Commit 9ed5af268e ("SUNRPC: Clean up the handling of page padding
in rpc_prepare_reply_pages()") [Dec 2020] affects RPC Replies that
have a data payload (i.e., Write chunks).
rpcrdma_prepare_readch(), as its name suggests, sets up Read chunks
which are data payloads within RPC Calls. Those payloads are
constructed by xdr_write_pages(), which continues to stuff the call
buffer's tail kvec with the payload's XDR roundup. Thus removing
the tail buffer logic in rpcrdma_prepare_readch() was the wrong
thing to do.
Fixes: 586a0787ce ("xprtrdma: Clean up rpcrdma_prepare_readch()")
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: Trond Myklebust <trond.myklebust@hammerspace.com>
Ensure that we fix the XPRT_CONGESTED starvation issue for RDMA as well
as socket based transports.
Ensure we always initialise the request after waking up from the backlog
list.
Fixes: e877a88d1f ("SUNRPC in case of backlog, hand free slots directly to waiting task")
Signed-off-by: Trond Myklebust <trond.myklebust@hammerspace.com>
If a disconnection occurs while we're trying to reply to a server
callback, then we may end up calling xs_tcp_send_request() with a NULL
value for transport->inet, which trips up the call to
tcp_sock_set_cork().
Fixes: d737e5d418 ("SUNRPC: Set TCP_CORK until the transmit queue is empty")
Cc: stable@vger.kernel.org
Signed-off-by: Trond Myklebust <trond.myklebust@hammerspace.com>
If sunrpc.tcp_max_slot_table_entries is small and there are tasks
on the backlog queue, then when a request completes it is freed and the
first task on the queue is woken. The expectation is that it will wake
and claim that request. However if it was a sync task and the waiting
process was killed at just that moment, it will wake and NOT claim the
request.
As long as TASK_CONGESTED remains set, requests can only be claimed by
tasks woken from the backlog, and they are woken only as requests are
freed, so when a task doesn't claim a request, no other task can ever
get that request until TASK_CONGESTED is cleared. Each time this
happens the number of available requests is decreased by one.
With a sufficiently high workload and sufficiently low setting of
max_slot (16 in the case where this was seen), TASK_CONGESTED can remain
set for an extended period, and the above scenario (of a process being
killed just as its task was woken) can repeat until no requests can be
allocated. Then traffic stops.
This patch addresses the problem by introducing a positive handover of a
request from a completing task to a backlog task - the request is never
freed when there is a backlog.
When a task is woken it might not already have a request attached in
which case it is *not* freed (as with current code) but is initialised
(if needed) and used. If it isn't used it will eventually be freed by
rpc_exit_task(). xprt_release() is enhanced to be able to correctly
release an uninitialised request.
Fixes: ba60eb25ff ("SUNRPC: Fix a livelock problem in the xprt->backlog queue")
Signed-off-by: NeilBrown <neilb@suse.de>
Signed-off-by: Trond Myklebust <trond.myklebust@hammerspace.com>
Highlights include:
Stable fixes:
- Add validation of the UDP retrans parameter to prevent shift out-of-bounds
- Don't discard pNFS layout segments that are marked for return
Bugfixes:
- Fix a NULL dereference crash in xprt_complete_bc_request() when the
NFSv4.1 server misbehaves.
- Fix the handling of NFS READDIR cookie verifiers
- Sundry fixes to ensure attribute revalidation works correctly when the
server does not return post-op attributes.
- nfs4_bitmask_adjust() must not change the server global bitmasks
- Fix major timeout handling in the RPC code.
- NFSv4.2 fallocate() fixes.
- Fix the NFSv4.2 SEEK_HOLE/SEEK_DATA end-of-file handling
- Copy offload attribute revalidation fixes
- Fix an incorrect filehandle size check in the pNFS flexfiles driver
- Fix several RDMA transport setup/teardown races
- Fix several RDMA queue wrapping issues
- Fix a misplaced memory read barrier in sunrpc's call_decode()
Features:
- Micro optimisation of the TCP transmission queue using TCP_CORK
- statx() performance improvements by further splitting up the tracking
of invalid cached file metadata.
- Support the NFSv4.2 "change_attr_type" attribute and use it to
optimise handling of change attribute updates.
