Linux Audio

Check our new training course

Loading...
v3.1
  1/*
  2 * Copyright © 2010 Daniel Vetter
 
  3 *
  4 * Permission is hereby granted, free of charge, to any person obtaining a
  5 * copy of this software and associated documentation files (the "Software"),
  6 * to deal in the Software without restriction, including without limitation
  7 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
  8 * and/or sell copies of the Software, and to permit persons to whom the
  9 * Software is furnished to do so, subject to the following conditions:
 10 *
 11 * The above copyright notice and this permission notice (including the next
 12 * paragraph) shall be included in all copies or substantial portions of the
 13 * Software.
 14 *
 15 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
 16 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
 17 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
 18 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
 19 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
 20 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
 21 * IN THE SOFTWARE.
 22 *
 23 */
 24
 25#include "drmP.h"
 26#include "drm.h"
 27#include "i915_drm.h"
 
 
 
 
 
 
 
 
 
 
 28#include "i915_drv.h"
 
 29#include "i915_trace.h"
 30#include "intel_drv.h"
 
 31
 32/* XXX kill agp_type! */
 33static unsigned int cache_level_to_agp_type(struct drm_device *dev,
 34					    enum i915_cache_level cache_level)
 35{
 36	switch (cache_level) {
 37	case I915_CACHE_LLC_MLC:
 38		if (INTEL_INFO(dev)->gen >= 6)
 39			return AGP_USER_CACHED_MEMORY_LLC_MLC;
 40		/* Older chipsets do not have this extra level of CPU
 41		 * cacheing, so fallthrough and request the PTE simply
 42		 * as cached.
 43		 */
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 44	case I915_CACHE_LLC:
 45		return AGP_USER_CACHED_MEMORY;
 
 
 
 
 46	default:
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 47	case I915_CACHE_NONE:
 48		return AGP_USER_MEMORY;
 
 
 
 49	}
 
 
 50}
 51
 52void i915_gem_restore_gtt_mappings(struct drm_device *dev)
 
 
 53{
 54	struct drm_i915_private *dev_priv = dev->dev_private;
 55	struct drm_i915_gem_object *obj;
 56
 57	/* First fill our portion of the GTT with scratch pages */
 58	intel_gtt_clear_range(dev_priv->mm.gtt_start / PAGE_SIZE,
 59			      (dev_priv->mm.gtt_end - dev_priv->mm.gtt_start) / PAGE_SIZE);
 
 
 
 
 
 60
 61	list_for_each_entry(obj, &dev_priv->mm.gtt_list, gtt_list) {
 62		i915_gem_clflush_object(obj);
 63		i915_gem_gtt_rebind_object(obj, obj->cache_level);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 64	}
 65
 66	intel_gtt_chipset_flush();
 67}
 68
 69int i915_gem_gtt_bind_object(struct drm_i915_gem_object *obj)
 70{
 71	struct drm_device *dev = obj->base.dev;
 72	struct drm_i915_private *dev_priv = dev->dev_private;
 73	unsigned int agp_type = cache_level_to_agp_type(dev, obj->cache_level);
 74	int ret;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 75
 76	if (dev_priv->mm.gtt->needs_dmar) {
 77		ret = intel_gtt_map_memory(obj->pages,
 78					   obj->base.size >> PAGE_SHIFT,
 79					   &obj->sg_list,
 80					   &obj->num_sg);
 81		if (ret != 0)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 82			return ret;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 83
 84		intel_gtt_insert_sg_entries(obj->sg_list,
 85					    obj->num_sg,
 86					    obj->gtt_space->start >> PAGE_SHIFT,
 87					    agp_type);
 88	} else
 89		intel_gtt_insert_pages(obj->gtt_space->start >> PAGE_SHIFT,
 90				       obj->base.size >> PAGE_SHIFT,
 91				       obj->pages,
 92				       agp_type);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 93
 94	return 0;
 95}
 96
 97void i915_gem_gtt_rebind_object(struct drm_i915_gem_object *obj,
 98				enum i915_cache_level cache_level)
 
 
 
 
 
 
 
 
 
 
 
 
 
 99{
100	struct drm_device *dev = obj->base.dev;
101	struct drm_i915_private *dev_priv = dev->dev_private;
102	unsigned int agp_type = cache_level_to_agp_type(dev, cache_level);
 
103
104	if (dev_priv->mm.gtt->needs_dmar) {
105		BUG_ON(!obj->sg_list);
 
106
107		intel_gtt_insert_sg_entries(obj->sg_list,
108					    obj->num_sg,
109					    obj->gtt_space->start >> PAGE_SHIFT,
110					    agp_type);
111	} else
112		intel_gtt_insert_pages(obj->gtt_space->start >> PAGE_SHIFT,
113				       obj->base.size >> PAGE_SHIFT,
114				       obj->pages,
115				       agp_type);
116}
117
118void i915_gem_gtt_unbind_object(struct drm_i915_gem_object *obj)
119{
120	intel_gtt_clear_range(obj->gtt_space->start >> PAGE_SHIFT,
121			      obj->base.size >> PAGE_SHIFT);
 
 
 
 
 
122
123	if (obj->sg_list) {
124		intel_gtt_unmap_memory(obj->sg_list, obj->num_sg);
125		obj->sg_list = NULL;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
126	}
 
 
 
 
 
 
 
