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/delay.h>
35 #include <linux/blkdev.h>
36 #include <linux/seq_file.h>
43 #define PRINTK(x...) printk(x)
49 * Number of guaranteed r1bios in case of extreme VM load:
51 #define NR_RAID1_BIOS 256
54 static void unplug_slaves(mddev_t *mddev);
56 static void allow_barrier(conf_t *conf);
57 static void lower_barrier(conf_t *conf);
59 static void * r1bio_pool_alloc(gfp_t gfp_flags, void *data)
61 struct pool_info *pi = data;
63 int size = offsetof(r1bio_t, bios[pi->raid_disks]);
65 /* allocate a r1bio with room for raid_disks entries in the bios array */
66 r1_bio = kzalloc(size, gfp_flags);
67 if (!r1_bio && pi->mddev)
68 unplug_slaves(pi->mddev);
73 static void r1bio_pool_free(void *r1_bio, void *data)
78 #define RESYNC_BLOCK_SIZE (64*1024)
79 //#define RESYNC_BLOCK_SIZE PAGE_SIZE
80 #define RESYNC_SECTORS (RESYNC_BLOCK_SIZE >> 9)
81 #define RESYNC_PAGES ((RESYNC_BLOCK_SIZE + PAGE_SIZE-1) / PAGE_SIZE)
82 #define RESYNC_WINDOW (2048*1024)
84 static void * r1buf_pool_alloc(gfp_t gfp_flags, void *data)
86 struct pool_info *pi = data;
92 r1_bio = r1bio_pool_alloc(gfp_flags, pi);
94 unplug_slaves(pi->mddev);
99 * Allocate bios : 1 for reading, n-1 for writing
101 for (j = pi->raid_disks ; j-- ; ) {
102 bio = bio_alloc(gfp_flags, RESYNC_PAGES);
105 r1_bio->bios[j] = bio;
108 * Allocate RESYNC_PAGES data pages and attach them to
110 * If this is a user-requested check/repair, allocate
111 * RESYNC_PAGES for each bio.
113 if (test_bit(MD_RECOVERY_REQUESTED, &pi->mddev->recovery))
118 bio = r1_bio->bios[j];
119 for (i = 0; i < RESYNC_PAGES; i++) {
120 page = alloc_page(gfp_flags);
124 bio->bi_io_vec[i].bv_page = page;
128 /* If not user-requests, copy the page pointers to all bios */
129 if (!test_bit(MD_RECOVERY_REQUESTED, &pi->mddev->recovery)) {
130 for (i=0; i<RESYNC_PAGES ; i++)
131 for (j=1; j<pi->raid_disks; j++)
132 r1_bio->bios[j]->bi_io_vec[i].bv_page =
133 r1_bio->bios[0]->bi_io_vec[i].bv_page;
136 r1_bio->master_bio = NULL;
141 for (j=0 ; j < pi->raid_disks; j++)
142 for (i=0; i < r1_bio->bios[j]->bi_vcnt ; i++)
143 put_page(r1_bio->bios[j]->bi_io_vec[i].bv_page);
146 while ( ++j < pi->raid_disks )
147 bio_put(r1_bio->bios[j]);
148 r1bio_pool_free(r1_bio, data);
152 static void r1buf_pool_free(void *__r1_bio, void *data)
154 struct pool_info *pi = data;
156 r1bio_t *r1bio = __r1_bio;
158 for (i = 0; i < RESYNC_PAGES; i++)
159 for (j = pi->raid_disks; j-- ;) {
161 r1bio->bios[j]->bi_io_vec[i].bv_page !=
162 r1bio->bios[0]->bi_io_vec[i].bv_page)
163 safe_put_page(r1bio->bios[j]->bi_io_vec[i].bv_page);
165 for (i=0 ; i < pi->raid_disks; i++)
166 bio_put(r1bio->bios[i]);
168 r1bio_pool_free(r1bio, data);
171 static void put_all_bios(conf_t *conf, r1bio_t *r1_bio)
175 for (i = 0; i < conf->raid_disks; i++) {
176 struct bio **bio = r1_bio->bios + i;
177 if (*bio && *bio != IO_BLOCKED)
183 static void free_r1bio(r1bio_t *r1_bio)
185 conf_t *conf = r1_bio->mddev->private;
188 * Wake up any possible resync thread that waits for the device
193 put_all_bios(conf, r1_bio);
194 mempool_free(r1_bio, conf->r1bio_pool);
197 static void put_buf(r1bio_t *r1_bio)
199 conf_t *conf = r1_bio->mddev->private;
202 for (i=0; i<conf->raid_disks; i++) {
203 struct bio *bio = r1_bio->bios[i];
205 rdev_dec_pending(conf->mirrors[i].rdev, r1_bio->mddev);
208 mempool_free(r1_bio, conf->r1buf_pool);
213 static void reschedule_retry(r1bio_t *r1_bio)
216 mddev_t *mddev = r1_bio->mddev;
217 conf_t *conf = mddev->private;
219 spin_lock_irqsave(&conf->device_lock, flags);
220 list_add(&r1_bio->retry_list, &conf->retry_list);
222 spin_unlock_irqrestore(&conf->device_lock, flags);
224 wake_up(&conf->wait_barrier);
225 md_wakeup_thread(mddev->thread);
229 * raid_end_bio_io() is called when we have finished servicing a mirrored
230 * operation and are ready to return a success/failure code to the buffer
233 static void raid_end_bio_io(r1bio_t *r1_bio)
235 struct bio *bio = r1_bio->master_bio;
237 /* if nobody has done the final endio yet, do it now */
238 if (!test_and_set_bit(R1BIO_Returned, &r1_bio->state)) {
239 PRINTK(KERN_DEBUG "raid1: sync end %s on sectors %llu-%llu\n",
240 (bio_data_dir(bio) == WRITE) ? "write" : "read",
241 (unsigned long long) bio->bi_sector,
242 (unsigned long long) bio->bi_sector +
243 (bio->bi_size >> 9) - 1);
246 test_bit(R1BIO_Uptodate, &r1_bio->state) ? 0 : -EIO);
252 * Update disk head position estimator based on IRQ completion info.
254 static inline void update_head_pos(int disk, r1bio_t *r1_bio)
256 conf_t *conf = r1_bio->mddev->private;
258 conf->mirrors[disk].head_position =
259 r1_bio->sector + (r1_bio->sectors);
262 static void raid1_end_read_request(struct bio *bio, int error)
264 int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
265 r1bio_t * r1_bio = (r1bio_t *)(bio->bi_private);
267 conf_t *conf = r1_bio->mddev->private;
269 mirror = r1_bio->read_disk;
271 * this branch is our 'one mirror IO has finished' event handler:
273 update_head_pos(mirror, r1_bio);
276 set_bit(R1BIO_Uptodate, &r1_bio->state);
278 /* If all other devices have failed, we want to return
279 * the error upwards rather than fail the last device.
280 * Here we redefine "uptodate" to mean "Don't want to retry"
283 spin_lock_irqsave(&conf->device_lock, flags);
284 if (r1_bio->mddev->degraded == conf->raid_disks ||
285 (r1_bio->mddev->degraded == conf->raid_disks-1 &&
286 !test_bit(Faulty, &conf->mirrors[mirror].rdev->flags)))
288 spin_unlock_irqrestore(&conf->device_lock, flags);
292 raid_end_bio_io(r1_bio);
297 char b[BDEVNAME_SIZE];
298 if (printk_ratelimit())
299 printk(KERN_ERR "raid1: %s: rescheduling sector %llu\n",
300 bdevname(conf->mirrors[mirror].rdev->bdev,b), (unsigned long long)r1_bio->sector);
301 reschedule_retry(r1_bio);
304 rdev_dec_pending(conf->mirrors[mirror].rdev, conf->mddev);
307 static void raid1_end_write_request(struct bio *bio, int error)
309 int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
310 r1bio_t * r1_bio = (r1bio_t *)(bio->bi_private);
311 int mirror, behind = test_bit(R1BIO_BehindIO, &r1_bio->state);
312 conf_t *conf = r1_bio->mddev->private;
313 struct bio *to_put = NULL;
316 for (mirror = 0; mirror < conf->raid_disks; mirror++)
317 if (r1_bio->bios[mirror] == bio)
320 if (error == -EOPNOTSUPP && test_bit(R1BIO_Barrier, &r1_bio->state)) {
321 set_bit(BarriersNotsupp, &conf->mirrors[mirror].rdev->flags);
322 set_bit(R1BIO_BarrierRetry, &r1_bio->state);
323 r1_bio->mddev->barriers_work = 0;
324 /* Don't rdev_dec_pending in this branch - keep it for the retry */
327 * this branch is our 'one mirror IO has finished' event handler:
329 r1_bio->bios[mirror] = NULL;
332 md_error(r1_bio->mddev, conf->mirrors[mirror].rdev);
333 /* an I/O failed, we can't clear the bitmap */
334 set_bit(R1BIO_Degraded, &r1_bio->state);
337 * Set R1BIO_Uptodate in our master bio, so that
338 * we will return a good error code for to the higher
339 * levels even if IO on some other mirrored buffer fails.
