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 431f6eb357 ("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>
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
-----BEGIN PGP SIGNATURE-----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=SxSD
-----END PGP SIGNATURE-----
Merge tag 'nfs-rdma-for-4.20-1' of 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 e2ac236c0b ("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 ce7c252a8c ("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 e7ce710a88 ("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>
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
-----BEGIN PGP SIGNATURE-----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=YWSl
-----END PGP SIGNATURE-----
Merge tag 'nfs-for-4.18-1' of git://git.linux-nfs.org/projects/trondmy/linux-nfs
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 5f83d86cf5 ("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 ab03eff58e ("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: 7a89f9c626 ("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>
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 ccede75985 ("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 ccede75985 ("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: ccede75985 ("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 16f906d66c ("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 ae72950abf ("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: 16f906d66c ("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: 9d6b040978 ("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>
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 d8f532d20e ("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: d8f532d20e ("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 655fec6987 ("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>
Clean up: Make it easier to pass the decoded XID, vers, credits, and
proc fields around by moving these variables into struct rpcrdma_rep.
Note: the credits field will be handled in a subsequent patch.
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: Anna Schumaker <Anna.Schumaker@Netapp.com>
A reply with an unrecognized value in the version field means the
transport header is potentially garbled and therefore all the fields
are untrustworthy.
Fixes: 59aa1f9a3c ("xprtrdma: Properly handle RDMA_ERROR ... ")
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: Anna Schumaker <Anna.Schumaker@Netapp.com>
Adopt the use of xprt_pin_rqst to eliminate contention between
Call-side users of rb_lock and the use of rb_lock in
rpcrdma_reply_handler.
This replaces the mechanism introduced in 431af645cf ("xprtrdma:
Fix client lock-up after application signal fires").
Use recv_lock to quickly find the completing rqst, pin it, then
drop the lock. At that point invalidation and pull-up of the Reply
XDR can be done. Both are often expensive operations.
Finally, take recv_lock again to signal completion to the RPC
layer. It also protects adjustment of "cwnd".
This greatly reduces the amount of time a lock is held by the
reply handler. Comparing lock_stat results shows a marked decrease
in contention on rb_lock and recv_lock.
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
[trond.myklebust@primarydata.com: Remove call to rpcrdma_buffer_put() from
the "out_norqst:" path in rpcrdma_reply_handler.]
Signed-off-by: Trond Myklebust <trond.myklebust@primarydata.com>
This further reduces contention with the transport_lock, and allows us
to convert to using a non-bh-safe spinlock, since the list is now never
accessed from a bh context.
Signed-off-by: Trond Myklebust <trond.myklebust@primarydata.com>
Re-arrange the pointer arithmetic in the chunk list encoders to
eliminate several more integer multiplication instructions during
Transport Header encoding.
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: Anna Schumaker <Anna.Schumaker@Netapp.com>
Re-arrange the pointer arithmetic in rpcrdma_convert_iovs() to
eliminate several integer multiplication instructions during
Transport Header encoding.
Also, array overflow does not occur outside development
environments, so replace overflow checking with one spot check
at the end. This reduces the number of conditional branches in
the common case.
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: Anna Schumaker <Anna.Schumaker@Netapp.com>
While marshaling chunk lists which are variable-length XDR objects,
check for XDR buffer overflow at every step. Measurements show no
significant changes in CPU utilization.
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: Anna Schumaker <Anna.Schumaker@Netapp.com>
Initialize an xdr_stream at the top of rpcrdma_marshal_req(), and
use it to encode the fixed transport header fields. This xdr_stream
will be used to encode the chunk lists in a subsequent patch.
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: Anna Schumaker <Anna.Schumaker@Netapp.com>
Clean up: Remove a variable whose result is no longer used.
Commit 655fec6987 ("xprtrdma: Use gathered Send for large inline
messages") should have removed it.
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: Anna Schumaker <Anna.Schumaker@Netapp.com>
Clean up: The caller already has rpcrdma_xprt, so pass that directly
instead. And provide a documenting comment for this critical
function.
