NFSoRDMA Client Updates for Linux 5.5
New Features:
- New tracepoints for congestion control and Local Invalidate WRs
Bugfixes and Cleanups:
- Eliminate log noise in call_reserveresult
- Fix unstable connections after a reconnect
- Clean up some code duplication
- Close race between waking a sender and posting a receive
- Fix MR list corruption, and clean up MR usage
- Remove unused rpcrdma_sendctx fields
- Try to avoid DMA mapping pages if it is too costly
- Wake pending tasks if connection fails
- Replace some dprintk()s with tracepoints
If nfs4_delegreturn_prepare needs to wait for a layoutreturn to complete
then make sure we drop the sequence slot if we hold it.
Fixes: 1c5bd76d17 ("pNFS: Enable layoutreturn operation for return-on-close")
Signed-off-by: Trond Myklebust <[email protected]>
If the server returns a bad or dead session error, the we don't want
to update the session slot number, but just immediately schedule
recovery and allow it to proceed.
We can/should then remove handling in other places
Fixes: 3453d5708b ("NFSv4.1: Avoid false retries when RPC calls are interrupted")
Signed-off-by: Trond Myklebust <[email protected]>
We shouldn't insert things into the NFSPROC4_CLNT enums, since that
causes the nfsstat array to be reordered.
Signed-off-by: Trond Myklebust <[email protected]>
Jon Hunter: "I have been tracking down another suspend/NFS related
issue where again I am seeing random delays exiting suspend. The delays
can be up to a couple minutes in the worst case and this is causing a
suspend test we have to fail."
Change the use of a deferrable work to a standard delayed one.
Reported-by: Jon Hunter <[email protected]>
Tested-by: Jon Hunter <[email protected]>
Fixes: 7e0a0e38fc ("SUNRPC: Replace the queue timer with a delayed work function")
Signed-off-by: Trond Myklebust <[email protected]>
If the server returns NFS4ERR_OLD_STATEID, then just skip retrying the
GETATTR when replaying the delegreturn compound. We know nothing will
have changed on the server.
Signed-off-by: Trond Myklebust <[email protected]>
If the server returns NFS4ERR_OLD_STATEID in response to our delegreturn,
we want to sync to the most recent seqid for the delegation stateid. However
if we are already at the most recent, we have two possibilities:
- an OPEN reply is still outstanding and will return a new seqid
- an earlier OPEN reply was dropped on the floor due to a timeout.
In the latter case, we may end up unable to complete the delegreturn,
so we want to bump the seqid to a value greater than the cached value.
While this may cause us to lose the delegation in the former case,
it should now be safe to assume that the client will replay the OPEN
if necessary in order to get a new valid stateid.
Signed-off-by: Trond Myklebust <[email protected]>
If the server returns the same delegation in an open that we just used
in a delegreturn, we need to ensure we don't apply that stateid if
the delegreturn has freed it on the server.
To do so, we ensure that we do not free the storage for the delegation
until either it is replaced by a new one, or we throw the inode out of
cache.
Signed-off-by: Trond Myklebust <[email protected]>
In nfs_inode_find_state_and_recover() we want to mark for recovery
only those stateids that match or are older than the supplied
stateid parameter.
Signed-off-by: Trond Myklebust <[email protected]>
Fix the checks in nfs4_inode_make_writeable() to ignore the case where
we hold no delegations. Currently, in such a case, we automatically
flush writes.
Signed-off-by: Trond Myklebust <[email protected]>
Ensure that we check that the delegation is valid in
nfs4_return_incompatible_delegation() before we try to return it.
Signed-off-by: Trond Myklebust <[email protected]>
If the delegation was revoked, or is already being returned, just
clear the NFS_DELEGATION_RETURN and NFS_DELEGATION_RETURN_IF_CLOSED
flags and keep going.
Signed-off-by: Trond Myklebust <[email protected]>
If we revoke a delegation, but the stateid's seqid is newer, then
ensure we update the seqid when marking the delegation as revoked.
Signed-off-by: Trond Myklebust <[email protected]>
If the server sent us a new delegation stateid that is more recent than
the one that got revoked, then clear the NFS_DELEGATION_REVOKED flag.
Signed-off-by: Trond Myklebust <[email protected]>
Add a check to ensure that we haven't already removed the delegation
from the inode after we take all the relevant locks.
Signed-off-by: Trond Myklebust <[email protected]>
Rename nfs_inode_return_delegation_noreclaim() to
nfs_inode_evict_delegation(), which better describes what it
does.
