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 <linux/slab.h>
35 #include <linux/delay.h>
36 #include <linux/blkdev.h>
37 #include <linux/seq_file.h>
44 #define PRINTK(x...) printk(x)
50 * Number of guaranteed r1bios in case of extreme VM load:
52 #define NR_RAID1_BIOS 256
55 static void unplug_slaves(mddev_t *mddev);
57 static void allow_barrier(conf_t *conf);
58 static void lower_barrier(conf_t *conf);
60 static void * r1bio_pool_alloc(gfp_t gfp_flags, void *data)
62 struct pool_info *pi = data;
64 int size = offsetof(r1bio_t, bios[pi->raid_disks]);
66 /* allocate a r1bio with room for raid_disks entries in the bios array */
67 r1_bio = kzalloc(size, gfp_flags);
68 if (!r1_bio && pi->mddev)
69 unplug_slaves(pi->mddev);
74 static void r1bio_pool_free(void *r1_bio, void *data)
79 #define RESYNC_BLOCK_SIZE (64*1024)
80 //#define RESYNC_BLOCK_SIZE PAGE_SIZE
81 #define RESYNC_SECTORS (RESYNC_BLOCK_SIZE >> 9)
82 #define RESYNC_PAGES ((RESYNC_BLOCK_SIZE + PAGE_SIZE-1) / PAGE_SIZE)
83 #define RESYNC_WINDOW (2048*1024)
85 static void * r1buf_pool_alloc(gfp_t gfp_flags, void *data)
87 struct pool_info *pi = data;
93 r1_bio = r1bio_pool_alloc(gfp_flags, pi);
95 unplug_slaves(pi->mddev);
100 * Allocate bios : 1 for reading, n-1 for writing
102 for (j = pi->raid_disks ; j-- ; ) {
103 bio = bio_kmalloc(gfp_flags, RESYNC_PAGES);
106 r1_bio->bios[j] = bio;
109 * Allocate RESYNC_PAGES data pages and attach them to
111 * If this is a user-requested check/repair, allocate
112 * RESYNC_PAGES for each bio.
114 if (test_bit(MD_RECOVERY_REQUESTED, &pi->mddev->recovery))
119 bio = r1_bio->bios[j];
120 for (i = 0; i < RESYNC_PAGES; i++) {
121 page = alloc_page(gfp_flags);
125 bio->bi_io_vec[i].bv_page = page;
129 /* If not user-requests, copy the page pointers to all bios */
130 if (!test_bit(MD_RECOVERY_REQUESTED, &pi->mddev->recovery)) {
131 for (i=0; i<RESYNC_PAGES ; i++)
132 for (j=1; j<pi->raid_disks; j++)
133 r1_bio->bios[j]->bi_io_vec[i].bv_page =
134 r1_bio->bios[0]->bi_io_vec[i].bv_page;
137 r1_bio->master_bio = NULL;
142 for (j=0 ; j < pi->raid_disks; j++)
143 for (i=0; i < r1_bio->bios[j]->bi_vcnt ; i++)
144 put_page(r1_bio->bios[j]->bi_io_vec[i].bv_page);
147 while ( ++j < pi->raid_disks )
148 bio_put(r1_bio->bios[j]);
149 r1bio_pool_free(r1_bio, data);
153 static void r1buf_pool_free(void *__r1_bio, void *data)
155 struct pool_info *pi = data;
157 r1bio_t *r1bio = __r1_bio;
159 for (i = 0; i < RESYNC_PAGES; i++)
160 for (j = pi->raid_disks; j-- ;) {
162 r1bio->bios[j]->bi_io_vec[i].bv_page !=
163 r1bio->bios[0]->bi_io_vec[i].bv_page)
164 safe_put_page(r1bio->bios[j]->bi_io_vec[i].bv_page);
166 for (i=0 ; i < pi->raid_disks; i++)
167 bio_put(r1bio->bios[i]);
169 r1bio_pool_free(r1bio, data);
172 static void put_all_bios(conf_t *conf, r1bio_t *r1_bio)
176 for (i = 0; i < conf->raid_disks; i++) {
177 struct bio **bio = r1_bio->bios + i;
178 if (*bio && *bio != IO_BLOCKED)
184 static void free_r1bio(r1bio_t *r1_bio)
186 conf_t *conf = r1_bio->mddev->private;
189 * Wake up any possible resync thread that waits for the device
194 put_all_bios(conf, r1_bio);
195 mempool_free(r1_bio, conf->r1bio_pool);
198 static void put_buf(r1bio_t *r1_bio)
200 conf_t *conf = r1_bio->mddev->private;
203 for (i=0; i<conf->raid_disks; i++) {
204 struct bio *bio = r1_bio->bios[i];
206 rdev_dec_pending(conf->mirrors[i].rdev, r1_bio->mddev);
209 mempool_free(r1_bio, conf->r1buf_pool);
214 static void reschedule_retry(r1bio_t *r1_bio)
217 mddev_t *mddev = r1_bio->mddev;
218 conf_t *conf = mddev->private;
220 spin_lock_irqsave(&conf->device_lock, flags);
221 list_add(&r1_bio->retry_list, &conf->retry_list);
223 spin_unlock_irqrestore(&conf->device_lock, flags);
225 wake_up(&conf->wait_barrier);
226 md_wakeup_thread(mddev->thread);
230 * raid_end_bio_io() is called when we have finished servicing a mirrored
231 * operation and are ready to return a success/failure code to the buffer
234 static void raid_end_bio_io(r1bio_t *r1_bio)
236 struct bio *bio = r1_bio->master_bio;
238 /* if nobody has done the final endio yet, do it now */
239 if (!test_and_set_bit(R1BIO_Returned, &r1_bio->state)) {
240 PRINTK(KERN_DEBUG "raid1: sync end %s on sectors %llu-%llu\n",
241 (bio_data_dir(bio) == WRITE) ? "write" : "read",
242 (unsigned long long) bio->bi_sector,
243 (unsigned long long) bio->bi_sector +
244 (bio->bi_size >> 9) - 1);
247 test_bit(R1BIO_Uptodate, &r1_bio->state) ? 0 : -EIO);
253 * Update disk head position estimator based on IRQ completion info.
255 static inline void update_head_pos(int disk, r1bio_t *r1_bio)
257 conf_t *conf = r1_bio->mddev->private;
259 conf->mirrors[disk].head_position =
260 r1_bio->sector + (r1_bio->sectors);
263 static void raid1_end_read_request(struct bio *bio, int error)
265 int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
266 r1bio_t *r1_bio = bio->bi_private;
268 conf_t *conf = r1_bio->mddev->private;
270 mirror = r1_bio->read_disk;
272 * this branch is our 'one mirror IO has finished' event handler:
274 update_head_pos(mirror, r1_bio);
277 set_bit(R1BIO_Uptodate, &r1_bio->state);
279 /* If all other devices have failed, we want to return
280 * the error upwards rather than fail the last device.
281 * Here we redefine "uptodate" to mean "Don't want to retry"
284 spin_lock_irqsave(&conf->device_lock, flags);
285 if (r1_bio->mddev->degraded == conf->raid_disks ||
286 (r1_bio->mddev->degraded == conf->raid_disks-1 &&
287 !test_bit(Faulty, &conf->mirrors[mirror].rdev->flags)))
289 spin_unlock_irqrestore(&conf->device_lock, flags);
293 raid_end_bio_io(r1_bio);
298 char b[BDEVNAME_SIZE];
299 if (printk_ratelimit())
300 printk(KERN_ERR "md/raid1:%s: %s: rescheduling sector %llu\n",
302 bdevname(conf->mirrors[mirror].rdev->bdev,b), (unsigned long long)r1_bio->sector);
303 reschedule_retry(r1_bio);
306 rdev_dec_pending(conf->mirrors[mirror].rdev, conf->mddev);
309 static void r1_bio_write_done(r1bio_t *r1_bio, int vcnt, struct bio_vec *bv,
312 if (atomic_dec_and_test(&r1_bio->remaining))
314 /* it really is the end of this request */
315 if (test_bit(R1BIO_BehindIO, &r1_bio->state)) {
316 /* free extra copy of the data pages */
319 safe_put_page(bv[i].bv_page);
321 /* clear the bitmap if all writes complete successfully */
322 bitmap_endwrite(r1_bio->mddev->bitmap, r1_bio->sector,
324 !test_bit(R1BIO_Degraded, &r1_bio->state),
326 md_write_end(r1_bio->mddev);
327 raid_end_bio_io(r1_bio);
331 static void raid1_end_write_request(struct bio *bio, int error)
333 int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
334 r1bio_t *r1_bio = bio->bi_private;
335 int mirror, behind = test_bit(R1BIO_BehindIO, &r1_bio->state);
336 conf_t *conf = r1_bio->mddev->private;
337 struct bio *to_put = NULL;
340 for (mirror = 0; mirror < conf->raid_disks; mirror++)
341 if (r1_bio->bios[mirror] == bio)
345 * 'one mirror IO has finished' event handler:
347 r1_bio->bios[mirror] = NULL;
350 md_error(r1_bio->mddev, conf->mirrors[mirror].rdev);
351 /* an I/O failed, we can't clear the bitmap */
352 set_bit(R1BIO_Degraded, &r1_bio->state);
355 * Set R1BIO_Uptodate in our master bio, so that we
356 * will return a good error code for to the higher
357 * levels even if IO on some other mirrored buffer
360 * The 'master' represents the composite IO operation
361 * to user-side. So if something waits for IO, then it
362 * will wait for the 'master' bio.
