2 * raid5.c : Multiple Devices driver for Linux
3 * Copyright (C) 1996, 1997 Ingo Molnar, Miguel de Icaza, Gadi Oxman
4 * Copyright (C) 1999, 2000 Ingo Molnar
5 * Copyright (C) 2002, 2003 H. Peter Anvin
7 * RAID-4/5/6 management functions.
8 * Thanks to Penguin Computing for making the RAID-6 development possible
9 * by donating a test server!
11 * This program is free software; you can redistribute it and/or modify
12 * it under the terms of the GNU General Public License as published by
13 * the Free Software Foundation; either version 2, or (at your option)
16 * You should have received a copy of the GNU General Public License
17 * (for example /usr/src/linux/COPYING); if not, write to the Free
18 * Software Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
24 * The sequencing for updating the bitmap reliably is a little
25 * subtle (and I got it wrong the first time) so it deserves some
28 * We group bitmap updates into batches. Each batch has a number.
29 * We may write out several batches at once, but that isn't very important.
30 * conf->bm_write is the number of the last batch successfully written.
31 * conf->bm_flush is the number of the last batch that was closed to
33 * When we discover that we will need to write to any block in a stripe
34 * (in add_stripe_bio) we update the in-memory bitmap and record in sh->bm_seq
35 * the number of the batch it will be in. This is bm_flush+1.
36 * When we are ready to do a write, if that batch hasn't been written yet,
37 * we plug the array and queue the stripe for later.
38 * When an unplug happens, we increment bm_flush, thus closing the current
40 * When we notice that bm_flush > bm_write, we write out all pending updates
41 * to the bitmap, and advance bm_write to where bm_flush was.
42 * This may occasionally write a bit out twice, but is sure never to
46 #include <linux/module.h>
47 #include <linux/slab.h>
48 #include <linux/highmem.h>
49 #include <linux/bitops.h>
50 #include <linux/kthread.h>
51 #include <asm/atomic.h>
54 #include <linux/raid/bitmap.h>
60 #define NR_STRIPES 256
61 #define STRIPE_SIZE PAGE_SIZE
62 #define STRIPE_SHIFT (PAGE_SHIFT - 9)
63 #define STRIPE_SECTORS (STRIPE_SIZE>>9)
64 #define IO_THRESHOLD 1
65 #define NR_HASH (PAGE_SIZE / sizeof(struct hlist_head))
66 #define HASH_MASK (NR_HASH - 1)
68 #define stripe_hash(conf, sect) (&((conf)->stripe_hashtbl[((sect) >> STRIPE_SHIFT) & HASH_MASK]))
70 /* bio's attached to a stripe+device for I/O are linked together in bi_sector
71 * order without overlap. There may be several bio's per stripe+device, and
72 * a bio could span several devices.
73 * When walking this list for a particular stripe+device, we must never proceed
74 * beyond a bio that extends past this device, as the next bio might no longer
76 * This macro is used to determine the 'next' bio in the list, given the sector
77 * of the current stripe+device
79 #define r5_next_bio(bio, sect) ( ( (bio)->bi_sector + ((bio)->bi_size>>9) < sect + STRIPE_SECTORS) ? (bio)->bi_next : NULL)
81 * The following can be used to debug the driver
84 #define RAID5_PARANOIA 1
85 #if RAID5_PARANOIA && defined(CONFIG_SMP)
86 # define CHECK_DEVLOCK() assert_spin_locked(&conf->device_lock)
88 # define CHECK_DEVLOCK()
91 #define PRINTK(x...) ((void)(RAID5_DEBUG && printk(x)))
97 #if !RAID6_USE_EMPTY_ZERO_PAGE
98 /* In .bss so it's zeroed */
99 const char raid6_empty_zero_page[PAGE_SIZE] __attribute__((aligned(256)));
102 static inline int raid6_next_disk(int disk, int raid_disks)
105 return (disk < raid_disks) ? disk : 0;
107 static void print_raid5_conf (raid5_conf_t *conf);
109 static void __release_stripe(raid5_conf_t *conf, struct stripe_head *sh)
111 if (atomic_dec_and_test(&sh->count)) {
112 BUG_ON(!list_empty(&sh->lru));
113 BUG_ON(atomic_read(&conf->active_stripes)==0);
114 if (test_bit(STRIPE_HANDLE, &sh->state)) {
115 if (test_bit(STRIPE_DELAYED, &sh->state)) {
116 list_add_tail(&sh->lru, &conf->delayed_list);
117 blk_plug_device(conf->mddev->queue);
118 } else if (test_bit(STRIPE_BIT_DELAY, &sh->state) &&
119 sh->bm_seq - conf->seq_write > 0) {
120 list_add_tail(&sh->lru, &conf->bitmap_list);
121 blk_plug_device(conf->mddev->queue);
123 clear_bit(STRIPE_BIT_DELAY, &sh->state);
124 list_add_tail(&sh->lru, &conf->handle_list);
126 md_wakeup_thread(conf->mddev->thread);
128 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
129 atomic_dec(&conf->preread_active_stripes);
130 if (atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD)
131 md_wakeup_thread(conf->mddev->thread);
133 atomic_dec(&conf->active_stripes);
134 if (!test_bit(STRIPE_EXPANDING, &sh->state)) {
135 list_add_tail(&sh->lru, &conf->inactive_list);
136 wake_up(&conf->wait_for_stripe);
137 if (conf->retry_read_aligned)
138 md_wakeup_thread(conf->mddev->thread);
143 static void release_stripe(struct stripe_head *sh)
145 raid5_conf_t *conf = sh->raid_conf;
148 spin_lock_irqsave(&conf->device_lock, flags);
149 __release_stripe(conf, sh);
150 spin_unlock_irqrestore(&conf->device_lock, flags);
153 static inline void remove_hash(struct stripe_head *sh)
155 PRINTK("remove_hash(), stripe %llu\n", (unsigned long long)sh->sector);
157 hlist_del_init(&sh->hash);
160 static inline void insert_hash(raid5_conf_t *conf, struct stripe_head *sh)
162 struct hlist_head *hp = stripe_hash(conf, sh->sector);
164 PRINTK("insert_hash(), stripe %llu\n", (unsigned long long)sh->sector);
167 hlist_add_head(&sh->hash, hp);
171 /* find an idle stripe, make sure it is unhashed, and return it. */
172 static struct stripe_head *get_free_stripe(raid5_conf_t *conf)
174 struct stripe_head *sh = NULL;
175 struct list_head *first;
178 if (list_empty(&conf->inactive_list))
180 first = conf->inactive_list.next;
181 sh = list_entry(first, struct stripe_head, lru);
182 list_del_init(first);
184 atomic_inc(&conf->active_stripes);
189 static void shrink_buffers(struct stripe_head *sh, int num)
194 for (i=0; i<num ; i++) {
198 sh->dev[i].page = NULL;
203 static int grow_buffers(struct stripe_head *sh, int num)
207 for (i=0; i<num; i++) {
210 if (!(page = alloc_page(GFP_KERNEL))) {
213 sh->dev[i].page = page;
218 static void raid5_build_block (struct stripe_head *sh, int i);
220 static void init_stripe(struct stripe_head *sh, sector_t sector, int pd_idx, int disks)
222 raid5_conf_t *conf = sh->raid_conf;
225 BUG_ON(atomic_read(&sh->count) != 0);
226 BUG_ON(test_bit(STRIPE_HANDLE, &sh->state));
229 PRINTK("init_stripe called, stripe %llu\n",
230 (unsigned long long)sh->sector);
240 for (i = sh->disks; i--; ) {
241 struct r5dev *dev = &sh->dev[i];
243 if (dev->toread || dev->towrite || dev->written ||
244 test_bit(R5_LOCKED, &dev->flags)) {
245 printk("sector=%llx i=%d %p %p %p %d\n",
246 (unsigned long long)sh->sector, i, dev->toread,
247 dev->towrite, dev->written,
248 test_bit(R5_LOCKED, &dev->flags));
252 raid5_build_block(sh, i);
254 insert_hash(conf, sh);
257 static struct stripe_head *__find_stripe(raid5_conf_t *conf, sector_t sector, int disks)
259 struct stripe_head *sh;
260 struct hlist_node *hn;
263 PRINTK("__find_stripe, sector %llu\n", (unsigned long long)sector);
264 hlist_for_each_entry(sh, hn, stripe_hash(conf, sector), hash)
265 if (sh->sector == sector && sh->disks == disks)
267 PRINTK("__stripe %llu not in cache\n", (unsigned long long)sector);
271 static void unplug_slaves(mddev_t *mddev);
272 static void raid5_unplug_device(request_queue_t *q);
274 static struct stripe_head *get_active_stripe(raid5_conf_t *conf, sector_t sector, int disks,
275 int pd_idx, int noblock)
277 struct stripe_head *sh;
279 PRINTK("get_stripe, sector %llu\n", (unsigned long long)sector);
281 spin_lock_irq(&conf->device_lock);
284 wait_event_lock_irq(conf->wait_for_stripe,
286 conf->device_lock, /* nothing */);
287 sh = __find_stripe(conf, sector, disks);
289 if (!conf->inactive_blocked)
290 sh = get_free_stripe(conf);
291 if (noblock && sh == NULL)
294 conf->inactive_blocked = 1;
295 wait_event_lock_irq(conf->wait_for_stripe,
296 !list_empty(&conf->inactive_list) &&
297 (atomic_read(&conf->active_stripes)
298 < (conf->max_nr_stripes *3/4)
299 || !conf->inactive_blocked),
301 raid5_unplug_device(conf->mddev->queue)
303 conf->inactive_blocked = 0;
305 init_stripe(sh, sector, pd_idx, disks);
307 if (atomic_read(&sh->count)) {
308 BUG_ON(!list_empty(&sh->lru));
310 if (!test_bit(STRIPE_HANDLE, &sh->state))
311 atomic_inc(&conf->active_stripes);
312 if (list_empty(&sh->lru) &&
313 !test_bit(STRIPE_EXPANDING, &sh->state))
315 list_del_init(&sh->lru);
318 } while (sh == NULL);
321 atomic_inc(&sh->count);
323 spin_unlock_irq(&conf->device_lock);
327 static int grow_one_stripe(raid5_conf_t *conf)
329 struct stripe_head *sh;
330 sh = kmem_cache_alloc(conf->slab_cache, GFP_KERNEL);
333 memset(sh, 0, sizeof(*sh) + (conf->raid_disks-1)*sizeof(struct r5dev));
334 sh->raid_conf = conf;
335 spin_lock_init(&sh->lock);
337 if (grow_buffers(sh, conf->raid_disks)) {
338 shrink_buffers(sh, conf->raid_disks);
339 kmem_cache_free(conf->slab_cache, sh);
342 sh->disks = conf->raid_disks;
343 /* we just created an active stripe so... */
344 atomic_set(&sh->count, 1);
345 atomic_inc(&conf->active_stripes);
346 INIT_LIST_HEAD(&sh->lru);
351 static int grow_stripes(raid5_conf_t *conf, int num)
353 struct kmem_cache *sc;
354 int devs = conf->raid_disks;
356 sprintf(conf->cache_name[0], "raid5/%s", mdname(conf->mddev));
357 sprintf(conf->cache_name[1], "raid5/%s-alt", mdname(conf->mddev));
358 conf->active_name = 0;
359 sc = kmem_cache_create(conf->cache_name[conf->active_name],
360 sizeof(struct stripe_head)+(devs-1)*sizeof(struct r5dev),
364 conf->slab_cache = sc;
365 conf->pool_size = devs;
367 if (!grow_one_stripe(conf))
372 #ifdef CONFIG_MD_RAID5_RESHAPE
373 static int resize_stripes(raid5_conf_t *conf, int newsize)
375 /* Make all the stripes able to hold 'newsize' devices.
376 * New slots in each stripe get 'page' set to a new page.
378 * This happens in stages:
379 * 1/ create a new kmem_cache and allocate the required number of
381 * 2/ gather all the old stripe_heads and tranfer the pages across
382 * to the new stripe_heads. This will have the side effect of
383 * freezing the array as once all stripe_heads have been collected,
384 * no IO will be possible. Old stripe heads are freed once their
385 * pages have been transferred over, and the old kmem_cache is
386 * freed when all stripes are done.
387 * 3/ reallocate conf->disks to be suitable bigger. If this fails,
388 * we simple return a failre status - no need to clean anything up.
389 * 4/ allocate new pages for the new slots in the new stripe_heads.
390 * If this fails, we don't bother trying the shrink the
391 * stripe_heads down again, we just leave them as they are.
392 * As each stripe_head is processed the new one is released into
395 * Once step2 is started, we cannot afford to wait for a write,
396 * so we use GFP_NOIO allocations.
398 struct stripe_head *osh, *nsh;
399 LIST_HEAD(newstripes);
400 struct disk_info *ndisks;
402 struct kmem_cache *sc;
405 if (newsize <= conf->pool_size)
406 return 0; /* never bother to shrink */
409 sc = kmem_cache_create(conf->cache_name[1-conf->active_name],
410 sizeof(struct stripe_head)+(newsize-1)*sizeof(struct r5dev),
415 for (i = conf->max_nr_stripes; i; i--) {
416 nsh = kmem_cache_alloc(sc, GFP_KERNEL);
420 memset(nsh, 0, sizeof(*nsh) + (newsize-1)*sizeof(struct r5dev));
422 nsh->raid_conf = conf;
423 spin_lock_init(&nsh->lock);
425 list_add(&nsh->lru, &newstripes);
428 /* didn't get enough, give up */
429 while (!list_empty(&newstripes)) {
430 nsh = list_entry(newstripes.next, struct stripe_head, lru);
432 kmem_cache_free(sc, nsh);
434 kmem_cache_destroy(sc);
437 /* Step 2 - Must use GFP_NOIO now.
