| Age | Commit message (Collapse) | Author | Files | Lines |
|---|
|
There are several different contexts xfs_bmap_btalloc() handles, and
large chunks of the code execute independent allocation contexts.
Try to untangle this mess a bit.
Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Darrick J. Wong <djwong@kernel.org>
|
|
Change obvious callers of single AG allocation to use
xfs_alloc_vextent_this_ag(). Drive the per-ag grabbing out to the
callers, too, so that callers with active references don't need
to do new lookups just for an allocation in a context that already
has a perag reference.
The only remaining caller that does single AG allocation through
xfs_alloc_vextent() is xfs_bmap_btalloc() with
XFS_ALLOCTYPE_NEAR_BNO. That is going to need more untangling before
it can be converted cleanly.
Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Darrick J. Wong <djwong@kernel.org>
|
|
There's a bit of a recursive conundrum around
xfs_alloc_ag_vextent(). We can't first call xfs_alloc_ag_vextent()
without preparing the AGFL for the allocation, and preparing the
AGFL calls xfs_alloc_ag_vextent() to prepare the AGFL for the
allocation. This "double allocation" requirement is not really clear
from the current xfs_alloc_fix_freelist() calls that are sprinkled
through the allocation code.
It's not helped that xfs_alloc_ag_vextent() can actually allocate
from the AGFL itself, but there's special code to prevent AGFL prep
allocations from allocating from the free list it's trying to prep.
The naming is also not consistent: args->wasfromfl is true when we
allocated _from_ the free list, but the indication that we are
allocating _for_ the free list is via checking that (args->resv ==
XFS_AG_RESV_AGFL).
So, lets make this "allocation required for allocation" situation
clear by moving it all inside xfs_alloc_ag_vextent(). The freelist
allocation is a specific XFS_ALLOCTYPE_THIS_AG allocation, which
translated directly to xfs_alloc_ag_vextent_size() allocation.
This enables us to replace __xfs_alloc_vextent_this_ag() with a call
to xfs_alloc_ag_vextent(), and we drive the freelist fixing further
into the per-ag allocation algorithm.
Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Darrick J. Wong <djwong@kernel.org>
|
|
The core of the per-ag iteration is effectively doing a "this ag"
allocation on one AG at a time. Use the same code to implement the
core "this ag" allocation in both xfs_alloc_vextent_this_ag()
and xfs_alloc_vextent_iterate_ags().
This means we only call xfs_alloc_ag_vextent() from one place so we
can easily collapse the call stack in future patches.
Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Darrick J. Wong <djwong@kernel.org>
|
|
It's a multiplexing mess that can be greatly simplified, and really
needs to be simplified to allow active per-ag references to
propagate from initial AG selection code the the bmapi code.
This splits the code out into separate a parameter checking
function, an iterator function, and allocation completion functions
and then implements the individual policies using these functions.
Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Darrick J. Wong <djwong@kernel.org>
|
|
In several places we iterate every AG from a specific start agno and
wrap back to the first AG when we reach the end of the filesystem to
continue searching. We don't have a primitive for this iteration
yet, so add one for conversion of these algorithms to per-ag based
iteration.
The filestream AG select code is a mess, and this initially makes it
worse. The per-ag selection needs to be driven completely into the
filestream code to clean this up and it will be done in a future
patch that makes the filestream allocator use active per-ag
references correctly.
Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Darrick J. Wong <djwong@kernel.org>
|
|
We currently don't have any flags or operational state in the
xfs_perag except for the pagf_init and pagi_init flags. And the
agflreset flag. Oh, there's also the pagf_metadata and pagi_inodeok
flags, too.
For controlling per-ag operations, we are going to need some atomic
state flags. Hence add an opstate field similar to what we already
have in the mount and log, and convert all these state flags across
to atomic bit operations.
Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Allison Henderson <allison.henderson@oracle.com>
Reviewed-by: Darrick J. Wong <djwong@kernel.org>
|
|
This is currently a spinlock lock protected rotor which can be
implemented with a single atomic operation. Change it to be more
efficient and get rid of the m_agirotor_lock. Noticed while
converting the inode allocation AG selection loop to active perag
references.
Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Allison Henderson <allison.henderson@oracle.com>
Reviewed-by: Darrick J. Wong <djwong@kernel.org>
|
|
Lots of code in the inobt infrastructure is passed both xfs_mount
and perags. We only need perags for the per-ag inode allocation
code, so reduce the duplication by passing only the perags as the
primary object.
This ends up reducing the code size by a bit:
text data bss dec hex filename
orig 1138878 323979 548 1463405 16546d (TOTALS)
patched 1138709 323979 548 1463236 1653c4 (TOTALS)
Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Allison Henderson <allison.henderson@oracle.com>
Reviewed-by: Darrick J. Wong <djwong@kernel.org>
|
|
Convert the inode allocation routines to use active perag references
or references held by callers rather than grab their own. Also drive
the perag further inwards to replace xfs_mounts when doing
operations on a specific AG.
Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Allison Henderson <allison.henderson@oracle.com>
Reviewed-by: Darrick J. Wong <djwong@kernel.org>
|
|
Callers have referenced perags but they don't pass it into
xfs_imap() so it takes it's own reference. Fix that so we can change
inode allocation over to using active references.
Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Darrick J. Wong <djwong@kernel.org>
|
|
So that they all output the same information in the traces to make
debugging refcount issues easier.
This means that all the lookup/drop functions no longer need to use
the full memory barrier atomic operations (atomic*_return()) so
will have less overhead when tracing is off. The set/clear tag
tracepoints no longer abuse the reference count to pass the tag -
the tag being cleared is obvious from the _RET_IP_ that is recorded
in the trace point.
Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Allison Henderson <allison.henderson@oracle.com>
Reviewed-by: Darrick J. Wong <djwong@kernel.org>
|
|
We need to be able to dynamically remove instantiated AGs from
memory safely, either for shrinking the filesystem or paging AG
state in and out of memory (e.g. supporting millions of AGs). This
means we need to be able to safely exclude operations from accessing
perags while dynamic removal is in progress.
To do this, introduce the concept of active and passive references.
