[PATCH bpf-next v2 0/4] Optimize bpf_csum_diff() and homogenize for all archs

Puranjay Mohan posted 4 patches 1 month ago
There is a newer version of this series
arch/parisc/lib/checksum.c                    |  13 +-
include/net/checksum.h                        |   6 +
lib/checksum.c                                |  11 +-
net/core/filter.c                             |  37 +-
.../selftests/bpf/prog_tests/test_csum_diff.c | 408 ++++++++++++++++++
.../selftests/bpf/progs/csum_diff_test.c      |  42 ++
.../bpf/progs/verifier_array_access.c         |   3 +-
7 files changed, 469 insertions(+), 51 deletions(-)
create mode 100644 tools/testing/selftests/bpf/prog_tests/test_csum_diff.c
create mode 100644 tools/testing/selftests/bpf/progs/csum_diff_test.c
[PATCH bpf-next v2 0/4] Optimize bpf_csum_diff() and homogenize for all archs
Posted by Puranjay Mohan 1 month ago
Changes in v2:
v1: https://lore.kernel.org/all/20241021122112.101513-1-puranjay@kernel.org/
- Remove the patch that adds the benchmark as it is not useful enough to be
  added to the tree.
- Fixed a sparse warning in patch 1.
- Add reviewed-by and acked-by tags.

NOTE: There are some sparse warning in net/core/filter.c but those are not
worth fixing because bpf helpers take and return u64 values and using them
in csum related functions that take and return __sum16 / __wsum would need
a lot of casts everywhere.

The bpf_csum_diff() helper currently returns different values on different
architectures because it calls csum_partial() that is either implemented by
the architecture like x86_64, arm, etc or uses the generic implementation
in lib/checksum.c like arm64, riscv, etc.

The implementation in lib/checksum.c returns the folded result that is
16-bit long, but the architecture specific implementation can return an
unfolded value that is larger than 16-bits.

The helper uses a per-cpu scratchpad buffer for copying the data and then
computing the csum on this buffer. This can be optimised by utilising some
mathematical properties of csum.

The patch 1 in this series does preparatory work for homogenizing the
helper. patch 2 does the changes to the helper itself. The performance gain
can be seen in the tables below that are generated using the benchmark
built in patch 4 of v1 of this series:

  x86-64:
  +-------------+------------------+------------------+-------------+
  | Buffer Size |      Before      |      After       | Improvement |
  +-------------+------------------+------------------+-------------+
  |      4      | 2.296 ± 0.066M/s | 3.415 ± 0.001M/s |   48.73  %  |
  |      8      | 2.320 ± 0.003M/s | 3.409 ± 0.003M/s |   46.93  %  |
  |      16     | 2.315 ± 0.001M/s | 3.414 ± 0.003M/s |   47.47  %  |
  |      20     | 2.318 ± 0.001M/s | 3.416 ± 0.001M/s |   47.36  %  |
  |      32     | 2.308 ± 0.003M/s | 3.413 ± 0.003M/s |   47.87  %  |
  |      40     | 2.300 ± 0.029M/s | 3.413 ± 0.003M/s |   48.39  %  |
  |      64     | 2.286 ± 0.001M/s | 3.410 ± 0.001M/s |   49.16  %  |
  |      128    | 2.250 ± 0.001M/s | 3.404 ± 0.001M/s |   51.28  %  |
  |      256    | 2.173 ± 0.001M/s | 3.383 ± 0.001M/s |   55.68  %  |
  |      512    | 2.023 ± 0.055M/s | 3.340 ± 0.001M/s |   65.10  %  |
  +-------------+------------------+------------------+-------------+

  ARM64:
  +-------------+------------------+------------------+-------------+
  | Buffer Size |      Before      |      After       | Improvement |
  +-------------+------------------+------------------+-------------+
  |      4      | 1.397 ± 0.005M/s | 1.493 ± 0.005M/s |    6.87  %  |
  |      8      | 1.402 ± 0.002M/s | 1.489 ± 0.002M/s |    6.20  %  |
  |      16     | 1.391 ± 0.001M/s | 1.481 ± 0.001M/s |    6.47  %  |
  |      20     | 1.379 ± 0.001M/s | 1.477 ± 0.001M/s |    7.10  %  |
  |      32     | 1.358 ± 0.001M/s | 1.469 ± 0.002M/s |    8.17  %  |
  |      40     | 1.339 ± 0.001M/s | 1.462 ± 0.002M/s |    9.18  %  |
  |      64     | 1.302 ± 0.002M/s | 1.449 ± 0.003M/s |    11.29 %  |
  |      128    | 1.214 ± 0.001M/s | 1.443 ± 0.003M/s |    18.86 %  |
  |      256    | 1.080 ± 0.001M/s | 1.423 ± 0.001M/s |    31.75 %  |
  |      512    | 0.887 ± 0.001M/s | 1.411 ± 0.002M/s |    59.07 %  |
  +-------------+------------------+------------------+-------------+

Patch 3 reverts a hack that was done to make the selftest pass on all
architectures.

Patch 4 adds a selftest for this helper to verify the results produced by
this helper in multiple modes and edge cases.

Puranjay Mohan (4):
  net: checksum: move from32to16() to generic header
  bpf: bpf_csum_diff: optimize and homogenize for all archs
  selftests/bpf: don't mask result of bpf_csum_diff() in test_verifier
  selftests/bpf: Add a selftest for bpf_csum_diff()

 arch/parisc/lib/checksum.c                    |  13 +-
 include/net/checksum.h                        |   6 +
 lib/checksum.c                                |  11 +-
 net/core/filter.c                             |  37 +-
 .../selftests/bpf/prog_tests/test_csum_diff.c | 408 ++++++++++++++++++
 .../selftests/bpf/progs/csum_diff_test.c      |  42 ++
 .../bpf/progs/verifier_array_access.c         |   3 +-
 7 files changed, 469 insertions(+), 51 deletions(-)
 create mode 100644 tools/testing/selftests/bpf/prog_tests/test_csum_diff.c
 create mode 100644 tools/testing/selftests/bpf/progs/csum_diff_test.c

-- 
2.40.1