GitLab

While sending each RPC Reply, svc_rdma_sendto allocates and DMA-
maps a separate buffer where the RPC/RDMA transport header is
constructed. The buffer is unmapped and released in the Send
completion handler. This is significant per-RPC overhead,
especially for small RPCs.

Instead, allocate and DMA-map a buffer, and cache it in each
svc_rdma_send_ctxt. This buffer and its mapping can be re-used
for each RPC, saving the cost of memory allocation and DMA
mapping.
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: J. Bruce Fields <bfields@redhat.com>

Clean up: No current caller of svc_rdma_send's passes in a chained
WR. The logic that counts the chain length can be replaced with a
constant (1).
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: J. Bruce Fields <bfields@redhat.com>

Clean up: Now that the send_wr is part of the svc_rdma_send_ctxt,
svc_rdma_post_send_wr is nearly empty.
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: J. Bruce Fields <bfields@redhat.com>

Receive buffers are always the same size, but each Send WR has a
variable number of SGEs, based on the contents of the xdr_buf being
sent.

While assembling a Send WR, keep track of the number of SGEs so that
we don't exceed the device's maximum, or walk off the end of the
Send SGE array.

For now the Send path just fails if it exceeds the maximum.

The current logic in svc_rdma_accept bases the maximum number of
Send SGEs on the largest NFS request that can be sent or received.
In the transport layer, the limit is actually based on the
capabilities of the underlying device, not on properties of the
Upper Layer Protocol.
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: J. Bruce Fields <bfields@redhat.com>

svc_rdma_op_ctxt's are pre-allocated and maintained on a per-xprt
free list. This eliminates the overhead of calling kmalloc / kfree,
both of which grab a globally shared lock that disables interrupts.
Introduce a replacement to svc_rdma_op_ctxt's that is built
especially for the svcrdma Send path.

Subsequent patches will take advantage of this new structure by
allocating real resources which are then cached in these objects.
The allocations are freed when the transport is torn down.

I've renamed the structure so that static type checking can be used
to ensure that uses of op_ctxt and send_ctxt are not confused. As an
additional clean up, structure fields are renamed to conform with
kernel coding conventions.

Additional clean ups:
- Handle svc_rdma_send_ctxt_get allocation failure at each call
  site, rather than pre-allocating and hoping we guessed correctly
- All send_ctxt_put call-sites request page freeing, so remove
  the @free_pages argument
- All send_ctxt_put call-sites unmap SGEs, so fold that into
  svc_rdma_send_ctxt_put
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: J. Bruce Fields <bfields@redhat.com>

Clean up: Since there's already a svc_rdma_op_ctxt being passed
around with the running count of mapped SGEs, drop unneeded
parameters to svc_rdma_post_send_wr().
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: J. Bruce Fields <bfields@redhat.com>

Clean up: svc_rdma_dma_map_buf does mostly the same thing as
svc_rdma_dma_map_page, so let's fold these together.
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: J. Bruce Fields <bfields@redhat.com>

The current Receive path uses an array of pages which are allocated
and DMA mapped when each Receive WR is posted, and then handed off
to the upper layer in rqstp::rq_arg. The page flip releases unused
pages in the rq_pages pagelist. This mechanism introduces a
significant amount of overhead.

So instead, kmalloc the Receive buffer, and leave it DMA-mapped
while the transport remains connected. This confers a number of
benefits:

* Each Receive WR requires only one receive SGE, no matter how large
  the inline threshold is. This helps the server-side NFS/RDMA
  transport operate on less capable RDMA devices.

* The Receive buffer is left allocated and mapped all the time. This
  relieves svc_rdma_post_recv from the overhead of allocating and
  DMA-mapping a fresh buffer.

* svc_rdma_wc_receive no longer has to DMA unmap the Receive buffer.
  It has to DMA sync only the number of bytes that were received.

* svc_rdma_build_arg_xdr no longer has to free a page in rq_pages
  for each page in the Receive buffer, making it a constant-time
  function.

