Skip to contentSkip to conversation

xfs/530: Do not pass block size argument to _scratch_mkfs

PATCH In review Chandan Babu R 3 participants 11 messages Jul 26, 2021 – Aug 1, 2021

lore
Share & tools
lore
#0 lore raw

PATCH 1/3 xfs/530: Do not pass block size argument to _scratch_mkfs

_scratch_do_mkfs constructs a mkfs command line by concatenating the values of
1. $mkfs_cmd
2. $MKFS_OPTIONS
3. $extra_mkfs_options

The block size argument passed by xfs/530 to _scratch_mkfs() will cause mkfs.xfs to fail if $MKFS_OPTIONS also has a block size specified. In such a case, _scratch_do_mkfs() will construct and invoke an mkfs command line without including the value of $MKFS_OPTIONS.

To prevent such silent failures, this commit removes the block size option that was being explicitly passed to _scratch_mkfs().

---
 tests/xfs/530 | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
tests/xfs/530 +1 -1
index 4d168ac5..16dc426c 100755
@@ -60,7 +60,7 @@ echo "Format and mount rt volume"
 
 export USE_EXTERNAL=yes
 export SCRATCH_RTDEV=$rtdev
-_scratch_mkfs -d size=$((1024 * 1024 * 1024)) -b size=${dbsize} \
+_scratch_mkfs -d size=$((1024 * 1024 * 1024)) \
 	      -r size=${rtextsz},extsize=${rtextsz} >> $seqres.full
 _try_scratch_mount || _notrun "Couldn't mount fs with synthetic rt volume"
 
Signature · 1 line
2.30.2
#1 lore raw

PATCH 2/3 common/xfs: Add helpers to obtain reflink/rmapbt status of a filesystem

This commit adds helpers to obtain status of reflink and rmapbt features of a filesystem. The status of these features are obtained by invoking $XFS_INFO_PROG program.

---
 common/xfs | 34 ++++++++++++++++++++++++++++++++++
 1 file changed, 34 insertions(+)
common/xfs +34 -0
index c5e39427..e9f84b56 100644
@@ -1099,6 +1099,40 @@ _xfs_mount_agcount()
 	$XFS_INFO_PROG "$1" | grep agcount= | sed -e 's/^.*agcount=\([0-9]*\),.*$/\1/g'
 }
 
+# Get reflink status of a filesystem
+_xfs_is_reflink_enabled()
+{
+	local status
+
+	$XFS_INFO_PROG "$1" | grep -q reflink >> $seqres.full
+	[[ $? != 0 ]] && return 1
+
+	status=$($XFS_INFO_PROG "$1" | grep reflink= | \
+			 sed -e 's/^.*reflink=\([0-1]\).*$/\1/g')
+	if [[ $status == 1 ]]; then
+		return 0
+	else
+		return 1
+	fi
+}
+
+# Get rmapbt status of a filesystem
+_xfs_is_rmapbt_enabled()
+{
+	local status
+
+	$XFS_INFO_PROG "$1" | grep -q rmapbt >> $seqres.full
+	[[ $? != 0 ]] && return 1
+
+	status=$($XFS_INFO_PROG "$1" | grep rmapbt= | \
+			 sed -e 's/^.*rmapbt=\([0-1]\).*$/\1/g')
+	if [[ $status == 1 ]]; then
+		return 0
+	else
+		return 1
+	fi
+}
+
 # Wipe the superblock of each XFS AGs
 _try_wipe_scratch_xfs()
 {
Signature · 1 line
2.30.2
#2 lore raw

PATCH 3/3 xfs/530: Bail out if either of reflink or rmapbt is enabled

In reply to Chandan Babu R

_scratch_do_mkfs constructs a mkfs command line by concatenating the values of
1. $mkfs_cmd
2. $MKFS_OPTIONS
3. $extra_mkfs_options

The corresponding mkfs command line fails if $MKFS_OPTIONS enables either reflink or rmapbt feature. The failure occurs because the test tries to create a filesystem with realtime device enabled. In such a case, _scratch_do_mkfs() will construct and invoke an mkfs command line without including the value of $MKFS_OPTIONS.

To prevent such silent failures, this commit causes the test to exit if it detects either reflink or rmapbt feature being enabled.

