Close-to-open cache consistency rules really only require us to flush out
writes on calls to close(), and require us to revalidate attributes on the
very last close of the file.
Currently we appear to be doing a lot of extra attribute revalidation
and cache flushes.
Signed-off-by: Trond Myklebust <Trond.Myklebust@netapp.com>
The following patch is a combination of a patch by myself and Peter
Staubach.
Trond: If we allow other processes to dirty pages while a process is doing
a consistency sync to disk, we can end up never making progress.
Peter: Attached is a patch which addresses a continuing problem with
the NFS client generating out of order WRITE requests. While
this is compliant with all of the current protocol
specifications, there are servers in the market which can not
handle out of order WRITE requests very well. Also, this may
lead to sub-optimal block allocations in the underlying file
system on the server. This may cause the read throughputs to
be reduced when reading the file from the server.
Peter: There has been a lot of work recently done to address out of
order issues on a systemic level. However, the NFS client is
still susceptible to the problem. Out of order WRITE
requests can occur when pdflush is in the middle of writing
out pages while the process dirtying the pages calls
generic_file_buffered_write which calls
generic_perform_write which calls
balance_dirty_pages_rate_limited which ends up calling
writeback_inodes which ends up calling back into the NFS
client to writes out dirty pages for the same file that
pdflush happens to be working with.
Signed-off-by: Peter Staubach <staubach@redhat.com>
[modification by Trond to merge the two similar patches]
Signed-off-by: Trond Myklebust <Trond.Myklebust@netapp.com>
Currently, filling struct nfs_fattr is more or less an all or nothing
operation, since NFSv2 and NFSv3 have only mandatory attributes.
In NFSv4, some attributes are optional, and so we may simply not be able to
fill in those fields. Furthermore, NFSv4 allows you to specify which
attributes you are interested in retrieving, thus permitting you to
optimise away retrieval of attributes that you know will no change...
Signed-off-by: Trond Myklebust <Trond.Myklebust@netapp.com>
If cached directory contents becomes incorrect, there is no way to
flush the contents. This contrasts with files where file locking is
the recommended way to ensure cache consistency between multiple
applications (a read-lock always flushes the cache).
Also while changes to files often change the size of the file (thus
triggering a cache flush), changes to directories often do not change
the apparent size (as the size is often rounded to a block size).
So it is particularly important with directories to avoid the
possibility of an incorrect cache wherever possible.
When the link count on a directory changes it implies a change in the
number of child directories, and so a change in the contents of this
directory. So use that as a trigger to flush cached contents.
When the ctime changes but the mtime does not, there are two possible
reasons.
1/ The owner/mode information has been changed.
2/ utimes has been used to set the mtime backwards.
In the first case, a data-cache flush is not required.
In the second case it is.
So on the basis that correctness trumps performance, flush the
directory contents cache in this case also.
Signed-off-by: NeilBrown <neilb@suse.de>
Signed-off-by: Trond Myklebust <Trond.Myklebust@netapp.com>
Remove redundant NFS_STALE() check, a leftover due to the commit
691beb13cd
Signed-off-by: Suresh Jayaraman <sjayaraman@suse.de>
Signed-off-by: Trond Myklebust <Trond.Myklebust@netapp.com>
Hi.
I've been looking at a bugzilla which describes a problem where
a customer was advised to use either the "noac" or "actimeo=0"
mount options to solve a consistency problem that they were
seeing in the file attributes. It turned out that this solution
did not work reliably for them because sometimes, the local
attribute cache was believed to be valid and not timed out.
(With an attribute cache timeout of 0, the cache should always
appear to be timed out.)
In looking at this situation, it appears to me that the problem
is that the attribute cache timeout code has an off-by-one
error in it. It is assuming that the cache is valid in the
region, [read_cache_jiffies, read_cache_jiffies + attrtimeo]. The
cache should be considered valid only in the region,
[read_cache_jiffies, read_cache_jiffies + attrtimeo). With this
change, the options, "noac" and "actimeo=0", work as originally
expected.
This problem was previously addressed by special casing the
attrtimeo == 0 case. However, since the problem is only an off-
by-one error, the cleaner solution is address the off-by-one
error and thus, not require the special case.
Thanx...
ps
Signed-off-by: Peter Staubach <staubach@redhat.com>
Signed-off-by: Trond Myklebust <Trond.Myklebust@netapp.com>
The most important property we need from nfs_attr_generation_counter is
monotonicity, which is not guaranteed by the current system of smp memory
barriers. We should convert it to an atomic_long_t, and drop the memory
barriers.
Signed-off-by: Trond Myklebust <Trond.Myklebust@netapp.com>
The cache_change_attribute is used to decide whether or not a directory has
changed, in which case we may need to look it up again. Again, the use of
'jiffies' leads to an issue of resolution.
