Send an RDMA-CM private message on connect, and look for one during
a connection-established event.

Both sides can communicate their various implementation limits.
Implementations that don't support this sideband protocol ignore it.

Once the client knows the server's inline threshold maxima, it can
adjust the use of Reply chunks, and eliminate most use of Position
Zero Read chunks. Moderately-sized I/O can be done using a pure
inline RDMA Send instead of RDMA operations that require memory
registration.
Signed-off-by: NChuck Lever <chuck.lever@oracle.com>
Signed-off-by: NAnna Schumaker <Anna.Schumaker@Netapp.com>

Instead of placing registered MWs sparsely into the rl_segments
array, place these MWs on a per-req list.

ro_unmap_{sync,safe} can then simply pull those MWs off the list
instead of walking through the array.

This change significantly reduces the size of struct rpcrdma_req
by removing nsegs and rl_mw from every array element.

As an additional clean-up, chunk co-ordinates are returned in the
"*mw" output argument so they are no longer needed in every
array element.
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>

Instead of leaving orphaned MRs to be released when the transport
is destroyed, release them immediately. The MR free list can now be
replenished if it becomes exhausted.
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>

Frequent MR list exhaustion can impact I/O throughput, so enough MRs
are always created during transport set-up to prevent running out.
This means more MRs are created than most workloads need.

Commit 94f58c58 ("xprtrdma: Allow Read list and Reply chunk
simultaneously") introduced support for sending two chunk lists per
RPC, which consumes more MRs per RPC.

Instead of trying to provision more MRs, introduce a mechanism for
allocating MRs on demand. A few MRs are allocated during transport
set-up to kick things off.

This significantly reduces the average number of MRs per transport
while allowing the MR count to grow for workloads or devices that
need more MRs.

FRWR with mlx4 allocated almost 400 MRs per transport before this
patch. Now it starts with 32.
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, based on code audit: Remove the possibility that the
chunk list XDR encoders can return zero, which would be interpreted
as a NULL.
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>

Commit c93c6223 ("xprtrdma: Disconnect on registration failure")
added a disconnect for some RPC marshaling failures. This is needed
only in a handful of cases, but it was triggering for simple stuff
like temporary resource shortages. Try to straighten this out.

Fix up the lower layers so they don't return -ENOMEM or other error
codes that the RPC client's FSM doesn't explicitly recognize.

Also fix up the places in the send_request path that do want a
disconnect. For example, when ib_post_send or ib_post_recv fail,
this is a sign that there is a send or receive queue resource
miscalculation. That should be rare, and is a sign of a software
bug. But xprtrdma can recover: disconnect to reset the transport and
start over.
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: Move device capability detection into memreg-specific
source files.
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>

Based on code audit.
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: Moving these helpers in a separate patch makes later
patches more readable.
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: FMR is about to replace the rpcrdma_map_one code with
scatterlists. Move the scatterlist fields out of the FRWR-specific
union and into the generic part of rpcrdma_mw.

One minor change: -EIO is now returned if FRWR registration fails.
The RPC is terminated immediately, since the problem is likely due
to a software bug, thus retrying likely won't help.
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: The ro_unmap method is no longer used.
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>

There needs to be a safe method of releasing registered memory
resources when an RPC terminates. Safe can mean a number of things:

+ Doesn't have to sleep

+ Doesn't rely on having a QP in RTS

ro_unmap_safe will be that safe method. It can be used in cases
where synchronous memory invalidation can deadlock, or needs to have
an active QP.

The important case is fencing an RPC's memory regions after it is
signaled (^C) and before it exits. If this is not done, there is a
window where the server can write an RPC reply into memory that the
client has released and re-used for some other purpose.

Note that this is a full solution for FRWR, but FMR and physical
still have some gaps where a particularly bad server can wreak
some havoc on the client. These gaps are not made worse by this
patch and are expected to be exceptionally rare and timing-based.
They are noted in documenting comments.
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>

In a subsequent patch, the fr_xprt and fr_worker fields will be
needed by another memory registration mode. Move them into the
generic rpcrdma_mw structure that wraps struct rpcrdma_frmr.
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>

Maintain the order of invalidation and DMA unmapping when doing
a background MR reset.
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>

frwr_op_unmap_sync() is now invoked in a workqueue context, the same
as __frwr_queue_recovery(). There's no need to defer MR reset if
posting LOCAL_INV MRs fails.

This means that even when ib_post_send() fails (which should occur
very rarely) the invalidation and DMA unmapping steps are still done
in the correct order.
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>

Move the the I/O direction field from rpcrdma_mr_seg into the
rpcrdma_frmr.

This makes it possible to DMA-unmap the frwr long after an RPC has
exited and its rpcrdma_mr_seg array has been released and re-used.
This might occur if an RPC times out while waiting for a new
connection to be established.
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: Follow same naming convention as other fields in struct
rpcrdma_frwr.
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>

When deciding whether to send a Call inline, rpcrdma_marshal_req
doesn't take into account header bytes consumed by chunk lists.
This results in Call messages on the wire that are sometimes larger
than the inline threshold.

Likewise, when a Write list or Reply chunk is in play, the server's
reply has to emit an RDMA Send that includes a larger-than-minimal
RPC-over-RDMA header.

The actual size of a Call message cannot be estimated until after
the chunk lists have been registered. Thus the size of each
RPC-over-RDMA header can be estimated only after chunks are
registered; but the decision to register chunks is based on the size
of that header. Chicken, meet egg.

