Re: Future of target/arm "max" CPU type

["Daniel P. Berrangé" <berrange@redhat.com>]

On Thu, May 07, 2026 at 03:20:27PM +0200, Philippe Mathieu-Daudé wrote:
> On 7/5/26 12:29, Daniel P. Berrangé wrote:
> > On Thu, May 07, 2026 at 12:17:32PM +0200, Philippe Mathieu-Daudé wrote:
> > > On 7/5/26 00:17, Pierrick Bouvier wrote:
> > > > Hi Peter,
> > > > 
> > > > On 5/1/2026 6:26 AM, Peter Maydell wrote:
> > > > > This is something we've discussed before. I don't think I have a very
> > > > > solid grip on it and I suspect I may be missing some parts of the
> > > > > "problem description" (i.e. what are we trying to achieve) and perhaps
> > > > > some implications of the choice. Overall it feels to me like a "no
> > > > > obviously correct answer, make a tradeoff" question. I've written this
> > > > > up to try to clarify my thinking and canvass for opinions and for
> > > > > people to point out what I'm missing.
> > > > > 
> > > > > For a long time now the arm targets have defined a "max" CPU type
> > > > > whose definition is roughly "all the features that QEMU
> > > > > implements". This is helpful for:
> > > > > 
> > > > >    - giving us a way to provide newly emulated CPU features to users
> > > > >      without having to define a specific named CPU type matching some
> > > > >      hardware that implements the feature (which might not exist at the
> > > > >      point we add the emulation, or which might also need some other
> > > > >      features we don't yet support)
> > > > >    - giving users a way to say "I just want everything"
> > > > > 
> > > > > Mostly this has worked pretty well, I think. (It can cause problems
> > > > > for some workloads that break when a new feature is added; TF-A
> > > > > firmware seems particularly prone to this.)
> > > > > 
> > > > > However, it doesn't work so well if the architecture decides to drop a
> > > > > feature and mandate that newer implementations may not implement
> > > > > it. Up til now that hasn't been a big deal, as only a few very niche
> > > > > items (like FEAT_DoubleLock, an old mechanism for quiescing external
> > > > > debug before a CPU power-down) have ever been removed.
> > > > > 
> > > > > With the Armv9.0 architecture, a big feature gets removed: support for
> > > > > running in 32-bit mode at anything above userspace (EL1, EL2, EL3) is
> > > > > not allowed to be implemented.  We've got away so far with ignoring
> > > > > that requirement, but as new v9.x features are added, they are
> > > > > designed under the assumption that there is no 32-bit system mode
> > > > > code. For instance, the GICv5 CPU interface defines only 64-bit system
> > > > > registers, and our non-compliant "32-bit is actually there"
> > > > > implementation would give you a 32-bit OS that couldn't handle
> > > > > interrupts. Other similar (or harder to spot) issues are likely to
> > > > > also turn up.
> > > > > 
> > > > > So my suggestion is that we should have two new CPU types:
> > > > > 
> > > > >    v8-max : "all the features we implement that are valid
> > > > >      for the v8.x architecture"
> > > > >    v9-max : "all the features we implement that are valid
> > > > >      for the v9.x architecture"
> > > > > 
> > > > > v9-max would not implement AArch32 EL1/2/3; v8-max wouldn't implement
> > > > > any feature the architecture says is v9-or-later.
> > > 
> > > Certainly clearer than what we currently have.
> > > 
> > > > > For the existing 'max' CPU:
> > > > >    - on qemu-system-arm this would be a synonym for v8-max
> > > > >      (since it must be 32-bit)
> > > > >    - on qemu-aarch64 and qemu-system-aarch64 this would be a synonym
> > > > >      for v9-max
> > > > >    - on qemu-arm it can also be v9-max, since v9 doesn't drop 32-bit
> > > > >      usermode support. (Though we don't currently implement qemu-arm
> > > > >      max this way, so it might end up being effectively v8-max until
> > > > >      we get round to implementing support for selecting AArch64 CPUs in
> > > > >      qemu-arm.)
> > > > > 
> > > > 
> > > > This seems like a good approach. A side effect is to solve the
> > > > duplication we encounter for the single-binary also, by correctly
> > > > deduplicating QOM types associated.
> > > > 
> > > > The concept of having a "max" alias that maps to a specific type (in
> > > > this case, distinct depending on the binary) could be used for other
> > > > arch if needed also.
> > > > 
> > > > > I don't think it's worth claiming 'max' as deprecated and dropping it:
> > > > > as an alias it's trivial to support, and there are a lot of workloads
> > > > > that won't ever notice this change.
> > > 
> > > We could display a note though.
> > > 
> > >    "Using 'v9-max' for requested 'max' CPU".
> > 
> > Please don't make use of "-cpu max" print warnings.
> > 
> > IMHO "max" should remain a first class citizen across all targets,
> > because it provides a consistent approach for users.
> > 
> > What Peter describes with resolving "max" to either "v9-max" or
> > "v8-max" depending on arm vs aarch64  makes sense to me.
> 
> Effectively postponing the problem until ARMv10 gets released
> with v10-max diverging from v9-max.
> 
> Nevermind, we'll keep handling "max" alias at the machine level.

Tieing to machine type is tricky, however, as v10-max vs v9-max
will have different "runability" depending on what the host OS
is

With regards,
Daniel
-- 
|: https://berrange.com       ~~        https://hachyderm.io/@berrange :|
|: https://libvirt.org          ~~          https://entangle-photo.org :|
|: https://pixelfed.art/berrange   ~~    https://fstop138.berrange.com :|