438 * OK, we have enough stripes, start collecting inactive
439 * stripes and copying them over
441 list_for_each_entry(nsh, &newstripes, lru) {
442 spin_lock_irq(&conf->device_lock);
443 wait_event_lock_irq(conf->wait_for_stripe,
444 !list_empty(&conf->inactive_list),
446 unplug_slaves(conf->mddev)
448 osh = get_free_stripe(conf);
449 spin_unlock_irq(&conf->device_lock);
450 atomic_set(&nsh->count, 1);
451 for(i=0; i<conf->pool_size; i++)
452 nsh->dev[i].page = osh->dev[i].page;
453 for( ; i<newsize; i++)
454 nsh->dev[i].page = NULL;
455 kmem_cache_free(conf->slab_cache, osh);
457 kmem_cache_destroy(conf->slab_cache);
460 * At this point, we are holding all the stripes so the array
461 * is completely stalled, so now is a good time to resize
464 ndisks = kzalloc(newsize * sizeof(struct disk_info), GFP_NOIO);
466 for (i=0; i<conf->raid_disks; i++)
467 ndisks[i] = conf->disks[i];
469 conf->disks = ndisks;
473 /* Step 4, return new stripes to service */
474 while(!list_empty(&newstripes)) {
475 nsh = list_entry(newstripes.next, struct stripe_head, lru);
476 list_del_init(&nsh->lru);
477 for (i=conf->raid_disks; i < newsize; i++)
478 if (nsh->dev[i].page == NULL) {
479 struct page *p = alloc_page(GFP_NOIO);
480 nsh->dev[i].page = p;
486 /* critical section pass, GFP_NOIO no longer needed */
488 conf->slab_cache = sc;
489 conf->active_name = 1-conf->active_name;
490 conf->pool_size = newsize;
495 static int drop_one_stripe(raid5_conf_t *conf)
497 struct stripe_head *sh;
499 spin_lock_irq(&conf->device_lock);
500 sh = get_free_stripe(conf);
501 spin_unlock_irq(&conf->device_lock);
504 BUG_ON(atomic_read(&sh->count));
505 shrink_buffers(sh, conf->pool_size);
506 kmem_cache_free(conf->slab_cache, sh);
507 atomic_dec(&conf->active_stripes);
511 static void shrink_stripes(raid5_conf_t *conf)
513 while (drop_one_stripe(conf))
516 if (conf->slab_cache)
517 kmem_cache_destroy(conf->slab_cache);
518 conf->slab_cache = NULL;
521 static int raid5_end_read_request(struct bio * bi, unsigned int bytes_done,
524 struct stripe_head *sh = bi->bi_private;
525 raid5_conf_t *conf = sh->raid_conf;
526 int disks = sh->disks, i;
527 int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
528 char b[BDEVNAME_SIZE];
534 for (i=0 ; i<disks; i++)
535 if (bi == &sh->dev[i].req)
538 PRINTK("end_read_request %llu/%d, count: %d, uptodate %d.\n",
539 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
550 spin_lock_irqsave(&conf->device_lock, flags);
551 /* we can return a buffer if we bypassed the cache or
552 * if the top buffer is not in highmem. If there are
553 * multiple buffers, leave the extra work to
556 buffer = sh->bh_read[i];
558 (!PageHighMem(buffer->b_page)
559 || buffer->b_page == bh->b_page )
561 sh->bh_read[i] = buffer->b_reqnext;
562 buffer->b_reqnext = NULL;
565 spin_unlock_irqrestore(&conf->device_lock, flags);
566 if (sh->bh_page[i]==bh->b_page)
567 set_buffer_uptodate(bh);
569 if (buffer->b_page != bh->b_page)
570 memcpy(buffer->b_data, bh->b_data, bh->b_size);
571 buffer->b_end_io(buffer, 1);
574 set_bit(R5_UPTODATE, &sh->dev[i].flags);
576 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
577 rdev = conf->disks[i].rdev;
578 printk(KERN_INFO "raid5:%s: read error corrected (%lu sectors at %llu on %s)\n",
579 mdname(conf->mddev), STRIPE_SECTORS,
580 (unsigned long long)sh->sector + rdev->data_offset,
581 bdevname(rdev->bdev, b));
582 clear_bit(R5_ReadError, &sh->dev[i].flags);
583 clear_bit(R5_ReWrite, &sh->dev[i].flags);
585 if (atomic_read(&conf->disks[i].rdev->read_errors))
586 atomic_set(&conf->disks[i].rdev->read_errors, 0);
588 const char *bdn = bdevname(conf->disks[i].rdev->bdev, b);
590 rdev = conf->disks[i].rdev;
592 clear_bit(R5_UPTODATE, &sh->dev[i].flags);
593 atomic_inc(&rdev->read_errors);
594 if (conf->mddev->degraded)
595 printk(KERN_WARNING "raid5:%s: read error not correctable (sector %llu on %s).\n",
597 (unsigned long long)sh->sector + rdev->data_offset,
599 else if (test_bit(R5_ReWrite, &sh->dev[i].flags))
601 printk(KERN_WARNING "raid5:%s: read error NOT corrected!! (sector %llu on %s).\n",
603 (unsigned long long)sh->sector + rdev->data_offset,
605 else if (atomic_read(&rdev->read_errors)
606 > conf->max_nr_stripes)
608 "raid5:%s: Too many read errors, failing device %s.\n",
609 mdname(conf->mddev), bdn);
613 set_bit(R5_ReadError, &sh->dev[i].flags);
615 clear_bit(R5_ReadError, &sh->dev[i].flags);
616 clear_bit(R5_ReWrite, &sh->dev[i].flags);
617 md_error(conf->mddev, rdev);
620 rdev_dec_pending(conf->disks[i].rdev, conf->mddev);
622 /* must restore b_page before unlocking buffer... */
623 if (sh->bh_page[i] != bh->b_page) {
624 bh->b_page = sh->bh_page[i];
625 bh->b_data = page_address(bh->b_page);
626 clear_buffer_uptodate(bh);
629 clear_bit(R5_LOCKED, &sh->dev[i].flags);
630 set_bit(STRIPE_HANDLE, &sh->state);
635 static int raid5_end_write_request (struct bio *bi, unsigned int bytes_done,
638 struct stripe_head *sh = bi->bi_private;
639 raid5_conf_t *conf = sh->raid_conf;
640 int disks = sh->disks, i;
641 int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
646 for (i=0 ; i<disks; i++)
647 if (bi == &sh->dev[i].req)
650 PRINTK("end_write_request %llu/%d, count %d, uptodate: %d.\n",
651 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
659 md_error(conf->mddev, conf->disks[i].rdev);
661 rdev_dec_pending(conf->disks[i].rdev, conf->mddev);
663 clear_bit(R5_LOCKED, &sh->dev[i].flags);
664 set_bit(STRIPE_HANDLE, &sh->state);
670 static sector_t compute_blocknr(struct stripe_head *sh, int i);
672 static void raid5_build_block (struct stripe_head *sh, int i)
674 struct r5dev *dev = &sh->dev[i];
677 dev->req.bi_io_vec = &dev->vec;
679 dev->req.bi_max_vecs++;
680 dev->vec.bv_page = dev->page;
681 dev->vec.bv_len = STRIPE_SIZE;
682 dev->vec.bv_offset = 0;
684 dev->req.bi_sector = sh->sector;
685 dev->req.bi_private = sh;
688 dev->sector = compute_blocknr(sh, i);
691 static void error(mddev_t *mddev, mdk_rdev_t *rdev)
693 char b[BDEVNAME_SIZE];
694 raid5_conf_t *conf = (raid5_conf_t *) mddev->private;
695 PRINTK("raid5: error called\n");
697 if (!test_bit(Faulty, &rdev->flags)) {
698 set_bit(MD_CHANGE_DEVS, &mddev->flags);
699 if (test_and_clear_bit(In_sync, &rdev->flags)) {
701 spin_lock_irqsave(&conf->device_lock, flags);
703 spin_unlock_irqrestore(&conf->device_lock, flags);
705 * if recovery was running, make sure it aborts.
707 set_bit(MD_RECOVERY_ERR, &mddev->recovery);
709 set_bit(Faulty, &rdev->flags);
711 "raid5: Disk failure on %s, disabling device."
712 " Operation continuing on %d devices\n",
713 bdevname(rdev->bdev,b), conf->raid_disks - mddev->degraded);
718 * Input: a 'big' sector number,
719 * Output: index of the data and parity disk, and the sector # in them.
721 static sector_t raid5_compute_sector(sector_t r_sector, unsigned int raid_disks,
722 unsigned int data_disks, unsigned int * dd_idx,
723 unsigned int * pd_idx, raid5_conf_t *conf)
726 unsigned long chunk_number;
727 unsigned int chunk_offset;
729 int sectors_per_chunk = conf->chunk_size >> 9;
731 /* First compute the information on this sector */
734 * Compute the chunk number and the sector offset inside the chunk
736 chunk_offset = sector_div(r_sector, sectors_per_chunk);
737 chunk_number = r_sector;
738 BUG_ON(r_sector != chunk_number);
741 * Compute the stripe number
743 stripe = chunk_number / data_disks;
746 * Compute the data disk and parity disk indexes inside the stripe
748 *dd_idx = chunk_number % data_disks;
751 * Select the parity disk based on the user selected algorithm.
753 switch(conf->level) {
755 *pd_idx = data_disks;
758 switch (conf->algorithm) {
759 case ALGORITHM_LEFT_ASYMMETRIC:
760 *pd_idx = data_disks - stripe % raid_disks;
761 if (*dd_idx >= *pd_idx)
764 case ALGORITHM_RIGHT_ASYMMETRIC:
765 *pd_idx = stripe % raid_disks;
766 if (*dd_idx >= *pd_idx)
769 case ALGORITHM_LEFT_SYMMETRIC:
770 *pd_idx = data_disks - stripe % raid_disks;
771 *dd_idx = (*pd_idx + 1 + *dd_idx) % raid_disks;
773 case ALGORITHM_RIGHT_SYMMETRIC:
774 *pd_idx = stripe % raid_disks;
775 *dd_idx = (*pd_idx + 1 + *dd_idx) % raid_disks;
778 printk(KERN_ERR "raid5: unsupported algorithm %d\n",
785 switch (conf->algorithm) {
786 case ALGORITHM_LEFT_ASYMMETRIC:
787 *pd_idx = raid_disks - 1 - (stripe % raid_disks);
788 if (*pd_idx == raid_disks-1)
789 (*dd_idx)++; /* Q D D D P */
790 else if (*dd_idx >= *pd_idx)
791 (*dd_idx) += 2; /* D D P Q D */
793 case ALGORITHM_RIGHT_ASYMMETRIC:
794 *pd_idx = stripe % raid_disks;
795 if (*pd_idx == raid_disks-1)
796 (*dd_idx)++; /* Q D D D P */
797 else if (*dd_idx >= *pd_idx)
798 (*dd_idx) += 2; /* D D P Q D */
800 case ALGORITHM_LEFT_SYMMETRIC:
801 *pd_idx = raid_disks - 1 - (stripe % raid_disks);
802 *dd_idx = (*pd_idx + 2 + *dd_idx) % raid_disks;
804 case ALGORITHM_RIGHT_SYMMETRIC:
805 *pd_idx = stripe % raid_disks;
806 *dd_idx = (*pd_idx + 2 + *dd_idx) % raid_disks;
809 printk (KERN_CRIT "raid6: unsupported algorithm %d\n",
816 * Finally, compute the new sector number
818 new_sector = (sector_t)stripe * sectors_per_chunk + chunk_offset;
823 static sector_t compute_blocknr(struct stripe_head *sh, int i)
825 raid5_conf_t *conf = sh->raid_conf;
826 int raid_disks = sh->disks, data_disks = raid_disks - 1;
827 sector_t new_sector = sh->sector, check;
828 int sectors_per_chunk = conf->chunk_size >> 9;
831 int chunk_number, dummy1, dummy2, dd_idx = i;
835 chunk_offset = sector_div(new_sector, sectors_per_chunk);
837 BUG_ON(new_sector != stripe);
841 switch(conf->level) {
844 switch (conf->algorithm) {
845 case ALGORITHM_LEFT_ASYMMETRIC:
846 case ALGORITHM_RIGHT_ASYMMETRIC:
850 case ALGORITHM_LEFT_SYMMETRIC:
851 case ALGORITHM_RIGHT_SYMMETRIC:
854 i -= (sh->pd_idx + 1);
857 printk(KERN_ERR "raid5: unsupported algorithm %d\n",
862 data_disks = raid_disks - 2;
863 if (i == raid6_next_disk(sh->pd_idx, raid_disks))
864 return 0; /* It is the Q disk */
865 switch (conf->algorithm) {
866 case ALGORITHM_LEFT_ASYMMETRIC:
867 case ALGORITHM_RIGHT_ASYMMETRIC:
868 if (sh->pd_idx == raid_disks-1)
870 else if (i > sh->pd_idx)
871 i -= 2; /* D D P Q D */
873 case ALGORITHM_LEFT_SYMMETRIC:
874 case ALGORITHM_RIGHT_SYMMETRIC:
875 if (sh->pd_idx == raid_disks-1)
881 i -= (sh->pd_idx + 2);
885 printk (KERN_CRIT "raid6: unsupported algorithm %d\n",
891 chunk_number = stripe * data_disks + i;
892 r_sector = (sector_t)chunk_number * sectors_per_chunk + chunk_offset;
894 check = raid5_compute_sector (r_sector, raid_disks, data_disks, &dummy1, &dummy2, conf);
895 if (check != sh->sector || dummy1 != dd_idx || dummy2 != sh->pd_idx) {
896 printk(KERN_ERR "compute_blocknr: map not correct\n");
905 * Copy data between a page in the stripe cache, and one or more bion
906 * The page could align with the middle of the bio, or there could be
907 * several bion, each with several bio_vecs, which cover part of the page
908 * Multiple bion are linked together on bi_next. There may be extras
909 * at the end of this list. We ignore them.
911 static void copy_data(int frombio, struct bio *bio,
915 char *pa = page_address(page);
920 if (bio->bi_sector >= sector)
921 page_offset = (signed)(bio->bi_sector - sector) * 512;
923 page_offset = (signed)(sector - bio->bi_sector) * -512;
924 bio_for_each_segment(bvl, bio, i) {
925 int len = bio_iovec_idx(bio,i)->bv_len;
929 if (page_offset < 0) {
930 b_offset = -page_offset;
931 page_offset += b_offset;
935 if (len > 0 && page_offset + len > STRIPE_SIZE)
936 clen = STRIPE_SIZE - page_offset;
940 char *ba = __bio_kmap_atomic(bio, i, KM_USER0);
942 memcpy(pa+page_offset, ba+b_offset, clen);
944 memcpy(ba+b_offset, pa+page_offset, clen);
945 __bio_kunmap_atomic(ba, KM_USER0);
947 if (clen < len) /* hit end of page */
953 #define check_xor() do { \
954 if (count == MAX_XOR_BLOCKS) { \
955 xor_block(count, STRIPE_SIZE, ptr); \
961 static void compute_block(struct stripe_head *sh, int dd_idx)
963 int i, count, disks = sh->disks;
964 void *ptr[MAX_XOR_BLOCKS], *p;
966 PRINTK("compute_block, stripe %llu, idx %d\n",
967 (unsigned long long)sh->sector, dd_idx);
969 ptr[0] = page_address(sh->dev[dd_idx].page);
970 memset(ptr[0], 0, STRIPE_SIZE);
972 for (i = disks ; i--; ) {
975 p = page_address(sh->dev[i].page);
976 if (test_bit(R5_UPTODATE, &sh->dev[i].flags))
979 printk(KERN_ERR "compute_block() %d, stripe %llu, %d"
980 " not present\n", dd_idx,
981 (unsigned long long)sh->sector, i);
986 xor_block(count, STRIPE_SIZE, ptr);
987 set_bit(R5_UPTODATE, &sh->dev[dd_idx].flags);
990 static void compute_parity5(struct stripe_head *sh, int method)
992 raid5_conf_t *conf = sh->raid_conf;
993 int i, pd_idx = sh->pd_idx, disks = sh->disks, count;
994 void *ptr[MAX_XOR_BLOCKS];
997 PRINTK("compute_parity5, stripe %llu, method %d\n",
998 (unsigned long long)sh->sector, method);
1001 ptr[0] = page_address(sh->dev[pd_idx].page);
1003 case READ_MODIFY_WRITE:
1004 BUG_ON(!test_bit(R5_UPTODATE, &sh->dev[pd_idx].flags));
1005 for (i=disks ; i-- ;) {
1008 if (sh->dev[i].towrite &&
1009 test_bit(R5_UPTODATE, &sh->dev[i].flags)) {
1010 ptr[count++] = page_address(sh->dev[i].page);
1011 chosen = sh->dev[i].towrite;
1012 sh->dev[i].towrite = NULL;
1014 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
1015 wake_up(&conf->wait_for_overlap);
1017 BUG_ON(sh->dev[i].written);
1018 sh->dev[i].written = chosen;
1023 case RECONSTRUCT_WRITE:
1024 memset(ptr[0], 0, STRIPE_SIZE);
1025 for (i= disks; i-- ;)
1026 if (i!=pd_idx && sh->dev[i].towrite) {
1027 chosen = sh->dev[i].towrite;
1028 sh->dev[i].towrite = NULL;
1030 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
1031 wake_up(&conf->wait_for_overlap);
1033 BUG_ON(sh->dev[i].written);
1034 sh->dev[i].written = chosen;
1041 xor_block(count, STRIPE_SIZE, ptr);
1045 for (i = disks; i--;)
1046 if (sh->dev[i].written) {
1047 sector_t sector = sh->dev[i].sector;
1048 struct bio *wbi = sh->dev[i].written;
1049 while (wbi && wbi->bi_sector < sector + STRIPE_SECTORS) {
1050 copy_data(1, wbi, sh->dev[i].page, sector);
