Fix invariant violation for single-value tnums
Subsystems (4)
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.
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 | 52 +++++++++++ kernel/bpf/verifier.c | 16 ++++ .../selftests/bpf/prog_tests/reg_bounds.c | 2 +- .../selftests/bpf/progs/verifier_bounds.c | 91 +++++++++++++++++++ 5 files changed, 163 insertions(+), 1 deletion(-)
Signature · 1 line
2.43.0
PATCH 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..0This 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
---------
.0.1.101This 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 | 52 ++++++++++++++++++++++++++++++++++++++++++++ 2 files changed, 55 insertions(+)
include/linux/tnum.h +3 -0
@@ -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 +52 -0
55 lines not on this page · raw message
Signature · 1 line
2.43.0
[email protected]
AI reviewRe: [PATCH bpf 1/4] bpf: Introduce tnum_step to step through tnum's members
When t.value has bit 63 set and z does not (reachable when z < tmin, which the function comment says is handled), fls64(p) returns 64, making this U64_MAX << 64.
Quoted 5 lines
kernel/bpf/tnum.c +0
-0@@ -269,3 +269,55 @@ struct tnum tnum_bswap64(struct tnum a)
[ ... ]
Quoted 8 lines
kernel/bpf/tnum.c +8
-0+u64 tnum_step (struct tnum t,u64 z) +{ +u64 tmax, j, p, q, r, s, v, u, w, res; +u8 k; + + tmax = t.value | t.mask; + +/* if z >= largest member of t, return largest member of t */
kernel/bpf/tnum.c struct tnum tnum_bswap64(struct tnum a) · line 304
Quoted 10 lines
kernel/bpf/tnum.c +10
-0+if (z >= tmax) +return tmax; + +/* keep t's known bits, and match all unknown bits to z */ + j = t.value | (z & t.mask); + +if (j > z) { + p = ~z & t.value & ~t.mask; + k =fls64 (p);/* k is the most-significant 0-to-1 flip */ + q =U64_MAX << k;
When t.value has bit 63 set and z does not (reachable when z < tmin, which the function comment says is handled), fls64(p) returns 64, making this U64_MAX << 64.
Shifting a 64-bit value by 64 is undefined behavior in C. tnum_range() in the same file already guards against this:
kernel/bpf/tnum.c:tnum_range() {
...
/* special case, needed because 1ULL << 64 is undefined */
if (bits > 63)
return tnum_unknown;
...
}Should tnum_step() have a similar guard when k == 64?
Quoted 5 lines
kernel/bpf/tnum.c +5
-0+ r = q & z;/* positions > k matched to z */ + s = ~q & t.value;/* positions <= k matched to t.value */ + v = r | s; + res = v; + }else {
[ ... ]
kernel/bpf/tnum.c struct tnum tnum_bswap64(struct tnum a) · line 323
Quoted 3 lines
kernel/bpf/tnum.c +3
-0+ } +return res; +}
---
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/22092666641
AI-authorship-score: low
AI-authorship-explanation: Detailed hand-crafted mathematical proofs with academic co-authors from Rutgers University and references to SMT solver verification are consistent with human academic research.
issues-found: 1
issue-severity-score: medium
issue-severity-explanation: Undefined behavior from shifting a 64-bit value by 64 in the j > z branch; currently unreachable from the only caller in the series but reachable through the documented API contract.
Re: [PATCH bpf 1/4] bpf: Introduce tnum_step to step through tnum's members
Quoted 23 lines · On Tue, Feb 17, 2026 at 4:44 AM <[email protected]> wrote:
Quoted 6 lines
kernel/bpf/tnum.c +0
-0@@ -269,3 +269,55 @@ struct tnum tnum_bswap64(struct tnum a)[ ... ]
Quoted 14 lines
kernel/bpf/tnum.c +14
-0+u64 tnum_step (struct tnum t,u64 z) +{ +u64 tmax, j, p, q, r, s, v, u, w, res; +u8 k; + + tmax = t.value | t.mask; + +/* if z >= largest member of t, return largest member of t */ +if (z >= tmax) +return tmax; + +/* keep t's known bits, and match all unknown bits to z */ + j = t.value | (z & t.mask); +
Quoted 8 lines
kernel/bpf/tnum.c struct tnum tnum_bswap64(struct tnum a) · line 304
Quoted 4 lines
kernel/bpf/tnum.c +4
-0+if (j > z) { + p = ~z & t.value & ~t.mask; + k =fls64 (p);/* k is the most-significant 0-to-1 flip */ + q =U64_MAX << k;When t.value has bit 63 set and z does not (reachable when z < tmin, which the function comment says is handled), fls64(p) returns 64, making this
U64_MAX<< 64.
