| Age | Commit message (Collapse) | Author |
|---|
|
Add a helper 'sync_directory()' function to synchronize directories.
Signed-off-by: Artem Bityutskiy <Artem.Bityutskiy@nokia.com>
|
|
In 'dir_check()' print name of the directory if we fail to open it.
Signed-off-by: Artem Bityutskiy <Artem.Bityutskiy@nokia.com>
|
|
Currently when we are doing UBIFS power cut emulation testing we assume that
EROFS is returned in case of an emulated power cut event. However, this is
not true and sometimes Linux returns EIO to user-space instead. Namely, this
happens when we are waiting for write-back to finish. Here is a call trace
form kernel 2.6.39-rc7:
Call Trace:
[<ffffffff810aaa77>] filemap_write_and_wait_range+0x78/0xb1
[<ffffffff8110f34d>] vfs_fsync_range+0x38/0xd3
[<ffffffff8110f42b>] generic_write_sync+0x43/0x65
[<ffffffff810ab0c9>] generic_file_aio_write+0xaa/0xf2
[<ffffffffa02224c9>] ubifs_aio_write+0x13a/0x1a1 [ubifs]
[<ffffffff810ed71f>] do_sync_write+0xcb/0x141
[<ffffffff810ee016>] vfs_write+0xb0/0x150
[<ffffffff810ee175>] sys_write+0x48/0x72
[<ffffffff813e0692>] system_call_fastpath+0x16/0x1b
'filemap_fdatawait_range()' returns hard-coded -EIO if a page write-back
happens with whatever error code.
Thus, this patch teaches integck to handle EIO when doing power cut emulation
testing.
Signed-off-by: Artem Bityutskiy <Artem.Bityutskiy@nokia.com>
|
|
This patch adds -m <mtdnum> option to integck and teaches integck to re-attach
the MTD device to UBI in case of an emulated power cut event. This is needed
for the new UBI power cut and unstable bits emulation infrastructure: when UBI
emulates a power cut the only way to recover form this is to re-attach the MTD
device.
Signed-off-by: Artem Bityutskiy <Artem.Bityutskiy@nokia.com>
|
|
When we are in power cut mode read all files after re-mounting. This will
check that FS recovery worked well and all files are readable after recovery.
Signed-off-by: Artem Bityutskiy <Artem.Bityutskiy@nokia.com>
|
|
Currently we always verify holes when truncating, but we have -v switch for
this and should do this only if -v was given.
Signed-off-by: Artem Bityutskiy <Artem.Bityutskiy@nokia.com>
|
|
Call 'fsync()' for directories sometimes as well, and check synchronized
directories after re-mount - cound of directory entries has to be correct.
Signed-off-by: Artem Bityutskiy <Artem.Bityutskiy@nokia.com>
|
|
If 'fsync()' or 'fdatasync()' fail we have to return an error, but we by
mistake were returning 0 (success).
Signed-off-by: Artem Bityutskiy <Artem.Bityutskiy@nokia.com>
|
|
Currently integck calls 'fsync()' very rarely - with 0.1% probability. Make
this happen more often - with 1% probability.
Signed-off-by: Artem Bityutskiy <Artem.Bityutskiy@nokia.com>
|
|
Allow verification for clean files, i.e. the files which had been fsync()'ed
before the emulated power cut happened.
Signed-off-by: Artem Bityutskiy <Artem.Bityutskiy@nokia.com>
|
|
Make the verify option -v work with power cut emulation option -e. We can
verify all operations except of writing to the file, because of write-back.
Indeed, even if we write successfully, the pages may still be in the page-cache
and if an emulated power cut happens they will be dropped, and our verification
will fail.
I am not 100% sure it is save to verify non-write operation in other FSes, but
only UBIFS has power cut emulation mode now, so we do not really care about
other FSes at this point.
