Re: [Lse-tech] Priority access to NUMA-aware locks

[Kanoj S. <kan...@ya...>]

--- Paul McKenney <Pau...@us...> wrote:
> At OLS, Nick passed on  a question from John as to
> whether the NUMA-aware
> locks (ported from ptx by Jack Vogel and Swami
> Sivasubramanian) could
> support high-priority access, so that a
> high-priority request would get the
> lock next, rather than having to wait for its turn.
> 

Hmm, spinlocks (and by extension, numalocks) should
not need to be priority aware. Imagine a simple SMP
machine that implements spinlocks using bus based
schemes. I haven't seen any hardware implementation
that allows priority. Also, these locks are meant to
be held for a short time, so the extra overhead of
doing priority, boosts etc, are too costly. If you
really wanted a priority based lock, why not add in 
the concept of priority to sleeping locks, which 
themselves are built on spinlocks/numalocks? The only
drawback with this scheme is you can not order
(prioritywise) normal process context and intr
context,
unless you can push in most of isr code into sleepable
regions (bottom halves etc).

Kanoj

__________________________________________________
Do You Yahoo!?
Make international calls for as low as $.04/minute with Yahoo! Messenger
http://phonecard.yahoo.com/