Skip to contentSkip to conversation

bpf, riscv: Add support for signed arena loads

PATCH Superseded Chen Pei 2 participants 5 messages Aug 19

Subsystems (5)
Kselftest BPF RISC-V BPF selftests +1 more subsystems

v2 of 3 Superseded Latest is v3 →

lore
Share & tools
lore
author
#1 lore raw

PATCH v2 1/2 bpf, riscv: Add support for signed arena loads

In reply to Chen Pei

Signed loads from arena memory are currently rejected on riscv64, as bpf_jit_supports_insn() refuses BPF_MEMSX loads when in_arena is set, while x86 and arm64 gained support for them in v6.18. Compilers such as GCC-14 are free to generate signed loads into arena memory, which breaks loading of otherwise valid BPF programs on riscv64.

Implement BPF_PROBE_MEM32SX support in the RV64 JIT by reusing the existing arena handling: the arena base (RV_REG_ARENA) is added to the source register and the load is emitted with sign extension (lb/lh/lw). Add BPF_PROBE_MEM32SX to the add_exception_handler() mode gate so that faulting loads get an exception table entry which clears the destination register and resumes execution. Since BPF_PROBE_MEM32SX shares its mode value (0xc0) with BPF_ATOMIC, the gate accepts it only for LDX class instructions so that plain atomic instructions do not register exception table entries.

Verified by running the arena LDSX selftests (arena_ldsx_disasm, arena_ldsx_exception, arena_ldsx_s8/s16/s32) and the full arena_atomics test suite on riscv64 QEMU, all passing.

---

Changes in v2:
- Fix extable entry overflow breaking arena_atomics load

 arch/riscv/net/bpf_jit_comp64.c | 16 ++++++++++------
 1 file changed, 10 insertions(+), 6 deletions(-)
arch/riscv/net/bpf_jit_comp64.c +10 -6
index f9d5347ba966..acc1b16d2e2f 100644
@@ -777,6 +777,8 @@ static int add_exception_handler(const struct bpf_insn *insn, int dst_reg,
 	if (BPF_MODE(insn->code) != BPF_PROBE_MEM &&
 	    BPF_MODE(insn->code) != BPF_PROBE_MEMSX &&
 	    BPF_MODE(insn->code) != BPF_PROBE_MEM32 &&
+	    !(BPF_MODE(insn->code) == BPF_PROBE_MEM32SX &&
+	      BPF_CLASS(insn->code) == BPF_LDX) &&
 	    BPF_MODE(insn->code) != BPF_PROBE_ATOMIC)
 		return 0;
 
@@ -1902,13 +1904,19 @@ int bpf_jit_emit_insn(const struct bpf_insn *insn, struct rv_jit_context *ctx,
 	case BPF_LDX | BPF_PROBE_MEM32 | BPF_H:
 	case BPF_LDX | BPF_PROBE_MEM32 | BPF_W:
 	case BPF_LDX | BPF_PROBE_MEM32 | BPF_DW:
+	/* LDX | PROBE_MEM32SX: sign-extending arena load */
+	case BPF_LDX | BPF_PROBE_MEM32SX | BPF_B:
+	case BPF_LDX | BPF_PROBE_MEM32SX | BPF_H:
+	case BPF_LDX | BPF_PROBE_MEM32SX | BPF_W:
 	{
 		bool sign_ext;
 
 		sign_ext = BPF_MODE(insn->code) == BPF_MEMSX ||
-			   BPF_MODE(insn->code) == BPF_PROBE_MEMSX;
+			   BPF_MODE(insn->code) == BPF_PROBE_MEMSX ||
+			   BPF_MODE(insn->code) == BPF_PROBE_MEM32SX;
 
-		if (BPF_MODE(insn->code) == BPF_PROBE_MEM32) {
+		if (BPF_MODE(insn->code) == BPF_PROBE_MEM32 ||
+		    BPF_MODE(insn->code) == BPF_PROBE_MEM32SX) {
 			emit_add(RV_REG_T2, rs, RV_REG_ARENA, ctx);
 			rs = RV_REG_T2;
 		}
@@ -2126,10 +2134,6 @@ bool bpf_jit_supports_insn(struct bpf_insn *insn, bool in_arena)
 			if (insn->imm == BPF_CMPXCHG)
 				return rv_ext_enabled(ZACAS);
 			break;
-		case BPF_LDX | BPF_MEMSX | BPF_B:
-		case BPF_LDX | BPF_MEMSX | BPF_H:
-		case BPF_LDX | BPF_MEMSX | BPF_W:
-			return false;
 		}
 	}
 
Signature · 1 line
2.50.1

Chen Pei

author
#0 lore raw

v2 cover bpf, riscv: Add support for signed arena loads

Hi,

Signed loads from arena memory are currently unsupported on riscv64:
bpf_jit_supports_insn() rejects BPF_MEMSX loads when in_arena is set, so the verifier fails such programs with "sign extending loads from arena are not supported yet". The x86 and arm64 JITs gained support for them in v6.18 (a91ae3c89311, eab2a71f3a6a). Since compilers are free to generate signed loads into arena memory (e.g. GCC-14 was reported to do so), otherwise valid BPF programs fail to load on riscv64.

