Skip to contentSkip to conversation

Fix invariant violation for single-value tnums

PATCH Applied Paul Chaignon 3 participants 8 messages Feb 27 – Mar 25

Subsystems (4)
lore
Share & tools
lore
#0 lore raw

v3 cover Fix invariant violation for single-value tnums

We're hitting an invariant violation in Cilium that sometimes leads to BPF programs being rejected and Cilium failing to start [1]. As far as I know this is the first case of invariant violation found in a real program (i.e., not by a fuzzer). The following extract from verifier logs shows what's happening:

  from 201 to 236: R1=0 R6=ctx() R7=1 R9=scalar(smin=umin=smin32=umin32=3584,smax=umax=smax32=umax32=3840,var_off=(0xe00; 0x100)) R10=fp0
  236: R1=0 R6=ctx() R7=1 R9=scalar(smin=umin=smin32=umin32=3584,smax=umax=smax32=umax32=3840,var_off=(0xe00; 0x100)) R10=fp0
  ; if (magic == MARK_MAGIC_HOST || magic == MARK_MAGIC_OVERLAY || magic == MARK_MAGIC_ENCRYPT) @ bpf_host.c:1337
  236: (16) if w9 == 0xe00 goto pc+45   ; R9=scalar(smin=umin=smin32=umin32=3585,smax=umax=smax32=umax32=3840,var_off=(0xe00; 0x100))
  237: (16) if w9 == 0xf00 goto pc+1
  verifier bug: REG INVARIANTS VIOLATION (false_reg1): range bounds violation u64=[0xe01, 0xe00] s64=[0xe01, 0xe00] u32=[0xe01, 0xe00] s32=[0xe01, 0xe00] var_off=(0xe00, 0x0)

More details are given in the second patch, but in short, the verifier should be able to detect that the false branch of instruction 237 is never true. After instruction 236, the u64 range and the tnum overlap in a single value, 0xf00.

The long-term solution to invariant violation is likely to rely on the refinement + invariant violation check to detect dead branches, as started by Eduard. To fix the current issue, we need something with less refactoring that we can backport to affected kernels.

The solution implemented in the second patch is to improve the bounds refinement to avoid this case. It relies on a new tnum helper, tnum_step, first sent as an RFC in [2]. The last two patches extend and update the selftests.

Changes in v3:
- Fix commit description error spotted by AI bot.
- Simplify constants in first two tests (Eduard).
- Rework comment on third test (Eduard).
- Add two new negative test cases (Eduard).
- Rebased.
Changes in v2:
- Add guard suggested by Hari in tnum_step, to avoid undefined behavior spotted by AI code review.
- Add explanation diagrams in code as suggested by Eduard.
- Rework conditions for readability as suggested by Eduard.
- Updated reference to SMT formula.
- Rebased.

Harishankar Vishwanathan (1):
bpf: Introduce tnum_step to step through tnum's members

Paul Chaignon (3):
bpf: Improve bounds when tnum has a single possible value selftests/bpf: Test refinement of single-value tnum selftests/bpf: Avoid simplification of crafted bounds test

 include/linux/tnum.h                          |   3 +
 kernel/bpf/tnum.c                             |  56 +++++++
 kernel/bpf/verifier.c                         |  30 ++++
 .../selftests/bpf/prog_tests/reg_bounds.c     |   2 +-
 .../selftests/bpf/progs/verifier_bounds.c     | 137 ++++++++++++++++++
 5 files changed, 227 insertions(+), 1 deletion(-)
Signature · 1 line
2.43.0
#1 lore raw

PATCH v3 1/4 bpf: Introduce tnum_step to step through tnum's members

From: Harishankar Vishwanathan <[email protected]>

This commit introduces tnum_step(), a function that, when given t, and a number z returns the smallest member of t larger than z. The number z must be greater or equal to the smallest member of t and less than the largest member of t.

The first step is to compute j, a number that keeps all of t's known bits, and matches all unknown bits to z's bits. Since j is a member of the t, it is already a candidate for result. However, we want our result to be (minimally) greater than z.