341 * The 'master' represents the composite IO operation to
342 * user-side. So if something waits for IO, then it will
343 * wait for the 'master' bio.
345 set_bit(R1BIO_Uptodate, &r1_bio->state);
347 update_head_pos(mirror, r1_bio);
350 if (test_bit(WriteMostly, &conf->mirrors[mirror].rdev->flags))
351 atomic_dec(&r1_bio->behind_remaining);
353 /* In behind mode, we ACK the master bio once the I/O has safely
354 * reached all non-writemostly disks. Setting the Returned bit
355 * ensures that this gets done only once -- we don't ever want to
356 * return -EIO here, instead we'll wait */
358 if (atomic_read(&r1_bio->behind_remaining) >= (atomic_read(&r1_bio->remaining)-1) &&
359 test_bit(R1BIO_Uptodate, &r1_bio->state)) {
360 /* Maybe we can return now */
361 if (!test_and_set_bit(R1BIO_Returned, &r1_bio->state)) {
362 struct bio *mbio = r1_bio->master_bio;
363 PRINTK(KERN_DEBUG "raid1: behind end write sectors %llu-%llu\n",
364 (unsigned long long) mbio->bi_sector,
365 (unsigned long long) mbio->bi_sector +
366 (mbio->bi_size >> 9) - 1);
371 rdev_dec_pending(conf->mirrors[mirror].rdev, conf->mddev);
375 * Let's see if all mirrored write operations have finished
378 if (atomic_dec_and_test(&r1_bio->remaining)) {
379 if (test_bit(R1BIO_BarrierRetry, &r1_bio->state))
380 reschedule_retry(r1_bio);
382 /* it really is the end of this request */
383 if (test_bit(R1BIO_BehindIO, &r1_bio->state)) {
384 /* free extra copy of the data pages */
385 int i = bio->bi_vcnt;
387 safe_put_page(bio->bi_io_vec[i].bv_page);
389 /* clear the bitmap if all writes complete successfully */
390 bitmap_endwrite(r1_bio->mddev->bitmap, r1_bio->sector,
392 !test_bit(R1BIO_Degraded, &r1_bio->state),
394 md_write_end(r1_bio->mddev);
395 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->private;
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 struct request_queue *r_queue = bdev_get_queue(rdev->bdev);
553 atomic_inc(&rdev->nr_pending);
558 rdev_dec_pending(rdev, mddev);
565 static void raid1_unplug(struct request_queue *q)
567 mddev_t *mddev = q->queuedata;
569 unplug_slaves(mddev);
570 md_wakeup_thread(mddev->thread);
573 static int raid1_congested(void *data, int bits)
575 mddev_t *mddev = data;
576 conf_t *conf = mddev->private;
579 if (mddev_congested(mddev, bits))
583 for (i = 0; i < mddev->raid_disks; i++) {
584 mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[i].rdev);
585 if (rdev && !test_bit(Faulty, &rdev->flags)) {
586 struct request_queue *q = bdev_get_queue(rdev->bdev);
588 /* Note the '|| 1' - when read_balance prefers
589 * non-congested targets, it can be removed
591 if ((bits & (1<<BDI_async_congested)) || 1)
592 ret |= bdi_congested(&q->backing_dev_info, bits);
594 ret &= bdi_congested(&q->backing_dev_info, bits);
602 static int flush_pending_writes(conf_t *conf)
604 /* Any writes that have been queued but are awaiting
605 * bitmap updates get flushed here.
606 * We return 1 if any requests were actually submitted.
610 spin_lock_irq(&conf->device_lock);
612 if (conf->pending_bio_list.head) {
614 bio = bio_list_get(&conf->pending_bio_list);
615 blk_remove_plug(conf->mddev->queue);
616 spin_unlock_irq(&conf->device_lock);
617 /* flush any pending bitmap writes to
618 * disk before proceeding w/ I/O */
619 bitmap_unplug(conf->mddev->bitmap);
621 while (bio) { /* submit pending writes */
622 struct bio *next = bio->bi_next;
624 generic_make_request(bio);
629 spin_unlock_irq(&conf->device_lock);
634 * Sometimes we need to suspend IO while we do something else,
635 * either some resync/recovery, or reconfigure the array.
636 * To do this we raise a 'barrier'.
637 * The 'barrier' is a counter that can be raised multiple times
638 * to count how many activities are happening which preclude
640 * We can only raise the barrier if there is no pending IO.
641 * i.e. if nr_pending == 0.
642 * We choose only to raise the barrier if no-one is waiting for the
643 * barrier to go down. This means that as soon as an IO request
644 * is ready, no other operations which require a barrier will start
645 * until the IO request has had a chance.
647 * So: regular IO calls 'wait_barrier'. When that returns there
648 * is no backgroup IO happening, It must arrange to call
649 * allow_barrier when it has finished its IO.
650 * backgroup IO calls must call raise_barrier. Once that returns
651 * there is no normal IO happeing. It must arrange to call
652 * lower_barrier when the particular background IO completes.
654 #define RESYNC_DEPTH 32
656 static void raise_barrier(conf_t *conf)
658 spin_lock_irq(&conf->resync_lock);
660 /* Wait until no block IO is waiting */
661 wait_event_lock_irq(conf->wait_barrier, !conf->nr_waiting,
663 raid1_unplug(conf->mddev->queue));
665 /* block any new IO from starting */
668 /* No wait for all pending IO to complete */
669 wait_event_lock_irq(conf->wait_barrier,
670 !conf->nr_pending && conf->barrier < RESYNC_DEPTH,
672 raid1_unplug(conf->mddev->queue));
674 spin_unlock_irq(&conf->resync_lock);
677 static void lower_barrier(conf_t *conf)
680 BUG_ON(conf->barrier <= 0);
681 spin_lock_irqsave(&conf->resync_lock, flags);
683 spin_unlock_irqrestore(&conf->resync_lock, flags);
684 wake_up(&conf->wait_barrier);
687 static void wait_barrier(conf_t *conf)
689 spin_lock_irq(&conf->resync_lock);
692 wait_event_lock_irq(conf->wait_barrier, !conf->barrier,
694 raid1_unplug(conf->mddev->queue));
698 spin_unlock_irq(&conf->resync_lock);
701 static void allow_barrier(conf_t *conf)
704 spin_lock_irqsave(&conf->resync_lock, flags);
706 spin_unlock_irqrestore(&conf->resync_lock, flags);
707 wake_up(&conf->wait_barrier);
710 static void freeze_array(conf_t *conf)
712 /* stop syncio and normal IO and wait for everything to
714 * We increment barrier and nr_waiting, and then
715 * wait until nr_pending match nr_queued+1
716 * This is called in the context of one normal IO request
717 * that has failed. Thus any sync request that might be pending
718 * will be blocked by nr_pending, and we need to wait for
719 * pending IO requests to complete or be queued for re-try.
