Allow the getattr() callback to check things like whether or not we hold
a delegation so that it can adjust the attributes that it is asking for.
Signed-off-by: Trond Myklebust <trond.myklebust@hammerspace.com>
When we hold a delegation, we should not need to request attributes such
as the file size or the change attribute. For some servers, avoiding
asking for these unneeded attributes can improve the overall system
performance.
Signed-off-by: Trond Myklebust <trond.myklebust@hammerspace.com>
If the server recalls the layout that was just handed out, we risk hitting
a race as described in RFC5661 Section 2.10.6.3 unless we ensure that we
release the sequence slot after processing the LAYOUTGET operation that
was sent as part of the OPEN compound.
Signed-off-by: Trond Myklebust <trond.myklebust@hammerspace.com>
If the layoutget on open call failed, we can't really commit the inode,
so don't bother calling it.
Signed-off-by: Trond Myklebust <trond.myklebust@hammerspace.com>
If we're only opening the file for reading, and the file is empty and/or
we already have cached data, then heuristically optimise away the
LAYOUTGET.
Signed-off-by: Trond Myklebust <trond.myklebust@hammerspace.com>
Ensure that we only switch off the LAYOUTGET operation in the OPEN
compound when the server is truly broken, and/or it is complaining
that the compound is too large.
Currently, we end up turning off the functionality permanently,
even for transient errors such as EACCES or ENOSPC.
Signed-off-by: Trond Myklebust <trond.myklebust@hammerspace.com>
We need to ensure that pnfs_parse_lgopen() doesn't try to parse a
struct nfs4_layoutget_res that was not filled by a successful call
to decode_layoutget(). This can happen if we performed a cached open,
or if either the OP_ACCESS or OP_GETATTR operations preceding the
OP_LAYOUTGET in the compound returned an error.
By initialising the 'status' field to NFS4ERR_DELAY, we ensure that
pnfs_parse_lgopen() won't try to interpret the structure.
Signed-off-by: Trond Myklebust <trond.myklebust@hammerspace.com>
The flag was not always being cleared after LAYOUTGET on OPEN.
Signed-off-by: Fred Isaman <fred.isaman@gmail.com>
Signed-off-by: Trond Myklebust <trond.myklebust@hammerspace.com>
Since the LAYOUTGET on OPEN can be sent without prior inode information,
existing methods to prevent LAYOUTGET from being sent while processing
CB_LAYOUTRECALL don't work. Track if a recall occurred while LAYOUTGET
was being sent, and if so ignore the results.
Signed-off-by: Fred Isaman <fred.isaman@gmail.com>
Signed-off-by: Trond Myklebust <trond.myklebust@hammerspace.com>
Move the actual freeing of the struct nfs4_layoutget into fs/nfs/pnfs.c
where it can be reused by the layoutget on open code.
Signed-off-by: Trond Myklebust <trond.myklebust@hammerspace.com>
This triggers when have no pre-existing inode to attach to.
The preexisting case is saved for later.
Signed-off-by: Fred Isaman <fred.isaman@gmail.com>
Signed-off-by: Trond Myklebust <trond.myklebust@hammerspace.com>
Don't send in a layout, instead use the (possibly NULL) inode.
This is needed for LAYOUTGET attached to an OPEN where the inode is not
yet set.
Signed-off-by: Fred Isaman <fred.isaman@gmail.com>
Signed-off-by: Trond Myklebust <trond.myklebust@hammerspace.com>
It will be needed now by the pnfs code.
Signed-off-by: Fred Isaman <fred.isaman@gmail.com>
Signed-off-by: Trond Myklebust <trond.myklebust@hammerspace.com>
They work better in the new alloc_init function.
Signed-off-by: Fred Isaman <fred.isaman@gmail.com>
Signed-off-by: Trond Myklebust <trond.myklebust@hammerspace.com>
Pull out the alloc/init part for eventual reuse by OPEN.
Signed-off-by: Fred Isaman <fred.isaman@gmail.com>
Signed-off-by: Trond Myklebust <trond.myklebust@hammerspace.com>
Driver can set flag to allow LAYOUTGET to be sent with OPEN.
Signed-off-by: Fred Isaman <fred.isaman@gmail.com>
Signed-off-by: Trond Myklebust <trond.myklebust@hammerspace.com>
Preparing to add conditional LAYOUTGET to OPEN rpc, the LAYOUTGET
will need the ctx info.