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: Anna Schumaker <Anna.Schumaker@Netapp.com>
This field is no longer used outside the Receive completion handler.
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: Anna Schumaker <Anna.Schumaker@Netapp.com>
Clean up chunk list decoding by using the xdr_stream set up in
rpcrdma_reply_handler. This hardens decoding by checking for buffer
overflow at every step while unmarshaling variable-length XDR
objects.
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: Anna Schumaker <Anna.Schumaker@Netapp.com>
Refactor the reply handler's transport header decoding logic to make
it easier to understand and update.
Convert some of the handler to use xdr_streams, which will enable
stricter validation of input data and enable the eventual addition
of support for new combinations of chunks, such as "Write + Reply"
or "PZRC + normal Read".
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: Anna Schumaker <Anna.Schumaker@Netapp.com>
Transport header decoding deals with untrusted input data, therefore
decoding this header needs to be hardened.
Adopt the same infrastructure that is used when XDR decoding NFS
replies. This is slightly more CPU-intensive than the replaced code,
but we're not adding new atomics, locking, or context switches. The
cost is manageable.
Start by initializing an xdr_stream in rpcrdma_reply_handler().
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: Anna Schumaker <Anna.Schumaker@Netapp.com>
After a signal, the RPC client aborts synchronous RPCs running on
behalf of the signaled application.
The server is still executing those RPCs, and will write the results
back into the client's memory when it's done. By the time the server
writes the results, that memory is likely being used for other
purposes. Therefore xprtrdma has to immediately invalidate all
memory regions used by those aborted RPCs to prevent the server's
writes from clobbering that re-used memory.
With FMR memory registration, invalidation takes a relatively long
time. In fact, the invalidation is often still running when the
server tries to write the results into the memory regions that are
being invalidated.
This sets up a race between two processes:
1. After the signal, xprt_rdma_free calls ro_unmap_safe.
2. While ro_unmap_safe is still running, the server replies and
rpcrdma_reply_handler runs, calling ro_unmap_sync.
Both processes invoke ib_unmap_fmr on the same FMR.
The mlx4 driver allows two ib_unmap_fmr calls on the same FMR at
the same time, but HCAs generally don't tolerate this. Sometimes
this can result in a system crash.
If the HCA happens to survive, rpcrdma_reply_handler continues. It
removes the rpc_rqst from rq_list and releases the transport_lock.
This enables xprt_rdma_free to run in another process, and the
rpc_rqst is released while rpcrdma_reply_handler is still waiting
for the ib_unmap_fmr call to finish.
But further down in rpcrdma_reply_handler, the transport_lock is
taken again, and "rqst" is dereferenced. If "rqst" has already been
released, this triggers a general protection fault. Since bottom-
halves are disabled, the system locks up.
Address both issues by reversing the order of the xprt_lookup_rqst
call and the ro_unmap_sync call. Introduce a separate lookup
mechanism for rpcrdma_req's to enable calling ro_unmap_sync before
xprt_lookup_rqst. Now the handler takes the transport_lock once
and holds it for the XID lookup and RPC completion.
BugLink: https://bugzilla.linux-nfs.org/show_bug.cgi?id=305
Fixes: 68791649a7 ('xprtrdma: Invalidate in the RPC reply ... ')
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: Anna Schumaker <Anna.Schumaker@Netapp.com>
There are rare cases where an rpcrdma_req can be re-used (via
rpcrdma_buffer_put) while the RPC reply handler is still running.
This is due to a signal firing at just the wrong instant.
Since commit 9d6b040978 ("xprtrdma: Place registered MWs on a
per-req list"), rpcrdma_mws are self-contained; ie., they fully
describe an MR and scatterlist, and no part of that information is
stored in struct rpcrdma_req.
As part of closing the above race window, pass only the req's list
of registered MRs to ro_unmap_sync, rather than the rpcrdma_req
itself.
Some extra transport header sanity checking is removed. Since the
client depends on its own recollection of what memory had been
registered, there doesn't seem to be a way to abuse this change.