 
127}
v4.17
   1/*
   2 * Copyright © 2010 Daniel Vetter
   3 * Copyright © 2011-2014 Intel Corporation
   4 *
   5 * Permission is hereby granted, free of charge, to any person obtaining a
   6 * copy of this software and associated documentation files (the "Software"),
   7 * to deal in the Software without restriction, including without limitation
   8 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
   9 * and/or sell copies of the Software, and to permit persons to whom the
  10 * Software is furnished to do so, subject to the following conditions:
  11 *
  12 * The above copyright notice and this permission notice (including the next
  13 * paragraph) shall be included in all copies or substantial portions of the
  14 * Software.
  15 *
  16 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  17 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  18 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
  19 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
  20 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
  21 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
  22 * IN THE SOFTWARE.
  23 *
  24 */
  25
  26#include <linux/slab.h> /* fault-inject.h is not standalone! */
  27
  28#include <linux/fault-inject.h>
  29#include <linux/log2.h>
  30#include <linux/random.h>
  31#include <linux/seq_file.h>
  32#include <linux/stop_machine.h>
  33
  34#include <asm/set_memory.h>
  35
  36#include <drm/drmP.h>
  37#include <drm/i915_drm.h>
  38
  39#include "i915_drv.h"
  40#include "i915_vgpu.h"
  41#include "i915_trace.h"
  42#include "intel_drv.h"
  43#include "intel_frontbuffer.h"
  44
  45#define I915_GFP_DMA (GFP_KERNEL | __GFP_HIGHMEM)
  46
  47/**
  48 * DOC: Global GTT views
  49 *
  50 * Background and previous state
  51 *
  52 * Historically objects could exists (be bound) in global GTT space only as
  53 * singular instances with a view representing all of the object's backing pages
  54 * in a linear fashion. This view will be called a normal view.
  55 *
  56 * To support multiple views of the same object, where the number of mapped
  57 * pages is not equal to the backing store, or where the layout of the pages
  58 * is not linear, concept of a GGTT view was added.
  59 *
  60 * One example of an alternative view is a stereo display driven by a single
  61 * image. In this case we would have a framebuffer looking like this
  62 * (2x2 pages):
  63 *
  64 *    12
  65 *    34
  66 *
  67 * Above would represent a normal GGTT view as normally mapped for GPU or CPU
  68 * rendering. In contrast, fed to the display engine would be an alternative
  69 * view which could look something like this:
  70 *
  71 *   1212
  72 *   3434
  73 *
  74 * In this example both the size and layout of pages in the alternative view is
  75 * different from the normal view.
  76 *
  77 * Implementation and usage
  78 *
  79 * GGTT views are implemented using VMAs and are distinguished via enum
  80 * i915_ggtt_view_type and struct i915_ggtt_view.
  81 *
  82 * A new flavour of core GEM functions which work with GGTT bound objects were
  83 * added with the _ggtt_ infix, and sometimes with _view postfix to avoid
  84 * renaming  in large amounts of code. They take the struct i915_ggtt_view
  85 * parameter encapsulating all metadata required to implement a view.
  86 *
  87 * As a helper for callers which are only interested in the normal view,
  88 * globally const i915_ggtt_view_normal singleton instance exists. All old core
  89 * GEM API functions, the ones not taking the view parameter, are operating on,
  90 * or with the normal GGTT view.
  91 *
  92 * Code wanting to add or use a new GGTT view needs to:
  93 *
  94 * 1. Add a new enum with a suitable name.
  95 * 2. Extend the metadata in the i915_ggtt_view structure if required.
  96 * 3. Add support to i915_get_vma_pages().
  97 *
  98 * New views are required to build a scatter-gather table from within the
  99 * i915_get_vma_pages function. This table is stored in the vma.ggtt_view and
 100 * exists for the lifetime of an VMA.
 101 *
 102 * Core API is designed to have copy semantics which means that passed in
 103 * struct i915_ggtt_view does not need to be persistent (left around after
 104 * calling the core API functions).
 105 *
 106 */
 107
 108static int
 109i915_get_ggtt_vma_pages(struct i915_vma *vma);
 110
 111static void gen6_ggtt_invalidate(struct drm_i915_private *dev_priv)
 112{
 113	/* Note that as an uncached mmio write, this should flush the
 114	 * WCB of the writes into the GGTT before it triggers the invalidate.
 115	 */
 116	I915_WRITE(GFX_FLSH_CNTL_GEN6, GFX_FLSH_CNTL_EN);
 117}
 118
 119static void guc_ggtt_invalidate(struct drm_i915_private *dev_priv)
 120{
 121	gen6_ggtt_invalidate(dev_priv);
 122	I915_WRITE(GEN8_GTCR, GEN8_GTCR_INVALIDATE);
 123}
 124
 125static void gmch_ggtt_invalidate(struct drm_i915_private *dev_priv)
 126{
 127	intel_gtt_chipset_flush();
 128}
 129
 130static inline void i915_ggtt_invalidate(struct drm_i915_private *i915)
 131{
 132	i915->ggtt.invalidate(i915);
 133}
 134
 135int intel_sanitize_enable_ppgtt(struct drm_i915_private *dev_priv,
 136			       	int enable_ppgtt)
 137{
 138	bool has_full_ppgtt;
 139	bool has_full_48bit_ppgtt;
 140
 141	if (!dev_priv->info.has_aliasing_ppgtt)
 142		return 0;
 143
 144	has_full_ppgtt = dev_priv->info.has_full_ppgtt;
 145	has_full_48bit_ppgtt = dev_priv->info.has_full_48bit_ppgtt;
 146
 147	if (intel_vgpu_active(dev_priv)) {
 148		/* GVT-g has no support for 32bit ppgtt */
 149		has_full_ppgtt = false;
 150		has_full_48bit_ppgtt = intel_vgpu_has_full_48bit_ppgtt(dev_priv);
 151	}
 152
 153	/*
 154	 * We don't allow disabling PPGTT for gen9+ as it's a requirement for
 155	 * execlists, the sole mechanism available to submit work.
 156	 */
 157	if (enable_ppgtt == 0 && INTEL_GEN(dev_priv) < 9)
 158		return 0;
 159
 160	if (enable_ppgtt == 1)
 161		return 1;
 162
 163	if (enable_ppgtt == 2 && has_full_ppgtt)
 164		return 2;
 165
 166	if (enable_ppgtt == 3 && has_full_48bit_ppgtt)
 167		return 3;
 168
 169	/* Disable ppgtt on SNB if VT-d is on. */
 170	if (IS_GEN6(dev_priv) && intel_vtd_active()) {
 171		DRM_INFO("Disabling PPGTT because VT-d is on\n");
 172		return 0;
 173	}
 174
 175	/* Early VLV doesn't have this */
 176	if (IS_VALLEYVIEW(dev_priv) && dev_priv->drm.pdev->revision < 0xb) {
 177		DRM_DEBUG_DRIVER("disabling PPGTT on pre-B3 step VLV\n");
 178		return 0;
 179	}
 180
 181	if (HAS_LOGICAL_RING_CONTEXTS(dev_priv)) {
 182		if (has_full_48bit_ppgtt)
 183			return 3;
 184
 185		if (has_full_ppgtt)
 186			return 2;
 187	}
 188
 189	return 1;
 190}
 191
 192static int ppgtt_bind_vma(struct i915_vma *vma,
 193			  enum i915_cache_level cache_level,
 194			  u32 unused)
 195{
 196	u32 pte_flags;
 197	int ret;
 198
 199	if (!(vma->flags & I915_VMA_LOCAL_BIND)) {
 200		ret = vma->vm->allocate_va_range(vma->vm, vma->node.start,
 201						 vma->size);
 202		if (ret)
 203			return ret;
 204	}
 205
 206	/* Currently applicable only to VLV */
 207	pte_flags = 0;
 208	if (vma->obj->gt_ro)
 209		pte_flags |= PTE_READ_ONLY;
 210
 211	vma->vm->insert_entries(vma->vm, vma, cache_level, pte_flags);
 212
 213	return 0;
 214}
 215
 216static void ppgtt_unbind_vma(struct i915_vma *vma)
 217{
 218	vma->vm->clear_range(vma->vm, vma->node.start, vma->size);
 219}
 220
 221static int ppgtt_set_pages(struct i915_vma *vma)
 222{
 223	GEM_BUG_ON(vma->pages);
 224
 225	vma->pages = vma->obj->mm.pages;
 226
 227	vma->page_sizes = vma->obj->mm.page_sizes;
 228
 229	return 0;
 230}
 231
 232static void clear_pages(struct i915_vma *vma)
 233{
 234	GEM_BUG_ON(!vma->pages);
 235
 236	if (vma->pages != vma->obj->mm.pages) {
 237		sg_free_table(vma->pages);
 238		kfree(vma->pages);
 239	}
 240	vma->pages = NULL;
 241
 242	memset(&vma->page_sizes, 0, sizeof(vma->page_sizes));
 243}
 244
 245static gen8_pte_t gen8_pte_encode(dma_addr_t addr,
 246				  enum i915_cache_level level)
 247{
 248	gen8_pte_t pte = _PAGE_PRESENT | _PAGE_RW;
 249	pte |= addr;
 250
 251	switch (level) {
 252	case I915_CACHE_NONE:
 253		pte |= PPAT_UNCACHED;
 254		break;
 255	case I915_CACHE_WT:
 256		pte |= PPAT_DISPLAY_ELLC;
 257		break;
 258	default:
 259		pte |= PPAT_CACHED;
 260		break;
 261	}
 262
 263	return pte;
 264}
 265
 266static gen8_pde_t gen8_pde_encode(const dma_addr_t addr,
 267				  const enum i915_cache_level level)
 268{
 269	gen8_pde_t pde = _PAGE_PRESENT | _PAGE_RW;
 270	pde |= addr;
 271	if (level != I915_CACHE_NONE)
 272		pde |= PPAT_CACHED_PDE;
 273	else
 274		pde |= PPAT_UNCACHED;
 275	return pde;
 276}
 277
 278#define gen8_pdpe_encode gen8_pde_encode
 279#define gen8_pml4e_encode gen8_pde_encode
 280
 281static gen6_pte_t snb_pte_encode(dma_addr_t addr,
 282				 enum i915_cache_level level,
 283				 u32 unused)
 284{
 285	gen6_pte_t pte = GEN6_PTE_VALID;
 286	pte |= GEN6_PTE_ADDR_ENCODE(addr);
 287
 288	switch (level) {
 289	case I915_CACHE_L3_LLC:
 290	case I915_CACHE_LLC:
 291		pte |= GEN6_PTE_CACHE_LLC;
 292		break;
 293	case I915_CACHE_NONE:
 294		pte |= GEN6_PTE_UNCACHED;
 295		break;
 296	default:
 297		MISSING_CASE(level);
 298	}
 299
 300	return pte;
 301}
 302
 303static gen6_pte_t ivb_pte_encode(dma_addr_t addr,
 304				 enum i915_cache_level level,
 305				 u32 unused)
 306{
 307	gen6_pte_t pte = GEN6_PTE_VALID;
 308	pte |= GEN6_PTE_ADDR_ENCODE(addr);
 309
 310	switch (level) {
 311	case I915_CACHE_L3_LLC:
 312		pte |= GEN7_PTE_CACHE_L3_LLC;
 313		break;
 314	case I915_CACHE_LLC:
 315		pte |= GEN6_PTE_CACHE_LLC;
 316		break;
 317	case I915_CACHE_NONE:
 318		pte |= GEN6_PTE_UNCACHED;
 319		break;
 320	default:
 321		MISSING_CASE(level);
 322	}
 323
 324	return pte;
 325}
 326
 327static gen6_pte_t byt_pte_encode(dma_addr_t addr,
 328				 enum i915_cache_level level,
 329				 u32 flags)
 330{
 331	gen6_pte_t pte = GEN6_PTE_VALID;
 332	pte |= GEN6_PTE_ADDR_ENCODE(addr);
 333
 334	if (!(flags & PTE_READ_ONLY))
 335		pte |= BYT_PTE_WRITEABLE;
 336
 337	if (level != I915_CACHE_NONE)
 338		pte |= BYT_PTE_SNOOPED_BY_CPU_CACHES;
 339
 340	return pte;
 341}
 342
 343static gen6_pte_t hsw_pte_encode(dma_addr_t addr,
 344				 enum i915_cache_level level,
 345				 u32 unused)
 346{
