Skip to contentSkip to conversation

bpf, riscv: Register extable entries at probe insn call sites

PATCH In review Chen Pei 4 participants 6 messages Sep 1 – Sep 7

Subsystems (3)
lore
Share & tools
lore

Chen Pei

author
#0 lore raw

PATCH 1/1 bpf, riscv: Register extable entries at probe insn call sites

add_exception_handler() filters insns by matching BPF_MODE() against the probe mode values. These values occupy unused opcode slots and collide with legitimate encodings: BPF_PROBE_MEM32SX (0xc0) shares its mode value with BPF_ATOMIC. Relying on the mode alone therefore risks registering exception table entries for the wrong instructions, as would happen for plain atomics without the LDX class check.

Make the decision explicit at the call sites instead: register an entry only when emitting a probe load, a PROBE_MEM32 store, or a PROBE_ATOMIC atomic, and drop the fragile mode gate from add_exception_handler() along with the now unused insn argument. No functional change intended.

---
base-commit: d761934c9483ecde93fe99d8705282f716dfee50

 arch/riscv/net/bpf_jit_comp64.c | 50 ++++++++++++++++++---------------
 1 file changed, 28 insertions(+), 22 deletions(-)
arch/riscv/net/bpf_jit_comp64.c +28 -22
index e7378be171a9..726c0169fc60 100644
@@ -731,9 +731,8 @@ bool ex_handler_bpf(const struct exception_table_entry *ex,
 	return true;
 }
 
-/* For accesses to BTF pointers, add an entry to the exception table */
-static int add_exception_handler(const struct bpf_insn *insn, int dst_reg,
-				 struct rv_jit_context *ctx)
+/* Add an entry to the exception table for a faulting probe insn */
+static int add_exception_handler(int dst_reg, struct rv_jit_context *ctx)
 {
 	struct exception_table_entry *ex;
 	unsigned long pc;
@@ -744,14 +743,6 @@ static int add_exception_handler(const struct bpf_insn *insn, int dst_reg,
 	    ctx->ex_insn_off <= 0 || ctx->ex_jmp_off <= 0)
 		return 0;
 
-	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;
-
 	if (WARN_ON_ONCE(ctx->nexentries >= ctx->prog->aux->num_exentries))
 		return -EINVAL;
 
@@ -1360,6 +1351,15 @@ int arch_prepare_bpf_trampoline(struct bpf_tramp_image *im, void *ro_image,
 	return ret < 0 ? ret : size;
 }
 
+/* Probe loads fault on unmapped memory and need an exception table entry */
+static bool insn_is_probe_ldx(const struct bpf_insn *insn)
+{
+	return 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;
+}
+
 int bpf_jit_emit_insn(const struct bpf_insn *insn, struct rv_jit_context *ctx,
 		      bool extra_pass)
 {
@@ -1929,9 +1929,11 @@ int bpf_jit_emit_insn(const struct bpf_insn *insn, struct rv_jit_context *ctx,
 
 		emit_ldx(rd, off, rs, BPF_SIZE(code), sign_ext, ctx);
 
-		ret = add_exception_handler(insn, rd, ctx);
-		if (ret)
-			return ret;
+		if (insn_is_probe_ldx(insn)) {
+			ret = add_exception_handler(rd, ctx);
+			if (ret)
+				return ret;
+		}
 
 		if (BPF_SIZE(code) != BPF_DW && insn_is_zext(&insn[1]))
 			return 1;
@@ -1959,9 +1961,11 @@ int bpf_jit_emit_insn(const struct bpf_insn *insn, struct rv_jit_context *ctx,
 
 		emit_st(rd, off, imm, BPF_SIZE(code), ctx);
 
-		ret = add_exception_handler(insn, REG_DONT_CLEAR_MARKER, ctx);
-		if (ret)
-			return ret;
+		if (BPF_MODE(insn->code) == BPF_PROBE_MEM32) {
+			ret = add_exception_handler(REG_DONT_CLEAR_MARKER, ctx);
+			if (ret)
+				return ret;
+		}
 		break;
 
 	/* STX: *(size *)(dst + off) = src */
@@ -1981,9 +1985,11 @@ int bpf_jit_emit_insn(const struct bpf_insn *insn, struct rv_jit_context *ctx,
 
 		emit_stx(rd, off, rs, BPF_SIZE(code), ctx);
 
-		ret = add_exception_handler(insn, REG_DONT_CLEAR_MARKER, ctx);
-		if (ret)
-			return ret;
+		if (BPF_MODE(insn->code) == BPF_PROBE_MEM32) {
+			ret = add_exception_handler(REG_DONT_CLEAR_MARKER, ctx);
+			if (ret)
+				return ret;
+		}
 		break;
 
