Re: [PATCH bpf-next 1/4] bpf: enable task local storage for tracing programs

[Yonghong Song <yhs@fb.com>]

On 1/15/21 5:12 PM, Song Liu wrote:
> 
> 
>> On Jan 15, 2021, at 4:55 PM, Yonghong Song <yhs@fb.com> wrote:
>>
>>
>>
>> On 1/15/21 3:34 PM, Song Liu wrote:
>>>> On Jan 12, 2021, at 8:53 AM, KP Singh <kpsingh@kernel.org> wrote:
>>>>
>>>> On Tue, Jan 12, 2021 at 5:32 PM Yonghong Song <yhs@fb.com> wrote:
>>>>>
>>>>>
>>>>>
>>>>> On 1/11/21 3:45 PM, Song Liu wrote:
>>>>>>
>>>>>>
>>>>>>> On Jan 11, 2021, at 1:58 PM, Martin Lau <kafai@fb.com> wrote:
>>>>>>>
>>>>>>> On Mon, Jan 11, 2021 at 10:35:43PM +0100, KP Singh wrote:
>>>>>>>> On Mon, Jan 11, 2021 at 7:57 PM Martin KaFai Lau <kafai@fb.com> wrote:
>>>>>>>>>
>>>>>>>>> On Fri, Jan 08, 2021 at 03:19:47PM -0800, Song Liu wrote:
>>>>>>>>>
>>>>>>>>> [ ... ]
>>>>>>>>>
>>>>>>>>>> diff --git a/kernel/bpf/bpf_local_storage.c b/kernel/bpf/bpf_local_storage.c
>>>>>>>>>> index dd5aedee99e73..9bd47ad2b26f1 100644
>>>>>>>>>> --- a/kernel/bpf/bpf_local_storage.c
>>>>>>>>>> +++ b/kernel/bpf/bpf_local_storage.c
>>>>
>>>> [...]
>>>>
>>>>>>>>>> +#include <linux/bpf.h>
>>>>>>>>>>
>>>>>>>>>> #include <asm/pgalloc.h>
>>>>>>>>>> #include <linux/uaccess.h>
>>>>>>>>>> @@ -734,6 +735,7 @@ void __put_task_struct(struct task_struct *tsk)
>>>>>>>>>>       cgroup_free(tsk);
>>>>>>>>>>       task_numa_free(tsk, true);
>>>>>>>>>>       security_task_free(tsk);
>>>>>>>>>> +     bpf_task_storage_free(tsk);
>>>>>>>>>>       exit_creds(tsk);
>>>>>>>>> If exit_creds() is traced by a bpf and this bpf is doing
>>>>>>>>> bpf_task_storage_get(..., BPF_LOCAL_STORAGE_GET_F_CREATE),
>>>>>>>>> new task storage will be created after bpf_task_storage_free().
>>>>>>>>>
>>>>>>>>> I recalled there was an earlier discussion with KP and KP mentioned
>>>>>>>>> BPF_LSM will not be called with a task that is going away.
>>>>>>>>> It seems enabling bpf task storage in bpf tracing will break
>>>>>>>>> this assumption and needs to be addressed?
>>>>>>>>
>>>>>>>> For tracing programs, I think we will need an allow list where
>>>>>>>> task local storage can be used.
>>>>>>> Instead of whitelist, can refcount_inc_not_zero(&tsk->usage) be used?
>>>>>>
>>>>>> I think we can put refcount_inc_not_zero() in bpf_task_storage_get, like:
>>>>>>
>>>>>> diff --git i/kernel/bpf/bpf_task_storage.c w/kernel/bpf/bpf_task_storage.c
>>>>>> index f654b56907b69..93d01b0a010e6 100644
>>>>>> --- i/kernel/bpf/bpf_task_storage.c
>>>>>> +++ w/kernel/bpf/bpf_task_storage.c
>>>>>> @@ -216,6 +216,9 @@ BPF_CALL_4(bpf_task_storage_get, struct bpf_map *, map, struct task_struct *,
>>>>>>           * by an RCU read-side critical section.
>>>>>>           */
>>>>>>          if (flags & BPF_LOCAL_STORAGE_GET_F_CREATE) {
>>>>>> +               if (!refcount_inc_not_zero(&task->usage))
>>>>>> +                       return -EBUSY;
>>>>>> +
>>>>>>                  sdata = bpf_local_storage_update(
>>>>>>                          task, (struct bpf_local_storage_map *)map, value,
>>>>>>                          BPF_NOEXIST);
>>>>>>
>>>>>> But where shall we add the refcount_dec()? IIUC, we cannot add it to
