Linux-Raid Archives on lore.kernel.org
 help / color / Atom feed
From: Xiao Ni <xni@redhat.com>
To: songliubraving@fb.com
Cc: linux-raid@vger.kernel.org, matthew.ruffell@canonical.com,
	colyli@suse.de, guoqing.jiang@cloud.ionos.com,
	ncroxon@redhat.com
Subject: [PATCH 4/5] md/raid10: improve raid10 discard request
Date: Wed,  3 Feb 2021 21:45:30 +0800
Message-ID: <1612359931-24209-5-git-send-email-xni@redhat.com> (raw)
In-Reply-To: <1612359931-24209-1-git-send-email-xni@redhat.com>

Now the discard request is split by chunk size. So it takes a long time
to finish mkfs on disks which support discard function. This patch improve
handling raid10 discard request. It uses the similar way with patch
29efc390b (md/md0: optimize raid0 discard handling).

But it's a little complex than raid0. Because raid10 has different layout.
If raid10 is offset layout and the discard request is smaller than stripe
size. There are some holes when we submit discard bio to underlayer disks.

For example: five disks (disk1 - disk5)
D01 D02 D03 D04 D05
D05 D01 D02 D03 D04
D06 D07 D08 D09 D10
D10 D06 D07 D08 D09
The discard bio just wants to discard from D03 to D10. For disk3, there is
a hole between D03 and D08. For disk4, there is a hole between D04 and D09.
D03 is a chunk, raid10_write_request can handle one chunk perfectly. So
the part that is not aligned with stripe size is still handled by
raid10_write_request.

If reshape is running when discard bio comes and the discard bio spans the
reshape position, raid10_write_request is responsible to handle this
discard bio.

I did a test with this patch set.
Without patch:
time mkfs.xfs /dev/md0
real4m39.775s
user0m0.000s
sys0m0.298s

With patch:
time mkfs.xfs /dev/md0
real0m0.105s
user0m0.000s
sys0m0.007s

nvme3n1           259:1    0   477G  0 disk
└─nvme3n1p1       259:10   0    50G  0 part
nvme4n1           259:2    0   477G  0 disk
└─nvme4n1p1       259:11   0    50G  0 part
nvme5n1           259:6    0   477G  0 disk
└─nvme5n1p1       259:12   0    50G  0 part
nvme2n1           259:9    0   477G  0 disk
└─nvme2n1p1       259:15   0    50G  0 part
nvme0n1           259:13   0   477G  0 disk
└─nvme0n1p1       259:14   0    50G  0 part

Reviewed-by: Coly Li <colyli@suse.de>
Reviewed-by: Guoqing Jiang <guoqing.jiang@cloud.ionos.com>
Signed-off-by: Xiao Ni <xni@redhat.com>
---
 drivers/md/raid10.c | 262 +++++++++++++++++++++++++++++++++++++++++++++++++++-
 1 file changed, 261 insertions(+), 1 deletion(-)

diff --git a/drivers/md/raid10.c b/drivers/md/raid10.c
index 388eb3f..e059796 100644
--- a/drivers/md/raid10.c
+++ b/drivers/md/raid10.c
@@ -1515,6 +1515,262 @@ static void __make_request(struct mddev *mddev, struct bio *bio, int sectors)
 		raid10_write_request(mddev, bio, r10_bio);
 }
 
