| Age | Commit message (Collapse) | Author | Files | Lines |
|---|
|
If a reply has been processed but the RPC is later retransmitted
anyway, the req->rl_reply field still contains the only pointer to
the old rpcrdma rep. When the next reply comes in, the reply handler
will stomp on the rl_reply field, leaking the old rep.
A trace event is added to capture such leaks.
This problem seems to be worsened by the restructuring of the RPC
Call path in v4.20. Fully addressing this issue will require at
least a re-architecture of the disconnect logic, which is not
appropriate during -rc.
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: Anna Schumaker <Anna.Schumaker@Netapp.com>
|
|
These are rare, but can be helpful at tracking down DMAR and other
problems.
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: Anna Schumaker <Anna.Schumaker@Netapp.com>
|
|
The chunk-related trace points capture nearly the same information
as the MR-related trace points.
Also, rename them so globbing can be used to enable or disable
these trace points more easily.
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: Anna Schumaker <Anna.Schumaker@Netapp.com>
|
|
Clean up: Remove dprintk() call sites that report rare or impossible
errors. Leave a few that display high-value low noise status
information.
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: Anna Schumaker <Anna.Schumaker@Netapp.com>
|
|
For better observability of parsing errors, return the error code
generated in the decoders to the upper layer consumer.
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: Anna Schumaker <Anna.Schumaker@Netapp.com>
|
|
Commit 431f6eb3570f ("SUNRPC: Add a label for RPC calls that require
allocation on receive") didn't update similar logic in rpc_rdma.c.
I don't think this is a bug, per-se; the commit just adds more
careful checking for broken upper layer behavior.
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: Anna Schumaker <Anna.Schumaker@Netapp.com>
|
|
Place the associated RPC transaction's XID in the upper 32 bits of
each RDMA segment's rdma_offset field. There are two reasons to do
this:
- The R_key only has 8 bits that are different from registration to
registration. The XID adds more uniqueness to each RDMA segment to
reduce the likelihood of a software bug on the server reading from
or writing into memory it's not supposed to.
- On-the-wire RDMA Read and Write requests do not otherwise carry
any identifier that matches them up to an RPC. The XID in the
upper 32 bits will act as an eye-catcher in network captures.
Suggested-by: Tom Talpey <ttalpey@microsoft.com>
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: Anna Schumaker <Anna.Schumaker@Netapp.com>
|
|
Clean up: Now that there is only FRWR, there is no need for a memory
registration switch. The indirect calls to the memreg operations can
be replaced with faster direct calls.
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: Anna Schumaker <Anna.Schumaker@Netapp.com>
|
|
To address a connection-close ordering problem, we need the ability
to drain the RPC completions running on rpcrdma_receive_wq for just
one transport. Give each transport its own RPC completion workqueue,
and drain that workqueue when disconnecting the transport.
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: Anna Schumaker <Anna.Schumaker@Netapp.com>
|
|
Clean up: Divide the work cleanly:
- rpcrdma_wc_receive is responsible only for RDMA Receives
- rpcrdma_reply_handler is responsible only for RPC Replies
- the posted send and receive counts both belong in rpcrdma_ep
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: Anna Schumaker <Anna.Schumaker@Netapp.com>
|
|
git://git.linux-nfs.org/projects/anna/linux-nfs
NFS RDMA client updates for Linux 4.20
Stable bugfixes:
- Reset credit grant properly after a disconnect
Other bugfixes and cleanups:
- xprt_release_rqst_cong is called outside of transport_lock
- Create more MRs at a time and toss out old ones during recovery
- Various improvements to the RDMA connection and disconnection code:
- Improve naming of trace events, functions, and variables
- Add documenting comments
- Fix metrics and stats reporting
- Fix a tracepoint sparse warning
Signed-off-by: Trond Myklebust <trond.myklebust@hammerspace.com>
|
|
When a memory operation fails, the MR's driver state might not match
its hardware state. The only reliable recourse is to dereg the MR.
This is done in ->ro_recover_mr, which then attempts to allocate a
fresh MR to replace the released MR.
