drm/i915: Move fence tracking from object to vma

In order to handle tiled partial GTT mmappings, we need to associate the
fence with an individual vma.

v2: A couple of silly drops replaced spotted by Joonas

Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Reviewed-by: Joonas Lahtinen <joonas.lahtinen@linux.intel.com>
Link: http://patchwork.freedesktop.org/patch/msgid/20160818161718.27187-21-chris@chris-wilson.co.uk

drivers/gpu/drm/i915/i915_debugfs.c

@@ -152,11 +152,9 @@ describe_obj(struct seq_file *m, struct drm_i915_gem_object *obj)
 		seq_printf(m, "%x ",
 			   i915_gem_active_get_seqno(&obj->last_read[id],
 						     &obj->base.dev->struct_mutex));
-	seq_printf(m, "] %x %x%s%s%s",
+	seq_printf(m, "] %x %s%s%s",
 		   i915_gem_active_get_seqno(&obj->last_write,
 					     &obj->base.dev->struct_mutex),
-		   i915_gem_active_get_seqno(&obj->last_fence,
-					     &obj->base.dev->struct_mutex),
 		   i915_cache_level_str(to_i915(obj->base.dev), obj->cache_level),
 		   obj->dirty ? " dirty" : "",
 		   obj->madv == I915_MADV_DONTNEED ? " purgeable" : "");
@@ -169,8 +167,6 @@ describe_obj(struct seq_file *m, struct drm_i915_gem_object *obj)
 	seq_printf(m, " (pinned x %d)", pin_count);
 	if (obj->pin_display)
 		seq_printf(m, " (display)");
-	if (obj->fence_reg != I915_FENCE_REG_NONE)
-		seq_printf(m, " (fence: %d)", obj->fence_reg);
 	list_for_each_entry(vma, &obj->vma_list, obj_link) {
 		if (!drm_mm_node_allocated(&vma->node))
 			continue;
@@ -180,6 +176,10 @@ describe_obj(struct seq_file *m, struct drm_i915_gem_object *obj)
 			   vma->node.start, vma->node.size);
 		if (i915_vma_is_ggtt(vma))
 			seq_printf(m, ", type: %u", vma->ggtt_view.type);
+		if (vma->fence)
+			seq_printf(m, " , fence: %d%s",
+				   vma->fence->id,
+				   i915_gem_active_isset(&vma->last_fence) ? "*" : "");
 		seq_puts(m, ")");
 	}
 	if (obj->stolen)
@@ -938,14 +938,14 @@ static int i915_gem_fence_regs_info(struct seq_file *m, void *data)
 
 	seq_printf(m, "Total fences = %d\n", dev_priv->num_fence_regs);
 	for (i = 0; i < dev_priv->num_fence_regs; i++) {
-		struct drm_i915_gem_object *obj = dev_priv->fence_regs[i].obj;
+		struct i915_vma *vma = dev_priv->fence_regs[i].vma;
 
 		seq_printf(m, "Fence %d, pin count = %d, object = ",
 			   i, dev_priv->fence_regs[i].pin_count);
-		if (obj == NULL)
+		if (!vma)
 			seq_puts(m, "unused");
 		else
-			describe_obj(m, obj);
+			describe_obj(m, vma->obj);
 		seq_putc(m, '\n');
 	}
 

drivers/gpu/drm/i915/i915_drv.h

@@ -455,15 +455,21 @@ struct intel_opregion {
 struct intel_overlay;
 struct intel_overlay_error_state;
 
-#define I915_FENCE_REG_NONE -1
-#define I915_MAX_NUM_FENCES 32
-/* 32 fences + sign bit for FENCE_REG_NONE */
-#define I915_MAX_NUM_FENCE_BITS 6
-
 struct drm_i915_fence_reg {
 	struct list_head link;
-	struct drm_i915_gem_object *obj;
+	struct drm_i915_private *i915;
+	struct i915_vma *vma;
 	int pin_count;
+	int id;
+	/**
+	 * Whether the tiling parameters for the currently
+	 * associated fence register have changed. Note that
+	 * for the purposes of tracking tiling changes we also
+	 * treat the unfenced register, the register slot that
+	 * the object occupies whilst it executes a fenced
+	 * command (such as BLT on gen2/3), as a "fence".
+	 */
+	bool dirty;
 };
 
 struct sdvo_device_mapping {
@@ -2171,27 +2177,11 @@ struct drm_i915_gem_object {
 	 */
 	unsigned int dirty:1;
 
-	/**
-	 * Fence register bits (if any) for this object.  Will be set
-	 * as needed when mapped into the GTT.
-	 * Protected by dev->struct_mutex.
-	 */
-	signed int fence_reg:I915_MAX_NUM_FENCE_BITS;
-
 	/**
 	 * Advice: are the backing pages purgeable?
 	 */
 	unsigned int madv:2;
 
-	/**
-	 * Whether the tiling parameters for the currently associated fence
-	 * register have changed. Note that for the purposes of tracking
-	 * tiling changes we also treat the unfenced register, the register
-	 * slot that the object occupies whilst it executes a fenced
-	 * command (such as BLT on gen2/3), as a "fence".
-	 */
-	unsigned int fence_dirty:1;
-
 	/**
 	 * Whether the current gtt mapping needs to be mappable (and isn't just
 	 * mappable by accident). Track pin and fault separate for a more
@@ -2240,7 +2230,6 @@ struct drm_i915_gem_object {
 	 */
 	struct i915_gem_active last_read[I915_NUM_ENGINES];
 	struct i915_gem_active last_write;
-	struct i915_gem_active last_fence;
 
 	/** References from framebuffers, locks out tiling changes. */
 	unsigned long framebuffer_references;
@@ -3343,11 +3332,50 @@ i915_gem_object_ggtt_offset(struct drm_i915_gem_object *o,
 }
 
