rpcrdma_buffer_get acquires an rpcrdma_req and rep for each RPC.
Currently this is done in the call_allocate action, and sometimes it
can fail if there are many outstanding RPCs.
When call_allocate fails, the RPC task is put on the delayq. It is
awoken a few milliseconds later, but there's no guarantee it will
get a buffer at that time. The RPC task can be repeatedly put back
to sleep or even starved.
The call_allocate action should rarely fail. The delayq mechanism is
not meant to deal with transport congestion.
In the current sunrpc stack, there is a friendlier way to deal with
this situation. These objects are actually tantamount to an RPC
slot (rpc_rqst) and there is a separate FSM action, distinct from
call_allocate, for allocating slot resources. This is the
call_reserve action.
When allocation fails during this action, the RPC is placed on the
transport's backlog queue. The backlog mechanism provides a stronger
guarantee that when the RPC is awoken, a buffer will be available
for it; and backlogged RPCs are awoken one-at-a-time.
To make slot resource allocation occur in the call_reserve action,
create special ->alloc_slot and ->free_slot call-outs for xprtrdma.
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: Anna Schumaker <Anna.Schumaker@Netapp.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>
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>
Refactor: Both rpcrdma_create_req call sites have to allocate the
buffer where the transport header is built, so just move that
allocation into rpcrdma_create_req.
This buffer is a fixed size. There's no needed information available
in call_allocate that is not also available when the transport is
created.
The original purpose for allocating these buffers on demand was to
reduce the possibility that an allocation failure during transport
creation will hork the mount operation during low memory scenarios.
Some relief for this rare possibility is coming up in the next few
patches.
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: Anna Schumaker <Anna.Schumaker@Netapp.com>
RPC-over-RDMA version 1 credit accounting relies on there being a
response message for every RPC Call. This means that RPC procedures
that have no reply will disrupt credit accounting, just in the same
way as a retransmit would (since it is sent because no reply has
arrived). Deal with the "no reply" case the same way.
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: Anna Schumaker <Anna.Schumaker@Netapp.com>
Currently, when the MR free list is exhausted during marshaling, the
RPC/RDMA transport places the RPC task on the delayq, which forces a
wait for HZ >> 2 before the marshal and send is retried.
With this change, the transport now places such an RPC task on the
pending queue, and wakes it just as soon as more MRs have been
created. Creating more MRs typically takes less than a millisecond,
and this waking mechanism is less deadlock-prone.
Moreover, the waiting RPC task is holding the transport's write
lock, which blocks the transport from sending RPCs. Therefore faster
recovery from MR exhaustion is desirable.
This is the same mechanism that the TCP transport utilizes when
handling write buffer space exhaustion.
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: Anna Schumaker <Anna.Schumaker@Netapp.com>
Clean up: The generic rq_connect_cookie is sufficient to detect RPC
Call retransmission.
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: Anna Schumaker <Anna.Schumaker@Netapp.com>
With v4.15, on one of my NFS/RDMA clients I measured a nearly
doubling in the latency of small read and write system calls. There
was no change in server round trip time. The extra latency appears
in the whole RPC execution path.
"git bisect" settled on commit ccede75985 ("xprtrdma: Spread reply
processing over more CPUs") .
After some experimentation, I found that leaving the WQ bound and
allowing the scheduler to pick the dispatch CPU seems to eliminate
the long latencies, and it does not introduce any new regressions.
The fix is implemented by reverting only the part of
commit ccede75985 ("xprtrdma: Spread reply processing over more
CPUs") that dispatches RPC replies specifically on the CPU where the
matching RPC call was made.
Interestingly, saving the CPU number and later queuing reply
processing there was effective _only_ for a NFS READ and WRITE
request. On my NUMA client, in-kernel RPC reply processing for
asynchronous RPCs was dispatched on the same CPU where the RPC call
was made, as expected. However synchronous RPCs seem to get their
reply dispatched on some other CPU than where the call was placed,
every time.
Fixes: ccede75985 ("xprtrdma: Spread reply processing over ... ")
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Cc: stable@vger.kernel.org # v4.15+
Signed-off-by: Anna Schumaker <Anna.Schumaker@Netapp.com>
Track RPC timeouts: report the XID and the server address to match
the content of network capture.
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: Anna Schumaker <Anna.Schumaker@Netapp.com>
Clean up: Over time, the industry has adopted the term "frwr"
instead of "frmr". The term "frwr" is now more widely recognized.
