To support device hotplug and migrating a connection between devices
of different capabilities, we have to guarantee that all in-kernel
devices can support the same max NFS payload size (1 megabyte).

This means that possibly one or two in-tree devices are no longer
supported for NFS/RDMA because they cannot support 1MB rsize/wsize.
The only one I confirmed was cxgb3, but it has already been removed
from the kernel.
Signed-off-by: NChuck Lever <chuck.lever@oracle.com>
Signed-off-by: NAnna Schumaker <Anna.Schumaker@Netapp.com>

Clean up: there is no need to keep two copies of the same value.
Also, in subsequent patches, rpcrdma_ep_create() will be called in
the connect worker rather than at set-up time.

Minor fix: Initialize the transport's sendctx to the value based on
the capabilities of the underlying device, not the maximum setting.
Signed-off-by: NChuck Lever <chuck.lever@oracle.com>
Signed-off-by: NAnna Schumaker <Anna.Schumaker@Netapp.com>

Clean-up. The max_send_sge value also happens to be stored in
ep->rep_attr. Let's keep just a single copy.
Signed-off-by: NChuck Lever <chuck.lever@oracle.com>
Signed-off-by: NAnna Schumaker <Anna.Schumaker@Netapp.com>

Clean up: Use a single trace point to record each connection's
negotiated inline thresholds and the computed maximum byte size
of transport headers.
Signed-off-by: NChuck Lever <chuck.lever@oracle.com>
Signed-off-by: NAnna Schumaker <Anna.Schumaker@Netapp.com>

Slightly reduce overhead and display more useful information.
Signed-off-by: NChuck Lever <chuck.lever@oracle.com>
Signed-off-by: NAnna Schumaker <Anna.Schumaker@Netapp.com>

On some platforms, DMA mapping part of a page is more costly than
copying bytes. Restore the pull-up code and use that when we
think it's going to be faster. The heuristic for now is to pull-up
when the size of the RPC message body fits in the buffer underlying
the head iovec.

Indeed, not involving the I/O MMU can help the RPC/RDMA transport
scale better for tiny I/Os across more RDMA devices. This is because
interaction with the I/O MMU is eliminated, as is handling a Send
completion, for each of these small I/Os. Without the explicit
unmapping, the NIC no longer needs to do a costly internal TLB shoot
down for buffers that are just a handful of bytes.
Signed-off-by: NChuck Lever <chuck.lever@oracle.com>
Signed-off-by: NAnna Schumaker <Anna.Schumaker@Netapp.com>

Refactor: Replace spaghetti with code that makes it plain what needs
to be done for each rtype. This makes it easier to add features and
optimizations.
Signed-off-by: NChuck Lever <chuck.lever@oracle.com>
Signed-off-by: NAnna Schumaker <Anna.Schumaker@Netapp.com>

Clean up: This field is not needed in the Send completion handler,
so it can be moved to struct rpcrdma_req to reduce the size of
struct rpcrdma_sendctx, and to reduce the amount of memory that
is sloshed between the sending process and the Send completion
process.
Signed-off-by: NChuck Lever <chuck.lever@oracle.com>
Signed-off-by: NAnna Schumaker <Anna.Schumaker@Netapp.com>

Micro-optimization: Save eight bytes in a frequently allocated
structure.
Signed-off-by: NChuck Lever <chuck.lever@oracle.com>
Signed-off-by: NAnna Schumaker <Anna.Schumaker@Netapp.com>

MRs are now allocated on demand so we can safely throw them away on
disconnect. This way an idle transport can disconnect and it won't
pin hardware MR resources.

Two additional changes:

- Now that all MRs are destroyed on disconnect, there's no need to
  check during header marshaling if a req has MRs to recycle. Each
  req is sent only once per connection, and now rl_registered is
  guaranteed to be empty when rpcrdma_marshal_req is invoked.

- Because MRs are now destroyed in a WQ_MEM_RECLAIM context, they
  also must be allocated in a WQ_MEM_RECLAIM context. This reduces
  the likelihood that device driver memory allocation will trigger
  memory reclaim during NFS writeback.
Signed-off-by: NChuck Lever <chuck.lever@oracle.com>
Signed-off-by: NAnna Schumaker <Anna.Schumaker@Netapp.com>

A recent clean up attempted to separate Receive handling and RPC
Reply processing, in the name of clean layering.

Unfortunately, we can't do this because the Receive Queue has to be
refilled _after_ the most recent credit update from the responder
is parsed from the transport header, but _before_ we wake up the
next RPC sender. That is right in the middle of
rpcrdma_reply_handler().