It is possible that the shift value k can be = 64. This happens when
(1) j > z (i.e.,
this cannot happen in the else branch when j <= z), (2) the 63rd bit position of z is 0,
(3) the 63rd position of t is a known 1 (t.value = 0).
If I recall correctly, the existing code assumed that the shift by 64 would return a 0,
and that consequently, the remaining computation would set res to tmin (Re the comment:
"if this function is called with z < tmin, the algorithm already returns tmin").
Quoted 11 lines
Shifting a 64-bit value by 64 is undefined behavior in C.
tnum_range()in the same file already guards against this:kernel/bpf/tnum.c:tnum_range() { ... /* special case, needed because 1ULL << 64 is undefined */ if (bits > 63) return tnum_unknown; ... }Should
tnum_step()have a similar guard when k == 64?
A cleaner way to address this is to make the idea of the comment explicit, with an early return.
/* if z < smallest member of t, return smallest member of t */ if (z < tmin)
return tmin;
Without z < tmin, it is not possible to satisfy necessary conditions
(2) and (3) above.
[ ... ]
As an aside, this didn't show up during SMT verification because the bvshl (bitvector shift left)
operation is defined such that a shift value greater than the width returns exactly 0.
Quoted 11 lines
---
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.mdCI run summary: https://github.com/kernel-patches/bpf/actions/runs/22092666641
AI-authorship-score: low
AI-authorship-explanation: Detailed hand-crafted mathematical proofs with academic co-authors from Rutgers University and references to SMT solver verification are consistent with human academic research.
issues-found: 1
issue-severity-score: medium
issue-severity-explanation: Undefined behavior from shifting a 64-bit value by 64 in the j > z branch; currently unreachable from the only caller in the series but reachable through the documented API contract.
Re: [PATCH bpf 1/4] bpf: Introduce tnum_step to step through tnum's members
Quoted 6 lines · Harishankar Vishwanathan
As an aside, this didn't show up during SMT verification because the bvshl (bitvector shift left)
operation is defined such that a shift value greater than the width returns exactly 0.
Something that you should keep in mind when relying so much on SMT. Shift by 64 on x86 is a nop while on arm64 it results in zero.
Re: [PATCH bpf 1/4] bpf: Introduce tnum_step to step through tnum's members
[ ... ]
Quoted 2 lines
Something that you should keep in mind when relying so much on SMT. Shift by 64 on x86 is a nop while on arm64 it results in zero.
That is a fair point; a good reminder of the semantic gap between SMT abstractions
and real hardware.
I missed a standard constraint here: explicitly model the undefined behavior by adding
an assertion that the shift amount is strictly < width. That would have forced the solver
to flag any execution with shift amount >= 64 as a verification failure.
Having said that, I will say that SMT was still instrumental in developing this; it did help
cover cases that would have been difficult to hit with testing or fuzzing.