For the past couple of years I've attempted to add new code using
"frwr" , but there still remains plenty of older code that still
uses "frmr". Replace all usage of "frmr" to avoid confusion.
While we're churning code, rename variables unhelpfully called "f"
to "frwr", to improve code clarity.
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: Anna Schumaker <Anna.Schumaker@Netapp.com>
Clean up. @rqst is set up differently for backchannel Replies. For
example, rqst->rq_task and task->tk_client are both NULL. So it is
easier to understand and maintain this code path if it is separated.
Also, we can get rid of the confusing rl_connect_cookie hack in
rpcrdma_bc_receive_call.
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: Anna Schumaker <Anna.Schumaker@Netapp.com>
Since commit 5a6d1db455 ("SUNRPC: Add a transport-specific private
field in rpc_rqst"), the rpc_rqst's for RPC-over-RDMA backchannel
operations leave rq_buffer set to NULL.
xprt_release does not invoke ->op->buf_free when rq_buffer is NULL.
The RPCRDMA_REQ_F_BACKCHANNEL check in xprt_rdma_free is therefore
redundant because xprt_rdma_free is not invoked for backchannel
requests.
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: Anna Schumaker <Anna.Schumaker@Netapp.com>
Clean up. This logic is related to marshaling the request, and I'd
like to keep everything that touches req->rl_registered close
together, for CPU cache efficiency.
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: Anna Schumaker <Anna.Schumaker@Netapp.com>
Clean up a harmless oversight. xprtrdma's ->set_port method has
never properly supported IPv6.
This issue has never been a problem because NFS/RDMA mounts have
always required "port=20049", thus so far, rpcbind is not invoked
for these mounts.
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: Anna Schumaker <Anna.Schumaker@Netapp.com>
Save more space in struct rpcrdma_xprt by removing the redundant
"addr" field from struct rpcrdma_create_data_internal. Wherever
we have rpcrdma_xprt, we also have the rpc_xprt, which has a
sockaddr_storage field with the same content.
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: Anna Schumaker <Anna.Schumaker@Netapp.com>
This makes the address strings available for debugging messages in
earlier stages of transport set up.
The first benefit is to get rid of the single-use rep_remote_addr
field, saving 128+ bytes in struct rpcrdma_ep.
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: Anna Schumaker <Anna.Schumaker@Netapp.com>
Clean up. Remove fields that should have been removed by
commit b3221d6a53 ("xprtrdma: Remove logic that constructs
RDMA_MSGP type calls").
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: Anna Schumaker <Anna.Schumaker@Netapp.com>
The rpcrdma_req is not shared yet, and its associated Send hasn't
been posted, thus RMW should be safe. There's no need for the
expense of a lock cycle here.
Fixes: 0ba6f37012 ("xprtrdma: Refactor rpcrdma_deferred_completion")
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: Anna Schumaker <Anna.Schumaker@Netapp.com>
This leak has been around forever, and is exceptionally rare.
EINVAL causes mount to fail with "an incorrect mount option was
specified" although it's not likely that one of the mount
options is incorrect. Instead, return ENODEV in this case, as this
appears to be an issue with system or device configuration rather
than a specific mount option.
Some obsolete comments are also removed.
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: Anna Schumaker <Anna.Schumaker@Netapp.com>
Commit d8f532d20e ("xprtrdma: Invoke rpcrdma_reply_handler
directly from RECV completion") introduced a performance regression
for NFS I/O small enough to not need memory registration. In multi-
threaded benchmarks that generate primarily small I/O requests,
IOPS throughput is reduced by nearly a third. This patch restores
the previous level of throughput.
Because workqueues are typically BOUND (in particular ib_comp_wq,
nfsiod_workqueue, and rpciod_workqueue), NFS/RDMA workloads tend
to aggregate on the CPU that is handling Receive completions.
The usual approach to addressing this problem is to create a QP
and CQ for each CPU, and then schedule transactions on the QP
for the CPU where you want the transaction to complete. The
transaction then does not require an extra context switch during
completion to end up on the same CPU where the transaction was
started.
This approach doesn't work for the Linux NFS/RDMA client because
currently the Linux NFS client does not support multiple connections
per client-server pair, and the RDMA core API does not make it
straightforward for ULPs to determine which CPU is responsible for
handling Receive completions for a CQ.
