Re: [PATCH] signal: optimise signal_pending()

[Pavel Begunkov]

On 5/17/21 6:22 PM, Eric W. Biederman wrote:
> Pavel Begunkov <asml.silence@xxxxxxxxx> writes:
> 
>> Optimise signal_pending() by checking both TIF_SIGPENDING and
>> TIF_NOTIFY_SIGNAL at once. Saves quite a bit of generated instructions,
>> e.g. sheds 240B from io_uring alone, some including ones in hot paths.
>>
>>    text    data     bss     dec     hex filename
>>   84087   12414       8   96509   178fd ./fs/io_uring.o
>>   83847   12414       8   96269   1780d ./fs/io_uring.o
> 
> I believe the atomic test_bit is pretty fundamental, especially with
> it's implied barriers.  I believe you are optimizing out the code
> that will makes signal_pending work in a loop.

Hmm, does it? I agree that at least it should volatile, but unlike
set_bit(), which is in atomic.h and has a non atomic __set_bit()
counter part, test_bit() is bitops/non-atomic.h, and I don't see
any implementation of test_bit() or arch_test_bit() having any
barriers.

READ_ONCE() should cover volatile, would it be better?

or test_mask() operating withing one word?

> I have tried looking and I really don't understand why TIF_NOTIFY_SIGNAL
> was added.  Perhaps instead of trying to optimize the test, you should
> optimize by combining TIF_NOTIFY_SIGNAL with TIF_SIGPENDING.

I'm speculating, but it looks to me that TIF_NOTIFY_SIGNAL was
specifically added to not rely on and separate from TIF_SIGPENDING
for task_work notification delivery.

> 
> Perhaps set_notify_signal could be optimized to set both.  I think I
> only see 4 calls in the tree.
> 
>> Signed-off-by: Pavel Begunkov <asml.silence@xxxxxxxxx>
>> ---
>>
>> Suggestions on how to make it less disruptive to abstractions are most
>> welcome, as even the one below fails to generated anything sane because
>> of test_bit()
>>
>> return unlikely(test_ti_thread_flag(ti, TIF_SIGPENDING) |
>> 		test_ti_thread_flag(ti, TIF_SIGPENDING));
>>
>>  include/linux/sched/signal.h | 6 +++---
>>  1 file changed, 3 insertions(+), 3 deletions(-)
>>
>> diff --git a/include/linux/sched/signal.h b/include/linux/sched/signal.h
>> index 3f6a0fcaa10c..97e1963a13fc 100644
>> --- a/include/linux/sched/signal.h
>> +++ b/include/linux/sched/signal.h
>> @@ -361,14 +361,14 @@ static inline int task_sigpending(struct task_struct *p)
>>  
>>  static inline int signal_pending(struct task_struct *p)
>>  {
>> +	struct thread_info *ti = task_thread_info(p);
>> +
>>  	/*
>>  	 * TIF_NOTIFY_SIGNAL isn't really a signal, but it requires the same
>>  	 * behavior in terms of ensuring that we break out of wait loops
>>  	 * so that notify signal callbacks can be processed.
>>  	 */
>> -	if (unlikely(test_tsk_thread_flag(p, TIF_NOTIFY_SIGNAL)))
>> -		return 1;
>> -	return task_sigpending(p);
>> +	return unlikely(ti->flags & (_TIF_SIGPENDING | _TIF_NOTIFY_SIGNAL));
>>  }
>>  
>>  static inline int __fatal_signal_pending(struct task_struct *p)

-- 
Pavel Begunkov