md/raid5: remove some sparse warnings.
[linux-2.6/linux-acpi-2.6/ibm-acpi-2.6.git] / drivers / md / raid5.c
blob34cb065f6d668b0102a2acf82d5459defd79ccb9
1 /*
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)
14 * any later version.
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.
22 * BITMAP UNPLUGGING:
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
26 * explanation.
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
32 * new additions.
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
39 * batch.
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
43 * miss any bits.
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/async.h>
51 #include <linux/seq_file.h>
52 #include <linux/cpu.h>
53 #include "md.h"
54 #include "raid5.h"
55 #include "bitmap.h"
58 * Stripe cache
61 #define NR_STRIPES 256
62 #define STRIPE_SIZE PAGE_SIZE
63 #define STRIPE_SHIFT (PAGE_SHIFT - 9)
64 #define STRIPE_SECTORS (STRIPE_SIZE>>9)
65 #define IO_THRESHOLD 1
66 #define BYPASS_THRESHOLD 1
67 #define NR_HASH (PAGE_SIZE / sizeof(struct hlist_head))
68 #define HASH_MASK (NR_HASH - 1)
70 #define stripe_hash(conf, sect) (&((conf)->stripe_hashtbl[((sect) >> STRIPE_SHIFT) & HASH_MASK]))
72 /* bio's attached to a stripe+device for I/O are linked together in bi_sector
73 * order without overlap. There may be several bio's per stripe+device, and
74 * a bio could span several devices.
75 * When walking this list for a particular stripe+device, we must never proceed
76 * beyond a bio that extends past this device, as the next bio might no longer
77 * be valid.
78 * This macro is used to determine the 'next' bio in the list, given the sector
79 * of the current stripe+device
81 #define r5_next_bio(bio, sect) ( ( (bio)->bi_sector + ((bio)->bi_size>>9) < sect + STRIPE_SECTORS) ? (bio)->bi_next : NULL)
83 * The following can be used to debug the driver
85 #define RAID5_PARANOIA 1
86 #if RAID5_PARANOIA && defined(CONFIG_SMP)
87 # define CHECK_DEVLOCK() assert_spin_locked(&conf->device_lock)
88 #else
89 # define CHECK_DEVLOCK()
90 #endif
92 #ifdef DEBUG
93 #define inline
94 #define __inline__
95 #endif
97 #define printk_rl(args...) ((void) (printk_ratelimit() && printk(args)))
100 * We maintain a biased count of active stripes in the bottom 16 bits of
101 * bi_phys_segments, and a count of processed stripes in the upper 16 bits
103 static inline int raid5_bi_phys_segments(struct bio *bio)
105 return bio->bi_phys_segments & 0xffff;
108 static inline int raid5_bi_hw_segments(struct bio *bio)
110 return (bio->bi_phys_segments >> 16) & 0xffff;
113 static inline int raid5_dec_bi_phys_segments(struct bio *bio)
115 --bio->bi_phys_segments;
116 return raid5_bi_phys_segments(bio);
119 static inline int raid5_dec_bi_hw_segments(struct bio *bio)
121 unsigned short val = raid5_bi_hw_segments(bio);
123 --val;
124 bio->bi_phys_segments = (val << 16) | raid5_bi_phys_segments(bio);
125 return val;
128 static inline void raid5_set_bi_hw_segments(struct bio *bio, unsigned int cnt)
130 bio->bi_phys_segments = raid5_bi_phys_segments(bio) || (cnt << 16);
133 /* Find first data disk in a raid6 stripe */
134 static inline int raid6_d0(struct stripe_head *sh)
136 if (sh->ddf_layout)
137 /* ddf always start from first device */
138 return 0;
139 /* md starts just after Q block */
140 if (sh->qd_idx == sh->disks - 1)
141 return 0;
142 else
143 return sh->qd_idx + 1;
145 static inline int raid6_next_disk(int disk, int raid_disks)
147 disk++;
148 return (disk < raid_disks) ? disk : 0;
151 /* When walking through the disks in a raid5, starting at raid6_d0,
152 * We need to map each disk to a 'slot', where the data disks are slot
153 * 0 .. raid_disks-3, the parity disk is raid_disks-2 and the Q disk
154 * is raid_disks-1. This help does that mapping.
156 static int raid6_idx_to_slot(int idx, struct stripe_head *sh,
157 int *count, int syndrome_disks)
159 int slot = *count;
161 if (sh->ddf_layout)
162 (*count)++;
163 if (idx == sh->pd_idx)
164 return syndrome_disks;
165 if (idx == sh->qd_idx)
166 return syndrome_disks + 1;
167 if (!sh->ddf_layout)
168 (*count)++;
169 return slot;
172 static void return_io(struct bio *return_bi)
174 struct bio *bi = return_bi;
175 while (bi) {
177 return_bi = bi->bi_next;
178 bi->bi_next = NULL;
179 bi->bi_size = 0;
180 bio_endio(bi, 0);
181 bi = return_bi;
185 static void print_raid5_conf (raid5_conf_t *conf);
187 static int stripe_operations_active(struct stripe_head *sh)
189 return sh->check_state || sh->reconstruct_state ||
190 test_bit(STRIPE_BIOFILL_RUN, &sh->state) ||
191 test_bit(STRIPE_COMPUTE_RUN, &sh->state);
194 static void __release_stripe(raid5_conf_t *conf, struct stripe_head *sh)
196 if (atomic_dec_and_test(&sh->count)) {
197 BUG_ON(!list_empty(&sh->lru));
198 BUG_ON(atomic_read(&conf->active_stripes)==0);
199 if (test_bit(STRIPE_HANDLE, &sh->state)) {
200 if (test_bit(STRIPE_DELAYED, &sh->state)) {
201 list_add_tail(&sh->lru, &conf->delayed_list);
202 blk_plug_device(conf->mddev->queue);
203 } else if (test_bit(STRIPE_BIT_DELAY, &sh->state) &&
204 sh->bm_seq - conf->seq_write > 0) {
205 list_add_tail(&sh->lru, &conf->bitmap_list);
206 blk_plug_device(conf->mddev->queue);
207 } else {
208 clear_bit(STRIPE_BIT_DELAY, &sh->state);
209 list_add_tail(&sh->lru, &conf->handle_list);
211 md_wakeup_thread(conf->mddev->thread);
212 } else {
213 BUG_ON(stripe_operations_active(sh));
214 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
215 atomic_dec(&conf->preread_active_stripes);
216 if (atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD)
217 md_wakeup_thread(conf->mddev->thread);
219 atomic_dec(&conf->active_stripes);
220 if (!test_bit(STRIPE_EXPANDING, &sh->state)) {
221 list_add_tail(&sh->lru, &conf->inactive_list);
222 wake_up(&conf->wait_for_stripe);
223 if (conf->retry_read_aligned)
224 md_wakeup_thread(conf->mddev->thread);
230 static void release_stripe(struct stripe_head *sh)
232 raid5_conf_t *conf = sh->raid_conf;
233 unsigned long flags;
235 spin_lock_irqsave(&conf->device_lock, flags);
236 __release_stripe(conf, sh);
237 spin_unlock_irqrestore(&conf->device_lock, flags);
240 static inline void remove_hash(struct stripe_head *sh)
242 pr_debug("remove_hash(), stripe %llu\n",
243 (unsigned long long)sh->sector);
245 hlist_del_init(&sh->hash);
248 static inline void insert_hash(raid5_conf_t *conf, struct stripe_head *sh)
250 struct hlist_head *hp = stripe_hash(conf, sh->sector);
252 pr_debug("insert_hash(), stripe %llu\n",
253 (unsigned long long)sh->sector);
255 CHECK_DEVLOCK();
256 hlist_add_head(&sh->hash, hp);
260 /* find an idle stripe, make sure it is unhashed, and return it. */
261 static struct stripe_head *get_free_stripe(raid5_conf_t *conf)
263 struct stripe_head *sh = NULL;
264 struct list_head *first;
266 CHECK_DEVLOCK();
267 if (list_empty(&conf->inactive_list))
268 goto out;
269 first = conf->inactive_list.next;
270 sh = list_entry(first, struct stripe_head, lru);
271 list_del_init(first);
272 remove_hash(sh);
273 atomic_inc(&conf->active_stripes);
274 out:
275 return sh;
278 static void shrink_buffers(struct stripe_head *sh, int num)
280 struct page *p;
281 int i;
283 for (i=0; i<num ; i++) {
284 p = sh->dev[i].page;
285 if (!p)
286 continue;
287 sh->dev[i].page = NULL;
288 put_page(p);
292 static int grow_buffers(struct stripe_head *sh, int num)
294 int i;
296 for (i=0; i<num; i++) {
297 struct page *page;
299 if (!(page = alloc_page(GFP_KERNEL))) {
300 return 1;
302 sh->dev[i].page = page;
304 return 0;
307 static void raid5_build_block(struct stripe_head *sh, int i, int previous);
308 static void stripe_set_idx(sector_t stripe, raid5_conf_t *conf, int previous,
309 struct stripe_head *sh);
311 static void init_stripe(struct stripe_head *sh, sector_t sector, int previous)
313 raid5_conf_t *conf = sh->raid_conf;
314 int i;
316 BUG_ON(atomic_read(&sh->count) != 0);
317 BUG_ON(test_bit(STRIPE_HANDLE, &sh->state));
318 BUG_ON(stripe_operations_active(sh));
320 CHECK_DEVLOCK();
321 pr_debug("init_stripe called, stripe %llu\n",
322 (unsigned long long)sh->sector);
324 remove_hash(sh);
326 sh->generation = conf->generation - previous;
327 sh->disks = previous ? conf->previous_raid_disks : conf->raid_disks;
328 sh->sector = sector;
329 stripe_set_idx(sector, conf, previous, sh);
330 sh->state = 0;
333 for (i = sh->disks; i--; ) {
334 struct r5dev *dev = &sh->dev[i];
336 if (dev->toread || dev->read || dev->towrite || dev->written ||
337 test_bit(R5_LOCKED, &dev->flags)) {
338 printk(KERN_ERR "sector=%llx i=%d %p %p %p %p %d\n",
339 (unsigned long long)sh->sector, i, dev->toread,
340 dev->read, dev->towrite, dev->written,
341 test_bit(R5_LOCKED, &dev->flags));
342 BUG();
344 dev->flags = 0;
345 raid5_build_block(sh, i, previous);
347 insert_hash(conf, sh);
350 static struct stripe_head *__find_stripe(raid5_conf_t *conf, sector_t sector,
351 short generation)
353 struct stripe_head *sh;
354 struct hlist_node *hn;
356 CHECK_DEVLOCK();
357 pr_debug("__find_stripe, sector %llu\n", (unsigned long long)sector);
358 hlist_for_each_entry(sh, hn, stripe_hash(conf, sector), hash)
359 if (sh->sector == sector && sh->generation == generation)
360 return sh;
361 pr_debug("__stripe %llu not in cache\n", (unsigned long long)sector);
362 return NULL;
365 static void unplug_slaves(mddev_t *mddev);
366 static void raid5_unplug_device(struct request_queue *q);
368 static struct stripe_head *
369 get_active_stripe(raid5_conf_t *conf, sector_t sector,
370 int previous, int noblock, int noquiesce)
372 struct stripe_head *sh;
374 pr_debug("get_stripe, sector %llu\n", (unsigned long long)sector);
376 spin_lock_irq(&conf->device_lock);
378 do {
379 wait_event_lock_irq(conf->wait_for_stripe,
380 conf->quiesce == 0 || noquiesce,
381 conf->device_lock, /* nothing */);
382 sh = __find_stripe(conf, sector, conf->generation - previous);
383 if (!sh) {
384 if (!conf->inactive_blocked)
385 sh = get_free_stripe(conf);
386 if (noblock && sh == NULL)
387 break;
388 if (!sh) {
389 conf->inactive_blocked = 1;
390 wait_event_lock_irq(conf->wait_for_stripe,
391 !list_empty(&conf->inactive_list) &&
392 (atomic_read(&conf->active_stripes)
393 < (conf->max_nr_stripes *3/4)
394 || !conf->inactive_blocked),
395 conf->device_lock,
396 raid5_unplug_device(conf->mddev->queue)
398 conf->inactive_blocked = 0;
399 } else
400 init_stripe(sh, sector, previous);
401 } else {
402 if (atomic_read(&sh->count)) {
403 BUG_ON(!list_empty(&sh->lru)
404 && !test_bit(STRIPE_EXPANDING, &sh->state));
405 } else {
406 if (!test_bit(STRIPE_HANDLE, &sh->state))
407 atomic_inc(&conf->active_stripes);
408 if (list_empty(&sh->lru) &&
409 !test_bit(STRIPE_EXPANDING, &sh->state))
410 BUG();
411 list_del_init(&sh->lru);
414 } while (sh == NULL);
416 if (sh)
417 atomic_inc(&sh->count);
419 spin_unlock_irq(&conf->device_lock);
420 return sh;
423 static void
424 raid5_end_read_request(struct bio *bi, int error);
425 static void
426 raid5_end_write_request(struct bio *bi, int error);
428 static void ops_run_io(struct stripe_head *sh, struct stripe_head_state *s)
430 raid5_conf_t *conf = sh->raid_conf;
431 int i, disks = sh->disks;
433 might_sleep();
435 for (i = disks; i--; ) {
436 int rw;
437 struct bio *bi;
438 mdk_rdev_t *rdev;
439 if (test_and_clear_bit(R5_Wantwrite, &sh->dev[i].flags))
440 rw = WRITE;
441 else if (test_and_clear_bit(R5_Wantread, &sh->dev[i].flags))
442 rw = READ;
443 else
444 continue;
446 bi = &sh->dev[i].req;
448 bi->bi_rw = rw;
449 if (rw == WRITE)
450 bi->bi_end_io = raid5_end_write_request;
451 else
452 bi->bi_end_io = raid5_end_read_request;
454 rcu_read_lock();
455 rdev = rcu_dereference(conf->disks[i].rdev);
456 if (rdev && test_bit(Faulty, &rdev->flags))
457 rdev = NULL;
458 if (rdev)
459 atomic_inc(&rdev->nr_pending);
460 rcu_read_unlock();
462 if (rdev) {
463 if (s->syncing || s->expanding || s->expanded)
464 md_sync_acct(rdev->bdev, STRIPE_SECTORS);
466 set_bit(STRIPE_IO_STARTED, &sh->state);
468 bi->bi_bdev = rdev->bdev;
469 pr_debug("%s: for %llu schedule op %ld on disc %d\n",
470 __func__, (unsigned long long)sh->sector,
471 bi->bi_rw, i);
472 atomic_inc(&sh->count);
473 bi->bi_sector = sh->sector + rdev->data_offset;
474 bi->bi_flags = 1 << BIO_UPTODATE;
475 bi->bi_vcnt = 1;
476 bi->bi_max_vecs = 1;
477 bi->bi_idx = 0;
478 bi->bi_io_vec = &sh->dev[i].vec;
479 bi->bi_io_vec[0].bv_len = STRIPE_SIZE;
480 bi->bi_io_vec[0].bv_offset = 0;
481 bi->bi_size = STRIPE_SIZE;
482 bi->bi_next = NULL;
483 if (rw == WRITE &&
484 test_bit(R5_ReWrite, &sh->dev[i].flags))
485 atomic_add(STRIPE_SECTORS,
486 &rdev->corrected_errors);
487 generic_make_request(bi);
488 } else {
489 if (rw == WRITE)
490 set_bit(STRIPE_DEGRADED, &sh->state);
491 pr_debug("skip op %ld on disc %d for sector %llu\n",
492 bi->bi_rw, i, (unsigned long long)sh->sector);
493 clear_bit(R5_LOCKED, &sh->dev[i].flags);
494 set_bit(STRIPE_HANDLE, &sh->state);
499 static struct dma_async_tx_descriptor *
500 async_copy_data(int frombio, struct bio *bio, struct page *page,
501 sector_t sector, struct dma_async_tx_descriptor *tx)
503 struct bio_vec *bvl;
504 struct page *bio_page;
505 int i;
506 int page_offset;
507 struct async_submit_ctl submit;
508 enum async_tx_flags flags = 0;
510 if (bio->bi_sector >= sector)
511 page_offset = (signed)(bio->bi_sector - sector) * 512;
512 else
513 page_offset = (signed)(sector - bio->bi_sector) * -512;
515 if (frombio)
516 flags |= ASYNC_TX_FENCE;
517 init_async_submit(&submit, flags, tx, NULL, NULL, NULL);
519 bio_for_each_segment(bvl, bio, i) {
520 int len = bio_iovec_idx(bio, i)->bv_len;
521 int clen;
522 int b_offset = 0;
524 if (page_offset < 0) {
525 b_offset = -page_offset;
526 page_offset += b_offset;
527 len -= b_offset;
530 if (len > 0 && page_offset + len > STRIPE_SIZE)
531 clen = STRIPE_SIZE - page_offset;
532 else
533 clen = len;
535 if (clen > 0) {
536 b_offset += bio_iovec_idx(bio, i)->bv_offset;
537 bio_page = bio_iovec_idx(bio, i)->bv_page;
538 if (frombio)
539 tx = async_memcpy(page, bio_page, page_offset,
540 b_offset, clen, &submit);
541 else
542 tx = async_memcpy(bio_page, page, b_offset,
543 page_offset, clen, &submit);
545 /* chain the operations */
546 submit.depend_tx = tx;
548 if (clen < len) /* hit end of page */
549 break;
550 page_offset += len;
553 return tx;
556 static void ops_complete_biofill(void *stripe_head_ref)
558 struct stripe_head *sh = stripe_head_ref;
559 struct bio *return_bi = NULL;
560 raid5_conf_t *conf = sh->raid_conf;
561 int i;
563 pr_debug("%s: stripe %llu\n", __func__,
564 (unsigned long long)sh->sector);
566 /* clear completed biofills */
567 spin_lock_irq(&conf->device_lock);
568 for (i = sh->disks; i--; ) {
569 struct r5dev *dev = &sh->dev[i];
571 /* acknowledge completion of a biofill operation */
572 /* and check if we need to reply to a read request,
573 * new R5_Wantfill requests are held off until
574 * !STRIPE_BIOFILL_RUN
576 if (test_and_clear_bit(R5_Wantfill, &dev->flags)) {
577 struct bio *rbi, *rbi2;
579 BUG_ON(!dev->read);
580 rbi = dev->read;
581 dev->read = NULL;
582 while (rbi && rbi->bi_sector <
583 dev->sector + STRIPE_SECTORS) {
584 rbi2 = r5_next_bio(rbi, dev->sector);
585 if (!raid5_dec_bi_phys_segments(rbi)) {
586 rbi->bi_next = return_bi;
587 return_bi = rbi;
589 rbi = rbi2;
593 spin_unlock_irq(&conf->device_lock);
594 clear_bit(STRIPE_BIOFILL_RUN, &sh->state);
596 return_io(return_bi);
598 set_bit(STRIPE_HANDLE, &sh->state);
599 release_stripe(sh);
602 static void ops_run_biofill(struct stripe_head *sh)
604 struct dma_async_tx_descriptor *tx = NULL;
605 raid5_conf_t *conf = sh->raid_conf;
606 struct async_submit_ctl submit;
607 int i;
609 pr_debug("%s: stripe %llu\n", __func__,
610 (unsigned long long)sh->sector);
612 for (i = sh->disks; i--; ) {
613 struct r5dev *dev = &sh->dev[i];
614 if (test_bit(R5_Wantfill, &dev->flags)) {
615 struct bio *rbi;
616 spin_lock_irq(&conf->device_lock);
617 dev->read = rbi = dev->toread;
618 dev->toread = NULL;
619 spin_unlock_irq(&conf->device_lock);
620 while (rbi && rbi->bi_sector <
621 dev->sector + STRIPE_SECTORS) {
622 tx = async_copy_data(0, rbi, dev->page,
623 dev->sector, tx);
624 rbi = r5_next_bio(rbi, dev->sector);
629 atomic_inc(&sh->count);
630 init_async_submit(&submit, ASYNC_TX_ACK, tx, ops_complete_biofill, sh, NULL);
631 async_trigger_callback(&submit);
634 static void mark_target_uptodate(struct stripe_head *sh, int target)
636 struct r5dev *tgt;
638 if (target < 0)
639 return;
641 tgt = &sh->dev[target];
642 set_bit(R5_UPTODATE, &tgt->flags);
643 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
644 clear_bit(R5_Wantcompute, &tgt->flags);
647 static void ops_complete_compute(void *stripe_head_ref)
649 struct stripe_head *sh = stripe_head_ref;
651 pr_debug("%s: stripe %llu\n", __func__,
652 (unsigned long long)sh->sector);
654 /* mark the computed target(s) as uptodate */
655 mark_target_uptodate(sh, sh->ops.target);
656 mark_target_uptodate(sh, sh->ops.target2);
658 clear_bit(STRIPE_COMPUTE_RUN, &sh->state);
659 if (sh->check_state == check_state_compute_run)
660 sh->check_state = check_state_compute_result;
661 set_bit(STRIPE_HANDLE, &sh->state);
662 release_stripe(sh);
665 /* return a pointer to the address conversion region of the scribble buffer */
666 static addr_conv_t *to_addr_conv(struct stripe_head *sh,
667 struct raid5_percpu *percpu)
669 return percpu->scribble + sizeof(struct page *) * (sh->disks + 2);
672 static struct dma_async_tx_descriptor *
673 ops_run_compute5(struct stripe_head *sh, struct raid5_percpu *percpu)
675 int disks = sh->disks;
676 struct page **xor_srcs = percpu->scribble;
677 int target = sh->ops.target;
678 struct r5dev *tgt = &sh->dev[target];
679 struct page *xor_dest = tgt->page;
680 int count = 0;
681 struct dma_async_tx_descriptor *tx;
682 struct async_submit_ctl submit;
683 int i;
685 pr_debug("%s: stripe %llu block: %d\n",
686 __func__, (unsigned long long)sh->sector, target);
687 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
689 for (i = disks; i--; )
690 if (i != target)
691 xor_srcs[count++] = sh->dev[i].page;
693 atomic_inc(&sh->count);
695 init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST, NULL,
696 ops_complete_compute, sh, to_addr_conv(sh, percpu));
697 if (unlikely(count == 1))
698 tx = async_memcpy(xor_dest, xor_srcs[0], 0, 0, STRIPE_SIZE, &submit);
699 else
700 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
702 return tx;
705 /* set_syndrome_sources - populate source buffers for gen_syndrome
706 * @srcs - (struct page *) array of size sh->disks
707 * @sh - stripe_head to parse
709 * Populates srcs in proper layout order for the stripe and returns the
710 * 'count' of sources to be used in a call to async_gen_syndrome. The P
711 * destination buffer is recorded in srcs[count] and the Q destination
712 * is recorded in srcs[count+1]].