So for the moment, record the CPU number in the rpcrdma_req before
the transport sends each RPC Call. Then during Receive completion,
queue the RPC completion on that same CPU.
Additionally, move all RPC completion processing to the deferred
handler so that even RPCs with simple small replies complete on
the CPU that sent the corresponding RPC Call.
Fixes: d8f532d20e ("xprtrdma: Invoke rpcrdma_reply_handler ...")
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: Anna Schumaker <Anna.Schumaker@Netapp.com>
Credit work contributed by Oracle engineers since 2014.
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: Anna Schumaker <Anna.Schumaker@Netapp.com>
When an RPC Call includes a file data payload, that payload can come
from pages in the page cache, or a user buffer (for direct I/O).
If the payload can fit inline, xprtrdma includes it in the Send
using a scatter-gather technique. xprtrdma mustn't allow the RPC
consumer to re-use the memory where that payload resides before the
Send completes. Otherwise, the new contents of that memory would be
exposed by an HCA retransmit of the Send operation.
So, block RPC completion on Send completion, but only in the case
where a separate file data payload is part of the Send. This
prevents the reuse of that memory while it is still part of a Send
operation without an undue cost to other cases.
Waiting is avoided in the common case because typically the Send
will have completed long before the RPC Reply arrives.
These days, an RPC timeout will trigger a disconnect, which tears
down the QP. The disconnect flushes all waiting Sends. This bounds
the amount of time the reply handler has to wait for a Send
completion.
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: Anna Schumaker <Anna.Schumaker@Netapp.com>
Invoke a common routine for releasing hardware resources (for
example, invalidating MRs). This needs to be done whether an
RPC Reply has arrived or the RPC was terminated early.
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: Anna Schumaker <Anna.Schumaker@Netapp.com>
We have one boolean flag in rpcrdma_req today. I'd like to add more
flags, so convert that boolean to a bit flag.
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: Anna Schumaker <Anna.Schumaker@Netapp.com>
Problem statement:
Recently Sagi Grimberg <sagi@grimberg.me> observed that kernel RDMA-
enabled storage initiators don't handle delayed Send completion
correctly. If Send completion is delayed beyond the end of a ULP
transaction, the ULP may release resources that are still being used
by the HCA to complete a long-running Send operation.
This is a common design trait amongst our initiators. Most Send
operations are faster than the ULP transaction they are part of.
Waiting for a completion for these is typically unnecessary.
Infrequently, a network partition or some other problem crops up
where an ordering problem can occur. In NFS parlance, the RPC Reply
arrives and completes the RPC, but the HCA is still retrying the
Send WR that conveyed the RPC Call. In this case, the HCA can try
to use memory that has been invalidated or DMA unmapped, and the
connection is lost. If that memory has been re-used for something
else (possibly not related to NFS), and the Send retransmission
exposes that data on the wire.
Thus we cannot assume that it is safe to release Send-related
resources just because a ULP reply has arrived.
After some analysis, we have determined that the completion
housekeeping will not be difficult for xprtrdma:
- Inline Send buffers are registered via the local DMA key, and
are already left DMA mapped for the lifetime of a transport
connection, thus no additional handling is necessary for those
- Gathered Sends involving page cache pages _will_ need to
DMA unmap those pages after the Send completes. But like
inline send buffers, they are registered via the local DMA key,
and thus will not need to be invalidated
In addition, RPC completion will need to wait for Send completion
in the latter case. However, nearly always, the Send that conveys
the RPC Call will have completed long before the RPC Reply
arrives, and thus no additional latency will be accrued.
Design notes:
In this patch, the rpcrdma_sendctx object is introduced, and a
lock-free circular queue is added to manage a set of them per
transport.
The RPC client's send path already prevents sending more than one
RPC Call at the same time. This allows us to treat the consumer
side of the queue (rpcrdma_sendctx_get_locked) as if there is a
single consumer thread.
The producer side of the queue (rpcrdma_sendctx_put_locked) is
invoked only from the Send completion handler, which is a single
thread of execution (soft IRQ).
The only care that needs to be taken is with the tail index, which
is shared between the producer and consumer. Only the producer
updates the tail index. The consumer compares the head with the
tail to ensure that the a sendctx that is in use is never handed
out again (or, expressed more conventionally, the queue is empty).
When the sendctx queue empties completely, there are enough Sends
outstanding that posting more Send operations can result in a Send
Queue overflow. In this case, the ULP is told to wait and try again.