 /* i915_gem_fence.c */
-int __must_check i915_gem_object_get_fence(struct drm_i915_gem_object *obj);
-int __must_check i915_gem_object_put_fence(struct drm_i915_gem_object *obj);
+int __must_check i915_vma_get_fence(struct i915_vma *vma);
+int __must_check i915_vma_put_fence(struct i915_vma *vma);
+
+/**
+ * i915_vma_pin_fence - pin fencing state
+ * @vma: vma to pin fencing for
+ *
+ * This pins the fencing state (whether tiled or untiled) to make sure the
+ * vma (and its object) is ready to be used as a scanout target. Fencing
+ * status must be synchronize first by calling i915_vma_get_fence():
+ *
+ * The resulting fence pin reference must be released again with
+ * i915_vma_unpin_fence().
+ *
+ * Returns:
+ *
+ * True if the vma has a fence, false otherwise.
+ */
+static inline bool
+i915_vma_pin_fence(struct i915_vma *vma)
+{
+	if (vma->fence) {
+		vma->fence->pin_count++;
+		return true;
+	} else
+		return false;
+}
 
-bool i915_gem_object_pin_fence(struct drm_i915_gem_object *obj);
-void i915_gem_object_unpin_fence(struct drm_i915_gem_object *obj);
+/**
+ * i915_vma_unpin_fence - unpin fencing state
+ * @vma: vma to unpin fencing for
+ *
+ * This releases the fence pin reference acquired through
+ * i915_vma_pin_fence. It will handle both objects with and without an
+ * attached fence correctly, callers do not need to distinguish this.
+ */
+static inline void
+i915_vma_unpin_fence(struct i915_vma *vma)
+{
+	if (vma->fence) {
+		GEM_BUG_ON(vma->fence->pin_count <= 0);
+		vma->fence->pin_count--;
+	}
+}
 
 void i915_gem_restore_fences(struct drm_device *dev);
 

drivers/gpu/drm/i915/i915_gem.c

@@ -829,7 +829,7 @@ i915_gem_gtt_pread(struct drm_device *dev,
 	if (!IS_ERR(vma)) {
 		node.start = i915_ggtt_offset(vma);
 		node.allocated = false;
-		ret = i915_gem_object_put_fence(obj);
+		ret = i915_vma_put_fence(vma);
 		if (ret) {
 			i915_vma_unpin(vma);
 			vma = ERR_PTR(ret);
@@ -1131,7 +1131,7 @@ i915_gem_gtt_pwrite_fast(struct drm_i915_private *i915,
 	if (!IS_ERR(vma)) {
 		node.start = i915_ggtt_offset(vma);
 		node.allocated = false;
-		ret = i915_gem_object_put_fence(obj);
+		ret = i915_vma_put_fence(vma);
 		if (ret) {
 			i915_vma_unpin(vma);
 			vma = ERR_PTR(ret);
@@ -1751,7 +1751,7 @@ int i915_gem_fault(struct vm_area_struct *area, struct vm_fault *vmf)
 	if (ret)
 		goto err_unpin;
 
-	ret = i915_gem_object_get_fence(obj);
+	ret = i915_vma_get_fence(vma);
 	if (ret)
 		goto err_unpin;
 
@@ -2903,7 +2903,7 @@ int i915_vma_unbind(struct i915_vma *vma)
 		i915_gem_object_finish_gtt(obj);
 
 		/* release the fence reg _after_ flushing */
-		ret = i915_gem_object_put_fence(obj);
+		ret = i915_vma_put_fence(vma);
 		if (ret)
 			return ret;
 
@@ -3385,9 +3385,11 @@ restart:
 			 * dropped the fence as all snoopable access is
 			 * supposed to be linear.
 			 */
-			ret = i915_gem_object_put_fence(obj);
-			if (ret)
-				return ret;
+			list_for_each_entry(vma, &obj->vma_list, obj_link) {
+				ret = i915_vma_put_fence(vma);
+				if (ret)
+					return ret;
+			}
 		} else {
 			/* We either have incoherent backing store and
 			 * so no GTT access or the architecture is fully
@@ -4065,14 +4067,12 @@ void i915_gem_object_init(struct drm_i915_gem_object *obj,
 				    i915_gem_object_retire__read);
 	init_request_active(&obj->last_write,
 			    i915_gem_object_retire__write);
-	init_request_active(&obj->last_fence, NULL);
 	INIT_LIST_HEAD(&obj->obj_exec_link);
 	INIT_LIST_HEAD(&obj->vma_list);
 	INIT_LIST_HEAD(&obj->batch_pool_link);
 
 	obj->ops = ops;
 
-	obj->fence_reg = I915_FENCE_REG_NONE;
 	obj->madv = I915_MADV_WILLNEED;
 
 	i915_gem_info_add_obj(to_i915(obj->base.dev), obj->base.size);
@@ -4502,6 +4502,7 @@ void
 i915_gem_load_init_fences(struct drm_i915_private *dev_priv)
 {
 	struct drm_device *dev = &dev_priv->drm;
+	int i;
 
 	if (INTEL_INFO(dev_priv)->gen >= 7 && !IS_VALLEYVIEW(dev_priv) &&
 	    !IS_CHERRYVIEW(dev_priv))
@@ -4517,6 +4518,13 @@ i915_gem_load_init_fences(struct drm_i915_private *dev_priv)
 				I915_READ(vgtif_reg(avail_rs.fence_num));
 
 	/* Initialize fence registers to zero */
+	for (i = 0; i < dev_priv->num_fence_regs; i++) {
+		struct drm_i915_fence_reg *fence = &dev_priv->fence_regs[i];
+
+		fence->i915 = dev_priv;
+		fence->id = i;
+		list_add_tail(&fence->link, &dev_priv->mm.fence_list);
+	}
 	i915_gem_restore_fences(dev);
 
 	i915_gem_detect_bit_6_swizzle(dev);
@@ -4552,8 +4560,6 @@ i915_gem_load_init(struct drm_device *dev)
 	INIT_LIST_HEAD(&dev_priv->mm.fence_list);
 	for (i = 0; i < I915_NUM_ENGINES; i++)
 		init_engine_lists(&dev_priv->engine[i]);
-	for (i = 0; i < I915_MAX_NUM_FENCES; i++)
-		INIT_LIST_HEAD(&dev_priv->fence_regs[i].link);
 	INIT_DELAYED_WORK(&dev_priv->gt.retire_work,
 			  i915_gem_retire_work_handler);
 	INIT_DELAYED_WORK(&dev_priv->gt.idle_work,
@@ -4563,8 +4569,6 @@ i915_gem_load_init(struct drm_device *dev)
 
 	dev_priv->relative_constants_mode = I915_EXEC_CONSTANTS_REL_GENERAL;
 