PATCH 2/4 bpf: Improve bounds when tnum has a single possible value
In reply to Paul ChaignonWe'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
u64: ---[xxxxxx]-----
tnum: xx----x-------x-
To detect these three cases, we call tnum_step(tnum, umin), which returns the largest member of the tnum greater than umin, called tnum_next here. We're in case (1) if umin uis 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 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 | 16 ++++++++++++++++ 1 file changed, 16 insertions(+)
kernel/bpf/verifier.c +16 -0
@@ -2379,6 +2379,8 @@ static void __update_reg32_bounds(struct bpf_reg_state *reg)static void __update_reg64_bounds (struct bpf_reg_state *reg) { +u64 tnum_next; +/* 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 +2390,20 @@ 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); +if ((reg->umin_value & ~reg->var_off.mask) == reg->var_off.value) { +/* The only overlap is umin */ +if (tnum_next > reg->umax_value) +___mark_reg_known (reg, reg->umin_value); + }else if (tnum_next == (reg->var_off.value | reg->var_off.mask)) { +/* The only overlap is tmax */ +___mark_reg_known (reg, tnum_next); + }else if (tnum_next <= reg->umax_value && +tnum_step (reg->var_off, tnum_next) > reg->umax_value) { +___mark_reg_known (reg, tnum_next); + } }static void __update_reg_bounds (struct bpf_reg_state *reg)
Signature · 1 line
2.43.0
Re: [PATCH bpf 2/4] bpf: Improve bounds when tnum has a single possible value
Quoted 104 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]. 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,
0xe00and0xf00. 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_stephelper 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
u64: ---[xxxxxx]-----
tnum: xx----x-------x-To detect these three cases, we call
tnum_step(tnum, umin), which returns the largest member of the tnum greater than umin, calledtnum_nexthere. We're in case (1) if umin uis part of the tnum andtnum_nextis greater than umax. We're in case (2) if umin is not part of the tnum andtnum_nextis equal to tmax. Finally, we're in case (3) iftnum_nextis inferior or equal to umax AND callingtnum_stepa 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=38404: (15) if r0 ==
0xf00goto pc+1
4: R0=3840
6: (95) exitIn 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.Link https://github.com/bpfverif/agni [3]Link https://pastebin.com/raw/9BJAi6Rw [4]Co-developed-by Harishankar Vishwanathan <[email protected]>Signed-off-by Harishankar Vishwanathan <[email protected]>Signed-off-by Paul Chaignon <[email protected]>--- kernel/bpf/verifier.c | 16 ++++++++++++++++ 1 file changed, 16 insertions(+)kernel/bpf/verifier.c +7
-0@@ -2379,6 +2379,8 @@ static void __update_reg32_bounds(struct bpf_reg_state *reg)static void __update_reg64_bounds (struct bpf_reg_state *reg) { +u64 tnum_next; +/* 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 +2390,20 @@ 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); +if ((reg->umin_value & ~reg->var_off.mask) == reg->var_off.value) { +/* The only overlap is umin */
kernel/bpf/verifier.c static void __update_reg64_bounds(struct bpf_reg_state *reg) · line 2407
Quoted 10 lines
kernel/bpf/verifier.c +9
-0+if (tnum_next > reg->umax_value) +___mark_reg_known (reg, reg->umin_value); + }else if (tnum_next == (reg->var_off.value | reg->var_off.mask)) { +/* The only overlap is tmax */ +___mark_reg_known (reg, tnum_next); + }else if (tnum_next <= reg->umax_value && +tnum_step (reg->var_off, tnum_next) > reg->umax_value) { +___mark_reg_known (reg, tnum_next); + } }
Looks like it's causing a big regression in one of sched_ext progs:
scx_cosmos-5c590c8858df35e1_out_bpf.bpf.o cosmos_select_cpu success +542.86 %
Please investigate. It needs to be understood.
Eduard Zingerman
maintainerRe: [PATCH bpf 2/4] bpf: Improve bounds when tnum has a single possible value
In reply to Paul ChaignonOn Tue, 2026-02-17 at 10:01 +0100, Paul Chaignon wrote:
[...]
Quoted 8 lines
This patch uses the
tnum_stephelper 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-
Quoted 8 lines
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
u64: ---[xxxxxx]-----
tnum: xx----x-------x-
Nit: please put these diagrams in the source code.
[...]
Quoted 22 lines
kernel/bpf/verifier.c +7
-0@@ -2379,6 +2379,8 @@ static void __update_reg32_bounds(struct bpf_reg_state *reg)static void __update_reg64_bounds (struct bpf_reg_state *reg) { +u64 tnum_next; +/* 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 +2390,20 @@ 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); +if ((reg->umin_value & ~reg->var_off.mask) == reg->var_off.value) { +/* The only overlap is umin */
kernel/bpf/verifier.c static void __update_reg64_bounds(struct bpf_reg_state *reg) · line 2407
Quoted 10 lines
kernel/bpf/verifier.c +9
-0+if (tnum_next > reg->umax_value) +___mark_reg_known (reg, reg->umin_value); + }else if (tnum_next == (reg->var_off.value | reg->var_off.mask)) { +/* The only overlap is tmax */ +___mark_reg_known (reg, tnum_next); + }else if (tnum_next <= reg->umax_value && +tnum_step (reg->var_off, tnum_next) > reg->umax_value) { +___mark_reg_known (reg, tnum_next); + } }
I think this hunk should be rewritten as follows:
tnum_next = tnum_step(reg->var_off, reg->umin_value);
tnum_max = reg->var_off.value | reg->var_off.mask;
tnum_min = reg->var_off.value;
if (tnum_next > reg->umax_value) {
/* The only overlap is umin */
___mark_reg_known(reg, tnum_min);
} else if (tnum_min < reg->umin_value && tnum_next == tnum_max) {
/* The only overlap is tmax */
___mark_reg_known(reg, tnum_next);
} else if (tnum_next <= reg->umax_value &&
tnum_step(reg->var_off, tnum_next) > reg->umax_value) {
___mark_reg_known(reg, tnum_next);
}- At-least to me, it easier to understand this way.