364 set_bit(R1BIO_Uptodate, &r1_bio->state);
366 update_head_pos(mirror, r1_bio);
369 if (test_bit(WriteMostly, &conf->mirrors[mirror].rdev->flags))
370 atomic_dec(&r1_bio->behind_remaining);
373 * In behind mode, we ACK the master bio once the I/O
374 * has safely reached all non-writemostly
375 * disks. Setting the Returned bit ensures that this
376 * gets done only once -- we don't ever want to return
377 * -EIO here, instead we'll wait
379 if (atomic_read(&r1_bio->behind_remaining) >= (atomic_read(&r1_bio->remaining)-1) &&
380 test_bit(R1BIO_Uptodate, &r1_bio->state)) {
381 /* Maybe we can return now */
382 if (!test_and_set_bit(R1BIO_Returned, &r1_bio->state)) {
383 struct bio *mbio = r1_bio->master_bio;
384 PRINTK(KERN_DEBUG "raid1: behind end write sectors %llu-%llu\n",
385 (unsigned long long) mbio->bi_sector,
386 (unsigned long long) mbio->bi_sector +
387 (mbio->bi_size >> 9) - 1);
392 rdev_dec_pending(conf->mirrors[mirror].rdev, conf->mddev);
395 * Let's see if all mirrored write operations have finished
398 r1_bio_write_done(r1_bio, bio->bi_vcnt, bio->bi_io_vec, behind);
406 * This routine returns the disk from which the requested read should
407 * be done. There is a per-array 'next expected sequential IO' sector
408 * number - if this matches on the next IO then we use the last disk.
409 * There is also a per-disk 'last know head position' sector that is
410 * maintained from IRQ contexts, both the normal and the resync IO
411 * completion handlers update this position correctly. If there is no
412 * perfect sequential match then we pick the disk whose head is closest.
414 * If there are 2 mirrors in the same 2 devices, performance degrades
415 * because position is mirror, not device based.
417 * The rdev for the device selected will have nr_pending incremented.
419 static int read_balance(conf_t *conf, r1bio_t *r1_bio)
421 const sector_t this_sector = r1_bio->sector;
422 int new_disk = conf->last_used, disk = new_disk;
424 const int sectors = r1_bio->sectors;
425 sector_t new_distance, current_distance;
430 * Check if we can balance. We can balance on the whole
431 * device if no resync is going on, or below the resync window.
432 * We take the first readable disk when above the resync window.
435 if (conf->mddev->recovery_cp < MaxSector &&
436 (this_sector + sectors >= conf->next_resync)) {
437 /* Choose the first operational device, for consistancy */
440 for (rdev = rcu_dereference(conf->mirrors[new_disk].rdev);
441 r1_bio->bios[new_disk] == IO_BLOCKED ||
442 !rdev || !test_bit(In_sync, &rdev->flags)
443 || test_bit(WriteMostly, &rdev->flags);
444 rdev = rcu_dereference(conf->mirrors[++new_disk].rdev)) {
446 if (rdev && test_bit(In_sync, &rdev->flags) &&
447 r1_bio->bios[new_disk] != IO_BLOCKED)
448 wonly_disk = new_disk;
450 if (new_disk == conf->raid_disks - 1) {
451 new_disk = wonly_disk;
459 /* make sure the disk is operational */
460 for (rdev = rcu_dereference(conf->mirrors[new_disk].rdev);
461 r1_bio->bios[new_disk] == IO_BLOCKED ||
462 !rdev || !test_bit(In_sync, &rdev->flags) ||
463 test_bit(WriteMostly, &rdev->flags);
464 rdev = rcu_dereference(conf->mirrors[new_disk].rdev)) {
466 if (rdev && test_bit(In_sync, &rdev->flags) &&
467 r1_bio->bios[new_disk] != IO_BLOCKED)
468 wonly_disk = new_disk;
471 new_disk = conf->raid_disks;
473 if (new_disk == disk) {
474 new_disk = wonly_disk;
483 /* now disk == new_disk == starting point for search */
486 * Don't change to another disk for sequential reads:
488 if (conf->next_seq_sect == this_sector)
490 if (this_sector == conf->mirrors[new_disk].head_position)
493 current_distance = abs(this_sector - conf->mirrors[disk].head_position);
495 /* Find the disk whose head is closest */
499 disk = conf->raid_disks;
502 rdev = rcu_dereference(conf->mirrors[disk].rdev);
504 if (!rdev || r1_bio->bios[disk] == IO_BLOCKED ||
505 !test_bit(In_sync, &rdev->flags) ||
506 test_bit(WriteMostly, &rdev->flags))
509 if (!atomic_read(&rdev->nr_pending)) {
513 new_distance = abs(this_sector - conf->mirrors[disk].head_position);
514 if (new_distance < current_distance) {
515 current_distance = new_distance;
518 } while (disk != conf->last_used);
524 rdev = rcu_dereference(conf->mirrors[new_disk].rdev);
527 atomic_inc(&rdev->nr_pending);
528 if (!test_bit(In_sync, &rdev->flags)) {
529 /* cannot risk returning a device that failed
530 * before we inc'ed nr_pending
532 rdev_dec_pending(rdev, conf->mddev);
535 conf->next_seq_sect = this_sector + sectors;
536 conf->last_used = new_disk;
543 static void unplug_slaves(mddev_t *mddev)
545 conf_t *conf = mddev->private;
549 for (i=0; i<mddev->raid_disks; i++) {
550 mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[i].rdev);
551 if (rdev && !test_bit(Faulty, &rdev->flags) && atomic_read(&rdev->nr_pending)) {
552 struct request_queue *r_queue = bdev_get_queue(rdev->bdev);
554 atomic_inc(&rdev->nr_pending);
559 rdev_dec_pending(rdev, mddev);
566 static void raid1_unplug(struct request_queue *q)
568 mddev_t *mddev = q->queuedata;
570 unplug_slaves(mddev);
571 md_wakeup_thread(mddev->thread);
574 static int raid1_congested(void *data, int bits)
576 mddev_t *mddev = data;
577 conf_t *conf = mddev->private;
580 if (mddev_congested(mddev, bits))
584 for (i = 0; i < mddev->raid_disks; i++) {
585 mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[i].rdev);
586 if (rdev && !test_bit(Faulty, &rdev->flags)) {
587 struct request_queue *q = bdev_get_queue(rdev->bdev);
589 /* Note the '|| 1' - when read_balance prefers
590 * non-congested targets, it can be removed
592 if ((bits & (1<<BDI_async_congested)) || 1)
593 ret |= bdi_congested(&q->backing_dev_info, bits);
595 ret &= bdi_congested(&q->backing_dev_info, bits);
603 static int flush_pending_writes(conf_t *conf)
605 /* Any writes that have been queued but are awaiting
606 * bitmap updates get flushed here.
607 * We return 1 if any requests were actually submitted.