There are only two possible cases:

(1) Case j <= z. In this case, we want to increase the value of j and make it > z.
(2) Case j > z. In this case, we want to decrease the value of j while keeping it > z.

(Case 1) j <= z

t = xx11x0x0
z = 10111101 (189)
j = 10111000 (184)

         ^
         k

(Case 1.1) Let's first consider the case where j < z. We will address j == z later.

Since z > j, there had to be a bit position that was 1 in z and a 0 in j, beyond which all positions of higher significance are equal in j and z. Further, this position could not have been unknown in a, because the unknown positions of a match z. This position had to be a 1 in z and known 0 in t.

Let k be position of the most significant 1-to-0 flip. In our example, k = 3 (starting the count at 1 at the least significant bit). Setting (to
1) the unknown bits of t in positions of significance smaller than k will not produce a result > z. Hence, we must set/unset the unknown bits at positions of significance higher than k. Specifically, we look for the next larger combination of 1s and 0s to place in those positions, relative to the combination that exists in z. We can achieve this by concatenating bits at unknown positions of t into an integer, adding 1, and writing the bits of that result back into the corresponding bit positions previously extracted from z.

Quoted 1 line

From our example, considering only positions of significance greater

than k:

t = xx..x
z = 10..1

    +    1
     -----
     11..0

This is the exact combination 1s and 0s we need at the unknown bits of t in positions of significance greater than k. Further, our result must only increase the value minimally above z. Hence, unknown bits in positions of significance smaller than k should remain 0. We finally have,

result = 11110000 (240)

(Case 1.2) Now consider the case when j = z, for example

t = 1x1x0xxx
z = 10110100 (180)
j = 10110100 (180)

Matching the unknown bits of the t to the bits of z yielded exactly z. To produce a number greater than z, we must set/unset the unknown bits in t, and *all* the unknown bits of t candidates for being set/unset. We can do this similar to Case 1.1, by adding 1 to the bits extracted from the masked bit positions of z. Essentially, this case is equivalent to Case 1.1, with k = 0.

t = 1x1x0xxx
z = .0.1.100

    +       1
    ---------
     .0.1.101

This is the exact combination of bits needed in the unknown positions of t. After recalling the known positions of t, we get

result = 10110101 (181)

(Case 2) j > z

t = x00010x1
z = 10000010 (130)
j = 10001011 (139)

	^
	k

Since j > z, there had to be a bit position which was 0 in z, and a 1 in j, beyond which all positions of higher significance are equal in j and z. This position had to be a 0 in z and known 1 in t. Let k be the position of the most significant 0-to-1 flip. In our example, k = 4.

Because of the 0-to-1 flip at position k, a member of t can become greater than z if the bits in positions greater than k are themselves >= to z. To make that member *minimally* greater than z, the bits in positions greater than k must be exactly = z. Hence, we simply match all of t's unknown bits in positions more significant than k to z's bits. In positions less significant than k, we set all t's unknown bits to 0 to retain minimality.

In our example, in positions of greater significance than k (=4), t=x000. These positions are matched with z (1000) to produce 1000. In positions of lower significance than k, t=10x1. All unknown bits are set to 0 to produce 1001. The final result is:

result = 10001001 (137)

This concludes the computation for a result > z that is a member of t.

The procedure for tnum_step() in this commit implements the idea described above. As a proof of correctness, we verified the algorithm against a logical specification of tnum_step. The specification asserts the following about the inputs t, z and output res that:

1. res is a member of t, and
2. res is strictly greater than z, and
3. there does not exist another value res2 such that 3a. res2 is also a member of t, and 3b. res2 is greater than z
3c. res2 is smaller than res

We checked the implementation against this logical specification using an SMT solver. The verification formula in SMTLIB format is available at [1]. The verification returned an "unsat": indicating that no input assignment exists for which the implementation and the specification produce different outputs.

In addition, we also automatically generated the logical encoding of the C implementation using Agni [2] and verified it against the same specification. This verification also returned an "unsat", confirming that the implementation is equivalent to the specification. The formula for this check is also available at [3].

