Re: [Qemu-devel] [PATCH] x86: Increase max vcpu number to 352

[Radim Krčmář]

2017-08-10 15:16-0300, Eduardo Habkost:
> On Thu, Aug 10, 2017 at 02:41:03PM +0200, Radim Krčmář wrote:
> > 2017-08-10 19:02+0800, Lan Tianyu:
> > > On 2017年08月10日 18:26, Daniel P. Berrange wrote:
> > >> On Thu, Aug 10, 2017 at 06:08:07PM +0800, Lan Tianyu wrote:
> > >>> Intel Xeon phi chip will support 352 logical threads. For HPC
> > >>> usage case, it will create a huge VM with vcpus number as same as host
> > >>> cpus. This patch is to increase max vcpu number to 352.
> > >> 
> > >> If we pick arbitray limits based on size of physical CPUs that happen
> > >> to be shipping today, we'll continue the cat+mouse game forever trailing
> > >> latest CPUs that vendors ship.
> > >> 
> > >> IMHO we should pick a higher number influenced by technical constraints
> > >> of the q35 impl instead. eg can we go straight to something like 512 or
> > >> 1024  ?
> > > 
> > > Sure. 512 should be enough and some arrays is defined according to max
> > > vcpu number.
> > 
> > Hm, which arrays are that?  I was thinking it is safe to bump it to
> > INT_MAX as the number is only used when setting global max_cpus.
> 
> We had a MAX_CPUMASK_BITS macro, and bitmaps whose sizes were
> defined at compile time based on it.  But commit
> cdda2018e3b9ce0c18938767dfdb1e05a05b67ca removed it.  Probably
> those arrays all use max_cpus, by now (and the default for
> max_cpus is smp_cpus, not MachineClass::max_cpus).

Ah, thanks.

> Anyway, if we set it to INT_MAX, there are some cases where more
> appropriate error checking/reporting could be required because
> they won't handle overflow very well:
> * pcms->apic_id_limit initialization at pc_cpus_init()
> * ACPI code that assumes possible_cpus->cpus[i].arch_id fits
>   in a 32-bit integer
> * Other x86_cpu_apic_id_from_index() calls in PC code
>   (especially the initialization of possible_cpus->cpus[i].arch_id).
>   Note that x86_cpu_apic_id_from_index(cpu_index) might not fit
>   in 32 bits even if cpu_index <= UINT32_MAX.

Good point, looks like it all comes to x86_cpu_apic_id_from_index().
Each level of the topology has at most one underutilized bit, so
2^(32 - 3) would be safe.

It is still needlessly large for the foreseeable future, but 512 is
going to be surpassed pretty soon, so I think that jumping at least to
8k would be better.
(8k the current default maximum for Linux and the resulting overcommit
 of ~20 is bearable for smoke testing on current hardware.)