Clean up: ALLPHYSICAL is gone and FMR has been converted to use
scatterlists. There are no more users of these functions.

This patch shrinks the size of struct rpcrdma_req by about 3500
bytes on x86_64. There is one of these structs for each RPC credit
(128 credits per transport connection).
Signed-off-by: NChuck Lever <chuck.lever@oracle.com>
Tested-by: NSteve Wise <swise@opengridcomputing.com>
Signed-off-by: NAnna Schumaker <Anna.Schumaker@Netapp.com>

No HCA or RNIC in the kernel tree requires the use of ALLPHYSICAL.

ALLPHYSICAL advertises in the clear on the network fabric an R_key
that is good for all of the client's memory. No known exploit
exists, but theoretically any user on the server can use that R_key
on the client's QP to read or update any part of the client's memory.

ALLPHYSICAL exposes the client to server bugs, including:
 o base/bounds errors causing data outside the i/o buffer to be
   accessed
 o RDMA access after reply causing data corruption and/or integrity
   fail

ALLPHYSICAL can't protect application memory regions from server
update after a local signal or soft timeout has terminated an RPC.

ALLPHYSICAL chunks are no larger than a page. Special cases to
handle small chunks and long chunk lists have been a source of
implementation complexity and bugs.
Signed-off-by: NChuck Lever <chuck.lever@oracle.com>
Tested-by: NSteve Wise <swise@opengridcomputing.com>
Signed-off-by: NAnna Schumaker <Anna.Schumaker@Netapp.com>

I found that commit ead3f26e ("xprtrdma: Add ro_unmap_safe
memreg method"), which introduces ro_unmap_safe, never wired up the
FMR recovery worker.

The FMR and FRWR recovery work queues both do the same thing.
Instead of setting up separate individual work queues for this,
schedule a delayed worker to deal with them, since recovering MRs is
not performance-critical.

Fixes: ead3f26e ("xprtrdma: Add ro_unmap_safe memreg method")
Signed-off-by: NChuck Lever <chuck.lever@oracle.com>
Tested-by: NSteve Wise <swise@opengridcomputing.com>
Signed-off-by: NAnna Schumaker <Anna.Schumaker@Netapp.com>

Clean up.

After "xprtrdma: Remove ro_unmap() from all registration modes",
there are no longer any sites that take rpcrdma_ia::qplock for read.
The one site that takes it for write is always single-threaded. It
is safe to remove it.
Signed-off-by: NChuck Lever <chuck.lever@oracle.com>
Signed-off-by: NAnna Schumaker <Anna.Schumaker@Netapp.com>

In a cluster failover scenario, it is desirable for the client to
attempt to reconnect quickly, as an alternate NFS server is already
waiting to take over for the down server. The client can't see that
a server IP address has moved to a new server until the existing
connection is gone.

For fabrics and devices where it is meaningful, set a definite upper
bound on the amount of time before it is determined that a
connection is no longer valid. This allows the RPC client to detect
connection loss in a timely matter, then perform a fresh resolution
of the server GUID in case it has changed (cluster failover).
Signed-off-by: NChuck Lever <chuck.lever@oracle.com>
Tested-by: NSteve Wise <swise@opengridcomputing.com>
Signed-off-by: NAnna Schumaker <Anna.Schumaker@Netapp.com>

Clean up: Replace rpcrdma_flush_cqs() and rpcrdma_clean_cqs() with
the new ib_drain_qp() API.
Signed-off-by: NChuck Lever <chuck.lever@oracle.com>
Reviewed-By: NLeon Romanovsky <leonro@mellanox.com>
Tested-by: NSteve Wise <swise@opengridcomputing.com>
Reviewed-by: NSagi Grimberg <sagi@grimberg.me>
Signed-off-by: NAnna Schumaker <Anna.Schumaker@Netapp.com>

Send buffer space is shared between the RPC-over-RDMA header and
an RPC message. A large RPC-over-RDMA header means less space is
available for the associated RPC message, which then has to be
moved via an RDMA Read or Write.

As more segments are added to the chunk lists, the header increases
in size.  Typical modern hardware needs only a few segments to
convey the maximum payload size, but some devices and registration
modes may need a lot of segments to convey data payload. Sometimes
so many are needed that the remaining space in the Send buffer is
not enough for the RPC message. Sending such a message usually
fails.

