Re: [PATCH] mm: remove unnecessary __get_user_pages_unlocked() calls

[Lorenzo Stoakes]

On Wed, Oct 26, 2016 at 11:15:43AM +0200, Michal Hocko wrote:
> On Wed 26-10-16 00:46:31, Lorenzo Stoakes wrote:
> > The holdout for unexporting __get_user_pages_unlocked() is its invocation in
> > mm/process_vm_access.c: process_vm_rw_single_vec(), as this definitely _does_
> > seem to invoke VM_FAULT_RETRY behaviour which get_user_pages_remote() will not
> > trigger if we were to replace it with the latter.
>
> I am not sure I understand. Prior to 1e9877902dc7e this used
> get_user_pages_unlocked. What prevents us from reintroducing it with
> FOLL_REMOVE which was meant to be added by the above commit?
>
> Or am I missing your point?

The issue isn't the flags being passed, rather that in this case:

a. Replacing __get_user_pages_unlocked() with get_user_pages_unlocked() won't
   work as the latter assumes task = current and mm = current->mm but
   process_vm_rw_single_vec() needs to pass different task, mm.

b. Moving to get_user_pages_remote() _will_ allow us to pass different task, mm
   but won't however match existing behaviour precisely, since
   __get_user_pages_unlocked() acquires mmap_sem then passes a pointer to a
   local 'locked' variable to __get_user_pages_locked() which allows
   VM_FAULT_RETRY to trigger.

The main issue I had here was not being sure whether we care about the
VM_FAULT_RETRY functionality being used here or not, if we don't care then we
can just move to get_user_pages_remote(), otherwise perhaps this should be left
alone or maybe we need to consider adjusting the API to allow for remote access
with VM_FAULT_RETRY functionality.