Return-Path: Received: (majordomo@vger.kernel.org) by vger.kernel.org via listexpand id S1752119AbdHGVrK (ORCPT ); Mon, 7 Aug 2017 17:47:10 -0400 Received: from mail-io0-f174.google.com ([209.85.223.174]:33837 "EHLO mail-io0-f174.google.com" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S1751662AbdHGVrJ (ORCPT ); Mon, 7 Aug 2017 17:47:09 -0400 MIME-Version: 1.0 In-Reply-To: References: <20170807203948.GA22298@beast> From: Kees Cook Date: Mon, 7 Aug 2017 14:47:08 -0700 X-Google-Sender-Auth: AkK77yMQpcN_hOVZdEHYnLfkrHY Message-ID: Subject: Re: [kernel-hardening] [PATCH] lkdtm: Test VMAP_STACK allocates leading/trailing guard pages To: Ard Biesheuvel Cc: "linux-kernel@vger.kernel.org" , Mark Rutland , Catalin Marinas , James Morse , Laura Abbott , Andy Lutomirski , Matt Fleming , Will Deacon , Kernel Hardening , "linux-arm-kernel@lists.infradead.org" Content-Type: text/plain; charset="UTF-8" Sender: linux-kernel-owner@vger.kernel.org List-ID: X-Mailing-List: linux-kernel@vger.kernel.org Content-Length: 1726 Lines: 41 On Mon, Aug 7, 2017 at 2:46 PM, Ard Biesheuvel wrote: > On 7 August 2017 at 22:44, Kees Cook wrote: >> On Mon, Aug 7, 2017 at 2:27 PM, Ard Biesheuvel >> wrote: >>> On 7 August 2017 at 21:39, Kees Cook wrote: >>>> Two new tests STACK_GUARD_PAGE_LEADING and STACK_GUARD_PAGE_TRAILING >>>> attempt to read the byte before and after, respectively, of the current >>>> stack frame, which should fault under VMAP_STACK. >>>> >>>> Signed-off-by: Kees Cook >>>> --- >>>> Do these tests both trip with the new arm64 VMAP_STACK code? >>> >>> Probably not. On arm64, the registers are stacked by software at >>> exception entry, and so we decrement the sp first by the size of the >>> register file, and if the resulting value overflows the stack, the >>> situation is handled as if the exception was caused by a faulting >>> stack access while it may be caused by something else in reality. >>> Since the act of handling the exception is guaranteed to overflow the >>> stack anyway, this does not really make a huge difference, and it >>> prevents the recursive fault from wiping the context that we need to >>> produce the diagnostics. >>> >>> This means an illegal access right above the stack will go undetected. >> >> I thought vmap entries provided guard pages around allocations? >> Shouldn't that catch it? >> > > Ah yes, so we will fault. We should probably double check whether we > will not misidentify the fault because of the subtraction we do first, > but that should be trivial to add. Okay, cool. I'd be curious to see what the lkdtm tests show for you. -Kees -- Kees Cook Pixel Security