And, the check was not terribly effective. If the client had sent
Read chunks, the "list_empty" test is negative in both of the
removed cases, which are actually looking for Write or Reply
chunks.
BugLink: https://bugzilla.linux-nfs.org/show_bug.cgi?id=305
Fixes: 68791649a7 ('xprtrdma: Invalidate in the RPC reply ... ')
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: Anna Schumaker <Anna.Schumaker@Netapp.com>
There are rare cases where an rpcrdma_req and its matched
rpcrdma_rep can be re-used, via rpcrdma_buffer_put, while the RPC
reply handler is still using that req. This is typically due to a
signal firing at just the wrong instant.
As part of closing this race window, avoid using the wrong
rpcrdma_rep to detect remotely invalidated MRs. Mark MRs as
invalidated while we are sure the rep is still OK to use.
BugLink: https://bugzilla.linux-nfs.org/show_bug.cgi?id=305
Fixes: 68791649a7 ('xprtrdma: Invalidate in the RPC reply ... ')
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: Anna Schumaker <Anna.Schumaker@Netapp.com>
When ro_map is out of buffers, that's not a permanent error, so
don't report a problem.
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: Anna Schumaker <Anna.Schumaker@Netapp.com>
When the underlying device driver is reloaded, ia->ri_device will be
replaced. All cached copies of that device pointer have to be
updated as well.
Commit 54cbd6b0c6 ("xprtrdma: Delay DMA mapping Send and Receive
buffers") added the rg_device field to each regbuf. As part of
handling a device removal, rpcrdma_dma_unmap_regbuf is invoked on
all regbufs for a transport.
Simply calling rpcrdma_dma_map_regbuf for each Receive buffer after
the driver has been reloaded should reinitialize rg_device correctly
for every case except rpcrdma_wc_receive, which still uses
rpcrdma_rep::rr_device.
Ensure the same device that was used to map a Receive buffer is also
used to sync it in rpcrdma_wc_receive by using rg_device there
instead of rr_device.
This is the only use of rr_device, so it can be removed.
The use of regbufs in the send path is also updated, for
completeness.
Fixes: 54cbd6b0c6 ("xprtrdma: Delay DMA mapping Send and ... ")
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: Anna Schumaker <Anna.Schumaker@Netapp.com>
Sriharsha (sriharsha.basavapatna@broadcom.com) reports an occasional
double DMA unmap of an FRWR MR when a connection is lost. I see one
way this can happen.
When a request requires more than one segment or chunk,
rpcrdma_marshal_req loops, invoking ->frwr_op_map for each segment
(MR) in each chunk. Each call posts a FASTREG Work Request to
register one MR.
Now suppose that the transport connection is lost part-way through
marshaling this request. As part of recovering and resetting that
req, rpcrdma_marshal_req invokes ->frwr_op_unmap_safe, which hands
all the req's registered FRWRs to the MR recovery thread.
But note: FRWR registration is asynchronous. So it's possible that
some of these "already registered" FRWRs are fully registered, and
some are still waiting for their FASTREG WR to complete.
When the connection is lost, the "already registered" frmrs are
marked FRMR_IS_VALID, and the "still waiting" WRs flush. Then
frwr_wc_fastreg marks these frmrs FRMR_FLUSHED_FR.
But thanks to ->frwr_op_unmap_safe, the MR recovery thread is doing
an unreg / alloc_mr, a DMA unmap, and marking each of these frwrs
FRMR_IS_INVALID, at the same time frwr_wc_fastreg might be running.
- If the recovery thread runs last, then the frmr is marked
FRMR_IS_INVALID, and life continues.
- If frwr_wc_fastreg runs last, the frmr is marked FRMR_FLUSHED_FR,
but the recovery thread has already DMA unmapped that MR. When
->frwr_op_map later re-uses this frmr, it sees it is not marked
FRMR_IS_INVALID, and tries to recover it before using it, resulting
in a second DMA unmap of the same MR.
The fix is to guarantee in-flight FASTREG WRs have flushed before MR
recovery runs on those FRWRs. Thus we depend on ro_unmap_safe
(called from xprt_rdma_send_request on retransmit, or from
xprt_rdma_free) to clean up old registrations as needed.