-----BEGIN PGP SIGNATURE-----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=h01b
-----END PGP SIGNATURE-----
Merge tag 'nfs-for-5.13-1' of git://git.linux-nfs.org/projects/trondmy/linux-nfs
Pull NFS client updates from Trond Myklebust:
"Highlights include:
Stable fixes:
- Add validation of the UDP retrans parameter to prevent shift
out-of-bounds
- Don't discard pNFS layout segments that are marked for return
Bugfixes:
- Fix a NULL dereference crash in xprt_complete_bc_request() when the
NFSv4.1 server misbehaves.
- Fix the handling of NFS READDIR cookie verifiers
- Sundry fixes to ensure attribute revalidation works correctly when
the server does not return post-op attributes.
- nfs4_bitmask_adjust() must not change the server global bitmasks
- Fix major timeout handling in the RPC code.
- NFSv4.2 fallocate() fixes.
- Fix the NFSv4.2 SEEK_HOLE/SEEK_DATA end-of-file handling
- Copy offload attribute revalidation fixes
- Fix an incorrect filehandle size check in the pNFS flexfiles driver
- Fix several RDMA transport setup/teardown races
- Fix several RDMA queue wrapping issues
- Fix a misplaced memory read barrier in sunrpc's call_decode()
Features:
- Micro optimisation of the TCP transmission queue using TCP_CORK
- statx() performance improvements by further splitting up the
tracking of invalid cached file metadata.
- Support the NFSv4.2 'change_attr_type' attribute and use it to
optimise handling of change attribute updates"
* tag 'nfs-for-5.13-1' of git://git.linux-nfs.org/projects/trondmy/linux-nfs: (85 commits)
xprtrdma: Fix a NULL dereference in frwr_unmap_sync()
sunrpc: Fix misplaced barrier in call_decode
NFSv4.2: Remove ifdef CONFIG_NFSD from NFSv4.2 client SSC code.
xprtrdma: Move fr_mr field to struct rpcrdma_mr
xprtrdma: Move the Work Request union to struct rpcrdma_mr
xprtrdma: Move fr_linv_done field to struct rpcrdma_mr
xprtrdma: Move cqe to struct rpcrdma_mr
xprtrdma: Move fr_cid to struct rpcrdma_mr
xprtrdma: Remove the RPC/RDMA QP event handler
xprtrdma: Don't display r_xprt memory addresses in tracepoints
xprtrdma: Add an rpcrdma_mr_completion_class
xprtrdma: Add tracepoints showing FastReg WRs and remote invalidation
xprtrdma: Avoid Send Queue wrapping
xprtrdma: Do not wake RPC consumer on a failed LocalInv
xprtrdma: Do not recycle MR after FastReg/LocalInv flushes
xprtrdma: Clarify use of barrier in frwr_wc_localinv_done()
xprtrdma: Rename frwr_release_mr()
xprtrdma: rpcrdma_mr_pop() already does list_del_init()
xprtrdma: Delete rpcrdma_recv_buffer_put()
xprtrdma: Fix cwnd update ordering
...
including a fix to grant read delegations for files open for
writing.
-----BEGIN PGP SIGNATURE-----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=HB+N
-----END PGP SIGNATURE-----
Merge tag 'nfsd-5.13-1' of git://git.kernel.org/pub/scm/linux/kernel/git/cel/linux
Pull more nfsd updates from Chuck Lever:
"Additional fixes and clean-ups for NFSD since tags/nfsd-5.13,
including a fix to grant read delegations for files open for writing"
* tag 'nfsd-5.13-1' of git://git.kernel.org/pub/scm/linux/kernel/git/cel/linux:
SUNRPC: Fix null pointer dereference in svc_rqst_free()
SUNRPC: fix ternary sign expansion bug in tracing
nfsd: Fix fall-through warnings for Clang
nfsd: grant read delegations to clients holding writes
nfsd: reshuffle some code
nfsd: track filehandle aliasing in nfs4_files
nfsd: hash nfs4_files by inode number
nfsd: ensure new clients break delegations
nfsd: removed unused argument in nfsd_startup_generic()
nfsd: remove unused function
svcrdma: Pass a useful error code to the send_err tracepoint
svcrdma: Rename goto labels in svc_rdma_sendto()
svcrdma: Don't leak send_ctxt on Send errors
The normal mechanism that invalidates and unmaps MRs is
frwr_unmap_async(). frwr_unmap_sync() is used only when an RPC
Reply bearing Write or Reply chunks has been lost (ie, almost
never).
