[Qemu-devel] Preliminary results of unit testing on qemu.git

[Qemu-devel] Preliminary results of unit testing on qemu.git

[Lucas Meneghel Rodrigues]

Hi folks,

I've recently sent a patch porting the -testdev device from qemu-kvm.git 
to qemu.git, and I went ahead and started running the unittests on 
qemu.git. It's still not integrated to our test grid, I just ran it in 
my laptop, which runs a fairly recent linux 3.0 kernel (which in Fedora 
15 they are calling 2.6.40 :))


We had a fair number of failures. Right now, avi's kernel + qemu-kvm.git 
only yields an apic failure, so I guess we need to start investigating 
the failures. If we are going to get qemu-kvm and qemu to merge 
eventually we need to sort out all the issues.

> 08/26 17:12:39 ERROR|       kvm:0102| Test failed: TestFail: Unit tests failed: apic vmexit access hypercall idt_test msr sieve xsave svm kvmclock_test

Due to some disk quota problems with my people.redhat.com page, I could 
not post the full test logs, so if anyone needs it (~600KB .tar.bz2), 
please let me know.

Cheers,

Lucas

Re: [Qemu-devel] Preliminary results of unit testing on qemu.git

[Jan Kiszka]

[-- Attachment #1: Type: text/plain, Size: 1932 bytes --]

On 2011-08-26 23:31, Lucas Meneghel Rodrigues wrote:
> Hi folks,
> 
> I've recently sent a patch porting the -testdev device from qemu-kvm.git
> to qemu.git, and I went ahead and started running the unittests on
> qemu.git. It's still not integrated to our test grid, I just ran it in
> my laptop, which runs a fairly recent linux 3.0 kernel (which in Fedora
> 15 they are calling 2.6.40 :))
> 
> 
> We had a fair number of failures. Right now, avi's kernel + qemu-kvm.git
> only yields an apic failure,

Not for me. Or did you run the test with -no-kvm-irqchip (see below)?

> so I guess we need to start investigating
> the failures. If we are going to get qemu-kvm and qemu to merge
> eventually we need to sort out all the issues.
> 
>> 08/26 17:12:39 ERROR|       kvm:0102| Test failed: TestFail: Unit
>> tests failed: apic vmexit access hypercall idt_test msr sieve xsave
>> svm kvmclock_test

apic:	First error is related to user space APIC exposing a different
        version than the in-kernel APIC. The version is coupled to the
	CPU model, but neither KVM nor QEMU get this right so far. I
	think we already had complaints about the too old version QEMU
	returns but no patches to fix it properly.

	Then there is a self-ipi issue under KVM but not TCG (or
        in-kernel irqchip), need to look closer. Also affects eventinj.

	Finally, this test stumbles over the bogus isa_get_irq in
	testdev.

sieve:	Works but takes a bit longer under TCG.

svm:	Can't test with KVM on Intel, shows one bug and a lockup in TCG
	mode.

All others mentioned above work fine with KVM but have issues under TCG.
I had a brief look at the obvious cases:

kvmclock_test: Can't work as we do not emulate the PV clock in TCG
	mode.

access:	Can't work as TCG's x86 model ignores segment limits
	(long-known limitation).

The others are likely bugs/shortcomings in the TCG model.

Jan


[-- Attachment #2: OpenPGP digital signature --]
[-- Type: application/pgp-signature, Size: 262 bytes --]

Re: [Qemu-devel] Preliminary results of unit testing on qemu.git

[Jan Kiszka]

[-- Attachment #1: Type: text/plain, Size: 863 bytes --]

On 2011-08-27 11:16, Jan Kiszka wrote:
> apic:	First error is related to user space APIC exposing a different
>         version than the in-kernel APIC. The version is coupled to the
> 	CPU model, but neither KVM nor QEMU get this right so far. I
> 	think we already had complaints about the too old version QEMU
> 	returns but no patches to fix it properly.
> 
> 	Then there is a self-ipi issue under KVM but not TCG (or
>         in-kernel irqchip), need to look closer. Also affects eventinj.

This issue is caused by a bug in the user space irqchip chain:
pic_update_irq triggered by the test case's mask_pic_interrupts fails to
reset the CPU IRQ on x86.

I've applied the testdev patch to an ancient i8259&Co. cleanup queue of
mine, and that actually fixes the issue. I guess I should rebase those
patches and post them eventually.

Jan


[-- Attachment #2: OpenPGP digital signature --]
[-- Type: application/pgp-signature, Size: 262 bytes --]