Skip to contentSkip to conversation

btrfs: Limit size of bios submitted from writeback

PATCH In review Jan Kara 3 participants 6 messages Apr 23 – Apr 28 Btrfs file system

lore
Share & tools
lore

Jan Kara

author
#0 lore raw

PATCH v2 1/1 btrfs: Limit size of bios submitted from writeback

Currently btrfs_writepages() just accumulates as large bio as possible (within writeback_control constraints) and then submits it. This can however lead to significant latency in writeback IO submission (I have observed tens of miliseconds) because the submitted bio easily has over hundred of megabytes. Consequently this leads to IO pipeline stalls and reduced throughput.

At the same time beyond certain size submitting so large bio provides diminishing returns because the bio is split by the block layer immediately anyway. So compute (estimate of) bio size beyond which we are unlikely to improve performance and just submit the bio for writeback once we accumulate that much to keep the IO pipeline busy. This improves writeback throughput for sequential writes by about 15% on the test machine I was using.

---
 fs/btrfs/disk-io.c   |  7 +++++++
 fs/btrfs/extent_io.c | 10 ++++++++++
 fs/btrfs/fs.h        |  1 +
 fs/btrfs/volumes.c   | 29 +++++++++++++++++++++++++++++
 fs/btrfs/volumes.h   |  1 +
 5 files changed, 48 insertions(+)

Changes since v1:
- moved limit checks to submit_extent_folio
- simplified computation of maximum bio size

fs/btrfs/disk-io.c +7 -0
index 8a11be02eeb9..f063595d0cee 100644
@@ -3591,6 +3591,13 @@ int __cold open_ctree(struct super_block *sb, struct btrfs_fs_devices *fs_device
 		}
 	}
 
+	ret = btrfs_init_writeback_bio_size(fs_info);
+	if (ret) {
+		btrfs_err(fs_info, "failed to get optimum writeback size: %d",
+			  ret);
+		goto fail_sysfs;
+	}
+
 	btrfs_free_zone_cache(fs_info);
 
 	btrfs_check_active_zone_reservation(fs_info);
fs/btrfs/extent_io.c +10 -0
index ca3e4b99aec2..d13d7eb95d44 100644
@@ -857,6 +857,16 @@ static void submit_extent_folio(struct btrfs_bio_ctrl *bio_ctrl,
 		/* Ordered extent boundary: move on to a new bio. */
 		if (bio_ctrl->len_to_oe_boundary == 0)
 			submit_one_bio(bio_ctrl);
+		/*
+		 * If we have accumulated decent amount of IO, send it to the
+		 * block layer so that IO can run while we are accumulating
+		 * more folios to write.
+		 */
+		else if (bio_ctrl->wbc &&
+			 bio_ctrl->bbio->bio.bi_iter.bi_size >=
+			    inode->root->fs_info->writeback_bio_size)
+			submit_one_bio(bio_ctrl);
+
 	} while (size);
 }
 
fs/btrfs/fs.h +1 -0
index a4758d94b32e..19e02452ab96 100644
@@ -880,6 +880,7 @@ struct btrfs_fs_info {
 	u32 block_min_order;
 	u32 block_max_order;
 	u32 stripesize;
+	u32 writeback_bio_size;
 	u32 csum_size;
 	u32 csums_per_leaf;
 	u32 csum_type;
fs/btrfs/volumes.c +29 -0
index a88e68f90564..c27614a23ffb 100644
@@ -8179,6 +8179,35 @@ int btrfs_init_dev_stats(struct btrfs_fs_info *fs_info)
 	return ret;
 }
 
