Changes in v4 -> v5:
v4: https://lore.kernel.org/all/[email protected]/
- Implement the inlining of the bpf_get_smp_processor_id() in the JIT.
NOTE: This needs to be based on:
https://lore.kernel.org/all/[email protected]/
to be built.
Manual run of bpf-ci with this series rebased on above:
https://github.com/kernel-patches/bpf/pull/6929
Changes in v3 -> v4:
v3: https://lore.kernel.org/all/[email protected]/
- Fix coding style issue related to C89 standards.
Changes in v2 -> v3:
v2: https://lore.kernel.org/all/[email protected]/
- Fixed the xlated dump of percpu mov to "r0 = &(void __percpu *)(r0)"
- Made ARM64 and x86-64 use the same code for inlining. The only difference
that remains is the per-cpu address of the cpu_number.
Changes in v1 -> v2:
v1: https://lore.kernel.org/all/[email protected]/
- Add a patch to inline bpf_get_smp_processor_id()
- Fix an issue in MRS instruction encoding as pointed out by Will
- Remove CONFIG_SMP check because arm64 kernel always compiles with CONFIG_SMP
This series adds the support of internal only per-CPU instructions and
inlines the bpf_get_smp_processor_id() helper call for ARM64 BPF JIT.
Here is an example of calls to bpf_get_smp_processor_id() and
percpu_array_map_lookup_elem() before and after this series.
BPF
=====
BEFORE AFTER
-------- -------
int cpu = bpf_get_smp_processor_id(); int cpu = bpf_get_smp_processor_id();
(85) call bpf_get_smp_processor_id#229032 (85) call bpf_get_smp_processor_id#8
p = bpf_map_lookup_elem(map, &zero); p = bpf_map_lookup_elem(map, &zero);
(18) r1 = map[id:78] (18) r1 = map[id:153]
(18) r2 = map[id:82][0]+65536 (18) r2 = map[id:157][0]+65536
(85) call percpu_array_map_lookup_elem#313512 (07) r1 += 496
(61) r0 = *(u32 *)(r2 +0)
(35) if r0 >= 0x1 goto pc+5
(67) r0 <<= 3
(0f) r0 += r1
(79) r0 = *(u64 *)(r0 +0)
(bf) r0 = &(void __percpu *)(r0)
(05) goto pc+1
(b7) r0 = 0
ARM64 JIT
===========
BEFORE AFTER
-------- -------
int cpu = bpf_get_smp_processor_id(); int cpu = bpf_get_smp_processor_id();
mov x10, #0xfffffffffffff4d0 mrs x10, sp_el0
movk x10, #0x802b, lsl #16 ldr w7, [x10, #24]
movk x10, #0x8000, lsl #32
blr x10
add x7, x0, #0x0
p = bpf_map_lookup_elem(map, &zero); p = bpf_map_lookup_elem(map, &zero);
mov x0, #0xffff0003ffffffff mov x0, #0xffff0003ffffffff
movk x0, #0xce5c, lsl #16 movk x0, #0xe0f3, lsl #16
movk x0, #0xca00 movk x0, #0x7c00
mov x1, #0xffff8000ffffffff mov x1, #0xffff8000ffffffff
movk x1, #0x8bdb, lsl #16 movk x1, #0xb0c7, lsl #16
movk x1, #0x6000 movk x1, #0xe000
mov x10, #0xffffffffffff3ed0 add x0, x0, #0x1f0
movk x10, #0x802d, lsl #16 ldr w7, [x1]
movk x10, #0x8000, lsl #32 cmp x7, #0x1
blr x10 b.cs 0x0000000000000090
add x7, x0, #0x0 lsl x7, x7, #3
add x7, x7, x0
ldr x7, [x7]
mrs x10, tpidr_el1
add x7, x7, x10
b 0x0000000000000094
mov x7, #0x0
Performance improvement found using benchmark[1]
/benchs/run_bench_trigger.sh glob-arr-inc arr-inc hash-inc
+---------------+-------------------+-------------------+--------------+
| Name | Before | After | % change |
|---------------+-------------------+-------------------+--------------|
| glob-arr-inc | 23.380 ± 1.675M/s | 25.893 ± 0.026M/s | + 10.74% |
| arr-inc | 23.928 ± 0.034M/s | 25.213 ± 0.063M/s | + 5.37% |
| hash-inc | 12.352 ± 0.005M/s | 12.609 ± 0.013M/s | + 2.08% |
+---------------+-------------------+-------------------+--------------+
[1] https://github.com/anakryiko/linux/commit/8dec900975ef
Puranjay Mohan (2):
arm64, bpf: add internal-only MOV instruction to resolve per-CPU addrs
bpf, arm64: inline bpf_get_smp_processor_id() helper
arch/arm64/include/asm/insn.h | 8 ++++++++
arch/arm64/lib/insn.c | 11 +++++++++++
arch/arm64/net/bpf_jit.h | 8 ++++++++
arch/arm64/net/bpf_jit_comp.c | 37 +++++++++++++++++++++++++++++++++++
4 files changed, 64 insertions(+)
--
2.40.1
From: Puranjay Mohan <[email protected]>
Support an instruction for resolving absolute addresses of per-CPU
data from their per-CPU offsets. This instruction is internal-only and
users are not allowed to use them directly. They will only be used for
internal inlining optimizations for now between BPF verifier and BPF
JITs.