To ensure a transport can always make forward progress, cap the
number of RDMA segments that are allowed in chunk lists. This
prevents less-capable devices and memory registrations from
consuming a large portion of the Send buffer by reducing the
maximum data payload that can be conveyed with such devices.

For now I choose an arbitrary maximum of 8 RDMA segments. This
allows a maximum size RPC-over-RDMA header to fit nicely in the
current 1024 byte inline threshold with over 700 bytes remaining
for an inline RPC message.

The current maximum data payload of NFS READ or WRITE requests is
one megabyte. To convey that payload on a client with 4KB pages,
each chunk segment would need to handle 32 or more data pages. This
is well within the capabilities of FMR. For physical registration,
the maximum payload size on platforms with 4KB pages is reduced to
32KB.

For FRWR, a device's maximum page list depth would need to be at
least 34 to support the maximum 1MB payload. A device with a smaller
maximum page list depth means the maximum data payload is reduced
when using that device.
Signed-off-by: NChuck Lever <chuck.lever@oracle.com>
Tested-by: NSteve Wise <swise@opengridcomputing.com>
Reviewed-by: NSagi Grimberg <sagi@grimberg.me>
Signed-off-by: NAnna Schumaker <Anna.Schumaker@Netapp.com>

Calling ib_poll_cq() to sort through WCs during a completion is a
common pattern amongst RDMA consumers. Since commit 14d3a3b2
("IB: add a proper completion queue abstraction"), WC sorting can
be handled by the IB core.

By converting to this new API, xprtrdma is made a better neighbor to
other RDMA consumers, as it allows the core to schedule the delivery
of completions more fairly amongst all active consumers.

Because each ib_cqe carries a pointer to a completion method, the
core can now post its own operations on a consumer's QP, and handle
the completions itself, without changes to the consumer.

Send completions were previously handled entirely in the completion
upcall handler (ie, deferring to a process context is unneeded).
Thus IB_POLL_SOFTIRQ is a direct replacement for the current
xprtrdma send code path.
Signed-off-by: NChuck Lever <chuck.lever@oracle.com>
Reviewed-by: NDevesh Sharma <devesh.sharma@broadcom.com>
Reviewed-by: NSagi Grimberg <sagig@mellanox.com>
Signed-off-by: NAnna Schumaker <Anna.Schumaker@Netapp.com>

Calling ib_poll_cq() to sort through WCs during a completion is a
common pattern amongst RDMA consumers. Since commit 14d3a3b2
("IB: add a proper completion queue abstraction"), WC sorting can
be handled by the IB core.

By converting to this new API, xprtrdma is made a better neighbor to
other RDMA consumers, as it allows the core to schedule the delivery
of completions more fairly amongst all active consumers.

Because each ib_cqe carries a pointer to a completion method, the
core can now post its own operations on a consumer's QP, and handle
the completions itself, without changes to the consumer.

xprtrdma's reply processing is already handled in a work queue, but
there is some initial order-dependent processing that is done in the
soft IRQ context before a work item is scheduled.

IB_POLL_SOFTIRQ is a direct replacement for the current xprtrdma
receive code path.
Signed-off-by: NChuck Lever <chuck.lever@oracle.com>
Reviewed-by: NDevesh Sharma <devesh.sharma@broadcom.com>
Reviewed-by: NSagi Grimberg <sagig@mellanox.com>
Signed-off-by: NAnna Schumaker <Anna.Schumaker@Netapp.com>

Commit fe97b47c ("xprtrdma: Use workqueue to process RPC/RDMA
replies") replaced the reply tasklet with a workqueue that allows
RPC replies to be processed in parallel. Thus the credit values in
RPC-over-RDMA replies can be applied in a different order than in
which the server sent them.

To fix this, revert commit eba8ff66 ("xprtrdma: Move credit
update to RPC reply handler"). Reverting is done by hand to
accommodate code changes that have occurred since then.

Fixes: fe97b47c ("xprtrdma: Use workqueue to process . . .")
Signed-off-by: NChuck Lever <chuck.lever@oracle.com>
Reviewed-by: NSagi Grimberg <sagig@mellanox.com>
Signed-off-by: NAnna Schumaker <Anna.Schumaker@Netapp.com>

Instead, use the cached copy of the attributes present on the device.
Signed-off-by: NOr Gerlitz <ogerlitz@mellanox.com>
Signed-off-by: NDoug Ledford <dledford@redhat.com>

The root of the problem was that sends (especially unsignalled
FASTREG and LOCAL_INV Work Requests) were not properly flow-
controlled, which allowed a send queue overrun.

