When the socket is closed, we currently send an EAGAIN error to all
pending requests in order to ask them to retransmit. Use ENOTCONN
instead, to ensure that they try to reconnect before attempting to
transmit.
This also helps SOFTCONN tasks to behave correctly in this
situation.
Signed-off-by: Trond Myklebust <trond.myklebust@hammerspace.com>
If a layout segment gets invalidated while a pNFS I/O operation
is queued for transmission, then we ideally want to abort
immediately. This is particularly the case when there is a large
number of I/O related RPCs queued in the RPC layer, and the layout
segment gets invalidated due to an ENOSPC error, or an EACCES (because
the client was fenced). We may end up forced to spam the MDS with a
lot of otherwise unnecessary LAYOUTERRORs after that I/O fails.
Signed-off-by: Trond Myklebust <trond.myklebust@hammerspace.com>
Set memalloc_nofs_save() on all the rpciod/xprtiod jobs so that we
ensure memory allocations for asynchronous rpc calls don't ever end
up recursing back to the NFS layer for memory reclaim.
Signed-off-by: Trond Myklebust <trond.myklebust@hammerspace.com>
Add infrastructure for trace points in the RPC_AUTH_GSS kernel
module, and add a few sample trace points. These report exceptional
or unexpected events, and observe the assignment of GSS sequence
numbers.
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: Anna Schumaker <Anna.Schumaker@Netapp.com>
When using Kerberos with v4.20, I've observed frequent connection
loss on heavy workloads. I traced it down to the client underrunning
the GSS sequence number window -- NFS servers are required to drop
the RPC with the low sequence number, and also drop the connection
to signal that an RPC was dropped.
Bisected to commit 918f3c1fe8 ("SUNRPC: Improve latency for
interactive tasks").
I've got a one-line workaround for this issue, which is easy to
backport to v4.20 while a more permanent solution is being derived.
Essentially, tk_owner-based sorting is disabled for RPCs that carry
a GSS sequence number.
Fixes: 918f3c1fe8 ("SUNRPC: Improve latency for interactive ... ")
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: Anna Schumaker <Anna.Schumaker@Netapp.com>
When we resend a request, ensure that the 'rq_bytes_sent' is reset
to zero.
Signed-off-by: Trond Myklebust <trond.myklebust@hammerspace.com>
Signed-off-by: Anna Schumaker <Anna.Schumaker@Netapp.com>
Over the years, xprt_connect_status() has been superseded by
call_connect_status(), which now handles all the errors that
xprt_connect_status() does and more. Since the latter converts
all errors that it doesn't recognise to EIO, then it is time
for it to be retired.
Reported-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: Trond Myklebust <trond.myklebust@hammerspace.com>
Tested-by: Chuck Lever <chuck.lever@oracle.com>
When the socket is closed, we need to call xprt_disconnect_done() in order
to clean up the XPRT_WRITE_SPACE flag, and wake up the sleeping tasks.
However, we also want to ensure that we don't wake them up before the socket
is closed, since that would cause thundering herd issues with everyone
piling up to retransmit before the TCP shutdown dance has completed.
Only the task that holds XPRT_LOCKED needs to wake up early in order to
allow the close to complete.
Reported-by: Dave Wysochanski <dwysocha@redhat.com>
Reported-by: Scott Mayhew <smayhew@redhat.com>
Cc: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: Trond Myklebust <trond.myklebust@hammerspace.com>
Tested-by: Chuck Lever <chuck.lever@oracle.com>
If an asynchronous connection attempt completes while another task is
in xprt_connect(), then the call to rpc_sleep_on() could end up
racing with the call to xprt_wake_pending_tasks().
So add a second test of the connection state after we've put the
task to sleep and set the XPRT_CONNECTING flag, when we know that there
can be no asynchronous connection attempts still in progress.
Fixes: 0b9e794313 ("SUNRPC: Move the test for XPRT_CONNECTING into...")
Signed-off-by: Trond Myklebust <trond.myklebust@hammerspace.com>
If we retransmit an RPC request, we currently end up clobbering the
value of req->rq_rcv_buf.bvec that was allocated by the initial call to
xprt_request_prepare(req).
