[PATCH bpf-next v3 0/2] bpf: Optimize recursion detection on arm64

[Puranjay Mohan <puranjay@kernel.org>]

V2: https://lore.kernel.org/all/20251217233608.2374187-1-puranjay@kernel.org/
Changes in v2->v3:
- Added acked by Yonghong
- Patch 2:
        - Change alignment of active from 8 to 4
        - Use le32_to_cpu in place of get_unaligned_le32()

V1: https://lore.kernel.org/all/20251217162830.2597286-1-puranjay@kernel.org/
Changes in V1->V2:
- Patch 2:
        - Put preempt_enable()/disable() around RMW accesses to mitigate
          race conditions. Because on CONFIG_PREEMPT_RCU and sleepable
	  bpf programs, preemption can cause no bpf prog to execute in
	  case of recursion.

BPF programs detect recursion using a per-CPU 'active' flag in struct
bpf_prog. The trampoline currently sets/clears this flag with atomic
operations.

On some arm64 platforms (e.g., Neoverse V2 with LSE), per-CPU atomic
operations are relatively slow. Unlike x86_64 - where per-CPU updates
can avoid cross-core atomicity, arm64 LSE atomics are always atomic
across all cores, which is unnecessary overhead for strictly per-CPU
state.

This patch removes atomics from the recursion detection path on arm64.

It was discovered in [1] that per-CPU atomics that don't return a value
were extremely slow on some arm64 platforms, Catalin added a fix in
commit 535fdfc5a228 ("arm64: Use load LSE atomics for the non-return
per-CPU atomic operations") to solve this issue, but it seems to have
caused a regression on the fentry benchmark.

Using the fentry benchmark from the bpf selftests shows the following:

  ./tools/testing/selftests/bpf/bench trig-fentry

 +---------------------------------------------+------------------------+
 |               Configuration                 | Total Operations (M/s) |
 +---------------------------------------------+------------------------+
 | bpf-next/master with Catalin’s fix reverted |         51.770         |
 |---------------------------------------------|------------------------|
 | bpf-next/master                             |         43.271         |
 | bpf-next/master with this change            |         43.271         |
 +---------------------------------------------+------------------------+

All benchmarks were run on a KVM based vm with Neoverse-V2 and 8 cpus.

This patch yields a 25% improvement in this benchmark compared to
bpf-next. Notably, reverting Catalin's fix also results in a performance
gain for this benchmark, which is interesting but expected.

For completeness, this benchmark was also run with the change enabled on
x86-64, which resulted in a 30% regression in the fentry benchmark. So,
it is only enabled on arm64.

P.S. - Here is more data with other program types:

 +-----------------+-----------+-----------+----------+
 |     Metric      |  Before   |   After   | % Diff   |
 +-----------------+-----------+-----------+----------+
 | fentry          |   43.149  |   53.948  | +25.03%  |
 | fentry.s        |   41.831  |   50.937  | +21.76%  |
 | rawtp           |   50.834  |   58.731  | +15.53%  |
 | fexit           |   31.118  |   34.360  | +10.42%  |
 | tp              |   39.536  |   41.632  |  +5.30%  |
 | syscall-count   |    8.053  |    8.305  |  +3.13%  |
 | fmodret         |   33.940  |   34.769  |  +2.44%  |
 | kprobe          |    9.970  |    9.998  |  +0.28%  |
 | usermode-count  |  224.886  |  224.839  |  -0.02%  |
 | kernel-count    |  154.229  |  153.043  |  -0.77%  |
 +-----------------+-----------+-----------+----------+

[1] https://lore.kernel.org/all/e7d539ed-ced0-4b96-8ecd-048a5b803b85@paulmck-laptop/

Puranjay Mohan (2):
  bpf: move recursion detection logic to helpers
  bpf: arm64: Optimize recursion detection by not using atomics

 include/linux/bpf.h      | 38 +++++++++++++++++++++++++++++++++++++-
 kernel/bpf/core.c        |  3 ++-
 kernel/bpf/trampoline.c  |  8 ++++----
 kernel/trace/bpf_trace.c |  4 ++--
 4 files changed, 45 insertions(+), 8 deletions(-)


base-commit: ec439c38013550420aecc15988ae6acb670838c1
-- 
2.47.3