md/raid1 has handled barriers for quite some time, but the other md/raid levels don't. The recent thread about raid5 being fundamentally broken (http://lwn.net/Articles/349970/) made me decide it was about time to actually do something about this (it was always my hope that filesystems would do the sensible thing and all 'flush' at appropriate times, but it doesn't look like that is going to happen). This patch is the result (though it won't apply without one or two others, so safest to pull from git://neil.brown.name/md md-scratch ). It adds barrier support for all other levels by: - sending zero-length barriers to all devices - sending the original request - sending another load of zero-length barriers New requests are blocked while we wait for the barriers to all get processed. This ensures the barrier request is properly ordered w.r.t. all other requests (but possibly makes things horribly slow...). I would really appreciate review of the approach, and testing of the result, if anyone is interested. How to actually test that your barriers are doing the right thing is not clear to be, and at least testing that it doesn't break things or kill performance would be good. Thanks, NeilBrown From: NeilBrown <neilb@xxxxxxx> Date: Fri, 18 Sep 2009 15:55:09 +1000 Subject: [PATCH] md: support barrier requests on all personalities. Previously barriers were only supported on RAID1. This is because other levels requires synchronisation across all devices and so needed a different approach. Here is that approach. When a barrier arrives, we send a zero-length barrier to every active device. When that completes we submit the barrier request itself (with the barrier flag cleared) and the submit a fresh load of zero length barriers. The barrier request itself is asynchronous, but any subsequent request will block until the barrier completes. The reason for clearing the barrier flag is that a barrier request is allowed to fail. If we pass a non-empty barrier through a striping raid level it is conceivable that part of it could succeed and part could fail. That would be way to hard to deal with. So if the first run of zero length barriers succeed, we assume all is sufficiently well that we send the request and ignore errors in the second run of barriers. RAID5 needs extra care as write requests may not have been submitted to the underlying devices yet. So we flush the stripe cache before proceeding with the barrier. Signed-off-by: NeilBrown <neilb@xxxxxxx> --- drivers/md/linear.c | 2 +- drivers/md/md.c | 111 +++++++++++++++++++++++++++++++++++++++++++++++- drivers/md/md.h | 12 +++++ drivers/md/multipath.c | 2 +- drivers/md/raid0.c | 2 +- drivers/md/raid10.c | 2 +- drivers/md/raid5.c | 8 +++- 7 files changed, 132 insertions(+), 7 deletions(-) diff --git a/drivers/md/linear.c b/drivers/md/linear.c index 53607c1..80504a8 100644 --- a/drivers/md/linear.c +++ b/drivers/md/linear.c @@ -292,7 +292,7 @@ static int linear_make_request (struct request_queue *q, struct bio *bio) int cpu; if (unlikely(bio_barrier(bio))) { - bio_endio(bio, -EOPNOTSUPP); + md_barrier_request(mddev, bio); return 0; } diff --git a/drivers/md/md.c b/drivers/md/md.c index 8c41efe..820ae3b 100644 --- a/drivers/md/md.c +++ b/drivers/md/md.c @@ -217,12 +217,12 @@ static int md_make_request(struct request_queue *q, struct bio *bio) return 0; } rcu_read_lock(); - if (mddev->suspended) { + if (mddev->suspended || mddev->barrier) { DEFINE_WAIT(__wait); for (;;) { prepare_to_wait(&mddev->sb_wait, &__wait, TASK_UNINTERRUPTIBLE); - if (!mddev->suspended) + if (!mddev->suspended && !mddev->barrier) break; rcu_read_unlock(); schedule(); @@ -264,10 +264,116 @@ static void mddev_resume(mddev_t *mddev) int mddev_congested(mddev_t *mddev, int bits) { + if (mddev->barrier) + return 1; return mddev->suspended; } EXPORT_SYMBOL(mddev_congested); +/* + * Generic barrier handling for md + */ + +static void md_end_barrier(struct bio *bio, int err) +{ + mdk_rdev_t *rdev = bio->bi_private; + mddev_t *mddev = rdev->mddev; + if (err == -EOPNOTSUPP && mddev->barrier != (void*)1) + set_bit(BIO_EOPNOTSUPP, &mddev->barrier->bi_flags); + + rdev_dec_pending(rdev, mddev); + + if (atomic_dec_and_test(&mddev->flush_pending)) { + if (mddev->barrier == (void*)1) { + mddev->barrier = NULL; + wake_up(&mddev->sb_wait); + } else + schedule_work(&mddev->barrier_work); + } + bio_put(bio); +} + +static void md_submit_barrier(struct work_struct *ws) +{ + mddev_t *mddev = container_of(ws, mddev_t, barrier_work); + struct bio *bio = mddev->barrier; + + atomic_set(&mddev->flush_pending, 1); + + if (!test_bit(BIO_EOPNOTSUPP, &bio->bi_flags)) { + mdk_rdev_t *rdev; + + bio->bi_rw &= ~(1<<BIO_RW_BARRIER); + if (mddev->pers->make_request(mddev->queue, bio)) + generic_make_request(bio); + mddev->barrier = (void*)1; + rcu_read_lock(); + list_for_each_entry_rcu(rdev, &mddev->disks, same_set) + if (rdev->raid_disk >= 0 && + !test_bit(Faulty, &rdev->flags)) { + /* Take two references, one is dropped + * when request finishes, one after + * we reclaim rcu_read_lock + */ + struct bio *bi; + atomic_inc(&rdev->nr_pending); + atomic_inc(&rdev->nr_pending); + rcu_read_unlock(); + bi = bio_alloc(GFP_KERNEL, 0); + bi->bi_end_io = md_end_barrier; + bi->bi_private = rdev; + bi->bi_bdev = rdev->bdev; + atomic_inc(&mddev->flush_pending); + submit_bio(WRITE_BARRIER, bi); + rcu_read_lock(); + rdev_dec_pending(rdev, mddev); + } + rcu_read_unlock(); + } else + bio_endio(bio, -EOPNOTSUPP); + if (atomic_dec_and_test(&mddev->flush_pending)) { + mddev->barrier = NULL; + wake_up(&mddev->sb_wait); + } +} + +void md_barrier_request(mddev_t *mddev, struct bio *bio) +{ + mdk_rdev_t *rdev; + + spin_lock_irq(&mddev->write_lock); + wait_event_lock_irq(mddev->sb_wait, + !mddev->barrier, + mddev->write_lock, /*nothing*/); + mddev->barrier = bio; + spin_unlock_irq(&mddev->write_lock); + + atomic_set(&mddev->flush_pending, 1); + INIT_WORK(&mddev->barrier_work, md_submit_barrier); + + rcu_read_lock(); + list_for_each_entry_rcu(rdev, &mddev->disks, same_set) + if (rdev->raid_disk >= 0 && + !test_bit(Faulty, &rdev->flags)) { + struct bio *bi; + + atomic_inc(&rdev->nr_pending); + atomic_inc(&rdev->nr_pending); + rcu_read_unlock(); + bi = bio_alloc(GFP_KERNEL, 0); + bi->bi_end_io = md_end_barrier; + bi->bi_private = rdev; + bi->bi_bdev = rdev->bdev; + atomic_inc(&mddev->flush_pending); + submit_bio(WRITE_BARRIER, bi); + rcu_read_lock(); + rdev_dec_pending(rdev, mddev); + } + rcu_read_unlock(); + if (atomic_dec_and_test(&mddev->flush_pending)) + schedule_work(&mddev->barrier_work); +} +EXPORT_SYMBOL(md_barrier_request); static inline mddev_t *mddev_get(mddev_t *mddev) { @@ -374,6 +480,7 @@ static mddev_t * mddev_find(dev_t unit) atomic_set(&new->openers, 0); atomic_set(&new->active_io, 0); spin_lock_init(&new->write_lock); + atomic_set(&new->flush_pending, 0); init_waitqueue_head(&new->sb_wait); init_waitqueue_head(&new->recovery_wait); new->reshape_position = MaxSector; diff --git a/drivers/md/md.h b/drivers/md/md.h index 5de3902..57ccc6f 100644 --- a/drivers/md/md.h +++ b/drivers/md/md.h @@ -291,6 +291,17 @@ struct mddev_s */ struct list_head all_mddevs; + + /* Generic barrier handling. + * If there is a pending barrier request, all other + * writes are blocked while the devices are flushed. + * The last to finish a flush schedules a worker to + * submit the barrier request (without the barrier flag), + * then submit more flush requests. + */ + struct bio *barrier; + atomic_t flush_pending; + struct work_struct barrier_work; }; @@ -431,6 +442,7 @@ extern void md_done_sync(mddev_t *mddev, int blocks, int ok); extern void md_error(mddev_t *mddev, mdk_rdev_t *rdev); extern int mddev_congested(mddev_t *mddev, int bits); +extern void md_barrier_request(mddev_t *mddev, struct bio *bio); extern void md_super_write(mddev_t *mddev, mdk_rdev_t *rdev, sector_t sector, int size, struct page *page); extern void md_super_wait(mddev_t *mddev); diff --git a/drivers/md/multipath.c b/drivers/md/multipath.c index ce4b4cd..f2bb20c 100644 --- a/drivers/md/multipath.c +++ b/drivers/md/multipath.c @@ -145,7 +145,7 @@ static int multipath_make_request (struct request_queue *q, struct bio * bio) int cpu; if (unlikely(bio_barrier(bio))) { - bio_endio(bio, -EOPNOTSUPP); + md_barrier_request(mddev, bio); return 0; } diff --git a/drivers/md/raid0.c b/drivers/md/raid0.c index 08adcb0..87325c5 100644 --- a/drivers/md/raid0.c +++ b/drivers/md/raid0.c @@ -453,7 +453,7 @@ static int raid0_make_request(struct request_queue *q, struct bio *bio) int cpu; if (unlikely(bio_barrier(bio))) { - bio_endio(bio, -EOPNOTSUPP); + md_barrier_request(mddev, bio); return 0; } diff --git a/drivers/md/raid10.c b/drivers/md/raid10.c index 3868880..113b672 100644 --- a/drivers/md/raid10.c +++ b/drivers/md/raid10.c @@ -804,7 +804,7 @@ static int make_request(struct request_queue *q, struct bio * bio) mdk_rdev_t *blocked_rdev; if (unlikely(bio_barrier(bio))) { - bio_endio(bio, -EOPNOTSUPP); + md_barrier_request(mddev, bio); return 0; } diff --git a/drivers/md/raid5.c b/drivers/md/raid5.c index 584abea..7b6544e 100644 --- a/drivers/md/raid5.c +++ b/drivers/md/raid5.c @@ -3611,7 +3611,13 @@ static int make_request(struct request_queue *q, struct bio * bi) int cpu, remaining; if (unlikely(bio_barrier(bi))) { - bio_endio(bi, -EOPNOTSUPP); + /* Drain all pending writes. We only really need + * to ensure they have been submitted, but this is + * easier. + */ + mddev->pers->quiesce(mddev, 1); + mddev->pers->quiesce(mddev, 0); + md_barrier_request(mddev, bi); return 0; } -- 1.6.3.3 -- To unsubscribe from this list: send the line "unsubscribe linux-raid" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