This series adds BPF_PROBE_MEM32SX support to the RV64 JIT and enables the corresponding selftests on riscv64:

1 implements signed arena loads in the RV64 JIT. The verifier
already converts MEMSX loads from PTR_TO_ARENA to BPF_PROBE_MEM32SX once bpf_jit_supports_insn() allows them, so the JIT reuses the existing arena handling: the arena base (RV_REG_ARENA) is added to the source register and the load is emitted with sign extension (lb/lh/lw). BPF_PROBE_MEM32SX is also added to the add_exception_handler() mode gate so faulting loads register an exception table entry that clears the destination register and resumes execution.

2 enables the arena LDSX tests on riscv64: JIT disassembly
assertions are added to arena_ldsx_disasm, and arena_ldsx_exception/s8/s16/s32 are now run on riscv64.

The series was verified on riscv64 with QEMU (-M virt): all five arena_ldsx tests pass, including the exception path (load from unallocated arena memory returns 0) and the sign-extension values (s8/s16/s32 tests return -1 as expected). The arena_atomics test suite also loads and runs cleanly.

Changes in v2:
- Fix an extable entry overflow breaking the arena_atomics load.
- Collect Reviewed-by on patch 2.

Note: the __jited assertions in patch 2 were derived from the JIT register allocation (R0->a5, R1->a0, R8->s3, R9->s4, arena base in s7) and the emit_ldx() code paths; happy to adjust them if a disassembler output detail differs.

Thanks,
Pei

Chen Pei (2):
bpf, riscv: Add support for signed arena loads selftests/bpf: Enable arena LDSX tests for riscv64

 arch/riscv/net/bpf_jit_comp64.c                 | 16 ++++++++++------
 .../testing/selftests/bpf/progs/verifier_ldsx.c | 17 +++++++++++++++++
 2 files changed, 27 insertions(+), 6 deletions(-)
Signature · 1 line
2.50.1
author
#4 lore raw

PATCH v2 2/2 selftests/bpf: Enable arena LDSX tests for riscv64

Now that the riscv64 JIT supports signed arena loads (BPF_PROBE_MEM32SX), enable the arena LDSX tests on riscv64:
add JIT disassembly assertions for arena_ldsx_disasm (arena base in s7, add into t2, sign-extending lw/lh/lb loads) and run arena_ldsx_exception and arena_ldsx_s8/s16/s32 on riscv64.

---
 .../testing/selftests/bpf/progs/verifier_ldsx.c | 17 +++++++++++++++++
 1 file changed, 17 insertions(+)
tools/testing/selftests/bpf/progs/verifier_ldsx.c +17 -0
index 41340877dc9d..ed0a0f159bc1 100644
@@ -286,6 +286,19 @@ __jited("add	x11, x0, x28")
 __jited("ldrsh	x22, [x11, #0x18]")
 __jited("add	x11, x0, x28")
 __jited("ldrsb	x22, [x11, #0x20]")
+__arch_riscv64
+__jited("add	t2, a5, s7")
+__jited("lw	s3, 0x10(t2)")
+__jited("add	t2, a5, s7")
+__jited("lh	s3, 0x18(t2)")
+__jited("add	t2, a5, s7")
+__jited("lb	s3, 0x20(t2)")
+__jited("add	t2, a0, s7")
+__jited("lw	s4, 0x10(t2)")
+__jited("add	t2, a0, s7")
+__jited("lh	s4, 0x18(t2)")
+__jited("add	t2, a0, s7")
+__jited("lb	s4, 0x20(t2)")
 __naked void arena_ldsx_disasm(void *ctx)
 {
 	asm volatile (
@@ -317,6 +330,7 @@ __description("Arena LDSX Exception")
 __success __retval(0)
 __arch_x86_64
 __arch_arm64
+__arch_riscv64
 __naked void arena_ldsx_exception(void *ctx)
 {
 	asm volatile (
@@ -338,6 +352,7 @@ __description("Arena LDSX, S8")
 __success __retval(-1)
 __arch_x86_64
 __arch_arm64
+__arch_riscv64
 __naked void arena_ldsx_s8(void *ctx)
 {
 	asm volatile (
@@ -369,6 +384,7 @@ __description("Arena LDSX, S16")
 __success __retval(-1)
 __arch_x86_64
 __arch_arm64
+__arch_riscv64
 __naked void arena_ldsx_s16(void *ctx)
 {
 	asm volatile (
@@ -400,6 +416,7 @@ __description("Arena LDSX, S32")
 __success __retval(-1)
 __arch_x86_64
 __arch_arm64
+__arch_riscv64
 __naked void arena_ldsx_s32(void *ctx)
 {
 	asm volatile (
Signature · 1 line
2.50.1
#2 lore raw
In reply to Chen Pei
This isn't a bug, but would the comment fit the surrounding style better as an expression? The neighbouring case-group comments all spell the access out in pseudo-code form:
Quoted 4 lines
arch/riscv/net/bpf_jit_comp64.c +0 -0
index 74efe4b138d2..56196b001286 100644