Coverity found that after commit 9a301cafc8 ("xprtrdma: Move
fr_linv_done field to struct rpcrdma_mr"), the while() loop in
frwr_unmap_sync() exits only once @mr is NULL, unconditionally
causing subsequent dereferences of @mr to Oops.
I've tested this fix by creating a client that skips invoking
frwr_unmap_async() when RPC Replies complete. That forces all
invalidation tasks to fall upon frwr_unmap_sync(). Simple workloads
with this fix applied to the adulterated client work as designed.
Reported-by: coverity-bot <keescook+coverity-bot@chromium.org>
Addresses-Coverity-ID: 1504556 ("Null pointer dereferences")
Fixes: 9a301cafc8 ("xprtrdma: Move fr_linv_done field to struct rpcrdma_mr")
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: Trond Myklebust <trond.myklebust@hammerspace.com>
Fix a misplaced barrier in call_decode. The struct rpc_rqst is modified
as follows by xprt_complete_rqst:
req->rq_private_buf.len = copied;
/* Ensure all writes are done before we update */
/* req->rq_reply_bytes_recvd */
smp_wmb();
req->rq_reply_bytes_recvd = copied;
And currently read as follows by call_decode:
smp_rmb(); // misplaced
if (!req->rq_reply_bytes_recvd)
goto out;
req->rq_rcv_buf.len = req->rq_private_buf.len;
This patch places the smp_rmb after the if to ensure that
rq_reply_bytes_recvd and rq_private_buf.len are read in order.
Fixes: 9ba828861c ("SUNRPC: Don't try to parse incomplete RPC messages")
Signed-off-by: Baptiste Lepers <baptiste.lepers@gmail.com>
Signed-off-by: Trond Myklebust <trond.myklebust@hammerspace.com>
Reduce the rate at which nfsd threads hammer on the page allocator. This
improves throughput scalability by enabling the threads to run more
independently of each other.
[mgorman: Update interpretation of alloc_pages_bulk return value]
Link: https://lkml.kernel.org/r/20210325114228.27719-8-mgorman@techsingularity.net
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: Mel Gorman <mgorman@techsingularity.net>
Reviewed-by: Alexander Lobakin <alobakin@pm.me>
Cc: Alexander Duyck <alexander.duyck@gmail.com>
Cc: Christoph Hellwig <hch@infradead.org>
Cc: David Miller <davem@davemloft.net>
Cc: Ilias Apalodimas <ilias.apalodimas@linaro.org>
Cc: Jesper Dangaard Brouer <brouer@redhat.com>
Cc: Matthew Wilcox (Oracle) <willy@infradead.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Patch series "SUNRPC consumer for the bulk page allocator"
This patch set and the measurements below are based on yesterday's
bulk allocator series:
git://git.kernel.org/pub/scm/linux/kernel/git/mel/linux.git mm-bulk-rebase-v5r9
The patches change SUNRPC to invoke the array-based bulk allocator
instead of alloc_page().
The micro-benchmark results are promising. I ran a mixture of 256KB
reads and writes over NFSv3. The server's kernel is built with KASAN
enabled, so the comparison is exaggerated but I believe it is still
valid.
I instrumented svc_recv() to measure the latency of each call to
svc_alloc_arg() and report it via a trace point. The following results
are averages across the trace events.
Single page: 25.007 us per call over 532,571 calls
Bulk list: 6.258 us per call over 517,034 calls
Bulk array: 4.590 us per call over 517,442 calls
This patch (of 2)
Refactor:
I'm about to use the loop variable @i for something else.
As far as the "i++" is concerned, that is a post-increment. The
value of @i is not used subsequently, so the increment operator
is unnecessary and can be removed.
Also note that nfsd_read_actor() was renamed nfsd_splice_actor()
by commit cf8208d0ea ("sendfile: convert nfsd to
splice_direct_to_actor()").
