[PATCH bpf-next v2 0/2] bpf: Inline helper in powerpc JIT

[PATCH bpf-next v2 0/2] bpf: Inline helper in powerpc JIT

[Saket Kumar Bhaskar]

This series add support for internal only per-CPU instructions,
inlines the bpf_get_smp_processor_id() and bpf_get_current_task()
helper calls for powerpc BPF JIT.

Changes since v1:
* Addressed Christophe's comments.
* Inlined bpf_get_current_task() as well.

v1: https://lore.kernel.org/all/20250311160955.825647-1-skb99@linux.ibm.com/ 

Saket Kumar Bhaskar (2):
  powerpc64/bpf: Support internal-only MOV instruction to resolve
    per-CPU addrs
  powerpc64/bpf: Inline bpf_get_smp_processor_id() and
    bpf_get_current_task()

 arch/powerpc/net/bpf_jit_comp.c   | 16 ++++++++++++++++
 arch/powerpc/net/bpf_jit_comp64.c | 19 +++++++++++++++++++
 2 files changed, 35 insertions(+)

-- 
2.51.0

[PATCH bpf-next v2 1/2] powerpc64/bpf: Support internal-only MOV instruction to resolve per-CPU addrs

[Saket Kumar Bhaskar]

With the introduction of commit 7bdbf7446305 ("bpf: add special
internal-only MOV instruction to resolve per-CPU addrs"),
a new BPF instruction BPF_MOV64_PERCPU_REG has been added to
resolve absolute addresses of per-CPU data from their per-CPU
offsets. This update requires enabling support for this
instruction in the powerpc JIT compiler.

As of commit 7a0268fa1a36 ("[PATCH] powerpc/64: per cpu data
optimisations"), the per-CPU data offset for the CPU is stored in
the paca.

To support this BPF instruction in the powerpc JIT, the following
powerpc instructions are emitted:

ld tmp1_reg, 48(13)		//Load per-CPU data offset from paca(r13) in tmp1_reg.
add dst_reg, src_reg, tmp1_reg	//Add the per cpu offset to the dst.
mr dst_reg, src_reg		//Move src_reg to dst_reg, if src_reg != dst_reg

To evaluate the performance improvements introduced by this change,
the benchmark described in [1] was employed.

Before Change:
glob-arr-inc   :   41.580 ± 0.034M/s
arr-inc        :   39.592 ± 0.055M/s
hash-inc       :   25.873 ± 0.012M/s

After Change:
glob-arr-inc   :   42.024 ± 0.049M/s
arr-inc        :   55.447 ± 0.031M/s
hash-inc       :   26.565 ± 0.014M/s

[1] https://github.com/anakryiko/linux/commit/8dec900975ef

Signed-off-by: Saket Kumar Bhaskar <skb99@linux.ibm.com>
---
 arch/powerpc/net/bpf_jit_comp.c   | 5 +++++
 arch/powerpc/net/bpf_jit_comp64.c | 9 +++++++++
 2 files changed, 14 insertions(+)

diff --git a/arch/powerpc/net/bpf_jit_comp.c b/arch/powerpc/net/bpf_jit_comp.c
index 88ad5ba7b87f..2f2230ae2145 100644
--- a/arch/powerpc/net/bpf_jit_comp.c
+++ b/arch/powerpc/net/bpf_jit_comp.c
@@ -466,6 +466,11 @@ bool bpf_jit_supports_insn(struct bpf_insn *insn, bool in_arena)
 	return true;
 }
 
+bool bpf_jit_supports_percpu_insn(void)
+{
+	return IS_ENABLED(CONFIG_PPC64);
+}
+
 void *arch_alloc_bpf_trampoline(unsigned int size)
 {
 	return bpf_prog_pack_alloc(size, bpf_jit_fill_ill_insns);
diff --git a/arch/powerpc/net/bpf_jit_comp64.c b/arch/powerpc/net/bpf_jit_comp64.c
index 1fe37128c876..21486706b5ea 100644
--- a/arch/powerpc/net/bpf_jit_comp64.c
+++ b/arch/powerpc/net/bpf_jit_comp64.c
@@ -918,6 +918,15 @@ int bpf_jit_build_body(struct bpf_prog *fp, u32 *image, u32 *fimage, struct code
 		case BPF_ALU | BPF_MOV | BPF_X: /* (u32) dst = src */
 		case BPF_ALU64 | BPF_MOV | BPF_X: /* dst = src */
 
