Skip to contentSkip to conversation

RV64 BPF JIT Optimizations

PATCH Applied Luke Nelson 3 participants 7 messages May 6, 2020

Subsystems (5)
lore
Share & tools
lore
#0 lore raw

cover RV64 BPF JIT Optimizations

This patch series introduces a set of optimizations to the BPF JIT on RV64. The optimizations are related to the verifier zero-extension optimization and BPF_JMP BPF_K.

We tested the optimizations on a QEMU riscv64 virt machine, using lib/test_bpf and test_verifier, and formally verified their correctness using Serval.

Luke Nelson (4):
bpf, riscv: Enable missing verifier_zext optimizations on RV64 bpf, riscv: Optimize FROM_LE using verifier_zext on RV64 bpf, riscv: Optimize BPF_JMP BPF_K when imm == 0 on RV64 bpf, riscv: Optimize BPF_JSET BPF_K using andi on RV64

 arch/riscv/net/bpf_jit_comp64.c | 64 ++++++++++++++++++++++-----------
 1 file changed, 44 insertions(+), 20 deletions(-)
Cc Xi Wang <[email protected]>
Signature · 1 line
2.17.1
#1 lore raw

PATCH 1/4 bpf, riscv: Enable missing verifier_zext optimizations on RV64

Commit 66d0d5a854a6 ("riscv: bpf: eliminate zero extension code-gen") added support for the verifier zero-extension optimization on RV64 and commit 46dd3d7d287b ("bpf, riscv: Enable zext optimization for more RV64G ALU ops") enabled it for more instruction cases.

However, BPF_LSH BPF_X and BPF_{LSH,RSH,ARSH} BPF_K are still missing the optimization.

This patch enables the zero-extension optimization for these remaining cases.

---
 arch/riscv/net/bpf_jit_comp64.c | 8 ++++----
 1 file changed, 4 insertions(+), 4 deletions(-)
arch/riscv/net/bpf_jit_comp64.c +4 -4
index d208a9fd6c52..e2636902a74e 100644
@@ -515,7 +515,7 @@ int bpf_jit_emit_insn(const struct bpf_insn *insn, struct rv_jit_context *ctx,
 	case BPF_ALU | BPF_LSH | BPF_X:
 	case BPF_ALU64 | BPF_LSH | BPF_X:
 		emit(is64 ? rv_sll(rd, rd, rs) : rv_sllw(rd, rd, rs), ctx);
-		if (!is64)
+		if (!is64 && !aux->verifier_zext)
 			emit_zext_32(rd, ctx);
 		break;
 	case BPF_ALU | BPF_RSH | BPF_X:
@@ -692,19 +692,19 @@ int bpf_jit_emit_insn(const struct bpf_insn *insn, struct rv_jit_context *ctx,
 	case BPF_ALU | BPF_LSH | BPF_K:
 	case BPF_ALU64 | BPF_LSH | BPF_K:
 		emit(is64 ? rv_slli(rd, rd, imm) : rv_slliw(rd, rd, imm), ctx);
-		if (!is64)
+		if (!is64 && !aux->verifier_zext)
 			emit_zext_32(rd, ctx);
 		break;
 	case BPF_ALU | BPF_RSH | BPF_K:
 	case BPF_ALU64 | BPF_RSH | BPF_K:
 		emit(is64 ? rv_srli(rd, rd, imm) : rv_srliw(rd, rd, imm), ctx);
-		if (!is64)
+		if (!is64 && !aux->verifier_zext)
 			emit_zext_32(rd, ctx);
 		break;
 	case BPF_ALU | BPF_ARSH | BPF_K:
 	case BPF_ALU64 | BPF_ARSH | BPF_K:
 		emit(is64 ? rv_srai(rd, rd, imm) : rv_sraiw(rd, rd, imm), ctx);
-		if (!is64)
+		if (!is64 && !aux->verifier_zext)
 			emit_zext_32(rd, ctx);
 		break;
 
Signature · 1 line
2.17.1
#2 lore raw

PATCH 2/4 bpf, riscv: Optimize FROM_LE using verifier_zext on RV64

In reply to Luke Nelson

This patch adds two optimizations for BPF_ALU BPF_END BPF_FROM_LE in the RV64 BPF JIT.

First, it enables the verifier zero-extension optimization to avoid zero extension when imm == 32. Second, it avoids generating code for imm == 64, since it is equivalent to a no-op.