Link: https://lkml.kernel.org/r/20210325114228.27719-7-mgorman@techsingularity.net
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: Mel Gorman <mgorman@techsingularity.net>
Reviewed-by: Alexander Lobakin <alobakin@pm.me>
Cc: Alexander Duyck <alexander.duyck@gmail.com>
Cc: Christoph Hellwig <hch@infradead.org>
Cc: David Miller <davem@davemloft.net>
Cc: Ilias Apalodimas <ilias.apalodimas@linaro.org>
Cc: Jesper Dangaard Brouer <brouer@redhat.com>
Cc: Matthew Wilcox (Oracle) <willy@infradead.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Clean up: The last remaining field in struct rpcrdma_frwr has been
removed, so the struct can be eliminated.
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: Trond Myklebust <trond.myklebust@hammerspace.com>
Clean up: Move more of struct rpcrdma_frwr into its parent.
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: Trond Myklebust <trond.myklebust@hammerspace.com>
Clean up.
- Simplify variable initialization in the completion handlers.
- Move another field out of struct rpcrdma_frwr.
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: Trond Myklebust <trond.myklebust@hammerspace.com>
Clean up (for several purposes):
- The MR's cid is initialized sooner so that tracepoints can show
something reasonable even if the MR is never posted.
- The MR's res.id doesn't change so the cid won't change either.
Initializing the cid once is sufficient.
- struct rpcrdma_frwr is going away soon.
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: Trond Myklebust <trond.myklebust@hammerspace.com>
Clean up: The handler only recorded a trace event. If indeed no
action is needed by the RPC/RDMA consumer, then the event can be
ignored.
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: Trond Myklebust <trond.myklebust@hammerspace.com>
The Send signaling logic is a little subtle, so add some
observability around it. For every xprtrdma_mr_fastreg event, there
should be an xprtrdma_mr_localinv or xprtrdma_mr_reminv event.
When these tracepoints are enabled, we can see exactly when an MR is
DMA-mapped, registered, invalidated (either locally or remotely) and
then DMA-unmapped.
kworker/u25:2-190 [000] 787.979512: xprtrdma_mr_map: task:351@5 mr.id=4 nents=2 5608@0x8679e0c8f6f56000:0x00000503 (TO_DEVICE)
kworker/u25:2-190 [000] 787.979515: xprtrdma_chunk_read: task:351@5 pos=148 5608@0x8679e0c8f6f56000:0x00000503 (last)
kworker/u25:2-190 [000] 787.979519: xprtrdma_marshal: task:351@5 xid=0x8679e0c8: hdr=52 xdr=148/5608/0 read list/inline
kworker/u25:2-190 [000] 787.979525: xprtrdma_mr_fastreg: task:351@5 mr.id=4 nents=2 5608@0x8679e0c8f6f56000:0x00000503 (TO_DEVICE)
kworker/u25:2-190 [000] 787.979526: xprtrdma_post_send: task:351@5 cq.id=0 cid=73 (2 SGEs)
...
kworker/5:1H-219 [005] 787.980567: xprtrdma_wc_receive: cq.id=1 cid=161 status=SUCCESS (0/0x0) received=164
kworker/5:1H-219 [005] 787.980571: xprtrdma_post_recvs: peer=[192.168.100.55]:20049 r_xprt=0xffff8884974d4000: 0 new recvs, 70 active (rc 0)
kworker/5:1H-219 [005] 787.980573: xprtrdma_reply: task:351@5 xid=0x8679e0c8 credits=64
kworker/5:1H-219 [005] 787.980576: xprtrdma_mr_reminv: task:351@5 mr.id=4 nents=2 5608@0x8679e0c8f6f56000:0x00000503 (TO_DEVICE)
kworker/5:1H-219 [005] 787.980577: xprtrdma_mr_unmap: mr.id=4 nents=2 5608@0x8679e0c8f6f56000:0x00000503 (TO_DEVICE)
Note that I've moved the xprtrdma_post_send tracepoint so that event
always appears after the xprtrdma_mr_fastreg tracepoint. Otherwise
the event log looks counterintuitive (FastReg is always supposed to
happen before Send).
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: Trond Myklebust <trond.myklebust@hammerspace.com>
Send WRs can be signalled or unsignalled. A signalled Send WR
always has a matching Send completion, while a unsignalled Send
has a completion only if the Send WR fails.
xprtrdma has a Send account mechanism that is designed to reduce
the number of signalled Send WRs. This in turn mitigates the
interrupt rate of the underlying device.