Since commit e2ac236c0b651 ("xprtrdma: Allocate MRs on demand"),
xprtrdma dynamically allocates MRs. It can add more MRs whenever
they are needed.
That makes it possible to simply release an MR when a memory
operation fails, instead of "recovering" it. It will automatically
be replaced by the on-demand MR allocator.
This commit is a little larger than I wanted, but it replaces
->ro_recover_mr, rb_recovery_lock, rb_recovery_worker, and the
rb_stale_mrs list with a generic work queue.
Since MRs are no longer orphaned, the mrs_orphaned metric is no
longer used.
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: Anna Schumaker <Anna.Schumaker@Netapp.com>
|
|
Some devices require more than 3 MRs to build a single 1MB I/O.
Ensure that rpcrdma_mrs_create() will add enough MRs to build that
I/O.
In a subsequent patch I'm changing the MR recovery logic to just
toss out the MRs. In that case it's possible for ->send_request to
loop acquiring some MRs, not getting enough, getting called again,
recycling the previous MRs, then not getting enough, lather rinse
repeat. Thus first we need to ensure enough MRs are created to
prevent that loop.
I'm "reusing" ia->ri_max_segs. All of its accessors seem to want the
maximum number of data segments plus two, so I'm going to bake that
into the initial calculation.
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: Anna Schumaker <Anna.Schumaker@Netapp.com>
|
|
Since commit ce7c252a8c74 ("SUNRPC: Add a separate spinlock to
protect the RPC request receive list") the RPC/RDMA reply handler
has been calling xprt_release_rqst_cong without holding
xprt->transport_lock.
I think the only way this call is ever made is if the credit grant
increases and there are RPCs pending. Current server implementations
do not change their credit grant during operation (except at
connect time).
Commit e7ce710a8802 ("xprtrdma: Avoid deadlock when credit window is
reset") added the ->release_rqst call because UDP invokes
xprt_adjust_cwnd(), which calls __xprt_put_cong() after adjusting
xprt->cwnd. Both xprt_release() and ->xprt_release_xprt already wake
another task in this case, so it is safe to remove this call from
the reply handler.
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: Anna Schumaker <Anna.Schumaker@Netapp.com>
|
|
Treat socket write space handling in the same way we now treat transport
congestion: by denying the XPRT_LOCK until the transport signals that it
has free buffer space.
Signed-off-by: Trond Myklebust <trond.myklebust@hammerspace.com>
|
|
We will use the same lock to protect both the transmit and receive queues.
Signed-off-by: Trond Myklebust <trond.myklebust@hammerspace.com>
|
|
Pull NFS client updates from Trond Myklebust:
"Highlights include:
Stable fixes:
- Fix a 1-byte stack overflow in nfs_idmap_read_and_verify_message
- Fix a hang due to incorrect error returns in rpcrdma_convert_iovs()
- Revert an incorrect change to the NFSv4.1 callback channel
- Fix a bug in the NFSv4.1 sequence error handling
Features and optimisations:
- Support for piggybacking a LAYOUTGET operation to the OPEN compound
- RDMA performance enhancements to deal with transport congestion
- Add proper SPDX tags for NetApp-contributed RDMA source
- Do not request delegated file attributes (size+change) from the
server
- Optimise away a GETATTR in the lookup revalidate code when doing
NFSv4 OPEN
- Optimise away unnecessary lookups for rename targets
- Misc performance improvements when freeing NFSv4 delegations
Bugfixes and cleanups:
- Try to fail quickly if proto=rdma
- Clean up RDMA receive trace points
- Fix sillyrename to return the delegation when appropriate
- Misc attribute revalidation fixes
- Immediately clear the pNFS layout on a file when the server returns
ESTALE
- Return NFS4ERR_DELAY when delegation/layout recalls fail due to
igrab()
- Fix the client behaviour on NFS4ERR_SEQ_FALSE_RETRY"
* tag 'nfs-for-4.18-1' of git://git.linux-nfs.org/projects/trondmy/linux-nfs: (80 commits)
skip LAYOUTRETURN if layout is invalid
NFSv4.1: Fix the client behaviour on NFS4ERR_SEQ_FALSE_RETRY
NFSv4: Fix a typo in nfs41_sequence_process
NFSv4: Revert commit 5f83d86cf531d ("NFSv4.x: Fix wraparound issues..")
