[PATCH] help guest boot up on AArch64 host with GICv2

[Marc Zyngier]

On 28/01/16 20:12, Chris Metcalf wrote:
> On 01/27/2016 04:12 AM, Marc Zyngier wrote:
>> On 26/01/16 20:43, Chris Metcalf wrote:
>>> On 01/18/2016 04:28 AM, Marc Zyngier wrote:
>>>> Hi Chris,
>>>>
>>>> On 15/01/16 20:02, Chris Metcalf wrote:
>>>>> We are using GICv2 compatibility mode in the Fast Models/Foundation
>>>>> Models simulations we are running because the boot code (ATF/UEFI)
>>>>> doesn't support GICv3 in our system at the moment.
>>>>>
>>>>> However, starting with kernel 4.2, the guest couldn't boot up because it
>>>>> wasn't getting timer interrupts.  I tracked this down to a kernel commit
>>>>> that switched to using the "alternatives" mechanism -- rather than
>>>>> seeing either a GICv2 or GICv3 and configuring appropriately, the KVM
>>>>> code just configured the code that saves/restores the vgic state based
>>>>> on the presence of the system register interface to the GIC CPU
>>>>> interface.  See the attached patch for a fix that manages this
>>>>> differently and allows me to boot up the guest in this configuration.
>>>>>
>>>>> However, even assuming this patch can be taken into an upstream tree, I
>>>>> still have a couple of additional problems:
>>>>>
>>>>> - I can boot up with the Foundation Models using this change, but not
>>>>> with the Fast Models (again, using a v3 GIC but in v2 compatibility mode
>>>>> in the device tree).  The Fast Models dts looks like it has the same
>>>>> configuration for the GIC and the timers so I'm not sure what's going on
>>>>> here.  Any suggestions appreciated.
>>>>>
>>>>> - Without this change, I could only boot kernels up to 4.1.  With the
>>>>> change, I can boot kernels up to 4.3.  But 4.4 won't boot for me either;
>>>>> I haven't bisected it down yet.  So any suggestions on what might be
>>>>> going wrong here would also be appreciated.
>>>>>
>>>>> We are planning to eventually use GICv3 mode in our software stack but
>>>>> for the time being I assume it is interesting to resolve issues with GIC
>>>>> v2 compatibility mode on GIC v3.
>>>>>
>>>> I'm afraid that this is the wrong approach. Whilst 4.2 was a bit too
>>>> eager to use GICv3 (only checking the CPU capability and ignoring the
>>>> actual state of the EL2/EL3 SRE bits), the fact that 4.4 doesn't boot is
>>>> probably the sign of a broken firmware that enables the system register
>>>> interface at EL3, letting the rest of the software stack to use GICv3 in
>>>> native mode, and yet providing a GICv2 DT.
>>>>
>>>> This combination is unpredictable, and is likely to  cause issues on
>>>> some HW implementations.
>>>>
>>>> Could you please point me to the firmware you're using?
>>>>
>>>> Also, please check the following patches:
>>>>
>>>> 6d32ab2 arm64: Update booting requirements for GICv3 in GICv2 mode
>>>> 76e52dd irqchip/gic: Warn if GICv3 system registers are enabled
>>>> 963fcd4 arm64: cpufeatures: Check ICC_EL1_SRE.SRE before enabling
>>>> ARM64_HAS_SYSREG_GIC_CPUIF
>>>> 7cabd00 irqchip/gic-v3: Make gic_enable_sre an inline function
>>>> d271976 arm64: el2_setup: Make sure ICC_SRE_EL2.SRE sticks before using
>>>> GICv3 sysregs
>>>>
>>>> Can you point me to the one that prevents you from booting?
>>> The problematic commit is 963fcd4, because it calls gic_enable_sre()
>>> in the host kernel even with a GICv2 DT specified, and this seems to
>>> put things in a state such that we don't receive virtual timer
>>> interrupts in the guest when we boot it up.  (I'm not that familiar with
>>> the QEMU DT but it is providing a GIC v2 to the guest.)
>>>
>>> With a v4.5-rc1 host, if I "return false" before the code in gic_enable_sre()
>>> that tries to actually enable the SRE, and then hardcode the
>>> __vgic_v2_XXX_state() save/restore calls into the __vgic_XXX_state()
>>> routines, then my guest boots up OK.
>> What if you just do the "return false"? I bet that it will work as well...
> 
> Yes, that also works for my case.
> 
>>> We are using a modified ARM version of EDK v3.0-rc0, and a modified
>>> ARM Trusted Firmware based on commit 963fcd4 (between v1.1 and 1.2).
>> Are you sure of that commit? It looks suspiciously like the ID ftom the
>> kernel tree...
> 
> Hah, good catch!  The double-click-to-copy behavior is kind of flakey
> on RHEL 6's default terminal, and I bet that bit me.  It's 41099f4e.
> 
>>> We certainly haven't touched any of the GIC code in either one.
>>>
>>> I tried to modify the host DT to enable GICv3, but then the host itself
>>> hangs on boot, so clearly more is needed.  (To be fair I've only tested
>>> v4.4 in that configuration, not v4.5-rc1.)  The firmware isn't yet using
>>> GICv3 so perhaps that is part of the problem.
>> That's indeed part of the problem. The firmware running at EL3 insists
>> on using GICv2, but still let EL2 (and EL1) use GICv3 system registers.
>> Could you please dump the content of ICC_SRE_EL3 just before entering
>> the kernel at EL2? If you see ICC_SRE_EL3.SRE being set, then this would
>> indicate a firmware bug (and leave the system in an unpredictable
>> configuration).
> 
> Well, the firmware clearly does this intentionally.  In ATF's
> drivers/arm/giv/arm_gic.c, the gicv3_cpuif_setup() function has
> a comment that reads:
> 
> /*******************************************************************************
>   * This function does some minimal GICv3 configuration. The Firmware itself does
>   * not fully support GICv3 at this time and relies on GICv2 emulation as
>   * provided by GICv3. This function allows software (like Linux) in later stages
>   * to use full GICv3 features.
>   ******************************************************************************/
> 
> and the function ends with:
> 
> 	val = read_icc_sre_el3();
> 	write_icc_sre_el3(val | ICC_SRE_EN | ICC_SRE_SRE);
> 
> In our build environment, if I comment out those two lines, that
> fixes the guest boot problem (without any hacking on the Linux side),
> so that's good anyway.  With this change it works for me in the
> Fast Models as well as Foundation Models, too.

By the look of it, you're trying to use a GICv3 firmware, and pass a
GICv2 DT to the kernel. Do not do that. Either you use a GICv2 firmware
(having spoken to the ATF guys, there is a GICv2 driver in there that
should work for your case) and pass a GICv2 DT, or you go GICv3 all the way.

A mix of the two things is completely unsupported on the model, and
solidly places you in the UNPREDICTABLE category when running that on
actual HW...

Thanks,

	M.
-- 
Jazz is not dead. It just smells funny...