[Cake] issue with Cake and bpf filter

[Toke Høiland-Jørgensen]

Jonathan Morton <chromatix99 at gmail.com> writes:

>> On 13 Aug, 2018, at 12:34 am, Jonathan Morton <chromatix99 at gmail.com> wrote:
>> 
>>> On 12 Aug, 2018, at 10:42 pm, Toke Høiland-Jørgensen <toke at toke.dk> wrote:
>>> 
>>> Yes it does; setting tc_classid is one way to set the flow class that is
>>> read by the call to tcf_classify(). I guess the problem is that
>>> cake_hash() is bypassed, so the *host_refcnt variables are not
>>> incremented.
>>> 
>>> There's been another report of the same issue on github; haven't had
>>> time to look into it yet, but I guess this is the reason...
>> 
>> I'll see what I can do tonight, now that I understand the problem.
>
> Oh, this is a mess. Ultimately the problem stems from having
> previously factored out choosing tin and flow to two distinct
> functions, but now there's an external mechanism for overriding both
> at once, which can only hook into one or the other.
>
> The easiest fix is to just remove the broken support for setting the
> flow ID via a filter, but leave in the support for setting the tin. I
> think that's the most useful "simple" fix. It's perhaps worth noting
> that this was the first thing I *removed* when reworking fq_codel into
> the first version of Cake, because I couldn't see a valid use for it.

Well, seeing as lots of people have expressed interest in the feature
since we put it back in it seems that there *is* a use for it :)

> The next simplest fix is to ignore the flow ID override unless we're
> in "flows" mode. We can then make valid assumptions about what should
> go into the host tables.
>
> The *right* fix, if we want to maximise functionality, would be to
> pass the result struct by reference into cake_hash(), where it can
> override the *host* IDs (not the flow ID). Users can then choose
> between using the override as a flow ID (by setting "hosts" mode
> instead of "flows"), or retaining the default host-isolation semantics
> with a revised definition of "host".

Ah, making it possible to override both host and flow mode is a great
idea! I guess we could use the major/minor distinction in the class to
steer this. I'll see if I can't integrate this.

> I feel compelled to point out that baggage like this is probably at
> least some of why Cake is no longer faster than HTB+fq_codel, as it
> used to be.

The benchmarks showing HTB to be faster were performed before the
classification went in, so I very much doubt that :)

-Toke