Re: [RFC, v3 9/9] media: platform: Add Mediatek ISP P1 shared memory device

[Tomasz Figa <tfiga@chromium.org>]

On Fri, Jul 26, 2019 at 8:59 PM Jungo Lin <jungo.lin@mediatek.com> wrote:
>
> Hi Robin:
>
> On Fri, 2019-07-26 at 12:04 +0100, Robin Murphy wrote:
> > On 26/07/2019 08:42, Tomasz Figa wrote:
> > > On Fri, Jul 26, 2019 at 4:41 PM Christoph Hellwig <hch@infradead.org> wrote:
> > >>
> > >> On Fri, Jul 26, 2019 at 02:15:14PM +0900, Tomasz Figa wrote:
> > >>> Could you try dma_get_sgtable() with the SCP struct device and then
> > >>> dma_map_sg() with the P1 struct device?
> > >>
> > >> Please don't do that.  dma_get_sgtable is a pretty broken API (see
> > >> the common near the arm implementation) and we should not add more
> > >> users of it.  If you want a piece of memory that can be mapped to
> > >> multiple devices allocate it using alloc_pages and then just map
> > >> it to each device.
> > >
> > > Thanks for taking a look at this thread.
> > >
> > > Unfortunately that wouldn't work. We have a specific reserved memory
> > > pool that is the only memory area accessible to one of the devices.
> > > Any idea how to handle this?
> >
> > If it's reserved in the sense of being outside struct-page-backed
> > "kernel memory", then provided you have a consistent CPU physical
> > address it might be reasonable for other devices to access it via
> > dma_map_resource().
> >
> > Robin.
>
> Thank you for your suggestion.
>
> After revising to use dma_map_resource(), it is worked. Below is the
> current implementation. Pleas kindly help us to check if there is any
> misunderstanding.
>
> #define MTK_ISP_COMPOSER_MEM_SIZE               0x200000
>
>         /*
>          * Allocate coherent reserved memory for SCP firmware usage.
>          * The size of SCP composer's memory is fixed to 0x200000
>          * for the requirement of firmware.
>          */
>         ptr = dma_alloc_coherent(p1_dev->cam_dev.smem_dev,
>                                  MTK_ISP_COMPOSER_MEM_SIZE, &addr, GFP_KERNEL);
>         if (!ptr) {
>                 dev_err(dev, "failed to allocate compose memory\n");
>                 return -ENOMEM;
>         }
>         p1_dev->composer_scp_addr = addr;
>         p1_dev->composer_virt_addr = ptr;
>         dev_dbg(dev, "scp addr:%pad va:%pK\n", &addr, ptr);
>
>         /*
>          * This reserved memory is also be used by ISP P1 HW.
>          * Need to get iova address for ISP P1 DMA.
>          */
>         addr = dma_map_resource(dev, addr, MTK_ISP_COMPOSER_MEM_SIZE,
>                                 DMA_BIDIRECTIONAL, DMA_ATTR_SKIP_CPU_SYNC);

This is still incorrect, because addr is a DMA address, but the second
argument to dma_map_resource() is a physical address.

>         if (dma_mapping_error(dev, addr)) {
>                 dev_err(dev, "Failed to map scp iova\n");
>                 ret = -ENOMEM;
>                 goto fail_free_mem;
>         }
>         p1_dev->composer_iova = addr;
>         dev_info(dev, "scp iova addr:%pad\n", &addr);
>
> Moreover, appropriate Tomasz & Christoph's help on this issue.

Robin, the memory is specified using the reserved-memory DT binding
and managed by the coherent DMA pool framework. We can allocate from
it using dma_alloc_coherent(), which gives us a DMA address, not CPU
physial address (although in practice on this platform they are equal
numerically).

Best regards,
Tomasz