NFSv4: Return NFS4ERR_DELAY when a layout recall fails due to igrab()
NFSv4: Return NFS4ERR_DELAY when a delegation recall fails due to igrab()
NFSv4.0: Remove transport protocol name from non-UCS client ID
NFSv4.0: Remove cl_ipaddr from non-UCS client ID
NFSv4: Fix a compiler warning when CONFIG_NFS_V4_1 is undefined
NFS: Filter cache invalidation when holding a delegation
NFS: Ignore NFS_INO_REVAL_FORCED in nfs_check_inode_attributes()
NFS: Improve caching while holding a delegation
NFS: Fix attribute revalidation
NFS: fix up nfs_setattr_update_inode
NFSv4: Ensure the inode is clean when we set a delegation
NFSv4: Ignore NFS_INO_REVAL_FORCED in nfs4_proc_access
NFSv4: Don't ask for delegated attributes when adding a hard link
NFSv4: Don't ask for delegated attributes when revalidating the inode
NFS: Pass the inode down to the getattr() callback
NFSv4: Don't request size+change attribute if they are delegated to us
...
|
|
Clean up: This array was used in a dprintk that was replaced by a
trace point in commit ab03eff58eb5 ("xprtrdma: Add trace points in
RPC Call transmit paths").
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: Anna Schumaker <Anna.Schumaker@Netapp.com>
|
|
Currently, when the sendctx queue is exhausted during marshaling, the
RPC/RDMA transport places the RPC task on the delayq, which forces a
wait for HZ >> 2 before the marshal and send is retried.
With this change, the transport now places such an RPC task on the
pending queue, and wakes it just as soon as more sendctxs become
available. This typically takes less than a millisecond, and the
write_space waking mechanism is less deadlock-prone.
Moreover, the waiting RPC task is holding the transport's write
lock, which blocks the transport from sending RPCs. Therefore faster
recovery from sendctx queue exhaustion is desirable.
Cf. commit 5804891455d5 ("xprtrdma: ->send_request returns -EAGAIN
when there are no free MRs").
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: Anna Schumaker <Anna.Schumaker@Netapp.com>
|
|
Clean up: The logic to wait for write space is common to a bunch of
the encoding helper functions. Lift it out and put it in the tail
of rpcrdma_marshal_req().
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: Anna Schumaker <Anna.Schumaker@Netapp.com>
|
|
The use of -EAGAIN in rpcrdma_convert_iovs() is a latent bug: the
transport never calls xprt_write_space() when more pages become
available. -ENOBUFS will trigger the correct "delay briefly and call
again" logic.
Fixes: 7a89f9c626e3 ("xprtrdma: Honor ->send_request API contract")
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Cc: stable@vger.kernel.org # 4.8+
Signed-off-by: Anna Schumaker <Anna.Schumaker@Netapp.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>
|
|
Clean up: Move #include <trace/events/rpcrdma.h> into source files,
similar to how it is done with trace/events/sunrpc.h.
Server-side trace points will be part of the rpcrdma subsystem,
just like the client-side trace points.
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: J. Bruce Fields <bfields@redhat.com>
|
|
Receive completion and Reply handling are done by a BOUND
workqueue, meaning they run on only one CPU.
Posting receives is currently done in the send_request path, which
on large systems is typically done on a different CPU than the one
handling Receive completions. This results in movement of
Receive-related cachelines between the sending and receiving CPUs.
More importantly, it means that currently Receives are posted while
the transport's write lock is held, which is unnecessary and costly.
Finally, allocation of Receive buffers is performed on-demand in
the Receive completion handler. This helps guarantee that they are
allocated on the same NUMA node as the CPU that handles Receive
completions.