Active references are required for high level operations that make
use of an AG for a given operation (e.g. allocation) and pin the
perag in memory for the duration of the operation that is operating
on the perag (e.g. transaction scope). This means we can fail to get
an active reference to an AG, hence callers of the new active
reference API must be able to handle lookup failure gracefully.
Passive references are used in low level code, where we might need
to access the perag structure for the purposes of completing high
level operations. For example, buffers need to use passive
references because:
- we need to be able to do metadata IO during operations like grow
and shrink transactions where high level active references to the
AG have already been blocked
- buffers need to pin the perag until they are reclaimed from
memory, something that high level code has no direct control over.
- unused cached buffers should not prevent a shrink from being
started.
Hence we have active references that will form exclusion barriers
for operations to be performed on an AG, and passive references that
will prevent reclaim of the perag until all objects with passive
references have been reclaimed themselves.
This patch introduce xfs_perag_grab()/xfs_perag_rele() as the API
for active AG reference functionality. We also need to convert the
for_each_perag*() iterators to use active references, which will
start the process of converting high level code over to using active
references. Conversion of non-iterator based code to active
references will be done in followup patches.
Note that the implementation using reference counting is really just
a development vehicle for the API to ensure we don't have any leaks
in the callers. Once we need to remove perag structures from memory
dyanmically, we will need a much more robust per-ag state transition
mechanism for preventing new references from being taken while we
wait for existing references to drain before removal from memory can
occur....
Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Allison Henderson <allison.henderson@oracle.com>
Reviewed-by: Darrick J. Wong <djwong@kernel.org>
|
|
The tp->t_firstblock field is now raelly tracking the highest AG we
have locked, not the block number of the highest allocation we've
made. It's purpose is to prevent AGF locking deadlocks, so rename it
to "highest AG" and simplify the implementation to just track the
agno rather than a fsbno.
Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Allison Henderson <allison.henderson@oracle.com>
Reviewed-by: Darrick J. Wong <djwong@kernel.org>
|
|
Now that xfs_alloc_vextent() does all the AGF deadlock prevention
filtering for multiple allocations in a single transaction, we no
longer need the allocation setup code to care about what AGs we
might already have locked.
Hence we can remove all the "nullfb" conditional logic in places
like xfs_bmap_btalloc() and instead have them focus simply on
setting up locality constraints. If the allocation fails due to
AGF lock filtering in xfs_alloc_vextent, then we just fall back as
we normally do to more relaxed allocation constraints.
As a result, any allocation that allows AG scanning (i.e. not
confined to a single AG) and does not force a worst case full
filesystem scan will now be able to attempt allocation from AGs
lower than that defined by tp->t_firstblock. This is because
xfs_alloc_vextent() allows try-locking of the AGFs and hence enables
low space algorithms to at least -try- to get space from AGs lower
than the one that we have currently locked and allocated from. This
is a significant improvement in the low space allocation algorithm.
Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Allison Henderson <allison.henderson@oracle.com>
Reviewed-by: Darrick J. Wong <djwong@kernel.org>
|
|
When we enter xfs_bmbt_alloc_block() without having first allocated
a data extent (i.e. tp->t_firstblock == NULLFSBLOCK) because we
are doing something like unwritten extent conversion, the transaction
block reservation is used as the minleft value.
This works for operations like unwritten extent conversion, but it
assumes that the block reservation is only for a BMBT split. THis is
not always true, and sometimes results in larger than necessary
minleft values being set. We only actually need enough space for a
btree split, something we already handle correctly in
xfs_bmapi_write() via the xfs_bmapi_minleft() calculation.
We should use xfs_bmapi_minleft() in xfs_bmbt_alloc_block() to
calculate the number of blocks a BMBT split on this inode is going to
require, not use the transaction block reservation that contains the
maximum number of blocks this transaction may consume in it...
Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Allison Henderson <allison.henderson@oracle.com>
Reviewed-by: Darrick J. Wong <djwong@kernel.org>
|
|
When an XFS filesystem has free inodes in chunks already allocated
on disk, it will still allocate new inode chunks if the target AG
has no free inodes in it. Normally, this is a good idea as it
preserves locality of all the inodes in a given directory.
However, at ENOSPC this can lead to using the last few remaining
free filesystem blocks to allocate a new chunk when there are many,
many free inodes that could be allocated without consuming free
space. This results in speeding up the consumption of the last few
blocks and inode create operations then returning ENOSPC when there
free inodes available because we don't have enough block left in the
filesystem for directory creation reservations to proceed.
Hence when we are near ENOSPC, we should be attempting to preserve
the remaining blocks for directory block allocation rather than
using them for unnecessary inode chunk creation.
This particular behaviour is exposed by xfs/294, when it drives to
ENOSPC on empty file creation whilst there are still thousands of
free inodes available for allocation in other AGs in the filesystem.
Hence, when we are within 1% of ENOSPC, change the inode allocation
behaviour to prefer to use existing free inodes over allocating new
inode chunks, even though it results is poorer locality of the data
set. It is more important for the allocations to be space efficient
near ENOSPC than to have optimal locality for performance, so lets
modify the inode AG selection code to reflect that fact.
This allows generic/294 to not only pass with this allocator rework
patchset, but to increase the number of post-ENOSPC empty inode
allocations to from ~600 to ~9080 before we hit ENOSPC on the
directory create transaction reservation.
Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Allison Henderson <allison.henderson@oracle.com>
Reviewed-by: Darrick J. Wong <djwong@kernel.org>
|
|
I've recently encountered an ABBA deadlock with g/476. The upcoming
changes seem to make this much easier to hit, but the underlying
problem is a pre-existing one.
Essentially, if we select an AG for allocation, then lock the AGF
and then fail to allocate for some reason (e.g. minimum length
requirements cannot be satisfied), then we drop out of the
allocation with the AGF still locked.
The caller then modifies the allocation constraints - usually
loosening them up - and tries again. This can result in trying to
access AGFs that are lower than the AGF we already have locked from
the failed attempt. e.g. the failed attempt skipped several AGs
before failing, so we have locks an AG higher than the start AG.
Retrying the allocation from the start AG then causes us to violate
AGF lock ordering and this can lead to deadlocks.
