netdev - [RFC PATCH bpf-next 0/2] bpf: keep track of prog verification stats

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

Message-ID: <20210920151112.3770991-1-davemarchevsky@fb.com>
Date:   Mon, 20 Sep 2021 08:11:10 -0700
From:   Dave Marchevsky <davemarchevsky@...com>
To:     <bpf@...r.kernel.org>
CC:     <netdev@...r.kernel.org>, Alexei Starovoitov <ast@...nel.org>,
        Daniel Borkmann <daniel@...earbox.net>,
        Andrii Nakryiko <andrii@...nel.org>,
        Yonghong Song <yhs@...com>,
        Dave Marchevsky <davemarchevsky@...com>
Subject: [RFC PATCH bpf-next 0/2] bpf: keep track of prog verification stats

The verifier currently logs some useful statistics in
print_verification_stats. Although the text log is an effective feedback
tool for an engineer iterating on a single application, it would also be
useful to enable tracking these stats in a more structured form for
fleetwide or historical analysis, which this patchset attempts to do.

A concrete motivating usecase which came up in recent weeks:

A team owns a complex BPF program, with various folks extending its
functionality over the years. An engineer tries to make a relatively
simple addition but encounters "BPF program is too large. Processed
1000001 insn". 

Their changes bumped the processed insns from 700k to over the limit and
there's no obvious way to simplify. They must now consider a large
refactor in order to incorporate the new feature. What if there was some
previous change which bumped processed insns from 200k->700k which
_could_ be modified to stress verifier less? Tracking historical
verifier stats for each version of the program over the years would
reduce manual work necessary to find such a change.

Although parsing the text log could work for this scenario, a solution
that's resilient to log format and other verifier changes would be
preferable.

This patchset adds a bpf_prog_verif_stats struct - containing the same
data logged by print_verification_stats - which can be retrieved as part
of bpf_prog_info. Looking for general feedback on approach and a few
specific areas before fleshing it out further:

* None of my usecases require storing verif_stats for the lifetime of a
  loaded prog, but adding to bpf_prog_aux felt more correct than trying
  to pass verif_stats back as part of BPF_PROG_LOAD
* The verif_stats are probably not generally useful enough to warrant
  inclusion in fdinfo, but hoping to get confirmation before removing
  that change in patch 1
* processed_insn, verification_time, and total_states are immediately
  useful for me, rest were added for parity with
	print_verification_stats. Can remove.
* Perhaps a version field would be useful in verif_stats in case future
  verifier changes make some current stats meaningless
* Note: stack_depth stat was intentionally skipped to keep patch 1
  simple. Will add if approach looks good.

Dave Marchevsky (2):
  bpf: add verifier stats to bpf_prog_info and fdinfo
  selftests/bpf: add verif_stats test

 include/linux/bpf.h                           |  1 +
 include/uapi/linux/bpf.h                      | 10 ++++++
 kernel/bpf/syscall.c                          | 20 +++++++++--
 kernel/bpf/verifier.c                         | 13 +++++++
 tools/include/uapi/linux/bpf.h                | 10 ++++++
 .../selftests/bpf/prog_tests/verif_stats.c    | 34 +++++++++++++++++++
 6 files changed, 86 insertions(+), 2 deletions(-)
 create mode 100644 tools/testing/selftests/bpf/prog_tests/verif_stats.c

-- 
2.30.2