[PATCH v4 00/16] kvm/arm: Align the VMID allocation with the arm64 ASID one

[Shameerali Kolothum Thodi]

> -----Original Message-----
> From: Will Deacon [mailto:will at kernel.org]
> Sent: 22 April 2021 18:09
> To: Shameerali Kolothum Thodi <shameerali.kolothum.thodi at huawei.com>
> Cc: linux-arm-kernel at lists.infradead.org; kvmarm at lists.cs.columbia.edu;
> linux-kernel at vger.kernel.org; maz at kernel.org; catalin.marinas at arm.com;
> james.morse at arm.com; julien.thierry.kdev at gmail.com;
> suzuki.poulose at arm.com; jean-philippe at linaro.org; julien at xen.org; Linuxarm
> <linuxarm at huawei.com>
> Subject: Re: [PATCH v4 00/16] kvm/arm: Align the VMID allocation with the
> arm64 ASID one
> 
> On Wed, Apr 14, 2021 at 12:22:56PM +0100, Shameer Kolothum wrote:
> > Hi,
> >
> > This is an attempt to revive this series originally posted by
> > Julien Grall[1]. The main motive to work on this now is because
> > of the requirement to have Pinned KVM VMIDs and the RFC discussion
> > for the same basically suggested[2] to have a common/better vmid
> > allocator for KVM which this series provides.
> >
> > Major Changes from v3:
> >
> > -Changes related to Pinned ASID support.
> > -Changes to take care KPTI related bits reservation.
> > -Dropped support for 32 bit KVM.
> > -Rebase to 5.12-rc7
> >
> > Individual patches have change history for any major changes
> > from v3.
> >
> > Tests were performed on a HiSilicon D06 platform and so far not observed
> > any regressions.
> >
> > For ASID allocation,
> >
> > Avg of 10 runs(hackbench -s 512 -l 200 -g 300 -f 25 -P),
> > 5.12-rc7: Time:18.8119
> > 5.12-rc7+v4: Time: 18.459
> >
> > ~1.8% improvement.
> >
> > For KVM VMID,
> >
> > The measurement was made with maxcpus set to 8 and with the
> > number of VMID limited to 4-bit. The test involves running
> > concurrently 40 guests with 2 vCPUs. Each guest will then
> > execute hackbench 5 times before exiting.
> >
> > The performance difference between the current algo and the
> > new one are(ag. of 10 runs):
> >     - 1.9% less exit from the guest
> >     - 0.7% faster
> >
> > For complete series, please see,
> >  https://github.com/hisilicon/kernel-dev/tree/private-v5.12-rc7-asid-v4
> >
> > Please take a look and let me know your feedback.
> 
> Although I think aligning the two algorithms makes sense, I'm not completely
> sold on the need to abstract all this into a library and whether the
> additional indirection is justified.
> 
> It would be great to compare this approach with one where portions of the
> code are duplicated into a separate VMID allocator. Have you tried that to
> see what it looks like? Doesn't need to be a proper patch set, but comparing
> the end result might help to evaluate the proposal here.

Ok. I will give it a go and get back.

Thanks,
Shameer