2 * raid1.c : Multiple Devices driver for Linux
4 * Copyright (C) 1999, 2000, 2001 Ingo Molnar, Red Hat
6 * Copyright (C) 1996, 1997, 1998 Ingo Molnar, Miguel de Icaza, Gadi Oxman
8 * RAID-1 management functions.
10 * Better read-balancing code written by Mika Kuoppala <miku@iki.fi>, 2000
12 * Fixes to reconstruction by Jakob Østergaard" <jakob@ostenfeld.dk>
13 * Various fixes by Neil Brown <neilb@cse.unsw.edu.au>
15 * Changes by Peter T. Breuer <ptb@it.uc3m.es> 31/1/2003 to support
16 * bitmapped intelligence in resync:
18 * - bitmap marked during normal i/o
19 * - bitmap used to skip nondirty blocks during sync
21 * Additions to bitmap code, (C) 2003-2004 Paul Clements, SteelEye Technology:
22 * - persistent bitmap code
24 * This program is free software; you can redistribute it and/or modify
25 * it under the terms of the GNU General Public License as published by
26 * the Free Software Foundation; either version 2, or (at your option)
29 * You should have received a copy of the GNU General Public License
30 * (for example /usr/src/linux/COPYING); if not, write to the Free
31 * Software Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
34 #include "dm-bio-list.h"
35 #include <linux/raid/raid1.h>
36 #include <linux/raid/bitmap.h>
40 #define PRINTK(x...) printk(x)
46 * Number of guaranteed r1bios in case of extreme VM load:
48 #define NR_RAID1_BIOS 256
51 static void unplug_slaves(mddev_t *mddev);
53 static void allow_barrier(conf_t *conf);
54 static void lower_barrier(conf_t *conf);
56 static void * r1bio_pool_alloc(gfp_t gfp_flags, void *data)
58 struct pool_info *pi = data;
60 int size = offsetof(r1bio_t, bios[pi->raid_disks]);
62 /* allocate a r1bio with room for raid_disks entries in the bios array */
63 r1_bio = kzalloc(size, gfp_flags);
65 unplug_slaves(pi->mddev);
70 static void r1bio_pool_free(void *r1_bio, void *data)
75 #define RESYNC_BLOCK_SIZE (64*1024)
76 //#define RESYNC_BLOCK_SIZE PAGE_SIZE
77 #define RESYNC_SECTORS (RESYNC_BLOCK_SIZE >> 9)
78 #define RESYNC_PAGES ((RESYNC_BLOCK_SIZE + PAGE_SIZE-1) / PAGE_SIZE)
79 #define RESYNC_WINDOW (2048*1024)
81 static void * r1buf_pool_alloc(gfp_t gfp_flags, void *data)
83 struct pool_info *pi = data;
89 r1_bio = r1bio_pool_alloc(gfp_flags, pi);
91 unplug_slaves(pi->mddev);
96 * Allocate bios : 1 for reading, n-1 for writing
98 for (j = pi->raid_disks ; j-- ; ) {
99 bio = bio_alloc(gfp_flags, RESYNC_PAGES);
102 r1_bio->bios[j] = bio;
105 * Allocate RESYNC_PAGES data pages and attach them to
107 * If this is a user-requested check/repair, allocate
108 * RESYNC_PAGES for each bio.
110 if (test_bit(MD_RECOVERY_REQUESTED, &pi->mddev->recovery))
115 bio = r1_bio->bios[j];
116 for (i = 0; i < RESYNC_PAGES; i++) {
117 page = alloc_page(gfp_flags);
121 bio->bi_io_vec[i].bv_page = page;
124 /* If not user-requests, copy the page pointers to all bios */
125 if (!test_bit(MD_RECOVERY_REQUESTED, &pi->mddev->recovery)) {
126 for (i=0; i<RESYNC_PAGES ; i++)
127 for (j=1; j<pi->raid_disks; j++)
128 r1_bio->bios[j]->bi_io_vec[i].bv_page =
129 r1_bio->bios[0]->bi_io_vec[i].bv_page;
132 r1_bio->master_bio = NULL;
137 for (i=0; i < RESYNC_PAGES ; i++)
138 for (j=0 ; j < pi->raid_disks; j++)
139 safe_put_page(r1_bio->bios[j]->bi_io_vec[i].bv_page);
142 while ( ++j < pi->raid_disks )
143 bio_put(r1_bio->bios[j]);
144 r1bio_pool_free(r1_bio, data);
148 static void r1buf_pool_free(void *__r1_bio, void *data)
150 struct pool_info *pi = data;
152 r1bio_t *r1bio = __r1_bio;
154 for (i = 0; i < RESYNC_PAGES; i++)
155 for (j = pi->raid_disks; j-- ;) {
157 r1bio->bios[j]->bi_io_vec[i].bv_page !=
158 r1bio->bios[0]->bi_io_vec[i].bv_page)
159 safe_put_page(r1bio->bios[j]->bi_io_vec[i].bv_page);
161 for (i=0 ; i < pi->raid_disks; i++)
162 bio_put(r1bio->bios[i]);
164 r1bio_pool_free(r1bio, data);
167 static void put_all_bios(conf_t *conf, r1bio_t *r1_bio)
171 for (i = 0; i < conf->raid_disks; i++) {
172 struct bio **bio = r1_bio->bios + i;
173 if (*bio && *bio != IO_BLOCKED)
179 static void free_r1bio(r1bio_t *r1_bio)
181 conf_t *conf = mddev_to_conf(r1_bio->mddev);
184 * Wake up any possible resync thread that waits for the device
189 put_all_bios(conf, r1_bio);
190 mempool_free(r1_bio, conf->r1bio_pool);
193 static void put_buf(r1bio_t *r1_bio)
195 conf_t *conf = mddev_to_conf(r1_bio->mddev);
198 for (i=0; i<conf->raid_disks; i++) {
199 struct bio *bio = r1_bio->bios[i];
201 rdev_dec_pending(conf->mirrors[i].rdev, r1_bio->mddev);
204 mempool_free(r1_bio, conf->r1buf_pool);
209 static void reschedule_retry(r1bio_t *r1_bio)
212 mddev_t *mddev = r1_bio->mddev;
213 conf_t *conf = mddev_to_conf(mddev);
215 spin_lock_irqsave(&conf->device_lock, flags);
216 list_add(&r1_bio->retry_list, &conf->retry_list);
218 spin_unlock_irqrestore(&conf->device_lock, flags);
220 wake_up(&conf->wait_barrier);
221 md_wakeup_thread(mddev->thread);
225 * raid_end_bio_io() is called when we have finished servicing a mirrored
226 * operation and are ready to return a success/failure code to the buffer
229 static void raid_end_bio_io(r1bio_t *r1_bio)
231 struct bio *bio = r1_bio->master_bio;
233 /* if nobody has done the final endio yet, do it now */
234 if (!test_and_set_bit(R1BIO_Returned, &r1_bio->state)) {
235 PRINTK(KERN_DEBUG "raid1: sync end %s on sectors %llu-%llu\n",
236 (bio_data_dir(bio) == WRITE) ? "write" : "read",
237 (unsigned long long) bio->bi_sector,
238 (unsigned long long) bio->bi_sector +
239 (bio->bi_size >> 9) - 1);
241 bio_endio(bio, bio->bi_size,
242 test_bit(R1BIO_Uptodate, &r1_bio->state) ? 0 : -EIO);
248 * Update disk head position estimator based on IRQ completion info.
250 static inline void update_head_pos(int disk, r1bio_t *r1_bio)
252 conf_t *conf = mddev_to_conf(r1_bio->mddev);
254 conf->mirrors[disk].head_position =
255 r1_bio->sector + (r1_bio->sectors);
258 static int raid1_end_read_request(struct bio *bio, unsigned int bytes_done, int error)
260 int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
261 r1bio_t * r1_bio = (r1bio_t *)(bio->bi_private);
263 conf_t *conf = mddev_to_conf(r1_bio->mddev);
268 mirror = r1_bio->read_disk;
270 * this branch is our 'one mirror IO has finished' event handler:
272 update_head_pos(mirror, r1_bio);
274 if (uptodate || conf->working_disks <= 1) {
276 * Set R1BIO_Uptodate in our master bio, so that
277 * we will return a good error code for to the higher
278 * levels even if IO on some other mirrored buffer fails.
