On 20.09.2016 23:45, Benjamin Herrenschmidt wrote: > On Tue, 2016-09-20 at 13:44 +0200, Thomas Huth wrote: >> >> Seems like KVM PR is using the "degraded" ISA variants (without the >> 1TB >> segments), but the new POWERPC_MMU_64K flag has not been added to >> those. >> Has this been done on purpose, or was this just by accident? >> I can make KVM PR working again with the following patch: >> >> diff --git a/target-ppc/cpu-qom.h b/target-ppc/cpu-qom.h >> index 2864105..36694cb 100644 >> --- a/target-ppc/cpu-qom.h >> +++ b/target-ppc/cpu-qom.h >> @@ -81,6 +81,7 @@ enum powerpc_mmu_t { >> | POWERPC_MMU_AMR | 0x00000003, >> /* Architecture 2.06 "degraded" (no 1T segments) */ >> POWERPC_MMU_2_06a = POWERPC_MMU_64 | POWERPC_MMU_AMR >> + | POWERPC_MMU_64K >> | 0x00000003, >> /* Architecture 2.07 variant */ >> POWERPC_MMU_2_07 = POWERPC_MMU_64 | POWERPC_MMU_1TSEG >> @@ -88,6 +89,7 @@ enum powerpc_mmu_t { >> | POWERPC_MMU_AMR | 0x00000004, >> /* Architecture 2.07 "degraded" (no 1T segments) */ >> POWERPC_MMU_2_07a = POWERPC_MMU_64 | POWERPC_MMU_AMR >> + | POWERPC_MMU_64K >> | 0x00000004, >> }; >> >> However, not sure whether this is the right fix ... Cédric, Ben, any >> ideas? > > Oh I thought I had removed the degraded variants ... Definitely looks > like an accident. I *think* PR KVM supports 64K pages, no ? If not, > then we shouldn't enable the flag.. somebody needs to check the kernel.
I've now added some debug printf statements to kvm_fixup_page_sizes() in QEMU, and it seems that at least my current downstream kernel does not report support for 64K page sizes. So the right fix is likely to disable the POWERPC_MMU_64K bit there in env->mmu_model if the kernel does not report support for 64K pages. I'll try to come up with a patch... But actually, now I wonder why my kernel does not support this page size. There was a kernel patch from Paul (a4a0f2524acc2c6 - "KVM: PPC: Book3S PR: Allow guest to use 64k pages") which added this page size to KVM-PR, so IMHO it should work ... Seem like I need to do some more debugging here... Thomas