net: sockmap, fix missing MSG_MORE causing TCP disruptions

PATCH In review John Fastabend 3 participants 4 messages May 3, 2024 – May 4, 2024

Subsystems (2)

Stable branch BPF

lore

Share & tools

lore

John Fastabend

author

2024-05-03 16:48 UTC #0 lore raw

PATCH 1/1 net: sockmap, fix missing MSG_MORE causing TCP disruptions

[ 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 <[email protected]> # 6.1.x

Fixes 604326b41a6fb ("tcp_bpf, smc, tls, espintcp, siw: Reduce MSG_SENDPAGE_NOTLAST usage")

Signed-off-by John Fastabend <[email protected]>

---
 net/ipv4/tcp_bpf.c | 3 +++
 1 file changed, 3 insertions(+)

net/ipv4/tcp_bpf.c +3 -0

index f8037d142bb7..20d94f67fde2 100644
@@ -105,6 +105,9 @@ static int tcp_bpf_push(struct sock *sk, struct sk_msg *msg, u32 apply_bytes,
 
 		tcp_rate_check_app_limited(sk);
 retry:
+		if (size < sge->length && msg->sg.start != msg->sg.end)
+			flags |= MSG_SENDPAGE_NOTLAST;
+
 		has_tx_ulp = tls_sw_has_ctx_tx(sk);
 		if (has_tx_ulp) {
 			flags |= MSG_SENDPAGE_NOPOLICY;

Signature · 1 line

2.33.0

Daniel Borkmann

maintainer Acked-by on v1 1/1

2024-05-03 18:16 UTC #1 lore raw

Quoted 51 lines · John Fastabend

[ 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.

Quoted 3 lines

Cc <[email protected]> # 6.1.x
Fixes 604326b41a6fb ("tcp_bpf, smc, tls, espintcp, siw: Reduce MSG_SENDPAGE_NOTLAST usage")
Signed-off-by John Fastabend <[email protected]>

Acked-by Daniel Borkmann <[email protected]>

Greg KH

maintainer

2024-05-04 07:30 UTC #2 lore raw

In reply to John Fastabend

Quoted 2 lines · John Fastabend

[ Upstream commit ebf2e8860eea66e2c4764316b80c6a5ee5f336ee] [ Upstream commit f8dd95b29d7ef08c19ec9720564acf72243ddcf6]

Why are you mushing 2 patches together? Why can't we just take the two as-is instead? That makes tracking everything much simpler and possible.

Quoted 42 lines

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.

Quoted 5 lines

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.

Can we just take the two original patches instead?

thanks,

greg k-h

John Fastabend

author

2024-05-04 16:48 UTC #3 lore raw

Quoted 7 lines · Greg KH

Quoted 2 lines · John Fastabend
[ Upstream commit ebf2e8860eea66e2c4764316b80c6a5ee5f336ee] [ Upstream commit f8dd95b29d7ef08c19ec9720564acf72243ddcf6]
Why are you mushing 2 patches together? Why can't we just take the two as-is instead? That makes tracking everything much simpler and possible.

OK the thought was to get the minimal diff needed. But that is problematic. We can take the first one as-is and then the second one will have a couple chunks that don't apply but we don't need those chunks because the infiniband part it touches doesn't have the same issue in 6.1.

Quoted 2 lines

Quoted 2 lines
In the first patch,

[...]

Quoted 7 lines

Quoted 5 lines
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.
Can we just take the two original patches instead?

Yes minus the couple chunks that don't apply on the second one. I'll do some testing and resend thanks.

Quoted 4 lines

thanks,
greg k-h