611 spin_lock_irq(&conf->device_lock);
613 if (conf->pending_bio_list.head) {
615 bio = bio_list_get(&conf->pending_bio_list);
616 blk_remove_plug(conf->mddev->queue);
617 spin_unlock_irq(&conf->device_lock);
618 /* flush any pending bitmap writes to
619 * disk before proceeding w/ I/O */
620 bitmap_unplug(conf->mddev->bitmap);
622 while (bio) { /* submit pending writes */
623 struct bio *next = bio->bi_next;
625 generic_make_request(bio);
630 spin_unlock_irq(&conf->device_lock);
635 * Sometimes we need to suspend IO while we do something else,
636 * either some resync/recovery, or reconfigure the array.
637 * To do this we raise a 'barrier'.
638 * The 'barrier' is a counter that can be raised multiple times
639 * to count how many activities are happening which preclude
641 * We can only raise the barrier if there is no pending IO.
642 * i.e. if nr_pending == 0.
643 * We choose only to raise the barrier if no-one is waiting for the
644 * barrier to go down. This means that as soon as an IO request
645 * is ready, no other operations which require a barrier will start
646 * until the IO request has had a chance.
648 * So: regular IO calls 'wait_barrier'. When that returns there
649 * is no backgroup IO happening, It must arrange to call
650 * allow_barrier when it has finished its IO.
651 * backgroup IO calls must call raise_barrier. Once that returns
652 * there is no normal IO happeing. It must arrange to call
653 * lower_barrier when the particular background IO completes.
655 #define RESYNC_DEPTH 32
657 static void raise_barrier(conf_t *conf)
659 spin_lock_irq(&conf->resync_lock);
661 /* Wait until no block IO is waiting */
662 wait_event_lock_irq(conf->wait_barrier, !conf->nr_waiting,
664 raid1_unplug(conf->mddev->queue));
666 /* block any new IO from starting */
669 /* No wait for all pending IO to complete */
670 wait_event_lock_irq(conf->wait_barrier,
671 !conf->nr_pending && conf->barrier < RESYNC_DEPTH,
673 raid1_unplug(conf->mddev->queue));
675 spin_unlock_irq(&conf->resync_lock);
678 static void lower_barrier(conf_t *conf)
681 BUG_ON(conf->barrier <= 0);
682 spin_lock_irqsave(&conf->resync_lock, flags);
684 spin_unlock_irqrestore(&conf->resync_lock, flags);
685 wake_up(&conf->wait_barrier);
688 static void wait_barrier(conf_t *conf)
690 spin_lock_irq(&conf->resync_lock);
693 wait_event_lock_irq(conf->wait_barrier, !conf->barrier,
695 raid1_unplug(conf->mddev->queue));
699 spin_unlock_irq(&conf->resync_lock);
702 static void allow_barrier(conf_t *conf)
705 spin_lock_irqsave(&conf->resync_lock, flags);
707 spin_unlock_irqrestore(&conf->resync_lock, flags);
708 wake_up(&conf->wait_barrier);
711 static void freeze_array(conf_t *conf)
713 /* stop syncio and normal IO and wait for everything to
715 * We increment barrier and nr_waiting, and then
716 * wait until nr_pending match nr_queued+1
717 * This is called in the context of one normal IO request
718 * that has failed. Thus any sync request that might be pending
719 * will be blocked by nr_pending, and we need to wait for
720 * pending IO requests to complete or be queued for re-try.
721 * Thus the number queued (nr_queued) plus this request (1)
722 * must match the number of pending IOs (nr_pending) before
725 spin_lock_irq(&conf->resync_lock);
728 wait_event_lock_irq(conf->wait_barrier,
729 conf->nr_pending == conf->nr_queued+1,
731 ({ flush_pending_writes(conf);
732 raid1_unplug(conf->mddev->queue); }));
733 spin_unlock_irq(&conf->resync_lock);
735 static void unfreeze_array(conf_t *conf)
737 /* reverse the effect of the freeze */
738 spin_lock_irq(&conf->resync_lock);
741 wake_up(&conf->wait_barrier);
742 spin_unlock_irq(&conf->resync_lock);
746 /* duplicate the data pages for behind I/O
747 * We return a list of bio_vec rather than just page pointers
748 * as it makes freeing easier
750 static struct bio_vec *alloc_behind_pages(struct bio *bio)
753 struct bio_vec *bvec;
754 struct bio_vec *pages = kzalloc(bio->bi_vcnt * sizeof(struct bio_vec),
756 if (unlikely(!pages))
759 bio_for_each_segment(bvec, bio, i) {
760 pages[i].bv_page = alloc_page(GFP_NOIO);
761 if (unlikely(!pages[i].bv_page))
763 memcpy(kmap(pages[i].bv_page) + bvec->bv_offset,
764 kmap(bvec->bv_page) + bvec->bv_offset, bvec->bv_len);
765 kunmap(pages[i].bv_page);
766 kunmap(bvec->bv_page);
773 for (i = 0; i < bio->bi_vcnt && pages[i].bv_page; i++)
774 put_page(pages[i].bv_page);
776 PRINTK("%dB behind alloc failed, doing sync I/O\n", bio->bi_size);
780 static int make_request(mddev_t *mddev, struct bio * bio)
782 conf_t *conf = mddev->private;
783 mirror_info_t *mirror;
785 struct bio *read_bio;
786 int i, targets = 0, disks;
787 struct bitmap *bitmap;
789 struct bio_vec *behind_pages = NULL;
790 const int rw = bio_data_dir(bio);
791 const unsigned long do_sync = (bio->bi_rw & REQ_SYNC);
792 const unsigned long do_flush_fua = (bio->bi_rw & (REQ_FLUSH | REQ_FUA));
793 mdk_rdev_t *blocked_rdev;
796 * Register the new request and wait if the reconstruction
797 * thread has put up a bar for new requests.
798 * Continue immediately if no resync is active currently.
801 md_write_start(mddev, bio); /* wait on superblock update early */
803 if (bio_data_dir(bio) == WRITE &&
804 bio->bi_sector + bio->bi_size/512 > mddev->suspend_lo &&
805 bio->bi_sector < mddev->suspend_hi) {
806 /* As the suspend_* range is controlled by
807 * userspace, we want an interruptible
812 flush_signals(current);
813 prepare_to_wait(&conf->wait_barrier,
814 &w, TASK_INTERRUPTIBLE);
815 if (bio->bi_sector + bio->bi_size/512 <= mddev->suspend_lo ||
816 bio->bi_sector >= mddev->suspend_hi)
820 finish_wait(&conf->wait_barrier, &w);
825 bitmap = mddev->bitmap;
828 * make_request() can abort the operation when READA is being
829 * used and no empty request is available.