1051 wbi = r5_next_bio(wbi, sector);
1054 set_bit(R5_LOCKED, &sh->dev[i].flags);
1055 set_bit(R5_UPTODATE, &sh->dev[i].flags);
1059 case RECONSTRUCT_WRITE:
1063 ptr[count++] = page_address(sh->dev[i].page);
1067 case READ_MODIFY_WRITE:
1068 for (i = disks; i--;)
1069 if (sh->dev[i].written) {
1070 ptr[count++] = page_address(sh->dev[i].page);
1075 xor_block(count, STRIPE_SIZE, ptr);
1077 if (method != CHECK_PARITY) {
1078 set_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
1079 set_bit(R5_LOCKED, &sh->dev[pd_idx].flags);
1081 clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
1084 static void compute_parity6(struct stripe_head *sh, int method)
1086 raid6_conf_t *conf = sh->raid_conf;
1087 int i, pd_idx = sh->pd_idx, qd_idx, d0_idx, disks = conf->raid_disks, count;
1089 /**** FIX THIS: This could be very bad if disks is close to 256 ****/
1092 qd_idx = raid6_next_disk(pd_idx, disks);
1093 d0_idx = raid6_next_disk(qd_idx, disks);
1095 PRINTK("compute_parity, stripe %llu, method %d\n",
1096 (unsigned long long)sh->sector, method);
1099 case READ_MODIFY_WRITE:
1100 BUG(); /* READ_MODIFY_WRITE N/A for RAID-6 */
1101 case RECONSTRUCT_WRITE:
1102 for (i= disks; i-- ;)
1103 if ( i != pd_idx && i != qd_idx && sh->dev[i].towrite ) {
1104 chosen = sh->dev[i].towrite;
1105 sh->dev[i].towrite = NULL;
1107 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
1108 wake_up(&conf->wait_for_overlap);
1110 BUG_ON(sh->dev[i].written);
1111 sh->dev[i].written = chosen;
1115 BUG(); /* Not implemented yet */
1118 for (i = disks; i--;)
1119 if (sh->dev[i].written) {
1120 sector_t sector = sh->dev[i].sector;
1121 struct bio *wbi = sh->dev[i].written;
1122 while (wbi && wbi->bi_sector < sector + STRIPE_SECTORS) {
1123 copy_data(1, wbi, sh->dev[i].page, sector);
1124 wbi = r5_next_bio(wbi, sector);
1127 set_bit(R5_LOCKED, &sh->dev[i].flags);
1128 set_bit(R5_UPTODATE, &sh->dev[i].flags);
1132 // case RECONSTRUCT_WRITE:
1133 // case CHECK_PARITY:
1134 // case UPDATE_PARITY:
1135 /* Note that unlike RAID-5, the ordering of the disks matters greatly. */
1136 /* FIX: Is this ordering of drives even remotely optimal? */
1140 ptrs[count++] = page_address(sh->dev[i].page);
1141 if (count <= disks-2 && !test_bit(R5_UPTODATE, &sh->dev[i].flags))
1142 printk("block %d/%d not uptodate on parity calc\n", i,count);
1143 i = raid6_next_disk(i, disks);
1144 } while ( i != d0_idx );
1148 raid6_call.gen_syndrome(disks, STRIPE_SIZE, ptrs);
1151 case RECONSTRUCT_WRITE:
1152 set_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
1153 set_bit(R5_UPTODATE, &sh->dev[qd_idx].flags);
1154 set_bit(R5_LOCKED, &sh->dev[pd_idx].flags);
1155 set_bit(R5_LOCKED, &sh->dev[qd_idx].flags);
1158 set_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
1159 set_bit(R5_UPTODATE, &sh->dev[qd_idx].flags);
1165 /* Compute one missing block */
1166 static void compute_block_1(struct stripe_head *sh, int dd_idx, int nozero)
1168 raid6_conf_t *conf = sh->raid_conf;
1169 int i, count, disks = conf->raid_disks;
1170 void *ptr[MAX_XOR_BLOCKS], *p;
1171 int pd_idx = sh->pd_idx;
1172 int qd_idx = raid6_next_disk(pd_idx, disks);
1174 PRINTK("compute_block_1, stripe %llu, idx %d\n",
1175 (unsigned long long)sh->sector, dd_idx);
1177 if ( dd_idx == qd_idx ) {
1178 /* We're actually computing the Q drive */
1179 compute_parity6(sh, UPDATE_PARITY);
1181 ptr[0] = page_address(sh->dev[dd_idx].page);
1182 if (!nozero) memset(ptr[0], 0, STRIPE_SIZE);
1184 for (i = disks ; i--; ) {
1185 if (i == dd_idx || i == qd_idx)
1187 p = page_address(sh->dev[i].page);
1188 if (test_bit(R5_UPTODATE, &sh->dev[i].flags))
1191 printk("compute_block() %d, stripe %llu, %d"
1192 " not present\n", dd_idx,
1193 (unsigned long long)sh->sector, i);
1198 xor_block(count, STRIPE_SIZE, ptr);
1199 if (!nozero) set_bit(R5_UPTODATE, &sh->dev[dd_idx].flags);
1200 else clear_bit(R5_UPTODATE, &sh->dev[dd_idx].flags);
1204 /* Compute two missing blocks */
1205 static void compute_block_2(struct stripe_head *sh, int dd_idx1, int dd_idx2)
1207 raid6_conf_t *conf = sh->raid_conf;
1208 int i, count, disks = conf->raid_disks;
1209 int pd_idx = sh->pd_idx;
1210 int qd_idx = raid6_next_disk(pd_idx, disks);
1211 int d0_idx = raid6_next_disk(qd_idx, disks);
1214 /* faila and failb are disk numbers relative to d0_idx */
1215 /* pd_idx become disks-2 and qd_idx become disks-1 */
1216 faila = (dd_idx1 < d0_idx) ? dd_idx1+(disks-d0_idx) : dd_idx1-d0_idx;
1217 failb = (dd_idx2 < d0_idx) ? dd_idx2+(disks-d0_idx) : dd_idx2-d0_idx;
1219 BUG_ON(faila == failb);
1220 if ( failb < faila ) { int tmp = faila; faila = failb; failb = tmp; }
1222 PRINTK("compute_block_2, stripe %llu, idx %d,%d (%d,%d)\n",
1223 (unsigned long long)sh->sector, dd_idx1, dd_idx2, faila, failb);
1225 if ( failb == disks-1 ) {
1226 /* Q disk is one of the missing disks */
1227 if ( faila == disks-2 ) {
1228 /* Missing P+Q, just recompute */
1229 compute_parity6(sh, UPDATE_PARITY);
1232 /* We're missing D+Q; recompute D from P */
1233 compute_block_1(sh, (dd_idx1 == qd_idx) ? dd_idx2 : dd_idx1, 0);
1234 compute_parity6(sh, UPDATE_PARITY); /* Is this necessary? */
1239 /* We're missing D+P or D+D; build pointer table */
1241 /**** FIX THIS: This could be very bad if disks is close to 256 ****/
1247 ptrs[count++] = page_address(sh->dev[i].page);
1248 i = raid6_next_disk(i, disks);
1249 if (i != dd_idx1 && i != dd_idx2 &&
1250 !test_bit(R5_UPTODATE, &sh->dev[i].flags))
1251 printk("compute_2 with missing block %d/%d\n", count, i);
1252 } while ( i != d0_idx );
1254 if ( failb == disks-2 ) {
1255 /* We're missing D+P. */
1256 raid6_datap_recov(disks, STRIPE_SIZE, faila, ptrs);
1258 /* We're missing D+D. */
1259 raid6_2data_recov(disks, STRIPE_SIZE, faila, failb, ptrs);
1262 /* Both the above update both missing blocks */
1263 set_bit(R5_UPTODATE, &sh->dev[dd_idx1].flags);
1264 set_bit(R5_UPTODATE, &sh->dev[dd_idx2].flags);
1271 * Each stripe/dev can have one or more bion attached.
1272 * toread/towrite point to the first in a chain.
1273 * The bi_next chain must be in order.
1275 static int add_stripe_bio(struct stripe_head *sh, struct bio *bi, int dd_idx, int forwrite)
1278 raid5_conf_t *conf = sh->raid_conf;
1281 PRINTK("adding bh b#%llu to stripe s#%llu\n",
1282 (unsigned long long)bi->bi_sector,
1283 (unsigned long long)sh->sector);
1286 spin_lock(&sh->lock);
1287 spin_lock_irq(&conf->device_lock);
1289 bip = &sh->dev[dd_idx].towrite;
1290 if (*bip == NULL && sh->dev[dd_idx].written == NULL)
1293 bip = &sh->dev[dd_idx].toread;
1294 while (*bip && (*bip)->bi_sector < bi->bi_sector) {
1295 if ((*bip)->bi_sector + ((*bip)->bi_size >> 9) > bi->bi_sector)
1297 bip = & (*bip)->bi_next;
1299 if (*bip && (*bip)->bi_sector < bi->bi_sector + ((bi->bi_size)>>9))
1302 BUG_ON(*bip && bi->bi_next && (*bip) != bi->bi_next);
1306 bi->bi_phys_segments ++;
1307 spin_unlock_irq(&conf->device_lock);
1308 spin_unlock(&sh->lock);
1310 PRINTK("added bi b#%llu to stripe s#%llu, disk %d.\n",
1311 (unsigned long long)bi->bi_sector,
1312 (unsigned long long)sh->sector, dd_idx);
1314 if (conf->mddev->bitmap && firstwrite) {
1315 bitmap_startwrite(conf->mddev->bitmap, sh->sector,
1317 sh->bm_seq = conf->seq_flush+1;
1318 set_bit(STRIPE_BIT_DELAY, &sh->state);
1322 /* check if page is covered */
1323 sector_t sector = sh->dev[dd_idx].sector;
1324 for (bi=sh->dev[dd_idx].towrite;
1325 sector < sh->dev[dd_idx].sector + STRIPE_SECTORS &&
1326 bi && bi->bi_sector <= sector;
1327 bi = r5_next_bio(bi, sh->dev[dd_idx].sector)) {
1328 if (bi->bi_sector + (bi->bi_size>>9) >= sector)
1329 sector = bi->bi_sector + (bi->bi_size>>9);
1331 if (sector >= sh->dev[dd_idx].sector + STRIPE_SECTORS)
1332 set_bit(R5_OVERWRITE, &sh->dev[dd_idx].flags);
1337 set_bit(R5_Overlap, &sh->dev[dd_idx].flags);
1338 spin_unlock_irq(&conf->device_lock);
1339 spin_unlock(&sh->lock);
1343 static void end_reshape(raid5_conf_t *conf);
1345 static int page_is_zero(struct page *p)
1347 char *a = page_address(p);
1348 return ((*(u32*)a) == 0 &&
1349 memcmp(a, a+4, STRIPE_SIZE-4)==0);
1352 static int stripe_to_pdidx(sector_t stripe, raid5_conf_t *conf, int disks)
1354 int sectors_per_chunk = conf->chunk_size >> 9;
1356 int chunk_offset = sector_div(stripe, sectors_per_chunk);
1358 raid5_compute_sector(stripe*(disks-1)*sectors_per_chunk
1359 + chunk_offset, disks, disks-1, &dd_idx, &pd_idx, conf);
1365 * handle_stripe - do things to a stripe.
1367 * We lock the stripe and then examine the state of various bits
1368 * to see what needs to be done.
1370 * return some read request which now have data
1371 * return some write requests which are safely on disc
1372 * schedule a read on some buffers
1373 * schedule a write of some buffers
1374 * return confirmation of parity correctness
1376 * Parity calculations are done inside the stripe lock
1377 * buffers are taken off read_list or write_list, and bh_cache buffers
1378 * get BH_Lock set before the stripe lock is released.
1382 static void handle_stripe5(struct stripe_head *sh)
1384 raid5_conf_t *conf = sh->raid_conf;
1385 int disks = sh->disks;
1386 struct bio *return_bi= NULL;
1389 int syncing, expanding, expanded;
1390 int locked=0, uptodate=0, to_read=0, to_write=0, failed=0, written=0;
1391 int non_overwrite = 0;
1395 PRINTK("handling stripe %llu, cnt=%d, pd_idx=%d\n",
1396 (unsigned long long)sh->sector, atomic_read(&sh->count),
1399 spin_lock(&sh->lock);
1400 clear_bit(STRIPE_HANDLE, &sh->state);
1401 clear_bit(STRIPE_DELAYED, &sh->state);
1403 syncing = test_bit(STRIPE_SYNCING, &sh->state);
1404 expanding = test_bit(STRIPE_EXPAND_SOURCE, &sh->state);
1405 expanded = test_bit(STRIPE_EXPAND_READY, &sh->state);
1406 /* Now to look around and see what can be done */
1409 for (i=disks; i--; ) {
1412 clear_bit(R5_Insync, &dev->flags);
1414 PRINTK("check %d: state 0x%lx read %p write %p written %p\n",
1415 i, dev->flags, dev->toread, dev->towrite, dev->written);
1416 /* maybe we can reply to a read */
1417 if (test_bit(R5_UPTODATE, &dev->flags) && dev->toread) {
1418 struct bio *rbi, *rbi2;
1419 PRINTK("Return read for disc %d\n", i);
1420 spin_lock_irq(&conf->device_lock);
1423 if (test_and_clear_bit(R5_Overlap, &dev->flags))
1424 wake_up(&conf->wait_for_overlap);
1425 spin_unlock_irq(&conf->device_lock);
1426 while (rbi && rbi->bi_sector < dev->sector + STRIPE_SECTORS) {
1427 copy_data(0, rbi, dev->page, dev->sector);
1428 rbi2 = r5_next_bio(rbi, dev->sector);
1429 spin_lock_irq(&conf->device_lock);
1430 if (--rbi->bi_phys_segments == 0) {
1431 rbi->bi_next = return_bi;
1434 spin_unlock_irq(&conf->device_lock);
1439 /* now count some things */
1440 if (test_bit(R5_LOCKED, &dev->flags)) locked++;
1441 if (test_bit(R5_UPTODATE, &dev->flags)) uptodate++;
1444 if (dev->toread) to_read++;
1447 if (!test_bit(R5_OVERWRITE, &dev->flags))
1450 if (dev->written) written++;
1451 rdev = rcu_dereference(conf->disks[i].rdev);
1452 if (!rdev || !test_bit(In_sync, &rdev->flags)) {
1453 /* The ReadError flag will just be confusing now */
1454 clear_bit(R5_ReadError, &dev->flags);
1455 clear_bit(R5_ReWrite, &dev->flags);
1457 if (!rdev || !test_bit(In_sync, &rdev->flags)
1458 || test_bit(R5_ReadError, &dev->flags)) {
1462 set_bit(R5_Insync, &dev->flags);
1465 PRINTK("locked=%d uptodate=%d to_read=%d"
1466 " to_write=%d failed=%d failed_num=%d\n",
1467 locked, uptodate, to_read, to_write, failed, failed_num);
1468 /* check if the array has lost two devices and, if so, some requests might
1471 if (failed > 1 && to_read+to_write+written) {
1472 for (i=disks; i--; ) {
1475 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
1478 rdev = rcu_dereference(conf->disks[i].rdev);
1479 if (rdev && test_bit(In_sync, &rdev->flags))
1480 /* multiple read failures in one stripe */
1481 md_error(conf->mddev, rdev);
1485 spin_lock_irq(&conf->device_lock);
1486 /* fail all writes first */
1487 bi = sh->dev[i].towrite;
1488 sh->dev[i].towrite = NULL;
1489 if (bi) { to_write--; bitmap_end = 1; }
1491 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
1492 wake_up(&conf->wait_for_overlap);
1494 while (bi && bi->bi_sector < sh->dev[i].sector + STRIPE_SECTORS){
1495 struct bio *nextbi = r5_next_bio(bi, sh->dev[i].sector);
1496 clear_bit(BIO_UPTODATE, &bi->bi_flags);
1497 if (--bi->bi_phys_segments == 0) {
1498 md_write_end(conf->mddev);
1499 bi->bi_next = return_bi;
1504 /* and fail all 'written' */
1505 bi = sh->dev[i].written;
1506 sh->dev[i].written = NULL;
1507 if (bi) bitmap_end = 1;
1508 while (bi && bi->bi_sector < sh->dev[i].sector + STRIPE_SECTORS) {
1509 struct bio *bi2 = r5_next_bio(bi, sh->dev[i].sector);
1510 clear_bit(BIO_UPTODATE, &bi->bi_flags);
1511 if (--bi->bi_phys_segments == 0) {
1512 md_write_end(conf->mddev);
1513 bi->bi_next = return_bi;
1519 /* fail any reads if this device is non-operational */
1520 if (!test_bit(R5_Insync, &sh->dev[i].flags) ||
1521 test_bit(R5_ReadError, &sh->dev[i].flags)) {
1522 bi = sh->dev[i].toread;
1523 sh->dev[i].toread = NULL;
1524 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
1525 wake_up(&conf->wait_for_overlap);
1527 while (bi && bi->bi_sector < sh->dev[i].sector + STRIPE_SECTORS){
1528 struct bio *nextbi = r5_next_bio(bi, sh->dev[i].sector);
1529 clear_bit(BIO_UPTODATE, &bi->bi_flags);
1530 if (--bi->bi_phys_segments == 0) {
1531 bi->bi_next = return_bi;
1537 spin_unlock_irq(&conf->device_lock);
1539 bitmap_endwrite(conf->mddev->bitmap, sh->sector,
1540 STRIPE_SECTORS, 0, 0);
1543 if (failed > 1 && syncing) {
1544 md_done_sync(conf->mddev, STRIPE_SECTORS,0);
1545 clear_bit(STRIPE_SYNCING, &sh->state);
1549 /* might be able to return some write requests if the parity block
1550 * is safe, or on a failed drive
1552 dev = &sh->dev[sh->pd_idx];
1554 ( (test_bit(R5_Insync, &dev->flags) && !test_bit(R5_LOCKED, &dev->flags) &&
1555 test_bit(R5_UPTODATE, &dev->flags))
1556 || (failed == 1 && failed_num == sh->pd_idx))
1558 /* any written block on an uptodate or failed drive can be returned.
