drm/amdgpu: enable huge page handling in the VM v5
authorAlex Deucher <alexander.deucher@amd.com>
Tue, 25 Jul 2017 20:35:38 +0000 (16:35 -0400)
committerAlex Deucher <alexander.deucher@amd.com>
Tue, 25 Jul 2017 20:35:38 +0000 (16:35 -0400)
The hardware can use huge pages to map 2MB of address space with only one PDE.

v2: few cleanups and rebased
v3: skip PT updates if we are using the PDE
v4: rebased, added support for CPU based updates
v5: fix CPU based updates once more
v6: fix ndw estimation

Signed-off-by: Christian König <christian.koenig@amd.com>
Reviewed-and-tested-by: Felix Kuehling <Felix.Kuehling@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c
drivers/gpu/drm/amd/amdgpu/amdgpu_vm.h

index 088c8504dbe7906ae84e78ec0eae69a9937847cf..250c8e80e646bbf800143937b6c96f6e065235ca 100644 (file)
@@ -352,6 +352,7 @@ static int amdgpu_vm_alloc_levels(struct amdgpu_device *adev,
 
                        entry->bo = pt;
                        entry->addr = 0;
+                       entry->huge_page = false;
                }
 
                if (level < adev->vm_manager.num_level) {
@@ -1117,7 +1118,8 @@ static int amdgpu_vm_update_level(struct amdgpu_device *adev,
 
                pt = amdgpu_bo_gpu_offset(bo);
                pt = amdgpu_gart_get_vm_pde(adev, pt);
-               if (parent->entries[pt_idx].addr == pt)
+               if (parent->entries[pt_idx].addr == pt ||
+                   parent->entries[pt_idx].huge_page)
                        continue;
 
                parent->entries[pt_idx].addr = pt;
@@ -1258,29 +1260,95 @@ int amdgpu_vm_update_directories(struct amdgpu_device *adev,
 }
 
 /**
- * amdgpu_vm_find_pt - find the page table for an address
+ * amdgpu_vm_find_entry - find the entry for an address
  *
  * @p: see amdgpu_pte_update_params definition
  * @addr: virtual address in question
+ * @entry: resulting entry or NULL
+ * @parent: parent entry
  *
- * Find the page table BO for a virtual address, return NULL when none found.
+ * Find the vm_pt entry and it's parent for the given address.
  */
-static struct amdgpu_bo *amdgpu_vm_get_pt(struct amdgpu_pte_update_params *p,
-                                         uint64_t addr)
+void amdgpu_vm_get_entry(struct amdgpu_pte_update_params *p, uint64_t addr,
+                        struct amdgpu_vm_pt **entry,
+                        struct amdgpu_vm_pt **parent)
 {
-       struct amdgpu_vm_pt *entry = &p->vm->root;
        unsigned idx, level = p->adev->vm_manager.num_level;
 
-       while (entry->entries) {
+       *parent = NULL;
+       *entry = &p->vm->root;
+       while ((*entry)->entries) {
                idx = addr >> (p->adev->vm_manager.block_size * level--);
-               idx %= amdgpu_bo_size(entry->bo) / 8;
-               entry = &entry->entries[idx];
+               idx %= amdgpu_bo_size((*entry)->bo) / 8;
+               *parent = *entry;
+               *entry = &(*entry)->entries[idx];
        }
 
        if (level)
-               return NULL;
+               *entry = NULL;
+}
+
+/**
+ * amdgpu_vm_handle_huge_pages - handle updating the PD with huge pages
+ *
+ * @p: see amdgpu_pte_update_params definition
+ * @entry: vm_pt entry to check
+ * @parent: parent entry
+ * @nptes: number of PTEs updated with this operation
+ * @dst: destination address where the PTEs should point to
+ * @flags: access flags fro the PTEs
+ *
+ * Check if we can update the PD with a huge page.
+ */
+static int amdgpu_vm_handle_huge_pages(struct amdgpu_pte_update_params *p,
+                                      struct amdgpu_vm_pt *entry,
+                                      struct amdgpu_vm_pt *parent,
+                                      unsigned nptes, uint64_t dst,
+                                      uint64_t flags)
+{
+       bool use_cpu_update = (p->func == amdgpu_vm_cpu_set_ptes);
+       uint64_t pd_addr, pde;
+       int r;
+
+       /* In the case of a mixed PT the PDE must point to it*/
+       if (p->adev->asic_type < CHIP_VEGA10 ||
+           nptes != AMDGPU_VM_PTE_COUNT(p->adev) ||
+           p->func == amdgpu_vm_do_copy_ptes ||
+           !(flags & AMDGPU_PTE_VALID)) {
+
+               dst = amdgpu_bo_gpu_offset(entry->bo);
+               dst = amdgpu_gart_get_vm_pde(p->adev, dst);
+               flags = AMDGPU_PTE_VALID;
+       } else {
+               flags |= AMDGPU_PDE_PTE;
+       }
+
+       if (entry->addr == dst &&
+           entry->huge_page == !!(flags & AMDGPU_PDE_PTE))
+               return 0;
+
+       entry->addr = dst;
+       entry->huge_page = !!(flags & AMDGPU_PDE_PTE);
+
+       if (use_cpu_update) {
+               r = amdgpu_bo_kmap(parent->bo, (void *)&pd_addr);
+               if (r)
+                       return r;
+
+               pde = pd_addr + (entry - parent->entries) * 8;
+               amdgpu_vm_cpu_set_ptes(p, pde, dst, 1, 0, flags);
+       } else {
+               if (parent->bo->shadow) {
+                       pd_addr = amdgpu_bo_gpu_offset(parent->bo->shadow);
+                       pde = pd_addr + (entry - parent->entries) * 8;
+                       amdgpu_vm_do_set_ptes(p, pde, dst, 1, 0, flags);
+               }
+               pd_addr = amdgpu_bo_gpu_offset(parent->bo);
+               pde = pd_addr + (entry - parent->entries) * 8;
+               amdgpu_vm_do_set_ptes(p, pde, dst, 1, 0, flags);
+       }
 