 347	gen6_pte_t pte = GEN6_PTE_VALID;
 348	pte |= HSW_PTE_ADDR_ENCODE(addr);
 349
 350	if (level != I915_CACHE_NONE)
 351		pte |= HSW_WB_LLC_AGE3;
 352
 353	return pte;
 354}
 355
 356static gen6_pte_t iris_pte_encode(dma_addr_t addr,
 357				  enum i915_cache_level level,
 358				  u32 unused)
 359{
 360	gen6_pte_t pte = GEN6_PTE_VALID;
 361	pte |= HSW_PTE_ADDR_ENCODE(addr);
 362
 363	switch (level) {
 364	case I915_CACHE_NONE:
 365		break;
 366	case I915_CACHE_WT:
 367		pte |= HSW_WT_ELLC_LLC_AGE3;
 368		break;
 369	default:
 370		pte |= HSW_WB_ELLC_LLC_AGE3;
 371		break;
 372	}
 373
 374	return pte;
 375}
 376
 377static struct page *vm_alloc_page(struct i915_address_space *vm, gfp_t gfp)
 378{
 379	struct pagevec *pvec = &vm->free_pages;
 380	struct pagevec stash;
 381
 382	if (I915_SELFTEST_ONLY(should_fail(&vm->fault_attr, 1)))
 383		i915_gem_shrink_all(vm->i915);
 384
 385	if (likely(pvec->nr))
 386		return pvec->pages[--pvec->nr];
 387
 388	if (!vm->pt_kmap_wc)
 389		return alloc_page(gfp);
 390
 391	/* A placeholder for a specific mutex to guard the WC stash */
 392	lockdep_assert_held(&vm->i915->drm.struct_mutex);
 393
 394	/* Look in our global stash of WC pages... */
 395	pvec = &vm->i915->mm.wc_stash;
 396	if (likely(pvec->nr))
 397		return pvec->pages[--pvec->nr];
 398
 399	/*
 400	 * Otherwise batch allocate pages to amoritize cost of set_pages_wc.
 401	 *
 402	 * We have to be careful as page allocation may trigger the shrinker
 403	 * (via direct reclaim) which will fill up the WC stash underneath us.
 404	 * So we add our WB pages into a temporary pvec on the stack and merge
 405	 * them into the WC stash after all the allocations are complete.
 406	 */
 407	pagevec_init(&stash);
 408	do {
 409		struct page *page;
 410
 411		page = alloc_page(gfp);
 412		if (unlikely(!page))
 413			break;
 414
 415		stash.pages[stash.nr++] = page;
 416	} while (stash.nr < pagevec_space(pvec));
 417
 418	if (stash.nr) {
 419		int nr = min_t(int, stash.nr, pagevec_space(pvec));
 420		struct page **pages = stash.pages + stash.nr - nr;
 421
 422		if (nr && !set_pages_array_wc(pages, nr)) {
 423			memcpy(pvec->pages + pvec->nr,
 424			       pages, sizeof(pages[0]) * nr);
 425			pvec->nr += nr;
 426			stash.nr -= nr;
 427		}
 428
 429		pagevec_release(&stash);
 430	}
 431
 432	return likely(pvec->nr) ? pvec->pages[--pvec->nr] : NULL;
 433}
 434
 435static void vm_free_pages_release(struct i915_address_space *vm,
 436				  bool immediate)
 437{
 438	struct pagevec *pvec = &vm->free_pages;
 439
 440	GEM_BUG_ON(!pagevec_count(pvec));
 441
 442	if (vm->pt_kmap_wc) {
 443		struct pagevec *stash = &vm->i915->mm.wc_stash;
 444
 445		/* When we use WC, first fill up the global stash and then
 446		 * only if full immediately free the overflow.
 447		 */
 448
 449		lockdep_assert_held(&vm->i915->drm.struct_mutex);
 450		if (pagevec_space(stash)) {
 451			do {
 452				stash->pages[stash->nr++] =
 453					pvec->pages[--pvec->nr];
 454				if (!pvec->nr)
 455					return;
 456			} while (pagevec_space(stash));
 457
 458			/* As we have made some room in the VM's free_pages,
 459			 * we can wait for it to fill again. Unless we are
 460			 * inside i915_address_space_fini() and must
 461			 * immediately release the pages!
 462			 */
 463			if (!immediate)
 464				return;
 465		}
 466
 467		set_pages_array_wb(pvec->pages, pvec->nr);
 468	}
 469
 470	__pagevec_release(pvec);
 471}
 472
 473static void vm_free_page(struct i915_address_space *vm, struct page *page)
 474{
 475	/*
 476	 * On !llc, we need to change the pages back to WB. We only do so
 477	 * in bulk, so we rarely need to change the page attributes here,
 478	 * but doing so requires a stop_machine() from deep inside arch/x86/mm.
 479	 * To make detection of the possible sleep more likely, use an
 480	 * unconditional might_sleep() for everybody.
 481	 */
 482	might_sleep();
 483	if (!pagevec_add(&vm->free_pages, page))
 484		vm_free_pages_release(vm, false);
 485}
 486
 487static int __setup_page_dma(struct i915_address_space *vm,
 488			    struct i915_page_dma *p,
 489			    gfp_t gfp)
 490{
 491	p->page = vm_alloc_page(vm, gfp | __GFP_NOWARN | __GFP_NORETRY);
 492	if (unlikely(!p->page))
 493		return -ENOMEM;
 494
 495	p->daddr = dma_map_page(vm->dma, p->page, 0, PAGE_SIZE,
 496				PCI_DMA_BIDIRECTIONAL);
 497	if (unlikely(dma_mapping_error(vm->dma, p->daddr))) {
 498		vm_free_page(vm, p->page);
 499		return -ENOMEM;
 500	}
 501
 502	return 0;
 503}
 504
 505static int setup_page_dma(struct i915_address_space *vm,
 506			  struct i915_page_dma *p)
 507{
 508	return __setup_page_dma(vm, p, I915_GFP_DMA);
 509}
 510
 511static void cleanup_page_dma(struct i915_address_space *vm,
 512			     struct i915_page_dma *p)
 513{
 514	dma_unmap_page(vm->dma, p->daddr, PAGE_SIZE, PCI_DMA_BIDIRECTIONAL);
 515	vm_free_page(vm, p->page);
 516}
 517
 518#define kmap_atomic_px(px) kmap_atomic(px_base(px)->page)
 519
 520#define setup_px(vm, px) setup_page_dma((vm), px_base(px))
 521#define cleanup_px(vm, px) cleanup_page_dma((vm), px_base(px))
 522#define fill_px(ppgtt, px, v) fill_page_dma((vm), px_base(px), (v))
 523#define fill32_px(ppgtt, px, v) fill_page_dma_32((vm), px_base(px), (v))
 524
 525static void fill_page_dma(struct i915_address_space *vm,
 526			  struct i915_page_dma *p,
 527			  const u64 val)
 528{
 529	u64 * const vaddr = kmap_atomic(p->page);
 530
 531	memset64(vaddr, val, PAGE_SIZE / sizeof(val));
 532
 533	kunmap_atomic(vaddr);
 534}
 535
 536static void fill_page_dma_32(struct i915_address_space *vm,
 537			     struct i915_page_dma *p,
 538			     const u32 v)
 539{
 540	fill_page_dma(vm, p, (u64)v << 32 | v);
 541}
 542
 543static int
 544setup_scratch_page(struct i915_address_space *vm, gfp_t gfp)
 545{
 546	unsigned long size;
 547
 548	/*
 549	 * In order to utilize 64K pages for an object with a size < 2M, we will
 550	 * need to support a 64K scratch page, given that every 16th entry for a
 551	 * page-table operating in 64K mode must point to a properly aligned 64K
 552	 * region, including any PTEs which happen to point to scratch.
 553	 *
 554	 * This is only relevant for the 48b PPGTT where we support
 555	 * huge-gtt-pages, see also i915_vma_insert().
 556	 *
 557	 * TODO: we should really consider write-protecting the scratch-page and
 558	 * sharing between ppgtt
 559	 */
 560	size = I915_GTT_PAGE_SIZE_4K;
 561	if (i915_vm_is_48bit(vm) &&
 562	    HAS_PAGE_SIZES(vm->i915, I915_GTT_PAGE_SIZE_64K)) {
 563		size = I915_GTT_PAGE_SIZE_64K;
 564		gfp |= __GFP_NOWARN;
 565	}
 566	gfp |= __GFP_ZERO | __GFP_RETRY_MAYFAIL;
 567
 568	do {
 569		int order = get_order(size);
 570		struct page *page;
 571		dma_addr_t addr;
 572
 573		page = alloc_pages(gfp, order);
 574		if (unlikely(!page))
 575			goto skip;
 576
 577		addr = dma_map_page(vm->dma, page, 0, size,
 578				    PCI_DMA_BIDIRECTIONAL);
 579		if (unlikely(dma_mapping_error(vm->dma, addr)))
 580			goto free_page;
 581
 582		if (unlikely(!IS_ALIGNED(addr, size)))
 583			goto unmap_page;
 584
 585		vm->scratch_page.page = page;
 586		vm->scratch_page.daddr = addr;
 587		vm->scratch_page.order = order;
 588		return 0;
 589
 590unmap_page:
 591		dma_unmap_page(vm->dma, addr, size, PCI_DMA_BIDIRECTIONAL);
 592free_page:
 593		__free_pages(page, order);
 594skip:
 595		if (size == I915_GTT_PAGE_SIZE_4K)
 596			return -ENOMEM;
 597
 598		size = I915_GTT_PAGE_SIZE_4K;
 599		gfp &= ~__GFP_NOWARN;
 600	} while (1);
 601}
 602
 603static void cleanup_scratch_page(struct i915_address_space *vm)
 604{
 605	struct i915_page_dma *p = &vm->scratch_page;
 606
 607	dma_unmap_page(vm->dma, p->daddr, BIT(p->order) << PAGE_SHIFT,
 608		       PCI_DMA_BIDIRECTIONAL);
 609	__free_pages(p->page, p->order);
 610}
 611
 612static struct i915_page_table *alloc_pt(struct i915_address_space *vm)
 613{
 614	struct i915_page_table *pt;
 615
 616	pt = kmalloc(sizeof(*pt), GFP_KERNEL | __GFP_NOWARN);
 617	if (unlikely(!pt))
 618		return ERR_PTR(-ENOMEM);
 619
 620	if (unlikely(setup_px(vm, pt))) {
 621		kfree(pt);
 622		return ERR_PTR(-ENOMEM);
 623	}
 624
 625	pt->used_ptes = 0;
 626	return pt;
 627}
 628
 629static void free_pt(struct i915_address_space *vm, struct i915_page_table *pt)
 630{
 631	cleanup_px(vm, pt);
 632	kfree(pt);
 633}
 634
 635static void gen8_initialize_pt(struct i915_address_space *vm,
 636			       struct i915_page_table *pt)
 637{
 638	fill_px(vm, pt,
 639		gen8_pte_encode(vm->scratch_page.daddr, I915_CACHE_LLC));
 640}
 641
 642static void gen6_initialize_pt(struct i915_address_space *vm,
 643			       struct i915_page_table *pt)
 644{
 645	fill32_px(vm, pt,
 646		  vm->pte_encode(vm->scratch_page.daddr, I915_CACHE_LLC, 0));
 647}
 648
 649static struct i915_page_directory *alloc_pd(struct i915_address_space *vm)
 650{
 651	struct i915_page_directory *pd;
 652
 653	pd = kzalloc(sizeof(*pd), GFP_KERNEL | __GFP_NOWARN);
 654	if (unlikely(!pd))
 655		return ERR_PTR(-ENOMEM);
 656
 657	if (unlikely(setup_px(vm, pd))) {
 658		kfree(pd);
 659		return ERR_PTR(-ENOMEM);
 660	}
 661
 662	pd->used_pdes = 0;
 663	return pd;
 664}
 665
 666static void free_pd(struct i915_address_space *vm,
 667		    struct i915_page_directory *pd)
 668{
 669	cleanup_px(vm, pd);
 670	kfree(pd);
 671}
 672
 673static void gen8_initialize_pd(struct i915_address_space *vm,
 674			       struct i915_page_directory *pd)
 675{
 676	fill_px(vm, pd,
 677		gen8_pde_encode(px_dma(vm->scratch_pt), I915_CACHE_LLC));
 678	memset_p((void **)pd->page_table, vm->scratch_pt, I915_PDES);
 679}
 680
 681static int __pdp_init(struct i915_address_space *vm,
 682		      struct i915_page_directory_pointer *pdp)
 683{
 684	const unsigned int pdpes = i915_pdpes_per_pdp(vm);
 685
 686	pdp->page_directory = kmalloc_array(pdpes, sizeof(*pdp->page_directory),
 687					    GFP_KERNEL | __GFP_NOWARN);
 688	if (unlikely(!pdp->page_directory))
 689		return -ENOMEM;
 690
 691	memset_p((void **)pdp->page_directory, vm->scratch_pd, pdpes);
 692
 693	return 0;
 694}
 695
 696static void __pdp_fini(struct i915_page_directory_pointer *pdp)
 697{
 698	kfree(pdp->page_directory);
 699	pdp->page_directory = NULL;
 700}
 701
 702static inline bool use_4lvl(const struct i915_address_space *vm)