---
 arch/riscv/net/bpf_jit_comp64.c | 20 ++++++++++++++------
 1 file changed, 14 insertions(+), 6 deletions(-)
arch/riscv/net/bpf_jit_comp64.c +14 -6
index e2636902a74e..c3ce9a911b66 100644
@@ -542,13 +542,21 @@ int bpf_jit_emit_insn(const struct bpf_insn *insn, struct rv_jit_context *ctx,
 
 	/* dst = BSWAP##imm(dst) */
 	case BPF_ALU | BPF_END | BPF_FROM_LE:
-	{
-		int shift = 64 - imm;
-
-		emit(rv_slli(rd, rd, shift), ctx);
-		emit(rv_srli(rd, rd, shift), ctx);
+		switch (imm) {
+		case 16:
+			emit(rv_slli(rd, rd, 48), ctx);
+			emit(rv_srli(rd, rd, 48), ctx);
+			break;
+		case 32:
+			if (!aux->verifier_zext)
+				emit_zext_32(rd, ctx);
+			break;
+		case 64:
+			/* Do nothing */
+			break;
+		}
 		break;
-	}
+
 	case BPF_ALU | BPF_END | BPF_FROM_BE:
 		emit(rv_addi(RV_REG_T2, RV_REG_ZERO, 0), ctx);
 
Signature · 1 line
2.17.1
#3 lore raw

PATCH 3/4 bpf, riscv: Optimize BPF_JMP BPF_K when imm == 0 on RV64

In reply to Luke Nelson

This patch adds an optimization to BPF_JMP (32- and 64-bit) BPF_K for when the BPF immediate is zero.

When the immediate is zero, the code can directly use the RISC-V zero register instead of loading a zero immediate to a temporary register first.

---
 arch/riscv/net/bpf_jit_comp64.c | 15 ++++++++++-----
 1 file changed, 10 insertions(+), 5 deletions(-)
arch/riscv/net/bpf_jit_comp64.c +10 -5
index c3ce9a911b66..b07cef952019 100644
@@ -796,7 +796,13 @@ int bpf_jit_emit_insn(const struct bpf_insn *insn, struct rv_jit_context *ctx,
 	case BPF_JMP32 | BPF_JSET | BPF_K:
 		rvoff = rv_offset(i, off, ctx);
 		s = ctx->ninsns;
-		emit_imm(RV_REG_T1, imm, ctx);
+		if (imm) {
+			emit_imm(RV_REG_T1, imm, ctx);
+			rs = RV_REG_T1;
+		} else {
+			/* If imm is 0, simply use zero register. */
+			rs = RV_REG_ZERO;
+		}
 		if (!is64) {
 			if (is_signed_bpf_cond(BPF_OP(code)))
 				emit_sext_32_rd(&rd, ctx);
@@ -811,11 +817,10 @@ int bpf_jit_emit_insn(const struct bpf_insn *insn, struct rv_jit_context *ctx,
 		if (BPF_OP(code) == BPF_JSET) {
 			/* Adjust for and */
 			rvoff -= 4;
-			emit(rv_and(RV_REG_T1, rd, RV_REG_T1), ctx);
-			emit_branch(BPF_JNE, RV_REG_T1, RV_REG_ZERO, rvoff,
-				    ctx);
+			emit(rv_and(rs, rd, rs), ctx);
+			emit_branch(BPF_JNE, rs, RV_REG_ZERO, rvoff, ctx);
 		} else {
-			emit_branch(BPF_OP(code), rd, RV_REG_T1, rvoff, ctx);
+			emit_branch(BPF_OP(code), rd, rs, rvoff, ctx);
 		}
 		break;
 
Signature · 1 line
2.17.1
#4 lore raw

PATCH 4/4 bpf, riscv: Optimize BPF_JSET BPF_K using andi on RV64

In reply to Luke Nelson

This patch optimizes BPF_JSET BPF_K by using a RISC-V andi instruction when the BPF immediate fits in 12 bits, instead of first loading the immediate to a temporary register.

Examples of generated code with and without this optimization:

BPF_JMP_IMM(BPF_JSET, R1, 2, 1) without optimization:

  20: li    t1,2
  24: and   t1,a0,t1
  28: bnez  t1,0x30

BPF_JMP_IMM(BPF_JSET, R1, 2, 1) with optimization:

20: andi t1,a0,2
24: bnez t1,0x2c

BPF_JMP32_IMM(BPF_JSET, R1, 2, 1) without optimization:

  20: li    t1,2
  24: mv    t2,a0
  28: slli  t2,t2,0x20
  2c: srli  t2,t2,0x20
  30: slli  t1,t1,0x20
  34: srli  t1,t1,0x20
  38: and   t1,t2,t1
  3c: bnez  t1,0x44

BPF_JMP32_IMM(BPF_JSET, R1, 2, 1) with optimization:

20: andi t1,a0,2
24: bnez t1,0x2c

In these examples, because the upper 32 bits of the sign-extended immediate are 0, BPF_JMP BPF_JSET and BPF_JMP32 BPF_JSET are equivalent and therefore the JIT produces identical code for them.

