gpu: nvgpu: Add sub-nvgpu_mem

Add an API for creating a special sub-nvgpu_mem struct. This struct
comes with some fairly important caveats but is very useful for the
semaphore code.

Also, make sure that in nvgpu_mem_begin() and nvgpu_mem_end() no
additional mappings are made if not necessary. More importantly during
nvgpu_mem_end() it would be possible to vunmap() a CPU mapping of a
DMA allocation that does not expect this to happen.

JIRA NVGPU-12
JIRA NVGPU-30

Change-Id: I579429da9ff7288488753a113bafc558e0f17a0f
Signed-off-by: Alex Waterman <alexw@nvidia.com>
Reviewed-on: http://git-master/r/1464077
Reviewed-by: Automatic_Commit_Validation_User
GVS: Gerrit_Virtual_Submit
Reviewed-by: Terje Bergstrom <tbergstrom@nvidia.com>
Reviewed-by: svccoveritychecker <svccoveritychecker@nvidia.com>

drivers/gpu/nvgpu/common/linux/dma.c

@@ -334,7 +334,8 @@ static void nvgpu_dma_free_sys(struct gk20a *g, struct nvgpu_mem *mem)
 {
 	struct device *d = dev_from_gk20a(g);
 
-	if (mem->cpu_va || mem->priv.pages) {
+	if (!(mem->mem_flags & NVGPU_MEM_FLAG_SHADOW_COPY) &&
+	    (mem->cpu_va || mem->priv.pages)) {
 		if (mem->priv.flags) {
 			DEFINE_DMA_ATTRS(dma_attrs);
 

drivers/gpu/nvgpu/common/linux/nvgpu_mem.c

@@ -14,6 +14,7 @@
  * along with this program.  If not, see <http://www.gnu.org/licenses/>.
  */
 
+#include <nvgpu/dma.h>
 #include <nvgpu/nvgpu_mem.h>
 #include <nvgpu/page_allocator.h>
 #include <nvgpu/log.h>
@@ -52,6 +53,14 @@ int nvgpu_mem_begin(struct gk20a *g, struct nvgpu_mem *mem)
 	if (mem->aperture != APERTURE_SYSMEM || g->mm.force_pramin)
 		return 0;
 
+	/*
+	 * A CPU mapping is implicitly made for all SYSMEM DMA allocations that
+	 * don't have NVGPU_DMA_NO_KERNEL_MAPPING. Thus we don't need to make
+	 * another CPU mapping.
+	 */
+	if (!(mem->priv.flags & NVGPU_DMA_NO_KERNEL_MAPPING))
+		return 0;
+
 	if (WARN_ON(mem->cpu_va)) {
 		nvgpu_warn(g, "nested");
 		return -EBUSY;
@@ -73,6 +82,13 @@ void nvgpu_mem_end(struct gk20a *g, struct nvgpu_mem *mem)
 	if (mem->aperture != APERTURE_SYSMEM || g->mm.force_pramin)
 		return;
 
