gpu: host1x: Utilize IOMMU mapping for firewall-copied buffers

Map firewall-copied buffers into Host1x's IOVA space, otherwise Host1x CDMA can't access the command buffers and all submitted jobs fail if IOMMU and Host1x firewall are enabled in the kernels config.

Signed-off-by Dmitry Osipenko <[email protected]>

---
 drivers/gpu/host1x/job.c | 58 +++++++++++++++++++++++++++++++++++-----
 include/linux/host1x.h   |  4 ++-
 2 files changed, 55 insertions(+), 7 deletions(-)

drivers/gpu/host1x/job.c +52 -6

index e2f4a4d93d20..57384a5b5059 100644
@@ -449,10 +449,13 @@ static int validate(struct host1x_firewall *fw, struct host1x_job_gather *g)
 
 static inline int copy_gathers(struct host1x_job *job, struct device *dev)
 {
+	struct host1x *host = dev_get_drvdata(job->channel->dev->parent);
 	struct host1x_firewall fw;
+	dma_addr_t dma_addr;
 	size_t size = 0;
 	size_t offset = 0;
 	unsigned int i;
+	int err;
 
 	fw.job = job;
 	fw.dev = dev;

@@ -466,23 +469,55 @@ static inline int copy_gathers(struct host1x_job *job, struct device *dev)
 		size += g->words * sizeof(u32);
 	}
 
+	if (host->domain)
+		size = iova_align(&host->iova, size);
+
 	/*
 	 * Try a non-blocking allocation from a higher priority pools first,
 	 * as awaiting for the allocation here is a major performance hit.
 	 */
-	job->gather_copy_mapped = dma_alloc_wc(dev, size, &job->gather_copy,
-					       GFP_NOWAIT);
+	job->gather_copy_mapped = dma_alloc_wc(dev, size,
+					       &job->gather_copy_phys,
+ 					       GFP_NOWAIT);
 
 	/* the higher priority allocation failed, try the generic-blocking */
 	if (!job->gather_copy_mapped)
 		job->gather_copy_mapped = dma_alloc_wc(dev, size,
-						       &job->gather_copy,
+						       &job->gather_copy_phys,
 						       GFP_KERNEL);
 	if (!job->gather_copy_mapped)
 		return -ENOMEM;
 
 	job->gather_copy_size = size;
 
+	if (host->domain) {
+		unsigned long shift;
+
+		shift = iova_shift(&host->iova);
+		job->gather_copy_iova_alloc = alloc_iova(
+					&host->iova, size >> shift,
+					host->iova_end >> shift, true);
+		if (!job->gather_copy_iova_alloc)
+			return -ENOMEM;
+
+		job->gather_copy_iova = iova_dma_addr(
+				&host->iova, job->gather_copy_iova_alloc);
+
+		err = iommu_map(host->domain,
+				job->gather_copy_iova,
+				job->gather_copy_phys,
+				size, IOMMU_READ);
+		if (err) {
+			__free_iova(&host->iova, job->gather_copy_iova_alloc);
+			job->gather_copy_iova_alloc = NULL;
+			return err;
+		}
+
+		dma_addr = job->gather_copy_iova;
+	} else {
+		dma_addr = job->gather_copy_phys;
+	}
+
 	for (i = 0; i < job->num_gathers; i++) {
 		struct host1x_job_gather *g = &job->gathers[i];
 		void *gather;

@@ -494,7 +529,7 @@ static inline int copy_gathers(struct host1x_job *job, struct device *dev)
 		host1x_bo_munmap(g->bo, gather);
 
 		/* Store the location in the buffer */
-		g->base = job->gather_copy;
+		g->base = dma_addr;
 		g->offset = offset;
 
 		/* Validate the job */

@@ -582,9 +617,20 @@ void host1x_job_unpin(struct host1x_job *job)
 
 	job->num_unpins = 0;
 
-	if (job->gather_copy_size)
+	if (job->gather_copy_size) {
 		dma_free_wc(job->channel->dev, job->gather_copy_size,
-			    job->gather_copy_mapped, job->gather_copy);
+			    job->gather_copy_mapped, job->gather_copy_phys);
+
+		if (job->gather_copy_iova_alloc) {
+			iommu_unmap(host->domain,
+				    job->gather_copy_iova,
+				    job->gather_copy_size);
+
+			__free_iova(&host->iova, job->gather_copy_iova_alloc);
+
+			job->gather_copy_iova_alloc = NULL;
+		}
+	}
 }
 EXPORT_SYMBOL(host1x_job_unpin);
 

include/linux/host1x.h +0 -0

index 57d26406bdfd..536a678f81d4 100644
@@ -235,8 +235,10 @@ struct host1x_job {
 	unsigned int num_slots;