On 25/10/24 06:09, Alexei Starovoitov wrote:
> On Mon, Oct 21, 2024 at 6:39 AM Leon Hwang <leon.hwang@xxxxxxxxx> wrote:
>>
>> In the x86_64 JIT, when calling a function, tailcall info is propagated if
>> the program is tail_call_reachable, regardless of whether the function is a
>> subprog, helper, or kfunc. However, this propagation is unnecessary for
>> not-tail_call_reachable subprogs, helpers, or kfuncs.
>>
>> The verifier can determine if a subprog is tail_call_reachable. Therefore,
>> it can be optimized to only propagate tailcall info when the callee is
>> subprog and the subprog is actually tail_call_reachable.
>>
>> Signed-off-by: Leon Hwang <leon.hwang@xxxxxxxxx>
>> ---
>> arch/x86/net/bpf_jit_comp.c | 4 +++-
>> kernel/bpf/verifier.c | 6 ++++++
>> 2 files changed, 9 insertions(+), 1 deletion(-)
>>
>> diff --git a/arch/x86/net/bpf_jit_comp.c b/arch/x86/net/bpf_jit_comp.c
>> index 06b080b61aa57..6ad6886ecfc88 100644
>> --- a/arch/x86/net/bpf_jit_comp.c
>> +++ b/arch/x86/net/bpf_jit_comp.c
>> @@ -2124,10 +2124,12 @@ st: if (is_imm8(insn->off))
>>
>> /* call */
>> case BPF_JMP | BPF_CALL: {
>> + bool pseudo_call = src_reg == BPF_PSEUDO_CALL;
>> + bool subprog_tail_call_reachable = dst_reg;
>> u8 *ip = image + addrs[i - 1];
>>
>> func = (u8 *) __bpf_call_base + imm32;
>> - if (tail_call_reachable) {
>> + if (pseudo_call && subprog_tail_call_reachable) {
>> LOAD_TAIL_CALL_CNT_PTR(bpf_prog->aux->stack_depth);
>> ip += 7;
>> }
>> diff --git a/kernel/bpf/verifier.c b/kernel/bpf/verifier.c
>> index f514247ba8ba8..6e7e42c7bc7b1 100644
>> --- a/kernel/bpf/verifier.c
>> +++ b/kernel/bpf/verifier.c
>> @@ -19990,6 +19990,12 @@ static int jit_subprogs(struct bpf_verifier_env *env)
>> insn[0].imm = (u32)addr;
>> insn[1].imm = addr >> 32;
>> }
>> +
>> + if (bpf_pseudo_call(insn))
>> + /* In the x86_64 JIT, tailcall information can only be
>> + * propagated if the subprog is tail_call_reachable.
>> + */
>> + insn->dst_reg = env->subprog_info[subprog].tail_call_reachable;
>
> I really don't like hacking flags into dst_reg.
> We already abuse insn->off which is ugly too,
> but at least we clean insns later after JIT.
>
> I'd rather live with this tail call inefficiency than abuse insns
> fields further.
>
OK, let us use 'pseudo_call && tail_call_reachable' in x86 JIT to avoid
touching 'insn->dst_reg'.
Thanks,
Leon
> pw-bot: cr
