| Message ID | 20190117125906.15479-1-jthumshirn@suse.de (mailing list archive) |
|---|---|
| State | New, archived |
| Headers | show |
| Series | [blktests,1/2] nvme: enable buffered_io for file-backed namespaces | expand |
This will force the user to run all tests in the buffered_io mode. What we really need is a config for the nvme category option which user can set and decide whether to run buffered_io or direct_io, you can set by default to the buffered_io mode. From: linux-block-owner@vger.kernel.org <linux-block-owner@vger.kernel.org> on behalf of Johannes Thumshirn <jthumshirn@suse.de> Sent: Thursday, January 17, 2019 4:59 AM To: Omar Sandoval Cc: Linux Block Layer Mailinglist; Linux NVMe Mailinglist; Anthony Iliopoulos; Johannes Thumshirn Subject: [PATCH blktests 1/2] nvme: enable buffered_io for file-backed namespaces All file-based NVMe over Fabrics tests fail in my test environment as I'm running these tests form within an initramfs. It turns out the file backing store of the NVMe target defaults to using direct I/O which is not available on tmpfs for obvious reasons. If the backing store of an nvme target is file-based enable the use of buffered I/O. Signed-off-by: Johannes Thumshirn <jthumshirn@suse.de> --- tests/nvme/rc | 7 +++++-- 1 file changed, 5 insertions(+), 2 deletions(-) diff --git a/tests/nvme/rc b/tests/nvme/rc index eff1dd992460..ec92e41396be 100644 --- a/tests/nvme/rc +++ b/tests/nvme/rc @@ -48,7 +48,7 @@ _remove_nvmet_port() { _create_nvmet_ns() { local nvmet_subsystem="$1" local nsid="$2" - local blkdev="$3" + local backstore="$3" local uuid="00000000-0000-0000-0000-000000000000" local subsys_path="${NVMET_CFS}/subsystems/${nvmet_subsystem}" local ns_path="${subsys_path}/namespaces/${nsid}" @@ -58,8 +58,11 @@ _create_nvmet_ns() { fi mkdir "${ns_path}" - printf "%s" "${blkdev}" > "${ns_path}/device_path" + printf "%s" "${backstore}" > "${ns_path}/device_path" printf "%s" "${uuid}" > "${ns_path}/device_uuid" + if [[ -f ${backstore} ]]; then + printf 1 > "${ns_path}/buffered_io" + fi printf 1 > "${ns_path}/enable" }
On 18/01/2019 00:46, Chaitanya Kulkarni wrote: > This will force the user to run all tests in the buffered_io mode. > > What we really need is a config for the nvme category option which user can set and decide > whether to run buffered_io or direct_io, you can set by default to the buffered_io mode. We could also introduce a check if we can use direct IO on the underlying FS, but the only way I can think of from a shell script is doing a 'dd of=$TMPDIR/testfile oflag=direct' and then populate the result to the helper functions. Chaitanya, Omar any ideas?
xfstest has the following nicely written routine may be we can re-use that?
from git://git.kernel.org/pub/scm/fs/xfs/xfstests-dev.git, ${XFSTEST_HOME}/common/rc :-
2185 # check that kernel and filesystem support direct I/O
2186 _require_odirect()
2187 {
2188 if [ $FSTYP = "ext4" ] ; then
2189 if echo "$MOUNT_OPTIONS" | grep -q "test_dummy_encryption"; then
2190 _notrun "ext4 encryption doesn't support O_DIRECT"
2191 elif echo "$MOUNT_OPTIONS" | grep -q "data=journal"; then
2192 _notrun "ext4 data journaling doesn't support O_DIRECT"
2193 fi
2194 fi
2195 local testfile=$TEST_DIR/$$.direct
2196 $XFS_IO_PROG -F -f -d -c "pwrite 0 20k" $testfile > /dev/null 2>&1
2197 if [ $? -ne 0 ]; then
2198 _notrun "O_DIRECT is not supported"
2199 fi
2200 rm -f $testfile 2>&1 > /dev/null
2201 }
2202
This will require to have xfs_io program.
From: Johannes Thumshirn <jthumshirn@suse.de>
Sent: Thursday, January 17, 2019 11:17 PM
To: Chaitanya Kulkarni; Omar Sandoval
Cc: Linux Block Layer Mailinglist; Linux NVMe Mailinglist; Anthony Iliopoulos
Subject: Re: [PATCH blktests 1/2] nvme: enable buffered_io for file-backed namespaces
On 18/01/2019 00:46, Chaitanya Kulkarni wrote:
> This will force the user to run all tests in the buffered_io mode.
>
> What we really need is a config for the nvme category option which user can set and decide
> whether to run buffered_io or direct_io, you can set by default to the buffered_io mode.
We could also introduce a check if we can use direct IO on the
underlying FS, but the only way I can think of from a shell script is
doing a 'dd of=$TMPDIR/testfile oflag=direct' and then populate the
result to the helper functions.
Chaitanya, Omar any ideas?
diff --git a/tests/nvme/rc b/tests/nvme/rc index eff1dd992460..ec92e41396be 100644 --- a/tests/nvme/rc +++ b/tests/nvme/rc @@ -48,7 +48,7 @@ _remove_nvmet_port() { _create_nvmet_ns() { local nvmet_subsystem="$1" local nsid="$2" - local blkdev="$3" + local backstore="$3" local uuid="00000000-0000-0000-0000-000000000000" local subsys_path="${NVMET_CFS}/subsystems/${nvmet_subsystem}" local ns_path="${subsys_path}/namespaces/${nsid}" @@ -58,8 +58,11 @@ _create_nvmet_ns() { fi mkdir "${ns_path}" - printf "%s" "${blkdev}" > "${ns_path}/device_path" + printf "%s" "${backstore}" > "${ns_path}/device_path" printf "%s" "${uuid}" > "${ns_path}/device_uuid" + if [[ -f ${backstore} ]]; then + printf 1 > "${ns_path}/buffered_io" + fi printf 1 > "${ns_path}/enable" }
All file-based NVMe over Fabrics tests fail in my test environment as I'm running these tests form within an initramfs. It turns out the file backing store of the NVMe target defaults to using direct I/O which is not available on tmpfs for obvious reasons. If the backing store of an nvme target is file-based enable the use of buffered I/O. Signed-off-by: Johannes Thumshirn <jthumshirn@suse.de> --- tests/nvme/rc | 7 +++++-- 1 file changed, 5 insertions(+), 2 deletions(-)