bpf: make kprobe_multi_link_prog_run always_inline

Make kprobe_multi_link_prog_run() always inline to obtain better
performance. Before this patch, the bench performance is:

./bench trig-kprobe-multi
Setting up benchmark 'trig-kprobe-multi'...
Benchmark 'trig-kprobe-multi' started.
Iter   0 ( 95.485us): hits   62.462M/s ( 62.462M/prod), [...]
Iter   1 (-80.054us): hits   62.486M/s ( 62.486M/prod), [...]
Iter   2 ( 13.572us): hits   62.287M/s ( 62.287M/prod), [...]
Iter   3 ( 76.961us): hits   62.293M/s ( 62.293M/prod), [...]
Iter   4 (-77.698us): hits   62.394M/s ( 62.394M/prod), [...]
Iter   5 (-13.399us): hits   62.319M/s ( 62.319M/prod), [...]
Iter   6 ( 77.573us): hits   62.250M/s ( 62.250M/prod), [...]
Summary: hits   62.338 ± 0.083M/s ( 62.338M/prod)

And after this patch, the performance is:

Iter   0 (454.148us): hits   66.900M/s ( 66.900M/prod), [...]
Iter   1 (-435.540us): hits   68.925M/s ( 68.925M/prod), [...]
Iter   2 (  8.223us): hits   68.795M/s ( 68.795M/prod), [...]
Iter   3 (-12.347us): hits   68.880M/s ( 68.880M/prod), [...]
Iter   4 (  2.291us): hits   68.767M/s ( 68.767M/prod), [...]
Iter   5 ( -1.446us): hits   68.756M/s ( 68.756M/prod), [...]
Iter   6 ( 13.882us): hits   68.657M/s ( 68.657M/prod), [...]
Summary: hits   68.792 ± 0.087M/s ( 68.792M/prod)

As we can see, the performance of kprobe-multi increase from 62M/s to
68M/s.

Signed-off-by: Menglong Dong <dongml2@chinatelecom.cn>
Link: https://lore.kernel.org/r/20251126085246.309942-1-dongml2@chinatelecom.cn
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
This commit is contained in:
Menglong Dong
2025-11-26 16:52:46 +08:00
committed by Alexei Starovoitov
parent 85bdeeb48f
commit c1af4465b9

View File

@@ -2529,7 +2529,7 @@ static u64 bpf_kprobe_multi_entry_ip(struct bpf_run_ctx *ctx)
return run_ctx->entry_ip;
}
static int
static __always_inline int
kprobe_multi_link_prog_run(struct bpf_kprobe_multi_link *link,
unsigned long entry_ip, struct ftrace_regs *fregs,
bool is_return, void *data)