 703{
 704	return i915_vm_is_48bit(vm);
 705}
 706
 707static struct i915_page_directory_pointer *
 708alloc_pdp(struct i915_address_space *vm)
 709{
 710	struct i915_page_directory_pointer *pdp;
 711	int ret = -ENOMEM;
 712
 713	GEM_BUG_ON(!use_4lvl(vm));
 714
 715	pdp = kzalloc(sizeof(*pdp), GFP_KERNEL);
 716	if (!pdp)
 717		return ERR_PTR(-ENOMEM);
 718
 719	ret = __pdp_init(vm, pdp);
 720	if (ret)
 721		goto fail_bitmap;
 722
 723	ret = setup_px(vm, pdp);
 724	if (ret)
 725		goto fail_page_m;
 726
 727	return pdp;
 728
 729fail_page_m:
 730	__pdp_fini(pdp);
 731fail_bitmap:
 732	kfree(pdp);
 733
 734	return ERR_PTR(ret);
 735}
 736
 737static void free_pdp(struct i915_address_space *vm,
 738		     struct i915_page_directory_pointer *pdp)
 739{
 740	__pdp_fini(pdp);
 741
 742	if (!use_4lvl(vm))
 743		return;
 744
 745	cleanup_px(vm, pdp);
 746	kfree(pdp);
 747}
 748
 749static void gen8_initialize_pdp(struct i915_address_space *vm,
 750				struct i915_page_directory_pointer *pdp)
 751{
 752	gen8_ppgtt_pdpe_t scratch_pdpe;
 753
 754	scratch_pdpe = gen8_pdpe_encode(px_dma(vm->scratch_pd), I915_CACHE_LLC);
 755
 756	fill_px(vm, pdp, scratch_pdpe);
 757}
 758
 759static void gen8_initialize_pml4(struct i915_address_space *vm,
 760				 struct i915_pml4 *pml4)
 761{
 762	fill_px(vm, pml4,
 763		gen8_pml4e_encode(px_dma(vm->scratch_pdp), I915_CACHE_LLC));
 764	memset_p((void **)pml4->pdps, vm->scratch_pdp, GEN8_PML4ES_PER_PML4);
 765}
 766
 767/* Broadwell Page Directory Pointer Descriptors */
 768static int gen8_write_pdp(struct i915_request *rq,
 769			  unsigned entry,
 770			  dma_addr_t addr)
 771{
 772	struct intel_engine_cs *engine = rq->engine;
 773	u32 *cs;
 774
 775	BUG_ON(entry >= 4);
 776
 777	cs = intel_ring_begin(rq, 6);
 778	if (IS_ERR(cs))
 779		return PTR_ERR(cs);
 780
 781	*cs++ = MI_LOAD_REGISTER_IMM(1);
 782	*cs++ = i915_mmio_reg_offset(GEN8_RING_PDP_UDW(engine, entry));
 783	*cs++ = upper_32_bits(addr);
 784	*cs++ = MI_LOAD_REGISTER_IMM(1);
 785	*cs++ = i915_mmio_reg_offset(GEN8_RING_PDP_LDW(engine, entry));
 786	*cs++ = lower_32_bits(addr);
 787	intel_ring_advance(rq, cs);
 788
 789	return 0;
 790}
 791
 792static int gen8_mm_switch_3lvl(struct i915_hw_ppgtt *ppgtt,
 793			       struct i915_request *rq)
 794{
 795	int i, ret;
 796
 797	for (i = GEN8_3LVL_PDPES - 1; i >= 0; i--) {
 798		const dma_addr_t pd_daddr = i915_page_dir_dma_addr(ppgtt, i);
 799
 800		ret = gen8_write_pdp(rq, i, pd_daddr);
 801		if (ret)
 802			return ret;
 803	}
 804
 805	return 0;
 806}
 807
 808static int gen8_mm_switch_4lvl(struct i915_hw_ppgtt *ppgtt,
 809			       struct i915_request *rq)
 810{
 811	return gen8_write_pdp(rq, 0, px_dma(&ppgtt->pml4));
 812}
 813
 814/* PDE TLBs are a pain to invalidate on GEN8+. When we modify
 815 * the page table structures, we mark them dirty so that
 816 * context switching/execlist queuing code takes extra steps
 817 * to ensure that tlbs are flushed.
 818 */
 819static void mark_tlbs_dirty(struct i915_hw_ppgtt *ppgtt)
 820{
 821	ppgtt->pd_dirty_rings = INTEL_INFO(ppgtt->base.i915)->ring_mask;
 822}
 823
 824/* Removes entries from a single page table, releasing it if it's empty.
 825 * Caller can use the return value to update higher-level entries.
 826 */
 827static bool gen8_ppgtt_clear_pt(struct i915_address_space *vm,
 828				struct i915_page_table *pt,
 829				u64 start, u64 length)
 830{
 831	unsigned int num_entries = gen8_pte_count(start, length);
 832	unsigned int pte = gen8_pte_index(start);
 833	unsigned int pte_end = pte + num_entries;
 834	const gen8_pte_t scratch_pte =
 835		gen8_pte_encode(vm->scratch_page.daddr, I915_CACHE_LLC);
 836	gen8_pte_t *vaddr;
 837
 838	GEM_BUG_ON(num_entries > pt->used_ptes);
 839
 840	pt->used_ptes -= num_entries;
 841	if (!pt->used_ptes)
 842		return true;
 843
 844	vaddr = kmap_atomic_px(pt);
 845	while (pte < pte_end)
 846		vaddr[pte++] = scratch_pte;
 847	kunmap_atomic(vaddr);
 848
 849	return false;
 850}
 851
 852static void gen8_ppgtt_set_pde(struct i915_address_space *vm,
 853			       struct i915_page_directory *pd,
 854			       struct i915_page_table *pt,
 855			       unsigned int pde)
 856{
 857	gen8_pde_t *vaddr;
 858
 859	pd->page_table[pde] = pt;
 860
 861	vaddr = kmap_atomic_px(pd);
 862	vaddr[pde] = gen8_pde_encode(px_dma(pt), I915_CACHE_LLC);
 863	kunmap_atomic(vaddr);
 864}
 865
 866static bool gen8_ppgtt_clear_pd(struct i915_address_space *vm,
 867				struct i915_page_directory *pd,
 868				u64 start, u64 length)
 869{
 870	struct i915_page_table *pt;
 871	u32 pde;
 872
 873	gen8_for_each_pde(pt, pd, start, length, pde) {
 874		GEM_BUG_ON(pt == vm->scratch_pt);
 875
 876		if (!gen8_ppgtt_clear_pt(vm, pt, start, length))
 877			continue;
 878
 879		gen8_ppgtt_set_pde(vm, pd, vm->scratch_pt, pde);
 880		GEM_BUG_ON(!pd->used_pdes);
 881		pd->used_pdes--;
 882
 883		free_pt(vm, pt);
 884	}
 885
 886	return !pd->used_pdes;
 887}
 888
 889static void gen8_ppgtt_set_pdpe(struct i915_address_space *vm,
 890				struct i915_page_directory_pointer *pdp,
 891				struct i915_page_directory *pd,
 892				unsigned int pdpe)
 893{
 894	gen8_ppgtt_pdpe_t *vaddr;
 895
 896	pdp->page_directory[pdpe] = pd;
 897	if (!use_4lvl(vm))
 898		return;
 899
 900	vaddr = kmap_atomic_px(pdp);
 901	vaddr[pdpe] = gen8_pdpe_encode(px_dma(pd), I915_CACHE_LLC);
 902	kunmap_atomic(vaddr);
 903}
 904
 905/* Removes entries from a single page dir pointer, releasing it if it's empty.
 906 * Caller can use the return value to update higher-level entries
 907 */
 908static bool gen8_ppgtt_clear_pdp(struct i915_address_space *vm,
 909				 struct i915_page_directory_pointer *pdp,
 910				 u64 start, u64 length)
 911{
 912	struct i915_page_directory *pd;
 913	unsigned int pdpe;
 914
 915	gen8_for_each_pdpe(pd, pdp, start, length, pdpe) {
 916		GEM_BUG_ON(pd == vm->scratch_pd);
 917
 918		if (!gen8_ppgtt_clear_pd(vm, pd, start, length))
 919			continue;
 920
 921		gen8_ppgtt_set_pdpe(vm, pdp, vm->scratch_pd, pdpe);
 922		GEM_BUG_ON(!pdp->used_pdpes);
 923		pdp->used_pdpes--;
 924
 925		free_pd(vm, pd);
 926	}
 927
 928	return !pdp->used_pdpes;
 929}
 930
 931static void gen8_ppgtt_clear_3lvl(struct i915_address_space *vm,
 932				  u64 start, u64 length)
 933{
 934	gen8_ppgtt_clear_pdp(vm, &i915_vm_to_ppgtt(vm)->pdp, start, length);
 935}
 936
 937static void gen8_ppgtt_set_pml4e(struct i915_pml4 *pml4,
 938				 struct i915_page_directory_pointer *pdp,
 939				 unsigned int pml4e)
 940{
 941	gen8_ppgtt_pml4e_t *vaddr;
 942
 943	pml4->pdps[pml4e] = pdp;
 944
 945	vaddr = kmap_atomic_px(pml4);
 946	vaddr[pml4e] = gen8_pml4e_encode(px_dma(pdp), I915_CACHE_LLC);
 947	kunmap_atomic(vaddr);
 948}
 949
 950/* Removes entries from a single pml4.
 951 * This is the top-level structure in 4-level page tables used on gen8+.
 952 * Empty entries are always scratch pml4e.
 953 */
 954static void gen8_ppgtt_clear_4lvl(struct i915_address_space *vm,
 955				  u64 start, u64 length)
 956{
 957	struct i915_hw_ppgtt *ppgtt = i915_vm_to_ppgtt(vm);
 958	struct i915_pml4 *pml4 = &ppgtt->pml4;
 959	struct i915_page_directory_pointer *pdp;
 960	unsigned int pml4e;
 961
 962	GEM_BUG_ON(!use_4lvl(vm));
 963
 964	gen8_for_each_pml4e(pdp, pml4, start, length, pml4e) {
 965		GEM_BUG_ON(pdp == vm->scratch_pdp);
 966
 967		if (!gen8_ppgtt_clear_pdp(vm, pdp, start, length))
 968			continue;
 969
 970		gen8_ppgtt_set_pml4e(pml4, vm->scratch_pdp, pml4e);
 971
 972		free_pdp(vm, pdp);
 973	}
 974}
 975
 976static inline struct sgt_dma {
 977	struct scatterlist *sg;
 978	dma_addr_t dma, max;
 979} sgt_dma(struct i915_vma *vma) {
 980	struct scatterlist *sg = vma->pages->sgl;
 981	dma_addr_t addr = sg_dma_address(sg);
 982	return (struct sgt_dma) { sg, addr, addr + sg->length };
 983}
 984
 985struct gen8_insert_pte {
 986	u16 pml4e;
 987	u16 pdpe;
 988	u16 pde;
 989	u16 pte;
 990};
 991
 992static __always_inline struct gen8_insert_pte gen8_insert_pte(u64 start)
 993{
 994	return (struct gen8_insert_pte) {
 995		 gen8_pml4e_index(start),
 996		 gen8_pdpe_index(start),
 997		 gen8_pde_index(start),
 998		 gen8_pte_index(start),
 999	};
1000}
1001
1002static __always_inline bool
1003gen8_ppgtt_insert_pte_entries(struct i915_hw_ppgtt *ppgtt,
1004			      struct i915_page_directory_pointer *pdp,
1005			      struct sgt_dma *iter,
1006			      struct gen8_insert_pte *idx,
1007			      enum i915_cache_level cache_level)
1008{
1009	struct i915_page_directory *pd;
1010	const gen8_pte_t pte_encode = gen8_pte_encode(0, cache_level);
1011	gen8_pte_t *vaddr;
1012	bool ret;
1013
1014	GEM_BUG_ON(idx->pdpe >= i915_pdpes_per_pdp(&ppgtt->base));
1015	pd = pdp->page_directory[idx->pdpe];
1016	vaddr = kmap_atomic_px(pd->page_table[idx->pde]);
1017	do {
1018		vaddr[idx->pte] = pte_encode | iter->dma;
1019
1020		iter->dma += PAGE_SIZE;
1021		if (iter->dma >= iter->max) {
1022			iter->sg = __sg_next(iter->sg);
1023			if (!iter->sg) {
1024				ret = false;
1025				break;
1026			}
1027
1028			iter->dma = sg_dma_address(iter->sg);
1029			iter->max = iter->dma + iter->sg->length;
1030		}
1031
1032		if (++idx->pte == GEN8_PTES) {
1033			idx->pte = 0;
1034
1035			if (++idx->pde == I915_PDES) {
1036				idx->pde = 0;
1037
1038				/* Limited by sg length for 3lvl */
1039				if (++idx->pdpe == GEN8_PML4ES_PER_PML4) {
1040					idx->pdpe = 0;
1041					ret = true;
1042					break;
1043				}
1044
1045				GEM_BUG_ON(idx->pdpe >= i915_pdpes_per_pdp(&ppgtt->base));
1046				pd = pdp->page_directory[idx->pdpe];
1047			}
1048
1049			kunmap_atomic(vaddr);
1050			vaddr = kmap_atomic_px(pd->page_table[idx->pde]);
1051		}
1052	} while (1);
1053	kunmap_atomic(vaddr);
1054
1055	return ret;
1056}
1057
1058static void gen8_ppgtt_insert_3lvl(struct i915_address_space *vm,
1059				   struct i915_vma *vma,
1060				   enum i915_cache_level cache_level,
1061				   u32 unused)
1062{
1063	struct i915_hw_ppgtt *ppgtt = i915_vm_to_ppgtt(vm);
1064	struct sgt_dma iter = sgt_dma(vma);
1065	struct gen8_insert_pte idx = gen8_insert_pte(vma->node.start);
1066
1067	gen8_ppgtt_insert_pte_entries(ppgtt, &ppgtt->pdp, &iter, &idx,
1068				      cache_level);
1069
1070	vma->page_sizes.gtt = I915_GTT_PAGE_SIZE;
1071}
1072