832 r1_bio = mempool_alloc(conf->r1bio_pool, GFP_NOIO);
834 r1_bio->master_bio = bio;
835 r1_bio->sectors = bio->bi_size >> 9;
837 r1_bio->mddev = mddev;
838 r1_bio->sector = bio->bi_sector;
842 * read balancing logic:
844 int rdisk = read_balance(conf, r1_bio);
847 /* couldn't find anywhere to read from */
848 raid_end_bio_io(r1_bio);
851 mirror = conf->mirrors + rdisk;
853 if (test_bit(WriteMostly, &mirror->rdev->flags) &&
855 /* Reading from a write-mostly device must
856 * take care not to over-take any writes
859 wait_event(bitmap->behind_wait,
860 atomic_read(&bitmap->behind_writes) == 0);
862 r1_bio->read_disk = rdisk;
864 read_bio = bio_clone_mddev(bio, GFP_NOIO, mddev);
866 r1_bio->bios[rdisk] = read_bio;
868 read_bio->bi_sector = r1_bio->sector + mirror->rdev->data_offset;
869 read_bio->bi_bdev = mirror->rdev->bdev;
870 read_bio->bi_end_io = raid1_end_read_request;
871 read_bio->bi_rw = READ | do_sync;
872 read_bio->bi_private = r1_bio;
874 generic_make_request(read_bio);
881 /* first select target devices under spinlock and
882 * inc refcount on their rdev. Record them by setting
885 disks = conf->raid_disks;
889 for (i = 0; i < disks; i++) {
890 mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[i].rdev);
891 if (rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
892 atomic_inc(&rdev->nr_pending);
896 if (rdev && !test_bit(Faulty, &rdev->flags)) {
897 atomic_inc(&rdev->nr_pending);
898 if (test_bit(Faulty, &rdev->flags)) {
899 rdev_dec_pending(rdev, mddev);
900 r1_bio->bios[i] = NULL;
902 r1_bio->bios[i] = bio;
906 r1_bio->bios[i] = NULL;
910 if (unlikely(blocked_rdev)) {
911 /* Wait for this device to become unblocked */
914 for (j = 0; j < i; j++)
916 rdev_dec_pending(conf->mirrors[j].rdev, mddev);
919 md_wait_for_blocked_rdev(blocked_rdev, mddev);
924 BUG_ON(targets == 0); /* we never fail the last device */
926 if (targets < conf->raid_disks) {
927 /* array is degraded, we will not clear the bitmap
928 * on I/O completion (see raid1_end_write_request) */
929 set_bit(R1BIO_Degraded, &r1_bio->state);
933 * Not if there are too many, or cannot allocate memory,
934 * or a reader on WriteMostly is waiting for behind writes
937 (atomic_read(&bitmap->behind_writes)
938 < mddev->bitmap_info.max_write_behind) &&
939 !waitqueue_active(&bitmap->behind_wait) &&
940 (behind_pages = alloc_behind_pages(bio)) != NULL)
941 set_bit(R1BIO_BehindIO, &r1_bio->state);
943 atomic_set(&r1_bio->remaining, 1);
944 atomic_set(&r1_bio->behind_remaining, 0);
946 bitmap_startwrite(bitmap, bio->bi_sector, r1_bio->sectors,
947 test_bit(R1BIO_BehindIO, &r1_bio->state));
948 for (i = 0; i < disks; i++) {
950 if (!r1_bio->bios[i])
953 mbio = bio_clone_mddev(bio, GFP_NOIO, mddev);
954 r1_bio->bios[i] = mbio;
956 mbio->bi_sector = r1_bio->sector + conf->mirrors[i].rdev->data_offset;
957 mbio->bi_bdev = conf->mirrors[i].rdev->bdev;
958 mbio->bi_end_io = raid1_end_write_request;
959 mbio->bi_rw = WRITE | do_flush_fua | do_sync;
960 mbio->bi_private = r1_bio;
963 struct bio_vec *bvec;
966 /* Yes, I really want the '__' version so that
967 * we clear any unused pointer in the io_vec, rather
968 * than leave them unchanged. This is important
969 * because when we come to free the pages, we won't
970 * know the originial bi_idx, so we just free
973 __bio_for_each_segment(bvec, mbio, j, 0)
974 bvec->bv_page = behind_pages[j].bv_page;
975 if (test_bit(WriteMostly, &conf->mirrors[i].rdev->flags))
976 atomic_inc(&r1_bio->behind_remaining);
979 atomic_inc(&r1_bio->remaining);
980 spin_lock_irqsave(&conf->device_lock, flags);
981 bio_list_add(&conf->pending_bio_list, mbio);
982 blk_plug_device(mddev->queue);
983 spin_unlock_irqrestore(&conf->device_lock, flags);
985 r1_bio_write_done(r1_bio, bio->bi_vcnt, behind_pages, behind_pages != NULL);
986 kfree(behind_pages); /* the behind pages are attached to the bios now */
988 /* In case raid1d snuck in to freeze_array */
989 wake_up(&conf->wait_barrier);
992 md_wakeup_thread(mddev->thread);
997 static void status(struct seq_file *seq, mddev_t *mddev)
999 conf_t *conf = mddev->private;
1002 seq_printf(seq, " [%d/%d] [", conf->raid_disks,
1003 conf->raid_disks - mddev->degraded);
1005 for (i = 0; i < conf->raid_disks; i++) {
1006 mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[i].rdev);
1007 seq_printf(seq, "%s",
1008 rdev && test_bit(In_sync, &rdev->flags) ? "U" : "_");
1011 seq_printf(seq, "]");
1015 static void error(mddev_t *mddev, mdk_rdev_t *rdev)
1017 char b[BDEVNAME_SIZE];
1018 conf_t *conf = mddev->private;
1021 * If it is not operational, then we have already marked it as dead
1022 * else if it is the last working disks, ignore the error, let the
1023 * next level up know.
1024 * else mark the drive as failed
1026 if (test_bit(In_sync, &rdev->flags)
1027 && (conf->raid_disks - mddev->degraded) == 1) {
1029 * Don't fail the drive, act as though we were just a
1030 * normal single drive.
1031 * However don't try a recovery from this drive as
1032 * it is very likely to fail.
1034 mddev->recovery_disabled = 1;
1037 if (test_and_clear_bit(In_sync, &rdev->flags)) {
1038 unsigned long flags;
1039 spin_lock_irqsave(&conf->device_lock, flags);
1041 set_bit(Faulty, &rdev->flags);
1042 spin_unlock_irqrestore(&conf->device_lock, flags);
1044 * if recovery is running, make sure it aborts.
1046 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1048 set_bit(Faulty, &rdev->flags);
1049 set_bit(MD_CHANGE_DEVS, &mddev->flags);
1050 printk(KERN_ALERT "md/raid1:%s: Disk failure on %s, disabling device.\n"
1051 KERN_ALERT "md/raid1:%s: Operation continuing on %d devices.\n",
1052 mdname(mddev), bdevname(rdev->bdev, b),
1053 mdname(mddev), conf->raid_disks - mddev->degraded);
1056 static void print_conf(conf_t *conf)
1060 printk(KERN_DEBUG "RAID1 conf printout:\n");
1062 printk(KERN_DEBUG "(!conf)\n");
1065 printk(KERN_DEBUG " --- wd:%d rd:%d\n", conf->raid_disks - conf->mddev->degraded,
1069 for (i = 0; i < conf->raid_disks; i++) {
1070 char b[BDEVNAME_SIZE];
1071 mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[i].rdev);
1073 printk(KERN_DEBUG " disk %d, wo:%d, o:%d, dev:%s\n",
1074 i, !test_bit(In_sync, &rdev->flags),
1075 !test_bit(Faulty, &rdev->flags),
1076 bdevname(rdev->bdev,b));
1081 static void close_sync(conf_t *conf)
1084 allow_barrier(conf);
1086 mempool_destroy(conf->r1buf_pool);
1087 conf->r1buf_pool = NULL;
1090 static int raid1_spare_active(mddev_t *mddev)
1093 conf_t *conf = mddev->private;
1095 unsigned long flags;
1098 * Find all failed disks within the RAID1 configuration
1099 * and mark them readable.
1100 * Called under mddev lock, so rcu protection not needed.
1102 for (i = 0; i < conf->raid_disks; i++) {
1103 mdk_rdev_t *rdev = conf->mirrors[i].rdev;
1105 && !test_bit(Faulty, &rdev->flags)
1106 && !test_and_set_bit(In_sync, &rdev->flags)) {
1108 sysfs_notify_dirent(rdev->sysfs_state);
1111 spin_lock_irqsave(&conf->device_lock, flags);
1112 mddev->degraded -= count;
1113 spin_unlock_irqrestore(&conf->device_lock, flags);
1120 static int raid1_add_disk(mddev_t *mddev, mdk_rdev_t *rdev)
1122 conf_t *conf = mddev->private;
1127 int last = mddev->raid_disks - 1;
1129 if (rdev->raid_disk >= 0)
1130 first = last = rdev->raid_disk;
1132 for (mirror = first; mirror <= last; mirror++)
1133 if ( !(p=conf->mirrors+mirror)->rdev) {
1135 disk_stack_limits(mddev->gendisk, rdev->bdev,
1136 rdev->data_offset << 9);
1137 /* as we don't honour merge_bvec_fn, we must
1138 * never risk violating it, so limit
1139 * ->max_segments to one lying with a single
1140 * page, as a one page request is never in
1143 if (rdev->bdev->bd_disk->queue->merge_bvec_fn) {
1144 blk_queue_max_segments(mddev->queue, 1);
1145 blk_queue_segment_boundary(mddev->queue,
1146 PAGE_CACHE_SIZE - 1);
1149 p->head_position = 0;
1150 rdev->raid_disk = mirror;
1152 /* As all devices are equivalent, we don't need a full recovery
1153 * if this was recently any drive of the array
1155 if (rdev->saved_raid_disk < 0)
1157 rcu_assign_pointer(p->rdev, rdev);
1160 md_integrity_add_rdev(rdev, mddev);
1165 static int raid1_remove_disk(mddev_t *mddev, int number)
1167 conf_t *conf = mddev->private;
1170 mirror_info_t *p = conf->mirrors+ number;
1175 if (test_bit(In_sync, &rdev->flags) ||
1176 atomic_read(&rdev->nr_pending)) {
1180 /* Only remove non-faulty devices is recovery
1183 if (!test_bit(Faulty, &rdev->flags) &&
1184 mddev->degraded < conf->raid_disks) {
1190 if (atomic_read(&rdev->nr_pending)) {
1191 /* lost the race, try later */
1196 md_integrity_register(mddev);
1205 static void end_sync_read(struct bio *bio, int error)
1207 r1bio_t *r1_bio = bio->bi_private;
1210 for (i=r1_bio->mddev->raid_disks; i--; )
1211 if (r1_bio->bios[i] == bio)
1214 update_head_pos(i, r1_bio);
1216 * we have read a block, now it needs to be re-written,
1217 * or re-read if the read failed.