+static struct bio *raid10_split_bio(struct r10conf *conf,
+			struct bio *bio, sector_t sectors, bool want_first)
+{
+	struct bio *split;
+
+	split = bio_split(bio, sectors,	GFP_NOIO, &conf->bio_split);
+	bio_chain(split, bio);
+	allow_barrier(conf);
+	if (want_first) {
+		submit_bio_noacct(bio);
+		bio = split;
+	} else
+		submit_bio_noacct(split);
+	wait_barrier(conf);
+
+	return bio;
+}
+
+static void raid10_end_discard_request(struct bio *bio)
+{
+	struct r10bio *r10_bio = bio->bi_private;
+	struct r10conf *conf = r10_bio->mddev->private;
+	struct md_rdev *rdev = NULL;
+	int dev;
+	int slot, repl;
+
+	/*
+	 * We don't care the return value of discard bio
+	 */
+	if (!test_bit(R10BIO_Uptodate, &r10_bio->state))
+		set_bit(R10BIO_Uptodate, &r10_bio->state);
+
+	dev = find_bio_disk(conf, r10_bio, bio, &slot, &repl);
+	if (repl)
+		rdev = conf->mirrors[dev].replacement;
+	if (!rdev) {
+		/* raid10_remove_disk uses smp_mb to make sure rdev is set to
+		 * replacement before setting replacement to NULL. It can read
+		 * rdev first without barrier protect even replacment is NULL
+		 */
+		smp_rmb();
+		rdev = conf->mirrors[dev].rdev;
+	}
+
+	if (atomic_dec_and_test(&r10_bio->remaining)) {
+		md_write_end(r10_bio->mddev);
+		raid_end_bio_io(r10_bio);
+	}
+
+	rdev_dec_pending(rdev, conf->mddev);
+}
+
+/* There are some limitations to handle discard bio
+ * 1st, the discard size is bigger than stripe_size*2.
+ * 2st, if the discard bio spans reshape progress, we use the old way to
+ * handle discard bio
+ */
+static int raid10_handle_discard(struct mddev *mddev, struct bio *bio)
+{
+	struct r10conf *conf = mddev->private;
+	struct geom *geo = &conf->geo;
+	struct r10bio *r10_bio;
+
+	int disk;
+	sector_t chunk;
+	unsigned int stripe_size;
+	unsigned int stripe_data_disks;
+	sector_t split_size;
+
+	sector_t bio_start, bio_end;
+	sector_t first_stripe_index, last_stripe_index;
+	sector_t start_disk_offset;
+	unsigned int start_disk_index;
+	sector_t end_disk_offset;
+	unsigned int end_disk_index;
+	unsigned int remainder;
+
+	if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
+		return -EAGAIN;
+
+	wait_barrier(conf);
+
+	/* Check reshape again to avoid reshape happens after checking
+	 * MD_RECOVERY_RESHAPE and before wait_barrier
+	 */
+	if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
+		goto out;
+
+	stripe_data_disks = geo->near_copies ?
+				geo->raid_disks / geo->near_copies +
+				geo->raid_disks % geo->near_copies :
+				geo->raid_disks;
+	stripe_size = stripe_data_disks << geo->chunk_shift;
+
+	bio_start = bio->bi_iter.bi_sector;
+	bio_end = bio_end_sector(bio);
+
+	/* Maybe one discard bio is smaller than strip size or across one stripe
+	 * and discard region is larger than one stripe size. For far offset layout,
+	 * if the discard region is not aligned with stripe size, there is hole
+	 * when we submit discard bio to member disk. For simplicity, we only
+	 * handle discard bio which discard region is bigger than stripe_size*2
+	 */
+	if (bio_sectors(bio) < stripe_size*2)
+		goto out;
+
+	/*
+	 * Keep bio aligned with strip size.
+	 */
+	div_u64_rem(bio_start, stripe_size, &remainder);
+	if (remainder) {
+		split_size = stripe_size - remainder;
+		bio = raid10_split_bio(conf, bio, split_size, false);
+	}
+	div_u64_rem(bio_end, stripe_size, &remainder);
+	if (remainder) {
+		split_size = bio_sectors(bio) - remainder;
+		bio = raid10_split_bio(conf, bio, split_size, true);
+	}
+
+	r10_bio = mempool_alloc(&conf->r10bio_pool, GFP_NOIO);
+	r10_bio->mddev = mddev;
+	r10_bio->state = 0;
+	r10_bio->sectors = 0;
+	memset(r10_bio->devs, 0, sizeof(r10_bio->devs[0]) * geo->raid_disks);
+
+	wait_blocked_dev(mddev, r10_bio);
+
+	r10_bio->master_bio = bio;
+
+	bio_start = bio->bi_iter.bi_sector;
+	bio_end = bio_end_sector(bio);
+
+	/* raid10 uses chunk as the unit to store data. It's similar like raid0.
+	 * One stripe contains the chunks from all member disk (one chunk from
+	 * one disk at the same HBA address). For layout detail, see 'man md 4'
+	 */
+	chunk = bio_start >> geo->chunk_shift;
