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>

This is because xprt_request_get_cong() is allowing more than one
RPC Call to be transmitted before the first Receive on the new
connection. The first Receive fills the Receive Queue based on the
server's credit grant. Before that Receive, there is only a single
Receive WR posted because the client doesn't know the server's
credit grant.

Solution is to clear rq_cong on all outstanding rpc_rqsts when the
the cwnd is reset. This is because an RPC/RDMA credit is good for
one connection instance only.

Fixes: 75891f50 ("SUNRPC: Support for congestion control ... ")
Signed-off-by: NChuck Lever <chuck.lever@oracle.com>
Signed-off-by: NAnna Schumaker <Anna.Schumaker@Netapp.com>

Eli Dorfman reports that after a series of idle disconnects, an
RPC/RDMA transport becomes unusable (rdma_create_qp returns
-ENOMEM). Problem was tracked down to increasing Send Queue size
after each reconnect.

The rdma_create_qp() API does not promise to leave its @qp_init_attr
parameter unaltered. In fact, some drivers do modify one or more of
its fields. Thus our calls to rdma_create_qp must use a fresh copy
of ib_qp_init_attr each time.

This fix is appropriate for kernels dating back to late 2007, though
it will have to be adapted, as the connect code has changed over the
years.
Reported-by: NEli Dorfman <eli@vastdata.com>
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>

Micro-optimization: In rpcrdma_post_recvs, since commit e340c2d6
("xprtrdma: Reduce the doorbell rate (Receive)"), the common case is
to return without doing anything. Found with perf.
Signed-off-by: NChuck Lever <chuck.lever@oracle.com>
Signed-off-by: NAnna Schumaker <Anna.Schumaker@Netapp.com>

rpcrdma_rep objects are removed from their free list by only a
single thread: the Receive completion handler. Thus that free list
can be converted to an llist, where a single-threaded consumer and
a multi-threaded producer (rpcrdma_buffer_put) can both access the
llist without the need for any serialization.

This eliminates spin lock contention between the Receive completion
handler and rpcrdma_buffer_get, and makes the rep consumer wait-
free.
Signed-off-by: NChuck Lever <chuck.lever@oracle.com>
Signed-off-by: NAnna Schumaker <Anna.Schumaker@Netapp.com>

Clean up: Now that the free list is used sparingly, get rid of the
separate spin lock protecting it.
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>

Probably would be good to also pass GFP flags to ib_alloc_mr.
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. There is only one remaining rpcrdma_mr_put call site, and
it can be directly replaced with unmap_and_put because mr->mr_dir is
set to DMA_NONE just before the call.

Now all the call sites do a DMA unmap, and we can just rename
mr_unmap_and_put to mr_put, which nicely matches mr_get.
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: There are other "all" list heads. For code clarity
distinguish this one as for use only for MRs by renaming it.
Signed-off-by: NChuck Lever <chuck.lever@oracle.com>
Signed-off-by: NAnna Schumaker <Anna.Schumaker@Netapp.com>

Make the field name the same for all trace points that handle
pointers to struct rpcrdma_rep. That makes it easy to grep for
matching rep points in trace output.
Signed-off-by: NChuck Lever <chuck.lever@oracle.com>
Signed-off-by: NAnna Schumaker <Anna.Schumaker@Netapp.com>

Although I haven't seen any performance results that justify it,
I've received several complaints that NFS/RDMA no longer supports
a maximum rsize and wsize of 1MB. These days it is somewhat smaller.

To simplify the logic that determines whether a chunk list is
necessary, the implementation uses a fixed maximum size of the
transport header. Currently that maximum size is 256 bytes, one
quarter of the default inline threshold size for RPC/RDMA v1.

Since commit a7886849 ("xprtrdma: Reduce max_frwr_depth"), the
size of chunks is also smaller to take advantage of inline page
lists in device internal MR data structures.