714 static int set_syndrome_sources(struct page **srcs, struct stripe_head *sh)
716 int disks = sh->disks;
717 int syndrome_disks = sh->ddf_layout ? disks : (disks - 2);
718 int d0_idx = raid6_d0(sh);
719 int count;
720 int i;
722 for (i = 0; i < disks; i++)
723 srcs[i] = NULL;
725 count = 0;
726 i = d0_idx;
727 do {
728 int slot = raid6_idx_to_slot(i, sh, &count, syndrome_disks);
730 srcs[slot] = sh->dev[i].page;
731 i = raid6_next_disk(i, disks);
732 } while (i != d0_idx);
734 return syndrome_disks;
737 static struct dma_async_tx_descriptor *
738 ops_run_compute6_1(struct stripe_head *sh, struct raid5_percpu *percpu)
740 int disks = sh->disks;
741 struct page **blocks = percpu->scribble;
742 int target;
743 int qd_idx = sh->qd_idx;
744 struct dma_async_tx_descriptor *tx;
745 struct async_submit_ctl submit;
746 struct r5dev *tgt;
747 struct page *dest;
748 int i;
749 int count;
751 if (sh->ops.target < 0)
752 target = sh->ops.target2;
753 else if (sh->ops.target2 < 0)
754 target = sh->ops.target;
755 else
756 /* we should only have one valid target */
757 BUG();
758 BUG_ON(target < 0);
759 pr_debug("%s: stripe %llu block: %d\n",
760 __func__, (unsigned long long)sh->sector, target);
762 tgt = &sh->dev[target];
763 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
764 dest = tgt->page;
766 atomic_inc(&sh->count);
768 if (target == qd_idx) {
769 count = set_syndrome_sources(blocks, sh);
770 blocks[count] = NULL; /* regenerating p is not necessary */
771 BUG_ON(blocks[count+1] != dest); /* q should already be set */
772 init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
773 ops_complete_compute, sh,
774 to_addr_conv(sh, percpu));
775 tx = async_gen_syndrome(blocks, 0, count+2, STRIPE_SIZE, &submit);
776 } else {
777 /* Compute any data- or p-drive using XOR */
778 count = 0;
779 for (i = disks; i-- ; ) {
780 if (i == target || i == qd_idx)
781 continue;
782 blocks[count++] = sh->dev[i].page;
785 init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST,
786 NULL, ops_complete_compute, sh,
787 to_addr_conv(sh, percpu));
788 tx = async_xor(dest, blocks, 0, count, STRIPE_SIZE, &submit);
791 return tx;
794 static struct dma_async_tx_descriptor *
795 ops_run_compute6_2(struct stripe_head *sh, struct raid5_percpu *percpu)
797 int i, count, disks = sh->disks;
798 int syndrome_disks = sh->ddf_layout ? disks : disks-2;
799 int d0_idx = raid6_d0(sh);
800 int faila = -1, failb = -1;
801 int target = sh->ops.target;
802 int target2 = sh->ops.target2;
803 struct r5dev *tgt = &sh->dev[target];
804 struct r5dev *tgt2 = &sh->dev[target2];
805 struct dma_async_tx_descriptor *tx;
806 struct page **blocks = percpu->scribble;
807 struct async_submit_ctl submit;
809 pr_debug("%s: stripe %llu block1: %d block2: %d\n",
810 __func__, (unsigned long long)sh->sector, target, target2);
811 BUG_ON(target < 0 || target2 < 0);
812 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
813 BUG_ON(!test_bit(R5_Wantcompute, &tgt2->flags));
815 /* we need to open-code set_syndrome_sources to handle the
816 * slot number conversion for 'faila' and 'failb'
818 for (i = 0; i < disks ; i++)
819 blocks[i] = NULL;
820 count = 0;
821 i = d0_idx;
822 do {
823 int slot = raid6_idx_to_slot(i, sh, &count, syndrome_disks);
825 blocks[slot] = sh->dev[i].page;
827 if (i == target)
828 faila = slot;
829 if (i == target2)
830 failb = slot;
831 i = raid6_next_disk(i, disks);
832 } while (i != d0_idx);
834 BUG_ON(faila == failb);
835 if (failb < faila)
836 swap(faila, failb);
837 pr_debug("%s: stripe: %llu faila: %d failb: %d\n",
838 __func__, (unsigned long long)sh->sector, faila, failb);
840 atomic_inc(&sh->count);
842 if (failb == syndrome_disks+1) {
843 /* Q disk is one of the missing disks */
844 if (faila == syndrome_disks) {
845 /* Missing P+Q, just recompute */
846 init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
847 ops_complete_compute, sh,
848 to_addr_conv(sh, percpu));
849 return async_gen_syndrome(blocks, 0, syndrome_disks+2,
850 STRIPE_SIZE, &submit);
851 } else {
852 struct page *dest;
853 int data_target;
854 int qd_idx = sh->qd_idx;
856 /* Missing D+Q: recompute D from P, then recompute Q */
857 if (target == qd_idx)
858 data_target = target2;
859 else
860 data_target = target;
862 count = 0;
863 for (i = disks; i-- ; ) {
864 if (i == data_target || i == qd_idx)
865 continue;
866 blocks[count++] = sh->dev[i].page;
868 dest = sh->dev[data_target].page;
869 init_async_submit(&submit,
870 ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST,
871 NULL, NULL, NULL,
872 to_addr_conv(sh, percpu));
873 tx = async_xor(dest, blocks, 0, count, STRIPE_SIZE,
874 &submit);
876 count = set_syndrome_sources(blocks, sh);
877 init_async_submit(&submit, ASYNC_TX_FENCE, tx,
878 ops_complete_compute, sh,
879 to_addr_conv(sh, percpu));
880 return async_gen_syndrome(blocks, 0, count+2,
881 STRIPE_SIZE, &submit);
883 } else {
884 init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
885 ops_complete_compute, sh,
886 to_addr_conv(sh, percpu));
887 if (failb == syndrome_disks) {
888 /* We're missing D+P. */
889 return async_raid6_datap_recov(syndrome_disks+2,
890 STRIPE_SIZE, faila,
891 blocks, &submit);
892 } else {
893 /* We're missing D+D. */
894 return async_raid6_2data_recov(syndrome_disks+2,
895 STRIPE_SIZE, faila, failb,
896 blocks, &submit);
902 static void ops_complete_prexor(void *stripe_head_ref)
904 struct stripe_head *sh = stripe_head_ref;
906 pr_debug("%s: stripe %llu\n", __func__,
907 (unsigned long long)sh->sector);
910 static struct dma_async_tx_descriptor *
911 ops_run_prexor(struct stripe_head *sh, struct raid5_percpu *percpu,
912 struct dma_async_tx_descriptor *tx)
914 int disks = sh->disks;
915 struct page **xor_srcs = percpu->scribble;
916 int count = 0, pd_idx = sh->pd_idx, i;
917 struct async_submit_ctl submit;
919 /* existing parity data subtracted */
920 struct page *xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
922 pr_debug("%s: stripe %llu\n", __func__,
923 (unsigned long long)sh->sector);
925 for (i = disks; i--; ) {
926 struct r5dev *dev = &sh->dev[i];
927 /* Only process blocks that are known to be uptodate */
928 if (test_bit(R5_Wantdrain, &dev->flags))
929 xor_srcs[count++] = dev->page;
932 init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_DROP_DST, tx,
933 ops_complete_prexor, sh, to_addr_conv(sh, percpu));
934 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
936 return tx;
939 static struct dma_async_tx_descriptor *
940 ops_run_biodrain(struct stripe_head *sh, struct dma_async_tx_descriptor *tx)
942 int disks = sh->disks;
943 int i;
945 pr_debug("%s: stripe %llu\n", __func__,
946 (unsigned long long)sh->sector);
948 for (i = disks; i--; ) {
949 struct r5dev *dev = &sh->dev[i];
950 struct bio *chosen;
952 if (test_and_clear_bit(R5_Wantdrain, &dev->flags)) {
953 struct bio *wbi;
955 spin_lock(&sh->lock);
956 chosen = dev->towrite;
957 dev->towrite = NULL;
958 BUG_ON(dev->written);
959 wbi = dev->written = chosen;
960 spin_unlock(&sh->lock);
962 while (wbi && wbi->bi_sector <
963 dev->sector + STRIPE_SECTORS) {
964 tx = async_copy_data(1, wbi, dev->page,
965 dev->sector, tx);
966 wbi = r5_next_bio(wbi, dev->sector);
971 return tx;
974 static void ops_complete_reconstruct(void *stripe_head_ref)
976 struct stripe_head *sh = stripe_head_ref;
977 int disks = sh->disks;
978 int pd_idx = sh->pd_idx;
979 int qd_idx = sh->qd_idx;
980 int i;
982 pr_debug("%s: stripe %llu\n", __func__,
983 (unsigned long long)sh->sector);
985 for (i = disks; i--; ) {
986 struct r5dev *dev = &sh->dev[i];
988 if (dev->written || i == pd_idx || i == qd_idx)
989 set_bit(R5_UPTODATE, &dev->flags);
992 if (sh->reconstruct_state == reconstruct_state_drain_run)
993 sh->reconstruct_state = reconstruct_state_drain_result;
994 else if (sh->reconstruct_state == reconstruct_state_prexor_drain_run)
995 sh->reconstruct_state = reconstruct_state_prexor_drain_result;
996 else {
997 BUG_ON(sh->reconstruct_state != reconstruct_state_run);
998 sh->reconstruct_state = reconstruct_state_result;
1001 set_bit(STRIPE_HANDLE, &sh->state);
1002 release_stripe(sh);
1005 static void
1006 ops_run_reconstruct5(struct stripe_head *sh, struct raid5_percpu *percpu,
1007 struct dma_async_tx_descriptor *tx)
1009 int disks = sh->disks;
1010 struct page **xor_srcs = percpu->scribble;
1011 struct async_submit_ctl submit;
1012 int count = 0, pd_idx = sh->pd_idx, i;
1013 struct page *xor_dest;
1014 int prexor = 0;
1015 unsigned long flags;
1017 pr_debug("%s: stripe %llu\n", __func__,
1018 (unsigned long long)sh->sector);
1020 /* check if prexor is active which means only process blocks
1021 * that are part of a read-modify-write (written)
1023 if (sh->reconstruct_state == reconstruct_state_prexor_drain_run) {
1024 prexor = 1;
1025 xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
1026 for (i = disks; i--; ) {
1027 struct r5dev *dev = &sh->dev[i];
1028 if (dev->written)
1029 xor_srcs[count++] = dev->page;
1031 } else {
1032 xor_dest = sh->dev[pd_idx].page;
1033 for (i = disks; i--; ) {
1034 struct r5dev *dev = &sh->dev[i];
1035 if (i != pd_idx)
1036 xor_srcs[count++] = dev->page;
1040 /* 1/ if we prexor'd then the dest is reused as a source
1041 * 2/ if we did not prexor then we are redoing the parity
1042 * set ASYNC_TX_XOR_DROP_DST and ASYNC_TX_XOR_ZERO_DST
1043 * for the synchronous xor case
1045 flags = ASYNC_TX_ACK |
1046 (prexor ? ASYNC_TX_XOR_DROP_DST : ASYNC_TX_XOR_ZERO_DST);
1048 atomic_inc(&sh->count);
1050 init_async_submit(&submit, flags, tx, ops_complete_reconstruct, sh,
1051 to_addr_conv(sh, percpu));
1052 if (unlikely(count == 1))
1053 tx = async_memcpy(xor_dest, xor_srcs[0], 0, 0, STRIPE_SIZE, &submit);
1054 else
1055 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
1058 static void
1059 ops_run_reconstruct6(struct stripe_head *sh, struct raid5_percpu *percpu,
1060 struct dma_async_tx_descriptor *tx)
1062 struct async_submit_ctl submit;
1063 struct page **blocks = percpu->scribble;
1064 int count;
1066 pr_debug("%s: stripe %llu\n", __func__, (unsigned long long)sh->sector);
1068 count = set_syndrome_sources(blocks, sh);
1070 atomic_inc(&sh->count);
1072 init_async_submit(&submit, ASYNC_TX_ACK, tx, ops_complete_reconstruct,
1073 sh, to_addr_conv(sh, percpu));
1074 async_gen_syndrome(blocks, 0, count+2, STRIPE_SIZE, &submit);
1077 static void ops_complete_check(void *stripe_head_ref)
1079 struct stripe_head *sh = stripe_head_ref;
1081 pr_debug("%s: stripe %llu\n", __func__,
1082 (unsigned long long)sh->sector);
1084 sh->check_state = check_state_check_result;
1085 set_bit(STRIPE_HANDLE, &sh->state);
1086 release_stripe(sh);
1089 static void ops_run_check_p(struct stripe_head *sh, struct raid5_percpu *percpu)
1091 int disks = sh->disks;
1092 int pd_idx = sh->pd_idx;
1093 int qd_idx = sh->qd_idx;
1094 struct page *xor_dest;
1095 struct page **xor_srcs = percpu->scribble;
1096 struct dma_async_tx_descriptor *tx;
1097 struct async_submit_ctl submit;
1098 int count;
1099 int i;
1101 pr_debug("%s: stripe %llu\n", __func__,
1102 (unsigned long long)sh->sector);
1104 count = 0;
1105 xor_dest = sh->dev[pd_idx].page;
1106 xor_srcs[count++] = xor_dest;
1107 for (i = disks; i--; ) {
1108 if (i == pd_idx || i == qd_idx)
1109 continue;
1110 xor_srcs[count++] = sh->dev[i].page;
1113 init_async_submit(&submit, 0, NULL, NULL, NULL,
1114 to_addr_conv(sh, percpu));
1115 tx = async_xor_val(xor_dest, xor_srcs, 0, count, STRIPE_SIZE,
1116 &sh->ops.zero_sum_result, &submit);
1118 atomic_inc(&sh->count);
1119 init_async_submit(&submit, ASYNC_TX_ACK, tx, ops_complete_check, sh, NULL);
1120 tx = async_trigger_callback(&submit);
1123 static void ops_run_check_pq(struct stripe_head *sh, struct raid5_percpu *percpu, int checkp)
1125 struct page **srcs = percpu->scribble;
1126 struct async_submit_ctl submit;
1127 int count;
1129 pr_debug("%s: stripe %llu checkp: %d\n", __func__,
1130 (unsigned long long)sh->sector, checkp);
1132 count = set_syndrome_sources(srcs, sh);
1133 if (!checkp)
1134 srcs[count] = NULL;
1136 atomic_inc(&sh->count);
1137 init_async_submit(&submit, ASYNC_TX_ACK, NULL, ops_complete_check,
1138 sh, to_addr_conv(sh, percpu));
1139 async_syndrome_val(srcs, 0, count+2, STRIPE_SIZE,
1140 &sh->ops.zero_sum_result, percpu->spare_page, &submit);
1143 static void __raid_run_ops(struct stripe_head *sh, unsigned long ops_request)
1145 int overlap_clear = 0, i, disks = sh->disks;
1146 struct dma_async_tx_descriptor *tx = NULL;
1147 raid5_conf_t *conf = sh->raid_conf;
1148 int level = conf->level;
1149 struct raid5_percpu *percpu;
1150 unsigned long cpu;
1152 cpu = get_cpu();
1153 percpu = per_cpu_ptr(conf->percpu, cpu);
1154 if (test_bit(STRIPE_OP_BIOFILL, &ops_request)) {
1155 ops_run_biofill(sh);
1156 overlap_clear++;
1159 if (test_bit(STRIPE_OP_COMPUTE_BLK, &ops_request)) {
1160 if (level < 6)
1161 tx = ops_run_compute5(sh, percpu);
1162 else {
1163 if (sh->ops.target2 < 0 || sh->ops.target < 0)
1164 tx = ops_run_compute6_1(sh, percpu);
1165 else
1166 tx = ops_run_compute6_2(sh, percpu);
1168 /* terminate the chain if reconstruct is not set to be run */
1169 if (tx && !test_bit(STRIPE_OP_RECONSTRUCT, &ops_request))
1170 async_tx_ack(tx);
1173 if (test_bit(STRIPE_OP_PREXOR, &ops_request))
1174 tx = ops_run_prexor(sh, percpu, tx);
1176 if (test_bit(STRIPE_OP_BIODRAIN, &ops_request)) {
1177 tx = ops_run_biodrain(sh, tx);
1178 overlap_clear++;
1181 if (test_bit(STRIPE_OP_RECONSTRUCT, &ops_request)) {
1182 if (level < 6)
1183 ops_run_reconstruct5(sh, percpu, tx);
1184 else
1185 ops_run_reconstruct6(sh, percpu, tx);
1188 if (test_bit(STRIPE_OP_CHECK, &ops_request)) {
1189 if (sh->check_state == check_state_run)
1190 ops_run_check_p(sh, percpu);
1191 else if (sh->check_state == check_state_run_q)
1192 ops_run_check_pq(sh, percpu, 0);
1193 else if (sh->check_state == check_state_run_pq)
1194 ops_run_check_pq(sh, percpu, 1);
1195 else
1196 BUG();
1199 if (overlap_clear)
1200 for (i = disks; i--; ) {
1201 struct r5dev *dev = &sh->dev[i];
1202 if (test_and_clear_bit(R5_Overlap, &dev->flags))
1203 wake_up(&sh->raid_conf->wait_for_overlap);
1205 put_cpu();
1208 #ifdef CONFIG_MULTICORE_RAID456
1209 static void async_run_ops(void *param, async_cookie_t cookie)
1211 struct stripe_head *sh = param;
1212 unsigned long ops_request = sh->ops.request;
1214 clear_bit_unlock(STRIPE_OPS_REQ_PENDING, &sh->state);
1215 wake_up(&sh->ops.wait_for_ops);
1217 __raid_run_ops(sh, ops_request);
1218 release_stripe(sh);
1221 static void raid_run_ops(struct stripe_head *sh, unsigned long ops_request)
1223 /* since handle_stripe can be called outside of raid5d context
1224 * we need to ensure sh->ops.request is de-staged before another
1225 * request arrives
1227 wait_event(sh->ops.wait_for_ops,
1228 !test_and_set_bit_lock(STRIPE_OPS_REQ_PENDING, &sh->state));
1229 sh->ops.request = ops_request;
1231 atomic_inc(&sh->count);
1232 async_schedule(async_run_ops, sh);
1234 #else
1235 #define raid_run_ops __raid_run_ops
1236 #endif
1238 static int grow_one_stripe(raid5_conf_t *conf)
1240 struct stripe_head *sh;
1241 int disks = max(conf->raid_disks, conf->previous_raid_disks);
1242 sh = kmem_cache_alloc(conf->slab_cache, GFP_KERNEL);
1243 if (!sh)
1244 return 0;
1245 memset(sh, 0, sizeof(*sh) + (disks-1)*sizeof(struct r5dev));
1246 sh->raid_conf = conf;
1247 spin_lock_init(&sh->lock);
1248 #ifdef CONFIG_MULTICORE_RAID456
1249 init_waitqueue_head(&sh->ops.wait_for_ops);
1250 #endif
1252 if (grow_buffers(sh, disks)) {
1253 shrink_buffers(sh, disks);
1254 kmem_cache_free(conf->slab_cache, sh);
1255 return 0;
1257 /* we just created an active stripe so... */
1258 atomic_set(&sh->count, 1);
1259 atomic_inc(&conf->active_stripes);
1260 INIT_LIST_HEAD(&sh->lru);
1261 release_stripe(sh);
1262 return 1;
1265 static int grow_stripes(raid5_conf_t *conf, int num)
1267 struct kmem_cache *sc;
1268 int devs = max(conf->raid_disks, conf->previous_raid_disks);
1270 sprintf(conf->cache_name[0],
1271 "raid%d-%s", conf->level, mdname(conf->mddev));
1272 sprintf(conf->cache_name[1],
1273 "raid%d-%s-alt", conf->level, mdname(conf->mddev));
1274 conf->active_name = 0;
1275 sc = kmem_cache_create(conf->cache_name[conf->active_name],
1276 sizeof(struct stripe_head)+(devs-1)*sizeof(struct r5dev),
1277 0, 0, NULL);
1278 if (!sc)
1279 return 1;
1280 conf->slab_cache = sc;
1281 conf->pool_size = devs;
1282 while (num--)
1283 if (!grow_one_stripe(conf))
1284 return 1;
1285 return 0;
1289 * scribble_len - return the required size of the scribble region
1290 * @num - total number of disks in the array
1292 * The size must be enough to contain:
1293 * 1/ a struct page pointer for each device in the array +2
1294 * 2/ room to convert each entry in (1) to its corresponding dma
1295 * (dma_map_page()) or page (page_address()) address.
1297 * Note: the +2 is for the destination buffers of the ddf/raid6 case where we
1298 * calculate over all devices (not just the data blocks), using zeros in place
1299 * of the P and Q blocks.
1301 static size_t scribble_len(int num)
1303 size_t len;
1305 len = sizeof(struct page *) * (num+2) + sizeof(addr_conv_t) * (num+2);
1307 return len;
1310 static int resize_stripes(raid5_conf_t *conf, int newsize)
1312 /* Make all the stripes able to hold 'newsize' devices.
1313 * New slots in each stripe get 'page' set to a new page.
1315 * This happens in stages:
1316 * 1/ create a new kmem_cache and allocate the required number of
1317 * stripe_heads.
1318 * 2/ gather all the old stripe_heads and tranfer the pages across
1319 * to the new stripe_heads. This will have the side effect of
1320 * freezing the array as once all stripe_heads have been collected,
1321 * no IO will be possible. Old stripe heads are freed once their
1322 * pages have been transferred over, and the old kmem_cache is
1323 * freed when all stripes are done.
1324 * 3/ reallocate conf->disks to be suitable bigger. If this fails,
1325 * we simple return a failre status - no need to clean anything up.
1326 * 4/ allocate new pages for the new slots in the new stripe_heads.
1327 * If this fails, we don't bother trying the shrink the
1328 * stripe_heads down again, we just leave them as they are.
1329 * As each stripe_head is processed the new one is released into
1330 * active service.
1332 * Once step2 is started, we cannot afford to wait for a write,
1333 * so we use GFP_NOIO allocations.
1335 struct stripe_head *osh, *nsh;
1336 LIST_HEAD(newstripes);
1337 struct disk_info *ndisks;
1338 unsigned long cpu;
1339 int err;
1340 struct kmem_cache *sc;
1341 int i;
1343 if (newsize <= conf->pool_size)
1344 return 0; /* never bother to shrink */
1346 err = md_allow_write(conf->mddev);
1347 if (err)
1348 return err;
1350 /* Step 1 */
1351 sc = kmem_cache_create(conf->cache_name[1-conf->active_name],
1352 sizeof(struct stripe_head)+(newsize-1)*sizeof(struct r5dev),
1353 0, 0, NULL);
1354 if (!sc)
1355 return -ENOMEM;
1357 for (i = conf->max_nr_stripes; i; i--) {
1358 nsh = kmem_cache_alloc(sc, GFP_KERNEL);
1359 if (!nsh)
1360 break;
1362 memset(nsh, 0, sizeof(*nsh) + (newsize-1)*sizeof(struct r5dev));
1364 nsh->raid_conf = conf;
1365 spin_lock_init(&nsh->lock);
1366 #ifdef CONFIG_MULTICORE_RAID456
1367 init_waitqueue_head(&nsh->ops.wait_for_ops);
1368 #endif
1370 list_add(&nsh->lru, &newstripes);
1372 if (i) {
1373 /* didn't get enough, give up */
1374 while (!list_empty(&newstripes)) {
1375 nsh = list_entry(newstripes.next, struct stripe_head, lru);
1376 list_del(&nsh->lru);
1377 kmem_cache_free(sc, nsh);
1379 kmem_cache_destroy(sc);
1380 return -ENOMEM;
1382 /* Step 2 - Must use GFP_NOIO now.
1383 * OK, we have enough stripes, start collecting inactive
1384 * stripes and copying them over
1386 list_for_each_entry(nsh, &newstripes, lru) {
1387 spin_lock_irq(&conf->device_lock);
1388 wait_event_lock_irq(conf->wait_for_stripe,
1389 !list_empty(&conf->inactive_list),
1390 conf->device_lock,
1391 unplug_slaves(conf->mddev)
1393 osh = get_free_stripe(conf);
1394 spin_unlock_irq(&conf->device_lock);
1395 atomic_set(&nsh->count, 1);
1396 for(i=0; i<conf->pool_size; i++)
1397 nsh->dev[i].page = osh->dev[i].page;
1398 for( ; i<newsize; i++)
1399 nsh->dev[i].page = NULL;
1400 kmem_cache_free(conf->slab_cache, osh);
1402 kmem_cache_destroy(conf->slab_cache);
1404 /* Step 3.
1405 * At this point, we are holding all the stripes so the array
1406 * is completely stalled, so now is a good time to resize
1407 * conf->disks and the scribble region
1409 ndisks = kzalloc(newsize * sizeof(struct disk_info), GFP_NOIO);
1410 if (ndisks) {
1411 for (i=0; i<conf->raid_disks; i++)
1412 ndisks[i] = conf->disks[i];
1413 kfree(conf->disks);
1414 conf->disks = ndisks;
1415 } else
1416 err = -ENOMEM;
1418 get_online_cpus();
1419 conf->scribble_len = scribble_len(newsize);
1420 for_each_present_cpu(cpu) {
1421 struct raid5_percpu *percpu;
1422 void *scribble;
1424 percpu = per_cpu_ptr(conf->percpu, cpu);
1425 scribble = kmalloc(conf->scribble_len, GFP_NOIO);
1427 if (scribble) {
1428 kfree(percpu->scribble);
1429 percpu->scribble = scribble;
1430 } else {
1431 err = -ENOMEM;
1432 break;
1435 put_online_cpus();
1437 /* Step 4, return new stripes to service */
1438 while(!list_empty(&newstripes)) {
1439 nsh = list_entry(newstripes.next, struct stripe_head, lru);
1440 list_del_init(&nsh->lru);
1442 for (i=conf->raid_disks; i < newsize; i++)
1443 if (nsh->dev[i].page == NULL) {
1444 struct page *p = alloc_page(GFP_NOIO);
1445 nsh->dev[i].page = p;
1446 if (!p)
1447 err = -ENOMEM;
1449 release_stripe(nsh);
1451 /* critical section pass, GFP_NOIO no longer needed */
1453 conf->slab_cache = sc;
1454 conf->active_name = 1-conf->active_name;
1455 conf->pool_size = newsize;
1456 return err;
1459 static int drop_one_stripe(raid5_conf_t *conf)
1461 struct stripe_head *sh;
1463 spin_lock_irq(&conf->device_lock);
1464 sh = get_free_stripe(conf);
1465 spin_unlock_irq(&conf->device_lock);
1466 if (!sh)
1467 return 0;
1468 BUG_ON(atomic_read(&sh->count));
1469 shrink_buffers(sh, conf->pool_size);
1470 kmem_cache_free(conf->slab_cache, sh);
1471 atomic_dec(&conf->active_stripes);
1472 return 1;
1475 static void shrink_stripes(raid5_conf_t *conf)
1477 while (drop_one_stripe(conf))
1480 if (conf->slab_cache)
1481 kmem_cache_destroy(conf->slab_cache);
1482 conf->slab_cache = NULL;
1485 static void raid5_end_read_request(struct bio * bi, int error)
1487 struct stripe_head *sh = bi->bi_private;
1488 raid5_conf_t *conf = sh->raid_conf;
1489 int disks = sh->disks, i;
1490 int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
1491 char b[BDEVNAME_SIZE];
1492 mdk_rdev_t *rdev;
1495 for (i=0 ; i<disks; i++)
1496 if (bi == &sh->dev[i].req)
1497 break;
1499 pr_debug("end_read_request %llu/%d, count: %d, uptodate %d.\n",
1500 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
1501 uptodate);
1502 if (i == disks) {
1503 BUG();
1504 return;
1507 if (uptodate) {
1508 set_bit(R5_UPTODATE, &sh->dev[i].flags);
1509 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
1510 rdev = conf->disks[i].rdev;
1511 printk_rl(KERN_INFO "raid5:%s: read error corrected"
1512 " (%lu sectors at %llu on %s)\n",
1513 mdname(conf->mddev), STRIPE_SECTORS,
1514 (unsigned long long)(sh->sector
1515 + rdev->data_offset),
1516 bdevname(rdev->bdev, b));
1517 clear_bit(R5_ReadError, &sh->dev[i].flags);
1518 clear_bit(R5_ReWrite, &sh->dev[i].flags);
1520 if (atomic_read(&conf->disks[i].rdev->read_errors))
1521 atomic_set(&conf->disks[i].rdev->read_errors, 0);
1522 } else {
1523 const char *bdn = bdevname(conf->disks[i].rdev->bdev, b);
1524 int retry = 0;
1525 rdev = conf->disks[i].rdev;
1527 clear_bit(R5_UPTODATE, &sh->dev[i].flags);
1528 atomic_inc(&rdev->read_errors);
1529 if (conf->mddev->degraded)
1530 printk_rl(KERN_WARNING
1531 "raid5:%s: read error not correctable "
1532 "(sector %llu on %s).\n",
1533 mdname(conf->mddev),
1534 (unsigned long long)(sh->sector
1535 + rdev->data_offset),
1536 bdn);
1537 else if (test_bit(R5_ReWrite, &sh->dev[i].flags))
1538 /* Oh, no!!! */
1539 printk_rl(KERN_WARNING
1540 "raid5:%s: read error NOT corrected!! "
1541 "(sector %llu on %s).\n",
1542 mdname(conf->mddev),
1543 (unsigned long long)(sh->sector
1544 + rdev->data_offset),
1545 bdn);
1546 else if (atomic_read(&rdev->read_errors)
1547 > conf->max_nr_stripes)
1548 printk(KERN_WARNING
1549 "raid5:%s: Too many read errors, failing device %s.\n",
1550 mdname(conf->mddev), bdn);
1551 else
1552 retry = 1;
1553 if (retry)
1554 set_bit(R5_ReadError, &sh->dev[i].flags);
1555 else {
1556 clear_bit(R5_ReadError, &sh->dev[i].flags);
1557 clear_bit(R5_ReWrite, &sh->dev[i].flags);
1558 md_error(conf->mddev, rdev);
1561 rdev_dec_pending(conf->disks[i].rdev, conf->mddev);
1562 clear_bit(R5_LOCKED, &sh->dev[i].flags);
1563 set_bit(STRIPE_HANDLE, &sh->state);
1564 release_stripe(sh);
1567 static void raid5_end_write_request(struct bio *bi, int error)
1569 struct stripe_head *sh = bi->bi_private;
1570 raid5_conf_t *conf = sh->raid_conf;
1571 int disks = sh->disks, i;
1572 int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
1574 for (i=0 ; i<disks; i++)
1575 if (bi == &sh->dev[i].req)
1576 break;
1578 pr_debug("end_write_request %llu/%d, count %d, uptodate: %d.\n",
1579 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
1580 uptodate);
1581 if (i == disks) {
1582 BUG();
1583 return;
1586 if (!uptodate)
1587 md_error(conf->mddev, conf->disks[i].rdev);
1589 rdev_dec_pending(conf->disks[i].rdev, conf->mddev);
1591 clear_bit(R5_LOCKED, &sh->dev[i].flags);
1592 set_bit(STRIPE_HANDLE, &sh->state);
1593 release_stripe(sh);
1597 static sector_t compute_blocknr(struct stripe_head *sh, int i, int previous);
1599 static void raid5_build_block(struct stripe_head *sh, int i, int previous)
1601 struct r5dev *dev = &sh->dev[i];
1603 bio_init(&dev->req);
1604 dev->req.bi_io_vec = &dev->vec;
1605 dev->req.bi_vcnt++;
1606 dev->req.bi_max_vecs++;
1607 dev->vec.bv_page = dev->page;
1608 dev->vec.bv_len = STRIPE_SIZE;
1609 dev->vec.bv_offset = 0;
1611 dev->req.bi_sector = sh->sector;
1612 dev->req.bi_private = sh;
1614 dev->flags = 0;
1615 dev->sector = compute_blocknr(sh, i, previous);
1618 static void error(mddev_t *mddev, mdk_rdev_t *rdev)
1620 char b[BDEVNAME_SIZE];
1621 raid5_conf_t *conf = (raid5_conf_t *) mddev->private;
1622 pr_debug("raid5: error called\n");
1624 if (!test_bit(Faulty, &rdev->flags)) {
1625 set_bit(MD_CHANGE_DEVS, &mddev->flags);
1626 if (test_and_clear_bit(In_sync, &rdev->flags)) {
1627 unsigned long flags;
1628 spin_lock_irqsave(&conf->device_lock, flags);
1629 mddev->degraded++;
1630 spin_unlock_irqrestore(&conf->device_lock, flags);
1632 * if recovery was running, make sure it aborts.