Now that the RPC/RDMA reply handler waits for invalidation to
complete, the send queue is properly flow-controlled. Thus this
limit is no longer necessary.
Signed-off-by: NChuck Lever <chuck.lever@oracle.com>
Tested-by: NDevesh Sharma <devesh.sharma@avagotech.com>
Signed-off-by: NAnna Schumaker <Anna.Schumaker@Netapp.com>

Clean up.

rb_lock critical sections added in rpcrdma_ep_post_extra_recv()
should have first been converted to use normal spin_lock now that
the reply handler is a work queue.

The backchannel set up code should use the appropriate helper
instead of open-coding a rb_recv_bufs list add.

Problem introduced by glib patch re-ordering on my part.

Fixes: f531a5db ('xprtrdma: Pre-allocate backward rpc_rqst')
Signed-off-by: NChuck Lever <chuck.lever@oracle.com>
Tested-by: NDevesh Sharma <devesh.sharma@avagotech.com>
Signed-off-by: NAnna Schumaker <Anna.Schumaker@Netapp.com>

It doesn't matter either way, but the curly braces were clearly intended
here.  It causes a Smatch warning.
Signed-off-by: NDan Carpenter <dan.carpenter@oracle.com>
Signed-off-by: NAnna Schumaker <Anna.Schumaker@Netapp.com>

Pre-allocate extra send and receive Work Requests needed to handle
backchannel receives and sends.

The transport doesn't know how many extra WRs to pre-allocate until
the xprt_setup_backchannel() call, but that's long after the WRs are
allocated during forechannel setup.

So, use a fixed value for now.
Signed-off-by: NChuck Lever <chuck.lever@oracle.com>
Reviewed-by: NSagi Grimberg <sagig@mellanox.com>
Tested-By: NDevesh Sharma <devesh.sharma@avagotech.com>
Signed-off-by: NAnna Schumaker <Anna.Schumaker@Netapp.com>

xprtrdma's backward direction send and receive buffers are the same
size as the forechannel's inline threshold, and must be pre-
registered.

The consumer has no control over which receive buffer the adapter
chooses to catch an incoming backwards-direction call. Any receive
buffer can be used for either a forward reply or a backward call.
Thus both types of RPC message must all be the same size.
Signed-off-by: NChuck Lever <chuck.lever@oracle.com>
Reviewed-by: NSagi Grimberg <sagig@mellanox.com>
Tested-By: NDevesh Sharma <devesh.sharma@avagotech.com>
Signed-off-by: NAnna Schumaker <Anna.Schumaker@Netapp.com>

Now that RPC replies are processed in a workqueue, there's no need
to disable IRQs when managing send and receive buffers. This saves
noticeable overhead per RPC.
Signed-off-by: NChuck Lever <chuck.lever@oracle.com>
Reviewed-by: NSagi Grimberg <sagig@mellanox.com>
Tested-By: NDevesh Sharma <devesh.sharma@avagotech.com>
Signed-off-by: NAnna Schumaker <Anna.Schumaker@Netapp.com>

Clean up: The reply tasklet is no longer used.
Signed-off-by: NChuck Lever <chuck.lever@oracle.com>
Reviewed-by: NSagi Grimberg <sagig@mellanox.com>
Tested-By: NDevesh Sharma <devesh.sharma@avagotech.com>
Signed-off-by: NAnna Schumaker <Anna.Schumaker@Netapp.com>

The reply tasklet is fast, but it's single threaded. After reply
traffic saturates a single CPU, there's no more reply processing
capacity.

Replace the tasklet with a workqueue to spread reply handling across
all CPUs.  This also moves RPC/RDMA reply handling out of the soft
IRQ context and into a context that allows sleeps.
Signed-off-by: NChuck Lever <chuck.lever@oracle.com>
Reviewed-by: NSagi Grimberg <sagig@mellanox.com>
Tested-By: NDevesh Sharma <devesh.sharma@avagotech.com>
Signed-off-by: NAnna Schumaker <Anna.Schumaker@Netapp.com>

The rb_send_bufs and rb_recv_bufs arrays are used to implement a
pair of stacks for keeping track of free rpcrdma_req and rpcrdma_rep
structs. Replace those arrays with free lists.

