Re: [PATCH v3] mm: memdup_user*() should use same gfp flags

[Michal Hocko <mhocko@suse.com>]

On Wed 27-01-21 15:19:40, Andrew Morton wrote:
> On Wed, 27 Jan 2021 23:03:33 +0900 Tetsuo Handa <penguin-kernel@i-love.sakura.ne.jp> wrote:
> 
> > On 2021/01/27 21:17, Michal Hocko wrote:
> > > On Wed 27-01-21 12:59:28, Michal Hocko wrote:
> > >> On Wed 27-01-21 19:55:38, Tetsuo Handa wrote:
> > >>> syzbot is reporting that memdup_user_nul() which receives user-controlled
> > >>> size (which can be up to (INT_MAX & PAGE_MASK)) via vfs_write() will hit
> > >>> order >= MAX_ORDER path [1].
> > >>>
> > >>> Making costly allocations (order > PAGE_ALLOC_COSTLY_ORDER) naturally fail
> > >>> should be better than trying to enforce PAGE_SIZE upper limit, for some of
> > >>> callers accept space-delimited list arguments.
> > >>>
> > >>> Therefore, let's add __GFP_NOWARN to memdup_user_nul() as with
> > >>> commit 6c8fcc096be9d02f ("mm: don't let userspace spam allocations
> > >>> warnings"). Also use GFP_USER as with other userspace-controllable
> > >>> allocations like memdup_user().
> > >>
> > >> I absolutely detest hiding this behind __GFP_NOWARN. There should be no
> > >> reason to even try hard for memdup_user_nul. Can you explain why this
> > > 
> > > this should have been "try hard to get a physicaly contiguous memory for memdup_user_nul"
> > > 
> > >> cannot use kvmalloc instead?
> > > 
> > 
> > There is no point with allowing userspace to allocate 2GB of physically non-contiguous
> > memory using kvmalloc(). Size is controlled by userspace, and memdup_user_nul() is used
> > for allocating temporary memory which will be released before returning to userspace.
> > 
> > Sane userspace processes should allocate only one or a few pages using memdup_user_nul().
> > Just making insane user processes (like fuzzer) fail memory allocation requests is a
> > reasonable decision.
> 
> (cc Casey)
> 
> I'd say that the immediate problem is in smk_write_syslog().  Obviously
> it was implemented expecting small writes, but the fuzzer is passing it a
> huge write and things fall apart.

I am not familiar with this particular caller and having a limit check
which suits that particular usage is a reasonable thing to do.

I do argue two things
- using NOWARN to work around potentially buggy callers is just sweeping
  the mess under the rug and opens
- these helper functions are to help copy user input and that doesn't
  really need physically contiguous pages. This can even become
  dangerous as a higher order depleting vector and DoS via OOM in  the
  worst case.

From that it sounds natural that the helper should be using kvmalloc.
This will not solve a due size check on the caller side but that is not
possible from a generic helper library function anyway. But it will
provide a reasonable allocation policy.
-- 
Michal Hocko
SUSE Labs