fix(socket): restore buffer capacities before reusing recvmmsg headers - #2813
Open
maxlapshin wants to merge 2 commits into
Open
fix(socket): restore buffer capacities before reusing recvmmsg headers#2813maxlapshin wants to merge 2 commits into
maxlapshin wants to merge 2 commits into
Conversation
maxlapshin
added a commit
to maxlapshin/nix
that referenced
this pull request
Aug 16, 2026
pacak
reviewed
Aug 25, 2026
pacak
left a comment
pacak
left a comment
Contributor
There was a problem hiding this comment.
I think sendmmsg suffers from the same problem - it can write control messages to return timestamps for sent messages.
| _cmsg_buffers: Option<Box<[u8]>>, | ||
| msg_controllen: usize, | ||
| // the capacity of every address buffer, needed to restore `msg_namelen` | ||
| // before reusing the headers, see `reset_for_receive` |
| let p = &mut mmsghdr.msg_hdr; | ||
| p.msg_iov = slice.as_mut().as_mut_ptr().cast(); | ||
| p.msg_iovlen = slice.as_mut().len() as _; | ||
| p.msg_namelen = msg_namelen; |
Contributor
There was a problem hiding this comment.
Name length is restored, but there's no test for it. Doing something with unix sockets might do the trick.
Author
There was a problem hiding this comment.
Added test_recvmmsg_address_after_reuse
`msg_namelen` and `msg_controllen` are in-out parameters of `recvmsg(2)`: the kernel replaces the capacity with the length it actually used, and for the address it also copies no more than the capacity it was given. Since `recvmmsg` reuses the same `MultiHeaders` across calls and only refilled `msg_iov`, a slot that once received a datagram without control messages was left with a control capacity of zero, and every later datagram in that slot came back with `MSG_CTRUNC` and no control message at all; a slot that once received from a peer with a short address truncated the address of every later peer. `sendmmsg` never sets `msg_controllen` either, so headers that had gone through `recvmmsg` made it encode control messages into no room at all. Refill both capacities, and clear `msg_flags`, before every call of both.
maxlapshin
force-pushed
the
recvmmsg-loses-cmsg
branch
from
August 29, 2026 06:42
8400c9b to
1ecff38
Compare
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Description
recvmmsg loses control messages when MultiHeaders are reused
Expected
MultiHeaders is documented as a preallocated structure to be reused across recvmmsg calls, so every call should be able to receive as much control data as the buffer given to preallocate holds.
Actual
only the first call can. msg_namelen and msg_controllen are in-out parameters of recvmsg(2): going in they are the capacity of the buffers, coming out the kernel replaces them with the number of bytes it actually wrote. recvmmsg refills only msg_iov/msg_iovlen between calls, so once a slot receives a datagram carrying no control message, its control capacity stays at zero forever. Every later datagram in that slot comes back with MSG_CTRUNC and cmsgs() returns ENOBUFS — and code that reads control data as an optional extra (.cmsgs().ok()) simply stops seeing it.
Reproduction (added as test_recvmmsg_cmsgs_after_reuse)
bind a UDP socket, send and receive one datagram with ReceiveTimestampns off, then turn it on and receive a second one through the same MultiHeaders. The timestamp never arrives.
Found in the wild
a WebRTC load generator that reads with UDP_GRO — the segment size arrives in a control message. After the first non-coalesced datagram the size stopped arriving, each coalesced batch was counted as a single datagram, and the measurement reported 72% packet loss that did not exist.
Versions
reproduced on 0.29.0 and on master (0.31.3), Linux 6.8 and 7.0.
Fix
remember the capacities in MultiHeaders at preallocate time and restore msg_namelen/msg_controllen (and clear msg_flags) in the recvmmsg loop. msg_namelen matters on the BSDs, where the input value bounds the address copy; on Linux the kernel resets it itself.
Checklist:
CONTRIBUTING.md