Re: [PATCH] bpf: Prevent invalid u32 bounds in __reg32_deduce_bounds()

[Eduard Zingerman <eddyz87@gmail.com>]

On Fri, 2026-02-20 at 20:07 +0100, Andrea Righi wrote:
> When refining register bounds, __reg32_deduce_bounds can derive u32
> min/max from 64-bit or s32 bounds and assign them with max_t/min_t.
> 
> If the existing u32 range and the derived range do not overlap (e.g.,
> u64 says [0, 1] while u32 was [2, 2] from an earlier path), the
> intersection is empty and the new u32_min_value can end up greater than
> u32_max_value, triggering the following warning:
> 
>   verifier bug: REG INVARIANTS VIOLATION (false_reg1): range bounds violation
>     u64=[0x0, 0x1] s64=[0x0, 0x1] u32=[0x3, 0x1] s32=[0x0, 0x1] var_off=(0x0, 0x1)
>   WARNING: kernel/bpf/verifier.c:2742 at reg_bounds_sanity_check
>   Call Trace:
>     reg_bounds_sanity_check+0xbc/0x1e0
>     reg_set_min_max+0x1a2/0x1f0
>     check_cond_jmp_op+0x5d2/0x1980
>     do_check_common+0x2b0f/0x3410
>     do_check_subprogs+0xcd/0x180
>     bpf_check+0x33fe/0x3850
>     bpf_prog_load+0x7d7/0xee0
>     __sys_bpf+0xea2/0x2e30
> 
> This was triggered by the scx CI while loading the scx_layered sched_ext
> scheduler [1].
> 
> Fix by only applying the derived u32 bounds when the resulting range is
> valid (u32_min <= u32_max).
> 
> [1] https://github.com/sched-ext/scx/pull/3349
> 
> Fixes: c1efab6468fd5 ("bpf: derive subreg bounds from full bounds when upper 32 bits are constant")
> Signed-off-by: Andrea Righi <arighi@nvidia.com>
> ---

Hi Andrea,

Sorry for delayed response.

Currently the assumption is that all ranges/tnum tracked as a part of
a register state are valid, and I don't think we'd like to sidestep
this assumption. Eventually is_scalar_branch_taken() and
adjust_scalar_min_max_vals() should be unified to avoid possibility
for mismatch between is_scalar_branch_taken() predictions (when it
assumes that some branch can be taken by failing to predict jump)
and adjust_scalar_min_max_vals() adjustments (when it infers that in
fact register state is empty for the branch).

As it turns out, the problem you report is a bit easier to fix.
After minimizing the failing file with Emil's help I derived the
following test case that captures the problem:

SEC("socket")
__success
__flag(BPF_F_TEST_REG_INVARIANTS)
__naked void signed_unsined_intersection32(void *ctx)
{
	asm volatile("			\
	call %[bpf_get_prandom_u32];	\
	w0 &= 0xffffffff;		\
	if w0 < 0x3 goto 1f;		\ /* on fall-through u32 range [3..U32_MAX] */
	if w0 s> 0x1 goto 1f;		\ /* on fall-through s32 range [S32_MIN..1] *
	if w0 s< 0x0 goto 1f;		\ /* the above ranges can be narrowed to [S32_MIN..-1] */
	r10 = 0;			\ /* meaning that condition can be predicted, */
1:	exit;				\ /* but verifier is not smart enough at the moment. */
"	:
	: __imm(bpf_get_prandom_u32)
	: __clobber_all);
}

This branch has verifier modifications making it smart enough to
process this case: https://github.com/eddyz87/bpf/tree/cnum-sync-bounds .
With some additional verification: https://github.com/eddyz87/cnum-verif .
I'm working towards submitting the fix, need to decide on what to do
with reg_bounds.c, as this test is now not smart enough.

Thanks,
Eduard

[...]