[ ... ]

Quoted 12 lines
arch/riscv/net/bpf_jit_comp64.c +4 -0
@@ -1908,13 +1910,19 @@ int bpf_jit_emit_insn(const struct bpf_insn *insn, struct rv_jit_context *ctx,
 	case BPF_LDX | BPF_PROBE_MEM32 | BPF_H:
 	case BPF_LDX | BPF_PROBE_MEM32 | BPF_W:
 	case BPF_LDX | BPF_PROBE_MEM32 | BPF_DW:
+	/* LDX | PROBE_MEM32SX: sign-extending arena load */
+	case BPF_LDX | BPF_PROBE_MEM32SX | BPF_B:
+	case BPF_LDX | BPF_PROBE_MEM32SX | BPF_H:
+	case BPF_LDX | BPF_PROBE_MEM32SX | BPF_W:
 	{
 		bool sign_ext;
 
 		sign_ext = BPF_MODE(insn->code) == BPF_MEMSX ||

arch/riscv/net/bpf_jit_comp64.c int bpf_jit_emit_insn(const struct bpf_insn *insn, struct rv_jit_context *ctx, · line 1928

Quoted 10 lines
arch/riscv/net/bpf_jit_comp64.c +4 -2
-			   BPF_MODE(insn->code) == BPF_PROBE_MEMSX;
+			   BPF_MODE(insn->code) == BPF_PROBE_MEMSX ||
+			   BPF_MODE(insn->code) == BPF_PROBE_MEM32SX;
 
-		if (BPF_MODE(insn->code) == BPF_PROBE_MEM32) {
+		if (BPF_MODE(insn->code) == BPF_PROBE_MEM32 ||
+		    BPF_MODE(insn->code) == BPF_PROBE_MEM32SX) {
 			emit_add(RV_REG_T2, rs, RV_REG_ARENA, ctx);
 			rs = RV_REG_T2;
 		}

This isn't a bug, but would the comment fit the surrounding style better as an expression? The neighbouring case-group comments all spell the access out in pseudo-code form:

    /* LDX: dst = *(unsigned size *)(src + off) */
    /* LDSX: dst = *(signed size *)(src + off) */
    /* LDX | PROBE_MEM32: dst = *(unsigned size *)(src + RV_REG_ARENA + off) */
    /* ST | PROBE_MEM32: *(size *)(dst + RV_REG_ARENA + off) = imm */
    /* STX | PROBE_MEM32: *(size *)(dst + RV_REG_ARENA + off) = src */

Something like this would match the pattern:

    /* LDX | PROBE_MEM32SX: dst = *(signed size *)(src + RV_REG_ARENA + off) */

---
AI reviewed your patch. Please fix the bug or email reply why it's not a bug.
See: https://github.com/kernel-patches/vmtest/blob/master/ci/claude/README.md

CI run summary: https://github.com/kernel-patches/bpf/actions/runs/32235574103

author
#3 lore raw

Hi,

Quoted 10 lines · On 2026/8/19 17:37, [email protected] wrote:

This isn't a bug, but would the comment fit the surrounding style better as an expression? The neighbouring case-group comments all spell the access out in pseudo-code form:

    /* LDX: dst = *(unsigned size *)(src + off) */
    /* LDSX: dst = *(signed size *)(src + off) */
    /* LDX | PROBE_MEM32: dst = *(unsigned size *)(src + RV_REG_ARENA + off) */
    /* ST | PROBE_MEM32: *(size *)(dst + RV_REG_ARENA + off) = imm */
    /* STX | PROBE_MEM32: *(size *)(dst + RV_REG_ARENA + off) = src */
Quoted 3 lines

Something like this would match the pattern:

    /* LDX | PROBE_MEM32SX: dst = *(signed size *)(src + RV_REG_ARENA + off) */

Agreed, consistency with the neighbouring comments is more valuable here. The pseudo-code form will be restored in v3.

Best regards,
Pei