /* If any of this stripe is beyond the end of the old
3683 * array, then we need to zero those blocks
3685 for (j=sh->disks; j--;) {
3687 if (j == sh->pd_idx)
3689 if (conf->level == 6 &&
3690 j == raid6_next_disk(sh->pd_idx, sh->disks))
3692 s = compute_blocknr(sh, j);
3693 if (s < (mddev->array_size<<1)) {
3697 memset(page_address(sh->dev[j].page), 0, STRIPE_SIZE);
3698 set_bit(R5_Expanded, &sh->dev[j].flags);
3699 set_bit(R5_UPTODATE, &sh->dev[j].flags);
3702 set_bit(STRIPE_EXPAND_READY, &sh->state);
3703 set_bit(STRIPE_HANDLE, &sh->state);
3707 spin_lock_irq(&conf->device_lock);
3708 conf->expand_progress = (sector_nr + i) * new_data_disks;
3709 spin_unlock_irq(&conf->device_lock);
3710 /* Ok, those stripe are ready. We can start scheduling
3711 * reads on the source stripes.
3712 * The source stripes are determined by mapping the first and last
3713 * block on the destination stripes.
3716 raid5_compute_sector(sector_nr*(new_data_disks),
3717 raid_disks, data_disks,
3718 &dd_idx, &pd_idx, conf);
3720 raid5_compute_sector((sector_nr+conf->chunk_size/512)
3721 *(new_data_disks) -1,
3722 raid_disks, data_disks,
3723 &dd_idx, &pd_idx, conf);
3724 if (last_sector >= (mddev->size<<1))
3725 last_sector = (mddev->size<<1)-1;
3726 while (first_sector <= last_sector) {
3727 pd_idx = stripe_to_pdidx(first_sector, conf,
3728 conf->previous_raid_disks);
3729 sh = get_active_stripe(conf, first_sector,
3730 conf->previous_raid_disks, pd_idx, 0);
3731 set_bit(STRIPE_EXPAND_SOURCE, &sh->state);
3732 set_bit(STRIPE_HANDLE, &sh->state);
3734 first_sector += STRIPE_SECTORS;
3736 /* If this takes us to the resync_max point where we have to pause,
3737 * then we need to write out the superblock.
3739 sector_nr += conf->chunk_size>>9;
3740 if (sector_nr >= mddev->resync_max) {
3741 /* Cannot proceed until we've updated the superblock... */
3742 wait_event(conf->wait_for_overlap,
3743 atomic_read(&conf->reshape_stripes) == 0);
3744 mddev->reshape_position = conf->expand_progress;
3745 set_bit(MD_CHANGE_DEVS, &mddev->flags);
3746 md_wakeup_thread(mddev->thread);
3747 wait_event(mddev->sb_wait,
3748 !test_bit(MD_CHANGE_DEVS, &mddev->flags)
3749 || kthread_should_stop());
3750 spin_lock_irq(&conf->device_lock);
3751 conf->expand_lo = mddev->reshape_position;
3752 spin_unlock_irq(&conf->device_lock);
3753 wake_up(&conf->wait_for_overlap);
3755 return conf->chunk_size>>9;
3758 /* FIXME go_faster isn't used */
3759 static inline sector_t sync_request(mddev_t *mddev, sector_t sector_nr, int *skipped, int go_faster)
3761 raid5_conf_t *conf = (raid5_conf_t *) mddev->private;
3762 struct stripe_head *sh;
3764 int raid_disks = conf->raid_disks;
3765 sector_t max_sector = mddev->size << 1;
3767 int still_degraded = 0;
3770 if (sector_nr >= max_sector) {
3771 /* just being told to finish up .. nothing much to do */
3772 unplug_slaves(mddev);
3773 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery)) {
3778 if (mddev->curr_resync < max_sector) /* aborted */
3779 bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
3781 else /* completed sync */
3783 bitmap_close_sync(mddev->bitmap);
3788 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
3789 return reshape_request(mddev, sector_nr, skipped);
3791 /* No need to check resync_max as we never do more than one
3792 * stripe, and as resync_max will always be on a chunk boundary,
3793 * if the check in md_do_sync didn't fire, there is no chance
3794 * of overstepping resync_max here
3797 /* if there is too many failed drives and we are trying
3798 * to resync, then assert that we are finished, because there is
3799 * nothing we can do.