---
 arch/riscv/net/bpf_jit_comp64.c | 27 +++++++++++++++++++--------
 1 file changed, 19 insertions(+), 8 deletions(-)
arch/riscv/net/bpf_jit_comp64.c +19 -8
index b07cef952019..6cfd164cbe88 100644
@@ -792,8 +792,6 @@ int bpf_jit_emit_insn(const struct bpf_insn *insn, struct rv_jit_context *ctx,
 	case BPF_JMP32 | BPF_JSGE | BPF_K:
 	case BPF_JMP | BPF_JSLE | BPF_K:
 	case BPF_JMP32 | BPF_JSLE | BPF_K:
-	case BPF_JMP | BPF_JSET | BPF_K:
-	case BPF_JMP32 | BPF_JSET | BPF_K:
 		rvoff = rv_offset(i, off, ctx);
 		s = ctx->ninsns;
 		if (imm) {
@@ -813,15 +811,28 @@ int bpf_jit_emit_insn(const struct bpf_insn *insn, struct rv_jit_context *ctx,
 
 		/* Adjust for extra insns */
 		rvoff -= (e - s) << 2;
+		emit_branch(BPF_OP(code), rd, rs, rvoff, ctx);
+		break;
 
-		if (BPF_OP(code) == BPF_JSET) {
-			/* Adjust for and */
-			rvoff -= 4;
-			emit(rv_and(rs, rd, rs), ctx);
-			emit_branch(BPF_JNE, rs, RV_REG_ZERO, rvoff, ctx);
+	case BPF_JMP | BPF_JSET | BPF_K:
+	case BPF_JMP32 | BPF_JSET | BPF_K:
+		rvoff = rv_offset(i, off, ctx);
+		s = ctx->ninsns;
+		if (is_12b_int(imm)) {
+			emit(rv_andi(RV_REG_T1, rd, imm), ctx);
 		} else {
-			emit_branch(BPF_OP(code), rd, rs, rvoff, ctx);
+			emit_imm(RV_REG_T1, imm, ctx);
+			emit(rv_and(RV_REG_T1, rd, RV_REG_T1), ctx);
 		}
+		/* For jset32, we should clear the upper 32 bits of t1, but
+		 * sign-extension is sufficient here and saves one instruction,
+		 * as t1 is used only in comparison against zero.
+		 */
+		if (!is64 && imm < 0)
+			emit(rv_addiw(RV_REG_T1, RV_REG_T1, 0), ctx);
+		e = ctx->ninsns;
+		rvoff -= (e - s) << 2;
+		emit_branch(BPF_JNE, RV_REG_T1, RV_REG_ZERO, rvoff, ctx);
 		break;
 
 	/* function call */
Signature · 1 line
2.17.1
Reviewed-by Acked-by on the v1 series
#5 lore raw
In reply to Luke Nelson
Quoted 9 lines · Luke Nelson

This patch series introduces a set of optimizations to the BPF JIT on RV64. The optimizations are related to the verifier zero-extension optimization and BPF_JMP BPF_K.

We tested the optimizations on a QEMU riscv64 virt machine, using lib/test_bpf and test_verifier, and formally verified their correctness using Serval.

Luke and Xi,

Thanks a lot for working on this! Very nice series!

For the series:

Quoted 14 lines

Luke Nelson (4):
bpf, riscv: Enable missing verifier_zext optimizations on RV64 bpf, riscv: Optimize FROM_LE using verifier_zext on RV64 bpf, riscv: Optimize BPF_JMP BPF_K when imm == 0 on RV64 bpf, riscv: Optimize BPF_JSET BPF_K using andi on RV64

 arch/riscv/net/bpf_jit_comp64.c | 64 ++++++++++++++++++++++-----------
 1 file changed, 44 insertions(+), 20 deletions(-)
Cc Xi Wang <[email protected]>

--
2.17.1

maintainer Applied
#6 lore raw
Quoted 11 lines · Björn Töpel
Quoted 9 lines · Luke Nelson

This patch series introduces a set of optimizations to the BPF JIT on RV64. The optimizations are related to the verifier zero-extension optimization and BPF_JMP BPF_K.

We tested the optimizations on a QEMU riscv64 virt machine, using lib/test_bpf and test_verifier, and formally verified their correctness using Serval.

Quoted 18 lines

Luke and Xi,

Thanks a lot for working on this! Very nice series!

For the series:

Quoted 10 lines

Luke Nelson (4):
bpf, riscv: Enable missing verifier_zext optimizations on RV64 bpf, riscv: Optimize FROM_LE using verifier_zext on RV64 bpf, riscv: Optimize BPF_JMP BPF_K when imm == 0 on RV64 bpf, riscv: Optimize BPF_JSET BPF_K using andi on RV64

  arch/riscv/net/bpf_jit_comp64.c | 64 ++++++++++++++++++++++-----------
  1 file changed, 44 insertions(+), 20 deletions(-)
Cc Xi Wang <[email protected]>

Applied, thanks everyone!