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: Anna Schumaker <Anna.Schumaker@Netapp.com>
|
|
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: Anna Schumaker <Anna.Schumaker@Netapp.com>
|
|
Currently, when the MR free list is exhausted during marshaling, the
RPC/RDMA transport places the RPC task on the delayq, which forces a
wait for HZ >> 2 before the marshal and send is retried.
With this change, the transport now places such an RPC task on the
pending queue, and wakes it just as soon as more MRs have been
created. Creating more MRs typically takes less than a millisecond,
and this waking mechanism is less deadlock-prone.
Moreover, the waiting RPC task is holding the transport's write
lock, which blocks the transport from sending RPCs. Therefore faster
recovery from MR exhaustion is desirable.
This is the same mechanism that the TCP transport utilizes when
handling write buffer space exhaustion.
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: Anna Schumaker <Anna.Schumaker@Netapp.com>
|
|
With v4.15, on one of my NFS/RDMA clients I measured a nearly
doubling in the latency of small read and write system calls. There
was no change in server round trip time. The extra latency appears
in the whole RPC execution path.
"git bisect" settled on commit ccede7598588 ("xprtrdma: Spread reply
processing over more CPUs") .
After some experimentation, I found that leaving the WQ bound and
allowing the scheduler to pick the dispatch CPU seems to eliminate
the long latencies, and it does not introduce any new regressions.
The fix is implemented by reverting only the part of
commit ccede7598588 ("xprtrdma: Spread reply processing over more
CPUs") that dispatches RPC replies specifically on the CPU where the
matching RPC call was made.
Interestingly, saving the CPU number and later queuing reply
processing there was effective _only_ for a NFS READ and WRITE
request. On my NUMA client, in-kernel RPC reply processing for
asynchronous RPCs was dispatched on the same CPU where the RPC call
was made, as expected. However synchronous RPCs seem to get their
reply dispatched on some other CPU than where the call was placed,
every time.
Fixes: ccede7598588 ("xprtrdma: Spread reply processing over ... ")
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Cc: stable@vger.kernel.org # v4.15+
Signed-off-by: Anna Schumaker <Anna.Schumaker@Netapp.com>
|
|
Commit 16f906d66cd7 ("xprtrdma: Reduce required number of send
SGEs") introduced the rpcrdma_ia::ri_max_send_sges field. This fixes
a problem where xprtrdma would not work if the device's max_sge
capability was small (low single digits).
At least RPCRDMA_MIN_SEND_SGES are needed for the inline parts of
each RPC. ri_max_send_sges is set to this value:
ia->ri_max_send_sges = max_sge - RPCRDMA_MIN_SEND_SGES;
Then when marshaling each RPC, rpcrdma_args_inline uses that value
to determine whether the device has enough Send SGEs to convey an
NFS WRITE payload inline, or whether instead a Read chunk is
required.
More recently, commit ae72950abf99 ("xprtrdma: Add data structure to
manage RDMA Send arguments") used the ri_max_send_sges value to
calculate the size of an array, but that commit erroneously assumed
ri_max_send_sges contains a value similar to the device's max_sge,
and not one that was reduced by the minimum SGE count.
This assumption results in the calculated size of the sendctx's
Send SGE array to be too small. When the array is used to marshal
an RPC, the code can write Send SGEs into the following sendctx
element in that array, corrupting it. When the device's max_sge is
large, this issue is entirely harmless; but it results in an oops
in the provider's post_send method, if dev.attrs.max_sge is small.
So let's straighten this out: ri_max_send_sges will now contain a
value with the same meaning as dev.attrs.max_sge, which makes
the code easier to understand, and enables rpcrdma_sendctx_create
to calculate the size of the SGE array correctly.
Reported-by: Michal Kalderon <Michal.Kalderon@cavium.com>
Fixes: 16f906d66cd7 ("xprtrdma: Reduce required number of send SGEs")
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Tested-by: Michal Kalderon <Michal.Kalderon@cavium.com>
Cc: stable@vger.kernel.org # v4.10+
Signed-off-by: Anna Schumaker <Anna.Schumaker@Netapp.com>
|
|
The contents of seg->mr_len changed when ->ro_map stopped returning
the full chunk length in the first segment. Count the full length of
each Write chunk, not the length of the first segment (which now can
only be as large as a page).