3801 if (mddev->degraded >= conf->max_degraded &&
3802 test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
3803 sector_t rv = (mddev->size << 1) - sector_nr;
3807 if (!bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
3808 !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
3809 !conf->fullsync && sync_blocks >= STRIPE_SECTORS) {
3810 /* we can skip this block, and probably more */
3811 sync_blocks /= STRIPE_SECTORS;
3813 return sync_blocks * STRIPE_SECTORS; /* keep things rounded to whole stripes */
3817 bitmap_cond_end_sync(mddev->bitmap, sector_nr);
3819 pd_idx = stripe_to_pdidx(sector_nr, conf, raid_disks);
3820 sh = get_active_stripe(conf, sector_nr, raid_disks, pd_idx, 1);
3822 sh = get_active_stripe(conf, sector_nr, raid_disks, pd_idx, 0);
3823 /* make sure we don't swamp the stripe cache if someone else
3824 * is trying to get access
3826 schedule_timeout_uninterruptible(1);
3828 /* Need to check if array will still be degraded after recovery/resync
3829 * We don't need to check the 'failed' flag as when that gets set,
3832 for (i=0; i<mddev->raid_disks; i++)
3833 if (conf->disks[i].rdev == NULL)
3836 bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, still_degraded);
3838 spin_lock(&sh->lock);
3839 set_bit(STRIPE_SYNCING, &sh->state);
3840 clear_bit(STRIPE_INSYNC, &sh->state);
3841 spin_unlock(&sh->lock);
3843 handle_stripe(sh, NULL);
3846 return STRIPE_SECTORS;
3849 static int retry_aligned_read(raid5_conf_t *conf, struct bio *raid_bio)
3851 /* We may not be able to submit a whole bio at once as there
3852 * may not be enough stripe_heads available.
3853 * We cannot pre-allocate enough stripe_heads as we may need
3854 * more than exist in the cache (if we allow ever large chunks).
3855 * So we do one stripe head at a time and record in
3856 * ->bi_hw_segments how many have been done.
3858 * We *know* that this entire raid_bio is in one chunk, so
3859 * it will be only one 'dd_idx' and only need one call to raid5_compute_sector.
3861 struct stripe_head *sh;
3863 sector_t sector, logical_sector, last_sector;
3868 logical_sector = raid_bio->bi_sector & ~((sector_t)STRIPE_SECTORS-1);
3869 sector = raid5_compute_sector( logical_sector,
3871 conf->raid_disks - conf->max_degraded,
3875 last_sector = raid_bio->bi_sector + (raid_bio->bi_size>>9);
3877 for (; logical_sector < last_sector;
3878 logical_sector += STRIPE_SECTORS,
3879 sector += STRIPE_SECTORS,
3882 if (scnt < raid_bio->bi_hw_segments)
3883 /* already done this stripe */
3886 sh = get_active_stripe(conf, sector, conf->raid_disks, pd_idx, 1);
3889 /* failed to get a stripe - must wait */
3890 raid_bio->bi_hw_segments = scnt;
3891 conf->retry_read_aligned = raid_bio;
3895 set_bit(R5_ReadError, &sh->dev[dd_idx].flags);
3896 if (!add_stripe_bio(sh, raid_bio, dd_idx, 0)) {
3898 raid_bio->bi_hw_segments = scnt;
3899 conf->retry_read_aligned = raid_bio;
3903 handle_stripe(sh, NULL);
3907 spin_lock_irq(&conf->device_lock);
3908 remaining = --raid_bio->bi_phys_segments;
3909 spin_unlock_irq(&conf->device_lock);
3911 bio_endio(raid_bio, 0);
3912 if (atomic_dec_and_test(&conf->active_aligned_reads))
3913 wake_up(&conf->wait_for_stripe);
3920 * This is our raid5 kernel thread.
3922 * We scan the hash table for stripes which can be handled now.
3923 * During the scan, completed stripes are saved for us by the interrupt
3924 * handler, so that they will not have to wait for our next wakeup.
