drm/i915: Split out i915_gem_object_set_tiling()
authorChris Wilson <chris@chris-wilson.co.uk>
Tue, 10 Jan 2017 12:10:45 +0000 (12:10 +0000)
committerChris Wilson <chris@chris-wilson.co.uk>
Tue, 10 Jan 2017 15:29:14 +0000 (15:29 +0000)
Expose an interface for changing the tiling and stride on an object,
that includes the complexity of checking for conflicting bindings and
fence registers.

Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Reviewed-by: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
Link: http://patchwork.freedesktop.org/patch/msgid/20170110121045.27144-2-chris@chris-wilson.co.uk
drivers/gpu/drm/i915/i915_gem_object.h
drivers/gpu/drm/i915/i915_gem_tiling.c

index ed3f4d5..290eaa7 100644 (file)
@@ -337,6 +337,9 @@ i915_gem_object_get_tile_row_size(struct drm_i915_gem_object *obj)
                i915_gem_object_get_tile_height(obj));
 }
 
+int i915_gem_object_set_tiling(struct drm_i915_gem_object *obj,
+                              unsigned int tiling, unsigned int stride);
+
 static inline struct intel_engine_cs *
 i915_gem_object_last_write_engine(struct drm_i915_gem_object *obj)
 {
index 59c010d..4f83e33 100644 (file)
@@ -129,61 +129,56 @@ u32 i915_gem_fence_alignment(struct drm_i915_private *i915, u32 size,
 
 /* Check pitch constriants for all chips & tiling formats */
 static bool
-i915_tiling_ok(struct drm_i915_private *dev_priv,
-              int stride, int size, int tiling_mode)
+i915_tiling_ok(struct drm_i915_gem_object *obj,
+              unsigned int tiling, unsigned int stride)
 {
-       int tile_width;
+       struct drm_i915_private *i915 = to_i915(obj->base.dev);
+       unsigned int tile_width;
 
        /* Linear is always fine */
-       if (tiling_mode == I915_TILING_NONE)
+       if (tiling == I915_TILING_NONE)
                return true;
 
-       if (tiling_mode > I915_TILING_LAST)
+       if (tiling > I915_TILING_LAST)
                return false;
 
-       if (IS_GEN2(dev_priv) ||
-           (tiling_mode == I915_TILING_Y && HAS_128_BYTE_Y_TILING(dev_priv)))
-               tile_width = 128;
-       else
-               tile_width = 512;
-
        /* check maximum stride & object size */
        /* i965+ stores the end address of the gtt mapping in the fence
         * reg, so dont bother to check the size */
-       if (INTEL_GEN(dev_priv) >= 7) {
+       if (INTEL_GEN(i915) >= 7) {
                if (stride / 128 > GEN7_FENCE_MAX_PITCH_VAL)
                        return false;
-       } else if (INTEL_GEN(dev_priv) >= 4) {
+       } else if (INTEL_GEN(i915) >= 4) {
                if (stride / 128 > I965_FENCE_MAX_PITCH_VAL)
                        return false;
        } else {
                if (stride > 8192)
                        return false;
 
-               if (IS_GEN3(dev_priv)) {
-                       if (size > I830_FENCE_MAX_SIZE_VAL << 20)
+               if (IS_GEN3(i915)) {
+                       if (obj->base.size > I830_FENCE_MAX_SIZE_VAL << 20)
                                return false;
                } else {
-                       if (size > I830_FENCE_MAX_SIZE_VAL << 19)
+                       if (obj->base.size > I830_FENCE_MAX_SIZE_VAL << 19)
                                return false;
                }
        }
 
-       if (stride < tile_width)
+       if (IS_GEN2(i915) ||
+           (tiling == I915_TILING_Y && HAS_128_BYTE_Y_TILING(i915)))
+               tile_width = 128;
+       else
+               tile_width = 512;
+
+       if (stride & (tile_width - 1))
                return false;
 
        /* 965+ just needs multiples of tile width */
-       if (INTEL_GEN(dev_priv) >= 4) {
-               if (stride & (tile_width - 1))
-                       return false;
+       if (INTEL_GEN(i915) >= 4)
                return true;
-       }
 
        /* Pre-965 needs power of two tile widths */
-       if (stride & (stride - 1))
-               return false;
-
-       return true;
+       return is_power_of_2(stride);
 }
 
 static bool i915_vma_fence_prepare(struct i915_vma *vma,
@@ -232,6 +227,98 @@ i915_gem_object_fence_prepare(struct drm_i915_gem_object *obj,
        return 0;
 }
 
