Kernel mode drops many DTMF events when DTMF packets are interleaved with audio packets

Start in kernel-module/xt_RTPENGINE.c at the pt_filter check (~lines 6268-6310) where packets with an unknown RTP payload type are dropped, then read daemon/media_socket.c (~1538-1560) where the kernel target's allowed payload-type list is built from payloads already observed. Reproduce with a g711 call carrying interleaved telephone-event packets (e.g. Cisco 7841). Done means DTMF still passes once the flow is kernelized, either by seeding the list from the SDP or by bubbling the unknown payload type up to userspace.

rtpengine version the issue has been seen with

13.4.1.6-1~bpo12+1

Used distribution and its version

Debian 12

Linux kernel version used

6.1.0-38-amd64

CPU architecture issue was seen on (see uname -m)

x86_64

Expected behaviour you didn't see

1:1 Proxying of DTMF Events

Unexpected behaviour you saw

In kernel mode many DTMF events are dropped when DTMF packets are interleaved with audio packets.

When running rtpengine only in user mode then interleaved DTMF is passed without problems.

Steps to reproduce the problem

Generate a g711 call with interleaved DMTF in the style as shown in the screenshot. E.g. with a CISCO 7841 phone. DTMF packets of one event are sent with the same timestamp and increasing duration until the DTMF end packet has been reached, while interleaved audio packets have increasing timestamps. Interleaving of packets is unusual but allowed according to rfc4733 and rfc2833.

Additional program output to the terminal or logs illustrating the issue

Anything else?

Claude AI Code Analysis

I'll scan the kernel module and related RTP/DTMF handling to see how interleaved RTP and RFC2833 DTMF are processed, focusing on sequence/timestamp ordering, payload-type switching, and packet coalescing that could drop events. Running broad semantic and exact searches in parallel.

I'll look specifically inside the kernel module sources for RFC2833/4733 handling, sequence/timestamp checks, and any jitter/ordering assumptions that might drop interleaved events. Then I'll open the key files it points to.

I looked through the kernel module path (kernel-module/xt_RTPENGINE.c) and how userspace populates it (daemon/media_socket.c). The module doesn’t do any DTMF-specific parsing; it just forwards RTP if the payload type is allowed for that target.

Key observations:

• Kernel forwarding only passes payload types that were registered at kernelization time. See pt_filter handling:

		// if RTP, only forward packets of known/passthrough payload types
		if (rtp_pt_idx < 0) {
			if (g->target.pt_filter)
				goto out;
		}

• The list of allowed payload types (g->target.pt_stats[]) is built in userspace from the stream’s observed RTP payloads, then sorted and pushed down to the kernel:

struct rtp_stats *payload_types[RTPE_NUM_PAYLOAD_TYPES];
...
while (t_hash_table_iter_next(&iter, NULL, &rs) && s->num_payload_types < RTPE_NUM_PAYLOAD_TYPES)
	s->payload_types[s->num_payload_types++] = rs;
qsort(s->payload_types, s->num_payload_types, sizeof(*s->payload_types), __rtp_stats_pt_sort);
...
if (!can_kernelize) {
	reti->pt_filter = 1;

• Critically, that list is derived from payload types already seen on the stream, not from the SDP. So if you kernelize when only audio has been observed, telephone-event may not be included. When DTMF later appears interleaved with audio, those packets have an “unknown” PT for the kernel target and are dropped (not propagated up to userspace), causing lost tones.
• There’s no fallback to userspace on unknown PT when pt_filter is set; packets are just dropped.
• Nothing else in the kernel path (seq/jitter tracking, SRTP re-encrypt, pattern replacement, min_payload_len) explains selective loss tied specifically to interleaving. The min_payload_len and blackhole gating only apply if configured, and don’t target DTMF by default.

Why interleaving triggers it:

• Audio continues, so the flow stays kernelized.
• First DTMF packets (start/continuations/end) come with PT telephone-event, which wasn’t registered; the kernel drops them due to pt_filter.
• If the system never bubbles those packets to userspace, the kernel target never gets updated, so the receiver misses some or all tones.

Practical fixes:

• Userspace: when creating the kernel target, include telephone-event proactively based on SDP, not just observed traffic. Or always include telephone-event if negotiated in the stream’s codec set.
• Userspace: if pt_filter is enabled, on the first unknown payload type observed in userspace (or via a signal from kernel), reconfigure the kernel target to add it.
• Kernel: instead of dropping unknown PT under pt_filter, “propagate” the first unknown PT up to userspace so that it can update the kernel target, or temporarily forward unknown PT for a grace period.
• Optional: for safety, treat telephone-event as allowed by default when RTP is allowed for the stream’s SSRC.

If you want, I can implement the userspace change in daemon/media_socket.c to seed the kernel target’s payload list from the SDP (including telephone-event) so interleaved DTMF isn’t dropped after kernelization.

		// if RTP, only forward packets of known/passthrough payload types
		if (rtp_pt_idx < 0) {
			if (g->target.pt_filter)
				goto out;
		}

struct rtp_stats *payload_types[RTPE_NUM_PAYLOAD_TYPES];
...
while (t_hash_table_iter_next(&iter, NULL, &rs) && s->num_payload_types < RTPE_NUM_PAYLOAD_TYPES)
	s->payload_types[s->num_payload_types++] = rs;
qsort(s->payload_types, s->num_payload_types, sizeof(*s->payload_types), __rtp_stats_pt_sort);
...
if (!can_kernelize) {
	reti->pt_filter = 1;