Re: [RFC PATCH v4 0/7] hw/arm/virt: Introduce cpu topology support

[Andrew Jones]

On Mon, Jun 28, 2021 at 04:43:05PM +0800, wangyanan (Y) wrote:
> Hi,
> On 2021/6/23 1:39, Daniel P. Berrangé wrote:
> > On Tue, Jun 22, 2021 at 07:29:34PM +0200, Andrew Jones wrote:
> > > On Tue, Jun 22, 2021 at 06:14:25PM +0100, Daniel P. Berrangé wrote:
> > > > On Tue, Jun 22, 2021 at 05:40:13PM +0200, Igor Mammedov wrote:
> > > > > On Tue, 22 Jun 2021 16:29:15 +0200
> > > > > Andrew Jones <drjones@redhat.com> wrote:
> > > > > 
> > > > > > On Tue, Jun 22, 2021 at 03:10:57PM +0100, Daniel P. Berrangé wrote:
> > > > > > > On Tue, Jun 22, 2021 at 10:04:52PM +0800, wangyanan (Y) wrote:
> > > > > > > > Hi Daniel,
> > > > > > > > 
> > > > > > > > On 2021/6/22 20:41, Daniel P. Berrangé wrote:
> > > > > > > > > On Tue, Jun 22, 2021 at 08:31:22PM +0800, wangyanan (Y) wrote:
> > > > > > > > > > On 2021/6/22 19:46, Andrew Jones wrote:
> > > > > > > > > > > On Tue, Jun 22, 2021 at 11:18:09AM +0100, Daniel P. 
> > > > > > > > > > > Berrangé wrote:
> > > > > > > > > > > > On Tue, Jun 22, 2021 at 05:34:06PM +0800, Yanan Wang 
> > > > > > > > > > > > wrote:
> > > > > > > > > > > > > Hi,
> > > > > > > > > > > > > 
> > > > > > > > > > > > > This is v4 of the series [1] that I posted to 
> > > > > > > > > > > > > introduce support for
> > > > > > > > > > > > > generating cpu topology descriptions to guest. 
> > > > > > > > > > > > > Comments are welcome!
> > > > > > > > > > > > > 
> > > > > > > > > > > > > Description:
> > > > > > > > > > > > > Once the view of an accurate virtual cpu topology is 
> > > > > > > > > > > > > provided to guest,
> > > > > > > > > > > > > with a well-designed vCPU pinning to the pCPU we may 
> > > > > > > > > > > > > get a huge benefit,
> > > > > > > > > > > > > e.g., the scheduling performance improvement. See 
> > > > > > > > > > > > > Dario Faggioli's
> > > > > > > > > > > > > research and the related performance tests in [2] for 
> > > > > > > > > > > > > reference. So here
> > > > > > > > > > > > > we go, this patch series introduces cpu topology 
> > > > > > > > > > > > > support for ARM platform.
> > > > > > > > > > > > > 
> > > > > > > > > > > > > In this series, instead of quietly enforcing the 
> > > > > > > > > > > > > support for the latest
> > > > > > > > > > > > > machine type, a new parameter "expose=on|off" in -smp 
> > > > > > > > > > > > > command line is
> > > > > > > > > > > > > introduced to leave QEMU users a choice to decide 
> > > > > > > > > > > > > whether to enable the
> > > > > > > > > > > > > feature or not. This will allow the feature to work 
> > > > > > > > > > > > > on different machine
> > > > > > > > > > > > > types and also ideally compat with already in-use 
> > > > > > > > > > > > > -smp command lines.
> > > > > > > > > > > > > Also we make much stricter requirement for the 
> > > > > > > > > > > > > topology configuration
> > > > > > > > > > > > > with "expose=on".
> > > > > > > > > > > > Seeing this 'expose=on' parameter feels to me like 
> > > > > > > > > > > > we're adding a
> > > > > > > > > > > > "make-it-work=yes" parameter. IMHO this is just 
> > > > > > > > > > > > something that should
> > > > > > > > > > > > be done by default for the current machine type version 
> > > > > > > > > > > > and beyond.
> > > > > > > > > > > > I don't see the need for a parameter to turnthis on, 
> > > > > > > > > > > > especially since
> > > > > > > > > > > > it is being made architecture specific.
> > > > > > > > > > > I agree.