>>>>>> __put_task_struct().
>>>>>
>>>>> Maybe put_task_struct()?
>>>>
>>>> Yeah, something like, or if you find a more elegant alternative :)
>>>>
>>>> --- a/include/linux/sched/task.h
>>>> +++ b/include/linux/sched/task.h
>>>> @@ -107,13 +107,20 @@ extern void __put_task_struct(struct task_struct *t);
>>>>
>>>> static inline void put_task_struct(struct task_struct *t)
>>>> {
>>>> -       if (refcount_dec_and_test(&t->usage))
>>>> +
>>>> +       if (rcu_access_pointer(t->bpf_storage)) {
>>>> +               if (refcount_sub_and_test(2, &t->usage))
>>>> +                       __put_task_struct(t);
>>>> +       } else if (refcount_dec_and_test(&t->usage))
>>>>                 __put_task_struct(t);
>>>> }
>>>>
>>>> static inline void put_task_struct_many(struct task_struct *t, int nr)
>>>> {
>>>> -       if (refcount_sub_and_test(nr, &t->usage))
>>>> +       if (rcu_access_pointer(t->bpf_storage)) {
>>>> +               if (refcount_sub_and_test(nr + 1, &t->usage))
>>>> +                       __put_task_struct(t);
>>>> +       } else if (refcount_sub_and_test(nr, &t->usage))
>>>>                 __put_task_struct(t);
>>>> }
>>> It is not ideal to leak bpf_storage here. How about we only add the
>>> following:
>>> diff --git i/kernel/bpf/bpf_task_storage.c w/kernel/bpf/bpf_task_storage.c
>>> index f654b56907b69..2811b9fc47233 100644
>>> --- i/kernel/bpf/bpf_task_storage.c
>>> +++ w/kernel/bpf/bpf_task_storage.c
>>> @@ -216,6 +216,10 @@ BPF_CALL_4(bpf_task_storage_get, struct bpf_map *, map, struct task_struct *,
>>>           * by an RCU read-side critical section.
>>>           */
>>>          if (flags & BPF_LOCAL_STORAGE_GET_F_CREATE) {
>>> +               /* the task_struct is being freed, fail over*/
>>> +               if (!refcount_read(&task->usage))
>>> +                       return -EBUSY;
>>
>> This may not work? Even we check here and task->usage is not 0, it could still become 0 immediately after the above refcount_read, right?
> 
> We call bpf_task_storage_get() with "task" that has valid BTF, so "task"
> should not go away during the BPF program? Whatever mechanism that

Oh, right. this is true. Otherwise, we cannot use task ptr in the helper.

> triggers the BPF program should either hold a reference to task (usage > 0)
> or be the only one owning it (usage == 0, in __put_task_struct). Did I miss
> anything?

Sorry. I think you are right. Not sure lsm requirement. There are two
more possible ways to check task is exiting which happens before 
__put_task_struct():
   . check task->exit_state
   . check task->flags & PF_EXITING (used in bpf_trace.c)

Not sure which condition is the correct one to check.

> 
> Thanks,
> Song
> 
>>
>>> +
>>>                  sdata = bpf_local_storage_update(
>>>                          task, (struct bpf_local_storage_map *)map, value,
>>>                          BPF_NOEXIST);
>>>>
>>>>
>>>> I may be missing something but shouldn't bpf_storage be an __rcu
>>>> member like we have for sk_bpf_storage?
>>> Good catch! I will fix this in v2.
>>> Thanks,
>>> Song
>