The deadlock exists even if allocation succeeds - we can do a
followup allocations in the same transaction for BMBT blocks that
aren't guaranteed to be in the same AG as the original, and can move
into higher AGs. Hence we really need to move the tp->t_firstblock
tracking down into xfs_alloc_vextent() where it can be set when we
exit with a locked AG.
xfs_alloc_vextent() can also check there if the requested
allocation falls within the allow range of AGs set by
tp->t_firstblock. If we can't allocate within the range set, we have
to fail the allocation. If we are allowed to to non-blocking AGF
locking, we can ignore the AG locking order limitations as we can
use try-locks for the first iteration over requested AG range.
This invalidates a set of post allocation asserts that check that
the allocation is always above tp->t_firstblock if it is set.
Because we can use try-locks to avoid the deadlock in some
circumstances, having a pre-existing locked AGF doesn't always
prevent allocation from lower order AGFs. Hence those ASSERTs need
to be removed.
Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Allison Henderson <allison.henderson@oracle.com>
Reviewed-by: Darrick J. Wong <djwong@kernel.org>
|
|
When we split a BMBT due to record insertion, we offload it to a
worker thread because we can be deep in the stack when we try to
allocate a new block for the BMBT. Allocation can use several
kilobytes of stack (full memory reclaim, swap and/or IO path can
end up on the stack during allocation) and we can already be several
kilobytes deep in the stack when we need to split the BMBT.
A recent workload demonstrated a deadlock in this BMBT split
offload. It requires several things to happen at once:
1. two inodes need a BMBT split at the same time, one must be
unwritten extent conversion from IO completion, the other must be
from extent allocation.
2. there must be a no available xfs_alloc_wq worker threads
available in the worker pool.
3. There must be sustained severe memory shortages such that new
kworker threads cannot be allocated to the xfs_alloc_wq pool for
both threads that need split work to be run
4. The split work from the unwritten extent conversion must run
first.
5. when the BMBT block allocation runs from the split work, it must
loop over all AGs and not be able to either trylock an AGF
successfully, or each AGF is is able to lock has no space available
for a single block allocation.
6. The BMBT allocation must then attempt to lock the AGF that the
second task queued to the rescuer thread already has locked before
it finds an AGF it can allocate from.
At this point, we have an ABBA deadlock between tasks queued on the
xfs_alloc_wq rescuer thread and a locked AGF. i.e. The queued task
holding the AGF lock can't be run by the rescuer thread until the
task the rescuer thread is runing gets the AGF lock....
This is a highly improbably series of events, but there it is.
There's a couple of ways to fix this, but the easiest way to ensure
that we only punt tasks with a locked AGF that holds enough space
for the BMBT block allocations to the worker thread.
This works for unwritten extent conversion in IO completion (which
doesn't have a locked AGF and space reservations) because we have
tight control over the IO completion stack. It is typically only 6
functions deep when xfs_btree_split() is called because we've
already offloaded the IO completion work to a worker thread and
hence we don't need to worry about stack overruns here.
The other place we can be called for a BMBT split without a
preceeding allocation is __xfs_bunmapi() when punching out the
center of an existing extent. We don't remove extents in the IO
path, so these operations don't tend to be called with a lot of
stack consumed. Hence we don't really need to ship the split off to
a worker thread in these cases, either.
Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Darrick J. Wong <djwong@kernel.org>
Signed-off-by: Darrick J. Wong <djwong@kernel.org>
|
|
Pass the incore refcount intent through the CUI logging code instead of
repeatedly boxing and unboxing parameters.
Signed-off-by: Darrick J. Wong <djwong@kernel.org>
|
|
Pass the incore rmap space mapping through the RUI logging code instead
of repeatedly boxing and unboxing parameters.
Signed-off-by: Darrick J. Wong <djwong@kernel.org>
|
|
Change the name of all pointers to xfs_extent_item structures to "xefi"
to make the name consistent and because the current selections ("new"
and "free") mean other things in C.
Signed-off-by: Darrick J. Wong <djwong@kernel.org>
|
|
Instead of repeatedly boxing and unboxing the incore extent mapping
structure as it passes through the BUI code, pass the pointer directly
through.
Signed-off-by: Darrick J. Wong <djwong@kernel.org>
|
|
Lately I've been stress-testing extreme-sized rmap btrees by using the
(new) xfs_db bmap_inflate command to clone bmbt mappings billions of
times and then using xfs_repair to build new rmap and refcount btrees.
This of course is /much/ faster than actually FICLONEing a file billions
of times.
Unfortunately, xfs_repair fails in xfs_btree_bload_compute_geometry with
EOVERFLOW, which indicates that xfs_mount.m_rmap_maxlevels is not
sufficiently large for the test scenario. For a 1TB filesystem (~67
million AG blocks, 4 AGs) the btheight command reports:
$ xfs_db -c 'btheight -n 4400801200 -w min rmapbt' /dev/sda
rmapbt: worst case per 4096-byte block: 84 records (leaf) / 45 keyptrs (node)
level 0: 4400801200 records, 52390491 blocks
level 1: 52390491 records, 1164234 blocks
level 2: 1164234 records, 25872 blocks
level 3: 25872 records, 575 blocks
level 4: 575 records, 13 blocks
level 5: 13 records, 1 block
6 levels, 53581186 blocks total
The AG is sufficiently large to build this rmap btree. Unfortunately,
m_rmap_maxlevels is 5. Augmenting the loop in the space->height
function to report height, node blocks, and blocks remaining produces
this:
ht 1 node_blocks 45 blockleft 67108863
ht 2 node_blocks 2025 blockleft 67108818
ht 3 node_blocks 91125 blockleft 67106793
ht 4 node_blocks 4100625 blockleft 67015668
final height: 5
The goal of this function is to compute the maximum height btree that
can be stored in the given number of ondisk fsblocks. Starting with the
top level of the tree, each iteration through the loop adds the fanout
factor of the next level down until we run out of blocks. IOWs, maximum
height is achieved by using the smallest fanout factor that can apply
to that level.
