Re: [PATCH bpf-next v5 07/11] bpf: Add instructions for atomic_[cmp]xchg
On 12/15/20 4:18 AM, Brendan Jackman wrote: This adds two atomic opcodes, both of which include the BPF_FETCH flag. XCHG without the BPF_FETCH flag would naturally encode atomic_set. This is not supported because it would be of limited value to userspace (it doesn't imply any barriers). CMPXCHG without BPF_FETCH woulud be an atomic compare-and-write. We don't have such an operation in the kernel so it isn't provided to BPF either. There are two significant design decisions made for the CMPXCHG instruction: - To solve the issue that this operation fundamentally has 3 operands, but we only have two register fields. Therefore the operand we compare against (the kernel's API calls it 'old') is hard-coded to be R0. x86 has similar design (and A64 doesn't have this problem). A potential alternative might be to encode the other operand's register number in the immediate field. - The kernel's atomic_cmpxchg returns the old value, while the C11 userspace APIs return a boolean indicating the comparison result. Which should BPF do? A64 returns the old value. x86 returns the old value in the hard-coded register (and also sets a flag). That means return-old-value is easier to JIT, so that's what we use. Signed-off-by: Brendan Jackman Ack with a minor comment below. Acked-by: Yonghong Song --- arch/x86/net/bpf_jit_comp.c| 8 include/linux/filter.h | 2 ++ include/uapi/linux/bpf.h | 4 +++- kernel/bpf/core.c | 20 kernel/bpf/disasm.c| 15 +++ kernel/bpf/verifier.c | 19 +-- tools/include/linux/filter.h | 2 ++ tools/include/uapi/linux/bpf.h | 4 +++- 8 files changed, 70 insertions(+), 4 deletions(-) diff --git a/arch/x86/net/bpf_jit_comp.c b/arch/x86/net/bpf_jit_comp.c index eea7d8b0bb12..308241187582 100644 --- a/arch/x86/net/bpf_jit_comp.c +++ b/arch/x86/net/bpf_jit_comp.c @@ -815,6 +815,14 @@ static int emit_atomic(u8 **pprog, u8 atomic_op, /* src_reg = atomic_fetch_add(dst_reg + off, src_reg); */ EMIT2(0x0F, 0xC1); break; + case BPF_XCHG: + /* src_reg = atomic_xchg(dst_reg + off, src_reg); */ + EMIT1(0x87); + break; + case BPF_CMPXCHG: + /* r0 = atomic_cmpxchg(dst_reg + off, r0, src_reg); */ + EMIT2(0x0F, 0xB1); + break; default: pr_err("bpf_jit: unknown atomic opcode %02x\n", atomic_op); return -EFAULT; diff --git a/include/linux/filter.h b/include/linux/filter.h index c3e87a63e0b8..16e0ba5e8937 100644 --- a/include/linux/filter.h +++ b/include/linux/filter.h @@ -265,6 +265,8 @@ static inline bool insn_is_zext(const struct bpf_insn *insn) * * BPF_ADD *(uint *) (dst_reg + off16) += src_reg * BPF_ADD | BPF_FETCH src_reg = atomic_fetch_add(dst_reg + off16, src_reg); + * BPF_XCHG src_reg = atomic_xchg(dst_reg + off16, src_reg) + * BPF_CMPXCHG r0 = atomic_cmpxchg(dst_reg + off16, r0, src_reg) */ #define BPF_ATOMIC_OP(SIZE, OP, DST, SRC, OFF) \ diff --git a/include/uapi/linux/bpf.h b/include/uapi/linux/bpf.h index 760ae333a5ed..538b95472c8f 100644 --- a/include/uapi/linux/bpf.h +++ b/include/uapi/linux/bpf.h @@ -45,7 +45,9 @@ #define BPF_EXIT 0x90/* function return */ /* atomic op type fields (stored in immediate) */ -#define BPF_FETCH 0x01/* fetch previous value into src reg */ +#define BPF_XCHG (0xe0 | BPF_FETCH) /* atomic exchange */ +#define BPF_CMPXCHG(0xf0 | BPF_FETCH) /* atomic compare-and-write */ +#define BPF_FETCH 0x01/* not an opcode on its own, used to build others */ Although the above code works fine, I would suggest to put BPF_FETCH definition before BPF_XCHG and BPF_CMPXCHG, which makes more sense intuitively. /* Register numbers */ enum { diff --git a/kernel/bpf/core.c b/kernel/bpf/core.c [...] \ diff --git a/tools/include/uapi/linux/bpf.h b/tools/include/uapi/linux/bpf.h index 760ae333a5ed..538b95472c8f 100644 --- a/tools/include/uapi/linux/bpf.h +++ b/tools/include/uapi/linux/bpf.h @@ -45,7 +45,9 @@ #define BPF_EXIT 0x90/* function return */ /* atomic op type fields (stored in immediate) */ -#define BPF_FETCH 0x01/* fetch previous value into src reg */ +#define BPF_XCHG (0xe0 | BPF_FETCH) /* atomic exchange */ +#define BPF_CMPXCHG(0xf0 | BPF_FETCH) /* atomic compare-and-write */ +#define BPF_FETCH 0x01/* not an opcode on its own, used to build others */ same here. /* Register numbers */ enum {
