On Thu, Jun 28, 2018 at 09:48:25PM +0200, Greg Kurz wrote: > On Thu, 28 Jun 2018 12:14:25 +0200 > Greg Kurz <gr...@kaod.org> wrote: > > > Since the recent cleanups to hide host configuration details from guests, > > it isn't possible to start an older machine type with HV KVM [*]: > > > > qemu-system-ppc64: KVM doesn't support for base page shift 34 > > > > This basically boils down to the fact that it isn't safe to call > > the kvmppc_hpt_needs_host_contiguous_pages() helper from a class > > init function because: > > - KVM isn't initialized yet, and kvm_enabled() always return false > > in this case. This causes kvmppc_hpt_needs_host_contiguous_pages() > > to do nothing and we end up choosing a 16G default page size > > which is not supported by KVM. > > - even if we drop kvm_enabled() we then have the issue that > > kvmppc_hpt_needs_host_contiguous_pages() assumes CPUs are > > created, which isn't the case either. > > > > The choice was made to initialize capabilities during machine > > init before creating the CPUs, and I don't think we should > > revert to the previous behavior. Let's go forward instead and > > ensure we can retrieve the MMU information from KVM before > > CPUs are created. > > > > To fix this, we first change kvm_get_smmu_info() so that it > > doesn't need a CPU object. This allows to stop using first_cpu > > in kvmppc_hpt_needs_host_contiguous_pages(). Then we delay > > the setting of the default value to machine init time, so > > that we're sure that KVM is fully initialized. > > > > As a bonus, the last patch is a tentative to be able to detect > > such misuse of *_enabled() accelerator helpers earlier. > > > > Please comment. > > > > [*] it also breaks PR KVM actually, but the error is different and > > I need to dig some more. > > > > With current master: > > 1) qemu-system-ppc64 -machine pseries,accel=kvm,kvm-type=PR > > The guest starts but its kernel oopses at some point: > > [ 0.011328] kernel tried to execute exec-protected page (c000000001611244) > -exploit attempt? (uid: 0) > [ 0.011379] Unable to handle kernel paging request for instruction fetch > [ 0.011416] Faulting instruction address: 0xc000000001611244 > [ 0.011453] Oops: Kernel access of bad area, sig: 11 [#1] > [ 0.011482] LE SMP NR_CPUS=1024 NUMA pSeries > [ 0.011512] Modules linked in: > [ 0.011557] CPU: 0 PID: 1 Comm: swapper/0 Not tainted > 4.17.2-200.fc28.ppc64le #1 > [ 0.011600] NIP: c000000001611244 LR: c00000000000acec CTR: > 0000000000000000 > [ 0.011643] REGS: c00000003fffba90 TRAP: 0400 Not tainted > (4.17.2-200.fc28.ppc64le) > [ 0.011694] MSR: b000000010001033 <SF,HV,ME,IR,DR,RI,LE> CR: 28000848 > XER: 20000000 > [ 0.011741] CFAR: 0000000000000000 SOFTE: 1 > [ 0.011741] GPR00: 0000000000000000 c00000003fffbd10 c000000001570b00 > c00000003fffbd80 > [ 0.011741] GPR04: c000000000034418 0000000048000000 000000000000000a > 000000004aa21de8 > [ 0.011741] GPR08: 000000007d410164 0000000000000000 0000000000000002 > 0000000000000900 > [ 0.011741] GPR12: b000000002009033 c000000001840000 c000000000071a2c > 00000000495de1a4 > [ 0.011741] GPR16: 0000000000000078 c00000000160fd10 c000000000e705e0 > 000000007c1b03a6 > [ 0.011741] GPR20: 000000007c1ffaa6 c0000000016125b8 c0000000014253e8 > 000000007c1303a6 > [ 0.011741] GPR24: 000000007c1643a6 000000007c1a03a6 c00000000160fd08 > ffffffffebc0f008 > [ 0.011741] GPR28: ffffffffebc0f000 c0000000000345d8 c0000000000345d8 > 0000000000000000 > [ 0.012138] NIP [c000000001611244] kvm_tmp+0x1534/0x100000 > [ 0.012170] LR [c00000000000acec] soft_nmi_common+0xcc/0xd0 > [ 0.012199] Call Trace: > [ 0.012214] Instruction dump: > [ 0.012236] XXXXXXXX XXXXXXXX XXXXXXXX XXXXXXXX XXXXXXXX XXXXXXXX XXXXXXXX > XXXXXXXX > [ 0.012289] XXXXXXXX XXXXXXXX XXXXXXXX XXXXXXXX XXXXXXXX XXXXXXXX XXXXXXXX > XXXXXXXX > [ 0.012334] ---[ end trace d2ee28832d481d2d ]--- > [ 0.012362] > [ 1.012387] kernel tried to execute exec-protected page (c000000001611808) > -exploit attempt? (uid: 0) > [ 1.012433] Unable to handle kernel paging request for instruction fetch > [ 1.012468] Faulting instruction address: 0xc000000001611808 > [ 1.012504] Oops: Kernel access of bad area, sig: 11 [#2] > [ 1.012532] LE SMP NR_CPUS=1024 NUMA pSeries > [ 1.012561] Modules linked in: > [ 1.012583] CPU: 0 PID: 1 Comm: swapper/0 Tainted: G D > 4.17.2-200.fc28.ppc64le #1 > [ 1.012641] NIP: c000000001611808 LR: c0000000001247fc CTR: > c000000001840000 > [ 1.012684] REGS: c00000003fffb5d0 TRAP: 0400 Tainted: G D > (4.17.2-200.fc28.ppc64le) > [ 1.012740] MSR: b000000010001033 <SF,HV,ME,IR,DR,RI,LE> CR: 48000224 > XER: 20000000 > [ 1.012785] CFAR: 0000000000000000 SOFTE: 0 > [ 1.012785] GPR00: c0000000001247fc c00000003fffb850 c000000001570b00 > 0000000000000000 > [ 1.012785] GPR04: 0000000000000000 c0000000fe9e4900 fffffffffffffffd > c0000000fe9e4900 > [ 1.012785] GPR08: 00000000fed50000 b000000000001033 0000000000000009 > c00000003fffb55f > [ 1.012785] GPR12: 0000000000000000 c000000001840000 c000000000071a2c > 00000000495de1a4 > [ 1.012785] GPR16: 0000000000000078 c00000000160fd10 c000000000e705e0 > 000000007c1b03a6 > [ 1.012785] GPR20: 000000007c1ffaa6 c0000000016125b8 c0000000014253e8 > 000000007c1303a6 > [ 1.012785] GPR24: 000000007c1643a6 000000007c1a03a6 c00000000160fd08 > ffffffffebc0f008 > [ 1.012785] GPR28: 0000000000000000 000000000000000b 000000000000000b > c0000000fe9e4900 > [ 1.013166] NIP [c000000001611808] kvm_tmp+0x1af8/0x100000 > [ 1.013196] LR [c0000000001247fc] do_exit+0x12c/0xd30 > [ 1.013224] Call Trace: > [ 1.013238] Instruction dump: > [ 1.013260] XXXXXXXX XXXXXXXX XXXXXXXX XXXXXXXX XXXXXXXX XXXXXXXX XXXXXXXX > XXXXXXXX > [ 1.013303] XXXXXXXX XXXXXXXX XXXXXXXX XXXXXXXX XXXXXXXX XXXXXXXX XXXXXXXX > XXXXXXXX > [ 1.013348] ---[ end trace d2ee28832d481d2e ]--- > [ 1.013375] > [ 2.013391] Fixing recursive fault but reboot is needed! > > and the guest gets unresponsive.
Huh, that's a bit weird.
> 2) qemu-system-ppc64 -machine pseries-2.12,accel=kvm,kvm-type=PR
>
> prints an error message and terminates right away:
>
> qemu-system-ppc64: KVM doesn't support page shift 24/12
>
> This error is expected: since PR KVM doesn't set KVM_PPC_PAGE_SIZES_REAL,
> ie, we choose to support all possible page sizes, but PR KVM doesn't
> support this page shift combination indeed. Unsurprisingly we get the
> same error with:
>
> -machine pseries,accel-kvm,kvm-type=PR,cap-hpt-max-page-size=${pagesize}
>
> if ${pagesize} is >= 16m. This is the result of PR KVM not supporting
> MPSS at all, even though it supports 16m pages in a 16m segment. We
> cannot really fix this in QEMU, unless we completely filter out MPSS
> in spapr_pagesize_cb() but I'm pretty sure we don't want that. :)
Yeah. I think sacrificing PR without special options (or fixing PR)
is the price we have to pay for sane behaviour otherwise here.
> But then, if we go for a 64k limit, we hit 1).
>
> An obvious change in the DT since the page size cleanup is:
>
> [4k seg [4k pg]] [64k seg [64k pg]] [16m
> seg [16m pg]]
> - ibm,segment-page-sizes = <0xc 0x0 0x1 0xc 0x0 0x10 0x110 0x1 0x10 0x1 0x18
> 0x100 0x1 0x18 0x0>;
> + ibm,segment-page-sizes = <0xc 0x0 0x1 0xc 0x0 0x10 0x110 0x1 0x10 0x1>;
> [4k seg [4k pg]] [64k seg [64k pg]]
>
> If I add the 16m entry back, the guest boots just fine.
>
> Not sure yet what's happening... any idea ?
No, not sure why lacking 16m pages would break PR.
--
David Gibson | I'll have my music baroque, and my code
david AT gibson.dropbear.id.au | minimalist, thank you. NOT _the_ _other_
| _way_ _around_!
http://www.ozlabs.org/~dgibson
signature.asc
Description: PGP signature