1634 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1636 set_bit(Faulty, &rdev->flags);
1637 printk(KERN_ALERT
1638 "raid5: Disk failure on %s, disabling device.\n"
1639 "raid5: Operation continuing on %d devices.\n",
1640 bdevname(rdev->bdev,b), conf->raid_disks - mddev->degraded);
1645 * Input: a 'big' sector number,
1646 * Output: index of the data and parity disk, and the sector # in them.
1648 static sector_t raid5_compute_sector(raid5_conf_t *conf, sector_t r_sector,
1649 int previous, int *dd_idx,
1650 struct stripe_head *sh)
1652 long stripe;
1653 unsigned long chunk_number;
1654 unsigned int chunk_offset;
1655 int pd_idx, qd_idx;
1656 int ddf_layout = 0;
1657 sector_t new_sector;
1658 int algorithm = previous ? conf->prev_algo
1659 : conf->algorithm;
1660 int sectors_per_chunk = previous ? conf->prev_chunk_sectors
1661 : conf->chunk_sectors;
1662 int raid_disks = previous ? conf->previous_raid_disks
1663 : conf->raid_disks;
1664 int data_disks = raid_disks - conf->max_degraded;
1666 /* First compute the information on this sector */
1669 * Compute the chunk number and the sector offset inside the chunk
1671 chunk_offset = sector_div(r_sector, sectors_per_chunk);
1672 chunk_number = r_sector;
1673 BUG_ON(r_sector != chunk_number);
1676 * Compute the stripe number
1678 stripe = chunk_number / data_disks;
1681 * Compute the data disk and parity disk indexes inside the stripe
1683 *dd_idx = chunk_number % data_disks;
1686 * Select the parity disk based on the user selected algorithm.
1688 pd_idx = qd_idx = ~0;
1689 switch(conf->level) {
1690 case 4:
1691 pd_idx = data_disks;
1692 break;
1693 case 5:
1694 switch (algorithm) {
1695 case ALGORITHM_LEFT_ASYMMETRIC:
1696 pd_idx = data_disks - stripe % raid_disks;
1697 if (*dd_idx >= pd_idx)
1698 (*dd_idx)++;
1699 break;
1700 case ALGORITHM_RIGHT_ASYMMETRIC:
1701 pd_idx = stripe % raid_disks;
1702 if (*dd_idx >= pd_idx)
1703 (*dd_idx)++;
1704 break;
1705 case ALGORITHM_LEFT_SYMMETRIC:
1706 pd_idx = data_disks - stripe % raid_disks;
1707 *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
1708 break;
1709 case ALGORITHM_RIGHT_SYMMETRIC:
1710 pd_idx = stripe % raid_disks;
1711 *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
1712 break;
1713 case ALGORITHM_PARITY_0:
1714 pd_idx = 0;
1715 (*dd_idx)++;
1716 break;
1717 case ALGORITHM_PARITY_N:
1718 pd_idx = data_disks;
1719 break;
1720 default:
1721 printk(KERN_ERR "raid5: unsupported algorithm %d\n",
1722 algorithm);
1723 BUG();
1725 break;
1726 case 6:
1728 switch (algorithm) {
1729 case ALGORITHM_LEFT_ASYMMETRIC:
1730 pd_idx = raid_disks - 1 - (stripe % raid_disks);
1731 qd_idx = pd_idx + 1;
1732 if (pd_idx == raid_disks-1) {
1733 (*dd_idx)++; /* Q D D D P */
1734 qd_idx = 0;
1735 } else if (*dd_idx >= pd_idx)
1736 (*dd_idx) += 2; /* D D P Q D */
1737 break;
1738 case ALGORITHM_RIGHT_ASYMMETRIC:
1739 pd_idx = stripe % raid_disks;
1740 qd_idx = pd_idx + 1;
1741 if (pd_idx == raid_disks-1) {
1742 (*dd_idx)++; /* Q D D D P */
1743 qd_idx = 0;
1744 } else if (*dd_idx >= pd_idx)
1745 (*dd_idx) += 2; /* D D P Q D */
1746 break;
1747 case ALGORITHM_LEFT_SYMMETRIC:
1748 pd_idx = raid_disks - 1 - (stripe % raid_disks);
1749 qd_idx = (pd_idx + 1) % raid_disks;
1750 *dd_idx = (pd_idx + 2 + *dd_idx) % raid_disks;
1751 break;
1752 case ALGORITHM_RIGHT_SYMMETRIC:
1753 pd_idx = stripe % raid_disks;
1754 qd_idx = (pd_idx + 1) % raid_disks;
1755 *dd_idx = (pd_idx + 2 + *dd_idx) % raid_disks;
1756 break;
1758 case ALGORITHM_PARITY_0:
1759 pd_idx = 0;
1760 qd_idx = 1;
1761 (*dd_idx) += 2;
1762 break;
1763 case ALGORITHM_PARITY_N:
1764 pd_idx = data_disks;
1765 qd_idx = data_disks + 1;
1766 break;
1768 case ALGORITHM_ROTATING_ZERO_RESTART:
1769 /* Exactly the same as RIGHT_ASYMMETRIC, but or
1770 * of blocks for computing Q is different.
1772 pd_idx = stripe % raid_disks;
1773 qd_idx = pd_idx + 1;
1774 if (pd_idx == raid_disks-1) {
1775 (*dd_idx)++; /* Q D D D P */
1776 qd_idx = 0;
1777 } else if (*dd_idx >= pd_idx)
1778 (*dd_idx) += 2; /* D D P Q D */
1779 ddf_layout = 1;
1780 break;
1782 case ALGORITHM_ROTATING_N_RESTART:
1783 /* Same a left_asymmetric, by first stripe is
1784 * D D D P Q rather than
1785 * Q D D D P
1787 pd_idx = raid_disks - 1 - ((stripe + 1) % raid_disks);
1788 qd_idx = pd_idx + 1;
1789 if (pd_idx == raid_disks-1) {
1790 (*dd_idx)++; /* Q D D D P */
1791 qd_idx = 0;
1792 } else if (*dd_idx >= pd_idx)
1793 (*dd_idx) += 2; /* D D P Q D */
1794 ddf_layout = 1;
1795 break;
1797 case ALGORITHM_ROTATING_N_CONTINUE:
1798 /* Same as left_symmetric but Q is before P */
1799 pd_idx = raid_disks - 1 - (stripe % raid_disks);
1800 qd_idx = (pd_idx + raid_disks - 1) % raid_disks;
1801 *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
1802 ddf_layout = 1;
1803 break;
1805 case ALGORITHM_LEFT_ASYMMETRIC_6:
1806 /* RAID5 left_asymmetric, with Q on last device */
1807 pd_idx = data_disks - stripe % (raid_disks-1);
1808 if (*dd_idx >= pd_idx)
1809 (*dd_idx)++;
1810 qd_idx = raid_disks - 1;
1811 break;
1813 case ALGORITHM_RIGHT_ASYMMETRIC_6:
1814 pd_idx = stripe % (raid_disks-1);
1815 if (*dd_idx >= pd_idx)
1816 (*dd_idx)++;
1817 qd_idx = raid_disks - 1;
1818 break;
1820 case ALGORITHM_LEFT_SYMMETRIC_6:
1821 pd_idx = data_disks - stripe % (raid_disks-1);
1822 *dd_idx = (pd_idx + 1 + *dd_idx) % (raid_disks-1);
1823 qd_idx = raid_disks - 1;
1824 break;
1826 case ALGORITHM_RIGHT_SYMMETRIC_6:
1827 pd_idx = stripe % (raid_disks-1);
1828 *dd_idx = (pd_idx + 1 + *dd_idx) % (raid_disks-1);
1829 qd_idx = raid_disks - 1;
1830 break;
1832 case ALGORITHM_PARITY_0_6:
1833 pd_idx = 0;
1834 (*dd_idx)++;
1835 qd_idx = raid_disks - 1;
1836 break;
1839 default:
1840 printk(KERN_CRIT "raid6: unsupported algorithm %d\n",
1841 algorithm);
1842 BUG();
1844 break;
1847 if (sh) {
1848 sh->pd_idx = pd_idx;
1849 sh->qd_idx = qd_idx;
1850 sh->ddf_layout = ddf_layout;
1853 * Finally, compute the new sector number
1855 new_sector = (sector_t)stripe * sectors_per_chunk + chunk_offset;
1856 return new_sector;
1860 static sector_t compute_blocknr(struct stripe_head *sh, int i, int previous)
1862 raid5_conf_t *conf = sh->raid_conf;
1863 int raid_disks = sh->disks;
1864 int data_disks = raid_disks - conf->max_degraded;
1865 sector_t new_sector = sh->sector, check;
1866 int sectors_per_chunk = previous ? conf->prev_chunk_sectors
1867 : conf->chunk_sectors;
1868 int algorithm = previous ? conf->prev_algo
1869 : conf->algorithm;
1870 sector_t stripe;
1871 int chunk_offset;
1872 int chunk_number, dummy1, dd_idx = i;
1873 sector_t r_sector;
1874 struct stripe_head sh2;
1877 chunk_offset = sector_div(new_sector, sectors_per_chunk);
1878 stripe = new_sector;
1879 BUG_ON(new_sector != stripe);
1881 if (i == sh->pd_idx)
1882 return 0;
1883 switch(conf->level) {
1884 case 4: break;
1885 case 5:
1886 switch (algorithm) {
1887 case ALGORITHM_LEFT_ASYMMETRIC:
1888 case ALGORITHM_RIGHT_ASYMMETRIC:
1889 if (i > sh->pd_idx)
1890 i--;
1891 break;
1892 case ALGORITHM_LEFT_SYMMETRIC:
1893 case ALGORITHM_RIGHT_SYMMETRIC:
1894 if (i < sh->pd_idx)
1895 i += raid_disks;
1896 i -= (sh->pd_idx + 1);
1897 break;
1898 case ALGORITHM_PARITY_0:
1899 i -= 1;
1900 break;
1901 case ALGORITHM_PARITY_N:
1902 break;
1903 default:
1904 printk(KERN_ERR "raid5: unsupported algorithm %d\n",
1905 algorithm);
1906 BUG();
1908 break;
1909 case 6:
1910 if (i == sh->qd_idx)
1911 return 0; /* It is the Q disk */
1912 switch (algorithm) {
1913 case ALGORITHM_LEFT_ASYMMETRIC:
1914 case ALGORITHM_RIGHT_ASYMMETRIC:
1915 case ALGORITHM_ROTATING_ZERO_RESTART:
1916 case ALGORITHM_ROTATING_N_RESTART:
1917 if (sh->pd_idx == raid_disks-1)
1918 i--; /* Q D D D P */
1919 else if (i > sh->pd_idx)
1920 i -= 2; /* D D P Q D */
1921 break;
1922 case ALGORITHM_LEFT_SYMMETRIC:
1923 case ALGORITHM_RIGHT_SYMMETRIC:
1924 if (sh->pd_idx == raid_disks-1)
1925 i--; /* Q D D D P */
1926 else {
1927 /* D D P Q D */
1928 if (i < sh->pd_idx)
1929 i += raid_disks;
1930 i -= (sh->pd_idx + 2);
1932 break;
1933 case ALGORITHM_PARITY_0:
1934 i -= 2;
1935 break;
1936 case ALGORITHM_PARITY_N:
1937 break;
1938 case ALGORITHM_ROTATING_N_CONTINUE:
1939 /* Like left_symmetric, but P is before Q */
1940 if (sh->pd_idx == 0)
1941 i--; /* P D D D Q */
1942 else {
1943 /* D D Q P D */
1944 if (i < sh->pd_idx)
1945 i += raid_disks;
1946 i -= (sh->pd_idx + 1);
1948 break;
1949 case ALGORITHM_LEFT_ASYMMETRIC_6:
1950 case ALGORITHM_RIGHT_ASYMMETRIC_6:
1951 if (i > sh->pd_idx)
1952 i--;
1953 break;
1954 case ALGORITHM_LEFT_SYMMETRIC_6:
1955 case ALGORITHM_RIGHT_SYMMETRIC_6:
1956 if (i < sh->pd_idx)
1957 i += data_disks + 1;
1958 i -= (sh->pd_idx + 1);
1959 break;
1960 case ALGORITHM_PARITY_0_6:
1961 i -= 1;
1962 break;
1963 default:
1964 printk(KERN_CRIT "raid6: unsupported algorithm %d\n",
1965 algorithm);
1966 BUG();
1968 break;
1971 chunk_number = stripe * data_disks + i;
1972 r_sector = (sector_t)chunk_number * sectors_per_chunk + chunk_offset;
1974 check = raid5_compute_sector(conf, r_sector,
1975 previous, &dummy1, &sh2);
1976 if (check != sh->sector || dummy1 != dd_idx || sh2.pd_idx != sh->pd_idx
1977 || sh2.qd_idx != sh->qd_idx) {
1978 printk(KERN_ERR "compute_blocknr: map not correct\n");
1979 return 0;
1981 return r_sector;
1985 static void
1986 schedule_reconstruction(struct stripe_head *sh, struct stripe_head_state *s,
1987 int rcw, int expand)
1989 int i, pd_idx = sh->pd_idx, disks = sh->disks;
1990 raid5_conf_t *conf = sh->raid_conf;
1991 int level = conf->level;
1993 if (rcw) {
1994 /* if we are not expanding this is a proper write request, and
1995 * there will be bios with new data to be drained into the
1996 * stripe cache
1998 if (!expand) {
1999 sh->reconstruct_state = reconstruct_state_drain_run;
2000 set_bit(STRIPE_OP_BIODRAIN, &s->ops_request);
2001 } else
2002 sh->reconstruct_state = reconstruct_state_run;
2004 set_bit(STRIPE_OP_RECONSTRUCT, &s->ops_request);
2006 for (i = disks; i--; ) {
2007 struct r5dev *dev = &sh->dev[i];
2009 if (dev->towrite) {
2010 set_bit(R5_LOCKED, &dev->flags);
2011 set_bit(R5_Wantdrain, &dev->flags);
2012 if (!expand)
2013 clear_bit(R5_UPTODATE, &dev->flags);
2014 s->locked++;
2017 if (s->locked + conf->max_degraded == disks)
2018 if (!test_and_set_bit(STRIPE_FULL_WRITE, &sh->state))
2019 atomic_inc(&conf->pending_full_writes);
2020 } else {
2021 BUG_ON(level == 6);
2022 BUG_ON(!(test_bit(R5_UPTODATE, &sh->dev[pd_idx].flags) ||
2023 test_bit(R5_Wantcompute, &sh->dev[pd_idx].flags)));
2025 sh->reconstruct_state = reconstruct_state_prexor_drain_run;
2026 set_bit(STRIPE_OP_PREXOR, &s->ops_request);
2027 set_bit(STRIPE_OP_BIODRAIN, &s->ops_request);
2028 set_bit(STRIPE_OP_RECONSTRUCT, &s->ops_request);
2030 for (i = disks; i--; ) {
2031 struct r5dev *dev = &sh->dev[i];
2032 if (i == pd_idx)
2033 continue;
2035 if (dev->towrite &&
2036 (test_bit(R5_UPTODATE, &dev->flags) ||
2037 test_bit(R5_Wantcompute, &dev->flags))) {
2038 set_bit(R5_Wantdrain, &dev->flags);
2039 set_bit(R5_LOCKED, &dev->flags);
2040 clear_bit(R5_UPTODATE, &dev->flags);
2041 s->locked++;
2046 /* keep the parity disk(s) locked while asynchronous operations
2047 * are in flight
2049 set_bit(R5_LOCKED, &sh->dev[pd_idx].flags);
2050 clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
2051 s->locked++;
2053 if (level == 6) {
2054 int qd_idx = sh->qd_idx;
2055 struct r5dev *dev = &sh->dev[qd_idx];
2057 set_bit(R5_LOCKED, &dev->flags);
2058 clear_bit(R5_UPTODATE, &dev->flags);
2059 s->locked++;
2062 pr_debug("%s: stripe %llu locked: %d ops_request: %lx\n",
2063 __func__, (unsigned long long)sh->sector,
2064 s->locked, s->ops_request);
2068 * Each stripe/dev can have one or more bion attached.
2069 * toread/towrite point to the first in a chain.
2070 * The bi_next chain must be in order.
2072 static int add_stripe_bio(struct stripe_head *sh, struct bio *bi, int dd_idx, int forwrite)
2074 struct bio **bip;
2075 raid5_conf_t *conf = sh->raid_conf;
2076 int firstwrite=0;
2078 pr_debug("adding bh b#%llu to stripe s#%llu\n",
2079 (unsigned long long)bi->bi_sector,
2080 (unsigned long long)sh->sector);
2083 spin_lock(&sh->lock);
2084 spin_lock_irq(&conf->device_lock);
2085 if (forwrite) {
2086 bip = &sh->dev[dd_idx].towrite;
2087 if (*bip == NULL && sh->dev[dd_idx].written == NULL)
2088 firstwrite = 1;
2089 } else
2090 bip = &sh->dev[dd_idx].toread;
2091 while (*bip && (*bip)->bi_sector < bi->bi_sector) {
2092 if ((*bip)->bi_sector + ((*bip)->bi_size >> 9) > bi->bi_sector)
2093 goto overlap;
2094 bip = & (*bip)->bi_next;
2096 if (*bip && (*bip)->bi_sector < bi->bi_sector + ((bi->bi_size)>>9))
2097 goto overlap;
2099 BUG_ON(*bip && bi->bi_next && (*bip) != bi->bi_next);
2100 if (*bip)
2101 bi->bi_next = *bip;
2102 *bip = bi;
2103 bi->bi_phys_segments++;
2104 spin_unlock_irq(&conf->device_lock);
2105 spin_unlock(&sh->lock);
2107 pr_debug("added bi b#%llu to stripe s#%llu, disk %d.\n",
2108 (unsigned long long)bi->bi_sector,
2109 (unsigned long long)sh->sector, dd_idx);
2111 if (conf->mddev->bitmap && firstwrite) {
2112 bitmap_startwrite(conf->mddev->bitmap, sh->sector,
2113 STRIPE_SECTORS, 0);
2114 sh->bm_seq = conf->seq_flush+1;
2115 set_bit(STRIPE_BIT_DELAY, &sh->state);
2118 if (forwrite) {
2119 /* check if page is covered */
2120 sector_t sector = sh->dev[dd_idx].sector;
2121 for (bi=sh->dev[dd_idx].towrite;
2122 sector < sh->dev[dd_idx].sector + STRIPE_SECTORS &&
2123 bi && bi->bi_sector <= sector;
2124 bi = r5_next_bio(bi, sh->dev[dd_idx].sector)) {
2125 if (bi->bi_sector + (bi->bi_size>>9) >= sector)
2126 sector = bi->bi_sector + (bi->bi_size>>9);
2128 if (sector >= sh->dev[dd_idx].sector + STRIPE_SECTORS)
2129 set_bit(R5_OVERWRITE, &sh->dev[dd_idx].flags);
2131 return 1;
2133 overlap:
2134 set_bit(R5_Overlap, &sh->dev[dd_idx].flags);
2135 spin_unlock_irq(&conf->device_lock);
2136 spin_unlock(&sh->lock);
2137 return 0;
2140 static void end_reshape(raid5_conf_t *conf);
2142 static void stripe_set_idx(sector_t stripe, raid5_conf_t *conf, int previous,
2143 struct stripe_head *sh)
2145 int sectors_per_chunk =
2146 previous ? conf->prev_chunk_sectors : conf->chunk_sectors;
2147 int dd_idx;
2148 int chunk_offset = sector_div(stripe, sectors_per_chunk);
2149 int disks = previous ? conf->previous_raid_disks : conf->raid_disks;
2151 raid5_compute_sector(conf,
2152 stripe * (disks - conf->max_degraded)
2153 *sectors_per_chunk + chunk_offset,
2154 previous,
2155 &dd_idx, sh);
2158 static void
2159 handle_failed_stripe(raid5_conf_t *conf, struct stripe_head *sh,
2160 struct stripe_head_state *s, int disks,
2161 struct bio **return_bi)
2163 int i;
2164 for (i = disks; i--; ) {
2165 struct bio *bi;
2166 int bitmap_end = 0;
2168 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
2169 mdk_rdev_t *rdev;
2170 rcu_read_lock();
2171 rdev = rcu_dereference(conf->disks[i].rdev);
2172 if (rdev && test_bit(In_sync, &rdev->flags))
2173 /* multiple read failures in one stripe */
2174 md_error(conf->mddev, rdev);
2175 rcu_read_unlock();
2177 spin_lock_irq(&conf->device_lock);
2178 /* fail all writes first */
2179 bi = sh->dev[i].towrite;
2180 sh->dev[i].towrite = NULL;
2181 if (bi) {
2182 s->to_write--;
2183 bitmap_end = 1;
2186 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
2187 wake_up(&conf->wait_for_overlap);
2189 while (bi && bi->bi_sector <
2190 sh->dev[i].sector + STRIPE_SECTORS) {
2191 struct bio *nextbi = r5_next_bio(bi, sh->dev[i].sector);
2192 clear_bit(BIO_UPTODATE, &bi->bi_flags);
2193 if (!raid5_dec_bi_phys_segments(bi)) {
2194 md_write_end(conf->mddev);
2195 bi->bi_next = *return_bi;
2196 *return_bi = bi;
2198 bi = nextbi;
2200 /* and fail all 'written' */
2201 bi = sh->dev[i].written;
2202 sh->dev[i].written = NULL;
2203 if (bi) bitmap_end = 1;
2204 while (bi && bi->bi_sector <
2205 sh->dev[i].sector + STRIPE_SECTORS) {
2206 struct bio *bi2 = r5_next_bio(bi, sh->dev[i].sector);
2207 clear_bit(BIO_UPTODATE, &bi->bi_flags);
2208 if (!raid5_dec_bi_phys_segments(bi)) {
2209 md_write_end(conf->mddev);
2210 bi->bi_next = *return_bi;
2211 *return_bi = bi;
2213 bi = bi2;
2216 /* fail any reads if this device is non-operational and
2217 * the data has not reached the cache yet.
2219 if (!test_bit(R5_Wantfill, &sh->dev[i].flags) &&
2220 (!test_bit(R5_Insync, &sh->dev[i].flags) ||
2221 test_bit(R5_ReadError, &sh->dev[i].flags))) {
2222 bi = sh->dev[i].toread;
2223 sh->dev[i].toread = NULL;
2224 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
2225 wake_up(&conf->wait_for_overlap);
2226 if (bi) s->to_read--;
2227 while (bi && bi->bi_sector <
2228 sh->dev[i].sector + STRIPE_SECTORS) {
2229 struct bio *nextbi =
2230 r5_next_bio(bi, sh->dev[i].sector);
2231 clear_bit(BIO_UPTODATE, &bi->bi_flags);
2232 if (!raid5_dec_bi_phys_segments(bi)) {
2233 bi->bi_next = *return_bi;
2234 *return_bi = bi;
2236 bi = nextbi;
2239 spin_unlock_irq(&conf->device_lock);
2240 if (bitmap_end)
2241 bitmap_endwrite(conf->mddev->bitmap, sh->sector,
2242 STRIPE_SECTORS, 0, 0);
2245 if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
2246 if (atomic_dec_and_test(&conf->pending_full_writes))
2247 md_wakeup_thread(conf->mddev->thread);
2250 /* fetch_block5 - checks the given member device to see if its data needs
2251 * to be read or computed to satisfy a request.
2253 * Returns 1 when no more member devices need to be checked, otherwise returns
2254 * 0 to tell the loop in handle_stripe_fill5 to continue
2256 static int fetch_block5(struct stripe_head *sh, struct stripe_head_state *s,
2257 int disk_idx, int disks)
2259 struct r5dev *dev = &sh->dev[disk_idx];
2260 struct r5dev *failed_dev = &sh->dev[s->failed_num];
2262 /* is the data in this block needed, and can we get it? */
2263 if (!test_bit(R5_LOCKED, &dev->flags) &&
2264 !test_bit(R5_UPTODATE, &dev->flags) &&
2265 (dev->toread ||
2266 (dev->towrite && !test_bit(R5_OVERWRITE, &dev->flags)) ||
2267 s->syncing || s->expanding ||
2268 (s->failed &&
2269 (failed_dev->toread ||
2270 (failed_dev->towrite &&
2271 !test_bit(R5_OVERWRITE, &failed_dev->flags)))))) {
2272 /* We would like to get this block, possibly by computing it,
2273 * otherwise read it if the backing disk is insync
2275 if ((s->uptodate == disks - 1) &&
2276 (s->failed && disk_idx == s->failed_num)) {
2277 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
2278 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
2279 set_bit(R5_Wantcompute, &dev->flags);
2280 sh->ops.target = disk_idx;
2281 sh->ops.target2 = -1;
2282 s->req_compute = 1;
2283 /* Careful: from this point on 'uptodate' is in the eye
2284 * of raid_run_ops which services 'compute' operations
2285 * before writes. R5_Wantcompute flags a block that will
2286 * be R5_UPTODATE by the time it is needed for a
2287 * subsequent operation.