Signed-off-by: Artem Bityutskiy <Artem.Bityutskiy@nokia.com>
|
|
Enable directory entry checking in power cut mode. This means that with this
patch integck will check that all directory entries in the file-system were
created by integck.
Signed-off-by: Artem Bityutskiy <Artem.Bityutskiy@nokia.com>
|
|
All errors but write errors have corresponding CHECK(), so if something
gets wrong we fail. But in case of I/O errors we do not fail if an
unexpected error happens, which makes it difficult to trace issues.
Inject the errno checking to the pvc macro to make sure we fail if
errno is not EROFS.
Signed-off-by: Artem Bityutskiy <Artem.Bityutskiy@nokia.com>
|
|
Commit 1394a46213e8180e1233ca6d4811e7c77d49b1e0 was a good idea, but it
introduced a segfault - files can be unlinked and 'file->links' is NULL, it is
bad idea to unconditionally dereference it with 'file->links->name'. This patch
introduces a helper function which takes care of the situation with unlinked
files, plus it adds few assertions.
Signed-off-by: Artem Bityutskiy <Artem.Bityutskiy@nokia.com>
|
|
Introduce per-file "clean" flag which tells whether the file is synchronized or
not. This flag is not really used so far.
Signed-off-by: Artem Bityutskiy <Artem.Bityutskiy@nokia.com>
|
|
In pcv() messages also print the error number - this is useful when
an error happens and it is not because of EROFS.
Signed-off-by: Artem Bityutskiy <Artem.Bityutskiy@nokia.com>
|
|
This is just a preparation - re-structure 'file_write()' to make one "success"
return point - we'll add synchronization stuff there. Also, while on it, make
it delete the file errored in truncation, not only in write. Also, move
the fsync() call to this function because it is common to all write paths.
Signed-off-by: Artem Bityutskiy <Artem.Bityutskiy@nokia.com>
|
|
I observes segfaults in integck test, and unfortunately I do not have the core
file to investigate the problem. But I see one possibility for the test to
segfault - it has unbounded recursion. Limit the maximum recursion depth.
Signed-off-by: Artem Bityutskiy <Artem.Bityutskiy@nokia.com>
|
|
If we are not in power cut testing mode (-p was not used) and we get an
unexpected EROFS, we silently exit. This patch fixes this behavior and make
the test loudly fail instead.
Signed-off-by: Artem Bityutskiy <Artem.Bityutskiy@nokia.com>
|
|
The test sometimes dies due to SIGBUS because a power cut event happens when
we are writing to an mmap()'ed area. SIGBUS is very difficult to ignore, so
fix the issue simply by avoiding doing 'mmap()' operations when in power cut
testing mode.
Signed-off-by: Artem Bityutskiy <Artem.Bityutskiy@nokia.com>
|
|
Similarly to the previous commit, do not verify operations when in
power-cut testing mode. Again, this is because write-back may fail
meanwhile.
Signed-off-by: Artem Bityutskiy <Artem.Bityutskiy@nokia.com>
|
|
In 'check_failed()' use the file and line number provided by the CHECK()
macro via the function parameters, rather than using __LINE__ directly.
Otherwise the funtion prints its own line, instead of the line where the
CHECK() failed.
Signed-off-by: Artem Bityutskiy <Artem.Bityutskiy@nokia.com>
|
|
Do not check the file-system when we are in power cut mode. Currently
we still check the FS if we did not encounter any error, and the test
fails sometimes because meanwhile write-back encounters an error and
drops pages. Disable the checking so far. It should be enabled when
we implement the 'clean' flag for files.
Signed-off-by: Artem Bityutskiy <Artem.Bityutskiy@nokia.com>
|
|
Introduce a 'check_tested_fs()' helper function.
Signed-off-by: Artem Bityutskiy <Artem.Bityutskiy@nokia.com>
|
|
Do not print extra messages when unable to re-mount - they are
already printed by the recover function.