-	INIT_LIST_HEAD(&dev_priv->mm.fence_list);
-
 	init_waitqueue_head(&dev_priv->pending_flip_queue);
 
 	dev_priv->mm.interruptible = true;

drivers/gpu/drm/i915/i915_gem_execbuffer.c

@@ -250,7 +250,6 @@ static void
 i915_gem_execbuffer_unreserve_vma(struct i915_vma *vma)
 {
 	struct drm_i915_gem_exec_object2 *entry;
-	struct drm_i915_gem_object *obj = vma->obj;
 
 	if (!drm_mm_node_allocated(&vma->node))
 		return;
@@ -258,7 +257,7 @@ i915_gem_execbuffer_unreserve_vma(struct i915_vma *vma)
 	entry = vma->exec_entry;
 
 	if (entry->flags & __EXEC_OBJECT_HAS_FENCE)
-		i915_gem_object_unpin_fence(obj);
+		i915_vma_unpin_fence(vma);
 
 	if (entry->flags & __EXEC_OBJECT_HAS_PIN)
 		__i915_vma_unpin(vma);
@@ -455,7 +454,7 @@ static void *reloc_iomap(struct drm_i915_gem_object *obj,
 			if (ret)
 				return ERR_PTR(ret);
 		} else {
-			ret = i915_gem_object_put_fence(obj);
+			ret = i915_vma_put_fence(vma);
 			if (ret) {
 				i915_vma_unpin(vma);
 				return ERR_PTR(ret);
@@ -811,11 +810,11 @@ i915_gem_execbuffer_reserve_vma(struct i915_vma *vma,
 	entry->flags |= __EXEC_OBJECT_HAS_PIN;
 
 	if (entry->flags & EXEC_OBJECT_NEEDS_FENCE) {
-		ret = i915_gem_object_get_fence(obj);
+		ret = i915_vma_get_fence(vma);
 		if (ret)
 			return ret;
 
-		if (i915_gem_object_pin_fence(obj))
+		if (i915_vma_pin_fence(vma))
 			entry->flags |= __EXEC_OBJECT_HAS_FENCE;
 	}
 
@@ -1305,15 +1304,8 @@ void i915_vma_move_to_active(struct i915_vma *vma,
 		obj->base.write_domain &= ~I915_GEM_GPU_DOMAINS;
 	}
 
-	if (flags & EXEC_OBJECT_NEEDS_FENCE) {
-		i915_gem_active_set(&obj->last_fence, req);
-		if (flags & __EXEC_OBJECT_HAS_FENCE) {
-			struct drm_i915_private *dev_priv = req->i915;
-
-			list_move_tail(&dev_priv->fence_regs[obj->fence_reg].link,
-				       &dev_priv->mm.fence_list);
-		}
-	}
+	if (flags & EXEC_OBJECT_NEEDS_FENCE)
+		i915_gem_active_set(&vma->last_fence, req);
 
 	i915_vma_set_active(vma, idx);
 	i915_gem_active_set(&vma->last_read[idx], req);

drivers/gpu/drm/i915/i915_gem_fence.c

@@ -55,74 +55,73 @@
  * CPU ptes into GTT mmaps (not the GTT ptes themselves) as needed.
  */
 
-static void i965_write_fence_reg(struct drm_device *dev, int reg,
-				 struct drm_i915_gem_object *obj)
+#define pipelined 0
+
+static void i965_write_fence_reg(struct drm_i915_fence_reg *fence,
+				 struct i915_vma *vma)
 {
-	struct drm_i915_private *dev_priv = to_i915(dev);
 	i915_reg_t fence_reg_lo, fence_reg_hi;
 	int fence_pitch_shift;
+	u64 val;
 
-	if (INTEL_INFO(dev)->gen >= 6) {
-		fence_reg_lo = FENCE_REG_GEN6_LO(reg);
-		fence_reg_hi = FENCE_REG_GEN6_HI(reg);
+	if (INTEL_INFO(fence->i915)->gen >= 6) {
+		fence_reg_lo = FENCE_REG_GEN6_LO(fence->id);
+		fence_reg_hi = FENCE_REG_GEN6_HI(fence->id);
 		fence_pitch_shift = GEN6_FENCE_PITCH_SHIFT;
+
 	} else {
-		fence_reg_lo = FENCE_REG_965_LO(reg);
-		fence_reg_hi = FENCE_REG_965_HI(reg);
+		fence_reg_lo = FENCE_REG_965_LO(fence->id);
+		fence_reg_hi = FENCE_REG_965_HI(fence->id);
 		fence_pitch_shift = I965_FENCE_PITCH_SHIFT;
 	}
 
-	/* To w/a incoherency with non-atomic 64-bit register updates,
-	 * we split the 64-bit update into two 32-bit writes. In order
-	 * for a partial fence not to be evaluated between writes, we
-	 * precede the update with write to turn off the fence register,
-	 * and only enable the fence as the last step.
-	 *
-	 * For extra levels of paranoia, we make sure each step lands
-	 * before applying the next step.
-	 */
-	I915_WRITE(fence_reg_lo, 0);
-	POSTING_READ(fence_reg_lo);
-
-	if (obj) {
-		struct i915_vma *vma = i915_gem_object_to_ggtt(obj, NULL);
-		unsigned int tiling = i915_gem_object_get_tiling(obj);
-		unsigned int stride = i915_gem_object_get_stride(obj);
-		u32 size = vma->node.size;
-		u32 row_size = stride * (tiling == I915_TILING_Y ? 32 : 8);
-		u64 val;
-
-		/* Adjust fence size to match tiled area */
-		size = rounddown(size, row_size);
+	val = 0;
+	if (vma) {
+		unsigned int tiling = i915_gem_object_get_tiling(vma->obj);
+		bool is_y_tiled = tiling == I915_TILING_Y;
+		unsigned int stride = i915_gem_object_get_stride(vma->obj);
+		u32 row_size = stride * (is_y_tiled ? 32 : 8);
+		u32 size = rounddown((u32)vma->node.size, row_size);
 
 		val = ((vma->node.start + size - 4096) & 0xfffff000) << 32;
 		val |= vma->node.start & 0xfffff000;
 		val |= (u64)((stride / 128) - 1) << fence_pitch_shift;
-		if (tiling == I915_TILING_Y)
-			val |= 1 << I965_FENCE_TILING_Y_SHIFT;
+		if (is_y_tiled)
+			val |= BIT(I965_FENCE_TILING_Y_SHIFT);
 		val |= I965_FENCE_REG_VALID;
+	}
 
