Unable to boot guest on kernel 2.6.29.1 with kvm-84 or kvm-85

DISCUSSION Kenni Lund 2 participants 3 messages May 1, 2009 – May 3, 2009

Subsystems (3)

x86 KVM KVM x86

lore

Share & tools

lore

Kenni Lund

author

2009-05-01 22:52 UTC #0 lore raw

Quoted 11 lines · Avi Kivity [email protected] wrote:

Quoted 1 line · Kenni Lund
Ok, but as I write in my message, I'm using the
KVM modules from the latest> upstream kernel, not the kvm-85 modules.>
Quoted 1 line
According to the KVM download page, http://www.linux-kvm.org/page/Downloads,> any kernel above 2.6.25 should work with the
latest KVM userspace. This has> been true until now in my case, but it breaks with 2.6.29.1 and that's the reason> why I'm posting this bug report.
Quoted 2 lines
Can you try a bisect?

Yes, sorry for the late reply. I did the bisect as requested and it returned the following results:

# bad: [8d7bff2d72660d9d60aa371ae3d1356bbf329a09] Linux 2.6.29.1
# good: [4a6908a3a050aacc9c3a2f36b276b46c0629ad91] Linux 2.6.28
git bisect start 'v2.6.29.1' 'v2.6.28' '--' 'arch/x86/kvm' 'virt/kvm'
# good: [b82091824ee4970adf92d5cd6d57b12273171625] KVM: Prevent trace call into unloaded module text git bisect good b82091824ee4970adf92d5cd6d57b12273171625
# good: [7f59f492da722eb3551bbe1f8f4450a21896f05d] KVM: use cpumask_var_t for cpus_hardware_enabled git bisect good 7f59f492da722eb3551bbe1f8f4450a21896f05d
# good: [19de40a8472fa64693eab844911eec277d489f6c] KVM: change KVM to use IOMMU API
git bisect good 19de40a8472fa64693eab844911eec277d489f6c
# good: [2aaf69dcee864f4fb6402638dd2f263324ac839f] KVM: MMU: Map device MMIO as UC in EPT git bisect good 2aaf69dcee864f4fb6402638dd2f263324ac839f
# good: [682edb4c01e690c7c7cd772dbd6f4e0fd74dc572] KVM: Fix assigned devices circular locking dependency git bisect good 682edb4c01e690c7c7cd772dbd6f4e0fd74dc572
# bad: [f438349efb8247cd0c1d453a4131b1f801bf5691] KVM: VMX: Don't allow uninhibited access to EFER on i386 git bisect bad f438349efb8247cd0c1d453a4131b1f801bf5691
# good: [516a1a7e9dc80358030fe01aabb3bedf882db9e2] KVM: VMX: Flush volatile msrs before emulating rdmsr git bisect good 516a1a7e9dc80358030fe01aabb3bedf882db9e2

----------------
And the final output:

f438349efb8247cd0c1d453a4131b1f801bf5691 is first bad commit commit f438349efb8247cd0c1d453a4131b1f801bf5691

Author: Avi Kivity 
Date:   Thu Mar 26 23:05:03 2009 +0000

KVM: VMX: Don't allow uninhibited access to EFER on i386

upstream commit: 16175a796d061833aacfbd9672235f2d2725df65

vmx_set_msr() does not allow i386 guests to touch EFER, but they can still do so through the default: label in the switch. If they set EFER_LME, they can oops the host.

Fix by having EFER access through the normal channel (which will check for EFER_LME) even on i386.

Reported-and-tested-by: Benjamin Gilbert
Cc: [email protected]
Signed-off-by: Avi Kivity
Signed-off-by: Chris Wright

:040000 040000 cf7848d35c136beee6665e67839080d450977af0 0a39980481dd346306b2ac54dbe916741515f1f1 M arch

----------------

FYI, I also tested 2.6.29.2 and the issue still exists.

Do you need more information?