- There is no need to gate the condition tnum_next > reg->umax_value, if next tnum overshoots the reg->umax_value then only tnum_min is left.
- Accidentally, it fixes the scx_cosmos regression
(probably, because first condition is relaxed).
Wdyt?
Re: [PATCH bpf 2/4] bpf: Improve bounds when tnum has a single possible value
On Tue, Feb 17, 2026 at 5:58 PM Eduard Zingerman <[email protected]> wrote:
[...]
Quoted 19 lines
Quoted 3 lines
1. The u64 range and the tnum only overlap in umin.
u64: ---[xxxxxx]-----
tnum: --xx----------x-I think this hunk should be rewritten as follows:
tnum_next = tnum_step(reg->var_off, reg->umin_value); tnum_max = reg->var_off.value | reg->var_off.mask; tnum_min = reg->var_off.value; if (tnum_next > reg->umax_value) { /* The only overlap is umin */ ___mark_reg_known(reg, tnum_min); } else if (tnum_min < reg->umin_value && tnum_next == tnum_max) { /* The only overlap is tmax */ ___mark_reg_known(reg, tnum_next); } else if (tnum_next <= reg->umax_value && tnum_step(reg->var_off, tnum_next) > reg->umax_value) { ___mark_reg_known(reg, tnum_next); }
Quoted 1 line
- At-least to me, it easier to understand this way.
We can also use tmin and tmax, like tnum_step from the previous commit does, to keep things consistent.
Quoted 2 lines
- There is no need to gate the condition
tnum_next > reg->umax_value, if next tnum overshoots thereg->umax_valuethen onlytnum_minis left.
The guard helps avoid the case where u64 and tnum have no intersection:
u64: ---[xxxxxx]----- tnum: x------------x-
The entire condition for "only overlap of tnum is with umin" needs to check both
(1) if umin is a member of var_off
(2) if the next member of var_off after umin > umax.
Quoted 4 lines
- Accidentally, it fixes the
scx_cosmosregression (probably, because first condition is relaxed).Wdyt?
Eduard Zingerman
maintainerRe: [PATCH bpf 2/4] bpf: Improve bounds when tnum has a single possible value
Quoted 21 lines · Harishankar Vishwanathan
On Tue, Feb 17, 2026 at 5:58 PM Eduard Zingerman <[email protected]> wrote:
[...]Quoted 19 lines
Quoted 3 lines
1. The u64 range and the tnum only overlap in umin.
u64: ---[xxxxxx]-----
tnum: --xx----------x-I think this hunk should be rewritten as follows:
tnum_next = tnum_step(reg->var_off, reg->umin_value); tnum_max = reg->var_off.value | reg->var_off.mask; tnum_min = reg->var_off.value; if (tnum_next > reg->umax_value) { /* The only overlap is umin */ ___mark_reg_known(reg, tnum_min); } else if (tnum_min < reg->umin_value && tnum_next == tnum_max) { /* The only overlap is tmax */ ___mark_reg_known(reg, tnum_next); } else if (tnum_next <= reg->umax_value && tnum_step(reg->var_off, tnum_next) > reg->umax_value) { ___mark_reg_known(reg, tnum_next); }
Quoted 8 lines
Quoted 1 line
- At-least to me, it easier to understand this way.
We can also use tmin and tmax, like
tnum_stepfrom the previous commit does, to keep things consistent.Quoted 2 lines
- There is no need to gate the condition
tnum_next > reg->umax_value, if next tnum overshoots thereg->umax_valuethen onlytnum_minis left.
Hi Harishankar,
Sorry for delayed response.