720 * Thus the number queued (nr_queued) plus this request (1)
721 * must match the number of pending IOs (nr_pending) before
724 spin_lock_irq(&conf->resync_lock);
727 wait_event_lock_irq(conf->wait_barrier,
728 conf->nr_pending == conf->nr_queued+1,
730 ({ flush_pending_writes(conf);
731 raid1_unplug(conf->mddev->queue); }));
732 spin_unlock_irq(&conf->resync_lock);
734 static void unfreeze_array(conf_t *conf)
736 /* reverse the effect of the freeze */
737 spin_lock_irq(&conf->resync_lock);
740 wake_up(&conf->wait_barrier);
741 spin_unlock_irq(&conf->resync_lock);
745 /* duplicate the data pages for behind I/O */
746 static struct page **alloc_behind_pages(struct bio *bio)
749 struct bio_vec *bvec;
750 struct page **pages = kzalloc(bio->bi_vcnt * sizeof(struct page *),
752 if (unlikely(!pages))
755 bio_for_each_segment(bvec, bio, i) {
756 pages[i] = alloc_page(GFP_NOIO);
757 if (unlikely(!pages[i]))
759 memcpy(kmap(pages[i]) + bvec->bv_offset,
760 kmap(bvec->bv_page) + bvec->bv_offset, bvec->bv_len);
762 kunmap(bvec->bv_page);
769 for (i = 0; i < bio->bi_vcnt && pages[i]; i++)
772 PRINTK("%dB behind alloc failed, doing sync I/O\n", bio->bi_size);
776 static int make_request(struct request_queue *q, struct bio * bio)
778 mddev_t *mddev = q->queuedata;
779 conf_t *conf = mddev->private;
780 mirror_info_t *mirror;
782 struct bio *read_bio;
783 int i, targets = 0, disks;
784 struct bitmap *bitmap;
787 struct page **behind_pages = NULL;
788 const int rw = bio_data_dir(bio);
789 const bool do_sync = bio_rw_flagged(bio, BIO_RW_SYNCIO);
792 mdk_rdev_t *blocked_rdev;
795 * Register the new request and wait if the reconstruction
796 * thread has put up a bar for new requests.
797 * Continue immediately if no resync is active currently.
798 * We test barriers_work *after* md_write_start as md_write_start
799 * may cause the first superblock write, and that will check out
803 md_write_start(mddev, bio); /* wait on superblock update early */
805 if (bio_data_dir(bio) == WRITE &&
806 bio->bi_sector + bio->bi_size/512 > mddev->suspend_lo &&
807 bio->bi_sector < mddev->suspend_hi) {
808 /* As the suspend_* range is controlled by
809 * userspace, we want an interruptible
814 flush_signals(current);
815 prepare_to_wait(&conf->wait_barrier,
816 &w, TASK_INTERRUPTIBLE);
817 if (bio->bi_sector + bio->bi_size/512 <= mddev->suspend_lo ||
818 bio->bi_sector >= mddev->suspend_hi)
822 finish_wait(&conf->wait_barrier, &w);
824 if (unlikely(!mddev->barriers_work &&
825 bio_rw_flagged(bio, BIO_RW_BARRIER))) {
828 bio_endio(bio, -EOPNOTSUPP);
834 bitmap = mddev->bitmap;
836 cpu = part_stat_lock();
837 part_stat_inc(cpu, &mddev->gendisk->part0, ios[rw]);
838 part_stat_add(cpu, &mddev->gendisk->part0, sectors[rw],
843 * make_request() can abort the operation when READA is being
844 * used and no empty request is available.
847 r1_bio = mempool_alloc(conf->r1bio_pool, GFP_NOIO);
849 r1_bio->master_bio = bio;
850 r1_bio->sectors = bio->bi_size >> 9;
852 r1_bio->mddev = mddev;
853 r1_bio->sector = bio->bi_sector;
857 * read balancing logic:
859 int rdisk = read_balance(conf, r1_bio);
862 /* couldn't find anywhere to read from */
863 raid_end_bio_io(r1_bio);
866 mirror = conf->mirrors + rdisk;
868 r1_bio->read_disk = rdisk;
870 read_bio = bio_clone(bio, GFP_NOIO);
872 r1_bio->bios[rdisk] = read_bio;
874 read_bio->bi_sector = r1_bio->sector + mirror->rdev->data_offset;
875 read_bio->bi_bdev = mirror->rdev->bdev;
876 read_bio->bi_end_io = raid1_end_read_request;
877 read_bio->bi_rw = READ | (do_sync << BIO_RW_SYNCIO);
878 read_bio->bi_private = r1_bio;
880 generic_make_request(read_bio);
887 /* first select target devices under spinlock and
888 * inc refcount on their rdev. Record them by setting
891 disks = conf->raid_disks;
893 { static int first=1;
894 if (first) printk("First Write sector %llu disks %d\n",
895 (unsigned long long)r1_bio->sector, disks);
902 for (i = 0; i < disks; i++) {
903 mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[i].rdev);
904 if (rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
905 atomic_inc(&rdev->nr_pending);
909 if (rdev && !test_bit(Faulty, &rdev->flags)) {
910 atomic_inc(&rdev->nr_pending);
911 if (test_bit(Faulty, &rdev->flags)) {
912 rdev_dec_pending(rdev, mddev);
913 r1_bio->bios[i] = NULL;
915 r1_bio->bios[i] = bio;
918 r1_bio->bios[i] = NULL;
922 if (unlikely(blocked_rdev)) {
923 /* Wait for this device to become unblocked */
926 for (j = 0; j < i; j++)
928 rdev_dec_pending(conf->mirrors[j].rdev, mddev);
931 md_wait_for_blocked_rdev(blocked_rdev, mddev);
936 BUG_ON(targets == 0); /* we never fail the last device */
938 if (targets < conf->raid_disks) {
939 /* array is degraded, we will not clear the bitmap
940 * on I/O completion (see raid1_end_write_request) */
941 set_bit(R1BIO_Degraded, &r1_bio->state);
944 /* do behind I/O ? */
946 atomic_read(&bitmap->behind_writes) < bitmap->max_write_behind &&
947 (behind_pages = alloc_behind_pages(bio)) != NULL)
948 set_bit(R1BIO_BehindIO, &r1_bio->state);
950 atomic_set(&r1_bio->remaining, 0);
951 atomic_set(&r1_bio->behind_remaining, 0);
953 do_barriers = bio_rw_flagged(bio, BIO_RW_BARRIER);
955 set_bit(R1BIO_Barrier, &r1_bio->state);
958 for (i = 0; i < disks; i++) {
960 if (!r1_bio->bios[i])
963 mbio = bio_clone(bio, GFP_NOIO);
964 r1_bio->bios[i] = mbio;
966 mbio->bi_sector = r1_bio->sector + conf->mirrors[i].rdev->data_offset;
967 mbio->bi_bdev = conf->mirrors[i].rdev->bdev;
968 mbio->bi_end_io = raid1_end_write_request;
969 mbio->bi_rw = WRITE | (do_barriers << BIO_RW_BARRIER) |
970 (do_sync << BIO_RW_SYNCIO);
971 mbio->bi_private = r1_bio;
974 struct bio_vec *bvec;
977 /* Yes, I really want the '__' version so that
978 * we clear any unused pointer in the io_vec, rather
979 * than leave them unchanged. This is important
980 * because when we come to free the pages, we won't
981 * know the originial bi_idx, so we just free
984 __bio_for_each_segment(bvec, mbio, j, 0)
985 bvec->bv_page = behind_pages[j];
986 if (test_bit(WriteMostly, &conf->mirrors[i].rdev->flags))
987 atomic_inc(&r1_bio->behind_remaining);
990 atomic_inc(&r1_bio->remaining);
992 bio_list_add(&bl, mbio);
994 kfree(behind_pages); /* the behind pages are attached to the bios now */
996 bitmap_startwrite(bitmap, bio->bi_sector, r1_bio->sectors,
997 test_bit(R1BIO_BehindIO, &r1_bio->state));
998 spin_lock_irqsave(&conf->device_lock, flags);
999 bio_list_merge(&conf->pending_bio_list, &bl);
1002 blk_plug_device(mddev->queue);
1003 spin_unlock_irqrestore(&conf->device_lock, flags);
1005 /* In case raid1d snuck into freeze_array */
1006 wake_up(&conf->wait_barrier);
1009 md_wakeup_thread(mddev->thread);
1011 while ((bio = bio_list_pop(&bl)) != NULL)
1012 generic_make_request(bio);
1018 static void status(struct seq_file *seq, mddev_t *mddev)
1020 conf_t *conf = mddev->private;
1023 seq_printf(seq, " [%d/%d] [", conf->raid_disks,
1024 conf->raid_disks - mddev->degraded);
1026 for (i = 0; i < conf->raid_disks; i++) {
1027 mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[i].rdev);
1028 seq_printf(seq, "%s",
1029 rdev && test_bit(In_sync, &rdev->flags) ? "U" : "_");
1032 seq_printf(seq, "]");
1036 static void error(mddev_t *mddev, mdk_rdev_t *rdev)
1038 char b[BDEVNAME_SIZE];
1039 conf_t *conf = mddev->private;
1042 * If it is not operational, then we have already marked it as dead
1043 * else if it is the last working disks, ignore the error, let the
1044 * next level up know.
