video: tegra: nvmap: implement new cache operation

Provide the ability to flush a list of handles. This code will
optimize the cache flush to be full inner/outer flushes if the
total size of the passed list of handles exceeds the inner cache
maintenance threshold.

Bug 1373180

Change-Id: I357b5d5d1a9697aa6487ad6ef6723448a368a45b
Signed-off-by: Alex Waterman <alexw@nvidia.com>
Signed-off-by: Sri Krishna chowdary <schowdary@nvidia.com>
Reviewed-on: http://git-master/r/330638
Reviewed-by: Hiroshi Doyu <hdoyu@nvidia.com>
Tested-by: Hiroshi Doyu <hdoyu@nvidia.com>

4 files changed, 87 insertions, 39 deletions

diff --git a/drivers/video/tegra/nvmap/nvmap_dmabuf.c b/drivers/video/tegra/nvmap/nvmap_dmabuf.c
index 370a6c702b26..46f4f63ca45a 100644
--- a/drivers/video/tegra/nvmap/nvmap_dmabuf.c
+++ b/drivers/video/tegra/nvmap/nvmap_dmabuf.c
@@ -458,8 +458,8 @@ static int nvmap_dmabuf_begin_cpu_access(struct dma_buf *dmabuf,
 	struct nvmap_handle_info *info = dmabuf->priv;
 
 	trace_nvmap_dmabuf_begin_cpu_access(dmabuf, start, len);
-	return __nvmap_cache_maint(NULL, info->handle, start, start + len,
-				   NVMAP_CACHE_OP_INV, 1);
+	return __nvmap_do_cache_maint(NULL, info->handle, start, start + len,
+				      NVMAP_CACHE_OP_INV, 1);
 }
 
 static void nvmap_dmabuf_end_cpu_access(struct dma_buf *dmabuf,
@@ -469,7 +469,7 @@ static void nvmap_dmabuf_end_cpu_access(struct dma_buf *dmabuf,
 	struct nvmap_handle_info *info = dmabuf->priv;
 
 	trace_nvmap_dmabuf_end_cpu_access(dmabuf, start, len);
-	__nvmap_cache_maint(NULL, info->handle, start, start + len,
+	__nvmap_do_cache_maint(NULL, info->handle, start, start + len,
 				   NVMAP_CACHE_OP_WB_INV, 1);
 
 }
diff --git a/drivers/video/tegra/nvmap/nvmap_ioctl.c b/drivers/video/tegra/nvmap/nvmap_ioctl.c
index b449d40051fe..5a89be557860 100644
--- a/drivers/video/tegra/nvmap/nvmap_ioctl.c
+++ b/drivers/video/tegra/nvmap/nvmap_ioctl.c
@@ -646,13 +646,9 @@ int nvmap_ioctl_rw_handle(struct file *filp, int is_read, void __user *arg,
 	return err;
 }
 
-int nvmap_ioctl_cache_maint(struct file *filp, void __user *arg, bool is32)
+static int __nvmap_cache_maint(struct nvmap_client *client,
+			       struct nvmap_cache_op *op)
 {
-	struct nvmap_client *client = filp->private_data;
-	struct nvmap_cache_op op;
-#ifdef CONFIG_COMPAT
-	struct nvmap_cache_op_32 op32;
-#endif
 	struct vm_area_struct *vma;
 	struct nvmap_vma_priv *vpriv;
 	struct nvmap_handle *handle;
@@ -660,31 +656,18 @@ int nvmap_ioctl_cache_maint(struct file *filp, void __user *arg, bool is32)
 	unsigned long end;
 	int err = 0;
 
