On Fri, May 09, 2014 at 09:19:32PM -0400, Waiman Long wrote: > On 05/08/2014 03:06 PM, Peter Zijlstra wrote: > >On Wed, May 07, 2014 at 11:01:37AM -0400, Waiman Long wrote: > >>If unfair lock is supported, the lock acquisition loop at the end of > >>the queue_spin_lock_slowpath() function may need to detect the fact > >>the lock can be stolen. Code are added for the stolen lock detection. > >> > >>A new qhead macro is also defined as a shorthand for mcs.locked. > >NAK, unfair should be a pure test-and-set lock. > > I have performance data showing that a simple test-and-set lock does not > scale well. That is the primary reason of ditching the test-and-set lock and > use a more complicated scheme which scales better. Nobody should give a fuck about scalability in this case anyway. Also, as I explained/asked earlier: lkml.kernel.org/r/20140314083001.GN27965@xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx Lock holder preemption is _way_ worse with any kind of queueing. You've not explained how the simple 3 cpu example in that email gets better performance than a test-and-set lock. > Also, it will be hard to > make the unfair test-and-set lock code to coexist nicely with PV spinlock > code. That's just complete crap as the test-and-set lock is like 3 lines of code.
Attachment:
pgpmUZtoXavpL.pgp
Description: PGP signature
_______________________________________________ Virtualization mailing list Virtualization@xxxxxxxxxxxxxxxxxxxxxxxxxx https://lists.linuxfoundation.org/mailman/listinfo/virtualization
