On 11/13/2026 4:19 AM, Daniel P. Berrangé wrote: > On Thu, Nov 12, 2026 at 11:08:02PM +0800, Yonggang Luo wrote: >> On Thu, Nov 12, 2026 at 9:31 PM Daniel P. Berrangé <[email protected]> >> wrote: >>> >>> When building a multi-target binary, QOM types from multiple targets >>> will be linked together. QOM types are required to have unique names >>> when registered, so linking multiple targets in turn requires QOM type >>> names to be unique across the QEMU codebase. >>> >>> Global QOM type name uniqueness is a desirable property for QEMU >>> regardless of multi-target work, so that QEMU presents a non-ambiguous >>> object model to applications regardless of context. >>> >>> In a multi-target binary, it is desirable to applications are able to >>> introspect all types linked into QEMU, when launching with '-machine >>> none'. Dynamic filtering of registered classes at runtime, to resolving >>> QOM naming conflicts is undesirable, as it requires introspection >>> to be repeated for each available target in order to discover the full >>> set of available types in a binary. >>> >>> Introspection is a fairly heavy weight process due to QEMU's vast size, >>> so enabling introspection to be done once only for all targets will be a >>> very significant benefit for management applications with a multi-target >>> binary. >>> >>> The glaring problem is that our historical design has led to multiple >>> targets all providing a 'virt' machine type, which leads to multiple QOM >>> types call 'virt-machine' which are all different. >>> >>> Having a common short name "virt" is a useful characteristic for users, >>> but this does not imply we need to use that as the basis for primary QOM >>> typename. It is sufficient to have an alias called 'virt'. >>> >>> In terms of implementation there are two core pieces >>> >>> * An 'Aliased' interface which can be implemented by QOM types which >>> want to support type aliases. >>> >>> * A 'TargetSpecific' interface which can be implemented by QOM types >>> which want to declare they can only be used with a subset of targets. >>> >>> With those foundational pieces >>> >>> * Machine implements Aliased and TargetSpecific >>> >>> * query-machines reports multiple aliases and declared supported >>> targets per machine for introspection >>> >>> * All the 'virt' machines are renamed to $target-virt, and aliases >>> added for the shorter 'virt' names to retain compatibility and >>> convenience when a target is selected >>> >>> * All the machines which have 'is_available' set now report their >>> targets via the TargetSpecific interface >>> >>> * Creating machines validates declared targets from TargetSpecific >>> >>> An apparent design mistake in TYPE_ACCEL_CPU is resolved which allows >> >> Target independent TYPE_ACCEL_CPU is my first choice, if the QOM name is >> do not have to be kept, I think this is better. >> >>> removal of the is_available hooks from target specific CPU impls, which >>> were working around the bug. >>> >>> The original proposed patches to unconditionally register some object >>> properties is pulled in. >>> >>> Overall this brings us closer to having a single unified object model >>> across all targets, without having to make runtime variations of what >>> is exposed/reported by a given QOM class name. >> >> >> Thanks, overall looks good to me. >> >> hw/microblaze: defer PetaLogix endianness default until init: >> https://lore.kernel.org/qemu-devel/[email protected]/ >> >> is also for avoid target_info in class_init > > I added a R-B to that. > >> The last gap is how to handling cpu class name conflict >> target/arm: register separate max-v8 CPU types for arm and aarch64: >> https://lore.kernel.org/qemu-devel/[email protected]/ > > I had a look at your earlier proposal here: > > > https://lore.kernel.org/qemu-devel/[email protected]/ > > and to me that looks like it should do what we need. > > * In existing qemu-system-arm and qemu-system-aarch64 it will > "just work" > > * In qemu-system-aarch64, users can still opt-in to 32-bit > mode only by using "-cpu max-v8:aarch64=off" > > So based on Peter's comment it does not look like we need to > have the ability to register two distnict "max-v8" CPU classes, > a single class can do the right thing at runtime in its > instance constructor. > > I see Pierrick commented that it would not work in multi-target: > > > https://lore.kernel.org/qemu-devel/[email protected]/ > > I'm not understanding why that is ? The choice of 32-bit vs 64-bit > is made at Object initialization time, so if the > > if (target_aarch64()) { > ... > > condition is not sufficient, it seems like we could add more logic > to that condition easily and to make it do the right thing at runtime. >
I suspect we'll have more issues with those kind of if in instance_init instead of class_init, but I don't have any concrete counter-example for now, so let's move with that. > With regards, > Daniel Regards, Pierrick