Quoted 7 lines
The guard helps avoid the case where u64 and tnum have no intersection:
u64: ---[xxxxxx]----- tnum: x------------x-The entire condition for "only overlap of tnum is with umin" needs to check both
(1) if umin is a member ofvar_off
(2) if the next member ofvar_offafter umin > umax.
Right before the hunk there is:
reg->umin_value = max(reg->umin_value, reg->var_off.value);
Which implies that:
tnum_min <= reg->umin_value
My mental model for register sync functions is that all ranges are considered valid, hence I suggested dropping the check.
If you think the check should be preserved, maybe still rewrite it like:
tnum_min == reg->umin_value && tnum_next > reg->umax_value
That's a bit easier to parse compared to '(reg->umin_value & ~reg->var_off.mask) == reg->var_off.value'. Wdyt?
Quoted 3 lines
Quoted 3 lines
- Accidentally, it fixes the
scx_cosmosregression (probably, because first condition is relaxed).
I double-checked scx_cosmos with current master, patch as posted and patch with my changes, in all three cases I see:
File Program Verdict Insns ---------------- ------------------ ------- ----- scx_cosmos.bpf.o cosmos_select_cpu success 475
So, I'm not sure why CI flagged this as +500%.
Re: [PATCH bpf 2/4] bpf: Improve bounds when tnum has a single possible value
Quoted 22 lines · Eduard Zingerman
Quoted 21 lines · Harishankar Vishwanathan
On Tue, Feb 17, 2026 at 5:58 PM Eduard Zingerman <[email protected]> wrote:
[...]Quoted 19 lines
Quoted 3 lines
1. The u64 range and the tnum only overlap in umin.
u64: ---[xxxxxx]-----
tnum: --xx----------x-I think this hunk should be rewritten as follows:
tnum_next = tnum_step(reg->var_off, reg->umin_value); tnum_max = reg->var_off.value | reg->var_off.mask; tnum_min = reg->var_off.value; if (tnum_next > reg->umax_value) { /* The only overlap is umin */ ___mark_reg_known(reg, tnum_min); } else if (tnum_min < reg->umin_value && tnum_next == tnum_max) { /* The only overlap is tmax */ ___mark_reg_known(reg, tnum_next); } else if (tnum_next <= reg->umax_value && tnum_step(reg->var_off, tnum_next) > reg->umax_value) { ___mark_reg_known(reg, tnum_next); }
Quoted 1 line
Quoted 1 line
Quoted 1 line
- At-least to me, it easier to understand this way.
Agree. That looks easier to parse to me too. Thanks!
Quoted 22 lines
Quoted 7 lines
We can also use tmin and tmax, like
tnum_stepfrom the previous commit does, to keep things consistent.Quoted 2 lines
- There is no need to gate the condition
tnum_next > reg->umax_value, if next tnum overshoots thereg->umax_valuethen onlytnum_minis left.Hi Harishankar,
Sorry for delayed response.
Quoted 7 lines
The guard helps avoid the case where u64 and tnum have no intersection:
u64: ---[xxxxxx]----- tnum: x------------x-The entire condition for "only overlap of tnum is with umin" needs to check both
(1) if umin is a member ofvar_off
(2) if the next member ofvar_offafter umin > umax.Right before the hunk there is:
Quoted 8 lines
reg->umin_value = max(reg->umin_value, reg->var_off.value);Which implies that:
tnum_min<=reg->umin_valueMy mental model for register sync functions is that all ranges are considered valid, hence I suggested dropping the check.
I think I agree with that, especially considering the reg->umin_value computation above. If we think we need extra checks to be on the safe side, they should probably live in reg_bounds_sanity_check.
Quoted 12 lines
If you think the check should be preserved, maybe still rewrite it like:
tnum_min == reg->umin_value && tnum_next > reg->umax_valueThat's a bit easier to parse compared to '(
reg->umin_value& ~reg->var_off.mask) ==reg->var_off.value'. Wdyt?Quoted 3 lines
Quoted 3 lines
- Accidentally, it fixes the
scx_cosmosregression (probably, because first condition is relaxed).
Quoted 8 lines
I double-checked
scx_cosmoswith current master, patch as posted and patch with my changes, in all three cases I see:File Program Verdict Insns ---------------- ------------------ ------- ----- scx_cosmos.bpf.o cosmos_select_cpu success 475So, I'm not sure why CI flagged this as +500%.
Ok, that's reassuring. I was going crazy trying to understand how your changes could have fixed this given the semantics should be unchanged.