3926 static void raid5d(mddev_t *mddev)
3928 struct stripe_head *sh;
3929 raid5_conf_t *conf = mddev_to_conf(mddev);
3932 pr_debug("+++ raid5d active\n");
3934 md_check_recovery(mddev);
3937 spin_lock_irq(&conf->device_lock);
3941 if (conf->seq_flush != conf->seq_write) {
3942 int seq = conf->seq_flush;
3943 spin_unlock_irq(&conf->device_lock);
3944 bitmap_unplug(mddev->bitmap);
3945 spin_lock_irq(&conf->device_lock);
3946 conf->seq_write = seq;
3947 activate_bit_delay(conf);
3950 while ((bio = remove_bio_from_retry(conf))) {
3952 spin_unlock_irq(&conf->device_lock);
3953 ok = retry_aligned_read(conf, bio);
3954 spin_lock_irq(&conf->device_lock);
3960 sh = __get_priority_stripe(conf);
3963 async_tx_issue_pending_all();
3966 spin_unlock_irq(&conf->device_lock);
3969 handle_stripe(sh, conf->spare_page);
3972 spin_lock_irq(&conf->device_lock);
3974 pr_debug("%d stripes handled\n", handled);
3976 spin_unlock_irq(&conf->device_lock);
3978 unplug_slaves(mddev);
3980 pr_debug("--- raid5d inactive\n");
3984 raid5_show_stripe_cache_size(mddev_t *mddev, char *page)
3986 raid5_conf_t *conf = mddev_to_conf(mddev);
3988 return sprintf(page, "%d\n", conf->max_nr_stripes);
3994 raid5_store_stripe_cache_size(mddev_t *mddev, const char *page, size_t len)
3996 raid5_conf_t *conf = mddev_to_conf(mddev);
3998 if (len >= PAGE_SIZE)
4003 if (strict_strtoul(page, 10, &new))
4005 if (new <= 16 || new > 32768)
4007 while (new < conf->max_nr_stripes) {
4008 if (drop_one_stripe(conf))
4009 conf->max_nr_stripes--;
4013 md_allow_write(mddev);
4014 while (new > conf->max_nr_stripes) {
4015 if (grow_one_stripe(conf))
4016 conf->max_nr_stripes++;
4022 static struct md_sysfs_entry
4023 raid5_stripecache_size = __ATTR(stripe_cache_size, S_IRUGO | S_IWUSR,
4024 raid5_show_stripe_cache_size,
4025 raid5_store_stripe_cache_size);
4028 raid5_show_preread_threshold(mddev_t *mddev, char *page)
4030 raid5_conf_t *conf = mddev_to_conf(mddev);
4032 return sprintf(page, "%d\n", conf->bypass_threshold);
4038 raid5_store_preread_threshold(mddev_t *mddev, const char *page, size_t len)
4040 raid5_conf_t *conf = mddev_to_conf(mddev);
4042 if (len >= PAGE_SIZE)
4047 if (strict_strtoul(page, 10, &new))
4049 if (new > conf->max_nr_stripes)
4051 conf->bypass_threshold = new;
4055 static struct md_sysfs_entry
4056 raid5_preread_bypass_threshold = __ATTR(preread_bypass_threshold,
4058 raid5_show_preread_threshold,
4059 raid5_store_preread_threshold);
4062 stripe_cache_active_show(mddev_t *mddev, char *page)
4064 raid5_conf_t *conf = mddev_to_conf(mddev);
4066 return sprintf(page, "%d\n", atomic_read(&conf->active_stripes));
4071 static struct md_sysfs_entry
4072 raid5_stripecache_active = __ATTR_RO(stripe_cache_active);
4074 static struct attribute *raid5_attrs[] = {
4075 &raid5_stripecache_size.attr,
4076 &raid5_stripecache_active.attr,
4077 &raid5_preread_bypass_threshold.attr,
4080 static struct attribute_group raid5_attrs_group = {
4082 .attrs = raid5_attrs,
4085 static int run(mddev_t *mddev)
4088 int raid_disk, memory;
4090 struct disk_info *disk;
4091 struct list_head *tmp;
4092 int working_disks = 0;
4094 if (mddev->level != 5 && mddev->level != 4 && mddev->level != 6) {
4095 printk(KERN_ERR "raid5: %s: raid level not set to 4/5/6 (%d)\n",
4096 mdname(mddev), mddev->level);
4100 if (mddev->reshape_position != MaxSector) {
4101 /* Check that we can continue the reshape.