280 * The 'master' represents the composite IO operation to
281 * user-side. So if something waits for IO, then it will
282 * wait for the 'master' bio.
285 set_bit(R1BIO_Uptodate, &r1_bio->state);
287 raid_end_bio_io(r1_bio);
292 char b[BDEVNAME_SIZE];
293 if (printk_ratelimit())
294 printk(KERN_ERR "raid1: %s: rescheduling sector %llu\n",
295 bdevname(conf->mirrors[mirror].rdev->bdev,b), (unsigned long long)r1_bio->sector);
296 reschedule_retry(r1_bio);
299 rdev_dec_pending(conf->mirrors[mirror].rdev, conf->mddev);
303 static int raid1_end_write_request(struct bio *bio, unsigned int bytes_done, int error)
305 int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
306 r1bio_t * r1_bio = (r1bio_t *)(bio->bi_private);
307 int mirror, behind = test_bit(R1BIO_BehindIO, &r1_bio->state);
308 conf_t *conf = mddev_to_conf(r1_bio->mddev);
309 struct bio *to_put = NULL;
314 for (mirror = 0; mirror < conf->raid_disks; mirror++)
315 if (r1_bio->bios[mirror] == bio)
318 if (error == -EOPNOTSUPP && test_bit(R1BIO_Barrier, &r1_bio->state)) {
319 set_bit(BarriersNotsupp, &conf->mirrors[mirror].rdev->flags);
320 set_bit(R1BIO_BarrierRetry, &r1_bio->state);
321 r1_bio->mddev->barriers_work = 0;
322 /* Don't rdev_dec_pending in this branch - keep it for the retry */
325 * this branch is our 'one mirror IO has finished' event handler:
327 r1_bio->bios[mirror] = NULL;
330 md_error(r1_bio->mddev, conf->mirrors[mirror].rdev);
331 /* an I/O failed, we can't clear the bitmap */
332 set_bit(R1BIO_Degraded, &r1_bio->state);
335 * Set R1BIO_Uptodate in our master bio, so that
336 * we will return a good error code for to the higher
337 * levels even if IO on some other mirrored buffer fails.
339 * The 'master' represents the composite IO operation to
340 * user-side. So if something waits for IO, then it will
341 * wait for the 'master' bio.
343 set_bit(R1BIO_Uptodate, &r1_bio->state);
345 update_head_pos(mirror, r1_bio);
348 if (test_bit(WriteMostly, &conf->mirrors[mirror].rdev->flags))
349 atomic_dec(&r1_bio->behind_remaining);
351 /* In behind mode, we ACK the master bio once the I/O has safely
352 * reached all non-writemostly disks. Setting the Returned bit
353 * ensures that this gets done only once -- we don't ever want to
354 * return -EIO here, instead we'll wait */
356 if (atomic_read(&r1_bio->behind_remaining) >= (atomic_read(&r1_bio->remaining)-1) &&
357 test_bit(R1BIO_Uptodate, &r1_bio->state)) {
358 /* Maybe we can return now */
359 if (!test_and_set_bit(R1BIO_Returned, &r1_bio->state)) {
360 struct bio *mbio = r1_bio->master_bio;
361 PRINTK(KERN_DEBUG "raid1: behind end write sectors %llu-%llu\n",
362 (unsigned long long) mbio->bi_sector,
363 (unsigned long long) mbio->bi_sector +
364 (mbio->bi_size >> 9) - 1);
365 bio_endio(mbio, mbio->bi_size, 0);
369 rdev_dec_pending(conf->mirrors[mirror].rdev, conf->mddev);
373 * Let's see if all mirrored write operations have finished
376 if (atomic_dec_and_test(&r1_bio->remaining)) {
377 if (test_bit(R1BIO_BarrierRetry, &r1_bio->state))
378 reschedule_retry(r1_bio);
380 /* it really is the end of this request */
381 if (test_bit(R1BIO_BehindIO, &r1_bio->state)) {
382 /* free extra copy of the data pages */
383 int i = bio->bi_vcnt;
385 safe_put_page(bio->bi_io_vec[i].bv_page);
387 /* clear the bitmap if all writes complete successfully */
388 bitmap_endwrite(r1_bio->mddev->bitmap, r1_bio->sector,
390 !test_bit(R1BIO_Degraded, &r1_bio->state),
392 md_write_end(r1_bio->mddev);
393 raid_end_bio_io(r1_bio);
405 * This routine returns the disk from which the requested read should
406 * be done. There is a per-array 'next expected sequential IO' sector
407 * number - if this matches on the next IO then we use the last disk.
408 * There is also a per-disk 'last know head position' sector that is
409 * maintained from IRQ contexts, both the normal and the resync IO
410 * completion handlers update this position correctly. If there is no
411 * perfect sequential match then we pick the disk whose head is closest.
413 * If there are 2 mirrors in the same 2 devices, performance degrades
414 * because position is mirror, not device based.
416 * The rdev for the device selected will have nr_pending incremented.
418 static int read_balance(conf_t *conf, r1bio_t *r1_bio)
420 const unsigned long this_sector = r1_bio->sector;
421 int new_disk = conf->last_used, disk = new_disk;
423 const int sectors = r1_bio->sectors;
424 sector_t new_distance, current_distance;
429 * Check if we can balance. We can balance on the whole
430 * device if no resync is going on, or below the resync window.
431 * We take the first readable disk when above the resync window.
434 if (conf->mddev->recovery_cp < MaxSector &&
435 (this_sector + sectors >= conf->next_resync)) {
436 /* Choose the first operation device, for consistancy */
439 for (rdev = rcu_dereference(conf->mirrors[new_disk].rdev);
440 r1_bio->bios[new_disk] == IO_BLOCKED ||
441 !rdev || !test_bit(In_sync, &rdev->flags)
442 || test_bit(WriteMostly, &rdev->flags);
443 rdev = rcu_dereference(conf->mirrors[++new_disk].rdev)) {
445 if (rdev && test_bit(In_sync, &rdev->flags) &&
446 r1_bio->bios[new_disk] != IO_BLOCKED)
447 wonly_disk = new_disk;
449 if (new_disk == conf->raid_disks - 1) {
450 new_disk = wonly_disk;
458 /* make sure the disk is operational */
459 for (rdev = rcu_dereference(conf->mirrors[new_disk].rdev);
460 r1_bio->bios[new_disk] == IO_BLOCKED ||
461 !rdev || !test_bit(In_sync, &rdev->flags) ||
462 test_bit(WriteMostly, &rdev->flags);
463 rdev = rcu_dereference(conf->mirrors[new_disk].rdev)) {
465 if (rdev && test_bit(In_sync, &rdev->flags) &&
466 r1_bio->bios[new_disk] != IO_BLOCKED)
467 wonly_disk = new_disk;
470 new_disk = conf->raid_disks;
472 if (new_disk == disk) {
473 new_disk = wonly_disk;
482 /* now disk == new_disk == starting point for search */
485 * Don't change to another disk for sequential reads:
487 if (conf->next_seq_sect == this_sector)
489 if (this_sector == conf->mirrors[new_disk].head_position)
492 current_distance = abs(this_sector - conf->mirrors[disk].head_position);
494 /* Find the disk whose head is closest */
498 disk = conf->raid_disks;
501 rdev = rcu_dereference(conf->mirrors[disk].rdev);
503 if (!rdev || r1_bio->bios[disk] == IO_BLOCKED ||
504 !test_bit(In_sync, &rdev->flags) ||
505 test_bit(WriteMostly, &rdev->flags))
508 if (!atomic_read(&rdev->nr_pending)) {
512 new_distance = abs(this_sector - conf->mirrors[disk].head_position);
513 if (new_distance < current_distance) {
514 current_distance = new_distance;
517 } while (disk != conf->last_used);
523 rdev = rcu_dereference(conf->mirrors[new_disk].rdev);
526 atomic_inc(&rdev->nr_pending);
527 if (!test_bit(In_sync, &rdev->flags)) {
528 /* cannot risk returning a device that failed
529 * before we inc'ed nr_pending
531 rdev_dec_pending(rdev, conf->mddev);
534 conf->next_seq_sect = this_sector + sectors;
535 conf->last_used = new_disk;
542 static void unplug_slaves(mddev_t *mddev)
544 conf_t *conf = mddev_to_conf(mddev);
548 for (i=0; i<mddev->raid_disks; i++) {
549 mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[i].rdev);
550 if (rdev && !test_bit(Faulty, &rdev->flags) && atomic_read(&rdev->nr_pending)) {
551 request_queue_t *r_queue = bdev_get_queue(rdev->bdev);
553 atomic_inc(&rdev->nr_pending);
556 if (r_queue->unplug_fn)
557 r_queue->unplug_fn(r_queue);
559 rdev_dec_pending(rdev, mddev);
566 static void raid1_unplug(request_queue_t *q)
568 mddev_t *mddev = q->queuedata;
570 unplug_slaves(mddev);
571 md_wakeup_thread(mddev->thread);
574 static int raid1_issue_flush(request_queue_t *q, struct gendisk *disk,
575 sector_t *error_sector)
577 mddev_t *mddev = q->queuedata;
578 conf_t *conf = mddev_to_conf(mddev);
582 for (i=0; i<mddev->raid_disks && ret == 0; i++) {
583 mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[i].rdev);
584 if (rdev && !test_bit(Faulty, &rdev->flags)) {
585 struct block_device *bdev = rdev->bdev;
586 request_queue_t *r_queue = bdev_get_queue(bdev);
588 if (!r_queue->issue_flush_fn)
591 atomic_inc(&rdev->nr_pending);
593 ret = r_queue->issue_flush_fn(r_queue, bdev->bd_disk,
595 rdev_dec_pending(rdev, mddev);
605 * Sometimes we need to suspend IO while we do something else,
606 * either some resync/recovery, or reconfigure the array.