1559 * Note that if we 'wrote' to a failed drive, it will be UPTODATE, but
1560 * never LOCKED, so we don't need to test 'failed' directly.
1562 for (i=disks; i--; )
1563 if (sh->dev[i].written) {
1565 if (!test_bit(R5_LOCKED, &dev->flags) &&
1566 test_bit(R5_UPTODATE, &dev->flags) ) {
1567 /* We can return any write requests */
1568 struct bio *wbi, *wbi2;
1570 PRINTK("Return write for disc %d\n", i);
1571 spin_lock_irq(&conf->device_lock);
1573 dev->written = NULL;
1574 while (wbi && wbi->bi_sector < dev->sector + STRIPE_SECTORS) {
1575 wbi2 = r5_next_bio(wbi, dev->sector);
1576 if (--wbi->bi_phys_segments == 0) {
1577 md_write_end(conf->mddev);
1578 wbi->bi_next = return_bi;
1583 if (dev->towrite == NULL)
1585 spin_unlock_irq(&conf->device_lock);
1587 bitmap_endwrite(conf->mddev->bitmap, sh->sector,
1589 !test_bit(STRIPE_DEGRADED, &sh->state), 0);
1594 /* Now we might consider reading some blocks, either to check/generate
1595 * parity, or to satisfy requests
1596 * or to load a block that is being partially written.
1598 if (to_read || non_overwrite || (syncing && (uptodate < disks)) || expanding) {
1599 for (i=disks; i--;) {
1601 if (!test_bit(R5_LOCKED, &dev->flags) && !test_bit(R5_UPTODATE, &dev->flags) &&
1603 (dev->towrite && !test_bit(R5_OVERWRITE, &dev->flags)) ||
1606 (failed && (sh->dev[failed_num].toread ||
1607 (sh->dev[failed_num].towrite && !test_bit(R5_OVERWRITE, &sh->dev[failed_num].flags))))
1610 /* we would like to get this block, possibly
1611 * by computing it, but we might not be able to
1613 if (uptodate == disks-1) {
1614 PRINTK("Computing block %d\n", i);
1615 compute_block(sh, i);
1617 } else if (test_bit(R5_Insync, &dev->flags)) {
1618 set_bit(R5_LOCKED, &dev->flags);
1619 set_bit(R5_Wantread, &dev->flags);
1621 /* if I am just reading this block and we don't have
1622 a failed drive, or any pending writes then sidestep the cache */
1623 if (sh->bh_read[i] && !sh->bh_read[i]->b_reqnext &&
1624 ! syncing && !failed && !to_write) {
1625 sh->bh_cache[i]->b_page = sh->bh_read[i]->b_page;
1626 sh->bh_cache[i]->b_data = sh->bh_read[i]->b_data;
1630 PRINTK("Reading block %d (sync=%d)\n",
1635 set_bit(STRIPE_HANDLE, &sh->state);
1638 /* now to consider writing and what else, if anything should be read */
1641 for (i=disks ; i--;) {
1642 /* would I have to read this buffer for read_modify_write */
1644 if ((dev->towrite || i == sh->pd_idx) &&
1645 (!test_bit(R5_LOCKED, &dev->flags)
1647 || sh->bh_page[i]!=bh->b_page
1650 !test_bit(R5_UPTODATE, &dev->flags)) {
1651 if (test_bit(R5_Insync, &dev->flags)
1652 /* && !(!mddev->insync && i == sh->pd_idx) */
1655 else rmw += 2*disks; /* cannot read it */
1657 /* Would I have to read this buffer for reconstruct_write */
1658 if (!test_bit(R5_OVERWRITE, &dev->flags) && i != sh->pd_idx &&
1659 (!test_bit(R5_LOCKED, &dev->flags)
1661 || sh->bh_page[i] != bh->b_page
1664 !test_bit(R5_UPTODATE, &dev->flags)) {
1665 if (test_bit(R5_Insync, &dev->flags)) rcw++;
1666 else rcw += 2*disks;
1669 PRINTK("for sector %llu, rmw=%d rcw=%d\n",
1670 (unsigned long long)sh->sector, rmw, rcw);
1671 set_bit(STRIPE_HANDLE, &sh->state);
1672 if (rmw < rcw && rmw > 0)
1673 /* prefer read-modify-write, but need to get some data */
1674 for (i=disks; i--;) {
1676 if ((dev->towrite || i == sh->pd_idx) &&
1677 !test_bit(R5_LOCKED, &dev->flags) && !test_bit(R5_UPTODATE, &dev->flags) &&
1678 test_bit(R5_Insync, &dev->flags)) {
1679 if (test_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
1681 PRINTK("Read_old block %d for r-m-w\n", i);
1682 set_bit(R5_LOCKED, &dev->flags);
1683 set_bit(R5_Wantread, &dev->flags);
1686 set_bit(STRIPE_DELAYED, &sh->state);
1687 set_bit(STRIPE_HANDLE, &sh->state);
1691 if (rcw <= rmw && rcw > 0)
1692 /* want reconstruct write, but need to get some data */
1693 for (i=disks; i--;) {
1695 if (!test_bit(R5_OVERWRITE, &dev->flags) && i != sh->pd_idx &&
1696 !test_bit(R5_LOCKED, &dev->flags) && !test_bit(R5_UPTODATE, &dev->flags) &&
1697 test_bit(R5_Insync, &dev->flags)) {
1698 if (test_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
1700 PRINTK("Read_old block %d for Reconstruct\n", i);
1701 set_bit(R5_LOCKED, &dev->flags);
1702 set_bit(R5_Wantread, &dev->flags);
1705 set_bit(STRIPE_DELAYED, &sh->state);
1706 set_bit(STRIPE_HANDLE, &sh->state);
1710 /* now if nothing is locked, and if we have enough data, we can start a write request */
1711 if (locked == 0 && (rcw == 0 ||rmw == 0) &&
1712 !test_bit(STRIPE_BIT_DELAY, &sh->state)) {
1713 PRINTK("Computing parity...\n");
1714 compute_parity5(sh, rcw==0 ? RECONSTRUCT_WRITE : READ_MODIFY_WRITE);
1715 /* now every locked buffer is ready to be written */
1717 if (test_bit(R5_LOCKED, &sh->dev[i].flags)) {
1718 PRINTK("Writing block %d\n", i);
1720 set_bit(R5_Wantwrite, &sh->dev[i].flags);
1721 if (!test_bit(R5_Insync, &sh->dev[i].flags)
1722 || (i==sh->pd_idx && failed == 0))
1723 set_bit(STRIPE_INSYNC, &sh->state);
1725 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
1726 atomic_dec(&conf->preread_active_stripes);
1727 if (atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD)
1728 md_wakeup_thread(conf->mddev->thread);
1733 /* maybe we need to check and possibly fix the parity for this stripe
1734 * Any reads will already have been scheduled, so we just see if enough data
1737 if (syncing && locked == 0 &&
1738 !test_bit(STRIPE_INSYNC, &sh->state)) {
1739 set_bit(STRIPE_HANDLE, &sh->state);
1741 BUG_ON(uptodate != disks);
1742 compute_parity5(sh, CHECK_PARITY);
1744 if (page_is_zero(sh->dev[sh->pd_idx].page)) {
1745 /* parity is correct (on disc, not in buffer any more) */
1746 set_bit(STRIPE_INSYNC, &sh->state);
1748 conf->mddev->resync_mismatches += STRIPE_SECTORS;
1749 if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery))
1750 /* don't try to repair!! */
1751 set_bit(STRIPE_INSYNC, &sh->state);
1753 compute_block(sh, sh->pd_idx);
1758 if (!test_bit(STRIPE_INSYNC, &sh->state)) {
1759 /* either failed parity check, or recovery is happening */
1761 failed_num = sh->pd_idx;
1762 dev = &sh->dev[failed_num];
1763 BUG_ON(!test_bit(R5_UPTODATE, &dev->flags));
1764 BUG_ON(uptodate != disks);
1766 set_bit(R5_LOCKED, &dev->flags);
1767 set_bit(R5_Wantwrite, &dev->flags);
1768 clear_bit(STRIPE_DEGRADED, &sh->state);
1770 set_bit(STRIPE_INSYNC, &sh->state);
1773 if (syncing && locked == 0 && test_bit(STRIPE_INSYNC, &sh->state)) {
1774 md_done_sync(conf->mddev, STRIPE_SECTORS,1);
1775 clear_bit(STRIPE_SYNCING, &sh->state);
1778 /* If the failed drive is just a ReadError, then we might need to progress
1779 * the repair/check process
1781 if (failed == 1 && ! conf->mddev->ro &&
1782 test_bit(R5_ReadError, &sh->dev[failed_num].flags)
1783 && !test_bit(R5_LOCKED, &sh->dev[failed_num].flags)
1784 && test_bit(R5_UPTODATE, &sh->dev[failed_num].flags)
1786 dev = &sh->dev[failed_num];
1787 if (!test_bit(R5_ReWrite, &dev->flags)) {
1788 set_bit(R5_Wantwrite, &dev->flags);
1789 set_bit(R5_ReWrite, &dev->flags);
1790 set_bit(R5_LOCKED, &dev->flags);
1793 /* let's read it back */
1794 set_bit(R5_Wantread, &dev->flags);
1795 set_bit(R5_LOCKED, &dev->flags);
1800 if (expanded && test_bit(STRIPE_EXPANDING, &sh->state)) {
1801 /* Need to write out all blocks after computing parity */
1802 sh->disks = conf->raid_disks;
1803 sh->pd_idx = stripe_to_pdidx(sh->sector, conf, conf->raid_disks);
1804 compute_parity5(sh, RECONSTRUCT_WRITE);
1805 for (i= conf->raid_disks; i--;) {
1806 set_bit(R5_LOCKED, &sh->dev[i].flags);
1808 set_bit(R5_Wantwrite, &sh->dev[i].flags);
1810 clear_bit(STRIPE_EXPANDING, &sh->state);
1811 } else if (expanded) {
1812 clear_bit(STRIPE_EXPAND_READY, &sh->state);
1813 atomic_dec(&conf->reshape_stripes);
1814 wake_up(&conf->wait_for_overlap);
1815 md_done_sync(conf->mddev, STRIPE_SECTORS, 1);
1818 if (expanding && locked == 0) {
1819 /* We have read all the blocks in this stripe and now we need to
1820 * copy some of them into a target stripe for expand.
1822 clear_bit(STRIPE_EXPAND_SOURCE, &sh->state);
1823 for (i=0; i< sh->disks; i++)
1824 if (i != sh->pd_idx) {
1825 int dd_idx, pd_idx, j;
1826 struct stripe_head *sh2;
1828 sector_t bn = compute_blocknr(sh, i);
1829 sector_t s = raid5_compute_sector(bn, conf->raid_disks,
1831 &dd_idx, &pd_idx, conf);
1832 sh2 = get_active_stripe(conf, s, conf->raid_disks, pd_idx, 1);
1834 /* so far only the early blocks of this stripe
1835 * have been requested. When later blocks
1836 * get requested, we will try again
1839 if(!test_bit(STRIPE_EXPANDING, &sh2->state) ||
1840 test_bit(R5_Expanded, &sh2->dev[dd_idx].flags)) {
1841 /* must have already done this block */
1842 release_stripe(sh2);
1845 memcpy(page_address(sh2->dev[dd_idx].page),
1846 page_address(sh->dev[i].page),
1848 set_bit(R5_Expanded, &sh2->dev[dd_idx].flags);
1849 set_bit(R5_UPTODATE, &sh2->dev[dd_idx].flags);
1850 for (j=0; j<conf->raid_disks; j++)
1851 if (j != sh2->pd_idx &&
1852 !test_bit(R5_Expanded, &sh2->dev[j].flags))
1854 if (j == conf->raid_disks) {
1855 set_bit(STRIPE_EXPAND_READY, &sh2->state);
1856 set_bit(STRIPE_HANDLE, &sh2->state);
1858 release_stripe(sh2);
1862 spin_unlock(&sh->lock);
1864 while ((bi=return_bi)) {
1865 int bytes = bi->bi_size;
1867 return_bi = bi->bi_next;
1870 bi->bi_end_io(bi, bytes, 0);
1872 for (i=disks; i-- ;) {
1876 if (test_and_clear_bit(R5_Wantwrite, &sh->dev[i].flags))
1878 else if (test_and_clear_bit(R5_Wantread, &sh->dev[i].flags))
1883 bi = &sh->dev[i].req;
1887 bi->bi_end_io = raid5_end_write_request;
1889 bi->bi_end_io = raid5_end_read_request;
1892 rdev = rcu_dereference(conf->disks[i].rdev);
1893 if (rdev && test_bit(Faulty, &rdev->flags))
1896 atomic_inc(&rdev->nr_pending);
1900 if (syncing || expanding || expanded)
1901 md_sync_acct(rdev->bdev, STRIPE_SECTORS);
1903 bi->bi_bdev = rdev->bdev;
1904 PRINTK("for %llu schedule op %ld on disc %d\n",
1905 (unsigned long long)sh->sector, bi->bi_rw, i);
1906 atomic_inc(&sh->count);
1907 bi->bi_sector = sh->sector + rdev->data_offset;
1908 bi->bi_flags = 1 << BIO_UPTODATE;
1910 bi->bi_max_vecs = 1;
1912 bi->bi_io_vec = &sh->dev[i].vec;
1913 bi->bi_io_vec[0].bv_len = STRIPE_SIZE;
1914 bi->bi_io_vec[0].bv_offset = 0;
1915 bi->bi_size = STRIPE_SIZE;
1918 test_bit(R5_ReWrite, &sh->dev[i].flags))
1919 atomic_add(STRIPE_SECTORS, &rdev->corrected_errors);
1920 generic_make_request(bi);
1923 set_bit(STRIPE_DEGRADED, &sh->state);
1924 PRINTK("skip op %ld on disc %d for sector %llu\n",
1925 bi->bi_rw, i, (unsigned long long)sh->sector);
1926 clear_bit(R5_LOCKED, &sh->dev[i].flags);
1927 set_bit(STRIPE_HANDLE, &sh->state);
1932 static void handle_stripe6(struct stripe_head *sh, struct page *tmp_page)
1934 raid6_conf_t *conf = sh->raid_conf;
1935 int disks = conf->raid_disks;
1936 struct bio *return_bi= NULL;
1940 int locked=0, uptodate=0, to_read=0, to_write=0, failed=0, written=0;
1941 int non_overwrite = 0;
1942 int failed_num[2] = {0, 0};
1943 struct r5dev *dev, *pdev, *qdev;
1944 int pd_idx = sh->pd_idx;
1945 int qd_idx = raid6_next_disk(pd_idx, disks);
1946 int p_failed, q_failed;
1948 PRINTK("handling stripe %llu, state=%#lx cnt=%d, pd_idx=%d, qd_idx=%d\n",
1949 (unsigned long long)sh->sector, sh->state, atomic_read(&sh->count),
1952 spin_lock(&sh->lock);
1953 clear_bit(STRIPE_HANDLE, &sh->state);
1954 clear_bit(STRIPE_DELAYED, &sh->state);
1956 syncing = test_bit(STRIPE_SYNCING, &sh->state);
1957 /* Now to look around and see what can be done */
1960 for (i=disks; i--; ) {
1963 clear_bit(R5_Insync, &dev->flags);
1965 PRINTK("check %d: state 0x%lx read %p write %p written %p\n",
1966 i, dev->flags, dev->toread, dev->towrite, dev->written);
1967 /* maybe we can reply to a read */
1968 if (test_bit(R5_UPTODATE, &dev->flags) && dev->toread) {
1969 struct bio *rbi, *rbi2;
1970 PRINTK("Return read for disc %d\n", i);
1971 spin_lock_irq(&conf->device_lock);
1974 if (test_and_clear_bit(R5_Overlap, &dev->flags))
1975 wake_up(&conf->wait_for_overlap);
1976 spin_unlock_irq(&conf->device_lock);
1977 while (rbi && rbi->bi_sector < dev->sector + STRIPE_SECTORS) {
1978 copy_data(0, rbi, dev->page, dev->sector);
1979 rbi2 = r5_next_bio(rbi, dev->sector);
1980 spin_lock_irq(&conf->device_lock);
1981 if (--rbi->bi_phys_segments == 0) {
1982 rbi->bi_next = return_bi;
1985 spin_unlock_irq(&conf->device_lock);
1990 /* now count some things */
1991 if (test_bit(R5_LOCKED, &dev->flags)) locked++;
1992 if (test_bit(R5_UPTODATE, &dev->flags)) uptodate++;
1995 if (dev->toread) to_read++;
1998 if (!test_bit(R5_OVERWRITE, &dev->flags))
2001 if (dev->written) written++;
2002 rdev = rcu_dereference(conf->disks[i].rdev);
2003 if (!rdev || !test_bit(In_sync, &rdev->flags)) {
2004 /* The ReadError flag will just be confusing now */
2005 clear_bit(R5_ReadError, &dev->flags);
2006 clear_bit(R5_ReWrite, &dev->flags);
2008 if (!rdev || !test_bit(In_sync, &rdev->flags)
2009 || test_bit(R5_ReadError, &dev->flags)) {
2011 failed_num[failed] = i;
2014 set_bit(R5_Insync, &dev->flags);
2017 PRINTK("locked=%d uptodate=%d to_read=%d"
2018 " to_write=%d failed=%d failed_num=%d,%d\n",
2019 locked, uptodate, to_read, to_write, failed,
2020 failed_num[0], failed_num[1]);
2021 /* check if the array has lost >2 devices and, if so, some requests might
2024 if (failed > 2 && to_read+to_write+written) {
2025 for (i=disks; i--; ) {
2028 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
2031 rdev = rcu_dereference(conf->disks[i].rdev);
2032 if (rdev && test_bit(In_sync, &rdev->flags))
2033 /* multiple read failures in one stripe */
2034 md_error(conf->mddev, rdev);
2038 spin_lock_irq(&conf->device_lock);
2039 /* fail all writes first */
2040 bi = sh->dev[i].towrite;
2041 sh->dev[i].towrite = NULL;
2042 if (bi) { to_write--; bitmap_end = 1; }
2044 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
2045 wake_up(&conf->wait_for_overlap);
2047 while (bi && bi->bi_sector < sh->dev[i].sector + STRIPE_SECTORS){
2048 struct bio *nextbi = r5_next_bio(bi, sh->dev[i].sector);
2049 clear_bit(BIO_UPTODATE, &bi->bi_flags);
2050 if (--bi->bi_phys_segments == 0) {
2051 md_write_end(conf->mddev);
2052 bi->bi_next = return_bi;
2057 /* and fail all 'written' */
2058 bi = sh->dev[i].written;
2059 sh->dev[i].written = NULL;
2060 if (bi) bitmap_end = 1;
2061 while (bi && bi->bi_sector < sh->dev[i].sector + STRIPE_SECTORS) {
2062 struct bio *bi2 = r5_next_bio(bi, sh->dev[i].sector);
2063 clear_bit(BIO_UPTODATE, &bi->bi_flags);
2064 if (--bi->bi_phys_segments == 0) {
2065 md_write_end(conf->mddev);
2066 bi->bi_next = return_bi;
2072 /* fail any reads if this device is non-operational */
2073 if (!test_bit(R5_Insync, &sh->dev[i].flags) ||
2074 test_bit(R5_ReadError, &sh->dev[i].flags)) {
2075 bi = sh->dev[i].toread;
2076 sh->dev[i].toread = NULL;
2077 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
2078 wake_up(&conf->wait_for_overlap);
2080 while (bi && bi->bi_sector < sh->dev[i].sector + STRIPE_SECTORS){
2081 struct bio *nextbi = r5_next_bio(bi, sh->dev[i].sector);
2082 clear_bit(BIO_UPTODATE, &bi->bi_flags);
2083 if (--bi->bi_phys_segments == 0) {
2084 bi->bi_next = return_bi;
2090 spin_unlock_irq(&conf->device_lock);
2092 bitmap_endwrite(conf->mddev->bitmap, sh->sector,
2093 STRIPE_SECTORS, 0, 0);
2096 if (failed > 2 && syncing) {
2097 md_done_sync(conf->mddev, STRIPE_SECTORS,0);
2098 clear_bit(STRIPE_SYNCING, &sh->state);
2103 * might be able to return some write requests if the parity blocks
2104 * are safe, or on a failed drive
2106 pdev = &sh->dev[pd_idx];
2107 p_failed = (failed >= 1 && failed_num[0] == pd_idx)
2108 || (failed >= 2 && failed_num[1] == pd_idx);
2109 qdev = &sh->dev[qd_idx];
2110 q_failed = (failed >= 1 && failed_num[0] == qd_idx)
2111 || (failed >= 2 && failed_num[1] == qd_idx);
2114 ( p_failed || ((test_bit(R5_Insync, &pdev->flags)
2115 && !test_bit(R5_LOCKED, &pdev->flags)
2116 && test_bit(R5_UPTODATE, &pdev->flags))) ) &&
2117 ( q_failed || ((test_bit(R5_Insync, &qdev->flags)
2118 && !test_bit(R5_LOCKED, &qdev->flags)
2119 && test_bit(R5_UPTODATE, &qdev->flags))) ) ) {
2120 /* any written block on an uptodate or failed drive can be
2121 * returned. Note that if we 'wrote' to a failed drive,
2122 * it will be UPTODATE, but never LOCKED, so we don't need
2123 * to test 'failed' directly.