Best Regards
Kenni Lund

Avi Kivity

2009-05-02 07:52 UTC #1 lore raw

Quoted 56 lines · Kenni Lund

Quoted 7 lines · Avi Kivity [email protected] wrote:
Quoted 2 lines · Kenni Lund
Ok, but as I write in my message, I'm using the
KVM modules from the latest> upstream kernel, not the kvm-85 modules.>
Quoted 9 lines
Quoted 2 lines
According to the KVM download page, http://www.linux-kvm.org/page/Downloads,> any kernel above 2.6.25 should work with the
latest KVM userspace. This has> been true until now in my case, but it breaks with 2.6.29.1 and that's the reason> why I'm posting this bug report.
Quoted 2 lines
Can you try a bisect?
Yes, sorry for the late reply. I did the bisect as requested and it returned the following results:
# bad: [8d7bff2d72660d9d60aa371ae3d1356bbf329a09] Linux 2.6.29.1
# good: [4a6908a3a050aacc9c3a2f36b276b46c0629ad91] Linux 2.6.28 git bisect start 'v2.6.29.1' 'v2.6.28' '--' 'arch/x86/kvm' 'virt/kvm' # good: [b82091824ee4970adf92d5cd6d57b12273171625] KVM: Prevent trace call into unloaded module text
git bisect good b82091824ee4970adf92d5cd6d57b12273171625 # good: [7f59f492da722eb3551bbe1f8f4450a21896f05d] KVM: use cpumask_var_t for cpus_hardware_enabled git bisect good 7f59f492da722eb3551bbe1f8f4450a21896f05d # good: [19de40a8472fa64693eab844911eec277d489f6c] KVM: change KVM to use IOMMU API git bisect good 19de40a8472fa64693eab844911eec277d489f6c # good: [2aaf69dcee864f4fb6402638dd2f263324ac839f] KVM: MMU: Map device MMIO as UC in EPT
git bisect good 2aaf69dcee864f4fb6402638dd2f263324ac839f
# good: [682edb4c01e690c7c7cd772dbd6f4e0fd74dc572] KVM: Fix assigned devices circular locking dependency
git bisect good 682edb4c01e690c7c7cd772dbd6f4e0fd74dc572
# bad: [f438349efb8247cd0c1d453a4131b1f801bf5691] KVM: VMX: Don't allow uninhibited access to EFER on i386
git bisect bad f438349efb8247cd0c1d453a4131b1f801bf5691
# good: [516a1a7e9dc80358030fe01aabb3bedf882db9e2] KVM: VMX: Flush volatile msrs before emulating rdmsr
git bisect good 516a1a7e9dc80358030fe01aabb3bedf882db9e2
----------------
And the final output:
f438349efb8247cd0c1d453a4131b1f801bf5691 is first bad commit commit f438349efb8247cd0c1d453a4131b1f801bf5691
Author: Avi Kivity Date: Thu Mar 26 23:05:03 2009 +0000
KVM: VMX: Don't allow uninhibited access to EFER on i386
upstream commit: 16175a796d061833aacfbd9672235f2d2725df65
vmx_set_msr() does not allow i386 guests to touch EFER, but they can still do so through the default: label in the switch. If they set EFER_LME, they can oops the host.
Fix by having EFER access through the normal channel (which will check for EFER_LME) even on i386.
Reported-and-tested-by: Benjamin Gilbert
Cc: [email protected]
Signed-off-by: Avi Kivity
Signed-off-by: Chris Wright

Quoted 9 lines

:040000 040000 cf7848d35c136beee6665e67839080d450977af0 0a39980481dd346306b2ac54dbe916741515f1f1 M arch
----------------
FYI, I also tested 2.6.29.2 and the issue still exists.
Do you need more information?

Please try the attached patch.