---
 include/linux/tnum.h |  3 +++
 kernel/bpf/tnum.c    | 56 ++++++++++++++++++++++++++++++++++++++++++++
 2 files changed, 59 insertions(+)
include/linux/tnum.h +3 -0
index fa4654ffb621..ca2cfec8de08 100644
@@ -131,4 +131,7 @@ static inline bool tnum_subreg_is_const(struct tnum a)
 	return !(tnum_subreg(a)).mask;
 }
 
+/* Returns the smallest member of t larger than z */
+u64 tnum_step(struct tnum t, u64 z);
+
 #endif /* _LINUX_TNUM_H */
kernel/bpf/tnum.c +56 -0

59 lines not on this page · raw message

Signature · 1 line
2.43.0
#2 lore raw

PATCH v3 2/4 bpf: Improve bounds when tnum has a single possible value

In reply to Paul Chaignon

We're hitting an invariant violation in Cilium that sometimes leads to BPF programs being rejected and Cilium failing to start [1]. The following extract from verifier logs shows what's happening:

  from 201 to 236: R1=0 R6=ctx() R7=1 R9=scalar(smin=umin=smin32=umin32=3584,smax=umax=smax32=umax32=3840,var_off=(0xe00; 0x100)) R10=fp0
  236: R1=0 R6=ctx() R7=1 R9=scalar(smin=umin=smin32=umin32=3584,smax=umax=smax32=umax32=3840,var_off=(0xe00; 0x100)) R10=fp0
  ; if (magic == MARK_MAGIC_HOST || magic == MARK_MAGIC_OVERLAY || magic == MARK_MAGIC_ENCRYPT) @ bpf_host.c:1337
  236: (16) if w9 == 0xe00 goto pc+45   ; R9=scalar(smin=umin=smin32=umin32=3585,smax=umax=smax32=umax32=3840,var_off=(0xe00; 0x100))
  237: (16) if w9 == 0xf00 goto pc+1
  verifier bug: REG INVARIANTS VIOLATION (false_reg1): range bounds violation u64=[0xe01, 0xe00] s64=[0xe01, 0xe00] u32=[0xe01, 0xe00] s32=[0xe01, 0xe00] var_off=(0xe00, 0x0)

We reach instruction 236 with two possible values for R9, 0xe00 and 0xf00. This is perfectly reflected in the tnum, but of course the ranges are less accurate and cover [0xe00; 0xf00]. Taking the fallthrough path at instruction 236 allows the verifier to reduce the range to [0xe01; 0xf00]. The tnum is however not updated.

With these ranges, at instruction 237, the verifier is not able to deduce that R9 is always equal to 0xf00. Hence the fallthrough pass is explored first, the verifier refines the bounds using the assumption that R9 != 0xf00, and ends up with an invariant violation.

This pattern of impossible branch + bounds refinement is common to all invariant violations seen so far. The long-term solution is likely to rely on the refinement + invariant violation check to detect dead branches, as started by Eduard. To fix the current issue, we need something with less refactoring that we can backport.

This patch uses the tnum_step helper introduced in the previous patch to detect the above situation. In particular, three cases are now detected in the bounds refinement:

1. The u64 range and the tnum only overlap in umin.
u64: ---[xxxxxx]-----
tnum: --xx----------x-

2. The u64 range and the tnum only overlap in the maximum value represented by the tnum, called tmax.
u64: ---[xxxxxx]-----
tnum: xx-----x--------

3. The u64 range and the tnum only overlap in between umin (excluded) and umax.
u64: ---[xxxxxx]-----
tnum: xx----x-------x-

To detect these three cases, we call tnum_step(tnum, umin), which returns the smallest member of the tnum greater than umin, called tnum_next here. We're in case (1) if umin is part of the tnum and tnum_next is greater than umax. We're in case (2) if umin is not part of the tnum and tnum_next is equal to tmax. Finally, we're in case (3) if umin is not part of the tnum, tnum_next is inferior or equal to umax, and calling tnum_step a second time gives us a value past umax.