1045 * else mark the drive as failed
1047 if (test_bit(In_sync, &rdev->flags)
1048 && (conf->raid_disks - mddev->degraded) == 1) {
1050 * Don't fail the drive, act as though we were just a
1051 * normal single drive.
1052 * However don't try a recovery from this drive as
1053 * it is very likely to fail.
1055 mddev->recovery_disabled = 1;
1058 if (test_and_clear_bit(In_sync, &rdev->flags)) {
1059 unsigned long flags;
1060 spin_lock_irqsave(&conf->device_lock, flags);
1062 set_bit(Faulty, &rdev->flags);
1063 spin_unlock_irqrestore(&conf->device_lock, flags);
1065 * if recovery is running, make sure it aborts.
1067 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1069 set_bit(Faulty, &rdev->flags);
1070 set_bit(MD_CHANGE_DEVS, &mddev->flags);
1071 printk(KERN_ALERT "raid1: Disk failure on %s, disabling device.\n"
1072 "raid1: Operation continuing on %d devices.\n",
1073 bdevname(rdev->bdev,b), conf->raid_disks - mddev->degraded);
1076 static void print_conf(conf_t *conf)
1080 printk("RAID1 conf printout:\n");
1082 printk("(!conf)\n");
1085 printk(" --- wd:%d rd:%d\n", conf->raid_disks - conf->mddev->degraded,
1089 for (i = 0; i < conf->raid_disks; i++) {
1090 char b[BDEVNAME_SIZE];
1091 mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[i].rdev);
1093 printk(" disk %d, wo:%d, o:%d, dev:%s\n",
1094 i, !test_bit(In_sync, &rdev->flags),
1095 !test_bit(Faulty, &rdev->flags),
1096 bdevname(rdev->bdev,b));
1101 static void close_sync(conf_t *conf)
1104 allow_barrier(conf);
1106 mempool_destroy(conf->r1buf_pool);
1107 conf->r1buf_pool = NULL;
1110 static int raid1_spare_active(mddev_t *mddev)
1113 conf_t *conf = mddev->private;
1116 * Find all failed disks within the RAID1 configuration
1117 * and mark them readable.
1118 * Called under mddev lock, so rcu protection not needed.
1120 for (i = 0; i < conf->raid_disks; i++) {
1121 mdk_rdev_t *rdev = conf->mirrors[i].rdev;
1123 && !test_bit(Faulty, &rdev->flags)
1124 && !test_and_set_bit(In_sync, &rdev->flags)) {
1125 unsigned long flags;
1126 spin_lock_irqsave(&conf->device_lock, flags);
1128 spin_unlock_irqrestore(&conf->device_lock, flags);
1137 static int raid1_add_disk(mddev_t *mddev, mdk_rdev_t *rdev)
1139 conf_t *conf = mddev->private;
1144 int last = mddev->raid_disks - 1;
1146 if (rdev->raid_disk >= 0)
1147 first = last = rdev->raid_disk;
1149 for (mirror = first; mirror <= last; mirror++)
1150 if ( !(p=conf->mirrors+mirror)->rdev) {
1152 disk_stack_limits(mddev->gendisk, rdev->bdev,
1153 rdev->data_offset << 9);
1154 /* as we don't honour merge_bvec_fn, we must never risk
1155 * violating it, so limit ->max_sector to one PAGE, as
1156 * a one page request is never in violation.
1158 if (rdev->bdev->bd_disk->queue->merge_bvec_fn &&
1159 queue_max_sectors(mddev->queue) > (PAGE_SIZE>>9))
1160 blk_queue_max_sectors(mddev->queue, PAGE_SIZE>>9);
1162 p->head_position = 0;
1163 rdev->raid_disk = mirror;
1165 /* As all devices are equivalent, we don't need a full recovery
1166 * if this was recently any drive of the array
1168 if (rdev->saved_raid_disk < 0)
1170 rcu_assign_pointer(p->rdev, rdev);
1173 md_integrity_add_rdev(rdev, mddev);
1178 static int raid1_remove_disk(mddev_t *mddev, int number)
1180 conf_t *conf = mddev->private;
1183 mirror_info_t *p = conf->mirrors+ number;
1188 if (test_bit(In_sync, &rdev->flags) ||
1189 atomic_read(&rdev->nr_pending)) {
1193 /* Only remove non-faulty devices is recovery
1196 if (!test_bit(Faulty, &rdev->flags) &&
1197 mddev->degraded < conf->raid_disks) {
1203 if (atomic_read(&rdev->nr_pending)) {
1204 /* lost the race, try later */
1209 md_integrity_register(mddev);
1218 static void end_sync_read(struct bio *bio, int error)
1220 r1bio_t * r1_bio = (r1bio_t *)(bio->bi_private);
1223 for (i=r1_bio->mddev->raid_disks; i--; )
1224 if (r1_bio->bios[i] == bio)
1227 update_head_pos(i, r1_bio);
1229 * we have read a block, now it needs to be re-written,
1230 * or re-read if the read failed.
1231 * We don't do much here, just schedule handling by raid1d
1233 if (test_bit(BIO_UPTODATE, &bio->bi_flags))
1234 set_bit(R1BIO_Uptodate, &r1_bio->state);
1236 if (atomic_dec_and_test(&r1_bio->remaining))
1237 reschedule_retry(r1_bio);
1240 static void end_sync_write(struct bio *bio, int error)
1242 int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
1243 r1bio_t * r1_bio = (r1bio_t *)(bio->bi_private);
1244 mddev_t *mddev = r1_bio->mddev;
1245 conf_t *conf = mddev->private;
1249 for (i = 0; i < conf->raid_disks; i++)
1250 if (r1_bio->bios[i] == bio) {
1255 int sync_blocks = 0;
1256 sector_t s = r1_bio->sector;
1257 long sectors_to_go = r1_bio->sectors;
1258 /* make sure these bits doesn't get cleared. */
1260 bitmap_end_sync(mddev->bitmap, s,
1263 sectors_to_go -= sync_blocks;
1264 } while (sectors_to_go > 0);
1265 md_error(mddev, conf->mirrors[mirror].rdev);
1268 update_head_pos(mirror, r1_bio);
1270 if (atomic_dec_and_test(&r1_bio->remaining)) {
1271 sector_t s = r1_bio->sectors;
1273 md_done_sync(mddev, s, uptodate);
1277 static void sync_request_write(mddev_t *mddev, r1bio_t *r1_bio)
1279 conf_t *conf = mddev->private;
1281 int disks = conf->raid_disks;
1282 struct bio *bio, *wbio;
1284 bio = r1_bio->bios[r1_bio->read_disk];
1287 if (test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
1288 /* We have read all readable devices. If we haven't
1289 * got the block, then there is no hope left.
1290 * If we have, then we want to do a comparison
1291 * and skip the write if everything is the same.
