Re: Fix for spllower()

[Manuel Bouyer <bouyer%antioche.eu.org@localhost>]

On Wed, May 03, 2006 at 11:44:52AM +0200, Mathieu Ropert wrote:
> Hi,
> 
> I've noticed a potential bug in the spllower() implementation for Xen:
> the original code saves flags through read_psl(), disable interrupts 
> ('cli' on x86) and then even run Xspllower() (if pending interrupts) or 
> set new level and restore flags (with write_psl()).
> On x86, saving and restoring EFLAGS includes restoring the 'IF' flag (if 
> it was previously set) cleared by disable_intr(), thus reenabling 
> interrupts.
> As the 'IF' flag isn't virtualized on Xen, we may end spllower() without 
> reenabling interrupts.

On Xen, read_psl() and write_psl() don't change the EFLAGS register, but
read or write the evtchn_upcall_mask shared variable (see include/cpufunc.h).
So the write_psl() call will really reenable interrupts if needed. I can't
see a problem here.


-- 
Manuel Bouyer <bouyer%antioche.eu.org@localhost>
     NetBSD: 26 ans d'experience feront toujours la difference
--