Since commit 7158627686f0 ("arm64: percpu: implement optimised pcpu
access using tpidr_el1"), the per-cpu offset for the CPU is stored in
the tpidr_el1/2 register of that CPU.
To support this BPF instruction in the ARM64 JIT, the following ARM64
instructions are emitted:
mov dst, src // Move src to dst, if src != dst
mrs tmp, tpidr_el1/2 // Move per-cpu offset of the current cpu in tmp.
add dst, dst, tmp // Add the per cpu offset to the dst.
To measure the performance improvement provided by this change, the
benchmark in [1] was used:
Before:
glob-arr-inc : 23.597 ± 0.012M/s
arr-inc : 23.173 ± 0.019M/s
hash-inc : 12.186 ± 0.028M/s
After:
glob-arr-inc : 23.819 ± 0.034M/s
arr-inc : 23.285 ± 0.017M/s
hash-inc : 12.419 ± 0.011M/s
[1] https://github.com/anakryiko/linux/commit/8dec900975ef
Signed-off-by: Puranjay Mohan <[email protected]>
---
arch/arm64/include/asm/insn.h | 7 +++++++
arch/arm64/lib/insn.c | 11 +++++++++++
arch/arm64/net/bpf_jit.h | 6 ++++++
arch/arm64/net/bpf_jit_comp.c | 14 ++++++++++++++
4 files changed, 38 insertions(+)
diff --git a/arch/arm64/include/asm/insn.h b/arch/arm64/include/asm/insn.h
index db1aeacd4cd9..8de0e39b29f3 100644
--- a/arch/arm64/include/asm/insn.h
+++ b/arch/arm64/include/asm/insn.h
@@ -135,6 +135,11 @@ enum aarch64_insn_special_register {
AARCH64_INSN_SPCLREG_SP_EL2 = 0xF210
};
+enum aarch64_insn_system_register {
+ AARCH64_INSN_SYSREG_TPIDR_EL1 = 0x4684,
+ AARCH64_INSN_SYSREG_TPIDR_EL2 = 0x6682,
+};
+
enum aarch64_insn_variant {
AARCH64_INSN_VARIANT_32BIT,
AARCH64_INSN_VARIANT_64BIT
@@ -686,6 +691,8 @@ u32 aarch64_insn_gen_cas(enum aarch64_insn_register result,
}
#endif
u32 aarch64_insn_gen_dmb(enum aarch64_insn_mb_type type);
+u32 aarch64_insn_gen_mrs(enum aarch64_insn_register result,
+ enum aarch64_insn_system_register sysreg);
s32 aarch64_get_branch_offset(u32 insn);
u32 aarch64_set_branch_offset(u32 insn, s32 offset);
diff --git a/arch/arm64/lib/insn.c b/arch/arm64/lib/insn.c
index a635ab83fee3..b008a9b46a7f 100644
--- a/arch/arm64/lib/insn.c
+++ b/arch/arm64/lib/insn.c
@@ -1515,3 +1515,14 @@ u32 aarch64_insn_gen_dmb(enum aarch64_insn_mb_type type)
return insn;
}
+
+u32 aarch64_insn_gen_mrs(enum aarch64_insn_register result,
+ enum aarch64_insn_system_register sysreg)
+{
+ u32 insn = aarch64_insn_get_mrs_value();
+
+ insn &= ~GENMASK(19, 0);
+ insn |= sysreg << 5;
+ return aarch64_insn_encode_register(AARCH64_INSN_REGTYPE_RT,
+ insn, result);
+}
diff --git a/arch/arm64/net/bpf_jit.h b/arch/arm64/net/bpf_jit.h
index 23b1b34db088..b627ef7188c7 100644
--- a/arch/arm64/net/bpf_jit.h
+++ b/arch/arm64/net/bpf_jit.h
@@ -297,4 +297,10 @@
#define A64_ADR(Rd, offset) \
aarch64_insn_gen_adr(0, offset, Rd, AARCH64_INSN_ADR_TYPE_ADR)
+/* MRS */
+#define A64_MRS_TPIDR_EL1(Rt) \
+ aarch64_insn_gen_mrs(Rt, AARCH64_INSN_SYSREG_TPIDR_EL1)
+#define A64_MRS_TPIDR_EL2(Rt) \
+ aarch64_insn_gen_mrs(Rt, AARCH64_INSN_SYSREG_TPIDR_EL2)
+
#endif /* _BPF_JIT_H */
diff --git a/arch/arm64/net/bpf_jit_comp.c b/arch/arm64/net/bpf_jit_comp.c
index 76b91f36c729..ed8f9716d9d5 100644
