Re: [PATCH bpf-next 1/2] bpf: Skip bounds adjustment for conditional jumps on same register

[Eduard Zingerman <eddyz87@gmail.com>]

On Wed, 2025-10-22 at 13:12 -0700, Alexei Starovoitov wrote:
> On Wed, Oct 22, 2025 at 12:46 PM Eduard Zingerman <eddyz87@gmail.com> wrote:
> > 
> > On Wed, 2025-10-22 at 11:14 -0700, Yonghong Song wrote:
> > > 
> > > On 10/22/25 9:44 AM, KaFai Wan wrote:
> > > > When conditional jumps are performed on the same register (e.g., r0 <= r0,
> > > > r0 > r0, r0 < r0) where the register holds a scalar with range, the verifier
> > > > incorrectly attempts to adjust the register's min/max bounds. This leads to
> > > > invalid range bounds and triggers a BUG warning:
> > > > 
> > > > verifier bug: REG INVARIANTS VIOLATION (true_reg1): range bounds violation u64=[0x1, 0x0] s64=[0x1, 0x0] u32=[0x1, 0x0] s32=[0x1, 0x0] var_off=(0x0, 0x0)
> > > > WARNING: CPU: 0 PID: 93 at kernel/bpf/verifier.c:2731 reg_bounds_sanity_check+0x163/0x220
> > > > Modules linked in:
> > > > CPU: 0 UID: 0 PID: 93 Comm: repro-x-3 Tainted: G        W           6.18.0-rc1-ge7586577b75f-dirty #218 PREEMPT(full)
> > > > Tainted: [W]=WARN
> > > > Hardware name: QEMU Ubuntu 24.04 PC (i440FX + PIIX, 1996), BIOS 1.16.3-debian-1.16.3-2 04/01/2014
> > > > RIP: 0010:reg_bounds_sanity_check+0x163/0x220
> > > > Call Trace:
> > > >   <TASK>
> > > >   reg_set_min_max.part.0+0x1b1/0x360
> > > >   check_cond_jmp_op+0x1195/0x1a60
> > > >   do_check_common+0x33ac/0x33c0
> > > >   ...
> > > > 
> > > > The issue occurs in reg_set_min_max() function where bounds adjustment logic
> > > > is applied even when both registers being compared are the same. Comparing a
> > > > register with itself should not change its bounds since the comparison result
> > > > is always known (e.g., r0 == r0 is always true, r0 < r0 is always false).
> > > > 
> > > > Fix this by adding an early return in reg_set_min_max() when false_reg1 and
> > > > false_reg2 point to the same register, skipping the unnecessary bounds
> > > > adjustment that leads to the verifier bug.
> > > > 
> > > > Reported-by: Kaiyan Mei <M202472210@hust.edu.cn>
> > > > Reported-by: Yinhao Hu <dddddd@hust.edu.cn>
> > > > Closes: https://lore.kernel.org/all/1881f0f5.300df.199f2576a01.Coremail.kaiyanm@hust.edu.cn/
> > > > Fixes: 0df1a55afa83 ("bpf: Warn on internal verifier errors")
> > > > Signed-off-by: KaFai Wan <kafai.wan@linux.dev>
> > > > ---
> > > >   kernel/bpf/verifier.c | 4 ++++
> > > >   1 file changed, 4 insertions(+)
> > > > 
> > > > diff --git a/kernel/bpf/verifier.c b/kernel/bpf/verifier.c
> > > > index 6d175849e57a..420ad512d1af 100644
> > > > --- a/kernel/bpf/verifier.c
> > > > +++ b/kernel/bpf/verifier.c
> > > > @@ -16429,6 +16429,10 @@ static int reg_set_min_max(struct bpf_verifier_env *env,
> > > >     if (false_reg1->type != SCALAR_VALUE || false_reg2->type != SCALAR_VALUE)
> > > >             return 0;
> > > > 
> > > > +   /* If conditional jumps on the same register, skip the adjustment */
> > > > +   if (false_reg1 == false_reg2)
> > > > +           return 0;
> > > 
> > > Your change looks good. But this is a special case and it should not
> > > happen for any compiler generated code. So could you investigate
> > > why regs_refine_cond_op() does not work? Since false_reg1 and false_reg2
> > > is the same, so register refinement should keep the same. Probably
> > > some minor change in regs_refine_cond_op(...) should work?
> > > 
> > > > +
> > > >     /* fallthrough (FALSE) branch */
> > > >     regs_refine_cond_op(false_reg1, false_reg2, rev_opcode(opcode), is_jmp32);
> > > >     reg_bounds_sync(false_reg1);
> > 
> > I think regs_refine_cond_op() is not written in a way to handle same
> > registers passed as reg1 and reg2. E.g. in this particular case the
> > condition is reformulated as "r0 < r0", and then the following branch
> > is taken:
> > 
> >    static void regs_refine_cond_op(struct bpf_reg_state *reg1, struct bpf_reg_state *reg2,
> >                                  u8 opcode, bool is_jmp32)
> >    {
> >         ...
> >          case BPF_JLT: // condition is rephrased as r0 < r0
> >                  if (is_jmp32) {
> >                          ...
> >                  } else {
> >                          reg1->umax_value = min(reg1->umax_value, reg2->umax_value - 1);
> >                          reg2->umin_value = max(reg1->umin_value + 1, reg2->umin_value);
> >                  }
> >                  break;
> >         ...
> >    }
> > 
> > Note that intent is to adjust umax of the LHS (reg1) register and umin
> > of the RHS (reg2) register. But here it ends up adjusting the same register.
> > 
> > (a) before refinement: u64=[0x0, 0x80000000] s64=[0x0, 0x80000000] u32=[0x0, 0x80000000] s32=[0x80000000, 0x0]
> > (b) after  refinement: u64=[0x1, 0x7fffffff] s64=[0x0, 0x80000000] u32=[0x0, 0x80000000] s32=[0x80000000, 0x0]
> > (c) after  sync      : u64=[0x1, 0x0] s64=[0x1, 0x0] u32=[0x1, 0x0] s32=[0x1, 0x0]
> > 
> > At (b) the u64 range translated to s32 is > 0, while s32 range is <= 0,
> > hence the invariant violation.
> > 
> > I think it's better to move the reg1 == reg2 check inside
> > regs_refine_cond_op(), or to handle this case in is_branch_taken().
> 
> hmm. bu then regs_refine_cond_op() will skip it, yet reg_set_min_max()
> will still be doing pointless work with reg_bounds_sync() and sanity check.
> The current patch makes more sense to me.

Well, if we want to avoid useless work, we need something like:

@@ -16173,6 +16173,25 @@ static int is_pkt_ptr_branch_taken(struct bpf_reg_state *dst_reg,
 static int is_branch_taken(struct bpf_reg_state *reg1, struct bpf_reg_state *reg2,
                           u8 opcode, bool is_jmp32)
 {
+       if (reg1 == reg2) {
+               switch (opcode) {
+               case BPF_JGE:
+               case BPF_JLE:
+               case BPF_JSGE:
+               case BPF_JSLE:
+               case BPF_JEQ:
+               case BPF_JSET:
+                       return 1;
+               case BPF_JGT:
+               case BPF_JLT:
+               case BPF_JSGT:
+               case BPF_JSLT:
+               case BPF_JNE:
+                       return 0;
+               default:
+                       return -1;
+               }
+       }

But that's too much code for an artificial case.
Idk, either way is fine with me.