Re: [PATCH v4] KVM: x86: avoid large stack allocations in em_fxrstor

["Radim Krčmář" <rkrcmar@redhat.com>]

2017-05-30 12:15+0200, Paolo Bonzini:
> On 30/05/2017 00:48, Nick Desaulniers wrote:
>> diff --git a/arch/x86/kvm/emulate.c b/arch/x86/kvm/emulate.c
>> @@ -3985,57 +3985,45 @@ static int em_fxsave(struct x86_emulate_ctxt *ctxt)
>>  static int em_fxrstor(struct x86_emulate_ctxt *ctxt)
>>  {
>>  	struct fxregs_state fx_state;
>>  	int rc;
>> +	unsigned int size;
>>  
>>  	rc = check_fxsr(ctxt);
>>  	if (rc != X86EMUL_CONTINUE)
>>  		return rc;
>>  
>> -	rc = segmented_read_std(ctxt, ctxt->memop.addr.mem, &fx_state, 512);
>> -	if (rc != X86EMUL_CONTINUE)
>> -		return rc;
>> +	ctxt->ops->get_fpu(ctxt);
>> +
>> +	if (ctxt->mode < X86EMUL_MODE_PROT64) {
>> +		rc = asm_safe("fxsave %[fx]", , [fx] "+m"(fx_state));
>> +		if (rc != X86EMUL_CONTINUE)
>> +			return rc;

Please call ctxt->ops->put_fpu() before returning.
(Don't be afraid to use 'goto', the will likely be more of them.)

>> +		/*
>> +		 * Hardware doesn't save and restore XMM 0-7 without
>> +		 * CR4.OSFXSR, but does save and restore MXCSR.
>> +		 */
>> +		if (ctxt->ops->get_cr(ctxt, 4) & X86_CR4_OSFXSR)
>> +			size = offsetof(struct fxregs_state, xmm_space[8]);

This should be the size of first 8 XMM registers, but xmm_space is of
type u32, so the correct size is
  xmm_space[8 * 16/sizeof(*fx_state.xmm_space)].

Adding a separate function to compute the size and call it from
em_fxrstor and em_fxsave would make sense at this point.

>> +		else
>> +			size = offsetof(struct fxregs_state, xmm_space[0]);
>> +		rc = segmented_read_std(ctxt, ctxt->memop.addr.mem, &fx_state,
>> +			size);
>> +		if (rc != X86EMUL_CONTINUE)
>> +			return rc;
>> +	} else if (ctxt->mode == X86EMUL_MODE_PROT64) {
>> +		size = offsetof(struct fxregs_state, xmm_space[16]);
>> +		rc = segmented_read_std(ctxt, ctxt->memop.addr.mem, &fx_state,
>> +			size);
>> +		if (rc != X86EMUL_CONTINUE)
>> +			return rc;
>> +	}
> 
> You can just remove the "if (ctxt->mode == X86EMUL_MODE_PROT64)".  This
> way, the call of segmented_read_std and the "if (rc !=
> X86EMUL_CONTINUE)" can move outside the conditional.

Good idea.  check_fxsr() doesn't allow X86EMUL_MODE_PROT64 and the
condition was an artefact from older iterations.

I'll refresh the kvm-unit-test ([kvm-unit-tests PATCH] x86: realmode:
add FXSR tests).