This change implements these three cases. With it, the above bytecode looks as follows:

  0: (85) call bpf_get_prandom_u32#7    ; R0=scalar()
  1: (47) r0 |= 3584                    ; R0=scalar(smin=0x8000000000000e00,umin=umin32=3584,smin32=0x80000e00,var_off=(0xe00; 0xfffffffffffff1ff))
  2: (57) r0 &= 3840                    ; R0=scalar(smin=umin=smin32=umin32=3584,smax=umax=smax32=umax32=3840,var_off=(0xe00; 0x100))
  3: (15) if r0 == 0xe00 goto pc+2      ; R0=3840

4: (15) if r0 == 0xf00 goto pc+1
4: R0=3840
6: (95) exit

In addition to the new selftests, this change was also verified with Agni [3]. For the record, the raw SMT is available at [4]. The property it verifies is that: If a concrete value x is contained in all input abstract values, after __update_reg_bounds, it will continue to be contained in all output abstract values.

---
 kernel/bpf/verifier.c | 30 ++++++++++++++++++++++++++++++
 1 file changed, 30 insertions(+)
kernel/bpf/verifier.c +30 -0
index bb12ba020649..401d6c4960ec 100644
@@ -2379,6 +2379,9 @@ static void __update_reg32_bounds(struct bpf_reg_state *reg)
 
 static void __update_reg64_bounds(struct bpf_reg_state *reg)
 {
+	u64 tnum_next, tmax;
+	bool umin_in_tnum;
+
 	/* min signed is max(sign bit) | min(other bits) */
 	reg->smin_value = max_t(s64, reg->smin_value,
 				reg->var_off.value | (reg->var_off.mask & S64_MIN));
@@ -2388,6 +2391,33 @@ static void __update_reg64_bounds(struct bpf_reg_state *reg)
 	reg->umin_value = max(reg->umin_value, reg->var_off.value);
 	reg->umax_value = min(reg->umax_value,
 			      reg->var_off.value | reg->var_off.mask);
+
+	/* Check if u64 and tnum overlap in a single value */
+	tnum_next = tnum_step(reg->var_off, reg->umin_value);
+	umin_in_tnum = (reg->umin_value & ~reg->var_off.mask) == reg->var_off.value;
+	tmax = reg->var_off.value | reg->var_off.mask;
+	if (umin_in_tnum && tnum_next > reg->umax_value) {
+		/* The u64 range and the tnum only overlap in umin.
+		 * u64:  ---[xxxxxx]-----
+		 * tnum: --xx----------x-
+		 */
+		___mark_reg_known(reg, reg->umin_value);
+	} else if (!umin_in_tnum && tnum_next == tmax) {
+		/* The u64 range and the tnum only overlap in the maximum value
+		 * represented by the tnum, called tmax.
+		 * u64:  ---[xxxxxx]-----
+		 * tnum: xx-----x--------
+		 */
+		___mark_reg_known(reg, tmax);
+	} else if (!umin_in_tnum && tnum_next <= reg->umax_value &&
+		   tnum_step(reg->var_off, tnum_next) > reg->umax_value) {
+		/* The u64 range and the tnum only overlap in between umin
+		 * (excluded) and umax.
+		 * u64:  ---[xxxxxx]-----
+		 * tnum: xx----x-------x-
+		 */
+		___mark_reg_known(reg, tnum_next);
+	}
 }
 
 static void __update_reg_bounds(struct bpf_reg_state *reg)
Signature · 1 line
2.43.0
#3 lore raw

PATCH v3 3/4 selftests/bpf: Test refinement of single-value tnum

In reply to Paul Chaignon

This patch introduces selftests to cover the new bounds refinement logic introduced in the previous patch. Without the previous patch, the first two tests fail because of the invariant violation they trigger. The last test fails because the R10 access is not detected as dead code. In addition, all three tests fail because of R0 having a non-constant value in the verifier logs.

In addition, the last two cases are covering the negative cases: when we shouldn't refine the bounds because the u64 and tnum overlap in at least two values.

