Function bpf_patch_insn_data() uses vzalloc/vfree pair to allocate
memory for updated insn_aux_data. These operations are expensive for
big programs where a lot of rewrites happen, e.g. for pyperf180 test
case. The pair can be replaced with a call to vrealloc in order to
reduce the number of actual memory allocations.
Perf stat w/o this patch:
$ perf stat -B --all-kernel -r10 -- ./veristat -q pyperf180.bpf.o
...
2201.25 msec task-clock # 0.973 CPUs utilized ( +- 2.20% )
188 context-switches # 85.406 /sec ( +- 9.29% )
15 cpu-migrations # 6.814 /sec ( +- 5.64% )
5 page-faults # 2.271 /sec ( +- 3.27% )
4315057974 instructions # 1.28 insn per cycle
# 0.33 stalled cycles per insn ( +- 0.03% )
3366141387 cycles # 1.529 GHz ( +- 0.21% )
1420810964 stalled-cycles-frontend # 42.21% frontend cycles idle ( +- 0.23% )
1049956791 branches # 476.981 M/sec ( +- 0.03% )
60591781 branch-misses # 5.77% of all branches ( +- 0.07% )
2.2632 +- 0.0527 seconds time elapsed ( +- 2.33% )
Perf stat with this patch:
1227.15 msec task-clock # 0.963 CPUs utilized ( +- 2.27% )
170 context-switches # 138.532 /sec ( +- 5.62% )
2 cpu-migrations # 1.630 /sec ( +- 33.37% )
5 page-faults # 4.074 /sec ( +- 4.47% )
3312254304 instructions # 2.17 insn per cycle
# 0.15 stalled cycles per insn ( +- 0.03% )
1528944717 cycles # 1.246 GHz ( +- 0.31% )
501475146 stalled-cycles-frontend # 32.80% frontend cycles idle ( +- 0.50% )
730426891 branches # 595.222 M/sec ( +- 0.03% )
17372363 branch-misses # 2.38% of all branches ( +- 0.16% )
1.2744 +- 0.0301 seconds time elapsed ( +- 2.36% )
Changelog:
v1: https://lore.kernel.org/bpf/20250806200928.3080531-1-eddyz87@xxxxxxxxx/T/#t
v1 -> v2:
- added missing memset(0) in adjust_insn_aux_data(),
this fixes CI failure reported in [1].
[1] https://github.com/kernel-patches/bpf/actions/runs/16787563163/job/47542309875
Eduard Zingerman (2):
bpf: removed unused 'env' parameter from is_reg64 and insn_has_def32
bpf: use realloc in bpf_patch_insn_data
kernel/bpf/verifier.c | 39 ++++++++++++++++++++-------------------
1 file changed, 20 insertions(+), 19 deletions(-)
--
2.47.3
