Relieve contention on sc_rw_ctxt_lock by converting rdma->sc_rw_ctxts
to an llist.
The goal is to reduce the average overhead of Send completions,
because a transport's completion handlers are single-threaded on
one CPU core. This change reduces CPU utilization of each Send
completion by 2-3% on my server.
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Reviewed-By: Tom Talpey <tom@talpey.com>
/proc/lock_stat indicates the the sc_send_lock is heavily
contended when the server is under load from a single client.
To address this, convert the send_ctxt free list to an llist.
Returning an item to the send_ctxt cache is now waitless, which
reduces the instruction path length in the single-threaded Send
handler (svc_rdma_wc_send).
The goal is to enable the ib_comp_wq worker to handle a higher
RPC/RDMA Send completion rate given the same CPU resources. This
change reduces CPU utilization of Send completion by 2-3% on my
server.
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Reviewed-By: Tom Talpey <tom@talpey.com>
Because wake_up() takes an IRQ-safe lock, it can be expensive,
especially to call inside of a single-threaded completion handler.
What's more, the Send wait queue almost never has waiters, so
most of the time, this is an expensive no-op.
As always, the goal is to reduce the average overhead of each
completion, because a transport's completion handlers are single-
threaded on one CPU core. This change reduces CPU utilization of
the Send completion thread by 2-3% on my server.
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Reviewed-By: Tom Talpey <tom@talpey.com>
These fields are no longer used.
The size of struct svc_rdma_recv_ctxt is now less than 300 bytes on
x86_64, down from 2440 bytes.
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Currently svc_rdma_sendto() migrates xdr_buf pages into a separate
page list and NULLs out a bunch of entries in rq_pages while the
pages are under I/O. The Send completion handler then frees those
pages later.
Instead, let's wait for the Send completion, then handle page
releasing in the nfsd thread. I'd like to avoid the cost of 250+
put_page() calls in the Send completion handler, which is single-
threaded.
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Post more Receives when the number of pending Receives drops below
a water mark. The batch mechanism is disabled if the underlying
device cannot support a reasonably-sized Receive Queue.
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
I tested commit 43042b90ca ("svcrdma: Reduce Receive doorbell
rate") with mlx4 (IB) and software iWARP and didn't find any
issues. However, I recently got my hardware iWARP setup back on
line (FastLinQ) and it's crashing hard on this commit (confirmed
via bisect).
The failure mode is complex.
- After a connection is established, the first Receive completes
normally.
- But the second and third Receives have garbage in their Receive
buffers. The server responds with ERR_VERS as a result.
- When the client tears down the connection to retry, a couple
of posted Receives flush twice, and that corrupts the recv_ctxt
free list.
- __svc_rdma_free then faults or loops infinitely while destroying
the xprt's recv_ctxts.
Since 43042b90ca ("svcrdma: Reduce Receive doorbell rate") does
not fix a bug but is a scalability enhancement, it's safe and
appropriate to revert it while working on a replacement.
Fixes: 43042b90ca ("svcrdma: Reduce Receive doorbell rate")
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
This is similar to commit e340c2d6ef ("xprtrdma: Reduce the
doorbell rate (Receive)") which added Receive batching to the
client.
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Clean up. We are not permitted to remove old proc files. Instead,
convert these variables to stubs that are only ever allowed to
display a value of zero.
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Now that we have an efficient mechanism to update these two stats,
let's start maintaining them again.
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Receives are frequent events. Avoid the overhead of a memory bus
lock cycle for counting a value that is hardly every used.
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
As a pre-requisite for handling multiple Read chunks in each Read
list, convert svc_rdma_recv_read_chunk() to use the new parsed Read
chunk list.
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Refactor svc_rdma_send_reply_chunk() so that it Sends only the parts
of rq_res that do not contain a result payload.
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Refactor: svc_rdma_map_reply_msg() is restructured to DMA map only
the parts of rq_res that do not contain a result payload.
This change has been tested to confirm that it does not cause a
regression in the no Write chunk and single Write chunk cases.
Multiple Write chunks have not been tested.