RDMA consumers can't leave all Sends unsignaled, however, because
providers rely on Send completions to maintain their Send Queue head
and tail pointers. xprtrdma counts the number of unsignaled Send WRs
that have been posted to ensure that Sends are signalled often
enough to prevent the Send Queue from wrapping.
This mechanism neglected to account for FastReg WRs, which are
posted on the Send Queue but never signalled. As a result, the
Send Queue wrapped on occasion, resulting in duplication completions
of FastReg and LocalInv WRs.
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: Trond Myklebust <trond.myklebust@hammerspace.com>
Throw away any reply where the LocalInv flushes or could not be
posted. The registered memory region is in an unknown state until
the disconnect completes.
rpcrdma_xprt_disconnect() will find and release the MR. No need to
put it back on the MR free list in this case.
The client retransmits pending RPC requests once it reestablishes a
fresh connection, so a replacement reply should be forthcoming on
the next connection instance.
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: Trond Myklebust <trond.myklebust@hammerspace.com>
Better not to touch MRs involved in a flush or post error until the
Send and Receive Queues are drained and the transport is fully
quiescent. Simply don't insert such MRs back onto the free list.
They remain on mr_all and will be released when the connection is
torn down.
I had thought that recycling would prevent hardware resources from
being tied up for a long time. However, since v5.7, a transport
disconnect destroys the QP and other hardware-owned resources. The
MRs get cleaned up nicely at that point.
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: Trond Myklebust <trond.myklebust@hammerspace.com>
Clean up: The comment and the placement of the memory barrier is
confusing. Humans want to read the function statements from head
to tail.
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: Trond Myklebust <trond.myklebust@hammerspace.com>
Clean up: To be consistent with other functions in this source file,
follow the naming convention of putting the object being acted upon
before the action itself.
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: Trond Myklebust <trond.myklebust@hammerspace.com>
The rpcrdma_mr_pop() earlier in the function has already cleared
out mr_list, so it must not be done again in the error path.
Fixes: 847568942f ("xprtrdma: Remove fr_state")
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: Trond Myklebust <trond.myklebust@hammerspace.com>
Clean up: The name recv_buffer_put() is a vestige of older code,
and the function is just a wrapper for the newer rpcrdma_rep_put().
In most of the existing call sites, a pointer to the owning
rpcrdma_buffer is already available.
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: Trond Myklebust <trond.myklebust@hammerspace.com>
After a reconnect, the reply handler is opening the cwnd (and thus
enabling more RPC Calls to be sent) /before/ rpcrdma_post_recvs()
can post enough Receive WRs to receive their replies. This causes an
RNR and the new connection is lost immediately.
The race is most clearly exposed when KASAN and disconnect injection
are enabled. This slows down rpcrdma_rep_create() enough to allow
the send side to post a bunch of RPC Calls before the Receive
completion handler can invoke ib_post_recv().
Fixes: 2ae50ad68c ("xprtrdma: Close window between waking RPC senders and posting Receives")
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: Trond Myklebust <trond.myklebust@hammerspace.com>
Defensive clean up: Protect the rb_all_reps list during rep
creation.
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: Trond Myklebust <trond.myklebust@hammerspace.com>
Currently rpcrdma_reps_destroy() assumes that, at transport
tear-down, the content of the rb_free_reps list is the same as the
content of the rb_all_reps list. Although that is usually true,
using the rb_all_reps list should be more reliable because of
the way it's managed. And, rpcrdma_reps_unmap() uses rb_all_reps;
these two functions should both traverse the "all" list.
Ensure that all rpcrdma_reps are always destroyed whether they are
on the rep free list or not.
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: Trond Myklebust <trond.myklebust@hammerspace.com>
Defer destruction of an rpcrdma_rep until transport tear-down to
preserve the rb_all_reps list while Receives flush.
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Reviewed-by: Tom Talpey <tom@talpey.com>
Signed-off-by: Trond Myklebust <trond.myklebust@hammerspace.com>
Currently the Receive completion handler refreshes the Receive Queue
whenever a successful Receive completion occurs.
On disconnect, xprtrdma drains the Receive Queue. The first few
Receive completions after a disconnect are typically successful,
until the first flushed Receive.
