Revert "Revert "gpu: nvgpu: Get coherency on gv100 + NVLINK working""

This reverts commit 5a35a95654d561fce09a3b9abf6b82bb7a29d74b.

JIRA EVLR-2333

Change-Id: I923c32496c343d39d34f6d406c38a9f6ce7dc6e0
Signed-off-by: Alex Waterman <alexw@nvidia.com>
Reviewed-on: https://git-master.nvidia.com/r/1667167
Reviewed-by: svc-mobile-coverity <svc-mobile-coverity@nvidia.com>
GVS: Gerrit_Virtual_Submit
Reviewed-by: Terje Bergstrom <tbergstrom@nvidia.com>
Reviewed-by: mobile promotions <svcmobile_promotions@nvidia.com>
Tested-by: mobile promotions <svcmobile_promotions@nvidia.com>

1 files changed, 15 insertions, 1 deletions

diff --git a/drivers/gpu/nvgpu/common/mm/gmmu.c b/drivers/gpu/nvgpu/common/mm/gmmu.c
index e1942cbd..3b57e781 100644
--- a/drivers/gpu/nvgpu/common/mm/gmmu.c
+++ b/drivers/gpu/nvgpu/common/mm/gmmu.c
@@ -79,6 +79,13 @@ static u64 __nvgpu_gmmu_map(struct vm_gk20a *vm,
         if (!sgt)
                 return -ENOMEM;
 
+        /*
+         * If the GPU is IO coherent and the DMA API is giving us IO coherent
+         * CPU mappings then we gotta make sure we use the IO coherent aperture.
+         */
+        if (nvgpu_is_enabled(g, NVGPU_USE_COHERENT_SYSMEM))
+                flags |= NVGPU_VM_MAP_IO_COHERENT;
+
         nvgpu_mutex_acquire(&vm->update_gmmu_lock);
         vaddr = g->ops.mm.gmmu_map(vm, addr,
                                    sgt,    /* sg list */
@@ -627,7 +634,7 @@ static int __nvgpu_gmmu_update_page_table(struct vm_gk20a *vm,
                    page_size >> 10,
                    nvgpu_gmmu_perm_str(attrs->rw_flag),
                    attrs->kind_v,
-                   nvgpu_aperture_str(attrs->aperture),
+                   nvgpu_aperture_str(g, attrs->aperture),
                    attrs->cacheable ? 'C' : '-',
                    attrs->sparse    ? 'S' : '-',
                    attrs->priv      ? 'P' : '-',
@@ -705,6 +712,13 @@ u64 gk20a_locked_gmmu_map(struct vm_gk20a *vm,
         attrs.l3_alloc = (bool)(flags & NVGPU_VM_MAP_L3_ALLOC);
 
         /*
+         * Handle the IO coherency aperture: make sure the .aperture field is
+         * correct based on the IO coherency flag.
+         */
+        if (attrs.coherent && attrs.aperture == APERTURE_SYSMEM)
+                attrs.aperture = __APERTURE_SYSMEM_COH;
+
+        /*
          * Only allocate a new GPU VA range if we haven't already been passed a
          * GPU VA range. This facilitates fixed mappings.
          */