> > > > > > > > > > > 
> > > > > > > > > > > Yanan, we never discussed an "expose" parameter in the 
> > > > > > > > > > > previous versions
> > > > > > > > > > > of this series. We discussed a "strict" parameter though, 
> > > > > > > > > > > which would
> > > > > > > > > > > allow existing command lines to "work" using assumptions 
> > > > > > > > > > > of what the user
> > > > > > > > > > > meant and strict=on users to get what they mean or an 
> > > > > > > > > > > error saying that
> > > > > > > > > > > they asked for something that won't work or would require 
> > > > > > > > > > > unreasonable
> > > > > > > > > > > assumptions. Why was this changed to an "expose" 
> > > > > > > > > > > parameter?
> > > > > > > > > > Yes, we indeed discuss a new "strict" parameter but not a 
> > > > > > > > > > "expose" in v2 [1]
> > > > > > > > > > of this series.
> > > > > > > > > > [1] 
> > > > > > > > > > https://patchwork.kernel.org/project/qemu-devel/patch/20210413080745.33004-6-wangyanan55@huawei.com/
> > > > > > > > > > 
> > > > > > > > > > And in the discussion, we hoped things would work like 
> > > > > > > > > > below with "strict"
> > > > > > > > > > parameter:
> > > > > > > > > > Users who want to describe cpu topology should provide 
> > > > > > > > > > cmdline like
> > > > > > > > > > 
> > > > > > > > > > -smp strict=on,cpus=4,sockets=2,cores=2,threads=1
> > > > > > > > > > 
> > > > > > > > > > and in this case we require an more accurate -smp 
> > > > > > > > > > configuration and
> > > > > > > > > > then generate the cpu topology description through ACPI/DT.
> > > > > > > > > > 
> > > > > > > > > > While without a strict description, no cpu topology 
> > > > > > > > > > description would
> > > > > > > > > > be generated, so they get nothing through ACPI/DT.
> > > > > > > > > > 
> > > > > > > > > > It seems to me that the "strict" parameter actually serves 
> > > > > > > > > > as a knob to
> > > > > > > > > > turn on/off the exposure of topology, and this is the 
> > > > > > > > > > reason I changed
> > > > > > > > > > the name.
> > > > > > > > > Yes, the use of 'strict=on' is no better than expose=on IMHO.
> > > > > > > > > 
> > > > > > > > > If I give QEMU a cli
> > > > > > > > > 
> > > > > > > > >     -smp cpus=4,sockets=2,cores=2,threads=1
> > > > > > > > > 
> > > > > > > > > then I expect that topology to be exposed to the guest. I 
> > > > > > > > > shouldn't
> > > > > > > > > have to add extra flags to make that happen.
> > > > > > > > > 
> > > > > > > > > Looking at the thread, it seems the concern was around the 
> > > > > > > > > fact that
> > > > > > > > > the settings were not honoured historically and thus the CLI 
> > > > > > > > > values
> > > > > > > > > could be garbage. ie  -smp cpus=4,sockets=8,cores=3,thread=9
> > > > > > > > This "-smp cpus=4,sockets=8,cores=3,threads=9" behaviors as a 
> > > > > > > > wrong
> > > > > > > > configuration, and the parsing function already report error 
> > > > > > > > for this case.
> > > > > > > > 
> > > > > > > > We hope more complete config like "-smp 
> > > > > > > > 4,sockets=2,cores=2,threads=1"
> > > > > > > > for exposure of topology, and the incomplete ones like "-smp 
> > > > > > > > 4,sockets=1"
> > > > > > > > or "-smp 4, cores=1" are not acceptable any more because we are 
> > > > > > > > starting
> > > > > > > > to expose the topology.
> > > > > > > Incomplete specified topologies *are* acceptable.
> > > > > > > 
> > > > > > > The smp_parse method will automatically fill in any missing 
> > > > > > > values.
> > > > > > > 
> > > > > > > ie,
> > > > > > > 
> > > > > > >    -smp 4,cores=1
> > > > > > >    -smp cores=1
> > > > > > >    -smp threads=1
> > > > > > >    -smp sockets=4
> > > > > > > 
> > > > > > > are all functionally identical to
> > > > > > > 
> > > > > > >    -smp 4,sockets=4,cores=1,dies=1,threads=1
> > > > > > > 
> > > > > > > 
> > > > > > > The QEMU man page says this explicitly
> > > > > > > 
> > > > > > >                   For the PC target, the number of cores per die, 
> > > > > > > the
> > > > > > >      number of threads per cores, the number of dies per packages 
> > > > > > > and the
> > > > > > >      total number of sockets can be specified. Missing values 
> > > > > > > will be
> > > > > > >      computed. If any on the three values is given, the total 
> > > > > > > number of
> > > > > > >      CPUs n can be omitted.