I'm re-testing the patchset on Cilium just to be sure, and I'll send a v2 afterward. Thanks for the review!
Re: [PATCH bpf 2/4] bpf: Improve bounds when tnum has a single possible value
Quoted 9 lines · Paul Chaignon
Quoted 8 lines · Eduard Zingerman
Quoted 7 lines · Harishankar Vishwanathan
On Tue, Feb 17, 2026 at 5:58 PM Eduard Zingerman <[email protected]> wrote:
[...]Quoted 5 lines
Quoted 3 lines
1. The u64 range and the tnum only overlap in umin.
u64: ---[xxxxxx]-----
tnum: --xx----------x-I think this hunk should be rewritten as follows:
Quoted 12 lines
Quoted 4 lines
Quoted 4 lines
Quoted 4 lines
tnum_next = tnum_step(reg->var_off, reg->umin_value); tnum_max = reg->var_off.value | reg->var_off.mask; tnum_min = reg->var_off.value; if (tnum_next > reg->umax_value) {Quoted 8 lines
Quoted 8 lines
Quoted 8 lines
/* The only overlap is umin */ ___mark_reg_known(reg, tnum_min); } else if (tnum_min < reg->umin_value && tnum_next == tnum_max) { /* The only overlap is tmax */ ___mark_reg_known(reg, tnum_next); } else if (tnum_next <= reg->umax_value && tnum_step(reg->var_off, tnum_next) > reg->umax_value) { ___mark_reg_known(reg, tnum_next);
Actually, that last condition is not enough. In the original patch we also would check that umin is not part of the tnum. That's needed for example if we have R3=(u64=[0; 1], var_off=unknown) (a case possible because __update_reg_bounds is the first refinement step in reg_bounds_sync). With those values, we would match the third condition and set R3=1.
Checking that tmin < reg->umin_value in the third condition would also not work. For example R3=(u64=[0xffff; 0x10000], var_off=(0; 0x1ffff)) would be incorrectly simplified to R3=0x10000 because tmin=0, umin=0xffff.
What we really want is to check that umin is not already part of the tnum, as in the original patch. I think we can however improve readability as you did. Here's what I have in mind:
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 (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 once in between umin
* (excluded) and umax.
* u64: ---[xxxxxx]-----
* tnum: xx----x-------x-
*/
___mark_reg_known(reg, tnum_next);
}Wdyt?
[...]
Eduard Zingerman
maintainerRe: [PATCH bpf 2/4] bpf: Improve bounds when tnum has a single possible value
Quoted 65 lines · Paul Chaignon
Quoted 9 lines · Paul Chaignon
Quoted 8 lines · Eduard Zingerman
Quoted 7 lines · Harishankar Vishwanathan
On Tue, Feb 17, 2026 at 5:58 PM Eduard Zingerman <[email protected]> wrote:
[...]Quoted 5 lines
Quoted 3 lines
1. The u64 range and the tnum only overlap in umin. u64: ---[xxxxxx]-----
tnum: --xx----------x-I think this hunk should be rewritten as follows:
Quoted 12 lines
Quoted 4 lines
Quoted 4 lines
Quoted 4 lines
tnum_next=tnum_step(reg->var_off, reg->umin_value);tnum_max=reg->var_off.value|reg->var_off.mask;tnum_min=reg->var_off.value;
if (tnum_next>reg->umax_value) {Quoted 8 lines
Quoted 8 lines
Quoted 8 lines
/* The only overlap is umin */
___mark_reg_known(reg, tnum_min);
} else if (tnum_min<reg->umin_value&&tnum_next==tnum_max) { /* The only overlap is tmax */___mark_reg_known(reg, tnum_next);
} else if (tnum_next<=reg->umax_value&&tnum_step(reg->var_off, tnum_next)>reg->umax_value) {___mark_reg_known(reg, tnum_next);Actually, that last condition is not enough. In the original patch we also would check that umin is not part of the tnum. That's needed for example if we have R3=(u64=[0; 1], var_off=unknown) (a case possible because
__update_reg_boundsis the first refinement step inreg_bounds_sync). With those values, we would match the third condition and set R3=1.Checking that tmin <
reg->umin_valuein the third condition would also not work. For example R3=(u64=[0xffff;0x10000], var_off=(0;0x1ffff)) would be incorrectly simplified to R3=0x10000 because tmin=0, umin=0xffff.What we really want is to check that umin is not already part of the tnum, as in the original patch. I think we can however improve readability as you did. Here's what I have in mind:
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 (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 once in between umin
* (excluded) and umax.
