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

[Al Viro <viro@ZenIV.linux.org.uk>]

On Fri, Apr 20, 2012 at 11:54:01AM -0700, Hugh Dickins wrote:
> I can see that the discussion has since moved on quite a way from here.
> 
> But it looks to me fairly easy for mm to stop doing fput() under mmap_sem.
> 
> That's already the case when exiting (no mmap_sem held), and shouldn't add
> observable cost when unmapping (we already work on a chain of vmas to be
> freed, and when unmapping that chain will usually just be of one: shouldn't
> matter to defer a final pass until after mmap_sem is dropped).  Unless I'm
> mistaken, the fput() buried in vma_adjust() can never be a final fput.
> 
> Is it worth my trying to implement that?  Or do you see an immediate
> gotcha that I'm missing?  Or are you happy enough with your deferred
> fput() ideas, that it would be a waste of time to rearrange the mm end?

Deferring the final pass after dropping ->mmap_sem is going to be
interesting; what would protect ->vm_next on those suckers?  Locking
rules are already bloody complicated in mm/*; this will only add more
subtle fun.  Note that right now both the dissolution of ->vm_next
list and freeing of VMAs happen under ->mmap_sem; changing that might
cost us a lot of headache...