+			if (insn_is_mov_percpu_addr(&insn[i])) {
+				if (IS_ENABLED(CONFIG_SMP)) {
+					EMIT(PPC_RAW_LD(tmp1_reg, _R13, offsetof(struct paca_struct, data_offset)));
+					EMIT(PPC_RAW_ADD(dst_reg, src_reg, tmp1_reg));
+				} else {
+					EMIT(PPC_RAW_MR(dst_reg, src_reg));
+				}
+			}
+
 			if (insn_is_cast_user(&insn[i])) {
 				EMIT(PPC_RAW_RLDICL_DOT(tmp1_reg, src_reg, 0, 32));
 				PPC_LI64(dst_reg, (ctx->user_vm_start & 0xffffffff00000000UL));
-- 
2.51.0

Re: [PATCH bpf-next v2 1/2] powerpc64/bpf: Support internal-only MOV instruction to resolve per-CPU addrs

[Christophe Leroy (CS GROUP)]

Le 17/11/2025 à 07:52, Saket Kumar Bhaskar a écrit :
> With the introduction of commit 7bdbf7446305 ("bpf: add special
> internal-only MOV instruction to resolve per-CPU addrs"),
> a new BPF instruction BPF_MOV64_PERCPU_REG has been added to
> resolve absolute addresses of per-CPU data from their per-CPU
> offsets. This update requires enabling support for this
> instruction in the powerpc JIT compiler.
> 
> As of commit 7a0268fa1a36 ("[PATCH] powerpc/64: per cpu data
> optimisations"), the per-CPU data offset for the CPU is stored in
> the paca.
> 
> To support this BPF instruction in the powerpc JIT, the following
> powerpc instructions are emitted:
> 
> ld tmp1_reg, 48(13)		//Load per-CPU data offset from paca(r13) in tmp1_reg.
> add dst_reg, src_reg, tmp1_reg	//Add the per cpu offset to the dst.
> mr dst_reg, src_reg		//Move src_reg to dst_reg, if src_reg != dst_reg

Must be something wrong here. The 'add' was done into the dst_reg so 
here you erase the result of the addition by the source register.

> 
> To evaluate the performance improvements introduced by this change,
> the benchmark described in [1] was employed.
> 
> Before Change:
> glob-arr-inc   :   41.580 ± 0.034M/s
> arr-inc        :   39.592 ± 0.055M/s
> hash-inc       :   25.873 ± 0.012M/s
> 
> After Change:
> glob-arr-inc   :   42.024 ± 0.049M/s
> arr-inc        :   55.447 ± 0.031M/s
> hash-inc       :   26.565 ± 0.014M/s
> 
> [1] https://eur01.safelinks.protection.outlook.com/?url=https%3A%2F%2Fgithub.com%2Fanakryiko%2Flinux%2Fcommit%2F8dec900975ef&data=05%7C02%7Cchristophe.leroy%40csgroup.eu%7C2f16cef7d35341c9683608de25a5ee3b%7C8b87af7d86474dc78df45f69a2011bb5%7C0%7C0%7C638989591756011820%7CUnknown%7CTWFpbGZsb3d8eyJFbXB0eU1hcGkiOnRydWUsIlYiOiIwLjAuMDAwMCIsIlAiOiJXaW4zMiIsIkFOIjoiTWFpbCIsIldUIjoyfQ%3D%3D%7C0%7C%7C%7C&sdata=dNHc3FAFJkZpHq%2Be1hTv5CfBrGEXxWTKrLGSHaUw%2BRk%3D&reserved=0
> 
> Signed-off-by: Saket Kumar Bhaskar <skb99@linux.ibm.com>
> ---
>   arch/powerpc/net/bpf_jit_comp.c   | 5 +++++
>   arch/powerpc/net/bpf_jit_comp64.c | 9 +++++++++
>   2 files changed, 14 insertions(+)
> 
> diff --git a/arch/powerpc/net/bpf_jit_comp.c b/arch/powerpc/net/bpf_jit_comp.c
> index 88ad5ba7b87f..2f2230ae2145 100644
> --- a/arch/powerpc/net/bpf_jit_comp.c
> +++ b/arch/powerpc/net/bpf_jit_comp.c
> @@ -466,6 +466,11 @@ bool bpf_jit_supports_insn(struct bpf_insn *insn, bool in_arena)
>   	return true;
>   }
>   
> +bool bpf_jit_supports_percpu_insn(void)
> +{
> +	return IS_ENABLED(CONFIG_PPC64);
> +}
> +
>   void *arch_alloc_bpf_trampoline(unsigned int size)
>   {
>   	return bpf_prog_pack_alloc(size, bpf_jit_fill_ill_insns);
> diff --git a/arch/powerpc/net/bpf_jit_comp64.c b/arch/powerpc/net/bpf_jit_comp64.c
> index 1fe37128c876..21486706b5ea 100644
> --- a/arch/powerpc/net/bpf_jit_comp64.c
> +++ b/arch/powerpc/net/bpf_jit_comp64.c
> @@ -918,6 +918,15 @@ int bpf_jit_build_body(struct bpf_prog *fp, u32 *image, u32 *fimage, struct code
>   		case BPF_ALU | BPF_MOV | BPF_X: /* (u32) dst = src */
>   		case BPF_ALU64 | BPF_MOV | BPF_X: /* dst = src */
>   
> +			if (insn_is_mov_percpu_addr(&insn[i])) {
> +				if (IS_ENABLED(CONFIG_SMP)) {
> +					EMIT(PPC_RAW_LD(tmp1_reg, _R13, offsetof(struct paca_struct, data_offset)));
> +					EMIT(PPC_RAW_ADD(dst_reg, src_reg, tmp1_reg));
> +				} else {
> +					EMIT(PPC_RAW_MR(dst_reg, src_reg));