Usually this isn't a problem because current responder
implementations don't vary their credit grant. The one exception is
when a connection is established: the grant goes from one to a much
larger number on the first Receive. The requester MUST post enough
Receives right then so that any outstanding requests can be sent
without risking RNR and connection loss.

Fixes: 6ceea368 ("xprtrdma: Refactor Receive accounting")
Signed-off-by: NChuck Lever <chuck.lever@oracle.com>
Signed-off-by: NAnna Schumaker <Anna.Schumaker@Netapp.com>

Clean up/code de-duplication.

Nit: RPC_CWNDSHIFT is incorrect as the initial value for xprt->cwnd.
This mistake does not appear to have operational consequences, since
the cwnd value is replaced with a valid value upon the first Receive
completion.
Signed-off-by: NChuck Lever <chuck.lever@oracle.com>
Signed-off-by: NAnna Schumaker <Anna.Schumaker@Netapp.com>

Ensure that the re-establishment delay does not grow exponentially
on each good reconnect. This probably should have been part of
commit 675dd90a ("xprtrdma: Modernize ops->connect").
Signed-off-by: NChuck Lever <chuck.lever@oracle.com>
Signed-off-by: NAnna Schumaker <Anna.Schumaker@Netapp.com>

The optimization done in "xprtrdma: Simplify rpcrdma_mr_pop" was a
bit too optimistic. MRs left over after a reconnect still need to
be recycled, not added back to the free list, since they could be
in flight or actually fully registered.
Signed-off-by: NChuck Lever <chuck.lever@oracle.com>
Signed-off-by: NAnna Schumaker <Anna.Schumaker@Netapp.com>

Micro-optimization: Save the cost of three function calls during
transport header encoding.

These were "noinline" before to generate more meaningful call stacks
during debugging, but this code is now pretty stable.
Signed-off-by: NChuck Lever <chuck.lever@oracle.com>
Signed-off-by: NAnna Schumaker <Anna.Schumaker@Netapp.com>

Instead of a globally-contended MR free list, cache MRs in each
rpcrdma_req as they are released. This means acquiring and releasing
an MR will be lock-free in the common case, even outside the
transport send lock.

The original idea of per-rpcrdma_req MR free lists was suggested by
Shirley Ma <shirley.ma@oracle.com> several years ago. I just now
figured out how to make that idea work with on-demand MR allocation.
Signed-off-by: NChuck Lever <chuck.lever@oracle.com>
Signed-off-by: NAnna Schumaker <Anna.Schumaker@Netapp.com>

Refactor: Retrieve an MR and handle error recovery entirely in
rpc_rdma.c, as this is not a device-specific function.

Note that since commit 89f90fe1 ("SUNRPC: Allow calls to
xprt_transmit() to drain the entire transmit queue"), the
xprt_transmit function handles the cond_resched. The transport no
longer has to do this itself.
Signed-off-by: NChuck Lever <chuck.lever@oracle.com>
Signed-off-by: NAnna Schumaker <Anna.Schumaker@Netapp.com>

Clean up: rpcrdma_mr_pop call sites check if the list is empty
first. Let's replace the list_empty with less costly logic.
Signed-off-by: NChuck Lever <chuck.lever@oracle.com>
Signed-off-by: NAnna Schumaker <Anna.Schumaker@Netapp.com>

Clean up.

Move the "not present" case into the individual chunk encoders. This
improves code organization and readability.

The reason for the original organization was to optimize for the
case where there there are no chunks. The optimization turned out to
be inconsequential, so let's err on the side of code readability.
Signed-off-by: NChuck Lever <chuck.lever@oracle.com>
Signed-off-by: NAnna Schumaker <Anna.Schumaker@Netapp.com>

Eliminate a context switch in the path that handles RPC wake-ups
when a Receive completion has to wait for a Send completion.
Signed-off-by: NChuck Lever <chuck.lever@oracle.com>
Signed-off-by: NAnna Schumaker <Anna.Schumaker@Netapp.com>

Since commit ba69cd12 ("xprtrdma: Remove support for FMR memory
registration"), FRWR is the only supported memory registration mode.

We can take advantage of the asynchronous nature of FRWR's LOCAL_INV
Work Requests to get rid of the completion wait by having the
LOCAL_INV completion handler take care of DMA unmapping MRs and
waking the upper layer RPC waiter.

This eliminates two context switches when local invalidation is
necessary. As a side benefit, we will no longer need the per-xprt
deferred completion work queue.
Signed-off-by: NChuck Lever <chuck.lever@oracle.com>
Signed-off-by: NAnna Schumaker <Anna.Schumaker@Netapp.com>

When a marshal operation fails, any MRs that were already set up for
that request are recycled. Recycling releases MRs and creates new
ones, which is expensive.