[PATCH bpf-next v5 07/11] bpf: Add instructions for atomic_[cmp]xchg
This adds two atomic opcodes, both of which include the BPF_FETCH flag. XCHG without the BPF_FETCH flag would naturally encode atomic_set. This is not supported because it would be of limited value to userspace (it doesn't imply any barriers). CMPXCHG without BPF_FETCH woulud be an atomic compare-and-write. We don't have such an operation in the kernel so it isn't provided to BPF either. There are two significant design decisions made for the CMPXCHG instruction: - To solve the issue that this operation fundamentally has 3 operands, but we only have two register fields. Therefore the operand we compare against (the kernel's API calls it 'old') is hard-coded to be R0. x86 has similar design (and A64 doesn't have this problem). A potential alternative might be to encode the other operand's register number in the immediate field. - The kernel's atomic_cmpxchg returns the old value, while the C11 userspace APIs return a boolean indicating the comparison result. Which should BPF do? A64 returns the old value. x86 returns the old value in the hard-coded register (and also sets a flag). That means return-old-value is easier to JIT, so that's what we use. Signed-off-by: Brendan Jackman --- arch/x86/net/bpf_jit_comp.c| 8 include/linux/filter.h | 2 ++ include/uapi/linux/bpf.h | 4 +++- kernel/bpf/core.c | 20 kernel/bpf/disasm.c| 15 +++ kernel/bpf/verifier.c | 19 +-- tools/include/linux/filter.h | 2 ++ tools/include/uapi/linux/bpf.h | 4 +++- 8 files changed, 70 insertions(+), 4 deletions(-) diff --git a/arch/x86/net/bpf_jit_comp.c b/arch/x86/net/bpf_jit_comp.c index eea7d8b0bb12..308241187582 100644 --- a/arch/x86/net/bpf_jit_comp.c +++ b/arch/x86/net/bpf_jit_comp.c @@ -815,6 +815,14 @@ static int emit_atomic(u8 **pprog, u8 atomic_op, /* src_reg = atomic_fetch_add(dst_reg + off, src_reg); */ EMIT2(0x0F, 0xC1); break; + case BPF_XCHG: + /* src_reg = atomic_xchg(dst_reg + off, src_reg); */ + EMIT1(0x87); + break; + case BPF_CMPXCHG: + /* r0 = atomic_cmpxchg(dst_reg + off, r0, src_reg); */ + EMIT2(0x0F, 0xB1); + break; default: pr_err("bpf_jit: unknown atomic opcode %02x\n", atomic_op); return -EFAULT; diff --git a/include/linux/filter.h b/include/linux/filter.h index c3e87a63e0b8..16e0ba5e8937 100644 --- a/include/linux/filter.h +++ b/include/linux/filter.h @@ -265,6 +265,8 @@ static inline bool insn_is_zext(const struct bpf_insn *insn) * * BPF_ADD *(uint *) (dst_reg + off16) += src_reg * BPF_ADD | BPF_FETCH src_reg = atomic_fetch_add(dst_reg + off16, src_reg); + * BPF_XCHG src_reg = atomic_xchg(dst_reg + off16, src_reg) + * BPF_CMPXCHG r0 = atomic_cmpxchg(dst_reg + off16, r0, src_reg) */ #define BPF_ATOMIC_OP(SIZE, OP, DST, SRC, OFF) \ diff --git a/include/uapi/linux/bpf.h b/include/uapi/linux/bpf.h index 760ae333a5ed..538b95472c8f 100644 --- a/include/uapi/linux/bpf.h +++ b/include/uapi/linux/bpf.h @@ -45,7 +45,9 @@ #define BPF_EXIT 0x90/* function return */ /* atomic op type fields (stored in immediate) */ -#define BPF_FETCH 0x01/* fetch previous value into src reg */ +#define BPF_XCHG (0xe0 | BPF_FETCH) /* atomic exchange */ +#define BPF_CMPXCHG(0xf0 | BPF_FETCH) /* atomic compare-and-write */ +#define BPF_FETCH 0x01/* not an opcode on its own, used to build others */ /* Register numbers */ enum { diff --git a/kernel/bpf/core.c b/kernel/bpf/core.c index 2b1d3fea03fc..4f22cff4491e 100644 --- a/kernel/bpf/core.c +++ b/kernel/bpf/core.c @@ -1630,6 +1630,16 @@ static u64 ___bpf_prog_run(u64 *regs, const struct bpf_insn *insn, u64 *stack) (u32) SRC, (atomic_t *)(unsigned long) (DST + insn->off)); break; + case BPF_XCHG: + SRC = (u32) atomic_xchg( + (atomic_t *)(unsigned long) (DST + insn->off), + (u32) SRC); + break; + case BPF_CMPXCHG: + BPF_R0 = (u32) atomic_cmpxchg( + (atomic_t *)(unsigned long) (DST + insn->off), + (u32) BPF_R0, (u32) SRC); + break; default: goto default_label; } @@ -1647,6 +1657,16 @@ static u64 ___bpf_prog_run(u64 *regs, const struct bpf_insn *insn, u64 *stack) (u64) SRC, (atomic64_t *)(unsigned long) (DST + insn->off)); break; + case