Some targets read one datagram per recvmsg call instead of using recvmmsg: Apple platforms without the fast-apple-datapath feature, OpenBSD, NetBSD, Solaris and illumos, and from 0.6 also DragonFly BSD, Redox and GNU/Hurd. On these targets the receive loop set up its msghdr only once and looked at the MSG_TRUNC flag before checking whether recvmsg had failed. After a datagram larger than the receive buffer was dropped, a following EAGAIN still saw the old truncation flag. The loop then kept issuing non-blocking receives instead of returning WouldBlock.
When the receive buffer is smaller than a datagram that arrives, a remote sender can keep the task that drives the socket busy, before any QUIC packet authentication. The loop ends only when a later datagram fits in the buffer. Whether an application is exposed therefore depends on the receive buffer size it uses. Linux, Android, FreeBSD and Windows use other receive paths and are not affected.
The fix rebuilds the header before every recvmsg call and returns receive errors before checking the truncation flag. It is in 0.5.16 and 0.6.3.
{
"license": "CC0-1.0"
}