607 * To do this we raise a 'barrier'.
608 * The 'barrier' is a counter that can be raised multiple times
609 * to count how many activities are happening which preclude
611 * We can only raise the barrier if there is no pending IO.
612 * i.e. if nr_pending == 0.
613 * We choose only to raise the barrier if no-one is waiting for the
614 * barrier to go down. This means that as soon as an IO request
615 * is ready, no other operations which require a barrier will start
616 * until the IO request has had a chance.
618 * So: regular IO calls 'wait_barrier'. When that returns there
619 * is no backgroup IO happening, It must arrange to call
620 * allow_barrier when it has finished its IO.
621 * backgroup IO calls must call raise_barrier. Once that returns
622 * there is no normal IO happeing. It must arrange to call
623 * lower_barrier when the particular background IO completes.
625 #define RESYNC_DEPTH 32
627 static void raise_barrier(conf_t *conf)
629 spin_lock_irq(&conf->resync_lock);
631 /* Wait until no block IO is waiting */
632 wait_event_lock_irq(conf->wait_barrier, !conf->nr_waiting,
634 raid1_unplug(conf->mddev->queue));
636 /* block any new IO from starting */
639 /* No wait for all pending IO to complete */
640 wait_event_lock_irq(conf->wait_barrier,
641 !conf->nr_pending && conf->barrier < RESYNC_DEPTH,
643 raid1_unplug(conf->mddev->queue));
645 spin_unlock_irq(&conf->resync_lock);
648 static void lower_barrier(conf_t *conf)
651 spin_lock_irqsave(&conf->resync_lock, flags);
653 spin_unlock_irqrestore(&conf->resync_lock, flags);
654 wake_up(&conf->wait_barrier);
657 static void wait_barrier(conf_t *conf)
659 spin_lock_irq(&conf->resync_lock);
662 wait_event_lock_irq(conf->wait_barrier, !conf->barrier,
664 raid1_unplug(conf->mddev->queue));
668 spin_unlock_irq(&conf->resync_lock);
671 static void allow_barrier(conf_t *conf)
674 spin_lock_irqsave(&conf->resync_lock, flags);
676 spin_unlock_irqrestore(&conf->resync_lock, flags);
677 wake_up(&conf->wait_barrier);
680 static void freeze_array(conf_t *conf)
682 /* stop syncio and normal IO and wait for everything to
684 * We increment barrier and nr_waiting, and then
685 * wait until barrier+nr_pending match nr_queued+2
687 spin_lock_irq(&conf->resync_lock);
690 wait_event_lock_irq(conf->wait_barrier,
691 conf->barrier+conf->nr_pending == conf->nr_queued+2,
693 raid1_unplug(conf->mddev->queue));
694 spin_unlock_irq(&conf->resync_lock);
696 static void unfreeze_array(conf_t *conf)
698 /* reverse the effect of the freeze */
699 spin_lock_irq(&conf->resync_lock);
702 wake_up(&conf->wait_barrier);
703 spin_unlock_irq(&conf->resync_lock);
707 /* duplicate the data pages for behind I/O */
708 static struct page **alloc_behind_pages(struct bio *bio)
711 struct bio_vec *bvec;
712 struct page **pages = kzalloc(bio->bi_vcnt * sizeof(struct page *),
714 if (unlikely(!pages))
717 bio_for_each_segment(bvec, bio, i) {
718 pages[i] = alloc_page(GFP_NOIO);
719 if (unlikely(!pages[i]))
721 memcpy(kmap(pages[i]) + bvec->bv_offset,
722 kmap(bvec->bv_page) + bvec->bv_offset, bvec->bv_len);
724 kunmap(bvec->bv_page);
731 for (i = 0; i < bio->bi_vcnt && pages[i]; i++)
734 PRINTK("%dB behind alloc failed, doing sync I/O\n", bio->bi_size);
738 static int make_request(request_queue_t *q, struct bio * bio)
740 mddev_t *mddev = q->queuedata;
741 conf_t *conf = mddev_to_conf(mddev);
742 mirror_info_t *mirror;
744 struct bio *read_bio;
745 int i, targets = 0, disks;
747 struct bitmap *bitmap = mddev->bitmap;
750 struct page **behind_pages = NULL;
751 const int rw = bio_data_dir(bio);
755 * Register the new request and wait if the reconstruction
756 * thread has put up a bar for new requests.
757 * Continue immediately if no resync is active currently.
758 * We test barriers_work *after* md_write_start as md_write_start
759 * may cause the first superblock write, and that will check out
763 md_write_start(mddev, bio); /* wait on superblock update early */
765 if (unlikely(!mddev->barriers_work && bio_barrier(bio))) {
768 bio_endio(bio, bio->bi_size, -EOPNOTSUPP);
774 disk_stat_inc(mddev->gendisk, ios[rw]);
775 disk_stat_add(mddev->gendisk, sectors[rw], bio_sectors(bio));
778 * make_request() can abort the operation when READA is being
779 * used and no empty request is available.
