2 * raid5.c : Multiple Devices driver for Linux
3 * Copyright (C) 1996, 1997 Ingo Molnar, Miguel de Icaza, Gadi Oxman
4 * Copyright (C) 1999, 2000 Ingo Molnar
5 * Copyright (C) 2002, 2003 H. Peter Anvin
7 * RAID-4/5/6 management functions.
8 * Thanks to Penguin Computing for making the RAID-6 development possible
9 * by donating a test server!
11 * This program is free software; you can redistribute it and/or modify
12 * it under the terms of the GNU General Public License as published by
13 * the Free Software Foundation; either version 2, or (at your option)
16 * You should have received a copy of the GNU General Public License
17 * (for example /usr/src/linux/COPYING); if not, write to the Free
18 * Software Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
24 * The sequencing for updating the bitmap reliably is a little
25 * subtle (and I got it wrong the first time) so it deserves some
28 * We group bitmap updates into batches. Each batch has a number.
29 * We may write out several batches at once, but that isn't very important.
30 * conf->bm_write is the number of the last batch successfully written.
31 * conf->bm_flush is the number of the last batch that was closed to
33 * When we discover that we will need to write to any block in a stripe
34 * (in add_stripe_bio) we update the in-memory bitmap and record in sh->bm_seq
35 * the number of the batch it will be in. This is bm_flush+1.
36 * When we are ready to do a write, if that batch hasn't been written yet,
37 * we plug the array and queue the stripe for later.
38 * When an unplug happens, we increment bm_flush, thus closing the current
40 * When we notice that bm_flush > bm_write, we write out all pending updates
41 * to the bitmap, and advance bm_write to where bm_flush was.
42 * This may occasionally write a bit out twice, but is sure never to
46 #include <linux/blkdev.h>
47 #include <linux/kthread.h>
48 #include <linux/raid/pq.h>
49 #include <linux/async_tx.h>
50 #include <linux/seq_file.h>
59 #define NR_STRIPES 256
60 #define STRIPE_SIZE PAGE_SIZE
61 #define STRIPE_SHIFT (PAGE_SHIFT - 9)
62 #define STRIPE_SECTORS (STRIPE_SIZE>>9)
63 #define IO_THRESHOLD 1
64 #define BYPASS_THRESHOLD 1
65 #define NR_HASH (PAGE_SIZE / sizeof(struct hlist_head))
66 #define HASH_MASK (NR_HASH - 1)
68 #define stripe_hash(conf, sect) (&((conf)->stripe_hashtbl[((sect) >> STRIPE_SHIFT) & HASH_MASK]))
70 /* bio's attached to a stripe+device for I/O are linked together in bi_sector
71 * order without overlap. There may be several bio's per stripe+device, and
72 * a bio could span several devices.
73 * When walking this list for a particular stripe+device, we must never proceed
74 * beyond a bio that extends past this device, as the next bio might no longer
76 * This macro is used to determine the 'next' bio in the list, given the sector
77 * of the current stripe+device
79 #define r5_next_bio(bio, sect) ( ( (bio)->bi_sector + ((bio)->bi_size>>9) < sect + STRIPE_SECTORS) ? (bio)->bi_next : NULL)
81 * The following can be used to debug the driver
83 #define RAID5_PARANOIA 1
84 #if RAID5_PARANOIA && defined(CONFIG_SMP)
85 # define CHECK_DEVLOCK() assert_spin_locked(&conf->device_lock)
87 # define CHECK_DEVLOCK()
95 #define printk_rl(args...) ((void) (printk_ratelimit() && printk(args)))
98 * We maintain a biased count of active stripes in the bottom 16 bits of
99 * bi_phys_segments, and a count of processed stripes in the upper 16 bits
101 static inline int raid5_bi_phys_segments(struct bio *bio)
103 return bio->bi_phys_segments & 0xffff;
106 static inline int raid5_bi_hw_segments(struct bio *bio)
108 return (bio->bi_phys_segments >> 16) & 0xffff;
111 static inline int raid5_dec_bi_phys_segments(struct bio *bio)
113 --bio->bi_phys_segments;
114 return raid5_bi_phys_segments(bio);
117 static inline int raid5_dec_bi_hw_segments(struct bio *bio)
119 unsigned short val = raid5_bi_hw_segments(bio);
122 bio->bi_phys_segments = (val << 16) | raid5_bi_phys_segments(bio);
126 static inline void raid5_set_bi_hw_segments(struct bio *bio, unsigned int cnt)
128 bio->bi_phys_segments = raid5_bi_phys_segments(bio) || (cnt << 16);
131 /* Find first data disk in a raid6 stripe */
132 static inline int raid6_d0(struct stripe_head *sh)
135 /* ddf always start from first device */
137 /* md starts just after Q block */
138 if (sh->qd_idx == sh->disks - 1)
141 return sh->qd_idx + 1;
143 static inline int raid6_next_disk(int disk, int raid_disks)
146 return (disk < raid_disks) ? disk : 0;
149 /* When walking through the disks in a raid5, starting at raid6_d0,
150 * We need to map each disk to a 'slot', where the data disks are slot
151 * 0 .. raid_disks-3, the parity disk is raid_disks-2 and the Q disk
152 * is raid_disks-1. This help does that mapping.
154 static int raid6_idx_to_slot(int idx, struct stripe_head *sh,
155 int *count, int syndrome_disks)
159 if (idx == sh->pd_idx)
160 return syndrome_disks;
161 if (idx == sh->qd_idx)
162 return syndrome_disks + 1;
167 static void return_io(struct bio *return_bi)
169 struct bio *bi = return_bi;
172 return_bi = bi->bi_next;
180 static void print_raid5_conf (raid5_conf_t *conf);
182 static int stripe_operations_active(struct stripe_head *sh)
184 return sh->check_state || sh->reconstruct_state ||
185 test_bit(STRIPE_BIOFILL_RUN, &sh->state) ||
186 test_bit(STRIPE_COMPUTE_RUN, &sh->state);
189 static void __release_stripe(raid5_conf_t *conf, struct stripe_head *sh)
191 if (atomic_dec_and_test(&sh->count)) {
192 BUG_ON(!list_empty(&sh->lru));
193 BUG_ON(atomic_read(&conf->active_stripes)==0);
194 if (test_bit(STRIPE_HANDLE, &sh->state)) {
195 if (test_bit(STRIPE_DELAYED, &sh->state)) {
196 list_add_tail(&sh->lru, &conf->delayed_list);
197 blk_plug_device(conf->mddev->queue);
198 } else if (test_bit(STRIPE_BIT_DELAY, &sh->state) &&
199 sh->bm_seq - conf->seq_write > 0) {
200 list_add_tail(&sh->lru, &conf->bitmap_list);
201 blk_plug_device(conf->mddev->queue);
203 clear_bit(STRIPE_BIT_DELAY, &sh->state);
204 list_add_tail(&sh->lru, &conf->handle_list);
206 md_wakeup_thread(conf->mddev->thread);
208 BUG_ON(stripe_operations_active(sh));
209 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
210 atomic_dec(&conf->preread_active_stripes);
211 if (atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD)
212 md_wakeup_thread(conf->mddev->thread);
214 atomic_dec(&conf->active_stripes);
215 if (!test_bit(STRIPE_EXPANDING, &sh->state)) {
216 list_add_tail(&sh->lru, &conf->inactive_list);
217 wake_up(&conf->wait_for_stripe);
218 if (conf->retry_read_aligned)
219 md_wakeup_thread(conf->mddev->thread);
225 static void release_stripe(struct stripe_head *sh)
227 raid5_conf_t *conf = sh->raid_conf;
230 spin_lock_irqsave(&conf->device_lock, flags);
231 __release_stripe(conf, sh);
232 spin_unlock_irqrestore(&conf->device_lock, flags);
235 static inline void remove_hash(struct stripe_head *sh)
237 pr_debug("remove_hash(), stripe %llu\n",
238 (unsigned long long)sh->sector);
240 hlist_del_init(&sh->hash);
243 static inline void insert_hash(raid5_conf_t *conf, struct stripe_head *sh)
245 struct hlist_head *hp = stripe_hash(conf, sh->sector);
247 pr_debug("insert_hash(), stripe %llu\n",
248 (unsigned long long)sh->sector);
251 hlist_add_head(&sh->hash, hp);
255 /* find an idle stripe, make sure it is unhashed, and return it. */
256 static struct stripe_head *get_free_stripe(raid5_conf_t *conf)
258 struct stripe_head *sh = NULL;
259 struct list_head *first;
262 if (list_empty(&conf->inactive_list))
264 first = conf->inactive_list.next;
265 sh = list_entry(first, struct stripe_head, lru);
266 list_del_init(first);
268 atomic_inc(&conf->active_stripes);
273 static void shrink_buffers(struct stripe_head *sh, int num)
278 for (i=0; i<num ; i++) {
282 sh->dev[i].page = NULL;
287 static int grow_buffers(struct stripe_head *sh, int num)
291 for (i=0; i<num; i++) {
294 if (!(page = alloc_page(GFP_KERNEL))) {
297 sh->dev[i].page = page;
302 static void raid5_build_block(struct stripe_head *sh, int i, int previous);
303 static void stripe_set_idx(sector_t stripe, raid5_conf_t *conf, int previous,
304 struct stripe_head *sh);
306 static void init_stripe(struct stripe_head *sh, sector_t sector, int previous)
308 raid5_conf_t *conf = sh->raid_conf;
311 BUG_ON(atomic_read(&sh->count) != 0);
312 BUG_ON(test_bit(STRIPE_HANDLE, &sh->state));
313 BUG_ON(stripe_operations_active(sh));
316 pr_debug("init_stripe called, stripe %llu\n",
317 (unsigned long long)sh->sector);
321 sh->generation = conf->generation - previous;
322 sh->disks = previous ? conf->previous_raid_disks : conf->raid_disks;
324 stripe_set_idx(sector, conf, previous, sh);
328 for (i = sh->disks; i--; ) {
329 struct r5dev *dev = &sh->dev[i];
331 if (dev->toread || dev->read || dev->towrite || dev->written ||
332 test_bit(R5_LOCKED, &dev->flags)) {
333 printk(KERN_ERR "sector=%llx i=%d %p %p %p %p %d\n",
334 (unsigned long long)sh->sector, i, dev->toread,
335 dev->read, dev->towrite, dev->written,
336 test_bit(R5_LOCKED, &dev->flags));
340 raid5_build_block(sh, i, previous);
342 insert_hash(conf, sh);
345 static struct stripe_head *__find_stripe(raid5_conf_t *conf, sector_t sector,
348 struct stripe_head *sh;
349 struct hlist_node *hn;
352 pr_debug("__find_stripe, sector %llu\n", (unsigned long long)sector);
353 hlist_for_each_entry(sh, hn, stripe_hash(conf, sector), hash)
354 if (sh->sector == sector && sh->generation == generation)
356 pr_debug("__stripe %llu not in cache\n", (unsigned long long)sector);
360 static void unplug_slaves(mddev_t *mddev);
361 static void raid5_unplug_device(struct request_queue *q);
363 static struct stripe_head *
364 get_active_stripe(raid5_conf_t *conf, sector_t sector,
365 int previous, int noblock)
367 struct stripe_head *sh;
369 pr_debug("get_stripe, sector %llu\n", (unsigned long long)sector);
371 spin_lock_irq(&conf->device_lock);
374 wait_event_lock_irq(conf->wait_for_stripe,
376 conf->device_lock, /* nothing */);
377 sh = __find_stripe(conf, sector, conf->generation - previous);
379 if (!conf->inactive_blocked)
380 sh = get_free_stripe(conf);
381 if (noblock && sh == NULL)
384 conf->inactive_blocked = 1;
385 wait_event_lock_irq(conf->wait_for_stripe,
386 !list_empty(&conf->inactive_list) &&
387 (atomic_read(&conf->active_stripes)
388 < (conf->max_nr_stripes *3/4)
389 || !conf->inactive_blocked),
391 raid5_unplug_device(conf->mddev->queue)
393 conf->inactive_blocked = 0;
395 init_stripe(sh, sector, previous);
397 if (atomic_read(&sh->count)) {
398 BUG_ON(!list_empty(&sh->lru)
399 && !test_bit(STRIPE_EXPANDING, &sh->state));
401 if (!test_bit(STRIPE_HANDLE, &sh->state))
402 atomic_inc(&conf->active_stripes);
403 if (list_empty(&sh->lru) &&
404 !test_bit(STRIPE_EXPANDING, &sh->state))
406 list_del_init(&sh->lru);
409 } while (sh == NULL);
412 atomic_inc(&sh->count);
414 spin_unlock_irq(&conf->device_lock);
419 raid5_end_read_request(struct bio *bi, int error);
421 raid5_end_write_request(struct bio *bi, int error);
423 static void ops_run_io(struct stripe_head *sh, struct stripe_head_state *s)
425 raid5_conf_t *conf = sh->raid_conf;
426 int i, disks = sh->disks;
430 for (i = disks; i--; ) {
434 if (test_and_clear_bit(R5_Wantwrite, &sh->dev[i].flags))
436 else if (test_and_clear_bit(R5_Wantread, &sh->dev[i].flags))
441 bi = &sh->dev[i].req;
445 bi->bi_end_io = raid5_end_write_request;
447 bi->bi_end_io = raid5_end_read_request;
450 rdev = rcu_dereference(conf->disks[i].rdev);
451 if (rdev && test_bit(Faulty, &rdev->flags))
454 atomic_inc(&rdev->nr_pending);
458 if (s->syncing || s->expanding || s->expanded)
459 md_sync_acct(rdev->bdev, STRIPE_SECTORS);
461 set_bit(STRIPE_IO_STARTED, &sh->state);
463 bi->bi_bdev = rdev->bdev;
464 pr_debug("%s: for %llu schedule op %ld on disc %d\n",
465 __func__, (unsigned long long)sh->sector,
467 atomic_inc(&sh->count);
468 bi->bi_sector = sh->sector + rdev->data_offset;
469 bi->bi_flags = 1 << BIO_UPTODATE;
473 bi->bi_io_vec = &sh->dev[i].vec;
474 bi->bi_io_vec[0].bv_len = STRIPE_SIZE;
475 bi->bi_io_vec[0].bv_offset = 0;
476 bi->bi_size = STRIPE_SIZE;
479 test_bit(R5_ReWrite, &sh->dev[i].flags))
480 atomic_add(STRIPE_SECTORS,
481 &rdev->corrected_errors);
482 generic_make_request(bi);
485 set_bit(STRIPE_DEGRADED, &sh->state);
486 pr_debug("skip op %ld on disc %d for sector %llu\n",
487 bi->bi_rw, i, (unsigned long long)sh->sector);
488 clear_bit(R5_LOCKED, &sh->dev[i].flags);
489 set_bit(STRIPE_HANDLE, &sh->state);
494 static struct dma_async_tx_descriptor *
495 async_copy_data(int frombio, struct bio *bio, struct page *page,
496 sector_t sector, struct dma_async_tx_descriptor *tx)
499 struct page *bio_page;
502 struct async_submit_ctl submit;
504 if (bio->bi_sector >= sector)
505 page_offset = (signed)(bio->bi_sector - sector) * 512;
507 page_offset = (signed)(sector - bio->bi_sector) * -512;
509 init_async_submit(&submit, 0, tx, NULL, NULL, NULL);
510 bio_for_each_segment(bvl, bio, i) {
511 int len = bio_iovec_idx(bio, i)->bv_len;
515 if (page_offset < 0) {
516 b_offset = -page_offset;
517 page_offset += b_offset;
521 if (len > 0 && page_offset + len > STRIPE_SIZE)
522 clen = STRIPE_SIZE - page_offset;
527 b_offset += bio_iovec_idx(bio, i)->bv_offset;
528 bio_page = bio_iovec_idx(bio, i)->bv_page;
530 tx = async_memcpy(page, bio_page, page_offset,
531 b_offset, clen, &submit);
533 tx = async_memcpy(bio_page, page, b_offset,
534 page_offset, clen, &submit);
536 /* chain the operations */
537 submit.depend_tx = tx;
539 if (clen < len) /* hit end of page */
547 static void ops_complete_biofill(void *stripe_head_ref)
549 struct stripe_head *sh = stripe_head_ref;
550 struct bio *return_bi = NULL;
551 raid5_conf_t *conf = sh->raid_conf;
554 pr_debug("%s: stripe %llu\n", __func__,
555 (unsigned long long)sh->sector);
557 /* clear completed biofills */
558 spin_lock_irq(&conf->device_lock);
559 for (i = sh->disks; i--; ) {
560 struct r5dev *dev = &sh->dev[i];
562 /* acknowledge completion of a biofill operation */
563 /* and check if we need to reply to a read request,
564 * new R5_Wantfill requests are held off until
565 * !STRIPE_BIOFILL_RUN
567 if (test_and_clear_bit(R5_Wantfill, &dev->flags)) {
568 struct bio *rbi, *rbi2;
573 while (rbi && rbi->bi_sector <
574 dev->sector + STRIPE_SECTORS) {
575 rbi2 = r5_next_bio(rbi, dev->sector);
576 if (!raid5_dec_bi_phys_segments(rbi)) {
577 rbi->bi_next = return_bi;
584 spin_unlock_irq(&conf->device_lock);
585 clear_bit(STRIPE_BIOFILL_RUN, &sh->state);
587 return_io(return_bi);
589 set_bit(STRIPE_HANDLE, &sh->state);
593 static void ops_run_biofill(struct stripe_head *sh)
595 struct dma_async_tx_descriptor *tx = NULL;
596 raid5_conf_t *conf = sh->raid_conf;
597 struct async_submit_ctl submit;
600 pr_debug("%s: stripe %llu\n", __func__,
601 (unsigned long long)sh->sector);
603 for (i = sh->disks; i--; ) {
604 struct r5dev *dev = &sh->dev[i];
605 if (test_bit(R5_Wantfill, &dev->flags)) {
607 spin_lock_irq(&conf->device_lock);
608 dev->read = rbi = dev->toread;
610 spin_unlock_irq(&conf->device_lock);
611 while (rbi && rbi->bi_sector <
612 dev->sector + STRIPE_SECTORS) {
613 tx = async_copy_data(0, rbi, dev->page,
615 rbi = r5_next_bio(rbi, dev->sector);
620 atomic_inc(&sh->count);
621 init_async_submit(&submit, ASYNC_TX_ACK, tx, ops_complete_biofill, sh, NULL);
622 async_trigger_callback(&submit);
625 static void ops_complete_compute5(void *stripe_head_ref)
627 struct stripe_head *sh = stripe_head_ref;
628 int target = sh->ops.target;
629 struct r5dev *tgt = &sh->dev[target];
631 pr_debug("%s: stripe %llu\n", __func__,
632 (unsigned long long)sh->sector);
634 set_bit(R5_UPTODATE, &tgt->flags);
635 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
636 clear_bit(R5_Wantcompute, &tgt->flags);
637 clear_bit(STRIPE_COMPUTE_RUN, &sh->state);
638 if (sh->check_state == check_state_compute_run)
639 sh->check_state = check_state_compute_result;
640 set_bit(STRIPE_HANDLE, &sh->state);
644 static struct dma_async_tx_descriptor *ops_run_compute5(struct stripe_head *sh)
646 /* kernel stack size limits the total number of disks */
647 int disks = sh->disks;
648 struct page *xor_srcs[disks];
649 int target = sh->ops.target;
650 struct r5dev *tgt = &sh->dev[target];
651 struct page *xor_dest = tgt->page;
653 struct dma_async_tx_descriptor *tx;
654 struct async_submit_ctl submit;
657 pr_debug("%s: stripe %llu block: %d\n",
658 __func__, (unsigned long long)sh->sector, target);
659 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
661 for (i = disks; i--; )
663 xor_srcs[count++] = sh->dev[i].page;
665 atomic_inc(&sh->count);
667 init_async_submit(&submit, ASYNC_TX_XOR_ZERO_DST, NULL,
668 ops_complete_compute5, sh, NULL);
669 if (unlikely(count == 1))
670 tx = async_memcpy(xor_dest, xor_srcs[0], 0, 0, STRIPE_SIZE, &submit);
672 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
677 static void ops_complete_prexor(void *stripe_head_ref)
679 struct stripe_head *sh = stripe_head_ref;
681 pr_debug("%s: stripe %llu\n", __func__,
682 (unsigned long long)sh->sector);
685 static struct dma_async_tx_descriptor *
686 ops_run_prexor(struct stripe_head *sh, struct dma_async_tx_descriptor *tx)
688 /* kernel stack size limits the total number of disks */
689 int disks = sh->disks;
690 struct page *xor_srcs[disks];
691 int count = 0, pd_idx = sh->pd_idx, i;
692 struct async_submit_ctl submit;
694 /* existing parity data subtracted */
695 struct page *xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
697 pr_debug("%s: stripe %llu\n", __func__,
698 (unsigned long long)sh->sector);
700 for (i = disks; i--; ) {
701 struct r5dev *dev = &sh->dev[i];
702 /* Only process blocks that are known to be uptodate */
703 if (test_bit(R5_Wantdrain, &dev->flags))
704 xor_srcs[count++] = dev->page;
707 init_async_submit(&submit, ASYNC_TX_XOR_DROP_DST, tx,
708 ops_complete_prexor, sh, NULL);
709 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
714 static struct dma_async_tx_descriptor *
715 ops_run_biodrain(struct stripe_head *sh, struct dma_async_tx_descriptor *tx)
717 int disks = sh->disks;
720 pr_debug("%s: stripe %llu\n", __func__,
721 (unsigned long long)sh->sector);
723 for (i = disks; i--; ) {
724 struct r5dev *dev = &sh->dev[i];
727 if (test_and_clear_bit(R5_Wantdrain, &dev->flags)) {
730 spin_lock(&sh->lock);
731 chosen = dev->towrite;
733 BUG_ON(dev->written);
734 wbi = dev->written = chosen;
735 spin_unlock(&sh->lock);
737 while (wbi && wbi->bi_sector <
738 dev->sector + STRIPE_SECTORS) {
739 tx = async_copy_data(1, wbi, dev->page,
741 wbi = r5_next_bio(wbi, dev->sector);
749 static void ops_complete_postxor(void *stripe_head_ref)
751 struct stripe_head *sh = stripe_head_ref;
752 int disks = sh->disks, i, pd_idx = sh->pd_idx;
754 pr_debug("%s: stripe %llu\n", __func__,
755 (unsigned long long)sh->sector);
757 for (i = disks; i--; ) {
758 struct r5dev *dev = &sh->dev[i];
759 if (dev->written || i == pd_idx)
760 set_bit(R5_UPTODATE, &dev->flags);
763 if (sh->reconstruct_state == reconstruct_state_drain_run)
764 sh->reconstruct_state = reconstruct_state_drain_result;
765 else if (sh->reconstruct_state == reconstruct_state_prexor_drain_run)
766 sh->reconstruct_state = reconstruct_state_prexor_drain_result;
768 BUG_ON(sh->reconstruct_state != reconstruct_state_run);
769 sh->reconstruct_state = reconstruct_state_result;
772 set_bit(STRIPE_HANDLE, &sh->state);
777 ops_run_postxor(struct stripe_head *sh, struct dma_async_tx_descriptor *tx)
779 /* kernel stack size limits the total number of disks */
780 int disks = sh->disks;
781 struct page *xor_srcs[disks];
782 struct async_submit_ctl submit;
783 int count = 0, pd_idx = sh->pd_idx, i;
784 struct page *xor_dest;
788 pr_debug("%s: stripe %llu\n", __func__,
789 (unsigned long long)sh->sector);
791 /* check if prexor is active which means only process blocks
792 * that are part of a read-modify-write (written)
794 if (sh->reconstruct_state == reconstruct_state_prexor_drain_run) {
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 /* 1/ if we prexor'd then the dest is reused as a source
812 * 2/ if we did not prexor then we are redoing the parity
813 * set ASYNC_TX_XOR_DROP_DST and ASYNC_TX_XOR_ZERO_DST
814 * for the synchronous xor case
816 flags = ASYNC_TX_ACK |
817 (prexor ? ASYNC_TX_XOR_DROP_DST : ASYNC_TX_XOR_ZERO_DST);
819 atomic_inc(&sh->count);
821 init_async_submit(&submit, flags, tx, ops_complete_postxor, sh, NULL);
822 if (unlikely(count == 1))
823 tx = async_memcpy(xor_dest, xor_srcs[0], 0, 0, STRIPE_SIZE, &submit);
825 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
828 static void ops_complete_check(void *stripe_head_ref)
830 struct stripe_head *sh = stripe_head_ref;
832 pr_debug("%s: stripe %llu\n", __func__,
833 (unsigned long long)sh->sector);
835 sh->check_state = check_state_check_result;
836 set_bit(STRIPE_HANDLE, &sh->state);
840 static void ops_run_check(struct stripe_head *sh)
842 /* kernel stack size limits the total number of disks */
843 int disks = sh->disks;
844 struct page *xor_srcs[disks];
845 struct dma_async_tx_descriptor *tx;
846 struct async_submit_ctl submit;
848 int count = 0, pd_idx = sh->pd_idx, i;
849 struct page *xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
851 pr_debug("%s: stripe %llu\n", __func__,
852 (unsigned long long)sh->sector);
854 for (i = disks; i--; ) {
855 struct r5dev *dev = &sh->dev[i];
857 xor_srcs[count++] = dev->page;
860 init_async_submit(&submit, 0, NULL, NULL, NULL, NULL);
861 tx = async_xor_val(xor_dest, xor_srcs, 0, count, STRIPE_SIZE,
862 &sh->ops.zero_sum_result, &submit);
864 atomic_inc(&sh->count);
865 init_async_submit(&submit, ASYNC_TX_ACK, tx, ops_complete_check, sh, NULL);
866 tx = async_trigger_callback(&submit);
869 static void raid5_run_ops(struct stripe_head *sh, unsigned long ops_request)
871 int overlap_clear = 0, i, disks = sh->disks;
872 struct dma_async_tx_descriptor *tx = NULL;
874 if (test_bit(STRIPE_OP_BIOFILL, &ops_request)) {
879 if (test_bit(STRIPE_OP_COMPUTE_BLK, &ops_request)) {
880 tx = ops_run_compute5(sh);
881 /* terminate the chain if postxor is not set to be run */
882 if (tx && !test_bit(STRIPE_OP_POSTXOR, &ops_request))
886 if (test_bit(STRIPE_OP_PREXOR, &ops_request))
887 tx = ops_run_prexor(sh, tx);
889 if (test_bit(STRIPE_OP_BIODRAIN, &ops_request)) {
890 tx = ops_run_biodrain(sh, tx);
894 if (test_bit(STRIPE_OP_POSTXOR, &ops_request))
895 ops_run_postxor(sh, tx);
897 if (test_bit(STRIPE_OP_CHECK, &ops_request))
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],
938 "raid%d-%s", conf->level, mdname(conf->mddev));
939 sprintf(conf->cache_name[1],
940 "raid%d-%s-alt", conf->level, mdname(conf->mddev));
941 conf->active_name = 0;
942 sc = kmem_cache_create(conf->cache_name[conf->active_name],
943 sizeof(struct stripe_head)+(devs-1)*sizeof(struct r5dev),
947 conf->slab_cache = sc;
948 conf->pool_size = devs;
950 if (!grow_one_stripe(conf))
955 static int resize_stripes(raid5_conf_t *conf, int newsize)
957 /* Make all the stripes able to hold 'newsize' devices.