This means the Receive completion handler continues to post more
Receive WRs after the drain sentinel has been posted. The late-
posted Receives flush after the drain sentinel has completed,
leading to a crash later in rpcrdma_xprt_disconnect().
To prevent this crash, xprtrdma has to ensure that the Receive
handler stops posting Receives before ib_drain_rq() posts its
drain sentinel.
Suggested-by: Tom Talpey <tom@talpey.com>
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: Trond Myklebust <trond.myklebust@hammerspace.com>
Commit e340c2d6ef ("xprtrdma: Reduce the doorbell rate (Receive)")
increased the number of Receive WRs that are posted by the client,
but did not increase the size of the Receive Queue allocated during
transport set-up.
This is usually not an issue because RPCRDMA_BACKWARD_WRS is defined
as (32) when SUNRPC_BACKCHANNEL is defined. In cases where it isn't,
there is a real risk of Receive Queue wrapping.
Fixes: e340c2d6ef ("xprtrdma: Reduce the doorbell rate (Receive)")
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Reviewed-by: Tom Talpey <tom@talpey.com>
Signed-off-by: Trond Myklebust <trond.myklebust@hammerspace.com>
When alloc_pages_node() returns null in svc_rqst_alloc(), the
null rq_scratch_page pointer will be dereferenced when calling
put_page() in svc_rqst_free(). Fix it by adding a null check.
Addresses-Coverity: ("Dereference after null check")
Fixes: 5191955d6f ("SUNRPC: Prepare for xdr_stream-style decoding on the server-side")
Signed-off-by: Yunjian Wang <wangyunjian@huawei.com>
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
This code is supposed to pass negative "err" values for tracing but it
passes positive values instead. The problem is that the
trace_svcsock_tcp_send() function takes a long but "err" is an int and
"sent" is a u32. The negative is first type promoted to u32 so it
becomes a high positive then it is promoted to long and it stays
positive.
Fix this by casting "err" directly to long.
Fixes: 998024dee1 ("SUNRPC: Add more svcsock tracepoints")
Signed-off-by: Dan Carpenter <dan.carpenter@oracle.com>
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Capture error codes in @ret, which is passed to the send_err
tracepoint, so that they can be logged when something goes awry.
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Currently if a major timeout value is reached, but the minor value has
not been reached, an ETIMEOUT will not be sent back to the caller.
This can occur if the v4 server is not responding to requests and
retrans is configured larger than the default of two.
For example, A TCP mount with a configured timeout value of 50 and a
retransmission count of 3 to a v4 server which is not responding:
1. Initial value and increment set to 5s, maxval set to 20s, retries at 3
2. Major timeout is set to 20s, minor timeout set to 5s initially
3. xport_adjust_timeout() is called after 5s, retry with 10s timeout,
minor timeout is bumped to 10s
4. And again after another 10s, 15s total time with minor timeout set
to 15s
5. After 20s total time xport_adjust_timeout is called as major timeout is
reached, but skipped because the minor timeout is not reached
- After this time the cpu spins continually calling
xport_adjust_timeout() and returning 0 for 10 seconds.
As seen on perf sched:
39243.913182 [0005] mount.nfs[3794] 4607.938 0.017 9746.863
6. This continues until the 15s minor timeout condition is reached (in
this case for 10 seconds). After which the ETIMEOUT is processed
back to the caller, the cpu spinning stops, and normal operations
continue
Fixes: 7de62bc09f ("SUNRPC dont update timeout value on connection reset")
Signed-off-by: Chris Dion <Christopher.Dion@dell.com>
Signed-off-by: Trond Myklebust <trond.myklebust@hammerspace.com>
This tracepoint can crash when dereferencing snd_task because
when some transports connect, they put a cookie in that field
instead of a pointer to an rpc_task.
