Avi,
Are you sure there is a sse-mmio branch? I could not find it anywhere...
Wei Xu
On 3/27/11 4:57 AM, "Avi Kivity" <avi@xxxxxxxxxx> wrote:
> On 03/26/2011 12:12 AM, Wei Xu wrote:
>> Jiri& Avi:
>>
>> I attached the patched I did for movq and movdqa emulation. Please note:
>> (1) I only implemented those two. Other instructions like addq may be
>> following same way.
>> (2) I use same guest_fx_image to hold value and fxsave/fxrstor to copy
>> to/from registers. This is not very efficient I admit.
>> Any suggestions let me know.
>>
>
> Patch is severely whitespace damaged. Please observe the kernel
> whitespace style.
>
> I just remembered that I implemented this once - see the (very old)
> branch sse-mmio in kvm.git.
>
>
>> Index: linux/contents/arch/x86/include/asm/kvm_emulate.h
>> ===================================================================
>> --- linux.orig/contents/arch/x86/include/asm/kvm_emulate.h 2010-07-19
>> 06:42:26.000000000 -0700
>> +++ linux/contents/arch/x86/include/asm/kvm_emulate.h 2011-03-21
>> 09:16:39.000000000 -0700
>> @@ -116,6 +116,7 @@
>> enum { OP_REG, OP_MEM, OP_IMM, OP_NONE } type;
>> unsigned int bytes;
>> unsigned long val, orig_val, *ptr;
>> + unsigned long val_simd[2];
>> };
>
> Breaks on i386 (ulong is 32-bit).
>
>>
>> if (c->src.type == OP_MEM) {
>> + void *val;
>> c->src.ptr = (unsigned long *)memop;
>> c->src.val = 0;
>> + if (c->src.bytes> 8) { /* movdq case */
>> + c->src.val_simd[0] = c->src.val_simd[1] = 0;
>> + val = c->src.val_simd;
>> + } else {
>> + val =&c->src.val;
>> + }
>
> We have a union there for that purpose.
>
>> @@ -2506,6 +2529,55 @@
>> if (!test_cc(c->b, ctxt->eflags))
>> c->dst.type = OP_NONE; /* no writeback */
>> break;
>> + case 0x6f: /* movq from mm/m64 to mm; movdqa from xmm/m128 to xmm */
>> + if (c->op_bytes == 8){
>> + ctxt->vcpu->arch.guest_fx_image.st_space[c->modrm_reg<<2] =
>> + (c->src.val& 0x0ffffffff);
>> + ctxt->vcpu->arch.guest_fx_image.st_space[(c->modrm_reg<<2)+1] =
>> + (c->src.val>> 32);
>> + kvm_fx_restore(&ctxt->vcpu->arch.guest_fx_image);
>> + c->dst.type = OP_NONE; /* Disable writeback. */
>> + break;
>> + } else { /* movdqa */
>> + ctxt->vcpu->arch.guest_fx_image.xmm_space[c->modrm_reg<<2] =
>> + (c->src.val_simd[0]& 0x0ffffffff);
>> + ctxt->vcpu->arch.guest_fx_image.xmm_space[(c->modrm_reg<<2)+1] =
>> + (c->src.val_simd[0]>> 32);
>> + ctxt->vcpu->arch.guest_fx_image.xmm_space[(c->modrm_reg<<2)+2] =
>> + (c->src.val_simd[1]& 0x0ffffffff);
>> + ctxt->vcpu->arch.guest_fx_image.xmm_space[(c->modrm_reg<<2)+3] =
>> + (c->src.val_simd[1]>> 32);
>> + kvm_fx_restore(&ctxt->vcpu->arch.guest_fx_image);
>> + c->dst.type = OP_NONE; /* Disable writeback. */
>> + break;
>> + }
>> + case 0x7f: /* movq from mm to mm/m64; movdqa from xmm to xmm/m128 */
>> + if (c->op_bytes == 8) { /* movq */
>> + kvm_fx_save(&ctxt->vcpu->arch.guest_fx_image);
>> + if (c->dst.type == OP_MEM) {
>> + unsigned long lval,uval;
>> + lval =
>> ctxt->vcpu->arch.guest_fx_image.st_space[c->modrm_reg<<2];
>> + uval =
>> ctxt->vcpu->arch.guest_fx_image.st_space[(c->modrm_reg<<2)+1];
>> + c->dst.val = (uval<<32) + lval;
>> + } else {
>> + c->dst.type = OP_NONE; /* Disable writeback. */
>> + }
>> + break;
>> + } else { /* movdqa */
>> + kvm_fx_save(&ctxt->vcpu->arch.guest_fx_image);
>> + if (c->dst.type == OP_MEM) {
>> + unsigned long lval,uval;
>> + lval =
>> ctxt->vcpu->arch.guest_fx_image.xmm_space[c->modrm_reg<<2];
>> + uval =
>> ctxt->vcpu->arch.guest_fx_image.xmm_space[(c->modrm_reg<<2)+1];
>> + c->dst.val_simd[0] = (uval<<32) + lval;
>> + lval =
>> ctxt->vcpu->arch.guest_fx_image.xmm_space[(c->modrm_reg<<2)+2];
>> + uval =
>> ctxt->vcpu->arch.guest_fx_image.xmm_space[(c->modrm_reg<<2)+3];
>> + c->dst.val_simd[1] = (uval<<32) + lval;
>> + } else {
>> + c->dst.type = OP_NONE; /* Disable writeback. */
>> + }
>> + break;
>> + }
>
> In my implementation, I just forced the guest mmu to be active, and used
> the sse instructions directly.
>> Index: linux/contents/include/linux/kvm.h
>> ===================================================================
>> --- linux.orig/contents/include/linux/kvm.h 2010-07-19 06:42:23.000000000
>> -0700
>> +++ linux/contents/include/linux/kvm.h 2011-03-21 09:16:39.000000000 -0700
>> @@ -152,7 +152,7 @@
>> /* KVM_EXIT_MMIO */
>> struct {
>> __u64 phys_addr;
>> - __u8 data[8];
>> + __u8 data[16];
>> __u32 len;
>> __u8 is_write;
>> } mmio;
>
> This breaks the userspace interface. My implementation split the I/O
> into two separate 64-bit writes.
>
> I guess I'll have to rebase it.
--
To unsubscribe from this list: send the line "unsubscribe kvm" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