958 * New slots in each stripe get 'page' set to a new page.
960 * This happens in stages:
961 * 1/ create a new kmem_cache and allocate the required number of
963 * 2/ gather all the old stripe_heads and tranfer the pages across
964 * to the new stripe_heads. This will have the side effect of
965 * freezing the array as once all stripe_heads have been collected,
966 * no IO will be possible. Old stripe heads are freed once their
967 * pages have been transferred over, and the old kmem_cache is
968 * freed when all stripes are done.
969 * 3/ reallocate conf->disks to be suitable bigger. If this fails,
970 * we simple return a failre status - no need to clean anything up.
971 * 4/ allocate new pages for the new slots in the new stripe_heads.
972 * If this fails, we don't bother trying the shrink the
973 * stripe_heads down again, we just leave them as they are.
974 * As each stripe_head is processed the new one is released into
977 * Once step2 is started, we cannot afford to wait for a write,
978 * so we use GFP_NOIO allocations.
980 struct stripe_head *osh, *nsh;
981 LIST_HEAD(newstripes);
982 struct disk_info *ndisks;
984 struct kmem_cache *sc;
987 if (newsize <= conf->pool_size)
988 return 0; /* never bother to shrink */
990 err = md_allow_write(conf->mddev);
995 sc = kmem_cache_create(conf->cache_name[1-conf->active_name],
996 sizeof(struct stripe_head)+(newsize-1)*sizeof(struct r5dev),
1001 for (i = conf->max_nr_stripes; i; i--) {
1002 nsh = kmem_cache_alloc(sc, GFP_KERNEL);
1006 memset(nsh, 0, sizeof(*nsh) + (newsize-1)*sizeof(struct r5dev));
1008 nsh->raid_conf = conf;
1009 spin_lock_init(&nsh->lock);
1011 list_add(&nsh->lru, &newstripes);
1014 /* didn't get enough, give up */
1015 while (!list_empty(&newstripes)) {
1016 nsh = list_entry(newstripes.next, struct stripe_head, lru);
1017 list_del(&nsh->lru);
1018 kmem_cache_free(sc, nsh);
1020 kmem_cache_destroy(sc);
1023 /* Step 2 - Must use GFP_NOIO now.
1024 * OK, we have enough stripes, start collecting inactive
1025 * stripes and copying them over
1027 list_for_each_entry(nsh, &newstripes, lru) {
1028 spin_lock_irq(&conf->device_lock);
1029 wait_event_lock_irq(conf->wait_for_stripe,
1030 !list_empty(&conf->inactive_list),
1032 unplug_slaves(conf->mddev)
1034 osh = get_free_stripe(conf);
1035 spin_unlock_irq(&conf->device_lock);
1036 atomic_set(&nsh->count, 1);
1037 for(i=0; i<conf->pool_size; i++)
1038 nsh->dev[i].page = osh->dev[i].page;
1039 for( ; i<newsize; i++)
1040 nsh->dev[i].page = NULL;
1041 kmem_cache_free(conf->slab_cache, osh);
1043 kmem_cache_destroy(conf->slab_cache);
1046 * At this point, we are holding all the stripes so the array
1047 * is completely stalled, so now is a good time to resize
1050 ndisks = kzalloc(newsize * sizeof(struct disk_info), GFP_NOIO);
1052 for (i=0; i<conf->raid_disks; i++)
1053 ndisks[i] = conf->disks[i];
1055 conf->disks = ndisks;
1059 /* Step 4, return new stripes to service */
1060 while(!list_empty(&newstripes)) {
1061 nsh = list_entry(newstripes.next, struct stripe_head, lru);
1062 list_del_init(&nsh->lru);
1063 for (i=conf->raid_disks; i < newsize; i++)
1064 if (nsh->dev[i].page == NULL) {
1065 struct page *p = alloc_page(GFP_NOIO);
1066 nsh->dev[i].page = p;
1070 release_stripe(nsh);
1072 /* critical section pass, GFP_NOIO no longer needed */
1074 conf->slab_cache = sc;
1075 conf->active_name = 1-conf->active_name;
1076 conf->pool_size = newsize;
1080 static int drop_one_stripe(raid5_conf_t *conf)
1082 struct stripe_head *sh;
1084 spin_lock_irq(&conf->device_lock);
1085 sh = get_free_stripe(conf);
1086 spin_unlock_irq(&conf->device_lock);
1089 BUG_ON(atomic_read(&sh->count));
1090 shrink_buffers(sh, conf->pool_size);
1091 kmem_cache_free(conf->slab_cache, sh);
1092 atomic_dec(&conf->active_stripes);
1096 static void shrink_stripes(raid5_conf_t *conf)
1098 while (drop_one_stripe(conf))
1101 if (conf->slab_cache)
1102 kmem_cache_destroy(conf->slab_cache);
1103 conf->slab_cache = NULL;
1106 static void raid5_end_read_request(struct bio * bi, int error)
1108 struct stripe_head *sh = bi->bi_private;
1109 raid5_conf_t *conf = sh->raid_conf;
1110 int disks = sh->disks, i;
1111 int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
1112 char b[BDEVNAME_SIZE];
1116 for (i=0 ; i<disks; i++)
1117 if (bi == &sh->dev[i].req)
1120 pr_debug("end_read_request %llu/%d, count: %d, uptodate %d.\n",
1121 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
1129 set_bit(R5_UPTODATE, &sh->dev[i].flags);
1130 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
1131 rdev = conf->disks[i].rdev;
1132 printk_rl(KERN_INFO "raid5:%s: read error corrected"
1133 " (%lu sectors at %llu on %s)\n",
1134 mdname(conf->mddev), STRIPE_SECTORS,
1135 (unsigned long long)(sh->sector
1136 + rdev->data_offset),
1137 bdevname(rdev->bdev, b));
1138 clear_bit(R5_ReadError, &sh->dev[i].flags);
1139 clear_bit(R5_ReWrite, &sh->dev[i].flags);
1141 if (atomic_read(&conf->disks[i].rdev->read_errors))
1142 atomic_set(&conf->disks[i].rdev->read_errors, 0);
1144 const char *bdn = bdevname(conf->disks[i].rdev->bdev, b);
1146 rdev = conf->disks[i].rdev;
1148 clear_bit(R5_UPTODATE, &sh->dev[i].flags);
1149 atomic_inc(&rdev->read_errors);
1150 if (conf->mddev->degraded)
1151 printk_rl(KERN_WARNING
1152 "raid5:%s: read error not correctable "
1153 "(sector %llu on %s).\n",
1154 mdname(conf->mddev),
1155 (unsigned long long)(sh->sector
1156 + rdev->data_offset),
1158 else if (test_bit(R5_ReWrite, &sh->dev[i].flags))
1160 printk_rl(KERN_WARNING
1161 "raid5:%s: read error NOT corrected!! "
1162 "(sector %llu on %s).\n",
1163 mdname(conf->mddev),
1164 (unsigned long long)(sh->sector
1165 + rdev->data_offset),
1167 else if (atomic_read(&rdev->read_errors)
1168 > conf->max_nr_stripes)
1170 "raid5:%s: Too many read errors, failing device %s.\n",
1171 mdname(conf->mddev), bdn);
1175 set_bit(R5_ReadError, &sh->dev[i].flags);
1177 clear_bit(R5_ReadError, &sh->dev[i].flags);
1178 clear_bit(R5_ReWrite, &sh->dev[i].flags);
1179 md_error(conf->mddev, rdev);
1182 rdev_dec_pending(conf->disks[i].rdev, conf->mddev);
1183 clear_bit(R5_LOCKED, &sh->dev[i].flags);
1184 set_bit(STRIPE_HANDLE, &sh->state);
1188 static void raid5_end_write_request(struct bio *bi, int error)
1190 struct stripe_head *sh = bi->bi_private;
1191 raid5_conf_t *conf = sh->raid_conf;
1192 int disks = sh->disks, i;
1193 int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
1195 for (i=0 ; i<disks; i++)
1196 if (bi == &sh->dev[i].req)
1199 pr_debug("end_write_request %llu/%d, count %d, uptodate: %d.\n",
1200 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
1208 md_error(conf->mddev, conf->disks[i].rdev);
1210 rdev_dec_pending(conf->disks[i].rdev, conf->mddev);
1212 clear_bit(R5_LOCKED, &sh->dev[i].flags);
1213 set_bit(STRIPE_HANDLE, &sh->state);
1218 static sector_t compute_blocknr(struct stripe_head *sh, int i, int previous);
1220 static void raid5_build_block(struct stripe_head *sh, int i, int previous)
1222 struct r5dev *dev = &sh->dev[i];
1224 bio_init(&dev->req);
1225 dev->req.bi_io_vec = &dev->vec;
1227 dev->req.bi_max_vecs++;
1228 dev->vec.bv_page = dev->page;
1229 dev->vec.bv_len = STRIPE_SIZE;
1230 dev->vec.bv_offset = 0;
1232 dev->req.bi_sector = sh->sector;
1233 dev->req.bi_private = sh;
1236 dev->sector = compute_blocknr(sh, i, previous);
1239 static void error(mddev_t *mddev, mdk_rdev_t *rdev)
1241 char b[BDEVNAME_SIZE];
1242 raid5_conf_t *conf = (raid5_conf_t *) mddev->private;
1243 pr_debug("raid5: error called\n");
1245 if (!test_bit(Faulty, &rdev->flags)) {
1246 set_bit(MD_CHANGE_DEVS, &mddev->flags);
1247 if (test_and_clear_bit(In_sync, &rdev->flags)) {
1248 unsigned long flags;
1249 spin_lock_irqsave(&conf->device_lock, flags);
1251 spin_unlock_irqrestore(&conf->device_lock, flags);
1253 * if recovery was running, make sure it aborts.
1255 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1257 set_bit(Faulty, &rdev->flags);
1259 "raid5: Disk failure on %s, disabling device.\n"
1260 "raid5: Operation continuing on %d devices.\n",
1261 bdevname(rdev->bdev,b), conf->raid_disks - mddev->degraded);
1266 * Input: a 'big' sector number,
1267 * Output: index of the data and parity disk, and the sector # in them.
1269 static sector_t raid5_compute_sector(raid5_conf_t *conf, sector_t r_sector,
1270 int previous, int *dd_idx,
1271 struct stripe_head *sh)
1274 unsigned long chunk_number;
1275 unsigned int chunk_offset;
1278 sector_t new_sector;
1279 int algorithm = previous ? conf->prev_algo
1281 int sectors_per_chunk = previous ? (conf->prev_chunk >> 9)
1282 : (conf->chunk_size >> 9);
1283 int raid_disks = previous ? conf->previous_raid_disks
1285 int data_disks = raid_disks - conf->max_degraded;
1287 /* First compute the information on this sector */
1290 * Compute the chunk number and the sector offset inside the chunk
1292 chunk_offset = sector_div(r_sector, sectors_per_chunk);
1293 chunk_number = r_sector;
1294 BUG_ON(r_sector != chunk_number);
1297 * Compute the stripe number
1299 stripe = chunk_number / data_disks;
1302 * Compute the data disk and parity disk indexes inside the stripe
1304 *dd_idx = chunk_number % data_disks;
1307 * Select the parity disk based on the user selected algorithm.
1309 pd_idx = qd_idx = ~0;
1310 switch(conf->level) {
1312 pd_idx = data_disks;
1315 switch (algorithm) {
1316 case ALGORITHM_LEFT_ASYMMETRIC:
1317 pd_idx = data_disks - stripe % raid_disks;
1318 if (*dd_idx >= pd_idx)
1321 case ALGORITHM_RIGHT_ASYMMETRIC:
1322 pd_idx = stripe % raid_disks;
1323 if (*dd_idx >= pd_idx)
1326 case ALGORITHM_LEFT_SYMMETRIC:
1327 pd_idx = data_disks - stripe % raid_disks;
1328 *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
1330 case ALGORITHM_RIGHT_SYMMETRIC:
1331 pd_idx = stripe % raid_disks;
1332 *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
1334 case ALGORITHM_PARITY_0:
1338 case ALGORITHM_PARITY_N:
1339 pd_idx = data_disks;
1342 printk(KERN_ERR "raid5: unsupported algorithm %d\n",
1349 switch (algorithm) {
1350 case ALGORITHM_LEFT_ASYMMETRIC:
1351 pd_idx = raid_disks - 1 - (stripe % raid_disks);
1352 qd_idx = pd_idx + 1;
1353 if (pd_idx == raid_disks-1) {
1354 (*dd_idx)++; /* Q D D D P */
1356 } else if (*dd_idx >= pd_idx)
1357 (*dd_idx) += 2; /* D D P Q D */
1359 case ALGORITHM_RIGHT_ASYMMETRIC:
1360 pd_idx = stripe % raid_disks;
1361 qd_idx = pd_idx + 1;
1362 if (pd_idx == raid_disks-1) {
1363 (*dd_idx)++; /* Q D D D P */
1365 } else if (*dd_idx >= pd_idx)
1366 (*dd_idx) += 2; /* D D P Q D */
1368 case ALGORITHM_LEFT_SYMMETRIC:
1369 pd_idx = raid_disks - 1 - (stripe % raid_disks);
1370 qd_idx = (pd_idx + 1) % raid_disks;
1371 *dd_idx = (pd_idx + 2 + *dd_idx) % raid_disks;
1373 case ALGORITHM_RIGHT_SYMMETRIC:
1374 pd_idx = stripe % raid_disks;
1375 qd_idx = (pd_idx + 1) % raid_disks;
1376 *dd_idx = (pd_idx + 2 + *dd_idx) % raid_disks;
1379 case ALGORITHM_PARITY_0:
1384 case ALGORITHM_PARITY_N:
1385 pd_idx = data_disks;
1386 qd_idx = data_disks + 1;
1389 case ALGORITHM_ROTATING_ZERO_RESTART:
1390 /* Exactly the same as RIGHT_ASYMMETRIC, but or
1391 * of blocks for computing Q is different.