2289 s->uptodate++;
2290 return 1; /* uptodate + compute == disks */
2291 } else if (test_bit(R5_Insync, &dev->flags)) {
2292 set_bit(R5_LOCKED, &dev->flags);
2293 set_bit(R5_Wantread, &dev->flags);
2294 s->locked++;
2295 pr_debug("Reading block %d (sync=%d)\n", disk_idx,
2296 s->syncing);
2300 return 0;
2304 * handle_stripe_fill5 - read or compute data to satisfy pending requests.
2306 static void handle_stripe_fill5(struct stripe_head *sh,
2307 struct stripe_head_state *s, int disks)
2309 int i;
2311 /* look for blocks to read/compute, skip this if a compute
2312 * is already in flight, or if the stripe contents are in the
2313 * midst of changing due to a write
2315 if (!test_bit(STRIPE_COMPUTE_RUN, &sh->state) && !sh->check_state &&
2316 !sh->reconstruct_state)
2317 for (i = disks; i--; )
2318 if (fetch_block5(sh, s, i, disks))
2319 break;
2320 set_bit(STRIPE_HANDLE, &sh->state);
2323 /* fetch_block6 - checks the given member device to see if its data needs
2324 * to be read or computed to satisfy a request.
2326 * Returns 1 when no more member devices need to be checked, otherwise returns
2327 * 0 to tell the loop in handle_stripe_fill6 to continue
2329 static int fetch_block6(struct stripe_head *sh, struct stripe_head_state *s,
2330 struct r6_state *r6s, int disk_idx, int disks)
2332 struct r5dev *dev = &sh->dev[disk_idx];
2333 struct r5dev *fdev[2] = { &sh->dev[r6s->failed_num[0]],
2334 &sh->dev[r6s->failed_num[1]] };
2336 if (!test_bit(R5_LOCKED, &dev->flags) &&
2337 !test_bit(R5_UPTODATE, &dev->flags) &&
2338 (dev->toread ||
2339 (dev->towrite && !test_bit(R5_OVERWRITE, &dev->flags)) ||
2340 s->syncing || s->expanding ||
2341 (s->failed >= 1 &&
2342 (fdev[0]->toread || s->to_write)) ||
2343 (s->failed >= 2 &&
2344 (fdev[1]->toread || s->to_write)))) {
2345 /* we would like to get this block, possibly by computing it,
2346 * otherwise read it if the backing disk is insync
2348 BUG_ON(test_bit(R5_Wantcompute, &dev->flags));
2349 BUG_ON(test_bit(R5_Wantread, &dev->flags));
2350 if ((s->uptodate == disks - 1) &&
2351 (s->failed && (disk_idx == r6s->failed_num[0] ||
2352 disk_idx == r6s->failed_num[1]))) {
2353 /* have disk failed, and we're requested to fetch it;
2354 * do compute it
2356 pr_debug("Computing stripe %llu block %d\n",
2357 (unsigned long long)sh->sector, disk_idx);
2358 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
2359 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
2360 set_bit(R5_Wantcompute, &dev->flags);
2361 sh->ops.target = disk_idx;
2362 sh->ops.target2 = -1; /* no 2nd target */
2363 s->req_compute = 1;
2364 s->uptodate++;
2365 return 1;
2366 } else if (s->uptodate == disks-2 && s->failed >= 2) {
2367 /* Computing 2-failure is *very* expensive; only
2368 * do it if failed >= 2
2370 int other;
2371 for (other = disks; other--; ) {
2372 if (other == disk_idx)
2373 continue;
2374 if (!test_bit(R5_UPTODATE,
2375 &sh->dev[other].flags))
2376 break;
2378 BUG_ON(other < 0);
2379 pr_debug("Computing stripe %llu blocks %d,%d\n",
2380 (unsigned long long)sh->sector,
2381 disk_idx, other);
2382 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
2383 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
2384 set_bit(R5_Wantcompute, &sh->dev[disk_idx].flags);
2385 set_bit(R5_Wantcompute, &sh->dev[other].flags);
2386 sh->ops.target = disk_idx;
2387 sh->ops.target2 = other;
2388 s->uptodate += 2;
2389 s->req_compute = 1;
2390 return 1;
2391 } else if (test_bit(R5_Insync, &dev->flags)) {
2392 set_bit(R5_LOCKED, &dev->flags);
2393 set_bit(R5_Wantread, &dev->flags);
2394 s->locked++;
2395 pr_debug("Reading block %d (sync=%d)\n",
2396 disk_idx, s->syncing);
2400 return 0;
2404 * handle_stripe_fill6 - read or compute data to satisfy pending requests.
2406 static void handle_stripe_fill6(struct stripe_head *sh,
2407 struct stripe_head_state *s, struct r6_state *r6s,
2408 int disks)
2410 int i;
2412 /* look for blocks to read/compute, skip this if a compute
2413 * is already in flight, or if the stripe contents are in the
2414 * midst of changing due to a write
2416 if (!test_bit(STRIPE_COMPUTE_RUN, &sh->state) && !sh->check_state &&
2417 !sh->reconstruct_state)
2418 for (i = disks; i--; )
2419 if (fetch_block6(sh, s, r6s, i, disks))
2420 break;
2421 set_bit(STRIPE_HANDLE, &sh->state);
2425 /* handle_stripe_clean_event
2426 * any written block on an uptodate or failed drive can be returned.
2427 * Note that if we 'wrote' to a failed drive, it will be UPTODATE, but
2428 * never LOCKED, so we don't need to test 'failed' directly.
2430 static void handle_stripe_clean_event(raid5_conf_t *conf,
2431 struct stripe_head *sh, int disks, struct bio **return_bi)
2433 int i;
2434 struct r5dev *dev;
2436 for (i = disks; i--; )
2437 if (sh->dev[i].written) {
2438 dev = &sh->dev[i];
2439 if (!test_bit(R5_LOCKED, &dev->flags) &&
2440 test_bit(R5_UPTODATE, &dev->flags)) {
2441 /* We can return any write requests */
2442 struct bio *wbi, *wbi2;
2443 int bitmap_end = 0;
2444 pr_debug("Return write for disc %d\n", i);
2445 spin_lock_irq(&conf->device_lock);
2446 wbi = dev->written;
2447 dev->written = NULL;
2448 while (wbi && wbi->bi_sector <
2449 dev->sector + STRIPE_SECTORS) {
2450 wbi2 = r5_next_bio(wbi, dev->sector);
2451 if (!raid5_dec_bi_phys_segments(wbi)) {
2452 md_write_end(conf->mddev);
2453 wbi->bi_next = *return_bi;
2454 *return_bi = wbi;
2456 wbi = wbi2;
2458 if (dev->towrite == NULL)
2459 bitmap_end = 1;
2460 spin_unlock_irq(&conf->device_lock);
2461 if (bitmap_end)
2462 bitmap_endwrite(conf->mddev->bitmap,
2463 sh->sector,
2464 STRIPE_SECTORS,
2465 !test_bit(STRIPE_DEGRADED, &sh->state),
2470 if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
2471 if (atomic_dec_and_test(&conf->pending_full_writes))
2472 md_wakeup_thread(conf->mddev->thread);
2475 static void handle_stripe_dirtying5(raid5_conf_t *conf,
2476 struct stripe_head *sh, struct stripe_head_state *s, int disks)
2478 int rmw = 0, rcw = 0, i;
2479 for (i = disks; i--; ) {
2480 /* would I have to read this buffer for read_modify_write */
2481 struct r5dev *dev = &sh->dev[i];
2482 if ((dev->towrite || i == sh->pd_idx) &&
2483 !test_bit(R5_LOCKED, &dev->flags) &&
2484 !(test_bit(R5_UPTODATE, &dev->flags) ||
2485 test_bit(R5_Wantcompute, &dev->flags))) {
2486 if (test_bit(R5_Insync, &dev->flags))
2487 rmw++;
2488 else
2489 rmw += 2*disks; /* cannot read it */
2491 /* Would I have to read this buffer for reconstruct_write */
2492 if (!test_bit(R5_OVERWRITE, &dev->flags) && i != sh->pd_idx &&
2493 !test_bit(R5_LOCKED, &dev->flags) &&
2494 !(test_bit(R5_UPTODATE, &dev->flags) ||
2495 test_bit(R5_Wantcompute, &dev->flags))) {
2496 if (test_bit(R5_Insync, &dev->flags)) rcw++;
2497 else
2498 rcw += 2*disks;
2501 pr_debug("for sector %llu, rmw=%d rcw=%d\n",
2502 (unsigned long long)sh->sector, rmw, rcw);
2503 set_bit(STRIPE_HANDLE, &sh->state);
2504 if (rmw < rcw && rmw > 0)
2505 /* prefer read-modify-write, but need to get some data */
2506 for (i = disks; i--; ) {
2507 struct r5dev *dev = &sh->dev[i];
2508 if ((dev->towrite || i == sh->pd_idx) &&
2509 !test_bit(R5_LOCKED, &dev->flags) &&
2510 !(test_bit(R5_UPTODATE, &dev->flags) ||
2511 test_bit(R5_Wantcompute, &dev->flags)) &&
2512 test_bit(R5_Insync, &dev->flags)) {
2513 if (
2514 test_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2515 pr_debug("Read_old block "
2516 "%d for r-m-w\n", i);
2517 set_bit(R5_LOCKED, &dev->flags);
2518 set_bit(R5_Wantread, &dev->flags);
2519 s->locked++;
2520 } else {
2521 set_bit(STRIPE_DELAYED, &sh->state);
2522 set_bit(STRIPE_HANDLE, &sh->state);
2526 if (rcw <= rmw && rcw > 0)
2527 /* want reconstruct write, but need to get some data */
2528 for (i = disks; i--; ) {
2529 struct r5dev *dev = &sh->dev[i];
2530 if (!test_bit(R5_OVERWRITE, &dev->flags) &&
2531 i != sh->pd_idx &&
2532 !test_bit(R5_LOCKED, &dev->flags) &&
2533 !(test_bit(R5_UPTODATE, &dev->flags) ||
2534 test_bit(R5_Wantcompute, &dev->flags)) &&
2535 test_bit(R5_Insync, &dev->flags)) {
2536 if (
2537 test_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2538 pr_debug("Read_old block "
2539 "%d for Reconstruct\n", i);
2540 set_bit(R5_LOCKED, &dev->flags);
2541 set_bit(R5_Wantread, &dev->flags);
2542 s->locked++;
2543 } else {
2544 set_bit(STRIPE_DELAYED, &sh->state);
2545 set_bit(STRIPE_HANDLE, &sh->state);
2549 /* now if nothing is locked, and if we have enough data,
2550 * we can start a write request
2552 /* since handle_stripe can be called at any time we need to handle the
2553 * case where a compute block operation has been submitted and then a
2554 * subsequent call wants to start a write request. raid_run_ops only
2555 * handles the case where compute block and reconstruct are requested
2556 * simultaneously. If this is not the case then new writes need to be
2557 * held off until the compute completes.
2559 if ((s->req_compute || !test_bit(STRIPE_COMPUTE_RUN, &sh->state)) &&
2560 (s->locked == 0 && (rcw == 0 || rmw == 0) &&
2561 !test_bit(STRIPE_BIT_DELAY, &sh->state)))
2562 schedule_reconstruction(sh, s, rcw == 0, 0);
2565 static void handle_stripe_dirtying6(raid5_conf_t *conf,
2566 struct stripe_head *sh, struct stripe_head_state *s,
2567 struct r6_state *r6s, int disks)
2569 int rcw = 0, pd_idx = sh->pd_idx, i;
2570 int qd_idx = sh->qd_idx;
2572 set_bit(STRIPE_HANDLE, &sh->state);
2573 for (i = disks; i--; ) {
2574 struct r5dev *dev = &sh->dev[i];
2575 /* check if we haven't enough data */
2576 if (!test_bit(R5_OVERWRITE, &dev->flags) &&
2577 i != pd_idx && i != qd_idx &&
2578 !test_bit(R5_LOCKED, &dev->flags) &&
2579 !(test_bit(R5_UPTODATE, &dev->flags) ||
2580 test_bit(R5_Wantcompute, &dev->flags))) {
2581 rcw++;
2582 if (!test_bit(R5_Insync, &dev->flags))
2583 continue; /* it's a failed drive */
2585 if (
2586 test_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2587 pr_debug("Read_old stripe %llu "
2588 "block %d for Reconstruct\n",
2589 (unsigned long long)sh->sector, i);
2590 set_bit(R5_LOCKED, &dev->flags);
2591 set_bit(R5_Wantread, &dev->flags);
2592 s->locked++;
2593 } else {
2594 pr_debug("Request delayed stripe %llu "
2595 "block %d for Reconstruct\n",
2596 (unsigned long long)sh->sector, i);
2597 set_bit(STRIPE_DELAYED, &sh->state);
2598 set_bit(STRIPE_HANDLE, &sh->state);
2602 /* now if nothing is locked, and if we have enough data, we can start a
2603 * write request
2605 if ((s->req_compute || !test_bit(STRIPE_COMPUTE_RUN, &sh->state)) &&
2606 s->locked == 0 && rcw == 0 &&
2607 !test_bit(STRIPE_BIT_DELAY, &sh->state)) {
2608 schedule_reconstruction(sh, s, 1, 0);
2612 static void handle_parity_checks5(raid5_conf_t *conf, struct stripe_head *sh,
2613 struct stripe_head_state *s, int disks)
2615 struct r5dev *dev = NULL;
2617 set_bit(STRIPE_HANDLE, &sh->state);
2619 switch (sh->check_state) {
2620 case check_state_idle:
2621 /* start a new check operation if there are no failures */
2622 if (s->failed == 0) {
2623 BUG_ON(s->uptodate != disks);
2624 sh->check_state = check_state_run;
2625 set_bit(STRIPE_OP_CHECK, &s->ops_request);
2626 clear_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags);
2627 s->uptodate--;
2628 break;
2630 dev = &sh->dev[s->failed_num];
2631 /* fall through */
2632 case check_state_compute_result:
2633 sh->check_state = check_state_idle;
2634 if (!dev)
2635 dev = &sh->dev[sh->pd_idx];
2637 /* check that a write has not made the stripe insync */
2638 if (test_bit(STRIPE_INSYNC, &sh->state))
2639 break;
2641 /* either failed parity check, or recovery is happening */
2642 BUG_ON(!test_bit(R5_UPTODATE, &dev->flags));
2643 BUG_ON(s->uptodate != disks);
2645 set_bit(R5_LOCKED, &dev->flags);
2646 s->locked++;
2647 set_bit(R5_Wantwrite, &dev->flags);
2649 clear_bit(STRIPE_DEGRADED, &sh->state);
2650 set_bit(STRIPE_INSYNC, &sh->state);
2651 break;
2652 case check_state_run:
2653 break; /* we will be called again upon completion */
2654 case check_state_check_result:
2655 sh->check_state = check_state_idle;
2657 /* if a failure occurred during the check operation, leave
2658 * STRIPE_INSYNC not set and let the stripe be handled again
2660 if (s->failed)
2661 break;
2663 /* handle a successful check operation, if parity is correct
2664 * we are done. Otherwise update the mismatch count and repair
2665 * parity if !MD_RECOVERY_CHECK
2667 if ((sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) == 0)
2668 /* parity is correct (on disc,
2669 * not in buffer any more)
2671 set_bit(STRIPE_INSYNC, &sh->state);
2672 else {
2673 conf->mddev->resync_mismatches += STRIPE_SECTORS;
2674 if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery))
2675 /* don't try to repair!! */
2676 set_bit(STRIPE_INSYNC, &sh->state);
2677 else {
2678 sh->check_state = check_state_compute_run;
2679 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
2680 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
2681 set_bit(R5_Wantcompute,
2682 &sh->dev[sh->pd_idx].flags);
2683 sh->ops.target = sh->pd_idx;
2684 sh->ops.target2 = -1;
2685 s->uptodate++;
2688 break;
2689 case check_state_compute_run:
2690 break;
2691 default:
2692 printk(KERN_ERR "%s: unknown check_state: %d sector: %llu\n",
2693 __func__, sh->check_state,
2694 (unsigned long long) sh->sector);
2695 BUG();
2700 static void handle_parity_checks6(raid5_conf_t *conf, struct stripe_head *sh,
2701 struct stripe_head_state *s,
2702 struct r6_state *r6s, int disks)
2704 int pd_idx = sh->pd_idx;
2705 int qd_idx = sh->qd_idx;
2706 struct r5dev *dev;
2708 set_bit(STRIPE_HANDLE, &sh->state);
2710 BUG_ON(s->failed > 2);
2712 /* Want to check and possibly repair P and Q.
2713 * However there could be one 'failed' device, in which
2714 * case we can only check one of them, possibly using the
2715 * other to generate missing data
2718 switch (sh->check_state) {
2719 case check_state_idle:
2720 /* start a new check operation if there are < 2 failures */
2721 if (s->failed == r6s->q_failed) {
2722 /* The only possible failed device holds Q, so it
2723 * makes sense to check P (If anything else were failed,
2724 * we would have used P to recreate it).
2726 sh->check_state = check_state_run;
2728 if (!r6s->q_failed && s->failed < 2) {
2729 /* Q is not failed, and we didn't use it to generate
2730 * anything, so it makes sense to check it
2732 if (sh->check_state == check_state_run)
2733 sh->check_state = check_state_run_pq;
2734 else
2735 sh->check_state = check_state_run_q;
2738 /* discard potentially stale zero_sum_result */
2739 sh->ops.zero_sum_result = 0;
2741 if (sh->check_state == check_state_run) {
2742 /* async_xor_zero_sum destroys the contents of P */
2743 clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
2744 s->uptodate--;
2746 if (sh->check_state >= check_state_run &&
2747 sh->check_state <= check_state_run_pq) {
2748 /* async_syndrome_zero_sum preserves P and Q, so
2749 * no need to mark them !uptodate here
2751 set_bit(STRIPE_OP_CHECK, &s->ops_request);
2752 break;
2755 /* we have 2-disk failure */
2756 BUG_ON(s->failed != 2);
2757 /* fall through */
2758 case check_state_compute_result:
2759 sh->check_state = check_state_idle;
2761 /* check that a write has not made the stripe insync */
2762 if (test_bit(STRIPE_INSYNC, &sh->state))
2763 break;
2765 /* now write out any block on a failed drive,
2766 * or P or Q if they were recomputed
2768 BUG_ON(s->uptodate < disks - 1); /* We don't need Q to recover */
2769 if (s->failed == 2) {
2770 dev = &sh->dev[r6s->failed_num[1]];
2771 s->locked++;
2772 set_bit(R5_LOCKED, &dev->flags);
2773 set_bit(R5_Wantwrite, &dev->flags);
2775 if (s->failed >= 1) {
2776 dev = &sh->dev[r6s->failed_num[0]];
2777 s->locked++;
2778 set_bit(R5_LOCKED, &dev->flags);
2779 set_bit(R5_Wantwrite, &dev->flags);
2781 if (sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) {
2782 dev = &sh->dev[pd_idx];
2783 s->locked++;
2784 set_bit(R5_LOCKED, &dev->flags);
2785 set_bit(R5_Wantwrite, &dev->flags);
2787 if (sh->ops.zero_sum_result & SUM_CHECK_Q_RESULT) {
2788 dev = &sh->dev[qd_idx];
2789 s->locked++;
2790 set_bit(R5_LOCKED, &dev->flags);
2791 set_bit(R5_Wantwrite, &dev->flags);
2793 clear_bit(STRIPE_DEGRADED, &sh->state);
2795 set_bit(STRIPE_INSYNC, &sh->state);
2796 break;
2797 case check_state_run:
2798 case check_state_run_q:
2799 case check_state_run_pq:
2800 break; /* we will be called again upon completion */
2801 case check_state_check_result:
2802 sh->check_state = check_state_idle;
2804 /* handle a successful check operation, if parity is correct
2805 * we are done. Otherwise update the mismatch count and repair
2806 * parity if !MD_RECOVERY_CHECK
2808 if (sh->ops.zero_sum_result == 0) {
2809 /* both parities are correct */
2810 if (!s->failed)
2811 set_bit(STRIPE_INSYNC, &sh->state);
2812 else {
2813 /* in contrast to the raid5 case we can validate
2814 * parity, but still have a failure to write
2815 * back
2817 sh->check_state = check_state_compute_result;
2818 /* Returning at this point means that we may go
2819 * off and bring p and/or q uptodate again so
2820 * we make sure to check zero_sum_result again
2821 * to verify if p or q need writeback
2824 } else {
2825 conf->mddev->resync_mismatches += STRIPE_SECTORS;
2826 if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery))
2827 /* don't try to repair!! */
2828 set_bit(STRIPE_INSYNC, &sh->state);
2829 else {
2830 int *target = &sh->ops.target;
2832 sh->ops.target = -1;
2833 sh->ops.target2 = -1;
2834 sh->check_state = check_state_compute_run;
2835 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
2836 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
2837 if (sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) {
2838 set_bit(R5_Wantcompute,
2839 &sh->dev[pd_idx].flags);
2840 *target = pd_idx;
2841 target = &sh->ops.target2;
2842 s->uptodate++;
2844 if (sh->ops.zero_sum_result & SUM_CHECK_Q_RESULT) {
2845 set_bit(R5_Wantcompute,
2846 &sh->dev[qd_idx].flags);
2847 *target = qd_idx;
2848 s->uptodate++;
2852 break;
2853 case check_state_compute_run:
2854 break;
2855 default:
2856 printk(KERN_ERR "%s: unknown check_state: %d sector: %llu\n",
2857 __func__, sh->check_state,
2858 (unsigned long long) sh->sector);
2859 BUG();
2863 static void handle_stripe_expansion(raid5_conf_t *conf, struct stripe_head *sh,
2864 struct r6_state *r6s)
2866 int i;
2868 /* We have read all the blocks in this stripe and now we need to
2869 * copy some of them into a target stripe for expand.
2871 struct dma_async_tx_descriptor *tx = NULL;
2872 clear_bit(STRIPE_EXPAND_SOURCE, &sh->state);
2873 for (i = 0; i < sh->disks; i++)
2874 if (i != sh->pd_idx && i != sh->qd_idx) {
2875 int dd_idx, j;
2876 struct stripe_head *sh2;
2877 struct async_submit_ctl submit;
2879 sector_t bn = compute_blocknr(sh, i, 1);
2880 sector_t s = raid5_compute_sector(conf, bn, 0,
2881 &dd_idx, NULL);
2882 sh2 = get_active_stripe(conf, s, 0, 1, 1);
2883 if (sh2 == NULL)
2884 /* so far only the early blocks of this stripe
2885 * have been requested. When later blocks
2886 * get requested, we will try again
2888 continue;
2889 if (!test_bit(STRIPE_EXPANDING, &sh2->state) ||
2890 test_bit(R5_Expanded, &sh2->dev[dd_idx].flags)) {
2891 /* must have already done this block */
2892 release_stripe(sh2);
2893 continue;
2896 /* place all the copies on one channel */
2897 init_async_submit(&submit, 0, tx, NULL, NULL, NULL);
2898 tx = async_memcpy(sh2->dev[dd_idx].page,
2899 sh->dev[i].page, 0, 0, STRIPE_SIZE,
2900 &submit);
2902 set_bit(R5_Expanded, &sh2->dev[dd_idx].flags);
2903 set_bit(R5_UPTODATE, &sh2->dev[dd_idx].flags);
2904 for (j = 0; j < conf->raid_disks; j++)
2905 if (j != sh2->pd_idx &&
2906 (!r6s || j != sh2->qd_idx) &&
2907 !test_bit(R5_Expanded, &sh2->dev[j].flags))
2908 break;
2909 if (j == conf->raid_disks) {
2910 set_bit(STRIPE_EXPAND_READY, &sh2->state);
2911 set_bit(STRIPE_HANDLE, &sh2->state);
2913 release_stripe(sh2);
2916 /* done submitting copies, wait for them to complete */
2917 if (tx) {
2918 async_tx_ack(tx);
2919 dma_wait_for_async_tx(tx);
2925 * handle_stripe - do things to a stripe.
2927 * We lock the stripe and then examine the state of various bits
2928 * to see what needs to be done.
2929 * Possible results:
2930 * return some read request which now have data
2931 * return some write requests which are safely on disc
2932 * schedule a read on some buffers
2933 * schedule a write of some buffers
2934 * return confirmation of parity correctness
2936 * buffers are taken off read_list or write_list, and bh_cache buffers
2937 * get BH_Lock set before the stripe lock is released.
