Skip to contentSkip to conversation

bpf: verifier: fix integer wrap in regs_refine_cond_op() BPF_JNE case

PATCH In review Ibrahim Zein 4 participants 4 messages Mar 18

Subsystems (2)
lore
Share & tools
lore
#0 lore raw

PATCH v3 1/1 bpf: verifier: fix integer wrap in regs_refine_cond_op() BPF_JNE case

From: Ibrahim Zein <[email protected]>

In regs_refine_cond_op(), the BPF_JNE case narrows register bounds without guarding against wrap-around at type boundaries. When umin_value equals U64_MAX, incrementing it wraps to 0, producing an inconsistent bounds state (umin=0, umax=U64_MAX-1).

This can be triggered by constructing a register state where:
r1 = unknown_byte | 0xFFFFFFFFFFFFFFFE
-> var_off = (value=U64_MAX-1, mask=1)
-> constrain with JGE U64_MAX -> umin=umax=U64_MAX, tnum non-const -> compare with BPF_JNE const=U64_MAX -> wrap occurs

The comment above these checks incorrectly claims this boundary condition is impossible. However, is_reg_const() uses tnum_is_const() which requires mask == 0, so a register can have umin == umax while tnum is non-const (mask != 0). Remove the misleading comment and add explicit boundary guards to prevent wrap-around in all {u,s}{min,max}{32,64} variants.

Fixes d028f87517d6 ("bpf: make the verifier tracks the "not equal" for regs")

---

kernel/bpf/verifier.c +16 -17
@@ -17046,31 +17046,30 @@
 		 */
 		val = reg_const_value(reg2, is_jmp32);
 		if (is_jmp32) {
-			/* u32_min_value is not equal to 0xffffffff at this point,
-			 * because otherwise u32_max_value is 0xffffffff as well,
-			 * in such a case both reg1 and reg2 would be constants,
-			 * jump would be predicted and reg_set_min_max() won't
-			 * be called.
-			 *
-			 * Same reasoning works for all {u,s}{min,max}{32,64} cases
-			 * below.
-			 */
-			if (reg1->u32_min_value == (u32)val)
+			if (reg1->u32_min_value == (u32)val &&
+			    reg1->u32_min_value != U32_MAX)
 				reg1->u32_min_value++;
-			if (reg1->u32_max_value == (u32)val)
+			if (reg1->u32_max_value == (u32)val &&
+			    reg1->u32_max_value != 0)
 				reg1->u32_max_value--;
-			if (reg1->s32_min_value == (s32)val)
+			if (reg1->s32_min_value == (s32)val &&
+			    reg1->s32_min_value != S32_MAX)
 				reg1->s32_min_value++;
-			if (reg1->s32_max_value == (s32)val)
+			if (reg1->s32_max_value == (s32)val &&
+			    reg1->s32_max_value != S32_MIN)
 				reg1->s32_max_value--;
 		} else {
-			if (reg1->umin_value == (u64)val)
+			if (reg1->umin_value == (u64)val &&
+			    reg1->umin_value != U64_MAX)
 				reg1->umin_value++;
-			if (reg1->umax_value == (u64)val)
+			if (reg1->umax_value == (u64)val &&
+			    reg1->umax_value != 0)
 				reg1->umax_value--;
-			if (reg1->smin_value == (s64)val)
+			if (reg1->smin_value == (s64)val &&
+			    reg1->smin_value != S64_MAX)
 				reg1->smin_value++;
-			if (reg1->smax_value == (s64)val)
+			if (reg1->smax_value == (s64)val &&
+			    reg1->smax_value != S64_MIN)
 				reg1->smax_value--;
 		}
 		break;
#1 lore raw
When umin == umax == val and val == U64_MAX, the umin guard prevents the increment (U64_MAX == U64_MAX), but the umax guard allows the decrement (U64_MAX != 0). This leaves umin_value at U64_MAX and…
Quoted 4 lines
kernel/bpf/verifier.c +0 -0
index 01c18f4268de..5e0fbeca8f37 100644

