On Mon, Oct 2, 2023 at 7:16 PM Hans Verkuil <hverkuil@xxxxxxxxx> wrote:
>
> Hi Ma Ke,
>
> On 24/09/2023 14:44, Ma Ke wrote:
> > In order to avoid error pointers from frame_vector_pages(), we could
> > use IS_ERR() to check the return value to fix this. This checking
> > operation could make sure that vector contains pages.
> >
> > Signed-off-by: Ma Ke <make_ruc2021@xxxxxxx>
> > ---
> > drivers/media/common/videobuf2/videobuf2-vmalloc.c | 1 +
> > 1 file changed, 1 insertion(+)
> >
> > diff --git a/drivers/media/common/videobuf2/videobuf2-vmalloc.c b/drivers/media/common/videobuf2/videobuf2-vmalloc.c
> > index 7c635e292106..c37775080aff 100644
> > --- a/drivers/media/common/videobuf2/videobuf2-vmalloc.c
> > +++ b/drivers/media/common/videobuf2/videobuf2-vmalloc.c
> > @@ -134,6 +134,7 @@ static void vb2_vmalloc_put_userptr(void *buf_priv)
> > if (!buf->vec->is_pfns) {
> > n_pages = frame_vector_count(buf->vec);
> > pages = frame_vector_pages(buf->vec);
> > + BUG_ON(IS_ERR(pages));
Can this even happen? We removed support for pfn maps from
get_vaddr_frames() quite long ago, so we should always have pages in
the frame vector if vb2_crate_framevec() in get_userptr() succeeds.
Or am I reading something wrong?
> > if (vaddr)
> > vm_unmap_ram((void *)vaddr, n_pages);
> > if (buf->dma_dir == DMA_FROM_DEVICE ||
>
> The use of BUG_ON is discouraged in the kernel. I did notice that is it
> also used in the put_userptr callback in videobuf2-dma-contig.c.
>
> I think it is much better to do something like this:
>
> if (!buf->vec->is_pfns) {
> n_pages = frame_vector_count(buf->vec);
> if (vaddr)
> vm_unmap_ram((void *)vaddr, n_pages);
> if (buf->dma_dir == DMA_FROM_DEVICE ||
> buf->dma_dir == DMA_BIDIRECTIONAL) {
> pages = frame_vector_pages(buf->vec);
> if (!WARN_ON_ONCE(IS_ERR(pages)))
> for (i = 0; i < n_pages; i++)
> set_page_dirty_lock(pages[i]);
> }
> } else {
>
> and do something similar in videobuf2-dma-contig.c.
>
> Regards,
>
> Hans
