Re: [PATCH v3 13/16] exit: Factor thread_group_exited out of pidfd_poll

[Daniel Borkmann <daniel@iogearbox.net>]

On 7/7/20 7:09 PM, Eric W. Biederman wrote:
> Christian Brauner <christian.brauner@ubuntu.com> writes:
>> On Fri, Jul 03, 2020 at 04:37:47PM -0500, Eric W. Biederman wrote:
>>> Alexei Starovoitov <alexei.starovoitov@gmail.com> writes:
>>>
>>>> On Thu, Jul 02, 2020 at 11:41:37AM -0500, Eric W. Biederman wrote:
>>>>> Create an independent helper thread_group_exited report return true
>>>>> when all threads have passed exit_notify in do_exit.  AKA all of the
>>>>> threads are at least zombies and might be dead or completely gone.
>>>>>
>>>>> Create this helper by taking the logic out of pidfd_poll where
>>>>> it is already tested, and adding a missing READ_ONCE on
>>>>> the read of task->exit_state.
>>>>>
>>>>> I will be changing the user mode driver code to use this same logic
>>>>> to know when a user mode driver needs to be restarted.
>>>>>
>>>>> Place the new helper thread_group_exited in kernel/exit.c and
>>>>> EXPORT it so it can be used by modules.
>>>>>
>>>>> Signed-off-by: "Eric W. Biederman" <ebiederm@xmission.com>
>>>>> ---
>>>>>   include/linux/sched/signal.h |  2 ++
>>>>>   kernel/exit.c                | 24 ++++++++++++++++++++++++
>>>>>   kernel/fork.c                |  6 +-----
>>>>>   3 files changed, 27 insertions(+), 5 deletions(-)
>>>>>
>>>>> diff --git a/include/linux/sched/signal.h b/include/linux/sched/signal.h
>>>>> index 0ee5e696c5d8..1bad18a1d8ba 100644
>>>>> --- a/include/linux/sched/signal.h
>>>>> +++ b/include/linux/sched/signal.h
>>>>> @@ -674,6 +674,8 @@ static inline int thread_group_empty(struct task_struct *p)
>>>>>   #define delay_group_leader(p) \
>>>>>   		(thread_group_leader(p) && !thread_group_empty(p))
>>>>>   
>>>>> +extern bool thread_group_exited(struct pid *pid);
>>>>> +
>>>>>   extern struct sighand_struct *__lock_task_sighand(struct task_struct *task,
>>>>>   							unsigned long *flags);
>>>>>   
>>>>> diff --git a/kernel/exit.c b/kernel/exit.c
>>>>> index d3294b611df1..a7f112feb0f6 100644
>>>>> --- a/kernel/exit.c
>>>>> +++ b/kernel/exit.c
>>>>> @@ -1713,6 +1713,30 @@ COMPAT_SYSCALL_DEFINE5(waitid,
>>>>>   }
>>>>>   #endif
>>>>>   
>>>>> +/**
>>>>> + * thread_group_exited - check that a thread group has exited
>>>>> + * @pid: tgid of thread group to be checked.
>>>>> + *
>>>>> + * Test if thread group is has exited (all threads are zombies, dead
>>>>> + * or completely gone).
>>>>> + *
>>>>> + * Return: true if the thread group has exited. false otherwise.
>>>>> + */
>>>>> +bool thread_group_exited(struct pid *pid)
>>>>> +{
>>>>> +	struct task_struct *task;
>>>>> +	bool exited;
>>>>> +
>>>>> +	rcu_read_lock();
>>>>> +	task = pid_task(pid, PIDTYPE_PID);
>>>>> +	exited = !task ||
>>>>> +		(READ_ONCE(task->exit_state) && thread_group_empty(task));
>>>>> +	rcu_read_unlock();
>>>>> +
>>>>> +	return exited;
>>>>> +}
>>>>
>>>> I'm not sure why you think READ_ONCE was missing.
>>>> It's different in wait_consider_task() where READ_ONCE is needed because
>>>> of multiple checks. Here it's done once.
>>>
>>> In practice it probably has no effect on the generated code.  But
>>> READ_ONCE is about telling the compiler not to be clever.  Don't use
>>> tearing loads or stores etc.  When all of the other readers are using
>>> READ_ONCE I just get nervous if we have a case that doesn't.
>>
>> That's not true. The only place where READ_ONCE(->exit_state) is used is
>> in wait_consider_task() and nowhere else. We had that discussion a while
>> ago where I or someone proposed to simply place a READ_ONCE() around all
>> accesses to exit_state for the sake of kcsan and we agreed that it's
>> unnecessary and not to do this.
>> But it obviously doesn't hurt to have it.
> 
> There is a larger discussion to be had around the proper handling of
> exit_state.
> 
> In this particular case because we are accessing exit_state with
> only rcu_read_lock protection, because the outcome of the read
> is about correctness, and because the compiler has nothing else
> telling it not to re-read exit_state, I believe we actually need
> the READ_ONCE.
> 
> At the same time it would take a pretty special compiler to want to
> reaccess that field in thread_group_exited.
> 
> I have looked through and I don't find any of the other access of
> exit_state where the result is about correctness (so that we care)
> and we don't hold tasklist_lock.
> 
> But I have removed the necessary wording from the commit comment.

Hey Eric, are you planning to push the final version into a topic branch
so it can be pulled into bpf-next as discussed earlier?

Thanks,
Daniel