2125 for (i=disks; i--; )
2126 if (sh->dev[i].written) {
2128 if (!test_bit(R5_LOCKED, &dev->flags) &&
2129 test_bit(R5_UPTODATE, &dev->flags) ) {
2130 /* We can return any write requests */
2132 struct bio *wbi, *wbi2;
2133 PRINTK("Return write for stripe %llu disc %d\n",
2134 (unsigned long long)sh->sector, i);
2135 spin_lock_irq(&conf->device_lock);
2137 dev->written = NULL;
2138 while (wbi && wbi->bi_sector < dev->sector + STRIPE_SECTORS) {
2139 wbi2 = r5_next_bio(wbi, dev->sector);
2140 if (--wbi->bi_phys_segments == 0) {
2141 md_write_end(conf->mddev);
2142 wbi->bi_next = return_bi;
2147 if (dev->towrite == NULL)
2149 spin_unlock_irq(&conf->device_lock);
2151 bitmap_endwrite(conf->mddev->bitmap, sh->sector,
2153 !test_bit(STRIPE_DEGRADED, &sh->state), 0);
2158 /* Now we might consider reading some blocks, either to check/generate
2159 * parity, or to satisfy requests
2160 * or to load a block that is being partially written.
2162 if (to_read || non_overwrite || (to_write && failed) || (syncing && (uptodate < disks))) {
2163 for (i=disks; i--;) {
2165 if (!test_bit(R5_LOCKED, &dev->flags) && !test_bit(R5_UPTODATE, &dev->flags) &&
2167 (dev->towrite && !test_bit(R5_OVERWRITE, &dev->flags)) ||
2169 (failed >= 1 && (sh->dev[failed_num[0]].toread || to_write)) ||
2170 (failed >= 2 && (sh->dev[failed_num[1]].toread || to_write))
2173 /* we would like to get this block, possibly
2174 * by computing it, but we might not be able to
2176 if (uptodate == disks-1) {
2177 PRINTK("Computing stripe %llu block %d\n",
2178 (unsigned long long)sh->sector, i);
2179 compute_block_1(sh, i, 0);
2181 } else if ( uptodate == disks-2 && failed >= 2 ) {
2182 /* Computing 2-failure is *very* expensive; only do it if failed >= 2 */
2184 for (other=disks; other--;) {
2187 if ( !test_bit(R5_UPTODATE, &sh->dev[other].flags) )
2191 PRINTK("Computing stripe %llu blocks %d,%d\n",
2192 (unsigned long long)sh->sector, i, other);
2193 compute_block_2(sh, i, other);
2195 } else if (test_bit(R5_Insync, &dev->flags)) {
2196 set_bit(R5_LOCKED, &dev->flags);
2197 set_bit(R5_Wantread, &dev->flags);
2199 /* if I am just reading this block and we don't have
2200 a failed drive, or any pending writes then sidestep the cache */
2201 if (sh->bh_read[i] && !sh->bh_read[i]->b_reqnext &&
2202 ! syncing && !failed && !to_write) {
2203 sh->bh_cache[i]->b_page = sh->bh_read[i]->b_page;
2204 sh->bh_cache[i]->b_data = sh->bh_read[i]->b_data;
2208 PRINTK("Reading block %d (sync=%d)\n",
2213 set_bit(STRIPE_HANDLE, &sh->state);
2216 /* now to consider writing and what else, if anything should be read */
2218 int rcw=0, must_compute=0;
2219 for (i=disks ; i--;) {
2221 /* Would I have to read this buffer for reconstruct_write */
2222 if (!test_bit(R5_OVERWRITE, &dev->flags)
2223 && i != pd_idx && i != qd_idx
2224 && (!test_bit(R5_LOCKED, &dev->flags)
2226 || sh->bh_page[i] != bh->b_page
2229 !test_bit(R5_UPTODATE, &dev->flags)) {
2230 if (test_bit(R5_Insync, &dev->flags)) rcw++;
2232 PRINTK("raid6: must_compute: disk %d flags=%#lx\n", i, dev->flags);
2237 PRINTK("for sector %llu, rcw=%d, must_compute=%d\n",
2238 (unsigned long long)sh->sector, rcw, must_compute);
2239 set_bit(STRIPE_HANDLE, &sh->state);
2242 /* want reconstruct write, but need to get some data */
2243 for (i=disks; i--;) {
2245 if (!test_bit(R5_OVERWRITE, &dev->flags)
2246 && !(failed == 0 && (i == pd_idx || i == qd_idx))
2247 && !test_bit(R5_LOCKED, &dev->flags) && !test_bit(R5_UPTODATE, &dev->flags) &&
2248 test_bit(R5_Insync, &dev->flags)) {
2249 if (test_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
2251 PRINTK("Read_old stripe %llu block %d for Reconstruct\n",
2252 (unsigned long long)sh->sector, i);
2253 set_bit(R5_LOCKED, &dev->flags);
2254 set_bit(R5_Wantread, &dev->flags);
2257 PRINTK("Request delayed stripe %llu block %d for Reconstruct\n",
2258 (unsigned long long)sh->sector, i);
2259 set_bit(STRIPE_DELAYED, &sh->state);
2260 set_bit(STRIPE_HANDLE, &sh->state);
2264 /* now if nothing is locked, and if we have enough data, we can start a write request */
2265 if (locked == 0 && rcw == 0 &&
2266 !test_bit(STRIPE_BIT_DELAY, &sh->state)) {
2267 if ( must_compute > 0 ) {
2268 /* We have failed blocks and need to compute them */
2271 case 1: compute_block_1(sh, failed_num[0], 0); break;
2272 case 2: compute_block_2(sh, failed_num[0], failed_num[1]); break;
2273 default: BUG(); /* This request should have been failed? */
2277 PRINTK("Computing parity for stripe %llu\n", (unsigned long long)sh->sector);
2278 compute_parity6(sh, RECONSTRUCT_WRITE);
2279 /* now every locked buffer is ready to be written */
2281 if (test_bit(R5_LOCKED, &sh->dev[i].flags)) {
2282 PRINTK("Writing stripe %llu block %d\n",
2283 (unsigned long long)sh->sector, i);
2285 set_bit(R5_Wantwrite, &sh->dev[i].flags);
2287 /* after a RECONSTRUCT_WRITE, the stripe MUST be in-sync */
2288 set_bit(STRIPE_INSYNC, &sh->state);
2290 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2291 atomic_dec(&conf->preread_active_stripes);
2292 if (atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD)
2293 md_wakeup_thread(conf->mddev->thread);
2298 /* maybe we need to check and possibly fix the parity for this stripe
2299 * Any reads will already have been scheduled, so we just see if enough data
2302 if (syncing && locked == 0 && !test_bit(STRIPE_INSYNC, &sh->state)) {
2303 int update_p = 0, update_q = 0;
2306 set_bit(STRIPE_HANDLE, &sh->state);
2309 BUG_ON(uptodate < disks);
2310 /* Want to check and possibly repair P and Q.
2311 * However there could be one 'failed' device, in which
2312 * case we can only check one of them, possibly using the
2313 * other to generate missing data
2316 /* If !tmp_page, we cannot do the calculations,
2317 * but as we have set STRIPE_HANDLE, we will soon be called
2318 * by stripe_handle with a tmp_page - just wait until then.
2321 if (failed == q_failed) {
2322 /* The only possible failed device holds 'Q', so it makes
2323 * sense to check P (If anything else were failed, we would
2324 * have used P to recreate it).
