Re: [PATCH] irqchip/gic-v4.1:Check whether indirect table is supported in allocate_vpe_l1_table

[Tangnianyao <tangnianyao@huawei.com>]

On 5/15/2024 17:34, Marc Zyngier wrote:
> On Wed, 15 May 2024 09:56:10 +0100,
> Tangnianyao <tangnianyao@huawei.com> wrote:
>>
>>
>> On 1/22/2024 22:02, Marc Zyngier wrote:
>>> On Mon, 22 Jan 2024 13:13:09 +0000,
>>> Tangnianyao <tangnianyao@huawei.com> wrote:
>>>> On 1/22/2024 17:00, Marc Zyngier wrote:
>>>>> [Fixing the LKML address, which has bits of Stephan's address embedded
>>>>> in it...]
>>>>>
>>>>> On Mon, 22 Jan 2024 16:06:07 +0000,
>>>>> Nianyao Tang <tangnianyao@huawei.com> wrote:
>>>>>> In allocate_vpe_l1_table, when we fail to inherit VPE table from other
>>>>>> redistributors or ITSs, and we allocate a new vpe table for current common 
>>>>>> affinity field without checking whether indirect table is supported.
>>>>>> Let's fix it.
>>>>> Is there an actual implementation that doesn't support the indirect
>>>>> property for the VPE table? I know this is allowed for consistency
>>>>> with the original revision of the architecture, but I never expected
>>>>> an actual GICv4.1 implementation to be *that* bad.
>>>>>
>>>>> If that's the case, I'm a bit puzzled/worried.
>>>> I met this problem in a developing implementation and find it's allowed by GIC spec.
>>>> In such environment,  in a common affinity field with only redistributors and without
>>>> any ITS in it, forcing its_vpe_id_alloc to allocate a large vpeid(like 65000), and there
>>>> comes an error message "VPE IRQ allocation failure". It originally comes from
>>>> allocate_vpe_l2_table, reading GICR_VPROPBASER with GICR_VPROPBASER_4_1_SIZE=1
>>>> and GICR_VPROPBASER_4_1_INDIRECT=0.
>>> Really, you should get your HW engineers to fix their GIC
>>> implementation.  I'm OK with working around this issue for
>>> completeness, but shipping such an implementation would be a mistake.
>>>
>>> [...]
>>>
>>>> I have another question here. The max number of pages  for GITS_BASER
>>>> and GICR_VPROPBASER is different here, while GITS_BASER.Size is
>>>> bit[7:0] with max 256, and GICR_4_1_VPROPBASER.Size is bit[6:0] with max 128.
>>>> Kernel usually probe ITS basers first and then probe GICR_4_1_VPROPBASER in
>>>> a common affinity group. Maybe we need to check this in "inherit_vpe_l1_table_from_its" ?
>>> This is because GITS_BASER[] is generic (also works for devices and
>>> collections), while GICR_VPROPBASER is tailored to the VPE table which
>>> is usually smaller.
>>>
>>> I would expect that GICD_TYPER2.VID reports something that cannot
>>> result in something going wrong (in this case, the L1 allocation
>>> cannot be more than 128 pages).
>>>
>>> Overall, the kernel isn't a validation suite for the HW, and we expect
>>> it to have some level of sanity. So if none of this is in shipping HW
>>> but only in some model with crazy parameters, I don't think we should
>>> go out of our way to support it.
>>>
>>> Thanks,
>>>
>>> 	M.
>>>
>> Hi Marc,
>> Friendly ping. Do we have plan to fix this problem on kernel, or any other plan ?
> Hi Nianyao,
>
> My earlier question still stand: is this something that affects a
> shipping implementation? If not, then I don't think we should support
> this upstream, as this doesn't seem like a realistic configuration.
>
> If your employer has actually built this (which I still consider as a
> mistake), then we can add the workaround I suggested.
>
> Thanks,
>
> 	M.
>
Hi Marc,

For GIC4.1 of HIP09,HIP10,HIP10C, it only support one-level vcpu table and GITS_BASER<n>.Indirect
is RAZ/WI. It implements page size 16KB and entry size 32B,  each page support 512 vpe, and our
clients have requirement 1 or 2 pages at most, so it just supports flat page to simplify implementation.

Our designer has confirmed with ARM architect that we can waive this rule:
Quote from GIC spec 12.19.1 GITS_BASER<n> description, if the maximum width of the scaling factor
that is identified by GITS_BASER<n>.Type and the smallest page size that is supported result in a single
level table that requires multiple pages, then implementing this bit as RAZ/WI is DEPRECATED.

We have actually built this in HIP09,HIP10,HIP10C and would like to fix this in kernel.
Can we merge the above bug-fix patch to kernel?  Or we need to fix with errata for HIP09,HIP10,HIP10C?

thanks for your opinions

Nianyao Tang