Re: [RFC] store function address in BTF

[Jiri Olsa <jolsa@redhat.com>]

On Wed, Oct 06, 2021 at 09:17:39AM -0700, Andrii Nakryiko wrote:
> On Wed, Oct 6, 2021 at 1:42 AM Jiri Olsa <jolsa@redhat.com> wrote:
> >
> > hi,
> > I'm hitting performance issue and soft lock ups with the new version
> > of the patchset and the reason seems to be kallsyms lookup that we
> > need to do for each btf id we want to attach
> >
> > I tried to change kallsyms_lookup_name linear search into rbtree search,
> > but it has its own pitfalls like duplicate function names and it still
> > seems not to be fast enough when you want to attach like 30k functions
> 
> How not fast enough is it exactly? How long does it take?

30k functions takes 75 seconds for me, it's loop calling bpf_check_attach_target

getting soft lock up messages:

krava33 login: [  168.896671] watchdog: BUG: soft lockup - CPU#1 stuck for 26s! [bpftrace:1087]


> 
> >
> > so I wonder we could 'fix this' by storing function address in BTF,
> > which would cut kallsyms lookup completely, because it'd be done in
> > compile time
> >
> > my first thought was to add extra BTF section for that, after discussion
> > with Arnaldo perhaps we could be able to store extra 8 bytes after
> > BTF_KIND_FUNC record, using one of the 'unused' bits in btf_type to
> > indicate that? or new BTF_KIND_FUNC2 type?
> >
> > thoughts?
> 
> I'm strongly against this, because (besides the BTF bloat reason) we
> need similar mass attachment functionality for kprobe/kretprobe and
> that one won't be relying on BTF FUNCs, so I think it's better to
> stick to the same mechanism for figuring out the address of the
> function.

ok

> 
> If RB tree is not feasible, we can do a linear search over unsorted
> kallsyms and do binary search over sorted function names (derived from
> BTF IDs). That would be O(Nlog(M)), where N is number of ksyms, M is
> number of BTF IDs/functions-to-be-attached-to. If we did have an RB
> tree for kallsyms (is it hard to support duplicates? why?) it could be
> even faster O(Mlog(N)).

I had issues with generic kallsyms rbtree in the post some time ago,
I'll revisit it to check on details.. but having the tree with just
btf id functions might clear that.. I'll check

thanks,
jirka

> 
> 
> >
> > thanks,
> > jirka
> >
>