[PATCH v3 seccomp 0/5] seccomp: Add bitmap cache of constant allow filter results

[YiFei Zhu <zhuyifei1999@gmail.com>]

From: YiFei Zhu <yifeifz2@illinois.edu>

Alternative: https://lore.kernel.org/lkml/20200923232923.3142503-1-keescook@chromium.org/T/

Major differences from the linked alternative by Kees:
* No x32 special-case handling -- not worth the complexity
* No caching of denylist -- not worth the complexity
* No seccomp arch pinning -- I think this is an independent feature
* The bitmaps are part of the filters rather than the task.
* Architectures supported by default through arch number array,
  except for MIPS with its sparse syscall numbers.
* Configurable per-build for future different cache modes.

This series adds a bitmap to cache seccomp filter results if the
result permits a syscall and is indepenent of syscall arguments.
This visibly decreases seccomp overhead for most common seccomp
filters with very little memory footprint.

The overhead of running Seccomp filters has been part of some past
discussions [1][2][3]. Oftentimes, the filters have a large number
of instructions that check syscall numbers one by one and jump based
on that. Some users chain BPF filters which further enlarge the
overhead. A recent work [6] comprehensively measures the Seccomp
overhead and shows that the overhead is non-negligible and has a
non-trivial impact on application performance.

We observed some common filters, such as docker's [4] or
systemd's [5], will make most decisions based only on the syscall
numbers, and as past discussions considered, a bitmap where each bit
represents a syscall makes most sense for these filters.

In order to build this bitmap at filter attach time, each filter is
emulated for every syscall (under each possible architecture), and
checked for any accesses of struct seccomp_data that are not the "arch"
nor "nr" (syscall) members. If only "arch" and "nr" are examined, and
the program returns allow, then we can be sure that the filter must
return allow independent from syscall arguments.

When it is concluded that an allow must occur for the given
architecture and syscall pair, seccomp will immediately allow
the syscall, bypassing further BPF execution.

Ongoing work is to further support arguments with fast hash table
lookups. We are investigating the performance of doing so [6], and how
to best integrate with the existing seccomp infrastructure.

Some benchmarks are performed with results in patch 5, copied below:
  Current BPF sysctl settings:
  net.core.bpf_jit_enable = 1
  net.core.bpf_jit_harden = 0
  Benchmarking 200000000 syscalls...
  129.359381409 - 0.008724424 = 129350656985 (129.4s)
  getpid native: 646 ns
  264.385890006 - 129.360453229 = 135025436777 (135.0s)
  getpid RET_ALLOW 1 filter (bitmap): 675 ns
  399.400511893 - 264.387045901 = 135013465992 (135.0s)
  getpid RET_ALLOW 2 filters (bitmap): 675 ns
  545.872866260 - 399.401718327 = 146471147933 (146.5s)
  getpid RET_ALLOW 3 filters (full): 732 ns
  696.337101319 - 545.874097681 = 150463003638 (150.5s)
  getpid RET_ALLOW 4 filters (full): 752 ns
  Estimated total seccomp overhead for 1 bitmapped filter: 29 ns
  Estimated total seccomp overhead for 2 bitmapped filters: 29 ns
  Estimated total seccomp overhead for 3 full filters: 86 ns
  Estimated total seccomp overhead for 4 full filters: 106 ns
  Estimated seccomp entry overhead: 29 ns
  Estimated seccomp per-filter overhead (last 2 diff): 20 ns
  Estimated seccomp per-filter overhead (filters / 4): 19 ns
  Expectations:
  	native ≤ 1 bitmap (646 ≤ 675): ✔️
  	native ≤ 1 filter (646 ≤ 732): ✔️
  	per-filter (last 2 diff) ≈ per-filter (filters / 4) (20 ≈ 19): ✔️
  	1 bitmapped ≈ 2 bitmapped (29 ≈ 29): ✔️
  	entry ≈ 1 bitmapped (29 ≈ 29): ✔️
  	entry ≈ 2 bitmapped (29 ≈ 29): ✔️
  	native + entry + (per filter * 4) ≈ 4 filters total (755 ≈ 752): ✔️

v2 -> v3:
* Added array_index_nospec guards
* No more syscall_arches[] array and expecting on loop unrolling. Arches
  are configured with per-arch seccomp.h.
* Moved filter emulation to attach time (from prepare time).
* Further simplified emulator, basing on Kees's code.
* Guard /proc/pid/seccomp_cache with CAP_SYS_ADMIN.

v1 -> v2:
* Corrected one outdated function documentation.

RFC -> v1:
* Config made on by default across all arches that could support it.
* Added arch numbers array and emulate filter for each arch number, and
  have a per-arch bitmap.
* Massively simplified the emulator so it would only support the common
  instructions in Kees's list.
* Fixed inheriting bitmap across filters (filter->prev is always NULL
  during prepare).
* Stole the selftest from Kees.
* Added a /proc/pid/seccomp_cache by Jann's suggestion.

Patch 1 adds the arch macros for x86.

Patch 2 implements the emulator that finds if a filter must return allow,

Patch 3 implements the test_bit against the bitmaps.

Patch 4 updates the selftest to better show the new semantics.

Patch 5 implements /proc/pid/seccomp_cache.

[1] https://lore.kernel.org/linux-security-module/c22a6c3cefc2412cad00ae14c1371711@huawei.com/T/
[2] https://lore.kernel.org/lkml/202005181120.971232B7B@keescook/T/
[3] https://github.com/seccomp/libseccomp/issues/116
[4] https://github.com/moby/moby/blob/ae0ef82b90356ac613f329a8ef5ee42ca923417d/profiles/seccomp/default.json
[5] https://github.com/systemd/systemd/blob/6743a1caf4037f03dc51a1277855018e4ab61957/src/shared/seccomp-util.c#L270
[6] Draco: Architectural and Operating System Support for System Call Security
    https://tianyin.github.io/pub/draco.pdf, MICRO-53, Oct. 2020

Kees Cook (2):
  x86: Enable seccomp architecture tracking
  selftests/seccomp: Compare bitmap vs filter overhead

YiFei Zhu (3):
  seccomp/cache: Add "emulator" to check if filter is constant allow
  seccomp/cache: Lookup syscall allowlist for fast path
  seccomp/cache: Report cache data through /proc/pid/seccomp_cache

 arch/Kconfig                                  |  49 ++++
 arch/x86/Kconfig                              |   1 +
 arch/x86/include/asm/seccomp.h                |  15 +
 fs/proc/base.c                                |   3 +
 include/linux/seccomp.h                       |   5 +
 kernel/seccomp.c                              | 265 +++++++++++++++++-
 .../selftests/seccomp/seccomp_benchmark.c     | 151 ++++++++--
 tools/testing/selftests/seccomp/settings      |   2 +-
 8 files changed, 467 insertions(+), 24 deletions(-)

--
2.28.0