-		I915_WRITE(fence_reg_hi, val >> 32);
-		POSTING_READ(fence_reg_hi);
+	if (!pipelined) {
+		struct drm_i915_private *dev_priv = fence->i915;
 
-		I915_WRITE(fence_reg_lo, val);
+		/* To w/a incoherency with non-atomic 64-bit register updates,
+		 * we split the 64-bit update into two 32-bit writes. In order
+		 * for a partial fence not to be evaluated between writes, we
+		 * precede the update with write to turn off the fence register,
+		 * and only enable the fence as the last step.
+		 *
+		 * For extra levels of paranoia, we make sure each step lands
+		 * before applying the next step.
+		 */
+		I915_WRITE(fence_reg_lo, 0);
+		POSTING_READ(fence_reg_lo);
+
+		I915_WRITE(fence_reg_hi, upper_32_bits(val));
+		I915_WRITE(fence_reg_lo, lower_32_bits(val));
 		POSTING_READ(fence_reg_lo);
-	} else {
-		I915_WRITE(fence_reg_hi, 0);
-		POSTING_READ(fence_reg_hi);
 	}
 }
 
-static void i915_write_fence_reg(struct drm_device *dev, int reg,
-				 struct drm_i915_gem_object *obj)
+static void i915_write_fence_reg(struct drm_i915_fence_reg *fence,
+				 struct i915_vma *vma)
 {
-	struct drm_i915_private *dev_priv = to_i915(dev);
 	u32 val;
 
-	if (obj) {
-		struct i915_vma *vma = i915_gem_object_to_ggtt(obj, NULL);
-		unsigned int tiling = i915_gem_object_get_tiling(obj);
-		unsigned int stride = i915_gem_object_get_stride(obj);
+	val = 0;
+	if (vma) {
+		unsigned int tiling = i915_gem_object_get_tiling(vma->obj);
+		bool is_y_tiled = tiling == I915_TILING_Y;
+		unsigned int stride = i915_gem_object_get_stride(vma->obj);
 		int pitch_val;
 		int tile_width;
 
@@ -134,7 +133,7 @@ static void i915_write_fence_reg(struct drm_device *dev, int reg,
 		     i915_vma_is_map_and_fenceable(vma),
 		     vma->node.size);
 
-		if (tiling == I915_TILING_Y && HAS_128_BYTE_Y_TILING(dev))
+		if (is_y_tiled && HAS_128_BYTE_Y_TILING(fence->i915))
 			tile_width = 128;
 		else
 			tile_width = 512;
@@ -144,28 +143,32 @@ static void i915_write_fence_reg(struct drm_device *dev, int reg,
 		pitch_val = ffs(pitch_val) - 1;
 
 		val = vma->node.start;
-		if (tiling == I915_TILING_Y)
-			val |= 1 << I830_FENCE_TILING_Y_SHIFT;
+		if (is_y_tiled)
+			val |= BIT(I830_FENCE_TILING_Y_SHIFT);
 		val |= I915_FENCE_SIZE_BITS(vma->node.size);
 		val |= pitch_val << I830_FENCE_PITCH_SHIFT;
 		val |= I830_FENCE_REG_VALID;
-	} else
-		val = 0;
+	}
 
-	I915_WRITE(FENCE_REG(reg), val);
-	POSTING_READ(FENCE_REG(reg));
+	if (!pipelined) {
+		struct drm_i915_private *dev_priv = fence->i915;
+		i915_reg_t reg = FENCE_REG(fence->id);
+
+		I915_WRITE(reg, val);
+		POSTING_READ(reg);
+	}
 }
 
-static void i830_write_fence_reg(struct drm_device *dev, int reg,
-				struct drm_i915_gem_object *obj)
+static void i830_write_fence_reg(struct drm_i915_fence_reg *fence,
+				 struct i915_vma *vma)
 {
-	struct drm_i915_private *dev_priv = to_i915(dev);
 	u32 val;
 
-	if (obj) {
-		struct i915_vma *vma = i915_gem_object_to_ggtt(obj, NULL);
-		unsigned int tiling = i915_gem_object_get_tiling(obj);
-		unsigned int stride = i915_gem_object_get_stride(obj);
+	val = 0;
+	if (vma) {
+		unsigned int tiling = i915_gem_object_get_tiling(vma->obj);
+		bool is_y_tiled = tiling == I915_TILING_Y;
+		unsigned int stride = i915_gem_object_get_stride(vma->obj);
 		u32 pitch_val;
 
 		WARN((vma->node.start & ~I830_FENCE_START_MASK) ||
@@ -178,104 +181,102 @@ static void i830_write_fence_reg(struct drm_device *dev, int reg,
 		pitch_val = ffs(pitch_val) - 1;
 
 		val = vma->node.start;
-		if (tiling == I915_TILING_Y)
-			val |= 1 << I830_FENCE_TILING_Y_SHIFT;
+		if (is_y_tiled)
+			val |= BIT(I830_FENCE_TILING_Y_SHIFT);
 		val |= I830_FENCE_SIZE_BITS(vma->node.size);
 		val |= pitch_val << I830_FENCE_PITCH_SHIFT;
 		val |= I830_FENCE_REG_VALID;
-	} else
-		val = 0;
+	}
 
-	I915_WRITE(FENCE_REG(reg), val);
-	POSTING_READ(FENCE_REG(reg));
-}
+	if (!pipelined) {
+		struct drm_i915_private *dev_priv = fence->i915;
+		i915_reg_t reg = FENCE_REG(fence->id);
 
