Re: [PATCH v5.5 10/10] mmap locking API: rename mmap_sem to mmap_lock

[John Hubbard]

On 2020-05-19 08:32, Matthew Wilcox wrote:
> On Tue, May 19, 2020 at 03:20:40PM +0200, Laurent Dufour wrote:
> > Le 19/05/2020 Ã 15:10, Michel Lespinasse a ÃcritÂ:
> > > On Mon, May 18, 2020 at 03:45:22PM +0200, Laurent Dufour wrote:
> > > > Le 24/04/2020 Ã 03:39, Michel Lespinasse a ÃcritÂ:
> > > > > Rename the mmap_sem field to mmap_lock. Any new uses of this lock
> > > > > should now go through the new mmap locking api. The mmap_lock is
> > > > > still implemented as a rwsem, though this could change in the future.
> > > > >
> > > > > diff --git a/drivers/gpu/drm/etnaviv/etnaviv_gem.c b/drivers/gpu/drm/etnaviv/etnaviv_gem.c
> > > > > index dc9ef302f517..701f3995f621 100644
> > > > > --- a/drivers/gpu/drm/etnaviv/etnaviv_gem.c
> > > > > +++ b/drivers/gpu/drm/etnaviv/etnaviv_gem.c
> > > > > @@ -661,7 +661,7 @@ static int etnaviv_gem_userptr_get_pages(struct etnaviv_gem_object *etnaviv_obj)
> > > > >     	struct etnaviv_gem_userptr *userptr = &etnaviv_obj->userptr;
> > > > >     	int ret, pinned = 0, npages = etnaviv_obj->base.size >> PAGE_SHIFT;
> > > > > -	might_lock_read(&current->mm->mmap_sem);
> > > > > +	might_lock_read(&current->mm->mmap_lock);
> > > >
> > > > Why not a mm_might_lock_read() new API to hide the mmap_lock, and add it to
> > > > the previous patch?
> > >
> > > I'm not sure why this is needed - we may rework the lock to be
> > > something else than rwsem, but might_lock_read should still apply to
> > > it and make sense ? I'm not sure what the extra API would bring...
> >
> > I guess at one time the API would become might_lock_read_a_range(), isn't it?
> > Furthermore this would hiding the lock's name which the goal of this series.
>
> I think this assertion should be deleted from this driver.  It's there
> in case get_user_pages_fast() takes the mmap sem.  It would make sense to
> have this assertion in get_user_pages_fast() in case we take the fast path
> which doesn't acquire the mmap_sem.  Something like this:
>
> +++ b/mm/gup.c
> @@ -2754,6 +2754,7 @@ static int internal_get_user_pages_fast(unsigned long start, int nr_pages,
>                                         FOLL_FORCE | FOLL_PIN | FOLL_GET)))
>                  return -EINVAL;
>   
> +       might_lock_read(&current->mm->mmap_lock);
>          start = untagged_addr(start) & PAGE_MASK;
>          addr = start;
>          len = (unsigned long) nr_pages << PAGE_SHIFT;

Hi Michel and Matthew and all,

There are a couple of recent developments in this code to keep in mind. I don't
*think* either one is a problem here, but just in case:

a) The latest version of the above routine [1] is on its way to mmotm as of
yesterday, and that version more firmly divides the fast and slow parts,
via a new FOLL_FAST_ONLY flag. The fall-back to slow/regular gup only occurs
if the caller does not set FOLL_FAST_ONLY. (Note that it's a gup.c internal
flag, btw.)

That gives you additional options inside internal_get_user_pages_fast(), such
as, approximately:

if (!(gup_flags & FOLL_FAST_ONLY))
	might_lock_read(&current->mm->mmap_lock);

...not that that is necessarily a great idea, seeing as how it merely changes
"might lock" into "maybe might lock".  :)

b) I've posted a small patch to that same etnaviv_gem.c file [2], over the
weekend, to convert from get_user_pages()/put_page(), to 
pin_user_pages()/unpin_user_pages(). It hasn't been merged yet, and it
shouldn't conflict either, but just one more reason to hope that the
etnaviv list can do some run time testing on the whole lot.

[1] https://lore.kernel.org/r/20200519002124.2025955-3-jhubbard@xxxxxxxxxx

[2] https://lore.kernel.org/r/20200518054315.2407093-1-jhubbard@xxxxxxxxxx


thanks,
--
John Hubbard
NVIDIA