To allow more than 512 RPCs in-flight at once, each of these arrays
would be larger than a page (assuming 8-byte addresses and 4KB
pages). Allowing up to 64K in-flight RPCs (as TCP now does), each
buffer array would have to be 128 pages. That's an order-6
allocation. (Not that we're going there.)

A list is easier to expand dynamically. Instead of allocating a
larger array of pointers and copying the existing pointers to the
new array, simply append more buffers to each list.

This also makes it simpler to manage receive buffers that might
catch backwards-direction calls, or to post receive buffers in
bulk to amortize the overhead of ib_post_recv.
Signed-off-by: NChuck Lever <chuck.lever@oracle.com>
Reviewed-by: NSagi Grimberg <sagig@mellanox.com>
Reviewed-by: NDevesh Sharma <devesh.sharma@avagotech.com>
Tested-By: NDevesh Sharma <devesh.sharma@avagotech.com>
Signed-off-by: NAnna Schumaker <Anna.Schumaker@Netapp.com>

Clean up: The error cases in rpcrdma_reply_handler() almost never
execute. Ensure the compiler places them out of the hot path.

No behavior change expected.
Signed-off-by: NChuck Lever <chuck.lever@oracle.com>
Reviewed-by: NSagi Grimberg <sagig@mellanox.com>
Reviewed-by: NDevesh Sharma <devesh.sharma@avagotech.com>
Tested-By: NDevesh Sharma <devesh.sharma@avagotech.com>
Signed-off-by: NAnna Schumaker <Anna.Schumaker@Netapp.com>

Commit 8301a2c0 ("xprtrdma: Limit work done by completion
handler") was supposed to prevent xprtrdma's upcall handlers from
starving other softIRQ work by letting them return to the provider
before all CQEs have been polled.

The logic assumes the provider will call the upcall handler again
immediately if the CQ is re-armed while there are still queued CQEs.

This assumption is invalid. The IBTA spec says that after a CQ is
armed, the hardware must interrupt only when a new CQE is inserted.
xprtrdma can't rely on the provider calling again, even though some
providers do.

Therefore, leaving CQEs on queue makes sense only when there is
another mechanism that ensures all remaining CQEs are consumed in a
timely fashion. xprtrdma does not have such a mechanism. If a CQE
remains queued, the transport can wait forever to send the next RPC.

Finally, move the wcs array back onto the stack to ensure that the
poll array is always local to the CPU where the completion upcall is
running.

Fixes: 8301a2c0 ("xprtrdma: Limit work done by completion ...")
Signed-off-by: NChuck Lever <chuck.lever@oracle.com>
Reviewed-by: NSagi Grimberg <sagig@mellanox.com>
Reviewed-by: NDevesh Sharma <devesh.sharma@avagotech.com>
Tested-By: NDevesh Sharma <devesh.sharma@avagotech.com>
Signed-off-by: NAnna Schumaker <Anna.Schumaker@Netapp.com>

ib_req_notify_cq(IB_CQ_REPORT_MISSED_EVENTS) returns a positive
value if WCs were added to a CQ after the last completion upcall
but before the CQ has been re-armed.

Commit 7f23f6f6 ("xprtrmda: Reduce lock contention in
completion handlers") assumed that when ib_req_notify_cq() returned
a positive RC, the CQ had also been successfully re-armed, making
it safe to return control to the provider without losing any
completion signals. That is an invalid assumption.

Change both completion handlers to continue polling while
ib_req_notify_cq() returns a positive value.

Fixes: 7f23f6f6 ("xprtrmda: Reduce lock contention in ...")
Signed-off-by: NChuck Lever <chuck.lever@oracle.com>
Reviewed-by: NSagi Grimberg <sagig@mellanox.com>
Reviewed-by: NDevesh Sharma <devesh.sharma@avagotech.com>
Tested-By: NDevesh Sharma <devesh.sharma@avagotech.com>
Signed-off-by: NAnna Schumaker <Anna.Schumaker@Netapp.com>

Add support for network namespaces in the ib_cma module. This is
accomplished by:

1. Adding network namespace parameter for rdma_create_id. This parameter is
   used to populate the network namespace field in rdma_id_private.
   rdma_create_id keeps a reference on the network namespace.
2. Using the network namespace from the rdma_id instead of init_net inside
   of ib_cma, when listening on an ID and when looking for an ID for an
   incoming request.
3. Decrementing the reference count for the appropriate network namespace
   when calling rdma_destroy_id.

