gpu: nvgpu: use NO_KERNEL_MAPPING for default vidmem

Commit 8f3875393e7a6bd0fc03afdb1fa99b7e33b71576 ("abstract away dma
alloc attrs") added an implicit NVGPU_DMA_NO_KERNEL_MAPPING for the
explicit vidmem buffer allocation path. The default allocation api that
chooses vidmem or sysmem based on available support still passed a zero
flag and produced a warning when the flag should have been there. Force
the NO_KERNEL_MAPPING flag on currently via the default-allocator api.

Commit 8a15e02ca92b83aa5a216ea9cd42680373212ecd ("gpu: nvgpu: add
NO_KERNEL_MAPPING for alloc_map_vid") did the same for flagless
alloc-and-map calls but this default alloc-only call got overlooked.

Also, store the flags in the mem_desc during allocation, just like in
the sysmem allocator, to be checked during freeing.

Bug 1896734
Bug 1853519

Change-Id: I4b4182b4fd52298bdd2a77197c095a23d2a67c4a
Signed-off-by: Konsta Holtta <kholtta@nvidia.com>
Reviewed-on: http://git-master/r/1331252
Reviewed-by: mobile promotions <svcmobile_promotions@nvidia.com>
Tested-by: mobile promotions <svcmobile_promotions@nvidia.com>

1 files changed, 10 insertions, 1 deletions

diff --git a/drivers/gpu/nvgpu/gk20a/mm_gk20a.c b/drivers/gpu/nvgpu/gk20a/mm_gk20a.c
index 709f2fd2..9a2dbeb9 100644
--- a/drivers/gpu/nvgpu/gk20a/mm_gk20a.c
+++ b/drivers/gpu/nvgpu/gk20a/mm_gk20a.c
@@ -2735,7 +2735,15 @@ int gk20a_gmmu_alloc_flags(struct gk20a *g, unsigned long flags, size_t size,
                 struct mem_desc *mem)
 {
         if (g->mm.vidmem_is_vidmem) {
-                int err = gk20a_gmmu_alloc_flags_vid(g, flags, size, mem);
+                /*
+                 * Force the no-kernel-mapping flag on because we don't support
+                 * the lack of it for vidmem - the user should not care when
+                 * using gk20a_gmmu_alloc_map and it's vidmem, or if there's a
+                 * difference, the user should use the flag explicitly anyway.
+                 */
+                int err = gk20a_gmmu_alloc_flags_vid(g,
+                                flags | NVGPU_DMA_NO_KERNEL_MAPPING,
+                                size, mem);
 
                 if (!err)
                         return 0;
@@ -3012,6 +3020,7 @@ int gk20a_gmmu_alloc_flags_vid_at(struct gk20a *g, unsigned long flags,
         mem->size = size;
         mem->aperture = APERTURE_VIDMEM;
         mem->allocator = vidmem_alloc;
+        mem->flags = flags;
 
         nvgpu_init_list_node(&mem->clear_list_entry);