linux-kernel - Re: [PATCH 1/2] scs: switch to vmapped shadow stacks

lists.openwall.net		lists / announce owl-users owl-dev john-users john-dev passwdqc-users yescrypt popa3d-users / oss-security kernel-hardening musl sabotage tlsify passwords / crypt-dev xvendor / Bugtraq Full-Disclosure linux-kernel linux-netdev linux-ext4 linux-hardening linux-cve-announce PHC
Open Source and information security mailing list archives
Hash Suite: Windows password security audit tool. GUI, reports in PDF.
[<prev] [next>] [<thread-prev] [day] [month] [year] [list]
Message-ID: <20201123110811.GA9957@willie-the-truck>
Date:   Mon, 23 Nov 2020 11:08:12 +0000
From:   Will Deacon <will@...nel.org>
To:     Sami Tolvanen <samitolvanen@...gle.com>
Cc:     Catalin Marinas <catalin.marinas@....com>,
        Mark Rutland <mark.rutland@....com>,
        James Morse <james.morse@....com>,
        Ard Biesheuvel <ard.biesheuvel@...aro.org>,
        Kees Cook <keescook@...omium.org>,
        linux-arm-kernel <linux-arm-kernel@...ts.infradead.org>,
        LKML <linux-kernel@...r.kernel.org>
Subject: Re: [PATCH 1/2] scs: switch to vmapped shadow stacks

Hi Sami,

On Fri, Nov 20, 2020 at 09:00:17AM -0800, Sami Tolvanen wrote:
> On Thu, Nov 19, 2020 at 5:00 AM Will Deacon <will@...nel.org> wrote:
> > On Thu, Oct 22, 2020 at 01:23:54PM -0700, Sami Tolvanen wrote:
> > > The kernel currently uses kmem_cache to allocate shadow call stacks,
> > > which means an overflow may not be immediately detected and can
> > > potentially result in another task's shadow stack to be overwritten.
> > >
> > > This change switches SCS to use virtually mapped shadow stacks,
> > > which increases shadow stack size to a full page and provides more
> > > robust overflow detection similarly to VMAP_STACK.
> > >
> > > Signed-off-by: Sami Tolvanen <samitolvanen@...gle.com>
> > > ---
> > >  include/linux/scs.h |  7 +----
> > >  kernel/scs.c        | 63 ++++++++++++++++++++++++++++++++++++++-------
> > >  2 files changed, 55 insertions(+), 15 deletions(-)
> >
> > Cheers for posting this. I _much_ prefer handling the SCS this way, but I
> > have some comments on the implementation below.
> >
> > > diff --git a/include/linux/scs.h b/include/linux/scs.h
> > > index 6dec390cf154..86e3c4b7b714 100644
> > > --- a/include/linux/scs.h
> > > +++ b/include/linux/scs.h
> > > @@ -15,12 +15,7 @@
> > >
> > >  #ifdef CONFIG_SHADOW_CALL_STACK
> > >
> > > -/*
> > > - * In testing, 1 KiB shadow stack size (i.e. 128 stack frames on a 64-bit
> > > - * architecture) provided ~40% safety margin on stack usage while keeping
> > > - * memory allocation overhead reasonable.
> > > - */
> > > -#define SCS_SIZE             SZ_1K
> > > +#define SCS_SIZE             PAGE_SIZE
> >
> > We could make this SCS_ORDER and then forget about alignment etc.
> 
> It's still convenient to have SCS_SIZE defined, I think. I can
> certainly define SCS_ORDER and use that to define SCS_SIZE, but do you
> think we'll need an order >0 here at some point in future?

I'm not daft enough to comment on SCS size again ;)
Let's define SCS_ORDER 0 and then SCS_SIZE in terms of that.

> 
> > >  #define GFP_SCS                      (GFP_KERNEL | __GFP_ZERO)
> > >
> > >  /* An illegal pointer value to mark the end of the shadow stack. */
> > > diff --git a/kernel/scs.c b/kernel/scs.c
> > > index 4ff4a7ba0094..2136edba548d 100644
> > > --- a/kernel/scs.c
> > > +++ b/kernel/scs.c
> > > @@ -5,50 +5,95 @@
> > >   * Copyright (C) 2019 Google LLC
> > >   */
> > >
> > > +#include <linux/cpuhotplug.h>
> > >  #include <linux/kasan.h>
> > >  #include <linux/mm.h>
> > >  #include <linux/scs.h>
> > > -#include <linux/slab.h>
> > > +#include <linux/vmalloc.h>
> > >  #include <linux/vmstat.h>
> > >
> > > -static struct kmem_cache *scs_cache;
> > > -
> > >  static void __scs_account(void *s, int account)
> > >  {
> > > -     struct page *scs_page = virt_to_page(s);
> > > +     struct page *scs_page = vmalloc_to_page(s);
> > >
> > >       mod_node_page_state(page_pgdat(scs_page), NR_KERNEL_SCS_KB,
> > >                           account * (SCS_SIZE / SZ_1K));
> > >  }
> > >
> > > +/* Matches NR_CACHED_STACKS for VMAP_STACK */
> > > +#define NR_CACHED_SCS 2
> > > +static DEFINE_PER_CPU(void *, scs_cache[NR_CACHED_SCS]);
> > > +
> > >  static void *scs_alloc(int node)
> > >  {
> > > -     void *s = kmem_cache_alloc_node(scs_cache, GFP_SCS, node);
> > > +     int i;
> > > +     void *s;
> > > +
> > > +     for (i = 0; i < NR_CACHED_SCS; i++) {
> > > +             s = this_cpu_xchg(scs_cache[i], NULL);
> > > +             if (s) {
> > > +                     memset(s, 0, SCS_SIZE);
> > > +                     goto out;
> > > +             }
> > > +     }
> > > +
> > > +     /*
> > > +      * We allocate a full page for the shadow stack, which should be
> > > +      * more than we need. Check the assumption nevertheless.
> > > +      */
> > > +     BUILD_BUG_ON(SCS_SIZE > PAGE_SIZE);i
> >
> > With SCS_ORDER, you can drop this.
> >
> > > +
> > > +     s = __vmalloc_node_range(PAGE_SIZE, SCS_SIZE,
> > > +                              VMALLOC_START, VMALLOC_END,
> > > +                              GFP_SCS, PAGE_KERNEL, 0,
> > > +                              node, __builtin_return_address(0));
> >
> > Do we actually need vmalloc here? If we used alloc_pages() + vmap()
> 
> Does it matter that vmap() always uses NUMA_NO_NODE? We'll also lose
> the ability to use vfree_atomic() in scs_release() unless we use
> VM_MAP_PUT_PAGES and allocate the page array passed to vmap() with
> kvmalloc(), which I think we need to do to avoid sleeping in
> scs_free().

Huh, I didn't realise we didn't have vunmap_atomic(). In which case, I take
that back -- let's stick with vmalloc() for now.

Cheers,

Will