1073static void gen8_ppgtt_insert_huge_entries(struct i915_vma *vma,
1074					   struct i915_page_directory_pointer **pdps,
1075					   struct sgt_dma *iter,
1076					   enum i915_cache_level cache_level)
1077{
1078	const gen8_pte_t pte_encode = gen8_pte_encode(0, cache_level);
1079	u64 start = vma->node.start;
1080	dma_addr_t rem = iter->sg->length;
1081
1082	do {
1083		struct gen8_insert_pte idx = gen8_insert_pte(start);
1084		struct i915_page_directory_pointer *pdp = pdps[idx.pml4e];
1085		struct i915_page_directory *pd = pdp->page_directory[idx.pdpe];
1086		unsigned int page_size;
1087		bool maybe_64K = false;
1088		gen8_pte_t encode = pte_encode;
1089		gen8_pte_t *vaddr;
1090		u16 index, max;
1091
1092		if (vma->page_sizes.sg & I915_GTT_PAGE_SIZE_2M &&
1093		    IS_ALIGNED(iter->dma, I915_GTT_PAGE_SIZE_2M) &&
1094		    rem >= I915_GTT_PAGE_SIZE_2M && !idx.pte) {
1095			index = idx.pde;
1096			max = I915_PDES;
1097			page_size = I915_GTT_PAGE_SIZE_2M;
1098
1099			encode |= GEN8_PDE_PS_2M;
1100
1101			vaddr = kmap_atomic_px(pd);
1102		} else {
1103			struct i915_page_table *pt = pd->page_table[idx.pde];
1104
1105			index = idx.pte;
1106			max = GEN8_PTES;
1107			page_size = I915_GTT_PAGE_SIZE;
1108
1109			if (!index &&
1110			    vma->page_sizes.sg & I915_GTT_PAGE_SIZE_64K &&
1111			    IS_ALIGNED(iter->dma, I915_GTT_PAGE_SIZE_64K) &&
1112			    (IS_ALIGNED(rem, I915_GTT_PAGE_SIZE_64K) ||
1113			     rem >= (max - index) << PAGE_SHIFT))
1114				maybe_64K = true;
1115
1116			vaddr = kmap_atomic_px(pt);
1117		}
1118
1119		do {
1120			GEM_BUG_ON(iter->sg->length < page_size);
1121			vaddr[index++] = encode | iter->dma;
1122
1123			start += page_size;
1124			iter->dma += page_size;
1125			rem -= page_size;
1126			if (iter->dma >= iter->max) {
1127				iter->sg = __sg_next(iter->sg);
1128				if (!iter->sg)
1129					break;
1130
1131				rem = iter->sg->length;
1132				iter->dma = sg_dma_address(iter->sg);
1133				iter->max = iter->dma + rem;
1134
1135				if (maybe_64K && index < max &&
1136				    !(IS_ALIGNED(iter->dma, I915_GTT_PAGE_SIZE_64K) &&
1137				      (IS_ALIGNED(rem, I915_GTT_PAGE_SIZE_64K) ||
1138				       rem >= (max - index) << PAGE_SHIFT)))
1139					maybe_64K = false;
1140
1141				if (unlikely(!IS_ALIGNED(iter->dma, page_size)))
1142					break;
1143			}
1144		} while (rem >= page_size && index < max);
1145
1146		kunmap_atomic(vaddr);
1147
1148		/*
1149		 * Is it safe to mark the 2M block as 64K? -- Either we have
1150		 * filled whole page-table with 64K entries, or filled part of
1151		 * it and have reached the end of the sg table and we have
1152		 * enough padding.
1153		 */
1154		if (maybe_64K &&
1155		    (index == max ||
1156		     (i915_vm_has_scratch_64K(vma->vm) &&
1157		      !iter->sg && IS_ALIGNED(vma->node.start +
1158					      vma->node.size,
1159					      I915_GTT_PAGE_SIZE_2M)))) {
1160			vaddr = kmap_atomic_px(pd);
1161			vaddr[idx.pde] |= GEN8_PDE_IPS_64K;
1162			kunmap_atomic(vaddr);
1163			page_size = I915_GTT_PAGE_SIZE_64K;
1164		}
1165
1166		vma->page_sizes.gtt |= page_size;
1167	} while (iter->sg);
1168}
1169
1170static void gen8_ppgtt_insert_4lvl(struct i915_address_space *vm,
1171				   struct i915_vma *vma,
1172				   enum i915_cache_level cache_level,
1173				   u32 unused)
1174{
1175	struct i915_hw_ppgtt *ppgtt = i915_vm_to_ppgtt(vm);
1176	struct sgt_dma iter = sgt_dma(vma);
1177	struct i915_page_directory_pointer **pdps = ppgtt->pml4.pdps;
1178
1179	if (vma->page_sizes.sg > I915_GTT_PAGE_SIZE) {
1180		gen8_ppgtt_insert_huge_entries(vma, pdps, &iter, cache_level);
1181	} else {
1182		struct gen8_insert_pte idx = gen8_insert_pte(vma->node.start);
1183
1184		while (gen8_ppgtt_insert_pte_entries(ppgtt, pdps[idx.pml4e++],
1185						     &iter, &idx, cache_level))
1186			GEM_BUG_ON(idx.pml4e >= GEN8_PML4ES_PER_PML4);
1187
1188		vma->page_sizes.gtt = I915_GTT_PAGE_SIZE;
1189	}
1190}
1191
1192static void gen8_free_page_tables(struct i915_address_space *vm,
1193				  struct i915_page_directory *pd)
1194{
1195	int i;
1196
1197	if (!px_page(pd))
1198		return;
1199
1200	for (i = 0; i < I915_PDES; i++) {
1201		if (pd->page_table[i] != vm->scratch_pt)
1202			free_pt(vm, pd->page_table[i]);
1203	}
1204}
1205
1206static int gen8_init_scratch(struct i915_address_space *vm)
1207{
1208	int ret;
1209
1210	ret = setup_scratch_page(vm, I915_GFP_DMA);
1211	if (ret)
1212		return ret;
1213
1214	vm->scratch_pt = alloc_pt(vm);
1215	if (IS_ERR(vm->scratch_pt)) {
1216		ret = PTR_ERR(vm->scratch_pt);
1217		goto free_scratch_page;
1218	}
1219
1220	vm->scratch_pd = alloc_pd(vm);
1221	if (IS_ERR(vm->scratch_pd)) {
1222		ret = PTR_ERR(vm->scratch_pd);
1223		goto free_pt;
1224	}
1225
1226	if (use_4lvl(vm)) {
1227		vm->scratch_pdp = alloc_pdp(vm);
1228		if (IS_ERR(vm->scratch_pdp)) {
1229			ret = PTR_ERR(vm->scratch_pdp);
1230			goto free_pd;
1231		}
1232	}
1233
1234	gen8_initialize_pt(vm, vm->scratch_pt);
1235	gen8_initialize_pd(vm, vm->scratch_pd);
1236	if (use_4lvl(vm))
1237		gen8_initialize_pdp(vm, vm->scratch_pdp);
1238
1239	return 0;
1240
1241free_pd:
1242	free_pd(vm, vm->scratch_pd);
1243free_pt:
1244	free_pt(vm, vm->scratch_pt);
1245free_scratch_page:
1246	cleanup_scratch_page(vm);
1247
1248	return ret;
1249}
1250
1251static int gen8_ppgtt_notify_vgt(struct i915_hw_ppgtt *ppgtt, bool create)
1252{
1253	struct i915_address_space *vm = &ppgtt->base;
1254	struct drm_i915_private *dev_priv = vm->i915;
1255	enum vgt_g2v_type msg;
1256	int i;
1257
1258	if (use_4lvl(vm)) {
1259		const u64 daddr = px_dma(&ppgtt->pml4);
1260
1261		I915_WRITE(vgtif_reg(pdp[0].lo), lower_32_bits(daddr));
1262		I915_WRITE(vgtif_reg(pdp[0].hi), upper_32_bits(daddr));
1263
1264		msg = (create ? VGT_G2V_PPGTT_L4_PAGE_TABLE_CREATE :
1265				VGT_G2V_PPGTT_L4_PAGE_TABLE_DESTROY);
1266	} else {
1267		for (i = 0; i < GEN8_3LVL_PDPES; i++) {
1268			const u64 daddr = i915_page_dir_dma_addr(ppgtt, i);
1269
1270			I915_WRITE(vgtif_reg(pdp[i].lo), lower_32_bits(daddr));
1271			I915_WRITE(vgtif_reg(pdp[i].hi), upper_32_bits(daddr));
1272		}
1273
1274		msg = (create ? VGT_G2V_PPGTT_L3_PAGE_TABLE_CREATE :
1275				VGT_G2V_PPGTT_L3_PAGE_TABLE_DESTROY);
1276	}
1277
1278	I915_WRITE(vgtif_reg(g2v_notify), msg);
1279
1280	return 0;
1281}
1282
1283static void gen8_free_scratch(struct i915_address_space *vm)
1284{
1285	if (use_4lvl(vm))
1286		free_pdp(vm, vm->scratch_pdp);
1287	free_pd(vm, vm->scratch_pd);
1288	free_pt(vm, vm->scratch_pt);
1289	cleanup_scratch_page(vm);
1290}
1291
1292static void gen8_ppgtt_cleanup_3lvl(struct i915_address_space *vm,
1293				    struct i915_page_directory_pointer *pdp)
1294{
1295	const unsigned int pdpes = i915_pdpes_per_pdp(vm);
1296	int i;
1297
1298	for (i = 0; i < pdpes; i++) {
1299		if (pdp->page_directory[i] == vm->scratch_pd)
1300			continue;
1301
1302		gen8_free_page_tables(vm, pdp->page_directory[i]);
1303		free_pd(vm, pdp->page_directory[i]);
1304	}
1305
1306	free_pdp(vm, pdp);
1307}
1308
1309static void gen8_ppgtt_cleanup_4lvl(struct i915_hw_ppgtt *ppgtt)
1310{
1311	int i;
1312
1313	for (i = 0; i < GEN8_PML4ES_PER_PML4; i++) {
1314		if (ppgtt->pml4.pdps[i] == ppgtt->base.scratch_pdp)
1315			continue;
1316
1317		gen8_ppgtt_cleanup_3lvl(&ppgtt->base, ppgtt->pml4.pdps[i]);
1318	}
1319
1320	cleanup_px(&ppgtt->base, &ppgtt->pml4);
1321}
1322
1323static void gen8_ppgtt_cleanup(struct i915_address_space *vm)
1324{
1325	struct drm_i915_private *dev_priv = vm->i915;
1326	struct i915_hw_ppgtt *ppgtt = i915_vm_to_ppgtt(vm);
1327
1328	if (intel_vgpu_active(dev_priv))
1329		gen8_ppgtt_notify_vgt(ppgtt, false);
1330
1331	if (use_4lvl(vm))
1332		gen8_ppgtt_cleanup_4lvl(ppgtt);
1333	else
1334		gen8_ppgtt_cleanup_3lvl(&ppgtt->base, &ppgtt->pdp);
1335
1336	gen8_free_scratch(vm);
1337}
1338
1339static int gen8_ppgtt_alloc_pd(struct i915_address_space *vm,
1340			       struct i915_page_directory *pd,
1341			       u64 start, u64 length)
1342{
1343	struct i915_page_table *pt;
1344	u64 from = start;
1345	unsigned int pde;
1346
1347	gen8_for_each_pde(pt, pd, start, length, pde) {
1348		int count = gen8_pte_count(start, length);
1349
1350		if (pt == vm->scratch_pt) {
1351			pd->used_pdes++;
1352
1353			pt = alloc_pt(vm);
1354			if (IS_ERR(pt)) {
1355				pd->used_pdes--;
1356				goto unwind;
1357			}
1358
1359			if (count < GEN8_PTES || intel_vgpu_active(vm->i915))
1360				gen8_initialize_pt(vm, pt);
1361
1362			gen8_ppgtt_set_pde(vm, pd, pt, pde);
1363			GEM_BUG_ON(pd->used_pdes > I915_PDES);
1364		}
1365
1366		pt->used_ptes += count;
1367	}
1368	return 0;
1369
1370unwind:
1371	gen8_ppgtt_clear_pd(vm, pd, from, start - from);
1372	return -ENOMEM;
1373}
1374
1375static int gen8_ppgtt_alloc_pdp(struct i915_address_space *vm,
1376				struct i915_page_directory_pointer *pdp,
1377				u64 start, u64 length)
1378{
1379	struct i915_page_directory *pd;
1380	u64 from = start;
1381	unsigned int pdpe;
1382	int ret;
1383
1384	gen8_for_each_pdpe(pd, pdp, start, length, pdpe) {
1385		if (pd == vm->scratch_pd) {
1386			pdp->used_pdpes++;
1387
1388			pd = alloc_pd(vm);
1389			if (IS_ERR(pd)) {
1390				pdp->used_pdpes--;
1391				goto unwind;
1392			}
1393
1394			gen8_initialize_pd(vm, pd);
1395			gen8_ppgtt_set_pdpe(vm, pdp, pd, pdpe);
1396			GEM_BUG_ON(pdp->used_pdpes > i915_pdpes_per_pdp(vm));
1397
1398			mark_tlbs_dirty(i915_vm_to_ppgtt(vm));
1399		}
1400
1401		ret = gen8_ppgtt_alloc_pd(vm, pd, start, length);
1402		if (unlikely(ret))
1403			goto unwind_pd;
1404	}
1405
1406	return 0;
1407
1408unwind_pd:
1409	if (!pd->used_pdes) {
1410		gen8_ppgtt_set_pdpe(vm, pdp, vm->scratch_pd, pdpe);
1411		GEM_BUG_ON(!pdp->used_pdpes);
1412		pdp->used_pdpes--;
1413		free_pd(vm, pd);
1414	}
1415unwind:
1416	gen8_ppgtt_clear_pdp(vm, pdp, from, start - from);
1417	return -ENOMEM;
1418}
1419
1420static int gen8_ppgtt_alloc_3lvl(struct i915_address_space *vm,
1421				 u64 start, u64 length)
1422{
1423	return gen8_ppgtt_alloc_pdp(vm,
1424				    &i915_vm_to_ppgtt(vm)->pdp, start, length);
1425}
1426
1427static int gen8_ppgtt_alloc_4lvl(struct i915_address_space *vm,
1428				 u64 start, u64 length)
1429{
1430	struct i915_hw_ppgtt *ppgtt = i915_vm_to_ppgtt(vm);
1431	struct i915_pml4 *pml4 = &ppgtt->pml4;
1432	struct i915_page_directory_pointer *pdp;
1433	u64 from = start;
1434	u32 pml4e;
1435	int ret;