1292 * If any blocks failed to read, then we need to
1293 * attempt an over-write
1296 if (!test_bit(R1BIO_Uptodate, &r1_bio->state)) {
1297 for (i=0; i<mddev->raid_disks; i++)
1298 if (r1_bio->bios[i]->bi_end_io == end_sync_read)
1299 md_error(mddev, conf->mirrors[i].rdev);
1301 md_done_sync(mddev, r1_bio->sectors, 1);
1305 for (primary=0; primary<mddev->raid_disks; primary++)
1306 if (r1_bio->bios[primary]->bi_end_io == end_sync_read &&
1307 test_bit(BIO_UPTODATE, &r1_bio->bios[primary]->bi_flags)) {
1308 r1_bio->bios[primary]->bi_end_io = NULL;
1309 rdev_dec_pending(conf->mirrors[primary].rdev, mddev);
1312 r1_bio->read_disk = primary;
1313 for (i=0; i<mddev->raid_disks; i++)
1314 if (r1_bio->bios[i]->bi_end_io == end_sync_read) {
1316 int vcnt = r1_bio->sectors >> (PAGE_SHIFT- 9);
1317 struct bio *pbio = r1_bio->bios[primary];
1318 struct bio *sbio = r1_bio->bios[i];
1320 if (test_bit(BIO_UPTODATE, &sbio->bi_flags)) {
1321 for (j = vcnt; j-- ; ) {
1323 p = pbio->bi_io_vec[j].bv_page;
1324 s = sbio->bi_io_vec[j].bv_page;
1325 if (memcmp(page_address(p),
1333 mddev->resync_mismatches += r1_bio->sectors;
1334 if (j < 0 || (test_bit(MD_RECOVERY_CHECK, &mddev->recovery)
1335 && test_bit(BIO_UPTODATE, &sbio->bi_flags))) {
1336 sbio->bi_end_io = NULL;
1337 rdev_dec_pending(conf->mirrors[i].rdev, mddev);
1339 /* fixup the bio for reuse */
1341 sbio->bi_vcnt = vcnt;
1342 sbio->bi_size = r1_bio->sectors << 9;
1344 sbio->bi_phys_segments = 0;
1345 sbio->bi_flags &= ~(BIO_POOL_MASK - 1);
1346 sbio->bi_flags |= 1 << BIO_UPTODATE;
1347 sbio->bi_next = NULL;
1348 sbio->bi_sector = r1_bio->sector +
1349 conf->mirrors[i].rdev->data_offset;
1350 sbio->bi_bdev = conf->mirrors[i].rdev->bdev;
1351 size = sbio->bi_size;
1352 for (j = 0; j < vcnt ; j++) {
1354 bi = &sbio->bi_io_vec[j];
1356 if (size > PAGE_SIZE)
1357 bi->bv_len = PAGE_SIZE;
1361 memcpy(page_address(bi->bv_page),
1362 page_address(pbio->bi_io_vec[j].bv_page),
1369 if (!test_bit(R1BIO_Uptodate, &r1_bio->state)) {
1370 /* ouch - failed to read all of that.
1371 * Try some synchronous reads of other devices to get
1372 * good data, much like with normal read errors. Only
1373 * read into the pages we already have so we don't
1374 * need to re-issue the read request.
1375 * We don't need to freeze the array, because being in an
1376 * active sync request, there is no normal IO, and
1377 * no overlapping syncs.
1379 sector_t sect = r1_bio->sector;
1380 int sectors = r1_bio->sectors;
1385 int d = r1_bio->read_disk;
1389 if (s > (PAGE_SIZE>>9))
1392 if (r1_bio->bios[d]->bi_end_io == end_sync_read) {
1393 /* No rcu protection needed here devices
1394 * can only be removed when no resync is
1395 * active, and resync is currently active
1397 rdev = conf->mirrors[d].rdev;
1398 if (sync_page_io(rdev->bdev,
1399 sect + rdev->data_offset,
1401 bio->bi_io_vec[idx].bv_page,
1408 if (d == conf->raid_disks)
1410 } while (!success && d != r1_bio->read_disk);
1414 /* write it back and re-read */
1415 set_bit(R1BIO_Uptodate, &r1_bio->state);
1416 while (d != r1_bio->read_disk) {
1418 d = conf->raid_disks;
1420 if (r1_bio->bios[d]->bi_end_io != end_sync_read)
1422 rdev = conf->mirrors[d].rdev;
1423 atomic_add(s, &rdev->corrected_errors);
1424 if (sync_page_io(rdev->bdev,
1425 sect + rdev->data_offset,
1427 bio->bi_io_vec[idx].bv_page,
1429 md_error(mddev, rdev);
1432 while (d != r1_bio->read_disk) {
1434 d = conf->raid_disks;
1436 if (r1_bio->bios[d]->bi_end_io != end_sync_read)
1438 rdev = conf->mirrors[d].rdev;
1439 if (sync_page_io(rdev->bdev,
1440 sect + rdev->data_offset,
1442 bio->bi_io_vec[idx].bv_page,
1444 md_error(mddev, rdev);
1447 char b[BDEVNAME_SIZE];
1448 /* Cannot read from anywhere, array is toast */
1449 md_error(mddev, conf->mirrors[r1_bio->read_disk].rdev);
1450 printk(KERN_ALERT "raid1: %s: unrecoverable I/O read error"
1451 " for block %llu\n",
1452 bdevname(bio->bi_bdev,b),
1453 (unsigned long long)r1_bio->sector);
1454 md_done_sync(mddev, r1_bio->sectors, 0);
1467 atomic_set(&r1_bio->remaining, 1);
1468 for (i = 0; i < disks ; i++) {
1469 wbio = r1_bio->bios[i];
1470 if (wbio->bi_end_io == NULL ||
1471 (wbio->bi_end_io == end_sync_read &&
1472 (i == r1_bio->read_disk ||
1473 !test_bit(MD_RECOVERY_SYNC, &mddev->recovery))))
1476 wbio->bi_rw = WRITE;
1477 wbio->bi_end_io = end_sync_write;
1478 atomic_inc(&r1_bio->remaining);
1479 md_sync_acct(conf->mirrors[i].rdev->bdev, wbio->bi_size >> 9);
1481 generic_make_request(wbio);
1484 if (atomic_dec_and_test(&r1_bio->remaining)) {
1485 /* if we're here, all write(s) have completed, so clean up */
1486 md_done_sync(mddev, r1_bio->sectors, 1);
1492 * This is a kernel thread which:
1494 * 1. Retries failed read operations on working mirrors.
1495 * 2. Updates the raid superblock when problems encounter.
1496 * 3. Performs writes following reads for array syncronising.
1499 static void fix_read_error(conf_t *conf, int read_disk,
1500 sector_t sect, int sectors)
1502 mddev_t *mddev = conf->mddev;
1510 if (s > (PAGE_SIZE>>9))
1514 /* Note: no rcu protection needed here
1515 * as this is synchronous in the raid1d thread
1516 * which is the thread that might remove
1517 * a device. If raid1d ever becomes multi-threaded....