Signed-off-by: Fred Isaman <fred.isaman@gmail.com>
Signed-off-by: Trond Myklebust <trond.myklebust@hammerspace.com>
This will be needed to seperate return value of OPEN and LAYOUTGET
when they are combined into a single RPC.
Signed-off-by: Fred Isaman <fred.isaman@gmail.com>
Signed-off-by: Trond Myklebust <trond.myklebust@hammerspace.com>
nfs_init_sequence() will clear this for us.
Signed-off-by: Fred Isaman <fred.isaman@gmail.com>
Signed-off-by: Trond Myklebust <trond.myklebust@hammerspace.com>
If the wait for a LOCK operation is interrupted, and then the file is
closed, the locks cleanup code will assume that no new locks will be added
to the inode after it has completed. We already have a mechanism to detect
if there was signal, so let's use that to avoid recreating the local lock
once the RPC completes. Also skip re-sending the LOCK operation for the
various error cases if we were signaled.
Signed-off-by: Benjamin Coddington <bcodding@redhat.com>
[Trond: Fix inverted test of locks_lock_inode_wait()]
Signed-off-by: Trond Myklebust <trond.myklebust@hammerspace.com>
If we get an ESTALE error in response to an RPC call operating on the
file on the MDS, we should immediately cancel the layout for that file.
Signed-off-by: Trond Myklebust <trond.myklebust@hammerspace.com>
In nfs_idmap_read_and_verify_message there is an incorrect sprintf '%d'
that converts the __u32 'im_id' from struct idmap_msg to 'id_str', which
is a stack char array variable of length NFS_UINT_MAXLEN == 11.
If a uid or gid value is > 2147483647 = 0x7fffffff, the conversion
overflows into a negative value, for example:
crash> p (unsigned) (0x80000000)
$1 = 2147483648
crash> p (signed) (0x80000000)
$2 = -2147483648
The '-' sign is written to the buffer and this causes a 1 byte overflow
when the NULL byte is written, which corrupts kernel stack memory. If
CONFIG_CC_STACKPROTECTOR_STRONG is set we see a stack-protector panic:
[11558053.616565] Kernel panic - not syncing: stack-protector: Kernel stack is corrupted in: ffffffffa05b8a8c
[11558053.639063] CPU: 6 PID: 9423 Comm: rpc.idmapd Tainted: G W ------------ T 3.10.0-514.el7.x86_64 #1
[11558053.641990] Hardware name: Red Hat OpenStack Compute, BIOS 1.10.2-3.el7_4.1 04/01/2014
[11558053.644462] ffffffff818c7bc0 00000000b1f3aec1 ffff880de0f9bd48 ffffffff81685eac
[11558053.646430] ffff880de0f9bdc8 ffffffff8167f2b3 ffffffff00000010 ffff880de0f9bdd8
[11558053.648313] ffff880de0f9bd78 00000000b1f3aec1 ffffffff811dcb03 ffffffffa05b8a8c
[11558053.650107] Call Trace:
[11558053.651347] [<ffffffff81685eac>] dump_stack+0x19/0x1b
[11558053.653013] [<ffffffff8167f2b3>] panic+0xe3/0x1f2
[11558053.666240] [<ffffffff811dcb03>] ? kfree+0x103/0x140
[11558053.682589] [<ffffffffa05b8a8c>] ? idmap_pipe_downcall+0x1cc/0x1e0 [nfsv4]
[11558053.689710] [<ffffffff810855db>] __stack_chk_fail+0x1b/0x30
[11558053.691619] [<ffffffffa05b8a8c>] idmap_pipe_downcall+0x1cc/0x1e0 [nfsv4]
[11558053.693867] [<ffffffffa00209d6>] rpc_pipe_write+0x56/0x70 [sunrpc]
[11558053.695763] [<ffffffff811fe12d>] vfs_write+0xbd/0x1e0
[11558053.702236] [<ffffffff810acccc>] ? task_work_run+0xac/0xe0
[11558053.704215] [<ffffffff811fec4f>] SyS_write+0x7f/0xe0
[11558053.709674] [<ffffffff816964c9>] system_call_fastpath+0x16/0x1b
Fix this by calling the internally defined nfs_map_numeric_to_string()
function which properly uses '%u' to convert this __u32. For consistency,
also replace the one other place where snprintf is called.
