Skip to contentSkip to conversation

common/xfs: Execute _xfs_check only when explicitly asked

PATCH In review Chandan Rajendra 3 participants 3 messages Mar 30, 2020 – Apr 7, 2020

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 common/xfs: Execute _xfs_check only when explicitly asked

fsstress when executed as part of some of the tests (e.g. generic/270) invokes chown() syscall many times by passing random integers as value for the uid argument. For each such syscall invocation for which there is no on-disk quota block, xfs invokes xfs_dquot_disk_alloc() which allocates a new block and instantiates all the quota structures mapped by the newly allocated block. For filesystems with 64k block size, the number of on-disk quota structures created will be 16 times more than that for a filesystem with 4k block size.

xfs_db's check command (executed after test script finishes execution) will try to read in all of the on-disk quota structures into memory. This causes the OOM event to be triggered when reading from filesystems with 64k block size. For filesystems with sufficiently large amount of system memory, this causes the test to execute for a very long time.

Due to the above stated reasons, this commit disables execution of xfs_db's check command unless explictly asked by the user by setting $EXECUTE_XFS_DB_CHECK variable.

---
 README     |  2 ++
 common/xfs | 17 +++++++++++++----
 2 files changed, 15 insertions(+), 4 deletions(-)
README +2 -0
index 094a7742..c47569cd 100644
@@ -88,6 +88,8 @@ Preparing system for tests:
                run xfs_repair -n to check the filesystem; xfs_repair to rebuild
                metadata indexes; and xfs_repair -n (a third time) to check the
                results of the rebuilding.
+	     - set EXECUTE_XFS_DB_CHECK=1 to have _check_xfs_filesystem
+               run xfs_db's check command on the filesystem.
              - xfs_scrub, if present, will always check the test and scratch
                filesystems if they are still online at the end of the test.
                It is no longer necessary to set TEST_XFS_SCRUB.
common/xfs +13 -4
index d9a9784f..93ebab75 100644
@@ -455,10 +455,19 @@ _check_xfs_filesystem()
 		ok=0
 	fi
 
-	# xfs_check runs out of memory on large files, so even providing the test
-	# option (-t) to avoid indexing the free space trees doesn't make it pass on
-	# large filesystems. Avoid it.
-	if [ "$LARGE_SCRATCH_DEV" != yes ]; then
+	dbsize="$($XFS_INFO_PROG "${device}" | grep data.*bsize | sed -e 's/^.*bsize=//g' -e 's/\([0-9]*\).*$/\1/g')"
+
+	# xfs_check runs out of memory,
+	# 1. On large files. So even providing the test option (-t) to
+	# avoid indexing the free space trees doesn't make it pass on
+	# large filesystems.
+	# 2. When checking filesystems with large number of quota
+	# structures. This case happens consistently with 64k blocksize when
+	# creating large number of on-disk quota structures whose quota ids
+	# are spread across a large integer range.
+	#
+	# Hence avoid executing it unless explicitly asked by user.
+	if [ -n "$EXECUTE_XFS_DB_CHECK" -a "$LARGE_SCRATCH_DEV" != yes ]; then
 		_xfs_check $extra_log_options $device 2>&1 > $tmp.fs_check
 	fi
 	if [ -s $tmp.fs_check ]; then
Signature · 1 line
2.19.1
#1 lore raw
Quoted 31 lines · Chandan Rajendra

fsstress when executed as part of some of the tests (e.g. generic/270) invokes chown() syscall many times by passing random integers as value for the uid argument. For each such syscall invocation for which there is no on-disk quota block, xfs invokes xfs_dquot_disk_alloc() which allocates a new block and instantiates all the quota structures mapped by the newly allocated block. For filesystems with 64k block size, the number of on-disk quota structures created will be 16 times more than that for a filesystem with 4k block size.

xfs_db's check command (executed after test script finishes execution) will try to read in all of the on-disk quota structures into memory. This causes the OOM event to be triggered when reading from filesystems with 64k block size. For filesystems with sufficiently large amount of system memory, this causes the test to execute for a very long time.

Due to the above stated reasons, this commit disables execution of xfs_db's check command unless explictly asked by the user by setting $EXECUTE_XFS_DB_CHECK variable.

---
 README     |  2 ++
 common/xfs | 17 +++++++++++++----
 2 files changed, 15 insertions(+), 4 deletions(-)
README +0 -0
index 094a7742..c47569cd 100644

README Preparing system for tests: · line 92

Quoted 5 lines
README +2 -0
@@ -88,6 +88,8 @@ Preparing system for tests:
                run xfs_repair -n to check the filesystem; xfs_repair to rebuild
                metadata indexes; and xfs_repair -n (a third time) to check the
                results of the rebuilding.
+	     - set EXECUTE_XFS_DB_CHECK=1 to have _check_xfs_filesystem
+               run xfs_db's check command on the filesystem.

It seems spaces are used for indention instead of tab in README.