--- a/arch/arm64/net/bpf_jit_comp.c
+++ b/arch/arm64/net/bpf_jit_comp.c
@@ -877,6 +877,15 @@ static int build_insn(const struct bpf_insn *insn, struct jit_ctx *ctx,
emit(A64_ORR(1, tmp, dst, tmp), ctx);
emit(A64_MOV(1, dst, tmp), ctx);
break;
+ } else if (insn_is_mov_percpu_addr(insn)) {
+ if (dst != src)
+ emit(A64_MOV(1, dst, src), ctx);
+ if (cpus_have_cap(ARM64_HAS_VIRT_HOST_EXTN))
+ emit(A64_MRS_TPIDR_EL2(tmp), ctx);
+ else
+ emit(A64_MRS_TPIDR_EL1(tmp), ctx);
+ emit(A64_ADD(1, dst, dst, tmp), ctx);
+ break;
}
switch (insn->off) {
case 0:
@@ -2527,6 +2536,11 @@ bool bpf_jit_supports_arena(void)
return true;
}
+bool bpf_jit_supports_percpu_insn(void)
+{
+ return true;
+}
+
void bpf_jit_free(struct bpf_prog *prog)
{
if (prog->jited) {
--
2.40.1
Inline calls to bpf_get_smp_processor_id() helper in the JIT by emitting
a read from struct thread_info. The SP_EL0 system register holds the
pointer to the task_struct and thread_info is the first member of this
struct. We can read the cpu number from the thread_info.
Here is how the ARM64 JITed assembly changes after this commit:
ARM64 JIT
===========
BEFORE AFTER
-------- -------
int cpu = bpf_get_smp_processor_id(); int cpu = bpf_get_smp_processor_id();
mov x10, #0xfffffffffffff4d0 mrs x10, sp_el0
movk x10, #0x802b, lsl #16 ldr w7, [x10, #24]
movk x10, #0x8000, lsl #32
blr x10
add x7, x0, #0x0
Performance improvement using benchmark[1]
/benchs/run_bench_trigger.sh glob-arr-inc arr-inc hash-inc
+---------------+-------------------+-------------------+--------------+
| Name | Before | After | % change |
|---------------+-------------------+-------------------+--------------|
| glob-arr-inc | 23.380 ± 1.675M/s | 25.893 ± 0.026M/s | + 10.74% |
| arr-inc | 23.928 ± 0.034M/s | 25.213 ± 0.063M/s | + 5.37% |
| hash-inc | 12.352 ± 0.005M/s | 12.609 ± 0.013M/s | + 2.08% |
+---------------+-------------------+-------------------+--------------+
[1] https://github.com/anakryiko/linux/commit/8dec900975ef
Signed-off-by: Puranjay Mohan <[email protected]>
---
arch/arm64/include/asm/insn.h | 1 +
arch/arm64/net/bpf_jit.h | 2 ++
arch/arm64/net/bpf_jit_comp.c | 23 +++++++++++++++++++++++
3 files changed, 26 insertions(+)
diff --git a/arch/arm64/include/asm/insn.h b/arch/arm64/include/asm/insn.h
index 8de0e39b29f3..8c0a36f72d6f 100644
--- a/arch/arm64/include/asm/insn.h
+++ b/arch/arm64/include/asm/insn.h
@@ -138,6 +138,7 @@ enum aarch64_insn_special_register {
enum aarch64_insn_system_register {
AARCH64_INSN_SYSREG_TPIDR_EL1 = 0x4684,
AARCH64_INSN_SYSREG_TPIDR_EL2 = 0x6682,
+ AARCH64_INSN_SYSREG_SP_EL0 = 0x4208,
};
enum aarch64_insn_variant {
diff --git a/arch/arm64/net/bpf_jit.h b/arch/arm64/net/bpf_jit.h
index b627ef7188c7..b22ab2f97a30 100644
--- a/arch/arm64/net/bpf_jit.h
+++ b/arch/arm64/net/bpf_jit.h
@@ -302,5 +302,7 @@
aarch64_insn_gen_mrs(Rt, AARCH64_INSN_SYSREG_TPIDR_EL1)
#define A64_MRS_TPIDR_EL2(Rt) \
aarch64_insn_gen_mrs(Rt, AARCH64_INSN_SYSREG_TPIDR_EL2)
+#define A64_MRS_SP_EL0(Rt) \
+ aarch64_insn_gen_mrs(Rt, AARCH64_INSN_SYSREG_SP_EL0)