 	/* Atomics */
@@ -2001,7 +2007,7 @@ int bpf_jit_emit_insn(const struct bpf_insn *insn, struct rv_jit_context *ctx,
 			ret = emit_atomic_rmw(rd, rs, insn, ctx);
 
 		/* ret can be 1 (skip-zext); extable entry still needs to be added */
-		if (ret >= 0) {
+		if (ret >= 0 && BPF_MODE(insn->code) == BPF_PROBE_ATOMIC) {
 			/*
 			 * A load-acquire reads into dst_reg, and a read-modify-write
 			 * carrying BPF_FETCH reads the old value into src_reg, or into
@@ -2010,7 +2016,7 @@ int bpf_jit_emit_insn(const struct bpf_insn *insn, struct rv_jit_context *ctx,
 			 */
 			int load_reg = bpf_atomic_load_reg(insn);
 
-			ret = add_exception_handler(insn, load_reg < 0 ?
+			ret = add_exception_handler(load_reg < 0 ?
 					REG_DONT_CLEAR_MARKER : regmap[load_reg],
 					ctx) ?: ret;
 		}
Signature · 1 line
2.50.1
#1 lore raw
Thank you for your contribution! Sashiko AI review found 1 potential issue(s) to consider:

Thank you for your contribution! Sashiko AI review found 1 potential issue(s) to consider:

Pre-existing issues:
- [High] BPF_PROBE_ATOMIC cmpxchg on RISC-V without ZACAS lacks exception table entries, leading to an unhandled page fault and kernel panic for unmapped arena accesses. --

commit 77ab37f77b219155ec1dc7c551ef131f6c642039
Author: Chen Pei <[email protected]>
bpf, riscv: Register extable entries at probe insn call sites

This commit removes the mode gate from add_exception_handler() and explicitly registers exception handlers only at probe instruction call sites. It aims to prevent collisions between probe mode values and legitimate instruction encodings.

Quoted 4 lines
arch/riscv/net/bpf_jit_comp64.c +0 -0
index e7378be171a9f..726c0169fc60f 100644

[ ... ]

Quoted 10 lines
arch/riscv/net/bpf_jit_comp64.c +1 -1
@@ -2001,7 +2007,7 @@ int bpf_jit_emit_insn(const struct bpf_insn *insn, struct rv_jit_context *ctx,
 			ret = emit_atomic_rmw(rd, rs, insn, ctx);
 
 		/* ret can be 1 (skip-zext); extable entry still needs to be added */
-		if (ret >= 0) {
+		if (ret >= 0 && BPF_MODE(insn->code) == BPF_PROBE_ATOMIC) {
 			/*
 			 * A load-acquire reads into dst_reg, and a read-modify-write
 			 * carrying BPF_FETCH reads the old value into src_reg, or into

arch/riscv/net/bpf_jit_comp64.c int bpf_jit_emit_insn(const struct bpf_insn *insn, struct rv_jit_context *ctx, · line 2022

Quoted 8 lines
arch/riscv/net/bpf_jit_comp64.c +1 -1
@@ -2010,7 +2016,7 @@ int bpf_jit_emit_insn(const struct bpf_insn *insn, struct rv_jit_context *ctx,
 			 */
 			int load_reg = bpf_atomic_load_reg(insn);
 
-			ret = add_exception_handler(insn, load_reg < 0 ?
+			ret = add_exception_handler(load_reg < 0 ?
 					REG_DONT_CLEAR_MARKER : regmap[load_reg],
 					ctx) ?: ret;
 		}

[Severity: High]
This isn't a bug introduced by this patch, but does emit_cmpxchg() correctly establish an exception handler boundary when ZACAS is not enabled?

If a BPF program uses BPF_PROBE_ATOMIC with BPF_CMPXCHG on an unmapped memory address on a RISC-V CPU without the ZACAS extension, emit_cmpxchg() falls back to a load-reserved/store-conditional loop:

arch/riscv/net/bpf_jit.h:emit_cmpxchg() {
    ...
	if (rv_ext_enabled(ZACAS)) {
		ctx->ex_insn_off = ctx->ninsns;
		emit(is64 ? rvzacas_amocas_d(r0, rs, rd, 1, 1) :
		     rvzacas_amocas_w(r0, rs, rd, 1, 1), ctx);
		ctx->ex_jmp_off = ctx->ninsns;
        ...
		return;
	}

	if (is64)
		emit_mv(RV_REG_T2, r0, ctx);
	else
		emit_addiw(RV_REG_T2, r0, 0, ctx);
	emit(is64 ? rv_lr_d(r0, 0, rd, 0, 0) :
	     rv_lr_w(r0, 0, rd, 0, 0), ctx);
    ...
}

In the non-ZACAS path, ctx->ex_insn_off and ctx->ex_jmp_off are not updated before emitting the faulting load-reserved instruction.