1393 pd_idx = stripe % raid_disks;
1394 qd_idx = pd_idx + 1;
1395 if (pd_idx == raid_disks-1) {
1396 (*dd_idx)++; /* Q D D D P */
1398 } else if (*dd_idx >= pd_idx)
1399 (*dd_idx) += 2; /* D D P Q D */
1403 case ALGORITHM_ROTATING_N_RESTART:
1404 /* Same a left_asymmetric, by first stripe is
1405 * D D D P Q rather than
1408 pd_idx = raid_disks - 1 - ((stripe + 1) % raid_disks);
1409 qd_idx = pd_idx + 1;
1410 if (pd_idx == raid_disks-1) {
1411 (*dd_idx)++; /* Q D D D P */
1413 } else if (*dd_idx >= pd_idx)
1414 (*dd_idx) += 2; /* D D P Q D */
1418 case ALGORITHM_ROTATING_N_CONTINUE:
1419 /* Same as left_symmetric but Q is before P */
1420 pd_idx = raid_disks - 1 - (stripe % raid_disks);
1421 qd_idx = (pd_idx + raid_disks - 1) % raid_disks;
1422 *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
1426 case ALGORITHM_LEFT_ASYMMETRIC_6:
1427 /* RAID5 left_asymmetric, with Q on last device */
1428 pd_idx = data_disks - stripe % (raid_disks-1);
1429 if (*dd_idx >= pd_idx)
1431 qd_idx = raid_disks - 1;
1434 case ALGORITHM_RIGHT_ASYMMETRIC_6:
1435 pd_idx = stripe % (raid_disks-1);
1436 if (*dd_idx >= pd_idx)
1438 qd_idx = raid_disks - 1;
1441 case ALGORITHM_LEFT_SYMMETRIC_6:
1442 pd_idx = data_disks - stripe % (raid_disks-1);
1443 *dd_idx = (pd_idx + 1 + *dd_idx) % (raid_disks-1);
1444 qd_idx = raid_disks - 1;
1447 case ALGORITHM_RIGHT_SYMMETRIC_6:
1448 pd_idx = stripe % (raid_disks-1);
1449 *dd_idx = (pd_idx + 1 + *dd_idx) % (raid_disks-1);
1450 qd_idx = raid_disks - 1;
1453 case ALGORITHM_PARITY_0_6:
1456 qd_idx = raid_disks - 1;
1461 printk(KERN_CRIT "raid6: unsupported algorithm %d\n",
1469 sh->pd_idx = pd_idx;
1470 sh->qd_idx = qd_idx;
1471 sh->ddf_layout = ddf_layout;
1474 * Finally, compute the new sector number
1476 new_sector = (sector_t)stripe * sectors_per_chunk + chunk_offset;
1481 static sector_t compute_blocknr(struct stripe_head *sh, int i, int previous)
1483 raid5_conf_t *conf = sh->raid_conf;
1484 int raid_disks = sh->disks;
1485 int data_disks = raid_disks - conf->max_degraded;
1486 sector_t new_sector = sh->sector, check;
1487 int sectors_per_chunk = previous ? (conf->prev_chunk >> 9)
1488 : (conf->chunk_size >> 9);
1489 int algorithm = previous ? conf->prev_algo
1493 int chunk_number, dummy1, dd_idx = i;
1495 struct stripe_head sh2;
1498 chunk_offset = sector_div(new_sector, sectors_per_chunk);
1499 stripe = new_sector;
1500 BUG_ON(new_sector != stripe);
1502 if (i == sh->pd_idx)
1504 switch(conf->level) {
1507 switch (algorithm) {
1508 case ALGORITHM_LEFT_ASYMMETRIC:
1509 case ALGORITHM_RIGHT_ASYMMETRIC:
1513 case ALGORITHM_LEFT_SYMMETRIC:
1514 case ALGORITHM_RIGHT_SYMMETRIC:
1517 i -= (sh->pd_idx + 1);
1519 case ALGORITHM_PARITY_0:
1522 case ALGORITHM_PARITY_N:
1525 printk(KERN_ERR "raid5: unsupported algorithm %d\n",
1531 if (i == sh->qd_idx)
1532 return 0; /* It is the Q disk */
1533 switch (algorithm) {
1534 case ALGORITHM_LEFT_ASYMMETRIC:
1535 case ALGORITHM_RIGHT_ASYMMETRIC:
1536 case ALGORITHM_ROTATING_ZERO_RESTART:
1537 case ALGORITHM_ROTATING_N_RESTART:
1538 if (sh->pd_idx == raid_disks-1)
1539 i--; /* Q D D D P */
1540 else if (i > sh->pd_idx)
1541 i -= 2; /* D D P Q D */
1543 case ALGORITHM_LEFT_SYMMETRIC:
1544 case ALGORITHM_RIGHT_SYMMETRIC:
1545 if (sh->pd_idx == raid_disks-1)
1546 i--; /* Q D D D P */
1551 i -= (sh->pd_idx + 2);
1554 case ALGORITHM_PARITY_0:
1557 case ALGORITHM_PARITY_N:
1559 case ALGORITHM_ROTATING_N_CONTINUE:
1560 if (sh->pd_idx == 0)
1561 i--; /* P D D D Q */
1562 else if (i > sh->pd_idx)
1563 i -= 2; /* D D Q P D */
1565 case ALGORITHM_LEFT_ASYMMETRIC_6:
1566 case ALGORITHM_RIGHT_ASYMMETRIC_6:
1570 case ALGORITHM_LEFT_SYMMETRIC_6:
1571 case ALGORITHM_RIGHT_SYMMETRIC_6:
1573 i += data_disks + 1;
1574 i -= (sh->pd_idx + 1);
1576 case ALGORITHM_PARITY_0_6:
1580 printk(KERN_CRIT "raid6: unsupported algorithm %d\n",
1587 chunk_number = stripe * data_disks + i;
1588 r_sector = (sector_t)chunk_number * sectors_per_chunk + chunk_offset;
1590 check = raid5_compute_sector(conf, r_sector,
1591 previous, &dummy1, &sh2);
1592 if (check != sh->sector || dummy1 != dd_idx || sh2.pd_idx != sh->pd_idx
1593 || sh2.qd_idx != sh->qd_idx) {
1594 printk(KERN_ERR "compute_blocknr: map not correct\n");
1603 * Copy data between a page in the stripe cache, and one or more bion
1604 * The page could align with the middle of the bio, or there could be
1605 * several bion, each with several bio_vecs, which cover part of the page
1606 * Multiple bion are linked together on bi_next. There may be extras
1607 * at the end of this list. We ignore them.
1609 static void copy_data(int frombio, struct bio *bio,
1613 char *pa = page_address(page);
1614 struct bio_vec *bvl;
1618 if (bio->bi_sector >= sector)
1619 page_offset = (signed)(bio->bi_sector - sector) * 512;
1621 page_offset = (signed)(sector - bio->bi_sector) * -512;
1622 bio_for_each_segment(bvl, bio, i) {
1623 int len = bio_iovec_idx(bio,i)->bv_len;
1627 if (page_offset < 0) {
1628 b_offset = -page_offset;
1629 page_offset += b_offset;
1633 if (len > 0 && page_offset + len > STRIPE_SIZE)
1634 clen = STRIPE_SIZE - page_offset;
1638 char *ba = __bio_kmap_atomic(bio, i, KM_USER0);
1640 memcpy(pa+page_offset, ba+b_offset, clen);
1642 memcpy(ba+b_offset, pa+page_offset, clen);
1643 __bio_kunmap_atomic(ba, KM_USER0);
1645 if (clen < len) /* hit end of page */
1651 #define check_xor() do { \
1652 if (count == MAX_XOR_BLOCKS) { \
1653 xor_blocks(count, STRIPE_SIZE, dest, ptr);\
1658 static void compute_parity6(struct stripe_head *sh, int method)
1660 raid5_conf_t *conf = sh->raid_conf;
1661 int i, pd_idx, qd_idx, d0_idx, disks = sh->disks, count;
1662 int syndrome_disks = sh->ddf_layout ? disks : (disks - 2);
1664 /**** FIX THIS: This could be very bad if disks is close to 256 ****/
1665 void *ptrs[syndrome_disks+2];
1667 pd_idx = sh->pd_idx;
1668 qd_idx = sh->qd_idx;
1669 d0_idx = raid6_d0(sh);
1671 pr_debug("compute_parity, stripe %llu, method %d\n",
1672 (unsigned long long)sh->sector, method);
1675 case READ_MODIFY_WRITE:
1676 BUG(); /* READ_MODIFY_WRITE N/A for RAID-6 */
1677 case RECONSTRUCT_WRITE:
1678 for (i= disks; i-- ;)
1679 if ( i != pd_idx && i != qd_idx && sh->dev[i].towrite ) {
1680 chosen = sh->dev[i].towrite;
1681 sh->dev[i].towrite = NULL;
1683 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
1684 wake_up(&conf->wait_for_overlap);
1686 BUG_ON(sh->dev[i].written);
1687 sh->dev[i].written = chosen;
1691 BUG(); /* Not implemented yet */
1694 for (i = disks; i--;)
1695 if (sh->dev[i].written) {
1696 sector_t sector = sh->dev[i].sector;
1697 struct bio *wbi = sh->dev[i].written;
1698 while (wbi && wbi->bi_sector < sector + STRIPE_SECTORS) {
1699 copy_data(1, wbi, sh->dev[i].page, sector);
1700 wbi = r5_next_bio(wbi, sector);
1703 set_bit(R5_LOCKED, &sh->dev[i].flags);
1704 set_bit(R5_UPTODATE, &sh->dev[i].flags);
1707 /* Note that unlike RAID-5, the ordering of the disks matters greatly.*/
1709 for (i = 0; i < disks; i++)
1710 ptrs[i] = (void *)raid6_empty_zero_page;
1715 int slot = raid6_idx_to_slot(i, sh, &count, syndrome_disks);
1717 ptrs[slot] = page_address(sh->dev[i].page);
1718 if (slot < syndrome_disks &&
1719 !test_bit(R5_UPTODATE, &sh->dev[i].flags)) {
1720 printk(KERN_ERR "block %d/%d not uptodate "
1721 "on parity calc\n", i, count);
1725 i = raid6_next_disk(i, disks);
1726 } while (i != d0_idx);
1727 BUG_ON(count != syndrome_disks);
1729 raid6_call.gen_syndrome(syndrome_disks+2, STRIPE_SIZE, ptrs);
1732 case RECONSTRUCT_WRITE:
1733 set_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
1734 set_bit(R5_UPTODATE, &sh->dev[qd_idx].flags);
1735 set_bit(R5_LOCKED, &sh->dev[pd_idx].flags);
1736 set_bit(R5_LOCKED, &sh->dev[qd_idx].flags);
1739 set_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
1740 set_bit(R5_UPTODATE, &sh->dev[qd_idx].flags);
1746 /* Compute one missing block */
1747 static void compute_block_1(struct stripe_head *sh, int dd_idx, int nozero)
1749 int i, count, disks = sh->disks;
1750 void *ptr[MAX_XOR_BLOCKS], *dest, *p;
1751 int qd_idx = sh->qd_idx;
1753 pr_debug("compute_block_1, stripe %llu, idx %d\n",
1754 (unsigned long long)sh->sector, dd_idx);
1756 if ( dd_idx == qd_idx ) {
1757 /* We're actually computing the Q drive */
1758 compute_parity6(sh, UPDATE_PARITY);
1760 dest = page_address(sh->dev[dd_idx].page);
1761 if (!nozero) memset(dest, 0, STRIPE_SIZE);
1763 for (i = disks ; i--; ) {
1764 if (i == dd_idx || i == qd_idx)
1766 p = page_address(sh->dev[i].page);
1767 if (test_bit(R5_UPTODATE, &sh->dev[i].flags))
1770 printk("compute_block() %d, stripe %llu, %d"
1771 " not present\n", dd_idx,
1772 (unsigned long long)sh->sector, i);
1777 xor_blocks(count, STRIPE_SIZE, dest, ptr);
1778 if (!nozero) set_bit(R5_UPTODATE, &sh->dev[dd_idx].flags);
1779 else clear_bit(R5_UPTODATE, &sh->dev[dd_idx].flags);
1783 /* Compute two missing blocks */
1784 static void compute_block_2(struct stripe_head *sh, int dd_idx1, int dd_idx2)
1786 int i, count, disks = sh->disks;
1787 int syndrome_disks = sh->ddf_layout ? disks : disks-2;
1788 int d0_idx = raid6_d0(sh);
1789 int faila = -1, failb = -1;
1790 /**** FIX THIS: This could be very bad if disks is close to 256 ****/
1791 void *ptrs[syndrome_disks+2];
1793 for (i = 0; i < disks ; i++)
1794 ptrs[i] = (void *)raid6_empty_zero_page;
1798 int slot = raid6_idx_to_slot(i, sh, &count, syndrome_disks);
1800 ptrs[slot] = page_address(sh->dev[i].page);
1806 i = raid6_next_disk(i, disks);
1807 } while (i != d0_idx);
1808 BUG_ON(count != syndrome_disks);
1810 BUG_ON(faila == failb);
1811 if ( failb < faila ) { int tmp = faila; faila = failb; failb = tmp; }
1813 pr_debug("compute_block_2, stripe %llu, idx %d,%d (%d,%d)\n",
1814 (unsigned long long)sh->sector, dd_idx1, dd_idx2,
1817 if (failb == syndrome_disks+1) {
1818 /* Q disk is one of the missing disks */
1819 if (faila == syndrome_disks) {
1820 /* Missing P+Q, just recompute */
1821 compute_parity6(sh, UPDATE_PARITY);
1824 /* We're missing D+Q; recompute D from P */
1825 compute_block_1(sh, ((dd_idx1 == sh->qd_idx) ?
1828 compute_parity6(sh, UPDATE_PARITY); /* Is this necessary? */
1833 /* We're missing D+P or D+D; */
1834 if (failb == syndrome_disks) {
1835 /* We're missing D+P. */
1836 raid6_datap_recov(syndrome_disks+2, STRIPE_SIZE, faila, ptrs);
1838 /* We're missing D+D. */
1839 raid6_2data_recov(syndrome_disks+2, STRIPE_SIZE, faila, failb,
1843 /* Both the above update both missing blocks */
1844 set_bit(R5_UPTODATE, &sh->dev[dd_idx1].flags);
1845 set_bit(R5_UPTODATE, &sh->dev[dd_idx2].flags);
1849 schedule_reconstruction5(struct stripe_head *sh, struct stripe_head_state *s,
1850 int rcw, int expand)
1852 int i, pd_idx = sh->pd_idx, disks = sh->disks;
1855 /* if we are not expanding this is a proper write request, and
1856 * there will be bios with new data to be drained into the
1860 sh->reconstruct_state = reconstruct_state_drain_run;
1861 set_bit(STRIPE_OP_BIODRAIN, &s->ops_request);
1863 sh->reconstruct_state = reconstruct_state_run;
1865 set_bit(STRIPE_OP_POSTXOR, &s->ops_request);
1867 for (i = disks; i--; ) {
1868 struct r5dev *dev = &sh->dev[i];
1871 set_bit(R5_LOCKED, &dev->flags);
1872 set_bit(R5_Wantdrain, &dev->flags);
1874 clear_bit(R5_UPTODATE, &dev->flags);
1878 if (s->locked + 1 == disks)
1879 if (!test_and_set_bit(STRIPE_FULL_WRITE, &sh->state))
1880 atomic_inc(&sh->raid_conf->pending_full_writes);
1882 BUG_ON(!(test_bit(R5_UPTODATE, &sh->dev[pd_idx].flags) ||
1883 test_bit(R5_Wantcompute, &sh->dev[pd_idx].flags)));
1885 sh->reconstruct_state = reconstruct_state_prexor_drain_run;
1886 set_bit(STRIPE_OP_PREXOR, &s->ops_request);
1887 set_bit(STRIPE_OP_BIODRAIN, &s->ops_request);
1888 set_bit(STRIPE_OP_POSTXOR, &s->ops_request);
1890 for (i = disks; i--; ) {
1891 struct r5dev *dev = &sh->dev[i];
1896 (test_bit(R5_UPTODATE, &dev->flags) ||
1897 test_bit(R5_Wantcompute, &dev->flags))) {
1898 set_bit(R5_Wantdrain, &dev->flags);
1899 set_bit(R5_LOCKED, &dev->flags);
1900 clear_bit(R5_UPTODATE, &dev->flags);
1906 /* keep the parity disk locked while asynchronous operations
1909 set_bit(R5_LOCKED, &sh->dev[pd_idx].flags);
1910 clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
1913 pr_debug("%s: stripe %llu locked: %d ops_request: %lx\n",
1914 __func__, (unsigned long long)sh->sector,
1915 s->locked, s->ops_request);
1919 * Each stripe/dev can have one or more bion attached.
1920 * toread/towrite point to the first in a chain.
1921 * The bi_next chain must be in order.
1923 static int add_stripe_bio(struct stripe_head *sh, struct bio *bi, int dd_idx, int forwrite)
1926 raid5_conf_t *conf = sh->raid_conf;
1929 pr_debug("adding bh b#%llu to stripe s#%llu\n",
1930 (unsigned long long)bi->bi_sector,
1931 (unsigned long long)sh->sector);
1934 spin_lock(&sh->lock);
1935 spin_lock_irq(&conf->device_lock);
1937 bip = &sh->dev[dd_idx].towrite;
1938 if (*bip == NULL && sh->dev[dd_idx].written == NULL)
1941 bip = &sh->dev[dd_idx].toread;
1942 while (*bip && (*bip)->bi_sector < bi->bi_sector) {
1943 if ((*bip)->bi_sector + ((*bip)->bi_size >> 9) > bi->bi_sector)
1945 bip = & (*bip)->bi_next;
1947 if (*bip && (*bip)->bi_sector < bi->bi_sector + ((bi->bi_size)>>9))
1950 BUG_ON(*bip && bi->bi_next && (*bip) != bi->bi_next);
1954 bi->bi_phys_segments++;
1955 spin_unlock_irq(&conf->device_lock);
1956 spin_unlock(&sh->lock);
1958 pr_debug("added bi b#%llu to stripe s#%llu, disk %d.\n",
1959 (unsigned long long)bi->bi_sector,
1960 (unsigned long long)sh->sector, dd_idx);
1962 if (conf->mddev->bitmap && firstwrite) {
1963 bitmap_startwrite(conf->mddev->bitmap, sh->sector,
1965 sh->bm_seq = conf->seq_flush+1;
1966 set_bit(STRIPE_BIT_DELAY, &sh->state);
1970 /* check if page is covered */
1971 sector_t sector = sh->dev[dd_idx].sector;
1972 for (bi=sh->dev[dd_idx].towrite;
1973 sector < sh->dev[dd_idx].sector + STRIPE_SECTORS &&
1974 bi && bi->bi_sector <= sector;
1975 bi = r5_next_bio(bi, sh->dev[dd_idx].sector)) {
1976 if (bi->bi_sector + (bi->bi_size>>9) >= sector)
1977 sector = bi->bi_sector + (bi->bi_size>>9);
1979 if (sector >= sh->dev[dd_idx].sector + STRIPE_SECTORS)
1980 set_bit(R5_OVERWRITE, &sh->dev[dd_idx].flags);
1985 set_bit(R5_Overlap, &sh->dev[dd_idx].flags);
1986 spin_unlock_irq(&conf->device_lock);
1987 spin_unlock(&sh->lock);
1991 static void end_reshape(raid5_conf_t *conf);
1993 static int page_is_zero(struct page *p)
1995 char *a = page_address(p);
1996 return ((*(u32*)a) == 0 &&
1997 memcmp(a, a+4, STRIPE_SIZE-4)==0);
2000 static void stripe_set_idx(sector_t stripe, raid5_conf_t *conf, int previous,
2001 struct stripe_head *sh)
2003 int sectors_per_chunk =
2004 previous ? (conf->prev_chunk >> 9)
2005 : (conf->chunk_size >> 9);
2007 int chunk_offset = sector_div(stripe, sectors_per_chunk);
2008 int disks = previous ? conf->previous_raid_disks : conf->raid_disks;
2010 raid5_compute_sector(conf,
2011 stripe * (disks - conf->max_degraded)
2012 *sectors_per_chunk + chunk_offset,
2018 handle_failed_stripe(raid5_conf_t *conf, struct stripe_head *sh,
2019 struct stripe_head_state *s, int disks,
2020 struct bio **return_bi)
2023 for (i = disks; i--; ) {
2027 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
2030 rdev = rcu_dereference(conf->disks[i].rdev);
2031 if (rdev && test_bit(In_sync, &rdev->flags))
2032 /* multiple read failures in one stripe */
2033 md_error(conf->mddev, rdev);
2036 spin_lock_irq(&conf->device_lock);
2037 /* fail all writes first */
2038 bi = sh->dev[i].towrite;
2039 sh->dev[i].towrite = NULL;
2045 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
2046 wake_up(&conf->wait_for_overlap);
2048 while (bi && bi->bi_sector <
2049 sh->dev[i].sector + STRIPE_SECTORS) {
2050 struct bio *nextbi = r5_next_bio(bi, sh->dev[i].sector);
2051 clear_bit(BIO_UPTODATE, &bi->bi_flags);
2052 if (!raid5_dec_bi_phys_segments(bi)) {
2053 md_write_end(conf->mddev);
2054 bi->bi_next = *return_bi;
2059 /* and fail all 'written' */
2060 bi = sh->dev[i].written;
2061 sh->dev[i].written = NULL;
2062 if (bi) bitmap_end = 1;
2063 while (bi && bi->bi_sector <
2064 sh->dev[i].sector + STRIPE_SECTORS) {
2065 struct bio *bi2 = r5_next_bio(bi, sh->dev[i].sector);
2066 clear_bit(BIO_UPTODATE, &bi->bi_flags);
2067 if (!raid5_dec_bi_phys_segments(bi)) {
2068 md_write_end(conf->mddev);
2069 bi->bi_next = *return_bi;
2075 /* fail any reads if this device is non-operational and
2076 * the data has not reached the cache yet.
2078 if (!test_bit(R5_Wantfill, &sh->dev[i].flags) &&
2079 (!test_bit(R5_Insync, &sh->dev[i].flags) ||
2080 test_bit(R5_ReadError, &sh->dev[i].flags))) {
2081 bi = sh->dev[i].toread;
2082 sh->dev[i].toread = NULL;
2083 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
2084 wake_up(&conf->wait_for_overlap);
2085 if (bi) s->to_read--;
2086 while (bi && bi->bi_sector <
2087 sh->dev[i].sector + STRIPE_SECTORS) {
2088 struct bio *nextbi =
2089 r5_next_bio(bi, sh->dev[i].sector);
2090 clear_bit(BIO_UPTODATE, &bi->bi_flags);
2091 if (!raid5_dec_bi_phys_segments(bi)) {
2092 bi->bi_next = *return_bi;
2098 spin_unlock_irq(&conf->device_lock);
2100 bitmap_endwrite(conf->mddev->bitmap, sh->sector,
2101 STRIPE_SECTORS, 0, 0);
2104 if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
2105 if (atomic_dec_and_test(&conf->pending_full_writes))
2106 md_wakeup_thread(conf->mddev->thread);
2109 /* fetch_block5 - checks the given member device to see if its data needs
2110 * to be read or computed to satisfy a request.