Signed-off-by: Trond Myklebust <trond.myklebust@hammerspace.com>
Stable bugfixes:
- Reset credit grant properly after a disconnect
Other bugfixes and cleanups:
- xprt_release_rqst_cong is called outside of transport_lock
- Create more MRs at a time and toss out old ones during recovery
- Various improvements to the RDMA connection and disconnection code:
- Improve naming of trace events, functions, and variables
- Add documenting comments
- Fix metrics and stats reporting
- Fix a tracepoint sparse warning
-----BEGIN PGP SIGNATURE-----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=SxSD
-----END PGP SIGNATURE-----
Merge tag 'nfs-rdma-for-4.20-1' of git://git.linux-nfs.org/projects/anna/linux-nfs
NFS RDMA client updates for Linux 4.20
Stable bugfixes:
- Reset credit grant properly after a disconnect
Other bugfixes and cleanups:
- xprt_release_rqst_cong is called outside of transport_lock
- Create more MRs at a time and toss out old ones during recovery
- Various improvements to the RDMA connection and disconnection code:
- Improve naming of trace events, functions, and variables
- Add documenting comments
- Fix metrics and stats reporting
- Fix a tracepoint sparse warning
Signed-off-by: Trond Myklebust <trond.myklebust@hammerspace.com>
For TCP, the logic in xprt_connect_status is currently never invoked
to record a successful connection. Commit 2a4919919a ("SUNRPC:
Return EAGAIN instead of ENOTCONN when waking up xprt->pending")
changed the way TCP xprt's are awoken after a connect succeeds.
Instead, change connection-oriented transports to bump connect_count
and compute connect_time the moment that XPRT_CONNECTED is set.
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: Anna Schumaker <Anna.Schumaker@Netapp.com>
Add a bvec array to struct xdr_buf, and have the client allocate it
when we need to receive data into pages.
Signed-off-by: Trond Myklebust <trond.myklebust@hammerspace.com>
We no longer need priority semantics on the xprt->sending queue, because
the order in which tasks are sent is now dictated by their position in
the send queue.
Note that the backlog queue remains a priority queue, meaning that
slot resources are still managed in order of task priority.
Signed-off-by: Trond Myklebust <trond.myklebust@hammerspace.com>
If the server is slow, we can find ourselves with quite a lot of entries
on the receive queue. Converting the search from an O(n) to O(log(n))
can make a significant difference, particularly since we have to hold
a number of locks while searching.
Signed-off-by: Trond Myklebust <trond.myklebust@hammerspace.com>
Treat socket write space handling in the same way we now treat transport
congestion: by denying the XPRT_LOCK until the transport signals that it
has free buffer space.
Signed-off-by: Trond Myklebust <trond.myklebust@hammerspace.com>
The theory was that we would need to grab the socket lock anyway, so we
might as well use it to gate the allocation of RPC slots for a TCP
socket.
Signed-off-by: Trond Myklebust <trond.myklebust@hammerspace.com>
Rather than forcing each and every RPC task to grab the socket write
lock in order to send itself, we allow whichever task is holding the
write lock to attempt to drain the entire transmit queue.
Signed-off-by: Trond Myklebust <trond.myklebust@hammerspace.com>
Avoid memory starvation by giving RPCs that are tagged with the
RPC_TASK_SWAPPER flag the highest priority.
Signed-off-by: Trond Myklebust <trond.myklebust@hammerspace.com>
Both RDMA and UDP transports require the request to get a "congestion control"
credit before they can be transmitted. Right now, this is done when
the request locks the socket. We'd like it to happen when a request attempts
to be transmitted for the first time.
In order to support retransmission of requests that already hold such
credits, we also want to ensure that they get queued first, so that we
don't deadlock with requests that have yet to obtain a credit.
Signed-off-by: Trond Myklebust <trond.myklebust@hammerspace.com>
One of the intentions with the priority queues was to ensure that no
single process can hog the transport. The field task->tk_owner therefore
identifies the RPC call's origin, and is intended to allow the RPC layer
to organise queues for fairness.
This commit therefore modifies the transmit queue to group requests
by task->tk_owner, and ensures that we round robin among those groups.
Signed-off-by: Trond Myklebust <trond.myklebust@hammerspace.com>
Remove the checks for whether or not we need to transmit, and whether
or not a reply has been received. Those are already handled in
call_transmit() itself.
Signed-off-by: Trond Myklebust <trond.myklebust@hammerspace.com>
When we shift to using the transmit queue, then the task that holds the
write lock will not necessarily be the same as the one being transmitted.
Signed-off-by: Trond Myklebust <trond.myklebust@hammerspace.com>
Fix up the back channel code to recognise that it has already been
transmitted, so does not need to be called again.
Also ensure that we set req->rq_task.
Signed-off-by: Trond Myklebust <trond.myklebust@hammerspace.com>
When storing a struct rpc_rqst on the slot allocation list, we currently
use the same field 'rq_list' as we use to store the request on the
receive queue. Since the structure is never on both lists at the same
time, this is OK.