-inline static bool i915_gem_object_needs_mb(struct drm_i915_gem_object *obj)
-{
-	return obj && obj->base.read_domains & I915_GEM_DOMAIN_GTT;
+		I915_WRITE(reg, val);
+		POSTING_READ(reg);
+	}
 }
 
-static void i915_gem_write_fence(struct drm_device *dev, int reg,
-				 struct drm_i915_gem_object *obj)
+static void fence_write(struct drm_i915_fence_reg *fence,
+			struct i915_vma *vma)
 {
-	struct drm_i915_private *dev_priv = to_i915(dev);
-
-	/* Ensure that all CPU reads are completed before installing a fence
-	 * and all writes before removing the fence.
+	/* Previous access through the fence register is marshalled by
+	 * the mb() inside the fault handlers (i915_gem_release_mmaps)
+	 * and explicitly managed for internal users.
 	 */
-	if (i915_gem_object_needs_mb(dev_priv->fence_regs[reg].obj))
-		mb();
-
-	WARN(obj &&
-	     (!i915_gem_object_get_stride(obj) ||
-	      !i915_gem_object_get_tiling(obj)),
-	     "bogus fence setup with stride: 0x%x, tiling mode: %i\n",
-	     i915_gem_object_get_stride(obj),
-	     i915_gem_object_get_tiling(obj));
-
-	if (IS_GEN2(dev))
-		i830_write_fence_reg(dev, reg, obj);
-	else if (IS_GEN3(dev))
-		i915_write_fence_reg(dev, reg, obj);
-	else if (INTEL_INFO(dev)->gen >= 4)
-		i965_write_fence_reg(dev, reg, obj);
-
-	/* And similarly be paranoid that no direct access to this region
-	 * is reordered to before the fence is installed.
+
+	if (IS_GEN2(fence->i915))
+		i830_write_fence_reg(fence, vma);
+	else if (IS_GEN3(fence->i915))
+		i915_write_fence_reg(fence, vma);
+	else
+		i965_write_fence_reg(fence, vma);
+
+	/* Access through the fenced region afterwards is
+	 * ordered by the posting reads whilst writing the registers.
 	 */
-	if (i915_gem_object_needs_mb(obj))
-		mb();
-}
 
-static inline int fence_number(struct drm_i915_private *dev_priv,
-			       struct drm_i915_fence_reg *fence)
-{
-	return fence - dev_priv->fence_regs;
+	fence->dirty = false;
 }
 
-static void i915_gem_object_update_fence(struct drm_i915_gem_object *obj,
-					 struct drm_i915_fence_reg *fence,
-					 bool enable)
+static int fence_update(struct drm_i915_fence_reg *fence,
+			struct i915_vma *vma)
 {
-	struct drm_i915_private *dev_priv = to_i915(obj->base.dev);
-	int reg = fence_number(dev_priv, fence);
+	int ret;
 
-	i915_gem_write_fence(obj->base.dev, reg, enable ? obj : NULL);
+	if (vma) {
+		if (!i915_vma_is_map_and_fenceable(vma))
+			return -EINVAL;
 
-	if (enable) {
-		obj->fence_reg = reg;
-		fence->obj = obj;
-		list_move_tail(&fence->link, &dev_priv->mm.fence_list);
-	} else {
-		obj->fence_reg = I915_FENCE_REG_NONE;
-		fence->obj = NULL;
-		list_del_init(&fence->link);
+		if (WARN(!i915_gem_object_get_stride(vma->obj) ||
+			 !i915_gem_object_get_tiling(vma->obj),
+			 "bogus fence setup with stride: 0x%x, tiling mode: %i\n",
+			 i915_gem_object_get_stride(vma->obj),
+			 i915_gem_object_get_tiling(vma->obj)))
+			return -EINVAL;
+
+		ret = i915_gem_active_retire(&vma->last_fence,
+					     &vma->obj->base.dev->struct_mutex);
+		if (ret)
+			return ret;
 	}
-	obj->fence_dirty = false;
-}
 
-static inline void i915_gem_object_fence_lost(struct drm_i915_gem_object *obj)
-{
-	if (i915_gem_object_is_tiled(obj))
-		i915_gem_release_mmap(obj);
+	if (fence->vma) {
+		ret = i915_gem_active_retire(&fence->vma->last_fence,
+				      &fence->vma->obj->base.dev->struct_mutex);
+		if (ret)
+			return ret;
+	}
 
-	/* As we do not have an associated fence register, we will force
-	 * a tiling change if we ever need to acquire one.
-	 */
-	obj->fence_dirty = false;
-	obj->fence_reg = I915_FENCE_REG_NONE;
-}
+	if (fence->vma && fence->vma != vma) {
+		/* Ensure that all userspace CPU access is completed before
+		 * stealing the fence.
+		 */
+		i915_gem_release_mmap(fence->vma->obj);
 
-static int
-i915_gem_object_wait_fence(struct drm_i915_gem_object *obj)
-{
-	return i915_gem_active_retire(&obj->last_fence,
-				      &obj->base.dev->struct_mutex);
+		fence->vma->fence = NULL;
+		fence->vma = NULL;
+
+		list_move(&fence->link, &fence->i915->mm.fence_list);
+	}
+
+	fence_write(fence, vma);
+
+	if (vma) {
+		if (fence->vma != vma) {
+			vma->fence = fence;
+			fence->vma = vma;
+		}
+
+		list_move_tail(&fence->link, &fence->i915->mm.fence_list);
+	}
+
+	return 0;
 }
 
 /**
- * i915_gem_object_put_fence - force-remove fence for an object
- * @obj: object to map through a fence reg
+ * i915_vma_put_fence - force-remove fence for a VMA
+ * @vma: vma to map linearly (not through a fence reg)
  *
  * This function force-removes any fence from the given object, which is useful
  * if the kernel wants to do untiled GTT access.
@@ -285,70 +286,40 @@ i915_gem_object_wait_fence(struct drm_i915_gem_object *obj)
  * 0 on success, negative error code on failure.
  */
 int
-i915_gem_object_put_fence(struct drm_i915_gem_object *obj)
+i915_vma_put_fence(struct i915_vma *vma)
 {
-	struct drm_i915_private *dev_priv = to_i915(obj->base.dev);
-	struct drm_i915_fence_reg *fence;
-	int ret;
-
-	ret = i915_gem_object_wait_fence(obj);
-	if (ret)
-		return ret;
+	struct drm_i915_fence_reg *fence = vma->fence;
 