---
 tests/xfs/530 | 6 ++++++
 1 file changed, 6 insertions(+)
tests/xfs/530 +6 -0
index 16dc426c..669b061d 100755
@@ -39,6 +39,12 @@ _require_scratch_nocheck
 echo "* Test extending rt inodes"
 
 _scratch_mkfs | _filter_mkfs >> $seqres.full 2> $tmp.mkfs
+
+_xfs_is_reflink_enabled $SCRATCH_DEV && \
+	_notrun "Realtime device cannot be used when reflink feature is enabled"
+_xfs_is_rmapbt_enabled $SCRATCH_DEV && \
+	_notrun "Realtime device cannot be used when rmapbt feature is enabled"
+
 . $tmp.mkfs
 
 echo "Create fake rt volume"
Signature · 1 line
2.30.2
Reviewed-by on v1 1/3
#9 lore raw
In reply to Chandan Babu R
Quoted 5 lines · Chandan Babu R

_scratch_do_mkfs constructs a mkfs command line by concatenating the values of
1. $mkfs_cmd
2. $MKFS_OPTIONS
3. $extra_mkfs_options

Quoted 7 lines

The block size argument passed by xfs/530 to _scratch_mkfs() will cause mkfs.xfs to fail if $MKFS_OPTIONS also has a block size specified. In such a case, _scratch_do_mkfs() will construct and invoke an mkfs command line without including the value of $MKFS_OPTIONS.

To prevent such silent failures, this commit removes the block size option that was being explicitly passed to _scratch_mkfs().

Yes, that makes sense.

--D

Quoted 22 lines
---
 tests/xfs/530 | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
tests/xfs/530 +1 -1
index 4d168ac5..16dc426c 100755
@@ -60,7 +60,7 @@ echo "Format and mount rt volume"
 
 export USE_EXTERNAL=yes
 export SCRATCH_RTDEV=$rtdev
-_scratch_mkfs -d size=$((1024 * 1024 * 1024)) -b size=${dbsize} \
+_scratch_mkfs -d size=$((1024 * 1024 * 1024)) \
 	      -r size=${rtextsz},extsize=${rtextsz} >> $seqres.full
 _try_scratch_mount || _notrun "Couldn't mount fs with synthetic rt volume"
 
Signature · 1 line
2.30.2
#3 lore raw

Re: [PATCH 3/3] xfs/530: Bail out if either of reflink or rmapbt is enabled

In reply to Chandan Babu R
Quoted 5 lines · Chandan Babu R

_scratch_do_mkfs constructs a mkfs command line by concatenating the values of
1. $mkfs_cmd
2. $MKFS_OPTIONS
3. $extra_mkfs_options

Quoted 8 lines

The corresponding mkfs command line fails if $MKFS_OPTIONS enables either reflink or rmapbt feature. The failure occurs because the test tries to create a filesystem with realtime device enabled. In such a case, _scratch_do_mkfs() will construct and invoke an mkfs command line without including the value of $MKFS_OPTIONS.

To prevent such silent failures, this commit causes the test to exit if it detects either reflink or rmapbt feature being enabled.

Er, what combinations of mkfs.xfs and MKFS_OPTIONS cause this result? What kind of fs configuration comes out of that?

Eventually, the plan is to support rmap[1] and reflink[2] on the realtime device, at which point this will have to be torn out and a better solution found.

--D

[1] https://git.kernel.org/pub/scm/linux/kernel/git/djwong/xfs-linux.git/log/?h=realtime-rmap
[2] https://git.kernel.org/pub/scm/linux/kernel/git/djwong/xfs-linux.git/log/?h=realtime-reflink

Quoted 25 lines
---
 tests/xfs/530 | 6 ++++++
 1 file changed, 6 insertions(+)
tests/xfs/530 +6 -0
index 16dc426c..669b061d 100755
@@ -39,6 +39,12 @@ _require_scratch_nocheck
 echo "* Test extending rt inodes"
 