+	chunk *= geo->near_copies;
+	first_stripe_index = chunk;
+	start_disk_index = sector_div(first_stripe_index, geo->raid_disks);
+	if (geo->far_offset)
+		first_stripe_index *= geo->far_copies;
+	start_disk_offset = (bio_start & geo->chunk_mask) +
+				(first_stripe_index << geo->chunk_shift);
+
+	chunk = bio_end >> geo->chunk_shift;
+	chunk *= geo->near_copies;
+	last_stripe_index = chunk;
+	end_disk_index = sector_div(last_stripe_index, geo->raid_disks);
+	if (geo->far_offset)
+		last_stripe_index *= geo->far_copies;
+	end_disk_offset = (bio_end & geo->chunk_mask) +
+				(last_stripe_index << geo->chunk_shift);
+
+	rcu_read_lock();
+	for (disk = 0; disk < geo->raid_disks; disk++) {
+		struct md_rdev *rdev = rcu_dereference(conf->mirrors[disk].rdev);
+		struct md_rdev *rrdev = rcu_dereference(
+			conf->mirrors[disk].replacement);
+
+		r10_bio->devs[disk].bio = NULL;
+		r10_bio->devs[disk].repl_bio = NULL;
+
+		if (rdev && (test_bit(Faulty, &rdev->flags)))
+			rdev = NULL;
+		if (rrdev && (test_bit(Faulty, &rrdev->flags)))
+			rrdev = NULL;
+		if (!rdev && !rrdev)
+			continue;
+
+		if (rdev) {
+			r10_bio->devs[disk].bio = bio;
+			atomic_inc(&rdev->nr_pending);
+		}
+		if (rrdev) {
+			r10_bio->devs[disk].repl_bio = bio;
+			atomic_inc(&rrdev->nr_pending);
+		}
+	}
+	rcu_read_unlock();
+
+	atomic_set(&r10_bio->remaining, 1);
+	for (disk = 0; disk < geo->raid_disks; disk++) {
+		sector_t dev_start, dev_end;
+		struct bio *mbio, *rbio = NULL;
+		struct md_rdev *rdev = rcu_dereference(conf->mirrors[disk].rdev);
+		struct md_rdev *rrdev = rcu_dereference(
+			conf->mirrors[disk].replacement);
+
+		/*
+		 * Now start to calculate the start and end address for each disk.
+		 * The space between dev_start and dev_end is the discard region.
+		 *
+		 * For dev_start, it needs to consider three conditions:
+		 * 1st, the disk is before start_disk, you can imagine the disk in
+		 * the next stripe. So the dev_start is the start address of next
+		 * stripe.
+		 * 2st, the disk is after start_disk, it means the disk is at the
+		 * same stripe of first disk
+		 * 3st, the first disk itself, we can use start_disk_offset directly
+		 */
+		if (disk < start_disk_index)
+			dev_start = (first_stripe_index + 1) * mddev->chunk_sectors;
+		else if (disk > start_disk_index)
+			dev_start = first_stripe_index * mddev->chunk_sectors;
+		else
+			dev_start = start_disk_offset;
+
+		if (disk < end_disk_index)
+			dev_end = (last_stripe_index + 1) * mddev->chunk_sectors;
+		else if (disk > end_disk_index)
+			dev_end = last_stripe_index * mddev->chunk_sectors;
+		else
+			dev_end = end_disk_offset;
+
+		/* It only handles discard bio which size is >= stripe size, so
+		 * dev_end > dev_start all the time
+		 */
+		if (r10_bio->devs[disk].bio) {
+			mbio = bio_clone_fast(bio, GFP_NOIO, &mddev->bio_set);
+			mbio->bi_end_io = raid10_end_discard_request;
+			mbio->bi_private = r10_bio;
+			r10_bio->devs[disk].bio = mbio;
+			r10_bio->devs[disk].devnum = disk;
+			atomic_inc(&r10_bio->remaining);
+			md_submit_discard_bio(mddev, rdev, mbio,
+					dev_start + choose_data_offset(r10_bio, rdev),
+					dev_end - dev_start);
+			bio_endio(mbio);
+		}
+		if (r10_bio->devs[disk].repl_bio) {
+			rbio = bio_clone_fast(bio, GFP_NOIO, &mddev->bio_set);
+			rbio->bi_end_io = raid10_end_discard_request;
+			rbio->bi_private = r10_bio;
+			r10_bio->devs[disk].repl_bio = rbio;
+			r10_bio->devs[disk].devnum = disk;
+			atomic_inc(&r10_bio->remaining);
+			md_submit_discard_bio(mddev, rrdev, rbio,
+					dev_start + choose_data_offset(r10_bio, rrdev),
+					dev_end - dev_start);
+			bio_endio(rbio);
+		}
+	}
+
+	if (atomic_dec_and_test(&r10_bio->remaining)) {
+		md_write_end(r10_bio->mddev);
+		raid_end_bio_io(r10_bio);
+	}
+
+	return 0;
+out:
+	allow_barrier(conf);
+	return -EAGAIN;
+}
+
 static bool raid10_make_request(struct mddev *mddev, struct bio *bio)
 {
 	struct r10conf *conf = mddev->private;
@@ -1529,6 +1785,10 @@ static bool raid10_make_request(struct mddev *mddev, struct bio *bio)
 	if (!md_write_start(mddev, bio))
 		return false;
 