1519 rdev = conf->mirrors[d].rdev;
1521 test_bit(In_sync, &rdev->flags) &&
1522 sync_page_io(rdev->bdev,
1523 sect + rdev->data_offset,
1525 conf->tmppage, READ))
1529 if (d == conf->raid_disks)
1532 } while (!success && d != read_disk);
1535 /* Cannot read from anywhere -- bye bye array */
1536 md_error(mddev, conf->mirrors[read_disk].rdev);
1539 /* write it back and re-read */
1541 while (d != read_disk) {
1543 d = conf->raid_disks;
1545 rdev = conf->mirrors[d].rdev;
1547 test_bit(In_sync, &rdev->flags)) {
1548 if (sync_page_io(rdev->bdev,
1549 sect + rdev->data_offset,
1550 s<<9, conf->tmppage, WRITE)
1552 /* Well, this device is dead */
1553 md_error(mddev, rdev);
1557 while (d != read_disk) {
1558 char b[BDEVNAME_SIZE];
1560 d = conf->raid_disks;
1562 rdev = conf->mirrors[d].rdev;
1564 test_bit(In_sync, &rdev->flags)) {
1565 if (sync_page_io(rdev->bdev,
1566 sect + rdev->data_offset,
1567 s<<9, conf->tmppage, READ)
1569 /* Well, this device is dead */
1570 md_error(mddev, rdev);
1572 atomic_add(s, &rdev->corrected_errors);
1574 "raid1:%s: read error corrected "
1575 "(%d sectors at %llu on %s)\n",
1577 (unsigned long long)(sect +
1579 bdevname(rdev->bdev, b));
1588 static void raid1d(mddev_t *mddev)
1592 unsigned long flags;
1593 conf_t *conf = mddev->private;
1594 struct list_head *head = &conf->retry_list;
1598 md_check_recovery(mddev);
1601 char b[BDEVNAME_SIZE];
1603 unplug += flush_pending_writes(conf);
1605 spin_lock_irqsave(&conf->device_lock, flags);
1606 if (list_empty(head)) {
1607 spin_unlock_irqrestore(&conf->device_lock, flags);
1610 r1_bio = list_entry(head->prev, r1bio_t, retry_list);
1611 list_del(head->prev);
1613 spin_unlock_irqrestore(&conf->device_lock, flags);
1615 mddev = r1_bio->mddev;
1616 conf = mddev->private;
1617 if (test_bit(R1BIO_IsSync, &r1_bio->state)) {
1618 sync_request_write(mddev, r1_bio);
1620 } else if (test_bit(R1BIO_BarrierRetry, &r1_bio->state)) {
1621 /* some requests in the r1bio were BIO_RW_BARRIER
1622 * requests which failed with -EOPNOTSUPP. Hohumm..
1623 * Better resubmit without the barrier.
1624 * We know which devices to resubmit for, because
1625 * all others have had their bios[] entry cleared.
1626 * We already have a nr_pending reference on these rdevs.
1629 const bool do_sync = bio_rw_flagged(r1_bio->master_bio, BIO_RW_SYNCIO);
1630 clear_bit(R1BIO_BarrierRetry, &r1_bio->state);
1631 clear_bit(R1BIO_Barrier, &r1_bio->state);
1632 for (i=0; i < conf->raid_disks; i++)
1633 if (r1_bio->bios[i])
1634 atomic_inc(&r1_bio->remaining);
1635 for (i=0; i < conf->raid_disks; i++)
1636 if (r1_bio->bios[i]) {
1637 struct bio_vec *bvec;
1640 bio = bio_clone(r1_bio->master_bio, GFP_NOIO);
1641 /* copy pages from the failed bio, as
1642 * this might be a write-behind device */
1643 __bio_for_each_segment(bvec, bio, j, 0)
1644 bvec->bv_page = bio_iovec_idx(r1_bio->bios[i], j)->bv_page;
1645 bio_put(r1_bio->bios[i]);
1646 bio->bi_sector = r1_bio->sector +
1647 conf->mirrors[i].rdev->data_offset;
1648 bio->bi_bdev = conf->mirrors[i].rdev->bdev;
1649 bio->bi_end_io = raid1_end_write_request;
1650 bio->bi_rw = WRITE |
1651 (do_sync << BIO_RW_SYNCIO);
1652 bio->bi_private = r1_bio;
1653 r1_bio->bios[i] = bio;
1654 generic_make_request(bio);
1659 /* we got a read error. Maybe the drive is bad. Maybe just
1660 * the block and we can fix it.
1661 * We freeze all other IO, and try reading the block from
1662 * other devices. When we find one, we re-write
1663 * and check it that fixes the read error.
1664 * This is all done synchronously while the array is
1667 if (mddev->ro == 0) {
1669 fix_read_error(conf, r1_bio->read_disk,
1672 unfreeze_array(conf);
1675 conf->mirrors[r1_bio->read_disk].rdev);
1677 bio = r1_bio->bios[r1_bio->read_disk];
1678 if ((disk=read_balance(conf, r1_bio)) == -1) {
1679 printk(KERN_ALERT "raid1: %s: unrecoverable I/O"
1680 " read error for block %llu\n",
1681 bdevname(bio->bi_bdev,b),
1682 (unsigned long long)r1_bio->sector);
1683 raid_end_bio_io(r1_bio);
1685 const bool do_sync = bio_rw_flagged(r1_bio->master_bio, BIO_RW_SYNCIO);
1686 r1_bio->bios[r1_bio->read_disk] =
1687 mddev->ro ? IO_BLOCKED : NULL;
1688 r1_bio->read_disk = disk;
1690 bio = bio_clone(r1_bio->master_bio, GFP_NOIO);
1691 r1_bio->bios[r1_bio->read_disk] = bio;
1692 rdev = conf->mirrors[disk].rdev;
1693 if (printk_ratelimit())
1694 printk(KERN_ERR "raid1: %s: redirecting sector %llu to"
1695 " another mirror\n",
1696 bdevname(rdev->bdev,b),
1697 (unsigned long long)r1_bio->sector);
1698 bio->bi_sector = r1_bio->sector + rdev->data_offset;
1699 bio->bi_bdev = rdev->bdev;
1700 bio->bi_end_io = raid1_end_read_request;
1701 bio->bi_rw = READ | (do_sync << BIO_RW_SYNCIO);
1702 bio->bi_private = r1_bio;
1704 generic_make_request(bio);
1710 unplug_slaves(mddev);
1714 static int init_resync(conf_t *conf)
1718 buffs = RESYNC_WINDOW / RESYNC_BLOCK_SIZE;
1719 BUG_ON(conf->r1buf_pool);
1720 conf->r1buf_pool = mempool_create(buffs, r1buf_pool_alloc, r1buf_pool_free,
1722 if (!conf->r1buf_pool)
1724 conf->next_resync = 0;
1729 * perform a "sync" on one "block"
1731 * We need to make sure that no normal I/O request - particularly write
1732 * requests - conflict with active sync requests.
1734 * This is achieved by tracking pending requests and a 'barrier' concept
1735 * that can be installed to exclude normal IO requests.
1738 static sector_t sync_request(mddev_t *mddev, sector_t sector_nr, int *skipped, int go_faster)
1740 conf_t *conf = mddev->private;
1743 sector_t max_sector, nr_sectors;
1747 int write_targets = 0, read_targets = 0;
1749 int still_degraded = 0;
1751 if (!conf->r1buf_pool)
1754 printk("sync start - bitmap %p\n", mddev->bitmap);
1756 if (init_resync(conf))
1760 max_sector = mddev->dev_sectors;
1761 if (sector_nr >= max_sector) {
1762 /* If we aborted, we need to abort the
1763 * sync on the 'current' bitmap chunk (there will
1764 * only be one in raid1 resync.
1765 * We can find the current addess in mddev->curr_resync
1767 if (mddev->curr_resync < max_sector) /* aborted */
1768 bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
1770 else /* completed sync */
1773 bitmap_close_sync(mddev->bitmap);
1778 if (mddev->bitmap == NULL &&
1779 mddev->recovery_cp == MaxSector &&
1780 !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
1781 conf->fullsync == 0) {
1783 return max_sector - sector_nr;
1785 /* before building a request, check if we can skip these blocks..
1786 * This call the bitmap_start_sync doesn't actually record anything
1788 if (!bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
1789 !conf->fullsync && !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
1790 /* We can skip this block, and probably several more */
1795 * If there is non-resync activity waiting for a turn,
1796 * and resync is going fast enough,
1797 * then let it though before starting on this new sync request.
1799 if (!go_faster && conf->nr_waiting)
1800 msleep_interruptible(1000);
1802 bitmap_cond_end_sync(mddev->bitmap, sector_nr);
1803 raise_barrier(conf);
1805 conf->next_resync = sector_nr;
1807 r1_bio = mempool_alloc(conf->r1buf_pool, GFP_NOIO);
1810 * If we get a correctably read error during resync or recovery,
1811 * we might want to read from a different device. So we
1812 * flag all drives that could conceivably be read from for READ,
1813 * and any others (which will be non-In_sync devices) for WRITE.