[ ... ]

Quoted 1 line
		val = reg_const_value(reg2, is_jmp32);

[ ... ]

Quoted 7 lines
		} else {
kernel/bpf/verifier.c +2 -0
+			if (reg1->umin_value == (u64)val &&
+			    reg1->umin_value != U64_MAX)
				reg1->umin_value++;
kernel/bpf/verifier.c +2 -0
+			if (reg1->umax_value == (u64)val &&
+			    reg1->umax_value != 0)
				reg1->umax_value--;

When umin == umax == val and val == U64_MAX, the umin guard prevents the increment (U64_MAX == U64_MAX), but the umax guard allows the decrement (U64_MAX != 0). This leaves umin_value at U64_MAX and umax_value at U64_MAX-1, creating an inverted range.

This state survives reg_bounds_sync() because
__update_reg64_bounds() recomputes:

    umin = max(umin, var_off.value)   -> still U64_MAX
    umax = min(umax, var_off.value | var_off.mask) -> still U64_MAX-1

and then reg_bounds_sanity_check() catches umin > umax, triggering verifier_bug() which calls BPF_WARN_ONCE().

The scenario from the commit message (var_off with mask=1 after JGE U64_MAX) reaches this path because is_scalar_branch_taken() returns -1 for BPF_JNE when tnum overlaps but is not const, so both branches are explored.

The same pattern applies symmetrically when val == 0 (umax guard fires, umin proceeds), and to all the s32/s64/u32 variants at their respective type boundaries.

Would it make sense to handle this dead-branch case before the individual adjustments? The BPF_JSET | BPF_X case in the same function already does this:

    /* Forget the ranges before narrowing tnums, to avoid invariant
     * violations if we're on a dead branch.
     */
    __mark_reg_unbounded(reg1);

A similar approach here, or skipping adjustments entirely when umin == umax == val, would avoid the verifier_bug() trigger on what is an architecturally expected dead branch.

Quoted 8 lines
kernel/bpf/verifier.c +2 -0
+			if (reg1->smin_value == (s64)val &&
+			    reg1->smin_value != S64_MAX)
				reg1->smin_value++;
kernel/bpf/verifier.c +2 -0
+			if (reg1->smax_value == (s64)val &&
+			    reg1->smax_value != S64_MIN)
				reg1->smax_value--;
		}
		break;

---
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/23223625211

#2 lore raw
In reply to Ibrahim Zein
Quoted 21 lines · Ibrahim Zein

From: Ibrahim Zein <[email protected]>

In regs_refine_cond_op(), the BPF_JNE case narrows register bounds without guarding against wrap-around at type boundaries. When umin_value equals U64_MAX, incrementing it wraps to 0, producing an inconsistent bounds state (umin=0, umax=U64_MAX-1).

This can be triggered by constructing a register state where:
r1 = unknown_byte | 0xFFFFFFFFFFFFFFFE
-> var_off = (value=U64_MAX-1, mask=1)
-> constrain with JGE U64_MAX -> umin=umax=U64_MAX, tnum non-const -> compare with BPF_JNE const=U64_MAX -> wrap occurs

The comment above these checks incorrectly claims this boundary condition is impossible. However, is_reg_const() uses tnum_is_const() which requires mask == 0, so a register can have umin == umax while tnum is non-const (mask != 0). Remove the misleading comment and add explicit boundary guards to prevent wrap-around in all {u,s}{min,max}{32,64} variants.

Quoted 3 lines
Fixes d028f87517d6 ("bpf: make the verifier tracks the "not equal" for regs")

Please stop resending incorrect patch.
Instead please create a selftest that demonstrates the issue.