 _scratch_mkfs | _filter_mkfs >> $seqres.full 2> $tmp.mkfs
+
+_xfs_is_reflink_enabled $SCRATCH_DEV && \
+	_notrun "Realtime device cannot be used when reflink feature is enabled"
+_xfs_is_rmapbt_enabled $SCRATCH_DEV && \
+	_notrun "Realtime device cannot be used when rmapbt feature is enabled"
+
 . $tmp.mkfs
 
 echo "Create fake rt volume"
Signature · 1 line
2.30.2
#4 lore raw

Re: [PATCH 3/3] xfs/530: Bail out if either of reflink or rmapbt is enabled

Quoted 12 lines · Darrick J. Wong
Quoted 11 lines · Chandan Babu R

_scratch_do_mkfs constructs a mkfs command line by concatenating the values of
1. $mkfs_cmd
2. $MKFS_OPTIONS
3. $extra_mkfs_options

The corresponding mkfs command line fails if $MKFS_OPTIONS enables either reflink or rmapbt feature. The failure occurs because the test tries to create a filesystem with realtime device enabled. In such a case, _scratch_do_mkfs() will construct and invoke an mkfs command line without including the value of $MKFS_OPTIONS.

Quoted 5 lines
Quoted 2 lines

To prevent such silent failures, this commit causes the test to exit if it detects either reflink or rmapbt feature being enabled.

Er, what combinations of mkfs.xfs and MKFS_OPTIONS cause this result? What kind of fs configuration comes out of that?

With MKFS_OPTIONS set as shown below,

export MKFS_OPTIONS="-m reflink=1 -b size=1k"

_scratch_do_mkfs() invokes mkfs.xfs with both realtime and reflink options enabled. Such an invocation of mkfs.xfs fails causing _scratch_do_mkfs() to ignore the contents of $MKFS_OPTIONS while constructing and invoking mkfs.xfs once again.

This time, the fs block size will however be set to 4k (the default block size). At the beginning of the test we would have obtained the block size of the filesystem as 1k and used it to compute the size of the realtime device required to overflow realtime bitmap inode's max pseudo extent count.

Invocation of xfs_growfs (made later in the test) ends up succeeding since a 4k fs block can accommodate more bits than a 1k fs block.

Quoted 10 lines

Eventually, the plan is to support rmap[1] and reflink[2] on the realtime device, at which point this will have to be torn out and a better solution found.

--D

[1] https://git.kernel.org/pub/scm/linux/kernel/git/djwong/xfs-linux.git/log/?h=realtime-rmap
[2] https://git.kernel.org/pub/scm/linux/kernel/git/djwong/xfs-linux.git/log/?h=realtime-reflink

--
chandan

#5 lore raw

Re: [PATCH 3/3] xfs/530: Bail out if either of reflink or rmapbt is enabled

Quoted 28 lines · Chandan Babu R
Quoted 12 lines · Darrick J. Wong
Quoted 11 lines · Chandan Babu R

_scratch_do_mkfs constructs a mkfs command line by concatenating the values of
1. $mkfs_cmd
2. $MKFS_OPTIONS
3. $extra_mkfs_options

The corresponding mkfs command line fails if $MKFS_OPTIONS enables either reflink or rmapbt feature. The failure occurs because the test tries to create a filesystem with realtime device enabled. In such a case, _scratch_do_mkfs() will construct and invoke an mkfs command line without including the value of $MKFS_OPTIONS.

Quoted 5 lines
Quoted 2 lines

To prevent such silent failures, this commit causes the test to exit if it detects either reflink or rmapbt feature being enabled.

Er, what combinations of mkfs.xfs and MKFS_OPTIONS cause this result? What kind of fs configuration comes out of that?

With MKFS_OPTIONS set as shown below,

export MKFS_OPTIONS="-m reflink=1 -b size=1k"

_scratch_do_mkfs() invokes mkfs.xfs with both realtime and reflink options enabled. Such an invocation of mkfs.xfs fails causing _scratch_do_mkfs() to ignore the contents of $MKFS_OPTIONS while constructing and invoking mkfs.xfs once again.

Quoted 7 lines

This time, the fs block size will however be set to 4k (the default block size). At the beginning of the test we would have obtained the block size of the filesystem as 1k and used it to compute the size of the realtime device required to overflow realtime bitmap inode's max pseudo extent count.

Invocation of xfs_growfs (made later in the test) ends up succeeding since a 4k fs block can accommodate more bits than a 1k fs block.

OK, now I think I've finally put all the pieces together. Both of these patches are fixing weirdness when MKFS_OPTIONS="-m reflink=1 -b size=1k".