However, the loop setup is not correct. Top level btree blocks are
allowed to contain fewer than minrecs items, so the computation is
incorrect because the first time through the loop it should be using a
fanout factor of 2. With this corrected, the above becomes:
ht 1 node_blocks 2 blockleft 67108863
ht 2 node_blocks 90 blockleft 67108861
ht 3 node_blocks 4050 blockleft 67108771
ht 4 node_blocks 182250 blockleft 67104721
ht 5 node_blocks 8201250 blockleft 66922471
final height: 6
Fixes: 9ec691205e7d ("xfs: compute the maximum height of the rmap btree when reflink enabled")
Signed-off-by: Darrick J. Wong <djwong@kernel.org>
Reviewed-by: Dave Chinner <dchinner@redhat.com>
|
|
Pull XFS updates from Darrick Wong:
"The highlight of this is a batch of fixes for the online metadata
checking code as we start the loooong march towards merging online
repair. I aim to merge that in time for the 2023 LTS.
There are also a large number of data corruption and race condition
fixes in this patchset. Most notably fixed are write() calls to
unwritten extents racing with writeback, which required some late(r
than I prefer) code changes to iomap to support the necessary
revalidations. I don't really like iomap changes going in past -rc4,
but Dave and I have been working on it long enough that I chose to
push it for 6.2 anyway.
There are also a number of other subtle problems fixed, including the
log racing with inode writeback to write inodes with incorrect link
count to disk; file data mapping corruptions as a result of incorrect
lock cycling when attaching dquots; refcount metadata corruption if
one actually manages to share a block 2^32 times; and the log
clobbering cow staging extents if they were formerly metadata blocks.
Summary:
- Fix a race condition w.r.t. percpu inode free counters
- Fix a broken error return in xfs_remove
- Print FS UUID at mount/unmount time
- Numerous fixes to the online fsck code
- Fix inode locking inconsistency problems when dealing with realtime
metadata files
- Actually merge pull requests so that we capture the cover letter
contents
- Fix a race between rebuilding VFS inode state and the AIL flushing
inodes that could cause corrupt inodes to be written to the
filesystem
- Fix a data corruption problem resulting from a write() to an
unwritten extent racing with writeback started on behalf of memory
reclaim changing the extent state
- Add debugging knobs so that we can test iomap invalidation
- Fix the blockdev pagecache contents being stale after unmounting
the filesystem, leading to spurious xfs_db errors and corrupt
metadumps
- Fix a file mapping corruption bug due to ilock cycling when
attaching dquots to a file during delalloc reservation
- Fix a refcount btree corruption problem due to the refcount
adjustment code not handling MAXREFCOUNT correctly, resulting in
unnecessary record splits
- Fix COW staging extent alloctions not being classified as USERDATA,
which results in filestreams being ignored and possible data
corruption if the allocation was filled from the AGFL and the block
buffer is still being tracked in the AIL
- Fix new duplicated includes
- Fix a race between the dquot shrinker and dquot freeing that could
cause a UAF"
* tag 'xfs-6.2-merge-8' of git://git.kernel.org/pub/scm/fs/xfs/xfs-linux: (50 commits)
xfs: dquot shrinker doesn't check for XFS_DQFLAG_FREEING
xfs: Remove duplicated include in xfs_iomap.c
xfs: invalidate xfs_bufs when allocating cow extents
xfs: get rid of assert from xfs_btree_islastblock
xfs: estimate post-merge refcounts correctly
xfs: hoist refcount record merge predicates
xfs: fix super block buf log item UAF during force shutdown
xfs: wait iclog complete before tearing down AIL
xfs: attach dquots to inode before reading data/cow fork mappings
xfs: shut up -Wuninitialized in xfsaild_push
xfs: use memcpy, not strncpy, to format the attr prefix during listxattr
xfs: invalidate block device page cache during unmount
xfs: add debug knob to slow down write for fun
xfs: add debug knob to slow down writeback for fun
xfs: drop write error injection is unfixable, remove it
xfs: use iomap_valid method to detect stale cached iomaps
iomap: write iomap validity checks
xfs: xfs_bmap_punch_delalloc_range() should take a byte range
iomap: buffered write failure should not truncate the page cache
xfs,iomap: move delalloc punching to iomap
...
|
|
While investigating test failures in xfs/17[1-3] in alwayscow mode, I
noticed through code inspection that xfs_bmap_alloc_userdata isn't
setting XFS_ALLOC_USERDATA when allocating extents for a file's CoW
fork. COW staging extents should be flagged as USERDATA, since user
data are persisted to these blocks before being remapped into a file.
This mis-classification has a few impacts on the behavior of the system.
First, the filestreams allocator is supposed to keep allocating from a
chosen AG until it runs out of space in that AG. However, it only does
that for USERDATA allocations, which means that COW allocations aren't
tied to the filestreams AG. Fortunately, few people use filestreams, so
nobody's noticed.
A more serious problem is that xfs_alloc_ag_vextent_small looks for a
buffer to invalidate *if* the USERDATA flag is set and the AG is so full
that the allocation had to come from the AGFL because the cntbt is
empty. The consequences of not invalidating the buffer are severe --
if the AIL incorrectly checkpoints a buffer that is now being used to
store user data, that action will clobber the user's written data.
Fix filestreams and yet another data corruption vector by flagging COW
allocations as USERDATA.
Signed-off-by: Darrick J. Wong <djwong@kernel.org>
Reviewed-by: Dave Chinner <dchinner@redhat.com>
|
|
xfs_btree_check_block contains debugging knobs. With XFS_DEBUG setting up,
turn on the debugging knob can trigger the assert of xfs_btree_islastblock,
test script as follows:
while true
do
mount $disk $mountpoint
fsstress -d $testdir -l 0 -n 10000 -p 4 >/dev/null
echo 1 > /sys/fs/xfs/sda/errortag/btree_chk_sblk
sleep 10
umount $mountpoint
done
Kick off fsstress and only *then* turn on the debugging knob. If it
happens that the knob gets turned on after the cntbt lookup succeeds
but before the call to xfs_btree_islastblock, then we *can* end up in
the situation where a previously checked btree block suddenly starts
returning EFSCORRUPTED from xfs_btree_check_block. Kaboom.
