[Intel-gfx] [PATCH v3 08/12] drm/ttm: Use drm_memcpy_from_wc_dbm for TTM bo moves

[Thomas Hellström]

On Mon, 2021-05-24 at 19:16 +0100, Matthew Auld wrote:
> On Fri, 21 May 2021 at 16:33, Thomas Hellström
> <thomas.hellstrom at linux.intel.com> wrote:
> > 
> > Use fast wc memcpy for reading out of wc memory for TTM bo moves.
> > 
> > Cc: Dave Airlie <airlied at gmail.com>
> > Cc: Christian König <christian.koenig at amd.com>
> > Cc: Daniel Vetter <daniel.vetter at ffwll.ch>
> > Signed-off-by: Thomas Hellström <thomas.hellstrom at linux.intel.com>
> > ---
> >  drivers/gpu/drm/ttm/ttm_bo_util.c | 9 ++++++++-
> >  1 file changed, 8 insertions(+), 1 deletion(-)
> > 
> > diff --git a/drivers/gpu/drm/ttm/ttm_bo_util.c
> > b/drivers/gpu/drm/ttm/ttm_bo_util.c
> > index 912cbe8e60a2..4a7d3d672f9a 100644
> > --- a/drivers/gpu/drm/ttm/ttm_bo_util.c
> > +++ b/drivers/gpu/drm/ttm/ttm_bo_util.c
> > @@ -31,6 +31,7 @@
> > 
> >  #include <drm/ttm/ttm_bo_driver.h>
> >  #include <drm/ttm/ttm_placement.h>
> > +#include <drm/drm_memcpy.h>
> >  #include <drm/drm_vma_manager.h>
> >  #include <linux/dma-buf-map.h>
> >  #include <linux/io.h>
> > @@ -91,6 +92,7 @@ void ttm_move_memcpy(struct ttm_buffer_object
> > *bo,
> >         const struct ttm_kmap_iter_ops *src_ops = src_iter->ops;
> >         struct ttm_tt *ttm = bo->ttm;
> >         struct dma_buf_map src_map, dst_map;
> > +       bool wc_memcpy;
> >         pgoff_t i;
> > 
> >         /* Single TTM move. NOP */
> > @@ -114,11 +116,16 @@ void ttm_move_memcpy(struct ttm_buffer_object
> > *bo,
> >                 return;
> >         }
> > 
> > +       wc_memcpy = ((!src_ops->maps_tt || ttm->caching !=
> > ttm_cached) &&
> 
> Why do we only consider the caching value for the maps_tt case? Or am
> I misreading this?

Hmm, you're right we should probably check also the iomem caching or
ignore the tt caching. Sometimes these special memcpys tend to be less
cpu cache thrashing and should be used wherever possible, but I guess
we should start out with only using it when source is WC.

> 
> > +                    drm_has_memcpy_from_wc());
> > +
> >         for (i = 0; i < dst_mem->num_pages; ++i) {
> >                 dst_ops->map_local(dst_iter, &dst_map, i);
> >                 src_ops->map_local(src_iter, &src_map, i);
> > 
> > -               if (!src_map.is_iomem && !dst_map.is_iomem) {
> > +               if (wc_memcpy) {
> > +                       drm_memcpy_from_wc_dbm(&dst_map, &src_map,
> > PAGE_SIZE);
> 
> Do we need to check the return value here? memcpy_from_wc expects
> certain address alignment, or is that always guaranteed here? Maybe
> throw a warning just for paranoia?

It should always be PAGE_SIZE aligned. But I guess it doesn't hurt to
do 
if (wc_memcpy && drm_memcpy_from_wc_dbm())
;

> 
> > +               } else if (!src_map.is_iomem && !dst_map.is_iomem)
> > {
> >                         memcpy(dst_map.vaddr, src_map.vaddr,
> > PAGE_SIZE);
> >                 } else if (!src_map.is_iomem) {
> >                         dma_buf_map_memcpy_to(&dst_map,
> > src_map.vaddr,
> > --
> > 2.31.1
> >