On Thu, May 01, 2014 at 10:57:39AM +0100, Peter Zijlstra wrote:

On Thu, May 01, 2014 at 10:52:47AM +0100, Catalin Marinas wrote:

quoted

Does gup_fast walking increment the mm_users? Or is it a requirement of
the calling code? I can't seem to find where this happens.

No, its not required at all. One should only walk current->mm with
gup_fast, any other usage is broken.

OK, I get it now.

And by delaying TLB shootdown, either through disabling IRQs and
stalling IPIs or by using RCU freeing, you're guaranteed your own page
tables won't disappear underneath your feet.

And for RCU to work, we still need to use the full tlb_remove_table()
logic (Steve's patches just use tlb_remove_page() for table freeing).

Thanks.

-- 
Catalin