On 13 August 2024 00:59:40 BST, Sean Christopherson <seanjc@xxxxxxxxxx> wrote: >On Fri, Aug 02, 2024, David Woodhouse wrote: >> On Fri, 2024-08-02 at 07:55 -0700, Sean Christopherson wrote: >> > On Fri, Aug 02, 2024, David Woodhouse wrote: >> > > On Thu, 2024-08-01 at 20:54 +0200, Thomas Gleixner wrote: >> > > > On Thu, Aug 01 2024 at 16:14, Michael Kelley wrote: >> > > > > I don't have a convenient way to test my sequence on KVM. >> > > > >> > > > But still fails in KVM >> > > >> > > By KVM you mean the in-kernel one that we want to kill because everyone >> > > should be using userspace IRQ chips these days? >> > >> > What exactly do you want to kill? In-kernel local APIC obviously needs to stay >> > for APICv/AVIC. >> >> The legacy PIT, PIC and I/O APIC. >> >> > And IMO, encouraging userspace I/O APIC emulation is a net negative for KVM and >> > the community as a whole, as the number of VMMs in use these days results in a >> > decent amount of duplicated work in userspace VMMs, especially when accounting >> > for hardware and software quirks. >> >> I don't particularly care, but I thought the general trend was towards >> split irqchip mode, with the local APIC in-kernel but i8259 PIC and I/O >> APIC (and the i8254 PIT, which was the topic of this discussion) being >> done in userspace. > >Yeah, that's where most everyone is headed, if not already there. Letting the >I/O APIC live in userspace is probably the right direction long term, I just don't >love that every VMM seems to have it's own slightly different version. But I think >the answer to that is to build a library for (legacy?) device emulation so that >VMMs can link to an implementation instead of copy+pasting from somwhere else and >inevitably ending up with code that's frozen in time. Some would say the right answer is to present a micro-vm machine model that doesn't have any of that crap at all. Sadly we're going in the wrong direction. For >255 vCPUs on AMD machines it looks like we even have to emulate a full virtual IOMMU with DMA translation support. Well done, AMD! (Linux is OK with the 15-bit Extended Destination ID, but not Windows)