Signature · 1 line

Do not meddle in the internals of kernels, for they are subtle and quick to panic.

arch/x86/kvm/vmx.c +1 -1

index e8a5649..6dcfab1 100644
@@ -905,9 +905,9 @@ static int vmx_get_msr(struct kvm_vcpu *vcpu, u32 msr_index, u64 *pdata)
 	case MSR_GS_BASE:
 		data = vmcs_readl(GUEST_GS_BASE);
 		break;
+#endif
 	case MSR_EFER:
 		return kvm_get_msr_common(vcpu, msr_index, pdata);
-#endif
 	case MSR_IA32_TIME_STAMP_COUNTER:
 		data = guest_read_tsc();
 		break;

Avi Kivity

2009-05-03 10:14 UTC #2 lore raw

Quoted 60 lines · Avi Kivity

Quoted 59 lines · Kenni Lund
Quoted 7 lines · Avi Kivity [email protected] wrote:
Quoted 2 lines · Kenni Lund
Ok, but as I write in my message, I'm using the
KVM modules from the latest> upstream kernel, not the kvm-85 modules.>
Quoted 8 lines
Quoted 2 lines
According to the KVM download page, http://www.linux-kvm.org/page/Downloads,> any kernel above 2.6.25 should work with the
latest KVM userspace. This has> been true until now in my case, but it breaks with 2.6.29.1 and that's the reason> why I'm posting this bug report.
Quoted 1 line
Can you try a bisect?
Yes, sorry for the late reply. I did the bisect as requested and it returned the following results:
# bad: [8d7bff2d72660d9d60aa371ae3d1356bbf329a09] Linux 2.6.29.1
# good: [4a6908a3a050aacc9c3a2f36b276b46c0629ad91] Linux 2.6.28 git bisect start 'v2.6.29.1' 'v2.6.28' '--' 'arch/x86/kvm' 'virt/kvm' # good: [b82091824ee4970adf92d5cd6d57b12273171625] KVM: Prevent trace call into unloaded module text
git bisect good b82091824ee4970adf92d5cd6d57b12273171625 # good: [7f59f492da722eb3551bbe1f8f4450a21896f05d] KVM: use cpumask_var_t for cpus_hardware_enabled git bisect good 7f59f492da722eb3551bbe1f8f4450a21896f05d # good: [19de40a8472fa64693eab844911eec277d489f6c] KVM: change KVM to use IOMMU API git bisect good 19de40a8472fa64693eab844911eec277d489f6c # good: [2aaf69dcee864f4fb6402638dd2f263324ac839f] KVM: MMU: Map device MMIO as UC in EPT
git bisect good 2aaf69dcee864f4fb6402638dd2f263324ac839f
# good: [682edb4c01e690c7c7cd772dbd6f4e0fd74dc572] KVM: Fix assigned devices circular locking dependency
git bisect good 682edb4c01e690c7c7cd772dbd6f4e0fd74dc572
# bad: [f438349efb8247cd0c1d453a4131b1f801bf5691] KVM: VMX: Don't allow uninhibited access to EFER on i386
git bisect bad f438349efb8247cd0c1d453a4131b1f801bf5691
# good: [516a1a7e9dc80358030fe01aabb3bedf882db9e2] KVM: VMX: Flush volatile msrs before emulating rdmsr
git bisect good 516a1a7e9dc80358030fe01aabb3bedf882db9e2
----------------
And the final output:
f438349efb8247cd0c1d453a4131b1f801bf5691 is first bad commit commit f438349efb8247cd0c1d453a4131b1f801bf5691
Author: Avi Kivity Date: Thu Mar 26 23:05:03 2009 +0000
KVM: VMX: Don't allow uninhibited access to EFER on i386
upstream commit: 16175a796d061833aacfbd9672235f2d2725df65
vmx_set_msr() does not allow i386 guests to touch EFER, but they
can still
do so through the default: label in the switch. If they set
EFER_LME, they
can oops the host.
Fix by having EFER access through the normal channel (which will
check for
EFER_LME) even on i386.
Reported-and-tested-by Benjamin Gilbert Cc: [email protected]
Signed-off-by Avi Kivity Signed-off-by: Chris Wright
:040000 040000 cf7848d35c136beee6665e67839080d450977af0 0a39980481dd346306b2ac54dbe916741515f1f1 M arch
----------------

Quoted 8 lines

Quoted 5 lines
FYI, I also tested 2.6.29.2 and the issue still exists.
Do you need more information?
Please try the attached patch.

It won't help - I reproduced the issue. Instead, try passing the parameter '-cpu qemu32' (or '-cpu qemu64,-nx').

Signature · 1 line

error compiling committee.c: too many arguments to function