The list alloc blocks have an array of blknos that are offset by a start
field in the block header. The print code wasn't using that and was
always referencing the beginning of the array, which could miss blocks.
Signed-off-by: Zach Brown <zab@versity.com>
Since kABI migrations across minor versions is a thing of the past going
forward, we now:
- Detect if we're on EL9
- If so, add a requirement on the various flavors of release package to
that specific major.minor version
This appropriately does not allow upgrades across minor versions.
Additional blkdev/bdev changes now cause this call to be removed as
well resulting in us having to use yet another API to do the same for
el9_5.
The changes are a little more subtle as now the bdev_mount() call passes
a custom bd_holder_ops that we must match or else throw a WARN_ON, so we
switch to using sbi as our holder arg instead.
Make sure to bdev_fput and not fput, since we don't want to have our
private data cleanup deferred, failing xfstests generic/604.
Signed-off-by: Auke Kok <auke.kok@versity.com>
The assignments to it is no longer needed at all. All references can be
dropped since v6.4-rc4-163-g0d625446d0a4.
Signed-off-by: Auke Kok <auke.kok@versity.com>
Since v6.5-rc1-7-g9b6304c1d537, current_time() is no longer
extern, so we need to update this grep regex to continue to match.
Signed-off-by: Auke Kok <auke.kok@versity.com>
Add a quick test that races readers and shrinking to stress lock object
refcount racing between concurrent lock request handling threads in the
lock server.
Signed-off-by: Zach Brown <zab@versity.com>
Right now a client requesting a null mode for a lock will cause
invalidations of all existing granted modes of the lock across the
cluster.
This is unneccessarily broad. The absolute requirement is that a null
request invalidates other existing granted modes on the client. That's
how the client safely resolves shrinking's desire to free locks while
the locks are in use. It relies on turning it into the race between use
and remote invalidation.
But that only requires invalidating existing grants from the requesting
client, not all clients. It is always safe for null grants to coexist
with all grants on other clients. Consider the existing mechanics
involving null modes. First, null locks are instatiated on the client
before sending any requests at all. At any given time newly allocated
null locks are coexisting with all existing locks across the cluster.
Second, the server frees the client entry tracking struct the moment it
sends a null grant to the client. From that point on the client's null
lock can not have any impact on the rest of the lock holders because the
server has forgotten about it.
So we add this case to the server's test that two client lock modes are
compatible. We take the opportunity to comment the heck out of this
function instead of making it a dense boolean composition. The only
functional change is the addition of this case, the existing cases are
refactored but unchanged.
Signed-off-by: Zach Brown <zab@versity.com>
When freeing acked reesponses in the net layer we sweep the send and
resend queues looking for queued responses up to the sequence number
we've had acked. The code that did this used a weird pattern of
returning ints and adding them which gave me pause. Clean it up to use
bools and or (not short-circuiting ||) to more obviously communicate
what's going on.
Signed-off-by: Zach Brown <zab@versity.com>
Over time some fields have been added to the lock struct which haven't
been added to the lock tracing output. Add some of the more relevant
lock fields to tracing.
Signed-off-by: Zach Brown <zab@versity.com>
Lock object lifetimes in the lock server are protected by reference
counts. References are acquired while holding a lock on an rbtree.
Unfortunately, the decision to free lock objects wasn't tested while
also holding that lock on the rbtree. A caller putting their object
would test the refcount, then wait to get the rbtree lock to remove it
from the tree.
There's a possible race where the decision is made to remove the object
but another reference is added before the object is removed. This was
seen in testing and manifest as an incoming request handling path adding
a request message to the object before it is freed, losing the message.
Clients would then hang on a lock that never saw a response because
their request was freed with the lock object.
The fix is to hold the rbtree lock when testing the refcount and
deciding to free. It adds a bit more contention but not significantly
so, given the wild existing contention on a per-fs spinlocked rbtree.
Signed-off-by: Zach Brown <zab@versity.com>
Previously, any t_skip would cause the final test result to be a failure
because up until now no test should have been skipped.
However, with format-version-forward-back not being compatible with el9,
we are going to rely on el7/8 testing for that test soleley, and
therefore we have to allow skipping of this test on el9 and newer OS
versions.
