Ensure that we get an error back on all fds when a block device is
open by multiple writers and writeback fails.
Signed-off-by: Jeff Layton <jlayton@redhat.com>
Reviewed-by: Eryu Guan <eguan@redhat.com>
Signed-off-by: Eryu Guan <eguan@redhat.com>
I'm working on a set of kernel patches to change how writeback errors
are handled and reported in the kernel. Instead of reporting a
writeback error to only the first fsync caller on the file, it has
the the kernel report them once on every file description that was
open at the time of the error.
This patch adds a test for this new behavior. Basically, open many fds
to the same file, turn on dm_error, write to each of the fds, and then
fsync them all to ensure that they all get an error back.
To do that, I'm adding a new tools/dmerror script that the C program
can use to load the error table from the script. It's also suitable for
setting up, frobbing and tearing down a dmerror device for by-hand testing.
For now, only ext2/3/4 and xfs are whitelisted on this test, since those
filesystems are included in the initial patchset. We can add to that as
we convert filesystems, and eventually make it a more general test.
Signed-off-by: Jeff Layton <jlayton@redhat.com>
Reviewed-by: Eryu Guan <eguan@redhat.com>
Signed-off-by: Eryu Guan <eguan@redhat.com>
The writeback error handling test requires that you put the journal on a
separate device. This allows us to use dmerror to simulate data
writeback failure, without affecting the journal.
xfs already has infrastructure for this (a'la $SCRATCH_LOGDEV), so wire
up the ext4 code so that it can do the same thing when _scratch_mkfs is
called.
Signed-off-by: Jeff Layton <jlayton@redhat.com>
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Eryu Guan <eguan@redhat.com>
Linux swapfiles use bmap which has no ability to tell the swap code
that the blocks its reporting aren't actually on the data device.
Therefore, swapon on a realtime file could corrupt random blocks on
the data device, so we must ensure that swapon cannot happen.
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 unmerged and impure upper directories may contain invalid
whiteouts when we umount && modify lowerdir(e.g. clean up dir) and
remount overlay. This may lead to whiteouts exposure and rmdir
failure.
This case test impure dir's whiteouts check in ovl_iterate() and
ovl_remove_xxx().
Signed-off-by: zhangyi (F) <yi.zhang@huawei.com>
Reviewed-by: Amir Goldstein <amir73il@gmail.com>
Signed-off-by: Eryu Guan <eguan@redhat.com>
If generic/397 is executed in an environment with SIGPIPE ignored,
it fails because the 'yes' program prints an error message:
yes: standard output: Broken pipe
yes: write error
This can be reproduced with:
trap '' SIGPIPE; ./check generic/397
Fix it by generating the string of 255 y's using just 'head' and
'tr' instead of 'yes', 'head', and 'tr'.
Although it's not really a good idea to execute xfstests with
SIGPIPE ignored, this is the only test I've noticed where it causes
a problem, so it might as well be fixed in the test.
It would be much nicer to prevent this problem for all tests by
making the 'check' script restore the default SIGPIPE handler. But
that isn't straightforward because bash's 'trap' builtin doesn't
allow un-ignoring signals that were ignored on entry to the shell.
[ eguan added more background infomation to commit log, which is
also from Eric.
I think it's an easy problem for others to run into, since sometimes
processes ignore SIGPIPE because they want to get write errors
instead, but then when doing fork() + exec() they forget to reset
the SIGPIPE handler. Notably, Python got this wrong and it wasn't
fixed until Python 3, so any programs executing the 'check' script
from a Python 2 script will usually get this wrong (see:
https://bugs.python.org/issue1652). And usually everything works
fine but every once in a while there is a weird problem like this
which has to be debugged. ]
Signed-off-by: Eric Biggers <ebiggers@google.com>
Reviewed-by: Eryu Guan <eguan@redhat.com>
Signed-off-by: Eryu Guan <eguan@redhat.com>
Test that an incremental send works if we rename some directory inode A
and then rename some file inode B to the name inode A had, for the case
where the directory inode A is an ancestor of inode B in the parent
snapshot.
This issue is fixed by the following patch for the linux kernel:
"Btrfs: incremental send, fix invalid path for unlink commands"
Signed-off-by: Filipe Manana <fdmanana@suse.com>
Reviewed-by: Eryu Guan <eguan@redhat.com>
Signed-off-by: Eryu Guan <eguan@redhat.com>
UDF does not support journalling. Make the appropriate feature test fail
for it.
Signed-off-by: Jan Kara <jack@suse.cz>
Reviewed-by: Eryu Guan <eguan@redhat.com>
Signed-off-by: Eryu Guan <eguan@redhat.com>
A test is creating symlink with a path containing name 1023 characters
long. Such file name is invalid for most filesystems (the limit on file
name lenght is mostly 255 characters) and UDF actually complains about
this and so the test fails. Since the point of this test is to verify
storage of the symlink, change the test to use a valid path where each
component name has only 254 characters.
