Re: [PATCH] MA-21654 Use dma_alloc_pages in vb2_dma_sg_alloc_compacted

From: Tomasz Figa
Date: Thu Sep 21 2023 - 13:52:54 EST


On Thu, Sep 21, 2023 at 1:54 AM Robin Murphy <robin.murphy@xxxxxxx> wrote:
>
> On 20/09/2023 8:41 am, Tomasz Figa wrote:
> > Hi Fang,
> >
> > On Thu, Sep 14, 2023 at 4:41 PM Fang Hui <hui.fang@xxxxxxx> wrote:
> >>
> >> On system with "CONFIG_ZONE_DMA32=y", if the allocated physical address is
> >
> > First of all, thanks a lot for the patch! Please check my review comments below.
> >
> > Is CONFIG_ZONE_DMA32 really the factor that triggers the problem? My
> > understanding was that the problem was that the hardware has 32-bit
> > DMA, but the system has physical memory at addresses beyond the first
> > 4G.
>
> Indeed, without ZONE-DMA32 it would be difficult for any allocator to
> support this at all. SWIOTLB is merely a symptom - if it wasn't enabled,
> the dma_map_sgtable() operation would just fail entirely when any page
> is beyond the device's reach.
>
> >> greater than 4G, swiotlb will be used. It will lead below defects.
> >> 1) Impact performance due to an extra memcpy.
> >> 2) May meet below error due to swiotlb_max_mapping_size()
> >> is 256K (IO_TLB_SIZE * IO_TLB_SEGSIZE).
> >> "swiotlb buffer is full (sz: 393216 bytes), total 65536 (slots),
> >> used 2358 (slots)"
> >>
> >> To avoid those defects, use dma_alloc_pages() instead of alloc_pages()
> >> in vb2_dma_sg_alloc_compacted().
> >>
> >> Suggested-by: Tomasz Figa <tfiga@xxxxxxxxxxxx>
> >> Signed-off-by: Fang Hui <hui.fang@xxxxxxx>
> >> ---
> >> drivers/media/common/videobuf2/videobuf2-dma-sg.c | 11 +++++++----
> >> 1 file changed, 7 insertions(+), 4 deletions(-)
> >>
> >
> > Please remove MA-21654 from the subject and prefix it with the right
> > tags for the path (`git log drivers/media/common/videobuf2` should be
> > helpful to find the right one).
> >
> >> diff --git a/drivers/media/common/videobuf2/videobuf2-dma-sg.c b/drivers/media/common/videobuf2/videobuf2-dma-sg.c
> >> index 28f3fdfe23a2..b938582c68f4 100644
> >> --- a/drivers/media/common/videobuf2/videobuf2-dma-sg.c
> >> +++ b/drivers/media/common/videobuf2/videobuf2-dma-sg.c
> >> @@ -58,7 +58,7 @@ struct vb2_dma_sg_buf {
> >> static void vb2_dma_sg_put(void *buf_priv);
> >>
> >> static int vb2_dma_sg_alloc_compacted(struct vb2_dma_sg_buf *buf,
> >> - gfp_t gfp_flags)
> >> + gfp_t gfp_flags, struct device *dev)
> >
> > FWIW buf->dev already points to the right device - although we would
> > need to move the assignment in vb2_dma_sg_alloc() to a place higher in
> > that function before calling this function.
> >
> >> {
> >> unsigned int last_page = 0;
> >> unsigned long size = buf->size;
> >> @@ -67,6 +67,7 @@ static int vb2_dma_sg_alloc_compacted(struct vb2_dma_sg_buf *buf,
> >> struct page *pages;
> >> int order;
> >> int i;
> >> + dma_addr_t dma_handle;
> >>
> >> order = get_order(size);
> >> /* Don't over allocate*/
> >> @@ -75,8 +76,9 @@ static int vb2_dma_sg_alloc_compacted(struct vb2_dma_sg_buf *buf,
> >>
> >> pages = NULL;
> >> while (!pages) {
> >> - pages = alloc_pages(GFP_KERNEL | __GFP_ZERO |
> >> - __GFP_NOWARN | gfp_flags, order);
> >> + pages = dma_alloc_pages(dev, PAGE_SIZE << order, &dma_handle,
> >
> > Hmm, when I was proposing dma_alloc_pages(), I missed that it returns
> > a DMA handle. That on its own can be handled by saving the returned
> > handles somewhere in struct vb2_dma_sg_buf, but there is a bigger
> > problem - the function would actually create a mapping if the DMA
> > device requires some mapping management (e.g. is behind an IOMMU),
> > which is undesirable, because we create the mapping ourselves below
> > anyway...
> >
> > @Christoph Hellwig @Robin Murphy I need your thoughts on this as
> > well. Would it make sense to have a variant of dma_alloc_pages() that
> > only allocates the pages, but doesn't perform the mapping? (Or a flag
> > that tells the implementation to skip creating a mapping.)
>
> As I mentioned before, I think it might make the most sense to make the
> whole thing into a "proper" dma_alloc_sgtable() function, which can then
> be used with dma_sync_sgtable_*() as dma_alloc_pages() is used with
> dma_sync_single_*() (and then dma_alloc_noncontiguous() clearly falls as
> the special in-between case).
>

Okay, so that is the same thing that I was proposing from the
beginning of the original thread that reported the swiotlb issues.
Somehow I got convinced that it wasn't well received. Thanks for
clarifying!

Then it sounds like we just need someone to implement it?

Let me CC +Sergey Senozhatsky for visibility as well.

Best regards,
Tomasz

> Thanks,
> Robin.
>
> >> + DMA_BIDIRECTIONAL,
> >
> > The right value should be already available in buf->dma_dir.
> >
> >> + GFP_KERNEL | __GFP_ZERO | __GFP_NOWARN | gfp_flags);
> >> if (pages)
> >> break;
> >>
> >> @@ -96,6 +98,7 @@ static int vb2_dma_sg_alloc_compacted(struct vb2_dma_sg_buf *buf,
> >> }
> >>
> >> return 0;
> >> +
> >
> > Unnecessary blank line.
> >
> >> }
> >>
> >> static void *vb2_dma_sg_alloc(struct vb2_buffer *vb, struct device *dev,
> >> @@ -130,7 +133,7 @@ static void *vb2_dma_sg_alloc(struct vb2_buffer *vb, struct device *dev,
> >> if (!buf->pages)
> >> goto fail_pages_array_alloc;
> >>
> >> - ret = vb2_dma_sg_alloc_compacted(buf, vb->vb2_queue->gfp_flags);
> >> + ret = vb2_dma_sg_alloc_compacted(buf, vb->vb2_queue->gfp_flags, dev);
> >> if (ret)
> >> goto fail_pages_alloc;
> >>
> >> --
> >> 2.17.1
> >>
> >
> > We also need to use dma_free_pages() to free the memory.
> >
> > Best regards,
> > Tomasz