Signed-off-by: Trond Myklebust <[email protected]>
If we're processsing a delegation recall, ignore the delegations that
have already been revoked or returned.
Signed-off-by: Trond Myklebust <[email protected]>
NFSv2, v3 and NFSv4 servers often have duplicate replay caches that look
at the source port when deciding whether or not an RPC call is a replay
of a previous call. This requires clients to perform strange TCP gymnastics
in order to ensure that when they reconnect to the server, they bind
to the same source port.
NFSv4.1 and NFSv4.2 have sessions that provide proper replay semantics,
that do not look at the source port of the connection. This patch therefore
ensures they can ignore the rebind requirement.
Signed-off-by: Trond Myklebust <[email protected]>
If a NFSv3 server is being used as both a DS and as a regular NFSv3 server,
we may want to keep the IO traffic on a separate TCP connection, since
it will typically have very different timeout characteristics.
This patch therefore sets up a flag to separate the two modes of operation
for the nfs_client.
Signed-off-by: Trond Myklebust <[email protected]>
Connecting to the DS is a non-interactive, asynchronous task, so there is
no reason to fire up an extra RPC null ping in order to ensure that the
server is up.
Signed-off-by: Trond Myklebust <[email protected]>
Add a flag to tell the nfs_client it should set RPC_CLNT_CREATE_NOPING when
creating the rpc client.
Signed-off-by: Trond Myklebust <[email protected]>
Simplify the struct iattr timestamp encoding by skipping the step of
an intermediate struct timespec.
Signed-off-by: Trond Myklebust <[email protected]>
Simplify the struct iattr timestamp encoding by skipping the step of
an intermediate struct timespec.
Signed-off-by: Trond Myklebust <[email protected]>
Convert the NFSv4 callbacks to use struct timestamp64, rather than
truncating times to 32-bit values.
Signed-off-by: Trond Myklebust <[email protected]>
NFSv4 supports 64-bit timestamps, so there is no point in converting
the struct iattr timestamps to 32-bits before encoding.
Signed-off-by: Trond Myklebust <[email protected]>
NFSv4 supports 64-bit times, so we should switch to using struct
timespec64 when decoding attributes.
Signed-off-by: Trond Myklebust <[email protected]>
If we set nfs_mountpoint_expiry_timeout to a negative value, then
allow that to imply that we do not expire NFSv4 submounts.
Signed-off-by: Trond Myklebust <[email protected]>
We want to throw out the attrbute if it refers to the mounted on fileid,
and not the real fileid. However we do not want to block cache consistency
updates from NFSv4 writes.
Reported-by: Murphy Zhou <[email protected]>
Fixes: 7e10cc25bf ("NFS: Don't refresh attributes with mounted-on-file...")
Signed-off-by: Trond Myklebust <[email protected]>
Ensure we update the write result count on success, since the
RPC call itself does not do so.
Reported-by: Jan Stancek <[email protected]>
Reported-by: Naresh Kamboju <[email protected]>
Signed-off-by: Trond Myklebust <[email protected]>
Tested-by: Jan Stancek <[email protected]>
If we received a reply from the server with a zero length read and
no error, then that implies we are at eof.
Signed-off-by: Trond Myklebust <[email protected]>
If writepage()/writepages() saw an error, but handled it without
reporting it, we should not be re-reporting that error on exit.
Signed-off-by: Trond Myklebust <[email protected]>
If the client attempts to read a page, but the read fails due to some
spurious error (e.g. an ACCESS error or a timeout, ...) then we need
to allow other processes to retry.
Also try to report errors correctly when doing a synchronous readpage.
Signed-off-by: Trond Myklebust <[email protected]>
If the mount is hard, we should ignore the 'io_maxretrans' module
parameter so that we always keep retrying.
Signed-off-by: Trond Myklebust <[email protected]>
If the connection breaks while we're waiting for a reply from the
server, then we want to immediately try to reconnect.
Fixes: ec6017d903 ("SUNRPC fix regression in umount of a secure mount")
Signed-off-by: Trond Myklebust <[email protected]>
This reverts commit a79f194aa4.
The mechanism for aborting I/O is racy, since we are not guaranteed that
the request is asleep while we're changing both task->tk_status and
task->tk_action.
Signed-off-by: Trond Myklebust <[email protected]>
Cc: [email protected] # v5.1
If a connect or bind attempt returns EADDRINUSE, that means we want to
retry with a different port. It is not a fatal connection error.