In order to preserve the current behavior init_net is passed when calling
from other modules.
Signed-off-by: NGuy Shapiro <guysh@mellanox.com>
Signed-off-by: NHaggai Eran <haggaie@mellanox.com>
Signed-off-by: NYotam Kenneth <yotamke@mellanox.com>
Signed-off-by: NShachar Raindel <raindel@mellanox.com>
Signed-off-by: NDoug Ledford <dledford@redhat.com>

There is no need to require LOCAL_DMA_LKEY support as the
PD allocation makes sure that there is a local_dma_lkey. Also
correctly set a return value in error path.

This caused a NULL pointer dereference in mlx5 which removed
the support for LOCAL_DMA_LKEY.

Fixes: bb6c96d7 ("xprtrdma: Replace global lkey with lkey local to PD")
Signed-off-by: NSagi Grimberg <sagig@mellanox.com>
Reviewed-by: NChuck Lever <chuck.lever@oracle.com>
Acked-by: NAnna Schumaker <Anna.Schumaker@Netapp.com>
Signed-off-by: NDoug Ledford <dledford@redhat.com>

Otherwise a FRMR completion can cause a touch-after-free crash.

In xprt_rdma_destroy(), call rpcrdma_buffer_destroy() only after calling
rpcrdma_ep_destroy().

In rpcrdma_ep_destroy(), disconnect the cm_id first which should flush the
qp, then drain the cqs, then destroy the qp, and finally destroy the cqs.
Signed-off-by: NSteve Wise <swise@opengridcomputing.com>
Tested-by: NChuck Lever <chuck.lever@oracle.com>
Signed-off-by: NAnna Schumaker <Anna.Schumaker@Netapp.com>

The core API has changed so that devices that do not have a global
DMA lkey automatically create an mr, per-PD, and make that lkey
available. The global DMA lkey interface is going away in favor of
the per-PD DMA lkey.

The per-PD DMA lkey is always available. Convert xprtrdma to use the
device's per-PD DMA lkey for regbufs, no matter which memory
registration scheme is in use.
Signed-off-by: NChuck Lever <chuck.lever@oracle.com>
Signed-off-by: NSagi Grimberg <sagig@mellanox.com>
Cc: linux-nfs <linux-nfs@vger.kernel.org>
Acked-by: NAnna Schumaker <Anna.Schumaker@Netapp.com>
Signed-off-by: NDoug Ledford <dledford@redhat.com>

The majority of callers never check the return value, and even if they
did, they can't do anything about a failure.

All possible failure cases represent a bug in the caller, so just
WARN_ON inside the function instead.

This fixes a few random errors:
 net/rd/iw.c infinite loops while it fails. (racing with EBUSY?)

This also lays the ground work to get rid of error return from the
drivers. Most drivers do not error, the few that do are broken since
it cannot be handled.

Since uverbs can legitimately make use of EBUSY, open code the
check.
Signed-off-by: NJason Gunthorpe <jgunthorpe@obsidianresearch.com>
Reviewed-by: NChuck Lever <chuck.lever@oracle.com>
Signed-off-by: NDoug Ledford <dledford@redhat.com>

This is a rework of the following patch sent almost a year back:
http://www.mail-archive.com/linux-rdma%40vger.kernel.org/msg20730.html

In presence of active mount if someone tries to rmmod vendor-driver, the
command remains stuck forever waiting for destruction of all rdma-cm-id.
in worst case client can crash during shutdown with active mounts.

The existing code assumes that ia->ri_id->device cannot change during
the lifetime of a transport. xprtrdma do not have support for
DEVICE_REMOVAL event either. Lifting that assumption and adding support
for DEVICE_REMOVAL event is a long chain of work, and is in plan.

The community decided that preventing the hang right now is more
important than waiting for architectural changes.