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
When counting the number of SGEs needed to construct a Send request,
do not count result payloads. And, when copying the Reply message
into the pull-up buffer, result payloads are not to be copied to the
Send buffer.
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Refactor: Instead of re-parsing the ingress RPC Call transport
header when constructing RDMA Writes, use the new parsed chunk lists
for the Write list and Reply chunk, which are version-agnostic and
already XDR-decoded.
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Refactor: Don't duplicate header decoding smarts here. Instead, use
the new parsed chunk lists.
Note that the XID sanity test is also removed. The XID is already
looked up by the cb handler, and is rejected if it's not recognized.
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
This simple data structure binds the location of each data payload
inside of an RPC message to the chunk that will be used to push it
to or pull it from the client.
There are several benefits to this small additional overhead:
* It enables support for more than one chunk in incoming Read and
Write lists.
* It translates the version-specific on-the-wire format into a
generic in-memory structure, enabling support for multiple
versions of the RPC/RDMA transport protocol.
* It enables the server to re-organize a chunk list if it needs to
adjust where Read chunk data lands in server memory without
altering the contents of the XDR-encoded Receive buffer.
Construction of these lists is done while sanity checking each
incoming RPC/RDMA header. Subsequent patches will make use of the
generated data structures.
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
The only RPC/RDMA ordering requirement between RDMA Writes and RDMA
Sends is that the responder must post the Writes on the Send queue
before posting the Send that conveys the RPC Reply for that Write
payload.
The Linux NFS server implementation now has a transport method that
can post result Payload Writes earlier than svc_rdma_sendto:
->xpo_result_payload()
This gets RDMA Writes going earlier so they are more likely to be
complete at the remote end before the Send completes.
Some care must be taken with pulled-up Replies. We don't want to
push the Write chunk and then send the same payload data via Send.
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Clean up: "result payload" is a less confusing name for these
payloads. "READ payload" reflects only the NFS usage.
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
First, refactor: Dereference the svc_rdma_send_ctxt inside
svc_rdma_send() instead of at every call site.
Then, it can be passed into trace_svcrdma_post_send() to get the
proper completion ID.
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Set up a completion ID in each svc_rdma_send_ctxt. The ID is used
to match an incoming Send completion to a transport and to a
previous ib_post_send().
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Set up a completion ID in each svc_rdma_recv_ctxt. The ID is used
to match an incoming Receive completion to a transport and to a
previous ib_post_recv().
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Clean up: Commit d21b05f101 ("rdma: SVCRMDA Header File")
introduced the SVCRDMA_DEBUG macro, but it doesn't seem to have been
used.
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Way back when I was writing the RPC/RDMA server-side backchannel
code, I misread the TCP backchannel reply handler logic. When
svc_tcp_recvfrom() successfully receives a backchannel reply, it
does not return -EAGAIN. It sets XPT_DATA and returns zero.
Update svc_rdma_recvfrom() to return zero. Here, XPT_DATA doesn't
need to be set again: it is set whenever a new message is received,
behind a spin lock in a single threaded context.
Also, if handling the cb reply is not successful, the message is
simply dropped. There's no special message framing to deal with as
there is in the TCP case.
Now that the handle_bc_reply() return value is ignored, I've removed
the dprintk call sites in the error exit of handle_bc_reply() in
favor of trace points in other areas that already report the error
cases.
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Utilize the xpo_release_rqst transport method to ensure that each
rqstp's svc_rdma_recv_ctxt object is released even when the server
cannot return a Reply for that rqstp.
Without this fix, each RPC whose Reply cannot be sent leaks one
svc_rdma_recv_ctxt. This is a 2.5KB structure, a 4KB DMA-mapped
Receive buffer, and any pages that might be part of the Reply
message.
The leak is infrequent unless the network fabric is unreliable or
Kerberos is in use, as GSS sequence window overruns, which result
in connection loss, are more common on fast transports.
