Kudos to Neil Brown for spotting the problem:
"in nfs_sync_inode, there is effectively the sequence:
nfs_wait_on_requests
nfs_flush_inode
nfs_commit_inode
This seems a bit racy to me as if the only requests are on the
->commit list, and nfs_commit_inode is called separately after
nfs_wait_on_requests completes, and before nfs_commit_inode start
(say: by nfs_write_inode) then none of these function will return
>0, yet there will be some pending request that aren't waited for."
The solution is to search for requests to wait upon, search for dirty
requests, and search for uncommitted requests while holding the
nfsi->req_lock
The patch also cleans up nfs_sync_inode(), getting rid of the redundant
FLUSH_WAIT flag. It turns out that we were always setting it.
Signed-off-by: Trond Myklebust <Trond.Myklebust@netapp.com>
Currently NFS O_DIRECT writes use FILE_SYNC so that a COMMIT is not
necessary. This simplifies the internal logic, but this could be a
difficult workload for some servers.
Instead, let's send UNSTABLE writes, and after they all complete, send a
COMMIT for the dirty range. After the COMMIT returns successfully, then do
the wake_up or fire off aio_complete().
Test plan:
Async direct I/O tests against Solaris (or any server that requires
committed unstable writes). Reboot server during test.
Based on an earlier patch by Chuck Lever <cel@netapp.com>
Signed-off-by: Trond Myklebust <Trond.Myklebust@netapp.com>
Duplicate infrastructure from direct read path that will allow write
path to generate multiple write requests concurrently. This will
enable us to add support for aio in this path.
Temporarily we will lose the ability to do UNSTABLE writes followed by
a COMMIT in the direct write path. However, all applications I am
aware of that use NFS O_DIRECT currently write in relatively small
chunks, so this should not be inconvenient in any way.
Test plan:
Millions of fsx-odirect ops. OraSim.
Signed-off-by: Chuck Lever <cel@netapp.com>
Signed-off-by: Trond Myklebust <Trond.Myklebust@netapp.com>
Same callback hierarchy inversion as for the NFS write calls. This patch is
not strictly speaking needed by the O_DIRECT code, but avoids confusing
differences between the asynchronous read and write code.
Signed-off-by: Trond Myklebust <Trond.Myklebust@netapp.com>
This patch inverts the callback hierarchy for NFS write calls.
Instead of having the NFSv2/v3/v4-specific code set up the RPC callback
ops, we allow the original caller to do so. This allows for more
flexibility w.r.t. how to set up and tear down the nfs_write_data
structure while still allowing the NFSv3/v4 code to perform error
handling.
The greater flexibility is needed by the asynchronous O_DIRECT code, which
wants to be able to hold on to the original nfs_write_data structures after
the WRITE RPC call has completed in order to be able to replay them if the
COMMIT call determines that the server has rebooted.
Signed-off-by: Trond Myklebust <Trond.Myklebust@netapp.com>
read_cache_mtime is no longer used in nfs_inode. This patch removes
references of read_cache_mtime in the code comments.
Signed-off-by: Goldwyn Rodrigues <rgoldwyn@gmail.com>
Signed-off-by: Trond Myklebust <Trond.Myklebust@netapp.com>
The nfs_open_context may live longer than the file descriptor that spawned
it, so it needs to carry a reference to the vfsmount. If not, then
generic_shutdown_super() may end up being called before reads and writes
have been flushed out.
Make a couple of functions static while we're at it...
Signed-off-by: Trond Myklebust <Trond.Myklebust@netapp.com>
The compat syscalls are added to sys_ni.c since they are not defined if the
above CONFIG options are off. Also, nfs would not build with CONFIG_SYSCTL
off.
Noticed by Arthur Othieno.
Signed-off-by: Stephen Rothwell <sfr@canb.auug.org.au>
Cc: "David S. Miller" <davem@davemloft.net>
Cc: Trond Myklebust <trond.myklebust@fys.uio.no>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
The SuS states that a call to write() will cause mtime to be updated on
the file. In order to satisfy that requirement, we need to flush out
any cached writes in nfs_getattr().
Speed things up slightly by not committing the writes.
Signed-off-by: Trond Myklebust <Trond.Myklebust@netapp.com>
Most NFS server implementations allow up to 64KB reads and writes on the
wire. The Solaris NFS server allows up to a megabyte, for instance.
Now the Linux NFS client supports transfer sizes up to 1MB, too. This will
help reduce protocol and context switch overhead on read/write intensive NFS
workloads, and support larger atomic read and write operations on servers
that support them.
