Re: [PATCH v2 15/30] xfs: Define usercopy region in xfs_inode slab cache

[Kees Cook <keescook@chromium.org>]

On Mon, Aug 28, 2017 at 9:47 PM, Darrick J. Wong
<darrick.wong@oracle.com> wrote:
> On Mon, Aug 28, 2017 at 02:57:14PM -0700, Kees Cook wrote:
>> On Mon, Aug 28, 2017 at 2:49 PM, Darrick J. Wong
>> <darrick.wong@oracle.com> wrote:
>> > On Mon, Aug 28, 2017 at 02:34:56PM -0700, Kees Cook wrote:
>> >> From: David Windsor <dave@nullcore.net>
>> >>
>> >> The XFS inline inode data, stored in struct xfs_inode_t field
>> >> i_df.if_u2.if_inline_data and therefore contained in the xfs_inode slab
>> >> cache, needs to be copied to/from userspace.
>> >>
>> >> cache object allocation:
>> >>     fs/xfs/xfs_icache.c:
>> >>         xfs_inode_alloc(...):
>> >>             ...
>> >>             ip = kmem_zone_alloc(xfs_inode_zone, KM_SLEEP);
>> >>
>> >>     fs/xfs/libxfs/xfs_inode_fork.c:
>> >>         xfs_init_local_fork(...):
>> >>             ...
>> >>             if (mem_size <= sizeof(ifp->if_u2.if_inline_data))
>> >>                     ifp->if_u1.if_data = ifp->if_u2.if_inline_data;
>> >
>> > Hmm, what happens when mem_size > sizeof(if_inline_data)?  A slab object
>> > will be allocated for ifp->if_u1.if_data which can then be used for
>> > readlink in the same manner as the example usage trace below.  Does
>> > that allocated object have a need for a usercopy annotation like
>> > the one we're adding for if_inline_data?  Or is that already covered
>> > elsewhere?
>>
>> Yeah, the xfs helper kmem_alloc() is used in the other case, which
>> ultimately boils down to a call to kmalloc(), which is entirely
>> whitelisted by an earlier patch in the series:
>>
>> https://lkml.org/lkml/2017/8/28/1026
>
> Ah.  It would've been helpful to have the first three patches cc'd to
> the xfs list.  So basically this series establishes the ability to set

I went back and forth on that, and given all the things it touched, it
seemed like too large a CC list. :) I can explicitly add the xfs list
to the first three for any future versions.

> regions within a slab object into which copy_to_user can copy memory
> contents, and vice versa.  Have you seen any runtime performance impact?
> The overhead looks like it ought to be minimal.

Under CONFIG_HARDENED_USERCOPY, there's no difference in performance
between the earlier bounds checking (of the whole slab object) vs the
new bounds checking (of the useroffset/usersize portion of the slab
object). Perf difference of CONFIG_HARDENED_USERCOPY itself has proven
hard to measure, which likely means it's very minimal.

>> (It's possible that at some future time we can start segregating
>> kernel-only kmallocs from usercopy-able kmallocs, but for now, there
>> are no plans for this.)
>
> A pity.  It would be interesting to create no-usercopy versions of the
> kmalloc-* slabs and see how much of XFS' memory consumption never
> touches userspace buffers. :)

There are plans for building either a new helper (kmalloc_usercopy())
or adding a new flag (GFP_USERCOPY), but I haven't had time yet to
come back around to it. I wanted to land this step first, and we could
then move forward on the rest in future.

-Kees

-- 
Kees Cook
Pixel Security

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@kvack.org.  For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"dont@kvack.org"> email@kvack.org </a>