Re: [PATCHv5 4/4] KVM: emulator: optimize "rep ins" handling.

[Gleb Natapov <gleb@redhat.com>]

On Mon, Aug 06, 2012 at 12:28:05PM +0300, Avi Kivity wrote:
> On 08/06/2012 11:58 AM, Gleb Natapov wrote:
> > On Mon, Aug 06, 2012 at 11:50:20AM +0300, Avi Kivity wrote:
> >> On 07/30/2012 05:38 PM, Gleb Natapov wrote:
> >> > Optimize "rep ins" by allowing emulator to write back more than one
> >> > datum at a time. Introduce new operand type OP_MEM_STR which tells
> >> > writeback() that dst contains pointer to an array that should be written
> >> > back as opposite to just one data element.
> >> > 
> >> >  	}
> >> >  
> >> > -	memcpy(dest, rc->data + rc->pos, size);
> >> > -	rc->pos += size;
> >> > +	if (ctxt->rep_prefix && !(ctxt->eflags & EFLG_DF)) {
> >> > +		ctxt->dst.data = rc->data + rc->pos;
> >> > +		ctxt->dst.type = OP_MEM_STR;
> >> > +		ctxt->dst.count = (rc->end - rc->pos) / size;
> >> > +		rc->pos = rc->end;
> >> 
> >> Should take into account the segment limit.
> >> 
> > It does. During write back. pio_in_emulated() should linearize() address
> > before calculating page boundary, but this is (minor) bug unrelated to the patch
> > series.
> 
> I see, yes, this problem preexists.
> 
> However, in normal conditions, non-repeating instructions will not reach
> the emulator at all since they will fault in the guest (or in the shadow
> mmu, which will reflect the fault to the guest).  Here, the first
> iteration may fit in the segment but the second will not, so this will fail.
> 
Correct. And this can happen with or without the patch series.

> It's not a huge problem since no guest does this.
> 
> >> > @@ -2732,7 +2747,7 @@ int emulator_task_switch(struct x86_emulate_ctxt *ctxt,
> >> >  static void string_addr_inc(struct x86_emulate_ctxt *ctxt, int reg,
> >> >  		struct operand *op)
> >> >  {
> >> > -	int df = (ctxt->eflags & EFLG_DF) ? -1 : 1;
> >> > +	int df = (ctxt->eflags & EFLG_DF) ? -op->count : op->count;
> >> >  
> >> >  	register_address_increment(ctxt, &ctxt->regs[reg], df * op->bytes);
> >> >  	op->addr.mem.ea = register_address(ctxt, ctxt->regs[reg]);
> >> > @@ -3672,7 +3687,7 @@ static struct opcode opcode_table[256] = {
> >> >  	I(DstReg | SrcMem | ModRM | Src2Imm, em_imul_3op),
> >> >  	I(SrcImmByte | Mov | Stack, em_push),
> >> >  	I(DstReg | SrcMem | ModRM | Src2ImmByte, em_imul_3op),
> >> > -	I2bvIP(DstDI | SrcDX | Mov | String, em_in, ins, check_perm_in), /* insb, insw/insd */
> >> > +	I2bvIP(DstDI | SrcDX | Mov | String | Unaligned, em_in, ins, check_perm_in), /* insb, insw/insd */
> >> 
> >> Eww.
> > This brings us back to the question what alignment check is doing in
> > linearize :)
> 
> It's checking alignment...
> 
It either check it in a wrong place or we need to mark all instructions
that do not care about alignment, so the patch is not "Eww" :)

> Let's see how we would fix this mess.  We need to move linearization
> (and virt->phys translation) to the decode stage, or perhaps the
> execution state, but before instruction dispatch.  This would cause all
> the various exceptions to be checked against before execution, and would
> avoid double translation for RMW operands.
> 
Execution state likely. String instruction works on segmented address
for instance (address increment/decrement). May be there are others.

> 
> >> >  		string_addr_inc(ctxt, VCPU_REGS_RDI, &ctxt->dst);
> >> >  
> >> >  	if (ctxt->rep_prefix && (ctxt->d & String)) {
> >> > +		unsigned int count;
> >> >  		struct read_cache *r = &ctxt->io_read;
> >> > -		register_address_increment(ctxt, &ctxt->regs[VCPU_REGS_RCX], -1);
> >> > +		if ((ctxt->d & SrcMask) == SrcSI)
> >> > +			count = ctxt->src.count;
> >> > +		else
> >> > +			count = ctxt->dst.count;
> >> 
> >> Does this work correctly for 'rep movs' and friends?
> >> 
> > (src|dst).count is initialized to 1 during decode, so anything that does
> > not touch "count" behaves exactly like before.
> 
> Ok.
> 
> 
> -- 
> error compiling committee.c: too many arguments to function

--
			Gleb.