4102 * Currently only disks can change, it must
4103 * increase, and we must be past the point where
4104 * a stripe over-writes itself
4106 sector_t here_new, here_old;
4108 int max_degraded = (mddev->level == 5 ? 1 : 2);
4110 if (mddev->new_level != mddev->level ||
4111 mddev->new_layout != mddev->layout ||
4112 mddev->new_chunk != mddev->chunk_size) {
4113 printk(KERN_ERR "raid5: %s: unsupported reshape "
4114 "required - aborting.\n",
4118 if (mddev->delta_disks <= 0) {
4119 printk(KERN_ERR "raid5: %s: unsupported reshape "
4120 "(reduce disks) required - aborting.\n",
4124 old_disks = mddev->raid_disks - mddev->delta_disks;
4125 /* reshape_position must be on a new-stripe boundary, and one
4126 * further up in new geometry must map after here in old
4129 here_new = mddev->reshape_position;
4130 if (sector_div(here_new, (mddev->chunk_size>>9)*
4131 (mddev->raid_disks - max_degraded))) {
4132 printk(KERN_ERR "raid5: reshape_position not "
4133 "on a stripe boundary\n");
4136 /* here_new is the stripe we will write to */
4137 here_old = mddev->reshape_position;
4138 sector_div(here_old, (mddev->chunk_size>>9)*
4139 (old_disks-max_degraded));
4140 /* here_old is the first stripe that we might need to read
4142 if (here_new >= here_old) {
4143 /* Reading from the same stripe as writing to - bad */
4144 printk(KERN_ERR "raid5: reshape_position too early for "
4145 "auto-recovery - aborting.\n");
4148 printk(KERN_INFO "raid5: reshape will continue\n");
4149 /* OK, we should be able to continue; */
4153 mddev->private = kzalloc(sizeof (raid5_conf_t), GFP_KERNEL);
4154 if ((conf = mddev->private) == NULL)
4156 if (mddev->reshape_position == MaxSector) {
4157 conf->previous_raid_disks = conf->raid_disks = mddev->raid_disks;
4159 conf->raid_disks = mddev->raid_disks;
4160 conf->previous_raid_disks = mddev->raid_disks - mddev->delta_disks;
4163 conf->disks = kzalloc(conf->raid_disks * sizeof(struct disk_info),
4168 conf->mddev = mddev;
4170 if ((conf->stripe_hashtbl = kzalloc(PAGE_SIZE, GFP_KERNEL)) == NULL)
4173 if (mddev->level == 6) {
4174 conf->spare_page = alloc_page(GFP_KERNEL);
4175 if (!conf->spare_page)
4178 spin_lock_init(&conf->device_lock);
4179 mddev->queue->queue_lock = &conf->device_lock;
4180 init_waitqueue_head(&conf->wait_for_stripe);
4181 init_waitqueue_head(&conf->wait_for_overlap);
4182 INIT_LIST_HEAD(&conf->handle_list);
4183 INIT_LIST_HEAD(&conf->hold_list);
4184 INIT_LIST_HEAD(&conf->delayed_list);
4185 INIT_LIST_HEAD(&conf->bitmap_list);
4186 INIT_LIST_HEAD(&conf->inactive_list);
4187 atomic_set(&conf->active_stripes, 0);
4188 atomic_set(&conf->preread_active_stripes, 0);
4189 atomic_set(&conf->active_aligned_reads, 0);
4190 conf->bypass_threshold = BYPASS_THRESHOLD;
4192 pr_debug("raid5: run(%s) called.\n", mdname(mddev));
4194 rdev_for_each(rdev, tmp, mddev) {
4195 raid_disk = rdev->raid_disk;
4196 if (raid_disk >= conf->raid_disks
4199 disk = conf->disks + raid_disk;
4203 if (test_bit(In_sync, &rdev->flags)) {
4204 char b[BDEVNAME_SIZE];
4205 printk(KERN_INFO "raid5: device %s operational as raid"
4206 " disk %d\n", bdevname(rdev->bdev,b),
4210 /* Cannot rely on bitmap to complete recovery */
4215 * 0 for a fully functional array, 1 or 2 for a degraded array.
4217 mddev->degraded = conf->raid_disks - working_disks;
4218 conf->mddev = mddev;
4219 conf->chunk_size = mddev->chunk_size;
4220 conf->level = mddev->level;
4221 if (conf->level == 6)
4222 conf->max_degraded = 2;
4224 conf->max_degraded = 1;
4225 conf->algorithm = mddev->layout;
4226 conf->max_nr_stripes = NR_STRIPES;
4227 conf->expand_progress = mddev->reshape_position;
4229 /* device size must be a multiple of chunk size */
4230 mddev->size &= ~(mddev->chunk_size/1024 -1);
4231 mddev->resync_max_sectors = mddev->size << 1;
4233 if (conf->level == 6 && conf->raid_disks < 4) {
4234 printk(KERN_ERR "raid6: not enough configured devices for %s (%d, minimum 4)\n",