2941 static void handle_stripe5(struct stripe_head *sh)
2943 raid5_conf_t *conf = sh->raid_conf;
2944 int disks = sh->disks, i;
2945 struct bio *return_bi = NULL;
2946 struct stripe_head_state s;
2947 struct r5dev *dev;
2948 mdk_rdev_t *blocked_rdev = NULL;
2949 int prexor;
2951 memset(&s, 0, sizeof(s));
2952 pr_debug("handling stripe %llu, state=%#lx cnt=%d, pd_idx=%d check:%d "
2953 "reconstruct:%d\n", (unsigned long long)sh->sector, sh->state,
2954 atomic_read(&sh->count), sh->pd_idx, sh->check_state,
2955 sh->reconstruct_state);
2957 spin_lock(&sh->lock);
2958 clear_bit(STRIPE_HANDLE, &sh->state);
2959 clear_bit(STRIPE_DELAYED, &sh->state);
2961 s.syncing = test_bit(STRIPE_SYNCING, &sh->state);
2962 s.expanding = test_bit(STRIPE_EXPAND_SOURCE, &sh->state);
2963 s.expanded = test_bit(STRIPE_EXPAND_READY, &sh->state);
2965 /* Now to look around and see what can be done */
2966 rcu_read_lock();
2967 for (i=disks; i--; ) {
2968 mdk_rdev_t *rdev;
2970 dev = &sh->dev[i];
2971 clear_bit(R5_Insync, &dev->flags);
2973 pr_debug("check %d: state 0x%lx toread %p read %p write %p "
2974 "written %p\n", i, dev->flags, dev->toread, dev->read,
2975 dev->towrite, dev->written);
2977 /* maybe we can request a biofill operation
2979 * new wantfill requests are only permitted while
2980 * ops_complete_biofill is guaranteed to be inactive
2982 if (test_bit(R5_UPTODATE, &dev->flags) && dev->toread &&
2983 !test_bit(STRIPE_BIOFILL_RUN, &sh->state))
2984 set_bit(R5_Wantfill, &dev->flags);
2986 /* now count some things */
2987 if (test_bit(R5_LOCKED, &dev->flags)) s.locked++;
2988 if (test_bit(R5_UPTODATE, &dev->flags)) s.uptodate++;
2989 if (test_bit(R5_Wantcompute, &dev->flags)) s.compute++;
2991 if (test_bit(R5_Wantfill, &dev->flags))
2992 s.to_fill++;
2993 else if (dev->toread)
2994 s.to_read++;
2995 if (dev->towrite) {
2996 s.to_write++;
2997 if (!test_bit(R5_OVERWRITE, &dev->flags))
2998 s.non_overwrite++;
3000 if (dev->written)
3001 s.written++;
3002 rdev = rcu_dereference(conf->disks[i].rdev);
3003 if (blocked_rdev == NULL &&
3004 rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
3005 blocked_rdev = rdev;
3006 atomic_inc(&rdev->nr_pending);
3008 if (!rdev || !test_bit(In_sync, &rdev->flags)) {
3009 /* The ReadError flag will just be confusing now */
3010 clear_bit(R5_ReadError, &dev->flags);
3011 clear_bit(R5_ReWrite, &dev->flags);
3013 if (!rdev || !test_bit(In_sync, &rdev->flags)
3014 || test_bit(R5_ReadError, &dev->flags)) {
3015 s.failed++;
3016 s.failed_num = i;
3017 } else
3018 set_bit(R5_Insync, &dev->flags);
3020 rcu_read_unlock();
3022 if (unlikely(blocked_rdev)) {
3023 if (s.syncing || s.expanding || s.expanded ||
3024 s.to_write || s.written) {
3025 set_bit(STRIPE_HANDLE, &sh->state);
3026 goto unlock;
3028 /* There is nothing for the blocked_rdev to block */
3029 rdev_dec_pending(blocked_rdev, conf->mddev);
3030 blocked_rdev = NULL;
3033 if (s.to_fill && !test_bit(STRIPE_BIOFILL_RUN, &sh->state)) {
3034 set_bit(STRIPE_OP_BIOFILL, &s.ops_request);
3035 set_bit(STRIPE_BIOFILL_RUN, &sh->state);
3038 pr_debug("locked=%d uptodate=%d to_read=%d"
3039 " to_write=%d failed=%d failed_num=%d\n",
3040 s.locked, s.uptodate, s.to_read, s.to_write,
3041 s.failed, s.failed_num);
3042 /* check if the array has lost two devices and, if so, some requests might
3043 * need to be failed
3045 if (s.failed > 1 && s.to_read+s.to_write+s.written)
3046 handle_failed_stripe(conf, sh, &s, disks, &return_bi);
3047 if (s.failed > 1 && s.syncing) {
3048 md_done_sync(conf->mddev, STRIPE_SECTORS,0);
3049 clear_bit(STRIPE_SYNCING, &sh->state);
3050 s.syncing = 0;
3053 /* might be able to return some write requests if the parity block
3054 * is safe, or on a failed drive
3056 dev = &sh->dev[sh->pd_idx];
3057 if ( s.written &&
3058 ((test_bit(R5_Insync, &dev->flags) &&
3059 !test_bit(R5_LOCKED, &dev->flags) &&
3060 test_bit(R5_UPTODATE, &dev->flags)) ||
3061 (s.failed == 1 && s.failed_num == sh->pd_idx)))
3062 handle_stripe_clean_event(conf, sh, disks, &return_bi);
3064 /* Now we might consider reading some blocks, either to check/generate
3065 * parity, or to satisfy requests
3066 * or to load a block that is being partially written.
3068 if (s.to_read || s.non_overwrite ||
3069 (s.syncing && (s.uptodate + s.compute < disks)) || s.expanding)
3070 handle_stripe_fill5(sh, &s, disks);
3072 /* Now we check to see if any write operations have recently
3073 * completed
3075 prexor = 0;
3076 if (sh->reconstruct_state == reconstruct_state_prexor_drain_result)
3077 prexor = 1;
3078 if (sh->reconstruct_state == reconstruct_state_drain_result ||
3079 sh->reconstruct_state == reconstruct_state_prexor_drain_result) {
3080 sh->reconstruct_state = reconstruct_state_idle;
3082 /* All the 'written' buffers and the parity block are ready to
3083 * be written back to disk
3085 BUG_ON(!test_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags));
3086 for (i = disks; i--; ) {
3087 dev = &sh->dev[i];
3088 if (test_bit(R5_LOCKED, &dev->flags) &&
3089 (i == sh->pd_idx || dev->written)) {
3090 pr_debug("Writing block %d\n", i);
3091 set_bit(R5_Wantwrite, &dev->flags);
3092 if (prexor)
3093 continue;
3094 if (!test_bit(R5_Insync, &dev->flags) ||
3095 (i == sh->pd_idx && s.failed == 0))
3096 set_bit(STRIPE_INSYNC, &sh->state);
3099 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
3100 atomic_dec(&conf->preread_active_stripes);
3101 if (atomic_read(&conf->preread_active_stripes) <
3102 IO_THRESHOLD)
3103 md_wakeup_thread(conf->mddev->thread);
3107 /* Now to consider new write requests and what else, if anything
3108 * should be read. We do not handle new writes when:
3109 * 1/ A 'write' operation (copy+xor) is already in flight.
3110 * 2/ A 'check' operation is in flight, as it may clobber the parity
3111 * block.
3113 if (s.to_write && !sh->reconstruct_state && !sh->check_state)
3114 handle_stripe_dirtying5(conf, sh, &s, disks);
3116 /* maybe we need to check and possibly fix the parity for this stripe
3117 * Any reads will already have been scheduled, so we just see if enough
3118 * data is available. The parity check is held off while parity
3119 * dependent operations are in flight.
3121 if (sh->check_state ||
3122 (s.syncing && s.locked == 0 &&
3123 !test_bit(STRIPE_COMPUTE_RUN, &sh->state) &&
3124 !test_bit(STRIPE_INSYNC, &sh->state)))
3125 handle_parity_checks5(conf, sh, &s, disks);
3127 if (s.syncing && s.locked == 0 && test_bit(STRIPE_INSYNC, &sh->state)) {
3128 md_done_sync(conf->mddev, STRIPE_SECTORS,1);
3129 clear_bit(STRIPE_SYNCING, &sh->state);
3132 /* If the failed drive is just a ReadError, then we might need to progress
3133 * the repair/check process
3135 if (s.failed == 1 && !conf->mddev->ro &&
3136 test_bit(R5_ReadError, &sh->dev[s.failed_num].flags)
3137 && !test_bit(R5_LOCKED, &sh->dev[s.failed_num].flags)
3138 && test_bit(R5_UPTODATE, &sh->dev[s.failed_num].flags)
3140 dev = &sh->dev[s.failed_num];
3141 if (!test_bit(R5_ReWrite, &dev->flags)) {
3142 set_bit(R5_Wantwrite, &dev->flags);
3143 set_bit(R5_ReWrite, &dev->flags);
3144 set_bit(R5_LOCKED, &dev->flags);
3145 s.locked++;
3146 } else {
3147 /* let's read it back */
3148 set_bit(R5_Wantread, &dev->flags);
3149 set_bit(R5_LOCKED, &dev->flags);
3150 s.locked++;
3154 /* Finish reconstruct operations initiated by the expansion process */
3155 if (sh->reconstruct_state == reconstruct_state_result) {
3156 struct stripe_head *sh2
3157 = get_active_stripe(conf, sh->sector, 1, 1, 1);
3158 if (sh2 && test_bit(STRIPE_EXPAND_SOURCE, &sh2->state)) {
3159 /* sh cannot be written until sh2 has been read.
3160 * so arrange for sh to be delayed a little
3162 set_bit(STRIPE_DELAYED, &sh->state);
3163 set_bit(STRIPE_HANDLE, &sh->state);
3164 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE,
3165 &sh2->state))
3166 atomic_inc(&conf->preread_active_stripes);
3167 release_stripe(sh2);
3168 goto unlock;
3170 if (sh2)
3171 release_stripe(sh2);
3173 sh->reconstruct_state = reconstruct_state_idle;
3174 clear_bit(STRIPE_EXPANDING, &sh->state);
3175 for (i = conf->raid_disks; i--; ) {
3176 set_bit(R5_Wantwrite, &sh->dev[i].flags);
3177 set_bit(R5_LOCKED, &sh->dev[i].flags);
3178 s.locked++;
3182 if (s.expanded && test_bit(STRIPE_EXPANDING, &sh->state) &&
3183 !sh->reconstruct_state) {
3184 /* Need to write out all blocks after computing parity */
3185 sh->disks = conf->raid_disks;
3186 stripe_set_idx(sh->sector, conf, 0, sh);
3187 schedule_reconstruction(sh, &s, 1, 1);
3188 } else if (s.expanded && !sh->reconstruct_state && s.locked == 0) {
3189 clear_bit(STRIPE_EXPAND_READY, &sh->state);
3190 atomic_dec(&conf->reshape_stripes);
3191 wake_up(&conf->wait_for_overlap);
3192 md_done_sync(conf->mddev, STRIPE_SECTORS, 1);
3195 if (s.expanding && s.locked == 0 &&
3196 !test_bit(STRIPE_COMPUTE_RUN, &sh->state))
3197 handle_stripe_expansion(conf, sh, NULL);
3199 unlock:
3200 spin_unlock(&sh->lock);
3202 /* wait for this device to become unblocked */
3203 if (unlikely(blocked_rdev))
3204 md_wait_for_blocked_rdev(blocked_rdev, conf->mddev);
3206 if (s.ops_request)
3207 raid_run_ops(sh, s.ops_request);
3209 ops_run_io(sh, &s);
3211 return_io(return_bi);
3214 static void handle_stripe6(struct stripe_head *sh)
3216 raid5_conf_t *conf = sh->raid_conf;
3217 int disks = sh->disks;
3218 struct bio *return_bi = NULL;
3219 int i, pd_idx = sh->pd_idx, qd_idx = sh->qd_idx;
3220 struct stripe_head_state s;
3221 struct r6_state r6s;
3222 struct r5dev *dev, *pdev, *qdev;
3223 mdk_rdev_t *blocked_rdev = NULL;
3225 pr_debug("handling stripe %llu, state=%#lx cnt=%d, "
3226 "pd_idx=%d, qd_idx=%d\n, check:%d, reconstruct:%d\n",
3227 (unsigned long long)sh->sector, sh->state,
3228 atomic_read(&sh->count), pd_idx, qd_idx,
3229 sh->check_state, sh->reconstruct_state);
3230 memset(&s, 0, sizeof(s));
3232 spin_lock(&sh->lock);
3233 clear_bit(STRIPE_HANDLE, &sh->state);
3234 clear_bit(STRIPE_DELAYED, &sh->state);
3236 s.syncing = test_bit(STRIPE_SYNCING, &sh->state);
3237 s.expanding = test_bit(STRIPE_EXPAND_SOURCE, &sh->state);
3238 s.expanded = test_bit(STRIPE_EXPAND_READY, &sh->state);
3239 /* Now to look around and see what can be done */
3241 rcu_read_lock();
3242 for (i=disks; i--; ) {
3243 mdk_rdev_t *rdev;
3244 dev = &sh->dev[i];
3245 clear_bit(R5_Insync, &dev->flags);
3247 pr_debug("check %d: state 0x%lx read %p write %p written %p\n",
3248 i, dev->flags, dev->toread, dev->towrite, dev->written);
3249 /* maybe we can reply to a read
3251 * new wantfill requests are only permitted while
3252 * ops_complete_biofill is guaranteed to be inactive
3254 if (test_bit(R5_UPTODATE, &dev->flags) && dev->toread &&
3255 !test_bit(STRIPE_BIOFILL_RUN, &sh->state))
3256 set_bit(R5_Wantfill, &dev->flags);
3258 /* now count some things */
3259 if (test_bit(R5_LOCKED, &dev->flags)) s.locked++;
3260 if (test_bit(R5_UPTODATE, &dev->flags)) s.uptodate++;
3261 if (test_bit(R5_Wantcompute, &dev->flags)) {
3262 s.compute++;
3263 BUG_ON(s.compute > 2);
3266 if (test_bit(R5_Wantfill, &dev->flags)) {
3267 s.to_fill++;
3268 } else if (dev->toread)
3269 s.to_read++;
3270 if (dev->towrite) {
3271 s.to_write++;
3272 if (!test_bit(R5_OVERWRITE, &dev->flags))
3273 s.non_overwrite++;
3275 if (dev->written)
3276 s.written++;
3277 rdev = rcu_dereference(conf->disks[i].rdev);
3278 if (blocked_rdev == NULL &&
3279 rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
3280 blocked_rdev = rdev;
3281 atomic_inc(&rdev->nr_pending);
3283 if (!rdev || !test_bit(In_sync, &rdev->flags)) {
3284 /* The ReadError flag will just be confusing now */
3285 clear_bit(R5_ReadError, &dev->flags);
3286 clear_bit(R5_ReWrite, &dev->flags);
3288 if (!rdev || !test_bit(In_sync, &rdev->flags)
3289 || test_bit(R5_ReadError, &dev->flags)) {
3290 if (s.failed < 2)
3291 r6s.failed_num[s.failed] = i;
3292 s.failed++;
3293 } else
3294 set_bit(R5_Insync, &dev->flags);
3296 rcu_read_unlock();
3298 if (unlikely(blocked_rdev)) {
3299 if (s.syncing || s.expanding || s.expanded ||
3300 s.to_write || s.written) {
3301 set_bit(STRIPE_HANDLE, &sh->state);
3302 goto unlock;
3304 /* There is nothing for the blocked_rdev to block */
3305 rdev_dec_pending(blocked_rdev, conf->mddev);
3306 blocked_rdev = NULL;
3309 if (s.to_fill && !test_bit(STRIPE_BIOFILL_RUN, &sh->state)) {
3310 set_bit(STRIPE_OP_BIOFILL, &s.ops_request);
3311 set_bit(STRIPE_BIOFILL_RUN, &sh->state);
3314 pr_debug("locked=%d uptodate=%d to_read=%d"
3315 " to_write=%d failed=%d failed_num=%d,%d\n",
3316 s.locked, s.uptodate, s.to_read, s.to_write, s.failed,
3317 r6s.failed_num[0], r6s.failed_num[1]);
3318 /* check if the array has lost >2 devices and, if so, some requests
3319 * might need to be failed
3321 if (s.failed > 2 && s.to_read+s.to_write+s.written)
3322 handle_failed_stripe(conf, sh, &s, disks, &return_bi);
3323 if (s.failed > 2 && s.syncing) {
3324 md_done_sync(conf->mddev, STRIPE_SECTORS,0);
3325 clear_bit(STRIPE_SYNCING, &sh->state);
3326 s.syncing = 0;
3330 * might be able to return some write requests if the parity blocks
3331 * are safe, or on a failed drive
3333 pdev = &sh->dev[pd_idx];
3334 r6s.p_failed = (s.failed >= 1 && r6s.failed_num[0] == pd_idx)
3335 || (s.failed >= 2 && r6s.failed_num[1] == pd_idx);
3336 qdev = &sh->dev[qd_idx];
3337 r6s.q_failed = (s.failed >= 1 && r6s.failed_num[0] == qd_idx)
3338 || (s.failed >= 2 && r6s.failed_num[1] == qd_idx);
3340 if ( s.written &&
3341 ( r6s.p_failed || ((test_bit(R5_Insync, &pdev->flags)
3342 && !test_bit(R5_LOCKED, &pdev->flags)
3343 && test_bit(R5_UPTODATE, &pdev->flags)))) &&
3344 ( r6s.q_failed || ((test_bit(R5_Insync, &qdev->flags)
3345 && !test_bit(R5_LOCKED, &qdev->flags)
3346 && test_bit(R5_UPTODATE, &qdev->flags)))))
3347 handle_stripe_clean_event(conf, sh, disks, &return_bi);
3349 /* Now we might consider reading some blocks, either to check/generate
3350 * parity, or to satisfy requests
3351 * or to load a block that is being partially written.
3353 if (s.to_read || s.non_overwrite || (s.to_write && s.failed) ||
3354 (s.syncing && (s.uptodate + s.compute < disks)) || s.expanding)
3355 handle_stripe_fill6(sh, &s, &r6s, disks);
3357 /* Now we check to see if any write operations have recently
3358 * completed
3360 if (sh->reconstruct_state == reconstruct_state_drain_result) {
3362 sh->reconstruct_state = reconstruct_state_idle;
3363 /* All the 'written' buffers and the parity blocks are ready to
3364 * be written back to disk
3366 BUG_ON(!test_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags));
3367 BUG_ON(!test_bit(R5_UPTODATE, &sh->dev[qd_idx].flags));
3368 for (i = disks; i--; ) {
3369 dev = &sh->dev[i];
3370 if (test_bit(R5_LOCKED, &dev->flags) &&
3371 (i == sh->pd_idx || i == qd_idx ||
3372 dev->written)) {
3373 pr_debug("Writing block %d\n", i);
3374 BUG_ON(!test_bit(R5_UPTODATE, &dev->flags));
3375 set_bit(R5_Wantwrite, &dev->flags);
3376 if (!test_bit(R5_Insync, &dev->flags) ||
3377 ((i == sh->pd_idx || i == qd_idx) &&
3378 s.failed == 0))
3379 set_bit(STRIPE_INSYNC, &sh->state);
3382 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
3383 atomic_dec(&conf->preread_active_stripes);
3384 if (atomic_read(&conf->preread_active_stripes) <
3385 IO_THRESHOLD)
3386 md_wakeup_thread(conf->mddev->thread);
3390 /* Now to consider new write requests and what else, if anything
3391 * should be read. We do not handle new writes when:
3392 * 1/ A 'write' operation (copy+gen_syndrome) is already in flight.
3393 * 2/ A 'check' operation is in flight, as it may clobber the parity
3394 * block.
3396 if (s.to_write && !sh->reconstruct_state && !sh->check_state)
3397 handle_stripe_dirtying6(conf, sh, &s, &r6s, disks);
3399 /* maybe we need to check and possibly fix the parity for this stripe
3400 * Any reads will already have been scheduled, so we just see if enough
3401 * data is available. The parity check is held off while parity
3402 * dependent operations are in flight.
3404 if (sh->check_state ||
3405 (s.syncing && s.locked == 0 &&
3406 !test_bit(STRIPE_COMPUTE_RUN, &sh->state) &&
3407 !test_bit(STRIPE_INSYNC, &sh->state)))
3408 handle_parity_checks6(conf, sh, &s, &r6s, disks);
3410 if (s.syncing && s.locked == 0 && test_bit(STRIPE_INSYNC, &sh->state)) {
3411 md_done_sync(conf->mddev, STRIPE_SECTORS,1);
3412 clear_bit(STRIPE_SYNCING, &sh->state);
3415 /* If the failed drives are just a ReadError, then we might need
3416 * to progress the repair/check process
3418 if (s.failed <= 2 && !conf->mddev->ro)
3419 for (i = 0; i < s.failed; i++) {
3420 dev = &sh->dev[r6s.failed_num[i]];
3421 if (test_bit(R5_ReadError, &dev->flags)
3422 && !test_bit(R5_LOCKED, &dev->flags)
3423 && test_bit(R5_UPTODATE, &dev->flags)
3425 if (!test_bit(R5_ReWrite, &dev->flags)) {
3426 set_bit(R5_Wantwrite, &dev->flags);
3427 set_bit(R5_ReWrite, &dev->flags);
3428 set_bit(R5_LOCKED, &dev->flags);
3429 s.locked++;
3430 } else {
3431 /* let's read it back */
3432 set_bit(R5_Wantread, &dev->flags);
3433 set_bit(R5_LOCKED, &dev->flags);
3434 s.locked++;
3439 /* Finish reconstruct operations initiated by the expansion process */
3440 if (sh->reconstruct_state == reconstruct_state_result) {
3441 sh->reconstruct_state = reconstruct_state_idle;
3442 clear_bit(STRIPE_EXPANDING, &sh->state);
3443 for (i = conf->raid_disks; i--; ) {
3444 set_bit(R5_Wantwrite, &sh->dev[i].flags);
3445 set_bit(R5_LOCKED, &sh->dev[i].flags);
3446 s.locked++;
3450 if (s.expanded && test_bit(STRIPE_EXPANDING, &sh->state) &&
3451 !sh->reconstruct_state) {
3452 struct stripe_head *sh2
3453 = get_active_stripe(conf, sh->sector, 1, 1, 1);
3454 if (sh2 && test_bit(STRIPE_EXPAND_SOURCE, &sh2->state)) {
3455 /* sh cannot be written until sh2 has been read.
3456 * so arrange for sh to be delayed a little
3458 set_bit(STRIPE_DELAYED, &sh->state);
3459 set_bit(STRIPE_HANDLE, &sh->state);
3460 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE,
3461 &sh2->state))
3462 atomic_inc(&conf->preread_active_stripes);
3463 release_stripe(sh2);
3464 goto unlock;
3466 if (sh2)
3467 release_stripe(sh2);
3469 /* Need to write out all blocks after computing P&Q */
3470 sh->disks = conf->raid_disks;
3471 stripe_set_idx(sh->sector, conf, 0, sh);
3472 schedule_reconstruction(sh, &s, 1, 1);
3473 } else if (s.expanded && !sh->reconstruct_state && s.locked == 0) {
3474 clear_bit(STRIPE_EXPAND_READY, &sh->state);
3475 atomic_dec(&conf->reshape_stripes);
3476 wake_up(&conf->wait_for_overlap);
3477 md_done_sync(conf->mddev, STRIPE_SECTORS, 1);
3480 if (s.expanding && s.locked == 0 &&
3481 !test_bit(STRIPE_COMPUTE_RUN, &sh->state))
3482 handle_stripe_expansion(conf, sh, &r6s);
3484 unlock:
3485 spin_unlock(&sh->lock);
3487 /* wait for this device to become unblocked */
3488 if (unlikely(blocked_rdev))
3489 md_wait_for_blocked_rdev(blocked_rdev, conf->mddev);
3491 if (s.ops_request)
3492 raid_run_ops(sh, s.ops_request);
3494 ops_run_io(sh, &s);
3496 return_io(return_bi);
3499 static void handle_stripe(struct stripe_head *sh)
3501 if (sh->raid_conf->level == 6)
3502 handle_stripe6(sh);
3503 else
3504 handle_stripe5(sh);
3507 static void raid5_activate_delayed(raid5_conf_t *conf)
3509 if (atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD) {
3510 while (!list_empty(&conf->delayed_list)) {
3511 struct list_head *l = conf->delayed_list.next;
3512 struct stripe_head *sh;
3513 sh = list_entry(l, struct stripe_head, lru);
3514 list_del_init(l);
3515 clear_bit(STRIPE_DELAYED, &sh->state);
3516 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
3517 atomic_inc(&conf->preread_active_stripes);
3518 list_add_tail(&sh->lru, &conf->hold_list);
3520 } else
3521 blk_plug_device(conf->mddev->queue);
3524 static void activate_bit_delay(raid5_conf_t *conf)
3526 /* device_lock is held */
3527 struct list_head head;
3528 list_add(&head, &conf->bitmap_list);
3529 list_del_init(&conf->bitmap_list);
3530 while (!list_empty(&head)) {
3531 struct stripe_head *sh = list_entry(head.next, struct stripe_head, lru);
3532 list_del_init(&sh->lru);
3533 atomic_inc(&sh->count);
3534 __release_stripe(conf, sh);
3538 static void unplug_slaves(mddev_t *mddev)
3540 raid5_conf_t *conf = mddev->private;
3541 int i;
3542 int devs = max(conf->raid_disks, conf->previous_raid_disks);
3544 rcu_read_lock();
3545 for (i = 0; i < devs; i++) {
3546 mdk_rdev_t *rdev = rcu_dereference(conf->disks[i].rdev);
3547 if (rdev && !test_bit(Faulty, &rdev->flags) && atomic_read(&rdev->nr_pending)) {
3548 struct request_queue *r_queue = bdev_get_queue(rdev->bdev);
3550 atomic_inc(&rdev->nr_pending);
3551 rcu_read_unlock();
3553 blk_unplug(r_queue);
3555 rdev_dec_pending(rdev, mddev);
3556 rcu_read_lock();
3559 rcu_read_unlock();
3562 static void raid5_unplug_device(struct request_queue *q)
3564 mddev_t *mddev = q->queuedata;
3565 raid5_conf_t *conf = mddev->private;
3566 unsigned long flags;
3568 spin_lock_irqsave(&conf->device_lock, flags);
3570 if (blk_remove_plug(q)) {
3571 conf->seq_flush++;
3572 raid5_activate_delayed(conf);
3574 md_wakeup_thread(mddev->thread);
3576 spin_unlock_irqrestore(&conf->device_lock, flags);
3578 unplug_slaves(mddev);
3581 static int raid5_congested(void *data, int bits)
3583 mddev_t *mddev = data;
3584 raid5_conf_t *conf = mddev->private;
3586 /* No difference between reads and writes. Just check
3587 * how busy the stripe_cache is
3590 if (mddev_congested(mddev, bits))
3591 return 1;
3592 if (conf->inactive_blocked)
3593 return 1;
3594 if (conf->quiesce)
3595 return 1;
3596 if (list_empty_careful(&conf->inactive_list))
3597 return 1;
3599 return 0;
3602 /* We want read requests to align with chunks where possible,
3603 * but write requests don't need to.
3605 static int raid5_mergeable_bvec(struct request_queue *q,
3606 struct bvec_merge_data *bvm,
3607 struct bio_vec *biovec)
3609 mddev_t *mddev = q->queuedata;
3610 sector_t sector = bvm->bi_sector + get_start_sect(bvm->bi_bdev);
3611 int max;
3612 unsigned int chunk_sectors = mddev->chunk_sectors;
3613 unsigned int bio_sectors = bvm->bi_size >> 9;
3615 if ((bvm->bi_rw & 1) == WRITE)
3616 return biovec->bv_len; /* always allow writes to be mergeable */
3618 if (mddev->new_chunk_sectors < mddev->chunk_sectors)
3619 chunk_sectors = mddev->new_chunk_sectors;
3620 max = (chunk_sectors - ((sector & (chunk_sectors - 1)) + bio_sectors)) << 9;
3621 if (max < 0) max = 0;
3622 if (max <= biovec->bv_len && bio_sectors == 0)
3623 return biovec->bv_len;
3624 else
3625 return max;
3629 static int in_chunk_boundary(mddev_t *mddev, struct bio *bio)
3631 sector_t sector = bio->bi_sector + get_start_sect(bio->bi_bdev);
3632 unsigned int chunk_sectors = mddev->chunk_sectors;
3633 unsigned int bio_sectors = bio->bi_size >> 9;
3635 if (mddev->new_chunk_sectors < mddev->chunk_sectors)
3636 chunk_sectors = mddev->new_chunk_sectors;
3637 return chunk_sectors >=
3638 ((sector & (chunk_sectors - 1)) + bio_sectors);
3642 * add bio to the retry LIFO ( in O(1) ... we are in interrupt )
3643 * later sampled by raid5d.
