Re: [PATCH bpf-next v3 2/8] samples: bpf: Add common infrastructure for XDP samples

["Toke Høiland-Jørgensen" <toke@redhat.com>]

Daniel Borkmann <daniel@iogearbox.net> writes:

> On 7/28/21 6:55 PM, Kumar Kartikeya Dwivedi wrote:
>> This file implements some common helpers to consolidate differences in
>> features and functionality between the various XDP samples and give them
>> a consistent look, feel, and reporting capabilities.
>> 
>> Some of the key features are:
>>   * A concise output format accompanied by helpful text explaining its
>>     fields.
>>   * An elaborate output format building upon the concise one, and folding
>>     out details in case of errors and staying out of view otherwise.
>>   * Extended reporting of redirect errors by capturing hits for each
>>     errno and displaying them inline (ENETDOWN, EINVAL, ENOSPC, etc.)
>>     to aid debugging.
>>   * Reporting of each xdp_exception action for all samples that use these
>>     helpers (XDP_ABORTED, etc.) to aid debugging.
>>   * Capturing per ifindex pair devmap_xmit counts for decomposing the
>>     total TX count per devmap redirection.
>>   * Ability to jump to source locations invoking tracepoints.
>>   * Faster retrieval of stats per polling interval using mmap'd eBPF
>>     array map (through .bss).
>>   * Printing driver names for devices redirecting packets.
>>   * Printing summarized total statistics for the entire session.
>>   * Ability to dynamically switch between concise and verbose mode, using
>>     SIGQUIT (Ctrl + \).
>> 
>> The goal is sharing these helpers that most of the XDP samples implement
>> in some form but differently for each, lacking in some respect compared
>> to one another.
>> 
>> Signed-off-by: Kumar Kartikeya Dwivedi <memxor@gmail.com>
>
> Overall I think it's okay to try to streamline the individual XDP tools, but I
> also tend to wonder whether we keep going beyond the original purpose of kernel
> samples where the main goal is to provide small, easy to hack & stand-alone code
> snippets (like in samples/seccomp ... no doubt we have it more complex in BPF
> land, but still); things people can take away and extend for their purpose. A big
> portion of the samples are still better off in selftests so they can be run in CI,
> and those that are not should generally be simplified for developers to rip out,
> modify, experiment, and build actual applications on top.

FWIW the idea of improving the samples came from Jesper and myself;
we've come to rely on them quite a bit for benchmarking, and our QE
folks run them for testing as well. And I've lost count of the number of
times I had to redo tests because something wasn't working correctly and
I didn't notice that the numbers were off. Kumar took the "improve the
XDP samples" idea and ran with it, and I think the result is much
improved; having it be immediately obvious when something is off is a
huge benefit!

So while I do share your concern about expanding the samples code too
much, in this instance I think it's an improvement. I've toyed with the
idea of also distributing some of the XDP samples with xdp-tools so they
are easier to install as standalone utilities, but I think that is a
secondary concern for later.

-Toke