* The Receive buffer is now plugged directly into the rq_arg's
  head[0].iov_vec, and can be larger than a page without spilling
  over into rq_arg's page list. This enables simplification of
  the RDMA Read path in subsequent patches.
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: J. Bruce Fields <bfields@redhat.com>

Rather than releasing the incoming svc_rdma_recv_ctxt at the end of
svc_rdma_recvfrom, hold onto it until svc_rdma_sendto.

This permits the contents of the Receive buffer to be preserved
through svc_process and then referenced directly in sendto as it
constructs Write and Reply chunks to return to the client.

The real changes will come in subsequent patches.

Note: I cannot use ->xpo_release_rqst for this purpose because that
is called _before_ ->xpo_sendto. svc_rdma_sendto uses information in
the received Call transport header to construct the Reply transport
header, which is preserved in the RPC's Receive buffer.

The historical comment in svc_send() isn't helpful: it is already
obvious that ->xpo_release_rqst is being called before ->xpo_sendto,
but there is no explanation for this ordering going back to the
beginning of the git era.
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: J. Bruce Fields <bfields@redhat.com>

svc_rdma_op_ctxt's are pre-allocated and maintained on a per-xprt
free list. This eliminates the overhead of calling kmalloc / kfree,
both of which grab a globally shared lock that disables interrupts.
To reduce contention further, separate the use of these objects in
the Receive and Send paths in svcrdma.

Subsequent patches will take advantage of this separation by
allocating real resources which are then cached in these objects.
The allocations are freed when the transport is torn down.

I've renamed the structure so that static type checking can be used
to ensure that uses of op_ctxt and recv_ctxt are not confused. As an
additional clean up, structure fields are renamed to conform with
kernel coding conventions.

As a final clean up, helpers related to recv_ctxt are moved closer
to the functions that use them.
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: J. Bruce Fields <bfields@redhat.com>

This includes:
  * Posting on the Send and Receive queues
  * Send, Receive, Read, and Write completion
  * Connect upcalls
  * QP errors
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: J. Bruce Fields <bfields@redhat.com>

This includes:
  * Transport accept and tear-down
  * Decisions about using Write and Reply chunks
  * Each RDMA segment that is handled
  * Whenever an RDMA_ERR is sent

As a clean-up, I've standardized the order of the includes, and
removed some now redundant dprintk call sites.
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: J. Bruce Fields <bfields@redhat.com>

Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: J. Bruce Fields <bfields@redhat.com>

This change improves Receive efficiency by posting Receives only
on the same CPU that handles Receive completion. Improved latency
and throughput has been noted with this change.
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: J. Bruce Fields <bfields@redhat.com>

The current check will always be true and will always jump to
err1, this looks dubious to me. I believe && should be used
instead of ||.

Detected by CoverityScan, CID#1450120 ("Logically Dead Code")

Fixes: 107c1d0a

 ("svcrdma: Avoid Send Queue overflow")
Signed-off-by: Colin Ian King <colin.king@canonical.com>
Reviewed-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: J. Bruce Fields <bfields@redhat.com>

Sanity case: Catch the case where more Work Requests are being
posted to the Send Queue than there are Send Queue Entries.

This might happen if a client sends a chunk with more segments than
there are SQEs for the transport. The server can't send that reply,
so the transport will deadlock unless the client drops the RPC.
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: J. Bruce Fields <bfields@redhat.com>

The server displays "svcrdma: failed to post Send WR (-107)" in the
kernel log when the client disconnects. This could flood the server's
log, so remove the message.
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: J. Bruce Fields <bfields@redhat.com>

req_maps are no longer used by the send path and can thus be removed.
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Reviewed-by: Sagi Grimberg <sagi@grimberg.me>
Signed-off-by: J. Bruce Fields <bfields@redhat.com>

Observed at Connectathon 2017.

If a client has underestimated the size of a Write or Reply chunk,
the Linux server writes as much payload data as it can, then it
recognizes there was a problem and closes the connection without
sending the transport header.

This creates a couple of problems:

<> The client never receives indication of the server-side failure,
   so it continues to retransmit the bad RPC. Forward progress on
   the transport is blocked.