2326 compute_block_1(sh, pd_idx, 1);
2327 if (!page_is_zero(sh->dev[pd_idx].page)) {
2328 compute_block_1(sh,pd_idx,0);
2332 if (!q_failed && failed < 2) {
2333 /* q is not failed, and we didn't use it to generate
2334 * anything, so it makes sense to check it
2336 memcpy(page_address(tmp_page),
2337 page_address(sh->dev[qd_idx].page),
2339 compute_parity6(sh, UPDATE_PARITY);
2340 if (memcmp(page_address(tmp_page),
2341 page_address(sh->dev[qd_idx].page),
2343 clear_bit(STRIPE_INSYNC, &sh->state);
2347 if (update_p || update_q) {
2348 conf->mddev->resync_mismatches += STRIPE_SECTORS;
2349 if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery))
2350 /* don't try to repair!! */
2351 update_p = update_q = 0;
2354 /* now write out any block on a failed drive,
2355 * or P or Q if they need it
2359 dev = &sh->dev[failed_num[1]];
2361 set_bit(R5_LOCKED, &dev->flags);
2362 set_bit(R5_Wantwrite, &dev->flags);
2365 dev = &sh->dev[failed_num[0]];
2367 set_bit(R5_LOCKED, &dev->flags);
2368 set_bit(R5_Wantwrite, &dev->flags);
2372 dev = &sh->dev[pd_idx];
2374 set_bit(R5_LOCKED, &dev->flags);
2375 set_bit(R5_Wantwrite, &dev->flags);
2378 dev = &sh->dev[qd_idx];
2380 set_bit(R5_LOCKED, &dev->flags);
2381 set_bit(R5_Wantwrite, &dev->flags);
2383 clear_bit(STRIPE_DEGRADED, &sh->state);
2385 set_bit(STRIPE_INSYNC, &sh->state);
2389 if (syncing && locked == 0 && test_bit(STRIPE_INSYNC, &sh->state)) {
2390 md_done_sync(conf->mddev, STRIPE_SECTORS,1);
2391 clear_bit(STRIPE_SYNCING, &sh->state);
2394 /* If the failed drives are just a ReadError, then we might need
2395 * to progress the repair/check process
2397 if (failed <= 2 && ! conf->mddev->ro)
2398 for (i=0; i<failed;i++) {
2399 dev = &sh->dev[failed_num[i]];
2400 if (test_bit(R5_ReadError, &dev->flags)
2401 && !test_bit(R5_LOCKED, &dev->flags)
2402 && test_bit(R5_UPTODATE, &dev->flags)
2404 if (!test_bit(R5_ReWrite, &dev->flags)) {
2405 set_bit(R5_Wantwrite, &dev->flags);
2406 set_bit(R5_ReWrite, &dev->flags);
2407 set_bit(R5_LOCKED, &dev->flags);
2409 /* let's read it back */
2410 set_bit(R5_Wantread, &dev->flags);
2411 set_bit(R5_LOCKED, &dev->flags);
2415 spin_unlock(&sh->lock);
2417 while ((bi=return_bi)) {
2418 int bytes = bi->bi_size;
2420 return_bi = bi->bi_next;
2423 bi->bi_end_io(bi, bytes, 0);
2425 for (i=disks; i-- ;) {
2429 if (test_and_clear_bit(R5_Wantwrite, &sh->dev[i].flags))
2431 else if (test_and_clear_bit(R5_Wantread, &sh->dev[i].flags))
2436 bi = &sh->dev[i].req;
2440 bi->bi_end_io = raid5_end_write_request;
2442 bi->bi_end_io = raid5_end_read_request;
2445 rdev = rcu_dereference(conf->disks[i].rdev);
2446 if (rdev && test_bit(Faulty, &rdev->flags))
2449 atomic_inc(&rdev->nr_pending);
2454 md_sync_acct(rdev->bdev, STRIPE_SECTORS);
2456 bi->bi_bdev = rdev->bdev;
2457 PRINTK("for %llu schedule op %ld on disc %d\n",
2458 (unsigned long long)sh->sector, bi->bi_rw, i);
2459 atomic_inc(&sh->count);
2460 bi->bi_sector = sh->sector + rdev->data_offset;
2461 bi->bi_flags = 1 << BIO_UPTODATE;
2463 bi->bi_max_vecs = 1;
2465 bi->bi_io_vec = &sh->dev[i].vec;
2466 bi->bi_io_vec[0].bv_len = STRIPE_SIZE;
2467 bi->bi_io_vec[0].bv_offset = 0;
2468 bi->bi_size = STRIPE_SIZE;
2471 test_bit(R5_ReWrite, &sh->dev[i].flags))
2472 atomic_add(STRIPE_SECTORS, &rdev->corrected_errors);
2473 generic_make_request(bi);
2476 set_bit(STRIPE_DEGRADED, &sh->state);
2477 PRINTK("skip op %ld on disc %d for sector %llu\n",
2478 bi->bi_rw, i, (unsigned long long)sh->sector);
2479 clear_bit(R5_LOCKED, &sh->dev[i].flags);
2480 set_bit(STRIPE_HANDLE, &sh->state);
2485 static void handle_stripe(struct stripe_head *sh, struct page *tmp_page)
2487 if (sh->raid_conf->level == 6)
2488 handle_stripe6(sh, tmp_page);
2495 static void raid5_activate_delayed(raid5_conf_t *conf)
2497 if (atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD) {
2498 while (!list_empty(&conf->delayed_list)) {
2499 struct list_head *l = conf->delayed_list.next;
2500 struct stripe_head *sh;
2501 sh = list_entry(l, struct stripe_head, lru);
2503 clear_bit(STRIPE_DELAYED, &sh->state);
2504 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
2505 atomic_inc(&conf->preread_active_stripes);
2506 list_add_tail(&sh->lru, &conf->handle_list);
2511 static void activate_bit_delay(raid5_conf_t *conf)
2513 /* device_lock is held */
2514 struct list_head head;
2515 list_add(&head, &conf->bitmap_list);
2516 list_del_init(&conf->bitmap_list);
2517 while (!list_empty(&head)) {
2518 struct stripe_head *sh = list_entry(head.next, struct stripe_head, lru);
2519 list_del_init(&sh->lru);
2520 atomic_inc(&sh->count);
2521 __release_stripe(conf, sh);
2525 static void unplug_slaves(mddev_t *mddev)
2527 raid5_conf_t *conf = mddev_to_conf(mddev);
2531 for (i=0; i<mddev->raid_disks; i++) {
2532 mdk_rdev_t *rdev = rcu_dereference(conf->disks[i].rdev);
2533 if (rdev && !test_bit(Faulty, &rdev->flags) && atomic_read(&rdev->nr_pending)) {
2534 request_queue_t *r_queue = bdev_get_queue(rdev->bdev);
2536 atomic_inc(&rdev->nr_pending);
2539 if (r_queue->unplug_fn)
2540 r_queue->unplug_fn(r_queue);
2542 rdev_dec_pending(rdev, mddev);
2549 static void raid5_unplug_device(request_queue_t *q)
2551 mddev_t *mddev = q->queuedata;
2552 raid5_conf_t *conf = mddev_to_conf(mddev);
2553 unsigned long flags;
2555 spin_lock_irqsave(&conf->device_lock, flags);
2557 if (blk_remove_plug(q)) {
2559 raid5_activate_delayed(conf);
2561 md_wakeup_thread(mddev->thread);
2563 spin_unlock_irqrestore(&conf->device_lock, flags);
2565 unplug_slaves(mddev);
2568 static int raid5_issue_flush(request_queue_t *q, struct gendisk *disk,
2569 sector_t *error_sector)
2571 mddev_t *mddev = q->queuedata;
2572 raid5_conf_t *conf = mddev_to_conf(mddev);
2576 for (i=0; i<mddev->raid_disks && ret == 0; i++) {
2577 mdk_rdev_t *rdev = rcu_dereference(conf->disks[i].rdev);
2578 if (rdev && !test_bit(Faulty, &rdev->flags)) {
2579 struct block_device *bdev = rdev->bdev;
2580 request_queue_t *r_queue = bdev_get_queue(bdev);
2582 if (!r_queue->issue_flush_fn)
2585 atomic_inc(&rdev->nr_pending);
2587 ret = r_queue->issue_flush_fn(r_queue, bdev->bd_disk,
2589 rdev_dec_pending(rdev, mddev);
2598 static int raid5_congested(void *data, int bits)
2600 mddev_t *mddev = data;
2601 raid5_conf_t *conf = mddev_to_conf(mddev);
2603 /* No difference between reads and writes. Just check
2604 * how busy the stripe_cache is
2606 if (conf->inactive_blocked)
2610 if (list_empty_careful(&conf->inactive_list))
2616 /* We want read requests to align with chunks where possible,
2617 * but write requests don't need to.
2619 static int raid5_mergeable_bvec(request_queue_t *q, struct bio *bio, struct bio_vec *biovec)
2621 mddev_t *mddev = q->queuedata;
2622 sector_t sector = bio->bi_sector + get_start_sect(bio->bi_bdev);
2624 unsigned int chunk_sectors = mddev->chunk_size >> 9;
2625 unsigned int bio_sectors = bio->bi_size >> 9;
2627 if (bio_data_dir(bio))
2628 return biovec->bv_len; /* always allow writes to be mergeable */
2630 max = (chunk_sectors - ((sector & (chunk_sectors - 1)) + bio_sectors)) << 9;
2631 if (max < 0) max = 0;
2632 if (max <= biovec->bv_len && bio_sectors == 0)
2633 return biovec->bv_len;
2639 static int in_chunk_boundary(mddev_t *mddev, struct bio *bio)
2641 sector_t sector = bio->bi_sector + get_start_sect(bio->bi_bdev);
2642 unsigned int chunk_sectors = mddev->chunk_size >> 9;
2643 unsigned int bio_sectors = bio->bi_size >> 9;
2645 return chunk_sectors >=
2646 ((sector & (chunk_sectors - 1)) + bio_sectors);
2650 * add bio to the retry LIFO ( in O(1) ... we are in interrupt )
2651 * later sampled by raid5d.
2653 static void add_bio_to_retry(struct bio *bi,raid5_conf_t *conf)
2655 unsigned long flags;
2657 spin_lock_irqsave(&conf->device_lock, flags);
2659 bi->bi_next = conf->retry_read_aligned_list;
2660 conf->retry_read_aligned_list = bi;
2662 spin_unlock_irqrestore(&conf->device_lock, flags);
2663 md_wakeup_thread(conf->mddev->thread);
2667 static struct bio *remove_bio_from_retry(raid5_conf_t *conf)
2671 bi = conf->retry_read_aligned;
2673 conf->retry_read_aligned = NULL;
2676 bi = conf->retry_read_aligned_list;
2678 conf->retry_read_aligned = bi->bi_next;
2680 bi->bi_phys_segments = 1; /* biased count of active stripes */
2681 bi->bi_hw_segments = 0; /* count of processed stripes */
2689 * The "raid5_align_endio" should check if the read succeeded and if it
2690 * did, call bio_endio on the original bio (having bio_put the new bio
2692 * If the read failed..
2694 static int raid5_align_endio(struct bio *bi, unsigned int bytes, int error)
2696 struct bio* raid_bi = bi->bi_private;
2699 int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
2706 mddev = raid_bi->bi_bdev->bd_disk->queue->queuedata;
2707 conf = mddev_to_conf(mddev);
2708 rdev = (void*)raid_bi->bi_next;
2709 raid_bi->bi_next = NULL;
2711 rdev_dec_pending(rdev, conf->mddev);
2713 if (!error && uptodate) {
2714 bio_endio(raid_bi, bytes, 0);
2715 if (atomic_dec_and_test(&conf->active_aligned_reads))
2716 wake_up(&conf->wait_for_stripe);
2721 PRINTK("raid5_align_endio : io error...handing IO for a retry\n");
2723 add_bio_to_retry(raid_bi, conf);
2727 static int chunk_aligned_read(request_queue_t *q, struct bio * raid_bio)
2729 mddev_t *mddev = q->queuedata;
2730 raid5_conf_t *conf = mddev_to_conf(mddev);
2731 const unsigned int raid_disks = conf->raid_disks;
2732 const unsigned int data_disks = raid_disks - conf->max_degraded;
2733 unsigned int dd_idx, pd_idx;
2734 struct bio* align_bi;
2737 if (!in_chunk_boundary(mddev, raid_bio)) {
2738 printk("chunk_aligned_read : non aligned\n");
2742 * use bio_clone to make a copy of the bio
2744 align_bi = bio_clone(raid_bio, GFP_NOIO);
2748 * set bi_end_io to a new function, and set bi_private to the
2751 align_bi->bi_end_io = raid5_align_endio;
2752 align_bi->bi_private = raid_bio;
2756 align_bi->bi_sector = raid5_compute_sector(raid_bio->bi_sector,
2764 rdev = rcu_dereference(conf->disks[dd_idx].rdev);
2765 if (rdev && test_bit(In_sync, &rdev->flags)) {
2766 atomic_inc(&rdev->nr_pending);
2768 raid_bio->bi_next = (void*)rdev;
2769 align_bi->bi_bdev = rdev->bdev;
2770 align_bi->bi_flags &= ~(1 << BIO_SEG_VALID);
2771 align_bi->bi_sector += rdev->data_offset;
2773 spin_lock_irq(&conf->device_lock);
2774 wait_event_lock_irq(conf->wait_for_stripe,
2776 conf->device_lock, /* nothing */);
2777 atomic_inc(&conf->active_aligned_reads);
2778 spin_unlock_irq(&conf->device_lock);
2780 generic_make_request(align_bi);
2790 static int make_request(request_queue_t *q, struct bio * bi)
2792 mddev_t *mddev = q->queuedata;
2793 raid5_conf_t *conf = mddev_to_conf(mddev);
2794 unsigned int dd_idx, pd_idx;
2795 sector_t new_sector;
2796 sector_t logical_sector, last_sector;
2797 struct stripe_head *sh;
2798 const int rw = bio_data_dir(bi);
2801 if (unlikely(bio_barrier(bi))) {
2802 bio_endio(bi, bi->bi_size, -EOPNOTSUPP);
2806 md_write_start(mddev, bi);
2808 disk_stat_inc(mddev->gendisk, ios[rw]);
2809 disk_stat_add(mddev->gendisk, sectors[rw], bio_sectors(bi));
2811 logical_sector = bi->bi_sector & ~((sector_t)STRIPE_SECTORS-1);
2812 last_sector = bi->bi_sector + (bi->bi_size>>9);
2814 bi->bi_phys_segments = 1; /* over-loaded to count active stripes */
2816 for (;logical_sector < last_sector; logical_sector += STRIPE_SECTORS) {
2818 int disks, data_disks;
2821 prepare_to_wait(&conf->wait_for_overlap, &w, TASK_UNINTERRUPTIBLE);
2822 if (likely(conf->expand_progress == MaxSector))
2823 disks = conf->raid_disks;
2825 /* spinlock is needed as expand_progress may be
2826 * 64bit on a 32bit platform, and so it might be
2827 * possible to see a half-updated value
2828 * Ofcourse expand_progress could change after
2829 * the lock is dropped, so once we get a reference
2830 * to the stripe that we think it is, we will have
2833 spin_lock_irq(&conf->device_lock);
2834 disks = conf->raid_disks;
2835 if (logical_sector >= conf->expand_progress)
2836 disks = conf->previous_raid_disks;
2838 if (logical_sector >= conf->expand_lo) {
2839 spin_unlock_irq(&conf->device_lock);
2844 spin_unlock_irq(&conf->device_lock);
2846 data_disks = disks - conf->max_degraded;
2848 new_sector = raid5_compute_sector(logical_sector, disks, data_disks,
2849 &dd_idx, &pd_idx, conf);
2850 PRINTK("raid5: make_request, sector %llu logical %llu\n",
2851 (unsigned long long)new_sector,
2852 (unsigned long long)logical_sector);
2854 sh = get_active_stripe(conf, new_sector, disks, pd_idx, (bi->bi_rw&RWA_MASK));
2856 if (unlikely(conf->expand_progress != MaxSector)) {
2857 /* expansion might have moved on while waiting for a
2858 * stripe, so we must do the range check again.
2859 * Expansion could still move past after this
2860 * test, but as we are holding a reference to
2861 * 'sh', we know that if that happens,
2862 * STRIPE_EXPANDING will get set and the expansion
2863 * won't proceed until we finish with the stripe.
2866 spin_lock_irq(&conf->device_lock);
2867 if (logical_sector < conf->expand_progress &&
2868 disks == conf->previous_raid_disks)
2869 /* mismatch, need to try again */
2871 spin_unlock_irq(&conf->device_lock);
2877 /* FIXME what if we get a false positive because these
2878 * are being updated.
2880 if (logical_sector >= mddev->suspend_lo &&
2881 logical_sector < mddev->suspend_hi) {
2887 if (test_bit(STRIPE_EXPANDING, &sh->state) ||
2888 !add_stripe_bio(sh, bi, dd_idx, (bi->bi_rw&RW_MASK))) {
2889 /* Stripe is busy expanding or
2890 * add failed due to overlap. Flush everything
2893 raid5_unplug_device(mddev->queue);
2898 finish_wait(&conf->wait_for_overlap, &w);
2899 handle_stripe(sh, NULL);
2902 /* cannot get stripe for read-ahead, just give-up */
2903 clear_bit(BIO_UPTODATE, &bi->bi_flags);
2904 finish_wait(&conf->wait_for_overlap, &w);
2909 spin_lock_irq(&conf->device_lock);
2910 remaining = --bi->bi_phys_segments;
2911 spin_unlock_irq(&conf->device_lock);
2912 if (remaining == 0) {
2913 int bytes = bi->bi_size;
2916 md_write_end(mddev);
2918 bi->bi_end_io(bi, bytes, 0);
2923 static sector_t reshape_request(mddev_t *mddev, sector_t sector_nr, int *skipped)
2925 /* reshaping is quite different to recovery/resync so it is
2926 * handled quite separately ... here.
2928 * On each call to sync_request, we gather one chunk worth of
2929 * destination stripes and flag them as expanding.
2930 * Then we find all the source stripes and request reads.
2931 * As the reads complete, handle_stripe will copy the data
2932 * into the destination stripe and release that stripe.
2934 raid5_conf_t *conf = (raid5_conf_t *) mddev->private;
2935 struct stripe_head *sh;
2937 sector_t first_sector, last_sector;
2942 sector_t writepos, safepos, gap;
2944 if (sector_nr == 0 &&
2945 conf->expand_progress != 0) {
2946 /* restarting in the middle, skip the initial sectors */
2947 sector_nr = conf->expand_progress;
2948 sector_div(sector_nr, conf->raid_disks-1);
2953 /* we update the metadata when there is more than 3Meg
2954 * in the block range (that is rather arbitrary, should
2955 * probably be time based) or when the data about to be
2956 * copied would over-write the source of the data at
2957 * the front of the range.
2958 * i.e. one new_stripe forward from expand_progress new_maps
2959 * to after where expand_lo old_maps to
2961 writepos = conf->expand_progress +
2962 conf->chunk_size/512*(conf->raid_disks-1);
2963 sector_div(writepos, conf->raid_disks-1);
2964 safepos = conf->expand_lo;
2965 sector_div(safepos, conf->previous_raid_disks-1);
2966 gap = conf->expand_progress - conf->expand_lo;
2968 if (writepos >= safepos ||
2969 gap > (conf->raid_disks-1)*3000*2 /*3Meg*/) {
2970 /* Cannot proceed until we've updated the superblock... */
2971 wait_event(conf->wait_for_overlap,
2972 atomic_read(&conf->reshape_stripes)==0);
2973 mddev->reshape_position = conf->expand_progress;
2974 set_bit(MD_CHANGE_DEVS, &mddev->flags);
2975 md_wakeup_thread(mddev->thread);
2976 wait_event(mddev->sb_wait, mddev->flags == 0 ||
2977 kthread_should_stop());
2978 spin_lock_irq(&conf->device_lock);
2979 conf->expand_lo = mddev->reshape_position;
2980 spin_unlock_irq(&conf->device_lock);
2981 wake_up(&conf->wait_for_overlap);
2984 for (i=0; i < conf->chunk_size/512; i+= STRIPE_SECTORS) {
2987 pd_idx = stripe_to_pdidx(sector_nr+i, conf, conf->raid_disks);
2988 sh = get_active_stripe(conf, sector_nr+i,
2989 conf->raid_disks, pd_idx, 0);
2990 set_bit(STRIPE_EXPANDING, &sh->state);
2991 atomic_inc(&conf->reshape_stripes);
2992 /* If any of this stripe is beyond the end of the old
2993 * array, then we need to zero those blocks
2995 for (j=sh->disks; j--;) {
2997 if (j == sh->pd_idx)
2999 s = compute_blocknr(sh, j);
3000 if (s < (mddev->array_size<<1)) {
3004 memset(page_address(sh->dev[j].page), 0, STRIPE_SIZE);
3005 set_bit(R5_Expanded, &sh->dev[j].flags);
3006 set_bit(R5_UPTODATE, &sh->dev[j].flags);
3009 set_bit(STRIPE_EXPAND_READY, &sh->state);
3010 set_bit(STRIPE_HANDLE, &sh->state);
3014 spin_lock_irq(&conf->device_lock);
3015 conf->expand_progress = (sector_nr + i)*(conf->raid_disks-1);
3016 spin_unlock_irq(&conf->device_lock);
3017 /* Ok, those stripe are ready. We can start scheduling
3018 * reads on the source stripes.