Fixes: 9d6b04097882 ("xprtrdma: Place registered MWs on a ... ")
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: Anna Schumaker <Anna.Schumaker@Netapp.com>
|
|
This includes decoding Write and Reply chunks, and fixing up inline
payloads.
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: Anna Schumaker <Anna.Schumaker@Netapp.com>
|
|
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: Anna Schumaker <Anna.Schumaker@Netapp.com>
|
|
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: Anna Schumaker <Anna.Schumaker@Netapp.com>
|
|
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: Anna Schumaker <Anna.Schumaker@Netapp.com>
|
|
Clean up: struct rpcrdma_mw was named after Memory Windows, but
xprtrdma no longer supports a Memory Window registration mode.
Rename rpcrdma_mw and its fields to reduce confusion and make
the code more sensible to read.
Renaming "mw" was suggested by Tom Talpey, the author of the
original xprtrdma implementation. It's a good idea, but I haven't
done this until now because it's a huge diffstat for no benefit
other than code readability.
However, I'm about to introduce static trace points that expose
a few of xprtrdma's internal data structures. They should make sense
in the trace report, and it's reasonable to treat trace points as a
kernel API contract which might be difficult to change later.
While I'm churning things up, two additional changes:
- rename variables unhelpfully called "r" to "mr", to improve code
clarity, and
- rename the MR-related helper functions using the form
"rpcrdma_mr_<verb>", to be consistent with other areas of the
code.
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: Anna Schumaker <Anna.Schumaker@Netapp.com>
|
|
Clean up. @rqst is set up differently for backchannel Replies. For
example, rqst->rq_task and task->tk_client are both NULL. So it is
easier to understand and maintain this code path if it is separated.
Also, we can get rid of the confusing rl_connect_cookie hack in
rpcrdma_bc_receive_call.
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: Anna Schumaker <Anna.Schumaker@Netapp.com>
|
|
Clean up. This logic is related to marshaling the request, and I'd
like to keep everything that touches req->rl_registered close
together, for CPU cache efficiency.
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: Anna Schumaker <Anna.Schumaker@Netapp.com>
|
|
Refactoring change: Remote Invalidation is particular to the memory
registration mode that is use. Use a callout instead of a generic
function to handle Remote Invalidation.
This gets rid of the 8-byte flags field in struct rpcrdma_mw, of
which only a single bit flag has been allocated.
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: Anna Schumaker <Anna.Schumaker@Netapp.com>
|
|
Commit d8f532d20ee4 ("xprtrdma: Invoke rpcrdma_reply_handler
directly from RECV completion") introduced a performance regression
for NFS I/O small enough to not need memory registration. In multi-
threaded benchmarks that generate primarily small I/O requests,
IOPS throughput is reduced by nearly a third. This patch restores
the previous level of throughput.
Because workqueues are typically BOUND (in particular ib_comp_wq,
nfsiod_workqueue, and rpciod_workqueue), NFS/RDMA workloads tend
to aggregate on the CPU that is handling Receive completions.
The usual approach to addressing this problem is to create a QP
and CQ for each CPU, and then schedule transactions on the QP
for the CPU where you want the transaction to complete. The
transaction then does not require an extra context switch during
completion to end up on the same CPU where the transaction was
started.
This approach doesn't work for the Linux NFS/RDMA client because
currently the Linux NFS client does not support multiple connections
per client-server pair, and the RDMA core API does not make it
straightforward for ULPs to determine which CPU is responsible for
handling Receive completions for a CQ.
So for the moment, record the CPU number in the rpcrdma_req before
the transport sends each RPC Call. Then during Receive completion,
queue the RPC completion on that same CPU.
Additionally, move all RPC completion processing to the deferred
handler so that even RPCs with simple small replies complete on
the CPU that sent the corresponding RPC Call.
