netdev - Hardcoded multicast queue length in macvlan.c driver causes poor multicast receive performance

lists.openwall.net		lists / announce owl-users owl-dev john-users john-dev passwdqc-users yescrypt popa3d-users / oss-security kernel-hardening musl sabotage tlsify passwords / crypt-dev xvendor / Bugtraq Full-Disclosure linux-kernel linux-netdev linux-ext4 linux-hardening linux-cve-announce PHC
Open Source and information security mailing list archives

Hash Suite: Windows password security audit tool. GUI, reports in PDF.

[<prev] [next>] [thread-next>] [day] [month] [year] [list]

Message-ID: <147b704ac1d5426fbaa8617289dad648@paneda.se>
Date:   Mon, 23 Nov 2020 14:22:31 +0000
From:   Thomas Karlsson <thomas.karlsson@...eda.se>
To:     "davem@...emloft.net" <davem@...emloft.net>,
        "kuba@...nel.org" <kuba@...nel.org>
CC:     "netdev@...r.kernel.org" <netdev@...r.kernel.org>
Subject: Hardcoded multicast queue length in macvlan.c driver causes poor
 multicast receive performance

Hello,

There is a special queue handling in macvlan.c for broadcast and multicast packages that was arbitrarily set to 1000 in commit 07d92d5cc977a7fe1e683e1d4a6f723f7f2778cb . While this is probably sufficient for most uses cases it is insufficient to support high packet rates. I currently have a setup with 144 000 multicast packets incoming per second (144 different live audio RTP streams) and suffer very frequent packet loss. With unicast this is not an issue and I can in addition to the 144kpps load the macvlan interface with another 450mbit/s using iperf.

In order to verify that the queue is the problem I edited the define to 100000 and recompiled the kernel module. After replacing it with rmmod/insmod I get 0 packet loss (measured over 2 days where I before had losses every other second or so) and can also load an additional 450 mbit/s multicast traffic using iperf without losses. So basically no change in performance between unicast/multicast when it comes to lost packets on my machine.

I think It would be best if this queue length was configurable somehow. Either an option when creating the macvlan (like how bridge/passthrough/etc are set) or at least when loading the module (for instance by using a config in /etc/modprobe.d). One size does not fit all in this situation.

Link to code in question using the define (on master):
https://github.com/torvalds/linux/blob/27bba9c532a8d21050b94224ffd310ad0058c353/drivers/net/macvlan.c#L357 

(re-sent in text/plain instead of html)

Best regards,
Thomas Karlsson