Re: [Qemu-devel] spapr guests warn "System page size 0x2000 is not enabled in page_size_mask (0x11000). Performance may be slow" on SPARC hosts

[Alexey Kardashevskiy]

On 20/06/17 11:02, David Gibson wrote:
> On Mon, Jun 19, 2017 at 12:32:48PM +0200, Thomas Huth wrote:
>> On 19.06.2017 11:39, Alexander Graf wrote:
>>> On 06/15/2017 03:33 PM, Peter Maydell wrote:
>>>> I've just noticed that on a SPARC host, some of the PPC guests
>>>> warn during make check:
>>>>
>>>>    /ppc64/prom-env/pseries:
>>>> qemu-system-ppc64: System page size 0x2000 is not enabled in
>>>> page_size_mask (0x11000). Performance may be slow
>>>>
>>>> Is this really a performance problem on a TCG guest?
>>>> It makes the 'make check' output a bit noisy, but not
>>>> unbearably so.
>>>
>>> From what I can tell this really only affects DDW which should be page
>>> size agnostic when only using emulated devices. Maybe Alexey thought of
>>> using VFIO-pci devices in a TCG guest?
>>
>> Maybe we could at least silence the warning by checking
>> qtest_enabled()?
> 
> 
> So,
> 
> 1) Yes, we should at least silence the warning.

Agree, wrapping it into kvm_enabled() should do it.


> 
> 2) Certainly the mentioned slowdown shouldn't be significant for TCG
> guests.

It is not significant at all. While in-kernel TCE handling improves
throughput for VFIO devices quite a lot (200mb/s -> 1000mb/s), TCG could
not do any better than 20mb/s anyway.


> 3) I _think_ the restriction which causes the slowdown doesn't even
> exist for TCG, but I haven't 100% convinced myself of this.  I've been
> sick lately, so my brain isn't working all that well.

Mismatch between system page size and IOMMU page size means H_PUT_TCE is
going to be called more often than needed; if VFIO is involved, then more
ioctl(vfio_container_fd, DMAP_MAP) calls. But - we cannot do 8K IOMMU pages
at all, SPARC does not support VFIO (or does it?), and TCG is slow anyway.



-- 
Alexey

signature.asc
Description: OpenPGP digital signature