Signed-off-by: Jan Kara <jack@suse.cz>
Reviewed-by: Eryu Guan <eguan@redhat.com>
Signed-off-by: Eryu Guan <eguan@redhat.com>
Add a regression test for a bug where ->lookup() in an encrypted
directory would incorrectly return EPERM, depending on which inodes
happened to have their keys still cached in memory following removal of
the keyring key. This bug was fixed in v4.12-rc1, v4.9.29, and v4.4.70.
Cc: linux-fscrypt@vger.kernel.org
Signed-off-by: Eric Biggers <ebiggers@google.com>
Reviewed-by: Eryu Guan <eguan@redhat.com>
Signed-off-by: Eryu Guan <eguan@redhat.com>
Test that an incremental send operation works correctly when an inode A
is renamed, a new hard link added to it and some other inode B is renamed
to the old name of inode A.
The btrfs bug is fixed by the following patch for the linux kernel:
"Btrfs: send, fix invalid path after renaming and linking file"
Signed-off-by: Filipe Manana <fdmanana@suse.com>
Reviewed-by: Eryu Guan <eguan@redhat.com>
Signed-off-by: Eryu Guan <eguan@redhat.com>
Some tests require the scratch volume to be a block device,
although they don't do anything block device specific like
the device mapper tests. They actually only require a local
device and fail with network or overlay pseudo mount devices.
The same tests also try to mount the scratch device manually
after running _scratch_mkfs. For UBIFS, _scratch_mkfs simply
truncates the UBI volume and relies on the kernel to re-format
the volume on the next mount. This fails if the fs type is not
specified explicitly when mounting.
This patch replaces the block device requirement with a
generic check for a local device and adds a filesystem
specification to all manual mounts of the scratch device
to a mount point.
Signed-off-by: David Oberhollenzer <david.oberhollenzer@sigma-star.at>
Reviewed-by: Eryu Guan <eguan@redhat.com>
Signed-off-by: Eryu Guan <eguan@redhat.com>
Some filesystems (e.g. UBIFS) fail with EPERM when trying to move a
file into an encrypted directory via cross rename, without having
access to the encryption key.
Since rename is perfectly allowed to return EPERM, which is also
tested for, and no precise specification seems to exist that
clarifies when to expect EPERM and when ENOKEY, this patch modifies
the generic/398 test to accept both, for the test case where the key
is not available.
Signed-off-by: David Oberhollenzer <david.oberhollenzer@sigma-star.at>
Reviewed-by: Eryu Guan <eguan@redhat.com>
Signed-off-by: Eryu Guan <eguan@redhat.com>
UBIFS is a filesystem for unmanaged flash memory devices. It works
on top of UBI (Unsorted Block Images) which is a wear leveling and
volume management layer on top of flash memory devices, which are
handled by the MTD subsystem (memory technology device).
Since the semantics of flash devices are drastically different from
regular block devices (blocks or "pages" must be erased before
writing, only larger groups of pages or "erase blocks" can be erased
at once, page write must be in order within an erase block, etc...)
it was decided to expose MTD devices as character devices with
ioctls for operations like erase.
Since erasing a flash erase block causes physical wear on the
device, eventually causing the erase blocks to go bad, the UBI layer
provides mainly transparent wear leveling on top of MTD devices. UBI
does not attempt to emulate a regular block device, but rather
something like a flash memory with idealized characteristics that
can be partitioned into multiple UBI volumes in a fashion somewhat
similar to LVM. UBI volumes are also exposed to user space as
character devices.
This patch mainly deals with some quirks of UBIFS like working on
top of character devices instead of block devices. Also UBIFS
automatically formats UBI devices when trying to mount an empty
device. The mkfs.ubifs program is mainly used for creating images.
This patch changes _scratch_mkfs and _scratch_mkfs_encrypted to
truncate the UBI volume instead, relying on the kernel to reformat
it on the next mount.
For _scratch_mkfs_encrypted this is actually required to get the
encryption tests to run, because mkfs.ubifs, at the time of writing
this, the kernel support for UBIFS encryption is fairly recent and
mkfs.ubifs does not have proper support yet.
The necessity of an additional -ubifs switch was discussed but auto
detection of UBIFS formated UBI devices could not be reproduced on
my end and is unlikely to work with empty UBI volumes anyway.
Signed-off-by: David Oberhollenzer <david.oberhollenzer@sigma-star.at>
Reviewed-by: Eryu Guan <eguan@redhat.com>
Signed-off-by: Eryu Guan <eguan@redhat.com>
Implement _is_char_dev similar to _is_block_dev to test for
character devices.
Add a _require_local_device test. This test is similar to
_require_block_device but checks if the path refers to a
block or a character device.
Signed-off-by: David Oberhollenzer <david.oberhollenzer@sigma-star.at>
Reviewed-by: Eryu Guan <eguan@redhat.com>
Signed-off-by: Eryu Guan <eguan@redhat.com>
This is a regression test for patch "Btrfs: fix delalloc accounting
leak caused by u32 overflow". It creates a bunch of delalloc extents
and merges them together to make sure the accounting is done right.