1218 * We don't do much here, just schedule handling by raid1d
1220 if (test_bit(BIO_UPTODATE, &bio->bi_flags))
1221 set_bit(R1BIO_Uptodate, &r1_bio->state);
1223 if (atomic_dec_and_test(&r1_bio->remaining))
1224 reschedule_retry(r1_bio);
1227 static void end_sync_write(struct bio *bio, int error)
1229 int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
1230 r1bio_t *r1_bio = bio->bi_private;
1231 mddev_t *mddev = r1_bio->mddev;
1232 conf_t *conf = mddev->private;
1236 for (i = 0; i < conf->raid_disks; i++)
1237 if (r1_bio->bios[i] == bio) {
1242 sector_t sync_blocks = 0;
1243 sector_t s = r1_bio->sector;
1244 long sectors_to_go = r1_bio->sectors;
1245 /* make sure these bits doesn't get cleared. */
1247 bitmap_end_sync(mddev->bitmap, s,
1250 sectors_to_go -= sync_blocks;
1251 } while (sectors_to_go > 0);
1252 md_error(mddev, conf->mirrors[mirror].rdev);
1255 update_head_pos(mirror, r1_bio);
1257 if (atomic_dec_and_test(&r1_bio->remaining)) {
1258 sector_t s = r1_bio->sectors;
1260 md_done_sync(mddev, s, uptodate);
1264 static void sync_request_write(mddev_t *mddev, r1bio_t *r1_bio)
1266 conf_t *conf = mddev->private;
1268 int disks = conf->raid_disks;
1269 struct bio *bio, *wbio;
1271 bio = r1_bio->bios[r1_bio->read_disk];
1274 if (test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
1275 /* We have read all readable devices. If we haven't
1276 * got the block, then there is no hope left.
1277 * If we have, then we want to do a comparison
1278 * and skip the write if everything is the same.
1279 * If any blocks failed to read, then we need to
1280 * attempt an over-write
1283 if (!test_bit(R1BIO_Uptodate, &r1_bio->state)) {
1284 for (i=0; i<mddev->raid_disks; i++)
1285 if (r1_bio->bios[i]->bi_end_io == end_sync_read)
1286 md_error(mddev, conf->mirrors[i].rdev);
1288 md_done_sync(mddev, r1_bio->sectors, 1);
1292 for (primary=0; primary<mddev->raid_disks; primary++)
1293 if (r1_bio->bios[primary]->bi_end_io == end_sync_read &&
1294 test_bit(BIO_UPTODATE, &r1_bio->bios[primary]->bi_flags)) {
1295 r1_bio->bios[primary]->bi_end_io = NULL;
1296 rdev_dec_pending(conf->mirrors[primary].rdev, mddev);
1299 r1_bio->read_disk = primary;
1300 for (i=0; i<mddev->raid_disks; i++)
1301 if (r1_bio->bios[i]->bi_end_io == end_sync_read) {
1303 int vcnt = r1_bio->sectors >> (PAGE_SHIFT- 9);
1304 struct bio *pbio = r1_bio->bios[primary];
1305 struct bio *sbio = r1_bio->bios[i];
1307 if (test_bit(BIO_UPTODATE, &sbio->bi_flags)) {
1308 for (j = vcnt; j-- ; ) {
1310 p = pbio->bi_io_vec[j].bv_page;
1311 s = sbio->bi_io_vec[j].bv_page;
1312 if (memcmp(page_address(p),
1320 mddev->resync_mismatches += r1_bio->sectors;
1321 if (j < 0 || (test_bit(MD_RECOVERY_CHECK, &mddev->recovery)
1322 && test_bit(BIO_UPTODATE, &sbio->bi_flags))) {
1323 sbio->bi_end_io = NULL;
1324 rdev_dec_pending(conf->mirrors[i].rdev, mddev);
1326 /* fixup the bio for reuse */
1328 sbio->bi_vcnt = vcnt;
1329 sbio->bi_size = r1_bio->sectors << 9;
1331 sbio->bi_phys_segments = 0;
1332 sbio->bi_flags &= ~(BIO_POOL_MASK - 1);
1333 sbio->bi_flags |= 1 << BIO_UPTODATE;
1334 sbio->bi_next = NULL;
1335 sbio->bi_sector = r1_bio->sector +
1336 conf->mirrors[i].rdev->data_offset;
1337 sbio->bi_bdev = conf->mirrors[i].rdev->bdev;
1338 size = sbio->bi_size;
1339 for (j = 0; j < vcnt ; j++) {
1341 bi = &sbio->bi_io_vec[j];
1343 if (size > PAGE_SIZE)
1344 bi->bv_len = PAGE_SIZE;
1348 memcpy(page_address(bi->bv_page),
1349 page_address(pbio->bi_io_vec[j].bv_page),
1356 if (!test_bit(R1BIO_Uptodate, &r1_bio->state)) {
1357 /* ouch - failed to read all of that.
1358 * Try some synchronous reads of other devices to get
1359 * good data, much like with normal read errors. Only
1360 * read into the pages we already have so we don't
1361 * need to re-issue the read request.
1362 * We don't need to freeze the array, because being in an
1363 * active sync request, there is no normal IO, and
1364 * no overlapping syncs.
1366 sector_t sect = r1_bio->sector;
1367 int sectors = r1_bio->sectors;
1372 int d = r1_bio->read_disk;
1376 if (s > (PAGE_SIZE>>9))
1379 if (r1_bio->bios[d]->bi_end_io == end_sync_read) {
1380 /* No rcu protection needed here devices
1381 * can only be removed when no resync is
1382 * active, and resync is currently active
1384 rdev = conf->mirrors[d].rdev;
1385 if (sync_page_io(rdev,
1386 sect + rdev->data_offset,
1388 bio->bi_io_vec[idx].bv_page,
1395 if (d == conf->raid_disks)
1397 } while (!success && d != r1_bio->read_disk);
1401 /* write it back and re-read */
1402 set_bit(R1BIO_Uptodate, &r1_bio->state);
1403 while (d != r1_bio->read_disk) {
1405 d = conf->raid_disks;
1407 if (r1_bio->bios[d]->bi_end_io != end_sync_read)
1409 rdev = conf->mirrors[d].rdev;
1410 atomic_add(s, &rdev->corrected_errors);
1411 if (sync_page_io(rdev,
1412 sect + rdev->data_offset,
1414 bio->bi_io_vec[idx].bv_page,
1416 md_error(mddev, rdev);
1419 while (d != r1_bio->read_disk) {
1421 d = conf->raid_disks;
1423 if (r1_bio->bios[d]->bi_end_io != end_sync_read)
1425 rdev = conf->mirrors[d].rdev;
1426 if (sync_page_io(rdev,
1427 sect + rdev->data_offset,
1429 bio->bi_io_vec[idx].bv_page,
1431 md_error(mddev, rdev);
1434 char b[BDEVNAME_SIZE];
1435 /* Cannot read from anywhere, array is toast */
1436 md_error(mddev, conf->mirrors[r1_bio->read_disk].rdev);
1437 printk(KERN_ALERT "md/raid1:%s: %s: unrecoverable I/O read error"
1438 " for block %llu\n",
1440 bdevname(bio->bi_bdev, b),
1441 (unsigned long long)r1_bio->sector);
1442 md_done_sync(mddev, r1_bio->sectors, 0);
1455 atomic_set(&r1_bio->remaining, 1);
1456 for (i = 0; i < disks ; i++) {
1457 wbio = r1_bio->bios[i];
1458 if (wbio->bi_end_io == NULL ||
1459 (wbio->bi_end_io == end_sync_read &&
1460 (i == r1_bio->read_disk ||
1461 !test_bit(MD_RECOVERY_SYNC, &mddev->recovery))))
1464 wbio->bi_rw = WRITE;
1465 wbio->bi_end_io = end_sync_write;
1466 atomic_inc(&r1_bio->remaining);
1467 md_sync_acct(conf->mirrors[i].rdev->bdev, wbio->bi_size >> 9);
1469 generic_make_request(wbio);
1472 if (atomic_dec_and_test(&r1_bio->remaining)) {
1473 /* if we're here, all write(s) have completed, so clean up */
1474 md_done_sync(mddev, r1_bio->sectors, 1);
1480 * This is a kernel thread which:
1482 * 1. Retries failed read operations on working mirrors.