2112 * Returns 1 when no more member devices need to be checked, otherwise returns
2113 * 0 to tell the loop in handle_stripe_fill5 to continue
2115 static int fetch_block5(struct stripe_head *sh, struct stripe_head_state *s,
2116 int disk_idx, int disks)
2118 struct r5dev *dev = &sh->dev[disk_idx];
2119 struct r5dev *failed_dev = &sh->dev[s->failed_num];
2121 /* is the data in this block needed, and can we get it? */
2122 if (!test_bit(R5_LOCKED, &dev->flags) &&
2123 !test_bit(R5_UPTODATE, &dev->flags) &&
2125 (dev->towrite && !test_bit(R5_OVERWRITE, &dev->flags)) ||
2126 s->syncing || s->expanding ||
2128 (failed_dev->toread ||
2129 (failed_dev->towrite &&
2130 !test_bit(R5_OVERWRITE, &failed_dev->flags)))))) {
2131 /* We would like to get this block, possibly by computing it,
2132 * otherwise read it if the backing disk is insync
2134 if ((s->uptodate == disks - 1) &&
2135 (s->failed && disk_idx == s->failed_num)) {
2136 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
2137 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
2138 set_bit(R5_Wantcompute, &dev->flags);
2139 sh->ops.target = disk_idx;
2141 /* Careful: from this point on 'uptodate' is in the eye
2142 * of raid5_run_ops which services 'compute' operations
2143 * before writes. R5_Wantcompute flags a block that will
2144 * be R5_UPTODATE by the time it is needed for a
2145 * subsequent operation.
2148 return 1; /* uptodate + compute == disks */
2149 } else if (test_bit(R5_Insync, &dev->flags)) {
2150 set_bit(R5_LOCKED, &dev->flags);
2151 set_bit(R5_Wantread, &dev->flags);
2153 pr_debug("Reading block %d (sync=%d)\n", disk_idx,
2162 * handle_stripe_fill5 - read or compute data to satisfy pending requests.
2164 static void handle_stripe_fill5(struct stripe_head *sh,
2165 struct stripe_head_state *s, int disks)
2169 /* look for blocks to read/compute, skip this if a compute
2170 * is already in flight, or if the stripe contents are in the
2171 * midst of changing due to a write
2173 if (!test_bit(STRIPE_COMPUTE_RUN, &sh->state) && !sh->check_state &&
2174 !sh->reconstruct_state)
2175 for (i = disks; i--; )
2176 if (fetch_block5(sh, s, i, disks))
2178 set_bit(STRIPE_HANDLE, &sh->state);
2181 static void handle_stripe_fill6(struct stripe_head *sh,
2182 struct stripe_head_state *s, struct r6_state *r6s,
2186 for (i = disks; i--; ) {
2187 struct r5dev *dev = &sh->dev[i];
2188 if (!test_bit(R5_LOCKED, &dev->flags) &&
2189 !test_bit(R5_UPTODATE, &dev->flags) &&
2190 (dev->toread || (dev->towrite &&
2191 !test_bit(R5_OVERWRITE, &dev->flags)) ||
2192 s->syncing || s->expanding ||
2194 (sh->dev[r6s->failed_num[0]].toread ||
2197 (sh->dev[r6s->failed_num[1]].toread ||
2199 /* we would like to get this block, possibly
2200 * by computing it, but we might not be able to
2202 if ((s->uptodate == disks - 1) &&
2203 (s->failed && (i == r6s->failed_num[0] ||
2204 i == r6s->failed_num[1]))) {
2205 pr_debug("Computing stripe %llu block %d\n",
2206 (unsigned long long)sh->sector, i);
2207 compute_block_1(sh, i, 0);
2209 } else if ( s->uptodate == disks-2 && s->failed >= 2 ) {
2210 /* Computing 2-failure is *very* expensive; only
2211 * do it if failed >= 2
2214 for (other = disks; other--; ) {
2217 if (!test_bit(R5_UPTODATE,
2218 &sh->dev[other].flags))
2222 pr_debug("Computing stripe %llu blocks %d,%d\n",
2223 (unsigned long long)sh->sector,
2225 compute_block_2(sh, i, other);
2227 } else if (test_bit(R5_Insync, &dev->flags)) {
2228 set_bit(R5_LOCKED, &dev->flags);
2229 set_bit(R5_Wantread, &dev->flags);
2231 pr_debug("Reading block %d (sync=%d)\n",
2236 set_bit(STRIPE_HANDLE, &sh->state);
2240 /* handle_stripe_clean_event
2241 * any written block on an uptodate or failed drive can be returned.
2242 * Note that if we 'wrote' to a failed drive, it will be UPTODATE, but
2243 * never LOCKED, so we don't need to test 'failed' directly.
2245 static void handle_stripe_clean_event(raid5_conf_t *conf,
2246 struct stripe_head *sh, int disks, struct bio **return_bi)
2251 for (i = disks; i--; )
2252 if (sh->dev[i].written) {
2254 if (!test_bit(R5_LOCKED, &dev->flags) &&
2255 test_bit(R5_UPTODATE, &dev->flags)) {
2256 /* We can return any write requests */
2257 struct bio *wbi, *wbi2;
2259 pr_debug("Return write for disc %d\n", i);
2260 spin_lock_irq(&conf->device_lock);
2262 dev->written = NULL;
2263 while (wbi && wbi->bi_sector <
2264 dev->sector + STRIPE_SECTORS) {
2265 wbi2 = r5_next_bio(wbi, dev->sector);
2266 if (!raid5_dec_bi_phys_segments(wbi)) {
2267 md_write_end(conf->mddev);
2268 wbi->bi_next = *return_bi;
2273 if (dev->towrite == NULL)
2275 spin_unlock_irq(&conf->device_lock);
2277 bitmap_endwrite(conf->mddev->bitmap,
2280 !test_bit(STRIPE_DEGRADED, &sh->state),
2285 if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
2286 if (atomic_dec_and_test(&conf->pending_full_writes))
2287 md_wakeup_thread(conf->mddev->thread);
2290 static void handle_stripe_dirtying5(raid5_conf_t *conf,
2291 struct stripe_head *sh, struct stripe_head_state *s, int disks)
2293 int rmw = 0, rcw = 0, i;
2294 for (i = disks; i--; ) {
2295 /* would I have to read this buffer for read_modify_write */
2296 struct r5dev *dev = &sh->dev[i];
2297 if ((dev->towrite || i == sh->pd_idx) &&
2298 !test_bit(R5_LOCKED, &dev->flags) &&
2299 !(test_bit(R5_UPTODATE, &dev->flags) ||
2300 test_bit(R5_Wantcompute, &dev->flags))) {
2301 if (test_bit(R5_Insync, &dev->flags))
2304 rmw += 2*disks; /* cannot read it */
2306 /* Would I have to read this buffer for reconstruct_write */
2307 if (!test_bit(R5_OVERWRITE, &dev->flags) && i != sh->pd_idx &&
2308 !test_bit(R5_LOCKED, &dev->flags) &&
2309 !(test_bit(R5_UPTODATE, &dev->flags) ||
2310 test_bit(R5_Wantcompute, &dev->flags))) {
2311 if (test_bit(R5_Insync, &dev->flags)) rcw++;
2316 pr_debug("for sector %llu, rmw=%d rcw=%d\n",
2317 (unsigned long long)sh->sector, rmw, rcw);
2318 set_bit(STRIPE_HANDLE, &sh->state);
2319 if (rmw < rcw && rmw > 0)
2320 /* prefer read-modify-write, but need to get some data */
2321 for (i = disks; i--; ) {
2322 struct r5dev *dev = &sh->dev[i];
2323 if ((dev->towrite || i == sh->pd_idx) &&
2324 !test_bit(R5_LOCKED, &dev->flags) &&
2325 !(test_bit(R5_UPTODATE, &dev->flags) ||
2326 test_bit(R5_Wantcompute, &dev->flags)) &&
2327 test_bit(R5_Insync, &dev->flags)) {
2329 test_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2330 pr_debug("Read_old block "
2331 "%d for r-m-w\n", i);
2332 set_bit(R5_LOCKED, &dev->flags);
2333 set_bit(R5_Wantread, &dev->flags);
2336 set_bit(STRIPE_DELAYED, &sh->state);
2337 set_bit(STRIPE_HANDLE, &sh->state);
2341 if (rcw <= rmw && rcw > 0)
2342 /* want reconstruct write, but need to get some data */
2343 for (i = disks; i--; ) {
2344 struct r5dev *dev = &sh->dev[i];
2345 if (!test_bit(R5_OVERWRITE, &dev->flags) &&
2347 !test_bit(R5_LOCKED, &dev->flags) &&
2348 !(test_bit(R5_UPTODATE, &dev->flags) ||
2349 test_bit(R5_Wantcompute, &dev->flags)) &&
2350 test_bit(R5_Insync, &dev->flags)) {
2352 test_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2353 pr_debug("Read_old block "
2354 "%d for Reconstruct\n", i);
2355 set_bit(R5_LOCKED, &dev->flags);
2356 set_bit(R5_Wantread, &dev->flags);
2359 set_bit(STRIPE_DELAYED, &sh->state);
2360 set_bit(STRIPE_HANDLE, &sh->state);
2364 /* now if nothing is locked, and if we have enough data,
2365 * we can start a write request
2367 /* since handle_stripe can be called at any time we need to handle the
2368 * case where a compute block operation has been submitted and then a
2369 * subsequent call wants to start a write request. raid5_run_ops only
2370 * handles the case where compute block and postxor are requested
2371 * simultaneously. If this is not the case then new writes need to be
2372 * held off until the compute completes.
2374 if ((s->req_compute || !test_bit(STRIPE_COMPUTE_RUN, &sh->state)) &&
2375 (s->locked == 0 && (rcw == 0 || rmw == 0) &&
2376 !test_bit(STRIPE_BIT_DELAY, &sh->state)))
2377 schedule_reconstruction5(sh, s, rcw == 0, 0);
2380 static void handle_stripe_dirtying6(raid5_conf_t *conf,
2381 struct stripe_head *sh, struct stripe_head_state *s,
2382 struct r6_state *r6s, int disks)
2384 int rcw = 0, must_compute = 0, pd_idx = sh->pd_idx, i;
2385 int qd_idx = sh->qd_idx;
2386 for (i = disks; i--; ) {
2387 struct r5dev *dev = &sh->dev[i];
2388 /* Would I have to read this buffer for reconstruct_write */
2389 if (!test_bit(R5_OVERWRITE, &dev->flags)
2390 && i != pd_idx && i != qd_idx
2391 && (!test_bit(R5_LOCKED, &dev->flags)
2393 !test_bit(R5_UPTODATE, &dev->flags)) {
2394 if (test_bit(R5_Insync, &dev->flags)) rcw++;
2396 pr_debug("raid6: must_compute: "
2397 "disk %d flags=%#lx\n", i, dev->flags);
2402 pr_debug("for sector %llu, rcw=%d, must_compute=%d\n",
2403 (unsigned long long)sh->sector, rcw, must_compute);
2404 set_bit(STRIPE_HANDLE, &sh->state);
2407 /* want reconstruct write, but need to get some data */
2408 for (i = disks; i--; ) {
2409 struct r5dev *dev = &sh->dev[i];
2410 if (!test_bit(R5_OVERWRITE, &dev->flags)
2411 && !(s->failed == 0 && (i == pd_idx || i == qd_idx))
2412 && !test_bit(R5_LOCKED, &dev->flags) &&
2413 !test_bit(R5_UPTODATE, &dev->flags) &&
2414 test_bit(R5_Insync, &dev->flags)) {
2416 test_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2417 pr_debug("Read_old stripe %llu "
2418 "block %d for Reconstruct\n",
2419 (unsigned long long)sh->sector, i);
2420 set_bit(R5_LOCKED, &dev->flags);
2421 set_bit(R5_Wantread, &dev->flags);
2424 pr_debug("Request delayed stripe %llu "
2425 "block %d for Reconstruct\n",
2426 (unsigned long long)sh->sector, i);
2427 set_bit(STRIPE_DELAYED, &sh->state);
2428 set_bit(STRIPE_HANDLE, &sh->state);
2432 /* now if nothing is locked, and if we have enough data, we can start a
2435 if (s->locked == 0 && rcw == 0 &&
2436 !test_bit(STRIPE_BIT_DELAY, &sh->state)) {
2437 if (must_compute > 0) {
2438 /* We have failed blocks and need to compute them */
2439 switch (s->failed) {
2443 compute_block_1(sh, r6s->failed_num[0], 0);
2446 compute_block_2(sh, r6s->failed_num[0],
2447 r6s->failed_num[1]);
2449 default: /* This request should have been failed? */
2454 pr_debug("Computing parity for stripe %llu\n",
2455 (unsigned long long)sh->sector);
2456 compute_parity6(sh, RECONSTRUCT_WRITE);
2457 /* now every locked buffer is ready to be written */
2458 for (i = disks; i--; )
2459 if (test_bit(R5_LOCKED, &sh->dev[i].flags)) {
2460 pr_debug("Writing stripe %llu block %d\n",
2461 (unsigned long long)sh->sector, i);
2463 set_bit(R5_Wantwrite, &sh->dev[i].flags);
2465 if (s->locked == disks)
2466 if (!test_and_set_bit(STRIPE_FULL_WRITE, &sh->state))
2467 atomic_inc(&conf->pending_full_writes);
2468 /* after a RECONSTRUCT_WRITE, the stripe MUST be in-sync */
2469 set_bit(STRIPE_INSYNC, &sh->state);
2471 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2472 atomic_dec(&conf->preread_active_stripes);
2473 if (atomic_read(&conf->preread_active_stripes) <
2475 md_wakeup_thread(conf->mddev->thread);
2480 static void handle_parity_checks5(raid5_conf_t *conf, struct stripe_head *sh,
2481 struct stripe_head_state *s, int disks)
2483 struct r5dev *dev = NULL;
2485 set_bit(STRIPE_HANDLE, &sh->state);
2487 switch (sh->check_state) {
2488 case check_state_idle:
2489 /* start a new check operation if there are no failures */
2490 if (s->failed == 0) {
2491 BUG_ON(s->uptodate != disks);
2492 sh->check_state = check_state_run;
2493 set_bit(STRIPE_OP_CHECK, &s->ops_request);
2494 clear_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags);
2498 dev = &sh->dev[s->failed_num];
2500 case check_state_compute_result:
2501 sh->check_state = check_state_idle;
2503 dev = &sh->dev[sh->pd_idx];
2505 /* check that a write has not made the stripe insync */
2506 if (test_bit(STRIPE_INSYNC, &sh->state))
2509 /* either failed parity check, or recovery is happening */
2510 BUG_ON(!test_bit(R5_UPTODATE, &dev->flags));
2511 BUG_ON(s->uptodate != disks);
2513 set_bit(R5_LOCKED, &dev->flags);
2515 set_bit(R5_Wantwrite, &dev->flags);
2517 clear_bit(STRIPE_DEGRADED, &sh->state);
2518 set_bit(STRIPE_INSYNC, &sh->state);
2520 case check_state_run:
2521 break; /* we will be called again upon completion */
2522 case check_state_check_result:
2523 sh->check_state = check_state_idle;
2525 /* if a failure occurred during the check operation, leave
2526 * STRIPE_INSYNC not set and let the stripe be handled again
2531 /* handle a successful check operation, if parity is correct
2532 * we are done. Otherwise update the mismatch count and repair
2533 * parity if !MD_RECOVERY_CHECK
2535 if (sh->ops.zero_sum_result == 0)
2536 /* parity is correct (on disc,
2537 * not in buffer any more)
2539 set_bit(STRIPE_INSYNC, &sh->state);
2541 conf->mddev->resync_mismatches += STRIPE_SECTORS;
2542 if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery))
2543 /* don't try to repair!! */
2544 set_bit(STRIPE_INSYNC, &sh->state);
2546 sh->check_state = check_state_compute_run;
2547 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
2548 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
2549 set_bit(R5_Wantcompute,
2550 &sh->dev[sh->pd_idx].flags);
2551 sh->ops.target = sh->pd_idx;
2556 case check_state_compute_run:
2559 printk(KERN_ERR "%s: unknown check_state: %d sector: %llu\n",
2560 __func__, sh->check_state,
2561 (unsigned long long) sh->sector);
2567 static void handle_parity_checks6(raid5_conf_t *conf, struct stripe_head *sh,
2568 struct stripe_head_state *s,
2569 struct r6_state *r6s, struct page *tmp_page,
2572 int update_p = 0, update_q = 0;
2574 int pd_idx = sh->pd_idx;
2575 int qd_idx = sh->qd_idx;
2577 set_bit(STRIPE_HANDLE, &sh->state);
2579 BUG_ON(s->failed > 2);
2580 BUG_ON(s->uptodate < disks);
2581 /* Want to check and possibly repair P and Q.
2582 * However there could be one 'failed' device, in which
2583 * case we can only check one of them, possibly using the
2584 * other to generate missing data
2587 /* If !tmp_page, we cannot do the calculations,
2588 * but as we have set STRIPE_HANDLE, we will soon be called
2589 * by stripe_handle with a tmp_page - just wait until then.
2592 if (s->failed == r6s->q_failed) {
2593 /* The only possible failed device holds 'Q', so it
2594 * makes sense to check P (If anything else were failed,
2595 * we would have used P to recreate it).
2597 compute_block_1(sh, pd_idx, 1);
2598 if (!page_is_zero(sh->dev[pd_idx].page)) {
2599 compute_block_1(sh, pd_idx, 0);
2603 if (!r6s->q_failed && s->failed < 2) {
2604 /* q is not failed, and we didn't use it to generate
2605 * anything, so it makes sense to check it
2607 memcpy(page_address(tmp_page),
2608 page_address(sh->dev[qd_idx].page),
2610 compute_parity6(sh, UPDATE_PARITY);
2611 if (memcmp(page_address(tmp_page),
2612 page_address(sh->dev[qd_idx].page),
2613 STRIPE_SIZE) != 0) {
2614 clear_bit(STRIPE_INSYNC, &sh->state);
2618 if (update_p || update_q) {
2619 conf->mddev->resync_mismatches += STRIPE_SECTORS;
2620 if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery))
2621 /* don't try to repair!! */
2622 update_p = update_q = 0;
2625 /* now write out any block on a failed drive,
2626 * or P or Q if they need it
2629 if (s->failed == 2) {
2630 dev = &sh->dev[r6s->failed_num[1]];
2632 set_bit(R5_LOCKED, &dev->flags);
2633 set_bit(R5_Wantwrite, &dev->flags);
2635 if (s->failed >= 1) {
2636 dev = &sh->dev[r6s->failed_num[0]];
2638 set_bit(R5_LOCKED, &dev->flags);
2639 set_bit(R5_Wantwrite, &dev->flags);
2643 dev = &sh->dev[pd_idx];
2645 set_bit(R5_LOCKED, &dev->flags);
2646 set_bit(R5_Wantwrite, &dev->flags);
2649 dev = &sh->dev[qd_idx];
2651 set_bit(R5_LOCKED, &dev->flags);
2652 set_bit(R5_Wantwrite, &dev->flags);
2654 clear_bit(STRIPE_DEGRADED, &sh->state);
2656 set_bit(STRIPE_INSYNC, &sh->state);
2660 static void handle_stripe_expansion(raid5_conf_t *conf, struct stripe_head *sh,
2661 struct r6_state *r6s)
2665 /* We have read all the blocks in this stripe and now we need to
2666 * copy some of them into a target stripe for expand.
2668 struct dma_async_tx_descriptor *tx = NULL;
2669 clear_bit(STRIPE_EXPAND_SOURCE, &sh->state);
2670 for (i = 0; i < sh->disks; i++)
2671 if (i != sh->pd_idx && i != sh->qd_idx) {
2673 struct stripe_head *sh2;
2674 struct async_submit_ctl submit;
2676 sector_t bn = compute_blocknr(sh, i, 1);
2677 sector_t s = raid5_compute_sector(conf, bn, 0,
2679 sh2 = get_active_stripe(conf, s, 0, 1);
2681 /* so far only the early blocks of this stripe
2682 * have been requested. When later blocks
2683 * get requested, we will try again
2686 if (!test_bit(STRIPE_EXPANDING, &sh2->state) ||
2687 test_bit(R5_Expanded, &sh2->dev[dd_idx].flags)) {
2688 /* must have already done this block */
2689 release_stripe(sh2);
2693 /* place all the copies on one channel */
2694 init_async_submit(&submit, 0, tx, NULL, NULL, NULL);
2695 tx = async_memcpy(sh2->dev[dd_idx].page,
2696 sh->dev[i].page, 0, 0, STRIPE_SIZE,
2699 set_bit(R5_Expanded, &sh2->dev[dd_idx].flags);
2700 set_bit(R5_UPTODATE, &sh2->dev[dd_idx].flags);
2701 for (j = 0; j < conf->raid_disks; j++)
2702 if (j != sh2->pd_idx &&
2703 (!r6s || j != sh2->qd_idx) &&
2704 !test_bit(R5_Expanded, &sh2->dev[j].flags))
2706 if (j == conf->raid_disks) {
2707 set_bit(STRIPE_EXPAND_READY, &sh2->state);
2708 set_bit(STRIPE_HANDLE, &sh2->state);
2710 release_stripe(sh2);
2713 /* done submitting copies, wait for them to complete */
2716 dma_wait_for_async_tx(tx);
2722 * handle_stripe - do things to a stripe.
2724 * We lock the stripe and then examine the state of various bits
2725 * to see what needs to be done.
2727 * return some read request which now have data
2728 * return some write requests which are safely on disc
2729 * schedule a read on some buffers
2730 * schedule a write of some buffers
2731 * return confirmation of parity correctness
2733 * buffers are taken off read_list or write_list, and bh_cache buffers
2734 * get BH_Lock set before the stripe lock is released.