-#ifdef CONFIG_COMPAT
-	if (is32) {
-		if (copy_from_user(&op32, arg, sizeof(op32)))
-			return -EFAULT;
-		op.addr = op32.addr;
-		op.handle = op32.handle;
-		op.len = op32.len;
-		op.op = op32.op;
-	} else
-#endif
-		if (copy_from_user(&op, arg, sizeof(op)))
-			return -EFAULT;
-
-	handle = unmarshal_user_handle(op.handle);
-	if (!handle || !op.addr || op.op < NVMAP_CACHE_OP_WB ||
-	    op.op > NVMAP_CACHE_OP_WB_INV)
+	handle = unmarshal_user_handle(op->handle);
+	if (!handle || !op->addr || op->op < NVMAP_CACHE_OP_WB ||
+	    op->op > NVMAP_CACHE_OP_WB_INV)
 		return -EINVAL;
 
 	down_read(&current->mm->mmap_sem);
 
-	vma = find_vma(current->active_mm, op.addr);
+	vma = find_vma(current->active_mm, (unsigned long)op->addr);
 	if (!vma || !is_nvmap_vma(vma) ||
-	    (ulong)op.addr < vma->vm_start ||
-	    (ulong)op.addr >= vma->vm_end ||
-	    op.len > vma->vm_end - (ulong)op.addr) {
+	    (ulong)op->addr < vma->vm_start ||
+	    (ulong)op->addr >= vma->vm_end ||
+	    op->len > vma->vm_end - (ulong)op->addr) {
 		err = -EADDRNOTAVAIL;
 		goto out;
 	}
@@ -696,16 +679,41 @@ int nvmap_ioctl_cache_maint(struct file *filp, void __user *arg, bool is32)
 		goto out;
 	}
 
-	start = op.addr - vma->vm_start + (vma->vm_pgoff << PAGE_SHIFT);
-	end = start + op.len;
+	start = (unsigned long)op->addr - vma->vm_start +
+		(vma->vm_pgoff << PAGE_SHIFT);
+	end = start + op->len;
 
-	err = __nvmap_cache_maint(client, vpriv->handle, start, end, op.op,
-				  CACHE_MAINT_ALLOW_DEFERRED);
+	err = __nvmap_do_cache_maint(client, vpriv->handle, start, end, op->op,
+				     CACHE_MAINT_ALLOW_DEFERRED);
 out:
 	up_read(&current->mm->mmap_sem);
 	return err;
 }
 
+int nvmap_ioctl_cache_maint(struct file *filp, void __user *arg, bool is32)
+{
+	struct nvmap_client *client = filp->private_data;
+	struct nvmap_cache_op op;
+#ifdef CONFIG_COMPAT
+	struct nvmap_cache_op_32 op32;
+#endif
+
+#ifdef CONFIG_COMPAT
+	if (is32) {
+		if (copy_from_user(&op32, arg, sizeof(op32)))
+			return -EFAULT;
+		op.addr = op32.addr;
+		op.handle = op32.handle;
+		op.len = op32.len;
+		op.op = op32.op;
+	} else
+#endif
+		if (copy_from_user(&op, arg, sizeof(op)))
+			return -EFAULT;
+
+	return __nvmap_cache_maint(client, &op);
+}
+
 int nvmap_ioctl_free(struct file *filp, unsigned long arg)
 {
 	struct nvmap_client *client = filp->private_data;
@@ -938,7 +946,7 @@ out:
 	return err;
 }
 
-int __nvmap_cache_maint(struct nvmap_client *client,
+int __nvmap_do_cache_maint(struct nvmap_client *client,
 			struct nvmap_handle *h,
 			unsigned long start, unsigned long end,
 			unsigned int op, unsigned int allow_deferred)
@@ -1048,7 +1056,7 @@ static ssize_t rw_handle(struct nvmap_client *client, struct nvmap_handle *h,
 			break;
 		}
 		if (is_read)
-			__nvmap_cache_maint(client, h, h_offs,
+			__nvmap_do_cache_maint(client, h, h_offs,
 				h_offs + elem_size, NVMAP_CACHE_OP_INV,
 				CACHE_MAINT_IMMEDIATE);
 
