From 0e6673fb002e18518dfcb89e88380b1cfb1cdc15 Mon Sep 17 00:00:00 2001
From: Tuetuopay <tuetuopay@me.com>
Date: Wed, 12 Oct 2022 11:41:28 +0200
Subject: [PATCH] Add ENOBUFS handling for unsolicited messages

This can happen when large burst of messages come all of a sudden, which
happen very easily when routing protocols are involved (e.g. BGP). The
current implementation incorrectly assumes that any failure to read from
the socket is akin to the socket closed. This is not the case.

This commit adds handling for this specific error by generating a
`NetlinkPayload::Overrun(_)` message that users receive on their
unsolicited message channel. Since this is just an additional message,
there is no breaking change for existing users and they are free to
ignore it if they do not want to handle it, or handle it by e.g. resyncing.
---
 src/framed.rs | 36 +++++++++++++++++++++++++++++++++++-
 1 file changed, 35 insertions(+), 1 deletion(-)

diff --git a/src/framed.rs b/src/framed.rs
index 5d9d301..f940a0d 100644
--- a/src/framed.rs
+++ b/src/framed.rs
@@ -16,7 +16,13 @@ use crate::{
     codecs::NetlinkMessageCodec,
     sys::{AsyncSocket, SocketAddr},
 };
-use netlink_packet_core::{NetlinkDeserializable, NetlinkMessage, NetlinkSerializable};
+use netlink_packet_core::{
+    NetlinkDeserializable, NetlinkMessage, NetlinkSerializable, NetlinkHeader,
+    NLMSG_OVERRUN, NetlinkPayload,
+};
+
+/// Buffer overrun condition
+const ENOBUFS: i32 = 105;
 
 pub struct NetlinkFramed<T, S, C> {
     socket: S,
@@ -63,6 +69,34 @@ where
 
             *in_addr = match ready!(socket.poll_recv_from(cx, reader)) {
                 Ok(addr) => addr,
+                // When receiving messages in multicast mode (i.e. we subscribed to
+                // notifications), the kernel will not wait for us to read datagrams before
+                // sending more. The receive buffer has a finite size, so once it is full (no
+                // more message can fit in), new messages will be dropped and recv calls will
+                // return `ENOBUFS`.
+                // This needs to be handled for applications to resynchronize with the contents
+                // of the kernel if necessary.
+                // We don't need to do anything special:
+                // - contents of the reader is still valid because we won't have partial messages
+                //   in there anyways (large enough buffer)
+                // - contents of the socket's internal buffer is still valid because the kernel
+                //   won't put partial data in it
+                Err(e) if e.raw_os_error() == Some(ENOBUFS) => {
+                    warn!("netlink socket buffer full");
+                    return Poll::Ready(Some((
+                        NetlinkMessage {
+                            header: NetlinkHeader {
+                                length: std::mem::size_of::<NetlinkHeader>() as u32,
+                                message_type: NLMSG_OVERRUN,
+                                flags: 0,
+                                sequence_number: 0,
+                                port_number: 0,
+                            },
+                            payload: NetlinkPayload::Overrun(Vec::new()),
+                        },
+                        SocketAddr::new(0, 0),
+                    )));
+                }
                 Err(e) => {
                     error!("failed to read from netlink socket: {:?}", e);
                     return Poll::Ready(None);