Signed-off-by: Dave Wysochanski <dwysocha@redhat.com>
Reported-by: Stephen Johnston <sjohnsto@redhat.com>
Fixes: cf4ab538f1 ("NFSv4: Fix the string length returned by the idmapper")
Cc: stable@vger.kernel.org # v3.4+
Signed-off-by: Trond Myklebust <trond.myklebust@hammerspace.com>
Ensure that a delegation doesn't cause us to skip initialising the inode
if it was incomplete when we exited nfs_fhget()
Signed-off-by: Trond Myklebust <trond.myklebust@hammerspace.com>
Ensure that we pass down the inode of the file being deleted so
that we can return any delegation being held.
Signed-off-by: Trond Myklebust <trond.myklebust@hammerspace.com>
Even then it isn't really necessary. The reason why we may not want to
pass in a stateid in other cases is that we cannot use the delegation
credential.
Signed-off-by: Trond Myklebust <trond.myklebust@hammerspace.com>
Having these exist as two functions doesn't seem to add anything useful,
and I think merging them together makes this easier to follow.
Signed-off-by: Anna Schumaker <Anna.Schumaker@Netapp.com>
Signed-off-by: Trond Myklebust <trond.myklebust@hammerspace.com>
We currently have a separate function just to set this, but I think it
makes more sense to set it at the same time as the other values in
nfs4_init_sequence()
Signed-off-by: Anna Schumaker <Anna.Schumaker@Netapp.com>
Signed-off-by: Trond Myklebust <trond.myklebust@hammerspace.com>
Rather than doing this in the generic NFS client code. Let's put this
with the other v4 stuff so it's all in one place.
Signed-off-by: Anna Schumaker <Anna.Schumaker@Netapp.com>
Signed-off-by: Trond Myklebust <trond.myklebust@hammerspace.com>
This doesn't really need to be in the generic NFS client code, and I
think it makes more sense to keep the v4 code in one place.
Signed-off-by: Anna Schumaker <Anna.Schumaker@Netapp.com>
Signed-off-by: Trond Myklebust <trond.myklebust@hammerspace.com>
There are three places that walk all delegation for an nfs_client and
restart whenever they find something interesting - potentially
resulting in a quadratic search: If there are 10,000 uninteresting
delegations followed by 10,000 interesting one, then the code
skips over 100,000,000 delegations, which can take a noticeable amount
of time.
Of these nfs_delegation_reap_unclaimed() and
nfs_reap_expired_delegations() are only called during unusual events:
a server reboots or reports expired delegations, probably due to a
network partition. Optimizing these is not particularly important.
The third, nfs_client_return_marked_delegations(), is called
periodically via nfs_expire_unreferenced_delegations(). It could
cause periodic problems on a busy server.
New delegations are added to the end of the list, so if there are
10,000 open files with delegations, and 10,000 more recently opened files
that received delegations but are now closed, then
nfs_client_return_marked_delegations() can take seconds to skip over
the 10,000 open files 10,000 times. That is a waste of time.
The avoid this waste a place-holder (an inode) is kept when locks are
dropped, so that the place can usually be found again after taking
rcu_readlock(). This place holder ensure that we find the right
starting point in the list of nfs_servers, and makes is probable that
we find the right starting point in the list of delegations.
We might need to occasionally restart at the head of that list.
It might be possible that the place_holder inode could lose its
delegation separately, and then get a new one using the same (freed
and then reallocated) 'struct nfs_delegation'. Were this to happen,
the new delegation would be at the end of the list and we would miss
returning some other delegations. This would have the effect of
unnecessarily delaying the return of some unused delegations until the
next time this function is called - typically 90 seconds later. As
this is not a correctness issue and is vanishingly unlikely to happen,
it does not seem worth addressing.
Signed-off-by: NeilBrown <neilb@suse.com>
Signed-off-by: Trond Myklebust <trond.myklebust@hammerspace.com>
In three places we walk the list of delegations for an nfs_client
until an interesting one is found, then we act of that delegation
and restart the walk.
New delegations are added to the end of a list and the interesting
delegations are usually old, so in many case we won't repeat
a long walk over and over again, but it is possible - particularly if
the first server in the list has a large number of uninteresting
delegations.