Similarly, ENOBUFS is not fatal, but just indicates a memory allocation
issue. Retry after a short delay.
Signed-off-by: Trond Myklebust <[email protected]>
The pNFS/flexfiles I/O requests are sent with the SOFTCONN flag set, so
they automatically time out if the connection breaks. It should
therefore not be necessary to have the soft flag set in addition.
Fixes: 5f01d95394 ("nfs41: create NFSv3 DS connection if specified")
Signed-off-by: Trond Myklebust <[email protected]>
Don't handle errors in call_bind_status()/call_connect_status()
if it turns out that a previous call caused it to succeed.
Signed-off-by: Trond Myklebust <[email protected]>
Cc: [email protected] # v5.1+
If the writeback error is fatal, we need to remove the tracking structures
(i.e. the nfs_page) from the inode.
Fixes: 6fbda89b25 ("NFS: Replace custom error reporting mechanism...")
Signed-off-by: Trond Myklebust <[email protected]>
Initialise the result count to 0 rather than initialising it to the
argument count. The reason is that we want to ensure we record the
I/O stats correctly in the case where an error is returned (for
instance in the layoutstats).
Signed-off-by: Trond Myklebust <[email protected]>
If the attempt to resend the I/O results in no bytes being read/written,
we must ensure that we report the error.
Signed-off-by: Trond Myklebust <[email protected]>
Fixes: 0a00b77b33 ("nfs: mirroring support for direct io")
Cc: [email protected] # v3.20+
If the attempt to resend the pages fails, we need to ensure that we
clean up those pages that were not transmitted.
Fixes: d600ad1f2b ("NFS41: pop some layoutget errors to application")
Signed-off-by: Trond Myklebust <[email protected]>
Cc: [email protected] # v4.5+
If the file turns out to be of the wrong type after opening, we want
to revalidate the path and retry, so return EOPENSTALE rather than
ESTALE.
Signed-off-by: Trond Myklebust <[email protected]>
Currently, we are translating RPC level errors such as timeouts,
as well as interrupts etc into EOPENSTALE, which forces a single
replay of the open attempt. What we actually want to do is
force the replay only in the cases where the returned error
indicates that the file may have changed on the server.
So the fix is to spell out the exact set of errors where we want
to return EOPENSTALE.
Signed-off-by: Trond Myklebust <[email protected]>
If we've been given the attributes of the mounted-on-file, then do not
use those to check or update the attributes on the application-visible
inode.
Signed-off-by: Trond Myklebust <[email protected]>
Ensure that the state recovery code handles ETIMEDOUT correctly,
and also that we set RPC_TASK_TIMEOUT when recovering open state.
Signed-off-by: Trond Myklebust <[email protected]>
People are reporing seeing fscache errors being reported concerning
duplicate cookies even in cases where they are not setting up fscache
at all. The rule needs to be that if fscache is not enabled, then it
should have no side effects at all.
To ensure this is the case, we disable fscache completely on all superblocks
for which the 'fsc' mount option was not set. In order to avoid issues
with '-oremount', we also disable the ability to turn fscache on via
remount.
Fixes: f1fe29b4a0 ("NFS: Use i_writecount to control whether...")
Link: https://bugzilla.kernel.org/show_bug.cgi?id=200145
Signed-off-by: Trond Myklebust <[email protected]>
Cc: Steve Dickson <[email protected]>
Cc: David Howells <[email protected]>
If the user specifies an open mode of 3, then we don't have a NFSv4 state
attached to the context, and so we Oops when we try to dereference it.
Reported-by: Olga Kornievskaia <[email protected]>
Fixes: 29b59f9416 ("NFSv4: change nfs4_do_setattr to take...")
Signed-off-by: Trond Myklebust <[email protected]>
Cc: [email protected] # v4.10: 991eedb137: NFSv4: Only pass the...
Cc: [email protected] # v4.10+
John Hubbard reports seeing the following stack trace:
nfs4_do_reclaim
rcu_read_lock /* we are now in_atomic() and must not sleep */
nfs4_purge_state_owners
nfs4_free_state_owner
nfs4_destroy_seqid_counter
rpc_destroy_wait_queue
cancel_delayed_work_sync
__cancel_work_timer
__flush_work
start_flush_work
might_sleep:
(kernel/workqueue.c:2975: BUG)
The solution is to separate out the freeing of the state owners
from nfs4_purge_state_owners(), and perform that outside the atomic
context.
