From: Russell King <rmk+ker...@arm.linux.org.uk>
We test for write-combine and non-cacheable mappings before calling the
DMA API. This is werid, because non-cacheable mappings are DMA coherent
by definition, whereas cacheable mappings need cache maintanence
provided by the DMA API.
This seems to be a typo: ETNA_BO_CACHED should be used rather than
ETNA_BO_UNCACHED.
Moreover, add a comment to the dma_unmap_sg() site so to remind people
about the data-corrupting implications of this call if it is abused (as
can happen with the etnaviv DRM code structure as it currently stands
with long-term mapping of the buffer.)
Signed-off-by: Russell King <rmk+kernel at arm.linux.org.uk>
---
drivers/staging/etnaviv/etnaviv_gem.c | 18 +++++++++++++++---
1 file changed, 15 insertions(+), 3 deletions(-)
diff --git a/drivers/staging/etnaviv/etnaviv_gem.c
b/drivers/staging/etnaviv/etnaviv_gem.c
index e508d64aa2d3..034ff732bdf4 100644
--- a/drivers/staging/etnaviv/etnaviv_gem.c
+++ b/drivers/staging/etnaviv/etnaviv_gem.c
@@ -48,7 +48,7 @@ static int etnaviv_gem_shmem_get_pages(struct
etnaviv_gem_object *etnaviv_obj)
/* For non-cached buffers, ensure the new pages are clean
* because display controller, GPU, etc. are not coherent:
*/
- if (etnaviv_obj->flags & (ETNA_BO_WC|ETNA_BO_UNCACHED))
+ if (etnaviv_obj->flags & (ETNA_BO_WC|ETNA_BO_CACHED))
dma_map_sg(dev->dev, etnaviv_obj->sgt->sgl,
etnaviv_obj->sgt->nents, DMA_BIDIRECTIONAL);
@@ -60,10 +60,22 @@ static void put_pages(struct etnaviv_gem_object
*etnaviv_obj)
if (etnaviv_obj->pages) {
struct drm_device *dev = etnaviv_obj->base.dev;
- /* For non-cached buffers, ensure the new pages are clean
+ /*
+ * For non-cached buffers, ensure the new pages are clean
* because display controller, GPU, etc. are not coherent:
+ *
+ * WARNING: The DMA API does not support concurrent CPU
+ * and device access to the memory area. With BIDIRECTIONAL,
+ * we will clean the cache lines which overlap the region,
+ * and invalidate all cache lines (partially) contained in
+ * the region.
+ *
+ * If you have dirty data in the overlapping cache lines,
+ * that will corrupt the GPU-written data. If you have
+ * written into the remainder of the region, this can
+ * discard those writes.
*/
- if (etnaviv_obj->flags & (ETNA_BO_WC|ETNA_BO_UNCACHED))
+ if (etnaviv_obj->flags & (ETNA_BO_WC|ETNA_BO_CACHED))
dma_unmap_sg(dev->dev, etnaviv_obj->sgt->sgl,
etnaviv_obj->sgt->nents,
DMA_BIDIRECTIONAL);
--
2.1.4
