lkml.org 
[lkml]   [2012]   [Dec]   [16]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
From
SubjectRe: [PATCH] mm: Downgrade mmap_sem before locking or populating on mmap
On Fri, Dec 14, 2012 at 08:12:45AM -0800, Andy Lutomirski wrote:
> On Fri, Dec 14, 2012 at 6:49 AM, Al Viro <viro@zeniv.linux.org.uk> wrote:
> > On Fri, Dec 14, 2012 at 03:14:50AM -0800, Andy Lutomirski wrote:
> >
> >> > Wait a minute. get_user_pages() relies on ->mmap_sem being held. Unless
> >> > I'm seriously misreading your patch it removes that protection. And yes,
> >> > I'm aware of execve-related exception; it's in special circumstances -
> >> > bprm->mm is guaranteed to be not shared (and we need to rearchitect that
> >> > area anyway, but that's a separate story).
> >>
> >> Unless I completely screwed up the patch, ->mmap_sem is still held for
> >> read (it's downgraded from write). It's just not held for write
> >> anymore.
> >
> > Huh? I'm talking about the call of get_user_pages() in aio_setup_ring().
> > With your patch it's done completely outside of ->mmap_sem, isn't it?
>
> Oh, /that/ call to get_user_pages. That would qualify as screwing up...
>
> Since dropping and reacquiring mmap_sem there is probably a bad idea
> there, I'll rework this and post a v2.

FWIW, I've done some checking of ->mmap_sem uses yesterday. Got further than
the last time; catch so far, just from find_vma() audit:
* arm swp_emulate.c - missing ->mmap_sem around find_vma(). Fix sent to
rmk.
* blackfin ptrace - find_vma() without any protection, definitely broken
* m68k sys_cacheflush() - ditto
* mips process_fpemu_return() - ditto
* mips octeon_flush_cache_sigtramp() - ditto
* omap_vout_uservirt_to_phys() - ditto, patch sent
* vb2_get_contig_userptr() - probaly a bug, unless I've misread the (very
twisty maze of) v4l2 code leading to it
* vb2_get_contig_userptr() - ditto
* gntdev_ioctl_get_offset_for_vaddr() - definitely broken
and there's a couple of dubious places in arch/* I hadn't finished with,
plus a lot in mm/* proper.

That's just from a couple of days of RTFS. The locking in there is far too
convoluted as it is; worse, it's not localized code-wise, so rechecking
correctness is going to remain a big time-sink ;-/

Making it *more* complex doesn't look like a good idea, TBH...

BTW, the __get_user_pages()/find_extend_vma()/mlock_vma_pages_range() pile is
really asking for trouble; sure, the recursion there is limited, but it
deserves a comment. Moreover, the damn thing is reachable from coredump
path and there we do *not* have ->mmap_sem held. We don't reach the
VM_BUG_ON() in __mlock_vma_pages_range(), but the reason for that also
deserves a comment, IMO.

Moreover, I'm not quite convinced that huge_memory.c and ksm.c can't run
into all kinds of interesting races with ongoing coredump. Looking into
it...


\
 
 \ /
  Last update: 2012-12-16 18:41    [W:0.063 / U:0.204 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site