4235 mdname(mddev), conf->raid_disks);
4238 if (!conf->chunk_size || conf->chunk_size % 4) {
4239 printk(KERN_ERR "raid5: invalid chunk size %d for %s\n",
4240 conf->chunk_size, mdname(mddev));
4243 if (conf->algorithm > ALGORITHM_RIGHT_SYMMETRIC) {
4245 "raid5: unsupported parity algorithm %d for %s\n",
4246 conf->algorithm, mdname(mddev));
4249 if (mddev->degraded > conf->max_degraded) {
4250 printk(KERN_ERR "raid5: not enough operational devices for %s"
4251 " (%d/%d failed)\n",
4252 mdname(mddev), mddev->degraded, conf->raid_disks);
4256 if (mddev->degraded > 0 &&
4257 mddev->recovery_cp != MaxSector) {
4258 if (mddev->ok_start_degraded)
4260 "raid5: starting dirty degraded array: %s"
4261 "- data corruption possible.\n",
4265 "raid5: cannot start dirty degraded array for %s\n",
4272 mddev->thread = md_register_thread(raid5d, mddev, "%s_raid5");
4273 if (!mddev->thread) {
4275 "raid5: couldn't allocate thread for %s\n",
4280 memory = conf->max_nr_stripes * (sizeof(struct stripe_head) +
4281 conf->raid_disks * ((sizeof(struct bio) + PAGE_SIZE))) / 1024;
4282 if (grow_stripes(conf, conf->max_nr_stripes)) {
4284 "raid5: couldn't allocate %dkB for buffers\n", memory);
4285 shrink_stripes(conf);
4286 md_unregister_thread(mddev->thread);
4289 printk(KERN_INFO "raid5: allocated %dkB for %s\n",
4290 memory, mdname(mddev));
4292 if (mddev->degraded == 0)
4293 printk("raid5: raid level %d set %s active with %d out of %d"
4294 " devices, algorithm %d\n", conf->level, mdname(mddev),
4295 mddev->raid_disks-mddev->degraded, mddev->raid_disks,
4298 printk(KERN_ALERT "raid5: raid level %d set %s active with %d"
4299 " out of %d devices, algorithm %d\n", conf->level,
4300 mdname(mddev), mddev->raid_disks - mddev->degraded,
4301 mddev->raid_disks, conf->algorithm);
4303 print_raid5_conf(conf);
4305 if (conf->expand_progress != MaxSector) {
4306 printk("...ok start reshape thread\n");
4307 conf->expand_lo = conf->expand_progress;
4308 atomic_set(&conf->reshape_stripes, 0);
4309 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
4310 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
4311 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
4312 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
4313 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
4317 /* read-ahead size must cover two whole stripes, which is
4318 * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
4321 int data_disks = conf->previous_raid_disks - conf->max_degraded;
4322 int stripe = data_disks *
4323 (mddev->chunk_size / PAGE_SIZE);
4324 if (mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
4325 mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
4328 /* Ok, everything is just fine now */
4329 if (sysfs_create_group(&mddev->kobj, &raid5_attrs_group))
4331 "raid5: failed to create sysfs attributes for %s\n",
4334 mddev->queue->unplug_fn = raid5_unplug_device;
4335 mddev->queue->backing_dev_info.congested_data = mddev;
4336 mddev->queue->backing_dev_info.congested_fn = raid5_congested;
4338 mddev->array_size = mddev->size * (conf->previous_raid_disks -
4339 conf->max_degraded);
4341 blk_queue_merge_bvec(mddev->queue, raid5_mergeable_bvec);
4346 print_raid5_conf(conf);
4347 safe_put_page(conf->spare_page);
4349 kfree(conf->stripe_hashtbl);
4352 mddev->private = NULL;
4353 printk(KERN_ALERT "raid5: failed to run raid set %s\n", mdname(mddev));
4359 static int stop(mddev_t *mddev)
4361 raid5_conf_t *conf = (raid5_conf_t *) mddev->private;
4363 md_unregister_thread(mddev->thread);
4364 mddev->thread = NULL;
4365 shrink_stripes(conf);
4366 kfree(conf->stripe_hashtbl);
4367 mddev->queue->backing_dev_info.congested_fn = NULL;
4368 blk_sync_queue(mddev->queue); /* the unplug fn references 'conf'*/
4369 sysfs_remove_group(&mddev->kobj, &raid5_attrs_group);
4372 mddev->private = NULL;