3645 static void add_bio_to_retry(struct bio *bi,raid5_conf_t *conf)
3647 unsigned long flags;
3649 spin_lock_irqsave(&conf->device_lock, flags);
3651 bi->bi_next = conf->retry_read_aligned_list;
3652 conf->retry_read_aligned_list = bi;
3654 spin_unlock_irqrestore(&conf->device_lock, flags);
3655 md_wakeup_thread(conf->mddev->thread);
3659 static struct bio *remove_bio_from_retry(raid5_conf_t *conf)
3661 struct bio *bi;
3663 bi = conf->retry_read_aligned;
3664 if (bi) {
3665 conf->retry_read_aligned = NULL;
3666 return bi;
3668 bi = conf->retry_read_aligned_list;
3669 if(bi) {
3670 conf->retry_read_aligned_list = bi->bi_next;
3671 bi->bi_next = NULL;
3673 * this sets the active strip count to 1 and the processed
3674 * strip count to zero (upper 8 bits)
3676 bi->bi_phys_segments = 1; /* biased count of active stripes */
3679 return bi;
3684 * The "raid5_align_endio" should check if the read succeeded and if it
3685 * did, call bio_endio on the original bio (having bio_put the new bio
3686 * first).
3687 * If the read failed..
3689 static void raid5_align_endio(struct bio *bi, int error)
3691 struct bio* raid_bi = bi->bi_private;
3692 mddev_t *mddev;
3693 raid5_conf_t *conf;
3694 int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
3695 mdk_rdev_t *rdev;
3697 bio_put(bi);
3699 mddev = raid_bi->bi_bdev->bd_disk->queue->queuedata;
3700 conf = mddev->private;
3701 rdev = (void*)raid_bi->bi_next;
3702 raid_bi->bi_next = NULL;
3704 rdev_dec_pending(rdev, conf->mddev);
3706 if (!error && uptodate) {
3707 bio_endio(raid_bi, 0);
3708 if (atomic_dec_and_test(&conf->active_aligned_reads))
3709 wake_up(&conf->wait_for_stripe);
3710 return;
3714 pr_debug("raid5_align_endio : io error...handing IO for a retry\n");
3716 add_bio_to_retry(raid_bi, conf);
3719 static int bio_fits_rdev(struct bio *bi)
3721 struct request_queue *q = bdev_get_queue(bi->bi_bdev);
3723 if ((bi->bi_size>>9) > queue_max_sectors(q))
3724 return 0;
3725 blk_recount_segments(q, bi);
3726 if (bi->bi_phys_segments > queue_max_phys_segments(q))
3727 return 0;
3729 if (q->merge_bvec_fn)
3730 /* it's too hard to apply the merge_bvec_fn at this stage,
3731 * just just give up
3733 return 0;
3735 return 1;
3739 static int chunk_aligned_read(struct request_queue *q, struct bio * raid_bio)
3741 mddev_t *mddev = q->queuedata;
3742 raid5_conf_t *conf = mddev->private;
3743 int dd_idx;
3744 struct bio* align_bi;
3745 mdk_rdev_t *rdev;
3747 if (!in_chunk_boundary(mddev, raid_bio)) {
3748 pr_debug("chunk_aligned_read : non aligned\n");
3749 return 0;
3752 * use bio_clone to make a copy of the bio
3754 align_bi = bio_clone(raid_bio, GFP_NOIO);
3755 if (!align_bi)
3756 return 0;
3758 * set bi_end_io to a new function, and set bi_private to the
3759 * original bio.
3761 align_bi->bi_end_io = raid5_align_endio;
3762 align_bi->bi_private = raid_bio;
3764 * compute position
3766 align_bi->bi_sector = raid5_compute_sector(conf, raid_bio->bi_sector,
3768 &dd_idx, NULL);
3770 rcu_read_lock();
3771 rdev = rcu_dereference(conf->disks[dd_idx].rdev);
3772 if (rdev && test_bit(In_sync, &rdev->flags)) {
3773 atomic_inc(&rdev->nr_pending);
3774 rcu_read_unlock();
3775 raid_bio->bi_next = (void*)rdev;
3776 align_bi->bi_bdev = rdev->bdev;
3777 align_bi->bi_flags &= ~(1 << BIO_SEG_VALID);
3778 align_bi->bi_sector += rdev->data_offset;
3780 if (!bio_fits_rdev(align_bi)) {
3781 /* too big in some way */
3782 bio_put(align_bi);
3783 rdev_dec_pending(rdev, mddev);
3784 return 0;
3787 spin_lock_irq(&conf->device_lock);
3788 wait_event_lock_irq(conf->wait_for_stripe,
3789 conf->quiesce == 0,
3790 conf->device_lock, /* nothing */);
3791 atomic_inc(&conf->active_aligned_reads);
3792 spin_unlock_irq(&conf->device_lock);
3794 generic_make_request(align_bi);
3795 return 1;
3796 } else {
3797 rcu_read_unlock();
3798 bio_put(align_bi);
3799 return 0;
3803 /* __get_priority_stripe - get the next stripe to process
3805 * Full stripe writes are allowed to pass preread active stripes up until
3806 * the bypass_threshold is exceeded. In general the bypass_count
3807 * increments when the handle_list is handled before the hold_list; however, it
3808 * will not be incremented when STRIPE_IO_STARTED is sampled set signifying a
3809 * stripe with in flight i/o. The bypass_count will be reset when the
3810 * head of the hold_list has changed, i.e. the head was promoted to the
3811 * handle_list.
3813 static struct stripe_head *__get_priority_stripe(raid5_conf_t *conf)
3815 struct stripe_head *sh;
3817 pr_debug("%s: handle: %s hold: %s full_writes: %d bypass_count: %d\n",
3818 __func__,
3819 list_empty(&conf->handle_list) ? "empty" : "busy",
3820 list_empty(&conf->hold_list) ? "empty" : "busy",
3821 atomic_read(&conf->pending_full_writes), conf->bypass_count);
3823 if (!list_empty(&conf->handle_list)) {
3824 sh = list_entry(conf->handle_list.next, typeof(*sh), lru);
3826 if (list_empty(&conf->hold_list))
3827 conf->bypass_count = 0;
3828 else if (!test_bit(STRIPE_IO_STARTED, &sh->state)) {
3829 if (conf->hold_list.next == conf->last_hold)
3830 conf->bypass_count++;
3831 else {
3832 conf->last_hold = conf->hold_list.next;
3833 conf->bypass_count -= conf->bypass_threshold;
3834 if (conf->bypass_count < 0)
3835 conf->bypass_count = 0;
3838 } else if (!list_empty(&conf->hold_list) &&
3839 ((conf->bypass_threshold &&
3840 conf->bypass_count > conf->bypass_threshold) ||
3841 atomic_read(&conf->pending_full_writes) == 0)) {
3842 sh = list_entry(conf->hold_list.next,
3843 typeof(*sh), lru);
3844 conf->bypass_count -= conf->bypass_threshold;
3845 if (conf->bypass_count < 0)
3846 conf->bypass_count = 0;
3847 } else
3848 return NULL;
3850 list_del_init(&sh->lru);
3851 atomic_inc(&sh->count);
3852 BUG_ON(atomic_read(&sh->count) != 1);
3853 return sh;
3856 static int make_request(struct request_queue *q, struct bio * bi)
3858 mddev_t *mddev = q->queuedata;
3859 raid5_conf_t *conf = mddev->private;
3860 int dd_idx;
3861 sector_t new_sector;
3862 sector_t logical_sector, last_sector;
3863 struct stripe_head *sh;
3864 const int rw = bio_data_dir(bi);
3865 int cpu, remaining;
3867 if (unlikely(bio_rw_flagged(bi, BIO_RW_BARRIER))) {
3868 bio_endio(bi, -EOPNOTSUPP);
3869 return 0;
3872 md_write_start(mddev, bi);
3874 cpu = part_stat_lock();
3875 part_stat_inc(cpu, &mddev->gendisk->part0, ios[rw]);
3876 part_stat_add(cpu, &mddev->gendisk->part0, sectors[rw],
3877 bio_sectors(bi));
3878 part_stat_unlock();
3880 if (rw == READ &&
3881 mddev->reshape_position == MaxSector &&
3882 chunk_aligned_read(q,bi))
3883 return 0;
3885 logical_sector = bi->bi_sector & ~((sector_t)STRIPE_SECTORS-1);
3886 last_sector = bi->bi_sector + (bi->bi_size>>9);
3887 bi->bi_next = NULL;
3888 bi->bi_phys_segments = 1; /* over-loaded to count active stripes */
3890 for (;logical_sector < last_sector; logical_sector += STRIPE_SECTORS) {
3891 DEFINE_WAIT(w);
3892 int disks, data_disks;
3893 int previous;
3895 retry:
3896 previous = 0;
3897 disks = conf->raid_disks;
3898 prepare_to_wait(&conf->wait_for_overlap, &w, TASK_UNINTERRUPTIBLE);
3899 if (unlikely(conf->reshape_progress != MaxSector)) {
3900 /* spinlock is needed as reshape_progress may be
3901 * 64bit on a 32bit platform, and so it might be
3902 * possible to see a half-updated value
3903 * Ofcourse reshape_progress could change after
3904 * the lock is dropped, so once we get a reference
3905 * to the stripe that we think it is, we will have
3906 * to check again.
3908 spin_lock_irq(&conf->device_lock);
3909 if (mddev->delta_disks < 0
3910 ? logical_sector < conf->reshape_progress
3911 : logical_sector >= conf->reshape_progress) {
3912 disks = conf->previous_raid_disks;
3913 previous = 1;
3914 } else {
3915 if (mddev->delta_disks < 0
3916 ? logical_sector < conf->reshape_safe
3917 : logical_sector >= conf->reshape_safe) {
3918 spin_unlock_irq(&conf->device_lock);
3919 schedule();
3920 goto retry;
3923 spin_unlock_irq(&conf->device_lock);
3925 data_disks = disks - conf->max_degraded;
3927 new_sector = raid5_compute_sector(conf, logical_sector,
3928 previous,
3929 &dd_idx, NULL);
3930 pr_debug("raid5: make_request, sector %llu logical %llu\n",
3931 (unsigned long long)new_sector,
3932 (unsigned long long)logical_sector);
3934 sh = get_active_stripe(conf, new_sector, previous,
3935 (bi->bi_rw&RWA_MASK), 0);
3936 if (sh) {
3937 if (unlikely(previous)) {
3938 /* expansion might have moved on while waiting for a
3939 * stripe, so we must do the range check again.
3940 * Expansion could still move past after this
3941 * test, but as we are holding a reference to
3942 * 'sh', we know that if that happens,
3943 * STRIPE_EXPANDING will get set and the expansion
3944 * won't proceed until we finish with the stripe.
3946 int must_retry = 0;
3947 spin_lock_irq(&conf->device_lock);
3948 if (mddev->delta_disks < 0
3949 ? logical_sector >= conf->reshape_progress
3950 : logical_sector < conf->reshape_progress)
3951 /* mismatch, need to try again */
3952 must_retry = 1;
3953 spin_unlock_irq(&conf->device_lock);
3954 if (must_retry) {
3955 release_stripe(sh);
3956 schedule();
3957 goto retry;
3961 if (bio_data_dir(bi) == WRITE &&
3962 logical_sector >= mddev->suspend_lo &&
3963 logical_sector < mddev->suspend_hi) {
3964 release_stripe(sh);
3965 /* As the suspend_* range is controlled by
3966 * userspace, we want an interruptible
3967 * wait.
3969 flush_signals(current);
3970 prepare_to_wait(&conf->wait_for_overlap,
3971 &w, TASK_INTERRUPTIBLE);
3972 if (logical_sector >= mddev->suspend_lo &&
3973 logical_sector < mddev->suspend_hi)
3974 schedule();
3975 goto retry;
3978 if (test_bit(STRIPE_EXPANDING, &sh->state) ||
3979 !add_stripe_bio(sh, bi, dd_idx, (bi->bi_rw&RW_MASK))) {
3980 /* Stripe is busy expanding or
3981 * add failed due to overlap. Flush everything
3982 * and wait a while
3984 raid5_unplug_device(mddev->queue);
3985 release_stripe(sh);
3986 schedule();
3987 goto retry;
3989 finish_wait(&conf->wait_for_overlap, &w);
3990 set_bit(STRIPE_HANDLE, &sh->state);
3991 clear_bit(STRIPE_DELAYED, &sh->state);
3992 release_stripe(sh);
3993 } else {
3994 /* cannot get stripe for read-ahead, just give-up */
3995 clear_bit(BIO_UPTODATE, &bi->bi_flags);
3996 finish_wait(&conf->wait_for_overlap, &w);
3997 break;
4001 spin_lock_irq(&conf->device_lock);
4002 remaining = raid5_dec_bi_phys_segments(bi);
4003 spin_unlock_irq(&conf->device_lock);
4004 if (remaining == 0) {
4006 if ( rw == WRITE )
4007 md_write_end(mddev);
4009 bio_endio(bi, 0);
4011 return 0;
4014 static sector_t raid5_size(mddev_t *mddev, sector_t sectors, int raid_disks);
4016 static sector_t reshape_request(mddev_t *mddev, sector_t sector_nr, int *skipped)
4018 /* reshaping is quite different to recovery/resync so it is
4019 * handled quite separately ... here.
4021 * On each call to sync_request, we gather one chunk worth of
4022 * destination stripes and flag them as expanding.
4023 * Then we find all the source stripes and request reads.
4024 * As the reads complete, handle_stripe will copy the data
4025 * into the destination stripe and release that stripe.
4027 raid5_conf_t *conf = (raid5_conf_t *) mddev->private;
4028 struct stripe_head *sh;
4029 sector_t first_sector, last_sector;
4030 int raid_disks = conf->previous_raid_disks;
4031 int data_disks = raid_disks - conf->max_degraded;
4032 int new_data_disks = conf->raid_disks - conf->max_degraded;
4033 int i;
4034 int dd_idx;
4035 sector_t writepos, readpos, safepos;
4036 sector_t stripe_addr;
4037 int reshape_sectors;
4038 struct list_head stripes;
4040 if (sector_nr == 0) {
4041 /* If restarting in the middle, skip the initial sectors */
4042 if (mddev->delta_disks < 0 &&
4043 conf->reshape_progress < raid5_size(mddev, 0, 0)) {
4044 sector_nr = raid5_size(mddev, 0, 0)
4045 - conf->reshape_progress;
4046 } else if (mddev->delta_disks >= 0 &&
4047 conf->reshape_progress > 0)
4048 sector_nr = conf->reshape_progress;
4049 sector_div(sector_nr, new_data_disks);
4050 if (sector_nr) {
4051 mddev->curr_resync_completed = sector_nr;
4052 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
4053 *skipped = 1;
4054 return sector_nr;
4058 /* We need to process a full chunk at a time.
4059 * If old and new chunk sizes differ, we need to process the
4060 * largest of these
4062 if (mddev->new_chunk_sectors > mddev->chunk_sectors)
4063 reshape_sectors = mddev->new_chunk_sectors;
4064 else
4065 reshape_sectors = mddev->chunk_sectors;
4067 /* we update the metadata when there is more than 3Meg
4068 * in the block range (that is rather arbitrary, should
4069 * probably be time based) or when the data about to be
4070 * copied would over-write the source of the data at
4071 * the front of the range.
4072 * i.e. one new_stripe along from reshape_progress new_maps
4073 * to after where reshape_safe old_maps to
4075 writepos = conf->reshape_progress;
4076 sector_div(writepos, new_data_disks);
4077 readpos = conf->reshape_progress;
4078 sector_div(readpos, data_disks);
4079 safepos = conf->reshape_safe;
4080 sector_div(safepos, data_disks);
4081 if (mddev->delta_disks < 0) {
4082 writepos -= min_t(sector_t, reshape_sectors, writepos);
4083 readpos += reshape_sectors;
4084 safepos += reshape_sectors;
4085 } else {
4086 writepos += reshape_sectors;
4087 readpos -= min_t(sector_t, reshape_sectors, readpos);
4088 safepos -= min_t(sector_t, reshape_sectors, safepos);
4091 /* 'writepos' is the most advanced device address we might write.
4092 * 'readpos' is the least advanced device address we might read.
4093 * 'safepos' is the least address recorded in the metadata as having
4094 * been reshaped.
4095 * If 'readpos' is behind 'writepos', then there is no way that we can
4096 * ensure safety in the face of a crash - that must be done by userspace
4097 * making a backup of the data. So in that case there is no particular
4098 * rush to update metadata.
4099 * Otherwise if 'safepos' is behind 'writepos', then we really need to
4100 * update the metadata to advance 'safepos' to match 'readpos' so that
4101 * we can be safe in the event of a crash.
4102 * So we insist on updating metadata if safepos is behind writepos and
4103 * readpos is beyond writepos.
4104 * In any case, update the metadata every 10 seconds.
4105 * Maybe that number should be configurable, but I'm not sure it is
4106 * worth it.... maybe it could be a multiple of safemode_delay???
4108 if ((mddev->delta_disks < 0
4109 ? (safepos > writepos && readpos < writepos)
4110 : (safepos < writepos && readpos > writepos)) ||
4111 time_after(jiffies, conf->reshape_checkpoint + 10*HZ)) {
4112 /* Cannot proceed until we've updated the superblock... */
4113 wait_event(conf->wait_for_overlap,
4114 atomic_read(&conf->reshape_stripes)==0);
4115 mddev->reshape_position = conf->reshape_progress;
4116 mddev->curr_resync_completed = mddev->curr_resync;
4117 conf->reshape_checkpoint = jiffies;
4118 set_bit(MD_CHANGE_DEVS, &mddev->flags);
4119 md_wakeup_thread(mddev->thread);
4120 wait_event(mddev->sb_wait, mddev->flags == 0 ||
4121 kthread_should_stop());
4122 spin_lock_irq(&conf->device_lock);
4123 conf->reshape_safe = mddev->reshape_position;
4124 spin_unlock_irq(&conf->device_lock);
4125 wake_up(&conf->wait_for_overlap);
4126 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
4129 if (mddev->delta_disks < 0) {
4130 BUG_ON(conf->reshape_progress == 0);
4131 stripe_addr = writepos;
4132 BUG_ON((mddev->dev_sectors &
4133 ~((sector_t)reshape_sectors - 1))
4134 - reshape_sectors - stripe_addr
4135 != sector_nr);
4136 } else {
4137 BUG_ON(writepos != sector_nr + reshape_sectors);
4138 stripe_addr = sector_nr;
4140 INIT_LIST_HEAD(&stripes);
4141 for (i = 0; i < reshape_sectors; i += STRIPE_SECTORS) {
4142 int j;
4143 int skipped_disk = 0;
4144 sh = get_active_stripe(conf, stripe_addr+i, 0, 0, 1);
4145 set_bit(STRIPE_EXPANDING, &sh->state);
4146 atomic_inc(&conf->reshape_stripes);
4147 /* If any of this stripe is beyond the end of the old
4148 * array, then we need to zero those blocks
4150 for (j=sh->disks; j--;) {
4151 sector_t s;
4152 if (j == sh->pd_idx)
4153 continue;
4154 if (conf->level == 6 &&
4155 j == sh->qd_idx)
4156 continue;
4157 s = compute_blocknr(sh, j, 0);
4158 if (s < raid5_size(mddev, 0, 0)) {
4159 skipped_disk = 1;
4160 continue;
4162 memset(page_address(sh->dev[j].page), 0, STRIPE_SIZE);
4163 set_bit(R5_Expanded, &sh->dev[j].flags);
4164 set_bit(R5_UPTODATE, &sh->dev[j].flags);
4166 if (!skipped_disk) {
4167 set_bit(STRIPE_EXPAND_READY, &sh->state);
4168 set_bit(STRIPE_HANDLE, &sh->state);
4170 list_add(&sh->lru, &stripes);
4172 spin_lock_irq(&conf->device_lock);
4173 if (mddev->delta_disks < 0)
4174 conf->reshape_progress -= reshape_sectors * new_data_disks;
4175 else
4176 conf->reshape_progress += reshape_sectors * new_data_disks;
4177 spin_unlock_irq(&conf->device_lock);
4178 /* Ok, those stripe are ready. We can start scheduling
4179 * reads on the source stripes.
4180 * The source stripes are determined by mapping the first and last
4181 * block on the destination stripes.
4183 first_sector =
4184 raid5_compute_sector(conf, stripe_addr*(new_data_disks),
4185 1, &dd_idx, NULL);
4186 last_sector =
4187 raid5_compute_sector(conf, ((stripe_addr+reshape_sectors)
4188 * new_data_disks - 1),
4189 1, &dd_idx, NULL);
4190 if (last_sector >= mddev->dev_sectors)
4191 last_sector = mddev->dev_sectors - 1;
4192 while (first_sector <= last_sector) {
4193 sh = get_active_stripe(conf, first_sector, 1, 0, 1);
4194 set_bit(STRIPE_EXPAND_SOURCE, &sh->state);
4195 set_bit(STRIPE_HANDLE, &sh->state);
4196 release_stripe(sh);
4197 first_sector += STRIPE_SECTORS;
4199 /* Now that the sources are clearly marked, we can release
4200 * the destination stripes
4202 while (!list_empty(&stripes)) {
4203 sh = list_entry(stripes.next, struct stripe_head, lru);
4204 list_del_init(&sh->lru);
4205 release_stripe(sh);
4207 /* If this takes us to the resync_max point where we have to pause,
4208 * then we need to write out the superblock.
4210 sector_nr += reshape_sectors;
4211 if ((sector_nr - mddev->curr_resync_completed) * 2
4212 >= mddev->resync_max - mddev->curr_resync_completed) {
4213 /* Cannot proceed until we've updated the superblock... */
4214 wait_event(conf->wait_for_overlap,
4215 atomic_read(&conf->reshape_stripes) == 0);
4216 mddev->reshape_position = conf->reshape_progress;
4217 mddev->curr_resync_completed = mddev->curr_resync + reshape_sectors;
4218 conf->reshape_checkpoint = jiffies;
4219 set_bit(MD_CHANGE_DEVS, &mddev->flags);
4220 md_wakeup_thread(mddev->thread);
4221 wait_event(mddev->sb_wait,
4222 !test_bit(MD_CHANGE_DEVS, &mddev->flags)
4223 || kthread_should_stop());
4224 spin_lock_irq(&conf->device_lock);
4225 conf->reshape_safe = mddev->reshape_position;
4226 spin_unlock_irq(&conf->device_lock);
4227 wake_up(&conf->wait_for_overlap);
4228 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
4230 return reshape_sectors;
4233 /* FIXME go_faster isn't used */
4234 static inline sector_t sync_request(mddev_t *mddev, sector_t sector_nr, int *skipped, int go_faster)
4236 raid5_conf_t *conf = (raid5_conf_t *) mddev->private;
4237 struct stripe_head *sh;
4238 sector_t max_sector = mddev->dev_sectors;
4239 int sync_blocks;
4240 int still_degraded = 0;
4241 int i;
4243 if (sector_nr >= max_sector) {
4244 /* just being told to finish up .. nothing much to do */
4245 unplug_slaves(mddev);
4247 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery)) {
4248 end_reshape(conf);
4249 return 0;
4252 if (mddev->curr_resync < max_sector) /* aborted */
4253 bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
4254 &sync_blocks, 1);
4255 else /* completed sync */
4256 conf->fullsync = 0;
4257 bitmap_close_sync(mddev->bitmap);
4259 return 0;
4262 /* Allow raid5_quiesce to complete */
4263 wait_event(conf->wait_for_overlap, conf->quiesce != 2);
4265 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
4266 return reshape_request(mddev, sector_nr, skipped);
4268 /* No need to check resync_max as we never do more than one
4269 * stripe, and as resync_max will always be on a chunk boundary,
4270 * if the check in md_do_sync didn't fire, there is no chance
4271 * of overstepping resync_max here
4274 /* if there is too many failed drives and we are trying
4275 * to resync, then assert that we are finished, because there is
4276 * nothing we can do.
4278 if (mddev->degraded >= conf->max_degraded &&
4279 test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
4280 sector_t rv = mddev->dev_sectors - sector_nr;
4281 *skipped = 1;
4282 return rv;
4284 if (!bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
4285 !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
4286 !conf->fullsync && sync_blocks >= STRIPE_SECTORS) {
4287 /* we can skip this block, and probably more */
4288 sync_blocks /= STRIPE_SECTORS;
4289 *skipped = 1;
4290 return sync_blocks * STRIPE_SECTORS; /* keep things rounded to whole stripes */
4294 bitmap_cond_end_sync(mddev->bitmap, sector_nr);
4296 sh = get_active_stripe(conf, sector_nr, 0, 1, 0);
4297 if (sh == NULL) {
4298 sh = get_active_stripe(conf, sector_nr, 0, 0, 0);
4299 /* make sure we don't swamp the stripe cache if someone else
4300 * is trying to get access
4302 schedule_timeout_uninterruptible(1);
4304 /* Need to check if array will still be degraded after recovery/resync
4305 * We don't need to check the 'failed' flag as when that gets set,
4306 * recovery aborts.
4308 for (i = 0; i < conf->raid_disks; i++)
4309 if (conf->disks[i].rdev == NULL)
4310 still_degraded = 1;
4312 bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, still_degraded);
4314 spin_lock(&sh->lock);
4315 set_bit(STRIPE_SYNCING, &sh->state);
4316 clear_bit(STRIPE_INSYNC, &sh->state);
4317 spin_unlock(&sh->lock);
4319 handle_stripe(sh);
4320 release_stripe(sh);
4322 return STRIPE_SECTORS;
4325 static int retry_aligned_read(raid5_conf_t *conf, struct bio *raid_bio)
4327 /* We may not be able to submit a whole bio at once as there
4328 * may not be enough stripe_heads available.
4329 * We cannot pre-allocate enough stripe_heads as we may need
4330 * more than exist in the cache (if we allow ever large chunks).
4331 * So we do one stripe head at a time and record in
4332 * ->bi_hw_segments how many have been done.
4334 * We *know* that this entire raid_bio is in one chunk, so
4335 * it will be only one 'dd_idx' and only need one call to raid5_compute_sector.
4337 struct stripe_head *sh;
4338 int dd_idx;
4339 sector_t sector, logical_sector, last_sector;
4340 int scnt = 0;
4341 int remaining;
4342 int handled = 0;
4344 logical_sector = raid_bio->bi_sector & ~((sector_t)STRIPE_SECTORS-1);
4345 sector = raid5_compute_sector(conf, logical_sector,
4346 0, &dd_idx, NULL);
4347 last_sector = raid_bio->bi_sector + (raid_bio->bi_size>>9);
4349 for (; logical_sector < last_sector;
4350 logical_sector += STRIPE_SECTORS,
4351 sector += STRIPE_SECTORS,
4352 scnt++) {
4354 if (scnt < raid5_bi_hw_segments(raid_bio))
4355 /* already done this stripe */
4356 continue;
4358 sh = get_active_stripe(conf, sector, 0, 1, 0);
4360 if (!sh) {
4361 /* failed to get a stripe - must wait */
4362 raid5_set_bi_hw_segments(raid_bio, scnt);
4363 conf->retry_read_aligned = raid_bio;
4364 return handled;
4367 set_bit(R5_ReadError, &sh->dev[dd_idx].flags);
4368 if (!add_stripe_bio(sh, raid_bio, dd_idx, 0)) {
4369 release_stripe(sh);
4370 raid5_set_bi_hw_segments(raid_bio, scnt);
4371 conf->retry_read_aligned = raid_bio;
4372 return handled;
4375 handle_stripe(sh);
4376 release_stripe(sh);
4377 handled++;
4379 spin_lock_irq(&conf->device_lock);
4380 remaining = raid5_dec_bi_phys_segments(raid_bio);
4381 spin_unlock_irq(&conf->device_lock);
4382 if (remaining == 0)
4383 bio_endio(raid_bio, 0);
4384 if (atomic_dec_and_test(&conf->active_aligned_reads))
4385 wake_up(&conf->wait_for_stripe);
4386 return handled;
4391 * This is our raid5 kernel thread.