2738 static bool handle_stripe5(struct stripe_head *sh)
2740 raid5_conf_t *conf = sh->raid_conf;
2741 int disks = sh->disks, i;
2742 struct bio *return_bi = NULL;
2743 struct stripe_head_state s;
2745 mdk_rdev_t *blocked_rdev = NULL;
2748 memset(&s, 0, sizeof(s));
2749 pr_debug("handling stripe %llu, state=%#lx cnt=%d, pd_idx=%d check:%d "
2750 "reconstruct:%d\n", (unsigned long long)sh->sector, sh->state,
2751 atomic_read(&sh->count), sh->pd_idx, sh->check_state,
2752 sh->reconstruct_state);
2754 spin_lock(&sh->lock);
2755 clear_bit(STRIPE_HANDLE, &sh->state);
2756 clear_bit(STRIPE_DELAYED, &sh->state);
2758 s.syncing = test_bit(STRIPE_SYNCING, &sh->state);
2759 s.expanding = test_bit(STRIPE_EXPAND_SOURCE, &sh->state);
2760 s.expanded = test_bit(STRIPE_EXPAND_READY, &sh->state);
2762 /* Now to look around and see what can be done */
2764 for (i=disks; i--; ) {
2766 struct r5dev *dev = &sh->dev[i];
2767 clear_bit(R5_Insync, &dev->flags);
2769 pr_debug("check %d: state 0x%lx toread %p read %p write %p "
2770 "written %p\n", i, dev->flags, dev->toread, dev->read,
2771 dev->towrite, dev->written);
2773 /* maybe we can request a biofill operation
2775 * new wantfill requests are only permitted while
2776 * ops_complete_biofill is guaranteed to be inactive
2778 if (test_bit(R5_UPTODATE, &dev->flags) && dev->toread &&
2779 !test_bit(STRIPE_BIOFILL_RUN, &sh->state))
2780 set_bit(R5_Wantfill, &dev->flags);
2782 /* now count some things */
2783 if (test_bit(R5_LOCKED, &dev->flags)) s.locked++;
2784 if (test_bit(R5_UPTODATE, &dev->flags)) s.uptodate++;
2785 if (test_bit(R5_Wantcompute, &dev->flags)) s.compute++;
2787 if (test_bit(R5_Wantfill, &dev->flags))
2789 else if (dev->toread)
2793 if (!test_bit(R5_OVERWRITE, &dev->flags))
2798 rdev = rcu_dereference(conf->disks[i].rdev);
2799 if (blocked_rdev == NULL &&
2800 rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
2801 blocked_rdev = rdev;
2802 atomic_inc(&rdev->nr_pending);
2804 if (!rdev || !test_bit(In_sync, &rdev->flags)) {
2805 /* The ReadError flag will just be confusing now */
2806 clear_bit(R5_ReadError, &dev->flags);
2807 clear_bit(R5_ReWrite, &dev->flags);
2809 if (!rdev || !test_bit(In_sync, &rdev->flags)
2810 || test_bit(R5_ReadError, &dev->flags)) {
2814 set_bit(R5_Insync, &dev->flags);
2818 if (unlikely(blocked_rdev)) {
2819 if (s.syncing || s.expanding || s.expanded ||
2820 s.to_write || s.written) {
2821 set_bit(STRIPE_HANDLE, &sh->state);
2824 /* There is nothing for the blocked_rdev to block */
2825 rdev_dec_pending(blocked_rdev, conf->mddev);
2826 blocked_rdev = NULL;
2829 if (s.to_fill && !test_bit(STRIPE_BIOFILL_RUN, &sh->state)) {
2830 set_bit(STRIPE_OP_BIOFILL, &s.ops_request);
2831 set_bit(STRIPE_BIOFILL_RUN, &sh->state);
2834 pr_debug("locked=%d uptodate=%d to_read=%d"
2835 " to_write=%d failed=%d failed_num=%d\n",
2836 s.locked, s.uptodate, s.to_read, s.to_write,
2837 s.failed, s.failed_num);
2838 /* check if the array has lost two devices and, if so, some requests might
2841 if (s.failed > 1 && s.to_read+s.to_write+s.written)
2842 handle_failed_stripe(conf, sh, &s, disks, &return_bi);
2843 if (s.failed > 1 && s.syncing) {
2844 md_done_sync(conf->mddev, STRIPE_SECTORS,0);
2845 clear_bit(STRIPE_SYNCING, &sh->state);
2849 /* might be able to return some write requests if the parity block
2850 * is safe, or on a failed drive
2852 dev = &sh->dev[sh->pd_idx];
2854 ((test_bit(R5_Insync, &dev->flags) &&
2855 !test_bit(R5_LOCKED, &dev->flags) &&
2856 test_bit(R5_UPTODATE, &dev->flags)) ||
2857 (s.failed == 1 && s.failed_num == sh->pd_idx)))
2858 handle_stripe_clean_event(conf, sh, disks, &return_bi);
2860 /* Now we might consider reading some blocks, either to check/generate
2861 * parity, or to satisfy requests
2862 * or to load a block that is being partially written.
2864 if (s.to_read || s.non_overwrite ||
2865 (s.syncing && (s.uptodate + s.compute < disks)) || s.expanding)
2866 handle_stripe_fill5(sh, &s, disks);
2868 /* Now we check to see if any write operations have recently
2872 if (sh->reconstruct_state == reconstruct_state_prexor_drain_result)
2874 if (sh->reconstruct_state == reconstruct_state_drain_result ||
2875 sh->reconstruct_state == reconstruct_state_prexor_drain_result) {
2876 sh->reconstruct_state = reconstruct_state_idle;
2878 /* All the 'written' buffers and the parity block are ready to
2879 * be written back to disk
2881 BUG_ON(!test_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags));
2882 for (i = disks; i--; ) {
2884 if (test_bit(R5_LOCKED, &dev->flags) &&
2885 (i == sh->pd_idx || dev->written)) {
2886 pr_debug("Writing block %d\n", i);
2887 set_bit(R5_Wantwrite, &dev->flags);
2890 if (!test_bit(R5_Insync, &dev->flags) ||
2891 (i == sh->pd_idx && s.failed == 0))
2892 set_bit(STRIPE_INSYNC, &sh->state);
2895 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2896 atomic_dec(&conf->preread_active_stripes);
2897 if (atomic_read(&conf->preread_active_stripes) <
2899 md_wakeup_thread(conf->mddev->thread);
2903 /* Now to consider new write requests and what else, if anything
2904 * should be read. We do not handle new writes when:
2905 * 1/ A 'write' operation (copy+xor) is already in flight.
2906 * 2/ A 'check' operation is in flight, as it may clobber the parity
2909 if (s.to_write && !sh->reconstruct_state && !sh->check_state)
2910 handle_stripe_dirtying5(conf, sh, &s, disks);
2912 /* maybe we need to check and possibly fix the parity for this stripe
2913 * Any reads will already have been scheduled, so we just see if enough
2914 * data is available. The parity check is held off while parity
2915 * dependent operations are in flight.
2917 if (sh->check_state ||
2918 (s.syncing && s.locked == 0 &&
2919 !test_bit(STRIPE_COMPUTE_RUN, &sh->state) &&
2920 !test_bit(STRIPE_INSYNC, &sh->state)))
2921 handle_parity_checks5(conf, sh, &s, disks);
2923 if (s.syncing && s.locked == 0 && test_bit(STRIPE_INSYNC, &sh->state)) {
2924 md_done_sync(conf->mddev, STRIPE_SECTORS,1);
2925 clear_bit(STRIPE_SYNCING, &sh->state);
2928 /* If the failed drive is just a ReadError, then we might need to progress
2929 * the repair/check process
2931 if (s.failed == 1 && !conf->mddev->ro &&
2932 test_bit(R5_ReadError, &sh->dev[s.failed_num].flags)
2933 && !test_bit(R5_LOCKED, &sh->dev[s.failed_num].flags)
2934 && test_bit(R5_UPTODATE, &sh->dev[s.failed_num].flags)
2936 dev = &sh->dev[s.failed_num];
2937 if (!test_bit(R5_ReWrite, &dev->flags)) {
2938 set_bit(R5_Wantwrite, &dev->flags);
2939 set_bit(R5_ReWrite, &dev->flags);
2940 set_bit(R5_LOCKED, &dev->flags);
2943 /* let's read it back */
2944 set_bit(R5_Wantread, &dev->flags);
2945 set_bit(R5_LOCKED, &dev->flags);
2950 /* Finish reconstruct operations initiated by the expansion process */
2951 if (sh->reconstruct_state == reconstruct_state_result) {
2952 struct stripe_head *sh2
2953 = get_active_stripe(conf, sh->sector, 1, 1);
2954 if (sh2 && test_bit(STRIPE_EXPAND_SOURCE, &sh2->state)) {
2955 /* sh cannot be written until sh2 has been read.
2956 * so arrange for sh to be delayed a little
2958 set_bit(STRIPE_DELAYED, &sh->state);
2959 set_bit(STRIPE_HANDLE, &sh->state);
2960 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE,
2962 atomic_inc(&conf->preread_active_stripes);
2963 release_stripe(sh2);
2967 release_stripe(sh2);
2969 sh->reconstruct_state = reconstruct_state_idle;
2970 clear_bit(STRIPE_EXPANDING, &sh->state);
2971 for (i = conf->raid_disks; i--; ) {
2972 set_bit(R5_Wantwrite, &sh->dev[i].flags);
2973 set_bit(R5_LOCKED, &sh->dev[i].flags);
2978 if (s.expanded && test_bit(STRIPE_EXPANDING, &sh->state) &&
2979 !sh->reconstruct_state) {
2980 /* Need to write out all blocks after computing parity */
2981 sh->disks = conf->raid_disks;
2982 stripe_set_idx(sh->sector, conf, 0, sh);
2983 schedule_reconstruction5(sh, &s, 1, 1);
2984 } else if (s.expanded && !sh->reconstruct_state && s.locked == 0) {
2985 clear_bit(STRIPE_EXPAND_READY, &sh->state);
2986 atomic_dec(&conf->reshape_stripes);
2987 wake_up(&conf->wait_for_overlap);
2988 md_done_sync(conf->mddev, STRIPE_SECTORS, 1);
2991 if (s.expanding && s.locked == 0 &&
2992 !test_bit(STRIPE_COMPUTE_RUN, &sh->state))
2993 handle_stripe_expansion(conf, sh, NULL);
2996 spin_unlock(&sh->lock);
2998 /* wait for this device to become unblocked */
2999 if (unlikely(blocked_rdev))
3000 md_wait_for_blocked_rdev(blocked_rdev, conf->mddev);
3003 raid5_run_ops(sh, s.ops_request);
3007 return_io(return_bi);
3009 return blocked_rdev == NULL;
3012 static bool handle_stripe6(struct stripe_head *sh, struct page *tmp_page)
3014 raid5_conf_t *conf = sh->raid_conf;
3015 int disks = sh->disks;
3016 struct bio *return_bi = NULL;
3017 int i, pd_idx = sh->pd_idx, qd_idx = sh->qd_idx;
3018 struct stripe_head_state s;
3019 struct r6_state r6s;
3020 struct r5dev *dev, *pdev, *qdev;
3021 mdk_rdev_t *blocked_rdev = NULL;
3023 pr_debug("handling stripe %llu, state=%#lx cnt=%d, "
3024 "pd_idx=%d, qd_idx=%d\n",
3025 (unsigned long long)sh->sector, sh->state,
3026 atomic_read(&sh->count), pd_idx, qd_idx);
3027 memset(&s, 0, sizeof(s));
3029 spin_lock(&sh->lock);
3030 clear_bit(STRIPE_HANDLE, &sh->state);
3031 clear_bit(STRIPE_DELAYED, &sh->state);
3033 s.syncing = test_bit(STRIPE_SYNCING, &sh->state);
3034 s.expanding = test_bit(STRIPE_EXPAND_SOURCE, &sh->state);
3035 s.expanded = test_bit(STRIPE_EXPAND_READY, &sh->state);
3036 /* Now to look around and see what can be done */
3039 for (i=disks; i--; ) {
3042 clear_bit(R5_Insync, &dev->flags);
3044 pr_debug("check %d: state 0x%lx read %p write %p written %p\n",
3045 i, dev->flags, dev->toread, dev->towrite, dev->written);
3046 /* maybe we can reply to a read */
3047 if (test_bit(R5_UPTODATE, &dev->flags) && dev->toread) {
3048 struct bio *rbi, *rbi2;
3049 pr_debug("Return read for disc %d\n", i);
3050 spin_lock_irq(&conf->device_lock);
3053 if (test_and_clear_bit(R5_Overlap, &dev->flags))
3054 wake_up(&conf->wait_for_overlap);
3055 spin_unlock_irq(&conf->device_lock);
3056 while (rbi && rbi->bi_sector < dev->sector + STRIPE_SECTORS) {
3057 copy_data(0, rbi, dev->page, dev->sector);
3058 rbi2 = r5_next_bio(rbi, dev->sector);
3059 spin_lock_irq(&conf->device_lock);
3060 if (!raid5_dec_bi_phys_segments(rbi)) {
3061 rbi->bi_next = return_bi;
3064 spin_unlock_irq(&conf->device_lock);
3069 /* now count some things */
3070 if (test_bit(R5_LOCKED, &dev->flags)) s.locked++;
3071 if (test_bit(R5_UPTODATE, &dev->flags)) s.uptodate++;
3078 if (!test_bit(R5_OVERWRITE, &dev->flags))
3083 rdev = rcu_dereference(conf->disks[i].rdev);
3084 if (blocked_rdev == NULL &&
3085 rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
3086 blocked_rdev = rdev;
3087 atomic_inc(&rdev->nr_pending);
3089 if (!rdev || !test_bit(In_sync, &rdev->flags)) {
3090 /* The ReadError flag will just be confusing now */
3091 clear_bit(R5_ReadError, &dev->flags);
3092 clear_bit(R5_ReWrite, &dev->flags);
3094 if (!rdev || !test_bit(In_sync, &rdev->flags)
3095 || test_bit(R5_ReadError, &dev->flags)) {
3097 r6s.failed_num[s.failed] = i;
3100 set_bit(R5_Insync, &dev->flags);
3104 if (unlikely(blocked_rdev)) {
3105 if (s.syncing || s.expanding || s.expanded ||
3106 s.to_write || s.written) {
3107 set_bit(STRIPE_HANDLE, &sh->state);
3110 /* There is nothing for the blocked_rdev to block */
3111 rdev_dec_pending(blocked_rdev, conf->mddev);
3112 blocked_rdev = NULL;
3115 pr_debug("locked=%d uptodate=%d to_read=%d"
3116 " to_write=%d failed=%d failed_num=%d,%d\n",
3117 s.locked, s.uptodate, s.to_read, s.to_write, s.failed,
3118 r6s.failed_num[0], r6s.failed_num[1]);
3119 /* check if the array has lost >2 devices and, if so, some requests
3120 * might need to be failed
3122 if (s.failed > 2 && s.to_read+s.to_write+s.written)
3123 handle_failed_stripe(conf, sh, &s, disks, &return_bi);
3124 if (s.failed > 2 && s.syncing) {
3125 md_done_sync(conf->mddev, STRIPE_SECTORS,0);
3126 clear_bit(STRIPE_SYNCING, &sh->state);
3131 * might be able to return some write requests if the parity blocks
3132 * are safe, or on a failed drive
3134 pdev = &sh->dev[pd_idx];
3135 r6s.p_failed = (s.failed >= 1 && r6s.failed_num[0] == pd_idx)
3136 || (s.failed >= 2 && r6s.failed_num[1] == pd_idx);
3137 qdev = &sh->dev[qd_idx];
3138 r6s.q_failed = (s.failed >= 1 && r6s.failed_num[0] == qd_idx)
3139 || (s.failed >= 2 && r6s.failed_num[1] == qd_idx);
3142 ( r6s.p_failed || ((test_bit(R5_Insync, &pdev->flags)
3143 && !test_bit(R5_LOCKED, &pdev->flags)
3144 && test_bit(R5_UPTODATE, &pdev->flags)))) &&
3145 ( r6s.q_failed || ((test_bit(R5_Insync, &qdev->flags)
3146 && !test_bit(R5_LOCKED, &qdev->flags)
3147 && test_bit(R5_UPTODATE, &qdev->flags)))))
3148 handle_stripe_clean_event(conf, sh, disks, &return_bi);
3150 /* Now we might consider reading some blocks, either to check/generate
3151 * parity, or to satisfy requests
3152 * or to load a block that is being partially written.
3154 if (s.to_read || s.non_overwrite || (s.to_write && s.failed) ||
3155 (s.syncing && (s.uptodate < disks)) || s.expanding)
3156 handle_stripe_fill6(sh, &s, &r6s, disks);
3158 /* now to consider writing and what else, if anything should be read */
3160 handle_stripe_dirtying6(conf, sh, &s, &r6s, disks);
3162 /* maybe we need to check and possibly fix the parity for this stripe
3163 * Any reads will already have been scheduled, so we just see if enough
3166 if (s.syncing && s.locked == 0 && !test_bit(STRIPE_INSYNC, &sh->state))
3167 handle_parity_checks6(conf, sh, &s, &r6s, tmp_page, disks);
3169 if (s.syncing && s.locked == 0 && test_bit(STRIPE_INSYNC, &sh->state)) {
3170 md_done_sync(conf->mddev, STRIPE_SECTORS,1);
3171 clear_bit(STRIPE_SYNCING, &sh->state);
3174 /* If the failed drives are just a ReadError, then we might need
3175 * to progress the repair/check process
3177 if (s.failed <= 2 && !conf->mddev->ro)
3178 for (i = 0; i < s.failed; i++) {
3179 dev = &sh->dev[r6s.failed_num[i]];
3180 if (test_bit(R5_ReadError, &dev->flags)
3181 && !test_bit(R5_LOCKED, &dev->flags)
3182 && test_bit(R5_UPTODATE, &dev->flags)
3184 if (!test_bit(R5_ReWrite, &dev->flags)) {
3185 set_bit(R5_Wantwrite, &dev->flags);
3186 set_bit(R5_ReWrite, &dev->flags);
3187 set_bit(R5_LOCKED, &dev->flags);
3189 /* let's read it back */
3190 set_bit(R5_Wantread, &dev->flags);
3191 set_bit(R5_LOCKED, &dev->flags);
3196 if (s.expanded && test_bit(STRIPE_EXPANDING, &sh->state)) {
3197 struct stripe_head *sh2
3198 = get_active_stripe(conf, sh->sector, 1, 1);
3199 if (sh2 && test_bit(STRIPE_EXPAND_SOURCE, &sh2->state)) {
3200 /* sh cannot be written until sh2 has been read.
3201 * so arrange for sh to be delayed a little
3203 set_bit(STRIPE_DELAYED, &sh->state);
3204 set_bit(STRIPE_HANDLE, &sh->state);
3205 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE,
3207 atomic_inc(&conf->preread_active_stripes);
3208 release_stripe(sh2);
3212 release_stripe(sh2);
3214 /* Need to write out all blocks after computing P&Q */
3215 sh->disks = conf->raid_disks;
3216 stripe_set_idx(sh->sector, conf, 0, sh);
3217 compute_parity6(sh, RECONSTRUCT_WRITE);
3218 for (i = conf->raid_disks ; i-- ; ) {
3219 set_bit(R5_LOCKED, &sh->dev[i].flags);
3221 set_bit(R5_Wantwrite, &sh->dev[i].flags);
3223 clear_bit(STRIPE_EXPANDING, &sh->state);
3224 } else if (s.expanded) {
3225 clear_bit(STRIPE_EXPAND_READY, &sh->state);
3226 atomic_dec(&conf->reshape_stripes);
3227 wake_up(&conf->wait_for_overlap);
3228 md_done_sync(conf->mddev, STRIPE_SECTORS, 1);
3231 if (s.expanding && s.locked == 0 &&
3232 !test_bit(STRIPE_COMPUTE_RUN, &sh->state))
3233 handle_stripe_expansion(conf, sh, &r6s);
3236 spin_unlock(&sh->lock);
3238 /* wait for this device to become unblocked */
3239 if (unlikely(blocked_rdev))
3240 md_wait_for_blocked_rdev(blocked_rdev, conf->mddev);
3244 return_io(return_bi);
3246 return blocked_rdev == NULL;
3249 /* returns true if the stripe was handled */
3250 static bool handle_stripe(struct stripe_head *sh, struct page *tmp_page)
3252 if (sh->raid_conf->level == 6)
3253 return handle_stripe6(sh, tmp_page);
3255 return handle_stripe5(sh);
3260 static void raid5_activate_delayed(raid5_conf_t *conf)
3262 if (atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD) {
3263 while (!list_empty(&conf->delayed_list)) {
3264 struct list_head *l = conf->delayed_list.next;
3265 struct stripe_head *sh;
3266 sh = list_entry(l, struct stripe_head, lru);
3268 clear_bit(STRIPE_DELAYED, &sh->state);
3269 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
3270 atomic_inc(&conf->preread_active_stripes);
3271 list_add_tail(&sh->lru, &conf->hold_list);
3274 blk_plug_device(conf->mddev->queue);
3277 static void activate_bit_delay(raid5_conf_t *conf)
3279 /* device_lock is held */
3280 struct list_head head;
3281 list_add(&head, &conf->bitmap_list);
3282 list_del_init(&conf->bitmap_list);
3283 while (!list_empty(&head)) {
3284 struct stripe_head *sh = list_entry(head.next, struct stripe_head, lru);
3285 list_del_init(&sh->lru);
3286 atomic_inc(&sh->count);
3287 __release_stripe(conf, sh);
3291 static void unplug_slaves(mddev_t *mddev)
3293 raid5_conf_t *conf = mddev_to_conf(mddev);
3297 for (i=0; i<mddev->raid_disks; i++) {
3298 mdk_rdev_t *rdev = rcu_dereference(conf->disks[i].rdev);
3299 if (rdev && !test_bit(Faulty, &rdev->flags) && atomic_read(&rdev->nr_pending)) {
3300 struct request_queue *r_queue = bdev_get_queue(rdev->bdev);
3302 atomic_inc(&rdev->nr_pending);
3305 blk_unplug(r_queue);
3307 rdev_dec_pending(rdev, mddev);
3314 static void raid5_unplug_device(struct request_queue *q)
3316 mddev_t *mddev = q->queuedata;
3317 raid5_conf_t *conf = mddev_to_conf(mddev);
3318 unsigned long flags;
3320 spin_lock_irqsave(&conf->device_lock, flags);
3322 if (blk_remove_plug(q)) {
3324 raid5_activate_delayed(conf);
3326 md_wakeup_thread(mddev->thread);
3328 spin_unlock_irqrestore(&conf->device_lock, flags);
3330 unplug_slaves(mddev);
3333 static int raid5_congested(void *data, int bits)
3335 mddev_t *mddev = data;
3336 raid5_conf_t *conf = mddev_to_conf(mddev);
3338 /* No difference between reads and writes. Just check
3339 * how busy the stripe_cache is
3341 if (conf->inactive_blocked)
3345 if (list_empty_careful(&conf->inactive_list))
3351 /* We want read requests to align with chunks where possible,
3352 * but write requests don't need to.
