gpu: nvgpu: Pass correct va_allocated field in .gmmu_unmap()

When nvgpu maps an nvgpu_mem struct the nvgpu driver has a choice of
either using a fixed or non-fixed mapping. For non-fixed mappings the
GMMU APIs allocate a VA space for the caller. In that case the GMMU
APIs must also free that VA range when nvgpu unmaps the nvgpu_mem.

For fixed mappings the GMMU APIs must instead not manage the life
time of the VA space. To support these two possibilities add a field
to nvgpu_mem that specifies whether the GMMU APIs must or must not
free the GPU VA range during the GMMU unmap operation.

Also fix a case in the nvgpu vm_area code that would double free a
VA allocation in some cases (sparse allocs).

Change-Id: Idc32dbb8208fa7c1c05823e67b54707fea51c6b7
Signed-off-by: Alex Waterman <alexw@nvidia.com>
Reviewed-on: https://git-master.nvidia.com/r/1669920
Reviewed-by: mobile promotions <svcmobile_promotions@nvidia.com>
Tested-by: mobile promotions <svcmobile_promotions@nvidia.com>

1 files changed, 13 insertions, 1 deletions

diff --git a/drivers/gpu/nvgpu/common/mm/gmmu.c b/drivers/gpu/nvgpu/common/mm/gmmu.c
index 44e540dc..89d71b3b 100644
--- a/drivers/gpu/nvgpu/common/mm/gmmu.c
+++ b/drivers/gpu/nvgpu/common/mm/gmmu.c
@@ -86,6 +86,18 @@ static u64 __nvgpu_gmmu_map(struct vm_gk20a *vm,
         if (nvgpu_is_enabled(g, NVGPU_USE_COHERENT_SYSMEM))
                 flags |= NVGPU_VM_MAP_IO_COHERENT;
 
+        /*
+         * Later on, when we free this nvgpu_mem's GPU mapping, we are going to
+         * potentially have to free the GPU VA space. If the address passed in
+         * is non-zero then this API is not expected to manage the VA space and
+         * therefor we should not try and free it. But otherwise, if we do
+         * manage the VA alloc, we obviously must free it.
+         */
+        if (addr != 0)
+                mem->free_gpu_va = false;
+        else
+                mem->free_gpu_va = true;
+
         nvgpu_mutex_acquire(&vm->update_gmmu_lock);
         vaddr = g->ops.mm.gmmu_map(vm, addr,
                                    sgt,    /* sg list */
@@ -152,7 +164,7 @@ void nvgpu_gmmu_unmap(struct vm_gk20a *vm, struct nvgpu_mem *mem, u64 gpu_va)
                              gpu_va,
                              mem->size,
                              gmmu_page_size_kernel,
-                             true, /*va_allocated */
+                             mem->free_gpu_va,
                              gk20a_mem_flag_none,
                              false,
                              NULL);