<> The reply payload pages are not moved out of the svc_rqst, thus
   they can be released by the RPC server before the RDMA Writes
   have completed.

The new rdma_rw-ized helpers return a distinct error code when a
Write/Reply chunk overrun occurs, so it's now easy for the caller
(svc_rdma_sendto) to recognize this case.

Instead of dropping the connection, post an RDMA_ERROR message. The
client now sees an RDMA_ERROR and can properly terminate the RPC
transaction.

As part of the new logic, set up the same delayed release for these
payload pages as would have occurred in the normal case.
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Reviewed-by: Sagi Grimberg <sagi@grimberg.me>
Signed-off-by: J. Bruce Fields <bfields@redhat.com>

Now that svc_rdma_sendto has been renovated, svc_rdma_send_error can
be refactored to reduce code duplication and remove C structure-
based XDR encoding. It is also relocated to the source file that
contains its only caller.

This is a refactoring change only.
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Reviewed-by: Sagi Grimberg <sagi@grimberg.me>
Signed-off-by: J. Bruce Fields <bfields@redhat.com>

The current svcrdma sendto code path posts one RDMA Write WR at a
time. Each of these Writes typically carries a small number of pages
(for instance, up to 30 pages for mlx4 devices). That means a 1MB
NFS READ reply requires 9 ib_post_send() calls for the Write WRs,
and one for the Send WR carrying the actual RPC Reply message.

Instead, use the new rdma_rw API. The details of Write WR chain
construction and memory registration are taken care of in the RDMA
core. svcrdma can focus on the details of the RPC-over-RDMA
protocol. This gives three main benefits:

1. All Write WRs for one RDMA segment are posted in a single chain.
As few as one ib_post_send() for each Write chunk.

2. The Write path can now use FRWR to register the Write buffers.
If the device's maximum page list depth is large, this means a
single Write WR is needed for each RPC's Write chunk data.

3. The new code introduces support for RPCs that carry both a Write
list and a Reply chunk. This combination can be used for an NFSv4
READ where the data payload is large, and thus is removed from the
Payload Stream, but the Payload Stream is still larger than the
inline threshold.
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: J. Bruce Fields <bfields@redhat.com>

Replace C structure-based XDR decoding with more portable code that
instead uses pointer arithmetic.

This is a refactoring change only.
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: J. Bruce Fields <bfields@redhat.com>

Clean up: extract the logic to save pages under I/O into a helper to
add a big documenting comment without adding clutter in the send
path.

This is a refactoring change only.
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Reviewed-by: Sagi Grimberg <sagi@grimberg.me>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: J. Bruce Fields <bfields@redhat.com>

Introduce a helper to DMA-map a reply's transport header before
sending it. This will in part replace the map vector cache.
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: J. Bruce Fields <bfields@redhat.com>

Clean up: Move the ib_send_wr off the stack, and move common code
to post a Send Work Request into a helper.

This is a refactoring change only.
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: J. Bruce Fields <bfields@redhat.com>

Replace C structure-based XDR decoding with pointer arithmetic.
Pointer arithmetic is considered more portable, and is used
throughout the kernel's existing XDR encoders. The gcc optimizer
generates similar assembler code either way.

Byte-swapping before a memory store on x86 typically results in an
instruction pipeline stall. Avoid byte-swapping when encoding a new
header.

svcrdma currently doesn't alter a connection's credit grant value
after the connection has been accepted, so it is effectively a
constant. Cache the byte-swapped value in a separate field.

Christoph suggested pulling the header encoding logic into the only
function that uses it.
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: J. Bruce Fields <bfields@redhat.com>

Commit 5fdca653

 ("svcrdma: Renovate sendto chunk list parsing")
missed a spot. svc_rdma_xdr_get_reply_hdr_len() also assumes the
Write list has only one Write chunk. There's no harm in making this
code more general.
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: J. Bruce Fields <bfields@redhat.com>

No longer any need for the dprintk().
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: J. Bruce Fields <bfields@redhat.com>

svcrdma's current SQ accounting algorithm takes sc_lock and disables
bottom-halves while posting all RDMA Read, Write, and Send WRs.

