gpu: nvgpu: Reduce usage of nvgpu_vidmem_get_page_alloc

Reduce the usage of nvgpu_vidmem_get_page_alloc() and friends as much
as possible. This reduces the dependency of nvgpu on Linux SGLs. SGLs
still need to be used, however, since sharing buffers in userspace is
done by dma_buf FD. The best way to pass the vidmem buf through the
dma_buf is by SGL pointer.

JIRA NVGPU-30
JIRA NVGPU-138

Change-Id: Ide0e9e5a557f00aa63b063be085042101a5b34ee
Signed-off-by: Alex Waterman <alexw@nvidia.com>
Reviewed-on: https://git-master.nvidia.com/r/1540709
Reviewed-by: mobile promotions <svcmobile_promotions@nvidia.com>
Tested-by: mobile promotions <svcmobile_promotions@nvidia.com>

7 files changed, 29 insertions, 20 deletions

diff --git a/drivers/gpu/nvgpu/common/linux/dma.c b/drivers/gpu/nvgpu/common/linux/dma.c
index ca657da2..b62c4593 100644
--- a/drivers/gpu/nvgpu/common/linux/dma.c
+++ b/drivers/gpu/nvgpu/common/linux/dma.c
@@ -28,9 +28,11 @@
 #include <nvgpu/vidmem.h>
 
 #include <nvgpu/linux/dma.h>
+#include <nvgpu/linux/vidmem.h>
 
 #include "gk20a/gk20a.h"
 #include "gk20a/platform_gk20a.h"
+
 #include "os_linux.h"
 
 /*
@@ -348,6 +350,7 @@ int nvgpu_dma_alloc_flags_vid_at(struct gk20a *g, unsigned long flags,
 
         mem->aligned_size = size;
         mem->aperture = APERTURE_VIDMEM;
+        mem->vidmem_alloc = (struct nvgpu_page_alloc *)(uintptr_t)addr;
         mem->allocator = vidmem_alloc;
         mem->priv.flags = flags;
 
diff --git a/drivers/gpu/nvgpu/common/linux/nvgpu_mem.c b/drivers/gpu/nvgpu/common/linux/nvgpu_mem.c
index e1f8a598..8740ac3d 100644
--- a/drivers/gpu/nvgpu/common/linux/nvgpu_mem.c
+++ b/drivers/gpu/nvgpu/common/linux/nvgpu_mem.c
@@ -25,6 +25,7 @@
 #include <nvgpu/vidmem.h>
 
 #include <nvgpu/linux/dma.h>
+#include <nvgpu/linux/vidmem.h>
 
 #include "os_linux.h"
 
@@ -355,7 +356,7 @@ u64 nvgpu_mem_get_addr(struct gk20a *g, struct nvgpu_mem *mem)
         /*
          * Otherwise get the vidmem address.
          */
-        alloc = nvgpu_vidmem_get_page_alloc(mem->priv.sgt->sgl);
+        alloc = mem->vidmem_alloc;
 
         /* This API should not be used with > 1 chunks */
         WARN_ON(alloc->nr_chunks != 1);
diff --git a/drivers/gpu/nvgpu/common/mm/vidmem.c b/drivers/gpu/nvgpu/common/mm/vidmem.c
index 18f46c03..d1c5a2e8 100644
--- a/drivers/gpu/nvgpu/common/mm/vidmem.c
+++ b/drivers/gpu/nvgpu/common/mm/vidmem.c
@@ -195,7 +195,7 @@ int nvgpu_vidmem_clear(struct gk20a *g, struct nvgpu_mem *mem)
         if (g->mm.vidmem.ce_ctx_id == (u32)~0)
                 return -EINVAL;
 
-        alloc = nvgpu_vidmem_get_page_alloc(mem->priv.sgt->sgl);
+        alloc = mem->vidmem_alloc;
 
         nvgpu_sgt_for_each_sgl(sgl, &alloc->sgt) {
                 if (gk20a_last_fence)
diff --git a/drivers/gpu/nvgpu/common/pramin.c b/drivers/gpu/nvgpu/common/pramin.c
index b7bc7439..b6166f51 100644
--- a/drivers/gpu/nvgpu/common/pramin.c
+++ b/drivers/gpu/nvgpu/common/pramin.c
@@ -23,7 +23,6 @@
 #include <nvgpu/pramin.h>
 #include <nvgpu/page_allocator.h>
 #include <nvgpu/enabled.h>
-#include <nvgpu/vidmem.h>
 
 #include "gk20a/gk20a.h"
 
@@ -55,7 +54,7 @@ void nvgpu_pramin_access_batched(struct gk20a *g, struct nvgpu_mem *mem,
         if (!g->regs && nvgpu_is_enabled(g, NVGPU_DRIVER_IS_DYING))
                 return;
 
-        alloc = nvgpu_vidmem_get_page_alloc(mem->priv.sgt->sgl);
+        alloc = mem->vidmem_alloc;
         sgt = &alloc->sgt;
 
         nvgpu_sgt_for_each_sgl(sgl, sgt) {
diff --git a/drivers/gpu/nvgpu/include/nvgpu/linux/vidmem.h b/drivers/gpu/nvgpu/include/nvgpu/linux/vidmem.h
index 76bbb05b..ec02faec 100644
--- a/drivers/gpu/nvgpu/include/nvgpu/linux/vidmem.h
+++ b/drivers/gpu/nvgpu/include/nvgpu/linux/vidmem.h
@@ -24,13 +24,18 @@ struct dma_buf;
 struct gk20a;
 