---
 .../selftests/bpf/progs/verifier_bounds.c     | 137 ++++++++++++++++++
 1 file changed, 137 insertions(+)
tools/testing/selftests/bpf/progs/verifier_bounds.c +137 -0
index 560531404bce..97065a26cf70 100644
@@ -1863,4 +1863,141 @@ l1_%=:	r0 = 1;				\
 	: __clobber_all);
 }
 
+/* This test covers the bounds deduction when the u64 range and the tnum
+ * overlap only at umax. After instruction 3, the ranges look as follows:
+ *
+ * 0    umin=0xe01     umax=0xf00                              U64_MAX
+ * |    [xxxxxxxxxxxxxx]                                       |
+ * |----------------------------|------------------------------|
+ * |   x               x                                       | tnum values
+ *
+ * The verifier can therefore deduce that the R0=0xf0=240.
+ */
+SEC("socket")
+__description("bounds refinement with single-value tnum on umax")
+__msg("3: (15) if r0 == 0xe0 {{.*}} R0=240")
+__success __log_level(2)
+__flag(BPF_F_TEST_REG_INVARIANTS)
+__naked void bounds_refinement_tnum_umax(void *ctx)
+{
+	asm volatile("			\
+	call %[bpf_get_prandom_u32];	\
+	r0 |= 0xe0;			\
+	r0 &= 0xf0;			\
+	if r0 == 0xe0 goto +2;		\
+	if r0 == 0xf0 goto +1;		\
+	r10 = 0;			\
+	exit;				\
+"	:
+	: __imm(bpf_get_prandom_u32)
+	: __clobber_all);
+}
+
+/* This test covers the bounds deduction when the u64 range and the tnum
+ * overlap only at umin. After instruction 3, the ranges look as follows:
+ *
+ * 0    umin=0xe00     umax=0xeff                              U64_MAX
+ * |    [xxxxxxxxxxxxxx]                                       |
+ * |----------------------------|------------------------------|
+ * |    x               x                                      | tnum values
+ *
+ * The verifier can therefore deduce that the R0=0xe0=224.
+ */
+SEC("socket")
+__description("bounds refinement with single-value tnum on umin")
+__msg("3: (15) if r0 == 0xf0 {{.*}} R0=224")
+__success __log_level(2)
+__flag(BPF_F_TEST_REG_INVARIANTS)
+__naked void bounds_refinement_tnum_umin(void *ctx)
+{
+	asm volatile("			\
+	call %[bpf_get_prandom_u32];	\
+	r0 |= 0xe0;			\
+	r0 &= 0xf0;			\
+	if r0 == 0xf0 goto +2;		\
+	if r0 == 0xe0 goto +1;		\
+	r10 = 0;			\
+	exit;				\
+"	:
+	: __imm(bpf_get_prandom_u32)
+	: __clobber_all);
+}
+
+/* This test covers the bounds deduction when the only possible tnum value is
+ * in the middle of the u64 range. After instruction 3, the ranges look as
+ * follows:
+ *
+ * 0    umin=0x7cf   umax=0x7df                                U64_MAX
+ * |    [xxxxxxxxxxxx]                                         |
+ * |----------------------------|------------------------------|
+ * |     x            x            x            x            x | tnum values
+ *       |            +--- 0x7e0
+ *       +--- 0x7d0
+ *
+ * Since the lower four bits are zero, the tnum and the u64 range only overlap
+ * in R0=0x7d0=2000. Instruction 5 is therefore dead code.
+ */
+SEC("socket")
+__description("bounds refinement with single-value tnum in middle of range")
+__msg("3: (a5) if r0 < 0x7cf {{.*}} R0=2000")
+__success __log_level(2)
+__naked void bounds_refinement_tnum_middle(void *ctx)
+{
+	asm volatile("			\
+	call %[bpf_get_prandom_u32];	\
+	if r0 & 0x0f goto +4;		\
+	if r0 > 0x7df goto +3;		\
+	if r0 < 0x7cf goto +2;		\
+	if r0 == 0x7d0 goto +1;		\
+	r10 = 0;			\
+	exit;				\
+"	:
+	: __imm(bpf_get_prandom_u32)
+	: __clobber_all);
+}
+
+/* This test cover the negative case for the tnum/u64 overlap. Since
+ * they contain the same two values (i.e., {0, 1}), we can't deduce
+ * anything more.
+ */
+SEC("socket")
+__description("bounds refinement: several overlaps between tnum and u64")
+__msg("2: (25) if r0 > 0x1 {{.*}} R0=scalar(smin=smin32=0,smax=umax=smax32=umax32=1,var_off=(0x0; 0x1))")
+__failure __log_level(2)
+__naked void bounds_refinement_several_overlaps(void *ctx)
+{
+	asm volatile("			\
+	call %[bpf_get_prandom_u32];	\
+	if r0 < 0 goto +3;		\
+	if r0 > 1 goto +2;		\
+	if r0 == 1 goto +1;		\
+	r10 = 0;			\
+	exit;				\
+"	:
+	: __imm(bpf_get_prandom_u32)
+	: __clobber_all);
+}
+
+/* This test cover the negative case for the tnum/u64 overlap. Since
+ * they overlap in the two values contained by the u64 range (i.e.,
+ * {0xf, 0x10}), we can't deduce anything more.
+ */
+SEC("socket")
+__description("bounds refinement: multiple overlaps between tnum and u64")
+__msg("2: (25) if r0 > 0x10 {{.*}} R0=scalar(smin=umin=smin32=umin32=15,smax=umax=smax32=umax32=16,var_off=(0x0; 0x1f))")
+__failure __log_level(2)
+__naked void bounds_refinement_multiple_overlaps(void *ctx)
+{
+	asm volatile("			\
+	call %[bpf_get_prandom_u32];	\
+	if r0 < 0xf goto +3;		\
+	if r0 > 0x10 goto +2;		\
+	if r0 == 0x10 goto +1;		\
+	r10 = 0;			\
+	exit;				\
+"	:
+	: __imm(bpf_get_prandom_u32)
+	: __clobber_all);
+}
+
 char _license[] SEC("license") = "GPL";
Signature · 1 line
2.43.0
#4 lore raw