You should make sure dst_reg is different from src_reg before emitting 
this, you may otherwise generate one of the following instructions that 
change the thread priority:

#define HMT_very_low()		asm volatile("or 31, 31, 31	# very low priority")
#define HMT_low()		asm volatile("or 1, 1, 1	# low priority")
#define HMT_medium_low()	asm volatile("or 6, 6, 6	# medium low priority")
#define HMT_medium()		asm volatile("or 2, 2, 2	# medium priority")
#define HMT_medium_high()	asm volatile("or 5, 5, 5	# medium high priority")
#define HMT_high()		asm volatile("or 3, 3, 3	# high priority")


> +				}
> +			}
> +
>   			if (insn_is_cast_user(&insn[i])) {
>   				EMIT(PPC_RAW_RLDICL_DOT(tmp1_reg, src_reg, 0, 32));
>   				PPC_LI64(dst_reg, (ctx->user_vm_start & 0xffffffff00000000UL));

[PATCH bpf-next v2 2/2] powerpc64/bpf: Inline bpf_get_smp_processor_id() and bpf_get_current_task()

[Saket Kumar Bhaskar]

Inline the calls to bpf_get_smp_processor_id()/bpf_get_current_task()
in the powerpc bpf jit.

powerpc saves the Logical processor number (paca_index) and pointer
to current task (__current) in paca.

Here is how the powerpc JITed assembly changes after this commit:

Before:

cpu = bpf_get_smp_processor_id();

addis 12, 2, -517
addi 12, 12, -29456
mtctr 12
bctrl
mr	8, 3

After:

cpu = bpf_get_smp_processor_id();

lhz 8, 8(13)

To evaluate the performance improvements introduced by this change,
the benchmark described in [1] was employed.

+---------------+-------------------+-------------------+--------------+
|      Name     |      Before       |        After      |   % change   |
|---------------+-------------------+-------------------+--------------|
| glob-arr-inc  | 40.701 ± 0.008M/s | 55.207 ± 0.021M/s |   + 35.64%   |
| arr-inc       | 39.401 ± 0.007M/s | 56.275 ± 0.023M/s |   + 42.42%   |
| hash-inc      | 24.944 ± 0.004M/s | 26.212 ± 0.003M/s |   +  5.08%   |
+---------------+-------------------+-------------------+--------------+

[1] https://github.com/anakryiko/linux/commit/8dec900975ef

Signed-off-by: Saket Kumar Bhaskar <skb99@linux.ibm.com>
---
 arch/powerpc/net/bpf_jit_comp.c   | 11 +++++++++++
 arch/powerpc/net/bpf_jit_comp64.c | 10 ++++++++++
 2 files changed, 21 insertions(+)

diff --git a/arch/powerpc/net/bpf_jit_comp.c b/arch/powerpc/net/bpf_jit_comp.c
index 2f2230ae2145..c88dfa1418ec 100644
--- a/arch/powerpc/net/bpf_jit_comp.c
+++ b/arch/powerpc/net/bpf_jit_comp.c
@@ -471,6 +471,17 @@ bool bpf_jit_supports_percpu_insn(void)
 	return IS_ENABLED(CONFIG_PPC64);
 }
 