> > > > > > It doesn't say how it will compute them though, which for the 
> > > > > > default
> > > > > > smp_parse and for x86 is to prefer sockets over cores over threads.
> > > > > > That's not necessarily what the user expects. IMO, we need a 
> > > > > > 'strict=on'
> > > > > > parameter that doesn't allow any collection of smp parameters which
> > > > > > require unreasonable assumptions. Reasonable assumptions are 
> > > > > > threads=1,
> > > > > > when threads is not specified and the rest of the math adds up. 
> > > > > > Also,
> > > > > > maxcpus == cpus when maxcpus isn't specified is reasonable. But, 
> > > > > > it's not
> > > > > > as reasonable to decide how to divide cores among sockets or to 
> > > > > > assume
> > > > > > threads=1 when only sockets and cores are given. How do we know the 
> > > > > > user
> > > > > > didn't forget to specify threads if we can't check the math?
> > > > > or just outlaw all invalid topologies incl. incomplete by default
> > > > > (without requiring extra option), and permit them only for old machine
> > > > > types ()using compat machinery) without topo info provided to guest.
> > > > > And maybe later deprecate invalid topologies altogether.
> > > > This feels like it is creating pain for users to fix a problem that
> > > > isn't shown to actually be causing any common issues.
> > > > 
> > > > We've supposed that users are having problems when forgetting to
> > > > specify "threads" and not having the compute value be desirable,
> > > > but where are the bug reports to back this up ?
> > > > 
> > > > The partial topologies are valid and have well defined semantics.
> > > > Those semantics may not match everyone's preference, but that
> > > > doesn't make them invalid.
> > > > 
> > > If we adopt the [undocumented] semantics of x86 for arm, then we may
> > > surprise some users that expect e.g. '-smp 16' to give them a single
> > > socket with 16 cores, because they'll start getting 16 sockets with 1
> > > core each. That's because if we don't describe a topology to an arm linux
> > > guest then it assumes cores. Maybe we shouldn't worry about this, but I'd
> > > prefer we require explicit inputs from users and, if necessary, for them
> > > to explicitly opt-in to requiring those explicit inputs.
> > Even for x86, defaulting to maximising sockets over cores is sub-optimal.
> > In real world x86 hardware it is very rare to have sockets > 2 or 4. For
> > large CPU counts, you generally have large cores-per-socket counts on x86.
> > 
> > The QEMU preference for sockets over cores on x86 (and PPC too IIUC)
> > is a fairly arbitrary historical decision.
> > 
> > It can cause problems with guest OS licensing because both Windows
> > and RHEL have been known to charge differently for sockets vs cores,
> > with high core counts being cheaper.
> > 
> > We are not tied into the precise behaviour of the computed topology
> > values, as we have no made any promises. All that's required is that
> > we keep ABI compat for existing machine types.
> If based on this point of view that we haven't made any promises for the
> precise behavior of the computed topology, things may get much easier.
> I have the following understanding (also a proposal):
> 
> We will introduce the support for exposing cpu topology since machine
> type 6.2 and we will also describe the computed topology for the guest.
> We will not make any stricter parsing logic, however the -smp content in
> qemu-options.hx should be rearranged to clearly explain how the missing
> values will exactly be computed. And this is what QEMU is responsible for.
> 
> We know that a well designed cpu topology configuration can gain much
> benefit for the guest, while a badly designed one will also probably cause
> negative impact. But the users should be responsible for the design of the
> -smp cmdlines. If they are using an incomplete cmdline for a 6.2 machine,
> then they should have known what the computed values will be and that
> the computed topology will be exposed to the guest.
> > 
> > So we could decide to change the computed topology so that it prefers
> > high core counts, over sockets, whem using new machine types only.
> > That would seem to benefit all arches, by making QEMU more reflective
> > of real world CPUs topology.
> If we really decide to prefer cores over sockets over threads for new
> machine
> types, then I think we should also record this change in qemu-option.hx.
>

I agree. The proposal sounds good to me. I'd like to hear Eduardo's
opinion too (CC'ed).

Thanks,
drew