From mboxrd@z Thu Jan 1 00:00:00 1970 Return-Path: Received: from mail-pf1-f199.google.com (mail-pf1-f199.google.com [209.85.210.199]) by kanga.kvack.org (Postfix) with ESMTP id 1A7EF6B7CF4 for ; Thu, 6 Dec 2018 18:08:08 -0500 (EST) Received: by mail-pf1-f199.google.com with SMTP id p9so1617155pfj.3 for ; Thu, 06 Dec 2018 15:08:08 -0800 (PST) Received: from mail-sor-f65.google.com (mail-sor-f65.google.com. [209.85.220.65]) by mx.google.com with SMTPS id j193sor3007109pge.29.2018.12.06.15.08.06 for (Google Transport Security); Thu, 06 Dec 2018 15:08:06 -0800 (PST) Content-Type: text/plain; charset=utf-8 Mime-Version: 1.0 (Mac OS X Mail 12.1 \(3445.101.1\)) Subject: Re: [PATCH 1/2] vmalloc: New flag for flush before releasing pages From: Nadav Amit In-Reply-To: Date: Thu, 6 Dec 2018 15:08:02 -0800 Content-Transfer-Encoding: quoted-printable Message-Id: References: <20181128000754.18056-1-rick.p.edgecombe@intel.com> <20181128000754.18056-2-rick.p.edgecombe@intel.com> <4883FED1-D0EC-41B0-A90F-1A697756D41D@gmail.com> <20181204160304.GB7195@arm.com> <51281e69a3722014f718a6840f43b2e6773eed90.camel@intel.com> <20181205114148.GA15160@arm.com> <20181206190115.GC10086@cisco> Sender: owner-linux-mm@kvack.org List-ID: To: Andy Lutomirski Cc: Tycho Andersen , Ard Biesheuvel , Will Deacon , Rick Edgecombe , LKML , Daniel Borkmann , Jessica Yu , Steven Rostedt , Alexei Starovoitov , Linux-MM , Jann Horn , "Dock, Deneen T" , Peter Zijlstra , Kristen Carlson Accardi , Andrew Morton , Ingo Molnar , Anil S Keshavamurthy , Kernel Hardening , Masami Hiramatsu , "Naveen N . Rao" , "David S. Miller" , Network Development , Dave Hansen , Igor Stoppa > On Dec 6, 2018, at 12:17 PM, Andy Lutomirski wrote: >=20 > On Thu, Dec 6, 2018 at 11:39 AM Nadav Amit = wrote: >>> On Dec 6, 2018, at 11:19 AM, Andy Lutomirski = wrote: >>>=20 >>> On Thu, Dec 6, 2018 at 11:01 AM Tycho Andersen = wrote: >>>> On Thu, Dec 06, 2018 at 10:53:50AM -0800, Andy Lutomirski wrote: >>>>>> If we are going to unmap the linear alias, why not do it at = vmalloc() >>>>>> time rather than vfree() time? >>>>>=20 >>>>> That=E2=80=99s not totally nuts. Do we ever have code that expects = __va() to >>>>> work on module data? Perhaps crypto code trying to encrypt static >>>>> data because our APIs don=E2=80=99t understand virtual addresses. = I guess if >>>>> highmem is ever used for modules, then we should be fine. >>>>>=20 >>>>> RO instead of not present might be safer. But I do like the idea = of >>>>> renaming Rick's flag to something like VM_XPFO or VM_NO_DIRECT_MAP = and >>>>> making it do all of this. >>>>=20 >>>> Yeah, doing it for everything automatically seemed like it was/is >>>> going to be a lot of work to debug all the corner cases where = things >>>> expect memory to be mapped but don't explicitly say it. And in >>>> particular, the XPFO series only does it for user memory, whereas = an >>>> additional flag like this would work for extra paranoid allocations >>>> of kernel memory too. >>>=20 >>> I just read the code, and I looks like vmalloc() is already using >>> highmem (__GFP_HIGH) if available, so, on big x86_32 systems, for >>> example, we already don't have modules in the direct map. >>>=20 >>> So I say we go for it. This should be quite simple to implement -- >>> the pageattr code already has almost all the needed logic on x86. = The >>> only arch support we should need is a pair of functions to remove a >>> vmalloc address range from the address map (if it was present in the >>> first place) and a function to put it back. On x86, this should = only >>> be a few lines of code. >>>=20 >>> What do you all think? This should solve most of the problems we = have. >>>=20 >>> If we really wanted to optimize this, we'd make it so that >>> module_alloc() allocates memory the normal way, then, later on, we >>> call some function that, all at once, removes the memory from the >>> direct map and applies the right permissions to the vmalloc alias = (or >>> just makes the vmalloc alias not-present so we can add permissions >>> later without flushing), and flushes the TLB. And we arrange for >>> vunmap to zap the vmalloc range, then put the memory back into the >>> direct map, then free the pages back to the page allocator, with the >>> flush in the appropriate place. >>>=20 >>> I don't see why the page allocator needs to know about any of this. >>> It's already okay with the permissions being changed out from under = it >>> on x86, and it seems fine. Rick, do you want to give some variant = of >>> this a try? >>=20 >> Setting it as read-only may work (and already happens for the = read-only >> module data). I am not sure about setting it as non-present. >>=20 >> At some point, a discussion about a threat-model, as Rick indicated, = would >> be required. I presume ROP attacks can easily call = set_all_modules_text_rw() >> and override all the protections. >=20 > I am far from an expert on exploit techniques, but here's a > potentially useful model: let's assume there's an attacker who can > write controlled data to a controlled kernel address but cannot > directly modify control flow. It would be nice for such an attacker > to have a very difficult time of modifying kernel text or of > compromising control flow. So we're assuming a feature like kernel > CET or that the attacker finds it very difficult to do something like > modifying some thread's IRET frame. >=20 > Admittedly, for the kernel, this is an odd threat model, since an > attacker can presumably quite easily learn the kernel stack address of > one of their tasks, do some syscall, and then modify their kernel > thread's stack such that it will IRET right back to a fully controlled > register state with RSP pointing at an attacker-supplied kernel stack. > So this threat model gives very strong ROP powers. unless we have > either CET or some software technique to harden all the RET > instructions in the kernel. >=20 > I wonder if there's a better model to use. Maybe with stack-protector > we get some degree of protection? Or is all of this is rather weak > until we have CET or a RAP-like feature. I believe that seeing the end-goal would make reasoning about patches easier, otherwise the complaint =E2=80=9Cbut anyhow it=E2=80=99s all = insecure=E2=80=9D keeps popping up. I=E2=80=99m not sure CET or other CFI would be enough even with this = threat-model. The page-tables (the very least) need to be write-protected, as = otherwise controlled data writes may just modify them. There are various possible solutions I presume: write_rare for page-tables, hypervisor-assisted security to obtain physical level NX/RO (a-la Microsoft VBS) or some = sort of hardware enclave. What do you think?=