PATCH v3 4/4 selftests/bpf: Avoid simplification of crafted bounds test

In reply to Paul Chaignon

The reg_bounds_crafted tests validate the verifier's range analysis logic. They focus on the actual ranges and thus ignore the tnum. As a consequence, they carry the assumption that the tested cases can be reproduced in userspace without using the tnum information.

Unfortunately, the previous change the refinement logic breaks that assumption for one test case:

(u64)2147483648 (u32)<op> [4294967294; 0x100000000]

The tested bytecode is shown below. Without our previous improvement, on the false branch of the condition, R7 is only known to have u64 range [0xfffffffe; 0x100000000]. With our improvement, and using the tnum information, we can deduce that R7 equals 0x100000000.

  19: (bc) w0 = w6                ; R6=0x80000000
  20: (bc) w0 = w7                ; R7=scalar(smin=umin=0xfffffffe,smax=umax=0x100000000,smin32=-2,smax32=0,var_off=(0x0; 0x1ffffffff))
  21: (be) if w6 <= w7 goto pc+3  ; R6=0x80000000 R7=0x100000000

R7's tnum is (0; 0x1ffffffff). On the false branch, regs_refine_cond_op refines R7's u32 range to [0; 0x7fffffff]. Then, __reg32_deduce_bounds refines the s32 range to 0 using u32 and finally also sets u32=0. From this, __reg_bound_offset improves the tnum to (0; 0x100000000). Finally, our previous patch uses this new tnum to deduce that it only intersect with u64=[0xfffffffe; 0x100000000] in a single value:
0x100000000.

Because the verifier uses the tnum to reach this constant value, the selftest is unable to reproduce it by only simulating ranges. The solution implemented in this patch is to change the test case such that there is more than one overlap value between u64 and the tnum. The max. u64 value is thus changed from 0x100000000 to 0x300000000.