782 r1_bio = mempool_alloc(conf->r1bio_pool, GFP_NOIO);
784 r1_bio->master_bio = bio;
785 r1_bio->sectors = bio->bi_size >> 9;
787 r1_bio->mddev = mddev;
788 r1_bio->sector = bio->bi_sector;
792 * read balancing logic:
794 int rdisk = read_balance(conf, r1_bio);
797 /* couldn't find anywhere to read from */
798 raid_end_bio_io(r1_bio);
801 mirror = conf->mirrors + rdisk;
803 r1_bio->read_disk = rdisk;
805 read_bio = bio_clone(bio, GFP_NOIO);
807 r1_bio->bios[rdisk] = read_bio;
809 read_bio->bi_sector = r1_bio->sector + mirror->rdev->data_offset;
810 read_bio->bi_bdev = mirror->rdev->bdev;
811 read_bio->bi_end_io = raid1_end_read_request;
812 read_bio->bi_rw = READ;
813 read_bio->bi_private = r1_bio;
815 generic_make_request(read_bio);
822 /* first select target devices under spinlock and
823 * inc refcount on their rdev. Record them by setting
826 disks = conf->raid_disks;
828 { static int first=1;
829 if (first) printk("First Write sector %llu disks %d\n",
830 (unsigned long long)r1_bio->sector, disks);
835 for (i = 0; i < disks; i++) {
836 if ((rdev=rcu_dereference(conf->mirrors[i].rdev)) != NULL &&
837 !test_bit(Faulty, &rdev->flags)) {
838 atomic_inc(&rdev->nr_pending);
839 if (test_bit(Faulty, &rdev->flags)) {
840 rdev_dec_pending(rdev, mddev);
841 r1_bio->bios[i] = NULL;
843 r1_bio->bios[i] = bio;
846 r1_bio->bios[i] = NULL;
850 BUG_ON(targets == 0); /* we never fail the last device */
852 if (targets < conf->raid_disks) {
853 /* array is degraded, we will not clear the bitmap
854 * on I/O completion (see raid1_end_write_request) */
855 set_bit(R1BIO_Degraded, &r1_bio->state);
858 /* do behind I/O ? */
860 atomic_read(&bitmap->behind_writes) < bitmap->max_write_behind &&
861 (behind_pages = alloc_behind_pages(bio)) != NULL)
862 set_bit(R1BIO_BehindIO, &r1_bio->state);
864 atomic_set(&r1_bio->remaining, 0);
865 atomic_set(&r1_bio->behind_remaining, 0);
867 do_barriers = bio_barrier(bio);
869 set_bit(R1BIO_Barrier, &r1_bio->state);
872 for (i = 0; i < disks; i++) {
874 if (!r1_bio->bios[i])
877 mbio = bio_clone(bio, GFP_NOIO);
878 r1_bio->bios[i] = mbio;
880 mbio->bi_sector = r1_bio->sector + conf->mirrors[i].rdev->data_offset;
881 mbio->bi_bdev = conf->mirrors[i].rdev->bdev;
882 mbio->bi_end_io = raid1_end_write_request;
883 mbio->bi_rw = WRITE | do_barriers;
884 mbio->bi_private = r1_bio;
887 struct bio_vec *bvec;
890 /* Yes, I really want the '__' version so that
891 * we clear any unused pointer in the io_vec, rather
892 * than leave them unchanged. This is important
893 * because when we come to free the pages, we won't
894 * know the originial bi_idx, so we just free
897 __bio_for_each_segment(bvec, mbio, j, 0)
898 bvec->bv_page = behind_pages[j];
899 if (test_bit(WriteMostly, &conf->mirrors[i].rdev->flags))
900 atomic_inc(&r1_bio->behind_remaining);
903 atomic_inc(&r1_bio->remaining);
905 bio_list_add(&bl, mbio);
907 kfree(behind_pages); /* the behind pages are attached to the bios now */
909 bitmap_startwrite(bitmap, bio->bi_sector, r1_bio->sectors,
910 test_bit(R1BIO_BehindIO, &r1_bio->state));
911 spin_lock_irqsave(&conf->device_lock, flags);
912 bio_list_merge(&conf->pending_bio_list, &bl);
915 blk_plug_device(mddev->queue);
916 spin_unlock_irqrestore(&conf->device_lock, flags);
919 while ((bio = bio_list_pop(&bl)) != NULL)
920 generic_make_request(bio);
926 static void status(struct seq_file *seq, mddev_t *mddev)
928 conf_t *conf = mddev_to_conf(mddev);
931 seq_printf(seq, " [%d/%d] [", conf->raid_disks,
932 conf->working_disks);
934 for (i = 0; i < conf->raid_disks; i++) {
935 mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[i].rdev);
936 seq_printf(seq, "%s",
937 rdev && test_bit(In_sync, &rdev->flags) ? "U" : "_");
940 seq_printf(seq, "]");
944 static void error(mddev_t *mddev, mdk_rdev_t *rdev)
946 char b[BDEVNAME_SIZE];
947 conf_t *conf = mddev_to_conf(mddev);
950 * If it is not operational, then we have already marked it as dead
951 * else if it is the last working disks, ignore the error, let the
952 * next level up know.
953 * else mark the drive as failed
955 if (test_bit(In_sync, &rdev->flags)
956 && conf->working_disks == 1)
958 * Don't fail the drive, act as though we were just a
959 * normal single drive
962 if (test_bit(In_sync, &rdev->flags)) {
964 conf->working_disks--;
966 * if recovery is running, make sure it aborts.
968 set_bit(MD_RECOVERY_ERR, &mddev->recovery);
970 clear_bit(In_sync, &rdev->flags);
971 set_bit(Faulty, &rdev->flags);
972 set_bit(MD_CHANGE_DEVS, &mddev->flags);
973 printk(KERN_ALERT "raid1: Disk failure on %s, disabling device. \n"
974 " Operation continuing on %d devices\n",
975 bdevname(rdev->bdev,b), conf->working_disks);
978 static void print_conf(conf_t *conf)
982 printk("RAID1 conf printout:\n");
987 printk(" --- wd:%d rd:%d\n", conf->working_disks,
991 for (i = 0; i < conf->raid_disks; i++) {
992 char b[BDEVNAME_SIZE];
993 mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[i].rdev);
995 printk(" disk %d, wo:%d, o:%d, dev:%s\n",
996 i, !test_bit(In_sync, &rdev->flags),
997 !test_bit(Faulty, &rdev->flags),
998 bdevname(rdev->bdev,b));
1003 static void close_sync(conf_t *conf)
1006 allow_barrier(conf);
1008 mempool_destroy(conf->r1buf_pool);
1009 conf->r1buf_pool = NULL;
1012 static int raid1_spare_active(mddev_t *mddev)
1015 conf_t *conf = mddev->private;
1018 * Find all failed disks within the RAID1 configuration
1019 * and mark them readable.
1020 * Called under mddev lock, so rcu protection not needed.
1022 for (i = 0; i < conf->raid_disks; i++) {
1023 mdk_rdev_t *rdev = conf->mirrors[i].rdev;
1025 && !test_bit(Faulty, &rdev->flags)
1026 && !test_bit(In_sync, &rdev->flags)) {
1027 conf->working_disks++;
1029 set_bit(In_sync, &rdev->flags);
1038 static int raid1_add_disk(mddev_t *mddev, mdk_rdev_t *rdev)
1040 conf_t *conf = mddev->private;
1045 for (mirror=0; mirror < mddev->raid_disks; mirror++)
1046 if ( !(p=conf->mirrors+mirror)->rdev) {
1048 blk_queue_stack_limits(mddev->queue,
1049 rdev->bdev->bd_disk->queue);
1050 /* as we don't honour merge_bvec_fn, we must never risk
1051 * violating it, so limit ->max_sector to one PAGE, as
1052 * a one page request is never in violation.
1054 if (rdev->bdev->bd_disk->queue->merge_bvec_fn &&
1055 mddev->queue->max_sectors > (PAGE_SIZE>>9))
1056 blk_queue_max_sectors(mddev->queue, PAGE_SIZE>>9);
1058 p->head_position = 0;
1059 rdev->raid_disk = mirror;
1061 /* As all devices are equivalent, we don't need a full recovery
1062 * if this was recently any drive of the array
1064 if (rdev->saved_raid_disk < 0)
1066 rcu_assign_pointer(p->rdev, rdev);
1074 static int raid1_remove_disk(mddev_t *mddev, int number)
1076 conf_t *conf = mddev->private;
1079 mirror_info_t *p = conf->mirrors+ number;
1084 if (test_bit(In_sync, &rdev->flags) ||
1085 atomic_read(&rdev->nr_pending)) {
1091 if (atomic_read(&rdev->nr_pending)) {
1092 /* lost the race, try later */
1104 static int end_sync_read(struct bio *bio, unsigned int bytes_done, int error)
1106 r1bio_t * r1_bio = (r1bio_t *)(bio->bi_private);
1112 for (i=r1_bio->mddev->raid_disks; i--; )
1113 if (r1_bio->bios[i] == bio)
1116 update_head_pos(i, r1_bio);
1118 * we have read a block, now it needs to be re-written,
1119 * or re-read if the read failed.
