On Tue, Jul 18, 2017 at 3:50 PM, Andrew Jones <drjones@xxxxxxxxxx> wrote:
> On Tue, Jul 18, 2017 at 03:31:03PM +0200, Christoffer Dall wrote:
>> On Tue, Jul 18, 2017 at 03:23:24PM +0200, Andrew Jones wrote:
>> > On Tue, Jul 18, 2017 at 03:01:53PM +0200, Christoffer Dall wrote:
>> > > On Tue, Jul 18, 2017 at 02:09:57PM +0200, Andrew Jones wrote:
>> > > > On Thu, Jul 13, 2017 at 09:20:09PM +0200, Christoffer Dall wrote:
>> > > > > Rearrange the code to be able to reuse as much as posible and add
>> > > > > support for testing the physical timer as well.
>> > > > >
>> > > > > Also change the default unittests configuration to run a separate vtimer
>> > > > > and ptimer test so that older kernels without ptimer support just show a
>> > > > > failure.
>> > > >
>> > > > We could run tests for both the ptimer and vtimer in a single execution,
>> > > > rather than splitting them and requiring the input, because the read of
>> > > > cntp_ctl_el0 will predictably cause an UNKNOWN exception. Also, by
>> > > > applying the errata framework we can ensure that if we expect the read
>> > > > to work, i.e. the host kernel is recent enough, then, if we still get
>> > > > an UNKNOWN exception, we can report FAIL instead of SKIP. Below is an
>> > > > add on patch that makes the conversion. Let me know what you think.
>> > >
>> > > The problem with this patch, is that we then report SKIP instead of
>> > > FAIL, when we regress the kernel and actually break physical counter
>> > > access (which I did while developing my series).
>> > >
>> >
>> > Well, as long as the cntp_ctl_el0 read isn't regressed into generating
>> > an unknown exception, then the ptimer tests will always be run, reporting
>> > failures as they should.
>> >
>>
>> And that is exactly what we've done a couple of times around, because
>> VHE changes the layout of the trap control register to EL2, and the way
>> we handle traps to KVM of the physical counter register is to inject an
>> undefined exception...
>>
>> > > I think something like this should be discovered by way of capabilities
>> > > or hardcoding a kernel version.
>> >
>> > That's possible already by making one more change (which I should have
>> > made in the first place)
>> >
>> > diff --git a/errata.txt b/errata.txt
>> > index 5608a308ce7c9..8859d4f1d3860 100644
>> > --- a/errata.txt
>> > +++ b/errata.txt
>> > @@ -4,4 +4,5 @@
>> > #---------------:-----------------------:--------------------------------------
>> > 9e3f7a296940 : 4.9 : arm64: KVM: pmu: Fix AArch32 cycle counter access
>> > 6c7a5dce22b3 : 4.12 : KVM: arm/arm64: fix races in kvm_psci_vcpu_on
>> > +7b6b46311a85 : 4.11 : KVM: arm/arm64: Emulate the EL1 phys timer registers
>> > #---------------:-----------------------:--------------------------------------
>> >
>> > With that change, when the test runtime system detects it's running on
>> > a host with at least a 4.11 kernel, then it will automatically set
>> > ERRATA_7b6b46311a85=y (unless overridden by the user). Having that
>> > errata set will even ensure the cntp_ctl_el0 read is tested.
>> >
>>
>> ah, ok then that makes perfect sense.
>>
>> I'm a little confused about the logic though, if we regress the physical
>> counter access on a newer kernel in a way that gives you an undefined
>> exception, will we get FAIL or SKIP?
>>
>> We should get FAIL.
>
> We'll get FAIL as long as the user doesn't override ERRATA_7b6b46311a85
> to be 'n'. See the if-else in set_ptimer_unsupported()
>
> if (ptimer_unsupported && !ERRATA(7b6b46311a85)) {
> report_skip("Skipping ptimer tests. Set ERRATA_7b6b46311a85=y to enable.");
> } else if (ptimer_unsupported) {
> report("ptimer: read CNTP_CTL_EL0", false);
> report_info("ptimer: skipping remaining tests");
> }
so report("...", false); means FAIL?
If so, ok :)
>
> I was guessing we'd want to skip the remaining tests if we can't
> even read the register, but it could be reworked to try every
> ptimer test when ERRATA_7b6b46311a85=y as well.
>
It makes sense to skip the rest, agreed.
Thanks,
-Christoffer
