drm/amdgpu: Support page table update via CPU
authorHarish Kasiviswanathan <Harish.Kasiviswanathan@amd.com>
Thu, 11 May 2017 23:47:22 +0000 (19:47 -0400)
committerAlex Deucher <alexander.deucher@amd.com>
Fri, 9 Jun 2017 15:30:04 +0000 (11:30 -0400)
v2: Fix logical mistake. If CPU update failed amdgpu_vm_bo_update_mapping()
would not return and instead fall through to SDMA update. Minor change due to
amdgpu_vm_bo_wait() prototype change

Signed-off-by: Harish Kasiviswanathan <Harish.Kasiviswanathan@amd.com>
Reviewed-by: Christian König <christian.koenig@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c

index caca6896476cf7a48ef0a47378cd03a8744077f3..c4f1a305c68c3a78ed3344fc0ee3082085a3d7ca 100644 (file)
@@ -79,6 +79,12 @@ struct amdgpu_pte_update_params {
                     uint64_t flags);
        /* indicate update pt or its shadow */
        bool shadow;
+       /* The next two are used during VM update by CPU
+        *  DMA addresses to use for mapping
+        *  Kernel pointer of PD/PT BO that needs to be updated
+        */
+       dma_addr_t *pages_addr;
+       void *kptr;
 };
 
 /* Helper to disable partial resident texture feature from a fence callback */
@@ -972,10 +978,14 @@ static void amdgpu_vm_cpu_set_ptes(struct amdgpu_pte_update_params *params,
                                   uint64_t flags)
 {
        unsigned int i;
+       uint64_t value;
 
        for (i = 0; i < count; i++) {
+               value = params->pages_addr ?
+                       amdgpu_vm_map_gart(params->pages_addr, addr) :
+                       addr;
                amdgpu_gart_set_pte_pde(params->adev, (void *)pe,
-                                       i, addr, flags);
+                                       i, value, flags);
                addr += incr;
        }
 
@@ -1253,6 +1263,59 @@ static struct amdgpu_bo *amdgpu_vm_get_pt(struct amdgpu_pte_update_params *p,
        return entry->bo;
 }
 
+/**
+ * amdgpu_vm_update_ptes_cpu - Update the page tables in the range
+ *  start - @end using CPU.
+ * See amdgpu_vm_update_ptes for parameter description.
+ *
+ */
+static int amdgpu_vm_update_ptes_cpu(struct amdgpu_pte_update_params *params,
+                                    uint64_t start, uint64_t end,
+                                    uint64_t dst, uint64_t flags)
+{
+       struct amdgpu_device *adev = params->adev;
+       const uint64_t mask = AMDGPU_VM_PTE_COUNT(adev) - 1;
+       void *pe_ptr;
+       uint64_t addr;
+       struct amdgpu_bo *pt;
+       unsigned int nptes;
+       int r;
+
+       /* initialize the variables */
+       addr = start;
+
+       /* walk over the address space and update the page tables */
+       while (addr < end) {
+               pt = amdgpu_vm_get_pt(params, addr);
+               if (!pt) {
+                       pr_err("PT not found, aborting update_ptes\n");
+                       return -EINVAL;
+               }
+
+               WARN_ON(params->shadow);
+
+               r = amdgpu_bo_kmap(pt, &pe_ptr);
+               if (r)
+                       return r;
+
+               pe_ptr += (addr & mask) * 8;
+
+               if ((addr & ~mask) == (end & ~mask))
+                       nptes = end - addr;
+               else
+                       nptes = AMDGPU_VM_PTE_COUNT(adev) - (addr & mask);
+
+               params->func(params, (uint64_t)pe_ptr, dst, nptes,
+                            AMDGPU_GPU_PAGE_SIZE, flags);
+
+               amdgpu_bo_kunmap(pt);
+               addr += nptes;
+               dst += nptes * AMDGPU_GPU_PAGE_SIZE;
+       }
+
+       return 0;
+}
+
 /**
  * amdgpu_vm_update_ptes - make sure that page tables are valid
  *
@@ -1277,6 +1340,10 @@ static int amdgpu_vm_update_ptes(struct amdgpu_pte_update_params *params,
        struct amdgpu_bo *pt;
        unsigned nptes;
 
+       if (params->func == amdgpu_vm_cpu_set_ptes)
+               return amdgpu_vm_update_ptes_cpu(params, start, end,
+                                                dst, flags);
+
        /* walk over the address space and update the page tables */
        for (addr = start; addr < end; addr += nptes) {
                pt = amdgpu_vm_get_pt(params, addr);
@@ -1418,6 +1485,25 @@ static int amdgpu_vm_bo_update_mapping(struct amdgpu_device *adev,
        params.vm = vm;
        params.src = src;
 
+       if (vm->use_cpu_for_update) {
+               /* params.src is used as flag to indicate system Memory */
+               if (pages_addr)
+                       params.src = ~0;
+
+               /* Wait for PT BOs to be free. PTs share the same resv. object
+                * as the root PD BO
+                */
+               r = amdgpu_vm_bo_wait(adev, vm->root.bo);
+               if (unlikely(r))
+                       return r;
+
+               params.func = amdgpu_vm_cpu_set_ptes;
+               params.pages_addr = pages_addr;
+               params.shadow = false;
+               return amdgpu_vm_frag_ptes(&params, start, last + 1,
+                                          addr, flags);
+       }
+
        ring = container_of(vm->entity.sched, struct amdgpu_ring, sched);
 
        /* sync to everything on unmapping */