4393 * We scan the hash table for stripes which can be handled now.
4394 * During the scan, completed stripes are saved for us by the interrupt
4395 * handler, so that they will not have to wait for our next wakeup.
4397 static void raid5d(mddev_t *mddev)
4399 struct stripe_head *sh;
4400 raid5_conf_t *conf = mddev->private;
4401 int handled;
4403 pr_debug("+++ raid5d active\n");
4405 md_check_recovery(mddev);
4407 handled = 0;
4408 spin_lock_irq(&conf->device_lock);
4409 while (1) {
4410 struct bio *bio;
4412 if (conf->seq_flush != conf->seq_write) {
4413 int seq = conf->seq_flush;
4414 spin_unlock_irq(&conf->device_lock);
4415 bitmap_unplug(mddev->bitmap);
4416 spin_lock_irq(&conf->device_lock);
4417 conf->seq_write = seq;
4418 activate_bit_delay(conf);
4421 while ((bio = remove_bio_from_retry(conf))) {
4422 int ok;
4423 spin_unlock_irq(&conf->device_lock);
4424 ok = retry_aligned_read(conf, bio);
4425 spin_lock_irq(&conf->device_lock);
4426 if (!ok)
4427 break;
4428 handled++;
4431 sh = __get_priority_stripe(conf);
4433 if (!sh)
4434 break;
4435 spin_unlock_irq(&conf->device_lock);
4437 handled++;
4438 handle_stripe(sh);
4439 release_stripe(sh);
4440 cond_resched();
4442 spin_lock_irq(&conf->device_lock);
4444 pr_debug("%d stripes handled\n", handled);
4446 spin_unlock_irq(&conf->device_lock);
4448 async_tx_issue_pending_all();
4449 unplug_slaves(mddev);
4451 pr_debug("--- raid5d inactive\n");
4454 static ssize_t
4455 raid5_show_stripe_cache_size(mddev_t *mddev, char *page)
4457 raid5_conf_t *conf = mddev->private;
4458 if (conf)
4459 return sprintf(page, "%d\n", conf->max_nr_stripes);
4460 else
4461 return 0;
4464 static ssize_t
4465 raid5_store_stripe_cache_size(mddev_t *mddev, const char *page, size_t len)
4467 raid5_conf_t *conf = mddev->private;
4468 unsigned long new;
4469 int err;
4471 if (len >= PAGE_SIZE)
4472 return -EINVAL;
4473 if (!conf)
4474 return -ENODEV;
4476 if (strict_strtoul(page, 10, &new))
4477 return -EINVAL;
4478 if (new <= 16 || new > 32768)
4479 return -EINVAL;
4480 while (new < conf->max_nr_stripes) {
4481 if (drop_one_stripe(conf))
4482 conf->max_nr_stripes--;
4483 else
4484 break;
4486 err = md_allow_write(mddev);
4487 if (err)
4488 return err;
4489 while (new > conf->max_nr_stripes) {
4490 if (grow_one_stripe(conf))
4491 conf->max_nr_stripes++;
4492 else break;
4494 return len;
4497 static struct md_sysfs_entry
4498 raid5_stripecache_size = __ATTR(stripe_cache_size, S_IRUGO | S_IWUSR,
4499 raid5_show_stripe_cache_size,
4500 raid5_store_stripe_cache_size);
4502 static ssize_t
4503 raid5_show_preread_threshold(mddev_t *mddev, char *page)
4505 raid5_conf_t *conf = mddev->private;
4506 if (conf)
4507 return sprintf(page, "%d\n", conf->bypass_threshold);
4508 else
4509 return 0;
4512 static ssize_t
4513 raid5_store_preread_threshold(mddev_t *mddev, const char *page, size_t len)
4515 raid5_conf_t *conf = mddev->private;
4516 unsigned long new;
4517 if (len >= PAGE_SIZE)
4518 return -EINVAL;
4519 if (!conf)
4520 return -ENODEV;
4522 if (strict_strtoul(page, 10, &new))
4523 return -EINVAL;
4524 if (new > conf->max_nr_stripes)
4525 return -EINVAL;
4526 conf->bypass_threshold = new;
4527 return len;
4530 static struct md_sysfs_entry
4531 raid5_preread_bypass_threshold = __ATTR(preread_bypass_threshold,
4532 S_IRUGO | S_IWUSR,
4533 raid5_show_preread_threshold,
4534 raid5_store_preread_threshold);
4536 static ssize_t
4537 stripe_cache_active_show(mddev_t *mddev, char *page)
4539 raid5_conf_t *conf = mddev->private;
4540 if (conf)
4541 return sprintf(page, "%d\n", atomic_read(&conf->active_stripes));
4542 else
4543 return 0;
4546 static struct md_sysfs_entry
4547 raid5_stripecache_active = __ATTR_RO(stripe_cache_active);
4549 static struct attribute *raid5_attrs[] = {
4550 &raid5_stripecache_size.attr,
4551 &raid5_stripecache_active.attr,
4552 &raid5_preread_bypass_threshold.attr,
4553 NULL,
4555 static struct attribute_group raid5_attrs_group = {
4556 .name = NULL,
4557 .attrs = raid5_attrs,
4560 static sector_t
4561 raid5_size(mddev_t *mddev, sector_t sectors, int raid_disks)
4563 raid5_conf_t *conf = mddev->private;
4565 if (!sectors)
4566 sectors = mddev->dev_sectors;
4567 if (!raid_disks)
4568 /* size is defined by the smallest of previous and new size */
4569 raid_disks = min(conf->raid_disks, conf->previous_raid_disks);
4571 sectors &= ~((sector_t)mddev->chunk_sectors - 1);
4572 sectors &= ~((sector_t)mddev->new_chunk_sectors - 1);
4573 return sectors * (raid_disks - conf->max_degraded);
4576 static void raid5_free_percpu(raid5_conf_t *conf)
4578 struct raid5_percpu *percpu;
4579 unsigned long cpu;
4581 if (!conf->percpu)
4582 return;
4584 get_online_cpus();
4585 for_each_possible_cpu(cpu) {
4586 percpu = per_cpu_ptr(conf->percpu, cpu);
4587 safe_put_page(percpu->spare_page);
4588 kfree(percpu->scribble);
4590 #ifdef CONFIG_HOTPLUG_CPU
4591 unregister_cpu_notifier(&conf->cpu_notify);
4592 #endif
4593 put_online_cpus();
4595 free_percpu(conf->percpu);
4598 static void free_conf(raid5_conf_t *conf)
4600 shrink_stripes(conf);
4601 raid5_free_percpu(conf);
4602 kfree(conf->disks);
4603 kfree(conf->stripe_hashtbl);
4604 kfree(conf);
4607 #ifdef CONFIG_HOTPLUG_CPU
4608 static int raid456_cpu_notify(struct notifier_block *nfb, unsigned long action,
4609 void *hcpu)
4611 raid5_conf_t *conf = container_of(nfb, raid5_conf_t, cpu_notify);
4612 long cpu = (long)hcpu;
4613 struct raid5_percpu *percpu = per_cpu_ptr(conf->percpu, cpu);
4615 switch (action) {
4616 case CPU_UP_PREPARE:
4617 case CPU_UP_PREPARE_FROZEN:
4618 if (conf->level == 6 && !percpu->spare_page)
4619 percpu->spare_page = alloc_page(GFP_KERNEL);
4620 if (!percpu->scribble)
4621 percpu->scribble = kmalloc(conf->scribble_len, GFP_KERNEL);
4623 if (!percpu->scribble ||
4624 (conf->level == 6 && !percpu->spare_page)) {
4625 safe_put_page(percpu->spare_page);
4626 kfree(percpu->scribble);
4627 pr_err("%s: failed memory allocation for cpu%ld\n",
4628 __func__, cpu);
4629 return NOTIFY_BAD;
4631 break;
4632 case CPU_DEAD:
4633 case CPU_DEAD_FROZEN:
4634 safe_put_page(percpu->spare_page);
4635 kfree(percpu->scribble);
4636 percpu->spare_page = NULL;
4637 percpu->scribble = NULL;
4638 break;
4639 default:
4640 break;
4642 return NOTIFY_OK;
4644 #endif
4646 static int raid5_alloc_percpu(raid5_conf_t *conf)
4648 unsigned long cpu;
4649 struct page *spare_page;
4650 struct raid5_percpu *allcpus;
4651 void *scribble;
4652 int err;
4654 allcpus = alloc_percpu(struct raid5_percpu);
4655 if (!allcpus)
4656 return -ENOMEM;
4657 conf->percpu = allcpus;
4659 get_online_cpus();
4660 err = 0;
4661 for_each_present_cpu(cpu) {
4662 if (conf->level == 6) {
4663 spare_page = alloc_page(GFP_KERNEL);
4664 if (!spare_page) {
4665 err = -ENOMEM;
4666 break;
4668 per_cpu_ptr(conf->percpu, cpu)->spare_page = spare_page;
4670 scribble = kmalloc(conf->scribble_len, GFP_KERNEL);
4671 if (!scribble) {
4672 err = -ENOMEM;
4673 break;
4675 per_cpu_ptr(conf->percpu, cpu)->scribble = scribble;
4677 #ifdef CONFIG_HOTPLUG_CPU
4678 conf->cpu_notify.notifier_call = raid456_cpu_notify;
4679 conf->cpu_notify.priority = 0;
4680 if (err == 0)
4681 err = register_cpu_notifier(&conf->cpu_notify);
4682 #endif
4683 put_online_cpus();
4685 return err;
4688 static raid5_conf_t *setup_conf(mddev_t *mddev)
4690 raid5_conf_t *conf;
4691 int raid_disk, memory, max_disks;
4692 mdk_rdev_t *rdev;
4693 struct disk_info *disk;
4695 if (mddev->new_level != 5
4696 && mddev->new_level != 4
4697 && mddev->new_level != 6) {
4698 printk(KERN_ERR "raid5: %s: raid level not set to 4/5/6 (%d)\n",
4699 mdname(mddev), mddev->new_level);
4700 return ERR_PTR(-EIO);
4702 if ((mddev->new_level == 5
4703 && !algorithm_valid_raid5(mddev->new_layout)) ||
4704 (mddev->new_level == 6
4705 && !algorithm_valid_raid6(mddev->new_layout))) {
4706 printk(KERN_ERR "raid5: %s: layout %d not supported\n",
4707 mdname(mddev), mddev->new_layout);
4708 return ERR_PTR(-EIO);
4710 if (mddev->new_level == 6 && mddev->raid_disks < 4) {
4711 printk(KERN_ERR "raid6: not enough configured devices for %s (%d, minimum 4)\n",
4712 mdname(mddev), mddev->raid_disks);
4713 return ERR_PTR(-EINVAL);
4716 if (!mddev->new_chunk_sectors ||
4717 (mddev->new_chunk_sectors << 9) % PAGE_SIZE ||
4718 !is_power_of_2(mddev->new_chunk_sectors)) {
4719 printk(KERN_ERR "raid5: invalid chunk size %d for %s\n",
4720 mddev->new_chunk_sectors << 9, mdname(mddev));
4721 return ERR_PTR(-EINVAL);
4724 conf = kzalloc(sizeof(raid5_conf_t), GFP_KERNEL);
4725 if (conf == NULL)
4726 goto abort;
4727 spin_lock_init(&conf->device_lock);
4728 init_waitqueue_head(&conf->wait_for_stripe);
4729 init_waitqueue_head(&conf->wait_for_overlap);
4730 INIT_LIST_HEAD(&conf->handle_list);
4731 INIT_LIST_HEAD(&conf->hold_list);
4732 INIT_LIST_HEAD(&conf->delayed_list);
4733 INIT_LIST_HEAD(&conf->bitmap_list);
4734 INIT_LIST_HEAD(&conf->inactive_list);
4735 atomic_set(&conf->active_stripes, 0);
4736 atomic_set(&conf->preread_active_stripes, 0);
4737 atomic_set(&conf->active_aligned_reads, 0);
4738 conf->bypass_threshold = BYPASS_THRESHOLD;
4740 conf->raid_disks = mddev->raid_disks;
4741 if (mddev->reshape_position == MaxSector)
4742 conf->previous_raid_disks = mddev->raid_disks;
4743 else
4744 conf->previous_raid_disks = mddev->raid_disks - mddev->delta_disks;
4745 max_disks = max(conf->raid_disks, conf->previous_raid_disks);
4746 conf->scribble_len = scribble_len(max_disks);
4748 conf->disks = kzalloc(max_disks * sizeof(struct disk_info),
4749 GFP_KERNEL);
4750 if (!conf->disks)
4751 goto abort;
4753 conf->mddev = mddev;
4755 if ((conf->stripe_hashtbl = kzalloc(PAGE_SIZE, GFP_KERNEL)) == NULL)
4756 goto abort;
4758 conf->level = mddev->new_level;
4759 if (raid5_alloc_percpu(conf) != 0)
4760 goto abort;
4762 pr_debug("raid5: run(%s) called.\n", mdname(mddev));
4764 list_for_each_entry(rdev, &mddev->disks, same_set) {
4765 raid_disk = rdev->raid_disk;
4766 if (raid_disk >= max_disks
4767 || raid_disk < 0)
4768 continue;
4769 disk = conf->disks + raid_disk;
4771 disk->rdev = rdev;
4773 if (test_bit(In_sync, &rdev->flags)) {
4774 char b[BDEVNAME_SIZE];
4775 printk(KERN_INFO "raid5: device %s operational as raid"
4776 " disk %d\n", bdevname(rdev->bdev,b),
4777 raid_disk);
4778 } else
4779 /* Cannot rely on bitmap to complete recovery */
4780 conf->fullsync = 1;
4783 conf->chunk_sectors = mddev->new_chunk_sectors;
4784 conf->level = mddev->new_level;
4785 if (conf->level == 6)
4786 conf->max_degraded = 2;
4787 else
4788 conf->max_degraded = 1;
4789 conf->algorithm = mddev->new_layout;
4790 conf->max_nr_stripes = NR_STRIPES;
4791 conf->reshape_progress = mddev->reshape_position;
4792 if (conf->reshape_progress != MaxSector) {
4793 conf->prev_chunk_sectors = mddev->chunk_sectors;
4794 conf->prev_algo = mddev->layout;
4797 memory = conf->max_nr_stripes * (sizeof(struct stripe_head) +
4798 max_disks * ((sizeof(struct bio) + PAGE_SIZE))) / 1024;
4799 if (grow_stripes(conf, conf->max_nr_stripes)) {
4800 printk(KERN_ERR
4801 "raid5: couldn't allocate %dkB for buffers\n", memory);
4802 goto abort;
4803 } else
4804 printk(KERN_INFO "raid5: allocated %dkB for %s\n",
4805 memory, mdname(mddev));
4807 conf->thread = md_register_thread(raid5d, mddev, NULL);
4808 if (!conf->thread) {
4809 printk(KERN_ERR
4810 "raid5: couldn't allocate thread for %s\n",
4811 mdname(mddev));
4812 goto abort;
4815 return conf;
4817 abort:
4818 if (conf) {
4819 free_conf(conf);
4820 return ERR_PTR(-EIO);
4821 } else
4822 return ERR_PTR(-ENOMEM);
4826 static int only_parity(int raid_disk, int algo, int raid_disks, int max_degraded)
4828 switch (algo) {
4829 case ALGORITHM_PARITY_0:
4830 if (raid_disk < max_degraded)
4831 return 1;
4832 break;
4833 case ALGORITHM_PARITY_N:
4834 if (raid_disk >= raid_disks - max_degraded)
4835 return 1;
4836 break;
4837 case ALGORITHM_PARITY_0_6:
4838 if (raid_disk == 0 ||
4839 raid_disk == raid_disks - 1)
4840 return 1;
4841 break;
4842 case ALGORITHM_LEFT_ASYMMETRIC_6:
4843 case ALGORITHM_RIGHT_ASYMMETRIC_6:
4844 case ALGORITHM_LEFT_SYMMETRIC_6:
4845 case ALGORITHM_RIGHT_SYMMETRIC_6:
4846 if (raid_disk == raid_disks - 1)
4847 return 1;
4849 return 0;
4852 static int run(mddev_t *mddev)
4854 raid5_conf_t *conf;
4855 int working_disks = 0, chunk_size;
4856 int dirty_parity_disks = 0;
4857 mdk_rdev_t *rdev;
4858 sector_t reshape_offset = 0;
4860 if (mddev->recovery_cp != MaxSector)
4861 printk(KERN_NOTICE "raid5: %s is not clean"
4862 " -- starting background reconstruction\n",
4863 mdname(mddev));
4864 if (mddev->reshape_position != MaxSector) {
4865 /* Check that we can continue the reshape.
4866 * Currently only disks can change, it must
4867 * increase, and we must be past the point where
4868 * a stripe over-writes itself
4870 sector_t here_new, here_old;
4871 int old_disks;
4872 int max_degraded = (mddev->level == 6 ? 2 : 1);
4874 if (mddev->new_level != mddev->level) {
4875 printk(KERN_ERR "raid5: %s: unsupported reshape "
4876 "required - aborting.\n",
4877 mdname(mddev));
4878 return -EINVAL;
4880 old_disks = mddev->raid_disks - mddev->delta_disks;
4881 /* reshape_position must be on a new-stripe boundary, and one
4882 * further up in new geometry must map after here in old
4883 * geometry.
4885 here_new = mddev->reshape_position;
4886 if (sector_div(here_new, mddev->new_chunk_sectors *
4887 (mddev->raid_disks - max_degraded))) {
4888 printk(KERN_ERR "raid5: reshape_position not "
4889 "on a stripe boundary\n");
4890 return -EINVAL;
4892 reshape_offset = here_new * mddev->new_chunk_sectors;
4893 /* here_new is the stripe we will write to */
4894 here_old = mddev->reshape_position;
4895 sector_div(here_old, mddev->chunk_sectors *
4896 (old_disks-max_degraded));
4897 /* here_old is the first stripe that we might need to read
4898 * from */
4899 if (mddev->delta_disks == 0) {
4900 /* We cannot be sure it is safe to start an in-place
4901 * reshape. It is only safe if user-space if monitoring
4902 * and taking constant backups.
4903 * mdadm always starts a situation like this in
4904 * readonly mode so it can take control before
4905 * allowing any writes. So just check for that.
4907 if ((here_new * mddev->new_chunk_sectors !=
4908 here_old * mddev->chunk_sectors) ||
4909 mddev->ro == 0) {
4910 printk(KERN_ERR "raid5: in-place reshape must be started"
4911 " in read-only mode - aborting\n");
4912 return -EINVAL;
4914 } else if (mddev->delta_disks < 0
4915 ? (here_new * mddev->new_chunk_sectors <=
4916 here_old * mddev->chunk_sectors)
4917 : (here_new * mddev->new_chunk_sectors >=
4918 here_old * mddev->chunk_sectors)) {
4919 /* Reading from the same stripe as writing to - bad */
4920 printk(KERN_ERR "raid5: reshape_position too early for "
4921 "auto-recovery - aborting.\n");
4922 return -EINVAL;
4924 printk(KERN_INFO "raid5: reshape will continue\n");
4925 /* OK, we should be able to continue; */
4926 } else {
4927 BUG_ON(mddev->level != mddev->new_level);
4928 BUG_ON(mddev->layout != mddev->new_layout);
4929 BUG_ON(mddev->chunk_sectors != mddev->new_chunk_sectors);
4930 BUG_ON(mddev->delta_disks != 0);
4933 if (mddev->private == NULL)
4934 conf = setup_conf(mddev);
4935 else
4936 conf = mddev->private;
4938 if (IS_ERR(conf))
4939 return PTR_ERR(conf);
4941 mddev->thread = conf->thread;
4942 conf->thread = NULL;
4943 mddev->private = conf;
4946 * 0 for a fully functional array, 1 or 2 for a degraded array.
4948 list_for_each_entry(rdev, &mddev->disks, same_set) {
4949 if (rdev->raid_disk < 0)
4950 continue;
4951 if (test_bit(In_sync, &rdev->flags))
4952 working_disks++;
4953 /* This disc is not fully in-sync. However if it
4954 * just stored parity (beyond the recovery_offset),
4955 * when we don't need to be concerned about the
4956 * array being dirty.
4957 * When reshape goes 'backwards', we never have
4958 * partially completed devices, so we only need
4959 * to worry about reshape going forwards.
4961 /* Hack because v0.91 doesn't store recovery_offset properly. */
4962 if (mddev->major_version == 0 &&
4963 mddev->minor_version > 90)
4964 rdev->recovery_offset = reshape_offset;
4966 printk("%d: w=%d pa=%d pr=%d m=%d a=%d r=%d op1=%d op2=%d\n",
4967 rdev->raid_disk, working_disks, conf->prev_algo,
4968 conf->previous_raid_disks, conf->max_degraded,
4969 conf->algorithm, conf->raid_disks,
4970 only_parity(rdev->raid_disk,
4971 conf->prev_algo,
4972 conf->previous_raid_disks,
4973 conf->max_degraded),
4974 only_parity(rdev->raid_disk,
4975 conf->algorithm,
4976 conf->raid_disks,
4977 conf->max_degraded));
4978 if (rdev->recovery_offset < reshape_offset) {
4979 /* We need to check old and new layout */
4980 if (!only_parity(rdev->raid_disk,
4981 conf->algorithm,
4982 conf->raid_disks,
4983 conf->max_degraded))
4984 continue;
4986 if (!only_parity(rdev->raid_disk,
4987 conf->prev_algo,
4988 conf->previous_raid_disks,
4989 conf->max_degraded))
4990 continue;
4991 dirty_parity_disks++;
4994 mddev->degraded = (max(conf->raid_disks, conf->previous_raid_disks)
4995 - working_disks);
4997 if (mddev->degraded > conf->max_degraded) {
4998 printk(KERN_ERR "raid5: not enough operational devices for %s"
4999 " (%d/%d failed)\n",
5000 mdname(mddev), mddev->degraded, conf->raid_disks);
5001 goto abort;
5004 /* device size must be a multiple of chunk size */
5005 mddev->dev_sectors &= ~(mddev->chunk_sectors - 1);
5006 mddev->resync_max_sectors = mddev->dev_sectors;
5008 if (mddev->degraded > dirty_parity_disks &&
5009 mddev->recovery_cp != MaxSector) {
5010 if (mddev->ok_start_degraded)
5011 printk(KERN_WARNING
5012 "raid5: starting dirty degraded array: %s"
5013 "- data corruption possible.\n",
5014 mdname(mddev));
5015 else {
5016 printk(KERN_ERR
5017 "raid5: cannot start dirty degraded array for %s\n",
5018 mdname(mddev));
5019 goto abort;
5023 if (mddev->degraded == 0)
5024 printk("raid5: raid level %d set %s active with %d out of %d"
5025 " devices, algorithm %d\n", conf->level, mdname(mddev),
5026 mddev->raid_disks-mddev->degraded, mddev->raid_disks,
5027 mddev->new_layout);
5028 else
5029 printk(KERN_ALERT "raid5: raid level %d set %s active with %d"
5030 " out of %d devices, algorithm %d\n", conf->level,
5031 mdname(mddev), mddev->raid_disks - mddev->degraded,
5032 mddev->raid_disks, mddev->new_layout);
5034 print_raid5_conf(conf);
5036 if (conf->reshape_progress != MaxSector) {
5037 printk("...ok start reshape thread\n");
5038 conf->reshape_safe = conf->reshape_progress;
5039 atomic_set(&conf->reshape_stripes, 0);
5040 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
5041 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
5042 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
5043 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
5044 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
5045 "reshape");
5048 /* read-ahead size must cover two whole stripes, which is
5049 * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
5052 int data_disks = conf->previous_raid_disks - conf->max_degraded;
5053 int stripe = data_disks *
5054 ((mddev->chunk_sectors << 9) / PAGE_SIZE);
5055 if (mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
5056 mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
5059 /* Ok, everything is just fine now */
5060 if (sysfs_create_group(&mddev->kobj, &raid5_attrs_group))
5061 printk(KERN_WARNING
5062 "raid5: failed to create sysfs attributes for %s\n",
5063 mdname(mddev));
5065 mddev->queue->queue_lock = &conf->device_lock;
5067 mddev->queue->unplug_fn = raid5_unplug_device;
5068 mddev->queue->backing_dev_info.congested_data = mddev;
5069 mddev->queue->backing_dev_info.congested_fn = raid5_congested;
5071 md_set_array_sectors(mddev, raid5_size(mddev, 0, 0));
5073 blk_queue_merge_bvec(mddev->queue, raid5_mergeable_bvec);
5074 chunk_size = mddev->chunk_sectors << 9;
5075 blk_queue_io_min(mddev->queue, chunk_size);
5076 blk_queue_io_opt(mddev->queue, chunk_size *
5077 (conf->raid_disks - conf->max_degraded));
5079 list_for_each_entry(rdev, &mddev->disks, same_set)
5080 disk_stack_limits(mddev->gendisk, rdev->bdev,
5081 rdev->data_offset << 9);
5083 return 0;
5084 abort:
5085 md_unregister_thread(mddev->thread);
5086 mddev->thread = NULL;
5087 if (conf) {
5088 print_raid5_conf(conf);
5089 free_conf(conf);
5091 mddev->private = NULL;
5092 printk(KERN_ALERT "raid5: failed to run raid set %s\n", mdname(mddev));
5093 return -EIO;
5098 static int stop(mddev_t *mddev)
5100 raid5_conf_t *conf = (raid5_conf_t *) mddev->private;
5102 md_unregister_thread(mddev->thread);
5103 mddev->thread = NULL;
5104 mddev->queue->backing_dev_info.congested_fn = NULL;
5105 blk_sync_queue(mddev->queue); /* the unplug fn references 'conf'*/
5106 sysfs_remove_group(&mddev->kobj, &raid5_attrs_group);
5107 free_conf(conf);
5108 mddev->private = NULL;
5109 return 0;
5112 #ifdef DEBUG
5113 static void print_sh(struct seq_file *seq, struct stripe_head *sh)
5115 int i;
5117 seq_printf(seq, "sh %llu, pd_idx %d, state %ld.\n",
5118 (unsigned long long)sh->sector, sh->pd_idx, sh->state);
5119 seq_printf(seq, "sh %llu, count %d.\n",
5120 (unsigned long long)sh->sector, atomic_read(&sh->count));
5121 seq_printf(seq, "sh %llu, ", (unsigned long long)sh->sector);
5122 for (i = 0; i < sh->disks; i++) {
5123 seq_printf(seq, "(cache%d: %p %ld) ",
5124 i, sh->dev[i].page, sh->dev[i].flags);
5126 seq_printf(seq, "\n");
5129 static void printall(struct seq_file *seq, raid5_conf_t *conf)
5131 struct stripe_head *sh;
5132 struct hlist_node *hn;
5133 int i;
5135 spin_lock_irq(&conf->device_lock);
5136 for (i = 0; i < NR_HASH; i++) {
5137 hlist_for_each_entry(sh, hn, &conf->stripe_hashtbl[i], hash) {
5138 if (sh->raid_conf != conf)
5139 continue;
5140 print_sh(seq, sh);
5143 spin_unlock_irq(&conf->device_lock);
5145 #endif
5147 static void status(struct seq_file *seq, mddev_t *mddev)
5149 raid5_conf_t *conf = (raid5_conf_t *) mddev->private;
5150 int i;
5152 seq_printf(seq, " level %d, %dk chunk, algorithm %d", mddev->level,
5153 mddev->chunk_sectors / 2, mddev->layout);
5154 seq_printf (seq, " [%d/%d] [", conf->raid_disks, conf->raid_disks - mddev->degraded);
5155 for (i = 0; i < conf->raid_disks; i++)
5156 seq_printf (seq, "%s",
5157 conf->disks[i].rdev &&
5158 test_bit(In_sync, &conf->disks[i].rdev->flags) ? "U" : "_");
5159 seq_printf (seq, "]");
5160 #ifdef DEBUG
5161 seq_printf (seq, "\n");
5162 printall(seq, conf);
5163 #endif
5166 static void print_raid5_conf (raid5_conf_t *conf)
5168 int i;
5169 struct disk_info *tmp;
5171 printk("RAID5 conf printout:\n");
5172 if (!conf) {
5173 printk("(conf==NULL)\n");
5174 return;
5176 printk(" --- rd:%d wd:%d\n", conf->raid_disks,
5177 conf->raid_disks - conf->mddev->degraded);
5179 for (i = 0; i < conf->raid_disks; i++) {
5180 char b[BDEVNAME_SIZE];
5181 tmp = conf->disks + i;
5182 if (tmp->rdev)
5183 printk(" disk %d, o:%d, dev:%s\n",
5184 i, !test_bit(Faulty, &tmp->rdev->flags),
5185 bdevname(tmp->rdev->bdev,b));
5189 static int raid5_spare_active(mddev_t *mddev)
5191 int i;
5192 raid5_conf_t *conf = mddev->private;
5193 struct disk_info *tmp;
5195 for (i = 0; i < conf->raid_disks; i++) {
5196 tmp = conf->disks + i;
5197 if (tmp->rdev
5198 && !test_bit(Faulty, &tmp->rdev->flags)
5199 && !test_and_set_bit(In_sync, &tmp->rdev->flags)) {
5200 unsigned long flags;
5201 spin_lock_irqsave(&conf->device_lock, flags);
5202 mddev->degraded--;
5203 spin_unlock_irqrestore(&conf->device_lock, flags);
5206 print_raid5_conf(conf);
5207 return 0;
5210 static int raid5_remove_disk(mddev_t *mddev, int number)
5212 raid5_conf_t *conf = mddev->private;
5213 int err = 0;
5214 mdk_rdev_t *rdev;
5215 struct disk_info *p = conf->disks + number;
5217 print_raid5_conf(conf);
5218 rdev = p->rdev;
5219 if (rdev) {
5220 if (number >= conf->raid_disks &&
5221 conf->reshape_progress == MaxSector)
5222 clear_bit(In_sync, &rdev->flags);
5224 if (test_bit(In_sync, &rdev->flags) ||
5225 atomic_read(&rdev->nr_pending)) {
5226 err = -EBUSY;
5227 goto abort;
5229 /* Only remove non-faulty devices if recovery
5230 * isn't possible.