Thus, this patch introduces a temporary workaround to acquire HCA driver
module reference count during the mount of a nfs-rdma mount point.
Signed-off-by: NDevesh Sharma <devesh.sharma@avagotech.com>
Signed-off-by: NChuck Lever <chuck.lever@oracle.com>
Reviewed-by: NSagi Grimberg <sagig@dev.mellanox.co.il>
Signed-off-by: NAnna Schumaker <Anna.Schumaker@Netapp.com>

RDMA_MSGP type calls insert a zero pad in the middle of the RPC
message to align the RPC request's data payload to the server's
alignment preferences. A server can then "page flip" the payload
into place to avoid a data copy in certain circumstances. However:

1. The client has to have a priori knowledge of the server's
   preferred alignment

2. Requests eligible for RDMA_MSGP are requests that are small
   enough to have been sent inline, and convey a data payload
   at the _end_ of the RPC message

Today 1. is done with a sysctl, and is a global setting that is
copied during mount. Linux does not support CCP to query the
server's preferences (RFC 5666, Section 6).

A small-ish NFSv3 WRITE might use RDMA_MSGP, but no NFSv4
compound fits bullet 2.

Thus the Linux client currently leaves RDMA_MSGP disabled. The
Linux server handles RDMA_MSGP, but does not use any special
page flipping, so it confers no benefit.

Clean up the marshaling code by removing the logic that constructs
RDMA_MSGP type calls. This also reduces the maximum send iovec size
from four to just two elements.

/proc/sys/sunrpc/rdma_inline_write_padding is a kernel API, and
thus is left in place.
Signed-off-by: NChuck Lever <chuck.lever@oracle.com>
Tested-by: NDevesh Sharma <devesh.sharma@avagotech.com>
Signed-off-by: NAnna Schumaker <Anna.Schumaker@Netapp.com>

Untangle the end of rpcrdma_ia_open() by moving DMA MR set-up, which
is different for each registration method, to the .ro_open functions.

This is refactoring only. No behavior change is expected.
Signed-off-by: NChuck Lever <chuck.lever@oracle.com>
Tested-by: NDevesh Sharma <devesh.sharma@avagotech.com>
Signed-off-by: NAnna Schumaker <Anna.Schumaker@Netapp.com>

All HCA providers have an ib_get_dma_mr() verb. Thus
rpcrdma_ia_open() will either grab the device's local_dma_key if one
is available, or it will call ib_get_dma_mr(). If ib_get_dma_mr()
fails, rpcrdma_ia_open() fails and no transport is created.

Therefore execution never reaches the ib_reg_phys_mr() call site in
rpcrdma_register_internal(), so it can be removed.

The remaining logic in rpcrdma_{de}register_internal() is folded
into rpcrdma_{alloc,free}_regbuf().

This is clean up only. No behavior change is expected.
Signed-off-by: NChuck Lever <chuck.lever@oracle.com>
Reviewed-by: NDevesh Sharma <devesh.sharma@avagotech.com>
Reviewed-By: NSagi Grimberg <sagig@mellanox.com>
Tested-by: NDevesh Sharma <devesh.sharma@avagotech.com>
Signed-off-by: NAnna Schumaker <Anna.Schumaker@Netapp.com>

PHYSICAL memory registration uses a single rkey for all of the
client's memory, thus is insecure. It is still useful in some cases
for testing.

Retain the ability to select PHYSICAL memory registration capability
via /proc/sys/sunrpc/rdma_memreg_strategy, but don't fall back to it
if the HCA does not support FRWR or FMR.

This means amso1100 no longer works out of the box with NFS/RDMA.
When using amso1100 HCAs, set the memreg_strategy sysctl to 6 before
performing NFS/RDMA mounts.
Signed-off-by: NChuck Lever <chuck.lever@oracle.com>
Tested-by: NDevesh Sharma <devesh.sharma@avagotech.com>
Reviewed-by: NChristoph Hellwig <hch@lst.de>
Signed-off-by: NAnna Schumaker <Anna.Schumaker@Netapp.com>

Currently, ib_create_cq uses cqe and comp_vecotr instead
of the extendible ib_cq_init_attr struct.

Earlier patches already changed the vendors to work with
ib_cq_init_attr. This patch changes the consumers too.
Signed-off-by: NMatan Barak <matanb@mellanox.com>
Signed-off-by: NOr Gerlitz <ogerlitz@mellanox.com>
Signed-off-by: NDoug Ledford <dledford@redhat.com>

/proc/lock_stat showed contention between rpcrdma_buffer_get/put
and the MR allocation functions during I/O intensive workloads.