1120 * We don't do much here, just schedule handling by raid1d
1122 if (test_bit(BIO_UPTODATE, &bio->bi_flags))
1123 set_bit(R1BIO_Uptodate, &r1_bio->state);
1125 if (atomic_dec_and_test(&r1_bio->remaining))
1126 reschedule_retry(r1_bio);
1130 static int end_sync_write(struct bio *bio, unsigned int bytes_done, int error)
1132 int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
1133 r1bio_t * r1_bio = (r1bio_t *)(bio->bi_private);
1134 mddev_t *mddev = r1_bio->mddev;
1135 conf_t *conf = mddev_to_conf(mddev);
1142 for (i = 0; i < conf->raid_disks; i++)
1143 if (r1_bio->bios[i] == bio) {
1148 int sync_blocks = 0;
1149 sector_t s = r1_bio->sector;
1150 long sectors_to_go = r1_bio->sectors;
1151 /* make sure these bits doesn't get cleared. */
1153 bitmap_end_sync(mddev->bitmap, s,
1156 sectors_to_go -= sync_blocks;
1157 } while (sectors_to_go > 0);
1158 md_error(mddev, conf->mirrors[mirror].rdev);
1161 update_head_pos(mirror, r1_bio);
1163 if (atomic_dec_and_test(&r1_bio->remaining)) {
1164 md_done_sync(mddev, r1_bio->sectors, uptodate);
1170 static void sync_request_write(mddev_t *mddev, r1bio_t *r1_bio)
1172 conf_t *conf = mddev_to_conf(mddev);
1174 int disks = conf->raid_disks;
1175 struct bio *bio, *wbio;
1177 bio = r1_bio->bios[r1_bio->read_disk];
1180 if (test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
1181 /* We have read all readable devices. If we haven't
1182 * got the block, then there is no hope left.
1183 * If we have, then we want to do a comparison
1184 * and skip the write if everything is the same.
1185 * If any blocks failed to read, then we need to
1186 * attempt an over-write
1189 if (!test_bit(R1BIO_Uptodate, &r1_bio->state)) {
1190 for (i=0; i<mddev->raid_disks; i++)
1191 if (r1_bio->bios[i]->bi_end_io == end_sync_read)
1192 md_error(mddev, conf->mirrors[i].rdev);
1194 md_done_sync(mddev, r1_bio->sectors, 1);
1198 for (primary=0; primary<mddev->raid_disks; primary++)
1199 if (r1_bio->bios[primary]->bi_end_io == end_sync_read &&
1200 test_bit(BIO_UPTODATE, &r1_bio->bios[primary]->bi_flags)) {
1201 r1_bio->bios[primary]->bi_end_io = NULL;
1202 rdev_dec_pending(conf->mirrors[primary].rdev, mddev);
1205 r1_bio->read_disk = primary;
1206 for (i=0; i<mddev->raid_disks; i++)
1207 if (r1_bio->bios[i]->bi_end_io == end_sync_read &&
1208 test_bit(BIO_UPTODATE, &r1_bio->bios[i]->bi_flags)) {
1210 int vcnt = r1_bio->sectors >> (PAGE_SHIFT- 9);
1211 struct bio *pbio = r1_bio->bios[primary];
1212 struct bio *sbio = r1_bio->bios[i];
1213 for (j = vcnt; j-- ; )
1214 if (memcmp(page_address(pbio->bi_io_vec[j].bv_page),
1215 page_address(sbio->bi_io_vec[j].bv_page),
1219 mddev->resync_mismatches += r1_bio->sectors;
1220 if (j < 0 || test_bit(MD_RECOVERY_CHECK, &mddev->recovery)) {
1221 sbio->bi_end_io = NULL;
1222 rdev_dec_pending(conf->mirrors[i].rdev, mddev);
1224 /* fixup the bio for reuse */
1225 sbio->bi_vcnt = vcnt;
1226 sbio->bi_size = r1_bio->sectors << 9;
1228 sbio->bi_phys_segments = 0;
1229 sbio->bi_hw_segments = 0;
1230 sbio->bi_hw_front_size = 0;
1231 sbio->bi_hw_back_size = 0;
1232 sbio->bi_flags &= ~(BIO_POOL_MASK - 1);
1233 sbio->bi_flags |= 1 << BIO_UPTODATE;
1234 sbio->bi_next = NULL;
1235 sbio->bi_sector = r1_bio->sector +
1236 conf->mirrors[i].rdev->data_offset;
1237 sbio->bi_bdev = conf->mirrors[i].rdev->bdev;
1241 if (!test_bit(R1BIO_Uptodate, &r1_bio->state)) {
1242 /* ouch - failed to read all of that.
1243 * Try some synchronous reads of other devices to get
1244 * good data, much like with normal read errors. Only
1245 * read into the pages we already have so we don't
1246 * need to re-issue the read request.
1247 * We don't need to freeze the array, because being in an
1248 * active sync request, there is no normal IO, and
1249 * no overlapping syncs.
1251 sector_t sect = r1_bio->sector;
1252 int sectors = r1_bio->sectors;
1257 int d = r1_bio->read_disk;
1261 if (s > (PAGE_SIZE>>9))
1264 if (r1_bio->bios[d]->bi_end_io == end_sync_read) {
1265 /* No rcu protection needed here devices
1266 * can only be removed when no resync is
1267 * active, and resync is currently active
1269 rdev = conf->mirrors[d].rdev;
1270 if (sync_page_io(rdev->bdev,
1271 sect + rdev->data_offset,
1273 bio->bi_io_vec[idx].bv_page,
1280 if (d == conf->raid_disks)
1282 } while (!success && d != r1_bio->read_disk);
1286 /* write it back and re-read */
1287 set_bit(R1BIO_Uptodate, &r1_bio->state);
1288 while (d != r1_bio->read_disk) {
1290 d = conf->raid_disks;
1292 if (r1_bio->bios[d]->bi_end_io != end_sync_read)
1294 rdev = conf->mirrors[d].rdev;
1295 atomic_add(s, &rdev->corrected_errors);
1296 if (sync_page_io(rdev->bdev,
1297 sect + rdev->data_offset,
1299 bio->bi_io_vec[idx].bv_page,
1301 md_error(mddev, rdev);
1304 while (d != r1_bio->read_disk) {
1306 d = conf->raid_disks;
1308 if (r1_bio->bios[d]->bi_end_io != end_sync_read)
1310 rdev = conf->mirrors[d].rdev;
1311 if (sync_page_io(rdev->bdev,
1312 sect + rdev->data_offset,
1314 bio->bi_io_vec[idx].bv_page,
1316 md_error(mddev, rdev);
1319 char b[BDEVNAME_SIZE];
1320 /* Cannot read from anywhere, array is toast */
1321 md_error(mddev, conf->mirrors[r1_bio->read_disk].rdev);
1322 printk(KERN_ALERT "raid1: %s: unrecoverable I/O read error"
1323 " for block %llu\n",
1324 bdevname(bio->bi_bdev,b),
1325 (unsigned long long)r1_bio->sector);
1326 md_done_sync(mddev, r1_bio->sectors, 0);
1339 atomic_set(&r1_bio->remaining, 1);
1340 for (i = 0; i < disks ; i++) {
1341 wbio = r1_bio->bios[i];
1342 if (wbio->bi_end_io == NULL ||
1343 (wbio->bi_end_io == end_sync_read &&
1344 (i == r1_bio->read_disk ||
1345 !test_bit(MD_RECOVERY_SYNC, &mddev->recovery))))
1348 wbio->bi_rw = WRITE;
1349 wbio->bi_end_io = end_sync_write;
1350 atomic_inc(&r1_bio->remaining);
1351 md_sync_acct(conf->mirrors[i].rdev->bdev, wbio->bi_size >> 9);
1353 generic_make_request(wbio);
1356 if (atomic_dec_and_test(&r1_bio->remaining)) {
1357 /* if we're here, all write(s) have completed, so clean up */
1358 md_done_sync(mddev, r1_bio->sectors, 1);
1364 * This is a kernel thread which:
1366 * 1. Retries failed read operations on working mirrors.
