drm/etnaviv: add cmdbuf suballocator

There are 3 big benefits to suballocating a single big DMA buffer
for command submission:

1. Avoid hammering CMA. The old way of allocating and freeing a DMA
   buffer for each submission was hitting some of the real slow
   pathes in CMA, as this allocator was not designed for a concurrent
   small buffers load.

2. Less TLB flushes on IOMMUv2. If a new command buffer is mapped into
   the GPU address space the MMU TLBs need to be flushed. By having
   one big buffer statically mapped to the GPU, a lot of those flushes
   can be avoided.

3. No funky workarounds for GC3000. The FE TLB flush on GC3000 isn't
   reliable. To work around that we tried to lay out the cmdbufs in
   the GPU address space in a way to avoid this issue. This hasn't
   always worked if the address space is crowded. A single statically
   mapped buffer avoids the erratum completely.

Signed-off-by: Lucas Stach <l.stach@pengutronix.de>
Reviewed-by: Christian Gmeiner <christian.gmeiner@gmail.com>

1 files changed, 20 insertions, 25 deletions

diff --git a/drivers/gpu/drm/etnaviv/etnaviv_mmu.c b/drivers/gpu/drm/etnaviv/etnaviv_mmu.c
index 056685bd33b2..dcc86d8eeb98 100644
--- a/drivers/gpu/drm/etnaviv/etnaviv_mmu.c
+++ b/drivers/gpu/drm/etnaviv/etnaviv_mmu.c
@@ -321,55 +321,50 @@ void etnaviv_iommu_restore(struct etnaviv_gpu *gpu)
 		etnaviv_iommuv2_restore(gpu);
 }
 
-u32 etnaviv_iommu_get_cmdbuf_va(struct etnaviv_gpu *gpu,
-				struct etnaviv_cmdbuf *buf)
+int etnaviv_iommu_get_suballoc_va(struct etnaviv_gpu *gpu, dma_addr_t paddr,
+				  struct drm_mm_node *vram_node, size_t size,
+				  u32 *iova)
 {
 	struct etnaviv_iommu *mmu = gpu->mmu;
 
 	if (mmu->version == ETNAVIV_IOMMU_V1) {
-		return buf->paddr - gpu->memory_base;
+		*iova = paddr - gpu->memory_base;
+		return 0;
 	} else {
 		int ret;
 
-		if (buf->vram_node.allocated)
-			return (u32)buf->vram_node.start;
-
 		mutex_lock(&mmu->lock);
-		ret = etnaviv_iommu_find_iova(mmu, &buf->vram_node,
-					      buf->size + SZ_64K);
+		ret = etnaviv_iommu_find_iova(mmu, vram_node, size);
 		if (ret < 0) {
 			mutex_unlock(&mmu->lock);
-			return 0;
+			return ret;
 		}
-		ret = iommu_map(mmu->domain, buf->vram_node.start, buf->paddr,
-				buf->size, IOMMU_READ);
+		ret = iommu_map(mmu->domain, vram_node->start, paddr, size,
+				IOMMU_READ);
 		if (ret < 0) {
-			drm_mm_remove_node(&buf->vram_node);
+			drm_mm_remove_node(vram_node);
 			mutex_unlock(&mmu->lock);
-			return 0;
+			return ret;
 		}
-		/*
-		 * At least on GC3000 the FE MMU doesn't properly flush old TLB
-		 * entries. Make sure to space the command buffers out in a way
-		 * that the FE MMU prefetch won't load invalid entries.
-		 */
-		mmu->last_iova = buf->vram_node.start + buf->size + SZ_64K;
+		mmu->last_iova = vram_node->start + size;
 		gpu->mmu->need_flush = true;
 		mutex_unlock(&mmu->lock);
 
-		return (u32)buf->vram_node.start;
+		*iova = (u32)vram_node->start;
+		return 0;
 	}
 }
 
-void etnaviv_iommu_put_cmdbuf_va(struct etnaviv_gpu *gpu,
-				 struct etnaviv_cmdbuf *buf)
+void etnaviv_iommu_put_suballoc_va(struct etnaviv_gpu *gpu,
+				   struct drm_mm_node *vram_node, size_t size,
+				   u32 iova)
 {
 	struct etnaviv_iommu *mmu = gpu->mmu;
 
-	if (mmu->version == ETNAVIV_IOMMU_V2 && buf->vram_node.allocated) {
+	if (mmu->version == ETNAVIV_IOMMU_V2) {
 		mutex_lock(&mmu->lock);
-		iommu_unmap(mmu->domain, buf->vram_node.start, buf->size);
-		drm_mm_remove_node(&buf->vram_node);
+		iommu_unmap(mmu->domain,iova, size);
+		drm_mm_remove_node(vram_node);
 		mutex_unlock(&mmu->lock);
 	}
 }