Since commit f2877623 ("xprtrdma: Chain Send to FastReg WRs")
was merged, recycling FRWRs is unnecessary. This is because before
that commit, frwr_map had already posted FAST_REG Work Requests,
so ownership of the MRs had already been passed to the NIC and thus
dealing with them had to be delayed until they completed.

Since that commit, however, FAST_REG WRs are posted at the same time
as the Send WR. This means that if marshaling fails, we are certain
the MRs are safe to simply unmap and place back on the free list
because neither the Send nor the FAST_REG WRs have been posted yet.
The kernel still has ownership of the MRs at this point.

This reduces the total number of MRs that the xprt has to create
under heavy workloads and makes the marshaling logic less brittle.
Signed-off-by: NChuck Lever <chuck.lever@oracle.com>
Signed-off-by: NAnna Schumaker <Anna.Schumaker@Netapp.com>

Now that both the Send and Receive completions are handled in
process context, it is safe to DMA unmap and return MRs to the
free or recycle lists directly in the completion handlers.

Doing this means rpcrdma_frwr no longer needs to track the state of
each MR, meaning that a VALID or FLUSHED MR can no longer appear on
an xprt's MR free list. Thus there is no longer a need to track the
MR's registration state in rpcrdma_frwr.
Signed-off-by: NChuck Lever <chuck.lever@oracle.com>
Signed-off-by: NAnna Schumaker <Anna.Schumaker@Netapp.com>

Commit 9590d083 ("xprtrdma: Use xprt_pin_rqst in
rpcrdma_reply_handler") pins incoming RPC/RDMA replies so they
can be left in the pending requests queue while they are being
processed without introducing a race between ->buf_free and the
transport's reply handler. Therefore RPCRDMA_REQ_F_PENDING is no
longer necessary.
Signed-off-by: NChuck Lever <chuck.lever@oracle.com>
Signed-off-by: NAnna Schumaker <Anna.Schumaker@Netapp.com>

Under high I/O workloads, I've noticed that an RPC/RDMA transport
occasionally deadlocks (IOPS goes to zero, and doesn't recover).
Diagnosis shows that the sendctx queue is empty, but when sendctxs
are returned to the queue, the xprt_write_space wake-up never
occurs. The wake-up logic in rpcrdma_sendctx_put_locked is racy.

I noticed that both EMPTY_SCQ and XPRT_WRITE_SPACE are implemented
via an atomic bit. Just one of those is sufficient. Removing
EMPTY_SCQ in favor of the generic bit mechanism makes the deadlock
un-reproducible.

Without EMPTY_SCQ, rpcrdma_buffer::rb_flags is no longer used and
is therefore removed.

Unfortunately this patch does not apply cleanly to stable. If
needed, someone will have to port it and test it.

Fixes: 2fad6592 ("xprtrdma: Wait on empty sendctx queue")
Signed-off-by: NChuck Lever <chuck.lever@oracle.com>
Signed-off-by: NAnna Schumaker <Anna.Schumaker@Netapp.com>

This is a latent bug. xdr_stream_pos works by subtracting
xdr_stream::nwords from xdr_buf::len. But xdr_stream::nwords is not
initialized by xdr_init_encode().

It works today only because all fields in rpcrdma_req::rl_stream
are initialized to zero by rpcrdma_req_create, making the
subtraction in xdr_stream_pos always a no-op.

I found this issue via code inspection. It was introduced by commit
39f4cd9e ("xprtrdma: Harden chunk list encoding against send
buffer overflow"), but the code has changed enough since then that
this fix can't be automatically applied to stable.
Signed-off-by: NChuck Lever <chuck.lever@oracle.com>
Signed-off-by: NAnna Schumaker <Anna.Schumaker@Netapp.com>

Signed-off-by: NTrond Myklebust <trond.myklebust@hammerspace.com>

Clean up.

The inline settings are actually a characteristic of the endpoint,
and not related to the device. They are also modified after the
transport instance is created, so they do not belong in the cdata
structure either.

Lastly, let's use names that are more natural to RDMA than to NFS:
inline_write -> inline_send and inline_read -> inline_recv. The
/proc files retain their names to avoid breaking user space.
Signed-off-by: NChuck Lever <chuck.lever@oracle.com>
Signed-off-by: NAnna Schumaker <Anna.Schumaker@Netapp.com>

Minor clean-ups I've stumbled on since sendctx was merged last year.
In particular, making Send completion processing more efficient
appears to have a measurable impact on IOPS throughput.

