Re: [PATCH bpf-next v8 08/22] bpf: Require an unwind to leave a frame holding what it entered with

[Yonghong Song <yonghong.song@linux.dev>]

On 10/1/26 7:31 AM, bot+bpf-ci@kernel.org wrote:
>> bpf: Require an unwind to leave a frame holding what it entered with
> TL;DR: call_may_unwind() treats every callx as possibly unwinding, so the
> verifier may wrongly reject valid programs that hold a reference or spin lock
> across a callx; the exact target is available from the register.
>
>> diff --git a/kernel/bpf/verifier.c b/kernel/bpf/verifier.c
>> index c7a350be538e..a7b25ab04051 100644
>> --- a/kernel/bpf/verifier.c
>> +++ b/kernel/bpf/verifier.c
> [ ... ]
>
>> @@ -19244,6 +19245,32 @@ static int unwind_out_of_global_call(struct bpf_verifier_env *env, int call_idx,
>>   	return INSN_IDX_UPDATED;
>>   }
>>
>> +/* Can an unwind come back out of this call? */
>> +static bool call_may_unwind(struct bpf_verifier_env *env, const struct bpf_insn *insn,
>> +			    int insn_idx)
>> +{
>> +	int subprog;
>> +
>> +	/* Which subprog a callx lands in is not known here, so any may be it. */
>> +	if (bpf_is_callx(insn))
>> +		return bpf_prog_may_unwind(env);
> Is it true that the target of a callx is not known here?

This is way to conservatile as if *any* subprog is marked as may_unwind, then
it will return bpf_is_callx(insn). The best thing is to follow callx itself
and find whether it is possible may_unwind or not.

>
> call_may_unwind() is reached from check_unwind_through_call() in
> do_check_insn(), just before check_func_callx().  check_func_callx() takes
> the exact target from the same register state a few lines later:
>
> 	reg = reg_state(env, insn->dst_reg);
> 	if (reg->type != PTR_TO_FUNC) ...
> 	/* PTR_TO_FUNC is a pointer to a static subprog */
> 	subprog = reg->subprogno;
>
> and check_func_ptr_read() pushes one state per possible table target, each
> with its own subprogno.  So each verified callx has exactly one target at
> this point.  Only the CFG pass has to be conservative, and
> mark_callx_might_unwind() already marks any target that itself contains a
> callx.
>
> With the current code, an uncovered callx is refused by
> bpf_exc_check_frame_balance() whenever the calling frame has changed its
> RCU, preempt, IRQ, spin lock or reference state since entry, as long as some
> unrelated subprog in the program is marked might_unwind.  For example:
>
> 	foo():
> 	    t = bpf_task_acquire(...);  /* id > entry_id_gen */
> 	    r1 = vtable[i];             /* PTR_TO_FUNC -> bar, no unwind in bar */
> 	    callx r1                    /* no cleanup record covers it */
> 	    bpf_task_release(t);
>
> 	baz():
> 	    bpf_unwind();               /* makes bpf_prog_may_unwind() true */
>
> Before this patch this program was accepted.  Now it is refused with "an
> unwind through this call keeps the reference id=N the frame acquired", even
> though no unwind can pass through that call.  callx is also explicitly
> allowed under a spin lock, so a frame that holds a bpf_spin_lock across a
> callx is refused as well with "an unwind through this call does not leave
> the frame's bpf_spin_lock state as it found it".
>
> Could this look at the register instead, something like:
>
> 	reg = &cur_regs(env)[insn->dst_reg];
> 	if (reg->type == PTR_TO_FUNC)
> 		return env->subprog_info[reg->subprogno].might_unwind;
>
> and leave it to check_func_callx() to refuse any other register type, as it
> does today?
>
>> +	if (insn->src_reg != BPF_PSEUDO_CALL)
>> +		return false;
>> +	subprog = bpf_find_subprog(env, insn_idx + insn->imm + 1);
>> +	return subprog >= 0 && env->subprog_info[subprog].might_unwind;
>> +}
>> +
> [ ... ]
>
>
> ---
> 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/36872142096