On Wed, Jul 7, 2021 at 10:29 PM Sergey Senozhatsky
<senozhatsky@xxxxxxxxxxxx> wrote:
>
> On (21/07/07 21:48), Tomasz Figa wrote:
> > > [..]
> > > > static void *vb2_dc_vaddr(struct vb2_buffer *vb, void *buf_priv)
> > > > {
> > > > struct vb2_dc_buf *buf = buf_priv;
> > > >
> > > > if (buf->vaddr)
> > > > return buf->vaddr;
> > > >
> > > > if (buf->db_attach) {
> > > > struct dma_buf_map map;
> > > >
> > > > if (!dma_buf_vmap(buf->db_attach->dmabuf, &map))
> > > > buf->vaddr = map.vaddr;
> > > >
> > > > return buf->vaddr;
> > > > }
> > > >
> > > > if (!buf->coherent_mem)
> > > > buf->vaddr = dma_vmap_noncontiguous(buf->dev, buf->size,
> > > > buf->dma_sgt);
> > > > return buf->vaddr;
> > > > }
> > > >
> > > > And in vb2_dc_alloc functions set vaddr for !DMA_ATTR_NO_KERNEL_MAPPING
> > > > in both coherent and non-coherent. So that we probably can have less
> > > > branches when ->vaddr is NULL for one type of allocations, and is not
> > > > NULL for another.
> >
> > I'd prefer if it stayed as is. This opportunistic mapping as in the
> > current revision is quite nice, because most of the drivers don't
> > bother to set DMA_ATTR_NO_KERNEL_MAPPING even if they don't need the
> > kernel mapping. Also, even if the driver itself doesn't need the
> > kernel mapping, we can still create one on demand if the DMA-buf
> > importer demands it from us.
>
> [..]
>
> > > > static int vb2_dc_alloc_coherent(struct vb2_dc_buf *buf)
> > > > {
> > > > struct vb2_queue *q = buf->vb->vb2_queue;
> > > >
> > > > buf->cookie = dma_alloc_attrs(buf->dev,
> > > > buf->size,
> > > > &buf->dma_addr,
> > > > GFP_KERNEL | q->gfp_flags,
> > > > buf->attrs);
> > > > if (!buf->cookie)
> > > > return -ENOMEM;
> > > >
> > > > if (q->dma_attrs & DMA_ATTR_NO_KERNEL_MAPPING)
> > > > return 0;
> > > >
> > > > buf->vaddr = buf->cookie;
> > > > return 0;
> > > > }
> > > >
> > > > static int vb2_dc_alloc_non_coherent(struct vb2_dc_buf *buf)
> > > > {
> > > > struct vb2_queue *q = buf->vb->vb2_queue;
> > > >
> > > > buf->dma_sgt = dma_alloc_noncontiguous(buf->dev,
> > > > buf->size,
> > > > buf->dma_dir,
> > > > GFP_KERNEL | q->gfp_flags,
> > > > buf->attrs);
> > > > if (!buf->dma_sgt)
> > > > return -ENOMEM;
> > > >
> > > > if (q->dma_attrs & DMA_ATTR_NO_KERNEL_MAPPING)
> > > > return 0;
> > > >
> > > > buf->vaddr = dma_vmap_noncontiguous(buf->dev, buf->size, buf->dma_sgt);
> > > > if (!buf->vaddr) {
> > > > dma_free_noncontiguous(buf->dev, buf->size,
> > > > buf->dma_sgt, buf->dma_addr);
> > > > return -ENOMEM;
> > > > }
> > > > return 0;
> > > > }
> > >
> > > I guess this should address the case when
> > >
> > > "after allocating the buffer, the buffer is exported as a dma_buf and
> > > another device calls dma_buf_ops vb2_dc_dmabuf_ops_vmap, which in turn
> > > calls dma_buf_map_set_vaddr(map, buf->vaddr); with a NULL buf->vaddr"
> >
> > Sorry, I fail to get what this is about. Where does this quote come from?
>
> Bottom half of https://lore.kernel.org/lkml/10a0903a-e295-5cba-683a-1eb89a0804ed@xxxxxxxxx/
I see, thanks for the pointer. Yes, vb2_dc_dmabuf_ops_vmap() needs to
be changed so that it calls vb2_dc_vaddr() internally instead of
relying on buf->vaddr directly.