#endif /* _BPF_JIT_H */
diff --git a/arch/arm64/net/bpf_jit_comp.c b/arch/arm64/net/bpf_jit_comp.c
index ed8f9716d9d5..8084f3e61e0b 100644
--- a/arch/arm64/net/bpf_jit_comp.c
+++ b/arch/arm64/net/bpf_jit_comp.c
@@ -1215,6 +1215,19 @@ static int build_insn(const struct bpf_insn *insn, struct jit_ctx *ctx,
const u8 r0 = bpf2a64[BPF_REG_0];
bool func_addr_fixed;
u64 func_addr;
+ u32 cpu_offset = offsetof(struct thread_info, cpu);
+
+ /* Implement helper call to bpf_get_smp_processor_id() inline */
+ if (insn->src_reg == 0 && insn->imm == BPF_FUNC_get_smp_processor_id) {
+ emit(A64_MRS_SP_EL0(tmp), ctx);
+ if (is_lsi_offset(cpu_offset, 2)) {
+ emit(A64_LDR32I(r0, tmp, cpu_offset), ctx);
+ } else {
+ emit_a64_mov_i(1, tmp2, cpu_offset, ctx);
+ emit(A64_LDR32(r0, tmp, tmp2), ctx);
+ }
+ break;
+ }
ret = bpf_jit_get_func_addr(ctx->prog, insn, extra_pass,
&func_addr, &func_addr_fixed);
@@ -2541,6 +2554,16 @@ bool bpf_jit_supports_percpu_insn(void)
return true;
}
+bool bpf_jit_inlines_helper_call(s32 imm)
+{
+ switch (imm) {
+ case BPF_FUNC_get_smp_processor_id:
+ return true;
+ }
+
+ return false;
+}
+
void bpf_jit_free(struct bpf_prog *prog)
{
if (prog->jited) {
--
2.40.1
On Tue, Apr 30, 2024 at 4:48 PM Puranjay Mohan <[email protected]> wrote:
>
> Inline calls to bpf_get_smp_processor_id() helper in the JIT by emitting
> a read from struct thread_info. The SP_EL0 system register holds the
> pointer to the task_struct and thread_info is the first member of this
> struct. We can read the cpu number from the thread_info.
>
> Here is how the ARM64 JITed assembly changes after this commit:
>
> ARM64 JIT
> ===========
>
> BEFORE AFTER
> -------- -------
>
> int cpu = bpf_get_smp_processor_id(); int cpu = bpf_get_smp_processor_id();
>
> mov x10, #0xfffffffffffff4d0 mrs x10, sp_el0
> movk x10, #0x802b, lsl #16 ldr w7, [x10, #24]
> movk x10, #0x8000, lsl #32
> blr x10
> add x7, x0, #0x0
>
> Performance improvement using benchmark[1]
>
> ./benchs/run_bench_trigger.sh glob-arr-inc arr-inc hash-inc
>
> +---------------+-------------------+-------------------+--------------+
> | Name | Before | After | % change |
> |---------------+-------------------+-------------------+--------------|
> | glob-arr-inc | 23.380 ± 1.675M/s | 25.893 ± 0.026M/s | + 10.74% |
> | arr-inc | 23.928 ± 0.034M/s | 25.213 ± 0.063M/s | + 537% |
> | hash-inc | 12.352 ± 0.005M/s | 12.609 ± 0.013M/s | + 208% |
> +---------------+-------------------+-------------------+--------------+
>
> [1] https://github.com/anakryiko/linux/commit/8dec900975ef
>
> Signed-off-by: Puranjay Mohan <[email protected]>
> ---
> arch/arm64/include/asm/insn.h | 1 +
> arch/arm64/net/bpf_jit.h | 2 ++
> arch/arm64/net/bpf_jit_comp.c | 23 +++++++++++++++++++++++
> 3 files changed, 26 insertions(+)
>
Nice improvements! I suggest combining arm64 and risc-v patches
together when resubmitting, so that we can land them in one go. This
one depends on RISC-V patches landing first to avoid a warning about
global function without a prototype, right?