Once again, the fix is to change nfs_inode->cache_change_attribute, and
just make it a simple counter.
Signed-off-by: Trond Myklebust <Trond.Myklebust@netapp.com>
It appears that 'jiffies' timestamps do not have high enough resolution for
nfs_inode_attrs_need_update(). One problem is that a GETATTR can be
launched within < 1 jiffy of the last operation that updated the attribute.
Another problem is that RPC calls can take < 1 jiffy to execute.
We can fix this by switching the variables to use a simple global counter
that gets incremented every time we start another GETATTR call.
Signed-off-by: Trond Myklebust <Trond.Myklebust@netapp.com>
This fixes a regression seen when running the Connectathon testsuite
against an ext3 filesystem. The reason was that the inode was constantly
being marked as 'just updated' by the jiffy wraparound test.
This again meant that newer GETATTR calls were failing to pass the
nfs_inode_attrs_need_update() test unless the changes caused a ctime update
on the server, since they were perceived as having been started before the
latest inode update.
Given that nfs_inode_attrs_need_update() already checks for wraparound
of nfsi->last_updated, we can drop the buggy "protection" in
nfs_update_inode().
Also make a slight micro-optimisation of nfs_inode_attrs_need_update(): we
are more often going to see time_after(fattr->time_start, nfsi->last_updated)
be true, rather than seeing an update of ctime/size, so put that test
first to ensure that we optimise away the ctime/size tests.
Signed-off-by: Trond Myklebust <Trond.Myklebust@netapp.com>
Currently, if two processes are both trying to revalidate metadata for the
same inode, they will find themselves being serialised. There is no good
justification for this now that we have improved our ability to detect
stale attribute data, so we should remove that serialisation.
Signed-off-by: Trond Myklebust <Trond.Myklebust@netapp.com>
If we're merely checking the inode attributes because we suspect that the
'updated' attributes returned by the RPC call are stale, then we shouldn't
be doing weak cache consistency updates or clearing the cache_validity
flags.
Signed-off-by: Trond Myklebust <Trond.Myklebust@netapp.com>
In the case where there are parallel RPC calls to the same inode, we may
receive stale metadata due to the lack of ordering, hence the sanity
checking of metadata in nfs_refresh_inode().
Currently, __nfs_revalidate_inode() is calling nfs_update_inode() directly,
without any further sanity checks, and hence may end up setting the inode
up with stale metadata.
Fix is to use nfs_refresh_inode() instead of nfs_update_inode().
Signed-off-by: Trond Myklebust <Trond.Myklebust@netapp.com>
If we believe that the attributes are old (see nfs_refresh_inode()), then
we shouldn't force an update.
Also ensure that we hold the inode->i_lock across attribute checks and the
call to nfs_refresh_inode_locked() to ensure that we don't race with other
attribute updates.
Signed-off-by: Trond Myklebust <Trond.Myklebust@netapp.com>
Currently nfs_refresh_inode() will only update the inode metadata if it
sees that the RPC call that returned the nfs_fattr was started
after the last update of the inode. This means that if we have parallel
RPC calls to the same inode (when sending WRITE calls, for instance), we
may often miss updates.
This patch attempts to recover those missed updates by also accepting
them if the ctime in the nfs_fattr is more recent than the inode's
cached ctime.
It also recovers the case where the file size has increased, but the
ctime has not been updated due to limited ctime resolution.
Signed-off-by: Trond Myklebust <Trond.Myklebust@netapp.com>
Try to avoid taking and dropping the inode->i_lock more than once. Do so by
moving the code in nfs_refresh_inode() that needs to be done under the
spinlock into a function nfs_refresh_inode_locked(), and then having both
nfs_refresh_inode() and nfs_post_op_update_inode() call it directly.
Signed-off-by: Trond Myklebust <Trond.Myklebust@netapp.com>
Kmem cache passed to constructor is only needed for constructors that are
themselves multiplexeres. Nobody uses this "feature", nor does anybody uses
passed kmem cache in non-trivial way, so pass only pointer to object.
Non-trivial places are:
arch/powerpc/mm/init_64.c
arch/powerpc/mm/hugetlbpage.c
This is flag day, yes.
Signed-off-by: Alexey Dobriyan <adobriyan@gmail.com>
Acked-by: Pekka Enberg <penberg@cs.helsinki.fi>
Acked-by: Christoph Lameter <cl@linux-foundation.org>
Cc: Jon Tollefson <kniht@linux.vnet.ibm.com>
Cc: Nick Piggin <nickpiggin@yahoo.com.au>
Cc: Matt Mackall <mpm@selenic.com>
[akpm@linux-foundation.org: fix arch/powerpc/mm/hugetlbpage.c]
[akpm@linux-foundation.org: fix mm/slab.c]
[akpm@linux-foundation.org: fix ubifs]
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
The main problem is dealing with inode->i_size: we need to set the
inode->i_lock on all attribute updates, and so vmtruncate won't cut it.
