This patch adds case to test fsync and fdatasync with power-cuts.
The rule to check is:
1) fsync should guarantee all the inode metadata after power-cut,
2) fdatasync should guarantee i_size and i_blocks at least after
power-cut.
Note that, in XFS, it allocates more blocks when doing writes, so it
should close the file before fsync in order to get actual i_blocks
after power-cut. Or, it needs to do truncate the file with a
specific size to turn it off at the beginning.
Suggested-by: Chao Yu <yuchao0@huawei.com>
Signed-off-by: Jaegeuk Kim <jaegeuk@kernel.org>
Reviewed-by: Eryu Guan <eguan@redhat.com>
Signed-off-by: Eryu Guan <eguan@redhat.com>
There have been a couple of logic bugs in `btrfs_get_extent()` which
could lead to spurious -EEXIST errors from read or write. This test
exercises those conditions by having two threads race to add an
extent to the extent map.
This is fixed by Linux commit 8dff9c853410 ("Btrfs: deal with
duplciates during extent_map insertion in btrfs_get_extent") and the
patch "Btrfs: deal with existing encompassing extent map in
btrfs_get_extent()"
(http://marc.info/?l=linux-btrfs&m=147873402311143&w=2).
Although the bug is Btrfs-specific, nothing about the test is.
Signed-off-by: Omar Sandoval <osandov@fb.com>
Reviewed-by: Eryu Guan <eguan@redhat.com>
Signed-off-by: Eryu Guan <eguan@redhat.com>
This updates generic/098 by adding a sync option, i.e. 'sync' after
the second write, and with btrfs's NO_HOLES, we could still get
wrong isize after remount.
This gets fixed by the patch
'Btrfs: fix truncate down when no_holes feature is enabled'
Signed-off-by: Liu Bo <bo.li.liu@oracle.com>
Reviewed-by: Eryu Guan <eguan@redhat.com>
Signed-off-by: Eryu Guan <eguan@redhat.com>
Some tests use "tmp/$$" as $tmp, and this would result in ENOENT
failure if $tmp is referenced in helpers or tests, because we have
no "tmp" dir in pwd.
Fix them to use "/tmp/$$" as all other tests do.
Reviewed-by: Dave Chinner <dchinner@redhat.com>
Signed-off-by: Eryu Guan <eguan@redhat.com>
generic/38[3-6] are made generic from xfs tests to test project
quota on other filesystems, e.g. ext4. But we should make sure
project quota is enabled in between of _scratch_enable_pquota and
_qmount. Otherwise mount will fail if there's no project quota
support on SCRATCH_DEV, e.g. ext2 and ext3.
Reviewed-by: Dave Chinner <dchinner@redhat.com>
Signed-off-by: Eryu Guan <eguan@redhat.com>
generic/373 and generic/374 test that cross-mountpoint reflink/dedupe
fail with EXDEV.
For this test, _scratch_mount is called a 2nd time with same mount
arguments, but a different value of SCRATCH_MNT.
This practice is incorrect for overlayfs, which should not be using
the same upper/work dirs on two different overlay super blocks.
Change the test to use bind mount instead, which results in the
expected setup for both block device mount and overlay mount.
Signed-off-by: Amir Goldstein <amir73il@gmail.com>
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Eryu Guan <eguan@redhat.com>
This is a testcase for a bug which goes way back; googling
"xfs_trans_log_inode NULL pointer dereference" yields sporadic
reports over several years.
The test sets up several two-extent files with speculative
preallocation on them, and then runs xfs_fsr. The kernelside
code ignores the preallocation, and therefore sets up the
temporary inode incorrectly after the inode fork swap.
It is a "dangerous" test because the extent mishandling on
the temporary inode causes a null pointer dereference and
oops when the inode's i_itemp pointer gets overwritten
and we blow up in logging code that tries to use it.
Signed-off-by: Eric Sandeen <sandeen@redhat.com>
Reviewed-by: Eryu Guan <eguan@redhat.com>
Signed-off-by: Eryu Guan <eguan@redhat.com>
Starting from xfsprogs commit 384283555871 ("xfs_db: print one array
element per line"), xfs_db prints one array element per line. This
breaks the filter in xfs/021, which now fails as:
hdr.freemap[0-2] = [base,size] [FREEMAP..]