1367 * 2. Updates the raid superblock when problems encounter.
1368 * 3. Performs writes following reads for array syncronising.
1371 static void fix_read_error(conf_t *conf, int read_disk,
1372 sector_t sect, int sectors)
1374 mddev_t *mddev = conf->mddev;
1382 if (s > (PAGE_SIZE>>9))
1386 /* Note: no rcu protection needed here
1387 * as this is synchronous in the raid1d thread
1388 * which is the thread that might remove
1389 * a device. If raid1d ever becomes multi-threaded....
1391 rdev = conf->mirrors[d].rdev;
1393 test_bit(In_sync, &rdev->flags) &&
1394 sync_page_io(rdev->bdev,
1395 sect + rdev->data_offset,
1397 conf->tmppage, READ))
1401 if (d == conf->raid_disks)
1404 } while (!success && d != read_disk);
1407 /* Cannot read from anywhere -- bye bye array */
1408 md_error(mddev, conf->mirrors[read_disk].rdev);
1411 /* write it back and re-read */
1413 while (d != read_disk) {
1415 d = conf->raid_disks;
1417 rdev = conf->mirrors[d].rdev;
1419 test_bit(In_sync, &rdev->flags)) {
1420 if (sync_page_io(rdev->bdev,
1421 sect + rdev->data_offset,
1422 s<<9, conf->tmppage, WRITE)
1424 /* Well, this device is dead */
1425 md_error(mddev, rdev);
1429 while (d != read_disk) {
1430 char b[BDEVNAME_SIZE];
1432 d = conf->raid_disks;
1434 rdev = conf->mirrors[d].rdev;
1436 test_bit(In_sync, &rdev->flags)) {
1437 if (sync_page_io(rdev->bdev,
1438 sect + rdev->data_offset,
1439 s<<9, conf->tmppage, READ)
1441 /* Well, this device is dead */
1442 md_error(mddev, rdev);
1444 atomic_add(s, &rdev->corrected_errors);
1446 "raid1:%s: read error corrected "
1447 "(%d sectors at %llu on %s)\n",
1449 (unsigned long long)sect +
1451 bdevname(rdev->bdev, b));
1460 static void raid1d(mddev_t *mddev)
1464 unsigned long flags;
1465 conf_t *conf = mddev_to_conf(mddev);
1466 struct list_head *head = &conf->retry_list;
1470 md_check_recovery(mddev);
1473 char b[BDEVNAME_SIZE];
1474 spin_lock_irqsave(&conf->device_lock, flags);
1476 if (conf->pending_bio_list.head) {
1477 bio = bio_list_get(&conf->pending_bio_list);
1478 blk_remove_plug(mddev->queue);
1479 spin_unlock_irqrestore(&conf->device_lock, flags);
1480 /* flush any pending bitmap writes to disk before proceeding w/ I/O */
1481 if (bitmap_unplug(mddev->bitmap) != 0)
1482 printk("%s: bitmap file write failed!\n", mdname(mddev));
1484 while (bio) { /* submit pending writes */
1485 struct bio *next = bio->bi_next;
1486 bio->bi_next = NULL;
1487 generic_make_request(bio);
1495 if (list_empty(head))
1497 r1_bio = list_entry(head->prev, r1bio_t, retry_list);
1498 list_del(head->prev);
1500 spin_unlock_irqrestore(&conf->device_lock, flags);
1502 mddev = r1_bio->mddev;
1503 conf = mddev_to_conf(mddev);
1504 if (test_bit(R1BIO_IsSync, &r1_bio->state)) {
1505 sync_request_write(mddev, r1_bio);
1507 } else if (test_bit(R1BIO_BarrierRetry, &r1_bio->state)) {
1508 /* some requests in the r1bio were BIO_RW_BARRIER
1509 * requests which failed with -EOPNOTSUPP. Hohumm..
1510 * Better resubmit without the barrier.
1511 * We know which devices to resubmit for, because
1512 * all others have had their bios[] entry cleared.
1513 * We already have a nr_pending reference on these rdevs.
1516 clear_bit(R1BIO_BarrierRetry, &r1_bio->state);
1517 clear_bit(R1BIO_Barrier, &r1_bio->state);
1518 for (i=0; i < conf->raid_disks; i++)
1519 if (r1_bio->bios[i])
1520 atomic_inc(&r1_bio->remaining);
1521 for (i=0; i < conf->raid_disks; i++)
1522 if (r1_bio->bios[i]) {
1523 struct bio_vec *bvec;
1526 bio = bio_clone(r1_bio->master_bio, GFP_NOIO);
1527 /* copy pages from the failed bio, as
1528 * this might be a write-behind device */
1529 __bio_for_each_segment(bvec, bio, j, 0)
1530 bvec->bv_page = bio_iovec_idx(r1_bio->bios[i], j)->bv_page;
1531 bio_put(r1_bio->bios[i]);
1532 bio->bi_sector = r1_bio->sector +
1533 conf->mirrors[i].rdev->data_offset;
1534 bio->bi_bdev = conf->mirrors[i].rdev->bdev;
1535 bio->bi_end_io = raid1_end_write_request;
1537 bio->bi_private = r1_bio;
1538 r1_bio->bios[i] = bio;
1539 generic_make_request(bio);
1544 /* we got a read error. Maybe the drive is bad. Maybe just
1545 * the block and we can fix it.
1546 * We freeze all other IO, and try reading the block from
1547 * other devices. When we find one, we re-write
1548 * and check it that fixes the read error.
1549 * This is all done synchronously while the array is
1552 if (mddev->ro == 0) {
1554 fix_read_error(conf, r1_bio->read_disk,
1557 unfreeze_array(conf);
1560 bio = r1_bio->bios[r1_bio->read_disk];
1561 if ((disk=read_balance(conf, r1_bio)) == -1) {
1562 printk(KERN_ALERT "raid1: %s: unrecoverable I/O"
1563 " read error for block %llu\n",
1564 bdevname(bio->bi_bdev,b),
1565 (unsigned long long)r1_bio->sector);
1566 raid_end_bio_io(r1_bio);
1568 r1_bio->bios[r1_bio->read_disk] =
1569 mddev->ro ? IO_BLOCKED : NULL;
1570 r1_bio->read_disk = disk;
1572 bio = bio_clone(r1_bio->master_bio, GFP_NOIO);
1573 r1_bio->bios[r1_bio->read_disk] = bio;
1574 rdev = conf->mirrors[disk].rdev;
1575 if (printk_ratelimit())
1576 printk(KERN_ERR "raid1: %s: redirecting sector %llu to"
1577 " another mirror\n",
1578 bdevname(rdev->bdev,b),
1579 (unsigned long long)r1_bio->sector);
1580 bio->bi_sector = r1_bio->sector + rdev->data_offset;
1581 bio->bi_bdev = rdev->bdev;
1582 bio->bi_end_io = raid1_end_read_request;
1584 bio->bi_private = r1_bio;
1586 generic_make_request(bio);
1590 spin_unlock_irqrestore(&conf->device_lock, flags);
1592 unplug_slaves(mddev);
1596 static int init_resync(conf_t *conf)
1600 buffs = RESYNC_WINDOW / RESYNC_BLOCK_SIZE;
1601 BUG_ON(conf->r1buf_pool);
1602 conf->r1buf_pool = mempool_create(buffs, r1buf_pool_alloc, r1buf_pool_free,
1604 if (!conf->r1buf_pool)
1606 conf->next_resync = 0;
1611 * perform a "sync" on one "block"
1613 * We need to make sure that no normal I/O request - particularly write
1614 * requests - conflict with active sync requests.