-	if (obj->fence_reg == I915_FENCE_REG_NONE)
+	if (!fence)
 		return 0;
 
-	fence = &dev_priv->fence_regs[obj->fence_reg];
-
 	if (fence->pin_count)
 		return -EBUSY;
 
-	i915_gem_object_fence_lost(obj);
-	i915_gem_object_update_fence(obj, fence, false);
-
-	return 0;
+	return fence_update(fence, NULL);
 }
 
-static struct drm_i915_fence_reg *
-i915_find_fence_reg(struct drm_device *dev)
+static struct drm_i915_fence_reg *fence_find(struct drm_i915_private *dev_priv)
 {
-	struct drm_i915_private *dev_priv = to_i915(dev);
-	struct drm_i915_fence_reg *reg, *avail;
-	int i;
-
-	/* First try to find a free reg */
-	avail = NULL;
-	for (i = 0; i < dev_priv->num_fence_regs; i++) {
-		reg = &dev_priv->fence_regs[i];
-		if (!reg->obj)
-			return reg;
-
-		if (!reg->pin_count)
-			avail = reg;
-	}
-
-	if (avail == NULL)
-		goto deadlock;
+	struct drm_i915_fence_reg *fence;
 
-	/* None available, try to steal one or wait for a user to finish */
-	list_for_each_entry(reg, &dev_priv->mm.fence_list, link) {
-		if (reg->pin_count)
+	list_for_each_entry(fence, &dev_priv->mm.fence_list, link) {
+		if (fence->pin_count)
 			continue;
 
-		return reg;
+		return fence;
 	}
 
-deadlock:
 	/* Wait for completion of pending flips which consume fences */
-	if (intel_has_pending_fb_unpin(dev))
+	if (intel_has_pending_fb_unpin(&dev_priv->drm))
 		return ERR_PTR(-EAGAIN);
 
 	return ERR_PTR(-EDEADLK);
 }
 
 /**
- * i915_gem_object_get_fence - set up fencing for an object
- * @obj: object to map through a fence reg
+ * i915_vma_get_fence - set up fencing for a vma
+ * @vma: vma to map through a fence reg
  *
  * When mapping objects through the GTT, userspace wants to be able to write
  * to them without having to worry about swizzling if the object is tiled.
@@ -365,93 +336,27 @@ deadlock:
  * 0 on success, negative error code on failure.
  */
 int
-i915_gem_object_get_fence(struct drm_i915_gem_object *obj)
+i915_vma_get_fence(struct i915_vma *vma)
 {
-	struct drm_device *dev = obj->base.dev;
-	struct drm_i915_private *dev_priv = to_i915(dev);
-	bool enable = i915_gem_object_is_tiled(obj);
-	struct drm_i915_fence_reg *reg;
-	int ret;
-
-	/* Have we updated the tiling parameters upon the object and so
-	 * will need to serialise the write to the associated fence register?
-	 */
-	if (obj->fence_dirty) {
-		ret = i915_gem_object_wait_fence(obj);
-		if (ret)
-			return ret;
-	}
+	struct drm_i915_fence_reg *fence;
+	struct i915_vma *set = i915_gem_object_is_tiled(vma->obj) ? vma : NULL;
 
 	/* Just update our place in the LRU if our fence is getting reused. */
-	if (obj->fence_reg != I915_FENCE_REG_NONE) {
-		reg = &dev_priv->fence_regs[obj->fence_reg];
-		if (!obj->fence_dirty) {
-			list_move_tail(&reg->link, &dev_priv->mm.fence_list);
+	if (vma->fence) {
+		fence = vma->fence;
+		if (!fence->dirty) {
+			list_move_tail(&fence->link,
+				       &fence->i915->mm.fence_list);
 			return 0;
 		}
-	} else if (enable) {
-		reg = i915_find_fence_reg(dev);
-		if (IS_ERR(reg))
-			return PTR_ERR(reg);
-
-		if (reg->obj) {
-			struct drm_i915_gem_object *old = reg->obj;
-
-			ret = i915_gem_object_wait_fence(old);
-			if (ret)
-				return ret;
-
-			i915_gem_object_fence_lost(old);
-		}
+	} else if (set) {
+		fence = fence_find(to_i915(vma->vm->dev));
+		if (IS_ERR(fence))
+			return PTR_ERR(fence);
 	} else
 		return 0;
 
-	i915_gem_object_update_fence(obj, reg, enable);
-
-	return 0;
-}
-
-/**
- * i915_gem_object_pin_fence - pin fencing state
- * @obj: object to pin fencing for
- *
- * This pins the fencing state (whether tiled or untiled) to make sure the
- * object is ready to be used as a scanout target. Fencing status must be
- * synchronize first by calling i915_gem_object_get_fence():
- *
- * The resulting fence pin reference must be released again with
- * i915_gem_object_unpin_fence().
- *
- * Returns:
- *
- * True if the object has a fence, false otherwise.
- */
-bool
-i915_gem_object_pin_fence(struct drm_i915_gem_object *obj)
-{
-	if (obj->fence_reg != I915_FENCE_REG_NONE) {
-		to_i915(obj->base.dev)->fence_regs[obj->fence_reg].pin_count++;
-		return true;
-	} else
-		return false;
-}
-
-/**
- * i915_gem_object_unpin_fence - unpin fencing state
- * @obj: object to unpin fencing for
- *
- * This releases the fence pin reference acquired through
- * i915_gem_object_pin_fence. It will handle both objects with and without an
- * attached fence correctly, callers do not need to distinguish this.
- */
-void
-i915_gem_object_unpin_fence(struct drm_i915_gem_object *obj)
-{
-	if (obj->fence_reg != I915_FENCE_REG_NONE) {
-		struct drm_i915_private *dev_priv = to_i915(obj->base.dev);
-		WARN_ON(dev_priv->fence_regs[obj->fence_reg].pin_count <= 0);
-		dev_priv->fence_regs[obj->fence_reg].pin_count--;
-	}
+	return fence_update(fence, set);
 }
 