---
 tools/testing/selftests/bpf/prog_tests/reg_bounds.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
tools/testing/selftests/bpf/prog_tests/reg_bounds.c +1 -1
index d93a0c7b1786..0322f817d07b 100644
@@ -2091,7 +2091,7 @@ static struct subtest_case crafted_cases[] = {
 	{U64, S64, {0, 0xffffffffULL}, {0x7fffffff, 0x7fffffff}},
 
 	{U64, U32, {0, 0x100000000}, {0, 0}},
-	{U64, U32, {0xfffffffe, 0x100000000}, {0x80000000, 0x80000000}},
+	{U64, U32, {0xfffffffe, 0x300000000}, {0x80000000, 0x80000000}},
 
 	{U64, S32, {0, 0xffffffff00000000ULL}, {0, 0}},
 	/* these are tricky cases where lower 32 bits allow to tighten 64
Signature · 1 line
2.43.0
#5 lore raw
In reply to Paul Chaignon
This series was applied to bpf/bpf.git (master) by Alexei Starovoitov <[email protected]>:

Hello:

This series was applied to bpf/bpf.git (master)
by Alexei Starovoitov <[email protected]>:

Quoted 6 lines · On Fri, 27 Feb 2026 22:30:53 +0100 you wrote:

We're hitting an invariant violation in Cilium that sometimes leads to BPF programs being rejected and Cilium failing to start [1]. As far as I know this is the first case of invariant violation found in a real program (i.e., not by a fuzzer). The following extract from verifier logs shows what's happening:

Quoted 8 lines
  from 201 to 236: R1=0 R6=ctx() R7=1 R9=scalar(smin=umin=smin32=umin32=3584,smax=umax=smax32=umax32=3840,var_off=(0xe00; 0x100)) R10=fp0
  236: R1=0 R6=ctx() R7=1 R9=scalar(smin=umin=smin32=umin32=3584,smax=umax=smax32=umax32=3840,var_off=(0xe00; 0x100)) R10=fp0
  ; if (magic == MARK_MAGIC_HOST || magic == MARK_MAGIC_OVERLAY || magic == MARK_MAGIC_ENCRYPT) @ bpf_host.c:1337
  236: (16) if w9 == 0xe00 goto pc+45   ; R9=scalar(smin=umin=smin32=umin32=3585,smax=umax=smax32=umax32=3840,var_off=(0xe00; 0x100))
  237: (16) if w9 == 0xf00 goto pc+1
  verifier bug: REG INVARIANTS VIOLATION (false_reg1): range bounds violation u64=[0xe01, 0xe00] s64=[0xe01, 0xe00] u32=[0xe01, 0xe00] s32=[0xe01, 0xe00] var_off=(0xe00, 0x0)

[...]

Here is the summary with links:
- [v3,bpf,1/4] bpf: Introduce tnum_step to step through tnum's members

    https://git.kernel.org/bpf/bpf/c/76e954155b45
  - [v3,bpf,2/4] bpf: Improve bounds when tnum has a single possible value
    https://git.kernel.org/bpf/bpf/c/efc11a667878
  - [v3,bpf,3/4] selftests/bpf: Test refinement of single-value tnum
    https://git.kernel.org/bpf/bpf/c/e6ad477d1bf8
  - [v3,bpf,4/4] selftests/bpf: Avoid simplification of crafted bounds test
    https://git.kernel.org/bpf/bpf/c/024cea2d647e

You are awesome, thank you!

Signature · 2 lines
Deet-doot-dot, I am a bot.
https://korg.docs.kernel.org/patchwork/pwbot.html
maintainer
#6 lore raw
In reply to Paul Chaignon
Quoted 23 lines · Paul Chaignon

We're hitting an invariant violation in Cilium that sometimes leads to BPF programs being rejected and Cilium failing to start [1]. As far as I know this is the first case of invariant violation found in a real program (i.e., not by a fuzzer). The following extract from verifier logs shows what's happening:

  from 201 to 236: R1=0 R6=ctx() R7=1 R9=scalar(smin=umin=smin32=umin32=3584,smax=umax=smax32=umax32=3840,var_off=(0xe00; 0x100)) R10=fp0
  236: R1=0 R6=ctx() R7=1 R9=scalar(smin=umin=smin32=umin32=3584,smax=umax=smax32=umax32=3840,var_off=(0xe00; 0x100)) R10=fp0
  ; if (magic == MARK_MAGIC_HOST || magic == MARK_MAGIC_OVERLAY || magic == MARK_MAGIC_ENCRYPT) @ bpf_host.c:1337
  236: (16) if w9 == 0xe00 goto pc+45   ; R9=scalar(smin=umin=smin32=umin32=3585,smax=umax=smax32=umax32=3840,var_off=(0xe00; 0x100))
  237: (16) if w9 == 0xf00 goto pc+1
  verifier bug: REG INVARIANTS VIOLATION (false_reg1): range bounds violation u64=[0xe01, 0xe00] s64=[0xe01, 0xe00] u32=[0xe01, 0xe00] s32=[0xe01, 0xe00] var_off=(0xe00, 0x0)