1814 * If a read fails, we try reading from something else for which READ
1818 r1_bio->mddev = mddev;
1819 r1_bio->sector = sector_nr;
1821 set_bit(R1BIO_IsSync, &r1_bio->state);
1823 for (i=0; i < conf->raid_disks; i++) {
1825 bio = r1_bio->bios[i];
1827 /* take from bio_init */
1828 bio->bi_next = NULL;
1829 bio->bi_flags |= 1 << BIO_UPTODATE;
1833 bio->bi_phys_segments = 0;
1835 bio->bi_end_io = NULL;
1836 bio->bi_private = NULL;
1838 rdev = rcu_dereference(conf->mirrors[i].rdev);
1840 test_bit(Faulty, &rdev->flags)) {
1843 } else if (!test_bit(In_sync, &rdev->flags)) {
1845 bio->bi_end_io = end_sync_write;
1848 /* may need to read from here */
1850 bio->bi_end_io = end_sync_read;
1851 if (test_bit(WriteMostly, &rdev->flags)) {
1860 atomic_inc(&rdev->nr_pending);
1861 bio->bi_sector = sector_nr + rdev->data_offset;
1862 bio->bi_bdev = rdev->bdev;
1863 bio->bi_private = r1_bio;
1868 r1_bio->read_disk = disk;
1870 if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery) && read_targets > 0)
1871 /* extra read targets are also write targets */
1872 write_targets += read_targets-1;
1874 if (write_targets == 0 || read_targets == 0) {
1875 /* There is nowhere to write, so all non-sync
1876 * drives must be failed - so we are finished
1878 sector_t rv = max_sector - sector_nr;
1884 if (max_sector > mddev->resync_max)
1885 max_sector = mddev->resync_max; /* Don't do IO beyond here */
1890 int len = PAGE_SIZE;
1891 if (sector_nr + (len>>9) > max_sector)
1892 len = (max_sector - sector_nr) << 9;
1895 if (sync_blocks == 0) {
1896 if (!bitmap_start_sync(mddev->bitmap, sector_nr,
1897 &sync_blocks, still_degraded) &&
1899 !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery))
1901 BUG_ON(sync_blocks < (PAGE_SIZE>>9));
1902 if (len > (sync_blocks<<9))
1903 len = sync_blocks<<9;
1906 for (i=0 ; i < conf->raid_disks; i++) {
1907 bio = r1_bio->bios[i];
1908 if (bio->bi_end_io) {
1909 page = bio->bi_io_vec[bio->bi_vcnt].bv_page;
1910 if (bio_add_page(bio, page, len, 0) == 0) {
1912 bio->bi_io_vec[bio->bi_vcnt].bv_page = page;
1915 bio = r1_bio->bios[i];
1916 if (bio->bi_end_io==NULL)
1918 /* remove last page from this bio */
1920 bio->bi_size -= len;
1921 bio->bi_flags &= ~(1<< BIO_SEG_VALID);
1927 nr_sectors += len>>9;
1928 sector_nr += len>>9;
1929 sync_blocks -= (len>>9);
1930 } while (r1_bio->bios[disk]->bi_vcnt < RESYNC_PAGES);
1932 r1_bio->sectors = nr_sectors;
1934 /* For a user-requested sync, we read all readable devices and do a
1937 if (test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
1938 atomic_set(&r1_bio->remaining, read_targets);
1939 for (i=0; i<conf->raid_disks; i++) {
1940 bio = r1_bio->bios[i];
1941 if (bio->bi_end_io == end_sync_read) {
1942 md_sync_acct(bio->bi_bdev, nr_sectors);
1943 generic_make_request(bio);
1947 atomic_set(&r1_bio->remaining, 1);
1948 bio = r1_bio->bios[r1_bio->read_disk];
1949 md_sync_acct(bio->bi_bdev, nr_sectors);
1950 generic_make_request(bio);
1956 static sector_t raid1_size(mddev_t *mddev, sector_t sectors, int raid_disks)
1961 return mddev->dev_sectors;
1964 static conf_t *setup_conf(mddev_t *mddev)
1968 mirror_info_t *disk;
1972 conf = kzalloc(sizeof(conf_t), GFP_KERNEL);
1976 conf->mirrors = kzalloc(sizeof(struct mirror_info)*mddev->raid_disks,
1981 conf->tmppage = alloc_page(GFP_KERNEL);
1985 conf->poolinfo = kzalloc(sizeof(*conf->poolinfo), GFP_KERNEL);
1986 if (!conf->poolinfo)
1988 conf->poolinfo->raid_disks = mddev->raid_disks;
1989 conf->r1bio_pool = mempool_create(NR_RAID1_BIOS, r1bio_pool_alloc,
1992 if (!conf->r1bio_pool)
1995 conf->poolinfo->mddev = mddev;
1997 spin_lock_init(&conf->device_lock);
1998 list_for_each_entry(rdev, &mddev->disks, same_set) {
1999 int disk_idx = rdev->raid_disk;
2000 if (disk_idx >= mddev->raid_disks
2003 disk = conf->mirrors + disk_idx;
2007 disk->head_position = 0;
2009 conf->raid_disks = mddev->raid_disks;
2010 conf->mddev = mddev;
2011 INIT_LIST_HEAD(&conf->retry_list);
2013 spin_lock_init(&conf->resync_lock);
2014 init_waitqueue_head(&conf->wait_barrier);
2016 bio_list_init(&conf->pending_bio_list);
2017 bio_list_init(&conf->flushing_bio_list);
2019 conf->last_used = -1;
2020 for (i = 0; i < conf->raid_disks; i++) {
2022 disk = conf->mirrors + i;
2025 !test_bit(In_sync, &disk->rdev->flags)) {
2026 disk->head_position = 0;
2029 } else if (conf->last_used < 0)
2031 * The first working device is used as a
2032 * starting point to read balancing.
2034 conf->last_used = i;
2038 if (conf->last_used < 0) {
2039 printk(KERN_ERR "raid1: no operational mirrors for %s\n",
2044 conf->thread = md_register_thread(raid1d, mddev, NULL);
2045 if (!conf->thread) {
2047 "raid1: couldn't allocate thread for %s\n",
2056 if (conf->r1bio_pool)
2057 mempool_destroy(conf->r1bio_pool);
2058 kfree(conf->mirrors);
2059 safe_put_page(conf->tmppage);
2060 kfree(conf->poolinfo);
2063 return ERR_PTR(err);
2066 static int run(mddev_t *mddev)
2072 if (mddev->level != 1) {
2073 printk("raid1: %s: raid level not set to mirroring (%d)\n",
2074 mdname(mddev), mddev->level);
2077 if (mddev->reshape_position != MaxSector) {
2078 printk("raid1: %s: reshape_position set but not supported\n",
2083 * copy the already verified devices into our private RAID1
2084 * bookkeeping area. [whatever we allocate in run(),
2085 * should be freed in stop()]
2087 if (mddev->private == NULL)
2088 conf = setup_conf(mddev);
2090 conf = mddev->private;
2093 return PTR_ERR(conf);
2095 mddev->queue->queue_lock = &conf->device_lock;
2096 list_for_each_entry(rdev, &mddev->disks, same_set) {
2097 disk_stack_limits(mddev->gendisk, rdev->bdev,
2098 rdev->data_offset << 9);
2099 /* as we don't honour merge_bvec_fn, we must never risk
2100 * violating it, so limit ->max_sector to one PAGE, as
2101 * a one page request is never in violation.