In each cache the work done on interesting delegations will often
complete without sleeping, so this could loop many times without
giving up the CPU.
So add a cond_resched() at an appropriate point to avoid hogging the
CPU for too long.
Signed-off-by: NeilBrown <neilb@suse.com>
Signed-off-by: Trond Myklebust <trond.myklebust@hammerspace.com>
There are 3 places where we walk the list of delegations
for an nfs_client.
In each case there are two nested loops, one for nfs_servers
and one for nfs_delegations.
When we find an interesting delegation we try to get an active
reference to the server. If that fails, it is pointless to
continue to look at the other delegation for the server as
we will never be able to get an active reference.
So instead of continuing in the inner loop, break out
and continue in the outer loop.
Signed-off-by: NeilBrown <neilb@suse.com>
Signed-off-by: Trond Myklebust <trond.myklebust@hammerspace.com>
We can optimise away any lookup for a rename target, unless we're
being asked to revalidate a dentry that might be in use.
Signed-off-by: Trond Myklebust <trond.myklebust@hammerspace.com>
If nfs_lookup_revalidate() is called with LOOKUP_REVAL because a
previous path lookup failed, then we ought to force a full lookup
of the component name.
Signed-off-by: Trond Myklebust <trond.myklebust@hammerspace.com>
NFSv4 should not need to perform an extra close-to-open GETATTR as part
of the process of looking up a regular file, since the OPEN call will
do that for us.
Signed-off-by: Trond Myklebust <trond.myklebust@hammerspace.com>
The 4.17-rc /proc/meminfo and /proc/<pid>/smaps look ugly: single-digit
numbers (commonly 0) are misaligned.
Remove seq_put_decimal_ull_width()'s leftover optimization for single
digits: it's wrong now that num_to_str() takes care of the width.
Link: http://lkml.kernel.org/r/alpine.LSU.2.11.1805241554210.1326@eggly.anvils
Fixes: d1be35cb6f ("proc: add seq_put_decimal_ull_width to speed up /proc/pid/smaps")
Signed-off-by: Hugh Dickins <hughd@google.com>
Cc: Andrei Vagin <avagin@openvz.org>
Cc: Alexey Dobriyan <adobriyan@gmail.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
- Remove bouncing addresses from the MAINTAINERS file
- Kernel oops and bad error handling fixes for hfi, i40iw, cxgb4, and hns drivers
- Various small LOC behavioral/operational bugs in mlx5, hns, qedr and i40iw drivers
- Two fixes for patches already sent during the merge window
- A long standing bug related to not decreasing the pinned pages count in the right
MM was found and fixed
-----BEGIN PGP SIGNATURE-----
Version: GnuPG v2
iQIcBAABCgAGBQJbByPQAAoJEDht9xV+IJsa164P/AihB/vbn9MBdK3pe1OSUGTm
tKZJ/Y6nY/Q/XTJSeM2wNECk8fOrZbKuLBz2XlPRsB2djp4ugC5WWfK9YbwWMGXG
I5B/lB8VTorQr8E5i9lqqMDQc8aF8VcGJtdqVE3nD4JsVTrQSGiSnw45/BARDUm3
OycJJMDOWhDj2wnNSa+JfjPemIMDM1jse7DnsJfDsGfTMS/G+6nyzjKIlEnnFZ8/
PBxhq0q7C5viNDwwn2GsAVUrATTlW48SY0WYhkgMdSl20d2th9wMZqNMqtniz8NP
lg87SrhzsAPOTlbSWlYYkAnzE7nEhfJyIfYUp2piNJeYuOohYPtO6w99Tqjl/GmU
uLIYIXtZCxAK1Zb/znc49HkRVL5YFDsQGXdtYy7tvRZPwwR32kowUtpKIWaZFz8O
BA/x+Zgqu9AlwqSWwQwxmMbUX42RRwhNJDVyTYlXQSSzhfgFaLIZARqb4K6HxeNN
vZN0BK+x6pX6FI7hpdsqNRtH1oo4SNUBxiuUsrZ7cy7GqYNdUJ6piygDgmERaJxU
svIUJof/+OoU1QyErQ0JgUEK/3jOHbjxSPb/rjQeqxAnCqhaGOuNGMtdfsGqgvBU
x/u3eDcbfi/LBErXR46gYtxnOQ8I2BB+m8erUc/GVvCzWrX+R7ELZYpBrP5Pcu/6
mr2D7hDqgZHbeU8aB8+D
=uFZh
-----END PGP SIGNATURE-----
Merge tag 'for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/rdma/rdma
Pull rdma fixes from Jason Gunthorpe:
"This is pretty much just the usual array of smallish driver bugs.