BUG: KASAN: use-after-free in trace_event_raw_event_xprt_writelock_event+0x141/0x18e [sunrpc]
Read of size 2 at addr ffff8881a83bd3a0 by task git/331872
CPU: 11 PID: 331872 Comm: git Tainted: G S 5.12.0-rc2-00007-g3ab6e585a7f9 #1453
Hardware name: Supermicro SYS-6028R-T/X10DRi, BIOS 1.1a 10/16/2015
Call Trace:
dump_stack+0x9c/0xcf
print_address_description.constprop.0+0x18/0x239
kasan_report+0x174/0x1b0
trace_event_raw_event_xprt_writelock_event+0x141/0x18e [sunrpc]
xprt_prepare_transmit+0x8e/0xc1 [sunrpc]
call_transmit+0x4d/0xc6 [sunrpc]
Fixes: 9ce07ae5eb ("SUNRPC: Replace dprintk() call site in xprt_prepare_transmit")
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: Trond Myklebust <trond.myklebust@hammerspace.com>
A separate tracepoint can be left enabled all the time to capture
rare but important retransmission events. So for example:
kworker/u26:3-568 [009] 156.967933: xprt_retransmit: task:44093@5 xid=0xa25dbc79 nfsv3 WRITE ntrans=2
Or, for example, enable all nfs and nfs4 tracepoints, and set up a
trigger to disable tracing when xprt_retransmit fires to capture
everything that leads up to it.
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: Trond Myklebust <trond.myklebust@hammerspace.com>
I've hit some crashes that occur in the xprt_rdma_inject_disconnect
path. It appears that, for some provides, rdma_disconnect() can
take so long that the transport can disconnect and release its
hardware resources while rdma_disconnect() is still running,
resulting in a UAF in the provider.
The transport's fault injection method may depend on the stability
of transport data structures. That means it needs to be invoked
only from contexts that hold the transport write lock.
Fixes: 4a06825839 ("SUNRPC: Transport fault injection")
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: Trond Myklebust <trond.myklebust@hammerspace.com>
If the server sends CB_ calls on a connection that is not associated
with the backchannel, refuse to process the call and shut down the
connection. This avoids a NULL dereference crash in
xprt_complete_bc_request(). There's not much more we can do in this
situation unless we want to look into allowing all connections to be
associated with the fore and back channel.
Signed-off-by: Benjamin Coddington <bcodding@redhat.com>
Signed-off-by: Trond Myklebust <trond.myklebust@hammerspace.com>
Currently rpcbind client is created without setting rpc timeout (thus
using the default value). But if the rpc_task already has a customized
timeout in its tk_client field, it's also ignored.
Let's use the same timeout setting in rpc_task->tk_client->cl_timeout
for rpcbind connection.
Signed-off-by: Eryu Guan <eguan@linux.alibaba.com>
Signed-off-by: Trond Myklebust <trond.myklebust@hammerspace.com>
When we have multiple RPC requests queued up, it makes sense to set the
TCP_CORK option while the transmit queue is non-empty.
Signed-off-by: Trond Myklebust <trond.myklebust@hammerspace.com>
This, to me, seems less cluttered and less redundant. I was hoping
it could help reduce lock contention on the dto_q lock by reducing
the size of the critical section, but alas, the only improvement is
readability.
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
These fields are no longer used.
The size of struct svc_rdma_recv_ctxt is now less than 300 bytes on
x86_64, down from 2440 bytes.
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Currently the generic RPC server layer calls svc_rdma_recvfrom()
twice to retrieve an RPC message that uses Read chunks. I'm not
exactly sure why this design was chosen originally.
Instead, let's wait for the Read chunk completion inline in the
first call to svc_rdma_recvfrom().
The goal is to eliminate some page allocator churn.
rdma_read_complete() replaces pages in the second svc_rqst by
calling put_page() repeatedly while the upper layer waits for the
request to be constructed, which adds unnecessary NFS WRITE round-
trip latency.
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Reviewed-by: Tom Talpey <tom@talpey.com>
Currently, XPT_BUSY is not cleared until xpo_recvfrom returns.
That effectively blocks the receipt and handling of the next RPC
message until the current one has been taken off the transport.
This strict ordering is a requirement for socket transports.
For our kernel RPC/RDMA transport implementation, however, dequeuing
an ingress message is nothing more than a list_del(). The transport
can safely be marked un-busy as soon as that is done.
To keep the changes simpler, this patch just moves the
svc_xprt_received() call site from svc_handle_xprt() into the
transports, so that the actual optimization can be done in a
subsequent patch.
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Prepare svc_xprt_received() to be called from transport code instead
of from generic RPC server code.
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Chuck Lever
Trond Myklebust
NeilBrown
Linus Torvalds
Baptiste Lepers
Yunjian Wang
Dan Carpenter
Chris Dion
Jiapeng Chong
Benjamin Coddington
Eryu Guan