Fixes: d8f532d20ee4 ("xprtrdma: Invoke rpcrdma_reply_handler ...")
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: Anna Schumaker <Anna.Schumaker@Netapp.com>
|
|
Credit work contributed by Oracle engineers since 2014.
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: Anna Schumaker <Anna.Schumaker@Netapp.com>
|
|
Clean up: C-structure style XDR encoding and decoding logic has
been replaced over the past several merge windows on both the
client and server. These data structures are no longer used.
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Reviewed-by: Devesh Sharma <devesh.sharma@broadcom.com>
Signed-off-by: Anna Schumaker <Anna.Schumaker@Netapp.com>
|
|
When an RPC Call includes a file data payload, that payload can come
from pages in the page cache, or a user buffer (for direct I/O).
If the payload can fit inline, xprtrdma includes it in the Send
using a scatter-gather technique. xprtrdma mustn't allow the RPC
consumer to re-use the memory where that payload resides before the
Send completes. Otherwise, the new contents of that memory would be
exposed by an HCA retransmit of the Send operation.
So, block RPC completion on Send completion, but only in the case
where a separate file data payload is part of the Send. This
prevents the reuse of that memory while it is still part of a Send
operation without an undue cost to other cases.
Waiting is avoided in the common case because typically the Send
will have completed long before the RPC Reply arrives.
These days, an RPC timeout will trigger a disconnect, which tears
down the QP. The disconnect flushes all waiting Sends. This bounds
the amount of time the reply handler has to wait for a Send
completion.
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: Anna Schumaker <Anna.Schumaker@Netapp.com>
|
|
Invoke a common routine for releasing hardware resources (for
example, invalidating MRs). This needs to be done whether an
RPC Reply has arrived or the RPC was terminated early.
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: Anna Schumaker <Anna.Schumaker@Netapp.com>
|
|
Problem statement:
Recently Sagi Grimberg <sagi@grimberg.me> observed that kernel RDMA-
enabled storage initiators don't handle delayed Send completion
correctly. If Send completion is delayed beyond the end of a ULP
transaction, the ULP may release resources that are still being used
by the HCA to complete a long-running Send operation.
This is a common design trait amongst our initiators. Most Send
operations are faster than the ULP transaction they are part of.
Waiting for a completion for these is typically unnecessary.
Infrequently, a network partition or some other problem crops up
where an ordering problem can occur. In NFS parlance, the RPC Reply
arrives and completes the RPC, but the HCA is still retrying the
Send WR that conveyed the RPC Call. In this case, the HCA can try
to use memory that has been invalidated or DMA unmapped, and the
connection is lost. If that memory has been re-used for something
else (possibly not related to NFS), and the Send retransmission
exposes that data on the wire.
Thus we cannot assume that it is safe to release Send-related
resources just because a ULP reply has arrived.
After some analysis, we have determined that the completion
housekeeping will not be difficult for xprtrdma:
- Inline Send buffers are registered via the local DMA key, and
are already left DMA mapped for the lifetime of a transport
connection, thus no additional handling is necessary for those
- Gathered Sends involving page cache pages _will_ need to
DMA unmap those pages after the Send completes. But like
inline send buffers, they are registered via the local DMA key,
and thus will not need to be invalidated
In addition, RPC completion will need to wait for Send completion
in the latter case. However, nearly always, the Send that conveys
the RPC Call will have completed long before the RPC Reply
arrives, and thus no additional latency will be accrued.
Design notes:
In this patch, the rpcrdma_sendctx object is introduced, and a
lock-free circular queue is added to manage a set of them per
transport.
The RPC client's send path already prevents sending more than one
RPC Call at the same time. This allows us to treat the consumer
side of the queue (rpcrdma_sendctx_get_locked) as if there is a
single consumer thread.
The producer side of the queue (rpcrdma_sendctx_put_locked) is
invoked only from the Send completion handler, which is a single
thread of execution (soft IRQ).
