Source-Changes-D archive
[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index][Old Index]
Re: CVS commit: src/sys/dev/nvmm/x86
hi,
sorry for very late reply.
On Sat, Jul 18, 2026 at 11:16 PM Taylor R Campbell <riastradh%netbsd.org@localhost> wrote:
>
> > Date: Sat, 18 Jul 2026 22:04:30 +0900
> > From: Takashi YAMAMOTO <yamt9999%gmail.com@localhost>
> >
> > On Wed, Jul 15, 2026 at 10:22 AM Taylor R Campbell <riastradh%netbsd.org@localhost> wrote:
> > >
> > > nvmm: Add support for extended CPU state (XSAVE) beyond x87/SSE.
> >
> > i have been using a similar patch for a while. but i haven't committed
> > it because
> > i was not happy to implement cpuid emulation bits in kernel like this.
>
> Sorry, didn't realize you were working on this already!
>
> > it's simpler to make qemu provide them via NVMM_VCPU_CONF_CPUID
> > or something similar. (in addition to xcr0)
> > cpuid emulation is quite complex and will likely be more complex in future.
> > it's better to leverage what qemu already has.
> > what do you think?
>
> No opinion, sounds reasonable on its face -- while it is sometimes
> nice to just expose all (safe) host features to the guest, I would
> also want it to be easy to specify the exact features up front for
> reliable migration. Would this require adding a hook to let qemu
> completely override the CPUID visible to the guest, or can qemu use
> the existing CPUID overrides to manage it? I think they might be
> limited in space.
>
> > > Instead, we can expose the XSAVE area in another location in the
> > > virtual machine uvm object, and create an ioctl for querying its
> > > object address and size so userland can mmap it. TBD.
> >
> > i was thinking of having a new NVMM_IOC_VCPU_xxx to get the extra MD state.
>
> Yes, that's what I had in mind. Something like:
>
> struct nvmm_ioc_vcpu_xstate xstate = {
> .thingamajig = NVMM_X86_VCPU_XSTATE_XSAVE,
> };
> void *vxsave;
> struct xsave_header *xsave;
>
> if (ioctl(nvmm_fd, NVMM_IOC_VCPU_MAP_XSTATE, &xstate) == -1)
> goto fail;
> vxsave = mmap(NULL, roundup(xstate.size, PAGE_SIZE),
> PROT_READ|PROT_WRITE, MAP_SHARED|MAP_FILE, nvmm_fd,
> xstate.offset);
> if (vxsave == MAP_FAILED)
> goto fail;
> xsave = vxsave;
>
> Then NVMM_IOC_VCPU_GETSTATE/NVMM_IOC_VCPU_SETSTATE with the flag
> NVMM_X86_STATE_XSAVE, or setting the flag in comm->state_wanted /
> comm->state_cached, would arrange to make it available in the mmapped
> region, just like the other state is already made available in the
> mmapped comm page.
>
> Except we should probably consider how it might look on Arm (with SVE
> and SME) or RISC-V before committing to this ioctl interface.
i meant something separeted from NVMM_IOC_VCPU_GETSTATE/NVMM_IOC_VCPU_SETSTATE,
because i don't see much benefits to put xstate in comm page.
puttinng something this big and even variable-sized in the comm page
without breaking abi seems
quite cumbersome. i'm not sure if it pays off.
(i personally think the "bump NVMM_KERN_VERSION and break abi"
approach is good enough for nvmm.
but you are not happy with it, right?)
btw, do you have qemu bits corresponding to this commit?
or whatever using NVMM_VCPU_CONF_XCR0_MASK.
Home |
Main Index |
Thread Index |
Old Index