Re: [PATCH] mm/gup: Remove the restriction on locked with FOLL_LONGTERM

[Jason Gunthorpe <jgg@nvidia.com>]

On Wed, Nov 16, 2022 at 02:11:04PM -0800, John Hubbard wrote:
> On 11/16/22 12:07, Jason Gunthorpe wrote:
> > This restriction was created because FOLL_LONGTERM used to scan the vma
> > list, so it could not tolerate becoming unlocked. That was fixed in commit
> > 52650c8b466b ("mm/gup: remove the vma allocation from
> > gup_longterm_locked()") and the restriction on !vma was removed.
> > 
> > However, the locked restriction remained, even though it isn't necessary
> > anymore.
> > 
> > Adjust __gup_longterm_locked() so it can handle the mmap_read_lock()
> > becoming unlocked while it is looping for migration. Migration does not
> > require the mmap_read_sem because it is only handling struct pages.  If we
> > had to unlock then ensure the whole thing returns unlocked.
> > 
> > Remove __get_user_pages_remote() and __gup_longterm_unlocked(). These
> > cases can now just directly call other functions.
> > 
> > Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
> > ---
> >   mm/gup.c | 109 ++++++++++++++-----------------------------------------
> >   1 file changed, 27 insertions(+), 82 deletions(-)
> 
> Looks nice.
> 
> Even after this cleanup, gup.c is still a bit of a ball of string, what
> with things like __gup_longterm_locked() actually handling
> !FOLL_LONGTERM, and a bunch of other vestigial stuff, but this takes us
> in the right direction.

What do you think of this:

https://github.com/jgunthorpe/linux/commits/gup

Jason