We add `t_skip_permitted` to signal this from the test case to the
run-tests.sh script. A new exit code is passed, and all accounting is
updated to reflect that a test was skipped, but this was permitted. We
modify format-version-forward-back to use this new exit path.
Signed-off-by: Auke Kok <auke.kok@versity.com>
I'm seeing more and more of these as audit is enabled in el8 and el9
images I am using for testing, and during ENOSPC tests this has a chance
of triggering process accounting suspension, and subsequent resume.
Signed-off-by: Auke Kok <auke.kok@versity.com>
In v1.18-10-g5507ee5, we changed the test code away from loopback
to device-mapper, which simplified our DUT setup code.
However, this results in the occasional `device changed size` messages
now being emitted by the `dm` driver instead of the `loop` kernel
module. We have to additionally ignore these kernel messages from now as
well.
Signed-off-by: Auke Kok <auke.kok@versity.com>
In v5.17-rc4-53-g3a3bae50af5d, we can no longer omit having this
method unhooked as the mm caller blindly calls it now. In-kernel
filesystems all were fixed in this change.
aops->invalidatepage was the old aops method that would free pages
with private attached data. This method is replaced with the
new invalidate_folio method. If this method is NULL, the memory
will become orphaned. (v5.17-rc4-29-gf50015a596fa)
Signed-off-by: Auke Kok <auke.kok@versity.com>
I've pushed a tag/release to scoutfs-xfstests-dev instead of a full
blown branch. This seems simpler and cleaner than using branches,
because we're going to end up rebasing these things a lot. However, we
can't --track tags, so, if the branch name passed to -x is actually a
tag instead of a branch, we have to omit the --track option here.
Signed-off-by: Auke Kok <auke.kok@versity.com>
CI testing needs to know which xfstests branch to use on all OSs.
We can't just use the el9 xfstests branch on el9 only, because we
need to run the same el9 xfstests on el8 and el7 as well, otherwise
testing will just fail.
So, we put a marker file in our git repo that tells us that we're
not going to use the default `scoutfs` branch from scoutfs-xfstests-dev
but our own special tag or branch. The CI job then should pass the
proper -x {branch} flag to the run-tests.sh script.
Signed-off-by: Auke Kok <auke.kok@versity.com>
The new version of xfstests adds a _lot_ more tests to our mix. Many
of the new ones will auto enable or auto skip as needed.
There are tests we can't or won't support that will be in future
xfstests. Disable them now so we can avoid dealing with them later.
Quite a few fall into "we don't support these types of mounting yet",
mostly bind-mount or dm-mapper things. We disable all the swapfile
tests flatout.
A few tests fail on el7 but not el8/9 but we don't have a way to run
them without failing yet, so disable them as well.
Update golden with the proper new array of tests. This all requires
the `auke/scoutfs-el9` branch in `versity/scoutfs-xfstests-dev`.
Signed-off-by: Auke Kok <auke.kok@versity.com>
Using t_skip, we just skip this test on el9.
If we ever want to add a formatversion 2->3 test, perhaps we should
just add a separate test script, instead of going over a static array.
But let's not worry about this too much right now.
Signed-off-by: Auke Kok <auke.kok@versity.com>
It turns out that on el9, `bash -c` prints out `bash: line 1: cd..`
instead of `line 0:` on el7 or el8. So discard all the stderr from
these `cd` lines entirely and just rely on the expected echo
output to stdout.
Signed-off-by: Auke Kok <auke.kok@versity.com>
There's filefrag already, and that works, but, it's output is very
inconsistent between various OS release versions, and it has already
meant that we'd needed to adjust tests to account for these little
but insignificant changes. A lot more work than useful. It's even
more changed in el9.
This adds `scoutfs get-fiemap FILE` and prints out block extent
info with flags that we care about as an abbreviated letter: U for
Unwritten, L for Last, and O for Unknown (as in, "offline").
The -P/--physical and -L/--logical options turn off logical or physical
offset display, in case you only want to see the offsets in either
units. You can pass -b/--byte to display offsets and lengths in
byte values. The block size will then be obtained from fstat() of
the queried file (4096 for scoutfs).