3019 * The source stripes are determined by mapping the first and last
3020 * block on the destination stripes.
3022 raid_disks = conf->previous_raid_disks;
3023 data_disks = raid_disks - 1;
3025 raid5_compute_sector(sector_nr*(conf->raid_disks-1),
3026 raid_disks, data_disks,
3027 &dd_idx, &pd_idx, conf);
3029 raid5_compute_sector((sector_nr+conf->chunk_size/512)
3030 *(conf->raid_disks-1) -1,
3031 raid_disks, data_disks,
3032 &dd_idx, &pd_idx, conf);
3033 if (last_sector >= (mddev->size<<1))
3034 last_sector = (mddev->size<<1)-1;
3035 while (first_sector <= last_sector) {
3036 pd_idx = stripe_to_pdidx(first_sector, conf, conf->previous_raid_disks);
3037 sh = get_active_stripe(conf, first_sector,
3038 conf->previous_raid_disks, pd_idx, 0);
3039 set_bit(STRIPE_EXPAND_SOURCE, &sh->state);
3040 set_bit(STRIPE_HANDLE, &sh->state);
3042 first_sector += STRIPE_SECTORS;
3044 return conf->chunk_size>>9;
3047 /* FIXME go_faster isn't used */
3048 static inline sector_t sync_request(mddev_t *mddev, sector_t sector_nr, int *skipped, int go_faster)
3050 raid5_conf_t *conf = (raid5_conf_t *) mddev->private;
3051 struct stripe_head *sh;
3053 int raid_disks = conf->raid_disks;
3054 sector_t max_sector = mddev->size << 1;
3056 int still_degraded = 0;
3059 if (sector_nr >= max_sector) {
3060 /* just being told to finish up .. nothing much to do */
3061 unplug_slaves(mddev);
3062 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery)) {
3067 if (mddev->curr_resync < max_sector) /* aborted */
3068 bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
3070 else /* completed sync */
3072 bitmap_close_sync(mddev->bitmap);
3077 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
3078 return reshape_request(mddev, sector_nr, skipped);
3080 /* if there is too many failed drives and we are trying
3081 * to resync, then assert that we are finished, because there is
3082 * nothing we can do.
3084 if (mddev->degraded >= conf->max_degraded &&
3085 test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
3086 sector_t rv = (mddev->size << 1) - sector_nr;
3090 if (!bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
3091 !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
3092 !conf->fullsync && sync_blocks >= STRIPE_SECTORS) {
3093 /* we can skip this block, and probably more */
3094 sync_blocks /= STRIPE_SECTORS;
3096 return sync_blocks * STRIPE_SECTORS; /* keep things rounded to whole stripes */
3099 pd_idx = stripe_to_pdidx(sector_nr, conf, raid_disks);
3100 sh = get_active_stripe(conf, sector_nr, raid_disks, pd_idx, 1);
3102 sh = get_active_stripe(conf, sector_nr, raid_disks, pd_idx, 0);
3103 /* make sure we don't swamp the stripe cache if someone else
3104 * is trying to get access
3106 schedule_timeout_uninterruptible(1);
3108 /* Need to check if array will still be degraded after recovery/resync
3109 * We don't need to check the 'failed' flag as when that gets set,
3112 for (i=0; i<mddev->raid_disks; i++)
3113 if (conf->disks[i].rdev == NULL)
3116 bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, still_degraded);
3118 spin_lock(&sh->lock);
3119 set_bit(STRIPE_SYNCING, &sh->state);
3120 clear_bit(STRIPE_INSYNC, &sh->state);
3121 spin_unlock(&sh->lock);
3123 handle_stripe(sh, NULL);
3126 return STRIPE_SECTORS;
3129 static int retry_aligned_read(raid5_conf_t *conf, struct bio *raid_bio)
3131 /* We may not be able to submit a whole bio at once as there
3132 * may not be enough stripe_heads available.
3133 * We cannot pre-allocate enough stripe_heads as we may need
3134 * more than exist in the cache (if we allow ever large chunks).
3135 * So we do one stripe head at a time and record in
3136 * ->bi_hw_segments how many have been done.
3138 * We *know* that this entire raid_bio is in one chunk, so
3139 * it will be only one 'dd_idx' and only need one call to raid5_compute_sector.
3141 struct stripe_head *sh;
3143 sector_t sector, logical_sector, last_sector;
3148 logical_sector = raid_bio->bi_sector & ~((sector_t)STRIPE_SECTORS-1);
3149 sector = raid5_compute_sector( logical_sector,
3151 conf->raid_disks - conf->max_degraded,
3155 last_sector = raid_bio->bi_sector + (raid_bio->bi_size>>9);
3157 for (; logical_sector < last_sector;
3158 logical_sector += STRIPE_SECTORS, scnt++) {
3160 if (scnt < raid_bio->bi_hw_segments)
3161 /* already done this stripe */
3164 sh = get_active_stripe(conf, sector, conf->raid_disks, pd_idx, 1);
3167 /* failed to get a stripe - must wait */
3168 raid_bio->bi_hw_segments = scnt;
3169 conf->retry_read_aligned = raid_bio;
3173 set_bit(R5_ReadError, &sh->dev[dd_idx].flags);
3174 add_stripe_bio(sh, raid_bio, dd_idx, 0);
3175 handle_stripe(sh, NULL);
3179 spin_lock_irq(&conf->device_lock);
3180 remaining = --raid_bio->bi_phys_segments;
3181 spin_unlock_irq(&conf->device_lock);
3182 if (remaining == 0) {
3183 int bytes = raid_bio->bi_size;
3185 raid_bio->bi_size = 0;
3186 raid_bio->bi_end_io(raid_bio, bytes, 0);
3188 if (atomic_dec_and_test(&conf->active_aligned_reads))
3189 wake_up(&conf->wait_for_stripe);
3196 * This is our raid5 kernel thread.
3198 * We scan the hash table for stripes which can be handled now.
3199 * During the scan, completed stripes are saved for us by the interrupt
3200 * handler, so that they will not have to wait for our next wakeup.
3202 static void raid5d (mddev_t *mddev)
3204 struct stripe_head *sh;
3205 raid5_conf_t *conf = mddev_to_conf(mddev);
3208 PRINTK("+++ raid5d active\n");
3210 md_check_recovery(mddev);
3213 spin_lock_irq(&conf->device_lock);
3215 struct list_head *first;
3218 if (conf->seq_flush != conf->seq_write) {
3219 int seq = conf->seq_flush;
3220 spin_unlock_irq(&conf->device_lock);
3221 bitmap_unplug(mddev->bitmap);
3222 spin_lock_irq(&conf->device_lock);
3223 conf->seq_write = seq;
3224 activate_bit_delay(conf);
3227 if (list_empty(&conf->handle_list) &&
3228 atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD &&
3229 !blk_queue_plugged(mddev->queue) &&
3230 !list_empty(&conf->delayed_list))
3231 raid5_activate_delayed(conf);
3233 while ((bio = remove_bio_from_retry(conf))) {
3235 spin_unlock_irq(&conf->device_lock);
3236 ok = retry_aligned_read(conf, bio);
3237 spin_lock_irq(&conf->device_lock);
3243 if (list_empty(&conf->handle_list))
3246 first = conf->handle_list.next;
3247 sh = list_entry(first, struct stripe_head, lru);
3249 list_del_init(first);
3250 atomic_inc(&sh->count);
3251 BUG_ON(atomic_read(&sh->count)!= 1);
3252 spin_unlock_irq(&conf->device_lock);
3255 handle_stripe(sh, conf->spare_page);
3258 spin_lock_irq(&conf->device_lock);
3260 PRINTK("%d stripes handled\n", handled);
3262 spin_unlock_irq(&conf->device_lock);
3264 unplug_slaves(mddev);
3266 PRINTK("--- raid5d inactive\n");
3270 raid5_show_stripe_cache_size(mddev_t *mddev, char *page)
3272 raid5_conf_t *conf = mddev_to_conf(mddev);
3274 return sprintf(page, "%d\n", conf->max_nr_stripes);
3280 raid5_store_stripe_cache_size(mddev_t *mddev, const char *page, size_t len)
3282 raid5_conf_t *conf = mddev_to_conf(mddev);
3285 if (len >= PAGE_SIZE)
3290 new = simple_strtoul(page, &end, 10);
3291 if (!*page || (*end && *end != '\n') )
3293 if (new <= 16 || new > 32768)
3295 while (new < conf->max_nr_stripes) {
3296 if (drop_one_stripe(conf))
3297 conf->max_nr_stripes--;
3301 while (new > conf->max_nr_stripes) {
3302 if (grow_one_stripe(conf))
3303 conf->max_nr_stripes++;
3309 static struct md_sysfs_entry
3310 raid5_stripecache_size = __ATTR(stripe_cache_size, S_IRUGO | S_IWUSR,
3311 raid5_show_stripe_cache_size,
3312 raid5_store_stripe_cache_size);
3315 stripe_cache_active_show(mddev_t *mddev, char *page)
3317 raid5_conf_t *conf = mddev_to_conf(mddev);
3319 return sprintf(page, "%d\n", atomic_read(&conf->active_stripes));
3324 static struct md_sysfs_entry
3325 raid5_stripecache_active = __ATTR_RO(stripe_cache_active);
3327 static struct attribute *raid5_attrs[] = {
3328 &raid5_stripecache_size.attr,
3329 &raid5_stripecache_active.attr,
3332 static struct attribute_group raid5_attrs_group = {
3334 .attrs = raid5_attrs,
3337 static int run(mddev_t *mddev)
3340 int raid_disk, memory;
3342 struct disk_info *disk;
3343 struct list_head *tmp;
3344 int working_disks = 0;
3346 if (mddev->level != 5 && mddev->level != 4 && mddev->level != 6) {
3347 printk(KERN_ERR "raid5: %s: raid level not set to 4/5/6 (%d)\n",
3348 mdname(mddev), mddev->level);
3352 if (mddev->reshape_position != MaxSector) {
3353 /* Check that we can continue the reshape.
3354 * Currently only disks can change, it must
3355 * increase, and we must be past the point where
3356 * a stripe over-writes itself
3358 sector_t here_new, here_old;
3361 if (mddev->new_level != mddev->level ||
3362 mddev->new_layout != mddev->layout ||
3363 mddev->new_chunk != mddev->chunk_size) {
3364 printk(KERN_ERR "raid5: %s: unsupported reshape required - aborting.\n",
3368 if (mddev->delta_disks <= 0) {
3369 printk(KERN_ERR "raid5: %s: unsupported reshape (reduce disks) required - aborting.\n",
3373 old_disks = mddev->raid_disks - mddev->delta_disks;
3374 /* reshape_position must be on a new-stripe boundary, and one
3375 * further up in new geometry must map after here in old geometry.
3377 here_new = mddev->reshape_position;
3378 if (sector_div(here_new, (mddev->chunk_size>>9)*(mddev->raid_disks-1))) {
3379 printk(KERN_ERR "raid5: reshape_position not on a stripe boundary\n");
3382 /* here_new is the stripe we will write to */
3383 here_old = mddev->reshape_position;
3384 sector_div(here_old, (mddev->chunk_size>>9)*(old_disks-1));
3385 /* here_old is the first stripe that we might need to read from */
3386 if (here_new >= here_old) {
3387 /* Reading from the same stripe as writing to - bad */
3388 printk(KERN_ERR "raid5: reshape_position too early for auto-recovery - aborting.\n");
3391 printk(KERN_INFO "raid5: reshape will continue\n");
3392 /* OK, we should be able to continue; */
3396 mddev->private = kzalloc(sizeof (raid5_conf_t), GFP_KERNEL);
3397 if ((conf = mddev->private) == NULL)
3399 if (mddev->reshape_position == MaxSector) {
3400 conf->previous_raid_disks = conf->raid_disks = mddev->raid_disks;
3402 conf->raid_disks = mddev->raid_disks;
3403 conf->previous_raid_disks = mddev->raid_disks - mddev->delta_disks;
3406 conf->disks = kzalloc(conf->raid_disks * sizeof(struct disk_info),
3411 conf->mddev = mddev;
3413 if ((conf->stripe_hashtbl = kzalloc(PAGE_SIZE, GFP_KERNEL)) == NULL)
3416 if (mddev->level == 6) {
3417 conf->spare_page = alloc_page(GFP_KERNEL);
3418 if (!conf->spare_page)
3421 spin_lock_init(&conf->device_lock);
3422 init_waitqueue_head(&conf->wait_for_stripe);
3423 init_waitqueue_head(&conf->wait_for_overlap);
3424 INIT_LIST_HEAD(&conf->handle_list);
3425 INIT_LIST_HEAD(&conf->delayed_list);
3426 INIT_LIST_HEAD(&conf->bitmap_list);
3427 INIT_LIST_HEAD(&conf->inactive_list);
3428 atomic_set(&conf->active_stripes, 0);
3429 atomic_set(&conf->preread_active_stripes, 0);
3430 atomic_set(&conf->active_aligned_reads, 0);
3432 PRINTK("raid5: run(%s) called.\n", mdname(mddev));
3434 ITERATE_RDEV(mddev,rdev,tmp) {
3435 raid_disk = rdev->raid_disk;
3436 if (raid_disk >= conf->raid_disks
3439 disk = conf->disks + raid_disk;
3443 if (test_bit(In_sync, &rdev->flags)) {
3444 char b[BDEVNAME_SIZE];
3445 printk(KERN_INFO "raid5: device %s operational as raid"
3446 " disk %d\n", bdevname(rdev->bdev,b),
3453 * 0 for a fully functional array, 1 or 2 for a degraded array.
