Re: [RFC] situation with fput() locking (was Re: [PULL REQUEST] : ima-appraisal patches)

[Hugh Dickins <hughd@google.com>]

On Fri, 20 Apr 2012, Linus Torvalds wrote:
> On Fri, Apr 20, 2012 at 12:04 PM, Al Viro <viro@zeniv.linux.org.uk> wrote:
> >
> > Deferring the final pass after dropping ->mmap_sem is going to be
> > interesting; what would protect ->vm_next on those suckers?
> 
> Just remove them from the active list, and keep them linked to each
> other using vm_next.
> 
> After all, the only case we care about is the case where the vma gets
> removed entirely, so we just put them on their own list.
> 
> In fact, that's what we already do for other reasons. See
> detach_vmas_to_be_unmapped().
> 
> So vm_next is actually entirely *private* by this time, and needs no
> locking at all.
> 
> As far as I can tell, we could just make do_munmap() return that
> private list, and then do the fput's and freeing of the list outside
> the mmap_sem lock.
> 
> That actually looks pretty simple. There are a fair number of callers,
> which looks to be the main annoyance. But fixing it up with some
> pattern of "do finish_munmap after drooping the mmap_sem" doesn't look
> *complicated*, just slightly annoying.
> 
> The *bigger* annoyance is actually "do_mmap()", which does a
> do_munmap() as part of it, so it needs the same cleanup too.
> 
> There might be other cases that do munmap as part of their operation
> (and that have the mmap_sem held outside of the caller), but
> do_munmap() and do_mmap() seem to be the two obvious ones.
> 
> Many of the callers seem to do the mmap_sem() right around the call,
> though (see binfmt_elf.c for example), so it really would be a rather
> local fixup.

Yes, that's exactly how I was thinking of it.

Hugh