On Fri, 2024-07-26 at 02:06 -0400, Michael S. Tsirkin wrote: > On Thu, Jul 25, 2024 at 11:20:56PM +0100, David Woodhouse wrote: > > We're rolling out the AMZNVCLK device for internal use cases, and plan > > to add it in public instances some time later. > > Let's be real. If amazon does something in its own hypervisor, and the > only way to use that is to expose the interface to userspace, there is > very little the linux community can do. Moreover, userspace will be > written to this ABI, and be locked in to the specific hypervisor. It > might be a win for amazon short term but long term you will want to > extend things and it will be a mess. > > So I feel you have chosen ACPI badly. It just does not have the APIs > that you need. Virtio does, and would not create a userpspace lock-in > to a specific hypervisor. It's not really virtio specific either, > you can write a bare pci device with a BAR and a bunch of msix > vectors and it will get you the same effect. I *am* as bad as the next person for taking the "I have a hammer, therefore everything is a nail" approach. For you that hammer is virtio, and I respect that. But mine isn't ACPI — quite the opposite, it's DT. I *hate* ACPI. I hate everything about it. I hate that Arm started using it for Arm64 instead of going with Device Tree. That's why we have the DSM method for obtaining properties, and the PRP0001 ACPI HID which means "look for the compatible property and treat it like a DT node". So people can make DT bindings and hey, if you're on a system which is afflicted with ACPI, you can still use them. Which I'm still proselytising today, as you saw. But for this use case, we only need a memory region that the hypervisor can update. We don't need any of that complexity of gratuitously interrupting all the vCPUs just to ensure that none of them can be running userspace while one of them does an update for itself, potentially translating from one ABI to another. The hypervisor can just update the user-visible memory in place. In this case, exposing a simple MMIO memory region in _CRS of an ACPI device was the simplest and most compatible solution. Yes, we can add a virtio transport for that where the hypervisor is invited to DMA into (unencrypted) guest memory, and it solves the PAGE_SIZE problem of the trivial ACPI method. But there's still a place in this world for the ACPI method, and it doesn't *hurt* virtio. The important part is the vmclock_abi structure; the transport is just fluff. And I do not agree that this is a lock-in to a specific hypervisor. I've literally rewritten the fields in the structure to align to what virtio-rtc does and accommodate Peter's feedback (to the dismay of my internal team who just wanted to stick with the initial straw man struct and didn't want to keep up, and haven't even engaged with the public threads which have been ongoing since March¹, even when I've beaten them with a big stick). I've added a QEMU implementation too. We absolutely *don't* want this to be hypervisor-specific. ¹ https://lore.kernel.org/all/0e21e3e2be26acd70b5575b9932b3a911c9fe721.camel@xxxxxxxxxxxxx
<<attachment: smime.p7s>>
