From mboxrd@z Thu Jan  1 00:00:00 1970
From: Paolo Bonzini <pbonzini@redhat.com>
Subject: Re: [PATCH 9/9] KVM: VMX: automatic PLE window maximum
Date: Wed, 20 Aug 2014 18:03:30 +0200
Message-ID: <53F4C6D2.9050209@redhat.com>
References: <1408480536-8240-1-git-send-email-rkrcmar@redhat.com> <1408480536-8240-10-git-send-email-rkrcmar@redhat.com> <53F44B40.6060806@redhat.com> <20140820124112.GC28873@potion.brq.redhat.com> <53F49F80.1090408@redhat.com> <20140820153115.GA20453@potion.brq.redhat.com> <53F4BFF3.3060606@redhat.com> <20140820160137.GB20453@potion.brq.redhat.com>
Mime-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: QUOTED-PRINTABLE
Cc: kvm@vger.kernel.org, linux-kernel@vger.kernel.org,
	Gleb Natapov <gleb@kernel.org>,
	Raghavendra KT <raghavendra.kt@linux.vnet.ibm.com>,
	Vinod Chegu <chegu_vinod@hp.com>, Hui-Zhi <hui-zhi.zhao@hp.com>
To: =?UTF-8?B?UmFkaW0gS3LEjW3DocWZ?= <rkrcmar@redhat.com>
Return-path: <linux-kernel-owner@vger.kernel.org>
In-Reply-To: <20140820160137.GB20453@potion.brq.redhat.com>
Sender: linux-kernel-owner@vger.kernel.org
List-Id: kvm.vger.kernel.org

Il 20/08/2014 18:01, Radim Kr=C4=8Dm=C3=A1=C5=99 ha scritto:
> 2014-08-20 17:34+0200, Paolo Bonzini:
>> Il 20/08/2014 17:31, Radim Kr=C4=8Dm=C3=A1=C5=99 ha scritto:
>>> Btw. without extra code, we are still going to overflow on races wh=
en
>>> changing PW_grow, should they be covered as well?
>>
>> You mean because there is no spinlock or similar protecting the chan=
ges?
>>  I guess you could use a seqlock.
>=20
> Yes, for example between a modification of ple_window
>   new =3D min(old, PW_actual_max) * PW_grow
> which gets compiled into something like this:
>   1) tmp =3D min(old, PW_actual_max)
>   2) new =3D tmp * PW_grow
> and a write to increase PW_grow
>   3) PW_actual_max =3D min(PW_max / new_PW_grow, PW_actual_max)
>   4) PW_grow =3D new_PW_grow
>   5) PW_actual_max =3D PW_max / new_PW_grow
>=20
> 3 and 4 can exectute between 1 and 2, which could overflow.
>=20
> I don't think they are important enough to warrant a significant
> performance hit of locking.

A seqlock just costs two memory accesses to the same (shared) cache lin=
e
as the PW data, and a non-taken branch.  I don't like code that is
unsafe by design...

Paolo

> Or even more checks that would prevent it in a lockless way.
>=20
> (I'd just see that the result is set to something legal and also drop
>  line 3, because it does not help things that much.)
>=20