Note: test_and_clear_bit() returns a value, thus an explicit memory
barrier is not necessary.
Signed-off-by: NChuck Lever <chuck.lever@oracle.com>
Signed-off-by: NAnna Schumaker <Anna.Schumaker@Netapp.com>

Record an event when rpcrdma_marshal_req returns a non-zero return
value to help track down why an xprt close might have occurred.
Signed-off-by: NChuck Lever <chuck.lever@oracle.com>
Signed-off-by: NAnna Schumaker <Anna.Schumaker@Netapp.com>

For code legibility, clean up the function names to be consistent
with the pattern: "rpcrdma" _ object-type _ action

Also rpcrdma_regbuf_alloc and rpcrdma_regbuf_free no longer have any
callers outside of verbs.c, and can thus be made static.
Signed-off-by: NChuck Lever <chuck.lever@oracle.com>
Signed-off-by: NAnna Schumaker <Anna.Schumaker@Netapp.com>

Allocate the struct rpcrdma_regbuf separately from the I/O buffer
to better guarantee the alignment of the I/O buffer and eliminate
the wasted space between the rpcrdma_regbuf metadata and the buffer
itself.
Signed-off-by: NChuck Lever <chuck.lever@oracle.com>
Signed-off-by: NAnna Schumaker <Anna.Schumaker@Netapp.com>

Page allocation requests made when the SPARSE_PAGES flag is set are
allowed to fail, and are not critical. No need to spend a rare
resource.
Signed-off-by: NChuck Lever <chuck.lever@oracle.com>
Signed-off-by: NAnna Schumaker <Anna.Schumaker@Netapp.com>

Having access to the controlling rpc_rqst means a trace point in the
XDR code can report:

 - the XID
 - the task ID and client ID
 - the p_name of RPC being processed

Subsequent patches will introduce such trace points.
Signed-off-by: NChuck Lever <chuck.lever@oracle.com>
Signed-off-by: NAnna Schumaker <Anna.Schumaker@Netapp.com>

In very rare cases, an NFS READ operation might predict that the
non-payload part of the RPC Call is large. For instance, an
NFSv4 COMPOUND with a large GETATTR result, in combination with a
large Kerberos credential, could push the non-payload part to be
several kilobytes.

If the non-payload part is larger than the connection's inline
threshold, the client is required to provision a Reply chunk. The
current Linux client does not check for this case. There are two
obvious ways to handle it:

a. Provision a Write chunk for the payload and a Reply chunk for
   the non-payload part

b. Provision a Reply chunk for the whole RPC Reply

Some testing at a recent NFS bake-a-thon showed that servers can
mostly handle a. but there are some corner cases that do not work
yet. b. already works (it has to, to handle krb5i/p), but could be
somewhat less efficient. However, I expect this scenario to be very
rare -- no-one has reported a problem yet.

So I'm going to implement b. Sometime later I will provide some
patches to help make b. a little more efficient by more carefully
choosing the Reply chunk's segment sizes to ensure the payload is
optimally aligned.
Signed-off-by: NChuck Lever <chuck.lever@oracle.com>
Signed-off-by: NAnna Schumaker <Anna.Schumaker@Netapp.com>

If a reply has been processed but the RPC is later retransmitted
anyway, the req->rl_reply field still contains the only pointer to
the old rpcrdma rep. When the next reply comes in, the reply handler
will stomp on the rl_reply field, leaking the old rep.

A trace event is added to capture such leaks.

This problem seems to be worsened by the restructuring of the RPC
Call path in v4.20. Fully addressing this issue will require at
least a re-architecture of the disconnect logic, which is not
appropriate during -rc.
Signed-off-by: NChuck Lever <chuck.lever@oracle.com>
Signed-off-by: NAnna Schumaker <Anna.Schumaker@Netapp.com>

These are rare, but can be helpful at tracking down DMAR and other
problems.
Signed-off-by: NChuck Lever <chuck.lever@oracle.com>
Signed-off-by: NAnna Schumaker <Anna.Schumaker@Netapp.com>

The chunk-related trace points capture nearly the same information
as the MR-related trace points.

Also, rename them so globbing can be used to enable or disable
these trace points more easily.
Signed-off-by: NChuck Lever <chuck.lever@oracle.com>
Signed-off-by: NAnna Schumaker <Anna.Schumaker@Netapp.com>

Clean up: Remove dprintk() call sites that report rare or impossible
errors. Leave a few that display high-value low noise status
information.
Signed-off-by: NChuck Lever <chuck.lever@oracle.com>
Signed-off-by: NAnna Schumaker <Anna.Schumaker@Netapp.com>

For better observability of parsing errors, return the error code
generated in the decoders to the upper layer consumer.
Signed-off-by: NChuck Lever <chuck.lever@oracle.com>
Signed-off-by: NAnna Schumaker <Anna.Schumaker@Netapp.com>