Inline bpf_get_current_task() and bpf_get_current_task_btf() for x86_64
to obtain better performance. The instruction we use here is:

  65 48 8B 04 25 [offset] // mov rax, gs:[offset]

Signed-off-by: Menglong Dong <dongml2@chinatelecom.cn>
---
v2:
- check the variable type in emit_ldx_percpu_r0 with __verify_pcpu_ptr
- remove the usage of const_current_task
---
 arch/x86/net/bpf_jit_comp.c | 36 ++++++++++++++++++++++++++++++++++++
 1 file changed, 36 insertions(+)

diff --git a/arch/x86/net/bpf_jit_comp.c b/arch/x86/net/bpf_jit_comp.c
index e3b1c4b1d550..f5ff7c77aad7 100644
--- a/arch/x86/net/bpf_jit_comp.c
+++ b/arch/x86/net/bpf_jit_comp.c
@@ -1300,6 +1300,25 @@ static void emit_st_r12(u8 **pprog, u32 size, u32 dst_reg, int off, int imm)
 	emit_st_index(pprog, size, dst_reg, X86_REG_R12, off, imm);
 }
 
+static void __emit_ldx_percpu_r0(u8 **pprog, __force unsigned long ptr)
+{
+	u8 *prog = *pprog;
+
+	/* mov rax, gs:[ptr] */
+	EMIT2(0x65, 0x48);
+	EMIT2(0x8B, 0x04);
+	EMIT1(0x25);
+	EMIT((u32)ptr, 4);
+
+	*pprog = prog;
+}
+
+#define emit_ldx_percpu_r0(prog, variable)					\
+	do {									\
+		__verify_pcpu_ptr(&(variable));					\
+		__emit_ldx_percpu_r0(&prog, (__force unsigned long)&(variable));\
+	} while (0)
+
 static int emit_atomic_rmw(u8 **pprog, u32 atomic_op,
 			   u32 dst_reg, u32 src_reg, s16 off, u8 bpf_size)
 {
@@ -2441,6 +2460,12 @@ st:			if (is_imm8(insn->off))
 		case BPF_JMP | BPF_CALL: {
 			u8 *ip = image + addrs[i - 1];
 
+			if (insn->src_reg == 0 && (insn->imm == BPF_FUNC_get_current_task ||
+						   insn->imm == BPF_FUNC_get_current_task_btf)) {
+				emit_ldx_percpu_r0(prog, current_task);
+				break;
+			}
+
 			func = (u8 *) __bpf_call_base + imm32;
 			if (src_reg == BPF_PSEUDO_CALL && tail_call_reachable) {
 				LOAD_TAIL_CALL_CNT_PTR(stack_depth);
@@ -4082,3 +4107,14 @@ bool bpf_jit_supports_timed_may_goto(void)
 {
 	return true;
 }
+
+bool bpf_jit_inlines_helper_call(s32 imm)
+{
+	switch (imm) {
+	case BPF_FUNC_get_current_task:
+	case BPF_FUNC_get_current_task_btf:
+		return true;
+	default:
+		return false;
+	}
+}
-- 
2.52.0

On Sun, Jan 4, 2026 at 5:17 AM Menglong Dong <menglong8.dong@gmail.com> wrote:
>
> Inline bpf_get_current_task() and bpf_get_current_task_btf() for x86_64
> to obtain better performance. The instruction we use here is:
>
>   65 48 8B 04 25 [offset] // mov rax, gs:[offset]
>
> Signed-off-by: Menglong Dong <dongml2@chinatelecom.cn>
> ---
> v2:
> - check the variable type in emit_ldx_percpu_r0 with __verify_pcpu_ptr
> - remove the usage of const_current_task
> ---
>  arch/x86/net/bpf_jit_comp.c | 36 ++++++++++++++++++++++++++++++++++++
>  1 file changed, 36 insertions(+)
>
> diff --git a/arch/x86/net/bpf_jit_comp.c b/arch/x86/net/bpf_jit_comp.c
> index e3b1c4b1d550..f5ff7c77aad7 100644
> --- a/arch/x86/net/bpf_jit_comp.c
> +++ b/arch/x86/net/bpf_jit_comp.c
> @@ -1300,6 +1300,25 @@ static void emit_st_r12(u8 **pprog, u32 size, u32 dst_reg, int off, int imm)
>         emit_st_index(pprog, size, dst_reg, X86_REG_R12, off, imm);
>  }
>
> +static void __emit_ldx_percpu_r0(u8 **pprog, __force unsigned long ptr)
> +{
> +       u8 *prog = *pprog;
> +
> +       /* mov rax, gs:[ptr] */
> +       EMIT2(0x65, 0x48);
> +       EMIT2(0x8B, 0x04);
> +       EMIT1(0x25);
> +       EMIT((u32)ptr, 4);
> +
> +       *pprog = prog;
> +}

