[PATCH] drm/radeon: Convert get_user_pages() --> pin_user_pages()
Souptick Joarder
jrdr.linux at gmail.com
Fri May 29 06:49:30 UTC 2020
On Thu, May 28, 2020 at 12:37 AM John Hubbard <jhubbard at nvidia.com> wrote:
>
> On 2020-05-27 01:51, Daniel Vetter wrote:
> > On Wed, May 27, 2020 at 10:48:52AM +0200, Daniel Vetter wrote:
> >> On Tue, May 26, 2020 at 03:57:45PM -0700, John Hubbard wrote:
> >>> On 2020-05-26 14:00, Souptick Joarder wrote:
> >>>> This code was using get_user_pages(), in a "Case 2" scenario
> >>>> (DMA/RDMA), using the categorization from [1]. That means that it's
> >>>> time to convert the get_user_pages() + release_pages() calls to
> >>>> pin_user_pages() + unpin_user_pages() calls.
> >>>>
> >>>> There is some helpful background in [2]: basically, this is a small
> >>>> part of fixing a long-standing disconnect between pinning pages, and
> >>>> file systems' use of those pages.
> >>>>
> >>>> [1] Documentation/core-api/pin_user_pages.rst
> >>>>
> >>>> [2] "Explicit pinning of user-space pages":
> >>>> https://lwn.net/Articles/807108/
> >>
> >> I don't think this is a case 2 here, nor is it any of the others. Feels
> >> like not covered at all by the doc.
> >>
> >> radeon has a mmu notifier (might be a bit broken, but hey whatever there's
> >> other drivers which have the same concept, but less broken). So when you
> >> do an munmap, radeon will release the page refcount.
> >
>
> Aha, thanks Daniel. I withdraw my misinformed ACK, then.
>
> > I forgot to add: It's also not case 3, since there's no hw page fault
> > support. It's all faked in software, and explicitly synchronizes against
> > pending io (or preempts it, that depends a bit upon the jobs running).
> >
>
> This is what case 3 was *intended* to cover, but it looks like case 3 needs to
> be written a little better. I'll attempt that, and Cc you on the actual patch
> to -mm. (I think we also need a case 5 for an unrelated scenario, too, so
> it's time.)
There were no *case 5* in the other patch posted in -mm. Do we need to add it ?
>
>
> thanks,
> --
> John Hubbard
> NVIDIA
>
>
> >> Which case it that?
> >>
> >> Note that currently only amdgpu doesn't work like that for gpu dma
> >> directly to userspace ranges, it uses hmm and afaiui doens't hold a full
> >> page pin refcount.
> >>
> >> Cheers, Daniel
> >>
> >>
> >>>>
> >>>> Signed-off-by: Souptick Joarder <jrdr.linux at gmail.com>
> >>>> Cc: John Hubbard <jhubbard at nvidia.com>
> >>>>
> >>>> Hi,
> >>>>
> >>>> I'm compile tested this, but unable to run-time test, so any testing
> >>>> help is much appriciated.
> >>>> ---
> >>>> drivers/gpu/drm/radeon/radeon_ttm.c | 6 +++---
> >>>> 1 file changed, 3 insertions(+), 3 deletions(-)
> >>>>
> >>>> diff --git a/drivers/gpu/drm/radeon/radeon_ttm.c b/drivers/gpu/drm/radeon/radeon_ttm.c
> >>>> index 5d50c9e..e927de2 100644
> >>>> --- a/drivers/gpu/drm/radeon/radeon_ttm.c
> >>>> +++ b/drivers/gpu/drm/radeon/radeon_ttm.c
> >>>> @@ -506,7 +506,7 @@ static int radeon_ttm_tt_pin_userptr(struct ttm_tt *ttm)
> >>>> uint64_t userptr = gtt->userptr + pinned * PAGE_SIZE;
> >>>> struct page **pages = ttm->pages + pinned;
> >>>> - r = get_user_pages(userptr, num_pages, write ? FOLL_WRITE : 0,
> >>>> + r = pin_user_pages(userptr, num_pages, write ? FOLL_WRITE : 0,
> >>>> pages, NULL);
> >>>> if (r < 0)
> >>>> goto release_pages;
> >>>> @@ -535,7 +535,7 @@ static int radeon_ttm_tt_pin_userptr(struct ttm_tt *ttm)
> >>>> kfree(ttm->sg);
> >>>> release_pages:
> >>>> - release_pages(ttm->pages, pinned);
> >>>> + unpin_user_pages(ttm->pages, pinned);
> >>>> return r;
> >>>> }
> >>>> @@ -562,7 +562,7 @@ static void radeon_ttm_tt_unpin_userptr(struct ttm_tt *ttm)
> >>>> set_page_dirty(page);
> >>>
> >>>
> >>> Maybe we also need a preceding patch, to fix the above? It should be
> >>> set_page_dirty_lock(), rather than set_page_dirty(), unless I'm overlooking
> >>> something (which is very possible!).
> >>>
> >>> Either way, from a tunnel vision perspective of changing gup to pup, this
> >>> looks good to me, so
> >>>
> >>> Acked-by: John Hubbard <jhubbard at nvidia.com>
> >>>
> >>>
> >>> thanks,
> >>> --
> >>> John Hubbard
> >>> NVIDIA
> >>>
> >>>> mark_page_accessed(page);
> >>>> - put_page(page);
> >>>> + unpin_user_page(page);
> >>>> }
> >>>> sg_free_table(ttm->sg);
> >>>>
> >>>
> >>
> >> --
> >> Daniel Vetter
> >> Software Engineer, Intel Corporation
> >> http://blog.ffwll.ch
> >
>
More information about the amd-gfx
mailing list