drm/i915/gtt: Setup phys pages for 3lvl pdps
authorMika Kuoppala <mika.kuoppala@linux.intel.com>
Fri, 5 Jul 2019 21:52:03 +0000 (22:52 +0100)
committerChris Wilson <chris@chris-wilson.co.uk>
Sat, 6 Jul 2019 09:20:49 +0000 (10:20 +0100)
If we setup backing phys page for 3lvl pdps, as they
are not used, we will lose 5 pages per ppgtt.

Trading this memory on bsw, we gain more common code paths for all
gen8+ directory manipulation. And those paths are now void of checks
for page directory type, making the hot paths faster.

v2: don't shortcut vm (Chris)

Signed-off-by: Mika Kuoppala <mika.kuoppala@linux.intel.com>
Reviewed-by: Chris Wilson <chris@chris-wilson.co.uk>
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Link: https://patchwork.freedesktop.org/patch/msgid/20190705215204.4559-3-chris@chris-wilson.co.uk
drivers/gpu/drm/i915/i915_gem_gtt.c

index f011ce1..0a55b09 100644 (file)
@@ -758,22 +758,14 @@ static struct i915_page_directory *alloc_pd(struct i915_address_space *vm)
        return pd;
 }
 
-static inline bool pd_has_phys_page(const struct i915_page_directory * const pd)
-{
-       return pd->base.page;
-}
-
 static void free_pd(struct i915_address_space *vm,
                    struct i915_page_directory *pd)
 {
-       if (likely(pd_has_phys_page(pd)))
-               cleanup_page_dma(vm, &pd->base);
-
+       cleanup_page_dma(vm, &pd->base);
        kfree(pd);
 }
 
 #define init_pd(vm, pd, to) {                                  \
-       GEM_DEBUG_BUG_ON(!pd_has_phys_page(pd));                \
        fill_px((vm), (pd), gen8_pde_encode(px_dma(to), I915_CACHE_LLC)); \
        memset_p((pd)->entry, (to), 512);                               \
 }
@@ -1604,6 +1596,50 @@ static void ppgtt_init(struct i915_ppgtt *ppgtt, struct intel_gt *gt)
        ppgtt->vm.vma_ops.clear_pages = clear_pages;
 }
 
+static void init_pd_n(struct i915_address_space *vm,
+                     struct i915_page_directory *pd,
+                     struct i915_page_directory *to,
+                     const unsigned int entries)
+{
+       const u64 daddr = gen8_pde_encode(px_dma(to), I915_CACHE_LLC);
+       u64 * const vaddr = kmap_atomic(pd->base.page);
+
+       memset64(vaddr, daddr, entries);
+       kunmap_atomic(vaddr);
+
+       memset_p(pd->entry, to, entries);
+}
+
+static struct i915_page_directory *
+gen8_alloc_top_pd(struct i915_address_space *vm)
+{
+       struct i915_page_directory *pd;
+
+       if (i915_vm_is_4lvl(vm)) {
+               pd = alloc_pd(vm);
+               if (!IS_ERR(pd))
+                       init_pd(vm, pd, vm->scratch_pdp);
+
+               return pd;
+       }
+
+       /* 3lvl */
+       pd = __alloc_pd();
+       if (!pd)
+               return ERR_PTR(-ENOMEM);
+
+       pd->entry[GEN8_3LVL_PDPES] = NULL;
+
+       if (unlikely(setup_page_dma(vm, &pd->base))) {
+               kfree(pd);
+               return ERR_PTR(-ENOMEM);
+       }
+
+       init_pd_n(vm, pd, vm->scratch_pd, GEN8_3LVL_PDPES);
+
+       return pd;
+}
+
 /*
  * GEN8 legacy ppgtt programming is accomplished through a max 4 PDP registers
  * with a net effect resembling a 2-level page table in normal x86 terms. Each
@@ -1640,34 +1676,21 @@ static struct i915_ppgtt *gen8_ppgtt_create(struct drm_i915_private *i915)
        if (err)
                goto err_free;
 
-       ppgtt->pd = __alloc_pd();
-       if (!ppgtt->pd) {
-               err = -ENOMEM;
+       ppgtt->pd = gen8_alloc_top_pd(&ppgtt->vm);
+       if (IS_ERR(ppgtt->pd)) {
+               err = PTR_ERR(ppgtt->pd);
                goto err_free_scratch;
        }
 
        if (i915_vm_is_4lvl(&ppgtt->vm)) {
-               err = setup_page_dma(&ppgtt->vm, &ppgtt->pd->base);
-               if (err)
-                       goto err_free_pdp;
-
-               init_pd(&ppgtt->vm, ppgtt->pd, ppgtt->vm.scratch_pdp);
-
                ppgtt->vm.allocate_va_range = gen8_ppgtt_alloc_4lvl;
                ppgtt->vm.insert_entries = gen8_ppgtt_insert_4lvl;
                ppgtt->vm.clear_range = gen8_ppgtt_clear_4lvl;
        } else {
-               /*
-                * We don't need to setup dma for top level pdp, only
-                * for entries. So point entries to scratch.
-                */
-               memset_p(ppgtt->pd->entry, ppgtt->vm.scratch_pd,
-                        GEN8_3LVL_PDPES);
-
                if (intel_vgpu_active(i915)) {
                        err = gen8_preallocate_top_level_pdp(ppgtt);
                        if (err)
-                               goto err_free_pdp;
+                               goto err_free_pd;
                }
 
                ppgtt->vm.allocate_va_range = gen8_ppgtt_alloc_3lvl;
@@ -1682,7 +1705,7 @@ static struct i915_ppgtt *gen8_ppgtt_create(struct drm_i915_private *i915)
 
        return ppgtt;
 
-err_free_pdp:
+err_free_pd:
        free_pd(&ppgtt->vm, ppgtt->pd);
 err_free_scratch:
        gen8_free_scratch(&ppgtt->vm);