The best a client can do is estimate header size based on the
largest header that might occur, and then ensure that inline content
is always smaller than that.
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>

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>

The SRP initiator allows to set max_sectors to a value that exceeds
the largest amount of data that can be mapped at once with an mlx4
HCA using fast registration and a page size of 4 KB. Hence modify
ib_map_mr_sg() such that it can map partial sg-elements. If an
sg-element has been mapped partially, let the caller know
which fraction has been mapped by adjusting *sg_offset.
Signed-off-by: NBart Van Assche <bart.vanassche@sandisk.com>
Tested-by: NLaurence Oberman <loberman@redhat.com>
Cc: Christoph Hellwig <hch@lst.de>
Cc: Sagi Grimberg <sagi@grimberg.me>
Signed-off-by: NDoug Ledford <dledford@redhat.com>

Signed-off-by: NChristoph Hellwig <hch@lst.de>
Tested-by: NSteve Wise <swise@opengridcomputing.com>
Reviewed-by: NBart Van Assche <bart.vanassche@sandisk.com>
Reviewed-by: NSagi Grimberg <sagi@grimberg.me>
Reviewed-by: NSteve Wise <swise@opengridcomputing.com>
Signed-off-by: NDoug Ledford <dledford@redhat.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>

Clean up: Make code more readable.
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>

If ib_post_send() in ro_unmap_sync() fails, the WRs have not been
posted, no completions will fire, and wait_for_completion() will
wait forever. Skip the wait in that case.

To ensure the MRs are invalid, disconnect.
Signed-off-by: NChuck Lever <chuck.lever@oracle.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>

FRWR's ro_unmap is asynchronous. The new ro_unmap_sync posts
LOCAL_INV Work Requests and waits for them to complete before
returning.

Note also, DMA unmapping is now done _after_ invalidation.
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>

For FRWR FASTREG and LOCAL_INV, move the ib_*_wr structure off
the stack. This allows frwr_op_map and frwr_op_unmap to chain
WRs together without limit to register or invalidate a set of MRs
with a single ib_post_send().

(This will be for chaining LOCAL_INV requests).
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>

Unsignaled send WRs can get flushed as part of normal unmount, so don't
log them as warnings.
Signed-off-by: NSteve Wise <swise@opengridcomputing.com>
Signed-off-by: NAnna Schumaker <Anna.Schumaker@Netapp.com>

Instead of maintaining a fastreg page list, keep an sg table
and convert an array of pages to a sg list. Then call ib_map_mr_sg
and construct ib_reg_wr.
Signed-off-by: NSagi Grimberg <sagig@mellanox.com>
Acked-by: NChristoph Hellwig <hch@lst.de>
Tested-by: NSteve Wise <swise@opengridcomputing.com>
Tested-by: NSelvin Xavier <selvin.xavier@avagotech.com>
Reviewed-by: NChuck Lever <chuck.lever@oracle.com>
Signed-off-by: NDoug Ledford <dledford@redhat.com>

This patch split up struct ib_send_wr so that all non-trivial verbs
use their own structure which embedds struct ib_send_wr.  This dramaticly
shrinks the size of a WR for most common operations:

sizeof(struct ib_send_wr) (old):	96

sizeof(struct ib_send_wr):		48
sizeof(struct ib_rdma_wr):		64
sizeof(struct ib_atomic_wr):		96
sizeof(struct ib_ud_wr):		88
sizeof(struct ib_fast_reg_wr):		88
sizeof(struct ib_bind_mw_wr):		96
sizeof(struct ib_sig_handover_wr):	80

And with Sagi's pending MR rework the fast registration WR will also be
down to a reasonable size:

sizeof(struct ib_fastreg_wr):		64
Signed-off-by: NChristoph Hellwig <hch@lst.de>
Reviewed-by: Bart Van Assche <bart.vanassche@sandisk.com> [srp, srpt]
Reviewed-by: Chuck Lever <chuck.lever@oracle.com> [sunrpc]
Tested-by: NHaggai Eran <haggaie@mellanox.com>
Tested-by: NSagi Grimberg <sagig@mellanox.com>
Tested-by: NSteve Wise <swise@opengridcomputing.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>

Signed-off-by: NSagi Grimberg <sagig@mellanox.com>
Signed-off-by: NDoug Ledford <dledford@redhat.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>

Reduce resource consumption per-transport to make way for increasing
the credit limit and maximum r/wsize. Pre-allocate fewer MRs.
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>

/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>

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>

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>

After a transport disconnect, FRMRs can be left in an undetermined
state. In particular, the MR's rkey is no good.

Currently, FRMRs are fixed up by the transport connect worker, but
that can race with ->ro_unmap if an RPC happens to exit while the
transport connect worker is running.

A better way of dealing with broken FRMRs is to detect them before
they are re-used by ->ro_map. Such FRMRs are either already invalid
or are owned by the sending RPC, and thus no race with ->ro_unmap
is possible.

Introduce a mechanism for handing broken FRMRs to a workqueue to be
reset in a context that is appropriate for allocating resources
(ie. an ib_alloc_fast_reg_mr() API call).

This mechanism is not yet used, but will be in subsequent patches.
Signed-off-by: NChuck Lever <chuck.lever@oracle.com>
Reviewed-by: NSteve Wise <swise@opengridcomputing.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>

The connect worker can replace ri_id, but prevents ri_id->device
from changing during the lifetime of a transport instance. The old
ID is kept around until a new ID is created and the ->device is
confirmed to be the same.

Cache a copy of ri_id->device in rpcrdma_ia and in rpcrdma_rep.
The cached copy can be used safely in code that does not serialize
with the connect worker.

Other code can use it to save an extra address generation (one
pointer dereference instead of two).
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>