netdev - Re: [RFC PATCH bpf-next 2/4] bpf: introduce BPF dispatcher

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

Date:   Thu, 14 Nov 2019 14:56:06 +0100
From:   Björn Töpel <bjorn.topel@...il.com>
To:     Daniel Borkmann <daniel@...earbox.net>
Cc:     Toke Høiland-Jørgensen <thoiland@...hat.com>,
        Netdev <netdev@...r.kernel.org>,
        Alexei Starovoitov <ast@...nel.org>,
        Björn Töpel <bjorn.topel@...el.com>,
        bpf <bpf@...r.kernel.org>,
        Magnus Karlsson <magnus.karlsson@...il.com>,
        "Karlsson, Magnus" <magnus.karlsson@...el.com>,
        Jonathan Lemon <jonathan.lemon@...il.com>
Subject: Re: [RFC PATCH bpf-next 2/4] bpf: introduce BPF dispatcher

On Thu, 14 Nov 2019 at 14:03, Daniel Borkmann <daniel@...earbox.net> wrote:
>
> On 11/14/19 1:31 PM, Toke Høiland-Jørgensen wrote:
> > Björn Töpel <bjorn.topel@...il.com> writes:
> >> From: Björn Töpel <bjorn.topel@...el.com>
> >>
> >> The BPF dispatcher builds on top of the BPF trampoline ideas;
> >> Introduce bpf_arch_text_poke() and (re-)use the BPF JIT generate
> >> code. The dispatcher builds a dispatch table for XDP programs, for
> >> retpoline avoidance. The table is a simple binary search model, so
> >> lookup is O(log n). Here, the dispatch table is limited to four
> >> entries (for laziness reason -- only 1B relative jumps :-P). If the
> >> dispatch table is full, it will fallback to the retpoline path.
> >
> > So it's O(log n) with n == 4? Have you compared the performance of just
> > doing four linear compare-and-jumps? Seems to me it may not be that big
> > of a difference for such a small N?
>
> Did you perform some microbenchmarks wrt search tree? Mainly wondering
> since for code emission for switch/case statements, clang/gcc turns off
> indirect calls entirely under retpoline, see [0] from back then.
>

As Toke stated, binsearch is not needed for 4 entries. I started out
with 16 (and explicit ids instead of pointers), and there it made more
sense. If folks think it's a good idea to move forward -- and with 4
entries, it makes sense to make the code generator easier, or maybe
based on static_calls like Ed did.

As for ubenchmarks I only compared with 1 cmp, vs 3 vs 4 + retpoline
stated in the cover. For a proper patch I can do more in-depth
analysis. Or was it anything particular you were looking for?

For switch/case code generation there's a great paper on that here [3]
from the 2008 GCC dev summit ("A Superoptimizer Analysis of Multiway
Branch Code Generation")

[3] http://citeseerx.ist.psu.edu/viewdoc/download;jsessionid=968AE756567863243AC7B1728915861A?doi=10.1.1.602.1875&rep=rep1&type=pdf


> >> An example: A module/driver allocates a dispatcher. The dispatcher is
> >> shared for all netdevs. Each netdev allocate a slot in the dispatcher
> >> and a BPF program. The netdev then uses the dispatcher to call the
> >> correct program with a direct call (actually a tail-call).
> >
> > Is it really accurate to call it a tail call? To me, that would imply
> > that it increments the tail call limit counter and all that? Isn't this
> > just a direct jump using the trampoline stuff?
>
> Not meant in BPF context here, but more general [1].
>
> (For actual BPF tail calls I have a series close to ready for getting
> rid of most indirect calls which I'll post later today.)
>

Thanks for the clarification, Daniel! (call vs jmp)

> Best,
> Daniel
>
>    [0] https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/commit/?id=a9d57ef15cbe327fe54416dd194ee0ea66ae53a4
>    [1] https://en.wikipedia.org/wiki/Tail_call