With current HEAD, we try to mkfs.xfs with double "-b size" arguments. That fails with 'option respecified', so fstests tries again without MKFS_OPTIONS, which means you don't get the filesystem that you want. If, say, MKFS_OPTIONS also contained bigtime=1, you won't get a bigtime filesystem.

So the first patch removes the double -bsize arguments. But you still have the problem that the reflink=1 in MKFS_OPTIONS still causes mkfs.xfs to fail (because we don't do rt and reflink yet), so fstests again drops MKFS_OPTIONS, and now you're testing the fs without a block size option at all. The test still regresses because the special rt geometry depends on the blocksize, and we didn't get all the geometry elements that we need to trip the growfs failure.

Does the following patch fix all that for you?

--D

tests/xfs/530 +13 -1
index 4d168ac5..9c6f44d7 100755
@@ -60,10 +60,22 @@ echo "Format and mount rt volume"
 
 export USE_EXTERNAL=yes
 export SCRATCH_RTDEV=$rtdev
-_scratch_mkfs -d size=$((1024 * 1024 * 1024)) -b size=${dbsize} \
+_scratch_mkfs -d size=$((1024 * 1024 * 1024)) \
 	      -r size=${rtextsz},extsize=${rtextsz} >> $seqres.full
 _try_scratch_mount || _notrun "Couldn't mount fs with synthetic rt volume"
 
+# If we didn't get the desired realtime volume and the same blocksize as the
+# first format (which we used to compute a specific rt geometry), skip the
+# test.  This can happen if the MKFS_OPTIONS conflict with the ones we passed
+# to _scratch_mkfs or do not result in a valid rt fs geometry.  In this case,
+# _scratch_mkfs will try to "succeed" at formatting by dropping MKFS_OPTIONS,
+# giving us the wrong geometry.
+formatted_blksz="$(_get_block_size $SCRATCH_MNT)"
+test "$formatted_blksz" -ne "$dbsize" && \
+	_notrun "Tried to format with $dbsize blocksize, got $formatted_blksz."
+$XFS_INFO_PROG $SCRATCH_MNT | egrep -q 'realtime.*blocks=0' && \
+	_notrun "Filesystem should have a realtime volume"
+
 echo "Consume free space"
 fillerdir=$SCRATCH_MNT/fillerdir
 nr_free_blks=$(stat -f -c '%f' $SCRATCH_MNT)
Quoted 13 lines
Quoted 10 lines

Eventually, the plan is to support rmap[1] and reflink[2] on the realtime device, at which point this will have to be torn out and a better solution found.

--D

[1] https://git.kernel.org/pub/scm/linux/kernel/git/djwong/xfs-linux.git/log/?h=realtime-rmap
[2] https://git.kernel.org/pub/scm/linux/kernel/git/djwong/xfs-linux.git/log/?h=realtime-reflink

--
chandan

author Reviewed-by
#6 lore raw

Re: [PATCH 3/3] xfs/530: Bail out if either of reflink or rmapbt is enabled

Quoted 73 lines · Darrick J. Wong
Quoted 28 lines · Chandan Babu R
Quoted 12 lines · Darrick J. Wong
Quoted 11 lines · Chandan Babu R

_scratch_do_mkfs constructs a mkfs command line by concatenating the values of
1. $mkfs_cmd
2. $MKFS_OPTIONS
3. $extra_mkfs_options

The corresponding mkfs command line fails if $MKFS_OPTIONS enables either reflink or rmapbt feature. The failure occurs because the test tries to create a filesystem with realtime device enabled. In such a case, _scratch_do_mkfs() will construct and invoke an mkfs command line without including the value of $MKFS_OPTIONS.

Quoted 5 lines
Quoted 2 lines

To prevent such silent failures, this commit causes the test to exit if it detects either reflink or rmapbt feature being enabled.

Er, what combinations of mkfs.xfs and MKFS_OPTIONS cause this result? What kind of fs configuration comes out of that?

With MKFS_OPTIONS set as shown below,

export MKFS_OPTIONS="-m reflink=1 -b size=1k"

_scratch_do_mkfs() invokes mkfs.xfs with both realtime and reflink options enabled. Such an invocation of mkfs.xfs fails causing _scratch_do_mkfs() to ignore the contents of $MKFS_OPTIONS while constructing and invoking mkfs.xfs once again.