More details are given in the second patch, but in short, the verifier should be able to detect that the false branch of instruction 237 is never true. After instruction 236, the u64 range and the tnum overlap in a single value, 0xf00.

The long-term solution to invariant violation is likely to rely on the refinement + invariant violation check to detect dead branches, as started by Eduard. To fix the current issue, we need something with less refactoring that we can backport to affected kernels.

Quoted 4 lines

The solution implemented in the second patch is to improve the bounds refinement to avoid this case. It relies on a new tnum helper, tnum_step, first sent as an RFC in [2]. The last two patches extend and update the selftests.

[...]

Hi Paul, Harishankar,

Apologies for resurrecting an old thread.
While reviewing the recent series "Fix invariant violations and improve branch detection". I notice that we have a single test case broken, with bisect pointing to merge of this branch. Details are below.

There is a slow mode for reg_bounds* tests,
it can be triggered like this:

env SLOW_TESTS=1 ./test_progs -t reg_bounds

In this mods reg_bounds* enumerates some cases not processed during the regular CI run. One of these cases fails with the following log:

...
19: (bc) w0 = w6 ; R0=scalar(smin=0,smax=umax=0xffffffff,var_off=(0x0; 0xffffffff))

                                          R6=scalar(smin=0xfffffffe00000001,smax=0xffffffff00000000,
					            umin=0xfffffffe00000001,umax=0xffffffff00000000,
						    var_off=(0xfffffffe00000000; 0x1ffffffff))

20: (bc) w0 = w7 ; R0=0 R7=0x8000000000000000
21: (1e) if w6 == w7 goto pc+3 ; ...
...

from 21 to 25: ...
25: (bc) w0 = w6 ; R0=0 R6=0xffffffff00000000 <---- unexpected value here
26: (bc) w0 = w7 ; R0=0 R7=0x8000000000000000
27: (95) exit
...

=====================
...
ACTUAL TRUE1: scalar(u64=0xffffffff00000000,u32=0,s64=0xffffffff00000000,s32=0)
EXPECTED TRUE1: scalar(u64=[0xfffffffe00000001; 0xffffffff00000000],u32=0,s64=[0xfffffffe00000001; 0xffffffff00000000],s32=0)
...
#319/1007 reg_bounds_gen_consts_s64_s32/(s64)[0xfffffffe00000001; 0xffffffff00000000] (s32)<op> S64_MIN:FAIL

The test case can be triggered manually as follows:

env SLOW_TESTS=1 ./test_progs -a 'reg_bounds_gen_consts_s64_s32/(s64)[0xfffffffe00000001; 0xffffffff00000000] (s32)<op> S64_MIN'

The verifier produces a more precise result here because of additional information inferred by tnum_step: lower 32-bits can be equal to zero only for r7 value of 0xffffffff00000000.

Which implies that reg_bounds* test logic needs further adjustment. Could you please take a look? Please let me know if you see any viable routes to amend this test.

Thanks,
Eduard

#7 lore raw
Quoted 1 line · Eduard Zingerman

On Fri, 2026-02-27 at 22:30 +0100, Paul Chaignon wrote:

[...]

Quoted 7 lines

The verifier produces a more precise result here because of additional information inferred by tnum_step: lower 32-bits can be equal to zero only for r7 value of 0xffffffff00000000.

Which implies that reg_bounds* test logic needs further adjustment. Could you please take a look? Please let me know if you see any viable routes to amend this test.

Hi Eduard,

Thanks for the report! I wasn't aware of this slow mode. I'll take a look when back from vacation, on Monday.

Paul

Quoted 3 lines

Thanks,
Eduard