1436
1437	gen8_for_each_pml4e(pdp, pml4, start, length, pml4e) {
1438		if (pml4->pdps[pml4e] == vm->scratch_pdp) {
1439			pdp = alloc_pdp(vm);
1440			if (IS_ERR(pdp))
1441				goto unwind;
1442
1443			gen8_initialize_pdp(vm, pdp);
1444			gen8_ppgtt_set_pml4e(pml4, pdp, pml4e);
1445		}
1446
1447		ret = gen8_ppgtt_alloc_pdp(vm, pdp, start, length);
1448		if (unlikely(ret))
1449			goto unwind_pdp;
1450	}
1451
1452	return 0;
1453
1454unwind_pdp:
1455	if (!pdp->used_pdpes) {
1456		gen8_ppgtt_set_pml4e(pml4, vm->scratch_pdp, pml4e);
1457		free_pdp(vm, pdp);
1458	}
1459unwind:
1460	gen8_ppgtt_clear_4lvl(vm, from, start - from);
1461	return -ENOMEM;
1462}
1463
1464static void gen8_dump_pdp(struct i915_hw_ppgtt *ppgtt,
1465			  struct i915_page_directory_pointer *pdp,
1466			  u64 start, u64 length,
1467			  gen8_pte_t scratch_pte,
1468			  struct seq_file *m)
1469{
1470	struct i915_address_space *vm = &ppgtt->base;
1471	struct i915_page_directory *pd;
1472	u32 pdpe;
1473
1474	gen8_for_each_pdpe(pd, pdp, start, length, pdpe) {
1475		struct i915_page_table *pt;
1476		u64 pd_len = length;
1477		u64 pd_start = start;
1478		u32 pde;
1479
1480		if (pdp->page_directory[pdpe] == ppgtt->base.scratch_pd)
1481			continue;
1482
1483		seq_printf(m, "\tPDPE #%d\n", pdpe);
1484		gen8_for_each_pde(pt, pd, pd_start, pd_len, pde) {
1485			u32 pte;
1486			gen8_pte_t *pt_vaddr;
1487
1488			if (pd->page_table[pde] == ppgtt->base.scratch_pt)
1489				continue;
1490
1491			pt_vaddr = kmap_atomic_px(pt);
1492			for (pte = 0; pte < GEN8_PTES; pte += 4) {
1493				u64 va = (pdpe << GEN8_PDPE_SHIFT |
1494					  pde << GEN8_PDE_SHIFT |
1495					  pte << GEN8_PTE_SHIFT);
1496				int i;
1497				bool found = false;
1498
1499				for (i = 0; i < 4; i++)
1500					if (pt_vaddr[pte + i] != scratch_pte)
1501						found = true;
1502				if (!found)
1503					continue;
1504
1505				seq_printf(m, "\t\t0x%llx [%03d,%03d,%04d]: =", va, pdpe, pde, pte);
1506				for (i = 0; i < 4; i++) {
1507					if (pt_vaddr[pte + i] != scratch_pte)
1508						seq_printf(m, " %llx", pt_vaddr[pte + i]);
1509					else
1510						seq_puts(m, "  SCRATCH ");
1511				}
1512				seq_puts(m, "\n");
1513			}
1514			kunmap_atomic(pt_vaddr);
1515		}
1516	}
1517}
1518
1519static void gen8_dump_ppgtt(struct i915_hw_ppgtt *ppgtt, struct seq_file *m)
1520{
1521	struct i915_address_space *vm = &ppgtt->base;
1522	const gen8_pte_t scratch_pte =
1523		gen8_pte_encode(vm->scratch_page.daddr, I915_CACHE_LLC);
1524	u64 start = 0, length = ppgtt->base.total;
1525
1526	if (use_4lvl(vm)) {
1527		u64 pml4e;
1528		struct i915_pml4 *pml4 = &ppgtt->pml4;
1529		struct i915_page_directory_pointer *pdp;
1530
1531		gen8_for_each_pml4e(pdp, pml4, start, length, pml4e) {
1532			if (pml4->pdps[pml4e] == ppgtt->base.scratch_pdp)
1533				continue;
1534
1535			seq_printf(m, "    PML4E #%llu\n", pml4e);
1536			gen8_dump_pdp(ppgtt, pdp, start, length, scratch_pte, m);
1537		}
1538	} else {
1539		gen8_dump_pdp(ppgtt, &ppgtt->pdp, start, length, scratch_pte, m);
1540	}
1541}
1542
1543static int gen8_preallocate_top_level_pdp(struct i915_hw_ppgtt *ppgtt)
1544{
1545	struct i915_address_space *vm = &ppgtt->base;
1546	struct i915_page_directory_pointer *pdp = &ppgtt->pdp;
1547	struct i915_page_directory *pd;
1548	u64 start = 0, length = ppgtt->base.total;
1549	u64 from = start;
1550	unsigned int pdpe;
1551
1552	gen8_for_each_pdpe(pd, pdp, start, length, pdpe) {
1553		pd = alloc_pd(vm);
1554		if (IS_ERR(pd))
1555			goto unwind;
1556
1557		gen8_initialize_pd(vm, pd);
1558		gen8_ppgtt_set_pdpe(vm, pdp, pd, pdpe);
1559		pdp->used_pdpes++;
1560	}
1561
1562	pdp->used_pdpes++; /* never remove */
1563	return 0;
1564
1565unwind:
1566	start -= from;
1567	gen8_for_each_pdpe(pd, pdp, from, start, pdpe) {
1568		gen8_ppgtt_set_pdpe(vm, pdp, vm->scratch_pd, pdpe);
1569		free_pd(vm, pd);
1570	}
1571	pdp->used_pdpes = 0;
1572	return -ENOMEM;
1573}
1574
1575/*
1576 * GEN8 legacy ppgtt programming is accomplished through a max 4 PDP registers
1577 * with a net effect resembling a 2-level page table in normal x86 terms. Each
1578 * PDP represents 1GB of memory 4 * 512 * 512 * 4096 = 4GB legacy 32b address
1579 * space.
1580 *
1581 */
1582static int gen8_ppgtt_init(struct i915_hw_ppgtt *ppgtt)
1583{
1584	struct i915_address_space *vm = &ppgtt->base;
1585	struct drm_i915_private *dev_priv = vm->i915;
1586	int ret;
1587
1588	ppgtt->base.total = USES_FULL_48BIT_PPGTT(dev_priv) ?
1589		1ULL << 48 :
1590		1ULL << 32;
1591
1592	/* There are only few exceptions for gen >=6. chv and bxt.
1593	 * And we are not sure about the latter so play safe for now.
1594	 */
1595	if (IS_CHERRYVIEW(dev_priv) || IS_BROXTON(dev_priv))
1596		ppgtt->base.pt_kmap_wc = true;
1597
1598	ret = gen8_init_scratch(&ppgtt->base);
1599	if (ret) {
1600		ppgtt->base.total = 0;
1601		return ret;
1602	}
1603
1604	if (use_4lvl(vm)) {
1605		ret = setup_px(&ppgtt->base, &ppgtt->pml4);
1606		if (ret)
1607			goto free_scratch;
1608
1609		gen8_initialize_pml4(&ppgtt->base, &ppgtt->pml4);
1610
1611		ppgtt->switch_mm = gen8_mm_switch_4lvl;
1612		ppgtt->base.allocate_va_range = gen8_ppgtt_alloc_4lvl;
1613		ppgtt->base.insert_entries = gen8_ppgtt_insert_4lvl;
1614		ppgtt->base.clear_range = gen8_ppgtt_clear_4lvl;
1615	} else {
1616		ret = __pdp_init(&ppgtt->base, &ppgtt->pdp);
1617		if (ret)
1618			goto free_scratch;
1619
1620		if (intel_vgpu_active(dev_priv)) {
1621			ret = gen8_preallocate_top_level_pdp(ppgtt);
1622			if (ret) {
1623				__pdp_fini(&ppgtt->pdp);
1624				goto free_scratch;
1625			}
1626		}
1627
1628		ppgtt->switch_mm = gen8_mm_switch_3lvl;
1629		ppgtt->base.allocate_va_range = gen8_ppgtt_alloc_3lvl;
1630		ppgtt->base.insert_entries = gen8_ppgtt_insert_3lvl;
1631		ppgtt->base.clear_range = gen8_ppgtt_clear_3lvl;
1632	}
1633
1634	if (intel_vgpu_active(dev_priv))
1635		gen8_ppgtt_notify_vgt(ppgtt, true);
1636
1637	ppgtt->base.cleanup = gen8_ppgtt_cleanup;
1638	ppgtt->base.unbind_vma = ppgtt_unbind_vma;
1639	ppgtt->base.bind_vma = ppgtt_bind_vma;
1640	ppgtt->base.set_pages = ppgtt_set_pages;
1641	ppgtt->base.clear_pages = clear_pages;
1642	ppgtt->debug_dump = gen8_dump_ppgtt;
1643
1644	return 0;
1645
1646free_scratch:
1647	gen8_free_scratch(&ppgtt->base);
1648	return ret;
1649}
1650
1651static void gen6_dump_ppgtt(struct i915_hw_ppgtt *ppgtt, struct seq_file *m)
1652{
1653	struct i915_address_space *vm = &ppgtt->base;
1654	struct i915_page_table *unused;
1655	gen6_pte_t scratch_pte;
1656	u32 pd_entry, pte, pde;
1657	u32 start = 0, length = ppgtt->base.total;
1658
1659	scratch_pte = vm->pte_encode(vm->scratch_page.daddr,
1660				     I915_CACHE_LLC, 0);
1661
1662	gen6_for_each_pde(unused, &ppgtt->pd, start, length, pde) {
1663		u32 expected;
1664		gen6_pte_t *pt_vaddr;
1665		const dma_addr_t pt_addr = px_dma(ppgtt->pd.page_table[pde]);
1666		pd_entry = readl(ppgtt->pd_addr + pde);
1667		expected = (GEN6_PDE_ADDR_ENCODE(pt_addr) | GEN6_PDE_VALID);
1668
1669		if (pd_entry != expected)
1670			seq_printf(m, "\tPDE #%d mismatch: Actual PDE: %x Expected PDE: %x\n",
1671				   pde,
1672				   pd_entry,
1673				   expected);
1674		seq_printf(m, "\tPDE: %x\n", pd_entry);
1675
1676		pt_vaddr = kmap_atomic_px(ppgtt->pd.page_table[pde]);
1677
1678		for (pte = 0; pte < GEN6_PTES; pte+=4) {
1679			unsigned long va =
1680				(pde * PAGE_SIZE * GEN6_PTES) +
1681				(pte * PAGE_SIZE);
1682			int i;
1683			bool found = false;
1684			for (i = 0; i < 4; i++)
1685				if (pt_vaddr[pte + i] != scratch_pte)
1686					found = true;
1687			if (!found)
1688				continue;
1689
1690			seq_printf(m, "\t\t0x%lx [%03d,%04d]: =", va, pde, pte);
1691			for (i = 0; i < 4; i++) {
1692				if (pt_vaddr[pte + i] != scratch_pte)
1693					seq_printf(m, " %08x", pt_vaddr[pte + i]);
1694				else
1695					seq_puts(m, "  SCRATCH ");
1696			}
1697			seq_puts(m, "\n");
1698		}
1699		kunmap_atomic(pt_vaddr);
1700	}
1701}
1702
1703/* Write pde (index) from the page directory @pd to the page table @pt */
1704static inline void gen6_write_pde(const struct i915_hw_ppgtt *ppgtt,
1705				  const unsigned int pde,
1706				  const struct i915_page_table *pt)
1707{
1708	/* Caller needs to make sure the write completes if necessary */
1709	writel_relaxed(GEN6_PDE_ADDR_ENCODE(px_dma(pt)) | GEN6_PDE_VALID,
1710		       ppgtt->pd_addr + pde);
1711}
1712
1713/* Write all the page tables found in the ppgtt structure to incrementing page
1714 * directories. */
1715static void gen6_write_page_range(struct i915_hw_ppgtt *ppgtt,
1716				  u32 start, u32 length)
1717{
1718	struct i915_page_table *pt;
1719	unsigned int pde;
1720
1721	gen6_for_each_pde(pt, &ppgtt->pd, start, length, pde)
1722		gen6_write_pde(ppgtt, pde, pt);
1723
1724	mark_tlbs_dirty(ppgtt);
1725	wmb();
1726}
1727
1728static inline u32 get_pd_offset(struct i915_hw_ppgtt *ppgtt)
1729{
1730	GEM_BUG_ON(ppgtt->pd.base.ggtt_offset & 0x3f);
1731	return ppgtt->pd.base.ggtt_offset << 10;
1732}
1733
1734static int hsw_mm_switch(struct i915_hw_ppgtt *ppgtt,
1735			 struct i915_request *rq)
1736{
1737	struct intel_engine_cs *engine = rq->engine;
1738	u32 *cs;
1739
1740	/* NB: TLBs must be flushed and invalidated before a switch */
1741	cs = intel_ring_begin(rq, 6);
1742	if (IS_ERR(cs))
1743		return PTR_ERR(cs);
1744
1745	*cs++ = MI_LOAD_REGISTER_IMM(2);
1746	*cs++ = i915_mmio_reg_offset(RING_PP_DIR_DCLV(engine));
1747	*cs++ = PP_DIR_DCLV_2G;
1748	*cs++ = i915_mmio_reg_offset(RING_PP_DIR_BASE(engine));
1749	*cs++ = get_pd_offset(ppgtt);
1750	*cs++ = MI_NOOP;
1751	intel_ring_advance(rq, cs);
1752
1753	return 0;
1754}
1755
1756static int gen7_mm_switch(struct i915_hw_ppgtt *ppgtt,
1757			  struct i915_request *rq)
1758{
1759	struct intel_engine_cs *engine = rq->engine;
1760	u32 *cs;
1761
1762	/* NB: TLBs must be flushed and invalidated before a switch */
1763	cs = intel_ring_begin(rq, 6);
1764	if (IS_ERR(cs))
1765		return PTR_ERR(cs);
1766
1767	*cs++ = MI_LOAD_REGISTER_IMM(2);
1768	*cs++ = i915_mmio_reg_offset(RING_PP_DIR_DCLV(engine));
1769	*cs++ = PP_DIR_DCLV_2G;
1770	*cs++ = i915_mmio_reg_offset(RING_PP_DIR_BASE(engine));
1771	*cs++ = get_pd_offset(ppgtt);
1772	*cs++ = MI_NOOP;
1773	intel_ring_advance(rq, cs);
1774
1775	return 0;
1776}
1777
1778static int gen6_mm_switch(struct i915_hw_ppgtt *ppgtt,
