On Fri, 28 Jul 2017 07:35:34 +0200 Cédric Le Goater <c...@kaod.org> wrote:
> On 07/28/2017 05:40 AM, David Gibson wrote: > > On Fri, Jul 28, 2017 at 01:27:05PM +1000, Alexey Kardashevskiy wrote: > >> On 28/07/17 02:39, Greg Kurz wrote: > >>> On Wed, 26 Jul 2017 17:31:17 -0300 > >>> Daniel Henrique Barboza <danie...@linux.vnet.ibm.com> wrote: > >>> > >>>> I've tested the patch set using Greg's Github branch. It worked fine in > >>>> my tests > >>>> using a Fedora 26 and an Ubuntu 17.04 guests. I have two observations > >>>> though: > >>>> > >>>> 1 - This is not related to this patch set per se because it is > >>>> reproducible on master, but > >>>> I think it is interfering with this new feature. There is a > >>>> warning/error message in > >>>> the kernel right after SLOF that goes: > >>>> > >>>> (...) > >>>> -> smp_release_cpus() > >>>> spinning_secondaries = 0 > >>>> <- smp_release_cpus() > >>>> Linux ppc64le > >>>> #1 SMP Mon Jul 1[ 0.030450] pci 0000:00:02.0: of_irq_parse_pci: > >>>> failed with rc=-22 > >>>> [ 0.030552] pci 0000:00:0f.0: of_irq_parse_pci: failed with rc=-22 > >>>> [ OK ] Started Security Auditing Service. > >>>> (...) > >>>> > >>> > >>> This is a regression in QEMU master introduced by this commit: > >>> > >>> commit b87680427e8a3ff682f66514e99a8344e7437247 > >>> Author: Cédric Le Goater <c...@kaod.org> > >>> Date: Wed Jul 5 19:13:15 2017 +0200 > >>> > >>> spapr: populate device tree depending on XIVE_EXPLOIT option > >>> > >>> When XIVE is supported, the device tree should be populated > >>> accordingly and the XIVE memory regions mapped to activate MMIOs. > >>> > >>> Depending on the design we choose, we could also allocate different > >>> ICS and ICP objects, or switch between objects. This needs to be > >>> discussed. > >>> > >>> Signed-off-by: Cédric Le Goater <c...@kaod.org> > >>> Signed-off-by: David Gibson <da...@gibson.dropbear.id.au> > >>> > >>> It is very similar to the issue that motivated the new > >>> KVMPPC_H_UPDATE_PHANDLE > >>> hcall (see patch 24 and 26 in this series): > >>> > >>> - QEMU creates an "interrupt-controller" node with a phandle property > >>> with the value 0x1111 > >>> - QEMU passes the FDT to SLOF > >>> - SLOF converts all references to the phandle to an SLOF internal value > >>> > >>> => from now on (ie, until the next machine reset), the guest only knows > >>> the OF phandle. > >>> > >>> - during CAS, if we go XICS, we send back an updated FDT with the > >>> phandle of the "interrupt-controller" node reverted to 0x1111 > >>> > >>> => the guest complains because all cold-plugged devices nodes refer > >>> to the OF phandle, not 0x1111 > >>> > >>> A solution is to use the value set by KVMPPC_H_UPDATE_PHANDLE during CAS > >>> instead of 0x1111. I could verify it makes the guest warning disappear. > >>> > >>> I'll send a dedicated patchset to fix this in 2.10. > >> > >> > >> The SLOF I pushed for 2.10 does not have it though. And the rest of XIVE is > >> not targeted for 2.10 anyway. So imho the solution is reverting "spapr: > >> populate device tree depending on XIVE_EXPLOIT option" for 2.10. > > > > I agree, I've applied the revert to ppc-for-2.10 now. > > Sure. > > Greg, have you found a solution to get around this problem or do we need > to populate the device tree with the interrupt controller node for SLOF ? > With commit b87680427e8a, spapr_dt_xics() is called twice actually: first time during machine reset and a second time during CAS. The first call means we start in XICS mode by default. But I don't understand why it is also called during CAS: if the guest didn't advertise XIVE, then we don't have anything to do, do we ? Cheers, -- Greg > Thanks, > > C. >
pgpt5h_2PFdmq.pgp
Description: OpenPGP digital signature
