Re: [PATCH 07/13] task_work: Introduce task_work_pending

[Eric W. Biederman]

Jens Axboe <axboe@xxxxxxxxx> writes:

> On 3/9/22 9:24 AM, Eric W. Biederman wrote:
>> diff --git a/include/linux/task_work.h b/include/linux/task_work.h
>> index 5b8a93f288bb..897494b597ba 100644
>> --- a/include/linux/task_work.h
>> +++ b/include/linux/task_work.h
>> @@ -19,6 +19,11 @@ enum task_work_notify_mode {
>>  	TWA_SIGNAL,
>>  };
>>  
>> +static inline bool task_work_pending(struct task_struct *task)
>> +{
>> +	return READ_ONCE(task->task_works);
>> +}
>> +
>
> Most of the checks for this is current, do we need READ_ONCE here?

There is a non-current use in fs/io_uring in __io_uring_show_fdinfo
and another in task_work_cancel_match.

Beyond that there are quite a few writes that are not at all from
current so even on current task->task_works can change if you look
twice.

So if only to keep it from making unwarranted assumptions I think
READ_ONCE makes sense.

Given that READ_ONCE is practically free I don't see where there is
any harm in using it to document the kind of code we expect the compiler
to generate.

Looking a second time I see all of the other reads of task->task_works
are already READ_ONCE in kernel/task_work.c, so really I think if we
don't want READ_ONCE we need a big fat comment about why it is safe
in a check like task_work_pending and while it is needed everywhere
else.  At the moment I am not smart enough to write that comment.

I will see about adding this bit of discussion in the commit comment to
make it a little clearer why I am introducing READ_ONCE.

Eric