This introduces strong Send Queue accounting to xprtrdma.
As a final touch, Jason Gunthorpe <jgunthorpe@obsidianresearch.com>
suggested a mechanism that does not require signaling every Send.
We signal once every N Sends, and perform SGE unmapping of N Send
operations during that one completion.
Reported-by: Sagi Grimberg <sagi@grimberg.me>
Suggested-by: Jason Gunthorpe <jgunthorpe@obsidianresearch.com>
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: Anna Schumaker <Anna.Schumaker@Netapp.com>
The "safe" version of ro_unmap is used here to avoid waiting
unnecessarily. However:
- It is safe to wait. After all, we have to wait anyway when using
FMR to register memory.
- This case is rare: it occurs only after a reconnect.
By switching this call site to ro_unmap_sync, the final use of
ro_unmap_safe is removed.
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: Anna Schumaker <Anna.Schumaker@Netapp.com>
In current kernels, waiting in xprt_release appears to be safe to
do. I had erroneously believed that for ASYNC RPCs, waiting of any
kind in xprt_release->xprt_rdma_free would result in deadlock. I've
done injection testing and consulted with Trond to confirm that
waiting in the RPC release path is safe.
For the very few times where RPC resources haven't yet been released
earlier by the reply handler, it is safe to wait synchronously in
xprt_rdma_free for invalidation rather than defering it to MR
recovery.
Note: When the QP is error state, posting a LocalInvalidate should
flush and mark the MR as bad. There is no way the remote HCA can
access that MR via a QP in error state, so it is effectively already
inaccessible and thus safe for the Upper Layer to access. The next
time the MR is used it should be recognized and cleaned up properly
by frwr_op_map.
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: Anna Schumaker <Anna.Schumaker@Netapp.com>
Adopt the use of xprt_pin_rqst to eliminate contention between
Call-side users of rb_lock and the use of rb_lock in
rpcrdma_reply_handler.
This replaces the mechanism introduced in 431af645cf ("xprtrdma:
Fix client lock-up after application signal fires").
Use recv_lock to quickly find the completing rqst, pin it, then
drop the lock. At that point invalidation and pull-up of the Reply
XDR can be done. Both are often expensive operations.
Finally, take recv_lock again to signal completion to the RPC
layer. It also protects adjustment of "cwnd".
This greatly reduces the amount of time a lock is held by the
reply handler. Comparing lock_stat results shows a marked decrease
in contention on rb_lock and recv_lock.
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
[trond.myklebust@primarydata.com: Remove call to rpcrdma_buffer_put() from
the "out_norqst:" path in rpcrdma_reply_handler.]
Signed-off-by: Trond Myklebust <trond.myklebust@primarydata.com>
Initialize an xdr_stream at the top of rpcrdma_marshal_req(), and
use it to encode the fixed transport header fields. This xdr_stream
will be used to encode the chunk lists in a subsequent patch.
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: Anna Schumaker <Anna.Schumaker@Netapp.com>
Clean up: The caller already has rpcrdma_xprt, so pass that directly
instead. And provide a documenting comment for this critical
function.
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: Anna Schumaker <Anna.Schumaker@Netapp.com>
After transport instance creation, these function pointers never
change. Mark them as constant to prevent their use as an attack
vector for code injections.
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: Anna Schumaker <Anna.Schumaker@Netapp.com>
After a signal, the RPC client aborts synchronous RPCs running on
behalf of the signaled application.
The server is still executing those RPCs, and will write the results
back into the client's memory when it's done. By the time the server
writes the results, that memory is likely being used for other
purposes. Therefore xprtrdma has to immediately invalidate all
memory regions used by those aborted RPCs to prevent the server's
writes from clobbering that re-used memory.
With FMR memory registration, invalidation takes a relatively long
time. In fact, the invalidation is often still running when the
server tries to write the results into the memory regions that are
being invalidated.
This sets up a race between two processes:
1. After the signal, xprt_rdma_free calls ro_unmap_safe.
2. While ro_unmap_safe is still running, the server replies and
rpcrdma_reply_handler runs, calling ro_unmap_sync.
Both processes invoke ib_unmap_fmr on the same FMR.
The mlx4 driver allows two ib_unmap_fmr calls on the same FMR at
the same time, but HCAs generally don't tolerate this. Sometimes
this can result in a system crash.