Signed-off-by: Artem Bityutskiy <Artem.Bityutskiy@nokia.com>
|
|
The integck test assumes that if the FS is rootfs then it cannot
re-mount it, otherwise it can. However, this is not true because the
tested FS can be not remountable even if it is not rootfs, e.g., if
'integck' lives on this FS and is executed from it.
This patch improves the integck logic and actually checks wheter it
is possible to re-mount the FS before starting the test.
Also, it makes sure that power cut testing is run only on re-mountable
FS, because we have to be able to clean the "corrupted" state of the
file-system.
Signed-off-by: Artem Bityutskiy <Artem.Bityutskiy@nokia.com>
|
|
Commit 08cfd9d7cacc3c0b97530ad7e13fa8db75e99947 introduced a bug:
we do not re-mount the FS before checking at the first iteration.
This patch fixes the issue.
Signed-off-by: Artem Bityutskiy <Artem.Bityutskiy@nokia.com>
|
|
-v (verbose) switch is supposed to turn power cut testing error
messages on/off. However, it does not work because the messages
are printed even if -v was not specified. This patch fixes it.
Signed-off-by: Artem Bityutskiy <Artem.Bityutskiy@nokia.com>
|
|
Instead of using rand_r with own seed for operations and rand() with
global state for data it is cleaner to use rand_r with own different
states for these things. This patch eliminates 'srand()' and 'rand()'
usage.
Signed-off-by: Artem Bityutskiy <Artem.Bityutskiy@nokia.com>
|
|
Currently integck uses the same global random generator for everything -
for choosing the operation, generating the data, and for checking. This
makes integck to become stuck sometimes. My guess this is because of
we somehow re-set it back with srand() when checking files.
This patch makes integck use different generators for data and for
choosing operations by using rand_r() with own seed for operations.
Signed-off-by: Artem Bityutskiy <Artem.Bityutskiy@nokia.com>
|
|
I often need to use gdb to debug integck, or sometimes even attach
already running integck. But when it is compiled with optimizations, it
is nearly impossible to use gdb. Introduce 'debug' target to compile
integck without optimizations.
Additionally, to make the stack backtraces work, add a dirty hack to
integck.c to remove all the static keywords - this turns static
functions to non-static and makes them appear in stack backtraces.
Signed-off-by: Artem Bityutskiy <Artem.Bityutskiy@nokia.com>
|
|
When -e option is used verify file holes after the truncate
operation. Re-use the existing 'file_check_hole()' function for
this.
Signed-off-by: Artem Bityutskiy <Artem.Bityutskiy@nokia.com>
|
|
When -e options is used - verify written data. Re-use the existing
file_check_data() function to implement this.
Signed-off-by: Artem Bityutskiy <Artem.Bityutskiy@nokia.com>
|
|
Improve the test and add -e option which forces integck to verify all
operation it performs. This patch does not implement file writes and
holes operations so far - will be done in subsequent patches.
Signed-off-by: Artem Bityutskiy <Artem.Bityutskiy@nokia.com>
|
|
This patch teaches the 'CHECK()' macro to print the back-trace in
case of failure which is sometimes very useful for debugging. This
patch also adds a helper function for the 'CHECK()' macro because
otherwise it becomes too large.
Note, since all functions in the test are static, the stack trace
does not really show symbols.
Signed-off-by: Artem Bityutskiy <Artem.Bityutskiy@nokia.com>
|
|
Fix typos, make it a bit more descriptive.
Signed-off-by: Artem Bityutskiy <Artem.Bityutskiy@nokia.com>
|
|
This is a dirty patch - it makes many changes, but I simply had no
time to split it. And because this test is something no one except me
and Adrian cares about, I do not want to spend too much time creating
nice patches. Anyway, this patch make the power cut testing mode work.
Now you can run the test like "./integck /mnt/ubifs -p" and it will
run recover from emulated power cut failures and continue forever,
untill an unexpected issue is met.