+	/*
+	 * Similar to nvgpu_mem_begin() we don't need to unmap the CPU mapping
+	 * already made by the DMA API.
+	 */
+	if (!(mem->priv.flags & NVGPU_DMA_NO_KERNEL_MAPPING))
+		return;
+
 	vunmap(mem->cpu_va);
 	mem->cpu_va = NULL;
 }
@@ -225,3 +241,57 @@ void nvgpu_memset(struct gk20a *g, struct nvgpu_mem *mem, u32 offset,
 		WARN_ON("Accessing unallocated nvgpu_mem");
 	}
 }
+
+/*
+ * Be careful how you use this! You are responsible for correctly freeing this
+ * memory.
+ */
+int nvgpu_mem_create_from_mem(struct gk20a *g,
+			      struct nvgpu_mem *dest, struct nvgpu_mem *src,
+			      int start_page, int nr_pages)
+{
+	int ret;
+	u64 start = start_page * PAGE_SIZE;
+	u64 size = nr_pages * PAGE_SIZE;
+	dma_addr_t new_iova;
+
+	if (src->aperture != APERTURE_SYSMEM)
+		return -EINVAL;
+
+	/* Some silly things a caller might do... */
+	if (size > src->size)
+		return -EINVAL;
+	if ((start + size) > src->size)
+		return -EINVAL;
+
+	dest->mem_flags = src->mem_flags | NVGPU_MEM_FLAG_SHADOW_COPY;
+	dest->aperture  = src->aperture;
+	dest->skip_wmb  = src->skip_wmb;
+	dest->size      = size;
+
+	/*
+	 * Re-use the CPU mapping only if the mapping was made by the DMA API.
+	 */
+	if (!(src->priv.flags & NVGPU_DMA_NO_KERNEL_MAPPING))
+		dest->cpu_va = src->cpu_va + (PAGE_SIZE * start_page);
+
+	dest->priv.pages = src->priv.pages + start_page;
+	dest->priv.flags = src->priv.flags;
+
+	new_iova = sg_dma_address(src->priv.sgt->sgl) ?
+		sg_dma_address(src->priv.sgt->sgl) + start : 0;
+
+	/*
+	 * Make a new SG table that is based only on the subset of pages that
+	 * is passed to us. This table gets freed by the dma free routines.
+	 */
+	if (src->priv.flags & NVGPU_DMA_NO_KERNEL_MAPPING)
+		ret = gk20a_get_sgtable_from_pages(g->dev, &dest->priv.sgt,
+						   src->priv.pages + start_page,
+						   new_iova, size);
+	else
+		ret = gk20a_get_sgtable(g->dev, &dest->priv.sgt, dest->cpu_va,
+					new_iova, size);
+
+	return ret;
+}

drivers/gpu/nvgpu/include/nvgpu/nvgpu_mem.h

@@ -48,6 +48,13 @@ struct nvgpu_mem {
 	u64					 gpu_va;
 	bool					 skip_wmb;
 
+	/*
+	 * Set when a nvgpu_mem struct is not a "real" nvgpu_mem struct. Instead
+	 * the struct is just a copy of another nvgpu_mem struct.
+	 */
+#define NVGPU_MEM_FLAG_SHADOW_COPY		 (1 << 0)
+	unsigned long				 mem_flags;
+
 	/*
 	 * Only populated for a sysmem allocation.
 	 */
@@ -86,6 +93,42 @@ static inline const char *nvgpu_aperture_str(enum nvgpu_aperture aperture)
 	return "UNKNOWN";
 }
 
+/**
+ * nvgpu_mem_create_from_mem - Create a new nvgpu_mem struct from an old one.
+ *
+ * @g          - The GPU.
+ * @dest       - Destination nvgpu_mem to hold resulting memory description.
+ * @src        - Source memory. Must be valid.
+ * @start_page - Starting page to use.
+ * @nr_pages   - Number of pages to place in the new nvgpu_mem.
+ *
+ * Create a new nvgpu_mem struct describing a subsection of the @src nvgpu_mem.
+ * This will create an nvpgu_mem object starting at @start_page and is @nr_pages
+ * long. This currently only works on SYSMEM nvgpu_mems. If this is called on a
+ * VIDMEM nvgpu_mem then this will return an error.
+ *
+ * There is a _major_ caveat to this API: if the source buffer is freed before
+ * the copy is freed then the copy will become invalid. This is a result from
+ * how typical DMA APIs work: we can't call free on the buffer multiple times.
+ * Nor can we call free on parts of a buffer. Thus the only way to ensure that
+ * the entire buffer is actually freed is to call free once on the source
+ * buffer. Since these nvgpu_mem structs are not ref-counted in anyway it is up
+ * to the caller of this API to _ensure_ that the resulting nvgpu_mem buffer
+ * from this API is freed before the source buffer. Otherwise there can and will
+ * be memory corruption.
+ *
+ * The resulting nvgpu_mem should be released with the nvgpu_dma_free() or the
+ * nvgpu_dma_unmap_free() function depending on whether or not the resulting
+ * nvgpu_mem has been mapped.
+ *
+ * This will return 0 on success. An error is returned if the resulting
+ * nvgpu_mem would not make sense or if a new scatter gather table cannot be
+ * created.
+ */
+int nvgpu_mem_create_from_mem(struct gk20a *g,
+			      struct nvgpu_mem *dest, struct nvgpu_mem *src,
+			      int start_page, int nr_pages);
+
 /*
  * Buffer accessors - wrap between begin() and end() if there is no permanent
  * kernel mapping for this buffer.