+int
+i915_gem_object_set_tiling(struct drm_i915_gem_object *obj,
+                          unsigned int tiling, unsigned int stride)
+{
+       struct drm_i915_private *i915 = to_i915(obj->base.dev);
+       struct i915_vma *vma;
+       int err;
+
+       /* Make sure we don't cross-contaminate obj->tiling_and_stride */
+       BUILD_BUG_ON(I915_TILING_LAST & STRIDE_MASK);
+
+       GEM_BUG_ON(!i915_tiling_ok(obj, tiling, stride));
+       GEM_BUG_ON(!stride ^ (tiling == I915_TILING_NONE));
+       lockdep_assert_held(&i915->drm.struct_mutex);
+
+       if ((tiling | stride) == obj->tiling_and_stride)
+               return 0;
+
+       if (obj->framebuffer_references)
+               return -EBUSY;
+
+       /* We need to rebind the object if its current allocation
+        * no longer meets the alignment restrictions for its new
+        * tiling mode. Otherwise we can just leave it alone, but
+        * need to ensure that any fence register is updated before
+        * the next fenced (either through the GTT or by the BLT unit
+        * on older GPUs) access.
+        *
+        * After updating the tiling parameters, we then flag whether
+        * we need to update an associated fence register. Note this
+        * has to also include the unfenced register the GPU uses
+        * whilst executing a fenced command for an untiled object.
+        */
+
+       err = i915_gem_object_fence_prepare(obj, tiling, stride);
+       if (err)
+               return err;
+
+       /* If the memory has unknown (i.e. varying) swizzling, we pin the
+        * pages to prevent them being swapped out and causing corruption
+        * due to the change in swizzling.
+        */
+       mutex_lock(&obj->mm.lock);
+       if (obj->mm.pages &&
+           obj->mm.madv == I915_MADV_WILLNEED &&
+           i915->quirks & QUIRK_PIN_SWIZZLED_PAGES) {
+               if (tiling == I915_TILING_NONE) {
+                       GEM_BUG_ON(!obj->mm.quirked);
+                       __i915_gem_object_unpin_pages(obj);
+                       obj->mm.quirked = false;
+               }
+               if (!i915_gem_object_is_tiled(obj)) {
+                       GEM_BUG_ON(!obj->mm.quirked);
+                       __i915_gem_object_pin_pages(obj);
+                       obj->mm.quirked = true;
+               }
+       }
+       mutex_unlock(&obj->mm.lock);
+
+       list_for_each_entry(vma, &obj->vma_list, obj_link) {
+               if (!i915_vma_is_ggtt(vma))
+                       break;
+
+               vma->fence_size =
+                       i915_gem_fence_size(i915, vma->size, tiling, stride);
+               vma->fence_alignment =
+                       i915_gem_fence_alignment(i915,
+                                                vma->size, tiling, stride);
+
+               if (vma->fence)
+                       vma->fence->dirty = true;
+       }
+
+       obj->tiling_and_stride = tiling | stride;
+
+       /* Force the fence to be reacquired for GTT access */
+       i915_gem_release_mmap(obj);
+
+       /* Try to preallocate memory required to save swizzling on put-pages */
+       if (i915_gem_object_needs_bit17_swizzle(obj)) {
+               if (!obj->bit_17) {
+                       obj->bit_17 = kcalloc(BITS_TO_LONGS(obj->base.size >> PAGE_SHIFT),
+                                             sizeof(long), GFP_KERNEL);
+               }
+       } else {
+               kfree(obj->bit_17);
+               obj->bit_17 = NULL;
+       }
+
+       return 0;
+}
+
 /**
  * i915_gem_set_tiling_ioctl - IOCTL handler to set tiling mode
  * @dev: DRM device
@@ -251,26 +338,15 @@ i915_gem_set_tiling_ioctl(struct drm_device *dev, void *data,
                          struct drm_file *file)
 {
        struct drm_i915_gem_set_tiling *args = data;
-       struct drm_i915_private *dev_priv = to_i915(dev);
        struct drm_i915_gem_object *obj;
-       int err = 0;
-
-       /* Make sure we don't cross-contaminate obj->tiling_and_stride */
-       BUILD_BUG_ON(I915_TILING_LAST & STRIDE_MASK);
+       int err;
 
        obj = i915_gem_object_lookup(file, args->handle);
        if (!obj)
                return -ENOENT;
 
-       if (!i915_tiling_ok(dev_priv,
-                           args->stride, obj->base.size, args->tiling_mode)) {
-               i915_gem_object_put(obj);
-               return -EINVAL;
-       }
-
-       mutex_lock(&dev->struct_mutex);
-       if (obj->pin_display || obj->framebuffer_references) {
-               err = -EBUSY;
+       if (!i915_tiling_ok(obj, args->tiling_mode, args->stride)) {
+               err = -EINVAL;
                goto err;
        }
 