Darrick give a very detailed explanation as follows:
Looking back at commit 27d9ee577dcce, I think the point of all this was
to make sure that the cursor has actually performed a lookup, and that
the btree block at whatever level we're asking about is ok.
If the caller hasn't ever done a lookup, the bc_levels array will be
empty, so cur->bc_levels[level].bp pointer will be NULL. The call to
xfs_btree_get_block will crash anyway, so the "ASSERT(block);" part is
pointless.
If the caller did a lookup but the lookup failed due to block
corruption, the corresponding cur->bc_levels[level].bp pointer will also
be NULL, and we'll still crash. The "ASSERT(xfs_btree_check_block);"
logic is also unnecessary.
If the cursor level points to an inode root, the block buffer will be
incore, so it had better always be consistent.
If the caller ignores a failed lookup after a successful one and calls
this function, the cursor state is garbage and the assert wouldn't have
tripped anyway. So get rid of the assert.
Fixes: 27d9ee577dcc ("xfs: actually check xfs_btree_check_block return in xfs_btree_islastblock")
Signed-off-by: Guo Xuenan <guoxuenan@huawei.com>
Reviewed-by: Darrick J. Wong <djwong@kernel.org>
Signed-off-by: Darrick J. Wong <djwong@kernel.org>
|
|
Upon enabling fsdax + reflink for XFS, xfs/179 began to report refcount
metadata corruptions after being run. Specifically, xfs_repair noticed
single-block refcount records that could be combined but had not been.
The root cause of this is improper MAXREFCOUNT edge case handling in
xfs_refcount_merge_extents. When we're trying to find candidates for a
refcount btree record merge, we compute the refcount attribute of the
merged record, but we fail to account for the fact that once a record
hits rc_refcount == MAXREFCOUNT, it is pinned that way forever. Hence
the computed refcount is wrong, and we fail to merge the extents.
Fix this by adjusting the merge predicates to compute the adjusted
refcount correctly.
Fixes: 3172725814f9 ("xfs: adjust refcount of an extent of blocks in refcount btree")
Signed-off-by: Darrick J. Wong <djwong@kernel.org>
Reviewed-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Xiao Yang <yangx.jy@fujitsu.com>
|
|
Hoist these multiline conditionals into separate static inline helpers
to improve readability and set the stage for corruption fixes that will
be introduced in the next patch.
Signed-off-by: Darrick J. Wong <djwong@kernel.org>
Reviewed-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Xiao Yang <yangx.jy@fujitsu.com>
|
|
Add a new error injection knob so that we can arbitrarily slow down
pagecache writes to test for race conditions and aberrant reclaim
behavior if the writeback mechanisms are slow to issue writeback. This
will enable functional testing for the ifork sequence counters
introduced in commit 304a68b9c63b ("xfs: use iomap_valid method to
detect stale cached iomaps") that fixes write racing with reclaim
writeback.
Signed-off-by: Darrick J. Wong <djwong@kernel.org>
Reviewed-by: Dave Chinner <dchinner@redhat.com>
|
|
Add a new error injection knob so that we can arbitrarily slow down
writeback to test for race conditions and aberrant reclaim behavior if
the writeback mechanisms are slow to issue writeback. This will enable
functional testing for the ifork sequence counters introduced in commit
745b3f76d1c8 ("xfs: maintain a sequence count for inode fork
manipulations").
Signed-off-by: Darrick J. Wong <djwong@kernel.org>
Reviewed-by: Dave Chinner <dchinner@redhat.com>
|
|
git://git.kernel.org/pub/scm/linux/kernel/git/dgc/linux-xfs into xfs-6.2-mergeB
xfs, iomap: fix data corruption due to stale cached iomaps
This patch series fixes a data corruption that occurs in a specific
multi-threaded write workload. The workload combined
racing unaligned adjacent buffered writes with low memory conditions
that caused both writeback and memory reclaim to race with the
writes.
The result of this was random partial blocks containing zeroes
instead of the correct data. The underlying problem is that iomap
caches the write iomap for the duration of the write() operation,
but it fails to take into account that the extent underlying the
iomap can change whilst the write is in progress.
The short story is that an iomap can span mutliple folios, and so
under low memory writeback can be cleaning folios the write()
overlaps. Whilst the overlapping data is cached in memory, this
isn't a problem, but because the folios are now clean they can be
reclaimed. Once reclaimed, the write() does the wrong thing when
re-instantiating partial folios because the iomap no longer reflects
the underlying state of the extent. e.g. it thinks the extent is
unwritten, so it zeroes the partial range, when in fact the
underlying extent is now written and so it should have read the data
from disk. This is how we get random zero ranges in the file
instead of the correct data.
The gory details of the race condition can be found here:
https://lore.kernel.org/linux-xfs/20220817093627.GZ3600936@dread.disaster.area/
Fixing the problem has two aspects. The first aspect of the problem
is ensuring that iomap can detect a stale cached iomap during a
write in a race-free manner. We already do this stale iomap
detection in the writeback path, so we have a mechanism for
detecting that the iomap backing the data range may have changed
and needs to be remapped.
In the case of the write() path, we have to ensure that the iomap is
validated at a point in time when the page cache is stable and
cannot be reclaimed from under us. We also need to validate the
extent before we start performing any modifications to the folio
state or contents. Combine these two requirements together, and the
only "safe" place to validate the iomap is after we have looked up
and locked the folio we are going to copy the data into, but before
we've performed any initialisation operations on that folio.
If the iomap fails validation, we then mark it stale, unlock the
folio and end the write. This effectively means a stale iomap
results in a short write. Filesystems should already be able to
handle this, as write operations can end short for many reasons and
need to iterate through another mapping cycle to be completed. Hence
the iomap changes needed to detect and handle stale iomaps during
write() operations is relatively simple...
However, the assumption is that filesystems should already be able
to handle write failures safely, and that's where the second
(first?) part of the problem exists. That is, handling a partial
write is harder than just "punching out the unused delayed
allocation extent". This is because mmap() based faults can race
with writes, and if they land in the delalloc region that the write
allocated, then punching out the delalloc region can cause data
corruption.