* u64: ---[xxxxxx]-----
* tnum: xx----x-------x-
*/
___mark_reg_known(reg, tnum_next);
}
Quoted 1 line
Wdyt?
This makes sense, I missed this nuance in the original patch. Maybe keep the first check as 'umin_in_tnum && tnum_next > reg->umax_value', then?
Re: [PATCH bpf 2/4] bpf: Improve bounds when tnum has a single possible value
On Thu, Feb 19, 2026 at 8:29 PM Eduard Zingerman <[email protected]> wrote:
[...]
Quoted 26 lines
Quoted 26 lines
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 (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 once in between umin * (excluded) and umax. * u64: ---[xxxxxx]----- * tnum: xx----x-------x- */ ___mark_reg_known(reg, tnum_next); }
Quoted 5 lines
Quoted 1 line
Wdyt?
This makes sense, I missed this nuance in the original patch. Maybe keep the first check as '
umin_in_tnum&&tnum_next>reg->umax_value', then?
I too think Eduard's suggestion makes sense. We will be explicit in the code that
we want umin to be part of the tnum.
Re: [PATCH bpf 2/4] bpf: Improve bounds when tnum has a single possible value
Quoted 30 lines · Harishankar Vishwanathan
On Thu, Feb 19, 2026 at 8:29 PM Eduard Zingerman <[email protected]> wrote:
[...]Quoted 28 lines
Quoted 27 lines
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 (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 once in between umin * (excluded) and umax. * u64: ---[xxxxxx]----- * tnum: xx----x-------x- */ ___mark_reg_known(reg, tnum_next); }Wdyt?
Quoted 7 lines
Quoted 3 lines
This makes sense, I missed this nuance in the original patch. Maybe keep the first check as '
umin_in_tnum&&tnum_next>reg->umax_value', then?I too think Eduard's suggestion makes sense. We will be explicit in the code that
we want umin to be part of the tnum.
Agree, it helps readability. I've sent the v2 with that change.
Thanks everyone!
PATCH 3/4 selftests/bpf: Test refinement of single-value tnum
In reply to Paul ChaignonThis 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 tests fail because of R0 having a non-constant value in the verifier logs.
--- .../selftests/bpf/progs/verifier_bounds.c | 91 +++++++++++++++++++ 1 file changed, 91 insertions(+)
tools/testing/selftests/bpf/progs/verifier_bounds.c +91 -0
@@ -1863,4 +1863,95 @@ 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=0xf00=3840. +*/ +SEC("socket") +__description("bounds refinement with single-value tnum on umax") +__msg("3: (15) if r0 == 0xe00 {{.*}} R0=3840") +__success __log_level(2) +__flag (BPF_F_TEST_REG_INVARIANTS) +__naked void bounds_refinement_tnum_umax(void *ctx) +{ + asmvolatile (" \ +call %[bpf_get_prandom_u32]; \ +r0 |= 0xe00; \ +r0 &= 0xf00; \ +if r0 == 0xe00 goto +2; \ +if r0 == 0xf00 goto +1; \ +r0 = 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=0xe00=3584. +*/ +SEC("socket") +__description("bounds refinement with single-value tnum on umin") +__msg("3: (15) if r0 == 0xf00 {{.*}} R0=3584") +__success __log_level(2) +__flag (BPF_F_TEST_REG_INVARIANTS) +__naked void bounds_refinement_tnum_umin(void *ctx) +{ + asmvolatile (" \ +call %[bpf_get_prandom_u32]; \ +r0 |= 0xe00; \ +r0 &= 0xf00; \ +if r0 == 0xf00 goto +2; \ +if r0 == 0xe00 goto +1; \ +r0 = 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 +* | [xxxxxxxxxxxxxx] | +* |----------------------------|------------------------------| +* | x x x x x| tnum values +* +* The verifier can therefore deduce that the 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) +{ + asmvolatile (" \ +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); +} + char _license[] SEC("license") = "GPL";
Signature · 1 line
2.43.0
PATCH 4/4 selftests/bpf: Avoid simplification of crafted bounds test
In reply to Paul ChaignonThe 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
@@ -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