common/rc: use directio mode for the loop device when possible

Recently, I've been observing very high runtimes of tests that
format a filesystem atop a loop device and write enough data to fill
memory, such as generic/590 and generic/361.  Logging into the test
VMs, I noticed that the writes to the file on the upper filesystem
started fast, but soon slowed down to about 500KB/s and stayed that
way for nearly 20 minutes.  Looking through the D-state processes on
the system revealed:

/proc/4350/comm = xfs_io
/proc/4350/stack : [<0>] balance_dirty_pages+0x332/0xda0
[<0>] balance_dirty_pages_ratelimited+0x304/0x400
[<0>] iomap_file_buffered_write+0x1ab/0x260
[<0>] xfs_file_buffered_write+0xba/0x330 [xfs]
[<0>] new_sync_write+0x119/0x1a0
[<0>] vfs_write+0x274/0x310
[<0>] __x64_sys_pwrite64+0x89/0xc0
[<0>] do_syscall_64+0x35/0x80
[<0>] entry_SYSCALL_64_after_hwframe+0x44/0xae

Here's the xfs_io process performing a buffered write to the file on the
upper filesystem, which at this point has dirtied enough pages to be
ratelimited.

/proc/28/comm = u10:0+flush-8:80
/proc/28/stack : [<0>] blk_mq_get_tag+0x11c/0x280
[<0>] __blk_mq_alloc_request+0xce/0xf0
[<0>] blk_mq_submit_bio+0x139/0x5b0
[<0>] submit_bio_noacct+0x3ba/0x430
[<0>] iomap_submit_ioend+0x4b/0x70
[<0>] xfs_vm_writepages+0x86/0x170 [xfs]
[<0>] do_writepages+0xcc/0x200
[<0>] __writeback_single_inode+0x3d/0x300
[<0>] writeback_sb_inodes+0x207/0x4a0
[<0>] __writeback_inodes_wb+0x4c/0xe0
[<0>] wb_writeback+0x1da/0x2c0
[<0>] wb_workfn+0x2ad/0x4f0
[<0>] process_one_work+0x1e2/0x3d0
[<0>] worker_thread+0x53/0x3c0
[<0>] kthread+0x149/0x170
[<0>] ret_from_fork+0x1f/0x30

This is a flusher thread that has invoked writeback on the upper
filesystem to try to clean memory pages.

/proc/89/comm = u10:7+loop0
/proc/89/stack : [<0>] balance_dirty_pages+0x332/0xda0
[<0>] balance_dirty_pages_ratelimited+0x304/0x400
[<0>] iomap_file_buffered_write+0x1ab/0x260
[<0>] xfs_file_buffered_write+0xba/0x330 [xfs]
[<0>] do_iter_readv_writev+0x14f/0x1a0
[<0>] do_iter_write+0x7b/0x1c0
[<0>] lo_write_bvec+0x62/0x1c0
[<0>] loop_process_work+0x3a4/0xba0
[<0>] process_one_work+0x1e2/0x3d0
[<0>] worker_thread+0x53/0x3c0
[<0>] kthread+0x149/0x170
[<0>] ret_from_fork+0x1f/0x30

Here's the loop device worker handling the writeback IO submitted by the
flusher thread.  Unfortunately, the loop device is using buffered write
mode, which means that /writeback/ is dirtying pages and being throttled
for that.  This is stupid.

Fix this by trying to enable "directio" mode on the loop device, which
delivers two performance benefits: setting directio mode also enables
async io mode, which will allow multiple IOs at once; and using directio
nearly eliminates the chance that writeback will get throttled.

On the author's system with fast storage, this reduces the runtime of
g/590 from 20 minutes to 12 seconds, and g/361 from ~30s to ~3s.

Signed-off-by: Darrick J. Wong <djwong@kernel.org>
Reviewed-by: Eryu Guan <guaneryu@gmail.com>
Signed-off-by: Eryu Guan <guaneryu@gmail.com>

diff --git a/common/rc b/common/rc
index 275b1f2421f7b60a9543dd394cad7b036a531d40..a174b695d9b8d84057086c966f7ea50dd341b8a7 100644 (file)
--- a/common/rc
+++ b/common/rc
@@ -3849,6 +3849,14 @@ _create_loop_device()
 {
        local file=$1 dev
        dev=`losetup -f --show $file` || _fail "Cannot assign $file to a loop device"
+
+       # Try to enable asynchronous directio mode on the loopback device so
+       # that writeback started by a filesystem mounted on the loop device
+       # won't be throttled by buffered writes to the lower filesystem.  This
+       # is a performance optimization for tests that want to write a lot of
+       # data, so it isn't required to work.
+       test -b "$dev" && losetup --direct-io=on $dev 2> /dev/null
+
        echo $dev
 }