1483 * 2. Updates the raid superblock when problems encounter.
1484 * 3. Performs writes following reads for array syncronising.
1487 static void fix_read_error(conf_t *conf, int read_disk,
1488 sector_t sect, int sectors)
1490 mddev_t *mddev = conf->mddev;
1498 if (s > (PAGE_SIZE>>9))
1502 /* Note: no rcu protection needed here
1503 * as this is synchronous in the raid1d thread
1504 * which is the thread that might remove
1505 * a device. If raid1d ever becomes multi-threaded....
1507 rdev = conf->mirrors[d].rdev;
1509 test_bit(In_sync, &rdev->flags) &&
1511 sect + rdev->data_offset,
1513 conf->tmppage, READ))
1517 if (d == conf->raid_disks)
1520 } while (!success && d != read_disk);
1523 /* Cannot read from anywhere -- bye bye array */
1524 md_error(mddev, conf->mirrors[read_disk].rdev);
1527 /* write it back and re-read */
1529 while (d != read_disk) {
1531 d = conf->raid_disks;
1533 rdev = conf->mirrors[d].rdev;
1535 test_bit(In_sync, &rdev->flags)) {
1536 if (sync_page_io(rdev,
1537 sect + rdev->data_offset,
1538 s<<9, conf->tmppage, WRITE)
1540 /* Well, this device is dead */
1541 md_error(mddev, rdev);
1545 while (d != read_disk) {
1546 char b[BDEVNAME_SIZE];
1548 d = conf->raid_disks;
1550 rdev = conf->mirrors[d].rdev;
1552 test_bit(In_sync, &rdev->flags)) {
1553 if (sync_page_io(rdev,
1554 sect + rdev->data_offset,
1555 s<<9, conf->tmppage, READ)
1557 /* Well, this device is dead */
1558 md_error(mddev, rdev);
1560 atomic_add(s, &rdev->corrected_errors);
1562 "md/raid1:%s: read error corrected "
1563 "(%d sectors at %llu on %s)\n",
1565 (unsigned long long)(sect +
1567 bdevname(rdev->bdev, b));
1576 static void raid1d(mddev_t *mddev)
1580 unsigned long flags;
1581 conf_t *conf = mddev->private;
1582 struct list_head *head = &conf->retry_list;
1586 md_check_recovery(mddev);
1589 char b[BDEVNAME_SIZE];
1591 unplug += flush_pending_writes(conf);
1593 spin_lock_irqsave(&conf->device_lock, flags);
1594 if (list_empty(head)) {
1595 spin_unlock_irqrestore(&conf->device_lock, flags);
1598 r1_bio = list_entry(head->prev, r1bio_t, retry_list);
1599 list_del(head->prev);
1601 spin_unlock_irqrestore(&conf->device_lock, flags);
1603 mddev = r1_bio->mddev;
1604 conf = mddev->private;
1605 if (test_bit(R1BIO_IsSync, &r1_bio->state)) {
1606 sync_request_write(mddev, r1_bio);
1611 /* we got a read error. Maybe the drive is bad. Maybe just
1612 * the block and we can fix it.
1613 * We freeze all other IO, and try reading the block from
1614 * other devices. When we find one, we re-write
1615 * and check it that fixes the read error.
1616 * This is all done synchronously while the array is
1619 if (mddev->ro == 0) {
1621 fix_read_error(conf, r1_bio->read_disk,
1624 unfreeze_array(conf);
1627 conf->mirrors[r1_bio->read_disk].rdev);
1629 bio = r1_bio->bios[r1_bio->read_disk];
1630 if ((disk=read_balance(conf, r1_bio)) == -1) {
1631 printk(KERN_ALERT "md/raid1:%s: %s: unrecoverable I/O"
1632 " read error for block %llu\n",
1634 bdevname(bio->bi_bdev,b),
1635 (unsigned long long)r1_bio->sector);
1636 raid_end_bio_io(r1_bio);
1638 const unsigned long do_sync = r1_bio->master_bio->bi_rw & REQ_SYNC;
1639 r1_bio->bios[r1_bio->read_disk] =
1640 mddev->ro ? IO_BLOCKED : NULL;
1641 r1_bio->read_disk = disk;
1643 bio = bio_clone_mddev(r1_bio->master_bio,
1645 r1_bio->bios[r1_bio->read_disk] = bio;
1646 rdev = conf->mirrors[disk].rdev;
1647 if (printk_ratelimit())
1648 printk(KERN_ERR "md/raid1:%s: redirecting sector %llu to"
1649 " other mirror: %s\n",
1651 (unsigned long long)r1_bio->sector,
1652 bdevname(rdev->bdev,b));
1653 bio->bi_sector = r1_bio->sector + rdev->data_offset;
1654 bio->bi_bdev = rdev->bdev;
1655 bio->bi_end_io = raid1_end_read_request;
1656 bio->bi_rw = READ | do_sync;
1657 bio->bi_private = r1_bio;
1659 generic_make_request(bio);
1665 unplug_slaves(mddev);
1669 static int init_resync(conf_t *conf)
1673 buffs = RESYNC_WINDOW / RESYNC_BLOCK_SIZE;
1674 BUG_ON(conf->r1buf_pool);
1675 conf->r1buf_pool = mempool_create(buffs, r1buf_pool_alloc, r1buf_pool_free,
1677 if (!conf->r1buf_pool)
1679 conf->next_resync = 0;
1684 * perform a "sync" on one "block"
1686 * We need to make sure that no normal I/O request - particularly write
1687 * requests - conflict with active sync requests.
1689 * This is achieved by tracking pending requests and a 'barrier' concept
1690 * that can be installed to exclude normal IO requests.
1693 static sector_t sync_request(mddev_t *mddev, sector_t sector_nr, int *skipped, int go_faster)
1695 conf_t *conf = mddev->private;
1698 sector_t max_sector, nr_sectors;
1702 int write_targets = 0, read_targets = 0;
1703 sector_t sync_blocks;
1704 int still_degraded = 0;
1706 if (!conf->r1buf_pool)
1707 if (init_resync(conf))
1710 max_sector = mddev->dev_sectors;
1711 if (sector_nr >= max_sector) {
1712 /* If we aborted, we need to abort the
1713 * sync on the 'current' bitmap chunk (there will
1714 * only be one in raid1 resync.
1715 * We can find the current addess in mddev->curr_resync
1717 if (mddev->curr_resync < max_sector) /* aborted */
1718 bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
1720 else /* completed sync */
1723 bitmap_close_sync(mddev->bitmap);
1728 if (mddev->bitmap == NULL &&
1729 mddev->recovery_cp == MaxSector &&
1730 !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
1731 conf->fullsync == 0) {
1733 return max_sector - sector_nr;
1735 /* before building a request, check if we can skip these blocks..
1736 * This call the bitmap_start_sync doesn't actually record anything
1738 if (!bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
1739 !conf->fullsync && !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
1740 /* We can skip this block, and probably several more */
1745 * If there is non-resync activity waiting for a turn,
1746 * and resync is going fast enough,
1747 * then let it though before starting on this new sync request.
1749 if (!go_faster && conf->nr_waiting)
1750 msleep_interruptible(1000);
1752 bitmap_cond_end_sync(mddev->bitmap, sector_nr);
1753 r1_bio = mempool_alloc(conf->r1buf_pool, GFP_NOIO);
1754 raise_barrier(conf);
1756 conf->next_resync = sector_nr;
1760 * If we get a correctably read error during resync or recovery,
1761 * we might want to read from a different device. So we
1762 * flag all drives that could conceivably be read from for READ,
1763 * and any others (which will be non-In_sync devices) for WRITE.