This data corruption problem is exposed by generic/346 when iomap is
converted to detect stale iomaps during write() operations. Hence
write failure in the filesytems needs to handle the fact that the
write() in progress doesn't necessarily own the data in the page
cache over the range of the delalloc extent it just allocated.
As a result, we can't just truncate the page cache over the range
the write() didn't reach and punch all the delalloc extent. We have
to walk the page cache over the untouched range and skip over any
dirty data region in the cache in that range. Which is ....
non-trivial.
That is, iterating the page cache has to handle partially populated
folios (i.e. block size < page size) that contain data. The data
might be discontiguous within a folio. Indeed, there might be
*multiple* discontiguous data regions within a single folio. And to
make matters more complex, multi-page folios mean we just don't know
how many sub-folio regions we might have to iterate to find all
these regions. All the corner cases between the conversions and
rounding between filesystem block size, folio size and multi-page
folio size combined with unaligned write offsets kept breaking my
brain.
However, if we convert the code to track the processed
write regions by byte ranges instead of fileystem block or page
cache index, we could simply use mapping_seek_hole_data() to find
the start and end of each discrete data region within the range we
needed to scan. SEEK_DATA finds the start of the cached data region,
SEEK_HOLE finds the end of the region. These are byte based
interfaces that understand partially uptodate folio regions, and so
can iterate discrete sub-folio data regions directly. This largely
solved the problem of discovering the dirty regions we need to keep
the delalloc extent over.
However, to use mapping_seek_hole_data() without needing to export
it, we have to move all the delalloc extent cleanup to the iomap
core and so now the iomap core can clean up delayed allocation
extents in a safe, sane and filesystem neutral manner.
With all this done, the original data corruption never occurs
anymore, and we now have a generic mechanism for ensuring that page
cache writes do not do the wrong thing when writeback and reclaim
change the state of the physical extent and/or page cache contents
whilst the write is in progress.
Signed-off-by: Dave Chinner <dchinner@redhat.com>
Signed-off-by: Darrick J. Wong <djwong@kernel.org>
* tag 'xfs-iomap-stale-fixes' of git://git.kernel.org/pub/scm/linux/kernel/git/dgc/linux-xfs:
xfs: drop write error injection is unfixable, remove it
xfs: use iomap_valid method to detect stale cached iomaps
iomap: write iomap validity checks
xfs: xfs_bmap_punch_delalloc_range() should take a byte range
iomap: buffered write failure should not truncate the page cache
xfs,iomap: move delalloc punching to iomap
xfs: use byte ranges for write cleanup ranges
xfs: punching delalloc extents on write failure is racy
xfs: write page faults in iomap are not buffered writes
|
|
With the changes to scan the page cache for dirty data to avoid data
corruptions from partial write cleanup racing with other page cache
operations, the drop writes error injection no longer works the same
way it used to and causes xfs/196 to fail. This is because xfs/196
writes to the file and populates the page cache before it turns on
the error injection and starts failing -overwrites-.
The result is that the original drop-writes code failed writes only
-after- overwriting the data in the cache, followed by invalidates
the cached data, then punching out the delalloc extent from under
that data.
On the surface, this looks fine. The problem is that page cache
invalidation *doesn't guarantee that it removes anything from the
page cache* and it doesn't change the dirty state of the folio. When
block size == page size and we do page aligned IO (as xfs/196 does)
everything happens to align perfectly and page cache invalidation
removes the single page folios that span the written data. Hence the
followup delalloc punch pass does not find cached data over that
range and it can punch the extent out.
IOWs, xfs/196 "works" for block size == page size with the new
code. I say "works", because it actually only works for the case
where IO is page aligned, and no data was read from disk before
writes occur. Because the moment we actually read data first, the
readahead code allocates multipage folios and suddenly the
invalidate code goes back to zeroing subfolio ranges without
changing dirty state.
Hence, with multipage folios in play, block size == page size is
functionally identical to block size < page size behaviour, and
drop-writes is manifestly broken w.r.t to this case. Invalidation of
a subfolio range doesn't result in the folio being removed from the
cache, just the range gets zeroed. Hence after we've sequentially
walked over a folio that we've dirtied (via write data) and then
invalidated, we end up with a dirty folio full of zeroed data.
And because the new code skips punching ranges that have dirty
folios covering them, we end up leaving the delalloc range intact
after failing all the writes. Hence failed writes now end up
writing zeroes to disk in the cases where invalidation zeroes folios
rather than removing them from cache.
This is a fundamental change of behaviour that is needed to avoid
the data corruption vectors that exist in the old write fail path,
and it renders the drop-writes injection non-functional and
unworkable as it stands.
As it is, I think the error injection is also now unnecessary, as
partial writes that need delalloc extent are going to be a lot more
common with stale iomap detection in place. Hence this patch removes
the drop-writes error injection completely. xfs/196 can remain for
testing kernels that don't have this data corruption fix, but those
that do will report:
xfs/196 3s ... [not run] XFS error injection drop_writes unknown on this kernel.
Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Darrick J. Wong <djwong@kernel.org>
|
|
Now that iomap supports a mechanism to validate cached iomaps for
buffered write operations, hook it up to the XFS buffered write ops
so that we can avoid data corruptions that result from stale cached
iomaps. See:
https://lore.kernel.org/linux-xfs/20220817093627.GZ3600936@dread.disaster.area/
or the ->iomap_valid() introduction commit for exact details of the
corruption vector.
The validity cookie we store in the iomap is based on the type of
iomap we return. It is expected that the iomap->flags we set in
xfs_bmbt_to_iomap() is not perturbed by the iomap core and are
returned to us in the iomap passed via the .iomap_valid() callback.
This ensures that the validity cookie is always checking the correct
inode fork sequence numbers to detect potential changes that affect
the extent cached by the iomap.
Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Darrick J. Wong <djwong@kernel.org>
|
|
This is a simple mechanical transformation done by:
@@
expression E;
@@
- prandom_u32_max
+ get_random_u32_below
(E)
Reviewed-by: Kees Cook <keescook@chromium.org>
Reviewed-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Acked-by: Darrick J. Wong <djwong@kernel.org> # for xfs
Reviewed-by: SeongJae Park <sj@kernel.org> # for damon
Reviewed-by: Jason Gunthorpe <jgg@nvidia.com> # for infiniband
Reviewed-by: Russell King (Oracle) <rmk+kernel@armlinux.org.uk> # for arm
Acked-by: Ulf Hansson <ulf.hansson@linaro.org> # for mmc
Signed-off-by: Jason A. Donenfeld <Jason@zx2c4.com>
|
|
When lazysbcount is enabled, fsstress and loop mount/unmount test report
the following problems:
XFS (loop0): SB summary counter sanity check failed
XFS (loop0): Metadata corruption detected at xfs_sb_write_verify+0x13b/0x460,
xfs_sb block 0x0
XFS (loop0): Unmount and run xfs_repair
XFS (loop0): First 128 bytes of corrupted metadata buffer:
00000000: 58 46 53 42 00 00 10 00 00 00 00 00 00 28 00 00 XFSB.........(..
00000010: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................
00000020: 69 fb 7c cd 5f dc 44 af 85 74 e0 cc d4 e3 34 5a i.|._.D..t....4Z
00000030: 00 00 00 00 00 20 00 06 00 00 00 00 00 00 00 80 ..... ..........
00000040: 00 00 00 00 00 00 00 81 00 00 00 00 00 00 00 82 ................
00000050: 00 00 00 01 00 0a 00 00 00 00 00 04 00 00 00 00 ................
00000060: 00 00 0a 00 b4 b5 02 00 02 00 00 08 00 00 00 00 ................
00000070: 00 00 00 00 00 00 00 00 0c 09 09 03 14 00 00 19 ................
XFS (loop0): Corruption of in-memory data (0x8) detected at _xfs_buf_ioapply
+0xe1e/0x10e0 (fs/xfs/xfs_buf.c:1580). Shutting down filesystem.
XFS (loop0): Please unmount the filesystem and rectify the problem(s)
XFS (loop0): log mount/recovery failed: error -117
XFS (loop0): log mount failed
This corruption will shutdown the file system and the file system will
no longer be mountable. The following script can reproduce the problem,
but it may take a long time.
#!/bin/bash
device=/dev/sda
testdir=/mnt/test
round=0
function fail()
{
echo "$*"
exit 1
}
mkdir -p $testdir
while [ $round -lt 10000 ]
do
echo "******* round $round ********"
mkfs.xfs -f $device
mount $device $testdir || fail "mount failed!"
fsstress -d $testdir -l 0 -n 10000 -p 4 >/dev/null &
sleep 4
killall -w fsstress
umount $testdir
xfs_repair -e $device > /dev/null
if [ $? -eq 2 ];then
echo "ERR CODE 2: Dirty log exception during repair."
exit 1
fi
round=$(($round+1))
done
With lazysbcount is enabled, There is no additional lock protection for
reading m_ifree and m_icount in xfs_log_sb(), if other cpu modifies the
m_ifree, this will make the m_ifree greater than m_icount. For example,
consider the following sequence and ifreedelta is postive:
CPU0 CPU1
xfs_log_sb xfs_trans_unreserve_and_mod_sb
---------- ------------------------------
percpu_counter_sum(&mp->m_icount)
percpu_counter_add_batch(&mp->m_icount,
idelta, XFS_ICOUNT_BATCH)
percpu_counter_add(&mp->m_ifree, ifreedelta);
percpu_counter_sum(&mp->m_ifree)
After this, incorrect inode count (sb_ifree > sb_icount) will be writen to
the log. In the subsequent writing of sb, incorrect inode count (sb_ifree >
sb_icount) will fail to pass the boundary check in xfs_validate_sb_write()
that cause the file system shutdown.
When lazysbcount is enabled, we don't need to guarantee that Lazy sb
counters are completely correct, but we do need to guarantee that sb_ifree
<= sb_icount. On the other hand, the constraint that m_ifree <= m_icount
must be satisfied any time that there /cannot/ be other threads allocating
or freeing inode chunks. If the constraint is violated under these
circumstances, sb_i{count,free} (the ondisk superblock inode counters)
maybe incorrect and need to be marked sick at unmount, the count will
be rebuilt on the next mount.
Fixes: 8756a5af1819 ("libxfs: add more bounds checking to sb sanity checks")
Signed-off-by: Long Li <leo.lilong@huawei.com>
Reviewed-by: Darrick J. Wong <djwong@kernel.org>
Signed-off-by: Darrick J. Wong <djwong@kernel.org>
|
|
We've been (ab)using XFS_REFC_COW_START as both an integer quantity and
a bit flag, even though it's *only* a bit flag. Rename the variable to
reflect its nature and update the cast target since we're not supposed
to be comparing it to xfs_agblock_t now.
Signed-off-by: Darrick J. Wong <djwong@kernel.org>
Reviewed-by: Dave Chinner <dchinner@redhat.com>
|
|
We're supposed to initialize the list head of an object before adding it
to another list. Fix that, and stop using the kmem_{alloc,free} calls
from the Irix days.
Fixes: 174edb0e46e5 ("xfs: store in-progress CoW allocations in the refcount btree")
Signed-off-by: Darrick J. Wong <djwong@kernel.org>
Reviewed-by: Dave Chinner <dchinner@redhat.com>
|
|
As we've seen, refcount records use the upper bit of the rc_startblock
field to ensure that all the refcount records are at the right side of
the refcount btree. This works because an AG is never allowed to have
more than (1U << 31) blocks in it. If we ever encounter a filesystem
claiming to have that many blocks, we absolutely do not want reflink
touching it at all.
However, this test at the start of xfs_refcount_recover_cow_leftovers is
slightly incorrect -- it /should/ be checking that agblocks isn't larger
than the XFS_MAX_CRC_AG_BLOCKS constant, and it should check that the
constant is never large enough to conflict with that CoW flag.
Note that the V5 superblock verifier has not historically rejected
filesystems where agblocks >= XFS_MAX_CRC_AG_BLOCKS, which is why this
ended up in the COW recovery routine.