4377 static void print_sh (struct seq_file *seq, struct stripe_head *sh)
4381 seq_printf(seq, "sh %llu, pd_idx %d, state %ld.\n",
4382 (unsigned long long)sh->sector, sh->pd_idx, sh->state);
4383 seq_printf(seq, "sh %llu, count %d.\n",
4384 (unsigned long long)sh->sector, atomic_read(&sh->count));
4385 seq_printf(seq, "sh %llu, ", (unsigned long long)sh->sector);
4386 for (i = 0; i < sh->disks; i++) {
4387 seq_printf(seq, "(cache%d: %p %ld) ",
4388 i, sh->dev[i].page, sh->dev[i].flags);
4390 seq_printf(seq, "\n");
4393 static void printall (struct seq_file *seq, raid5_conf_t *conf)
4395 struct stripe_head *sh;
4396 struct hlist_node *hn;
4399 spin_lock_irq(&conf->device_lock);
4400 for (i = 0; i < NR_HASH; i++) {
4401 hlist_for_each_entry(sh, hn, &conf->stripe_hashtbl[i], hash) {
4402 if (sh->raid_conf != conf)
4407 spin_unlock_irq(&conf->device_lock);
4411 static void status (struct seq_file *seq, mddev_t *mddev)
4413 raid5_conf_t *conf = (raid5_conf_t *) mddev->private;
4416 seq_printf (seq, " level %d, %dk chunk, algorithm %d", mddev->level, mddev->chunk_size >> 10, mddev->layout);
4417 seq_printf (seq, " [%d/%d] [", conf->raid_disks, conf->raid_disks - mddev->degraded);
4418 for (i = 0; i < conf->raid_disks; i++)
4419 seq_printf (seq, "%s",
4420 conf->disks[i].rdev &&
4421 test_bit(In_sync, &conf->disks[i].rdev->flags) ? "U" : "_");
4422 seq_printf (seq, "]");
4424 seq_printf (seq, "\n");
4425 printall(seq, conf);
4429 static void print_raid5_conf (raid5_conf_t *conf)
4432 struct disk_info *tmp;
4434 printk("RAID5 conf printout:\n");
4436 printk("(conf==NULL)\n");
4439 printk(" --- rd:%d wd:%d\n", conf->raid_disks,
4440 conf->raid_disks - conf->mddev->degraded);
4442 for (i = 0; i < conf->raid_disks; i++) {
4443 char b[BDEVNAME_SIZE];
4444 tmp = conf->disks + i;
4446 printk(" disk %d, o:%d, dev:%s\n",
4447 i, !test_bit(Faulty, &tmp->rdev->flags),
4448 bdevname(tmp->rdev->bdev,b));
4452 static int raid5_spare_active(mddev_t *mddev)
4455 raid5_conf_t *conf = mddev->private;
4456 struct disk_info *tmp;
4458 for (i = 0; i < conf->raid_disks; i++) {
4459 tmp = conf->disks + i;
4461 && !test_bit(Faulty, &tmp->rdev->flags)
4462 && !test_and_set_bit(In_sync, &tmp->rdev->flags)) {
4463 unsigned long flags;
4464 spin_lock_irqsave(&conf->device_lock, flags);
4466 spin_unlock_irqrestore(&conf->device_lock, flags);
4469 print_raid5_conf(conf);
4473 static int raid5_remove_disk(mddev_t *mddev, int number)
4475 raid5_conf_t *conf = mddev->private;
4478 struct disk_info *p = conf->disks + number;
4480 print_raid5_conf(conf);
4483 if (test_bit(In_sync, &rdev->flags) ||
4484 atomic_read(&rdev->nr_pending)) {
4488 /* Only remove non-faulty devices if recovery
4491 if (!test_bit(Faulty, &rdev->flags) &&
4492 mddev->degraded <= conf->max_degraded) {
4498 if (atomic_read(&rdev->nr_pending)) {
4499 /* lost the race, try later */
4506 print_raid5_conf(conf);
4510 static int raid5_add_disk(mddev_t *mddev, mdk_rdev_t *rdev)
4512 raid5_conf_t *conf = mddev->private;
4515 struct disk_info *p;
4517 int last = conf->raid_disks - 1;
4519 if (mddev->degraded > conf->max_degraded)
4520 /* no point adding a device */
4523 if (rdev->raid_disk >= 0)
4524 first = last = rdev->raid_disk;
4527 * find the disk ... but prefer rdev->saved_raid_disk
4530 if (rdev->saved_raid_disk >= 0 &&
4531 rdev->saved_raid_disk >= first &&
4532 conf->disks[rdev->saved_raid_disk].rdev == NULL)
4533 disk = rdev->saved_raid_disk;
4536 for ( ; disk <= last ; disk++)
4537 if ((p=conf->disks + disk)->rdev == NULL) {
4538 clear_bit(In_sync, &rdev->flags);
4539 rdev->raid_disk = disk;
4541 if (rdev->saved_raid_disk != disk)
4543 rcu_assign_pointer(p->rdev, rdev);
4546 print_raid5_conf(conf);
4550 static int raid5_resize(mddev_t *mddev, sector_t sectors)
4552 /* no resync is happening, and there is enough space
4553 * on all devices, so we can resize.
4554 * We need to make sure resync covers any new space.
