Re: [PATCH v2] arm64: gic: increase the number of IRQ descriptors

[Shanker Donthineni <sdonthineni@nvidia.com>]

On 1/5/23 04:59, Marc Zyngier wrote:
> External email: Use caution opening links or attachments
> 
> 
> On Wed, 04 Jan 2023 13:47:03 +0000,
> Shanker Donthineni <sdonthineni@nvidia.com> wrote:
>>
>> Hi Marc,
>>
>> On 1/4/23 03:14, Marc Zyngier wrote:
>>> External email: Use caution opening links or attachments
>>>
>>>
>>> On Wed, 04 Jan 2023 02:37:38 +0000,
>>> Shanker Donthineni <sdonthineni@nvidia.com> wrote:
>>>>
>>>> The default value of NR_IRQS is not sufficient to support GICv4.1
>>>> features and ~56K LPIs. This parameter would be too small for certain
>>>> server platforms where it has many IO devices and is capable of
>>>> direct injection of vSGI and vLPI features.
>>>>
>>>> Currently, maximum of 64 + 8192 (IRQ_BITMAP_BITS) IRQ descriptors
>>>> are allowed. The vCPU creation fails after reaching count ~400 with
>>>> kvm-arm.vgic_v4_enable=1.
>>>>
>>>> This patch increases NR_IRQS to 1^19 to cover 56K LPIs and 262144
>>>> vSGIs (16K vPEs x 16).
>>>>
>>>> Signed-off-by: Shanker Donthineni <sdonthineni@nvidia.com>
>>>> ---
>>>> Changes since v1:
>>>>    -create from v6.2-rc1 and edit commit text
>>>>
>>>>    arch/arm64/include/asm/irq.h | 4 ++++
>>>>    1 file changed, 4 insertions(+)
>>>>
>>>> diff --git a/arch/arm64/include/asm/irq.h b/arch/arm64/include/asm/irq.h
>>>> index fac08e18bcd5..3fffc0b8b704 100644
>>>> --- a/arch/arm64/include/asm/irq.h
>>>> +++ b/arch/arm64/include/asm/irq.h
>>>> @@ -4,6 +4,10 @@
>>>>
>>>>    #ifndef __ASSEMBLER__
>>>>
>>>> +#if defined(CONFIG_ARM_GIC_V3_ITS)
>>>> +#define  NR_IRQS  (1 << 19)
>>>> +#endif
>>>> +
>>>>    #include <asm-generic/irq.h>
>>>>
>>>>    struct pt_regs;
>>>
>>> Sorry, but I don't think this is an acceptable change. This is a large
>>> overhead that affects *everyone*, and that will eventually be too
>>> small anyway with larger systems and larger interrupt spaces.
>>>
>>> A better way to address this would be to move to a more dynamic
>>> allocation, converting the irqdesc rb-tree into an xarray, getting rid
>>> of the bitmaps (the allocation bitmap and the resend one), and track
>>> everything in the xarray.
>>
>> The actual memory allocation for IRQ descriptors is still dynamic for ARM64.
>> This change increases static memory for variable 'allocated_irqs' by 64KB,
>> feel not a noticeable overhead.
> 
> 64kB for each bitmap, so that's already 128kB (you missed the
> irqs_resend bitmap). And that's for a number of IRQs that is still way
> below what the GIC architecture supports today.
> 
> The architecture supports 32bit INTIDs, and that's 1GB worth of
> bitmaps, only for the physical side. Add the virtual stuff for which
> we create host-side descriptors, and we can go way beyond that.
> 
> So what happens next, once you exceed the arbitrary limit that only
> satisfies your own use case? We will bump it up again, and again,
> bloating the kernel with useless static data that *nobody* needs.
> Specially not the VMs that you plan to run.
> 
> So I'm putting my foot down right now, and saying that it needs to be
> fixed once and for all. The current scheme was OK for small interrupt
> spaces, but it isn't fit for purpose anymore, certainly not with
> things like the GICv4 architecture.
> 
> I'm happy to help with it, but I'm certainly not willing to accept any
> sort of new compile-time limit.

Thanks for helping with a scalable solution instead of static allocation.
Please include me whenever patches posted to LKML. I'm happy to verify
on NVIDIA server platforms and provide test feedback.