Re: [PATCH net v2] TUN/TAP: Improving throughput and latency by avoiding SKB drops

[Stephen Hemminger <stephen@networkplumber.org>]

On Tue, 12 Aug 2025 00:03:48 +0200
Simon Schippers <simon.schippers@tu-dortmund.de> wrote:

> This patch is the result of our paper with the title "The NODROP Patch:
> Hardening Secure Networking for Real-time Teleoperation by Preventing
> Packet Drops in the Linux TUN Driver" [1].
> It deals with the tun_net_xmit function which drops SKB's with the reason
> SKB_DROP_REASON_FULL_RING whenever the tx_ring (TUN queue) is full,
> resulting in reduced TCP performance and packet loss for bursty video
> streams when used over VPN's.
> 
> The abstract reads as follows:
> "Throughput-critical teleoperation requires robust and low-latency
> communication to ensure safety and performance. Often, these kinds of
> applications are implemented in Linux-based operating systems and transmit
> over virtual private networks, which ensure encryption and ease of use by
> providing a dedicated tunneling interface (TUN) to user space
> applications. In this work, we identified a specific behavior in the Linux
> TUN driver, which results in significant performance degradation due to
> the sender stack silently dropping packets. This design issue drastically
> impacts real-time video streaming, inducing up to 29 % packet loss with
> noticeable video artifacts when the internal queue of the TUN driver is
> reduced to 25 packets to minimize latency. Furthermore, a small queue
> length also drastically reduces the throughput of TCP traffic due to many
> retransmissions. Instead, with our open-source NODROP Patch, we propose
> generating backpressure in case of burst traffic or network congestion.
> The patch effectively addresses the packet-dropping behavior, hardening
> real-time video streaming and improving TCP throughput by 36 % in high
> latency scenarios."
> 
> In addition to the mentioned performance and latency improvements for VPN
> applications, this patch also allows the proper usage of qdisc's. For
> example a fq_codel can not control the queuing delay when packets are
> already dropped in the TUN driver. This issue is also described in [2].
> 
> The performance evaluation of the paper (see Fig. 4) showed a 4%
> performance hit for a single queue TUN with the default TUN queue size of
> 500 packets. However it is important to notice that with the proposed
> patch no packet drop ever occurred even with a TUN queue size of 1 packet.
> The utilized validation pipeline is available under [3].
> 
> As the reduction of the TUN queue to a size of down to 5 packets showed no
> further performance hit in the paper, a reduction of the default TUN queue
> size might be desirable accompanying this patch. A reduction would
> obviously reduce buffer bloat and memory requirements.
> 
> Implementation details:
> - The netdev queue start/stop flow control is utilized.
> - Compatible with multi-queue by only stopping/waking the specific
> netdevice subqueue.
> - No additional locking is used.
> 
> In the tun_net_xmit function:
> - Stopping the subqueue is done when the tx_ring gets full after inserting
> the SKB into the tx_ring.
> - In the unlikely case when the insertion with ptr_ring_produce fails, the
> old dropping behavior is used for this SKB.
> 
> In the tun_ring_recv function:
> - Waking the subqueue is done after consuming a SKB from the tx_ring when
> the tx_ring is empty. Waking the subqueue when the tx_ring has any
> available space, so when it is not full, showed crashes in our testing. We
> are open to suggestions.
> - When the tx_ring is configured to be small (for example to hold 1 SKB),
> queuing might be stopped in the tun_net_xmit function while at the same
> time, ptr_ring_consume is not able to grab a SKB. This prevents
> tun_net_xmit from being called again and causes tun_ring_recv to wait
> indefinitely for a SKB in the blocking wait queue. Therefore, the netdev
> queue is woken in the wait queue if it has stopped.
> - Because the tun_struct is required to get the tx_queue into the new txq
> pointer, the tun_struct is passed in tun_do_read aswell. This is likely
> faster then trying to get it via the tun_file tfile because it utilizes a
> rcu lock.
> 
> We are open to suggestions regarding the implementation :)
> Thank you for your work!
> 
> [1] Link:
> https://cni.etit.tu-dortmund.de/storages/cni-etit/r/Research/Publications/2025/Gebauer_2025_VTCFall/Gebauer_VTCFall2025_AuthorsVersion.pdf
> [2] Link:
> https://unix.stackexchange.com/questions/762935/traffic-shaping-ineffective-on-tun-device
> [3] Link: https://github.com/tudo-cni/nodrop
> 
> Co-developed-by: Tim Gebauer <tim.gebauer@tu-dortmund.de>
> Signed-off-by: Tim Gebauer <tim.gebauer@tu-dortmund.de>
> Signed-off-by: Simon Schippers <simon.schippers@tu-dortmund.de>

I wonder if it would be possible to implement BQL in TUN/TAP?

https://lwn.net/Articles/454390/

BQL provides a feedback mechanism to application when queue fills.