@@ -1059,7 +1067,7 @@ static ssize_t rw_handle(struct nvmap_client *client, struct nvmap_handle *h,
 			break;
 
 		if (!is_read)
-			__nvmap_cache_maint(client, h, h_offs,
+			__nvmap_do_cache_maint(client, h, h_offs,
 				h_offs + elem_size, NVMAP_CACHE_OP_WB_INV,
 				CACHE_MAINT_IMMEDIATE);
 
diff --git a/drivers/video/tegra/nvmap/nvmap_mm.c b/drivers/video/tegra/nvmap/nvmap_mm.c
index 2af163279fc7..fd29f8b551a9 100644
--- a/drivers/video/tegra/nvmap/nvmap_mm.c
+++ b/drivers/video/tegra/nvmap/nvmap_mm.c
@@ -83,3 +83,41 @@ void nvmap_flush_cache(struct page **pages, int numpages)
 #endif
 	}
 }
+
+/*
+ * Flush the list of passed handles. This will optimze the flush if it can.
+ * In the case that all the handles together are larger than the inner cache
+ * maint threshold it is possible to just do an entire inner cache flush.
+ */
+int nvmap_flush_cache_list(struct nvmap_handle **handles, int nr)
+{
+	int i, err = 0;
+	u64 total = 0;
+
+	for (i = 0; i < nr; i++)
+		total += handles[i]->size;
+
+	/* Full flush in the case the passed list is bigger than our
+	 * threshold. */
+	if (total >= cache_maint_inner_threshold) {
+		inner_flush_cache_all();
+		outer_flush_all();
+		nvmap_stats_inc(NS_CFLUSH_RQ, total);
+		nvmap_stats_inc(NS_CFLUSH_DONE, cache_maint_inner_threshold);
+		trace_nvmap_cache_flush(total,
+					nvmap_stats_read(NS_ALLOC),
+					nvmap_stats_read(NS_CFLUSH_RQ),
+					nvmap_stats_read(NS_CFLUSH_DONE));
+	} else {
+		for (i = 0; i < nr; i++) {
+			err = __nvmap_do_cache_maint(handles[i]->owner,
+						     handles[i], 0,
+						     handles[i]->size,
+						     NVMAP_CACHE_OP_WB_INV, 0);
+			if (err)
+				break;
+		}
+	}
+
+	return err;
+}
diff --git a/drivers/video/tegra/nvmap/nvmap_priv.h b/drivers/video/tegra/nvmap/nvmap_priv.h
index 13be4a1aa1f0..22e2df786442 100644
--- a/drivers/video/tegra/nvmap/nvmap_priv.h
+++ b/drivers/video/tegra/nvmap/nvmap_priv.h
@@ -410,6 +410,8 @@ void inner_flush_cache_all(void);
 void inner_clean_cache_all(void);
 void nvmap_flush_cache(struct page **pages, int numpages);
 
+int nvmap_flush_cache_list(struct nvmap_handle **handles, int nr);
+
 /* Internal API to support dmabuf */
 struct dma_buf *__nvmap_dmabuf_export(struct nvmap_client *client,
 				 struct nvmap_handle *handle);
@@ -426,9 +428,9 @@ void __nvmap_munmap(struct nvmap_handle *h, void *addr);
 int __nvmap_map(struct nvmap_handle *h, struct vm_area_struct *vma);
 int __nvmap_get_handle_param(struct nvmap_client *client,
 			     struct nvmap_handle *h, u32 param, u64 *result);
-int __nvmap_cache_maint(struct nvmap_client *client, struct nvmap_handle *h,
-			unsigned long start, unsigned long end,
-			unsigned int op, unsigned int allow_deferred);
+int __nvmap_do_cache_maint(struct nvmap_client *client, struct nvmap_handle *h,
+			   unsigned long start, unsigned long end,
+			   unsigned int op, unsigned int allow_deferred);
 struct nvmap_client *__nvmap_create_client(struct nvmap_device *dev,
 					   const char *name);
 struct dma_buf *__nvmap_dmabuf_export_from_ref(struct nvmap_handle_ref *ref);