Re: [External] Re: [PATCH bpf-next v2 5/6] bpf: teach the verifier to enforce css_iter and process_iter in RCU CS

[Chuyi Zhou <zhouchuyi@bytedance.com>]

Hello.

在 2023/9/15 07:26, Andrii Nakryiko 写道:
> On Thu, Sep 14, 2023 at 1:56 AM Chuyi Zhou <zhouchuyi@bytedance.com> wrote:
>>
>>
>>
>> 在 2023/9/13 21:53, Chuyi Zhou 写道:
>>> Hello.
>>>
>>> 在 2023/9/12 15:01, Chuyi Zhou 写道:
>>>> css_iter and process_iter should be used in rcu section. Specifically, in
>>>> sleepable progs explicit bpf_rcu_read_lock() is needed before use these
>>>> iters. In normal bpf progs that have implicit rcu_read_lock(), it's OK to
>>>> use them directly.
>>>>
>>>> This patch checks whether we are in rcu cs before we want to invoke
>>>> bpf_iter_process_new and bpf_iter_css_{pre, post}_new in
>>>> mark_stack_slots_iter(). If the rcu protection is guaranteed, we would
>>>> let st->type = PTR_TO_STACK | MEM_RCU. is_iter_reg_valid_init() will
>>>> reject if reg->type is UNTRUSTED.
>>>
>>> I use the following BPF Prog to test this patch:
>>>
>>> SEC("?fentry.s/" SYS_PREFIX "sys_getpgid")
>>> int iter_task_for_each_sleep(void *ctx)
>>> {
>>>       struct task_struct *task;
>>>       struct task_struct *cur_task = bpf_get_current_task_btf();
>>>
>>>       if (cur_task->pid != target_pid)
>>>           return 0;
>>>       bpf_rcu_read_lock();
>>>       bpf_for_each(process, task) {
>>>           bpf_rcu_read_unlock();
>>>           if (task->pid == target_pid)
>>>               process_cnt += 1;
>>>           bpf_rcu_read_lock();
>>>       }
>>>       bpf_rcu_read_unlock();
>>>       return 0;
>>> }
>>>
>>> Unfortunately, we can pass the verifier.
>>>
>>> Then I add some printk-messages before setting/clearing state to help
>>> debug:
>>>
>>> diff --git a/kernel/bpf/verifier.c b/kernel/bpf/verifier.c
>>> index d151e6b43a5f..35f3fa9471a9 100644
>>> --- a/kernel/bpf/verifier.c
>>> +++ b/kernel/bpf/verifier.c
>>> @@ -1200,7 +1200,7 @@ static int mark_stack_slots_iter(struct
>>> bpf_verifier_env *env,
>>>                   __mark_reg_known_zero(st);
>>>                   st->type = PTR_TO_STACK; /* we don't have dedicated reg
>>> type */
>>>                   if (is_iter_need_rcu(meta)) {
>>> +                       printk("mark reg_addr : %px", st);
>>>                           if (in_rcu_cs(env))
>>>                                   st->type |= MEM_RCU;
>>>                           else
>>> @@ -11472,8 +11472,8 @@ static int check_kfunc_call(struct
>>> bpf_verifier_env *env, struct bpf_insn *insn,
>>>                           return -EINVAL;
>>>                   } else if (rcu_unlock) {
>>>                           bpf_for_each_reg_in_vstate(env->cur_state,
>>> state, reg, ({
>>> +                               printk("clear reg_addr : %px MEM_RCU :
>>> %d PTR_UNTRUSTED : %d\n ", reg, reg->type & MEM_RCU, reg->type &
>>> PTR_UNTRUSTED);
>>>                                   if (reg->type & MEM_RCU) {
>>> -                                       printk("clear reg addr : %lld",
>>> reg);
>>>                                           reg->type &= ~(MEM_RCU |
>>> PTR_MAYBE_NULL);
>>>                                           reg->type |= PTR_UNTRUSTED;
>>>                                   }
>>>
>>>
>>> The demsg log:
>>>
>>> [  393.705324] mark reg_addr : ffff88814e40e200
>>>
>>> [  393.706883] clear reg_addr : ffff88814d5f8000 MEM_RCU : 0
>>> PTR_UNTRUSTED : 0
>>>
>>> [  393.707353] clear reg_addr : ffff88814d5f8078 MEM_RCU : 0
>>> PTR_UNTRUSTED : 0
>>>
>>> [  393.708099] clear reg_addr : ffff88814d5f80f0 MEM_RCU : 0
>>> PTR_UNTRUSTED : 0
>>> ....
>>> ....
>>>
>>> I didn't see ffff88814e40e200 is cleared as expected because
>>> bpf_for_each_reg_in_vstate didn't find it.
>>>
>>> It seems when we are doing bpf_read_unlock() in the middle of iteration
>>> and want to clearing state through bpf_for_each_reg_in_vstate, we can
>>> not find the previous reg which we marked MEM_RCU/PTR_UNTRUSTED in
>>> mark_stack_slots_iter().
>>>
>>
>> bpf_get_spilled_reg will skip slots if they are not STACK_SPILL, but in
>> mark_stack_slots_iter() we has marked the slots *STACK_ITER*
>>
>> With the following change, everything seems work OK.
>>
>> diff --git a/include/linux/bpf_verifier.h b/include/linux/bpf_verifier.h
>> index a3236651ec64..83c5ecccadb4 100644
>> --- a/include/linux/bpf_verifier.h
>> +++ b/include/linux/bpf_verifier.h
>> @@ -387,7 +387,7 @@ struct bpf_verifier_state {
>>
>>    #define bpf_get_spilled_reg(slot, frame)                               \
>>           (((slot < frame->allocated_stack / BPF_REG_SIZE) &&             \
>> -         (frame->stack[slot].slot_type[0] == STACK_SPILL))             \
>> +         (frame->stack[slot].slot_type[0] == STACK_SPILL ||
>> frame->stack[slot].slot_type[0] == STACK_ITER))            \
>>            ? &frame->stack[slot].spilled_ptr : NULL)
>>
>> I am not sure whether this would harm some logic implicitly when using
>> bpf_get_spilled_reg/bpf_for_each_spilled_reg in other place. If so,
>> maybe we should add a extra parameter to control the picking behaviour.
>>
>> #define bpf_get_spilled_reg(slot, frame, stack_type)
>>                          \
>>          (((slot < frame->allocated_stack / BPF_REG_SIZE) &&             \
>>            (frame->stack[slot].slot_type[0] == stack_type))              \
>>           ? &frame->stack[slot].spilled_ptr : NULL)
>>
>> Thanks.
> 
> I don't think it's safe to just make bpf_get_spilled_reg, and
> subsequently bpf_for_each_reg_in_vstate and bpf_for_each_spilled_reg
> just suddenly start iterating iterator states and/or dynptrs. At least
> some of existing uses of those assume they are really working just
> with registers.

IIUC, when we are doing bpf_rcu_unlock, we do need to clear the state of 
reg including STACK_ITER.

Maybe here we only need change the logic when using 
bpf_for_each_reg_in_vstate to clear state in bpf_rcu_unlock and keep 
everything else unchanged ?

Thanks.