On Mon, 14 Sep 2015 17:23:58 PDT (-0700), kirill@shutemov.name wrote:

On Mon, Sep 14, 2015 at 03:50:38PM -0700, Palmer Dabbelt wrote:

quoted

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"?

That was bad writing for me.  I'd originally written something like "I
believe this should have been visible to userspace all along", but
then added the ()'s.  I meant to say:

 * I think MAP_UNINITIALIZED should have been non-zero in userspace.
 * MAP_UNINITAILIZED was zero in userspace.
 * A zero MAP_UNINITIALIZED does nothing when OR'd in.

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.

This doesn't have MAP_UNINITIALIZED do anything by default, it just
defines the flag the same way on all systems.  I was under the
impression that this just happened if I set MAP_UNINITIALIZED.
Looking at MAP_HUGE_SHIFT it mmap.c, that's definitely why my mmap()
test case ignored the set MAP_UNINITIALIZED on my PC.

I'm going to make this

 #ifndef MAP_UNINITAILIZED
 #define MAP_UNINITAILIZED 0
 #endif

and then leave Xtensa's port alone.  This is what Arnd suggested
originally, sorry for the extra work!

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.

The whole goal here was to eliminate "#ifndef CONFIG_*" from the
user-visible headers.  This all started because I got bit by a very
similar-looking bug (see patch #1), so I'd prefer not to go down that
route.

P.S. MAP_UNINITIALIZED itself looks very broken to me. I probably need dig
mailing list on why it was allowed.
But that's other topic.