Please add my ack as well
Acked-by: Andrii Nakryiko <[email protected]>
> diff --git a/arch/arm64/include/asm/insn.h b/arch/arm64/include/asm/insn.h
> index 8de0e39b29f3..8c0a36f72d6f 100644
> --- a/arch/arm64/include/asm/insn.h
> +++ b/arch/arm64/include/asm/insn.h
> @@ -138,6 +138,7 @@ enum aarch64_insn_special_register {
> enum aarch64_insn_system_register {
> AARCH64_INSN_SYSREG_TPIDR_EL1 = 0x4684,
> AARCH64_INSN_SYSREG_TPIDR_EL2 = 0x6682,
> + AARCH64_INSN_SYSREG_SP_EL0 = 0x4208,
> };
>
> enum aarch64_insn_variant {
> diff --git a/arch/arm64/net/bpf_jit.h b/arch/arm64/net/bpf_jit.h
> index b627ef7188c7..b22ab2f97a30 100644
> --- a/arch/arm64/net/bpf_jit.h
> +++ b/arch/arm64/net/bpf_jit.h
> @@ -302,5 +302,7 @@
> aarch64_insn_gen_mrs(Rt, AARCH64_INSN_SYSREG_TPIDR_EL1)
> #define A64_MRS_TPIDR_EL2(Rt) \
> aarch64_insn_gen_mrs(Rt, AARCH64_INSN_SYSREG_TPIDR_EL2)
> +#define A64_MRS_SP_EL0(Rt) \
> + aarch64_insn_gen_mrs(Rt, AARCH64_INSN_SYSREG_SP_EL0)
>
> #endif /* _BPF_JIT_H */
> diff --git a/arch/arm64/net/bpf_jit_comp.c b/arch/arm64/net/bpf_jit_comp.c
> index ed8f9716d9d5..8084f3e61e0b 100644
> --- a/arch/arm64/net/bpf_jit_comp.c
> +++ b/arch/arm64/net/bpf_jit_comp.c
> @@ -1215,6 +1215,19 @@ static int build_insn(const struct bpf_insn *insn, struct jit_ctx *ctx,
> const u8 r0 = bpf2a64[BPF_REG_0];
> bool func_addr_fixed;
> u64 func_addr;
> + u32 cpu_offset = offsetof(struct thread_info, cpu);
> +
> + /* Implement helper call to bpf_get_smp_processor_id() inline */
> + if (insn->src_reg == 0 && insn->imm == BPF_FUNC_get_smp_processor_id) {
> + emit(A64_MRS_SP_EL0(tmp), ctx);
> + if (is_lsi_offset(cpu_offset, 2)) {
> + emit(A64_LDR32I(r0, tmp, cpu_offset), ctx);
> + } else {
> + emit_a64_mov_i(1, tmp2, cpu_offset, ctx);
> + emit(A64_LDR32(r0, tmp, tmp2), ctx);
> + }
> + break;
> + }
>
> ret = bpf_jit_get_func_addr(ctx->prog, insn, extra_pass,
> &func_addr, &func_addr_fixed);
> @@ -2541,6 +2554,16 @@ bool bpf_jit_supports_percpu_insn(void)
> return true;
> }
>
> +bool bpf_jit_inlines_helper_call(s32 imm)
> +{
> + switch (imm) {
> + case BPF_FUNC_get_smp_processor_id:
> + return true;
> + }
> +
> + return false;
same minor nit to use default: return false inside the switch itself
> +}
> +
> void bpf_jit_free(struct bpf_prog *prog)
> {
> if (prog->jited) {
> --
> 2.40.1
>