Why asm?
Let's use BPF_MOV64_PERCPU_REG() similar to the way
BPF_FUNC_get_smp_processor_id inlining is handled.

pw-bot: cr

On Sun, 2026-01-04 at 21:16 +0800, Menglong Dong wrote:
> Inline bpf_get_current_task() and bpf_get_current_task_btf() for x86_64
> to obtain better performance. The instruction we use here is:
> 
>   65 48 8B 04 25 [offset] // mov rax, gs:[offset]
> 
> Signed-off-by: Menglong Dong <dongml2@chinatelecom.cn>
> ---
> v2:
> - check the variable type in emit_ldx_percpu_r0 with __verify_pcpu_ptr
> - remove the usage of const_current_task
> ---
>  arch/x86/net/bpf_jit_comp.c | 36 ++++++++++++++++++++++++++++++++++++
>  1 file changed, 36 insertions(+)
> 
> diff --git a/arch/x86/net/bpf_jit_comp.c b/arch/x86/net/bpf_jit_comp.c
> index e3b1c4b1d550..f5ff7c77aad7 100644
> --- a/arch/x86/net/bpf_jit_comp.c
> +++ b/arch/x86/net/bpf_jit_comp.c
> @@ -1300,6 +1300,25 @@ static void emit_st_r12(u8 **pprog, u32 size, u32 dst_reg, int off, int imm)
>  	emit_st_index(pprog, size, dst_reg, X86_REG_R12, off, imm);
>  }
>  
> +static void __emit_ldx_percpu_r0(u8 **pprog, __force unsigned long ptr)
> +{
> +	u8 *prog = *pprog;
> +
> +	/* mov rax, gs:[ptr] */
> +	EMIT2(0x65, 0x48);
> +	EMIT2(0x8B, 0x04);
> +	EMIT1(0x25);
> +	EMIT((u32)ptr, 4);
> +
> +	*pprog = prog;
> +}
> +
> +#define emit_ldx_percpu_r0(prog, variable)					\
> +	do {									\
> +		__verify_pcpu_ptr(&(variable));					\
> +		__emit_ldx_percpu_r0(&prog, (__force unsigned long)&(variable));\
> +	} while (0)
> +
>  static int emit_atomic_rmw(u8 **pprog, u32 atomic_op,
>  			   u32 dst_reg, u32 src_reg, s16 off, u8 bpf_size)
>  {
> @@ -2441,6 +2460,12 @@ st:			if (is_imm8(insn->off))
>  		case BPF_JMP | BPF_CALL: {
>  			u8 *ip = image + addrs[i - 1];
>  
> +			if (insn->src_reg == 0 && (insn->imm == BPF_FUNC_get_current_task ||
> +						   insn->imm == BPF_FUNC_get_current_task_btf)) {

I think this should be guarded by IS_ENABLED(CONFIG_SMP).
The current.h:get_current() used
arch/x86/include/asm/percpu.h:this_cpu_read_stable() that is unrolled
to __raw_cpu_read_stable(), which uses __force_percpu_arg(), which uses
__force_percpu_prefix, which is defined differently depending on CONFIG_SMP.

