On 07/19/2013 05:11 PM, George Spelvin wrote:
What I have in mind is to have 2 separate rwlock initializers - one for
fair and one for reader-bias behavior. So the lock owners can decide
what behavior do they want with a one line change.
That's definitely a nicer patch, if it will work. I was imagining that,
even for a single (type of) lock, only a few uses require reader bias
(because they might be recursive, or are in an interrupt), but you'd
want most read_lock sites to be fair.
Yes, fair rwlock will be the default.
Deciding on a per-lock basis means that one potentially recursive call
means you can't use fair queueing anywhere.
I was hoping that the number of necessary unfair calls would
be small enough that making the read_lock default fair and
only marking the unfair call sites would be enough.
But I don't really know until doing a survey of the calls.
I think so. The queue read/write lock, if merged, will be an optional
feature for people to try out to see if they see any problem in any of
the existing rwlock. So far, I didn't encounter any problem in my testing.
BTW, I also tried my version of the rwlock without the waiting queue. In
high contention case, it performs slightly better than the
__read_lock_failed changes suggested by Ingo at least for the
reader-bias one. It is still not as good as the full version with the
waiting queue. I should be able to provide more performance data next week.
Regards,
Longman
--
To unsubscribe from this list: send the line "unsubscribe linux-arch" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