Reported-by: John Hubbard <[email protected]>
Fixes: 0aaaf5c424 ("NFS: Cache state owners after files are closed")
Signed-off-by: Trond Myklebust <[email protected]>
Ensure that we always check the return value of update_open_stateid()
so that we can retry if the update of local state failed. This fixes
infinite looping on state recovery.
Fixes: e23008ec81 ("NFSv4 reduce attribute requests for open reclaim")
Signed-off-by: Trond Myklebust <[email protected]>
Cc: [email protected] # v3.7+
Fix nfs_reap_expired_delegations() to ensure that we only reap delegations
that are actually expired, rather than triggering on random errors.
Fixes: 45870d6909 ("NFSv4.1: Test delegation stateids when server...")
Signed-off-by: Trond Myklebust <[email protected]>
The logic for checking in nfs41_check_open_stateid() whether the state
is supported by a delegation is inverted. In addition, it makes more
sense to perform that check before we check for expired locks.
Fixes: 8a64c4ef10 ("NFSv4.1: Even if the stateid is OK,...")
Signed-off-by: Trond Myklebust <[email protected]>
In pnfs_update_layout() ensure that we do report any fatal errors from
nfs4_select_rw_stateid().
Fixes: d9aba2b40d ("NFSv4: Don't use the zero stateid with layoutget")
Signed-off-by: Trond Myklebust <[email protected]>
If the server returns with EAGAIN when we're trying to recover from
a server reboot, we currently delay for 1 second, but then mark the
stateid as needing recovery after the grace period has expired.
Instead, we should just retry the same recovery process immediately
after the 1 second delay. Break out of the loop after 10 retries.
Fixes: 35a61606a6 ("NFS: Reduce indentation of the switch statement...")
Signed-off-by: Trond Myklebust <[email protected]>
Once we clear the NFS_DELEGATED_STATE flag, we're telling
nfs_delegation_claim_opens() that we're done recovering all open state
for that stateid, so we really need to ensure that we test for all
open modes that are currently cached and recover them before exiting
nfs4_open_delegation_recall().
Fixes: 24311f8841 ("NFSv4: Recovery of recalled read delegations...")
Signed-off-by: Trond Myklebust <[email protected]>
Cc: [email protected] # v4.3+
It is unsafe to dereference delegation outside the rcu lock, and in
any case, the refcount is guaranteed held if cred is non-zero.
Signed-off-by: Trond Myklebust <[email protected]>
Add tracepoints to allow debugging of the event chain leading to
a pnfs fallback to doing I/O through the MDS.
Signed-off-by: Trond Myklebust <[email protected]>
If the client has to stop in pnfs_update_layout() to wait for another
layoutget to complete, it currently exits and defaults to I/O through
the MDS if the layoutget was successful.
Fixes: d03360aaf5 ("pNFS: Ensure we return the error if someone kills...")
Signed-off-by: Trond Myklebust <[email protected]>
Cc: [email protected] # v4.20+
The bvec tracks the list of pages, so if the number of pages changes
due to a re-encode, we need to reset the bvec as well.
Fixes: 277e4ab7d5 ("SUNRPC: Simplify TCP receive code by switching...")
Signed-off-by: Trond Myklebust <[email protected]>
Cc: [email protected] # v4.20+
mirror->mirror_ds can be NULL if uninitialised, but can contain
a PTR_ERR() if call to GETDEVICEINFO failed.
Fixes: 65990d1afb ("pNFS/flexfiles: Fix a deadlock on LAYOUTGET")
Signed-off-by: Trond Myklebust <[email protected]>
Cc: [email protected] # 4.10+
The NFSv4.1 protocol explicitly forbids us from using the zero stateid
together with layoutget, so when we see that nfs4_select_rw_stateid()
is unable to return a valid delegation, lock or open stateid, then
we should initiate recovery and retry.
Signed-off-by: Trond Myklebust <[email protected]>
Add a per-transport maximum limit in the socket case, and add
helpers to allow the NFSv4 code to discover that limit.
Signed-off-by: Trond Myklebust <[email protected]>
When looking for the next transport to use for an RPC call, skip those
that are in the process of being destroyed and that have a zero refcount.
Signed-off-by: Trond Myklebust <[email protected]>
When checking whether or not a particular xprt queue length is shorter
than the average queue length for all xprts, prefer to use multiplication
rather than division for performance reasons.
Signed-off-by: Trond Myklebust <[email protected]>
If the stateid is the zero or invalid stateid, then it is pointless
to attempt to use it for recovery. In that case, try to fall back
to using the open state stateid, or just doing a general recovery
of all state on a given inode.