4555 * If the array is shrinking we should possibly wait until
4556 * any io in the removed space completes, but it hardly seems
4559 raid5_conf_t *conf = mddev_to_conf(mddev);
4561 sectors &= ~((sector_t)mddev->chunk_size/512 - 1);
4562 mddev->array_size = (sectors * (mddev->raid_disks-conf->max_degraded))>>1;
4563 set_capacity(mddev->gendisk, mddev->array_size << 1);
4565 if (sectors/2 > mddev->size && mddev->recovery_cp == MaxSector) {
4566 mddev->recovery_cp = mddev->size << 1;
4567 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
4569 mddev->size = sectors /2;
4570 mddev->resync_max_sectors = sectors;
4574 #ifdef CONFIG_MD_RAID5_RESHAPE
4575 static int raid5_check_reshape(mddev_t *mddev)
4577 raid5_conf_t *conf = mddev_to_conf(mddev);
4580 if (mddev->delta_disks < 0 ||
4581 mddev->new_level != mddev->level)
4582 return -EINVAL; /* Cannot shrink array or change level yet */
4583 if (mddev->delta_disks == 0)
4584 return 0; /* nothing to do */
4586 /* Can only proceed if there are plenty of stripe_heads.
4587 * We need a minimum of one full stripe,, and for sensible progress
4588 * it is best to have about 4 times that.
4589 * If we require 4 times, then the default 256 4K stripe_heads will
4590 * allow for chunk sizes up to 256K, which is probably OK.
4591 * If the chunk size is greater, user-space should request more
4592 * stripe_heads first.
4594 if ((mddev->chunk_size / STRIPE_SIZE) * 4 > conf->max_nr_stripes ||
4595 (mddev->new_chunk / STRIPE_SIZE) * 4 > conf->max_nr_stripes) {
4596 printk(KERN_WARNING "raid5: reshape: not enough stripes. Needed %lu\n",
4597 (mddev->chunk_size / STRIPE_SIZE)*4);
4601 err = resize_stripes(conf, conf->raid_disks + mddev->delta_disks);
4605 if (mddev->degraded > conf->max_degraded)
4607 /* looks like we might be able to manage this */
4611 static int raid5_start_reshape(mddev_t *mddev)
4613 raid5_conf_t *conf = mddev_to_conf(mddev);
4615 struct list_head *rtmp;
4617 int added_devices = 0;
4618 unsigned long flags;
4620 if (test_bit(MD_RECOVERY_RUNNING, &mddev->recovery))
4623 rdev_for_each(rdev, rtmp, mddev)
4624 if (rdev->raid_disk < 0 &&
4625 !test_bit(Faulty, &rdev->flags))
4628 if (spares - mddev->degraded < mddev->delta_disks - conf->max_degraded)
4629 /* Not enough devices even to make a degraded array
4634 atomic_set(&conf->reshape_stripes, 0);
4635 spin_lock_irq(&conf->device_lock);
4636 conf->previous_raid_disks = conf->raid_disks;
4637 conf->raid_disks += mddev->delta_disks;
4638 conf->expand_progress = 0;
4639 conf->expand_lo = 0;
4640 spin_unlock_irq(&conf->device_lock);
4642 /* Add some new drives, as many as will fit.
4643 * We know there are enough to make the newly sized array work.
4645 rdev_for_each(rdev, rtmp, mddev)
4646 if (rdev->raid_disk < 0 &&
4647 !test_bit(Faulty, &rdev->flags)) {
4648 if (raid5_add_disk(mddev, rdev) == 0) {
4650 set_bit(In_sync, &rdev->flags);
4652 rdev->recovery_offset = 0;
4653 sprintf(nm, "rd%d", rdev->raid_disk);
4654 if (sysfs_create_link(&mddev->kobj,
4657 "raid5: failed to create "
4658 " link %s for %s\n",
4664 spin_lock_irqsave(&conf->device_lock, flags);
4665 mddev->degraded = (conf->raid_disks - conf->previous_raid_disks) - added_devices;
4666 spin_unlock_irqrestore(&conf->device_lock, flags);
4667 mddev->raid_disks = conf->raid_disks;
4668 mddev->reshape_position = 0;
4669 set_bit(MD_CHANGE_DEVS, &mddev->flags);
4671 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
4672 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
4673 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
4674 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
4675 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
4677 if (!mddev->sync_thread) {
4678 mddev->recovery = 0;
4679 spin_lock_irq(&conf->device_lock);
4680 mddev->raid_disks = conf->raid_disks = conf->previous_raid_disks;
4681 conf->expand_progress = MaxSector;
4682 spin_unlock_irq(&conf->device_lock);
4685 md_wakeup_thread(mddev->sync_thread);
4686 md_new_event(mddev);
4691 static void end_reshape(raid5_conf_t *conf)
4693 struct block_device *bdev;
