[lkml]   [2004]   [Sep]   [12]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
Messages in this thread
SubjectRe: [PATCH] Yielding processor resources during lock contention

On Sun, 12 Sep 2004, Zwane Mwaikambo wrote:
> On i386 processors with PNI this is achieved by using the
> monitor/mwait opcodes to halt the processor until a write to the lock is
> done.

I'd seriously suggest you ask Intel for an official opinion on this. Last
I heard (and that was, I believe, before monitor/mwait had been officially
announced, so it's certainly a bit dated now) it wasn't architecturally
clear that it's a good idea using it for things like spinlocks.

In particular, if the CPU idly waits for a cacheline to be dirtied, it is
entirely possible that the other CPU that owns the lock and releases it
won't actually _tell_ the world that the lock has been released for quite
some time. After all, why should it - if it is the exclusive owner, and it
sees no memory traffic on the bus, it may have no reason to push out the
fact that it just released the lock. Just keep it dirty in its caches.

In other words: monitor/mwait on purpose obviously causes fewer bus
cycles. But that very fact may well mean (at least in theory) that you get
very high latencies. It could make spinlock contention very very unfair
(the original CPU keeps getting the lock over and over again, while the
monitor/mwait one never gets to play), and it might also make ping-pong
locking latency be extremely high.

Also, it's entirely possible that monitor/mwait ends up shutting down the
CPU to the point that getting out of a lower-power mode might have a
latency of microseconds or even milliseconds. External (or even internal)
voltage regulators and frequency changes are not instantaneous by any

In other words, I would strongly suggest you _not_ really consider this a
serious thing (feel free to play around with it and try to get some
numbers) with this without getting an answer from Intel about what the
_architected_ behaviour of monitor/mwait is, from a latency standpoint.

Because otherwise you might find that even if it performs like you want it
to on some existing machine, next year it would suck so badly that it's
not even funny.

To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to
More majordomo info at
Please read the FAQ at

 \ /
  Last update: 2009-11-18 23:46    [W:0.101 / U:5.960 seconds]
©2003-2018 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site