- remove bouncing addresses from the MAINTAINERS file
- kernel oops and bad error handling fixes for hfi, i40iw, cxgb4, and
hns drivers
- various small LOC behavioral/operational bugs in mlx5, hns, qedr
and i40iw drivers
- two fixes for patches already sent during the merge window
- a long-standing bug related to not decreasing the pinned pages
count in the right MM was found and fixed"
* tag 'for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/rdma/rdma: (28 commits)
RDMA/hns: Move the location for initializing tmp_len
RDMA/hns: Bugfix for cq record db for kernel
IB/uverbs: Fix uverbs_attr_get_obj
RDMA/qedr: Fix doorbell bar mapping for dpi > 1
IB/umem: Use the correct mm during ib_umem_release
iw_cxgb4: Fix an error handling path in 'c4iw_get_dma_mr()'
RDMA/i40iw: Avoid panic when reading back the IRQ affinity hint
RDMA/i40iw: Avoid reference leaks when processing the AEQ
RDMA/i40iw: Avoid panic when objects are being created and destroyed
RDMA/hns: Fix the bug with NULL pointer
RDMA/hns: Set NULL for __internal_mr
RDMA/hns: Enable inner_pa_vld filed of mpt
RDMA/hns: Set desc_dma_addr for zero when free cmq desc
RDMA/hns: Fix the bug with rq sge
RDMA/hns: Not support qp transition from reset to reset for hip06
RDMA/hns: Add return operation when configured global param fail
RDMA/hns: Update convert function of endian format
RDMA/hns: Load the RoCE dirver automatically
RDMA/hns: Bugfix for rq record db for kernel
RDMA/hns: Add rq inline flags judgement
...
-----BEGIN PGP SIGNATURE-----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=iYyx
-----END PGP SIGNATURE-----
Merge tag 'for-4.17-rc6-tag' of git://git.kernel.org/pub/scm/linux/kernel/git/kdave/linux
Pull btrfs fix from David Sterba:
"A one-liner that prevents leaking an internal error value 1 out of the
ftruncate syscall.
This has been observed in practice. The steps to reproduce make a
common pattern (open/write/fync/ftruncate) but also need the
application to not check only for negative values and happens only for
compressed inlined files.
The conditions are narrow but as this could break userspace I think
it's better to merge it now and not wait for the merge window"
* tag 'for-4.17-rc6-tag' of git://git.kernel.org/pub/scm/linux/kernel/git/kdave/linux:
Btrfs: fix error handling in btrfs_truncate()
Jun Wu at Facebook reported that an internal service was seeing a return
value of 1 from ftruncate() on Btrfs in some cases. This is coming from
the NEED_TRUNCATE_BLOCK return value from btrfs_truncate_inode_items().
btrfs_truncate() uses two variables for error handling, ret and err.
When btrfs_truncate_inode_items() returns non-zero, we set err to the
return value. However, NEED_TRUNCATE_BLOCK is not an error. Make sure we
only set err if ret is an error (i.e., negative).