Test-plan:
Connectathon and iozone on mount point with wsize=rsize>32768 over TCP.
Tests with NFS over UDP to verify the maximum RPC payload size cap.
Signed-off-by: Chuck Lever <cel@netapp.com>
Signed-off-by: Trond Myklebust <Trond.Myklebust@netapp.com>
Shrink the RPC task structure. Instead of storing separate pointers
for task->tk_exit and task->tk_release, put them in a structure.
Also pass the user data pointer as a parameter instead of passing it via
task->tk_calldata. This enables us to nest callbacks.
Signed-off-by: Trond Myklebust <Trond.Myklebust@netapp.com>
Ensure we call unmap_mapping_range() and sync dirty pages to disk before
doing an NFS direct write.
Signed-off-by: Trond Myklebust <Trond.Myklebust@netapp.com>
RFC 3530 states that for OPEN_DOWNGRADE "The share_access and share_deny
bits specified must be exactly equal to the union of the share_access and
share_deny bits specified for some subset of the OPENs in effect for
current openowner on the current file.
Setattr is currently violating the NFSv4 rules for OPEN_DOWNGRADE in that
it may cause a downgrade from OPEN4_SHARE_ACCESS_BOTH to
OPEN4_SHARE_ACCESS_WRITE despite the fact that there exists no open file
with O_WRONLY access mode.
Fix the problem by replacing nfs4_find_state() with a modified version of
nfs_find_open_context().
Signed-off-by: Trond Myklebust <Trond.Myklebust@netapp.com>
Since we almost always call nfs_end_data_update() after we called
nfs_refresh_inode(), we now end up marking the inode metadata
as needing revalidation immediately after having updated it.
This patch rearranges things so that we mark the inode as needing
revalidation _before_ we call nfs_refresh_inode() on those operations
that need it.
Signed-off-by: Trond Myklebust <Trond.Myklebust@netapp.com>
Allow nfs_refresh_inode() also to update attributes on the inode if the
RPC call was sent after the last call to nfs_update_inode().
Signed-off-by: Trond Myklebust <Trond.Myklebust@netapp.com>
Each transport implementation can now set unique bind, connect,
reestablishment, and idle timeout values. These are variables,
allowing the values to be modified dynamically. This permits
exponential backoff of any of these values, for instance.
As an example, we implement exponential backoff for the connection
reestablishment timeout.
Test-plan:
Destructive testing (unplugging the network temporarily). Connectathon
with UDP and TCP.
Signed-off-by: Chuck Lever <cel@netapp.com>
Signed-off-by: Trond Myklebust <Trond.Myklebust@netapp.com>
Down the road we want to eliminate the use of the global kernel lock entirely
from the NFS client. To do this, we need to protect the fields in the
nfs_inode structure adequately. Start by serializing updates to the
"cache_validity" field.
Note this change addresses an SMP hang found by njw@osdl.org, where processes
deadlock because nfs_end_data_update and nfs_revalidate_mapping update the
"cache_validity" field without proper serialization.
Test plan:
Millions of fsx ops on SMP clients. Run Nick Wilson's breaknfs program on
large SMP clients.
Signed-off-by: Chuck Lever <cel@netapp.com>
Cc: Trond Myklebust <trond.myklebust@fys.uio.no>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
Introduce atomic bitops to manipulate the bits in the nfs_inode structure's
"flags" field.
Using bitops means we can use a generic wait_on_bit call instead of an ad hoc
locking scheme in fs/nfs/inode.c, so we can remove the "nfs_i_wait" field from
nfs_inode at the same time.
The other new flags field will continue to use bitmask and logic AND and OR.
This permits several flags to be set at the same time efficiently. The
following patch adds a spin lock to protect these flags, and this spin lock
will later cover other fields in the nfs_inode structure, amortizing the cost
of using this type of serialization.
Test plan:
Millions of fsx ops on SMP clients.
Signed-off-by: Chuck Lever <cel@netapp.com>
Cc: Trond Myklebust <trond.myklebust@fys.uio.no>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
Certain bits in nfsi->flags can be manipulated with atomic bitops, and some
are better manipulated via logical bitmask operations.
This patch splits the flags field into two. The next patch introduces atomic
bitops for one of the fields.
Test plan:
Millions of fsx ops on SMP clients.