1764 * If a read fails, we try reading from something else for which READ
1768 r1_bio->mddev = mddev;
1769 r1_bio->sector = sector_nr;
1771 set_bit(R1BIO_IsSync, &r1_bio->state);
1773 for (i=0; i < conf->raid_disks; i++) {
1775 bio = r1_bio->bios[i];
1777 /* take from bio_init */
1778 bio->bi_next = NULL;
1779 bio->bi_flags &= ~(BIO_POOL_MASK-1);
1780 bio->bi_flags |= 1 << BIO_UPTODATE;
1781 bio->bi_comp_cpu = -1;
1785 bio->bi_phys_segments = 0;
1787 bio->bi_end_io = NULL;
1788 bio->bi_private = NULL;
1790 rdev = rcu_dereference(conf->mirrors[i].rdev);
1792 test_bit(Faulty, &rdev->flags)) {
1795 } else if (!test_bit(In_sync, &rdev->flags)) {
1797 bio->bi_end_io = end_sync_write;
1800 /* may need to read from here */
1802 bio->bi_end_io = end_sync_read;
1803 if (test_bit(WriteMostly, &rdev->flags)) {
1812 atomic_inc(&rdev->nr_pending);
1813 bio->bi_sector = sector_nr + rdev->data_offset;
1814 bio->bi_bdev = rdev->bdev;
1815 bio->bi_private = r1_bio;
1820 r1_bio->read_disk = disk;
1822 if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery) && read_targets > 0)
1823 /* extra read targets are also write targets */
1824 write_targets += read_targets-1;
1826 if (write_targets == 0 || read_targets == 0) {
1827 /* There is nowhere to write, so all non-sync
1828 * drives must be failed - so we are finished
1830 sector_t rv = max_sector - sector_nr;
1836 if (max_sector > mddev->resync_max)
1837 max_sector = mddev->resync_max; /* Don't do IO beyond here */
1842 int len = PAGE_SIZE;
1843 if (sector_nr + (len>>9) > max_sector)
1844 len = (max_sector - sector_nr) << 9;
1847 if (sync_blocks == 0) {
1848 if (!bitmap_start_sync(mddev->bitmap, sector_nr,
1849 &sync_blocks, still_degraded) &&
1851 !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery))
1853 BUG_ON(sync_blocks < (PAGE_SIZE>>9));
1854 if ((len >> 9) > sync_blocks)
1855 len = sync_blocks<<9;
1858 for (i=0 ; i < conf->raid_disks; i++) {
1859 bio = r1_bio->bios[i];
1860 if (bio->bi_end_io) {
1861 page = bio->bi_io_vec[bio->bi_vcnt].bv_page;
1862 if (bio_add_page(bio, page, len, 0) == 0) {
1864 bio->bi_io_vec[bio->bi_vcnt].bv_page = page;
1867 bio = r1_bio->bios[i];
1868 if (bio->bi_end_io==NULL)
1870 /* remove last page from this bio */
1872 bio->bi_size -= len;
1873 bio->bi_flags &= ~(1<< BIO_SEG_VALID);
1879 nr_sectors += len>>9;
1880 sector_nr += len>>9;
1881 sync_blocks -= (len>>9);
1882 } while (r1_bio->bios[disk]->bi_vcnt < RESYNC_PAGES);
1884 r1_bio->sectors = nr_sectors;
1886 /* For a user-requested sync, we read all readable devices and do a
1889 if (test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
1890 atomic_set(&r1_bio->remaining, read_targets);
1891 for (i=0; i<conf->raid_disks; i++) {
1892 bio = r1_bio->bios[i];
1893 if (bio->bi_end_io == end_sync_read) {
1894 md_sync_acct(bio->bi_bdev, nr_sectors);
1895 generic_make_request(bio);
1899 atomic_set(&r1_bio->remaining, 1);
1900 bio = r1_bio->bios[r1_bio->read_disk];
1901 md_sync_acct(bio->bi_bdev, nr_sectors);
1902 generic_make_request(bio);
1908 static sector_t raid1_size(mddev_t *mddev, sector_t sectors, int raid_disks)
1913 return mddev->dev_sectors;
1916 static conf_t *setup_conf(mddev_t *mddev)
1920 mirror_info_t *disk;
1924 conf = kzalloc(sizeof(conf_t), GFP_KERNEL);
1928 conf->mirrors = kzalloc(sizeof(struct mirror_info)*mddev->raid_disks,
1933 conf->tmppage = alloc_page(GFP_KERNEL);
1937 conf->poolinfo = kzalloc(sizeof(*conf->poolinfo), GFP_KERNEL);
1938 if (!conf->poolinfo)
1940 conf->poolinfo->raid_disks = mddev->raid_disks;
1941 conf->r1bio_pool = mempool_create(NR_RAID1_BIOS, r1bio_pool_alloc,
1944 if (!conf->r1bio_pool)
1947 conf->poolinfo->mddev = mddev;
1949 spin_lock_init(&conf->device_lock);
1950 list_for_each_entry(rdev, &mddev->disks, same_set) {
1951 int disk_idx = rdev->raid_disk;
1952 if (disk_idx >= mddev->raid_disks
1955 disk = conf->mirrors + disk_idx;
1959 disk->head_position = 0;
1961 conf->raid_disks = mddev->raid_disks;
1962 conf->mddev = mddev;
1963 INIT_LIST_HEAD(&conf->retry_list);
1965 spin_lock_init(&conf->resync_lock);
1966 init_waitqueue_head(&conf->wait_barrier);
1968 bio_list_init(&conf->pending_bio_list);
1970 conf->last_used = -1;
1971 for (i = 0; i < conf->raid_disks; i++) {
1973 disk = conf->mirrors + i;
1976 !test_bit(In_sync, &disk->rdev->flags)) {
1977 disk->head_position = 0;
1980 } else if (conf->last_used < 0)
1982 * The first working device is used as a
1983 * starting point to read balancing.
1985 conf->last_used = i;
1989 if (conf->last_used < 0) {
1990 printk(KERN_ERR "md/raid1:%s: no operational mirrors\n",
1995 conf->thread = md_register_thread(raid1d, mddev, NULL);
1996 if (!conf->thread) {
1998 "md/raid1:%s: couldn't allocate thread\n",
2007 if (conf->r1bio_pool)
2008 mempool_destroy(conf->r1bio_pool);
2009 kfree(conf->mirrors);
2010 safe_put_page(conf->tmppage);
2011 kfree(conf->poolinfo);
2014 return ERR_PTR(err);
2017 static int run(mddev_t *mddev)
2023 if (mddev->level != 1) {
2024 printk(KERN_ERR "md/raid1:%s: raid level not set to mirroring (%d)\n",
2025 mdname(mddev), mddev->level);
2028 if (mddev->reshape_position != MaxSector) {
2029 printk(KERN_ERR "md/raid1:%s: reshape_position set but not supported\n",
2034 * copy the already verified devices into our private RAID1
2035 * bookkeeping area. [whatever we allocate in run(),
2036 * should be freed in stop()]
2038 if (mddev->private == NULL)
2039 conf = setup_conf(mddev);
2041 conf = mddev->private;
2044 return PTR_ERR(conf);
2046 mddev->queue->queue_lock = &conf->device_lock;
2047 list_for_each_entry(rdev, &mddev->disks, same_set) {
2048 disk_stack_limits(mddev->gendisk, rdev->bdev,
2049 rdev->data_offset << 9);
2050 /* as we don't honour merge_bvec_fn, we must never risk
2051 * violating it, so limit ->max_segments to 1 lying within
2052 * a single page, as a one page request is never in violation.