 /**
@@ -473,12 +378,7 @@ void i915_gem_restore_fences(struct drm_device *dev)
 		 * Commit delayed tiling changes if we have an object still
 		 * attached to the fence, otherwise just clear the fence.
 		 */
-		if (reg->obj) {
-			i915_gem_object_update_fence(reg->obj, reg,
-						     i915_gem_object_get_tiling(reg->obj));
-		} else {
-			i915_gem_write_fence(dev, i, NULL);
-		}
+		fence_write(reg, reg->vma);
 	}
 }
 

drivers/gpu/drm/i915/i915_gem_gtt.c

@@ -3322,6 +3322,7 @@ void i915_vma_destroy(struct i915_vma *vma)
 	GEM_BUG_ON(vma->node.allocated);
 	GEM_BUG_ON(i915_vma_is_active(vma));
 	GEM_BUG_ON(!i915_vma_is_closed(vma));
+	GEM_BUG_ON(vma->fence);
 
 	list_del(&vma->vm_link);
 	if (!i915_vma_is_ggtt(vma))
@@ -3357,6 +3358,7 @@ __i915_vma_create(struct drm_i915_gem_object *obj,
 	INIT_LIST_HEAD(&vma->exec_list);
 	for (i = 0; i < ARRAY_SIZE(vma->last_read); i++)
 		init_request_active(&vma->last_read[i], i915_vma_retire);
+	init_request_active(&vma->last_fence, NULL);
 	list_add(&vma->vm_link, &vm->unbound_list);
 	vma->vm = vm;
 	vma->obj = obj;

drivers/gpu/drm/i915/i915_gem_gtt.h

@@ -38,7 +38,13 @@
 
 #include "i915_gem_request.h"
 
+#define I915_FENCE_REG_NONE -1
+#define I915_MAX_NUM_FENCES 32
+/* 32 fences + sign bit for FENCE_REG_NONE */
+#define I915_MAX_NUM_FENCE_BITS 6
+
 struct drm_i915_file_private;
+struct drm_i915_fence_reg;
 
 typedef uint32_t gen6_pte_t;
 typedef uint64_t gen8_pte_t;
@@ -174,6 +180,7 @@ struct i915_vma {
 	struct drm_mm_node node;
 	struct drm_i915_gem_object *obj;
 	struct i915_address_space *vm;
+	struct drm_i915_fence_reg *fence;
 	struct sg_table *pages;
 	void __iomem *iomap;
 	u64 size;
@@ -203,6 +210,7 @@ struct i915_vma {
 
 	unsigned int active;
 	struct i915_gem_active last_read[I915_NUM_ENGINES];
+	struct i915_gem_active last_fence;
 
 	/**
 	 * Support different GGTT views into the same object.

drivers/gpu/drm/i915/i915_gem_tiling.c

@@ -116,13 +116,39 @@ i915_tiling_ok(struct drm_device *dev, int stride, int size, int tiling_mode)
 	return true;
 }
 
+static bool i915_vma_fence_prepare(struct i915_vma *vma, int tiling_mode)
+{
+	struct drm_i915_private *dev_priv = to_i915(vma->vm->dev);
+	u32 size;
+
+	if (!i915_vma_is_map_and_fenceable(vma))
+		return true;
+
+	if (INTEL_GEN(dev_priv) == 3) {
+		if (vma->node.start & ~I915_FENCE_START_MASK)
+			return false;
+	} else {
+		if (vma->node.start & ~I830_FENCE_START_MASK)
+			return false;
+	}
+
+	size = i915_gem_get_ggtt_size(dev_priv, vma->size, tiling_mode);
+	if (vma->node.size < size)
+		return false;
+
+	if (vma->node.start & (size - 1))
+		return false;
+
+	return true;
+}
+
 /* Make the current GTT allocation valid for the change in tiling. */
 static int
 i915_gem_object_fence_prepare(struct drm_i915_gem_object *obj, int tiling_mode)
 {
 	struct drm_i915_private *dev_priv = to_i915(obj->base.dev);
 	struct i915_vma *vma;
-	u32 size;
+	int ret;
 
 	if (tiling_mode == I915_TILING_NONE)
 		return 0;
@@ -130,32 +156,16 @@ i915_gem_object_fence_prepare(struct drm_i915_gem_object *obj, int tiling_mode)
 	if (INTEL_GEN(dev_priv) >= 4)
 		return 0;
 
-	vma = i915_gem_object_to_ggtt(obj, NULL);
-	if (!vma)
-		return 0;
-
-	if (!i915_vma_is_map_and_fenceable(vma))
-		return 0;
+	list_for_each_entry(vma, &obj->vma_list, obj_link) {
+		if (i915_vma_fence_prepare(vma, tiling_mode))
+			continue;
 
-	if (IS_GEN3(dev_priv)) {
-		if (vma->node.start & ~I915_FENCE_START_MASK)
-			goto bad;
-	} else {
-		if (vma->node.start & ~I830_FENCE_START_MASK)
-			goto bad;
+		ret = i915_vma_unbind(vma);
+		if (ret)
+			return ret;
 	}
 
-	size = i915_gem_get_ggtt_size(dev_priv, vma->size, tiling_mode);
-	if (vma->node.size < size)
-		goto bad;
-
-	if (vma->node.start & (size - 1))
-		goto bad;
-
 	return 0;
-
-bad:
-	return i915_vma_unbind(vma);
 }
 
 /**
@@ -248,6 +258,8 @@ i915_gem_set_tiling(struct drm_device *dev, void *data,
 
 		err = i915_gem_object_fence_prepare(obj, args->tiling_mode);
 		if (!err) {
+			struct i915_vma *vma;
+
 			if (obj->pages &&
 			    obj->madv == I915_MADV_WILLNEED &&
 			    dev_priv->quirks & QUIRK_PIN_SWIZZLED_PAGES) {
@@ -257,11 +269,12 @@ i915_gem_set_tiling(struct drm_device *dev, void *data,
 					i915_gem_object_pin_pages(obj);
 			}
 
-			obj->fence_dirty =
-				!i915_gem_active_is_idle(&obj->last_fence,
-							 &dev->struct_mutex) ||
-				obj->fence_reg != I915_FENCE_REG_NONE;
+			list_for_each_entry(vma, &obj->vma_list, obj_link) {
+				if (!vma->fence)
+					continue;
 