If the HCA happens to survive, rpcrdma_reply_handler continues. It
removes the rpc_rqst from rq_list and releases the transport_lock.
This enables xprt_rdma_free to run in another process, and the
rpc_rqst is released while rpcrdma_reply_handler is still waiting
for the ib_unmap_fmr call to finish.
But further down in rpcrdma_reply_handler, the transport_lock is
taken again, and "rqst" is dereferenced. If "rqst" has already been
released, this triggers a general protection fault. Since bottom-
halves are disabled, the system locks up.
Address both issues by reversing the order of the xprt_lookup_rqst
call and the ro_unmap_sync call. Introduce a separate lookup
mechanism for rpcrdma_req's to enable calling ro_unmap_sync before
xprt_lookup_rqst. Now the handler takes the transport_lock once
and holds it for the XID lookup and RPC completion.
BugLink: https://bugzilla.linux-nfs.org/show_bug.cgi?id=305
Fixes: 68791649a7 ('xprtrdma: Invalidate in the RPC reply ... ')
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: Anna Schumaker <Anna.Schumaker@Netapp.com>
The device driver for the underlying physical device associated
with an RPC-over-RDMA transport can be removed while RPC-over-RDMA
transports are still in use (ie, while NFS filesystems are still
mounted and active). The IB core performs a connection event upcall
to request that consumers free all RDMA resources associated with
a transport.
There may be pending RPCs when this occurs. Care must be taken to
release associated resources without leaving references that can
trigger a subsequent crash if a signal or soft timeout occurs. We
rely on the caller of the transport's ->close method to ensure that
the previous RPC task has invoked xprt_release but the transport
remains write-locked.
A DEVICE_REMOVE upcall forces a disconnect then sleeps. When ->close
is invoked, it destroys the transport's H/W resources, then wakes
the upcall, which completes and allows the core driver unload to
continue.
BugLink: https://bugzilla.linux-nfs.org/show_bug.cgi?id=266
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: Anna Schumaker <Anna.Schumaker@Netapp.com>
In order to unload a device driver and reload it, xprtrdma will need
to close a transport's interface adapter, and then call
rpcrdma_ia_open again, possibly finding a different interface
adapter.
Make rpcrdma_ia_open safe to call on the same transport multiple
times.
This is a refactoring change only.
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: Anna Schumaker <Anna.Schumaker@Netapp.com>
Current NFS clients rely on connection loss to determine when to
retransmit. In particular, for protocols like NFSv4, clients no
longer rely on RPC timeouts to drive retransmission: NFSv4 servers
are required to terminate a connection when they need a client to
retransmit pending RPCs.
When a server is no longer reachable, either because it has crashed
or because the network path has broken, the server cannot actively
terminate a connection. Thus NFS clients depend on transport-level
keepalive to determine when a connection must be replaced and
pending RPCs retransmitted.
However, RDMA RC connections do not have a native keepalive
mechanism. If an NFS/RDMA server crashes after a client has sent
RPCs successfully (an RC ACK has been received for all OTW RDMA
requests), there is no way for the client to know the connection is
moribund.
In addition, new RDMA requests are subject to the RPC-over-RDMA
credit limit. If the client has consumed all granted credits with
NFS traffic, it is not allowed to send another RDMA request until
the server replies. Thus it has no way to send a true keepalive when
the workload has already consumed all credits with pending RPCs.
To address this, forcibly disconnect a transport when an RPC times
out. This prevents moribund connections from stopping the
detection of failover or other configuration changes on the server.
Note that even if the connection is still good, retransmitting
any RPC will trigger a disconnect thanks to this logic in
xprt_rdma_send_request:
/* Must suppress retransmit to maintain credits */
if (req->rl_connect_cookie == xprt->connect_cookie)
goto drop_connection;
req->rl_connect_cookie = xprt->connect_cookie;
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: Anna Schumaker <Anna.Schumaker@Netapp.com>
Sriharsha (sriharsha.basavapatna@broadcom.com) reports an occasional
double DMA unmap of an FRWR MR when a connection is lost. I see one
way this can happen.
When a request requires more than one segment or chunk,
rpcrdma_marshal_req loops, invoking ->frwr_op_map for each segment
(MR) in each chunk. Each call posts a FASTREG Work Request to
register one MR.