2103 if (rdev->bdev->bd_disk->queue->merge_bvec_fn &&
2104 queue_max_sectors(mddev->queue) > (PAGE_SIZE>>9))
2105 blk_queue_max_sectors(mddev->queue, PAGE_SIZE>>9);
2108 mddev->degraded = 0;
2109 for (i=0; i < conf->raid_disks; i++)
2110 if (conf->mirrors[i].rdev == NULL ||
2111 !test_bit(In_sync, &conf->mirrors[i].rdev->flags) ||
2112 test_bit(Faulty, &conf->mirrors[i].rdev->flags))
2115 if (conf->raid_disks - mddev->degraded == 1)
2116 mddev->recovery_cp = MaxSector;
2118 if (mddev->recovery_cp != MaxSector)
2119 printk(KERN_NOTICE "raid1: %s is not clean"
2120 " -- starting background reconstruction\n",
2123 "raid1: raid set %s active with %d out of %d mirrors\n",
2124 mdname(mddev), mddev->raid_disks - mddev->degraded,
2128 * Ok, everything is just fine now
2130 mddev->thread = conf->thread;
2131 conf->thread = NULL;
2132 mddev->private = conf;
2134 md_set_array_sectors(mddev, raid1_size(mddev, 0, 0));
2136 mddev->queue->unplug_fn = raid1_unplug;
2137 mddev->queue->backing_dev_info.congested_fn = raid1_congested;
2138 mddev->queue->backing_dev_info.congested_data = mddev;
2139 md_integrity_register(mddev);
2143 static int stop(mddev_t *mddev)
2145 conf_t *conf = mddev->private;
2146 struct bitmap *bitmap = mddev->bitmap;
2147 int behind_wait = 0;
2149 /* wait for behind writes to complete */
2150 while (bitmap && atomic_read(&bitmap->behind_writes) > 0) {
2152 printk(KERN_INFO "raid1: behind writes in progress on device %s, waiting to stop (%d)\n", mdname(mddev), behind_wait);
2153 set_current_state(TASK_UNINTERRUPTIBLE);
2154 schedule_timeout(HZ); /* wait a second */
2155 /* need to kick something here to make sure I/O goes? */
2158 raise_barrier(conf);
2159 lower_barrier(conf);
2161 md_unregister_thread(mddev->thread);
2162 mddev->thread = NULL;
2163 blk_sync_queue(mddev->queue); /* the unplug fn references 'conf'*/
2164 if (conf->r1bio_pool)
2165 mempool_destroy(conf->r1bio_pool);
2166 kfree(conf->mirrors);
2167 kfree(conf->poolinfo);
2169 mddev->private = NULL;
2173 static int raid1_resize(mddev_t *mddev, sector_t sectors)
2175 /* no resync is happening, and there is enough space
2176 * on all devices, so we can resize.
2177 * We need to make sure resync covers any new space.
2178 * If the array is shrinking we should possibly wait until
2179 * any io in the removed space completes, but it hardly seems
2182 md_set_array_sectors(mddev, raid1_size(mddev, sectors, 0));
2183 if (mddev->array_sectors > raid1_size(mddev, sectors, 0))
2185 set_capacity(mddev->gendisk, mddev->array_sectors);
2187 revalidate_disk(mddev->gendisk);
2188 if (sectors > mddev->dev_sectors &&
2189 mddev->recovery_cp == MaxSector) {
2190 mddev->recovery_cp = mddev->dev_sectors;
2191 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
2193 mddev->dev_sectors = sectors;
2194 mddev->resync_max_sectors = sectors;
2198 static int raid1_reshape(mddev_t *mddev)
2201 * 1/ resize the r1bio_pool
2202 * 2/ resize conf->mirrors
2204 * We allocate a new r1bio_pool if we can.
2205 * Then raise a device barrier and wait until all IO stops.
2206 * Then resize conf->mirrors and swap in the new r1bio pool.
2208 * At the same time, we "pack" the devices so that all the missing
2209 * devices have the higher raid_disk numbers.
2211 mempool_t *newpool, *oldpool;
2212 struct pool_info *newpoolinfo;
2213 mirror_info_t *newmirrors;
2214 conf_t *conf = mddev->private;
2215 int cnt, raid_disks;
2216 unsigned long flags;
2219 /* Cannot change chunk_size, layout, or level */
2220 if (mddev->chunk_sectors != mddev->new_chunk_sectors ||
2221 mddev->layout != mddev->new_layout ||
2222 mddev->level != mddev->new_level) {
2223 mddev->new_chunk_sectors = mddev->chunk_sectors;
2224 mddev->new_layout = mddev->layout;
2225 mddev->new_level = mddev->level;
2229 err = md_allow_write(mddev);
2233 raid_disks = mddev->raid_disks + mddev->delta_disks;
2235 if (raid_disks < conf->raid_disks) {
2237 for (d= 0; d < conf->raid_disks; d++)
2238 if (conf->mirrors[d].rdev)
2240 if (cnt > raid_disks)
2244 newpoolinfo = kmalloc(sizeof(*newpoolinfo), GFP_KERNEL);
2247 newpoolinfo->mddev = mddev;
2248 newpoolinfo->raid_disks = raid_disks;
2250 newpool = mempool_create(NR_RAID1_BIOS, r1bio_pool_alloc,
2251 r1bio_pool_free, newpoolinfo);
2256 newmirrors = kzalloc(sizeof(struct mirror_info) * raid_disks, GFP_KERNEL);
2259 mempool_destroy(newpool);
2263 raise_barrier(conf);
2265 /* ok, everything is stopped */
2266 oldpool = conf->r1bio_pool;
2267 conf->r1bio_pool = newpool;
2269 for (d = d2 = 0; d < conf->raid_disks; d++) {
2270 mdk_rdev_t *rdev = conf->mirrors[d].rdev;
2271 if (rdev && rdev->raid_disk != d2) {
2273 sprintf(nm, "rd%d", rdev->raid_disk);
2274 sysfs_remove_link(&mddev->kobj, nm);
2275 rdev->raid_disk = d2;
2276 sprintf(nm, "rd%d", rdev->raid_disk);
2277 sysfs_remove_link(&mddev->kobj, nm);
2278 if (sysfs_create_link(&mddev->kobj,
2281 "md/raid1: cannot register "
2286 newmirrors[d2++].rdev = rdev;
2288 kfree(conf->mirrors);
2289 conf->mirrors = newmirrors;
2290 kfree(conf->poolinfo);
2291 conf->poolinfo = newpoolinfo;
2293 spin_lock_irqsave(&conf->device_lock, flags);
2294 mddev->degraded += (raid_disks - conf->raid_disks);
2295 spin_unlock_irqrestore(&conf->device_lock, flags);
2296 conf->raid_disks = mddev->raid_disks = raid_disks;
2297 mddev->delta_disks = 0;
2299 conf->last_used = 0; /* just make sure it is in-range */
2300 lower_barrier(conf);
2302 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
2303 md_wakeup_thread(mddev->thread);
2305 mempool_destroy(oldpool);
2309 static void raid1_quiesce(mddev_t *mddev, int state)
2311 conf_t *conf = mddev->private;
2314 case 2: /* wake for suspend */
2315 wake_up(&conf->wait_barrier);
2318 raise_barrier(conf);
2321 lower_barrier(conf);
2326 static void *raid1_takeover(mddev_t *mddev)
2328 /* raid1 can take over:
2329 * raid5 with 2 devices, any layout or chunk size
2331 if (mddev->level == 5 && mddev->raid_disks == 2) {
2333 mddev->new_level = 1;
2334 mddev->new_layout = 0;
2335 mddev->new_chunk_sectors = 0;
2336 conf = setup_conf(mddev);
2341 return ERR_PTR(-EINVAL);
2344 static struct mdk_personality raid1_personality =
2348 .owner = THIS_MODULE,
2349 .make_request = make_request,
2353 .error_handler = error,
2354 .hot_add_disk = raid1_add_disk,
2355 .hot_remove_disk= raid1_remove_disk,
2356 .spare_active = raid1_spare_active,
2357 .sync_request = sync_request,
2358 .resize = raid1_resize,
2360 .check_reshape = raid1_reshape,
2361 .quiesce = raid1_quiesce,
2362 .takeover = raid1_takeover,
2365 static int __init raid_init(void)
2367 return register_md_personality(&raid1_personality);
2370 static void raid_exit(void)
2372 unregister_md_personality(&raid1_personality);
2375 module_init(raid_init);
2376 module_exit(raid_exit);
2377 MODULE_LICENSE("GPL");
2378 MODULE_ALIAS("md-personality-3"); /* RAID1 */
2379 MODULE_ALIAS("md-raid1");
2380 MODULE_ALIAS("md-level-1");