Re: [PATCH] x86/IRQ: bump max number of guests for a shared IRQ to 31

[Jan Beulich <jbeulich@xxxxxxxx>]

On 02.12.2020 15:53, Igor Druzhinin wrote:
> On 02/12/2020 09:25, Jan Beulich wrote:
>> On 01.12.2020 00:59, Igor Druzhinin wrote:
>>> Current limit of 7 is too restrictive for modern systems where one GSI
>>> could be shared by potentially many PCI INTx sources where each of them
>>> corresponds to a device passed through to its own guest. Some systems do not
>>> apply due dilligence in swizzling INTx links in case e.g. INTA is declared 
>>> as
>>> interrupt pin for the majority of PCI devices behind a single router,
>>> resulting in overuse of a GSI.
>>>
>>> Signed-off-by: Igor Druzhinin <igor.druzhinin@xxxxxxxxxx>
>>> ---
>>>
>>> If people think that would make sense - I can rework the array to a list of
>>> domain pointers to avoid the limit.
>>
>> Not sure about this. What is the biggest number you've found on any
>> system? (I assume the chosen value of 31 has some headroom.)
> 
> The value 31 was taken as a practical maximum for one specific HP system
> if all of the PCI slots in all of its riser cards are occupied with GPUs.
> The value is obtained by reverse engineering their ACPI tables. Currently
> we're only concerned with number 8 (our graphics vendors do not recommend
> installing more cards than that) but it's a matter of time it will grow.
> I'm also not sure why this routing scheme was chosen in the first place:
> is it dictated by hardware restrictions or firmware engineers being lazy - 
> we're working with HP to determine it.

Thanks for the insight.

>> Instead I'm wondering whether this wouldn't better be a Kconfig
>> setting (or even command line controllable). There don't look to be
>> any restrictions on the precise value chosen (i.e. 2**n-1 like is
>> the case for old and new values here, for whatever reason), so a
>> simple permitted range of like 4...64 would seem fine to specify.
>> Whether the default then would want to be 8 (close to the current
>> 7) or higher (around the actually observed maximum) is a different
>> question.
> 
> I'm in favor of a command line argument here - it would be much less trouble
> if a higher limit was suddenly necessary in the field. The default IMO
> should definitely be higher than 8 - I'd stick with number 32 which to me
> should cover our real world scenarios and apply some headroom for the future.

Well, I'm concerned of the extra memory overhead. Every IRQ,
sharable or not, will get the extra slots allocated with the
current scheme. Perhaps a prereq change then would be to only
allocate multi-guest arrays for sharable IRQs, effectively
shrinking the overhead in particular for all MSI ones?

Jan