However, for clarity, let's make that a union with different names for
the different lists so that we can more easily distinguish between
the two states.
Signed-off-by: Trond Myklebust <trond.myklebust@hammerspace.com>
Allow the caller in clnt.c to call into the code to wait for a reply
after calling xprt_transmit(). Again, the reason is that the backchannel
code does not need this functionality.
Signed-off-by: Trond Myklebust <trond.myklebust@hammerspace.com>
Separate out the action of adding a request to the reply queue so that the
backchannel code can simply skip calling it altogether.
Signed-off-by: Trond Myklebust <trond.myklebust@hammerspace.com>
Rather than waking up the entire queue of RPC messages a second time,
just wake up the task that was put to sleep.
Signed-off-by: Trond Myklebust <trond.myklebust@hammerspace.com>
If a message has been encoded using RPCSEC_GSS, the server is
maintaining a window of sequence numbers that it considers valid.
The client should normally be tracking that window, and needs to
verify that the sequence number used by the message being transmitted
still lies inside the window of validity.
So far, we've been able to assume this condition would be realised
automatically, since the client has been encoding the message only
after taking the socket lock. Once we change that condition, we
will need the explicit check.
Signed-off-by: Trond Myklebust <trond.myklebust@hammerspace.com>
Remove trailing whitespace and blank line at EOF
Signed-off-by: Stephen Hemminger <stephen@networkplumber.org>
Signed-off-by: Anna Schumaker <Anna.Schumaker@Netapp.com>
Krzysztof Kozlowski <krzk@kernel.org> reports that a heavy NFSv4
WRITE workload against a slow NFS server causes his Raspberry Pi
clients to stall. Krzysztof bisected it to commit 37ac86c3a7
("SUNRPC: Initialize rpc_rqst outside of xprt->reserve_lock") .
I was able to reproduce similar behavior and it appears that rarely
the RPC client layer is re-allocating an XID for an RPC that it has
already partially sent. This results in the client ignoring the
subsequent reply, which carries the original XID.
For various reasons, checking !req->rq_xmit_bytes_sent in
xprt_prepare_transmit is not a 100% reliable mechanism for
determining when a fresh XID is needed.
Trond's preference is to allocate the XID at the time each rpc_rqst
slot is initialized.
This patch should also address a gcc 4.1.2 complaint reported by
Geert Uytterhoeven <geert@linux-m68k.org>.
Reported-by: Krzysztof Kozlowski <krzk@kernel.org>
Fixes: 37ac86c3a7 ("SUNRPC: Initialize rpc_rqst outside of ... ")
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Tested-by: Krzysztof Kozlowski <krzk@kernel.org>
Signed-off-by: Trond Myklebust <trond.myklebust@hammerspace.com>
Refactor: xprtrdma needs to have better control over when RPCs are
awoken from the backlog queue, so replace xprt_free_slot with a
transport op callout.
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: Anna Schumaker <Anna.Schumaker@Netapp.com>
alloc_slot is a transport-specific op, but initializing an rpc_rqst
is common to all transports. In addition, the only part of initial-
izing an rpc_rqst that needs serialization is getting a fresh XID.
Move rpc_rqst initialization to common code in preparation for
adding a transport-specific alloc_slot to xprtrdma.
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: Anna Schumaker <Anna.Schumaker@Netapp.com>
If recording xprt->stat.max_slots is moved into xprt_alloc_slot,
then xprt->num_reqs is never manipulated outside
xprt->reserve_lock. There's no longer a need for xprt->num_reqs to
be atomic.
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: Anna Schumaker <Anna.Schumaker@Netapp.com>
Some RPC transports have more overhead in their send_request
callouts than others. For example, for RPC-over-RDMA:
- Marshaling an RPC often has to DMA map the RPC arguments
- Registration methods perform memory registration as part of
marshaling
To capture just server and network latencies more precisely: when
sending a Call, capture the rq_xtime timestamp _after_ the transport
header has been marshaled.
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: Anna Schumaker <Anna.Schumaker@Netapp.com>
Some RPC transports have more overhead in their reply handlers
than others. For example, for RPC-over-RDMA:
- RPC completion has to wait for memory invalidation, which is
not a part of the server/network round trip
- Recently a context switch was introduced into the reply handler,
which further artificially inflates the measure of RPC RTT
To capture just server and network latencies more precisely: when
receiving a reply, compute the RTT as soon as the XID is recognized
rather than at RPC completion time.
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: Anna Schumaker <Anna.Schumaker@Netapp.com>