1616 * This is achieved by tracking pending requests and a 'barrier' concept
1617 * that can be installed to exclude normal IO requests.
1620 static sector_t sync_request(mddev_t *mddev, sector_t sector_nr, int *skipped, int go_faster)
1622 conf_t *conf = mddev_to_conf(mddev);
1625 sector_t max_sector, nr_sectors;
1629 int write_targets = 0, read_targets = 0;
1631 int still_degraded = 0;
1633 if (!conf->r1buf_pool)
1636 printk("sync start - bitmap %p\n", mddev->bitmap);
1638 if (init_resync(conf))
1642 max_sector = mddev->size << 1;
1643 if (sector_nr >= max_sector) {
1644 /* If we aborted, we need to abort the
1645 * sync on the 'current' bitmap chunk (there will
1646 * only be one in raid1 resync.
1647 * We can find the current addess in mddev->curr_resync
1649 if (mddev->curr_resync < max_sector) /* aborted */
1650 bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
1652 else /* completed sync */
1655 bitmap_close_sync(mddev->bitmap);
1660 if (mddev->bitmap == NULL &&
1661 mddev->recovery_cp == MaxSector &&
1662 !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
1663 conf->fullsync == 0) {
1665 return max_sector - sector_nr;
1667 /* before building a request, check if we can skip these blocks..
1668 * This call the bitmap_start_sync doesn't actually record anything
1670 if (!bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
1671 !conf->fullsync && !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
1672 /* We can skip this block, and probably several more */
1677 * If there is non-resync activity waiting for a turn,
1678 * and resync is going fast enough,
1679 * then let it though before starting on this new sync request.
1681 if (!go_faster && conf->nr_waiting)
1682 msleep_interruptible(1000);
1684 raise_barrier(conf);
1686 conf->next_resync = sector_nr;
1688 r1_bio = mempool_alloc(conf->r1buf_pool, GFP_NOIO);
1691 * If we get a correctably read error during resync or recovery,
1692 * we might want to read from a different device. So we
1693 * flag all drives that could conceivably be read from for READ,
1694 * and any others (which will be non-In_sync devices) for WRITE.
1695 * If a read fails, we try reading from something else for which READ
1699 r1_bio->mddev = mddev;
1700 r1_bio->sector = sector_nr;
1702 set_bit(R1BIO_IsSync, &r1_bio->state);
1704 for (i=0; i < conf->raid_disks; i++) {
1706 bio = r1_bio->bios[i];
1708 /* take from bio_init */
1709 bio->bi_next = NULL;
1710 bio->bi_flags |= 1 << BIO_UPTODATE;
1714 bio->bi_phys_segments = 0;
1715 bio->bi_hw_segments = 0;
1717 bio->bi_end_io = NULL;
1718 bio->bi_private = NULL;
1720 rdev = rcu_dereference(conf->mirrors[i].rdev);
1722 test_bit(Faulty, &rdev->flags)) {
1725 } else if (!test_bit(In_sync, &rdev->flags)) {
1727 bio->bi_end_io = end_sync_write;
1730 /* may need to read from here */
1732 bio->bi_end_io = end_sync_read;
1733 if (test_bit(WriteMostly, &rdev->flags)) {
1742 atomic_inc(&rdev->nr_pending);
1743 bio->bi_sector = sector_nr + rdev->data_offset;
1744 bio->bi_bdev = rdev->bdev;
1745 bio->bi_private = r1_bio;
1750 r1_bio->read_disk = disk;
1752 if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery) && read_targets > 0)
1753 /* extra read targets are also write targets */
1754 write_targets += read_targets-1;
1756 if (write_targets == 0 || read_targets == 0) {
1757 /* There is nowhere to write, so all non-sync
1758 * drives must be failed - so we are finished
1760 sector_t rv = max_sector - sector_nr;
1770 int len = PAGE_SIZE;
1771 if (sector_nr + (len>>9) > max_sector)
1772 len = (max_sector - sector_nr) << 9;
1775 if (sync_blocks == 0) {
1776 if (!bitmap_start_sync(mddev->bitmap, sector_nr,
1777 &sync_blocks, still_degraded) &&
1779 !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery))
1781 BUG_ON(sync_blocks < (PAGE_SIZE>>9));
1782 if (len > (sync_blocks<<9))
1783 len = sync_blocks<<9;
1786 for (i=0 ; i < conf->raid_disks; i++) {
1787 bio = r1_bio->bios[i];
1788 if (bio->bi_end_io) {
1789 page = bio->bi_io_vec[bio->bi_vcnt].bv_page;
1790 if (bio_add_page(bio, page, len, 0) == 0) {
1792 bio->bi_io_vec[bio->bi_vcnt].bv_page = page;
1795 bio = r1_bio->bios[i];
1796 if (bio->bi_end_io==NULL)
1798 /* remove last page from this bio */
1800 bio->bi_size -= len;
1801 bio->bi_flags &= ~(1<< BIO_SEG_VALID);
1807 nr_sectors += len>>9;
1808 sector_nr += len>>9;
1809 sync_blocks -= (len>>9);
1810 } while (r1_bio->bios[disk]->bi_vcnt < RESYNC_PAGES);
1812 r1_bio->sectors = nr_sectors;
1814 /* For a user-requested sync, we read all readable devices and do a
1817 if (test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
1818 atomic_set(&r1_bio->remaining, read_targets);
1819 for (i=0; i<conf->raid_disks; i++) {
1820 bio = r1_bio->bios[i];
1821 if (bio->bi_end_io == end_sync_read) {
1822 md_sync_acct(bio->bi_bdev, nr_sectors);
1823 generic_make_request(bio);
1827 atomic_set(&r1_bio->remaining, 1);
1828 bio = r1_bio->bios[r1_bio->read_disk];
1829 md_sync_acct(bio->bi_bdev, nr_sectors);
1830 generic_make_request(bio);
1836 static int run(mddev_t *mddev)
1840 mirror_info_t *disk;
1842 struct list_head *tmp;
1844 if (mddev->level != 1) {
1845 printk("raid1: %s: raid level not set to mirroring (%d)\n",
1846 mdname(mddev), mddev->level);
1849 if (mddev->reshape_position != MaxSector) {
1850 printk("raid1: %s: reshape_position set but not supported\n",
1855 * copy the already verified devices into our private RAID1
1856 * bookkeeping area. [whatever we allocate in run(),
1857 * should be freed in stop()]
1859 conf = kzalloc(sizeof(conf_t), GFP_KERNEL);
1860 mddev->private = conf;
1864 conf->mirrors = kzalloc(sizeof(struct mirror_info)*mddev->raid_disks,
1869 conf->tmppage = alloc_page(GFP_KERNEL);
1873 conf->poolinfo = kmalloc(sizeof(*conf->poolinfo), GFP_KERNEL);
1874 if (!conf->poolinfo)
1876 conf->poolinfo->mddev = mddev;
1877 conf->poolinfo->raid_disks = mddev->raid_disks;
1878 conf->r1bio_pool = mempool_create(NR_RAID1_BIOS, r1bio_pool_alloc,
1881 if (!conf->r1bio_pool)
1884 ITERATE_RDEV(mddev, rdev, tmp) {
1885 disk_idx = rdev->raid_disk;
1886 if (disk_idx >= mddev->raid_disks
1889 disk = conf->mirrors + disk_idx;
1893 blk_queue_stack_limits(mddev->queue,
1894 rdev->bdev->bd_disk->queue);
1895 /* as we don't honour merge_bvec_fn, we must never risk
1896 * violating it, so limit ->max_sector to one PAGE, as
1897 * a one page request is never in violation.