The only care that needs to be taken is with the tail index, which
is shared between the producer and consumer. Only the producer
updates the tail index. The consumer compares the head with the
tail to ensure that the a sendctx that is in use is never handed
out again (or, expressed more conventionally, the queue is empty).
When the sendctx queue empties completely, there are enough Sends
outstanding that posting more Send operations can result in a Send
Queue overflow. In this case, the ULP is told to wait and try again.
This introduces strong Send Queue accounting to xprtrdma.
As a final touch, Jason Gunthorpe <jgunthorpe@obsidianresearch.com>
suggested a mechanism that does not require signaling every Send.
We signal once every N Sends, and perform SGE unmapping of N Send
operations during that one completion.
Reported-by: Sagi Grimberg <sagi@grimberg.me>
Suggested-by: Jason Gunthorpe <jgunthorpe@obsidianresearch.com>
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: Anna Schumaker <Anna.Schumaker@Netapp.com>
|
|
Commit 655fec6987be ("xprtrdma: Use gathered Send for large inline
messages") assumed that, since the zeroeth element of the Send SGE
array always pointed to req->rl_rdmabuf, it needed to be initialized
just once. This was a valid assumption because the Send SGE array
and rl_rdmabuf both live in the same rpcrdma_req.
In a subsequent patch, the Send SGE array will be separated from the
rpcrdma_req, so the zeroeth element of the SGE array needs to be
initialized every time.
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: Anna Schumaker <Anna.Schumaker@Netapp.com>
|
|
Clean up: Make rpcrdma_prepare_send_sges() return a negative errno
instead of a bool. Soon callers will want distinct treatments of
different types of failures.
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: Anna Schumaker <Anna.Schumaker@Netapp.com>
|
|
When this function fails, it needs to undo the DMA mappings it's
done so far. Otherwise these are leaked.
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: Anna Schumaker <Anna.Schumaker@Netapp.com>
|
|
Clean up. rpcrdma_prepare_hdr_sge() sets num_sge to one, then
rpcrdma_prepare_msg_sges() sets num_sge again to the count of SGEs
it added, plus one for the header SGE just mapped in
rpcrdma_prepare_hdr_sge(). This is confusing, and nails in an
assumption about when these functions are called.
Instead, maintain a running count that both functions can update
with just the number of SGEs they have added to the SGE array.
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: Anna Schumaker <Anna.Schumaker@Netapp.com>
|
|
We need to decode and save the incoming rdma_credits field _after_
we know that the direction of the message is "forward direction
Reply". Otherwise, the credits value in reverse direction Calls is
also used to update the forward direction credits.
It is safe to decode the rdma_credits field in rpcrdma_reply_handler
now that rpcrdma_reply_handler is single-threaded. Receives complete
in the same order as they were sent on the NFS server.
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: Anna Schumaker <Anna.Schumaker@Netapp.com>
|
|
I noticed that the soft IRQ thread looked pretty busy under heavy
I/O workloads. perf suggested one area that was expensive was the
queue_work() call in rpcrdma_wc_receive. That gave me some ideas.
Instead of scheduling a separate worker to process RPC Replies,
promote the Receive completion handler to IB_POLL_WORKQUEUE, and
invoke rpcrdma_reply_handler directly.
Note that the poll workqueue is single-threaded. In order to keep
memory invalidation from serializing all RPC Replies, handle any
necessary invalidation tasks in a separate multi-threaded workqueue.
This provides a two-tier scheme, similar to OS I/O interrupt
handlers: A fast interrupt handler that schedules the slow handler
and re-enables the interrupt, and a slower handler that is invoked
for any needed heavy lifting.
Benefits include:
- One less context switch for RPCs that don't register memory
- Receive completion handling is moved out of soft IRQ context to
make room for other users of soft IRQ
- The same CPU core now DMA syncs and XDR decodes the Receive buffer
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: Anna Schumaker <Anna.Schumaker@Netapp.com>
|
|
Clean up: I'd like to be able to invoke the tail of
rpcrdma_reply_handler in two different places. Split the tail out
into its own helper function.
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: Anna Schumaker <Anna.Schumaker@Netapp.com>
|