Now that MRs are no longer allocated in rpcrdma_buffer_get(),
there's no reason the rb_mws list has to be managed using the
same lock as the send/receive buffers. Split that lock. The
new lock does not need to disable interrupts because buffer
get/put is never called in an interrupt context.

struct rpcrdma_buffer is re-arranged to ensure rb_mwlock and rb_mws
are always in a different cacheline than rb_lock and the buffer
pointers.
Signed-off-by: NChuck Lever <chuck.lever@oracle.com>
Reviewed-by: NSteve Wise <swise@opengridcomputing.com>
Reviewed-by: NSagi Grimberg <sagig@mellanox.com>
Tested-By: NDevesh Sharma <devesh.sharma@avagotech.com>
Reviewed-by: NDoug Ledford <dledford@redhat.com>
Signed-off-by: NAnna Schumaker <Anna.Schumaker@Netapp.com>

Clean up: This field is no longer used.
Signed-off-by: NChuck Lever <chuck.lever@oracle.com>
Reviewed-by: NSteve Wise <swise@opengridcomputing.com>
Reviewed-by: NSagi Grimberg <sagig@mellanox.com>
Reviewed-by: NDevesh Sharma <devesh.sharma@avagotech.com>
Tested-By: NDevesh Sharma <devesh.sharma@avagotech.com>
Reviewed-by: NDoug Ledford <dledford@redhat.com>
Signed-off-by: NAnna Schumaker <Anna.Schumaker@Netapp.com>

An RPC can exit at any time. When it does so, xprt_rdma_free() is
called, and it calls ->op_unmap().

If ->ro_reset() is running due to a transport disconnect, the two
methods can race while processing the same rpcrdma_mw. The results
are unpredictable.

Because of this, in previous patches I've altered ->ro_map() to
handle MR reset. ->ro_reset() is no longer needed and can be
removed.
Signed-off-by: NChuck Lever <chuck.lever@oracle.com>
Reviewed-by: NSteve Wise <swise@opengridcomputing.com>
Reviewed-by: NSagi Grimberg <sagig@mellanox.com>
Reviewed-by: NDevesh Sharma <devesh.sharma@avagotech.com>
Tested-By: NDevesh Sharma <devesh.sharma@avagotech.com>
Reviewed-by: NDoug Ledford <dledford@redhat.com>
Signed-off-by: NAnna Schumaker <Anna.Schumaker@Netapp.com>

Clean up: Remove functions no longer used to recover broken FRMRs.
Signed-off-by: NChuck Lever <chuck.lever@oracle.com>
Reviewed-by: NSteve Wise <swise@opengridcomputing.com>
Reviewed-by: NSagi Grimberg <sagig@mellanox.com>
Reviewed-by: NDevesh Sharma <devesh.sharma@avagotech.com>
Tested-By: NDevesh Sharma <devesh.sharma@avagotech.com>
Reviewed-by: NDoug Ledford <dledford@redhat.com>
Signed-off-by: NAnna Schumaker <Anna.Schumaker@Netapp.com>

Acquiring 64 MRs in rpcrdma_buffer_get() while holding the buffer
pool lock is expensive, and unnecessary because most modern adapters
can transfer 100s of KBs of payload using just a single MR.

Instead, acquire MRs one-at-a-time as chunks are registered, and
return them to rb_mws immediately during deregistration.

Note: commit 539431a4 ("xprtrdma: Don't invalidate FRMRs if
registration fails") is reverted: There is now a valid case where
registration can fail (with -ENOMEM) but the QP is still in RTS.
Signed-off-by: NChuck Lever <chuck.lever@oracle.com>
Reviewed-by: NSteve Wise <swise@opengridcomputing.com>
Tested-By: NDevesh Sharma <devesh.sharma@avagotech.com>
Reviewed-by: NDoug Ledford <dledford@redhat.com>
Signed-off-by: NAnna Schumaker <Anna.Schumaker@Netapp.com>

Acquiring 64 FMRs in rpcrdma_buffer_get() while holding the buffer
pool lock is expensive, and unnecessary because FMR mode can
transfer up to a 1MB payload using just a single ib_fmr.

Instead, acquire ib_fmrs one-at-a-time as chunks are registered, and
return them to rb_mws immediately during deregistration.
Signed-off-by: NChuck Lever <chuck.lever@oracle.com>
Reviewed-by: NSteve Wise <swise@opengridcomputing.com>
Tested-By: NDevesh Sharma <devesh.sharma@avagotech.com>
Reviewed-by: NDoug Ledford <dledford@redhat.com>
Signed-off-by: NAnna Schumaker <Anna.Schumaker@Netapp.com>