powerpc/bpf/64: Add instructions for atomic_[cmp]xchg

This adds two atomic opcodes BPF_XCHG and BPF_CMPXCHG on ppc64, both
of which include the BPF_FETCH flag.  The kernel's atomic_cmpxchg
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 BPF_REG_R0. Also, kernel's
atomic_cmpxchg returns the previous value at dst_reg + off. JIT the
same for BPF too with return value put in BPF_REG_0.

  BPF_REG_R0 = atomic_cmpxchg(dst_reg + off, BPF_REG_R0, src_reg);
Signed-off-by: Hari Bathini <hbathini@linux.ibm.com>
Tested-by: Naveen N. Rao <naveen.n.rao@linux.vnet.ibm.com> (ppc64le)
Reviewed-by: Naveen N. Rao <naveen.n.rao@linux.vnet.ibm.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Link: https://lore.kernel.org/r/20220610155552.25892-4-hbathini@linux.ibm.com

arch/powerpc/net/bpf_jit_comp64.c

@@ -360,6 +360,7 @@ int bpf_jit_build_body(struct bpf_prog *fp, u32 *image, struct codegen_context *
 		u32 size = BPF_SIZE(code);
 		u32 tmp1_reg = bpf_to_ppc(TMP_REG_1);
 		u32 tmp2_reg = bpf_to_ppc(TMP_REG_2);
+		u32 save_reg, ret_reg;
 		s16 off = insn[i].off;
 		s32 imm = insn[i].imm;
 		bool func_addr_fixed;
@@ -778,6 +779,9 @@ int bpf_jit_build_body(struct bpf_prog *fp, u32 *image, struct codegen_context *
 		 */
 		case BPF_STX | BPF_ATOMIC | BPF_W:
 		case BPF_STX | BPF_ATOMIC | BPF_DW:
+			save_reg = tmp2_reg;
+			ret_reg = src_reg;
+
 			/* Get offset into TMP_REG_1 */
 			EMIT(PPC_RAW_LI(tmp1_reg, off));
 			tmp_idx = ctx->idx * 4;
@@ -808,6 +812,24 @@ int bpf_jit_build_body(struct bpf_prog *fp, u32 *image, struct codegen_context *
 			case BPF_XOR | BPF_FETCH:
 				EMIT(PPC_RAW_XOR(tmp2_reg, tmp2_reg, src_reg));
 				break;
+			case BPF_CMPXCHG:
+				/*
+				 * Return old value in BPF_REG_0 for BPF_CMPXCHG &
+				 * in src_reg for other cases.
+				 */
+				ret_reg = bpf_to_ppc(BPF_REG_0);
+
+				/* Compare with old value in BPF_R0 */
+				if (size == BPF_DW)
+					EMIT(PPC_RAW_CMPD(bpf_to_ppc(BPF_REG_0), tmp2_reg));
+				else
+					EMIT(PPC_RAW_CMPW(bpf_to_ppc(BPF_REG_0), tmp2_reg));
+				/* Don't set if different from old value */
+				PPC_BCC_SHORT(COND_NE, (ctx->idx + 3) * 4);
+				fallthrough;
+			case BPF_XCHG:
+				save_reg = src_reg;
+				break;
 			default:
 				pr_err_ratelimited(
 					"eBPF filter atomic op code %02x (@%d) unsupported\n",
@@ -817,15 +839,22 @@ int bpf_jit_build_body(struct bpf_prog *fp, u32 *image, struct codegen_context *
 
 			/* store new value */
 			if (size == BPF_DW)
-				EMIT(PPC_RAW_STDCX(tmp2_reg, tmp1_reg, dst_reg));
+				EMIT(PPC_RAW_STDCX(save_reg, tmp1_reg, dst_reg));
 			else
-				EMIT(PPC_RAW_STWCX(tmp2_reg, tmp1_reg, dst_reg));
+				EMIT(PPC_RAW_STWCX(save_reg, tmp1_reg, dst_reg));
 			/* we're done if this succeeded */
 			PPC_BCC_SHORT(COND_NE, tmp_idx);
 
-			/* For the BPF_FETCH variant, get old value into src_reg */
-			if (imm & BPF_FETCH)
-				EMIT(PPC_RAW_MR(src_reg, _R0));
+			if (imm & BPF_FETCH) {
+				EMIT(PPC_RAW_MR(ret_reg, _R0));
+				/*
+				 * Skip unnecessary zero-extension for 32-bit cmpxchg.
+				 * For context, see commit 39491867ace5.
+				 */
+				if (size != BPF_DW && imm == BPF_CMPXCHG &&
+				    insn_is_zext(&insn[i + 1]))
+					addrs[++i] = ctx->idx * 4;
+			}
 			break;
 
 		/*