Reported-by: Sriharsha Basavapatna <sriharsha.basavapatna@broadcom.com>
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Tested-by: Sriharsha Basavapatna <sriharsha.basavapatna@broadcom.com>
Signed-off-by: Anna Schumaker <Anna.Schumaker@Netapp.com>
The MAX_SEND_SGES check introduced in commit 655fec6987
("xprtrdma: Use gathered Send for large inline messages") fails
for devices that have a small max_sge.
Instead of checking for a large fixed maximum number of SGEs,
check for a minimum small number. RPC-over-RDMA will switch to
using a Read chunk if an xdr_buf has more pages than can fit in
the device's max_sge limit. This is considerably better than
failing all together to mount the server.
This fix supports devices that have as few as three send SGEs
available.
Reported-by: Selvin Xavier <selvin.xavier@broadcom.com>
Reported-by: Devesh Sharma <devesh.sharma@broadcom.com>
Reported-by: Honggang Li <honli@redhat.com>
Reported-by: Ram Amrani <Ram.Amrani@cavium.com>
Fixes: 655fec6987 ("xprtrdma: Use gathered Send for large ...")
Cc: stable@vger.kernel.org # v4.9+
Tested-by: Honggang Li <honli@redhat.com>
Tested-by: Ram Amrani <Ram.Amrani@cavium.com>
Tested-by: Steve Wise <swise@opengridcomputing.com>
Reviewed-by: Parav Pandit <parav@mellanox.com>
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: Anna Schumaker <Anna.Schumaker@Netapp.com>
Pad optimization is changed by echoing into
/proc/sys/sunrpc/rdma_pad_optimize. This is a global setting,
affecting all RPC-over-RDMA connections to all servers.
The marshaling code picks up that value and uses it for decisions
about how to construct each RPC-over-RDMA frame. Having it change
suddenly in mid-operation can result in unexpected failures. And
some servers a client mounts might need chunk round-up, while
others don't.
So instead, copy the pad_optimize setting into each connection's
rpcrdma_ia when the transport is created, and use the copy, which
can't change during the life of the connection, instead.
This also removes a hack: rpcrdma_convert_iovs was using
the remote-invalidation-expected flag to predict when it could leave
out Write chunk padding. This is because the Linux server handles
implicit XDR padding on Write chunks correctly, and only Linux
servers can set the connection's remote-invalidation-expected flag.
It's more sensible to use the pad optimization setting instead.
Fixes: 677eb17e94 ("xprtrdma: Fix XDR tail buffer marshalling")
Cc: stable@vger.kernel.org # v4.9+
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: Anna Schumaker <Anna.Schumaker@Netapp.com>
When pad optimization is disabled, rpcrdma_convert_iovs still
does not add explicit XDR round-up padding to a Read chunk.
Commit 677eb17e94 ("xprtrdma: Fix XDR tail buffer marshalling")
incorrectly short-circuited the test for whether round-up padding
is needed that appears later in rpcrdma_convert_iovs.
However, if this is indeed a regular Read chunk (and not a
Position-Zero Read chunk), the tail iovec _always_ contains the
chunk's padding, and never anything else.
So, it's easy to just skip the tail when padding optimization is
enabled, and add the tail in a subsequent Read chunk segment, if
disabled.
Fixes: 677eb17e94 ("xprtrdma: Fix XDR tail buffer marshalling")
Cc: stable@vger.kernel.org # v4.9+
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: Anna Schumaker <Anna.Schumaker@Netapp.com>
Clean up: offset and handle should be zero-filled, just like in the
chunk encoders.
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: Anna Schumaker <Anna.Schumaker@Netapp.com>
Clean up: the extra layer of indirection doesn't add value.
