Return-Path: Received: (majordomo@vger.kernel.org) by vger.kernel.org via listexpand id S1752518AbdFLOOt (ORCPT ); Mon, 12 Jun 2017 10:14:49 -0400 Received: from mail-wr0-f193.google.com ([209.85.128.193]:36355 "EHLO mail-wr0-f193.google.com" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S1752089AbdFLOOq (ORCPT ); Mon, 12 Jun 2017 10:14:46 -0400 Subject: Re: Speeding up VMX with GDT fixmap trickery? To: Andy Lutomirski Cc: X86 ML , "linux-kernel@vger.kernel.org" , kvm list , Borislav Petkov , Thomas Garnier , Juergen Gross , Andrew Cooper , Boris Ostrovsky References: From: Paolo Bonzini Message-ID: <1df52354-170f-956e-3203-897f5771319e@redhat.com> Date: Mon, 12 Jun 2017 16:14:35 +0200 User-Agent: Mozilla/5.0 (X11; Linux x86_64; rv:52.0) Gecko/20100101 Thunderbird/52.1.0 MIME-Version: 1.0 In-Reply-To: Content-Type: text/plain; charset=utf-8 Content-Language: en-US Content-Transfer-Encoding: 7bit Sender: linux-kernel-owner@vger.kernel.org List-ID: X-Mailing-List: linux-kernel@vger.kernel.org Content-Length: 2467 Lines: 70 On 09/06/2017 17:45, Andy Lutomirski wrote: > On Fri, Jun 9, 2017 at 12:14 AM, Paolo Bonzini wrote: >> >> >> On 09/06/2017 03:13, Andy Lutomirski wrote: >>> Hi all- >>> >>> As promised when Thomas did his GDT fixmap work, here is a draft patch >>> to speed up KVM by extending it. >>> >>> The downside of this patch is that it makes the fixmap significantly >>> larger on 64-bit systems if NR_CPUS is large (it adds 15 more pages >>> per CPU). I don't know if we care at all. It also bloats the kernel >>> image by 4k and wastes 4k of RAM for the entire time the system is >>> booted. We could avoid the latter bit (sort of) by not mapping the >>> extra fixmap pages at all and handling the resulting faults somehow. >>> That would scare me -- now we have IRET generating #PF when running >>> malicious , and that way lies utter madness. >>> >>> The upside is that we don't need to do LGDT after a vmexit on VMX. >>> LGDT is slooooooooooow. But no, I haven't benchmarked this yet. >>> >>> What do you all think? >>> >>> https://git.kernel.org/pub/scm/linux/kernel/git/luto/linux.git/commit/?h=x86/kvm&id=e249a09787d6956b52d8260b2326d8f12f768799 >> >> Not sure I understand this completely, but: >> >> /* Get the fixmap index for a specific processor */ >> static inline unsigned int get_cpu_gdt_ro_index(int cpu) >> { >> - return FIX_GDT_REMAP_BEGIN + cpu; >> + return FIX_GDT_REMAP_END - cpu * PAGES_PER_GDT; >> } >> >> isn't this off by one. I think it should be >> >> FIX_GDT_REMAP_END + 1 - cpu * PAGES_PER_GDT >> >> or just FIX_GDT_REMAP_BEGIN + cpu * PAGES_PER_GDT? That is for example: >> >> FIX_GDT_REMAP_BEGIN = 100 >> get_cpu_gdt_ro_index(0) = 100 >> get_cpu_gdt_ro_index(1) = 116 >> get_cpu_gdt_ro_index(2) = 132 >> get_cpu_gdt_ro_index(3) = 148 >> FIX_GDT_REMAP_END = 163 > > The issue here is that the fixmap is upside down: lower indices are > *higher* addresses, which means that, if we have a multi-page GDT, we > need get_cpu_gdt_ro_index() to return an index of the lowest page in > each GDT. The simplest way seems to be to put them in ascending > order. > > With the range of indices being 100 .. 163 (with 4 CPUs), we'd want > the GDTs to be at: > > 163..148 > 147..132 > 131..116 > 115..100 > > so FIX_GDT_REMAP_END - cpu * PAGES_PER_GDT is correct, I think. Or am > I still off by one? No, you're right. Thanks for explaining! Paolo