Re: [patch] new spinlock variant, spinlock-2.3.30-A4

Manfred (manfreds@colorfullife.com)
Fri, 3 Dec 1999 23:11:25 +0100

From: Andrea Arcangeli <andrea@suse.de>
> On Tue, 30 Nov 1999, Ingo Molnar wrote:
>
> >no, spin_lock() must guarantee that instructions within the critical
> >section only execute at one CPU at once. Ie. all effects of the critical
>
> Without the lock on the bus reads could happen before the spin_lock().
> Thus instructions into the critical section would execute on two CPUs at
> the same time.
>
Could you try to trigger this problem?

eg cpu1 owns the spinlock:
write DATA
spin_unlock, ie write LOCK_1

cpu2 acquires the lock:
"asm: read LOCK_1; read DATA", the cpu could reorder the read operation.

if cpu2 gets the lock, then it has seen the "write LOCK_1" from cpu1.
but this automagically means that it has also seen the "write DATA" from
cpu1.

It seems that the cpu discards speculative reads if it sees a write
operation on
the system bus (I tried to force a speculative read, and my proggy never
failed)

--
BUT: I don't know if we should rely on this: 15 ticks vs 22 ticks is not a
very large
difference, and I don't know if this behaviour is guaranteed.


Ingo, do your new spin_lock() functions clobber any registers? This would
further
increase the costs in "real life".


Cheers,
    Manfred




-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.rutgers.edu
Please read the FAQ at http://www.tux.org/lkml/