Re: [PATCH] percpu_ref: call wake_up_all() after percpu_ref_put() completes

[Qi Zheng <zhengqi.arch@bytedance.com>]

On 2022/4/9 8:40 AM, Qi Zheng wrote:
> 
> 
> On 2022/4/9 1:41 AM, Tejun Heo wrote:
>> Hello,
>>
>> On Thu, Apr 07, 2022 at 06:33:35PM +0800, Qi Zheng wrote:
>>> In the percpu_ref_call_confirm_rcu(), we call the wake_up_all()
>>> before calling percpu_ref_put(), which will cause the value of
>>> percpu_ref to be unstable when percpu_ref_switch_to_atomic_sync()
>>> returns.
>>>
>>>     CPU0                CPU1
>>>
>>> percpu_ref_switch_to_atomic_sync(&ref)
>>> --> percpu_ref_switch_to_atomic(&ref)
>>>      --> percpu_ref_get(ref);    /* put after confirmation */
>>>     call_rcu(&ref->data->rcu, percpu_ref_switch_to_atomic_rcu);
>>>
>>>                     percpu_ref_switch_to_atomic_rcu
>>>                     --> percpu_ref_call_confirm_rcu
>>>                         --> data->confirm_switch = NULL;
>>>                         wake_up_all(&percpu_ref_switch_waitq);
>>>
>>>      /* here waiting to wake up */
>>>      wait_event(percpu_ref_switch_waitq, !ref->data->confirm_switch);
>>>                         (A)percpu_ref_put(ref);
>>> /* The value of &ref is unstable! */
>>> percpu_ref_is_zero(&ref)
>>>                         (B)percpu_ref_put(ref);
>>>
>>> As shown above, assuming that the counts on each cpu add up to 0 before
>>> calling percpu_ref_switch_to_atomic_sync(), we expect that after 
>>> switching
>>> to atomic mode, percpu_ref_is_zero() can return true. But actually it 
>>> will
>>> return different values in the two cases of A and B, which is not what
>>> we expected.
>>>
>>> Maybe the original purpose of percpu_ref_switch_to_atomic_sync() is
>>> just to ensure that the conversion to atomic mode is completed, but it
>>> should not return with an extra reference count.
>>>
>>> Calling wake_up_all() after percpu_ref_put() ensures that the value of
>>> percpu_ref is stable after percpu_ref_switch_to_atomic_sync() returns.
>>> So just do it.
>>>
>>> Signed-off-by: Qi Zheng <zhengqi.arch@bytedance.com>
>>> ---
>>>   lib/percpu-refcount.c | 3 ++-
>>>   1 file changed, 2 insertions(+), 1 deletion(-)
>>>
>>> diff --git a/lib/percpu-refcount.c b/lib/percpu-refcount.c
>>> index af9302141bcf..b11b4152c8cd 100644
>>> --- a/lib/percpu-refcount.c
>>> +++ b/lib/percpu-refcount.c
>>> @@ -154,13 +154,14 @@ static void percpu_ref_call_confirm_rcu(struct 
>>> rcu_head *rcu)
>>>       data->confirm_switch(ref);
>>>       data->confirm_switch = NULL;
>>> -    wake_up_all(&percpu_ref_switch_waitq);
>>>       if (!data->allow_reinit)
>>>           __percpu_ref_exit(ref);
>>>       /* drop ref from percpu_ref_switch_to_atomic() */
>>>       percpu_ref_put(ref);
>>> +
>>> +    wake_up_all(&percpu_ref_switch_waitq);
>>
>> The interface, at least originally, doesn't give any guarantee over 
>> whether
>> there's gonna be a residual reference on it or not. There's nothing
>> necessarily wrong with guaranteeing that but it's rather unusual and 
>> given
>> that putting the base ref in a percpu_ref is a special "kill" 
>> operation and
>> a ref in percpu mode always returns %false on is_zero(), I'm not quite 
>> sure
>> how such semantics would be useful. Do you care to explain the use 
>> case with
>> concrete examples?
> 
> There are currently two users of percpu_ref_switch_to_atomic_sync(), and 
> both are used in the example, one is mddev->writes_pending in
> driver/md/md.c and the other is q->q_usage_counter in block/blk-pm.c.
> 
> The former discards the initial reference count after percpu_ref_init(),
> and the latter kills the initial reference count(by calling 
> percpu_ref_kill() in blk_freeze_queue_start()) before
> percpu_ref_switch_to_atomic_sync(). Looks like they all expect
> percpu_ref to be stable when percpu_ref_switch_to_atomic_sync() returns.
> 
>>
>> Also, the proposed patch is racy. There's nothing preventing
>> percpu_ref_switch_to_atomic_sync() from waking up early between
>> confirm_switch clearing and the wake_up_all, so the above change doesn't
>> guarantee what it tries to guarantee. For that, you'd have to move
>> confirm_switch clearing *after* percpu_ref_put() but then, you'd be
>> accessing the ref after its final ref is put which can lead to
>> use-after-free.
>>
> 
> Oh sorry, it is my bad missing.
> 
>> In fact, the whole premise seems wrong. The switching needs a 
>> reference to
>> the percpu_ref because it is accessing it asynchronously. The 
>> switching side
>> doesn't know when the ref is gonna go away once it puts its reference and
>> thus can't signal that they're done after putting their reference.