This is relatively heavyweight serialization. And note that Write and
Send are already fully serialized by the xpt_mutex.

Using a single atomic_t should be all that is necessary to guarantee
that ib_post_send() is called only when there is enough space on the
send queue. This is what the other RDMA-enabled storage targets do.
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: J. Bruce Fields <bfields@redhat.com>

The current sendto code appears to support clients that provide only
one of a Read list, a Write list, or a Reply chunk. My reading of
that code is that it doesn't support the following cases:

 - Read list + Write list
 - Read list + Reply chunk
 - Write list + Reply chunk
 - Read list + Write list + Reply chunk

The protocol allows more than one Read or Write chunk in those
lists. Some clients do send a Read list and Reply chunk
simultaneously. NFSv4 WRITE uses a Read list for the data payload,
and a Reply chunk because the GETATTR result in the reply can
contain a large object like an ACL.

Generalize one of the sendto code paths needed to support all of
the above cases, and attempt to ensure that only one pass is done
through the RPC Call's transport header to gather chunk list
information for building the reply.
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: J. Bruce Fields <bfields@redhat.com>

Support Remote Invalidation. A private message is exchanged with
the client upon RDMA transport connect that indicates whether
Send With Invalidation may be used by the server to send RPC
replies. The invalidate_rkey is arbitrarily chosen from among
rkeys present in the RPC-over-RDMA header's chunk lists.

Send With Invalidate improves performance only when clients can
recognize, while processing an RPC reply, that an rkey has already
been invalidated. That has been submitted as a separate change.

In the future, the RPC-over-RDMA protocol might support Remote
Invalidation properly. The protocol needs to enable signaling
between peers to indicate when Remote Invalidation can be used
for each individual RPC.
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Reviewed-by: Sagi Grimberg <sagi@grimberg.me>
Signed-off-by: J. Bruce Fields <bfields@redhat.com>

Message from syslogd@klimt at Aug 18 17:00:37 ...
 kernel:page:ffffea0020639b00 count:0 mapcount:0 mapping:          (null) index:0x0
Aug 18 17:00:37 klimt kernel: flags: 0x2fffff80000000()
Aug 18 17:00:37 klimt kernel: page dumped because: VM_BUG_ON_PAGE(page_ref_count(page) == 0)

Aug 18 17:00:37 klimt kernel: kernel BUG at /home/cel/src/linux/linux-2.6/include/linux/mm.h:445!
Aug 18 17:00:37 klimt kernel: RIP: 0010:[<ffffffffa05c21c1>] svc_rdma_sendto+0x641/0x820 [rpcrdma]

send_reply() assigns its page argument as the first page of ctxt. On
error, send_reply() already invokes svc_rdma_put_context(ctxt, 1);
which does a put_page() on that very page. No need to do that again
as svc_rdma_sendto exits.

Fixes: 3e1eeb98

 ("svcrdma: Close connection when a send error occurs")
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: J. Bruce Fields <bfields@redhat.com>

The ctxt's count field is overloaded to mean the number of pages in
the ctxt->page array and the number of SGEs in the ctxt->sge array.
Typically these two numbers are the same.

However, when an inline RPC reply is constructed from an xdr_buf
with a tail iovec, the head and tail often occupy the same page,
but each are DMA mapped independently. In that case, ->count equals
the number of pages, but it does not equal the number of SGEs.
There's one more SGE, for the tail iovec. Hence there is one more
DMA mapping than there are pages in the ctxt->page array.

This isn't a real problem until the server's iommu is enabled. Then
each RPC reply that has content in that iovec orphans a DMA mapping
that consists of real resources.

krb5i and krb5p always populate that tail iovec. After a couple
million sent krb5i/p RPC replies, the NFS server starts behaving
erratically. Reboot is needed to clear the problem.