4695 if (!test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery)) {
4696 conf->mddev->array_size = conf->mddev->size *
4697 (conf->raid_disks - conf->max_degraded);
4698 set_capacity(conf->mddev->gendisk, conf->mddev->array_size << 1);
4699 conf->mddev->changed = 1;
4701 bdev = bdget_disk(conf->mddev->gendisk, 0);
4703 mutex_lock(&bdev->bd_inode->i_mutex);
4704 i_size_write(bdev->bd_inode, (loff_t)conf->mddev->array_size << 10);
4705 mutex_unlock(&bdev->bd_inode->i_mutex);
4708 spin_lock_irq(&conf->device_lock);
4709 conf->expand_progress = MaxSector;
4710 spin_unlock_irq(&conf->device_lock);
4711 conf->mddev->reshape_position = MaxSector;
4713 /* read-ahead size must cover two whole stripes, which is
4714 * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
4717 int data_disks = conf->previous_raid_disks - conf->max_degraded;
4718 int stripe = data_disks *
4719 (conf->mddev->chunk_size / PAGE_SIZE);
4720 if (conf->mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
4721 conf->mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
4726 static void raid5_quiesce(mddev_t *mddev, int state)
4728 raid5_conf_t *conf = mddev_to_conf(mddev);
4731 case 2: /* resume for a suspend */
4732 wake_up(&conf->wait_for_overlap);
4735 case 1: /* stop all writes */
4736 spin_lock_irq(&conf->device_lock);
4738 wait_event_lock_irq(conf->wait_for_stripe,
4739 atomic_read(&conf->active_stripes) == 0 &&
4740 atomic_read(&conf->active_aligned_reads) == 0,
4741 conf->device_lock, /* nothing */);
4742 spin_unlock_irq(&conf->device_lock);
4745 case 0: /* re-enable writes */
4746 spin_lock_irq(&conf->device_lock);
4748 wake_up(&conf->wait_for_stripe);
4749 wake_up(&conf->wait_for_overlap);
4750 spin_unlock_irq(&conf->device_lock);
4755 static struct mdk_personality raid6_personality =
4759 .owner = THIS_MODULE,
4760 .make_request = make_request,
4764 .error_handler = error,
4765 .hot_add_disk = raid5_add_disk,
4766 .hot_remove_disk= raid5_remove_disk,
4767 .spare_active = raid5_spare_active,
4768 .sync_request = sync_request,
4769 .resize = raid5_resize,
4770 #ifdef CONFIG_MD_RAID5_RESHAPE
4771 .check_reshape = raid5_check_reshape,
4772 .start_reshape = raid5_start_reshape,
4774 .quiesce = raid5_quiesce,
4776 static struct mdk_personality raid5_personality =
4780 .owner = THIS_MODULE,
4781 .make_request = make_request,
4785 .error_handler = error,
4786 .hot_add_disk = raid5_add_disk,
4787 .hot_remove_disk= raid5_remove_disk,
4788 .spare_active = raid5_spare_active,
4789 .sync_request = sync_request,
4790 .resize = raid5_resize,
4791 #ifdef CONFIG_MD_RAID5_RESHAPE
4792 .check_reshape = raid5_check_reshape,
4793 .start_reshape = raid5_start_reshape,
4795 .quiesce = raid5_quiesce,
4798 static struct mdk_personality raid4_personality =
4802 .owner = THIS_MODULE,
4803 .make_request = make_request,
4807 .error_handler = error,
4808 .hot_add_disk = raid5_add_disk,
4809 .hot_remove_disk= raid5_remove_disk,
4810 .spare_active = raid5_spare_active,
4811 .sync_request = sync_request,
4812 .resize = raid5_resize,
4813 #ifdef CONFIG_MD_RAID5_RESHAPE
4814 .check_reshape = raid5_check_reshape,
4815 .start_reshape = raid5_start_reshape,
4817 .quiesce = raid5_quiesce,
4820 static int __init raid5_init(void)
4824 e = raid6_select_algo();
4827 register_md_personality(&raid6_personality);
4828 register_md_personality(&raid5_personality);
4829 register_md_personality(&raid4_personality);
4833 static void raid5_exit(void)
4835 unregister_md_personality(&raid6_personality);
4836 unregister_md_personality(&raid5_personality);
4837 unregister_md_personality(&raid4_personality);
4840 module_init(raid5_init);
4841 module_exit(raid5_exit);
4842 MODULE_LICENSE("GPL");
4843 MODULE_ALIAS("md-personality-4"); /* RAID5 */
4844 MODULE_ALIAS("md-raid5");
4845 MODULE_ALIAS("md-raid4");
4846 MODULE_ALIAS("md-level-5");
4847 MODULE_ALIAS("md-level-4");
4848 MODULE_ALIAS("md-personality-8"); /* RAID6 */
4849 MODULE_ALIAS("md-raid6");
4850 MODULE_ALIAS("md-level-6");
4852 /* This used to be two separate modules, they were: */
4853 MODULE_ALIAS("raid5");
4854 MODULE_ALIAS("raid6");