[PATCH v2 00/21] arm64: KVM: world switch in C

[Christoffer Dall]

On Tue, Dec 01, 2015 at 05:51:46PM +0000, Marc Zyngier wrote:
> On 01/12/15 12:00, Christoffer Dall wrote:
> > On Tue, Dec 01, 2015 at 09:58:23AM +0000, Marc Zyngier wrote:
> >> On 30/11/15 20:33, Christoffer Dall wrote:
> >>> On Fri, Nov 27, 2015 at 06:49:54PM +0000, Marc Zyngier wrote:
> >>>> Once upon a time, the KVM/arm64 world switch was a nice, clean, lean
> >>>> and mean piece of hand-crafted assembly code. Over time, features have
> >>>> crept in, the code has become harder to maintain, and the smallest
> >>>> change is a pain to introduce. The VHE patches are a prime example of
> >>>> why this doesn't work anymore.
> >>>>
> >>>> This series rewrites most of the existing assembly code in C, but keeps
> >>>> the existing code structure in place (most function names will look
> >>>> familiar to the reader). The biggest change is that we don't have to
> >>>> deal with a static register allocation (the compiler does it for us),
> >>>> we can easily follow structure and pointers, and only the lowest level
> >>>> is still in assembly code. Oh, and a negative diffstat.
> >>>>
> >>>> There is still a healthy dose of inline assembly (system register
> >>>> accessors, runtime code patching), but I've tried not to make it too
> >>>> invasive. The generated code, while not exactly brilliant, doesn't
> >>>> look too shaby. I do expect a small performance degradation, but I
> >>>> believe this is something we can improve over time (my initial
> >>>> measurements don't show any obvious regression though).
> >>>
> >>> I ran this through my experimental setup on m400 and got this:
> >>
> >> [...]
> >>
> >>> What this tells me is that we do take a noticable hit on the
> >>> world-switch path, which shows up in the TCP_RR and hackbench workloads,
> >>> which have a high precision in their output.
> >>>
> >>> Note that the memcached number is well within its variability between
> >>> individual benchmark runs, where it varies to 12% of its average in over
> >>> 80% of the executions.
> >>>
> >>> I don't think this is a showstopper thought, but we could consider
> >>> looking more closely at a breakdown of the world-switch path and verify
> >>> if/where we are really taking a hit.
> >>
> >> Thanks for doing so, very interesting. As a data point, what compiler
> >> are you using? I'd expect some variability based on the compiler version...
> >>
> > I used the following (compiling natively on the m400):
> > 
> > gcc version 4.8.2 (Ubuntu/Linaro 4.8.2-19ubuntu1)
> 
> For what it is worth, I've ran hackbench on my Seattle B0 (8xA57 2GHz),
> with a 4 vcpu VM and got the following results (10 runs per kernel
> version, same configuration):
> 
> v4.4-rc3-wsinc: Average 31.750
> 32.459
> 32.124
> 32.435
> 31.940
> 31.085
> 31.804
> 31.862
> 30.985
> 31.450
> 31.359
> 
> v4.4-rc3: Average 31.954
> 31.806
> 31.598
> 32.697
> 31.472
> 31.410
> 32.562
> 31.938
> 31.932
> 31.672
> 32.459
> 
> This is with GCC as produced by Linaro:
> aarch64-linux-gnu-gcc (Linaro GCC 5.1-2015.08) 5.1.1 20150608
> 
> It could well be that your compiler generates worse code than the one I
> use, or that the code it outputs is badly tuned for XGene. I guess I
> need to unearth my Mustang to find out...
> 
Worth investigating I suppose.  At any rate, the conclusion stays the
same; we should proceed with these patches.

-Christoffer