-       return entry->bo;
+       return 0;
 }
 
 /**
@@ -1307,21 +1375,31 @@ static int amdgpu_vm_update_ptes(struct amdgpu_pte_update_params *params,
        struct amdgpu_bo *pt;
        unsigned nptes;
        bool use_cpu_update = (params->func == amdgpu_vm_cpu_set_ptes);
-
+       int r;
 
        /* walk over the address space and update the page tables */
-       for (addr = start; addr < end; addr += nptes) {
-               pt = amdgpu_vm_get_pt(params, addr);
-               if (!pt) {
-                       pr_err("PT not found, aborting update_ptes\n");
-                       return -EINVAL;
-               }
+       for (addr = start; addr < end; addr += nptes,
+            dst += nptes * AMDGPU_GPU_PAGE_SIZE) {
+               struct amdgpu_vm_pt *entry, *parent;
+
+               amdgpu_vm_get_entry(params, addr, &entry, &parent);
+               if (!entry)
+                       return -ENOENT;
 
                if ((addr & ~mask) == (end & ~mask))
                        nptes = end - addr;
                else
                        nptes = AMDGPU_VM_PTE_COUNT(adev) - (addr & mask);
 
+               r = amdgpu_vm_handle_huge_pages(params, entry, parent,
+                                               nptes, dst, flags);
+               if (r)
+                       return r;
+
+               if (entry->huge_page)
+                       continue;
+
+               pt = entry->bo;
                if (use_cpu_update) {
                        pe_start = (unsigned long)pt->kptr;
                } else {
@@ -1337,8 +1415,6 @@ static int amdgpu_vm_update_ptes(struct amdgpu_pte_update_params *params,
                pe_start += (addr & mask) * 8;
                params->func(params, pe_start, dst, nptes,
                             AMDGPU_GPU_PAGE_SIZE, flags);
-
-               dst += nptes * AMDGPU_GPU_PAGE_SIZE;
        }
 
        return 0;
@@ -1490,6 +1566,9 @@ static int amdgpu_vm_bo_update_mapping(struct amdgpu_device *adev,
        /* padding, etc. */
        ndw = 64;
 
+       /* one PDE write for each huge page */
+       ndw += ((nptes >> adev->vm_manager.block_size) + 1) * 6;
+
        if (src) {
                /* only copy commands needed */
                ndw += ncmds * 7;
@@ -1569,6 +1648,7 @@ static int amdgpu_vm_bo_update_mapping(struct amdgpu_device *adev,
 
 error_free:
        amdgpu_job_free(job);
+       amdgpu_vm_invalidate_level(&vm->root);
        return r;
 }
 
index c4f5d1ff042e7680a7a775388bd55dd2552272d2..34d9174ebff22a540f25cf517465fc761a43f4a1 100644 (file)
@@ -70,6 +70,9 @@ struct amdgpu_bo_list_entry;
 /* TILED for VEGA10, reserved for older ASICs  */
 #define AMDGPU_PTE_PRT         (1ULL << 51)
 
+/* PDE is handled as PTE for VEGA10 */
+#define AMDGPU_PDE_PTE         (1ULL << 54)
+
 /* VEGA10 only */
 #define AMDGPU_PTE_MTYPE(a)    ((uint64_t)a << 57)
 #define AMDGPU_PTE_MTYPE_MASK  AMDGPU_PTE_MTYPE(3ULL)
@@ -100,6 +103,7 @@ struct amdgpu_bo_list_entry;
 struct amdgpu_vm_pt {
        struct amdgpu_bo        *bo;
        uint64_t                addr;
+       bool                    huge_page;
 
        /* array of page tables, one for each directory entry */
        struct amdgpu_vm_pt     *entries;