Make an NFS-private version of vmtruncate that has the necessary locking
semantics.
The result should be that the following inode attribute updates are
protected by inode->i_lock
nfsi->cache_validity
nfsi->read_cache_jiffies
nfsi->attrtimeo
nfsi->attrtimeo_timestamp
nfsi->change_attr
nfsi->last_updated
nfsi->cache_change_attribute
nfsi->access_cache
nfsi->access_cache_entry_lru
nfsi->access_cache_inode_lru
nfsi->acl_access
nfsi->acl_default
nfsi->nfs_page_tree
nfsi->ncommit
nfsi->npages
nfsi->open_files
nfsi->silly_list
nfsi->acl
nfsi->open_states
inode->i_size
inode->i_atime
inode->i_mtime
inode->i_ctime
inode->i_nlink
inode->i_uid
inode->i_gid
The following is protected by dir->i_mutex
nfsi->cookieverf
Signed-off-by: Trond Myklebust <Trond.Myklebust@netapp.com>
ftruncate() access checking is supposed to be performed at open() time,
just like reads and writes.
Signed-off-by: Trond Myklebust <Trond.Myklebust@netapp.com>
Here are some more places where path_{get,put}() can be used instead of
dput()/mntput() pair.
Signed-off-by: Jan Blunck <jblunck@suse.de>
Cc: Trond Myklebust <trond.myklebust@fys.uio.no>
Cc: "J. Bruce Fields" <bfields@fieldses.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Trond Myklebust <Trond.Myklebust@netapp.com>
Both the 'noac' and 'actimeo=0' mount options should ensure that attributes
are not cached, however a bug in nfs_attribute_timeout() means that
currently, the attributes may in fact get cached for up to one jiffy. This
has been seen to cause corruption in some applications.
The reason for the bug is that the time_in_range() test returns 'true' as
long as the current time lies between nfsi->read_cache_jiffies and
nfsi->read_cache_jiffies + nfsi->attrtimeo. In other words, if jiffies
equals nfsi->read_cache_jiffies, then we still cache the attribute data.
Signed-off-by: Trond Myklebust <Trond.Myklebust@netapp.com>
The nfs_open_context struct had a "flags" field added recently, but the
allocator isn't initializing it. It also looks like the allocator isn't
initializing the mode or list either, but they seem to be overwritten
by the caller, so that's less of an issue.
Signed-off-by: Jeff Layton <jlayton@redhat.com>
Signed-off-by: Trond Myklebust <Trond.Myklebust@netapp.com>
The NFSv4 protocol allows clients to negotiate security protocols on the
fly in the case where an administrator on the server changes the export
settings and/or in the case where we may have a filesystem migration event.
Instead of having the NFS client code cache credentials that are tied to a
particular AUTH method it is therefore preferable to have a generic credential
that can be converted into whatever AUTH is in use by the RPC client when
the read/write/sillyrename/... is put on the wire.
We do this by means of the new "generic" credential, which basically just
caches the minimal information that is needed to look up an RPCSEC_GSS,
AUTH_SYS, or AUTH_NULL credential.
Signed-off-by: Trond Myklebust <Trond.Myklebust@netapp.com>
When we detect that we've crossed a mountpoint on the remote server, we
must take care not to use that inode to revalidate the fsid on our
current superblock. To do so, we label the inode as a remote mountpoint,
and check for that in nfs_update_inode().
Signed-off-by: Trond Myklebust <Trond.Myklebust@netapp.com>
NFS post-rpciod cleanups often involve tasks that cannot be safely
performed within the rpciod context (due to deadlock concerns). We
therefore add a dedicated NFS workqueue that can perform tasks like
cleaning up state after an interrupted NFSv4 open() call, or calling
put_nfs_open_context() after an asynchronous read or write call.
Signed-off-by: Trond Myklebust <Trond.Myklebust@netapp.com>
We can't allow rpc callback functions like task->tk_ops->rpc_call_prepare()
and task->tk_ops->rpc_call_done() to call mntput() in any way, since
that will cause a deadlock when the call to rpc_shutdown_client() attempts
to wait on 'task' to complete.
We can avoid the above deadlock by moving calls to mntput to
task->tk_ops->rpc_release() callback, since at that time the task will be
marked as completed, and so rpc_shutdown_client won't attempt to wait on
it.
