| Message ID | 158507151689.15666.566796274289413203.stgit@john-Precision-5820-Tower |
|---|---|
| State | New |
| Headers | show |
| Series | [bpf-next,01/10] bpf: verifier, do_refine_retval_range may clamp umin to 0 incorrectly | expand |
diff --git a/kernel/bpf/verifier.c b/kernel/bpf/verifier.c index f7a34b1..fea9725 100644 --- a/kernel/bpf/verifier.c +++ b/kernel/bpf/verifier.c @@ -5202,6 +5202,7 @@ static int adjust_scalar_min_max_vals(struct bpf_verifier_env *env, coerce_reg_to_size(dst_reg, 4); } + __update_reg_bounds(dst_reg); __reg_deduce_bounds(dst_reg); __reg_bound_offset(dst_reg); return 0;
Currently, for all op verification we call __red_deduce_bounds() and __red_bound_offset() but we only call __update_reg_bounds() in bitwise ops. However, we could benefit from calling __update_reg_bounds() in BPF_ADD, BPF_SUB, and BPF_MUL cases as well. For example, a register with state 'R1_w=invP0' when we subtract from it, w1 -= 2 Before coerce we will now have an smin_value=S64_MIN, smax_value=U64_MAX and unsigned bounds umin_value=0, umax_value=U64_MAX. These will then be clamped to S32_MIN, U32_MAX values by coerce in the case of alu32 op as done in above example. However tnum will be a constant because the ALU op is done on a constant. Without update_reg_bounds() we have a scenario where tnum is a const but our unsigned bounds do not reflect this. By calling update_reg_bounds after coerce to 32bit we further refine the umin_value to U64_MAX in the alu64 case or U32_MAX in the alu32 case above. Signed-off-by: John Fastabend <john.fastabend@gmail.com> --- kernel/bpf/verifier.c | 1 + 1 file changed, 1 insertion(+)