Re: [PATCH bpf-next 0/7] bpf: Inline IRQ enter/exit instructions

["Kumar Kartikeya Dwivedi" <memxor@gmail.com>]

On Wed Oct 28, 2026 at 2:55 AM CET, Emil Tsalapatis wrote:
> On Tue, Oct 27, 2026 at 4:53 PM Alexei Starovoitov
> <alexei.starovoitov@gmail.com> wrote:
>>
>> On Tue, Oct 27, 2026 at 09:28 AM Emil Tsalapatis <emil@etsalapatis.com> wrote:
>> > Add inlining for those functions. The inlining happens in two stages.
>> > First replace the call itself with a special BPF instruction denoting
>> > an IRQ disable-and save, or restore-and-enable. Those instructions
>> > are translated by the architecture-specific JIT into the corresponding
>> > operation.
>>
>> No. That doesn't scale. We cannot add new insns for every to-be-inlined kfunc.
>> See bpf_jit_inlines_helper_call().
>>
>
> Ack, the code below seems strictly better than reserving an instruction.
>
>> Like:
>>
>> diff --git a/arch/x86/net/bpf_jit_comp.c b/arch/x86/net/bpf_jit_comp.c
>> index 6a2fe3d..c1c9a1c 100644
>> --- a/arch/x86/net/bpf_jit_comp.c
>> +++ b/arch/x86/net/bpf_jit_comp.c
>> @@ -3495,6 +3495,17 @@ populate_extable:
>>                         }
>>                         if (!imm32)
>>                                 return -EINVAL;
>> +                       if (src_reg == BPF_PSEUDO_KFUNC_CALL && bpf_jit_inlines_kfunc_call(func)) {
>> +                               if (func == (u8 *)bpf_local_irq_save) {
>> +                                       /* pushfq; pop [rdi]; cli */
>> +                                       EMIT4(0x9C, 0x8F, 0x07, 0xFA);
>> +                               } else {
>> +                                       /* test dword ptr [rdi], X86_EFLAGS_IF; je 1f; sti; 1: */
>> +                                       EMIT2_off32(0xF7, 0x07, X86_EFLAGS_IF);
>> +                                       EMIT3(X86_JE, 1, 0xFB);
>> +                               }
>> +                               break;
>> +                       }
>>                         if (src_reg == BPF_PSEUDO_KFUNC_CALL) {
>>                                 fm = bpf_jit_find_kfunc_model(bpf_prog, insn);
>>                                 if (!fm)
>> @@ -5116,6 +5127,14 @@ bool bpf_jit_supports_kfunc_call(void)
>>         return true;
>>  }
>>
>> +bool bpf_jit_inlines_kfunc_call(void *func)
>> +{
>> +       if (!IS_ENABLED(CONFIG_BPF_SYSCALL) || IS_ENABLED(CONFIG_TRACE_IRQFLAGS) ||
>> +           cpu_feature_enabled(X86_FEATURE_XENPV))
>> +               return false;
>> +       return func == bpf_local_irq_save || func == bpf_local_irq_restore;
>> +}
>>
>> ...
>>
>> diff --git a/kernel/bpf/helpers.c b/kernel/bpf/helpers.c
>> index 648abe1..26b6a9d 100644
>> --- a/kernel/bpf/helpers.c
>> +++ b/kernel/bpf/helpers.c
>> @@ -5197,8 +5197,8 @@ BTF_ID_FLAGS(func, bpf_get_kmem_cache, KF_PERFMON)
>>  BTF_ID_FLAGS(func, bpf_iter_kmem_cache_new, KF_ITER_NEW | KF_SLEEPABLE)
>>  BTF_ID_FLAGS(func, bpf_iter_kmem_cache_next, KF_ITER_NEXT | KF_RET_NULL | KF_SLEEPABLE)
>>  BTF_ID_FLAGS(func, bpf_iter_kmem_cache_destroy, KF_ITER_DESTROY | KF_SLEEPABLE)
>> -BTF_ID_FLAGS(func, bpf_local_irq_save)
>> -BTF_ID_FLAGS(func, bpf_local_irq_restore)
>> +BTF_ID_FLAGS(func, bpf_local_irq_save, KF_FASTCALL)
>> +BTF_ID_FLAGS(func, bpf_local_irq_restore, KF_FASTCALL)
>>
>> and corresponding bits in bpf_get_call_summary()...
>>
>> but before doing any of that there gotta be performance numbers that it's worth doing.
>>
>
> Makes sense, I assume you mean more macro numbers from
> production-grade programs apart from ubench numbers.
>

While you are it, it would be useful for bpf_preempt_{disable,enable}() to get
similar treatment.

>> pw-bot: cr