1779			  struct i915_request *rq)
1780{
1781	struct intel_engine_cs *engine = rq->engine;
1782	struct drm_i915_private *dev_priv = rq->i915;
1783
1784	I915_WRITE(RING_PP_DIR_DCLV(engine), PP_DIR_DCLV_2G);
1785	I915_WRITE(RING_PP_DIR_BASE(engine), get_pd_offset(ppgtt));
1786	return 0;
1787}
1788
1789static void gen8_ppgtt_enable(struct drm_i915_private *dev_priv)
1790{
1791	struct intel_engine_cs *engine;
1792	enum intel_engine_id id;
1793
1794	for_each_engine(engine, dev_priv, id) {
1795		u32 four_level = USES_FULL_48BIT_PPGTT(dev_priv) ?
1796				 GEN8_GFX_PPGTT_48B : 0;
1797		I915_WRITE(RING_MODE_GEN7(engine),
1798			   _MASKED_BIT_ENABLE(GFX_PPGTT_ENABLE | four_level));
1799	}
1800}
1801
1802static void gen7_ppgtt_enable(struct drm_i915_private *dev_priv)
1803{
1804	struct intel_engine_cs *engine;
1805	u32 ecochk, ecobits;
1806	enum intel_engine_id id;
1807
1808	ecobits = I915_READ(GAC_ECO_BITS);
1809	I915_WRITE(GAC_ECO_BITS, ecobits | ECOBITS_PPGTT_CACHE64B);
1810
1811	ecochk = I915_READ(GAM_ECOCHK);
1812	if (IS_HASWELL(dev_priv)) {
1813		ecochk |= ECOCHK_PPGTT_WB_HSW;
1814	} else {
1815		ecochk |= ECOCHK_PPGTT_LLC_IVB;
1816		ecochk &= ~ECOCHK_PPGTT_GFDT_IVB;
1817	}
1818	I915_WRITE(GAM_ECOCHK, ecochk);
1819
1820	for_each_engine(engine, dev_priv, id) {
1821		/* GFX_MODE is per-ring on gen7+ */
1822		I915_WRITE(RING_MODE_GEN7(engine),
1823			   _MASKED_BIT_ENABLE(GFX_PPGTT_ENABLE));
1824	}
1825}
1826
1827static void gen6_ppgtt_enable(struct drm_i915_private *dev_priv)
1828{
1829	u32 ecochk, gab_ctl, ecobits;
1830
1831	ecobits = I915_READ(GAC_ECO_BITS);
1832	I915_WRITE(GAC_ECO_BITS, ecobits | ECOBITS_SNB_BIT |
1833		   ECOBITS_PPGTT_CACHE64B);
1834
1835	gab_ctl = I915_READ(GAB_CTL);
1836	I915_WRITE(GAB_CTL, gab_ctl | GAB_CTL_CONT_AFTER_PAGEFAULT);
1837
1838	ecochk = I915_READ(GAM_ECOCHK);
1839	I915_WRITE(GAM_ECOCHK, ecochk | ECOCHK_SNB_BIT | ECOCHK_PPGTT_CACHE64B);
1840
1841	I915_WRITE(GFX_MODE, _MASKED_BIT_ENABLE(GFX_PPGTT_ENABLE));
1842}
1843
1844/* PPGTT support for Sandybdrige/Gen6 and later */
1845static void gen6_ppgtt_clear_range(struct i915_address_space *vm,
1846				   u64 start, u64 length)
1847{
1848	struct i915_hw_ppgtt *ppgtt = i915_vm_to_ppgtt(vm);
1849	unsigned int first_entry = start >> PAGE_SHIFT;
1850	unsigned int pde = first_entry / GEN6_PTES;
1851	unsigned int pte = first_entry % GEN6_PTES;
1852	unsigned int num_entries = length >> PAGE_SHIFT;
1853	gen6_pte_t scratch_pte =
1854		vm->pte_encode(vm->scratch_page.daddr, I915_CACHE_LLC, 0);
1855
1856	while (num_entries) {
1857		struct i915_page_table *pt = ppgtt->pd.page_table[pde++];
1858		unsigned int end = min(pte + num_entries, GEN6_PTES);
1859		gen6_pte_t *vaddr;
1860
1861		num_entries -= end - pte;
1862
1863		/* Note that the hw doesn't support removing PDE on the fly
1864		 * (they are cached inside the context with no means to
1865		 * invalidate the cache), so we can only reset the PTE
1866		 * entries back to scratch.
1867		 */
1868
1869		vaddr = kmap_atomic_px(pt);
1870		do {
1871			vaddr[pte++] = scratch_pte;
1872		} while (pte < end);
1873		kunmap_atomic(vaddr);
1874
1875		pte = 0;
1876	}
1877}
1878
1879static void gen6_ppgtt_insert_entries(struct i915_address_space *vm,
1880				      struct i915_vma *vma,
1881				      enum i915_cache_level cache_level,
1882				      u32 flags)
1883{
1884	struct i915_hw_ppgtt *ppgtt = i915_vm_to_ppgtt(vm);
1885	unsigned first_entry = vma->node.start >> PAGE_SHIFT;
1886	unsigned act_pt = first_entry / GEN6_PTES;
1887	unsigned act_pte = first_entry % GEN6_PTES;
1888	const u32 pte_encode = vm->pte_encode(0, cache_level, flags);
1889	struct sgt_dma iter = sgt_dma(vma);
1890	gen6_pte_t *vaddr;
1891
1892	vaddr = kmap_atomic_px(ppgtt->pd.page_table[act_pt]);
1893	do {
1894		vaddr[act_pte] = pte_encode | GEN6_PTE_ADDR_ENCODE(iter.dma);
1895
1896		iter.dma += PAGE_SIZE;
1897		if (iter.dma == iter.max) {
1898			iter.sg = __sg_next(iter.sg);
1899			if (!iter.sg)
1900				break;
1901
1902			iter.dma = sg_dma_address(iter.sg);
1903			iter.max = iter.dma + iter.sg->length;
1904		}
1905
1906		if (++act_pte == GEN6_PTES) {
1907			kunmap_atomic(vaddr);
1908			vaddr = kmap_atomic_px(ppgtt->pd.page_table[++act_pt]);
1909			act_pte = 0;
1910		}
1911	} while (1);
1912	kunmap_atomic(vaddr);
1913
1914	vma->page_sizes.gtt = I915_GTT_PAGE_SIZE;
1915}
1916
1917static int gen6_alloc_va_range(struct i915_address_space *vm,
1918			       u64 start, u64 length)
1919{
1920	struct i915_hw_ppgtt *ppgtt = i915_vm_to_ppgtt(vm);
1921	struct i915_page_table *pt;
1922	u64 from = start;
1923	unsigned int pde;
1924	bool flush = false;
1925
1926	gen6_for_each_pde(pt, &ppgtt->pd, start, length, pde) {
1927		if (pt == vm->scratch_pt) {
1928			pt = alloc_pt(vm);
1929			if (IS_ERR(pt))
1930				goto unwind_out;
1931
1932			gen6_initialize_pt(vm, pt);
1933			ppgtt->pd.page_table[pde] = pt;
1934			gen6_write_pde(ppgtt, pde, pt);
1935			flush = true;
1936		}
1937	}
1938
1939	if (flush) {
1940		mark_tlbs_dirty(ppgtt);
1941		wmb();
1942	}
1943
1944	return 0;
1945
1946unwind_out:
1947	gen6_ppgtt_clear_range(vm, from, start);
1948	return -ENOMEM;
1949}
1950
1951static int gen6_init_scratch(struct i915_address_space *vm)
1952{
1953	int ret;
1954
1955	ret = setup_scratch_page(vm, I915_GFP_DMA);
1956	if (ret)
1957		return ret;
1958
1959	vm->scratch_pt = alloc_pt(vm);
1960	if (IS_ERR(vm->scratch_pt)) {
1961		cleanup_scratch_page(vm);
1962		return PTR_ERR(vm->scratch_pt);
1963	}
1964
1965	gen6_initialize_pt(vm, vm->scratch_pt);
1966
1967	return 0;
1968}
1969
1970static void gen6_free_scratch(struct i915_address_space *vm)
1971{
1972	free_pt(vm, vm->scratch_pt);
1973	cleanup_scratch_page(vm);
1974}
1975
1976static void gen6_ppgtt_cleanup(struct i915_address_space *vm)
1977{
1978	struct i915_hw_ppgtt *ppgtt = i915_vm_to_ppgtt(vm);
1979	struct i915_page_directory *pd = &ppgtt->pd;
1980	struct i915_page_table *pt;
1981	u32 pde;
1982
1983	drm_mm_remove_node(&ppgtt->node);
1984
1985	gen6_for_all_pdes(pt, pd, pde)
1986		if (pt != vm->scratch_pt)
1987			free_pt(vm, pt);
1988
1989	gen6_free_scratch(vm);
1990}
1991
1992static int gen6_ppgtt_allocate_page_directories(struct i915_hw_ppgtt *ppgtt)
1993{
1994	struct i915_address_space *vm = &ppgtt->base;
1995	struct drm_i915_private *dev_priv = ppgtt->base.i915;
1996	struct i915_ggtt *ggtt = &dev_priv->ggtt;
1997	int ret;
1998
1999	/* PPGTT PDEs reside in the GGTT and consists of 512 entries. The
2000	 * allocator works in address space sizes, so it's multiplied by page
2001	 * size. We allocate at the top of the GTT to avoid fragmentation.
2002	 */
2003	BUG_ON(!drm_mm_initialized(&ggtt->base.mm));
2004
2005	ret = gen6_init_scratch(vm);
2006	if (ret)
2007		return ret;
2008
2009	ret = i915_gem_gtt_insert(&ggtt->base, &ppgtt->node,
2010				  GEN6_PD_SIZE, GEN6_PD_ALIGN,
2011				  I915_COLOR_UNEVICTABLE,
2012				  0, ggtt->base.total,
2013				  PIN_HIGH);
2014	if (ret)
2015		goto err_out;
2016
2017	if (ppgtt->node.start < ggtt->mappable_end)
2018		DRM_DEBUG("Forced to use aperture for PDEs\n");
2019
2020	ppgtt->pd.base.ggtt_offset =
2021		ppgtt->node.start / PAGE_SIZE * sizeof(gen6_pte_t);
2022
2023	ppgtt->pd_addr = (gen6_pte_t __iomem *)ggtt->gsm +
2024		ppgtt->pd.base.ggtt_offset / sizeof(gen6_pte_t);
2025
2026	return 0;
2027
2028err_out:
2029	gen6_free_scratch(vm);
2030	return ret;
2031}
2032
2033static int gen6_ppgtt_alloc(struct i915_hw_ppgtt *ppgtt)
2034{
2035	return gen6_ppgtt_allocate_page_directories(ppgtt);
2036}
2037
2038static void gen6_scratch_va_range(struct i915_hw_ppgtt *ppgtt,
2039				  u64 start, u64 length)
2040{
2041	struct i915_page_table *unused;
2042	u32 pde;
2043
2044	gen6_for_each_pde(unused, &ppgtt->pd, start, length, pde)
2045		ppgtt->pd.page_table[pde] = ppgtt->base.scratch_pt;
2046}
2047
2048static int gen6_ppgtt_init(struct i915_hw_ppgtt *ppgtt)
2049{
2050	struct drm_i915_private *dev_priv = ppgtt->base.i915;
2051	struct i915_ggtt *ggtt = &dev_priv->ggtt;
2052	int ret;
2053
2054	ppgtt->base.pte_encode = ggtt->base.pte_encode;
2055	if (intel_vgpu_active(dev_priv) || IS_GEN6(dev_priv))
2056		ppgtt->switch_mm = gen6_mm_switch;
2057	else if (IS_HASWELL(dev_priv))
2058		ppgtt->switch_mm = hsw_mm_switch;
2059	else if (IS_GEN7(dev_priv))
2060		ppgtt->switch_mm = gen7_mm_switch;
2061	else
2062		BUG();
2063
2064	ret = gen6_ppgtt_alloc(ppgtt);
2065	if (ret)
2066		return ret;
2067
2068	ppgtt->base.total = I915_PDES * GEN6_PTES * PAGE_SIZE;
2069
2070	gen6_scratch_va_range(ppgtt, 0, ppgtt->base.total);
2071	gen6_write_page_range(ppgtt, 0, ppgtt->base.total);
2072
2073	ret = gen6_alloc_va_range(&ppgtt->base, 0, ppgtt->base.total);
2074	if (ret) {
2075		gen6_ppgtt_cleanup(&ppgtt->base);
2076		return ret;
2077	}
2078
2079	ppgtt->base.clear_range = gen6_ppgtt_clear_range;
2080	ppgtt->base.insert_entries = gen6_ppgtt_insert_entries;
2081	ppgtt->base.unbind_vma = ppgtt_unbind_vma;
2082	ppgtt->base.bind_vma = ppgtt_bind_vma;
2083	ppgtt->base.set_pages = ppgtt_set_pages;
2084	ppgtt->base.clear_pages = clear_pages;
2085	ppgtt->base.cleanup = gen6_ppgtt_cleanup;
2086	ppgtt->debug_dump = gen6_dump_ppgtt;
2087
2088	DRM_DEBUG_DRIVER("Allocated pde space (%lldM) at GTT entry: %llx\n",
2089			 ppgtt->node.size >> 20,
2090			 ppgtt->node.start / PAGE_SIZE);
2091
2092	DRM_DEBUG_DRIVER("Adding PPGTT at offset %x\n",
2093			 ppgtt->pd.base.ggtt_offset << 10);
2094
2095	return 0;
2096}
2097
2098static int __hw_ppgtt_init(struct i915_hw_ppgtt *ppgtt,
2099			   struct drm_i915_private *dev_priv)
2100{
2101	ppgtt->base.i915 = dev_priv;
2102	ppgtt->base.dma = &dev_priv->drm.pdev->dev;
2103
2104	if (INTEL_GEN(dev_priv) < 8)
2105		return gen6_ppgtt_init(ppgtt);
2106	else
2107		return gen8_ppgtt_init(ppgtt);
2108}
2109
2110static void i915_address_space_init(struct i915_address_space *vm,
2111				    struct drm_i915_private *dev_priv,
2112				    const char *name)
2113{
2114	i915_gem_timeline_init(dev_priv, &vm->timeline, name);
2115
2116	drm_mm_init(&vm->mm, 0, vm->total);
2117	vm->mm.head_node.color = I915_COLOR_UNEVICTABLE;
2118
2119	INIT_LIST_HEAD(&vm->active_list);
2120	INIT_LIST_HEAD(&vm->inactive_list);
2121	INIT_LIST_HEAD(&vm->unbound_list);
2122
2123	list_add_tail(&vm->global_link, &dev_priv->vm_list);
2124	pagevec_init(&vm->free_pages);
 