Signed-off-by: Darrick J. Wong <djwong@kernel.org>
Reviewed-by: Dave Chinner <dchinner@redhat.com>
|
|
Now that we've separated the startblock and CoW/shared extent domain in
the incore refcount record structure, check the domain whenever we
retrieve a record to ensure that it's still in the domain that we want.
Depending on the circumstances, a change in domain either means we're
done processing or that we've found a corruption and need to fail out.
The refcount check in xchk_xref_is_cow_staging is redundant since
_get_rec has done that for a long time now, so we can get rid of it.
Signed-off-by: Darrick J. Wong <djwong@kernel.org>
Reviewed-by: Dave Chinner <dchinner@redhat.com>
|
|
Now that we have an explicit enum for shared and CoW staging extents, we
can get rid of the old FIND_RCEXT flags. Omit a couple of conversions
that disappear in the next patches.
Signed-off-by: Darrick J. Wong <djwong@kernel.org>
Reviewed-by: Dave Chinner <dchinner@redhat.com>
|
|
Create a helper function to ensure that CoW staging extent records have
a single refcount and that shared extent records have more than 1
refcount. We'll put this to more use in the next patch.
Signed-off-by: Darrick J. Wong <djwong@kernel.org>
Reviewed-by: Dave Chinner <dchinner@redhat.com>
|
|
Now that we've broken out the startblock and shared/cow domain in the
incore refcount extent record structure, update the tracepoints to
report the domain.
Signed-off-by: Darrick J. Wong <djwong@kernel.org>
Reviewed-by: Dave Chinner <dchinner@redhat.com>
|
|
Just prior to committing the reflink code into upstream, the xfs
maintainer at the time requested that I find a way to shard the refcount
records into two domains -- one for records tracking shared extents, and
a second for tracking CoW staging extents. The idea here was to
minimize mount time CoW reclamation by pushing all the CoW records to
the right edge of the keyspace, and it was accomplished by setting the
upper bit in rc_startblock. We don't allow AGs to have more than 2^31
blocks, so the bit was free.
Unfortunately, this was a very late addition to the codebase, so most of
the refcount record processing code still treats rc_startblock as a u32
and pays no attention to whether or not the upper bit (the cow flag) is
set. This is a weakness is theoretically exploitable, since we're not
fully validating the incoming metadata records.
Fuzzing demonstrates practical exploits of this weakness. If the cow
flag of a node block key record is corrupted, a lookup operation can go
to the wrong record block and start returning records from the wrong
cow/shared domain. This causes the math to go all wrong (since cow
domain is still implicit in the upper bit of rc_startblock) and we can
crash the kernel by tricking xfs into jumping into a nonexistent AG and
tripping over xfs_perag_get(mp, <nonexistent AG>) returning NULL.
To fix this, start tracking the domain as an explicit part of struct
xfs_refcount_irec, adjust all refcount functions to check the domain
of a returned record, and alter the function definitions to accept them
where necessary.
Found by fuzzing keys[2].cowflag = add in xfs/464.
Signed-off-by: Darrick J. Wong <djwong@kernel.org>
Reviewed-by: Dave Chinner <dchinner@redhat.com>
|
|
Structure definitions for incore objects do not belong in the ondisk
format header. Move them to the incore types header where they belong.
Signed-off-by: Darrick J. Wong <djwong@kernel.org>
Reviewed-by: Dave Chinner <dchinner@redhat.com>
|
|
If we're in the middle of a deferred refcount operation and decide to
roll the transaction to avoid overflowing the transaction space, we need
to check the new agbno/aglen parameters that we're about to record in
the new intent. Specifically, we need to check that the new extent is
completely within the filesystem, and that continuation does not put us
into a different AG.
If the keys of a node block are wrong, the lookup to resume an
xfs_refcount_adjust_extents operation can put us into the wrong record
block. If this happens, we might not find that we run out of aglen at
an exact record boundary, which will cause the loop control to do the
wrong thing.
The previous patch should take care of that problem, but let's add this
extra sanity check to stop corruption problems sooner than later.
Signed-off-by: Darrick J. Wong <djwong@kernel.org>
Reviewed-by: Dave Chinner <dchinner@redhat.com>
|
|
Create a predicate function to verify that a given agbno/blockcount pair
fit entirely within a single allocation group and don't suffer
mathematical overflows. Refactor the existng open-coded logic; we're
going to add more calls to this function in the next patch.
Signed-off-by: Darrick J. Wong <djwong@kernel.org>
Reviewed-by: Dave Chinner <dchinner@redhat.com>
|
|
Prior to calling xfs_refcount_adjust_extents, we trimmed agbno/aglen
such that the end of the range would not be in the middle of a refcount
record. If this is no longer the case, something is seriously wrong
with the btree. Bail out with a corruption error.
Signed-off-by: Darrick J. Wong <djwong@kernel.org>
Reviewed-by: Dave Chinner <dchinner@redhat.com>
|
|
Refactor all the open-coded sizeof logic for EFI/EFD log item and log
format structures into common helper functions whose names reflect the
struct names.
Signed-off-by: Darrick J. Wong <djwong@kernel.org>
Reviewed-by: Allison Henderson <allison.henderson@oracle.com>
Reviewed-by: Dave Chinner <dchinner@redhat.com>
|
|
Starting in 6.1, CONFIG_FORTIFY_SOURCE checks the length parameter of
memcpy. Since we're already fixing problems with BUI item copying, we
should fix it everything else.
An extra difficulty here is that the ef[id]_extents arrays are declared
as single-element arrays. This is not the convention for flex arrays in
the modern kernel, and it causes all manner of problems with static
checking tools, since they often cannot tell the difference between a
single element array and a flex array.
So for starters, change those array[1] declarations to array[]
declarations to signal that they are proper flex arrays and adjust all
the "size-1" expressions to fit the new declaration style.
Next, refactor the xfs_efi_copy_format function to handle the copying of
the head and the flex array members separately. While we're at it, fix
a minor validation deficiency in the recovery function.
Signed-off-by: Darrick J. Wong <djwong@kernel.org>
Reviewed-by: Kees Cook <keescook@chromium.org>
Reviewed-by: Allison Henderson <allison.henderson@oracle.com>
Reviewed-by: Dave Chinner <dchinner@redhat.com>
|
