On Mon, 18 Mar 2024 18:15:36 +0000, David Woodhouse <dwmw2@xxxxxxxxxxxxx> wrote: > > [1 <text/plain; UTF-8 (quoted-printable)>] > On Mon, 2024-03-18 at 17:41 +0000, Marc Zyngier wrote: > > On Mon, 18 Mar 2024 17:26:07 +0000, > > David Woodhouse <dwmw2@xxxxxxxxxxxxx> wrote: > > > > > > [1 <text/plain; UTF-8 (quoted-printable)>] > > > On Mon, 2024-03-18 at 16:57 +0000, Marc Zyngier wrote: > > > > > > > > > > > > > > There *is* a way for a VMM to opt *out* of newer PSCI versions... by > > > > > setting a per-vCPU "special" register that actually ends up setting the > > > > > PSCI version KVM-wide. Quite why this isn't just a simple KVM_CAP, I > > > > > have no idea. > > > > > > > > Because the expectations are that the VMM can blindly save/restore the > > > > guest's state, including the PSCI version, and restore that blindly. > > > > KVM CAPs are just a really bad design pattern for this sort of things. > > > > > > Hm, am I missing something here? Does the *guest* get to set the PSCI > > > version somehow, and opt into the latest version that it understands > > > regardless of what the firmware/host can support? > > > > No. The *VMM* sets the PSCI version by writing to a pseudo register. > > It means that when the guest migrates, the VMM saves and restores that > > version, and the guest doesn't see any change. > > And when you boot a guest image which has been working for years under > a new kernel+KVM, your guest suddenly experiences a new PSCI version. > As I said that's not just new optional functions; it's potentially even > returning new error codes to the functions that said guest was already > using. If you want to stick to a given PSCI version, you write the version you want. > > And when you *hibernate* a guest and then launch it again under a newer > kernel+KVM, it experiences the same incompatibility. > > Unless the VMM realises this problem and opts *out* of the newer KVM > behaviour, of course. This is very much unlike how we *normally* expose > new KVM capabilities. This was discussed at length 5 or 6 years ago (opt-in vs opt-out). The feedback from QEMU (which is the only public VMM that does anything remotely useful with this) was that opt-out was a better model, specially as PSCI is the conduit for advertising the Spectre mitigations and users (such as certain cloud vendors) were pretty keen on guests seeing the mitigations advertised *by default*. And if you can spot any form of "normality" in the KVM interface, I'll buy you whatever beer you want. It is all inconsistent crap, so I think we're in pretty good company here. > > > > I don't think we ever aspired to be able to hand an arbitrary KVM fd to > > > a userspace VMM and have the VMM be able to drive that VM without > > > having any a priori context, did we? > > > > Arbitrary? No. This is actually very specific and pretty well > > documented. > > > > Also, to answer your question about why we treat 0.1 differently from > > 0.2+: 0.1 didn't specify the PSCI SMC/HCR encoding, meaning that KVM > > implemented something that was never fully specified. The VMM has to > > provide firmware tables that describe that. With 0.2+, there is a > > standard encoding for all functions, and the VMM doesn't have to > > provide the encoding to the guest. > > Gotcha. So for that case we were *forced* to do things correctly and > allow userspace to opt-in to the capability. While for 0.2 onwards we > got away with this awfulness of silently upgrading the version without > VMM consent. > > I was hoping to just follow the existing model of SYSTEM_RESET2 and not > have to touch this awfulness with a barge-pole, but sure, whatever you > want. Unless I'm reading the whole thing wrong (which isn't impossible given that I'm jet-lagged to my eyeballs), SYSTEM_RESET2 doesn't have any form of configuration. If PSCI 1.1 is selected, SYSTEM_RESET2 is available. So that'd be the model to follow. M. -- Without deviation from the norm, progress is not possible.
