DMARC-Filter: OpenDMARC Filter v1.3.2 mx1.redhat.com EF6E7E233A
Date: Tue, 1 Aug 2017 19:37:39 +0200
From: Radim =?utf-8?B?S3LEjW3DocWZ?= <rkrcmar@redhat.com>
To: Paolo Bonzini <pbonzini@redhat.com>
Cc: David Hildenbrand <david@redhat.com>, linux-kernel@vger.kernel.org,
        kvm@vger.kernel.org
Subject: Re: [PATCH] KVM: nVMX: INVPCID support
Message-ID: <20170801173739.GA316@flask>
References: <1501161639-9707-1-git-send-email-pbonzini@redhat.com>
 <0b82ca85-7b3e-9dc4-a102-24d0e9033dcd@redhat.com>
 <42be1e05-b6e4-42db-5ed8-276d61c4334d@redhat.com>
 <60a8347a-744d-47d3-e53b-3218f5c50623@redhat.com>
 <ae1bebea-43d1-6c69-ff0d-c1168450a202@redhat.com>
MIME-Version: 1.0
Content-Type: text/plain; charset=us-ascii
Content-Disposition: inline
In-Reply-To: <ae1bebea-43d1-6c69-ff0d-c1168450a202@redhat.com>
Sender: linux-kernel-owner@vger.kernel.org
Content-Length: 2990
Lines: 86

2017-08-01 13:35+0200, Paolo Bonzini:
> On 01/08/2017 13:18, David Hildenbrand wrote:
> > 
> >>> Can't we rewrite that a little bit, avoiding that "best" handling
> >>> (introducing guest_cpuid_disable_invpcid() and guest_cpuid_has_invpcid())
> >>>
> >>> bool invpcid_enabled = guest_cpuid_has_pcid(vcpu) &&
> >>> 		       guest_cpuid_has_invpcid();
> >>>
> >>> if (!invpcid_enabled) {
> >>> 	secondary_exec_ctl &= ~SECONDARY_EXEC_ENABLE_INVPCID;
> >>> 	/* make sure there is no no INVPCID without PCID */
> >>> 	guest_cpuid_disable_invpcid(vcpu);
> >>> }
> >>
> >> I don't know... if you need a comment, it means the different structure
> >> of the code doesn't spare many doubts from the reader.  And the code
> >> doesn't become much simpler since you have to handle "nested" anyway.
> >> What I tried to do was to mimic as much as possible the rdtscp case, but
> >> it cannot be exactly the same due to the interaction between PCID and
> >> INVPCID.
> > 
> > It's more about the handling of best here, which can be avoided quite
> > easily as I showed (by encapsulating how cpuids are looked up/modified).
> 
> Yeah, I don't like either option. :)  Luckily there is a second maintainer!

With three people, we'll just have three options! :)
I'd go with Paolo's version, because it is efficient and also follows
patterns the bit clearing in kvm_update_cpuid().

---
I would like the wrappers if they were more generic, e.g.:

   guest_cpuid_has(vcpu, X86_FEATURE_INVPCID);
   guest_cpuid_clear(vcpu, X86_FEATURE_INVPCID);

To do that, we need a mapping from X86_FEATURE to (leaf, subleaf,
register) triple, which can be done with cpuid_leafs.

I haven't tried to compile the idea below, but if the optimizer is good,
then we should have only slightly worse byte code as we do now thanks to
x86_feature being a compile-time constant.

Do you it's less ugly than the other two options?


static inline bool x86_feature_cpuid(int x86_feature, int *id, int *count, int *register)
{
	static struct {int x86_leaf; int id; int count; int register;} cpuid[] = {
		{CPUID_1_EDX, 1, 0, 3},
		{CPUID_8000_0001_EDX, 0x80000001, 0, 3},
		{CPUID_8086_0001_EDX, 0x80860001, 0, 3},
		{CPUID_1_ECX, 1, 0, 2},
		... // you get the idea, it's tedious :)
		};
	
	for (size_t i = 0; i < ARRAY_SIZE(cpuid); i++)
		if (cpuid[i].x86_leaf == x86_feature / 32) {
			*id = cpuid[i].id;
			*count = cpuid[i].count;
			*register = cpuid[i].register;
			return true;
		}
	
	return false;
}

static inline bool guest_cpuid_has(struct kvm_vcpu *vcpu, int x86_feature)
{
	int id, count, register;
	u32 *reg;
	struct kvm_cpuid_entry2 *best;

	if (!x86_feature_cpuid(x86_feature, &id, &count, &register))
		return false;

	if (!(best = kvm_find_cpuid_entry(id, count)))
		return false;

	reg = &best->eax + register; // rewrite to be safer
	return *reg & bit(x86_feature);
}

Similar for guest_cpuid_clear and we can also factor out the common part
for getting the register (all up to the last line, with NULL as failure)