@@ -279,9 +355,9 @@ i915_gem_set_tiling_ioctl(struct drm_device *dev, void *data,
                args->stride = 0;
        } else {
                if (args->tiling_mode == I915_TILING_X)
-                       args->swizzle_mode = dev_priv->mm.bit_6_swizzle_x;
+                       args->swizzle_mode = to_i915(dev)->mm.bit_6_swizzle_x;
                else
-                       args->swizzle_mode = dev_priv->mm.bit_6_swizzle_y;
+                       args->swizzle_mode = to_i915(dev)->mm.bit_6_swizzle_y;
 
                /* Hide bit 17 swizzling from the user.  This prevents old Mesa
                 * from aborting the application on sw fallbacks to bit 17,
@@ -303,84 +379,19 @@ i915_gem_set_tiling_ioctl(struct drm_device *dev, void *data,
                }
        }
 
-       if (args->tiling_mode != i915_gem_object_get_tiling(obj) ||
-           args->stride != i915_gem_object_get_stride(obj)) {
-               /* We need to rebind the object if its current allocation
-                * no longer meets the alignment restrictions for its new
-                * tiling mode. Otherwise we can just leave it alone, but
-                * need to ensure that any fence register is updated before
-                * the next fenced (either through the GTT or by the BLT unit
-                * on older GPUs) access.
-                *
-                * After updating the tiling parameters, we then flag whether
-                * we need to update an associated fence register. Note this
-                * has to also include the unfenced register the GPU uses
-                * whilst executing a fenced command for an untiled object.
-                */
+       err = mutex_lock_interruptible(&dev->struct_mutex);
+       if (err)
+               goto err;
 
-               err = i915_gem_object_fence_prepare(obj,
-                                                   args->tiling_mode,
-                                                   args->stride);
-               if (!err) {
-                       struct i915_vma *vma;
-
-                       mutex_lock(&obj->mm.lock);
-                       if (obj->mm.pages &&
-                           obj->mm.madv == I915_MADV_WILLNEED &&
-                           dev_priv->quirks & QUIRK_PIN_SWIZZLED_PAGES) {
-                               if (args->tiling_mode == I915_TILING_NONE) {
-                                       GEM_BUG_ON(!obj->mm.quirked);
-                                       __i915_gem_object_unpin_pages(obj);
-                                       obj->mm.quirked = false;
-                               }
-                               if (!i915_gem_object_is_tiled(obj)) {
-                                       GEM_BUG_ON(!obj->mm.quirked);
-                                       __i915_gem_object_pin_pages(obj);
-                                       obj->mm.quirked = true;
-                               }
-                       }
-                       mutex_unlock(&obj->mm.lock);
-
-                       list_for_each_entry(vma, &obj->vma_list, obj_link) {
-                               if (!i915_vma_is_ggtt(vma))
-                                       break;
-
-                               vma->fence_size = i915_gem_fence_size(dev_priv, vma->size,
-                                                                     args->tiling_mode,
-                                                                     args->stride);
-                               vma->fence_alignment = i915_gem_fence_alignment(dev_priv, vma->size,
-                                                                               args->tiling_mode,
-                                                                               args->stride);
-
-                               if (vma->fence)
-                                       vma->fence->dirty = true;
-                       }
-                       obj->tiling_and_stride =
-                               args->stride | args->tiling_mode;
-
-                       /* Force the fence to be reacquired for GTT access */
-                       i915_gem_release_mmap(obj);
-               }
-       }
-       /* we have to maintain this existing ABI... */
+       err = i915_gem_object_set_tiling(obj, args->tiling_mode, args->stride);
+       mutex_unlock(&dev->struct_mutex);
+
+       /* We have to maintain this existing ABI... */
        args->stride = i915_gem_object_get_stride(obj);
        args->tiling_mode = i915_gem_object_get_tiling(obj);
 
-       /* Try to preallocate memory required to save swizzling on put-pages */
-       if (i915_gem_object_needs_bit17_swizzle(obj)) {
-               if (obj->bit_17 == NULL) {
-                       obj->bit_17 = kcalloc(BITS_TO_LONGS(obj->base.size >> PAGE_SHIFT),
-                                             sizeof(long), GFP_KERNEL);
-               }
-       } else {
-               kfree(obj->bit_17);
-               obj->bit_17 = NULL;
-       }
-
 err:
        i915_gem_object_put(obj);
-       mutex_unlock(&dev->struct_mutex);
-
        return err;
 }