+0:[104,1892]
+1:[0,0]
+2:[0,0]
entries[0-2] = [hashval,nameidx,incomplete,root,local] [ENTRIES..]
So we have extra lines that need to be filtered out,
Reviewed-by: Zorro Lang <zlang@redhat.com>
Signed-off-by: Eryu Guan <eguan@redhat.com>
Due to the fact that btrfs uses different max extent size for
compressed and non-compressed write, it calculates metadata
reservation incorrectly.
This can leads to false ENOSPC alert for compressed write.
This test case will check it by using small fs and do parallel write
to trigger it.
It's highly recommened to use MOUNT_OPTIONS="-o compress" and
MKFS_OPTIONS="-n 64k" to trigger it. Without compression, it won't
be triggered.
The fix is not merged and may change in the future, but the function
is good enough:
btrfs: improve inode's outstanding_extents computation
btrfs: fix false enospc for compression
Signed-off-by: Qu Wenruo <quwenruo@cn.fujitsu.com>
Reviewed-by: Eryu Guan <eguan@redhat.com>
Signed-off-by: Eryu Guan <eguan@redhat.com>
Before xfsprogs commit a872b62 (xfs_copy: band-aids for CRC
filesystems), xfs_copy requires the "-d" option to copy a V5 XFS,
because it can't rewrite the UUID of V5 XFS properly.
Now xfs_copy already full support to copy a V5 XFS. But for above
old problem, xfstests use below patch to make sure xfs_copy always
use "-d" option to copy a V5 XFS:
8346e53 common: append -d option to XFS_COPY_PROG when testing v5 xfs
That cause xfstests miss the coverage of copying a V5 XFS without
"-d". For test this feature I did below things:
1. Changed init_rc(), add "-d" to $XFS_COPY_PROG if xfs_copy can't
copy a V5 XFS properly.
2. xfs/073 test V4 xfs forcibly by specify "-m crc=0" in case. I
think it's useless now, so remove it.
3. Changed xfs/032. If xfs_copy full support to copy a V5 XFS, test
with and without "-d" option, or only test with "-d" option.
Signed-off-by: Zorro Lang <zlang@redhat.com>
Reviewed-by: Eryu Guan <eguan@redhat.com>
Signed-off-by: Eryu Guan <eguan@redhat.com>
Multi-threads freeze/unfreeze maybe trigger some bugs, e.g: panic,
hang or data corruption etc. It does't check the return value of
freeze/unfreeze, but it tries to make sure fsstress won't run fails,
and no any other bugs happen.
[eguan: add comment on why we check _scratch_mount failure]
Signed-off-by: Zorro Lang <zlang@redhat.com>
Reviewed-by: Eryu Guan <eguan@redhat.com>
Signed-off-by: Eryu Guan <eguan@redhat.com>
_test_inode_flag() and _test_inode_extsz() use "which $XFS_IO_PROG"
to check if xfs_io command is available. And "-i" option was added
to XFS_IO_PROG varibable by commit 54659ecdb5 ("fstests: run
xfs_io with -i option if supported"). So the command becomes "which
/usr/sbin/xfs_io -i", and it stops and waits for input from stdin,
because "-i" option of "which" means "Read aliases from stdin".
I've seen xfs/008 hang when testing with latest xfsprogs, where
xfs_io has "-i" support.
Fix it by removing the xfs_io command detections, and making xfs_io
mandatory in common/config.
Also fix the indentions in these functions, use tab instead of
spaces, while we're at it.
Reviewed-by: Dave Chinner <dchinner@redhat.com>
Signed-off-by: Eryu Guan <eguan@redhat.com>
Upstream xfs_io has been converted to always use LFS compliant
(i.e. 64 bit) pwrite() rather than pwrite64(). Similar changes have
been made for multiple syscalls that have "*64" variants. hence the
error output of all these commands has changed, such as "pwrite64:
..." to "pwrite: ....".
