Re: [PATCH 00/20] x86: ticket lock rewrite and paravirtualization

[Jeremy Fitzhardinge <jeremy@goop.org>]

On 11/15/2010 12:14 PM, Peter Zijlstra wrote:
> On Mon, 2010-11-15 at 12:03 -0800, H. Peter Anvin wrote:
>> On 11/15/2010 12:00 PM, Jeremy Fitzhardinge wrote:
>>> Another approach I discussed with PeterZ and Mathieu is to steal the LSB
>>> of the ticket counters (halving the max CPU count) to use as a "there is
>>> someone in slowpath waiting on this lock".  But I haven't spent the time
>>> to work out an algorithm to maintain that flag (or flags, since there
>>> are bits available) in a correct and efficient way.
>>>
>> Definitely worth pondering.
> Right, so the idea was to make the ticket increment 2, which would leave
> the LSB of both the head and tail available. I think that if one were to
> set both (using a cmpxchg), the ticket fast-path wouldn't need any
> changes since head==tail is still the correct condition for acquisition.
>
> Then the unlock needs an added conditional:
>   if (tail & 1) 
> 	unlock_slowpath()

The tricky part is knowing how to clear the bit(s) on the last person
dropping out of the slow path, and making that race-free with respect to
new lockers entering the slow path.  I guess you could leave it in
slowpath state until you're the last unlocker (ie, you're unlocking into
uncontended state), whereupon you also clear the bits; I guess that
would probably need a cmpxchg to make it safe WRT new lockers entering
slowpath.

As a heuristic, it shouldn't be too bad performancewise, since
(handwaving) if ticketholder N has entered the slowpath, then its likely
that N+1 will as well.

    J