summaryrefslogtreecommitdiff
diff options
context:
space:
mode:
authorMarek Szyprowski <m.szyprowski@samsung.com>2014-04-22 11:14:08 (GMT)
committerMarek Szyprowski <m.szyprowski@samsung.com>2014-04-23 07:13:31 (GMT)
commitbba75455d47e71e0662c06508035333445481375 (patch)
treea5ba3a04ae1b26ff6ef5d11400b79d66b288ae38
parent079dad139665e7d9331fcb9dc0f0bb3a2a4f4244 (diff)
downloadlinux-3.10-bba75455d47e71e0662c06508035333445481375.zip
linux-3.10-bba75455d47e71e0662c06508035333445481375.tar.gz
linux-3.10-bba75455d47e71e0662c06508035333445481375.tar.bz2
media: vb2-dma-contig: add a hack to speed up cache managementrefs/changes/85/19885/2
In most cases video processing is done with coherent (non-cachable) userspace mappings. In such case for USERPTR mode cache management can be skipped. This patch introduces code for such optimisation. It is called a hack, because it is not possible in generic way to detect if the given userspace mapping is coherent on not. Signed-off-by: Marek Szyprowski <m.szyprowski@samsung.com> Change-id: I86aa292455371f1b18eb3f6b765234dc04ff937b
-rw-r--r--drivers/media/v4l2-core/videobuf2-dma-contig.c29
1 files changed, 27 insertions, 2 deletions
diff --git a/drivers/media/v4l2-core/videobuf2-dma-contig.c b/drivers/media/v4l2-core/videobuf2-dma-contig.c
index efdaca4..4e4d175 100644
--- a/drivers/media/v4l2-core/videobuf2-dma-contig.c
+++ b/drivers/media/v4l2-core/videobuf2-dma-contig.c
@@ -40,6 +40,7 @@ struct vb2_dc_buf {
/* USERPTR related */
struct vm_area_struct *vma;
+ bool coherent;
/* DMABUF related */
struct dma_buf_attachment *db_attach;
@@ -114,7 +115,7 @@ static void vb2_dc_prepare(void *buf_priv)
struct sg_table *sgt = buf->dma_sgt;
/* DMABUF exporter will flush the cache for us */
- if (!sgt || buf->db_attach)
+ if (!sgt || buf->db_attach || buf->coherent)
return;
dma_sync_sg_for_device(buf->dev, sgt->sgl, sgt->nents, buf->dma_dir);
@@ -126,7 +127,7 @@ static void vb2_dc_finish(void *buf_priv)
struct sg_table *sgt = buf->dma_sgt;
/* DMABUF exporter will flush the cache for us */
- if (!sgt || buf->db_attach)
+ if (!sgt || buf->db_attach || buf->coherent)
return;
dma_sync_sg_for_cpu(buf->dev, sgt->sgl, sgt->nents, buf->dma_dir);
@@ -423,6 +424,28 @@ static inline int vma_is_io(struct vm_area_struct *vma)
return !!(vma->vm_flags & (VM_IO | VM_PFNMAP));
}
+#ifdef __arm__
+static inline int vma_is_coherent(struct vm_area_struct *vma)
+{
+ pgprot_t prot = vma->vm_page_prot;
+ /*
+ * this is a hack to speed up cache management on coherent arm mappings,
+ * other architectures rely on dma-mapping internal optimisations.
+ */
+ if (vma_is_io(vma) || prot == pgprot_noncached(prot) ||
+ prot == pgprot_writecombine(prot) ||
+ prot == pgprot_dmacoherent(prot))
+ return true;
+
+ return false;
+}
+#else
+static inline int vma_is_coherent(struct vm_area_struct *vma)
+{
+ return false;
+}
+#endif
+
static int vb2_dc_get_user_pfn(unsigned long start, int n_pages,
struct vm_area_struct *vma, unsigned long *res)
{
@@ -614,6 +637,8 @@ static void *vb2_dc_get_userptr(void *alloc_ctx, unsigned long vaddr,
goto fail_pages;
}
+ buf->coherent = vma_is_coherent(vma);
+
/* extract page list from userspace mapping */
ret = vb2_dc_get_user_pages(start, pages, n_pages, vma, write);
if (ret) {