Re: [Linux-kernel-mentees][PATCH v6 1/2] sgi-gru: Convert put_page() to put_user_page*()

[John Hubbard]

On 8/20/19 1:18 AM, Michal Hocko wrote:
> On Mon 19-08-19 12:30:18, John Hubbard wrote:
> > On 8/19/19 12:06 PM, Bharath Vedartham wrote:
> > > On Mon, Aug 19, 2019 at 07:56:11AM -0500, Dimitri Sivanich wrote:
...
> > Conversion of gup/put_page sites:
> >
> > Before:
> >
> > 	get_user_pages(...);
> > 	...
> > 	for each page:
> > 		put_page();
> >
> > After:
> > 	
> > 	gup_flags |= FOLL_PIN; (maybe FOLL_LONGTERM in some cases)
> > 	vaddr_pin_user_pages(...gup_flags...)
>
> I was hoping that FOLL_PIN would be handled by vaddr_pin_user_pages.

Good point: now that we've got the 4 cases summarized, it turns out
that either FOLL_PIN is required, or there is no need to call
vaddr_pin_user_pages() at all.  So we can go back to setting FOLL_PIN
inside it, which is of course much better for maintenance. Great!


> > 	...
> > 	vaddr_unpin_user_pages(); /* which invokes put_user_page() */
> >
> > Fortunately, it's not harmful for the simpler conversion from put_page()
> > to put_user_page() to happen first, and in fact those have usually led
> > to simplifications, paving the way to make it easier to call
> > vaddr_unpin_user_pages(), once it's ready. (And showing exactly what
> > to convert, too.)
>
> If that makes the later conversion easier then no real objections from
> me. Assuming that the current put_user_page conversions are correct of
> course (I have the mlock one and potentials that falls into the same
> category in mind).

Agreed: only correct conversions should be done. Not the incorrect
ones. ahem.  :)

thanks,
--
John Hubbard
NVIDIA