On Tue, Feb 13, 2024 at 3:32 PM Kuniyuki Iwashima <kuniyu@xxxxxxxxxx> wrote:
>
> Recently, commit 8e5443d2b866 ("net: remove SOCK_DEBUG leftovers")
> removed the last users of SOCK_DEBUG(), and commit b1dffcf0da22 ("net:
> remove SOCK_DEBUG macro") removed the macro.
>
> Now is the time to deprecate the oldest socket option.
>
> Signed-off-by: Kuniyuki Iwashima <kuniyu@xxxxxxxxxx>
> ---
I would like to kindly implore you to please not remove the
functionality of the SO_DEBUG socket option. This socket option is a
key mechanism that the Google TCP team uses for automated testing of
Linux TCP, including BBR congestion control.
Widely used tools like netperf allow users to enable the SO_DEBUG
socket option via the command line (-g in netperf). Then debugging
code in the kernel can use the SOCK_DBG bit to decide whether to take
special actions, such as logging debug information, which can be used
to generate graphs or assertions about correct internal behavior. For
example, the transperf network testing tool that our team open-sourced
- https://github.com/google/transperf - uses the netperf -g/SO_DEBUG
mechanism to trigger debug logging that we use for testing,
troubleshooting, analysis, and development.
The SO_DEBUG mechanism is nice in that it works well no matter what
policy an application or benchmarking tool uses for choosing other
attributes (like port numbers) that could conceivably be used to point
out connections that should receive debug treatment. For example, most
benchmarking or production workloads will effectively end up with
random port numbers, which makes port numbers hard to use for
triggering debug treatment.
This mechanism is very simple and battle-tested, it works well, and
IMHO it would be a tragedy to remove it. It would cause our team
meaningful headaches to replace it. Please keep the SO_DEBUG socket
option functionality as-is. :-)
Thanks for your consideration on this!
Best regards,
neal