+int btrfs_init_writeback_bio_size(struct btrfs_fs_info *fs_info)
+{
+	struct btrfs_fs_devices *fs_devices = fs_info->fs_devices;
+	struct btrfs_device *device;
+	u32 writeback_bio_size = fs_info->sectorsize;
+
+	mutex_lock(&fs_devices->device_list_mutex);
+	/*
+	 * Let's take maximum over optimal request sizes for all devices. For
+	 * RAID profiles writeback will submit stripe (64k) sized bios anyway
+	 * so our value doesn't matter and for simple profiles this is a good
+	 * approximation of sensible IO chunking.
+	 */
+	list_for_each_entry(device, &fs_devices->devices, dev_list) {
+		struct request_queue *queue;
+		unsigned int io_opt;
+
+		queue = bdev_get_queue(device->bdev);
+		io_opt = queue_io_opt(queue) ? :
+				queue_max_sectors(queue) << SECTOR_SHIFT;
+		writeback_bio_size = max(writeback_bio_size, io_opt);
+	}
+	mutex_unlock(&fs_devices->device_list_mutex);
+
+	fs_info->writeback_bio_size = writeback_bio_size;
+
+	return 0;
+}
+
 static int update_dev_stat_item(struct btrfs_trans_handle *trans,
 				struct btrfs_device *device)
 {
fs/btrfs/volumes.h +1 -0
index 0082c166af91..96904d18f686 100644
@@ -784,6 +784,7 @@ int btrfs_get_dev_stats(struct btrfs_fs_info *fs_info,
 			struct btrfs_ioctl_get_dev_stats *stats);
 int btrfs_init_devices_late(struct btrfs_fs_info *fs_info);
 int btrfs_init_dev_stats(struct btrfs_fs_info *fs_info);
+int btrfs_init_writeback_bio_size(struct btrfs_fs_info *fs_info);
 int btrfs_run_dev_stats(struct btrfs_trans_handle *trans);
 void btrfs_rm_dev_replace_remove_srcdev(struct btrfs_device *srcdev);
 void btrfs_rm_dev_replace_free_srcdev(struct btrfs_device *srcdev);
Signature · 1 line
2.51.0
Reviewed-by on v2 1/1
#1 lore raw
Quoted 15 lines · Jan Kara

Currently btrfs_writepages() just accumulates as large bio as possible (within writeback_control constraints) and then submits it. This can however lead to significant latency in writeback IO submission (I have observed tens of miliseconds) because the submitted bio easily has over hundred of megabytes. Consequently this leads to IO pipeline stalls and reduced throughput.

At the same time beyond certain size submitting so large bio provides diminishing returns because the bio is split by the block layer immediately anyway. So compute (estimate of) bio size beyond which we are unlikely to improve performance and just submit the bio for writeback once we accumulate that much to keep the IO pipeline busy. This improves writeback throughput for sequential writes by about 15% on the test machine I was using.

Quoted 1 line

Looks great!

Just one minor question inlined below.

[...]

Quoted 15 lines
fs/btrfs/volumes.c +15 -0
+int btrfs_init_writeback_bio_size(struct btrfs_fs_info *fs_info)
+{
+	struct btrfs_fs_devices *fs_devices = fs_info->fs_devices;
+	struct btrfs_device *device;
+	u32 writeback_bio_size = fs_info->sectorsize;
+
+	mutex_lock(&fs_devices->device_list_mutex);
+	/*
+	 * Let's take maximum over optimal request sizes for all devices. For
+	 * RAID profiles writeback will submit stripe (64k) sized bios anyway
+	 * so our value doesn't matter and for simple profiles this is a good
+	 * approximation of sensible IO chunking.
+	 */
+	list_for_each_entry(device, &fs_devices->devices, dev_list) {
+		struct request_queue *queue;

fs/btrfs/volumes.c int btrfs_init_dev_stats(struct btrfs_fs_info *fs_info) · line 8206

Quoted 10 lines
fs/btrfs/volumes.c +10 -0
+		unsigned int io_opt;
+
+		queue = bdev_get_queue(device->bdev);
+		io_opt = queue_io_opt(queue) ? :
+				queue_max_sectors(queue) << SECTOR_SHIFT;
+		writeback_bio_size = max(writeback_bio_size, io_opt);
+	}
+	mutex_unlock(&fs_devices->device_list_mutex);
+
+	fs_info->writeback_bio_size = writeback_bio_size;

With this simplified version of optimal io size detection, do we want to hook dev add/removal/replace to update the calculation?

I guess in the real world, the added/removed/replaced disks should have all the same performance parameter for server usages, so no difference there.

And for personal/pro users, I doubt if the original performance problem is even noticeable for most end users.

So overall I'm fine either way.

Thanks,
Qu

author
#2 lore raw
Quoted 18 lines · Qu Wenruo
Quoted 15 lines · Jan Kara

Currently btrfs_writepages() just accumulates as large bio as possible (within writeback_control constraints) and then submits it. This can however lead to significant latency in writeback IO submission (I have observed tens of miliseconds) because the submitted bio easily has over hundred of megabytes. Consequently this leads to IO pipeline stalls and reduced throughput.