Now suppose that the transport connection is lost part-way through
marshaling this request. As part of recovering and resetting that
req, rpcrdma_marshal_req invokes ->frwr_op_unmap_safe, which hands
all the req's registered FRWRs to the MR recovery thread.
But note: FRWR registration is asynchronous. So it's possible that
some of these "already registered" FRWRs are fully registered, and
some are still waiting for their FASTREG WR to complete.
When the connection is lost, the "already registered" frmrs are
marked FRMR_IS_VALID, and the "still waiting" WRs flush. Then
frwr_wc_fastreg marks these frmrs FRMR_FLUSHED_FR.
But thanks to ->frwr_op_unmap_safe, the MR recovery thread is doing
an unreg / alloc_mr, a DMA unmap, and marking each of these frwrs
FRMR_IS_INVALID, at the same time frwr_wc_fastreg might be running.
- If the recovery thread runs last, then the frmr is marked
FRMR_IS_INVALID, and life continues.
- If frwr_wc_fastreg runs last, the frmr is marked FRMR_FLUSHED_FR,
but the recovery thread has already DMA unmapped that MR. When
->frwr_op_map later re-uses this frmr, it sees it is not marked
FRMR_IS_INVALID, and tries to recover it before using it, resulting
in a second DMA unmap of the same MR.
The fix is to guarantee in-flight FASTREG WRs have flushed before MR
recovery runs on those FRWRs. Thus we depend on ro_unmap_safe
(called from xprt_rdma_send_request on retransmit, or from
xprt_rdma_free) to clean up old registrations as needed.
Reported-by: Sriharsha Basavapatna <sriharsha.basavapatna@broadcom.com>
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Tested-by: Sriharsha Basavapatna <sriharsha.basavapatna@broadcom.com>
Signed-off-by: Anna Schumaker <Anna.Schumaker@Netapp.com>
Commit d5440e27d3 ("xprtrdma: Enable pad optimization") made the
Linux client omit XDR round-up padding in normal Read and Write
chunks so that the client doesn't have to register and invalidate
3-byte memory regions that contain no real data.
Unfortunately, my cheery 2014 assessment that this optimization "is
supported now by both Linux and Solaris servers" was premature.
We've found bugs in Solaris in this area since commit d5440e27d3
("xprtrdma: Enable pad optimization") was merged (SYMLINK is the
main offender).
So for maximum interoperability, I'm disabling this optimization
again. If a CM private message is exchanged when connecting, the
client recognizes that the server is Linux, and enables the
optimization for that connection.
Until now the Solaris server bugs did not impact common operations,
and were thus largely benign. Soon, less capable devices on Linux
NFS/RDMA clients will make use of Read chunks more often, and these
Solaris bugs will prevent interoperation in more cases.
Fixes: 677eb17e94 ("xprtrdma: Fix XDR tail buffer marshalling")
Cc: stable@vger.kernel.org # v4.9+
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: Anna Schumaker <Anna.Schumaker@Netapp.com>
Micro-optimization: Most of the time, calls to ro_unmap_safe are
expensive no-ops. Call only when there is work to do.
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: Anna Schumaker <Anna.Schumaker@Netapp.com>
The Version One default inline threshold is still 1KB. But allow
testing with thresholds up to 64KB.
This maximum is somewhat arbitrary. There's no fundamental
architectural limit I'm aware of, but it's good to keep the size of
Receive buffers reasonable. Now that Send can use a s/g list, a
Send buffer is only as large as each RPC requires. Receive buffers
are always the size of the inline threshold, however.
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: Anna Schumaker <Anna.Schumaker@Netapp.com>
An RPC Call message that is sent inline but that has a data payload
(ie, one or more items in rq_snd_buf's page list) must be "pulled
up:"
- call_allocate has to reserve enough RPC Call buffer space to
accommodate the data payload
- call_transmit has to memcopy the rq_snd_buf's page list and tail
into its head iovec before it is sent
As the inline threshold is increased beyond its current 1KB default,
however, this means data payloads of more than a few KB are copied
by the host CPU. For example, if the inline threshold is increased
just to 4KB, then NFS WRITE requests up to 4KB would involve a
memcpy of the NFS WRITE's payload data into the RPC Call buffer.
This is an undesirable amount of participation by the host CPU.
The inline threshold may be much larger than 4KB in the future,
after negotiation with a peer server.
Instead of copying the components of rq_snd_buf into its head iovec,
construct a gather list of these components, and send them all in
place. The same approach is already used in the Linux server's
RPC-over-RDMA reply path.