+				vma->fence->dirty = true;
+			}
 			obj->tiling_and_stride =
 				args->stride | args->tiling_mode;
 

drivers/gpu/drm/i915/i915_gpu_error.c

@@ -797,7 +797,7 @@ static void capture_bo(struct drm_i915_error_buffer *err,
 	err->gtt_offset = vma->node.start;
 	err->read_domains = obj->base.read_domains;
 	err->write_domain = obj->base.write_domain;
-	err->fence_reg = obj->fence_reg;
+	err->fence_reg = vma->fence ? vma->fence->id : -1;
 	err->tiling = i915_gem_object_get_tiling(obj);
 	err->dirty = obj->dirty;
 	err->purgeable = obj->madv != I915_MADV_WILLNEED;

drivers/gpu/drm/i915/intel_display.c

@@ -2188,7 +2188,6 @@ intel_pin_and_fence_fb_obj(struct drm_framebuffer *fb, unsigned int rotation)
 	struct i915_ggtt_view view;
 	struct i915_vma *vma;
 	u32 alignment;
-	int ret;
 
 	WARN_ON(!mutex_is_locked(&dev->struct_mutex));
 
@@ -2214,43 +2213,33 @@ intel_pin_and_fence_fb_obj(struct drm_framebuffer *fb, unsigned int rotation)
 	intel_runtime_pm_get(dev_priv);
 
 	vma = i915_gem_object_pin_to_display_plane(obj, alignment, &view);
-	if (IS_ERR(vma)) {
-		ret = PTR_ERR(vma);
-		goto err_pm;
-	}
+	if (IS_ERR(vma))
+		goto err;
 
-	/* Install a fence for tiled scan-out. Pre-i965 always needs a
-	 * fence, whereas 965+ only requires a fence if using
-	 * framebuffer compression.  For simplicity, we always install
-	 * a fence as the cost is not that onerous.
-	 */
 	if (i915_vma_is_map_and_fenceable(vma)) {
-		ret = i915_gem_object_get_fence(obj);
-		if (ret == -EDEADLK) {
-			/*
-			 * -EDEADLK means there are no free fences
-			 * no pending flips.
-			 *
-			 * This is propagated to atomic, but it uses
-			 * -EDEADLK to force a locking recovery, so
-			 * change the returned error to -EBUSY.
-			 */
-			ret = -EBUSY;
-			goto err_unpin;
-		} else if (ret)
-			goto err_unpin;
-
-		i915_gem_object_pin_fence(obj);
+		/* Install a fence for tiled scan-out. Pre-i965 always needs a
+		 * fence, whereas 965+ only requires a fence if using
+		 * framebuffer compression.  For simplicity, we always, when
+		 * possible, install a fence as the cost is not that onerous.
+		 *
+		 * If we fail to fence the tiled scanout, then either the
+		 * modeset will reject the change (which is highly unlikely as
+		 * the affected systems, all but one, do not have unmappable
+		 * space) or we will not be able to enable full powersaving
+		 * techniques (also likely not to apply due to various limits
+		 * FBC and the like impose on the size of the buffer, which
+		 * presumably we violated anyway with this unmappable buffer).
+		 * Anyway, it is presumably better to stumble onwards with
+		 * something and try to run the system in a "less than optimal"
+		 * mode that matches the user configuration.
+		 */
+		if (i915_vma_get_fence(vma) == 0)
+			i915_vma_pin_fence(vma);
 	}
 
+err:
 	intel_runtime_pm_put(dev_priv);
 	return vma;
-
-err_unpin:
-	i915_gem_object_unpin_from_display_plane(vma);
-err_pm:
-	intel_runtime_pm_put(dev_priv);
-	return ERR_PTR(ret);
 }
 
 void intel_unpin_fb_obj(struct drm_framebuffer *fb, unsigned int rotation)
@@ -2264,9 +2253,7 @@ void intel_unpin_fb_obj(struct drm_framebuffer *fb, unsigned int rotation)
 	intel_fill_fb_ggtt_view(&view, fb, rotation);
 	vma = i915_gem_object_to_ggtt(obj, &view);
 
-	if (i915_vma_is_map_and_fenceable(vma))
-		i915_gem_object_unpin_fence(obj);
-
+	i915_vma_unpin_fence(vma);
 	i915_gem_object_unpin_from_display_plane(vma);
 }
 

drivers/gpu/drm/i915/intel_fbc.c

@@ -709,6 +709,14 @@ static bool intel_fbc_hw_tracking_covers_screen(struct intel_crtc *crtc)
 	return effective_w <= max_w && effective_h <= max_h;
 }
 
+/* XXX replace me when we have VMA tracking for intel_plane_state */
+static int get_fence_id(struct drm_framebuffer *fb)
+{
+	struct i915_vma *vma = i915_gem_object_to_ggtt(intel_fb_obj(fb), NULL);
+
+	return vma && vma->fence ? vma->fence->id : I915_FENCE_REG_NONE;
+}
+
 static void intel_fbc_update_state_cache(struct intel_crtc *crtc,
 					 struct intel_crtc_state *crtc_state,
 					 struct intel_plane_state *plane_state)
@@ -740,7 +748,7 @@ static void intel_fbc_update_state_cache(struct intel_crtc *crtc,
 		cache->fb.ilk_ggtt_offset = i915_gem_object_ggtt_offset(obj, NULL);
 	cache->fb.pixel_format = fb->pixel_format;
 	cache->fb.stride = fb->pitches[0];
-	cache->fb.fence_reg = obj->fence_reg;
+	cache->fb.fence_reg = get_fence_id(fb);
 	cache->fb.tiling_mode = i915_gem_object_get_tiling(obj);
 }
 

drivers/gpu/drm/i915/intel_overlay.c

@@ -760,7 +760,7 @@ static int intel_overlay_do_put_image(struct intel_overlay *overlay,
 	if (IS_ERR(vma))
 		return PTR_ERR(vma);
 
-	ret = i915_gem_object_put_fence(new_bo);
+	ret = i915_vma_put_fence(vma);
 	if (ret)
 		goto out_unpin;