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: Anna Schumaker <Anna.Schumaker@Netapp.com>
An RPC Call message that is sent inline but that has a data payload
(ie, one or more items in rq_snd_buf's page list) must be "pulled
up:"
- call_allocate has to reserve enough RPC Call buffer space to
accommodate the data payload
- call_transmit has to memcopy the rq_snd_buf's page list and tail
into its head iovec before it is sent
As the inline threshold is increased beyond its current 1KB default,
however, this means data payloads of more than a few KB are copied
by the host CPU. For example, if the inline threshold is increased
just to 4KB, then NFS WRITE requests up to 4KB would involve a
memcpy of the NFS WRITE's payload data into the RPC Call buffer.
This is an undesirable amount of participation by the host CPU.
The inline threshold may be much larger than 4KB in the future,
after negotiation with a peer server.
Instead of copying the components of rq_snd_buf into its head iovec,
construct a gather list of these components, and send them all in
place. The same approach is already used in the Linux server's
RPC-over-RDMA reply path.
This mechanism also eliminates the need for rpcrdma_tail_pullup,
which is used to manage the XDR pad and trailing inline content when
a Read list is present.
This requires that the pages in rq_snd_buf's page list be DMA-mapped
during marshaling, and unmapped when a data-bearing RPC is
completed. This is slightly less efficient for very small I/O
payloads, but significantly more efficient as data payload size and
inline threshold increase past a kilobyte.
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: Anna Schumaker <Anna.Schumaker@Netapp.com>
Have frwr's ro_unmap_sync recognize an invalidated rkey that appears
as part of a Receive completion. Local invalidation can be skipped
for that rkey.
Use an out-of-band signaling mechanism to indicate to the server
that the client is prepared to receive RDMA Send With Invalidate.
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: Anna Schumaker <Anna.Schumaker@Netapp.com>
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: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: Anna Schumaker <Anna.Schumaker@Netapp.com>
Clean up: Most of the fields in each send_wr do not vary. There is
no need to initialize them before each ib_post_send(). This removes
a large-ish data structure from the stack.
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: Anna Schumaker <Anna.Schumaker@Netapp.com>
Clean up.
Since commit fc66448549 ("xprtrdma: Split the completion queue"),
rpcrdma_ep_post_recv() no longer uses the "ep" argument.
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: Anna Schumaker <Anna.Schumaker@Netapp.com>
Currently, each regbuf is allocated and DMA mapped at the same time.
This is done during transport creation.
When a device driver is unloaded, every DMA-mapped buffer in use by
a transport has to be unmapped, and then remapped to the new
device if the driver is loaded again. Remapping will have to be done
_after_ the connect worker has set up the new device.
But there's an ordering problem:
call_allocate, which invokes xprt_rdma_allocate which calls
rpcrdma_alloc_regbuf to allocate Send buffers, happens _before_
the connect worker can run to set up the new device.
Instead, at transport creation, allocate each buffer, but leave it
unmapped. Once the RPC carries these buffers into ->send_request, by
which time a transport connection should have been established,
check to see that the RPC's buffers have been DMA mapped. If not,
map them there.
When device driver unplug support is added, it will simply unmap all
the transport's regbufs, but it doesn't have to deallocate the
underlying memory.
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: Anna Schumaker <Anna.Schumaker@Netapp.com>
Clean up: r_xprt is already available everywhere these macros are
invoked, so just dereference that directly.
RPCRDMA_INLINE_PAD_VALUE is no longer used, so it can simply be
removed.
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: Anna Schumaker <Anna.Schumaker@Netapp.com>
Direct data placement is not allowed when using flavors that
guarantee integrity or privacy. When such security flavors are in
effect, don't allow the use of Read and Write chunks for moving
individual data items. All messages larger than the inline threshold
are sent via Long Call or Long Reply.
On my systems (CX-3 Pro on FDR), for small I/O operations, the use
of Long messages adds only around 5 usecs of latency in each
direction.
Note that when integrity or encryption is used, the host CPU touches
every byte in these messages. Even if it could be used, data
movement offload doesn't buy much in this case.
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Tested-by: Steve Wise <swise@opengridcomputing.com>
Signed-off-by: Anna Schumaker <Anna.Schumaker@Netapp.com>
fixup_copy_count should count only the number of bytes copied to the
page list. The head and tail are now always handled without a data
copy.