I've removed all uses of filefrag from our scoutfs tests. Xfstests
still calls it but their internal diff takes care of that issue.
Where needed and appropriate, the tests are adjusted so that the output
of `scoutfs get-fiemap` is as close as it can to what it used to be,
so that reading the test results allows the quick view of what might
have been going wrong.
There are some output strings I have not bothered to update because
there's no real value to updating every output string to match,
and we just adjust the golden file accordingly.
Signed-off-by: Auke Kok <auke.kok@versity.com>
This isn't a simple case where we can use u64_region_wraps because
length is s32.
Let's actually test an overflow case instead of a case that doesn't
overflow, though. We still should properly add an overflow test here as
well.
Signed-off-by: Auke Kok <auke.kok@versity.com>
We use check_add_overflow(a, b, d) here to validate that (off, len)
pairs do not exceed the max value type. The kernel conveniently has
several macros to sort out the problems with signed or unsigned types.
However, we're not interested in purely seeing whether (a + b)
overflows, because we're using this for (off, len) overflow checks,
where the bytes we read are from 0 to len -1. We must therefore call
this check with (b) being "len - 1".
I've made sure that we don't accidentally fail when (len == 0)
in all cases by making sure we've already checked this condition
before, and moving code around as needed to ensure that (len > 0)
in all cases where we check.
The macro check_add_overflow requires a (d) argument in which
temporarily the result of the addition is stored and then checked to see
if an overflow occurred. We put a `tmp` variable on the stack of the
correct type as needed to make the checks function.
simple-release-extents test mistakenly relied on this buggy wrap code,
so it needs fixing. The move-blocks test also got it wrong.
Signed-off-by: Auke Kok <auke.kok@versity.com>
We consistently enter scoutfs_data_wait_check when len == 0 from
scoutfs_aio_write() which directly passes the i_size_read() value,
and for cases where we `echo >> $FILE` this is always reached.
This can cause the wrapping check to fail since `0 + (0 - 1) < 0` which
triggers the WARN_ON_ONCE wrap check that needs updating to allow
certain operations on huge files.
More importantly we can just omit all these checks if `len == 0` anyway,
since they should always succeed and never should require taking all the
locks.
Signed-off-by: Auke Kok <auke.kok@versity.com>
getline() allocates the space for the return value even if there is an
error, so when it returns an error, we still have to free() it.
In el9, when reading stdin we will get errno=0 returned (no error) when
we hit the end of stdin. This behavior is different from el7/8. We don't
want to throw an error here to avoid failing the test, since it doesn't.
Signed-off-by: Auke Kok <auke.kok@versity.com>
The warnings thrown by el9's version of xargs are unexpected output and
cause this test to fail. When using the -I option (replace) the -n 1
arguments are always assumed. In el7/8 no warnings were printed.
We can just remove `-n 1` since the argument is never needed.
Signed-off-by: Auke Kok <auke.kok@versity.com>
In el9 releases, our includes declare offsetof() before our header
chain includes stddef.h, which doesn't properly check if offsetof
is already defined, leading to a redefinition. Just include stddef
at all times here.
Signed-off-by: Auke Kok <auke.kok@versity.com>
Passing a holder ptr to these functions now replaces the FMODE_EXCL
flag. _put no longer needs flags for this reason, but the holder
instead.
Signed-off-by: Auke Kok <auke.kok@versity.com>
v6.4-rc2-198-g05bdb9965305 adds a new type for passing flags instead
of abusing fmode_t flags. They are essentially the same flags just
in a new type.
Signed-off-by: Auke Kok <auke.kok@versity.com>
v6.0-rc6-9-g863f144f12ad changes the VFS method to pass in a struct
file and not a dentry in preperation for tmpfile support in fuse.
Signed-off-by: Auke Kok <auke.kok@versity.com>
The current spec template can't handle future major el releases
gracefully and fails to build entirely. We isolate all changes
so that they are either "el7 specific" or generic. This rids us
entirely of el8 specific conditionals.
Signed-off-by: Auke Kok <auke.kok@versity.com>
Folios are the new data types used for passing pages. For now,
folios only appear to have a single page. Future kernels will
change that.
Signed-off-by: Auke Kok <auke.kok@versity.com>
