[PATCH 04/13] Always expose MAP_UNINITIALIZED to userspace

[Josh Triplett]

On Tue, Sep 15, 2015 at 12:42:00PM +0300, Kirill A. Shutemov wrote:
> On Mon, Sep 14, 2015 at 10:19:19PM -0700, Josh Triplett wrote:
> > On Tue, Sep 15, 2015 at 03:23:58AM +0300, Kirill A. Shutemov wrote:
> > > On Mon, Sep 14, 2015 at 03:50:38PM -0700, Palmer Dabbelt wrote:
> > > > This used to be hidden behind CONFIG_MMAP_ALLOW_UNINITIALIZED, so
> > > > userspace wouldn't actually ever see it be non-zero.  While I could
> > > > have kept hiding it, the man pages seem to indicate that
> > > > MAP_UNINITIALIZED should be visible:
> > > > 
> > > >   mmap(2)
> > > >   MAP_UNINITIALIZED (since Linux 2.6.33)
> > > >     Don't clear anonymous pages.  This flag is intended to improve
> > > >     performance on embedded devices.  This flag is honored only if the
> > > >     kernel was configured with the CONFIG_MMAP_ALLOW_UNINITIALIZED
> > > >     option.  Because of the security implications, that option is
> > > >     normally enabled only on embedded devices (i.e., devices where one
> > > >     has complete control of the contents of user memory).
> > > > 
> > > > and since the only time it shows up in my /usr/include is in this
> > > > header I believe this should have been visible to userspace (as
> > > > non-zero, which wouldn't do anything when or'd into the flags) all
> > > > along.
> > > 
> > > Are you sure about "wouldn't do anything"?
> > > Suspiciously, 0x4000000 is also (1 << MAP_HUGE_SHIFT). I'm not sure if any
> > > architecture has order-1 huge pages, but still looks like we have conflict
> > > here.
> > > 
> > > I think it's harmful to expose non-zero MAP_UNINITIALIZED to system which
> > > potentially can handle multiple users. Or non-trivial user space in
> > > general.
> > 
> > The flag should always exist.
> 
> Sure. And 0 is perfectly fine value for the flag. Like with MAP_FILE.

Rephrasing: the flag should always exist with the correct value.
Whether the kernel handles it or not, the kernel *headers* shouldn't
change to match the kernel, not least of which because they don't
necessarily match the running kernel.  Just like we define the
prototypes for syscalls that the running kernel may return ENOSYS for.

> > If it was defined to conflict with
> > something else, that's a serious ABI problem.  But the flag
> > should always exist, even if the kernel ends up ignoring it.
> > 
> > > Should we leave it at least under '#ifndef CONFIG_MMU'? I don't think it's
> > > possible to have single ABI for MMU and MMU-less systems anyway. And we
> > > can avoid conflict with MAP_HUGE_SHIFT this way.
> > 
> > No; even if you have an MMU (which is useful for things like fork()), a
> > system without user separation (for instance, without CONFIG_MULTIUSER)
> > can reasonably use MAP_UNINITIALIZED.
> 
> Can? Yes. Reasonably? I don't think so.

Not all systems care.  Otherwise you should be complaining more bitterly
about options like CONFIG_MMU=n, which (*gasp*) allow access to *arbitrary
memory*.

> > > P.S. MAP_UNINITIALIZED itself looks very broken to me. I probably need dig
> > > mailing list on why it was allowed.
> > 
> > That's what the config option *and* explicit flag are for; there are
> > more than enough warnings about the implications.
> 
> I think it's misdesigned. It doesn't require explicid opt-in from a
> process who owned the page allocated in MAP_UNINITIALIZED mapping before.
> 
> #define MAP_LEAK_ME_SOME_DATA MAP_UNINITIALIZED

Hence why it has a config option.

The userspace option exists primarily because otherwise userspace might
get surprised by receiving a non-zeroed page.  On a system with the
config option turned on, processes have access to arbitrary freed
memory, as long as they say they can handle not having their memory
pre-zeroed.

- Josh Triplett