Fixes: 9d11b51c

 ("svcrdma: Fix send_reply() scatter/gather set-up")
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: J. Bruce Fields <bfields@redhat.com>

Get a fresh op_ctxt in send_reply() instead of in svc_rdma_sendto().
This ensures that svc_rdma_put_context() is invoked only once if
send_reply() fails.
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: J. Bruce Fields <bfields@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, svcrdma 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.

This new API also aims each completion at a function that is
specific to the WR's opcode. Thus the ctxt->wr_op field and the
switch in process_context is replaced by a set of methods that
handle each completion type.

Because each ib_cqe carries a pointer to a completion method, the
core can now post operations on a consumer's QP, and handle the
completions itself.

The server's rdma_stat_sq_poll and rdma_stat_sq_prod metrics are no
longer updated.

As a clean up, the cq_event_handler, the dto_tasklet, and all
associated locking is removed, as they are no longer referenced or
used.
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Tested-by: Steve Wise <swise@opengridcomputing.com>
Signed-off-by: J. Bruce Fields <bfields@redhat.com>

Fix several issues with svc_rdma_send_error():

 - Post a receive buffer to replace the one that was consumed by
   the incoming request
 - Posting a send should use DMA_TO_DEVICE, not DMA_FROM_DEVICE
 - No need to put_page _and_ free pages in svc_rdma_put_context
 - Make sure the sge is set up completely in case the error
   path goes through svc_rdma_unmap_dma()
 - Replace the use of ENOSYS, which has a reserved meaning

Related fixes in svc_rdma_recvfrom():

 - Don't leak the ctxt associated with the incoming request
 - Don't close the connection after sending an error reply
 - Let svc_rdma_send_error() figure out the right header error code

As a last clean up, move svc_rdma_send_error() to svc_rdma_sendto.c
with other similar functions. There is some common logic in these
functions that could someday be combined to reduce code duplication.
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Reviewed-by: Devesh Sharma <devesh.sharma@broadcom.com>
Tested-by: Devesh Sharma <devesh.sharma@broadcom.com>
Signed-off-by: J. Bruce Fields <bfields@redhat.com>

Clean up: Most svc_rdma_post_recv() call sites close the transport
connection when a receive cannot be posted. Wrap that in a common
helper.
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Reviewed-by: Devesh Sharma <devesh.sharma@broadcom.com>
Tested-by: Devesh Sharma <devesh.sharma@broadcom.com>
Signed-off-by: J. Bruce Fields <bfields@redhat.com>

Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: J. Bruce Fields <bfields@redhat.com>

The NFS server's XDR encoders adds an XDR pad for content in the
xdr_buf page list at the beginning of the xdr_buf's tail buffer.

On RDMA transports, Write chunks are sent separately and without an
XDR pad.

If a Write chunk is being sent, strip off the pad in the tail buffer
so that inline content following the Write chunk remains XDR-aligned
when it is sent to the client.

BugLink: https://bugzilla.linux-nfs.org/show_bug.cgi?id=294

Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: J. Bruce Fields <bfields@redhat.com>

When the Linux NFS server writes an odd-length data item into a
Write chunk, it finishes with XDR pad bytes. If the data item is
smaller than the Write chunk, the pad bytes are written at the end
of the data item, but still inside the chunk (ie, in the
application's buffer). Since this is direct data placement, that
exposes the pad bytes.

XDR pad bytes are inserted in order to preserve the XDR alignment
of the next XDR data item in an XDR stream. But Write chunks do not
appear in the payload XDR stream, and only one data item is allowed
in each chunk. Thus XDR padding is not needed in a Write chunk.

With NFSv4, the Linux NFS server places the results of any
operations that follow an NFSv4 READ or READLINK in the xdr_buf's
tail. Those results also should never be sent as a part of a Write
chunk. The current logic in send_write_chunks() appears to assume
that the xdr_buf's tail contains only pad bytes (ie, NFSv3).

The server should write only the contents of the xdr_buf's page list
in a Write chunk. If there's more than an XDR pad in the tail, that
needs to go inline or in the Reply chunk.

BugLink: https://bugzilla.linux-nfs.org/show_bug.cgi?id=294

Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: J. Bruce Fields <bfields@redhat.com>