Signed-off-by: Trond Myklebust <[email protected]>
Ensure that we do the required accounting for the round robin queue
when the caller to rpc_init_task() has passed in a transport to be
used.
Reported-by: Olga Kornievskaia <[email protected]>
Reported-by: Neil Brown <[email protected]>
Signed-off-by: Trond Myklebust <[email protected]>
NFSoRDMA client updates for 5.3
New features:
- Add a way to place MRs back on the free list
- Reduce context switching
- Add new trace events
Bugfixes and cleanups:
- Fix a BUG when tracing is enabled with NFSv4.1
- Fix a use-after-free in rpcrdma_post_recvs
- Replace use of xdr_stream_pos in rpcrdma_marshal_req
- Fix occasional transport deadlock
- Fix show_nfs_errors macros, other tracing improvements
- Remove RPCRDMA_REQ_F_PENDING and fr_state
- Various simplifications and refactors
Now that the VM promises never to recurse back into the filesystem
layer on writeback, remove all the GFP_NOFS references etc from
the generic writeback code.
Signed-off-by: Trond Myklebust <[email protected]>
In order to identify containers to the NFS client, we add a per-net
sysfs attribute that udev can fill with the appropriate identifier.
The identifier could be a unique hostname, but in most cases it
will probably be a persisted uuid.
Signed-off-by: Trond Myklebust <[email protected]>
If the client detects that close-to-open cache consistency has been
violated, and that the file or directory has been changed on the
server, then do a cache invalidation when we're done working with
the file.
The reason we don't do an immediate cache invalidation is that we
want to avoid performance problems due to false positives. Also,
note that we cannot guarantee cache consistency in this situation
even if we do invalidate the cache.
Signed-off-by: Trond Myklebust <[email protected]>
The queue timer function, which walks the RPC queue in order to locate
candidates for waking up is one of the current constraints against
removing the bh-safe queue spin locks. Replace it with a delayed
work queue, so that we can do the actual rpc task wake ups from an
ordinary process context.
Signed-off-by: Trond Myklebust <[email protected]>
According to the open() manpage, Linux reserves the access mode 3
to mean "check for read and write permission on the file and return
a file descriptor that can't be used for reading or writing."
Currently, the NFSv4 code will ask the server to open the file,
and will use an incorrect share access mode of 0. Since it has
an incorrect share access mode, the client later forgets to send
a corresponding close, meaning it can leak stateids on the server.
Fixes: ce4ef7c0a8 ("NFS: Split out NFS v4 file operations")
Cc: [email protected] # 3.6+
Signed-off-by: Trond Myklebust <[email protected]>
When mapping the NFSv4 context to an open mode and access mode,
we need to treat the FMODE_EXEC flag differently. For the open
mode, FMODE_EXEC means we need read share access. For the access
mode checking, we need to verify that the user actually has
execute access.
Signed-off-by: Trond Myklebust <[email protected]>
The addition of rpc_check_timeout() to call_decode causes an Oops
when the RPCSEC_GSS credential is rejected.
The reason is that rpc_decode_header() will call xprt_release() in
order to free task->tk_rqstp, which is needed by rpc_check_timeout()
to check whether or not we should exit due to a soft timeout.
The fix is to move the call to xprt_release() into call_decode() so
we can perform it after rpc_check_timeout().
Reported-by: Olga Kornievskaia <[email protected]>
Reported-by: Nick Bowler <[email protected]>
Fixes: cea57789e4 ("SUNRPC: Clean up")
Cc: [email protected] # v5.1+
Signed-off-by: Trond Myklebust <[email protected]>
Signed-off-by: Anna Schumaker <[email protected]>
If a request transmission fails due to write space or slot unavailability
errors, but the queued task then gets transmitted before it has time to
process the error in call_transmit_status() or call_bc_transmit_status(),
we need to suppress the transmission error code to prevent it from leaking
out of the RPC layer.
Reported-by: Chuck Lever <[email protected]>
Signed-off-by: Trond Myklebust <[email protected]>
Tested-by: Chuck Lever <[email protected]>
This reverts commit 009a82f643.
The ability to optimise here relies on compiler being able to optimise
away tail calls to avoid stack overflows. Unfortunately, we are seeing
reports of problems, so let's just revert.
Reported-by: Daniel Mack <[email protected]>
Signed-off-by: Trond Myklebust <[email protected]>
Trond Myklebust