To reproduce the issue: mount a filesystem with -o compress-force=zstd
and the following program will encounter return value of 1 from
ftruncate:
int main(void) {
char buf[256] = { 0 };
int ret;
int fd;
fd = open("test", O_CREAT | O_WRONLY | O_TRUNC, 0666);
if (fd == -1) {
perror("open");
return EXIT_FAILURE;
}
if (write(fd, buf, sizeof(buf)) != sizeof(buf)) {
perror("write");
close(fd);
return EXIT_FAILURE;
}
if (fsync(fd) == -1) {
perror("fsync");
close(fd);
return EXIT_FAILURE;
}
ret = ftruncate(fd, 128);
if (ret) {
printf("ftruncate() returned %d\n", ret);
close(fd);
return EXIT_FAILURE;
}
close(fd);
return EXIT_SUCCESS;
}
Fixes: ddfae63cc8 ("btrfs: move btrfs_truncate_block out of trans handle")
CC: stable@vger.kernel.org # 4.15+
Reported-by: Jun Wu <quark@fb.com>
Signed-off-by: Omar Sandoval <osandov@fb.com>
Signed-off-by: David Sterba <dsterba@suse.com>
Pull vfs fixes from Al Viro:
"Assorted fixes all over the place"
* 'fixes' of git://git.kernel.org/pub/scm/linux/kernel/git/viro/vfs:
aio: fix io_destroy(2) vs. lookup_ioctx() race
ext2: fix a block leak
nfsd: vfs_mkdir() might succeed leaving dentry negative unhashed
cachefiles: vfs_mkdir() might succeed leaving dentry negative unhashed
unfuck sysfs_mount()
kernfs: deal with kernfs_fill_super() failures
cramfs: Fix IS_ENABLED typo
befs_lookup(): use d_splice_alias()
affs_lookup: switch to d_splice_alias()
affs_lookup(): close a race with affs_remove_link()
fix breakage caused by d_find_alias() semantics change
fs: don't scan the inode cache before SB_BORN is set
do d_instantiate/unlock_new_inode combinations safely
iov_iter: fix memory leak in pipe_get_pages_alloc()
iov_iter: fix return type of __pipe_get_pages()
kill_ioctx() used to have an explicit RCU delay between removing the
reference from ->ioctx_table and percpu_ref_kill() dropping the refcount.
At some point that delay had been removed, on the theory that
percpu_ref_kill() itself contained an RCU delay. Unfortunately, that was
the wrong kind of RCU delay and it didn't care about rcu_read_lock() used
by lookup_ioctx(). As the result, we could get ctx freed right under
lookup_ioctx(). Tejun has fixed that in a6d7cff472 ("fs/aio: Add explicit
RCU grace period when freeing kioctx"); however, that fix is not enough.
Suppose io_destroy() from one thread races with e.g. io_setup() from another;
CPU1 removes the reference from current->mm->ioctx_table[...] just as CPU2
has picked it (under rcu_read_lock()). Then CPU1 proceeds to drop the
refcount, getting it to 0 and triggering a call of free_ioctx_users(),
which proceeds to drop the secondary refcount and once that reaches zero
calls free_ioctx_reqs(). That does
INIT_RCU_WORK(&ctx->free_rwork, free_ioctx);
queue_rcu_work(system_wq, &ctx->free_rwork);
and schedules freeing the whole thing after RCU delay.
In the meanwhile CPU2 has gotten around to percpu_ref_get(), bumping the
refcount from 0 to 1 and returned the reference to io_setup().
Tejun's fix (that queue_rcu_work() in there) guarantees that ctx won't get
freed until after percpu_ref_get(). Sure, we'd increment the counter before
ctx can be freed. Now we are out of rcu_read_lock() and there's nothing to
stop freeing of the whole thing. Unfortunately, CPU2 assumes that since it
has grabbed the reference, ctx is *NOT* going away until it gets around to
dropping that reference.
The fix is obvious - use percpu_ref_tryget_live() and treat failure as miss.
It's not costlier than what we currently do in normal case, it's safe to
call since freeing *is* delayed and it closes the race window - either
lookup_ioctx() comes before percpu_ref_kill() (in which case ctx->users
won't reach 0 until the caller of lookup_ioctx() drops it) or lookup_ioctx()
fails, ctx->users is unaffected and caller of lookup_ioctx() doesn't see
the object in question at all.
Cc: stable@kernel.org
Fixes: a6d7cff472 "fs/aio: Add explicit RCU grace period when freeing kioctx"
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
open file, unlink it, then use ioctl(2) to make it immutable or
append only. Now close it and watch the blocks *not* freed...
Immutable/append-only checks belong in ->setattr().
Note: the bug is old and backport to anything prior to 737f2e93b9
("ext2: convert to use the new truncate convention") will need
these checks lifted into ext2_setattr().
Cc: stable@kernel.org
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
Trond Myklebust
Trond Myklebust
Fred Isaman
Benjamin Coddington
Dave Wysochanski
Anna Schumaker
NeilBrown
Hugh Dickins
Changwei Ge
Linus Torvalds
Omar Sandoval
Al Viro