+bool bpf_jit_inlines_helper_call(s32 imm)
+{
+	switch (imm) {
+	case BPF_FUNC_get_smp_processor_id:
+	case BPF_FUNC_get_current_task:
+		return true;
+	default:
+		return false;
+	}
+}
+
 void *arch_alloc_bpf_trampoline(unsigned int size)
 {
 	return bpf_prog_pack_alloc(size, bpf_jit_fill_ill_insns);
diff --git a/arch/powerpc/net/bpf_jit_comp64.c b/arch/powerpc/net/bpf_jit_comp64.c
index 21486706b5ea..4e1643422370 100644
--- a/arch/powerpc/net/bpf_jit_comp64.c
+++ b/arch/powerpc/net/bpf_jit_comp64.c
@@ -1399,6 +1399,16 @@ int bpf_jit_build_body(struct bpf_prog *fp, u32 *image, u32 *fimage, struct code
 		case BPF_JMP | BPF_CALL:
 			ctx->seen |= SEEN_FUNC;
 
+			if (insn[i].src_reg == BPF_REG_0) {
+				if (imm == BPF_FUNC_get_smp_processor_id) {
+					EMIT(PPC_RAW_LHZ(insn[i].src_reg, _R13, offsetof(struct paca_struct, paca_index)));
+					break;
+				} else if (imm == BPF_FUNC_get_current_task) {
+					EMIT(PPC_RAW_LD(insn[i].src_reg, _R13, offsetof(struct paca_struct, __current)));
+					break;
+				}
+			}
+
 			ret = bpf_jit_get_func_addr(fp, &insn[i], extra_pass,
 						    &func_addr, &func_addr_fixed);
 			if (ret < 0)
-- 
2.51.0

Re: [PATCH bpf-next v2 2/2] powerpc64/bpf: Inline bpf_get_smp_processor_id() and bpf_get_current_task()

[Christophe Leroy (CS GROUP)]

Le 17/11/2025 à 07:52, Saket Kumar Bhaskar a écrit :
> Inline the calls to bpf_get_smp_processor_id()/bpf_get_current_task()
> in the powerpc bpf jit.
> 
> powerpc saves the Logical processor number (paca_index) and pointer
> to current task (__current) in paca.
> 
> Here is how the powerpc JITed assembly changes after this commit:
> 
> Before:
> 
> cpu = bpf_get_smp_processor_id();
> 
> addis 12, 2, -517
> addi 12, 12, -29456
> mtctr 12
> bctrl
> mr	8, 3
> 
> After:
> 
> cpu = bpf_get_smp_processor_id();
> 
> lhz 8, 8(13)
> 
> To evaluate the performance improvements introduced by this change,
> the benchmark described in [1] was employed.
> 
> +---------------+-------------------+-------------------+--------------+
> |      Name     |      Before       |        After      |   % change   |
> |---------------+-------------------+-------------------+--------------|
> | glob-arr-inc  | 40.701 ± 0.008M/s | 55.207 ± 0.021M/s |   + 35.64%   |
> | arr-inc       | 39.401 ± 0.007M/s | 56.275 ± 0.023M/s |   + 42.42%   |
> | hash-inc      | 24.944 ± 0.004M/s | 26.212 ± 0.003M/s |   +  5.08%   |
> +---------------+-------------------+-------------------+--------------+
> 
> [1] https://eur01.safelinks.protection.outlook.com/?url=https%3A%2F%2Fgithub.com%2Fanakryiko%2Flinux%2Fcommit%2F8dec900975ef&data=05%7C02%7Cchristophe.leroy%40csgroup.eu%7C4a08a3af41ff4f9bc55d08de25a5f0ee%7C8b87af7d86474dc78df45f69a2011bb5%7C0%7C0%7C638989591794687135%7CUnknown%7CTWFpbGZsb3d8eyJFbXB0eU1hcGkiOnRydWUsIlYiOiIwLjAuMDAwMCIsIlAiOiJXaW4zMiIsIkFOIjoiTWFpbCIsIldUIjoyfQ%3D%3D%7C0%7C%7C%7C&sdata=FtfTYpm9VgLfO1Q3iZvyrE4QRG317%2B%2BjfPd66Wd%2FQP4%3D&reserved=0
> 
> Signed-off-by: Saket Kumar Bhaskar <skb99@linux.ibm.com>
> ---
>   arch/powerpc/net/bpf_jit_comp.c   | 11 +++++++++++
>   arch/powerpc/net/bpf_jit_comp64.c | 10 ++++++++++
>   2 files changed, 21 insertions(+)
> 
> diff --git a/arch/powerpc/net/bpf_jit_comp.c b/arch/powerpc/net/bpf_jit_comp.c
> index 2f2230ae2145..c88dfa1418ec 100644
> --- a/arch/powerpc/net/bpf_jit_comp.c
> +++ b/arch/powerpc/net/bpf_jit_comp.c
> @@ -471,6 +471,17 @@ bool bpf_jit_supports_percpu_insn(void)
>   	return IS_ENABLED(CONFIG_PPC64);
>   }
>   
> +bool bpf_jit_inlines_helper_call(s32 imm)
> +{
> +	switch (imm) {
> +	case BPF_FUNC_get_smp_processor_id:
> +	case BPF_FUNC_get_current_task:

What about BPF_FUNC_get_current_task_btf ?

> +		return true;
> +	default:
> +		return false;
> +	}
> +}
> +
>   void *arch_alloc_bpf_trampoline(unsigned int size)
>   {
>   	return bpf_prog_pack_alloc(size, bpf_jit_fill_ill_insns);
> diff --git a/arch/powerpc/net/bpf_jit_comp64.c b/arch/powerpc/net/bpf_jit_comp64.c
> index 21486706b5ea..4e1643422370 100644
> --- a/arch/powerpc/net/bpf_jit_comp64.c
> +++ b/arch/powerpc/net/bpf_jit_comp64.c
> @@ -1399,6 +1399,16 @@ int bpf_jit_build_body(struct bpf_prog *fp, u32 *image, u32 *fimage, struct code
>   		case BPF_JMP | BPF_CALL:
>   			ctx->seen |= SEEN_FUNC;
>   
> +			if (insn[i].src_reg == BPF_REG_0) {

Are you sure you want to use BPF_REG_0 here ? Is it the correct meaning 
? I see RISCV and ARM64 use 0 instead.

If you keep BPF_REG_0 I would have a preference for

		if (src_reg == bpf_to_ppc(BPF_REG_0))

> +				if (imm == BPF_FUNC_get_smp_processor_id) {
> +					EMIT(PPC_RAW_LHZ(insn[i].src_reg, _R13, offsetof(struct paca_struct, paca_index)));

This looks wrong, you can't use insn[i].src_reg to emit powerpc 
instructions, you must use the local src_reg which converts the register 
ID with bpf_to_ppc()

> +					break;
> +				} else if (imm == BPF_FUNC_get_current_task) {
> +					EMIT(PPC_RAW_LD(insn[i].src_reg, _R13, offsetof(struct paca_struct, __current)));

Same here.

> +					break;
> +				}
> +			}
> +
>   			ret = bpf_jit_get_func_addr(fp, &insn[i], extra_pass,
>   						    &func_addr, &func_addr_fixed);
>   			if (ret < 0)

Re: [PATCH bpf-next v2 0/2] bpf: Inline helper in powerpc JIT

[Puranjay Mohan]

Saket Kumar Bhaskar <skb99@linux.ibm.com> writes:

> This series add support for internal only per-CPU instructions,
> inlines the bpf_get_smp_processor_id() and bpf_get_current_task()
> helper calls for powerpc BPF JIT.
>
> Changes since v1:
> * Addressed Christophe's comments.
> * Inlined bpf_get_current_task() as well.
>
> v1: https://lore.kernel.org/all/20250311160955.825647-1-skb99@linux.ibm.com/ 
>
> Saket Kumar Bhaskar (2):
>   powerpc64/bpf: Support internal-only MOV instruction to resolve
>     per-CPU addrs
>   powerpc64/bpf: Inline bpf_get_smp_processor_id() and
>     bpf_get_current_task()
>
>  arch/powerpc/net/bpf_jit_comp.c   | 16 ++++++++++++++++
>  arch/powerpc/net/bpf_jit_comp64.c | 19 +++++++++++++++++++
>  2 files changed, 35 insertions(+)

For both the patches:

Reviewed-by: Puranjay Mohan <puranjay@kernel.org>