At the same time beyond certain size submitting so large bio provides diminishing returns because the bio is split by the block layer immediately anyway. So compute (estimate of) bio size beyond which we are unlikely to improve performance and just submit the bio for writeback once we accumulate that much to keep the IO pipeline busy. This improves writeback throughput for sequential writes by about 15% on the test machine I was using.

Quoted 7 lines
Quoted 1 line

Looks great!

Just one minor question inlined below.

Thanks!

Quoted 29 lines
Quoted 15 lines
fs/btrfs/volumes.c +15 -0
+int btrfs_init_writeback_bio_size(struct btrfs_fs_info *fs_info)
+{
+	struct btrfs_fs_devices *fs_devices = fs_info->fs_devices;
+	struct btrfs_device *device;
+	u32 writeback_bio_size = fs_info->sectorsize;
+
+	mutex_lock(&fs_devices->device_list_mutex);
+	/*
+	 * Let's take maximum over optimal request sizes for all devices. For
+	 * RAID profiles writeback will submit stripe (64k) sized bios anyway
+	 * so our value doesn't matter and for simple profiles this is a good
+	 * approximation of sensible IO chunking.
+	 */
+	list_for_each_entry(device, &fs_devices->devices, dev_list) {
+		struct request_queue *queue;

fs/btrfs/volumes.c int btrfs_init_dev_stats(struct btrfs_fs_info *fs_info) · line 8206

Quoted 10 lines
fs/btrfs/volumes.c +10 -0
+		unsigned int io_opt;
+
+		queue = bdev_get_queue(device->bdev);
+		io_opt = queue_io_opt(queue) ? :
+				queue_max_sectors(queue) << SECTOR_SHIFT;
+		writeback_bio_size = max(writeback_bio_size, io_opt);
+	}
+	mutex_unlock(&fs_devices->device_list_mutex);
+
+	fs_info->writeback_bio_size = writeback_bio_size;

With this simplified version of optimal io size detection, do we want to hook dev add/removal/replace to update the calculation?

Quoted 7 lines

I guess in the real world, the added/removed/replaced disks should have all the same performance parameter for server usages, so no difference there.

And for personal/pro users, I doubt if the original performance problem is even noticeable for most end users.

So overall I'm fine either way.

Yeah, at this point I'm not sure the complexity is worth it. Normally added disks have very similar parameters as existing ones, also even if they are somewhat different, it will cost you a few percent of writeback speed at worst which doesn't seem too bad and it will "fix" itself on next mount.

Honza

Signature · 2 lines
Jan Kara <[email protected]>
SUSE Labs, CR
#3 lore raw
Quoted 64 lines · Jan Kara
Quoted 18 lines · Qu Wenruo
Quoted 15 lines · Jan Kara

Currently btrfs_writepages() just accumulates as large bio as possible (within writeback_control constraints) and then submits it. This can however lead to significant latency in writeback IO submission (I have observed tens of miliseconds) because the submitted bio easily has over hundred of megabytes. Consequently this leads to IO pipeline stalls and reduced throughput.

At the same time beyond certain size submitting so large bio provides diminishing returns because the bio is split by the block layer immediately anyway. So compute (estimate of) bio size beyond which we are unlikely to improve performance and just submit the bio for writeback once we accumulate that much to keep the IO pipeline busy. This improves writeback throughput for sequential writes by about 15% on the test machine I was using.

Quoted 7 lines
Quoted 1 line

Looks great!

Just one minor question inlined below.

Thanks!

Quoted 35 lines
Quoted 15 lines
fs/btrfs/volumes.c +15 -0
+int btrfs_init_writeback_bio_size(struct btrfs_fs_info *fs_info)
+{
+	struct btrfs_fs_devices *fs_devices = fs_info->fs_devices;
+	struct btrfs_device *device;
+	u32 writeback_bio_size = fs_info->sectorsize;
+
+	mutex_lock(&fs_devices->device_list_mutex);
+	/*
+	 * Let's take maximum over optimal request sizes for all devices. For
+	 * RAID profiles writeback will submit stripe (64k) sized bios anyway
+	 * so our value doesn't matter and for simple profiles this is a good
+	 * approximation of sensible IO chunking.
+	 */
+	list_for_each_entry(device, &fs_devices->devices, dev_list) {
+		struct request_queue *queue;

fs/btrfs/volumes.c int btrfs_init_dev_stats(struct btrfs_fs_info *fs_info) · line 8206

Quoted 10 lines
fs/btrfs/volumes.c +10 -0
+		unsigned int io_opt;
+
+		queue = bdev_get_queue(device->bdev);
+		io_opt = queue_io_opt(queue) ? :
+				queue_max_sectors(queue) << SECTOR_SHIFT;
+		writeback_bio_size = max(writeback_bio_size, io_opt);
+	}
+	mutex_unlock(&fs_devices->device_list_mutex);
+
+	fs_info->writeback_bio_size = writeback_bio_size;

With this simplified version of optimal io size detection, do we want to hook dev add/removal/replace to update the calculation?