 #ifdef CONFIG_GK20A_VIDMEM
+
 struct gk20a *nvgpu_vidmem_buf_owner(struct dma_buf *dmabuf);
 int nvgpu_vidmem_export_linux(struct gk20a *g, size_t bytes);
 
+void nvgpu_vidmem_set_page_alloc(struct scatterlist *sgl, u64 addr);
+struct nvgpu_page_alloc *nvgpu_vidmem_get_page_alloc(struct scatterlist *sgl);
+
 int nvgpu_vidmem_buf_access_memory(struct gk20a *g, struct dma_buf *dmabuf,
                 void *buffer, u64 offset, u64 size, u32 cmd);
 
 #else /* !CONFIG_GK20A_VIDMEM */
+
 static inline struct gk20a *nvgpu_vidmem_buf_owner(struct dma_buf *dmabuf)
 {
         return NULL;
@@ -41,16 +46,28 @@ static inline int nvgpu_vidmem_export_linux(struct gk20a *g, size_t bytes)
         return -ENOSYS;
 }
 
+static inline void nvgpu_vidmem_set_page_alloc(struct scatterlist *sgl,
+                                               u64 addr)
+{
+}
+
+static inline struct nvgpu_page_alloc *nvgpu_vidmem_get_page_alloc(
+        struct scatterlist *sgl)
+{
+        return NULL;
+}
+
 static inline int nvgpu_vidmem_buf_access_memory(struct gk20a *g,
-                                             struct dma_buf *dmabuf,
-                                             void *buffer, u64 offset,
-                                             u64 size, u32 cmd)
+                                                 struct dma_buf *dmabuf,
+                                                 void *buffer, u64 offset,
+                                                 u64 size, u32 cmd)
 {
         return -ENOSYS;
 }
 
 #endif
 
+
 struct nvgpu_vidmem_linux {
         struct dma_buf  *dmabuf;
         void            *dmabuf_priv;
diff --git a/drivers/gpu/nvgpu/include/nvgpu/nvgpu_mem.h b/drivers/gpu/nvgpu/include/nvgpu/nvgpu_mem.h
index 23a1bad7..537409a8 100644
--- a/drivers/gpu/nvgpu/include/nvgpu/nvgpu_mem.h
+++ b/drivers/gpu/nvgpu/include/nvgpu/nvgpu_mem.h
@@ -39,6 +39,7 @@ struct nvgpu_sgt;
 struct gk20a;
 struct nvgpu_allocator;
 struct nvgpu_gmmu_attrs;
+struct nvgpu_page_alloc;
 
 #define NVGPU_MEM_DMA_ERROR             (~0ULL)
 
@@ -162,6 +163,7 @@ struct nvgpu_mem {
         /*
          * Fields only populated for vidmem allocations.
          */
+        struct nvgpu_page_alloc                 *vidmem_alloc;
         struct nvgpu_allocator                  *allocator;
         struct nvgpu_list_node                   clear_list_entry;
 
diff --git a/drivers/gpu/nvgpu/include/nvgpu/vidmem.h b/drivers/gpu/nvgpu/include/nvgpu/vidmem.h
index b89c710d..9e9f8301 100644
--- a/drivers/gpu/nvgpu/include/nvgpu/vidmem.h
+++ b/drivers/gpu/nvgpu/include/nvgpu/vidmem.h
@@ -27,7 +27,6 @@
 #include <nvgpu/errno.h>
 #include <nvgpu/nvgpu_mem.h>
 
-struct scatterlist;
 struct work_struct;
 
 struct gk20a;
@@ -74,8 +73,6 @@ struct nvgpu_vidmem_buf *nvgpu_vidmem_user_alloc(struct gk20a *g, size_t bytes);
 
 void nvgpu_vidmem_buf_free(struct gk20a *g, struct nvgpu_vidmem_buf *buf);
 
-struct nvgpu_page_alloc *nvgpu_vidmem_get_page_alloc(struct scatterlist *sgl);
-void nvgpu_vidmem_set_page_alloc(struct scatterlist *sgl, u64 addr);
 bool nvgpu_addr_is_vidmem_page_alloc(u64 addr);
 int nvgpu_vidmem_get_space(struct gk20a *g, u64 *space);
 
@@ -93,16 +90,6 @@ int nvgpu_vidmem_clear(struct gk20a *g, struct nvgpu_mem *mem);
  * When VIDMEM support is not present this interface is used.
  */
 
-static inline struct nvgpu_page_alloc *
-nvgpu_vidmem_get_page_alloc(struct scatterlist *sgl)
-{
-        return NULL;
-}
-
-static inline void nvgpu_vidmem_set_page_alloc(struct scatterlist *sgl, u64 addr)
-{
-}
-
 static inline bool nvgpu_addr_is_vidmem_page_alloc(u64 addr)
 {
         return false;