3354 static int raid5_mergeable_bvec(struct request_queue *q,
3355 struct bvec_merge_data *bvm,
3356 struct bio_vec *biovec)
3358 mddev_t *mddev = q->queuedata;
3359 sector_t sector = bvm->bi_sector + get_start_sect(bvm->bi_bdev);
3361 unsigned int chunk_sectors = mddev->chunk_size >> 9;
3362 unsigned int bio_sectors = bvm->bi_size >> 9;
3364 if ((bvm->bi_rw & 1) == WRITE)
3365 return biovec->bv_len; /* always allow writes to be mergeable */
3367 if (mddev->new_chunk < mddev->chunk_size)
3368 chunk_sectors = mddev->new_chunk >> 9;
3369 max = (chunk_sectors - ((sector & (chunk_sectors - 1)) + bio_sectors)) << 9;
3370 if (max < 0) max = 0;
3371 if (max <= biovec->bv_len && bio_sectors == 0)
3372 return biovec->bv_len;
3378 static int in_chunk_boundary(mddev_t *mddev, struct bio *bio)
3380 sector_t sector = bio->bi_sector + get_start_sect(bio->bi_bdev);
3381 unsigned int chunk_sectors = mddev->chunk_size >> 9;
3382 unsigned int bio_sectors = bio->bi_size >> 9;
3384 if (mddev->new_chunk < mddev->chunk_size)
3385 chunk_sectors = mddev->new_chunk >> 9;
3386 return chunk_sectors >=
3387 ((sector & (chunk_sectors - 1)) + bio_sectors);
3391 * add bio to the retry LIFO ( in O(1) ... we are in interrupt )
3392 * later sampled by raid5d.
3394 static void add_bio_to_retry(struct bio *bi,raid5_conf_t *conf)
3396 unsigned long flags;
3398 spin_lock_irqsave(&conf->device_lock, flags);
3400 bi->bi_next = conf->retry_read_aligned_list;
3401 conf->retry_read_aligned_list = bi;
3403 spin_unlock_irqrestore(&conf->device_lock, flags);
3404 md_wakeup_thread(conf->mddev->thread);
3408 static struct bio *remove_bio_from_retry(raid5_conf_t *conf)
3412 bi = conf->retry_read_aligned;
3414 conf->retry_read_aligned = NULL;
3417 bi = conf->retry_read_aligned_list;
3419 conf->retry_read_aligned_list = bi->bi_next;
3422 * this sets the active strip count to 1 and the processed
3423 * strip count to zero (upper 8 bits)
3425 bi->bi_phys_segments = 1; /* biased count of active stripes */
3433 * The "raid5_align_endio" should check if the read succeeded and if it
3434 * did, call bio_endio on the original bio (having bio_put the new bio
3436 * If the read failed..
3438 static void raid5_align_endio(struct bio *bi, int error)
3440 struct bio* raid_bi = bi->bi_private;
3443 int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
3448 mddev = raid_bi->bi_bdev->bd_disk->queue->queuedata;
3449 conf = mddev_to_conf(mddev);
3450 rdev = (void*)raid_bi->bi_next;
3451 raid_bi->bi_next = NULL;
3453 rdev_dec_pending(rdev, conf->mddev);
3455 if (!error && uptodate) {
3456 bio_endio(raid_bi, 0);
3457 if (atomic_dec_and_test(&conf->active_aligned_reads))
3458 wake_up(&conf->wait_for_stripe);
3463 pr_debug("raid5_align_endio : io error...handing IO for a retry\n");
3465 add_bio_to_retry(raid_bi, conf);
3468 static int bio_fits_rdev(struct bio *bi)
3470 struct request_queue *q = bdev_get_queue(bi->bi_bdev);
3472 if ((bi->bi_size>>9) > q->max_sectors)
3474 blk_recount_segments(q, bi);
3475 if (bi->bi_phys_segments > q->max_phys_segments)
3478 if (q->merge_bvec_fn)
3479 /* it's too hard to apply the merge_bvec_fn at this stage,
3488 static int chunk_aligned_read(struct request_queue *q, struct bio * raid_bio)
3490 mddev_t *mddev = q->queuedata;
3491 raid5_conf_t *conf = mddev_to_conf(mddev);
3492 unsigned int dd_idx;
3493 struct bio* align_bi;
3496 if (!in_chunk_boundary(mddev, raid_bio)) {
3497 pr_debug("chunk_aligned_read : non aligned\n");
3501 * use bio_clone to make a copy of the bio
3503 align_bi = bio_clone(raid_bio, GFP_NOIO);
3507 * set bi_end_io to a new function, and set bi_private to the
3510 align_bi->bi_end_io = raid5_align_endio;
3511 align_bi->bi_private = raid_bio;
3515 align_bi->bi_sector = raid5_compute_sector(conf, raid_bio->bi_sector,
3520 rdev = rcu_dereference(conf->disks[dd_idx].rdev);
3521 if (rdev && test_bit(In_sync, &rdev->flags)) {
3522 atomic_inc(&rdev->nr_pending);
3524 raid_bio->bi_next = (void*)rdev;
3525 align_bi->bi_bdev = rdev->bdev;
3526 align_bi->bi_flags &= ~(1 << BIO_SEG_VALID);
3527 align_bi->bi_sector += rdev->data_offset;
3529 if (!bio_fits_rdev(align_bi)) {
3530 /* too big in some way */
3532 rdev_dec_pending(rdev, mddev);
3536 spin_lock_irq(&conf->device_lock);
3537 wait_event_lock_irq(conf->wait_for_stripe,
3539 conf->device_lock, /* nothing */);
3540 atomic_inc(&conf->active_aligned_reads);
3541 spin_unlock_irq(&conf->device_lock);
3543 generic_make_request(align_bi);
3552 /* __get_priority_stripe - get the next stripe to process
3554 * Full stripe writes are allowed to pass preread active stripes up until
3555 * the bypass_threshold is exceeded. In general the bypass_count
3556 * increments when the handle_list is handled before the hold_list; however, it
3557 * will not be incremented when STRIPE_IO_STARTED is sampled set signifying a
3558 * stripe with in flight i/o. The bypass_count will be reset when the
3559 * head of the hold_list has changed, i.e. the head was promoted to the
3562 static struct stripe_head *__get_priority_stripe(raid5_conf_t *conf)
3564 struct stripe_head *sh;
3566 pr_debug("%s: handle: %s hold: %s full_writes: %d bypass_count: %d\n",
3568 list_empty(&conf->handle_list) ? "empty" : "busy",
3569 list_empty(&conf->hold_list) ? "empty" : "busy",
3570 atomic_read(&conf->pending_full_writes), conf->bypass_count);
3572 if (!list_empty(&conf->handle_list)) {
3573 sh = list_entry(conf->handle_list.next, typeof(*sh), lru);
3575 if (list_empty(&conf->hold_list))
3576 conf->bypass_count = 0;
3577 else if (!test_bit(STRIPE_IO_STARTED, &sh->state)) {
3578 if (conf->hold_list.next == conf->last_hold)
3579 conf->bypass_count++;
3581 conf->last_hold = conf->hold_list.next;
3582 conf->bypass_count -= conf->bypass_threshold;
3583 if (conf->bypass_count < 0)
3584 conf->bypass_count = 0;
3587 } else if (!list_empty(&conf->hold_list) &&
3588 ((conf->bypass_threshold &&
3589 conf->bypass_count > conf->bypass_threshold) ||
3590 atomic_read(&conf->pending_full_writes) == 0)) {
3591 sh = list_entry(conf->hold_list.next,
3593 conf->bypass_count -= conf->bypass_threshold;
3594 if (conf->bypass_count < 0)
3595 conf->bypass_count = 0;
3599 list_del_init(&sh->lru);
3600 atomic_inc(&sh->count);
3601 BUG_ON(atomic_read(&sh->count) != 1);
3605 static int make_request(struct request_queue *q, struct bio * bi)
3607 mddev_t *mddev = q->queuedata;
3608 raid5_conf_t *conf = mddev_to_conf(mddev);
3610 sector_t new_sector;
3611 sector_t logical_sector, last_sector;
3612 struct stripe_head *sh;
3613 const int rw = bio_data_dir(bi);
3616 if (unlikely(bio_barrier(bi))) {
3617 bio_endio(bi, -EOPNOTSUPP);
3621 md_write_start(mddev, bi);
3623 cpu = part_stat_lock();
3624 part_stat_inc(cpu, &mddev->gendisk->part0, ios[rw]);
3625 part_stat_add(cpu, &mddev->gendisk->part0, sectors[rw],
3630 mddev->reshape_position == MaxSector &&
3631 chunk_aligned_read(q,bi))
3634 logical_sector = bi->bi_sector & ~((sector_t)STRIPE_SECTORS-1);
3635 last_sector = bi->bi_sector + (bi->bi_size>>9);
3637 bi->bi_phys_segments = 1; /* over-loaded to count active stripes */
3639 for (;logical_sector < last_sector; logical_sector += STRIPE_SECTORS) {
3641 int disks, data_disks;
3646 disks = conf->raid_disks;
3647 prepare_to_wait(&conf->wait_for_overlap, &w, TASK_UNINTERRUPTIBLE);
3648 if (unlikely(conf->reshape_progress != MaxSector)) {
3649 /* spinlock is needed as reshape_progress may be
3650 * 64bit on a 32bit platform, and so it might be
3651 * possible to see a half-updated value
3652 * Ofcourse reshape_progress could change after
3653 * the lock is dropped, so once we get a reference
3654 * to the stripe that we think it is, we will have
3657 spin_lock_irq(&conf->device_lock);
3658 if (mddev->delta_disks < 0
3659 ? logical_sector < conf->reshape_progress
3660 : logical_sector >= conf->reshape_progress) {
3661 disks = conf->previous_raid_disks;
3664 if (mddev->delta_disks < 0
3665 ? logical_sector < conf->reshape_safe
3666 : logical_sector >= conf->reshape_safe) {
3667 spin_unlock_irq(&conf->device_lock);
3672 spin_unlock_irq(&conf->device_lock);
3674 data_disks = disks - conf->max_degraded;
3676 new_sector = raid5_compute_sector(conf, logical_sector,
3679 pr_debug("raid5: make_request, sector %llu logical %llu\n",
3680 (unsigned long long)new_sector,
3681 (unsigned long long)logical_sector);
3683 sh = get_active_stripe(conf, new_sector, previous,
3684 (bi->bi_rw&RWA_MASK));
3686 if (unlikely(previous)) {
3687 /* expansion might have moved on while waiting for a
3688 * stripe, so we must do the range check again.
3689 * Expansion could still move past after this
3690 * test, but as we are holding a reference to
3691 * 'sh', we know that if that happens,
3692 * STRIPE_EXPANDING will get set and the expansion
3693 * won't proceed until we finish with the stripe.
3696 spin_lock_irq(&conf->device_lock);
3697 if (mddev->delta_disks < 0
3698 ? logical_sector >= conf->reshape_progress
3699 : logical_sector < conf->reshape_progress)
3700 /* mismatch, need to try again */
3702 spin_unlock_irq(&conf->device_lock);
3708 /* FIXME what if we get a false positive because these
3709 * are being updated.
3711 if (logical_sector >= mddev->suspend_lo &&
3712 logical_sector < mddev->suspend_hi) {
3718 if (test_bit(STRIPE_EXPANDING, &sh->state) ||
3719 !add_stripe_bio(sh, bi, dd_idx, (bi->bi_rw&RW_MASK))) {
3720 /* Stripe is busy expanding or
3721 * add failed due to overlap. Flush everything
3724 raid5_unplug_device(mddev->queue);
3729 finish_wait(&conf->wait_for_overlap, &w);
3730 set_bit(STRIPE_HANDLE, &sh->state);
3731 clear_bit(STRIPE_DELAYED, &sh->state);
3734 /* cannot get stripe for read-ahead, just give-up */
3735 clear_bit(BIO_UPTODATE, &bi->bi_flags);
3736 finish_wait(&conf->wait_for_overlap, &w);
3741 spin_lock_irq(&conf->device_lock);
3742 remaining = raid5_dec_bi_phys_segments(bi);
3743 spin_unlock_irq(&conf->device_lock);
3744 if (remaining == 0) {
3747 md_write_end(mddev);
3754 static sector_t raid5_size(mddev_t *mddev, sector_t sectors, int raid_disks);
3756 static sector_t reshape_request(mddev_t *mddev, sector_t sector_nr, int *skipped)
3758 /* reshaping is quite different to recovery/resync so it is
3759 * handled quite separately ... here.
3761 * On each call to sync_request, we gather one chunk worth of
3762 * destination stripes and flag them as expanding.
3763 * Then we find all the source stripes and request reads.
3764 * As the reads complete, handle_stripe will copy the data
3765 * into the destination stripe and release that stripe.
3767 raid5_conf_t *conf = (raid5_conf_t *) mddev->private;
3768 struct stripe_head *sh;
3769 sector_t first_sector, last_sector;
3770 int raid_disks = conf->previous_raid_disks;
3771 int data_disks = raid_disks - conf->max_degraded;
3772 int new_data_disks = conf->raid_disks - conf->max_degraded;
3775 sector_t writepos, readpos, safepos;
3776 sector_t stripe_addr;
3777 int reshape_sectors;
3778 struct list_head stripes;
3780 if (sector_nr == 0) {
3781 /* If restarting in the middle, skip the initial sectors */
3782 if (mddev->delta_disks < 0 &&
3783 conf->reshape_progress < raid5_size(mddev, 0, 0)) {
3784 sector_nr = raid5_size(mddev, 0, 0)
3785 - conf->reshape_progress;
3786 } else if (mddev->delta_disks > 0 &&
3787 conf->reshape_progress > 0)
3788 sector_nr = conf->reshape_progress;
3789 sector_div(sector_nr, new_data_disks);
3796 /* We need to process a full chunk at a time.
3797 * If old and new chunk sizes differ, we need to process the
3800 if (mddev->new_chunk > mddev->chunk_size)
3801 reshape_sectors = mddev->new_chunk / 512;
3803 reshape_sectors = mddev->chunk_size / 512;
3805 /* we update the metadata when there is more than 3Meg
3806 * in the block range (that is rather arbitrary, should
3807 * probably be time based) or when the data about to be
3808 * copied would over-write the source of the data at
3809 * the front of the range.
3810 * i.e. one new_stripe along from reshape_progress new_maps
3811 * to after where reshape_safe old_maps to
3813 writepos = conf->reshape_progress;
3814 sector_div(writepos, new_data_disks);
3815 readpos = conf->reshape_progress;
3816 sector_div(readpos, data_disks);
3817 safepos = conf->reshape_safe;
3818 sector_div(safepos, data_disks);
3819 if (mddev->delta_disks < 0) {
3820 writepos -= reshape_sectors;
3821 readpos += reshape_sectors;
3822 safepos += reshape_sectors;
3824 writepos += reshape_sectors;
3825 readpos -= reshape_sectors;
3826 safepos -= reshape_sectors;
3829 /* 'writepos' is the most advanced device address we might write.
3830 * 'readpos' is the least advanced device address we might read.
3831 * 'safepos' is the least address recorded in the metadata as having
3833 * If 'readpos' is behind 'writepos', then there is no way that we can
3834 * ensure safety in the face of a crash - that must be done by userspace
3835 * making a backup of the data. So in that case there is no particular
3836 * rush to update metadata.
3837 * Otherwise if 'safepos' is behind 'writepos', then we really need to
3838 * update the metadata to advance 'safepos' to match 'readpos' so that
3839 * we can be safe in the event of a crash.
3840 * So we insist on updating metadata if safepos is behind writepos and
3841 * readpos is beyond writepos.
3842 * In any case, update the metadata every 10 seconds.
3843 * Maybe that number should be configurable, but I'm not sure it is
3844 * worth it.... maybe it could be a multiple of safemode_delay???
3846 if ((mddev->delta_disks < 0
3847 ? (safepos > writepos && readpos < writepos)
3848 : (safepos < writepos && readpos > writepos)) ||
3849 time_after(jiffies, conf->reshape_checkpoint + 10*HZ)) {
3850 /* Cannot proceed until we've updated the superblock... */
3851 wait_event(conf->wait_for_overlap,
3852 atomic_read(&conf->reshape_stripes)==0);
3853 mddev->reshape_position = conf->reshape_progress;
3854 conf->reshape_checkpoint = jiffies;
3855 set_bit(MD_CHANGE_DEVS, &mddev->flags);
3856 md_wakeup_thread(mddev->thread);
3857 wait_event(mddev->sb_wait, mddev->flags == 0 ||
3858 kthread_should_stop());
3859 spin_lock_irq(&conf->device_lock);
3860 conf->reshape_safe = mddev->reshape_position;
3861 spin_unlock_irq(&conf->device_lock);
3862 wake_up(&conf->wait_for_overlap);
3865 if (mddev->delta_disks < 0) {
3866 BUG_ON(conf->reshape_progress == 0);
3867 stripe_addr = writepos;
3868 BUG_ON((mddev->dev_sectors &
3869 ~((sector_t)reshape_sectors - 1))
3870 - reshape_sectors - stripe_addr
3873 BUG_ON(writepos != sector_nr + reshape_sectors);
3874 stripe_addr = sector_nr;
3876 INIT_LIST_HEAD(&stripes);
3877 for (i = 0; i < reshape_sectors; i += STRIPE_SECTORS) {
3880 sh = get_active_stripe(conf, stripe_addr+i, 0, 0);
3881 set_bit(STRIPE_EXPANDING, &sh->state);
3882 atomic_inc(&conf->reshape_stripes);
3883 /* If any of this stripe is beyond the end of the old
3884 * array, then we need to zero those blocks
3886 for (j=sh->disks; j--;) {
3888 if (j == sh->pd_idx)
3890 if (conf->level == 6 &&
3893 s = compute_blocknr(sh, j, 0);
3894 if (s < raid5_size(mddev, 0, 0)) {
3898 memset(page_address(sh->dev[j].page), 0, STRIPE_SIZE);
3899 set_bit(R5_Expanded, &sh->dev[j].flags);
3900 set_bit(R5_UPTODATE, &sh->dev[j].flags);
3903 set_bit(STRIPE_EXPAND_READY, &sh->state);
3904 set_bit(STRIPE_HANDLE, &sh->state);
3906 list_add(&sh->lru, &stripes);
3908 spin_lock_irq(&conf->device_lock);
3909 if (mddev->delta_disks < 0)
3910 conf->reshape_progress -= reshape_sectors * new_data_disks;
3912 conf->reshape_progress += reshape_sectors * new_data_disks;
3913 spin_unlock_irq(&conf->device_lock);
3914 /* Ok, those stripe are ready. We can start scheduling
3915 * reads on the source stripes.
3916 * The source stripes are determined by mapping the first and last
3917 * block on the destination stripes.
3920 raid5_compute_sector(conf, stripe_addr*(new_data_disks),
3923 raid5_compute_sector(conf, ((stripe_addr+conf->chunk_size/512)
3924 *(new_data_disks) - 1),
3926 if (last_sector >= mddev->dev_sectors)
3927 last_sector = mddev->dev_sectors - 1;
3928 while (first_sector <= last_sector) {
3929 sh = get_active_stripe(conf, first_sector, 1, 0);
3930 set_bit(STRIPE_EXPAND_SOURCE, &sh->state);
3931 set_bit(STRIPE_HANDLE, &sh->state);
3933 first_sector += STRIPE_SECTORS;
3935 /* Now that the sources are clearly marked, we can release
3936 * the destination stripes
3938 while (!list_empty(&stripes)) {
3939 sh = list_entry(stripes.next, struct stripe_head, lru);
3940 list_del_init(&sh->lru);
3943 /* If this takes us to the resync_max point where we have to pause,
3944 * then we need to write out the superblock.
3946 sector_nr += reshape_sectors;
3947 if (sector_nr >= mddev->resync_max) {
3948 /* Cannot proceed until we've updated the superblock... */
3949 wait_event(conf->wait_for_overlap,
3950 atomic_read(&conf->reshape_stripes) == 0);
3951 mddev->reshape_position = conf->reshape_progress;
3952 conf->reshape_checkpoint = jiffies;
3953 set_bit(MD_CHANGE_DEVS, &mddev->flags);
3954 md_wakeup_thread(mddev->thread);
3955 wait_event(mddev->sb_wait,
3956 !test_bit(MD_CHANGE_DEVS, &mddev->flags)
3957 || kthread_should_stop());
3958 spin_lock_irq(&conf->device_lock);
3959 conf->reshape_safe = mddev->reshape_position;
3960 spin_unlock_irq(&conf->device_lock);
3961 wake_up(&conf->wait_for_overlap);
3963 return reshape_sectors;
3966 /* FIXME go_faster isn't used */
3967 static inline sector_t sync_request(mddev_t *mddev, sector_t sector_nr, int *skipped, int go_faster)
3969 raid5_conf_t *conf = (raid5_conf_t *) mddev->private;
3970 struct stripe_head *sh;
3971 sector_t max_sector = mddev->dev_sectors;
3973 int still_degraded = 0;
3976 if (sector_nr >= max_sector) {
3977 /* just being told to finish up .. nothing much to do */
3978 unplug_slaves(mddev);
3980 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery)) {
3985 if (mddev->curr_resync < max_sector) /* aborted */
3986 bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
3988 else /* completed sync */
3990 bitmap_close_sync(mddev->bitmap);
3995 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
3996 return reshape_request(mddev, sector_nr, skipped);
3998 /* No need to check resync_max as we never do more than one
3999 * stripe, and as resync_max will always be on a chunk boundary,
4000 * if the check in md_do_sync didn't fire, there is no chance
4001 * of overstepping resync_max here
4004 /* if there is too many failed drives and we are trying
4005 * to resync, then assert that we are finished, because there is
4006 * nothing we can do.
