On 5/3/24 6:48 PM, John Fastabend wrote:
[ Upstream commit ebf2e8860eea66e2c4764316b80c6a5ee5f336ee]
[ Upstream commit f8dd95b29d7ef08c19ec9720564acf72243ddcf6]
In the first patch,
ebf2e8860eea ("tcp_bpf: Inline do_tcp_sendpages as it's now a wrapper around tcp_sendmsg")
This block of code is added to tcp_bpf_push(). The
tcp_bpf_push is the code used by BPF to submit messages into the TCP
stack.
if (flags & MSG_SENDPAGE_NOTLAST)
msghdr.msg_flags | MSG_MORE;
In the second patch,
f8dd95b29d7e ("tcp_bpf, smc, tls, espintcp, siw: Reduce MSG_SENDPAGE_NOTLAST usage")
this logic was further changed to,
if (flags & MSG_SENDPAGE_NOTLAST)
msghdr.msg_flags |= MSG_MORE
This was done as part of an improvement to use the sendmsg() callbacks
and remove the sendpage usage inside the various sub systems.
However, these two patches together fixed a bug. The issue is without
MSG_MORE set we will break a msg up into many smaller sends. In some
case a lot because the operation loops over the scatter gather list.
Without the MSG_MORE set (the current 6.1 case) we see stalls in data
send/recv and sometimes applications failing to receive data. This
generally is the result of an application that gives up after calling
recv() or similar too many times. We introduce this because of how
we incorrectly change the TCP send pattern.
Now that we have both 6.5 and 6.1 stable kernels deployed we've
observed a series of issues related to this in real deployments. In 6.5
kernels all the HTTP and other compliance tests pass and we are not
observing any other issues. On 6.1 various compliance tests fail
(nginx for example), but more importantly in these clusters without
the flag set we observe stalled applications and increased retries in
other applications. Openssl users where we have annotations to monitor
retries and failures observed a significant increase in retries for
example.
For the backport we isolated the fix to the two lines in the above
patches that fixed the code. With this patch we deployed the workloads
again and error rates and stalls went away and 6.1 stable kernels
perform similar to 6.5 stable kernels. Similarly the compliance tests
also passed.
Cc: <stable@xxxxxxxxxxxxxxx> # 6.1.x
Fixes: 604326b41a6fb ("tcp_bpf, smc, tls, espintcp, siw: Reduce MSG_SENDPAGE_NOTLAST usage")
Signed-off-by: John Fastabend <john.fastabend@xxxxxxxxx>
Acked-by: Daniel Borkmann <daniel@xxxxxxxxxxxxx>