Signed-off-by: Trond Myklebust <Trond.Myklebust@netapp.com>
* 'task_killable' of git://git.kernel.org/pub/scm/linux/kernel/git/willy/misc: (22 commits)
Remove commented-out code copied from NFS
NFS: Switch from intr mount option to TASK_KILLABLE
Add wait_for_completion_killable
Add wait_event_killable
Add schedule_timeout_killable
Use mutex_lock_killable in vfs_readdir
Add mutex_lock_killable
Use lock_page_killable
Add lock_page_killable
Add fatal_signal_pending
Add TASK_WAKEKILL
exit: Use task_is_*
signal: Use task_is_*
sched: Use task_contributes_to_load, TASK_ALL and TASK_NORMAL
ptrace: Use task_is_*
power: Use task_is_*
wait: Use TASK_NORMAL
proc/base.c: Use task_is_*
proc/array.c: Use TASK_REPORT
perfmon: Use task_is_*
...
Fixed up conflicts in NFS/sunrpc manually..
Otherwise, there is a potential deadlock if the last dput() from an NFSv4
close() or other asynchronous operation leads to nfs_clear_inode calling
the synchronous delegreturn.
Signed-off-by: Trond Myklebust <Trond.Myklebust@netapp.com>
POSIX requires that ctime and mtime, as reported by the stat(2) call,
reflect the activity of the most recent write(2). To that end, nfs_getattr()
flushes pending dirty writes to a file before doing a GETATTR to allow the
NFS server to set the file's size, ctime, and mtime properly.
However, nfs_getattr() can be starved when a constant stream of application
writes to a file prevents nfs_wb_nocommit() from completing. This usually
results in hangs of programs doing a stat against an NFS file that is being
written. "ls -l" is a common victim of this behavior.
To prevent starvation, hold the file's i_mutex in nfs_getattr() to
freeze applications writes temporarily so the client can more quickly obtain
clean values for a file's size, mtime, and ctime.
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: Trond Myklebust <Trond.Myklebust@netapp.com>
The nfs_wcc_update_inode() function omits logic to convert the type of
the NFS on-the-wire value of a file's size (__u64) to the type of file
size value stored in struct inode (loff_t, which is signed).
Everywhere else in the NFS client I checked already correctly converts the
file size type.
This effects only very large files.
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: Trond Myklebust <Trond.Myklebust@netapp.com>
By using the TASK_KILLABLE infrastructure, we can get rid of the 'intr'
mount option. We have to use _killable everywhere instead of _interruptible
as we get rid of rpc_clnt_sigmask/sigunmask.
Signed-off-by: Liam R. Howlett <howlett@gmail.com>
Signed-off-by: Matthew Wilcox <willy@linux.intel.com>
Otherwise, we do end up breaking close-to-open semantics. We also end up
breaking some of the silly-rename tests in Connectathon on some setups.
Please refer to the bug-report at
http://bugzilla.linux-nfs.org/show_bug.cgi?id=150
Signed-off-by: Trond Myklebust <Trond.Myklebust@netapp.com>
lookup() and sillyrename() can race one another because the sillyrename()
completion cannot take the parent directory's inode->i_mutex since the
latter may be held by whoever is calling dput().
We therefore have little option but to add extra locking to ensure that
nfs_lookup() and nfs_atomic_open() do not race with the sillyrename
completion.
If somebody has looked up the sillyrenamed file in the meantime, we just
transfer the sillydelete information to the new dentry.
Please refer to the bug-report at
http://bugzilla.linux-nfs.org/show_bug.cgi?id=150
Signed-off-by: Trond Myklebust <Trond.Myklebust@netapp.com>
If the ATTR_KILL_S*ID bits are set then any mode change is only for clearing
the setuid/setgid bits. For NFS, skip the mode change and let the server
handle it.
Signed-off-by: Jeff Layton <jlayton@redhat.com>
Cc: Trond Myklebust <trond.myklebust@fys.uio.no>
Cc: "J. Bruce Fields" <bfields@fieldses.org>
Cc: Christoph Hellwig <hch@lst.de>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Slab constructors currently have a flags parameter that is never used. And
the order of the arguments is opposite to other slab functions. The object
pointer is placed before the kmem_cache pointer.
Convert
ctor(void *object, struct kmem_cache *s, unsigned long flags)
to
ctor(struct kmem_cache *s, void *object)
throughout the kernel
[akpm@linux-foundation.org: coupla fixes]
Signed-off-by: Christoph Lameter <clameter@sgi.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
This boot parameter will allow legacy 32-bit applications which call stat()
to continue to function even if the NFSv3/v4 server uses 64-bit inode
numbers.
Signed-off-by: Trond Myklebust <Trond.Myklebust@netapp.com>
If the cached attributes match the ones supplied in the fattr, then assume
we've revalidated the inode.
Signed-off-by: Trond Myklebust <Trond.Myklebust@netapp.com>