And the debugging at the end of rpcrdma_inline_fixup() is also no
longer necessary, since copy_len will be non-zero when there is reply
data in the tail (a normal and valid case).
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Tested-by: Steve Wise <swise@opengridcomputing.com>
Signed-off-by: Anna Schumaker <Anna.Schumaker@Netapp.com>
Now that rpcrdma_inline_fixup() updates only two fields in
rq_rcv_buf, a full memcpy of that structure to rq_private_buf is
unwarranted. Updating rq_private_buf fields only where needed also
better documents what is going on.
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Tested-by: Steve Wise <swise@opengridcomputing.com>
Signed-off-by: Anna Schumaker <Anna.Schumaker@Netapp.com>
While trying NFSv4.0/RDMA with sec=krb5p, I noticed small NFS READ
operations failed. After the client unwrapped the NFS READ reply
message, the NFS READ XDR decoder was not able to decode the reply.
The message was "Server cheating in reply", with the reported
number of received payload bytes being zero. Applications reported
a read(2) that returned -1/EIO.
The problem is rpcrdma_inline_fixup() sets the tail.iov_len to zero
when the incoming reply fits entirely in the head iovec. The zero
tail.iov_len confused xdr_buf_trim(), which then mangled the actual
reply data instead of simply removing the trailing GSS checksum.
As near as I can tell, RPC transports are not supposed to update the
head.iov_len, page_len, or tail.iov_len fields in the receive XDR
buffer when handling an incoming RPC reply message. These fields
contain the length of each component of the XDR buffer, and hence
the maximum number of bytes of reply data that can be stored in each
XDR buffer component. I've concluded this because:
- This is how xdr_partial_copy_from_skb() appears to behave
- rpcrdma_inline_fixup() already does not alter page_len
- call_decode() compares rq_private_buf and rq_rcv_buf and WARNs
if they are not exactly the same
Unfortunately, as soon as I tried the simple fix to just remove the
line that sets tail.iov_len to zero, I saw that the logic that
appends the implicit Write chunk pad inline depends on inline_fixup
setting tail.iov_len to zero.
To address this, re-organize the tail iovec handling logic to use
the same approach as with the head iovec: simply point tail.iov_base
to the correct bytes in the receive buffer.
While I remember all this, write down the conclusion in documenting
comments.
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Tested-by: Steve Wise <swise@opengridcomputing.com>
Signed-off-by: Anna Schumaker <Anna.Schumaker@Netapp.com>
When the remaining length of an incoming reply is longer than the
XDR buf's page_len, switch over to the tail iovec instead of
copying more than page_len bytes into the page list.
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Tested-by: Steve Wise <swise@opengridcomputing.com>
Signed-off-by: Anna Schumaker <Anna.Schumaker@Netapp.com>
Currently, all three chunk list encoders each use a portion of the
one rl_segments array in rpcrdma_req. This is because the MWs for
each chunk list were preserved in rl_segments so that ro_unmap could
find and invalidate them after the RPC was complete.
However, now that MWs are placed on a per-req linked list as they
are registered, there is no longer any information in rpcrdma_mr_seg
that is shared between ro_map and ro_unmap_{sync,safe}, and thus
nothing in rl_segments needs to be preserved after
rpcrdma_marshal_req is complete.
Thus the rl_segments array can be used now just for the needs of
each rpcrdma_convert_iovs call. Once each chunk list is encoded, the
next chunk list encoder is free to re-use all of rl_segments.
This means all three chunk lists in one RPC request can now each
encode a full size data payload with no increase in the size of
rl_segments.
This is a key requirement for Kerberos support, since both the Call
and Reply for a single RPC transaction are conveyed via Long
messages (RDMA Read/Write). Both can be large.