The combination of these two design choices has reduced the maximum
NFS rsize and wsize that can be used for most RNIC/HCAs. Increasing
the maximum transport header size and the maximum number of RDMA
segments it can contain increases the negotiated maximum rsize/wsize
on common RNIC/HCAs.
Signed-off-by: NChuck Lever <chuck.lever@oracle.com>
Signed-off-by: NAnna Schumaker <Anna.Schumaker@Netapp.com>

Send and Receive completion is handled on a single CPU selected at
the time each Completion Queue is allocated. Typically this is when
an initiator instantiates an RDMA transport, or when a target
accepts an RDMA connection.

Some ULPs cannot open a connection per CPU to spread completion
workload across available CPUs and MSI vectors. For such ULPs,
provide an API that allows the RDMA core to select a completion
vector based on the device's complement of available comp_vecs.

ULPs that invoke ib_alloc_cq() with only comp_vector 0 are converted
to use the new API so that their completion workloads interfere less
with each other.
Suggested-by: NHåkon Bugge <haakon.bugge@oracle.com>
Signed-off-by: NChuck Lever <chuck.lever@oracle.com>
Reviewed-by: NLeon Romanovsky <leonro@mellanox.com>
Cc: <linux-cifs@vger.kernel.org>
Cc: <v9fs-developer@lists.sourceforge.net>
Link: https://lore.kernel.org/r/20190729171923.13428.52555.stgit@manet.1015granger.netSigned-off-by: NDoug Ledford <dledford@redhat.com>

Clean up.

There is only one remaining function, rpcrdma_buffer_put(), that
uses this field. Its caller can supply a pointer to the correct
rpcrdma_buffer, enabling the removal of an 8-byte pointer field
from a frequently-allocated shared data structure.
Signed-off-by: NChuck Lever <chuck.lever@oracle.com>
Signed-off-by: NAnna Schumaker <Anna.Schumaker@Netapp.com>

rb_lock is contended between rpcrdma_buffer_create,
rpcrdma_buffer_put, and rpcrdma_post_recvs.

Commit e340c2d6 ("xprtrdma: Reduce the doorbell rate (Receive)")
causes rpcrdma_post_recvs to take the rb_lock repeatedly when it
determines more Receives are needed. Streamline this code path so
it takes the lock just once in most cases to build the Receive
chain that is about to be posted.
Signed-off-by: NChuck Lever <chuck.lever@oracle.com>
Signed-off-by: NAnna Schumaker <Anna.Schumaker@Netapp.com>

Clean up.

Commit 7c8d9e7c ("xprtrdma: Move Receive posting to Receive
handler") reduced the number of rpcrdma_rep_create call sites to
one. After that commit, the backchannel code no longer invokes it.

Therefore the free list logic added by commit d698c4a0
("xprtrdma: Fix backchannel allocation of extra rpcrdma_reps") is
no longer necessary, and in fact adds some extra overhead that we
can do without.

Simply post any newly created reps. They will get added back to
the rb_recv_bufs list when they subsequently complete.
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>

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>

Dereference wr->next /before/ the memory backing wr has been
released. This issue was found by code inspection. It is not
expected to be a significant problem because it is in an error
path that is almost never executed.

Fixes: 7c8d9e7c ("xprtrdma: Move Receive posting to ... ")
Signed-off-by: NChuck Lever <chuck.lever@oracle.com>
Signed-off-by: NAnna Schumaker <Anna.Schumaker@Netapp.com>

One of the more common cases of allocation size calculations is finding
the size of a structure that has a zero-sized array at the end, along
with memory for some number of elements for that array. For example:

struct foo {
    int stuff;
    struct boo entry[];
};

instance = kzalloc(sizeof(struct foo) + count * sizeof(struct boo), GFP_KERNEL);

Instead of leaving these open-coded and prone to type mistakes, we can
now use the new struct_size() helper:

instance = kzalloc(struct_size(instance, entry, count), GFP_KERNEL);

This code was detected with the help of Coccinelle.
Signed-off-by: NGustavo A. R. Silva <gustavo@embeddedor.com>
Reviewed-by: NChuck Lever <chuck.lever@oracle.com>
Signed-off-by: NAnna Schumaker <Anna.Schumaker@Netapp.com>

