Re: [Xen-devel] Bug: Windows 2003 fails to install on xen-unstable tip

[George Dunlap <george.dunlap@xxxxxxxxxxxxx>]

On 23/04/13 15:55, Jan Beulich wrote:
> > > > On 23.04.13 at 16:21, George Dunlap <george.dunlap@xxxxxxxxxxxxx> wrote:
> > On 23/04/13 14:00, Jan Beulich wrote:
> > > > > > On 23.04.13 at 13:52, George Dunlap <george.dunlap@xxxxxxxxxxxxx> wrote:
> > > > On 23/04/13 12:33, Jan Beulich wrote:
> > > Just went through that change again: The only thing that changes
> > > is that while rtc_periodic_cb() (simply setting REG_C flags) got
> > > called at the end of pt_intr_post(), rtc_periodic_interrupt() now
> > > gets called from pt_update_irq(), and therefore takes care of
> > > asserting the IRQ itself (which originally happened inside
> > > pt_update_irq()) along with setting REG_C flags. Bottom line -
> > > all the patch changes is when exactly REG_PF (and possibly
> > > REG_IRQF) get set, and whether the IRQ actually gets asserted.
> > So that last bit seems to do these things:
> > a. Changes when the handling happens in the Xen interrupt handler
> > b. Causes assert/deassert to only happen if !(C.PF) && (B.PIE)
> >     (Before it happened unconditionally)
> > c. Causes C.IRQF=1 only if (same condition above)
> >    (Before it happened unconditionally)
> > d. Runs destroy_periodic_timer() if C.PF already
> >
> > So just to figure out what it was that w2k3 wanted, I tried a bunch of
> > variations:
> >
> > * All of above
> >    BAD
> > * a only; always assert/deassert + set C.IRQF
> >    GOOD
> > * always assert/deassert, but leave destroy_periodic_timer and IRQF
> > setting alone
> >    FAIL
> > * destroy if C.PF, assert/deassert, set IRQF if setting C.PF (don't
> > check B.PIE)
> >    FAIL
> > * destroy if C.PF, always assert/deassert + set C.IRQF
> >    FAIL
> > * never destroy, assert/deassert + set C.IRQF if setting C.PF
> >    FAIL
> > * never destroy, assert/deassert + set C.IRQF if !C.IRQF
> >    FAIL
> >
> > In short, it seems that w2k3 basically expects an unlimited number of
> > attempts to actually deliver the interrupt.
> Not always doing the deassert/assert pair was actually part of Tim's
> subsequent changes - before that it got called conditionally upon
> B.PIE, but always set C.IRQF and always deasserted and asserted.
> Since we know from the debug log that B.PIE is set, I fail to see how
> the code prior to "x86/hvm: Centralize and simplify the RTC IRQ logic"
> would have not worked, but the above case turned out GOOD.
>
> So did his earlier 3 changes perhaps fix the issue, and the fourth re-
> introduced it?
If you've got a git hash I can try to revert it.

 -George

_______________________________________________
Xen-devel mailing list
Xen-devel@xxxxxxxxxxxxx
http://lists.xen.org/xen-devel