I guess in the real world, the added/removed/replaced disks should have all the same performance parameter for server usages, so no difference there.

And for personal/pro users, I doubt if the original performance problem is even noticeable for most end users.

Quoted 6 lines
Quoted 1 line

So overall I'm fine either way.

Yeah, at this point I'm not sure the complexity is worth it. Normally added disks have very similar parameters as existing ones, also even if they are somewhat different, it will cost you a few percent of writeback speed at worst which doesn't seem too bad and it will "fix" itself on next mount.

Thanks, we're on the same page.

Now the patch is pushed to for-next branch, with one typo "miliseconds" fixed.

Thanks,
Qu

Quoted 2 lines

Honza

#4 lore raw

在 2026/4/27 19:20, Qu Wenruo 写道:
[...]

Quoted 11 lines
Quoted 4 lines

Yeah, at this point I'm not sure the complexity is worth it. Normally added disks have very similar parameters as existing ones, also even if they are somewhat different, it will cost you a few percent of writeback speed at worst which doesn't seem too bad and it will "fix" itself on next mount.

Thanks, we're on the same page.

Now the patch is pushed to for-next branch, with one typo "miliseconds" fixed.

Thanks,
Qu

Just a minor update.

The existing check is considering device->bdev always exists, but we can have missing devices which doesn't have device->bdev, and this will cause NULL pointer dereference during test cases like btrfs/027.

Fixed in for-next branch with the following small diff:

fs/btrfs/volumes.c +2 -0
index 8d36e8a9f0d9..93a923e4ecaf 100644
@@ -8224,6 +8224,8 @@ int btrfs_init_writeback_bio_size(struct btrfs_fs_info *fs_info)
                struct request_queue *queue;
                unsigned int io_opt;
 
+               if (!device->bdev || test_bit(BTRFS_DEV_STATE_MISSING, &device->dev_state))
+                       continue;
                queue = bdev_get_queue(device->bdev);
                io_opt = queue_io_opt(queue) ? :
                                queue_max_sectors(queue) << SECTOR_SHIFT;

Thanks,
Qu

Quoted 4 lines
Quoted 2 lines

                                Honza

author
#5 lore raw
Quoted 19 lines · Qu Wenruo

在 2026/4/27 19:20, Qu Wenruo 写道:
[...]

Quoted 7 lines
Quoted 6 lines

Yeah, at this point I'm not sure the complexity is worth it. Normally added
disks have very similar parameters as existing ones, also even if they are
somewhat different, it will cost you a few percent of writeback speed at worst which doesn't seem too bad and it will "fix" itself on next mount.

Quoted 7 lines

Thanks, we're on the same page.

Now the patch is pushed to for-next branch, with one typo "miliseconds" fixed.

Thanks,
Qu

Quoted 7 lines

Just a minor update.

The existing check is considering device->bdev always exists, but we can have missing devices which doesn't have device->bdev, and this will cause NULL pointer dereference during test cases like btrfs/027.

Fixed in for-next branch with the following small diff:

Thanks for the fixup!

Honza

Quoted 24 lines
fs/btrfs/volumes.c +2 -0
index 8d36e8a9f0d9..93a923e4ecaf 100644
@@ -8224,6 +8224,8 @@ int btrfs_init_writeback_bio_size(struct btrfs_fs_info
 *fs_info)
                struct request_queue *queue;
                unsigned int io_opt;
 
+               if (!device->bdev || test_bit(BTRFS_DEV_STATE_MISSING,
 &device->dev_state))
+                       continue;
                queue = bdev_get_queue(device->bdev);

io_opt = queue_io_opt(queue) ? :
queue_max_sectors(queue) << SECTOR_SHIFT;

Thanks,
Qu

Quoted 4 lines
Quoted 2 lines

                                Honza

Signature · 2 lines
Jan Kara <[email protected]>
SUSE Labs, CR