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Tested-by: Steve Wise <swise@opengridcomputing.com>
Signed-off-by: Anna 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: Chuck Lever <chuck.lever@oracle.com>
Tested-by: Steve Wise <swise@opengridcomputing.com>
Signed-off-by: Anna 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: Chuck Lever <chuck.lever@oracle.com>
Tested-by: Steve Wise <swise@opengridcomputing.com>
Signed-off-by: Anna Schumaker <Anna.Schumaker@Netapp.com>
Commit c93c62231c ("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: Chuck Lever <chuck.lever@oracle.com>
Tested-by: Steve Wise <swise@opengridcomputing.com>
Signed-off-by: Anna 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: Chuck Lever <chuck.lever@oracle.com>
Tested-by: Steve Wise <swise@opengridcomputing.com>
Signed-off-by: Anna Schumaker <Anna.Schumaker@Netapp.com>
rpcrdma_create_chunks() has been replaced, and can be removed.
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Tested-by: Steve Wise <swise@opengridcomputing.com>
Reviewed-by: Sagi Grimberg <sagi@grimberg.me>
Signed-off-by: Anna Schumaker <Anna.Schumaker@Netapp.com>
rpcrdma_marshal_req() makes a simplifying assumption: that NFS
operations with large Call messages have small Reply messages, and
vice versa. Therefore with RPC-over-RDMA, only one chunk type is
ever needed for each Call/Reply pair, because one direction needs
chunks, the other direction will always fit inline.
In fact, this assumption is asserted in the code:
if (rtype != rpcrdma_noch && wtype != rpcrdma_noch) {
dprintk("RPC: %s: cannot marshal multiple chunk lists\n",
__func__);
return -EIO;
}
But RPCGSS_SEC breaks this assumption. Because krb5i and krb5p
perform data transformation on RPC messages before they are
transmitted, direct data placement techniques cannot be used, thus
RPC messages must be sent via a Long call in both directions.
All such calls are sent with a Position Zero Read chunk, and all
such replies are handled with a Reply chunk. Thus the client must
provide every Call/Reply pair with both a Read list and a Reply
chunk.
Without any special security in effect, NFSv4 WRITEs may now also
use the Read list and provide a Reply chunk. The marshal_req
logic was preventing that, meaning an NFSv4 WRITE with a large
payload that included a GETATTR result larger than the inline
threshold would fail.
The code that encodes each chunk list is now completely contained in
its own function. There is some code duplication, but the trade-off
is that the overall logic should be more clear.
Note that all three chunk lists now share the rl_segments array.
Some additional per-req accounting is necessary to track this
usage. For the same reasons that the above simplifying assumption
has held true for so long, I don't expect more array elements are
needed at this time.
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Tested-by: Steve Wise <swise@opengridcomputing.com>
Reviewed-by: Sagi Grimberg <sagi@grimberg.me>
Signed-off-by: Anna Schumaker <Anna.Schumaker@Netapp.com>
Update documenting comments to reflect code changes over the past
year.
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Tested-by: Steve Wise <swise@opengridcomputing.com>
Reviewed-by: Sagi Grimberg <sagi@grimberg.me>
Signed-off-by: Anna Schumaker <Anna.Schumaker@Netapp.com>
Avoid the latency and interrupt overhead of registering a Write
chunk when handling NFS READ requests of a few hundred bytes or
less.
This change does not interoperate with Linux NFS/RDMA servers
that do not have commit 9d11b51ce7 ('svcrdma: Fix send_reply()
scatter/gather set-up'). Commit 9d11b51ce7 was introduced in v4.3,
and is included in 4.2.y, 4.1.y, and 3.18.y.
Oracle bug 22925946 has been filed to request that the above fix
be included in the Oracle Linux UEK4 NFS/RDMA server.
Red Hat bugzillas 1327280 and 1327554 have been filed to request
that RHEL NFS/RDMA server backports include the above fix.
Workaround: Replace the "proto=rdma,port=20049" mount options
with "proto=tcp" until commit 9d11b51ce7 is applied to your
NFS server.
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Tested-by: Steve Wise <swise@opengridcomputing.com>
Reviewed-by: Sagi Grimberg <sagi@grimberg.me>
Signed-off-by: Anna Schumaker <Anna.Schumaker@Netapp.com>