3455 mddev->degraded = conf->raid_disks - working_disks;
3456 conf->mddev = mddev;
3457 conf->chunk_size = mddev->chunk_size;
3458 conf->level = mddev->level;
3459 if (conf->level == 6)
3460 conf->max_degraded = 2;
3462 conf->max_degraded = 1;
3463 conf->algorithm = mddev->layout;
3464 conf->max_nr_stripes = NR_STRIPES;
3465 conf->expand_progress = mddev->reshape_position;
3467 /* device size must be a multiple of chunk size */
3468 mddev->size &= ~(mddev->chunk_size/1024 -1);
3469 mddev->resync_max_sectors = mddev->size << 1;
3471 if (conf->level == 6 && conf->raid_disks < 4) {
3472 printk(KERN_ERR "raid6: not enough configured devices for %s (%d, minimum 4)\n",
3473 mdname(mddev), conf->raid_disks);
3476 if (!conf->chunk_size || conf->chunk_size % 4) {
3477 printk(KERN_ERR "raid5: invalid chunk size %d for %s\n",
3478 conf->chunk_size, mdname(mddev));
3481 if (conf->algorithm > ALGORITHM_RIGHT_SYMMETRIC) {
3483 "raid5: unsupported parity algorithm %d for %s\n",
3484 conf->algorithm, mdname(mddev));
3487 if (mddev->degraded > conf->max_degraded) {
3488 printk(KERN_ERR "raid5: not enough operational devices for %s"
3489 " (%d/%d failed)\n",
3490 mdname(mddev), mddev->degraded, conf->raid_disks);
3494 if (mddev->degraded > 0 &&
3495 mddev->recovery_cp != MaxSector) {
3496 if (mddev->ok_start_degraded)
3498 "raid5: starting dirty degraded array: %s"
3499 "- data corruption possible.\n",
3503 "raid5: cannot start dirty degraded array for %s\n",
3510 mddev->thread = md_register_thread(raid5d, mddev, "%s_raid5");
3511 if (!mddev->thread) {
3513 "raid5: couldn't allocate thread for %s\n",
3518 memory = conf->max_nr_stripes * (sizeof(struct stripe_head) +
3519 conf->raid_disks * ((sizeof(struct bio) + PAGE_SIZE))) / 1024;
3520 if (grow_stripes(conf, conf->max_nr_stripes)) {
3522 "raid5: couldn't allocate %dkB for buffers\n", memory);
3523 shrink_stripes(conf);
3524 md_unregister_thread(mddev->thread);
3527 printk(KERN_INFO "raid5: allocated %dkB for %s\n",
3528 memory, mdname(mddev));
3530 if (mddev->degraded == 0)
3531 printk("raid5: raid level %d set %s active with %d out of %d"
3532 " devices, algorithm %d\n", conf->level, mdname(mddev),
3533 mddev->raid_disks-mddev->degraded, mddev->raid_disks,
3536 printk(KERN_ALERT "raid5: raid level %d set %s active with %d"
3537 " out of %d devices, algorithm %d\n", conf->level,
3538 mdname(mddev), mddev->raid_disks - mddev->degraded,
3539 mddev->raid_disks, conf->algorithm);
3541 print_raid5_conf(conf);
3543 if (conf->expand_progress != MaxSector) {
3544 printk("...ok start reshape thread\n");
3545 conf->expand_lo = conf->expand_progress;
3546 atomic_set(&conf->reshape_stripes, 0);
3547 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
3548 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
3549 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
3550 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
3551 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
3555 /* read-ahead size must cover two whole stripes, which is
3556 * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
3559 int data_disks = conf->previous_raid_disks - conf->max_degraded;
3560 int stripe = data_disks *
3561 (mddev->chunk_size / PAGE_SIZE);
3562 if (mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
3563 mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
3566 /* Ok, everything is just fine now */
3567 sysfs_create_group(&mddev->kobj, &raid5_attrs_group);
3569 mddev->queue->unplug_fn = raid5_unplug_device;
3570 mddev->queue->issue_flush_fn = raid5_issue_flush;
3571 mddev->queue->backing_dev_info.congested_fn = raid5_congested;
3572 mddev->queue->backing_dev_info.congested_data = mddev;
3574 mddev->array_size = mddev->size * (conf->previous_raid_disks -
3575 conf->max_degraded);
3577 blk_queue_merge_bvec(mddev->queue, raid5_mergeable_bvec);
3582 print_raid5_conf(conf);
3583 safe_put_page(conf->spare_page);
3585 kfree(conf->stripe_hashtbl);
3588 mddev->private = NULL;
3589 printk(KERN_ALERT "raid5: failed to run raid set %s\n", mdname(mddev));
3595 static int stop(mddev_t *mddev)
3597 raid5_conf_t *conf = (raid5_conf_t *) mddev->private;
3599 md_unregister_thread(mddev->thread);
3600 mddev->thread = NULL;
3601 shrink_stripes(conf);
3602 kfree(conf->stripe_hashtbl);
3603 blk_sync_queue(mddev->queue); /* the unplug fn references 'conf'*/
3604 sysfs_remove_group(&mddev->kobj, &raid5_attrs_group);
3607 mddev->private = NULL;
3612 static void print_sh (struct seq_file *seq, struct stripe_head *sh)
3616 seq_printf(seq, "sh %llu, pd_idx %d, state %ld.\n",
3617 (unsigned long long)sh->sector, sh->pd_idx, sh->state);
3618 seq_printf(seq, "sh %llu, count %d.\n",
3619 (unsigned long long)sh->sector, atomic_read(&sh->count));
3620 seq_printf(seq, "sh %llu, ", (unsigned long long)sh->sector);
3621 for (i = 0; i < sh->disks; i++) {
3622 seq_printf(seq, "(cache%d: %p %ld) ",
3623 i, sh->dev[i].page, sh->dev[i].flags);
3625 seq_printf(seq, "\n");
3628 static void printall (struct seq_file *seq, raid5_conf_t *conf)
3630 struct stripe_head *sh;
3631 struct hlist_node *hn;
3634 spin_lock_irq(&conf->device_lock);
3635 for (i = 0; i < NR_HASH; i++) {
3636 hlist_for_each_entry(sh, hn, &conf->stripe_hashtbl[i], hash) {
3637 if (sh->raid_conf != conf)
3642 spin_unlock_irq(&conf->device_lock);
3646 static void status (struct seq_file *seq, mddev_t *mddev)
3648 raid5_conf_t *conf = (raid5_conf_t *) mddev->private;
3651 seq_printf (seq, " level %d, %dk chunk, algorithm %d", mddev->level, mddev->chunk_size >> 10, mddev->layout);
3652 seq_printf (seq, " [%d/%d] [", conf->raid_disks, conf->raid_disks - mddev->degraded);
3653 for (i = 0; i < conf->raid_disks; i++)
3654 seq_printf (seq, "%s",
3655 conf->disks[i].rdev &&
3656 test_bit(In_sync, &conf->disks[i].rdev->flags) ? "U" : "_");
3657 seq_printf (seq, "]");
3659 seq_printf (seq, "\n");
3660 printall(seq, conf);
3664 static void print_raid5_conf (raid5_conf_t *conf)
3667 struct disk_info *tmp;
3669 printk("RAID5 conf printout:\n");
3671 printk("(conf==NULL)\n");
3674 printk(" --- rd:%d wd:%d\n", conf->raid_disks,
3675 conf->raid_disks - conf->mddev->degraded);
3677 for (i = 0; i < conf->raid_disks; i++) {
3678 char b[BDEVNAME_SIZE];
3679 tmp = conf->disks + i;
3681 printk(" disk %d, o:%d, dev:%s\n",
3682 i, !test_bit(Faulty, &tmp->rdev->flags),
3683 bdevname(tmp->rdev->bdev,b));
3687 static int raid5_spare_active(mddev_t *mddev)
3690 raid5_conf_t *conf = mddev->private;
3691 struct disk_info *tmp;
3693 for (i = 0; i < conf->raid_disks; i++) {
3694 tmp = conf->disks + i;
3696 && !test_bit(Faulty, &tmp->rdev->flags)
3697 && !test_and_set_bit(In_sync, &tmp->rdev->flags)) {
3698 unsigned long flags;
3699 spin_lock_irqsave(&conf->device_lock, flags);
3701 spin_unlock_irqrestore(&conf->device_lock, flags);
3704 print_raid5_conf(conf);
3708 static int raid5_remove_disk(mddev_t *mddev, int number)
3710 raid5_conf_t *conf = mddev->private;
3713 struct disk_info *p = conf->disks + number;
3715 print_raid5_conf(conf);
3718 if (test_bit(In_sync, &rdev->flags) ||
3719 atomic_read(&rdev->nr_pending)) {
3725 if (atomic_read(&rdev->nr_pending)) {
3726 /* lost the race, try later */
3733 print_raid5_conf(conf);
3737 static int raid5_add_disk(mddev_t *mddev, mdk_rdev_t *rdev)
3739 raid5_conf_t *conf = mddev->private;
3742 struct disk_info *p;
3744 if (mddev->degraded > conf->max_degraded)
3745 /* no point adding a device */
3749 * find the disk ... but prefer rdev->saved_raid_disk
3752 if (rdev->saved_raid_disk >= 0 &&
3753 conf->disks[rdev->saved_raid_disk].rdev == NULL)
3754 disk = rdev->saved_raid_disk;
3757 for ( ; disk < conf->raid_disks; disk++)
3758 if ((p=conf->disks + disk)->rdev == NULL) {
3759 clear_bit(In_sync, &rdev->flags);
3760 rdev->raid_disk = disk;
3762 if (rdev->saved_raid_disk != disk)
3764 rcu_assign_pointer(p->rdev, rdev);
3767 print_raid5_conf(conf);
3771 static int raid5_resize(mddev_t *mddev, sector_t sectors)
3773 /* no resync is happening, and there is enough space
3774 * on all devices, so we can resize.
3775 * We need to make sure resync covers any new space.
3776 * If the array is shrinking we should possibly wait until
3777 * any io in the removed space completes, but it hardly seems
3780 raid5_conf_t *conf = mddev_to_conf(mddev);
3782 sectors &= ~((sector_t)mddev->chunk_size/512 - 1);
3783 mddev->array_size = (sectors * (mddev->raid_disks-conf->max_degraded))>>1;
3784 set_capacity(mddev->gendisk, mddev->array_size << 1);
3786 if (sectors/2 > mddev->size && mddev->recovery_cp == MaxSector) {
3787 mddev->recovery_cp = mddev->size << 1;
3788 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
3790 mddev->size = sectors /2;
3791 mddev->resync_max_sectors = sectors;
3795 #ifdef CONFIG_MD_RAID5_RESHAPE
3796 static int raid5_check_reshape(mddev_t *mddev)
3798 raid5_conf_t *conf = mddev_to_conf(mddev);
3801 if (mddev->delta_disks < 0 ||
3802 mddev->new_level != mddev->level)
3803 return -EINVAL; /* Cannot shrink array or change level yet */
3804 if (mddev->delta_disks == 0)
3805 return 0; /* nothing to do */
3807 /* Can only proceed if there are plenty of stripe_heads.
3808 * We need a minimum of one full stripe,, and for sensible progress
3809 * it is best to have about 4 times that.
3810 * If we require 4 times, then the default 256 4K stripe_heads will
3811 * allow for chunk sizes up to 256K, which is probably OK.
3812 * If the chunk size is greater, user-space should request more
3813 * stripe_heads first.
3815 if ((mddev->chunk_size / STRIPE_SIZE) * 4 > conf->max_nr_stripes ||
3816 (mddev->new_chunk / STRIPE_SIZE) * 4 > conf->max_nr_stripes) {
3817 printk(KERN_WARNING "raid5: reshape: not enough stripes. Needed %lu\n",
3818 (mddev->chunk_size / STRIPE_SIZE)*4);
3822 err = resize_stripes(conf, conf->raid_disks + mddev->delta_disks);
3826 /* looks like we might be able to manage this */
3830 static int raid5_start_reshape(mddev_t *mddev)
3832 raid5_conf_t *conf = mddev_to_conf(mddev);
3834 struct list_head *rtmp;
3836 int added_devices = 0;
3837 unsigned long flags;
3839 if (mddev->degraded ||
3840 test_bit(MD_RECOVERY_RUNNING, &mddev->recovery))
3843 ITERATE_RDEV(mddev, rdev, rtmp)
3844 if (rdev->raid_disk < 0 &&
3845 !test_bit(Faulty, &rdev->flags))
3848 if (spares < mddev->delta_disks-1)
3849 /* Not enough devices even to make a degraded array
3854 atomic_set(&conf->reshape_stripes, 0);
3855 spin_lock_irq(&conf->device_lock);
3856 conf->previous_raid_disks = conf->raid_disks;
3857 conf->raid_disks += mddev->delta_disks;
3858 conf->expand_progress = 0;
3859 conf->expand_lo = 0;
3860 spin_unlock_irq(&conf->device_lock);
3862 /* Add some new drives, as many as will fit.
3863 * We know there are enough to make the newly sized array work.
3865 ITERATE_RDEV(mddev, rdev, rtmp)
3866 if (rdev->raid_disk < 0 &&
3867 !test_bit(Faulty, &rdev->flags)) {
3868 if (raid5_add_disk(mddev, rdev)) {
3870 set_bit(In_sync, &rdev->flags);
3872 rdev->recovery_offset = 0;
3873 sprintf(nm, "rd%d", rdev->raid_disk);
3874 sysfs_create_link(&mddev->kobj, &rdev->kobj, nm);
3879 spin_lock_irqsave(&conf->device_lock, flags);
3880 mddev->degraded = (conf->raid_disks - conf->previous_raid_disks) - added_devices;
3881 spin_unlock_irqrestore(&conf->device_lock, flags);
3882 mddev->raid_disks = conf->raid_disks;
3883 mddev->reshape_position = 0;
3884 set_bit(MD_CHANGE_DEVS, &mddev->flags);
3886 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
3887 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
3888 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
3889 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
3890 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
3892 if (!mddev->sync_thread) {
3893 mddev->recovery = 0;
3894 spin_lock_irq(&conf->device_lock);
3895 mddev->raid_disks = conf->raid_disks = conf->previous_raid_disks;
3896 conf->expand_progress = MaxSector;
3897 spin_unlock_irq(&conf->device_lock);
3900 md_wakeup_thread(mddev->sync_thread);
3901 md_new_event(mddev);
3906 static void end_reshape(raid5_conf_t *conf)
3908 struct block_device *bdev;
3910 if (!test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery)) {
3911 conf->mddev->array_size = conf->mddev->size * (conf->raid_disks-1);
3912 set_capacity(conf->mddev->gendisk, conf->mddev->array_size << 1);
3913 conf->mddev->changed = 1;
3915 bdev = bdget_disk(conf->mddev->gendisk, 0);
3917 mutex_lock(&bdev->bd_inode->i_mutex);
3918 i_size_write(bdev->bd_inode, (loff_t)conf->mddev->array_size << 10);
3919 mutex_unlock(&bdev->bd_inode->i_mutex);
3922 spin_lock_irq(&conf->device_lock);
3923 conf->expand_progress = MaxSector;
3924 spin_unlock_irq(&conf->device_lock);
3925 conf->mddev->reshape_position = MaxSector;
3927 /* read-ahead size must cover two whole stripes, which is
3928 * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
3931 int data_disks = conf->previous_raid_disks - conf->max_degraded;
3932 int stripe = data_disks *
3933 (conf->mddev->chunk_size / PAGE_SIZE);
3934 if (conf->mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
3935 conf->mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
3940 static void raid5_quiesce(mddev_t *mddev, int state)
3942 raid5_conf_t *conf = mddev_to_conf(mddev);
3945 case 2: /* resume for a suspend */
3946 wake_up(&conf->wait_for_overlap);
3949 case 1: /* stop all writes */
3950 spin_lock_irq(&conf->device_lock);
3952 wait_event_lock_irq(conf->wait_for_stripe,
3953 atomic_read(&conf->active_stripes) == 0 &&
3954 atomic_read(&conf->active_aligned_reads) == 0,
3955 conf->device_lock, /* nothing */);
3956 spin_unlock_irq(&conf->device_lock);
3959 case 0: /* re-enable writes */
3960 spin_lock_irq(&conf->device_lock);
3962 wake_up(&conf->wait_for_stripe);
3963 wake_up(&conf->wait_for_overlap);
3964 spin_unlock_irq(&conf->device_lock);
3969 static struct mdk_personality raid6_personality =
3973 .owner = THIS_MODULE,
3974 .make_request = make_request,
3978 .error_handler = error,
3979 .hot_add_disk = raid5_add_disk,
3980 .hot_remove_disk= raid5_remove_disk,
3981 .spare_active = raid5_spare_active,
3982 .sync_request = sync_request,
3983 .resize = raid5_resize,
3984 .quiesce = raid5_quiesce,
3986 static struct mdk_personality raid5_personality =
3990 .owner = THIS_MODULE,
3991 .make_request = make_request,
3995 .error_handler = error,
3996 .hot_add_disk = raid5_add_disk,
3997 .hot_remove_disk= raid5_remove_disk,
3998 .spare_active = raid5_spare_active,
3999 .sync_request = sync_request,
4000 .resize = raid5_resize,
4001 #ifdef CONFIG_MD_RAID5_RESHAPE
4002 .check_reshape = raid5_check_reshape,
4003 .start_reshape = raid5_start_reshape,
4005 .quiesce = raid5_quiesce,
4008 static struct mdk_personality raid4_personality =
4012 .owner = THIS_MODULE,
4013 .make_request = make_request,
4017 .error_handler = error,
4018 .hot_add_disk = raid5_add_disk,
4019 .hot_remove_disk= raid5_remove_disk,
4020 .spare_active = raid5_spare_active,
4021 .sync_request = sync_request,
4022 .resize = raid5_resize,
4023 .quiesce = raid5_quiesce,
4026 static int __init raid5_init(void)
4030 e = raid6_select_algo();
4033 register_md_personality(&raid6_personality);
4034 register_md_personality(&raid5_personality);
4035 register_md_personality(&raid4_personality);
4039 static void raid5_exit(void)
4041 unregister_md_personality(&raid6_personality);
4042 unregister_md_personality(&raid5_personality);
4043 unregister_md_personality(&raid4_personality);
4046 module_init(raid5_init);
4047 module_exit(raid5_exit);
4048 MODULE_LICENSE("GPL");
4049 MODULE_ALIAS("md-personality-4"); /* RAID5 */
4050 MODULE_ALIAS("md-raid5");
4051 MODULE_ALIAS("md-raid4");
4052 MODULE_ALIAS("md-level-5");
4053 MODULE_ALIAS("md-level-4");
4054 MODULE_ALIAS("md-personality-8"); /* RAID6 */
4055 MODULE_ALIAS("md-raid6");
4056 MODULE_ALIAS("md-level-6");
4058 /* This used to be two separate modules, they were: */
4059 MODULE_ALIAS("raid5");
4060 MODULE_ALIAS("raid6");