Quoted 7 lines

This time, the fs block size will however be set to 4k (the default block size). At the beginning of the test we would have obtained the block size of the filesystem as 1k and used it to compute the size of the realtime device required to overflow realtime bitmap inode's max pseudo extent count.

Invocation of xfs_growfs (made later in the test) ends up succeeding since a 4k fs block can accommodate more bits than a 1k fs block.

OK, now I think I've finally put all the pieces together. Both of these patches are fixing weirdness when MKFS_OPTIONS="-m reflink=1 -b size=1k".

With current HEAD, we try to mkfs.xfs with double "-b size" arguments. That fails with 'option respecified', so fstests tries again without MKFS_OPTIONS, which means you don't get the filesystem that you want. If, say, MKFS_OPTIONS also contained bigtime=1, you won't get a bigtime filesystem.

So the first patch removes the double -bsize arguments. But you still have the problem that the reflink=1 in MKFS_OPTIONS still causes mkfs.xfs to fail (because we don't do rt and reflink yet), so fstests again drops MKFS_OPTIONS, and now you're testing the fs without a block size option at all. The test still regresses because the special rt geometry depends on the blocksize, and we didn't get all the geometry elements that we need to trip the growfs failure.

Does the following patch fix all that for you?

--D

tests/xfs/530 +3 -1
index 4d168ac5..9c6f44d7 100755
@@ -60,10 +60,22 @@ echo "Format and mount rt volume"
 
 export USE_EXTERNAL=yes
 export SCRATCH_RTDEV=$rtdev
-_scratch_mkfs -d size=$((1024 * 1024 * 1024)) -b size=${dbsize} \
+_scratch_mkfs -d size=$((1024 * 1024 * 1024)) \
 	      -r size=${rtextsz},extsize=${rtextsz} >> $seqres.full
 _try_scratch_mount || _notrun "Couldn't mount fs with synthetic rt volume"
 
+# If we didn't get the desired realtime volume and the same blocksize as the
+# first format (which we used to compute a specific rt geometry), skip the

tests/xfs/530 echo "Format and mount rt volume" · line 78

Quoted 10 lines
tests/xfs/530 +10 -0
+# test.  This can happen if the MKFS_OPTIONS conflict with the ones we passed
+# to _scratch_mkfs or do not result in a valid rt fs geometry.  In this case,
+# _scratch_mkfs will try to "succeed" at formatting by dropping MKFS_OPTIONS,
+# giving us the wrong geometry.
+formatted_blksz="$(_get_block_size $SCRATCH_MNT)"
+test "$formatted_blksz" -ne "$dbsize" && \
+	_notrun "Tried to format with $dbsize blocksize, got $formatted_blksz."
+$XFS_INFO_PROG $SCRATCH_MNT | egrep -q 'realtime.*blocks=0' && \
+	_notrun "Filesystem should have a realtime volume"
+

The above solution indeed fixes the problems that patch 2 and 3 were attempting to solve. If you plan on posting it, you can add

Signature · 1 line
chandan
#10 lore raw
In reply to Darrick J. Wong
Quoted 6 lines · Darrick J. Wong
Quoted 5 lines · Chandan Babu R

_scratch_do_mkfs constructs a mkfs command line by concatenating the values of
1. $mkfs_cmd
2. $MKFS_OPTIONS
3. $extra_mkfs_options

Quoted 7 lines
Quoted 7 lines

The block size argument passed by xfs/530 to _scratch_mkfs() will cause mkfs.xfs to fail if $MKFS_OPTIONS also has a block size specified. In such a case, _scratch_do_mkfs() will construct and invoke an mkfs command line without including the value of $MKFS_OPTIONS.

To prevent such silent failures, this commit removes the block size option that was being explicitly passed to _scratch_mkfs().

Patch looks fine to me, and queued for update.

Quoted 4 lines

Yes, that makes sense.

I'm wondering if mkfs.xfs could restore the behavior that allows re-specified options, and the last specified wins, e.g.

mkfs -t xfs -b 1k -b 2k -f /dev/sda2

and mkfs.xfs didn't report failure but created xfs with 2k blocksize. This may cause less problems like this patch resolved.