> +				emit_ldx_percpu_r0(prog, current_task);
> +				break;
> +			}
> +
>  			func = (u8 *) __bpf_call_base + imm32;
>  			if (src_reg == BPF_PSEUDO_CALL && tail_call_reachable) {
>  				LOAD_TAIL_CALL_CNT_PTR(stack_depth);
> @@ -4082,3 +4107,14 @@ bool bpf_jit_supports_timed_may_goto(void)
>  {
>  	return true;
>  }
> +
> +bool bpf_jit_inlines_helper_call(s32 imm)
> +{
> +	switch (imm) {
> +	case BPF_FUNC_get_current_task:
> +	case BPF_FUNC_get_current_task_btf:
> +		return true;
> +	default:
> +		return false;
> +	}
> +}

On 2026/1/6 01:45 Eduard Zingerman <eddyz87@gmail.com> write:
> On Sun, 2026-01-04 at 21:16 +0800, Menglong Dong wrote:
> > Inline bpf_get_current_task() and bpf_get_current_task_btf() for x86_64
> > to obtain better performance. The instruction we use here is:
> > 
> >   65 48 8B 04 25 [offset] // mov rax, gs:[offset]
> > 
> > Signed-off-by: Menglong Dong <dongml2@chinatelecom.cn>
> > ---
> > v2:
> > - check the variable type in emit_ldx_percpu_r0 with __verify_pcpu_ptr
> > - remove the usage of const_current_task
> > ---
> >  arch/x86/net/bpf_jit_comp.c | 36 ++++++++++++++++++++++++++++++++++++
> >  1 file changed, 36 insertions(+)
> > 
> > diff --git a/arch/x86/net/bpf_jit_comp.c b/arch/x86/net/bpf_jit_comp.c
> > index e3b1c4b1d550..f5ff7c77aad7 100644
> > --- a/arch/x86/net/bpf_jit_comp.c
> > +++ b/arch/x86/net/bpf_jit_comp.c
> > @@ -1300,6 +1300,25 @@ static void emit_st_r12(u8 **pprog, u32 size, u32 dst_reg, int off, int imm)
> >  	emit_st_index(pprog, size, dst_reg, X86_REG_R12, off, imm);
> >  }
> >  
> > +static void __emit_ldx_percpu_r0(u8 **pprog, __force unsigned long ptr)
> > +{
> > +	u8 *prog = *pprog;
> > +
> > +	/* mov rax, gs:[ptr] */
> > +	EMIT2(0x65, 0x48);
> > +	EMIT2(0x8B, 0x04);
> > +	EMIT1(0x25);
> > +	EMIT((u32)ptr, 4);
> > +
> > +	*pprog = prog;
> > +}
> > +
> > +#define emit_ldx_percpu_r0(prog, variable)					\
> > +	do {									\
> > +		__verify_pcpu_ptr(&(variable));					\
> > +		__emit_ldx_percpu_r0(&prog, (__force unsigned long)&(variable));\
> > +	} while (0)
> > +
> >  static int emit_atomic_rmw(u8 **pprog, u32 atomic_op,
> >  			   u32 dst_reg, u32 src_reg, s16 off, u8 bpf_size)
> >  {
> > @@ -2441,6 +2460,12 @@ st:			if (is_imm8(insn->off))
> >  		case BPF_JMP | BPF_CALL: {
> >  			u8 *ip = image + addrs[i - 1];
> >  
> > +			if (insn->src_reg == 0 && (insn->imm == BPF_FUNC_get_current_task ||
> > +						   insn->imm == BPF_FUNC_get_current_task_btf)) {
> 
> I think this should be guarded by IS_ENABLED(CONFIG_SMP).
> The current.h:get_current() used
> arch/x86/include/asm/percpu.h:this_cpu_read_stable() that is unrolled
> to __raw_cpu_read_stable(), which uses __force_percpu_arg(), which uses
> __force_percpu_prefix, which is defined differently depending on CONFIG_SMP.

Yeah, I missed this part. I'll use BPF_MOV64_PERCPU_REG() in
the next version, which should avoid this problem.

Thanks!
Menglong Dong

> 
> > +				emit_ldx_percpu_r0(prog, current_task);
> > +				break;
> > +			}
> > +
> >  			func = (u8 *) __bpf_call_base + imm32;
> >  			if (src_reg == BPF_PSEUDO_CALL && tail_call_reachable) {
> >  				LOAD_TAIL_CALL_CNT_PTR(stack_depth);
> > @@ -4082,3 +4107,14 @@ bool bpf_jit_supports_timed_may_goto(void)
> >  {
> >  	return true;
> >  }
> > +
> > +bool bpf_jit_inlines_helper_call(s32 imm)
> > +{
> > +	switch (imm) {
> > +	case BPF_FUNC_get_current_task:
> > +	case BPF_FUNC_get_current_task_btf:
> > +		return true;
> > +	default:
> > +		return false;
> > +	}
> > +}
>