4008 if (mddev->degraded >= conf->max_degraded &&
4009 test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
4010 sector_t rv = mddev->dev_sectors - sector_nr;
4014 if (!bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
4015 !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
4016 !conf->fullsync && sync_blocks >= STRIPE_SECTORS) {
4017 /* we can skip this block, and probably more */
4018 sync_blocks /= STRIPE_SECTORS;
4020 return sync_blocks * STRIPE_SECTORS; /* keep things rounded to whole stripes */
4024 bitmap_cond_end_sync(mddev->bitmap, sector_nr);
4026 sh = get_active_stripe(conf, sector_nr, 0, 1);
4028 sh = get_active_stripe(conf, sector_nr, 0, 0);
4029 /* make sure we don't swamp the stripe cache if someone else
4030 * is trying to get access
4032 schedule_timeout_uninterruptible(1);
4034 /* Need to check if array will still be degraded after recovery/resync
4035 * We don't need to check the 'failed' flag as when that gets set,
4038 for (i=0; i<mddev->raid_disks; i++)
4039 if (conf->disks[i].rdev == NULL)
4042 bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, still_degraded);
4044 spin_lock(&sh->lock);
4045 set_bit(STRIPE_SYNCING, &sh->state);
4046 clear_bit(STRIPE_INSYNC, &sh->state);
4047 spin_unlock(&sh->lock);
4049 /* wait for any blocked device to be handled */
4050 while(unlikely(!handle_stripe(sh, NULL)))
4054 return STRIPE_SECTORS;
4057 static int retry_aligned_read(raid5_conf_t *conf, struct bio *raid_bio)
4059 /* We may not be able to submit a whole bio at once as there
4060 * may not be enough stripe_heads available.
4061 * We cannot pre-allocate enough stripe_heads as we may need
4062 * more than exist in the cache (if we allow ever large chunks).
4063 * So we do one stripe head at a time and record in
4064 * ->bi_hw_segments how many have been done.
4066 * We *know* that this entire raid_bio is in one chunk, so
4067 * it will be only one 'dd_idx' and only need one call to raid5_compute_sector.
4069 struct stripe_head *sh;
4071 sector_t sector, logical_sector, last_sector;
4076 logical_sector = raid_bio->bi_sector & ~((sector_t)STRIPE_SECTORS-1);
4077 sector = raid5_compute_sector(conf, logical_sector,
4079 last_sector = raid_bio->bi_sector + (raid_bio->bi_size>>9);
4081 for (; logical_sector < last_sector;
4082 logical_sector += STRIPE_SECTORS,
4083 sector += STRIPE_SECTORS,
4086 if (scnt < raid5_bi_hw_segments(raid_bio))
4087 /* already done this stripe */
4090 sh = get_active_stripe(conf, sector, 0, 1);
4093 /* failed to get a stripe - must wait */
4094 raid5_set_bi_hw_segments(raid_bio, scnt);
4095 conf->retry_read_aligned = raid_bio;
4099 set_bit(R5_ReadError, &sh->dev[dd_idx].flags);
4100 if (!add_stripe_bio(sh, raid_bio, dd_idx, 0)) {
4102 raid5_set_bi_hw_segments(raid_bio, scnt);
4103 conf->retry_read_aligned = raid_bio;
4107 handle_stripe(sh, NULL);
4111 spin_lock_irq(&conf->device_lock);
4112 remaining = raid5_dec_bi_phys_segments(raid_bio);
4113 spin_unlock_irq(&conf->device_lock);
4115 bio_endio(raid_bio, 0);
4116 if (atomic_dec_and_test(&conf->active_aligned_reads))
4117 wake_up(&conf->wait_for_stripe);
4124 * This is our raid5 kernel thread.
4126 * We scan the hash table for stripes which can be handled now.
4127 * During the scan, completed stripes are saved for us by the interrupt
4128 * handler, so that they will not have to wait for our next wakeup.
4130 static void raid5d(mddev_t *mddev)
4132 struct stripe_head *sh;
4133 raid5_conf_t *conf = mddev_to_conf(mddev);
4136 pr_debug("+++ raid5d active\n");
4138 md_check_recovery(mddev);
4141 spin_lock_irq(&conf->device_lock);
4145 if (conf->seq_flush != conf->seq_write) {
4146 int seq = conf->seq_flush;
4147 spin_unlock_irq(&conf->device_lock);
4148 bitmap_unplug(mddev->bitmap);
4149 spin_lock_irq(&conf->device_lock);
4150 conf->seq_write = seq;
4151 activate_bit_delay(conf);
4154 while ((bio = remove_bio_from_retry(conf))) {
4156 spin_unlock_irq(&conf->device_lock);
4157 ok = retry_aligned_read(conf, bio);
4158 spin_lock_irq(&conf->device_lock);
4164 sh = __get_priority_stripe(conf);
4168 spin_unlock_irq(&conf->device_lock);
4171 handle_stripe(sh, conf->spare_page);
4174 spin_lock_irq(&conf->device_lock);
4176 pr_debug("%d stripes handled\n", handled);
4178 spin_unlock_irq(&conf->device_lock);
4180 async_tx_issue_pending_all();
4181 unplug_slaves(mddev);
4183 pr_debug("--- raid5d inactive\n");
4187 raid5_show_stripe_cache_size(mddev_t *mddev, char *page)
4189 raid5_conf_t *conf = mddev_to_conf(mddev);
4191 return sprintf(page, "%d\n", conf->max_nr_stripes);
4197 raid5_store_stripe_cache_size(mddev_t *mddev, const char *page, size_t len)
4199 raid5_conf_t *conf = mddev_to_conf(mddev);
4203 if (len >= PAGE_SIZE)
4208 if (strict_strtoul(page, 10, &new))
4210 if (new <= 16 || new > 32768)
4212 while (new < conf->max_nr_stripes) {
4213 if (drop_one_stripe(conf))
4214 conf->max_nr_stripes--;
4218 err = md_allow_write(mddev);
4221 while (new > conf->max_nr_stripes) {
4222 if (grow_one_stripe(conf))
4223 conf->max_nr_stripes++;
4229 static struct md_sysfs_entry
4230 raid5_stripecache_size = __ATTR(stripe_cache_size, S_IRUGO | S_IWUSR,
4231 raid5_show_stripe_cache_size,
4232 raid5_store_stripe_cache_size);
4235 raid5_show_preread_threshold(mddev_t *mddev, char *page)
4237 raid5_conf_t *conf = mddev_to_conf(mddev);
4239 return sprintf(page, "%d\n", conf->bypass_threshold);
4245 raid5_store_preread_threshold(mddev_t *mddev, const char *page, size_t len)
4247 raid5_conf_t *conf = mddev_to_conf(mddev);
4249 if (len >= PAGE_SIZE)
4254 if (strict_strtoul(page, 10, &new))
4256 if (new > conf->max_nr_stripes)
4258 conf->bypass_threshold = new;
4262 static struct md_sysfs_entry
4263 raid5_preread_bypass_threshold = __ATTR(preread_bypass_threshold,
4265 raid5_show_preread_threshold,
4266 raid5_store_preread_threshold);
4269 stripe_cache_active_show(mddev_t *mddev, char *page)
4271 raid5_conf_t *conf = mddev_to_conf(mddev);
4273 return sprintf(page, "%d\n", atomic_read(&conf->active_stripes));
4278 static struct md_sysfs_entry
4279 raid5_stripecache_active = __ATTR_RO(stripe_cache_active);
4281 static struct attribute *raid5_attrs[] = {
4282 &raid5_stripecache_size.attr,
4283 &raid5_stripecache_active.attr,
4284 &raid5_preread_bypass_threshold.attr,
4287 static struct attribute_group raid5_attrs_group = {
4289 .attrs = raid5_attrs,
4293 raid5_size(mddev_t *mddev, sector_t sectors, int raid_disks)
4295 raid5_conf_t *conf = mddev_to_conf(mddev);
4298 sectors = mddev->dev_sectors;
4300 /* size is defined by the smallest of previous and new size */
4301 if (conf->raid_disks < conf->previous_raid_disks)
4302 raid_disks = conf->raid_disks;
4304 raid_disks = conf->previous_raid_disks;
4307 sectors &= ~((sector_t)mddev->chunk_size/512 - 1);
4308 sectors &= ~((sector_t)mddev->new_chunk/512 - 1);
4309 return sectors * (raid_disks - conf->max_degraded);
4312 static raid5_conf_t *setup_conf(mddev_t *mddev)
4315 int raid_disk, memory;
4317 struct disk_info *disk;
4319 if (mddev->new_level != 5
4320 && mddev->new_level != 4
4321 && mddev->new_level != 6) {
4322 printk(KERN_ERR "raid5: %s: raid level not set to 4/5/6 (%d)\n",
4323 mdname(mddev), mddev->new_level);
4324 return ERR_PTR(-EIO);
4326 if ((mddev->new_level == 5
4327 && !algorithm_valid_raid5(mddev->new_layout)) ||
4328 (mddev->new_level == 6
4329 && !algorithm_valid_raid6(mddev->new_layout))) {
4330 printk(KERN_ERR "raid5: %s: layout %d not supported\n",
4331 mdname(mddev), mddev->new_layout);
4332 return ERR_PTR(-EIO);
4334 if (mddev->new_level == 6 && mddev->raid_disks < 4) {
4335 printk(KERN_ERR "raid6: not enough configured devices for %s (%d, minimum 4)\n",
4336 mdname(mddev), mddev->raid_disks);
4337 return ERR_PTR(-EINVAL);
4340 if (!mddev->new_chunk || mddev->new_chunk % PAGE_SIZE) {
4341 printk(KERN_ERR "raid5: invalid chunk size %d for %s\n",
4342 mddev->new_chunk, mdname(mddev));
4343 return ERR_PTR(-EINVAL);
4346 conf = kzalloc(sizeof(raid5_conf_t), GFP_KERNEL);
4350 conf->raid_disks = mddev->raid_disks;
4351 if (mddev->reshape_position == MaxSector)
4352 conf->previous_raid_disks = mddev->raid_disks;
4354 conf->previous_raid_disks = mddev->raid_disks - mddev->delta_disks;
4356 conf->disks = kzalloc(conf->raid_disks * sizeof(struct disk_info),
4361 conf->mddev = mddev;
4363 if ((conf->stripe_hashtbl = kzalloc(PAGE_SIZE, GFP_KERNEL)) == NULL)
4366 if (mddev->new_level == 6) {
4367 conf->spare_page = alloc_page(GFP_KERNEL);
4368 if (!conf->spare_page)
4371 spin_lock_init(&conf->device_lock);
4372 init_waitqueue_head(&conf->wait_for_stripe);
4373 init_waitqueue_head(&conf->wait_for_overlap);
4374 INIT_LIST_HEAD(&conf->handle_list);
4375 INIT_LIST_HEAD(&conf->hold_list);
4376 INIT_LIST_HEAD(&conf->delayed_list);
4377 INIT_LIST_HEAD(&conf->bitmap_list);
4378 INIT_LIST_HEAD(&conf->inactive_list);
4379 atomic_set(&conf->active_stripes, 0);
4380 atomic_set(&conf->preread_active_stripes, 0);
4381 atomic_set(&conf->active_aligned_reads, 0);
4382 conf->bypass_threshold = BYPASS_THRESHOLD;
4384 pr_debug("raid5: run(%s) called.\n", mdname(mddev));
4386 list_for_each_entry(rdev, &mddev->disks, same_set) {
4387 raid_disk = rdev->raid_disk;
4388 if (raid_disk >= conf->raid_disks
4391 disk = conf->disks + raid_disk;
4395 if (test_bit(In_sync, &rdev->flags)) {
4396 char b[BDEVNAME_SIZE];
4397 printk(KERN_INFO "raid5: device %s operational as raid"
4398 " disk %d\n", bdevname(rdev->bdev,b),
4401 /* Cannot rely on bitmap to complete recovery */
4405 conf->chunk_size = mddev->new_chunk;
4406 conf->level = mddev->new_level;
4407 if (conf->level == 6)
4408 conf->max_degraded = 2;
4410 conf->max_degraded = 1;
4411 conf->algorithm = mddev->new_layout;
4412 conf->max_nr_stripes = NR_STRIPES;
4413 conf->reshape_progress = mddev->reshape_position;
4414 if (conf->reshape_progress != MaxSector) {
4415 conf->prev_chunk = mddev->chunk_size;
4416 conf->prev_algo = mddev->layout;
4419 memory = conf->max_nr_stripes * (sizeof(struct stripe_head) +
4420 conf->raid_disks * ((sizeof(struct bio) + PAGE_SIZE))) / 1024;
4421 if (grow_stripes(conf, conf->max_nr_stripes)) {
4423 "raid5: couldn't allocate %dkB for buffers\n", memory);
4426 printk(KERN_INFO "raid5: allocated %dkB for %s\n",
4427 memory, mdname(mddev));
4429 conf->thread = md_register_thread(raid5d, mddev, "%s_raid5");
4430 if (!conf->thread) {
4432 "raid5: couldn't allocate thread for %s\n",
4441 shrink_stripes(conf);
4442 safe_put_page(conf->spare_page);
4444 kfree(conf->stripe_hashtbl);
4446 return ERR_PTR(-EIO);
4448 return ERR_PTR(-ENOMEM);
4451 static int run(mddev_t *mddev)
4454 int working_disks = 0;
4457 if (mddev->reshape_position != MaxSector) {
4458 /* Check that we can continue the reshape.
4459 * Currently only disks can change, it must
4460 * increase, and we must be past the point where
4461 * a stripe over-writes itself
4463 sector_t here_new, here_old;
4465 int max_degraded = (mddev->level == 6 ? 2 : 1);
4467 if (mddev->new_level != mddev->level) {
4468 printk(KERN_ERR "raid5: %s: unsupported reshape "
4469 "required - aborting.\n",
4473 old_disks = mddev->raid_disks - mddev->delta_disks;
4474 /* reshape_position must be on a new-stripe boundary, and one
4475 * further up in new geometry must map after here in old
4478 here_new = mddev->reshape_position;
4479 if (sector_div(here_new, (mddev->new_chunk>>9)*
4480 (mddev->raid_disks - max_degraded))) {
4481 printk(KERN_ERR "raid5: reshape_position not "
4482 "on a stripe boundary\n");
4485 /* here_new is the stripe we will write to */
4486 here_old = mddev->reshape_position;
4487 sector_div(here_old, (mddev->chunk_size>>9)*
4488 (old_disks-max_degraded));
4489 /* here_old is the first stripe that we might need to read
4491 if (here_new >= here_old) {
4492 /* Reading from the same stripe as writing to - bad */
4493 printk(KERN_ERR "raid5: reshape_position too early for "
4494 "auto-recovery - aborting.\n");
4497 printk(KERN_INFO "raid5: reshape will continue\n");
4498 /* OK, we should be able to continue; */
4500 BUG_ON(mddev->level != mddev->new_level);
4501 BUG_ON(mddev->layout != mddev->new_layout);
4502 BUG_ON(mddev->chunk_size != mddev->new_chunk);
4503 BUG_ON(mddev->delta_disks != 0);
4506 if (mddev->private == NULL)
4507 conf = setup_conf(mddev);
4509 conf = mddev->private;
4512 return PTR_ERR(conf);
4514 mddev->thread = conf->thread;
4515 conf->thread = NULL;
4516 mddev->private = conf;
4519 * 0 for a fully functional array, 1 or 2 for a degraded array.
4521 list_for_each_entry(rdev, &mddev->disks, same_set)
4522 if (rdev->raid_disk >= 0 &&
4523 test_bit(In_sync, &rdev->flags))
4526 mddev->degraded = conf->raid_disks - working_disks;
4528 if (mddev->degraded > conf->max_degraded) {
4529 printk(KERN_ERR "raid5: not enough operational devices for %s"
4530 " (%d/%d failed)\n",
4531 mdname(mddev), mddev->degraded, conf->raid_disks);
4535 /* device size must be a multiple of chunk size */
4536 mddev->dev_sectors &= ~(mddev->chunk_size / 512 - 1);
4537 mddev->resync_max_sectors = mddev->dev_sectors;
4539 if (mddev->degraded > 0 &&
4540 mddev->recovery_cp != MaxSector) {
4541 if (mddev->ok_start_degraded)
4543 "raid5: starting dirty degraded array: %s"
4544 "- data corruption possible.\n",
4548 "raid5: cannot start dirty degraded array for %s\n",
4554 if (mddev->degraded == 0)
4555 printk("raid5: raid level %d set %s active with %d out of %d"
4556 " devices, algorithm %d\n", conf->level, mdname(mddev),
4557 mddev->raid_disks-mddev->degraded, mddev->raid_disks,
4560 printk(KERN_ALERT "raid5: raid level %d set %s active with %d"
4561 " out of %d devices, algorithm %d\n", conf->level,
4562 mdname(mddev), mddev->raid_disks - mddev->degraded,
4563 mddev->raid_disks, mddev->new_layout);
4565 print_raid5_conf(conf);
4567 if (conf->reshape_progress != MaxSector) {
4568 printk("...ok start reshape thread\n");
4569 conf->reshape_safe = conf->reshape_progress;
4570 atomic_set(&conf->reshape_stripes, 0);
4571 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
4572 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
4573 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
4574 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
4575 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
4579 /* read-ahead size must cover two whole stripes, which is
4580 * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
4583 int data_disks = conf->previous_raid_disks - conf->max_degraded;
4584 int stripe = data_disks *
4585 (mddev->chunk_size / PAGE_SIZE);
4586 if (mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
4587 mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
4590 /* Ok, everything is just fine now */
4591 if (sysfs_create_group(&mddev->kobj, &raid5_attrs_group))
4593 "raid5: failed to create sysfs attributes for %s\n",
4596 mddev->queue->queue_lock = &conf->device_lock;
4598 mddev->queue->unplug_fn = raid5_unplug_device;
4599 mddev->queue->backing_dev_info.congested_data = mddev;
4600 mddev->queue->backing_dev_info.congested_fn = raid5_congested;
4602 md_set_array_sectors(mddev, raid5_size(mddev, 0, 0));
4604 blk_queue_merge_bvec(mddev->queue, raid5_mergeable_bvec);
4608 md_unregister_thread(mddev->thread);
4609 mddev->thread = NULL;
4611 shrink_stripes(conf);
4612 print_raid5_conf(conf);
4613 safe_put_page(conf->spare_page);
4615 kfree(conf->stripe_hashtbl);
4618 mddev->private = NULL;
4619 printk(KERN_ALERT "raid5: failed to run raid set %s\n", mdname(mddev));
4625 static int stop(mddev_t *mddev)
4627 raid5_conf_t *conf = (raid5_conf_t *) mddev->private;
4629 md_unregister_thread(mddev->thread);
4630 mddev->thread = NULL;
4631 shrink_stripes(conf);
4632 kfree(conf->stripe_hashtbl);
4633 mddev->queue->backing_dev_info.congested_fn = NULL;
4634 blk_sync_queue(mddev->queue); /* the unplug fn references 'conf'*/
4635 sysfs_remove_group(&mddev->kobj, &raid5_attrs_group);
4638 mddev->private = NULL;
4643 static void print_sh(struct seq_file *seq, struct stripe_head *sh)
4647 seq_printf(seq, "sh %llu, pd_idx %d, state %ld.\n",
4648 (unsigned long long)sh->sector, sh->pd_idx, sh->state);
4649 seq_printf(seq, "sh %llu, count %d.\n",
4650 (unsigned long long)sh->sector, atomic_read(&sh->count));
4651 seq_printf(seq, "sh %llu, ", (unsigned long long)sh->sector);
4652 for (i = 0; i < sh->disks; i++) {
4653 seq_printf(seq, "(cache%d: %p %ld) ",
4654 i, sh->dev[i].page, sh->dev[i].flags);
4656 seq_printf(seq, "\n");
4659 static void printall(struct seq_file *seq, raid5_conf_t *conf)
4661 struct stripe_head *sh;
4662 struct hlist_node *hn;
4665 spin_lock_irq(&conf->device_lock);
4666 for (i = 0; i < NR_HASH; i++) {
4667 hlist_for_each_entry(sh, hn, &conf->stripe_hashtbl[i], hash) {
4668 if (sh->raid_conf != conf)
4673 spin_unlock_irq(&conf->device_lock);
4677 static void status(struct seq_file *seq, mddev_t *mddev)
4679 raid5_conf_t *conf = (raid5_conf_t *) mddev->private;
4682 seq_printf (seq, " level %d, %dk chunk, algorithm %d", mddev->level, mddev->chunk_size >> 10, mddev->layout);
4683 seq_printf (seq, " [%d/%d] [", conf->raid_disks, conf->raid_disks - mddev->degraded);
4684 for (i = 0; i < conf->raid_disks; i++)
4685 seq_printf (seq, "%s",
4686 conf->disks[i].rdev &&
4687 test_bit(In_sync, &conf->disks[i].rdev->flags) ? "U" : "_");
4688 seq_printf (seq, "]");
4690 seq_printf (seq, "\n");
4691 printall(seq, conf);
4695 static void print_raid5_conf (raid5_conf_t *conf)
4698 struct disk_info *tmp;
4700 printk("RAID5 conf printout:\n");
4702 printk("(conf==NULL)\n");
4705 printk(" --- rd:%d wd:%d\n", conf->raid_disks,
4706 conf->raid_disks - conf->mddev->degraded);
4708 for (i = 0; i < conf->raid_disks; i++) {
4709 char b[BDEVNAME_SIZE];
4710 tmp = conf->disks + i;
4712 printk(" disk %d, o:%d, dev:%s\n",
4713 i, !test_bit(Faulty, &tmp->rdev->flags),
4714 bdevname(tmp->rdev->bdev,b));
4718 static int raid5_spare_active(mddev_t *mddev)
4721 raid5_conf_t *conf = mddev->private;
4722 struct disk_info *tmp;
4724 for (i = 0; i < conf->raid_disks; i++) {
4725 tmp = conf->disks + i;
4727 && !test_bit(Faulty, &tmp->rdev->flags)
4728 && !test_and_set_bit(In_sync, &tmp->rdev->flags)) {
4729 unsigned long flags;
4730 spin_lock_irqsave(&conf->device_lock, flags);
4732 spin_unlock_irqrestore(&conf->device_lock, flags);
4735 print_raid5_conf(conf);
4739 static int raid5_remove_disk(mddev_t *mddev, int number)
4741 raid5_conf_t *conf = mddev->private;
4744 struct disk_info *p = conf->disks + number;
4746 print_raid5_conf(conf);
4749 if (number >= conf->raid_disks &&
4750 conf->reshape_progress == MaxSector)
4751 clear_bit(In_sync, &rdev->flags);
4753 if (test_bit(In_sync, &rdev->flags) ||
4754 atomic_read(&rdev->nr_pending)) {
4758 /* Only remove non-faulty devices if recovery
4761 if (!test_bit(Faulty, &rdev->flags) &&
4762 mddev->degraded <= conf->max_degraded &&
4763 number < conf->raid_disks) {
4769 if (atomic_read(&rdev->nr_pending)) {
4770 /* lost the race, try later */
4777 print_raid5_conf(conf);
4781 static int raid5_add_disk(mddev_t *mddev, mdk_rdev_t *rdev)
4783 raid5_conf_t *conf = mddev->private;
4786 struct disk_info *p;
4788 int last = conf->raid_disks - 1;
4790 if (mddev->degraded > conf->max_degraded)
4791 /* no point adding a device */
4794 if (rdev->raid_disk >= 0)
4795 first = last = rdev->raid_disk;
4798 * find the disk ... but prefer rdev->saved_raid_disk
4801 if (rdev->saved_raid_disk >= 0 &&
4802 rdev->saved_raid_disk >= first &&
4803 conf->disks[rdev->saved_raid_disk].rdev == NULL)
4804 disk = rdev->saved_raid_disk;
4807 for ( ; disk <= last ; disk++)
4808 if ((p=conf->disks + disk)->rdev == NULL) {
4809 clear_bit(In_sync, &rdev->flags);
4810 rdev->raid_disk = disk;
4812 if (rdev->saved_raid_disk != disk)
4814 rcu_assign_pointer(p->rdev, rdev);
4817 print_raid5_conf(conf);
4821 static int raid5_resize(mddev_t *mddev, sector_t sectors)
4823 /* no resync is happening, and there is enough space
4824 * on all devices, so we can resize.