Fixes: 3a88092ee3 ("svcrdma: Preserve Receive buffer until svc_rdma_sendto")
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
On some platforms, DMA mapping part of a page is more costly than
copying bytes. Indeed, not involving the I/O MMU can help the
RPC/RDMA transport scale better for tiny I/Os across more RDMA
devices. This is because interaction with the I/O MMU is eliminated
for each of these small I/Os. Without the explicit unmapping, the
NIC no longer needs to do a costly internal TLB shoot down for
buffers that are just a handful of bytes.
Since pull-up is now a more a frequent operation, I've introduced a
trace point in the pull-up path. It can be used for debugging or
user-space tools that count pull-up frequency.
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Performance optimization: Avoid syncing the transport buffer twice
when Reply buffer pull-up is necessary.
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Same idea as the receive-side changes I did a while back: use
xdr_stream helpers rather than open-coding the XDR chunk list
encoders. This builds the Reply transport header from beginning to
end without backtracking.
As additional clean-ups, fill in documenting comments for the XDR
encoders and sprinkle some trace points in the new encoding
functions.
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Preparing for subsequent patches, no behavior change expected.
Pass the RPC Call's svc_rdma_recv_ctxt deeper into the sendto()
path. This enables passing more information about Requester-
provided Write and Reply chunks into those lower-level functions.
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Preparing for subsequent patches, no behavior change expected.
Pass the RPC Call's svc_rdma_recv_ctxt deeper into the sendto()
path. This enables passing more information about Requester-
provided Write and Reply chunks into the lower-level send
functions.
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Cache the locations of the Requester-provided Write list and Reply
chunk so that the Send path doesn't need to parse the Call header
again.
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
The logic that checks incoming network headers has to be scrupulous.
De-duplicate: replace open-coded buffer overflow checks with the use
of xdr_stream helpers that are used most everywhere else XDR
decoding is done.
One minor change to the sanity checks: instead of checking the
length of individual segments, cap the length of the whole chunk
to be sure it can fit in the set of pages available in rq_pages.
This should be a better test of whether the server can handle the
chunks in each request.
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
svcrdma expects that the payload falls precisely into the xdr_buf
page vector. This does not seem to be the case for
nfsd4_encode_readv().
This code is called only when fops->splice_read is missing or when
RQ_SPLICE_OK is clear, so it's not a noticeable problem in many
common cases.
Add new transport method: ->xpo_read_payload so that when a READ
payload does not fit exactly in rq_res's page vector, the XDR
encoder can inform the RPC transport exactly where that payload is,
without the payload's XDR pad.
That way, when a Write chunk is present, the transport knows what
byte range in the Reply message is supposed to be matched with the
chunk.
Note that the Linux NFS server implementation of NFS/RDMA can
currently handle only one Write chunk per RPC-over-RDMA message.
This simplifies the implementation of this fix.
Fixes: b042098063 ("nfsd4: allow exotic read compounds")
Buglink: https://bugzilla.kernel.org/show_bug.cgi?id=198053
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Use a wait-free mechanism for managing the svc_rdma_recv_ctxts free
list. Subsequently, sc_recv_lock can be eliminated.
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: J. Bruce Fields <bfields@redhat.com>
Clean up: the system workqueue will work just as well.
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: J. Bruce Fields <bfields@redhat.com>
xpo_prep_reply_hdr are not used now.
It was defined for tcp transport only, however it cannot be
called indirectly, so let's move it to its caller and
remove unused callback.
Signed-off-by: Vasily Averin <vvs@virtuozzo.com>
Signed-off-by: J. Bruce Fields <bfields@redhat.com>
o Select the R_key to invalidate while the CPU cache still contains
the received RPC Call transport header, rather than waiting until
we're about to send the RPC Reply.
o Choose Send With Invalidate if there is exactly one distinct R_key
in the received transport header. If there's more than one, the
client will have to perform local invalidation after it has
already waited for remote invalidation.
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: J. Bruce Fields <bfields@redhat.com>
Reduce queuing on clients by allowing more credits by default.
64 is the default NFSv4.1 slot table size on Linux clients. This
size prevents the credit limit from putting RPC requests to sleep
again after they have already slept waiting for a session slot.
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>
Clean up: Eliminate a structure that is no longer used.
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: J. Bruce Fields <bfields@redhat.com>
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: 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>
Chuck Lever