From: Russell King rmk+kernel@arm.linux.org.uk
Bypass the iommu when we are dealing with single-entry scatterlists.
The etnaviv iommu code needs to be more inteligent: as it currently stands, it is unusable as it always allocates from the bottom upwards. This causes entries to be re-used without the MMU TLB being flushed, but in order to flush the MMU TLB, we have to insert a command into the GPU command stream. Doing this for every allocation/free is really sub-optimal.
To get things working as it currently stands, bypass this so that the armada DRM scanout buffer can at least be used with etnaviv DRM. This at least gets us /some/ usable acceleration on Dove.
To fix this properly, the MMU handing needs to be re-evaluated and probably rewritten.
Signed-off-by: Russell King rmk+kernel@arm.linux.org.uk --- drivers/staging/etnaviv/etnaviv_mmu.c | 16 +++++++++++++++- 1 file changed, 15 insertions(+), 1 deletion(-)
diff --git a/drivers/staging/etnaviv/etnaviv_mmu.c b/drivers/staging/etnaviv/etnaviv_mmu.c index 2effe75cb154..4589995b83ff 100644 --- a/drivers/staging/etnaviv/etnaviv_mmu.c +++ b/drivers/staging/etnaviv/etnaviv_mmu.c @@ -99,6 +99,20 @@ int etnaviv_iommu_map_gem(struct etnaviv_iommu *mmu, struct drm_mm_node *node = NULL; int ret;
+ /* v1 MMU can optimize single entry (contiguous) scatterlists */ + if (sgt->nents == 1) { + uint32_t iova; + + iova = sg_dma_address(sgt->sgl); + if (!iova) + iova = sg_phys(sgt->sgl) - sgt->sgl->offset; + + if (iova < 0x80000000 - sg_dma_len(sgt->sgl)) { + etnaviv_obj->iova = iova; + return 0; + } + } + node = kzalloc(sizeof(*node), GFP_KERNEL); if (!node) return -ENOMEM; @@ -123,7 +137,7 @@ int etnaviv_iommu_map_gem(struct etnaviv_iommu *mmu, void etnaviv_iommu_unmap_gem(struct etnaviv_iommu *mmu, struct etnaviv_gem_object *etnaviv_obj) { - if (etnaviv_obj->iova) { + if (etnaviv_obj->gpu_vram_node) { uint32_t offset = etnaviv_obj->gpu_vram_node->start;
etnaviv_iommu_unmap(mmu, offset, etnaviv_obj->sgt,