[eguan: use $XFS_IO_PROG instead of xfs_io]
Signed-off-by: Omar Sandoval <osandov@fb.com>
Reviewed-by: Eryu Guan <eguan@redhat.com>
Signed-off-by: Eryu Guan <eguan@redhat.com>
When the filesystem is force-killed in generic/019 and fsstress is
compiled with DEBUG then it's possible for check_cwd() to fail. In
this case fsstress would just die and invoking kill on its pid would
produce an error to the output file:
+./tests/generic/019: line 157: kill: (32750) - No such process
Fix this possibility by ignoring the result of the output command,
this won't affect the test coverage.
Signed-off-by: Nikolay Borisov <nborisov@suse.com>
Reviewed-by: Eryu Guan <eguan@redhat.com>
Signed-off-by: Eryu Guan <eguan@redhat.com>
Make _require_group yield the correct error message when the
requested group is not found.
Signed-off-by: Nikolay Borisov <nborisov@suse.com>
Reviewed-by: Eryu Guan <eguan@redhat.com>
Signed-off-by: Eryu Guan <eguan@redhat.com>
Test that if we punch a hole in a file, with either a range that goes
beyond the file's size or covers a file range that is already a hole,
and that if after we do some buffered write operations that cover
different parts of the hole, no warnings are emmitted in syslog/dmesg
and the file's content is correct after remounting the filesystem.
This test is motivated by a bug in btrfs that is manifested in kernel
4.12-rc1 onwards (the bug existed long time ago but was not so easy
to expose before 4.12-rc1). The btrfs patch that fixes the issue is
titled: "Btrfs: fix invalid extent maps due to hole punching".
Signed-off-by: Filipe Manana <fdmanana@suse.com>
Reviewed-by: Eryu Guan <eguan@redhat.com>
Signed-off-by: Eryu Guan <eguan@redhat.com>
listxattr reaturns a null-terminated list of entries that represent
the xattr names. However, if it is passed larger buffer than it
requires it won't zero-out the rest of the memory. The way the
loop iterator in listxattr.c is written makes it go print every
null-terminated entry up to bufsize (which is user passed parameter).
This can lead to a situation where listxattr users N bytes out of
M bytes big buffer ( M > N). This will leave the rest (M-N)
as garbage, which in turn will be printed by listxattr. Fix this
by converting the 'for' loop to 'while' and properly ensuring
we are reading at most howevermany elements the syscall reported
it returned
Signed-off-by: Nikolay Borisov <nborisov@suse.com>
Reviewed-by: Eryu Guan <eguan@redhat.com>
Signed-off-by: Eryu Guan <eguan@redhat.com>
Xiao Yang reported that fstests failed to build on RHEL6.9 hosts due
to old autoreconf didn't pass -I to aclocal -I. (This was fixed by
autoconf commit 44fbeef86d03 ("Pass autoreconf -I to aclocal -I"),
but not on RHEL6.9).
So call aclocal, autoheader and autoconf directly instead of
autoreconf, as what's done in xfsprogs Makefile.
Reported-by: Xiao Yang <yangx.jy@cn.fujitsu.com>
Reviewed-by: Xiao Yang <yangx.jy@cn.fujitsu.com>
Signed-off-by: Eryu Guan <eguan@redhat.com>
Add test checking for a race in ext4 writeback that could result in
zeroing too much from the tail page during writeback.
[eguan: removed from quick group, it needs longer time on xfs and
btrfs]
Signed-off-by: Jan Kara <jack@suse.cz>
Reviewed-by: Eryu Guan <eguan@redhat.com>
Signed-off-by: Eryu Guan <eguan@redhat.com>
Test nfs4_getfacl gets ACL list correctly from server when the ACL
length is close enough to the end of a page. On buggy NFS client
getxattr could return ERANGE. Upstream commit ed92d8c137b7 ("NFSv4:
fix getacl ERANGE for some ACL buffer sizes") fixed this bug in 4.11
kernel.
Note that this reproducer was originally written by J. Bruce Fields.
Cc: J. Bruce Fields <bfields@redhat.com>
Cc: Weston Andros Adamson <dros@primarydata.com>
Cc: linux-nfs@vger.kernel.org
Reviewed-by: J. Bruce Fields <bfields@redhat.com>
Signed-off-by: Eryu Guan <eguan@redhat.com>
The OpenSSL dependency was added for one program, fssum, and it needs
it only because it needs a md5 implementation. Use Solar Designer's
openssl compatible implementation of md5 so we no longer need to
depend on OpenSSL.
Since the OpenSSL libraries are not always available, we had to add
extra complexity to test to see whether fssum exists.
The other problem with depending on the OpenSSL libraries is that
shared library compatibility situation is terrible; a fssum binary
built on a system using libssl1.0.0 is *NOT* run on a system with
libssl1.0.2, since the shared libraries are incompatible even across a
minor version bump. (Sigh.)
Signed-off-by: Theodore Ts'o <tytso@mit.edu>
Reviewed-by: Eryu Guan <eguan@redhat.com>
Signed-off-by: Eryu Guan <eguan@redhat.com>
Jeff Layton
Darrick J. Wong
zhangyi (F)
Eric Biggers
Filipe Manana
Jan Kara
David Oberhollenzer
Omar Sandoval
Nikolay Borisov
Eryu Guan
Theodore Ts'o