When add_exception_handler() is subsequently called here in bpf_jit_emit_insn(), it will use stale offsets. Can this cause an unhandled page fault and a kernel panic when accessing unmapped arena memory?

Signature · 1 line
Sashiko AI review · https://sashiko.dev/#/patchset/[email protected]?part=1
#2 lore raw
In reply to Chen Pei
This looks like a bug fix for the fragile mode-based filtering in add_exception_handler(). Should this include:
Quoted 9 lines

bpf, riscv: Register extable entries at probe insn call sites

add_exception_handler() filters insns by matching BPF_MODE() against the probe mode values. These values occupy unused opcode slots and collide with legitimate encodings: BPF_PROBE_MEM32SX (0xc0) shares its mode value with BPF_ATOMIC. Relying on the mode alone therefore risks registering exception table entries for the wrong instructions, as would happen for plain atomics without the LDX class check.

Quoted 7 lines

Make the decision explicit at the call sites instead: register an entry only when emitting a probe load, a PROBE_MEM32 store, or a PROBE_ATOMIC atomic, and drop the fragile mode gate from add_exception_handler() along with the now unused insn argument. No functional change intended.

This looks like a bug fix for the fragile mode-based filtering in add_exception_handler(). Should this include:

Fixes b4757714cc63 ("bpf, riscv: Add support for signed arena loads")

---
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/33510775602

author
#3 lore raw
Quoted 4 lines

This looks like a bug fix for the fragile mode-based filtering in add_exception_handler(). Should this include:

Fixes b4757714cc63 ("bpf, riscv: Add support for signed arena loads")

No Fixes tag, because the mode/class collision has never been a live bug in any released tree. b4757714cc63 introduced both the BPF_PROBE_MEM32SX support and the LDX class guard in the mode gate, in the same commit, precisely to keep plain atomics (which share the 0xc0 mode value) out of the exception table. Its commit message says so explicitly:

"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."

For context: I noticed this fragile shared-state gate while working on v1 of "bpf, riscv: Add support for signed arena loads", where the class guard was added as a preventive measure. This patch now follows up with the actual cleanup, making the registration decision explicit at each call site so future probe modes cannot silently collide with legitimate encodings.

Also, regarding sashiko-bot's separate concern about BPF_PROBE_ATOMIC cmpxchg without ZACAS: that path is unreachable, as arena BPF_CMPXCHG is already gated on ZACAS by bpf_jit_supports_insn(), and add_exception_handler() resets its offsets after every registration, so stale offsets cannot be consumed. A separate fix I have prepared, "bpf, riscv: Make arena support depend on ZACAS", further rejects arena map creation on non-ZACAS systems altogether.

Best regards,
Pei

maintainer
#4 lore raw
In reply to Chen Pei
Quoted 37 lines · Chen Pei

add_exception_handler() filters insns by matching BPF_MODE() against the probe mode values. These values occupy unused opcode slots and collide with legitimate encodings: BPF_PROBE_MEM32SX (0xc0) shares its mode value with BPF_ATOMIC. Relying on the mode alone therefore risks registering exception table entries for the wrong instructions, as would happen for plain atomics without the LDX class check.

Make the decision explicit at the call sites instead: register an entry only when emitting a probe load, a PROBE_MEM32 store, or a PROBE_ATOMIC atomic, and drop the fragile mode gate from add_exception_handler() along with the now unused insn argument. No functional change intended.

---
base-commit: d761934c9483ecde93fe99d8705282f716dfee50

  arch/riscv/net/bpf_jit_comp64.c | 50 ++++++++++++++++++---------------
  1 file changed, 28 insertions(+), 22 deletions(-)
arch/riscv/net/bpf_jit_comp64.c +2 -3
index e7378be171a9..726c0169fc60 100644
@@ -731,9 +731,8 @@ bool ex_handler_bpf(const struct exception_table_entry *ex,
  	return true;
  }
 
-/* For accesses to BTF pointers, add an entry to the exception table */
-static int add_exception_handler(const struct bpf_insn *insn, int dst_reg,
-				 struct rv_jit_context *ctx)
+/* Add an entry to the exception table for a faulting probe insn */
+static int add_exception_handler(int dst_reg, struct rv_jit_context *ctx)
  {
  	struct exception_table_entry *ex;
  	unsigned long pc;

arch/riscv/net/bpf_jit_comp64.c static int add_exception_handler(const struct bpf_insn *insn, int dst_reg, · line 753

Quoted 10 lines
arch/riscv/net/bpf_jit_comp64.c +0 -7
@@ -744,14 +743,6 @@ static int add_exception_handler(const struct bpf_insn *insn, int dst_reg,
  	    ctx->ex_insn_off <= 0 || ctx->ex_jmp_off <= 0)
  		return 0;
 
-	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;

Hi Pei,

This patch looks fine, but I still tend not to introduce these changes. I think the function itself should handle its own defensive checks, and we should also avoid "shotgun surgery".