Quoted 8 lines
README +0 -0
              - xfs_scrub, if present, will always check the test and scratch
                filesystems if they are still online at the end of the test.
                It is no longer necessary to set TEST_XFS_SCRUB.
common/xfs +0 -0
index d9a9784f..93ebab75 100644

common/xfs _check_xfs_filesystem() · line 458

Quoted 8 lines
common/xfs +1 -4
@@ -455,10 +455,19 @@ _check_xfs_filesystem()
 		ok=0
 	fi
 
-	# xfs_check runs out of memory on large files, so even providing the test
-	# option (-t) to avoid indexing the free space trees doesn't make it pass on
-	# large filesystems. Avoid it.
-	if [ "$LARGE_SCRATCH_DEV" != yes ]; then
+	dbsize="$($XFS_INFO_PROG "${device}" | grep data.*bsize | sed -e 's/^.*bsize=//g' -e 's/\([0-9]*\).*$/\1/g')"

This is probably a left-over from v1 patch, which is not needed in v2.

common/xfs _check_xfs_filesystem() · line 471

Quoted 13 lines
common/xfs +12 -0
+
+	# xfs_check runs out of memory,
+	# 1. On large files. So even providing the test option (-t) to
+	# avoid indexing the free space trees doesn't make it pass on
+	# large filesystems.
+	# 2. When checking filesystems with large number of quota
+	# structures. This case happens consistently with 64k blocksize when
+	# creating large number of on-disk quota structures whose quota ids
+	# are spread across a large integer range.
+	#
+	# Hence avoid executing it unless explicitly asked by user.
+	if [ -n "$EXECUTE_XFS_DB_CHECK" -a "$LARGE_SCRATCH_DEV" != yes ]; then
 		_xfs_check $extra_log_options $device 2>&1 > $tmp.fs_check

Looks fine to me, I'd like to see xfs_check being disabled. But it'd be great if xfs folks could ack it as well.

Thanks,
Eryu

Quoted 5 lines
common/xfs +0 -0
 	fi
 	if [ -s $tmp.fs_check ]; then
Signature · 1 line
2.19.1
#2 lore raw
Quoted 59 lines · Eryu Guan
Quoted 31 lines · Chandan Rajendra

fsstress when executed as part of some of the tests (e.g. generic/270) invokes chown() syscall many times by passing random integers as value for the uid argument. For each such syscall invocation for which there is no on-disk quota block, xfs invokes xfs_dquot_disk_alloc() which allocates a new block and instantiates all the quota structures mapped by the newly allocated block. For filesystems with 64k block size, the number of on-disk quota structures created will be 16 times more than that for a filesystem with 4k block size.

xfs_db's check command (executed after test script finishes execution) will try to read in all of the on-disk quota structures into memory. This causes the OOM event to be triggered when reading from filesystems with 64k block size. For filesystems with sufficiently large amount of system memory, this causes the test to execute for a very long time.

Due to the above stated reasons, this commit disables execution of xfs_db's check command unless explictly asked by the user by setting $EXECUTE_XFS_DB_CHECK variable.

---
 README     |  2 ++
 common/xfs | 17 +++++++++++++----
 2 files changed, 15 insertions(+), 4 deletions(-)
README +0 -0
index 094a7742..c47569cd 100644

README Preparing system for tests: · line 92

Quoted 5 lines
README +2 -0
@@ -88,6 +88,8 @@ Preparing system for tests:
                run xfs_repair -n to check the filesystem; xfs_repair to rebuild
                metadata indexes; and xfs_repair -n (a third time) to check the
                results of the rebuilding.
+	     - set EXECUTE_XFS_DB_CHECK=1 to have _check_xfs_filesystem
+               run xfs_db's check command on the filesystem.

It seems spaces are used for indention instead of tab in README.

Quoted 8 lines
README +0 -0
              - xfs_scrub, if present, will always check the test and scratch
                filesystems if they are still online at the end of the test.
                It is no longer necessary to set TEST_XFS_SCRUB.
common/xfs +0 -0
index d9a9784f..93ebab75 100644

common/xfs _check_xfs_filesystem() · line 458

Quoted 8 lines
common/xfs +1 -4
@@ -455,10 +455,19 @@ _check_xfs_filesystem()
 		ok=0
 	fi
 
-	# xfs_check runs out of memory on large files, so even providing the test
-	# option (-t) to avoid indexing the free space trees doesn't make it pass on
-	# large filesystems. Avoid it.
-	if [ "$LARGE_SCRATCH_DEV" != yes ]; then
+	dbsize="$($XFS_INFO_PROG "${device}" | grep data.*bsize | sed -e 's/^.*bsize=//g' -e 's/\([0-9]*\).*$/\1/g')"

This is probably a left-over from v1 patch, which is not needed in v2.

Quoted 16 lines

common/xfs _check_xfs_filesystem() · line 471

Quoted 13 lines
common/xfs +12 -0
+
+	# xfs_check runs out of memory,
+	# 1. On large files. So even providing the test option (-t) to
+	# avoid indexing the free space trees doesn't make it pass on
+	# large filesystems.
+	# 2. When checking filesystems with large number of quota
+	# structures. This case happens consistently with 64k blocksize when
+	# creating large number of on-disk quota structures whose quota ids
+	# are spread across a large integer range.
+	#
+	# Hence avoid executing it unless explicitly asked by user.
+	if [ -n "$EXECUTE_XFS_DB_CHECK" -a "$LARGE_SCRATCH_DEV" != yes ]; then
 		_xfs_check $extra_log_options $device 2>&1 > $tmp.fs_check

Looks fine to me, I'd like to see xfs_check being disabled. But it'd be great if xfs folks could ack it as well.

I think I'd rather we just do the testing to see what things xfs_check catches that xfs_repair doesn't, and then we can deprecate running xfs_check in fstests (by default) entirely. In the long run it doesn't make sense to maintain /three/ separate metadata verification utilities.

--D

Quoted 8 lines

Thanks,
Eryu

Quoted 5 lines
common/xfs +0 -0
 	fi
 	if [ -s $tmp.fs_check ]; then
Signature · 1 line
2.19.1