Thanks,
Eryu

Quoted 25 lines

--D

Quoted 22 lines
---
 tests/xfs/530 | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
tests/xfs/530 +1 -1
index 4d168ac5..16dc426c 100755
@@ -60,7 +60,7 @@ echo "Format and mount rt volume"
 
 export USE_EXTERNAL=yes
 export SCRATCH_RTDEV=$rtdev
-_scratch_mkfs -d size=$((1024 * 1024 * 1024)) -b size=${dbsize} \
+_scratch_mkfs -d size=$((1024 * 1024 * 1024)) \
 	      -r size=${rtextsz},extsize=${rtextsz} >> $seqres.full
 _try_scratch_mount || _notrun "Couldn't mount fs with synthetic rt volume"
 
Signature · 1 line
2.30.2
#7 lore raw

Re: [PATCH 3/3] xfs/530: Bail out if either of reflink or rmapbt is enabled

In reply to Darrick J. Wong
Quoted 54 lines · Darrick J. Wong
Quoted 28 lines · Chandan Babu R
Quoted 12 lines · Darrick J. Wong
Quoted 11 lines · Chandan Babu R

_scratch_do_mkfs constructs a mkfs command line by concatenating the values of
1. $mkfs_cmd
2. $MKFS_OPTIONS
3. $extra_mkfs_options

The corresponding mkfs command line fails if $MKFS_OPTIONS enables either reflink or rmapbt feature. The failure occurs because the test tries to create a filesystem with realtime device enabled. In such a case, _scratch_do_mkfs() will construct and invoke an mkfs command line without including the value of $MKFS_OPTIONS.

Quoted 5 lines
Quoted 2 lines

To prevent such silent failures, this commit causes the test to exit if it detects either reflink or rmapbt feature being enabled.

Er, what combinations of mkfs.xfs and MKFS_OPTIONS cause this result? What kind of fs configuration comes out of that?

With MKFS_OPTIONS set as shown below,

export MKFS_OPTIONS="-m reflink=1 -b size=1k"

_scratch_do_mkfs() invokes mkfs.xfs with both realtime and reflink options enabled. Such an invocation of mkfs.xfs fails causing _scratch_do_mkfs() to ignore the contents of $MKFS_OPTIONS while constructing and invoking mkfs.xfs once again.

Quoted 7 lines

This time, the fs block size will however be set to 4k (the default block size). At the beginning of the test we would have obtained the block size of the filesystem as 1k and used it to compute the size of the realtime device required to overflow realtime bitmap inode's max pseudo extent count.

Invocation of xfs_growfs (made later in the test) ends up succeeding since a 4k fs block can accommodate more bits than a 1k fs block.

OK, now I think I've finally put all the pieces together. Both of these patches are fixing weirdness when MKFS_OPTIONS="-m reflink=1 -b size=1k".

With current HEAD, we try to mkfs.xfs with double "-b size" arguments. That fails with 'option respecified', so fstests tries again without MKFS_OPTIONS, which means you don't get the filesystem that you want. If, say, MKFS_OPTIONS also contained bigtime=1, you won't get a bigtime filesystem.

So the first patch removes the double -bsize arguments. But you still have the problem that the reflink=1 in MKFS_OPTIONS still causes mkfs.xfs to fail (because we don't do rt and reflink yet), so fstests again drops MKFS_OPTIONS, and now you're testing the fs without a block size option at all. The test still regresses because the special rt geometry depends on the blocksize, and we didn't get all the geometry elements that we need to trip the growfs failure.

Quoted 1 line

Does the following patch fix all that for you?

Do you have plan to post formal patch? I think both problems could be fixed in one patch like you did. I'll leave patch 1 for now.

Thanks,
Eryu

Quoted 45 lines

--D

tests/xfs/530 +13 -1
index 4d168ac5..9c6f44d7 100755
@@ -60,10 +60,22 @@ echo "Format and mount rt volume"
 
 export USE_EXTERNAL=yes
 export SCRATCH_RTDEV=$rtdev
-_scratch_mkfs -d size=$((1024 * 1024 * 1024)) -b size=${dbsize} \
+_scratch_mkfs -d size=$((1024 * 1024 * 1024)) \
 	      -r size=${rtextsz},extsize=${rtextsz} >> $seqres.full
 _try_scratch_mount || _notrun "Couldn't mount fs with synthetic rt volume"
 