2054 if (rdev->bdev->bd_disk->queue->merge_bvec_fn) {
2055 blk_queue_max_segments(mddev->queue, 1);
2056 blk_queue_segment_boundary(mddev->queue,
2057 PAGE_CACHE_SIZE - 1);
2061 mddev->degraded = 0;
2062 for (i=0; i < conf->raid_disks; i++)
2063 if (conf->mirrors[i].rdev == NULL ||
2064 !test_bit(In_sync, &conf->mirrors[i].rdev->flags) ||
2065 test_bit(Faulty, &conf->mirrors[i].rdev->flags))
2068 if (conf->raid_disks - mddev->degraded == 1)
2069 mddev->recovery_cp = MaxSector;
2071 if (mddev->recovery_cp != MaxSector)
2072 printk(KERN_NOTICE "md/raid1:%s: not clean"
2073 " -- starting background reconstruction\n",
2076 "md/raid1:%s: active with %d out of %d mirrors\n",
2077 mdname(mddev), mddev->raid_disks - mddev->degraded,
2081 * Ok, everything is just fine now
2083 mddev->thread = conf->thread;
2084 conf->thread = NULL;
2085 mddev->private = conf;
2087 md_set_array_sectors(mddev, raid1_size(mddev, 0, 0));
2089 mddev->queue->unplug_fn = raid1_unplug;
2090 mddev->queue->backing_dev_info.congested_fn = raid1_congested;
2091 mddev->queue->backing_dev_info.congested_data = mddev;
2092 md_integrity_register(mddev);
2096 static int stop(mddev_t *mddev)
2098 conf_t *conf = mddev->private;
2099 struct bitmap *bitmap = mddev->bitmap;
2101 /* wait for behind writes to complete */
2102 if (bitmap && atomic_read(&bitmap->behind_writes) > 0) {
2103 printk(KERN_INFO "md/raid1:%s: behind writes in progress - waiting to stop.\n",
2105 /* need to kick something here to make sure I/O goes? */
2106 wait_event(bitmap->behind_wait,
2107 atomic_read(&bitmap->behind_writes) == 0);
2110 raise_barrier(conf);
2111 lower_barrier(conf);
2113 md_unregister_thread(mddev->thread);
2114 mddev->thread = NULL;
2115 blk_sync_queue(mddev->queue); /* the unplug fn references 'conf'*/
2116 if (conf->r1bio_pool)
2117 mempool_destroy(conf->r1bio_pool);
2118 kfree(conf->mirrors);
2119 kfree(conf->poolinfo);
2121 mddev->private = NULL;
2125 static int raid1_resize(mddev_t *mddev, sector_t sectors)
2127 /* no resync is happening, and there is enough space
2128 * on all devices, so we can resize.
2129 * We need to make sure resync covers any new space.
2130 * If the array is shrinking we should possibly wait until
2131 * any io in the removed space completes, but it hardly seems
2134 md_set_array_sectors(mddev, raid1_size(mddev, sectors, 0));
2135 if (mddev->array_sectors > raid1_size(mddev, sectors, 0))
2137 set_capacity(mddev->gendisk, mddev->array_sectors);
2138 revalidate_disk(mddev->gendisk);
2139 if (sectors > mddev->dev_sectors &&
2140 mddev->recovery_cp == MaxSector) {
2141 mddev->recovery_cp = mddev->dev_sectors;
2142 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
2144 mddev->dev_sectors = sectors;
2145 mddev->resync_max_sectors = sectors;
2149 static int raid1_reshape(mddev_t *mddev)
2152 * 1/ resize the r1bio_pool
2153 * 2/ resize conf->mirrors
2155 * We allocate a new r1bio_pool if we can.
2156 * Then raise a device barrier and wait until all IO stops.
2157 * Then resize conf->mirrors and swap in the new r1bio pool.
2159 * At the same time, we "pack" the devices so that all the missing
2160 * devices have the higher raid_disk numbers.
2162 mempool_t *newpool, *oldpool;
2163 struct pool_info *newpoolinfo;
2164 mirror_info_t *newmirrors;
2165 conf_t *conf = mddev->private;
2166 int cnt, raid_disks;
2167 unsigned long flags;
2170 /* Cannot change chunk_size, layout, or level */
2171 if (mddev->chunk_sectors != mddev->new_chunk_sectors ||
2172 mddev->layout != mddev->new_layout ||
2173 mddev->level != mddev->new_level) {
2174 mddev->new_chunk_sectors = mddev->chunk_sectors;
2175 mddev->new_layout = mddev->layout;
2176 mddev->new_level = mddev->level;
2180 err = md_allow_write(mddev);
2184 raid_disks = mddev->raid_disks + mddev->delta_disks;
2186 if (raid_disks < conf->raid_disks) {
2188 for (d= 0; d < conf->raid_disks; d++)
2189 if (conf->mirrors[d].rdev)
2191 if (cnt > raid_disks)
2195 newpoolinfo = kmalloc(sizeof(*newpoolinfo), GFP_KERNEL);
2198 newpoolinfo->mddev = mddev;
2199 newpoolinfo->raid_disks = raid_disks;
2201 newpool = mempool_create(NR_RAID1_BIOS, r1bio_pool_alloc,
2202 r1bio_pool_free, newpoolinfo);
2207 newmirrors = kzalloc(sizeof(struct mirror_info) * raid_disks, GFP_KERNEL);
2210 mempool_destroy(newpool);
2214 raise_barrier(conf);
2216 /* ok, everything is stopped */
2217 oldpool = conf->r1bio_pool;
2218 conf->r1bio_pool = newpool;
2220 for (d = d2 = 0; d < conf->raid_disks; d++) {
2221 mdk_rdev_t *rdev = conf->mirrors[d].rdev;
2222 if (rdev && rdev->raid_disk != d2) {
2224 sprintf(nm, "rd%d", rdev->raid_disk);
2225 sysfs_remove_link(&mddev->kobj, nm);
2226 rdev->raid_disk = d2;
2227 sprintf(nm, "rd%d", rdev->raid_disk);
2228 sysfs_remove_link(&mddev->kobj, nm);
2229 if (sysfs_create_link(&mddev->kobj,
2232 "md/raid1:%s: cannot register "
2237 newmirrors[d2++].rdev = rdev;
2239 kfree(conf->mirrors);
2240 conf->mirrors = newmirrors;
2241 kfree(conf->poolinfo);
2242 conf->poolinfo = newpoolinfo;
2244 spin_lock_irqsave(&conf->device_lock, flags);
2245 mddev->degraded += (raid_disks - conf->raid_disks);
2246 spin_unlock_irqrestore(&conf->device_lock, flags);
2247 conf->raid_disks = mddev->raid_disks = raid_disks;
2248 mddev->delta_disks = 0;
2250 conf->last_used = 0; /* just make sure it is in-range */
2251 lower_barrier(conf);
2253 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
2254 md_wakeup_thread(mddev->thread);
2256 mempool_destroy(oldpool);
2260 static void raid1_quiesce(mddev_t *mddev, int state)
2262 conf_t *conf = mddev->private;
2265 case 2: /* wake for suspend */
2266 wake_up(&conf->wait_barrier);
2269 raise_barrier(conf);
2272 lower_barrier(conf);
2277 static void *raid1_takeover(mddev_t *mddev)
2279 /* raid1 can take over:
2280 * raid5 with 2 devices, any layout or chunk size
2282 if (mddev->level == 5 && mddev->raid_disks == 2) {
2284 mddev->new_level = 1;
2285 mddev->new_layout = 0;
2286 mddev->new_chunk_sectors = 0;
2287 conf = setup_conf(mddev);
2292 return ERR_PTR(-EINVAL);
2295 static struct mdk_personality raid1_personality =
2299 .owner = THIS_MODULE,
2300 .make_request = make_request,
2304 .error_handler = error,
2305 .hot_add_disk = raid1_add_disk,
2306 .hot_remove_disk= raid1_remove_disk,
2307 .spare_active = raid1_spare_active,
2308 .sync_request = sync_request,
2309 .resize = raid1_resize,
2311 .check_reshape = raid1_reshape,
2312 .quiesce = raid1_quiesce,
2313 .takeover = raid1_takeover,
2316 static int __init raid_init(void)
2318 return register_md_personality(&raid1_personality);
2321 static void raid_exit(void)
2323 unregister_md_personality(&raid1_personality);
2326 module_init(raid_init);
2327 module_exit(raid_exit);
2328 MODULE_LICENSE("GPL");
2329 MODULE_DESCRIPTION("RAID1 (mirroring) personality for MD");
2330 MODULE_ALIAS("md-personality-3"); /* RAID1 */
2331 MODULE_ALIAS("md-raid1");
2332 MODULE_ALIAS("md-level-1");