#3 lore raw
In reply to Ibrahim Zein

On 2026/3/18 08:41 Ibrahim Zein <[email protected]> write:

Quoted 13 lines

From: Ibrahim Zein <[email protected]>

In regs_refine_cond_op(), the BPF_JNE case narrows register bounds without guarding against wrap-around at type boundaries. When umin_value equals U64_MAX, incrementing it wraps to 0, producing an inconsistent bounds state (umin=0, umax=U64_MAX-1).

This can be triggered by constructing a register state where:
r1 = unknown_byte | 0xFFFFFFFFFFFFFFFE
-> var_off = (value=U64_MAX-1, mask=1)
-> constrain with JGE U64_MAX -> umin=umax=U64_MAX, tnum non-const -> compare with BPF_JNE const=U64_MAX -> wrap occurs

Quoted 6 lines

The comment above these checks incorrectly claims this boundary condition is impossible. However, is_reg_const() uses tnum_is_const() which requires mask == 0, so a register can have umin == umax while tnum is non-const (mask != 0). Remove the misleading comment and add explicit boundary guards to prevent wrap-around in all {u,s}{min,max}{32,64} variants.

Hi, Ibrahim. As Alexei said, your patch is incorrect, and you should generate it with:

    git format-patch --subject-prefix="PATCH bpf-next" -1

How did you find this issue? By AI? I ask this because I don't think this is a issue that can be triggered. As you said, the key problem is that r1 is "umin=umax=U64_MAX, tnum non-const", so it's a known value. Therefore, the return value of is_branch_taken() should always be 0 or 1 in check_cond_jmp_op, as r2 is a const in this case. Then, reg_set_min_max() won't be called, as we don't need to adjust the min/max for a known register. Right?

Thanks!
Menglong Dong

Quoted 57 lines
Fixes d028f87517d6 ("bpf: make the verifier tracks the "not equal" for regs")

---

kernel/bpf/verifier.c +16 -17
@@ -17046,31 +17046,30 @@
 		 */
 		val = reg_const_value(reg2, is_jmp32);
 		if (is_jmp32) {
-			/* u32_min_value is not equal to 0xffffffff at this point,
-			 * because otherwise u32_max_value is 0xffffffff as well,
-			 * in such a case both reg1 and reg2 would be constants,
-			 * jump would be predicted and reg_set_min_max() won't
-			 * be called.
-			 *
-			 * Same reasoning works for all {u,s}{min,max}{32,64} cases
-			 * below.
-			 */
-			if (reg1->u32_min_value == (u32)val)
+			if (reg1->u32_min_value == (u32)val &&
+			    reg1->u32_min_value != U32_MAX)
 				reg1->u32_min_value++;
-			if (reg1->u32_max_value == (u32)val)
+			if (reg1->u32_max_value == (u32)val &&
+			    reg1->u32_max_value != 0)
 				reg1->u32_max_value--;
-			if (reg1->s32_min_value == (s32)val)
+			if (reg1->s32_min_value == (s32)val &&
+			    reg1->s32_min_value != S32_MAX)
 				reg1->s32_min_value++;
-			if (reg1->s32_max_value == (s32)val)
+			if (reg1->s32_max_value == (s32)val &&
+			    reg1->s32_max_value != S32_MIN)
 				reg1->s32_max_value--;
 		} else {
-			if (reg1->umin_value == (u64)val)
+			if (reg1->umin_value == (u64)val &&
+			    reg1->umin_value != U64_MAX)
 				reg1->umin_value++;
-			if (reg1->umax_value == (u64)val)
+			if (reg1->umax_value == (u64)val &&
+			    reg1->umax_value != 0)
 				reg1->umax_value--;
-			if (reg1->smin_value == (s64)val)
+			if (reg1->smin_value == (s64)val &&
+			    reg1->smin_value != S64_MAX)
 				reg1->smin_value++;
-			if (reg1->smax_value == (s64)val)
+			if (reg1->smax_value == (s64)val &&
+			    reg1->smax_value != S64_MIN)
 				reg1->smax_value--;
 		}
 		break;