>>
>> We *can* make that work by putting the whole thing in its own critical
>> section so that we can make confirm_switch clearing atomic with the 
>> possibly
>> final put, but that's gonna add some complexity and begs the question why
>> we'd need such a thing.
> 
> How about moving the last percpu_ref_put() outside of the
> percpu_ref_switch_to_atomic_rcu() in sync mode like below? But this may 
> not be elegant.
> 
> diff --git a/include/linux/percpu-refcount.h 
> b/include/linux/percpu-refcount.h
> index d73a1c08c3e3..07f92e7e3e19 100644
> --- a/include/linux/percpu-refcount.h
> +++ b/include/linux/percpu-refcount.h
> @@ -98,6 +98,7 @@ struct percpu_ref_data {
>          percpu_ref_func_t       *confirm_switch;
>          bool                    force_atomic:1;
>          bool                    allow_reinit:1;
> +       bool                    sync;
>          struct rcu_head         rcu;
>          struct percpu_ref       *ref;
>   };
> @@ -123,7 +124,8 @@ int __must_check percpu_ref_init(struct percpu_ref 
> *ref,
>                                   gfp_t gfp);
>   void percpu_ref_exit(struct percpu_ref *ref);
>   void percpu_ref_switch_to_atomic(struct percpu_ref *ref,
> -                                percpu_ref_func_t *confirm_switch);
> +                                percpu_ref_func_t *confirm_switch,
> +                                bool sync);
>   void percpu_ref_switch_to_atomic_sync(struct percpu_ref *ref);
>   void percpu_ref_switch_to_percpu(struct percpu_ref *ref);
>   void percpu_ref_kill_and_confirm(struct percpu_ref *ref,
> diff --git a/lib/percpu-refcount.c b/lib/percpu-refcount.c
> index af9302141bcf..2a9d777bcf35 100644
> --- a/lib/percpu-refcount.c
> +++ b/lib/percpu-refcount.c
> @@ -99,6 +99,7 @@ int percpu_ref_init(struct percpu_ref *ref, 
> percpu_ref_func_t *release,
>          data->release = release;
>          data->confirm_switch = NULL;
>          data->ref = ref;
> +       data->sync = false;
>          ref->data = data;
>          return 0;
>   }
> @@ -146,21 +147,30 @@ void percpu_ref_exit(struct percpu_ref *ref)
>   }
>   EXPORT_SYMBOL_GPL(percpu_ref_exit);
> 
> +static inline void percpu_ref_switch_to_atomic_post(struct percpu_ref 
> *ref)
> +{
> +       struct percpu_ref_data *data = ref->data;
> +
> +       if (!data->allow_reinit)
> +               __percpu_ref_exit(ref);
> +
> +       /* drop ref from percpu_ref_switch_to_atomic() */
> +       percpu_ref_put(ref);
> +}
> +
>   static void percpu_ref_call_confirm_rcu(struct rcu_head *rcu)
>   {
>          struct percpu_ref_data *data = container_of(rcu,
>                          struct percpu_ref_data, rcu);
>          struct percpu_ref *ref = data->ref;
> +       bool need_put = !data->sync;
> 
>          data->confirm_switch(ref);
>          data->confirm_switch = NULL;
>          wake_up_all(&percpu_ref_switch_waitq);
> 
> -       if (!data->allow_reinit)
> -               __percpu_ref_exit(ref);
> -
> -       /* drop ref from percpu_ref_switch_to_atomic() */
> -       percpu_ref_put(ref);
> +       if (need_put)
> +               percpu_ref_switch_to_atomic_post(ref);
>   }
> 
>   static void percpu_ref_switch_to_atomic_rcu(struct rcu_head *rcu)
> @@ -302,12 +312,14 @@ static void __percpu_ref_switch_mode(struct 
> percpu_ref *ref,
>    * switching to atomic mode, this function can be called from any 
> context.
>    */
>   void percpu_ref_switch_to_atomic(struct percpu_ref *ref,
> -                                percpu_ref_func_t *confirm_switch)
> +                                percpu_ref_func_t *confirm_switch,
> +                                bool sync)
>   {
>          unsigned long flags;
> 
>          spin_lock_irqsave(&percpu_ref_switch_lock, flags);
> 
> +       ref->data->sync = sync;
>          ref->data->force_atomic = true;
>          __percpu_ref_switch_mode(ref, confirm_switch);
> 
> @@ -325,8 +337,9 @@ EXPORT_SYMBOL_GPL(percpu_ref_switch_to_atomic);
>    */
>   void percpu_ref_switch_to_atomic_sync(struct percpu_ref *ref)
>   {
> -       percpu_ref_switch_to_atomic(ref, NULL);
> +       percpu_ref_switch_to_atomic(ref, NULL, true);
>          wait_event(percpu_ref_switch_waitq, !ref->data->confirm_switch);
> +       percpu_ref_switch_to_atomic_post(ref);
>   }
>   EXPORT_SYMBOL_GPL(percpu_ref_switch_to_atomic_sync);

Hi all, is this fix ok? I can send the v2 version if it looks good.

Thanks,
Qi

> 
>>
>> Andrew, I don't think the patch as proposed makes much sense. Maybe 
>> it'd be
>> better to keep it out of the tree for the time being?
>>
>> Thanks.
>>
> 

-- 
Thanks,
Qi