5232 if (!test_bit(Faulty, &rdev->flags) &&
5233 mddev->degraded <= conf->max_degraded &&
5234 number < conf->raid_disks) {
5235 err = -EBUSY;
5236 goto abort;
5238 p->rdev = NULL;
5239 synchronize_rcu();
5240 if (atomic_read(&rdev->nr_pending)) {
5241 /* lost the race, try later */
5242 err = -EBUSY;
5243 p->rdev = rdev;
5246 abort:
5248 print_raid5_conf(conf);
5249 return err;
5252 static int raid5_add_disk(mddev_t *mddev, mdk_rdev_t *rdev)
5254 raid5_conf_t *conf = mddev->private;
5255 int err = -EEXIST;
5256 int disk;
5257 struct disk_info *p;
5258 int first = 0;
5259 int last = conf->raid_disks - 1;
5261 if (mddev->degraded > conf->max_degraded)
5262 /* no point adding a device */
5263 return -EINVAL;
5265 if (rdev->raid_disk >= 0)
5266 first = last = rdev->raid_disk;
5269 * find the disk ... but prefer rdev->saved_raid_disk
5270 * if possible.
5272 if (rdev->saved_raid_disk >= 0 &&
5273 rdev->saved_raid_disk >= first &&
5274 conf->disks[rdev->saved_raid_disk].rdev == NULL)
5275 disk = rdev->saved_raid_disk;
5276 else
5277 disk = first;
5278 for ( ; disk <= last ; disk++)
5279 if ((p=conf->disks + disk)->rdev == NULL) {
5280 clear_bit(In_sync, &rdev->flags);
5281 rdev->raid_disk = disk;
5282 err = 0;
5283 if (rdev->saved_raid_disk != disk)
5284 conf->fullsync = 1;
5285 rcu_assign_pointer(p->rdev, rdev);
5286 break;
5288 print_raid5_conf(conf);
5289 return err;
5292 static int raid5_resize(mddev_t *mddev, sector_t sectors)
5294 /* no resync is happening, and there is enough space
5295 * on all devices, so we can resize.
5296 * We need to make sure resync covers any new space.
5297 * If the array is shrinking we should possibly wait until
5298 * any io in the removed space completes, but it hardly seems
5299 * worth it.
5301 sectors &= ~((sector_t)mddev->chunk_sectors - 1);
5302 md_set_array_sectors(mddev, raid5_size(mddev, sectors,
5303 mddev->raid_disks));
5304 if (mddev->array_sectors >
5305 raid5_size(mddev, sectors, mddev->raid_disks))
5306 return -EINVAL;
5307 set_capacity(mddev->gendisk, mddev->array_sectors);
5308 mddev->changed = 1;
5309 revalidate_disk(mddev->gendisk);
5310 if (sectors > mddev->dev_sectors && mddev->recovery_cp == MaxSector) {
5311 mddev->recovery_cp = mddev->dev_sectors;
5312 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
5314 mddev->dev_sectors = sectors;
5315 mddev->resync_max_sectors = sectors;
5316 return 0;
5319 static int check_stripe_cache(mddev_t *mddev)
5321 /* Can only proceed if there are plenty of stripe_heads.
5322 * We need a minimum of one full stripe,, and for sensible progress
5323 * it is best to have about 4 times that.
5324 * If we require 4 times, then the default 256 4K stripe_heads will
5325 * allow for chunk sizes up to 256K, which is probably OK.
5326 * If the chunk size is greater, user-space should request more
5327 * stripe_heads first.
5329 raid5_conf_t *conf = mddev->private;
5330 if (((mddev->chunk_sectors << 9) / STRIPE_SIZE) * 4
5331 > conf->max_nr_stripes ||
5332 ((mddev->new_chunk_sectors << 9) / STRIPE_SIZE) * 4
5333 > conf->max_nr_stripes) {
5334 printk(KERN_WARNING "raid5: reshape: not enough stripes. Needed %lu\n",
5335 ((max(mddev->chunk_sectors, mddev->new_chunk_sectors) << 9)
5336 / STRIPE_SIZE)*4);
5337 return 0;
5339 return 1;
5342 static int check_reshape(mddev_t *mddev)
5344 raid5_conf_t *conf = mddev->private;
5346 if (mddev->delta_disks == 0 &&
5347 mddev->new_layout == mddev->layout &&
5348 mddev->new_chunk_sectors == mddev->chunk_sectors)
5349 return 0; /* nothing to do */
5350 if (mddev->bitmap)
5351 /* Cannot grow a bitmap yet */
5352 return -EBUSY;
5353 if (mddev->degraded > conf->max_degraded)
5354 return -EINVAL;
5355 if (mddev->delta_disks < 0) {
5356 /* We might be able to shrink, but the devices must
5357 * be made bigger first.
5358 * For raid6, 4 is the minimum size.
5359 * Otherwise 2 is the minimum
5361 int min = 2;
5362 if (mddev->level == 6)
5363 min = 4;
5364 if (mddev->raid_disks + mddev->delta_disks < min)
5365 return -EINVAL;
5368 if (!check_stripe_cache(mddev))
5369 return -ENOSPC;
5371 return resize_stripes(conf, conf->raid_disks + mddev->delta_disks);
5374 static int raid5_start_reshape(mddev_t *mddev)
5376 raid5_conf_t *conf = mddev->private;
5377 mdk_rdev_t *rdev;
5378 int spares = 0;
5379 int added_devices = 0;
5380 unsigned long flags;
5382 if (test_bit(MD_RECOVERY_RUNNING, &mddev->recovery))
5383 return -EBUSY;
5385 if (!check_stripe_cache(mddev))
5386 return -ENOSPC;
5388 list_for_each_entry(rdev, &mddev->disks, same_set)
5389 if (rdev->raid_disk < 0 &&
5390 !test_bit(Faulty, &rdev->flags))
5391 spares++;
5393 if (spares - mddev->degraded < mddev->delta_disks - conf->max_degraded)
5394 /* Not enough devices even to make a degraded array
5395 * of that size
5397 return -EINVAL;
5399 /* Refuse to reduce size of the array. Any reductions in
5400 * array size must be through explicit setting of array_size
5401 * attribute.
5403 if (raid5_size(mddev, 0, conf->raid_disks + mddev->delta_disks)
5404 < mddev->array_sectors) {
5405 printk(KERN_ERR "md: %s: array size must be reduced "
5406 "before number of disks\n", mdname(mddev));
5407 return -EINVAL;
5410 atomic_set(&conf->reshape_stripes, 0);
5411 spin_lock_irq(&conf->device_lock);
5412 conf->previous_raid_disks = conf->raid_disks;
5413 conf->raid_disks += mddev->delta_disks;
5414 conf->prev_chunk_sectors = conf->chunk_sectors;
5415 conf->chunk_sectors = mddev->new_chunk_sectors;
5416 conf->prev_algo = conf->algorithm;
5417 conf->algorithm = mddev->new_layout;
5418 if (mddev->delta_disks < 0)
5419 conf->reshape_progress = raid5_size(mddev, 0, 0);
5420 else
5421 conf->reshape_progress = 0;
5422 conf->reshape_safe = conf->reshape_progress;
5423 conf->generation++;
5424 spin_unlock_irq(&conf->device_lock);
5426 /* Add some new drives, as many as will fit.
5427 * We know there are enough to make the newly sized array work.
5429 list_for_each_entry(rdev, &mddev->disks, same_set)
5430 if (rdev->raid_disk < 0 &&
5431 !test_bit(Faulty, &rdev->flags)) {
5432 if (raid5_add_disk(mddev, rdev) == 0) {
5433 char nm[20];
5434 if (rdev->raid_disk >= conf->previous_raid_disks)
5435 set_bit(In_sync, &rdev->flags);
5436 else
5437 rdev->recovery_offset = 0;
5438 added_devices++;
5439 sprintf(nm, "rd%d", rdev->raid_disk);
5440 if (sysfs_create_link(&mddev->kobj,
5441 &rdev->kobj, nm))
5442 printk(KERN_WARNING
5443 "raid5: failed to create "
5444 " link %s for %s\n",
5445 nm, mdname(mddev));
5446 } else
5447 break;
5450 if (mddev->delta_disks > 0) {
5451 spin_lock_irqsave(&conf->device_lock, flags);
5452 mddev->degraded = (conf->raid_disks - conf->previous_raid_disks)
5453 - added_devices;
5454 spin_unlock_irqrestore(&conf->device_lock, flags);
5456 mddev->raid_disks = conf->raid_disks;
5457 mddev->reshape_position = conf->reshape_progress;
5458 set_bit(MD_CHANGE_DEVS, &mddev->flags);
5460 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
5461 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
5462 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
5463 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
5464 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
5465 "reshape");
5466 if (!mddev->sync_thread) {
5467 mddev->recovery = 0;
5468 spin_lock_irq(&conf->device_lock);
5469 mddev->raid_disks = conf->raid_disks = conf->previous_raid_disks;
5470 conf->reshape_progress = MaxSector;
5471 spin_unlock_irq(&conf->device_lock);
5472 return -EAGAIN;
5474 conf->reshape_checkpoint = jiffies;
5475 md_wakeup_thread(mddev->sync_thread);
5476 md_new_event(mddev);
5477 return 0;
5480 /* This is called from the reshape thread and should make any
5481 * changes needed in 'conf'
5483 static void end_reshape(raid5_conf_t *conf)
5486 if (!test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery)) {
5488 spin_lock_irq(&conf->device_lock);
5489 conf->previous_raid_disks = conf->raid_disks;
5490 conf->reshape_progress = MaxSector;
5491 spin_unlock_irq(&conf->device_lock);
5492 wake_up(&conf->wait_for_overlap);
5494 /* read-ahead size must cover two whole stripes, which is
5495 * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
5498 int data_disks = conf->raid_disks - conf->max_degraded;
5499 int stripe = data_disks * ((conf->chunk_sectors << 9)
5500 / PAGE_SIZE);
5501 if (conf->mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
5502 conf->mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
5507 /* This is called from the raid5d thread with mddev_lock held.
5508 * It makes config changes to the device.
5510 static void raid5_finish_reshape(mddev_t *mddev)
5512 raid5_conf_t *conf = mddev->private;
5514 if (!test_bit(MD_RECOVERY_INTR, &mddev->recovery)) {
5516 if (mddev->delta_disks > 0) {
5517 md_set_array_sectors(mddev, raid5_size(mddev, 0, 0));
5518 set_capacity(mddev->gendisk, mddev->array_sectors);
5519 mddev->changed = 1;
5520 revalidate_disk(mddev->gendisk);
5521 } else {
5522 int d;
5523 mddev->degraded = conf->raid_disks;
5524 for (d = 0; d < conf->raid_disks ; d++)
5525 if (conf->disks[d].rdev &&
5526 test_bit(In_sync,
5527 &conf->disks[d].rdev->flags))
5528 mddev->degraded--;
5529 for (d = conf->raid_disks ;
5530 d < conf->raid_disks - mddev->delta_disks;
5531 d++) {
5532 mdk_rdev_t *rdev = conf->disks[d].rdev;
5533 if (rdev && raid5_remove_disk(mddev, d) == 0) {
5534 char nm[20];
5535 sprintf(nm, "rd%d", rdev->raid_disk);
5536 sysfs_remove_link(&mddev->kobj, nm);
5537 rdev->raid_disk = -1;
5541 mddev->layout = conf->algorithm;
5542 mddev->chunk_sectors = conf->chunk_sectors;
5543 mddev->reshape_position = MaxSector;
5544 mddev->delta_disks = 0;
5548 static void raid5_quiesce(mddev_t *mddev, int state)
5550 raid5_conf_t *conf = mddev->private;
5552 switch(state) {
5553 case 2: /* resume for a suspend */
5554 wake_up(&conf->wait_for_overlap);
5555 break;
5557 case 1: /* stop all writes */
5558 spin_lock_irq(&conf->device_lock);
5559 /* '2' tells resync/reshape to pause so that all
5560 * active stripes can drain
5562 conf->quiesce = 2;
5563 wait_event_lock_irq(conf->wait_for_stripe,
5564 atomic_read(&conf->active_stripes) == 0 &&
5565 atomic_read(&conf->active_aligned_reads) == 0,
5566 conf->device_lock, /* nothing */);
5567 conf->quiesce = 1;
5568 spin_unlock_irq(&conf->device_lock);
5569 /* allow reshape to continue */
5570 wake_up(&conf->wait_for_overlap);
5571 break;
5573 case 0: /* re-enable writes */
5574 spin_lock_irq(&conf->device_lock);
5575 conf->quiesce = 0;
5576 wake_up(&conf->wait_for_stripe);
5577 wake_up(&conf->wait_for_overlap);
5578 spin_unlock_irq(&conf->device_lock);
5579 break;
5584 static void *raid5_takeover_raid1(mddev_t *mddev)
5586 int chunksect;
5588 if (mddev->raid_disks != 2 ||
5589 mddev->degraded > 1)
5590 return ERR_PTR(-EINVAL);
5592 /* Should check if there are write-behind devices? */
5594 chunksect = 64*2; /* 64K by default */
5596 /* The array must be an exact multiple of chunksize */
5597 while (chunksect && (mddev->array_sectors & (chunksect-1)))
5598 chunksect >>= 1;
5600 if ((chunksect<<9) < STRIPE_SIZE)
5601 /* array size does not allow a suitable chunk size */
5602 return ERR_PTR(-EINVAL);
5604 mddev->new_level = 5;
5605 mddev->new_layout = ALGORITHM_LEFT_SYMMETRIC;
5606 mddev->new_chunk_sectors = chunksect;
5608 return setup_conf(mddev);
5611 static void *raid5_takeover_raid6(mddev_t *mddev)
5613 int new_layout;
5615 switch (mddev->layout) {
5616 case ALGORITHM_LEFT_ASYMMETRIC_6:
5617 new_layout = ALGORITHM_LEFT_ASYMMETRIC;
5618 break;
5619 case ALGORITHM_RIGHT_ASYMMETRIC_6:
5620 new_layout = ALGORITHM_RIGHT_ASYMMETRIC;
5621 break;
5622 case ALGORITHM_LEFT_SYMMETRIC_6:
5623 new_layout = ALGORITHM_LEFT_SYMMETRIC;
5624 break;
5625 case ALGORITHM_RIGHT_SYMMETRIC_6:
5626 new_layout = ALGORITHM_RIGHT_SYMMETRIC;
5627 break;
5628 case ALGORITHM_PARITY_0_6:
5629 new_layout = ALGORITHM_PARITY_0;
5630 break;
5631 case ALGORITHM_PARITY_N:
5632 new_layout = ALGORITHM_PARITY_N;
5633 break;
5634 default:
5635 return ERR_PTR(-EINVAL);
5637 mddev->new_level = 5;
5638 mddev->new_layout = new_layout;
5639 mddev->delta_disks = -1;
5640 mddev->raid_disks -= 1;
5641 return setup_conf(mddev);
5645 static int raid5_check_reshape(mddev_t *mddev)
5647 /* For a 2-drive array, the layout and chunk size can be changed
5648 * immediately as not restriping is needed.
5649 * For larger arrays we record the new value - after validation
5650 * to be used by a reshape pass.
5652 raid5_conf_t *conf = mddev->private;
5653 int new_chunk = mddev->new_chunk_sectors;
5655 if (mddev->new_layout >= 0 && !algorithm_valid_raid5(mddev->new_layout))
5656 return -EINVAL;
5657 if (new_chunk > 0) {
5658 if (!is_power_of_2(new_chunk))
5659 return -EINVAL;
5660 if (new_chunk < (PAGE_SIZE>>9))
5661 return -EINVAL;
5662 if (mddev->array_sectors & (new_chunk-1))
5663 /* not factor of array size */
5664 return -EINVAL;
5667 /* They look valid */
5669 if (mddev->raid_disks == 2) {
5670 /* can make the change immediately */
5671 if (mddev->new_layout >= 0) {
5672 conf->algorithm = mddev->new_layout;
5673 mddev->layout = mddev->new_layout;
5675 if (new_chunk > 0) {
5676 conf->chunk_sectors = new_chunk ;
5677 mddev->chunk_sectors = new_chunk;
5679 set_bit(MD_CHANGE_DEVS, &mddev->flags);
5680 md_wakeup_thread(mddev->thread);
5682 return check_reshape(mddev);
5685 static int raid6_check_reshape(mddev_t *mddev)
5687 int new_chunk = mddev->new_chunk_sectors;
5689 if (mddev->new_layout >= 0 && !algorithm_valid_raid6(mddev->new_layout))
5690 return -EINVAL;
5691 if (new_chunk > 0) {
5692 if (!is_power_of_2(new_chunk))
5693 return -EINVAL;
5694 if (new_chunk < (PAGE_SIZE >> 9))
5695 return -EINVAL;
5696 if (mddev->array_sectors & (new_chunk-1))
5697 /* not factor of array size */
5698 return -EINVAL;
5701 /* They look valid */
5702 return check_reshape(mddev);
5705 static void *raid5_takeover(mddev_t *mddev)
5707 /* raid5 can take over:
5708 * raid0 - if all devices are the same - make it a raid4 layout
5709 * raid1 - if there are two drives. We need to know the chunk size
5710 * raid4 - trivial - just use a raid4 layout.
5711 * raid6 - Providing it is a *_6 layout
5714 if (mddev->level == 1)
5715 return raid5_takeover_raid1(mddev);
5716 if (mddev->level == 4) {
5717 mddev->new_layout = ALGORITHM_PARITY_N;
5718 mddev->new_level = 5;
5719 return setup_conf(mddev);
5721 if (mddev->level == 6)
5722 return raid5_takeover_raid6(mddev);
5724 return ERR_PTR(-EINVAL);
5728 static struct mdk_personality raid5_personality;
5730 static void *raid6_takeover(mddev_t *mddev)
5732 /* Currently can only take over a raid5. We map the
5733 * personality to an equivalent raid6 personality
5734 * with the Q block at the end.
5736 int new_layout;
5738 if (mddev->pers != &raid5_personality)
5739 return ERR_PTR(-EINVAL);
5740 if (mddev->degraded > 1)
5741 return ERR_PTR(-EINVAL);
5742 if (mddev->raid_disks > 253)
5743 return ERR_PTR(-EINVAL);
5744 if (mddev->raid_disks < 3)
5745 return ERR_PTR(-EINVAL);
5747 switch (mddev->layout) {
5748 case ALGORITHM_LEFT_ASYMMETRIC:
5749 new_layout = ALGORITHM_LEFT_ASYMMETRIC_6;
5750 break;
5751 case ALGORITHM_RIGHT_ASYMMETRIC:
5752 new_layout = ALGORITHM_RIGHT_ASYMMETRIC_6;
5753 break;
5754 case ALGORITHM_LEFT_SYMMETRIC:
5755 new_layout = ALGORITHM_LEFT_SYMMETRIC_6;
5756 break;
5757 case ALGORITHM_RIGHT_SYMMETRIC:
5758 new_layout = ALGORITHM_RIGHT_SYMMETRIC_6;
5759 break;
5760 case ALGORITHM_PARITY_0:
5761 new_layout = ALGORITHM_PARITY_0_6;
5762 break;
5763 case ALGORITHM_PARITY_N:
5764 new_layout = ALGORITHM_PARITY_N;
5765 break;
5766 default:
5767 return ERR_PTR(-EINVAL);
5769 mddev->new_level = 6;
5770 mddev->new_layout = new_layout;
5771 mddev->delta_disks = 1;
5772 mddev->raid_disks += 1;
5773 return setup_conf(mddev);
5777 static struct mdk_personality raid6_personality =
5779 .name = "raid6",
5780 .level = 6,
5781 .owner = THIS_MODULE,
5782 .make_request = make_request,
5783 .run = run,
5784 .stop = stop,
5785 .status = status,
5786 .error_handler = error,
5787 .hot_add_disk = raid5_add_disk,
5788 .hot_remove_disk= raid5_remove_disk,
5789 .spare_active = raid5_spare_active,
5790 .sync_request = sync_request,
5791 .resize = raid5_resize,
5792 .size = raid5_size,
5793 .check_reshape = raid6_check_reshape,
5794 .start_reshape = raid5_start_reshape,
5795 .finish_reshape = raid5_finish_reshape,
5796 .quiesce = raid5_quiesce,
5797 .takeover = raid6_takeover,
5799 static struct mdk_personality raid5_personality =
5801 .name = "raid5",
5802 .level = 5,
5803 .owner = THIS_MODULE,
5804 .make_request = make_request,
5805 .run = run,
5806 .stop = stop,
5807 .status = status,
5808 .error_handler = error,
5809 .hot_add_disk = raid5_add_disk,
5810 .hot_remove_disk= raid5_remove_disk,
5811 .spare_active = raid5_spare_active,
5812 .sync_request = sync_request,
5813 .resize = raid5_resize,
5814 .size = raid5_size,
5815 .check_reshape = raid5_check_reshape,
5816 .start_reshape = raid5_start_reshape,
5817 .finish_reshape = raid5_finish_reshape,
5818 .quiesce = raid5_quiesce,
5819 .takeover = raid5_takeover,
5822 static struct mdk_personality raid4_personality =
5824 .name = "raid4",
5825 .level = 4,
5826 .owner = THIS_MODULE,
5827 .make_request = make_request,
5828 .run = run,
5829 .stop = stop,
5830 .status = status,
5831 .error_handler = error,
5832 .hot_add_disk = raid5_add_disk,
5833 .hot_remove_disk= raid5_remove_disk,
5834 .spare_active = raid5_spare_active,
5835 .sync_request = sync_request,
5836 .resize = raid5_resize,
5837 .size = raid5_size,
5838 .check_reshape = raid5_check_reshape,
5839 .start_reshape = raid5_start_reshape,
5840 .finish_reshape = raid5_finish_reshape,
5841 .quiesce = raid5_quiesce,
5844 static int __init raid5_init(void)
5846 register_md_personality(&raid6_personality);
5847 register_md_personality(&raid5_personality);
5848 register_md_personality(&raid4_personality);
5849 return 0;
5852 static void raid5_exit(void)
5854 unregister_md_personality(&raid6_personality);
5855 unregister_md_personality(&raid5_personality);
5856 unregister_md_personality(&raid4_personality);
5859 module_init(raid5_init);
5860 module_exit(raid5_exit);
5861 MODULE_LICENSE("GPL");
5862 MODULE_ALIAS("md-personality-4"); /* RAID5 */
5863 MODULE_ALIAS("md-raid5");
5864 MODULE_ALIAS("md-raid4");
5865 MODULE_ALIAS("md-level-5");
5866 MODULE_ALIAS("md-level-4");
5867 MODULE_ALIAS("md-personality-8"); /* RAID6 */
5868 MODULE_ALIAS("md-raid6");
5869 MODULE_ALIAS("md-level-6");
5871 /* This used to be two separate modules, they were: */
5872 MODULE_ALIAS("raid5");
5873 MODULE_ALIAS("raid6");