Make a filter to catch the *64 variants and strip it, and
convert all the golden output to use the non-*64 variant. This will
make all golden output matching work correctly regardless of what
version of xfs_io is in use.
Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Eryu Guan <eguan@redhat.com>
In the test ROOT_INO is filtered out and/or replaced, but if
ROOT_INO is also 32, more "32"s are filtered and replaced than
expected. This happens to me when testing 512B block size XFS and 1k
block size CRC enabled XFS.
Fix it by filtering out only ROOT_INO at the beginning of a line,
and removing all "g" modifiers in sed expressions.
Also the ROOT_INO should be the root inode number of TEST_DIR not
SCRATCH_MNT.
Reviewed-by: Dave Chinner <dchinner@redhat.com>
Signed-off-by: Eryu Guan <eguan@redhat.com>
generic/314 no longer tests POSIX ACLs. It only tests the SGID bit.
Therefore, it should not be in the acl group anymore.
Signed-off-by: Eric Biggers <ebiggers@google.com>
Reviewed-by: Eryu Guan <eguan@redhat.com>
Signed-off-by: Eryu Guan <eguan@redhat.com>
Commit 4d0c5ba (vfs: do get_write_access() on upper layer of overlayfs)
that was merged for kernel v4.8 changes the behavior of execve+truncate
in a subtle way.
When file is executed from lower and then truncated, the truncation is
allowed, becasue it only affects the size of the copied up file and does
not affect the VM mapping of the executed lower file.
Signed-off-by: Amir Goldstein <amir73il@gmail.com>
Reviewed-by: Eryu Guan <eguan@redhat.com>
Signed-off-by: Eryu Guan <eguan@redhat.com>
test_upper file was supposed to be in upper layer,
but it was wrongly created in lower layer.
Signed-off-by: Amir Goldstein <amir73il@gmail.com>
Reviewed-by: Eryu Guan <eguan@redhat.com>
Signed-off-by: Eryu Guan <eguan@redhat.com>
When an O_TMPFILE file is linked into a directory with a POSIX
default ACL, the file should inherit the default ACL and the umask
needs to be ignored.
[eguan introduced testdir var and added test to acl group]
Signed-off-by: Andreas Gruenbacher <agruenba@redhat.com>
Reviewed-by: Eryu Guan <eguan@redhat.com>
Signed-off-by: Eryu Guan <eguan@redhat.com>
In commit c08ebd092 ("xfs: fix $XFS_DB_PROG usage"), the need for
callers to pass the device to populate into
__populate_check_xfs_dir() was removed. So we can now clean up all
the callers by removing the device parameter.
Signed-off-by: Xiao Yang <yangx.jy@cn.fujitsu.com>
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Eryu Guan <eguan@redhat.com>
Convert some more cases of 'xfs_db $SCRATCH_DEV' to _scratch_xfs_db
that were left out of the initial cleanup patch.
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Eryu Guan <eguan@redhat.com>
We re-added the UNSHARE flag to fallocate, so go make sure that all
the unshare tests actually check that the installed copy of xfs_io
supports the 'funshare' command and that the underlying filesystem
understands the flag, and change the tests themselves to use
funshare.
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Eryu Guan <eguan@redhat.com>
When we're setting up a fake cow extent in the refcountbt to test
cleanup of fake cow extents, set the cowflag in the record field
to reflect our new disk format of storing the staging extents in
the right side of the tree.
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Eryu Guan <eguan@redhat.com>
The value of "$XFS_IO_PROG" may contain extra flags after the
binary path (e.g. -F), so it is wrong to use the variable inside
quotes in xfs_io execution call sites.
This bug surfaced while testing the new xfs_io -i flag.
Signed-off-by: Amir Goldstein <amir73il@gmail.com>
Reviewed-by: Eryu Guan <eguan@redhat.com>
Signed-off-by: Eryu Guan <eguan@redhat.com>
Jaegeuk Kim
Qu Wenruo
Omar Sandoval
Liu Bo
Eryu Guan
Amir Goldstein
Eric Sandeen
Zorro Lang
Dave Chinner
Eric Biggers
Andreas Gruenbacher
Xiao Yang
Darrick J. Wong