Re: [RFC 09/15] PM / Hibernate: user, implement user_ops writer

[Jiri Slaby]

On 03/25/2010 11:14 PM, Rafael J. Wysocki wrote:
> On Wednesday 24 March 2010, Jiri Slaby wrote:
>> On 03/24/2010 09:42 PM, Pavel Machek wrote:
>>>> +	if (test_bit(TODO_CLOSED, to_do_flags))
>>>> +		return -EIO;
>>>> +
>>>> +	to_do_buf = buf;
>>>> +	wmb();
>>>> +	set_bit(TODO_WORK, to_do_flags);
>>>> +	wake_up_interruptible(&to_do_wait);
>>>
>>> Uhuh, open-coded barriers... these need to be commented, and I guess
>>> you just should not play this kind of trickery.
>>
>> It's just to ensure the to_do_buf store is not reordered with the
>> set_bit. I wanted to avoid locks as too heavy tools here.
> 
> No, please use them, at least in a prototype version.
> 
> We can always optimize things out later, but doing optimizations upfront
> doesn't really work well from my experience.
> 
> So, if you'd use a lock somewhere, please use it, or maybe use a completion if
> that fits the design better.

That's it, I don't think a lock is appropriate here (I didn't even think
of that) -- I don't know what to lock (OK, I see it, but it's not that
clear). There is no potential for race per se, I only need to disable
reordering (which locks do as a side-effect). I need the steps to be
done in the A-B order where there is a barrier appropriate. Here, A is
store to to_do_buf, B is set_bit. It's I set to_do_buf, flag that it may
be used, the consumer will see the flag and use to_do_buf, in this order.

Above that if I introduce locks the wait_event on the other side will
grow into an unreadable mess. I would need to hold a lock when checking
the condition and hold it until I reach to_do_buf use, but also unlock
it on all paths that do not reach that point. Yeah, it's indeed doable,
but I don't think, it will improve things.

I also don't think completion is appropriate here, as I have a condition
to check for and it differs over wake_up sites.

thanks,
-- 
js
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/