From: Russell King rmk+kernel@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@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);