Quoted 83 lines
arch/riscv/net/bpf_jit_comp64.c +26 -12
-
  	if (WARN_ON_ONCE(ctx->nexentries >= ctx->prog->aux->num_exentries))
  		return -EINVAL;
 
@@ -1360,6 +1351,15 @@ int arch_prepare_bpf_trampoline(struct bpf_tramp_image *im, void *ro_image,
  	return ret < 0 ? ret : size;
  }
 
+/* Probe loads fault on unmapped memory and need an exception table entry */
+static bool insn_is_probe_ldx(const struct bpf_insn *insn)
+{
+	return 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;
+}
+
  int bpf_jit_emit_insn(const struct bpf_insn *insn, struct rv_jit_context *ctx,
  		      bool extra_pass)
  {
@@ -1929,9 +1929,11 @@ int bpf_jit_emit_insn(const struct bpf_insn *insn, struct rv_jit_context *ctx,
 
  		emit_ldx(rd, off, rs, BPF_SIZE(code), sign_ext, ctx);
 
-		ret = add_exception_handler(insn, rd, ctx);
-		if (ret)
-			return ret;
+		if (insn_is_probe_ldx(insn)) {
+			ret = add_exception_handler(rd, ctx);
+			if (ret)
+				return ret;
+		}
 
  		if (BPF_SIZE(code) != BPF_DW && insn_is_zext(&insn[1]))
  			return 1;
@@ -1959,9 +1961,11 @@ int bpf_jit_emit_insn(const struct bpf_insn *insn, struct rv_jit_context *ctx,
 
  		emit_st(rd, off, imm, BPF_SIZE(code), ctx);
 
-		ret = add_exception_handler(insn, REG_DONT_CLEAR_MARKER, ctx);
-		if (ret)
-			return ret;
+		if (BPF_MODE(insn->code) == BPF_PROBE_MEM32) {
+			ret = add_exception_handler(REG_DONT_CLEAR_MARKER, ctx);
+			if (ret)
+				return ret;
+		}
  		break;
 
  	/* STX: *(size *)(dst + off) = src */
@@ -1981,9 +1985,11 @@ int bpf_jit_emit_insn(const struct bpf_insn *insn, struct rv_jit_context *ctx,
 
  		emit_stx(rd, off, rs, BPF_SIZE(code), ctx);
 
-		ret = add_exception_handler(insn, REG_DONT_CLEAR_MARKER, ctx);
-		if (ret)
-			return ret;
+		if (BPF_MODE(insn->code) == BPF_PROBE_MEM32) {
+			ret = add_exception_handler(REG_DONT_CLEAR_MARKER, ctx);
+			if (ret)
+				return ret;
+		}
  		break;
 
  	/* Atomics */
@@ -2001,7 +2007,7 @@ int bpf_jit_emit_insn(const struct bpf_insn *insn, struct rv_jit_context *ctx,
  			ret = emit_atomic_rmw(rd, rs, insn, ctx);
 
  		/* ret can be 1 (skip-zext); extable entry still needs to be added */
-		if (ret >= 0) {
+		if (ret >= 0 && BPF_MODE(insn->code) == BPF_PROBE_ATOMIC) {
  			/*
  			 * A load-acquire reads into dst_reg, and a read-modify-write
  			 * carrying BPF_FETCH reads the old value into src_reg, or into
@@ -2010,7 +2016,7 @@ int bpf_jit_emit_insn(const struct bpf_insn *insn, struct rv_jit_context *ctx,
  			 */
  			int load_reg = bpf_atomic_load_reg(insn);
 
-			ret = add_exception_handler(insn, load_reg < 0 ?
+			ret = add_exception_handler(load_reg < 0 ?
  					REG_DONT_CLEAR_MARKER : regmap[load_reg],
  					ctx) ?: ret;
  		}
author
#5 lore raw
Quoted 3 lines

This patch looks fine, but I still tend not to introduce these changes. I think the function itself should handle its own defensive checks, and we should also avoid "shotgun surgery".

Indeed, you have a point. Thanks for the review.

Best regards,
Pei