Commit e1ede312 ("xprtrdma: Fix helper that drains the
transport") replaced the ib_drain_qp() call, so update documenting
comments to reflect current operation.
Signed-off-by: NChuck Lever <chuck.lever@oracle.com>
Signed-off-by: NAnna Schumaker <Anna.Schumaker@Netapp.com>

Clean up: rely on the trace points instead.
Signed-off-by: NChuck Lever <chuck.lever@oracle.com>
Signed-off-by: NAnna Schumaker <Anna.Schumaker@Netapp.com>

Clean up.

Move the remaining field in rpcrdma_create_data_internal so the
structure can be removed.
Signed-off-by: NChuck Lever <chuck.lever@oracle.com>
Signed-off-by: NAnna Schumaker <Anna.Schumaker@Netapp.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>

Clean up.

Since commit 54cbd6b0 ("xprtrdma: Delay DMA mapping Send and
Receive buffers"), a pointer to the device is now saved in each
regbuf when it is DMA mapped.
Signed-off-by: NChuck Lever <chuck.lever@oracle.com>
Signed-off-by: NAnna Schumaker <Anna.Schumaker@Netapp.com>

Instead of using a fixed number, allow the amount of Send completion
batching to vary based on the client's maximum credit limit.

- A larger default gives a small boost to IOPS throughput

- Reducing it based on max_requests gives a safe result when the
  max credit limit is cranked down (eg. when the device has a small
  max_qp_wr).
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>

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>

Clean up by providing an API to do this common task.

At this point, the difference between rpcrdma_get_sendbuf and
rpcrdma_get_recvbuf has become tiny. These can be collapsed into a
single helper.
Signed-off-by: NChuck Lever <chuck.lever@oracle.com>
Signed-off-by: NAnna Schumaker <Anna.Schumaker@Netapp.com>

Allocating an rpcrdma_req's regbufs at xprt create time enables
a pair of micro-optimizations:

First, if these regbufs are always there, we can eliminate two
conditional branches from the hot xprt_rdma_allocate path.

Second, by allocating a 1KB buffer, it places a lower bound on the
size of these buffers, without adding yet another conditional
branch. The lower bound reduces the number of hardway re-
allocations. In fact, for some workloads it completely eliminates
hardway allocations.
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>

For code legibility, clean up the function names to be consistent
with the pattern: "rpcrdma" _ object-type _ action
Signed-off-by: NChuck Lever <chuck.lever@oracle.com>
Signed-off-by: NAnna Schumaker <Anna.Schumaker@Netapp.com>

Eventually, I'd like to invoke rpcrdma_create_req() during the
call_reserve step. Memory allocation there probably needs to use
GFP_NOIO. Therefore a set of GFP flags needs to be passed in.

As an additional clean up, just return a pointer or NULL, because
the only error return code here is -ENOMEM.

Lastly, clean up the function names to be consistent with the
pattern: "rpcrdma" _ object-type _ action
Signed-off-by: NChuck Lever <chuck.lever@oracle.com>
Signed-off-by: NAnna Schumaker <Anna.Schumaker@Netapp.com>

We want to drain only the RQ first. Otherwise the transport can
deadlock on ->close if there are outstanding Send completions.

Fixes: 6d2d0ee2 ("xprtrdma: Replace rpcrdma_receive_wq ... ")
Signed-off-by: NChuck Lever <chuck.lever@oracle.com>
Cc: stable@vger.kernel.org # v5.0+
Signed-off-by: NTrond Myklebust <trond.myklebust@hammerspace.com>

Post RECV WRs in batches to reduce the hardware doorbell rate per
transport. This helps the RPC-over-RDMA client scale better in
number of transports.
Signed-off-by: NChuck Lever <chuck.lever@oracle.com>
Signed-off-by: NAnna Schumaker <Anna.Schumaker@Netapp.com>