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

[Sami Tolvanen]

On Thu, Nov 19, 2020 at 5:00 AM Will Deacon <will@xxxxxxxxxx> wrote:
>
> Hi Sami,
>
> 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@xxxxxxxxxx>
> > ---
> >  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?

> >  #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().

> instead, then we could avoid the expensive call to vmalloc_to_page()
> in __scs_account().

We still need vmalloc_to_page() in scs_release(). I suppose we could
alternatively follow the example in kernel/fork.c and cache the
vm_struct from find_vm_area() and use vm->pages[0] for the accounting.
Thoughts?

>
> >
> >       if (!s)
> >               return NULL;
> >
> > +out:
> >       *__scs_magic(s) = SCS_END_MAGIC;
> >
> >       /*
> >        * Poison the allocation to catch unintentional accesses to
> >        * the shadow stack when KASAN is enabled.
> >        */
> > -     kasan_poison_object_data(scs_cache, s);
> > +     kasan_poison_vmalloc(s, SCS_SIZE);
> >       __scs_account(s, 1);
> >       return s;
> >  }
> >
> >  static void scs_free(void *s)
> >  {
> > +     int i;
> > +
> >       __scs_account(s, -1);
> > -     kasan_unpoison_object_data(scs_cache, s);
> > -     kmem_cache_free(scs_cache, s);
> > +     kasan_unpoison_vmalloc(s, SCS_SIZE);
>
> I don't see the point in unpoisoning here tbh; vfree_atomic() re-poisons
> almost immediately, so we should probably defer this to scs_alloc() and
> only when picking the stack out of the cache.

Sure, I'll change this in v2.

>
> > +
> > +     for (i = 0; i < NR_CACHED_SCS; i++)
>
> Can you add a comment about the re-entrancy here and why we're using
> this_cpu_cmpxchg() please?

I'll add a comment.

Sami