DMA mapping API abuse in exynos-drm

[Marek Szyprowski]

Hello,

On 5/6/2013 7:59 AM, Inki Dae wrote:
> 2013/5/5 Tomasz Figa <tomasz.figa at gmail.com 
> <mailto:tomasz.figa at gmail.com>>
>
>     Hi,
>
>     Recently I've been working a bit on a DRM driver for the GPU of
>     Samsung
>     S3C6410 SoCs, which required me to familiarize a bit with
>     exynos-drm, as
>     it already contains a KMS driver which is compatible with the SoC I'm
>     working with, making it a good place to put my driver in.
>
>     Reading through exynos_drm_buf.c I stumbled across following (a
>     bit long)
>     piece of code:
>
>             dma_set_attr(DMA_ATTR_NO_KERNEL_MAPPING, &buf->dma_attrs);
>
>             nr_pages = buf->size >> PAGE_SHIFT;
>
>             if (!is_drm_iommu_supported(dev)) {
>                     dma_addr_t start_addr;
>                     unsigned int i = 0;
>
>                     buf->pages = kzalloc(sizeof(struct page) * nr_pages,
>                                             GFP_KERNEL);
>                     if (!buf->pages) {
>                             DRM_ERROR("failed to allocate pages.\n");
>                             return -ENOMEM;
>                     }
>
>                     buf->kvaddr = dma_alloc_attrs(dev->dev, buf->size,
>     &buf->dma_addr, GFP_KERNEL,
>     &buf->dma_attrs);
>                     if (!buf->kvaddr) {
>                             DRM_ERROR("failed to allocate buffer.\n");
>                             kfree(buf->pages);
>                             return -ENOMEM;
>                     }
>
>                     start_addr = buf->dma_addr;
>                     while (i < nr_pages) {
>                             buf->pages[i] = phys_to_page(start_addr);
>                             start_addr += PAGE_SIZE;
>                             i++;
>                     }
>             } else {
>
>                     buf->pages = dma_alloc_attrs(dev->dev, buf->size,
>     &buf->dma_addr, GFP_KERNEL,
>     &buf->dma_attrs);
>                     if (!buf->pages) {
>                             DRM_ERROR("failed to allocate buffer.\n");
>                             return -ENOMEM;
>                     }
>             }
>
>             buf->sgt = drm_prime_pages_to_sg(buf->pages, nr_pages);
>             if (!buf->sgt) {
>                     DRM_ERROR("failed to get sg table.\n");
>                     ret = -ENOMEM;
>                     goto err_free_attrs;
>             }
>
>     Notice that the value returned by dma_alloc_attrs() is assumed by this
>     code to be either a kernel virtual address (in
>     !is_drm_iommu_supported()
>     case) or struct pages ** (in is_drm_iommu_supported() case), which
>     seemed
>     a bit worrying to me, making me do a more in depth research on how
>     dma_alloc_attrs works.
>
>     This, in turn, led me to following excerpt of Documentation/DMA-
>     attributes.txt :
>
>     DMA_ATTR_NO_KERNEL_MAPPING
>     --------------------------
>
>     DMA_ATTR_NO_KERNEL_MAPPING lets the platform to avoid creating a
>     kernel
>     virtual mapping for the allocated buffer. On some architectures
>     creating
>     such mapping is non-trivial task and consumes very limited resources
>     (like kernel virtual address space or dma consistent address space).
>     Buffers allocated with this attribute can be only passed to user space
>     by calling dma_mmap_attrs(). By using this API, you are guaranteeing
>     that you won't dereference the pointer returned by
>     dma_alloc_attr(). You
>     can threat it as a cookie that must be passed to dma_mmap_attrs() and
>     dma_free_attrs(). Make sure that both of these also get this attribute
>     set on each call.
>
>     of which the following sentence is the most relevant:
>
>     By using this API, you are guaranteeing that you won't dereference the
>     pointer returned by dma_alloc_attr().
>
>     This statement is obviously ignored by buffer allocation code of
>     exynos-
>     drm.
>
>     A simple git blame revealed that this brokenness has been
>     introduced by
>     commit:
>
>     4744ad2 drm/exynos: use DMA_ATTR_NO_KERNEL_MAPPING attribute
>
>
>     and later fixed a bit by following depending patches (because the
>     original
>     patch apparently broke any allocations without IOMMU):
>
>     c704f1b drm/exynos: consider no iommu support to console framebuffer
>     694be45 drm/exynos: consider buffer allocation without iommu
>
>
>     Now, the real question is whether my reasoning is incorrect (sorry
>     for the
>     noise then) or this is really a bug which needs to be fixed?
>
>
> Hi Tomasz,
>
> Your question is reasonable to me. And below is my opinion,
>
> First, also the below attribute says like below,
>
> DMA_ATTR_NO_KERNEL_MAPPING
> --------------------------
> ...
> Since it is optional for platforms to implement
> DMA_ATTR_NO_KERNEL_MAPPING, those that do not will simply ignore the
> attribute and exhibit default behavior.
>
>
> In case of ARM SoC, it seems like that it just ignores the attribute
> without iommu: in case of no iommu, dma_alloc_attr() always maps pages
> allocated from highmem with kernel space. So I think we make sure that
> exynos drm driver sets the attribute only with iommu to avoid such
> confusing. For this, will post it soon.

IMHO this case simply shows that ExynosDRM driver should manage IO address
space on it's own, because DMA-mapping api is too limited. ExynosDRM is
unlikely to be used on a hardware other than Samsung Exynos, so it can
better use hardware if it access IOMMU and CMA subsystems directly and
have its own address space manager to avoid hacking in internal structures
of dma-mapping.

Best regards
-- 
Marek Szyprowski
Samsung Poland R&D Center