Re: [PATCH] x86: Document the hotplug code is incompatible with x86 irq handling

From: Rafael J. Wysocki
Date: Tue Jun 12 2007 - 18:52:19 EST


On Wednesday, 13 June 2007 00:24, Siddha, Suresh B wrote:
> On Wed, Jun 13, 2007 at 12:16:08AM +0200, Rafael J. Wysocki wrote:
> > On Tuesday, 12 June 2007 23:56, Siddha, Suresh B wrote:
> > > On Tue, Jun 12, 2007 at 10:52:09PM +0200, Rafael J. Wysocki wrote:
> > > > On Tuesday, 12 June 2007 20:19, Eric W. Biederman wrote:
> > > > > Because you are calling unfixably broken code. That should be a decent
> > > > > incentive to do something else won't it?
> > > >
> > > > Can you please tell me _what_ else can be done?
> > > >
> > > > > IOAPICs do not support what the code is doing here. There is lots of
> > > > > practical evidence including bad experiences and practical tests that
> > > > > support this.
> > > >
> > > > Well, AFAICS, Suresh has tried to debug one failing case recently without
> > > > any consistent conclusions. I don't know of any other test cases (links,
> > > > please?).
> > >
> > > Rafael, Darrick Wong's issue looks different and hence I was motivated to
> > > look and fix if it was a SW issue. For now, I am not able to comprehend
> > > what is happening on Darrick Wong's system. Need more help from Darrick
> > > as he has the golden failing system.
> > >
> > > Meanwhile I talked to our hardware folks about the irq migration in general.
> > >
> > > One good news is that future versions of chipsets will have an Interrupt
> > > Remapping feature(for more details please refer to
> > > http://download.intel.com/technology/computing/vptech/Intel(r)_VT_for_Direct_IO.pdf)
> > > where in we can reliably migrate the irq to someother cpu in the process
> > > context.
> > >
> > > And for the existing platforms, chipset folks don't see a reason why the
> > > Eric's algorithm(specified below) should fail.
> > >
> > > Eric's algorithm for level triggered(edge triggered should be easier than
> > > level triggered):
> > > - Mask the irq in process context.
> > > - Poll RIRR until an ack of for the irq was not pending.
> > > - Migrate the irq.
> > >
> > > Eric had a problem of stuck remote IRR on E75xx chipset with this algorithm
> > > and my next step is to reproduce this issue on this platform and understand
> > > the behavior.
> >
> > OK
> >
> > In that case, do I understand correctly that we are going to implement the
> > Eric's algorithm above for the CPU hotunplugging on x86 once you've figured
> > out what's the E75xx issue?
>
> That is the idea. Thanks.

OK, thanks.

Eric, would you agree to follow this plan without making the entire CPU hotplug
(on x86) depend on BROKEN?

Rafael


--
"Premature optimization is the root of all evil." - Donald Knuth
-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/