Signed-off-by: Chuck Lever <cel@netapp.com>
Cc: Trond Myklebust <trond.myklebust@fys.uio.no>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
When the client performs an exclusive create and opens the file for writing,
a Netapp filer will first create the file using the mode 01777. It does this
since an NFSv3/v4 exclusive create cannot immediately set the mode bits.
The 01777 mode then gets put into the inode->i_mode. After the file creation
is successful, we then do a setattr to change the mode to the correct value
(as per the NFS spec).
The problem is that nfs_refresh_inode() no longer updates inode->i_mode, so
the latter retains the 01777 mode. A bit later, the VFS notices this, and calls
remove_suid(). This of course now resets the file mode to inode->i_mode & 0777.
Hey presto, the file mode on the server is now magically changed to 0777. Duh...
Fixes http://bugzilla.linux-nfs.org/show_bug.cgi?id=32
Signed-off-by: Trond Myklebust <Trond.Myklebust@netapp.com>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
Unless we're doing O_APPEND writes, we really don't care about revalidating
the file length. Just make sure that we catch any page cache invalidations.
Signed-off-by: Trond Myklebust <Trond.Myklebust@netapp.com>
Instead of looking at whether or not the file is open for writes before
we accept to update the length using the server value, we should rather
be looking at whether or not we are currently caching any writes.
Failure to do so means in particular that we're not updating the file
length correctly after obtaining a POSIX or BSD lock.
Signed-off-by: Trond Myklebust <Trond.Myklebust@netapp.com>
NFSv3 currently returns the unsigned 64-bit cookie directly to
userspace. The following patch causes the kernel to generate
loff_t offsets for the benefit of userland.
The current server-generated READDIR cookie is cached in the
nfs_open_context instead of in filp->f_pos, so we still end up work
correctly under directory insertions/deletion.
Signed-off-by: Olivier Galibert <galibert@pobox.com>
Signed-off-by: Trond Myklebust <Trond.Myklebust@netapp.com>
Attach acls to inodes in the icache to avoid unnecessary GETACL RPC
round-trips. As long as the client doesn't retrieve any acls itself, only the
default acls of exiting directories and the default and access acls of new
directories will end up in the cache, which preserves some memory compared to
always caching the access and default acl of all files.
Signed-off-by: Andreas Gruenbacher <agruen@suse.de>
Acked-by: Olaf Kirch <okir@suse.de>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Trond Myklebust <Trond.Myklebust@netapp.com>
NFSv3 has no concept of a umask on the server side: The client applies
the umask locally, and sends the effective permissions to the server.
This behavior is wrong when files are created in a directory that has a
default ACL. In this case, the umask is supposed to be ignored, and
only the default ACL determines the file's effective permissions.
Usually its the server's task to conditionally apply the umask. But
since the server knows nothing about the umask, we have to do it on the
client side. This patch tries to fetch the parent directory's default
ACL before creating a new file, computes the appropriate create mode to
send to the server, and finally sets the new file's access and default
acl appropriately.
Many thanks to Buck Huppmann <buchk@pobox.com> for sending the initial
version of this patch, as well as for arguing why we need this change.
Signed-off-by: Andreas Gruenbacher <agruen@suse.de>
Acked-by: Olaf Kirch <okir@suse.de>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Trond Myklebust <Trond.Myklebust@netapp.com>
This adds acl support fo nfs clients via the NFSACL protocol extension, by
implementing the getxattr, listxattr, setxattr, and removexattr iops for the
system.posix_acl_access and system.posix_acl_default attributes. This patch
implements a dumb version that uses no caching (and thus adds some overhead).
(Another patch in this patchset adds caching as well.)
Signed-off-by: Andreas Gruenbacher <agruen@suse.de>
Acked-by: Olaf Kirch <okir@suse.de>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Trond Myklebust <Trond.Myklebust@netapp.com>
Add nfs4_acl field to the nfs_inode, and use it to cache acls. Only cache
acls of size up to a page. Also prepare for up to a page of acl data even
when the user doesn't pass in a buffer, as when they want to get the acl
length to decide what size buffer to allocate.
Signed-off-by: J. Bruce Fields <bfields@citi.umich.edu>
Signed-off-by: Trond Myklebust <Trond.Myklebust@netapp.com>
Initial git repository build. I'm not bothering with the full history,
even though we have it. We can create a separate "historical" git
archive of that later if we want to, and in the meantime it's about
3.2GB when imported into git - space that would just make the early
git days unnecessarily complicated, when we don't have a lot of good
infrastructure for it.
Let it rip!