1899 if (rdev->bdev->bd_disk->queue->merge_bvec_fn &&
1900 mddev->queue->max_sectors > (PAGE_SIZE>>9))
1901 blk_queue_max_sectors(mddev->queue, PAGE_SIZE>>9);
1903 disk->head_position = 0;
1904 if (!test_bit(Faulty, &rdev->flags) && test_bit(In_sync, &rdev->flags))
1905 conf->working_disks++;
1907 conf->raid_disks = mddev->raid_disks;
1908 conf->mddev = mddev;
1909 spin_lock_init(&conf->device_lock);
1910 INIT_LIST_HEAD(&conf->retry_list);
1911 if (conf->working_disks == 1)
1912 mddev->recovery_cp = MaxSector;
1914 spin_lock_init(&conf->resync_lock);
1915 init_waitqueue_head(&conf->wait_barrier);
1917 bio_list_init(&conf->pending_bio_list);
1918 bio_list_init(&conf->flushing_bio_list);
1920 if (!conf->working_disks) {
1921 printk(KERN_ERR "raid1: no operational mirrors for %s\n",
1926 mddev->degraded = 0;
1927 for (i = 0; i < conf->raid_disks; i++) {
1929 disk = conf->mirrors + i;
1932 !test_bit(In_sync, &disk->rdev->flags)) {
1933 disk->head_position = 0;
1939 * find the first working one and use it as a starting point
1940 * to read balancing.
1942 for (j = 0; j < conf->raid_disks &&
1943 (!conf->mirrors[j].rdev ||
1944 !test_bit(In_sync, &conf->mirrors[j].rdev->flags)) ; j++)
1946 conf->last_used = j;
1949 mddev->thread = md_register_thread(raid1d, mddev, "%s_raid1");
1950 if (!mddev->thread) {
1952 "raid1: couldn't allocate thread for %s\n",
1958 "raid1: raid set %s active with %d out of %d mirrors\n",
1959 mdname(mddev), mddev->raid_disks - mddev->degraded,
1962 * Ok, everything is just fine now
1964 mddev->array_size = mddev->size;
1966 mddev->queue->unplug_fn = raid1_unplug;
1967 mddev->queue->issue_flush_fn = raid1_issue_flush;
1972 printk(KERN_ERR "raid1: couldn't allocate memory for %s\n",
1977 if (conf->r1bio_pool)
1978 mempool_destroy(conf->r1bio_pool);
1979 kfree(conf->mirrors);
1980 safe_put_page(conf->tmppage);
1981 kfree(conf->poolinfo);
1983 mddev->private = NULL;
1989 static int stop(mddev_t *mddev)
1991 conf_t *conf = mddev_to_conf(mddev);
1992 struct bitmap *bitmap = mddev->bitmap;
1993 int behind_wait = 0;
1995 /* wait for behind writes to complete */
1996 while (bitmap && atomic_read(&bitmap->behind_writes) > 0) {
1998 printk(KERN_INFO "raid1: behind writes in progress on device %s, waiting to stop (%d)\n", mdname(mddev), behind_wait);
1999 set_current_state(TASK_UNINTERRUPTIBLE);
2000 schedule_timeout(HZ); /* wait a second */
2001 /* need to kick something here to make sure I/O goes? */
2004 md_unregister_thread(mddev->thread);
2005 mddev->thread = NULL;
2006 blk_sync_queue(mddev->queue); /* the unplug fn references 'conf'*/
2007 if (conf->r1bio_pool)
2008 mempool_destroy(conf->r1bio_pool);
2009 kfree(conf->mirrors);
2010 kfree(conf->poolinfo);
2012 mddev->private = NULL;
2016 static int raid1_resize(mddev_t *mddev, sector_t sectors)
2018 /* no resync is happening, and there is enough space
2019 * on all devices, so we can resize.
2020 * We need to make sure resync covers any new space.
2021 * If the array is shrinking we should possibly wait until
2022 * any io in the removed space completes, but it hardly seems
2025 mddev->array_size = sectors>>1;
2026 set_capacity(mddev->gendisk, mddev->array_size << 1);
2028 if (mddev->array_size > mddev->size && mddev->recovery_cp == MaxSector) {
2029 mddev->recovery_cp = mddev->size << 1;
2030 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
2032 mddev->size = mddev->array_size;
2033 mddev->resync_max_sectors = sectors;
2037 static int raid1_reshape(mddev_t *mddev)
2040 * 1/ resize the r1bio_pool
2041 * 2/ resize conf->mirrors
2043 * We allocate a new r1bio_pool if we can.
2044 * Then raise a device barrier and wait until all IO stops.
2045 * Then resize conf->mirrors and swap in the new r1bio pool.
2047 * At the same time, we "pack" the devices so that all the missing
2048 * devices have the higher raid_disk numbers.
2050 mempool_t *newpool, *oldpool;
2051 struct pool_info *newpoolinfo;
2052 mirror_info_t *newmirrors;
2053 conf_t *conf = mddev_to_conf(mddev);
2054 int cnt, raid_disks;
2058 /* Cannot change chunk_size, layout, or level */
2059 if (mddev->chunk_size != mddev->new_chunk ||
2060 mddev->layout != mddev->new_layout ||
2061 mddev->level != mddev->new_level) {
2062 mddev->new_chunk = mddev->chunk_size;
2063 mddev->new_layout = mddev->layout;
2064 mddev->new_level = mddev->level;
2068 raid_disks = mddev->raid_disks + mddev->delta_disks;
2070 if (raid_disks < conf->raid_disks) {
2072 for (d= 0; d < conf->raid_disks; d++)
2073 if (conf->mirrors[d].rdev)
2075 if (cnt > raid_disks)
2079 newpoolinfo = kmalloc(sizeof(*newpoolinfo), GFP_KERNEL);
2082 newpoolinfo->mddev = mddev;
2083 newpoolinfo->raid_disks = raid_disks;
2085 newpool = mempool_create(NR_RAID1_BIOS, r1bio_pool_alloc,
2086 r1bio_pool_free, newpoolinfo);
2091 newmirrors = kzalloc(sizeof(struct mirror_info) * raid_disks, GFP_KERNEL);
2094 mempool_destroy(newpool);
2098 raise_barrier(conf);
2100 /* ok, everything is stopped */
2101 oldpool = conf->r1bio_pool;
2102 conf->r1bio_pool = newpool;
2104 for (d=d2=0; d < conf->raid_disks; d++)
2105 if (conf->mirrors[d].rdev) {
2106 conf->mirrors[d].rdev->raid_disk = d2;
2107 newmirrors[d2++].rdev = conf->mirrors[d].rdev;
2109 kfree(conf->mirrors);
2110 conf->mirrors = newmirrors;
2111 kfree(conf->poolinfo);
2112 conf->poolinfo = newpoolinfo;
2114 mddev->degraded += (raid_disks - conf->raid_disks);
2115 conf->raid_disks = mddev->raid_disks = raid_disks;
2116 mddev->delta_disks = 0;
2118 conf->last_used = 0; /* just make sure it is in-range */
2119 lower_barrier(conf);
2121 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
2122 md_wakeup_thread(mddev->thread);
2124 mempool_destroy(oldpool);
2128 static void raid1_quiesce(mddev_t *mddev, int state)
2130 conf_t *conf = mddev_to_conf(mddev);
2134 raise_barrier(conf);
2137 lower_barrier(conf);
2143 static struct mdk_personality raid1_personality =
2147 .owner = THIS_MODULE,
2148 .make_request = make_request,
2152 .error_handler = error,
2153 .hot_add_disk = raid1_add_disk,
2154 .hot_remove_disk= raid1_remove_disk,
2155 .spare_active = raid1_spare_active,
2156 .sync_request = sync_request,
2157 .resize = raid1_resize,
2158 .check_reshape = raid1_reshape,
2159 .quiesce = raid1_quiesce,
2162 static int __init raid_init(void)
2164 return register_md_personality(&raid1_personality);
2167 static void raid_exit(void)
2169 unregister_md_personality(&raid1_personality);
2172 module_init(raid_init);
2173 module_exit(raid_exit);
2174 MODULE_LICENSE("GPL");
2175 MODULE_ALIAS("md-personality-3"); /* RAID1 */
2176 MODULE_ALIAS("md-raid1");
2177 MODULE_ALIAS("md-level-1");