+# If we didn't get the desired realtime volume and the same blocksize as the
+# first format (which we used to compute a specific rt geometry), skip the
+# test.  This can happen if the MKFS_OPTIONS conflict with the ones we passed
+# to _scratch_mkfs or do not result in a valid rt fs geometry.  In this case,
+# _scratch_mkfs will try to "succeed" at formatting by dropping MKFS_OPTIONS,
+# giving us the wrong geometry.
+formatted_blksz="$(_get_block_size $SCRATCH_MNT)"
+test "$formatted_blksz" -ne "$dbsize" && \
+	_notrun "Tried to format with $dbsize blocksize, got $formatted_blksz."
+$XFS_INFO_PROG $SCRATCH_MNT | egrep -q 'realtime.*blocks=0' && \
+	_notrun "Filesystem should have a realtime volume"
+
 echo "Consume free space"
 fillerdir=$SCRATCH_MNT/fillerdir
 nr_free_blks=$(stat -f -c '%f' $SCRATCH_MNT)
Quoted 14 lines
Quoted 11 lines

Eventually, the plan is to support rmap[1] and reflink[2] on the realtime device, at which point this will have to be torn out and a better solution found.

--D

[1] https://git.kernel.org/pub/scm/linux/kernel/git/djwong/xfs-linux.git/log/?h=realtime-rmap
[2] https://git.kernel.org/pub/scm/linux/kernel/git/djwong/xfs-linux.git/log/?h=realtime-reflink

--
chandan

#8 lore raw

Re: [PATCH 3/3] xfs/530: Bail out if either of reflink or rmapbt is enabled

Quoted 55 lines · Eryu Guan
Quoted 54 lines · Darrick J. Wong
Quoted 28 lines · Chandan Babu R
Quoted 12 lines · Darrick J. Wong
Quoted 11 lines · Chandan Babu R

_scratch_do_mkfs constructs a mkfs command line by concatenating the values of
1. $mkfs_cmd
2. $MKFS_OPTIONS
3. $extra_mkfs_options

The corresponding mkfs command line fails if $MKFS_OPTIONS enables either reflink or rmapbt feature. The failure occurs because the test tries to create a filesystem with realtime device enabled. In such a case, _scratch_do_mkfs() will construct and invoke an mkfs command line without including the value of $MKFS_OPTIONS.

Quoted 5 lines
Quoted 2 lines

To prevent such silent failures, this commit causes the test to exit if it detects either reflink or rmapbt feature being enabled.

Er, what combinations of mkfs.xfs and MKFS_OPTIONS cause this result? What kind of fs configuration comes out of that?

With MKFS_OPTIONS set as shown below,

export MKFS_OPTIONS="-m reflink=1 -b size=1k"

_scratch_do_mkfs() invokes mkfs.xfs with both realtime and reflink options enabled. Such an invocation of mkfs.xfs fails causing _scratch_do_mkfs() to ignore the contents of $MKFS_OPTIONS while constructing and invoking mkfs.xfs once again.

Quoted 7 lines

This time, the fs block size will however be set to 4k (the default block size). At the beginning of the test we would have obtained the block size of the filesystem as 1k and used it to compute the size of the realtime device required to overflow realtime bitmap inode's max pseudo extent count.

Invocation of xfs_growfs (made later in the test) ends up succeeding since a 4k fs block can accommodate more bits than a 1k fs block.

OK, now I think I've finally put all the pieces together. Both of these patches are fixing weirdness when MKFS_OPTIONS="-m reflink=1 -b size=1k".

With current HEAD, we try to mkfs.xfs with double "-b size" arguments. That fails with 'option respecified', so fstests tries again without MKFS_OPTIONS, which means you don't get the filesystem that you want. If, say, MKFS_OPTIONS also contained bigtime=1, you won't get a bigtime filesystem.

So the first patch removes the double -bsize arguments. But you still have the problem that the reflink=1 in MKFS_OPTIONS still causes mkfs.xfs to fail (because we don't do rt and reflink yet), so fstests again drops MKFS_OPTIONS, and now you're testing the fs without a block size option at all. The test still regresses because the special rt geometry depends on the blocksize, and we didn't get all the geometry elements that we need to trip the growfs failure.

Quoted 4 lines
Quoted 1 line

Does the following patch fix all that for you?

Do you have plan to post formal patch? I think both problems could be fixed in one patch like you did. I'll leave patch 1 for now.

Ah, I saw the patch now, thanks!

Eryu