Re: [PATCH RFC] kvm: emulate avx vmovdq

[Keith Busch <kbusch@kernel.org>]

On Sun, Sep 22, 2024 at 05:57:05AM -0700, Sean Christopherson wrote:
> On Tue, Aug 20, 2024, Keith Busch wrote:
> > To test, I executed the following program against a qemu emulated pci
> > device resource. Prior to this kernel patch, it would fail with
> > 
> >   traps: vmovdq[378] trap invalid opcode ip:4006b2 sp:7ffe2f5bb680 error:0 in vmovdq[6b2,400000+1000]
>  
> ...
> 
> > +static const struct gprefix pfx_avx_0f_6f_0f_7f = {
> > +	N, I(Avx | Aligned, em_mov), N, I(Avx | Unaligned, em_mov),
> > +};
> > +
> > +static const struct opcode avx_0f_table[256] = {
> > +	/* 0x00 - 0x5f */
> > +	X16(N), X16(N), X16(N), X16(N), X16(N), X16(N),
> > +	/* 0x60 - 0x6F */
> > +	X8(N), X4(N), X2(N), N,
> > +	GP(SrcMem | DstReg | ModRM | Mov, &pfx_avx_0f_6f_0f_7f),
> > +	/* 0x70 - 0x7F */
> > +	X8(N), X4(N), X2(N), N,
> > +	GP(SrcReg | DstMem | ModRM | Mov, &pfx_avx_0f_6f_0f_7f),
> > +	/* 0x80 - 0xFF */
> > +	X16(N), X16(N), X16(N), X16(N), X16(N), X16(N), X16(N), X16(N),
> > +};
> 
> Mostly as an FYI, we're likely going to run into more than just VMOVDQU sooner
> rather than later.  E.g. gcc-13 with -march=x86-64-v3 (which per Vitaly is now
> the default gcc behavior for some distros[*]) compiles this chunk from KVM
> selftests' kvm_fixup_exception():
> 
> 	regs->rip = regs->r11;
> 	regs->r9 = regs->vector;
> 	regs->r10 = regs->error_code;
> 
> intto this monstronsity (which is clever, but oof).
> 
>   405313:       c4 e1 f9 6e c8          vmovq  %rax,%xmm1
>   405318:       48 89 68 08             mov    %rbp,0x8(%rax)
>   40531c:       48 89 e8                mov    %rbp,%rax
>   40531f:       c4 c3 f1 22 c4 01       vpinsrq $0x1,%r12,%xmm1,%xmm0
>   405325:       49 89 6d 38             mov    %rbp,0x38(%r13)
>   405329:       c5 fa 7f 45 00          vmovdqu %xmm0,0x0(%rbp)
> 
> I wouldn't be surprised if the same packing shenanigans get employed when generating
> code for a struct overlay of emulated MMIO.

Thanks for the notice. I'm hoping we can proceed with just the mov
instructions for now, unless someone already has a real use for these on
emulated MMIO. Otherwise, we can cross that bridge when we get there.

As it is, if just the vmovdq[u,a] are okay, I have a follow on for
vmovdqu64, though I'm currently having trouble adding AVX-512 registers.
Simply increasing the size of the struct x86_emulate_ctxt appears to
break something even without trying to emulate those instructions. But I
want to wait to see if this first part is okay before spending too much
time on it.