Signed-off-by: Artem Bityutskiy <Artem.Bityutskiy@nokia.com>
|
|
The 'errmsg()' macro adds "\n" at the end, so do not use it in our
'pcv()' and 'CHECK()' macros.
Signed-off-by: Artem Bityutskiy <Artem.Bityutskiy@nokia.com>
|
|
If a power cut emulation error happens and we are trying to re-mount
the file-system, do not assume it is mounted because the failure could
have happened after it had had been unmounted but before it was
successfully mounted.
Signed-off-by: Artem Bityutskiy <Artem.Bityutskiy@nokia.com>
|
|
When re-mountig to recover from the emulated power-cut failure then we should
do a bit less fancy things, e.g., we should not try to re-mount to R/O and R/W
modes forth and back, but we should instead just unmount directly.
Additionally, we should not die if mount fails because there may be a power
cut emulation during mounting. Thus, create a separate function for recovering
and try to recover in a loop.
Signed-off-by: Artem Bityutskiy <Artem.Bityutskiy@nokia.com>
|
|
Teach 'add_dir_entry()' to allocate the file/dir/symlink object for
the client. This simplifies client's code. The only place where we
do not want to do this is when we are creating a hardling pointing
to an existing file.
Teach 'remove_dir_entry()' to free the file/dir/symlink object for
the client. Similarly, this simplifies client's code. The only place
where we do not want to do this is the rename function - when we
rename an object we want to only change its direntry, but not the
object itself.
But the main motivation for this change is not just to clean-up, but
to fix another memory leak. In the re-name function we free the
'rename_entry', but we did not free the corresponding object. This
patch fixes the situation.
Signed-off-by: Artem Bityutskiy <Artem.Bityutskiy@nokia.com>
|
|
The test forgets to free the args.mount_opts string on exit - fix this.
Signed-off-by: Artem Bityutskiy <Artem.Bityutskiy@nokia.com>
|
|
If the test runs in power-cut testing mode, do not exit in case of
EROFS error - but instead re-mount and start again.
Signed-off-by: Artem Bityutskiy <Artem.Bityutskiy@nokia.com>
|
|
Refuse to test the file-system if it is mounted Read-only.
Signed-off-by: Artem Bityutskiy <Artem.Bityutskiy@nokia.com>
|
|
Signed-off-by: Artem Bityutskiy <Artem.Bityutskiy@nokia.com>
|
|
We always pass non-NULL 'parent' argument to 'new_dir()', so no need to
test it for NULL.
Signed-off-by: Artem Bityutskiy <Artem.Bityutskiy@nokia.com>
|
|
This patch kills the "name" field from 'struct file_info' where we stored
the original file name. We do not really need to keep this name around,
becaus it might have long gone and it is not very interesting. Besides,
we have several memory leaks where we leak 'file->name' strings, so this
is the easiest way to fix those leaks as well.
Signed-off-by: Artem Bityutskiy <Artem.Bityutskiy@nokia.com>
|
|
We do not need to store directory name in the 'struct dir_info' objects,
because we already have it in 'struct dir_entry_info'. So we duplicate
the names for directories in 'dir->name' and 'dir->entry->name'.
Signed-off-by: Artem Bityutskiy <Artem.Bityutskiy@nokia.com>
|
|
Make 'add_dir_entry()' consistent and allocate name for all types,
not only for 'd' and 's'. Also, since 'add_dir_entry()' sets the
parent - do not do this in the calling functions.
This is a clean-up which makes 'add_dir_entry()' more consistent.
Signed-off-by: Artem Bityutskiy <Artem.Bityutskiy@nokia.com>
|
|
When we create a symlink we also allocate a symlink_info structure
and the target path name. But when we remove a symlink - we do not
delete that memory. This patch fixes the issue.
Signed-off-by: Artem Bityutskiy <Artem.Bityutskiy@nokia.com>
|