4825 * We need to make sure resync covers any new space.
4826 * If the array is shrinking we should possibly wait until
4827 * any io in the removed space completes, but it hardly seems
4830 sectors &= ~((sector_t)mddev->chunk_size/512 - 1);
4831 md_set_array_sectors(mddev, raid5_size(mddev, sectors,
4832 mddev->raid_disks));
4833 if (mddev->array_sectors >
4834 raid5_size(mddev, sectors, mddev->raid_disks))
4836 set_capacity(mddev->gendisk, mddev->array_sectors);
4838 if (sectors > mddev->dev_sectors && mddev->recovery_cp == MaxSector) {
4839 mddev->recovery_cp = mddev->dev_sectors;
4840 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
4842 mddev->dev_sectors = sectors;
4843 mddev->resync_max_sectors = sectors;
4847 static int raid5_check_reshape(mddev_t *mddev)
4849 raid5_conf_t *conf = mddev_to_conf(mddev);
4851 if (mddev->delta_disks == 0 &&
4852 mddev->new_layout == mddev->layout &&
4853 mddev->new_chunk == mddev->chunk_size)
4854 return -EINVAL; /* nothing to do */
4856 /* Cannot grow a bitmap yet */
4858 if (mddev->degraded > conf->max_degraded)
4860 if (mddev->delta_disks < 0) {
4861 /* We might be able to shrink, but the devices must
4862 * be made bigger first.
4863 * For raid6, 4 is the minimum size.
4864 * Otherwise 2 is the minimum
4867 if (mddev->level == 6)
4869 if (mddev->raid_disks + mddev->delta_disks < min)
4873 /* Can only proceed if there are plenty of stripe_heads.
4874 * We need a minimum of one full stripe,, and for sensible progress
4875 * it is best to have about 4 times that.
4876 * If we require 4 times, then the default 256 4K stripe_heads will
4877 * allow for chunk sizes up to 256K, which is probably OK.
4878 * If the chunk size is greater, user-space should request more
4879 * stripe_heads first.
4881 if ((mddev->chunk_size / STRIPE_SIZE) * 4 > conf->max_nr_stripes ||
4882 (mddev->new_chunk / STRIPE_SIZE) * 4 > conf->max_nr_stripes) {
4883 printk(KERN_WARNING "raid5: reshape: not enough stripes. Needed %lu\n",
4884 (max(mddev->chunk_size, mddev->new_chunk)
4889 return resize_stripes(conf, conf->raid_disks + mddev->delta_disks);
4892 static int raid5_start_reshape(mddev_t *mddev)
4894 raid5_conf_t *conf = mddev_to_conf(mddev);
4897 int added_devices = 0;
4898 unsigned long flags;
4900 if (test_bit(MD_RECOVERY_RUNNING, &mddev->recovery))
4903 list_for_each_entry(rdev, &mddev->disks, same_set)
4904 if (rdev->raid_disk < 0 &&
4905 !test_bit(Faulty, &rdev->flags))
4908 if (spares - mddev->degraded < mddev->delta_disks - conf->max_degraded)
4909 /* Not enough devices even to make a degraded array
4914 /* Refuse to reduce size of the array. Any reductions in
4915 * array size must be through explicit setting of array_size
4918 if (raid5_size(mddev, 0, conf->raid_disks + mddev->delta_disks)
4919 < mddev->array_sectors) {
4920 printk(KERN_ERR "md: %s: array size must be reduced "
4921 "before number of disks\n", mdname(mddev));
4925 atomic_set(&conf->reshape_stripes, 0);
4926 spin_lock_irq(&conf->device_lock);
4927 conf->previous_raid_disks = conf->raid_disks;
4928 conf->raid_disks += mddev->delta_disks;
4929 conf->prev_chunk = conf->chunk_size;
4930 conf->chunk_size = mddev->new_chunk;
4931 conf->prev_algo = conf->algorithm;
4932 conf->algorithm = mddev->new_layout;
4933 if (mddev->delta_disks < 0)
4934 conf->reshape_progress = raid5_size(mddev, 0, 0);
4936 conf->reshape_progress = 0;
4937 conf->reshape_safe = conf->reshape_progress;
4939 spin_unlock_irq(&conf->device_lock);
4941 /* Add some new drives, as many as will fit.
4942 * We know there are enough to make the newly sized array work.
4944 list_for_each_entry(rdev, &mddev->disks, same_set)
4945 if (rdev->raid_disk < 0 &&
4946 !test_bit(Faulty, &rdev->flags)) {
4947 if (raid5_add_disk(mddev, rdev) == 0) {
4949 set_bit(In_sync, &rdev->flags);
4951 rdev->recovery_offset = 0;
4952 sprintf(nm, "rd%d", rdev->raid_disk);
4953 if (sysfs_create_link(&mddev->kobj,
4956 "raid5: failed to create "
4957 " link %s for %s\n",
4963 if (mddev->delta_disks > 0) {
4964 spin_lock_irqsave(&conf->device_lock, flags);
4965 mddev->degraded = (conf->raid_disks - conf->previous_raid_disks)
4967 spin_unlock_irqrestore(&conf->device_lock, flags);
4969 mddev->raid_disks = conf->raid_disks;
4970 mddev->reshape_position = 0;
4971 set_bit(MD_CHANGE_DEVS, &mddev->flags);
4973 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
4974 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
4975 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
4976 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
4977 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
4979 if (!mddev->sync_thread) {
4980 mddev->recovery = 0;
4981 spin_lock_irq(&conf->device_lock);
4982 mddev->raid_disks = conf->raid_disks = conf->previous_raid_disks;
4983 conf->reshape_progress = MaxSector;
4984 spin_unlock_irq(&conf->device_lock);
4987 conf->reshape_checkpoint = jiffies;
4988 md_wakeup_thread(mddev->sync_thread);
4989 md_new_event(mddev);
4993 /* This is called from the reshape thread and should make any
4994 * changes needed in 'conf'
4996 static void end_reshape(raid5_conf_t *conf)
4999 if (!test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery)) {
5001 spin_lock_irq(&conf->device_lock);
5002 conf->previous_raid_disks = conf->raid_disks;
5003 conf->reshape_progress = MaxSector;
5004 spin_unlock_irq(&conf->device_lock);
5005 wake_up(&conf->wait_for_overlap);
5007 /* read-ahead size must cover two whole stripes, which is
5008 * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
5011 int data_disks = conf->raid_disks - conf->max_degraded;
5012 int stripe = data_disks * (conf->chunk_size
5014 if (conf->mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
5015 conf->mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
5020 /* This is called from the raid5d thread with mddev_lock held.
5021 * It makes config changes to the device.
5023 static void raid5_finish_reshape(mddev_t *mddev)
5025 struct block_device *bdev;
5026 raid5_conf_t *conf = mddev_to_conf(mddev);
5028 if (!test_bit(MD_RECOVERY_INTR, &mddev->recovery)) {
5030 if (mddev->delta_disks > 0) {
5031 md_set_array_sectors(mddev, raid5_size(mddev, 0, 0));
5032 set_capacity(mddev->gendisk, mddev->array_sectors);
5035 bdev = bdget_disk(mddev->gendisk, 0);
5037 mutex_lock(&bdev->bd_inode->i_mutex);
5038 i_size_write(bdev->bd_inode,
5039 (loff_t)mddev->array_sectors << 9);
5040 mutex_unlock(&bdev->bd_inode->i_mutex);
5045 mddev->degraded = conf->raid_disks;
5046 for (d = 0; d < conf->raid_disks ; d++)
5047 if (conf->disks[d].rdev &&
5049 &conf->disks[d].rdev->flags))
5051 for (d = conf->raid_disks ;
5052 d < conf->raid_disks - mddev->delta_disks;
5054 raid5_remove_disk(mddev, d);
5056 mddev->layout = conf->algorithm;
5057 mddev->chunk_size = conf->chunk_size;
5058 mddev->reshape_position = MaxSector;
5059 mddev->delta_disks = 0;
5063 static void raid5_quiesce(mddev_t *mddev, int state)
5065 raid5_conf_t *conf = mddev_to_conf(mddev);
5068 case 2: /* resume for a suspend */
5069 wake_up(&conf->wait_for_overlap);
5072 case 1: /* stop all writes */
5073 spin_lock_irq(&conf->device_lock);
5075 wait_event_lock_irq(conf->wait_for_stripe,
5076 atomic_read(&conf->active_stripes) == 0 &&
5077 atomic_read(&conf->active_aligned_reads) == 0,
5078 conf->device_lock, /* nothing */);
5079 spin_unlock_irq(&conf->device_lock);
5082 case 0: /* re-enable writes */
5083 spin_lock_irq(&conf->device_lock);
5085 wake_up(&conf->wait_for_stripe);
5086 wake_up(&conf->wait_for_overlap);
5087 spin_unlock_irq(&conf->device_lock);
5093 static void *raid5_takeover_raid1(mddev_t *mddev)
5097 if (mddev->raid_disks != 2 ||
5098 mddev->degraded > 1)
5099 return ERR_PTR(-EINVAL);
5101 /* Should check if there are write-behind devices? */
5103 chunksect = 64*2; /* 64K by default */
5105 /* The array must be an exact multiple of chunksize */
5106 while (chunksect && (mddev->array_sectors & (chunksect-1)))
5109 if ((chunksect<<9) < STRIPE_SIZE)
5110 /* array size does not allow a suitable chunk size */
5111 return ERR_PTR(-EINVAL);
5113 mddev->new_level = 5;
5114 mddev->new_layout = ALGORITHM_LEFT_SYMMETRIC;
5115 mddev->new_chunk = chunksect << 9;
5117 return setup_conf(mddev);
5120 static void *raid5_takeover_raid6(mddev_t *mddev)
5124 switch (mddev->layout) {
5125 case ALGORITHM_LEFT_ASYMMETRIC_6:
5126 new_layout = ALGORITHM_LEFT_ASYMMETRIC;
5128 case ALGORITHM_RIGHT_ASYMMETRIC_6:
5129 new_layout = ALGORITHM_RIGHT_ASYMMETRIC;
5131 case ALGORITHM_LEFT_SYMMETRIC_6:
5132 new_layout = ALGORITHM_LEFT_SYMMETRIC;
5134 case ALGORITHM_RIGHT_SYMMETRIC_6:
5135 new_layout = ALGORITHM_RIGHT_SYMMETRIC;
5137 case ALGORITHM_PARITY_0_6:
5138 new_layout = ALGORITHM_PARITY_0;
5140 case ALGORITHM_PARITY_N:
5141 new_layout = ALGORITHM_PARITY_N;
5144 return ERR_PTR(-EINVAL);
5146 mddev->new_level = 5;
5147 mddev->new_layout = new_layout;
5148 mddev->delta_disks = -1;
5149 mddev->raid_disks -= 1;
5150 return setup_conf(mddev);
5154 static int raid5_reconfig(mddev_t *mddev, int new_layout, int new_chunk)
5156 /* For a 2-drive array, the layout and chunk size can be changed
5157 * immediately as not restriping is needed.
5158 * For larger arrays we record the new value - after validation
5159 * to be used by a reshape pass.
5161 raid5_conf_t *conf = mddev_to_conf(mddev);
5163 if (new_layout >= 0 && !algorithm_valid_raid5(new_layout))
5165 if (new_chunk > 0) {
5166 if (new_chunk & (new_chunk-1))
5167 /* not a power of 2 */
5169 if (new_chunk < PAGE_SIZE)
5171 if (mddev->array_sectors & ((new_chunk>>9)-1))
5172 /* not factor of array size */
5176 /* They look valid */
5178 if (mddev->raid_disks == 2) {
5180 if (new_layout >= 0) {
5181 conf->algorithm = new_layout;
5182 mddev->layout = mddev->new_layout = new_layout;
5184 if (new_chunk > 0) {
5185 conf->chunk_size = new_chunk;
5186 mddev->chunk_size = mddev->new_chunk = new_chunk;
5188 set_bit(MD_CHANGE_DEVS, &mddev->flags);
5189 md_wakeup_thread(mddev->thread);
5191 if (new_layout >= 0)
5192 mddev->new_layout = new_layout;
5194 mddev->new_chunk = new_chunk;
5199 static int raid6_reconfig(mddev_t *mddev, int new_layout, int new_chunk)
5201 if (new_layout >= 0 && !algorithm_valid_raid6(new_layout))
5203 if (new_chunk > 0) {
5204 if (new_chunk & (new_chunk-1))
5205 /* not a power of 2 */
5207 if (new_chunk < PAGE_SIZE)
5209 if (mddev->array_sectors & ((new_chunk>>9)-1))
5210 /* not factor of array size */
5214 /* They look valid */
5216 if (new_layout >= 0)
5217 mddev->new_layout = new_layout;
5219 mddev->new_chunk = new_chunk;
5224 static void *raid5_takeover(mddev_t *mddev)
5226 /* raid5 can take over:
5227 * raid0 - if all devices are the same - make it a raid4 layout
5228 * raid1 - if there are two drives. We need to know the chunk size
5229 * raid4 - trivial - just use a raid4 layout.
5230 * raid6 - Providing it is a *_6 layout
5232 * For now, just do raid1
5235 if (mddev->level == 1)
5236 return raid5_takeover_raid1(mddev);
5237 if (mddev->level == 4) {
5238 mddev->new_layout = ALGORITHM_PARITY_N;
5239 mddev->new_level = 5;
5240 return setup_conf(mddev);
5242 if (mddev->level == 6)
5243 return raid5_takeover_raid6(mddev);
5245 return ERR_PTR(-EINVAL);
5249 static struct mdk_personality raid5_personality;
5251 static void *raid6_takeover(mddev_t *mddev)
5253 /* Currently can only take over a raid5. We map the
5254 * personality to an equivalent raid6 personality
5255 * with the Q block at the end.
5259 if (mddev->pers != &raid5_personality)
5260 return ERR_PTR(-EINVAL);
5261 if (mddev->degraded > 1)
5262 return ERR_PTR(-EINVAL);
5263 if (mddev->raid_disks > 253)
5264 return ERR_PTR(-EINVAL);
5265 if (mddev->raid_disks < 3)
5266 return ERR_PTR(-EINVAL);
5268 switch (mddev->layout) {
5269 case ALGORITHM_LEFT_ASYMMETRIC:
5270 new_layout = ALGORITHM_LEFT_ASYMMETRIC_6;
5272 case ALGORITHM_RIGHT_ASYMMETRIC:
5273 new_layout = ALGORITHM_RIGHT_ASYMMETRIC_6;
5275 case ALGORITHM_LEFT_SYMMETRIC:
5276 new_layout = ALGORITHM_LEFT_SYMMETRIC_6;
5278 case ALGORITHM_RIGHT_SYMMETRIC:
5279 new_layout = ALGORITHM_RIGHT_SYMMETRIC_6;
5281 case ALGORITHM_PARITY_0:
5282 new_layout = ALGORITHM_PARITY_0_6;
5284 case ALGORITHM_PARITY_N:
5285 new_layout = ALGORITHM_PARITY_N;
5288 return ERR_PTR(-EINVAL);
5290 mddev->new_level = 6;
5291 mddev->new_layout = new_layout;
5292 mddev->delta_disks = 1;
5293 mddev->raid_disks += 1;
5294 return setup_conf(mddev);
5298 static struct mdk_personality raid6_personality =
5302 .owner = THIS_MODULE,
5303 .make_request = make_request,
5307 .error_handler = error,
5308 .hot_add_disk = raid5_add_disk,
5309 .hot_remove_disk= raid5_remove_disk,
5310 .spare_active = raid5_spare_active,
5311 .sync_request = sync_request,
5312 .resize = raid5_resize,
5314 .check_reshape = raid5_check_reshape,
5315 .start_reshape = raid5_start_reshape,
5316 .finish_reshape = raid5_finish_reshape,
5317 .quiesce = raid5_quiesce,
5318 .takeover = raid6_takeover,
5319 .reconfig = raid6_reconfig,
5321 static struct mdk_personality raid5_personality =
5325 .owner = THIS_MODULE,
5326 .make_request = make_request,
5330 .error_handler = error,
5331 .hot_add_disk = raid5_add_disk,
5332 .hot_remove_disk= raid5_remove_disk,
5333 .spare_active = raid5_spare_active,
5334 .sync_request = sync_request,
5335 .resize = raid5_resize,
5337 .check_reshape = raid5_check_reshape,
5338 .start_reshape = raid5_start_reshape,
5339 .finish_reshape = raid5_finish_reshape,
5340 .quiesce = raid5_quiesce,
5341 .takeover = raid5_takeover,
5342 .reconfig = raid5_reconfig,
5345 static struct mdk_personality raid4_personality =
5349 .owner = THIS_MODULE,
5350 .make_request = make_request,
5354 .error_handler = error,
5355 .hot_add_disk = raid5_add_disk,
5356 .hot_remove_disk= raid5_remove_disk,
5357 .spare_active = raid5_spare_active,
5358 .sync_request = sync_request,
5359 .resize = raid5_resize,
5361 .check_reshape = raid5_check_reshape,
5362 .start_reshape = raid5_start_reshape,
5363 .finish_reshape = raid5_finish_reshape,
5364 .quiesce = raid5_quiesce,
5367 static int __init raid5_init(void)
5369 register_md_personality(&raid6_personality);
5370 register_md_personality(&raid5_personality);
5371 register_md_personality(&raid4_personality);
5375 static void raid5_exit(void)
5377 unregister_md_personality(&raid6_personality);
5378 unregister_md_personality(&raid5_personality);
5379 unregister_md_personality(&raid4_personality);
5382 module_init(raid5_init);
5383 module_exit(raid5_exit);
5384 MODULE_LICENSE("GPL");
5385 MODULE_ALIAS("md-personality-4"); /* RAID5 */
5386 MODULE_ALIAS("md-raid5");
5387 MODULE_ALIAS("md-raid4");
5388 MODULE_ALIAS("md-level-5");
5389 MODULE_ALIAS("md-level-4");
5390 MODULE_ALIAS("md-personality-8"); /* RAID6 */
5391 MODULE_ALIAS("md-raid6");
5392 MODULE_ALIAS("md-level-6");
5394 /* This used to be two separate modules, they were: */
5395 MODULE_ALIAS("raid5");
5396 MODULE_ALIAS("raid6");