+	if (unlikely(bio_op(bio) == REQ_OP_DISCARD))
+		if (!raid10_handle_discard(mddev, bio))
+			return true;
+
 	/*
 	 * If this request crosses a chunk boundary, we need to split
 	 * it.
@@ -3768,7 +4028,7 @@ static int raid10_run(struct mddev *mddev)
 
 	if (mddev->queue) {
 		blk_queue_max_discard_sectors(mddev->queue,
-					      mddev->chunk_sectors);
+					      UINT_MAX);
 		blk_queue_max_write_same_sectors(mddev->queue, 0);
 		blk_queue_max_write_zeroes_sectors(mddev->queue, 0);
 		blk_queue_io_min(mddev->queue, mddev->chunk_sectors << 9);
-- 
2.7.5


  parent reply index

Thread overview: 16+ messages / expand[flat|nested]  mbox.gz  Atom feed  top
2021-02-03 13:45 [PATCH 0/5] md/raid10: Improve handling " Xiao Ni
2021-02-03 13:45 ` [PATCH 1/5] md: add md_submit_discard_bio() for submitting discard bio Xiao Ni
2021-02-03 15:44   ` Christoph Hellwig
2021-02-04  1:51     ` Xiao Ni
2021-02-04  4:07     ` Xiao Ni
2021-02-04  7:19       ` Christoph Hellwig
2021-02-04  7:37         ` Xiao Ni
2021-02-03 13:45 ` [PATCH 2/5] md/raid10: extend r10bio devs to raid disks Xiao Ni
2021-02-03 15:44   ` Christoph Hellwig
2021-02-03 13:45 ` [PATCH 3/5] md/raid10: pull codes that wait for blocked dev into one function Xiao Ni
2021-02-03 15:46   ` Christoph Hellwig
2021-02-04  1:54     ` Xiao Ni
2021-02-03 13:45 ` Xiao Ni [this message]
2021-02-03 15:50   ` [PATCH 4/5] md/raid10: improve raid10 discard request Christoph Hellwig
2021-02-04  2:00     ` Xiao Ni
2021-02-03 13:45 ` [PATCH 5/5] md/raid10: improve discard request for far layout Xiao Ni

Reply instructions:

You may reply publicly to this message via plain-text email
using any one of the following methods:

* Save the following mbox file, import it into your mail client,
  and reply-to-all from there: mbox

  Avoid top-posting and favor interleaved quoting:
  https://en.wikipedia.org/wiki/Posting_style#Interleaved_style

* Reply using the --to, --cc, and --in-reply-to
  switches of git-send-email(1):

  git send-email \
    --in-reply-to=1612359931-24209-5-git-send-email-xni@redhat.com \
    --to=xni@redhat.com \
    --cc=colyli@suse.de \
    --cc=guoqing.jiang@cloud.ionos.com \
    --cc=linux-raid@vger.kernel.org \
    --cc=matthew.ruffell@canonical.com \
    --cc=ncroxon@redhat.com \
    --cc=songliubraving@fb.com \
    /path/to/YOUR_REPLY

  https://kernel.org/pub/software/scm/git/docs/git-send-email.html

* If your mail client supports setting the In-Reply-To header
  via mailto: links, try the mailto: link

Linux-Raid Archives on lore.kernel.org

Archives are clonable:
	git clone --mirror https://lore.kernel.org/linux-raid/0 linux-raid/git/0.git

	# If you have public-inbox 1.1+ installed, you may
	# initialize and index your mirror using the following commands:
	public-inbox-init -V2 linux-raid linux-raid/ https://lore.kernel.org/linux-raid \
		linux-raid@vger.kernel.org
	public-inbox-index linux-raid

Example config snippet for mirrors

Newsgroup available over NNTP:
	nntp://nntp.lore.kernel.org/org.kernel.vger.linux-raid


AGPL code for this site: git clone https://public-inbox.org/public-inbox.git