This mechanism also eliminates the need for rpcrdma_tail_pullup,
which is used to manage the XDR pad and trailing inline content when
a Read list is present.
This requires that the pages in rq_snd_buf's page list be DMA-mapped
during marshaling, and unmapped when a data-bearing RPC is
completed. This is slightly less efficient for very small I/O
payloads, but significantly more efficient as data payload size and
inline threshold increase past a kilobyte.
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: Anna Schumaker <Anna.Schumaker@Netapp.com>
Have frwr's ro_unmap_sync recognize an invalidated rkey that appears
as part of a Receive completion. Local invalidation can be skipped
for that rkey.
Use an out-of-band signaling mechanism to indicate to the server
that the client is prepared to receive RDMA Send With Invalidate.
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: Anna Schumaker <Anna.Schumaker@Netapp.com>
Clean up. The "ia" argument is no longer used.
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: Anna Schumaker <Anna.Schumaker@Netapp.com>
The use of DMA_BIDIRECTIONAL is discouraged by DMA-API.txt.
Fortunately, xprtrdma now knows which direction I/O is going as
soon as it allocates each regbuf.
The RPC Call and Reply buffers are no longer the same regbuf. They
can each be labeled correctly now. The RPC Reply buffer is never
part of either a Send or Receive WR, but it can be part of Reply
chunk, which is mapped and registered via ->ro_map . So it is not
DMA mapped when it is allocated (DMA_NONE), to avoid a double-
mapping.
Since Receive buffers are no longer DMA_BIDIRECTIONAL and their
contents are never modified by the host CPU, DMA-API-HOWTO.txt
suggests that a DMA sync before posting each buffer should be
unnecessary. (See my_card_interrupt_handler).
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: Anna Schumaker <Anna.Schumaker@Netapp.com>
Commit 949317464b ("xprtrdma: Limit number of RDMA segments in
RPC-over-RDMA headers") capped the number of chunks that may appear
in RPC-over-RDMA headers. The maximum header size can be estimated
and fixed to avoid allocating buffer space that is never used.
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: Anna Schumaker <Anna.Schumaker@Netapp.com>
RPC-over-RDMA needs to separate its RPC call and reply buffers.
o When an RPC Call is sent, rq_snd_buf is DMA mapped for an RDMA
Send operation using DMA_TO_DEVICE
o If the client expects a large RPC reply, it DMA maps rq_rcv_buf
as part of a Reply chunk using DMA_FROM_DEVICE
The two mappings are for data movement in opposite directions.
DMA-API.txt suggests that if these mappings share a DMA cacheline,
bad things can happen. This could occur in the final bytes of
rq_snd_buf and the first bytes of rq_rcv_buf if the two buffers
happen to share a DMA cacheline.
On x86_64 the cacheline size is typically 8 bytes, and RPC call
messages are usually much smaller than the send buffer, so this
hasn't been a noticeable problem. But the DMA cacheline size can be
larger on other platforms.
Also, often rq_rcv_buf starts most of the way into a page, thus
an additional RDMA segment is needed to map and register the end of
that buffer. Try to avoid that scenario to reduce the cost of
registering and invalidating Reply chunks.
Instead of carrying a single regbuf that covers both rq_snd_buf and
rq_rcv_buf, each struct rpcrdma_req now carries one regbuf for
rq_snd_buf and one regbuf for rq_rcv_buf.
Some incidental changes worth noting:
- To clear out some spaghetti, refactor xprt_rdma_allocate.
- The value stored in rg_size is the same as the value stored in
the iov.length field, so eliminate rg_size
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: Anna Schumaker <Anna.Schumaker@Netapp.com>
Currently there's a hidden and indirect mechanism for finding the
rpcrdma_req that goes with an rpc_rqst. It depends on getting from
the rq_buffer pointer in struct rpc_rqst to the struct
rpcrdma_regbuf that controls that buffer, and then to the struct
rpcrdma_req it goes with.
This was done back in the day to avoid the need to add a per-rqst
pointer or to alter the buf_free API when support for RPC-over-RDMA
was introduced.
I'm about to change the way regbuf's work to support larger inline
thresholds. Now is a good time to replace this indirect mechanism
with something that is more straightforward. I guess this should be
considered a clean up.
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: Anna Schumaker <Anna.Schumaker@Netapp.com>
Chuck Lever