 
 
 
 
 
 
2125}
2126
2127static void i915_address_space_fini(struct i915_address_space *vm)
2128{
2129	if (pagevec_count(&vm->free_pages))
2130		vm_free_pages_release(vm, true);
2131
2132	i915_gem_timeline_fini(&vm->timeline);
2133	drm_mm_takedown(&vm->mm);
2134	list_del(&vm->global_link);
2135}
2136
2137static void gtt_write_workarounds(struct drm_i915_private *dev_priv)
2138{
2139	/* This function is for gtt related workarounds. This function is
2140	 * called on driver load and after a GPU reset, so you can place
2141	 * workarounds here even if they get overwritten by GPU reset.
2142	 */
2143	/* WaIncreaseDefaultTLBEntries:chv,bdw,skl,bxt,kbl,glk,cfl,cnl */
2144	if (IS_BROADWELL(dev_priv))
2145		I915_WRITE(GEN8_L3_LRA_1_GPGPU, GEN8_L3_LRA_1_GPGPU_DEFAULT_VALUE_BDW);
2146	else if (IS_CHERRYVIEW(dev_priv))
2147		I915_WRITE(GEN8_L3_LRA_1_GPGPU, GEN8_L3_LRA_1_GPGPU_DEFAULT_VALUE_CHV);
2148	else if (IS_GEN9_BC(dev_priv) || IS_GEN10(dev_priv))
2149		I915_WRITE(GEN8_L3_LRA_1_GPGPU, GEN9_L3_LRA_1_GPGPU_DEFAULT_VALUE_SKL);
2150	else if (IS_GEN9_LP(dev_priv))
2151		I915_WRITE(GEN8_L3_LRA_1_GPGPU, GEN9_L3_LRA_1_GPGPU_DEFAULT_VALUE_BXT);
2152
2153	/*
2154	 * To support 64K PTEs we need to first enable the use of the
2155	 * Intermediate-Page-Size(IPS) bit of the PDE field via some magical
2156	 * mmio, otherwise the page-walker will simply ignore the IPS bit. This
2157	 * shouldn't be needed after GEN10.
2158	 *
2159	 * 64K pages were first introduced from BDW+, although technically they
2160	 * only *work* from gen9+. For pre-BDW we instead have the option for
2161	 * 32K pages, but we don't currently have any support for it in our
2162	 * driver.
2163	 */
2164	if (HAS_PAGE_SIZES(dev_priv, I915_GTT_PAGE_SIZE_64K) &&
2165	    INTEL_GEN(dev_priv) <= 10)
2166		I915_WRITE(GEN8_GAMW_ECO_DEV_RW_IA,
2167			   I915_READ(GEN8_GAMW_ECO_DEV_RW_IA) |
2168			   GAMW_ECO_ENABLE_64K_IPS_FIELD);
2169}
2170
2171int i915_ppgtt_init_hw(struct drm_i915_private *dev_priv)
2172{
2173	gtt_write_workarounds(dev_priv);
2174
2175	/* In the case of execlists, PPGTT is enabled by the context descriptor
2176	 * and the PDPs are contained within the context itself.  We don't
2177	 * need to do anything here. */
2178	if (HAS_LOGICAL_RING_CONTEXTS(dev_priv))
2179		return 0;
2180
2181	if (!USES_PPGTT(dev_priv))
2182		return 0;
2183
2184	if (IS_GEN6(dev_priv))
2185		gen6_ppgtt_enable(dev_priv);
2186	else if (IS_GEN7(dev_priv))
2187		gen7_ppgtt_enable(dev_priv);
2188	else if (INTEL_GEN(dev_priv) >= 8)
2189		gen8_ppgtt_enable(dev_priv);
2190	else
2191		MISSING_CASE(INTEL_GEN(dev_priv));
2192
2193	return 0;
2194}
2195
2196struct i915_hw_ppgtt *
2197i915_ppgtt_create(struct drm_i915_private *dev_priv,
2198		  struct drm_i915_file_private *fpriv,
2199		  const char *name)
2200{
2201	struct i915_hw_ppgtt *ppgtt;
2202	int ret;
2203
2204	ppgtt = kzalloc(sizeof(*ppgtt), GFP_KERNEL);
2205	if (!ppgtt)
2206		return ERR_PTR(-ENOMEM);
2207
2208	ret = __hw_ppgtt_init(ppgtt, dev_priv);
2209	if (ret) {
2210		kfree(ppgtt);
2211		return ERR_PTR(ret);
2212	}
2213
2214	kref_init(&ppgtt->ref);
2215	i915_address_space_init(&ppgtt->base, dev_priv, name);
2216	ppgtt->base.file = fpriv;
2217
2218	trace_i915_ppgtt_create(&ppgtt->base);
2219
2220	return ppgtt;
2221}
2222
2223void i915_ppgtt_close(struct i915_address_space *vm)
2224{
2225	struct list_head *phases[] = {
2226		&vm->active_list,
2227		&vm->inactive_list,
2228		&vm->unbound_list,
2229		NULL,
2230	}, **phase;
2231
2232	GEM_BUG_ON(vm->closed);
2233	vm->closed = true;
2234
2235	for (phase = phases; *phase; phase++) {
2236		struct i915_vma *vma, *vn;
2237
2238		list_for_each_entry_safe(vma, vn, *phase, vm_link)
2239			if (!i915_vma_is_closed(vma))
2240				i915_vma_close(vma);
2241	}
2242}
2243
2244void i915_ppgtt_release(struct kref *kref)
2245{
2246	struct i915_hw_ppgtt *ppgtt =
2247		container_of(kref, struct i915_hw_ppgtt, ref);
2248
2249	trace_i915_ppgtt_release(&ppgtt->base);
2250
2251	/* vmas should already be unbound and destroyed */
2252	GEM_BUG_ON(!list_empty(&ppgtt->base.active_list));
2253	GEM_BUG_ON(!list_empty(&ppgtt->base.inactive_list));
2254	GEM_BUG_ON(!list_empty(&ppgtt->base.unbound_list));
2255
2256	ppgtt->base.cleanup(&ppgtt->base);
2257	i915_address_space_fini(&ppgtt->base);
2258	kfree(ppgtt);
2259}
2260
2261/* Certain Gen5 chipsets require require idling the GPU before
2262 * unmapping anything from the GTT when VT-d is enabled.
2263 */
2264static bool needs_idle_maps(struct drm_i915_private *dev_priv)
2265{
2266	/* Query intel_iommu to see if we need the workaround. Presumably that
2267	 * was loaded first.
2268	 */
2269	return IS_GEN5(dev_priv) && IS_MOBILE(dev_priv) && intel_vtd_active();
2270}
2271
2272static void gen6_check_and_clear_faults(struct drm_i915_private *dev_priv)
2273{
2274	struct intel_engine_cs *engine;
2275	enum intel_engine_id id;
2276	u32 fault;
2277
2278	for_each_engine(engine, dev_priv, id) {
2279		fault = I915_READ(RING_FAULT_REG(engine));
2280		if (fault & RING_FAULT_VALID) {
2281			DRM_DEBUG_DRIVER("Unexpected fault\n"
2282					 "\tAddr: 0x%08lx\n"
2283					 "\tAddress space: %s\n"
2284					 "\tSource ID: %d\n"
2285					 "\tType: %d\n",
2286					 fault & PAGE_MASK,
2287					 fault & RING_FAULT_GTTSEL_MASK ? "GGTT" : "PPGTT",
2288					 RING_FAULT_SRCID(fault),
2289					 RING_FAULT_FAULT_TYPE(fault));
2290			I915_WRITE(RING_FAULT_REG(engine),
2291				   fault & ~RING_FAULT_VALID);
2292		}
2293	}
2294
2295	POSTING_READ(RING_FAULT_REG(dev_priv->engine[RCS]));
2296}
2297
2298static void gen8_check_and_clear_faults(struct drm_i915_private *dev_priv)
2299{
2300	u32 fault = I915_READ(GEN8_RING_FAULT_REG);
2301
2302	if (fault & RING_FAULT_VALID) {
2303		u32 fault_data0, fault_data1;
2304		u64 fault_addr;
2305
2306		fault_data0 = I915_READ(GEN8_FAULT_TLB_DATA0);
2307		fault_data1 = I915_READ(GEN8_FAULT_TLB_DATA1);
2308		fault_addr = ((u64)(fault_data1 & FAULT_VA_HIGH_BITS) << 44) |
2309			     ((u64)fault_data0 << 12);
2310
2311		DRM_DEBUG_DRIVER("Unexpected fault\n"
2312				 "\tAddr: 0x%08x_%08x\n"
2313				 "\tAddress space: %s\n"
2314				 "\tEngine ID: %d\n"
2315				 "\tSource ID: %d\n"
2316				 "\tType: %d\n",
2317				 upper_32_bits(fault_addr),
2318				 lower_32_bits(fault_addr),
2319				 fault_data1 & FAULT_GTT_SEL ? "GGTT" : "PPGTT",
2320				 GEN8_RING_FAULT_ENGINE_ID(fault),
2321				 RING_FAULT_SRCID(fault),
2322				 RING_FAULT_FAULT_TYPE(fault));
2323		I915_WRITE(GEN8_RING_FAULT_REG,
2324			   fault & ~RING_FAULT_VALID);
2325	}
2326
2327	POSTING_READ(GEN8_RING_FAULT_REG);
2328}
2329
2330void i915_check_and_clear_faults(struct drm_i915_private *dev_priv)
2331{
2332	/* From GEN8 onwards we only have one 'All Engine Fault Register' */
2333	if (INTEL_GEN(dev_priv) >= 8)
2334		gen8_check_and_clear_faults(dev_priv);
2335	else if (INTEL_GEN(dev_priv) >= 6)
2336		gen6_check_and_clear_faults(dev_priv);
2337	else
2338		return;
2339}
2340
2341void i915_gem_suspend_gtt_mappings(struct drm_i915_private *dev_priv)
2342{
2343	struct i915_ggtt *ggtt = &dev_priv->ggtt;
2344
2345	/* Don't bother messing with faults pre GEN6 as we have little
2346	 * documentation supporting that it's a good idea.
2347	 */
2348	if (INTEL_GEN(dev_priv) < 6)
2349		return;
2350
2351	i915_check_and_clear_faults(dev_priv);
2352
2353	ggtt->base.clear_range(&ggtt->base, 0, ggtt->base.total);
2354
2355	i915_ggtt_invalidate(dev_priv);
2356}
2357
2358int i915_gem_gtt_prepare_pages(struct drm_i915_gem_object *obj,
2359			       struct sg_table *pages)
2360{
2361	do {
2362		if (dma_map_sg_attrs(&obj->base.dev->pdev->dev,
2363				     pages->sgl, pages->nents,
2364				     PCI_DMA_BIDIRECTIONAL,
2365				     DMA_ATTR_NO_WARN))
2366			return 0;
2367
2368		/* If the DMA remap fails, one cause can be that we have
2369		 * too many objects pinned in a small remapping table,
2370		 * such as swiotlb. Incrementally purge all other objects and
2371		 * try again - if there are no more pages to remove from
2372		 * the DMA remapper, i915_gem_shrink will return 0.
2373		 */
2374		GEM_BUG_ON(obj->mm.pages == pages);
2375	} while (i915_gem_shrink(to_i915(obj->base.dev),
2376				 obj->base.size >> PAGE_SHIFT, NULL,
2377				 I915_SHRINK_BOUND |
2378				 I915_SHRINK_UNBOUND |
2379				 I915_SHRINK_ACTIVE));
2380
2381	return -ENOSPC;
2382}
2383
2384static void gen8_set_pte(void __iomem *addr, gen8_pte_t pte)
2385{
2386	writeq(pte, addr);
2387}
2388
2389static void gen8_ggtt_insert_page(struct i915_address_space *vm,
2390				  dma_addr_t addr,
2391				  u64 offset,
2392				  enum i915_cache_level level,
2393				  u32 unused)
2394{
2395	struct i915_ggtt *ggtt = i915_vm_to_ggtt(vm);
2396	gen8_pte_t __iomem *pte =
2397		(gen8_pte_t __iomem *)ggtt->gsm + (offset >> PAGE_SHIFT);
2398
2399	gen8_set_pte(pte, gen8_pte_encode(addr, level));
2400
2401	ggtt->invalidate(vm->i915);
2402}
2403
2404static void gen8_ggtt_insert_entries(struct i915_address_space *vm,
2405				     struct i915_vma *vma,
2406				     enum i915_cache_level level,
2407				     u32 unused)
2408{
2409	struct i915_ggtt *ggtt = i915_vm_to_ggtt(vm);
2410	struct sgt_iter sgt_iter;
2411	gen8_pte_t __iomem *gtt_entries;
2412	const gen8_pte_t pte_encode = gen8_pte_encode(0, level);
2413	dma_addr_t addr;
2414
2415	gtt_entries = (gen8_pte_t __iomem *)ggtt->gsm;
2416	gtt_entries += vma->node.start >> PAGE_SHIFT;
2417	for_each_sgt_dma(addr, sgt_iter, vma->pages)
2418		gen8_set_pte(gtt_entries++, pte_encode | addr);
2419
2420	wmb();
2421
2422	/* This next bit makes the above posting read even more important. We
2423	 * want to flush the TLBs only after we're certain all the PTE updates
2424	 * have finished.
2425	 */
2426	ggtt->invalidate(vm->i915);
2427}
2428
2429static void gen6_ggtt_insert_page(struct i915_address_space *vm,
2430				  dma_addr_t addr,
2431				  u64 offset,
2432				  enum i915_cache_level level,
2433				  u32 flags)
2434{
2435	struct i915_ggtt *ggtt = i915_vm_to_ggtt(vm);
2436	gen6_pte_t __iomem *pte =
2437		(gen6_pte_t __iomem *)ggtt->gsm + (offset >> PAGE_SHIFT);
2438
2439	iowrite32(vm->pte_encode(addr, level, flags), pte);
2440
2441	ggtt->invalidate(vm->i915);
2442}
2443
2444/*
2445 * Binds an object into the global gtt with the specified cache level. The object
2446 * will be accessible to the GPU via commands whose operands reference offsets
2447 * within the global GTT as well as accessible by the GPU through the GMADR
2448 * mapped BAR (dev_priv->mm.gtt->gtt).
2449 */
2450static void gen6_ggtt_insert_entries(struct i915_address_space *vm,
2451				     struct i915_vma *vma,
2452				     enum i915_cache_level level,
2453				     u32 flags)
2454{
2455	struct i915_ggtt *ggtt = i915_vm_to_ggtt(vm);
2456	gen6_pte_t __iomem *entries = (gen6_pte_t __iomem *)ggtt->gsm;
2457	unsigned int i = vma->node.start >> PAGE_SHIFT;
2458	struct sgt_iter iter;
2459	dma_addr_t addr;
2460	for_each_sgt_dma(addr, iter, vma->pages)
2461		iowrite32(vm->pte_encode(addr, level, flags), &entries[i++]);
2462	wmb();
2463
2464	/* This next bit makes the above posting read even more important. We
2465	 * want to flush the TLBs only after we're certain all the PTE updates
2466	 * have finished.
2467	 */
2468	ggtt->invalidate(vm->i915);
2469}
2470
2471static void nop_clear_range(struct i915_address_space *vm,
2472			    u64 start, u64 length)
2473{
2474}
2475
2476static void gen8_ggtt_clear_range(struct i915_address_space *vm,
2477				  u64 start, u64 length)
2478{
2479	struct i915_ggtt *ggtt = i915_vm_to_ggtt(vm);
2480	unsigned first_entry = start >> PAGE_SHIFT;
2481	unsigned num_entries = length >> PAGE_SHIFT;
2482	const gen8_pte_t scratch_pte =
2483		gen8_pte_encode(vm->scratch_page.daddr, I915_CACHE_LLC);
2484	gen8_pte_t __iomem *gtt_base =
2485		(gen8_pte_t __iomem *)ggtt->gsm + first_entry;
2486	const int max_entries = ggtt_total_entries(ggtt) - first_entry;
2487	int i;
2488
2489	if (WARN(num_entries > max_entries,
2490		 "First entry = %d; Num entries = %d (max=%d)\n",
2491		 first_entry, num_entries, max_entries))
2492		num_entries = max_entries;
2493
2494	for (i = 0; i < num_entries; i++)
2495		gen8_set_pte(&gtt_base[i], scratch_pte);
2496}
2497
2498static void bxt_vtd_ggtt_wa(struct i915_address_space *vm)
2499{
2500	struct drm_i915_private *dev_priv = vm->i915;
2501
2502	/*
2503	 * Make sure the internal GAM fifo has been cleared of all GTT
2504	 * writes before exiting stop_machine(). This guarantees that
2505	 * any aperture accesses waiting to start in another process
2506	 * cannot back up behind the GTT writes causing a hang.
2507	 * The register can be any arbitrary GAM register.
2508	 */
2509	POSTING_READ(GFX_FLSH_CNTL_GEN6);
2510}
2511
2512struct insert_page {
2513	struct i915_address_space *vm;
2514	dma_addr_t addr;
2515	u64 offset;
2516	enum i915_cache_level level;
2517};
2518
2519static int bxt_vtd_ggtt_insert_page__cb(void *_arg)
2520{
2521	struct insert_page *arg = _arg;
2522
2523	gen8_ggtt_insert_page(arg->vm, arg->addr, arg->offset, arg->level, 0);
2524	bxt_vtd_ggtt_wa(arg->vm);
2525
2526	return 0;
2527}
2528
2529static void bxt_vtd_ggtt_insert_page__BKL(struct i915_address_space *vm,
2530					  dma_addr_t addr,
2531					  u64 offset,
2532					  enum i915_cache_level level,
2533					  u32 unused)
2534{
2535	struct insert_page arg = { vm, addr, offset, level };
2536
2537	stop_machine(bxt_vtd_ggtt_insert_page__cb, &arg, NULL);
2538}
2539
2540struct insert_entries {
2541	struct i915_address_space *vm;
2542	struct i915_vma *vma;
2543	enum i915_cache_level level;
2544};
2545
2546static int bxt_vtd_ggtt_insert_entries__cb(void *_arg)
2547{
2548	struct insert_entries *arg = _arg;
2549
2550	gen8_ggtt_insert_entries(arg->vm, arg->vma, arg->level, 0);
2551	bxt_vtd_ggtt_wa(arg->vm);
2552
2553	return 0;
2554}
2555
2556static void bxt_vtd_ggtt_insert_entries__BKL(struct i915_address_space *vm,
2557					     struct i915_vma *vma,
2558					     enum i915_cache_level level,
2559					     u32 unused)
2560{
2561	struct insert_entries arg = { vm, vma, level };
2562
2563	stop_machine(bxt_vtd_ggtt_insert_entries__cb, &arg, NULL);
2564}
2565
2566struct clear_range {
2567	struct i915_address_space *vm;
2568	u64 start;
2569	u64 length;
2570};
2571
2572static int bxt_vtd_ggtt_clear_range__cb(void *_arg)
2573{
2574	struct clear_range *arg = _arg;
2575
2576	gen8_ggtt_clear_range(arg->vm, arg->start, arg->length);
2577	bxt_vtd_ggtt_wa(arg->vm);
2578
2579	return 0;
2580}
2581
2582static void bxt_vtd_ggtt_clear_range__BKL(struct i915_address_space *vm,
2583					  u64 start,
2584					  u64 length)
2585{
2586	struct clear_range arg = { vm, start, length };
2587
2588	stop_machine(bxt_vtd_ggtt_clear_range__cb, &arg, NULL);
2589}
2590
2591static void gen6_ggtt_clear_range(struct i915_address_space *vm,
2592				  u64 start, u64 length)
2593{
2594	struct i915_ggtt *ggtt = i915_vm_to_ggtt(vm);
2595	unsigned first_entry = start >> PAGE_SHIFT;
2596	unsigned num_entries = length >> PAGE_SHIFT;
2597	gen6_pte_t scratch_pte, __iomem *gtt_base =
2598		(gen6_pte_t __iomem *)ggtt->gsm + first_entry;
2599	const int max_entries = ggtt_total_entries(ggtt) - first_entry;
2600	int i;
2601
2602	if (WARN(num_entries > max_entries,
2603		 "First entry = %d; Num entries = %d (max=%d)\n",
2604		 first_entry, num_entries, max_entries))
2605		num_entries = max_entries;
2606
2607	scratch_pte = vm->pte_encode(vm->scratch_page.daddr,
2608				     I915_CACHE_LLC, 0);
2609
2610	for (i = 0; i < num_entries; i++)
2611		iowrite32(scratch_pte, &gtt_base[i]);
2612}
2613
2614static void i915_ggtt_insert_page(struct i915_address_space *vm,
2615				  dma_addr_t addr,
2616				  u64 offset,
2617				  enum i915_cache_level cache_level,
2618				  u32 unused)
2619{
2620	unsigned int flags = (cache_level == I915_CACHE_NONE) ?
2621		AGP_USER_MEMORY : AGP_USER_CACHED_MEMORY;
2622
2623	intel_gtt_insert_page(addr, offset >> PAGE_SHIFT, flags);
2624}
2625
2626static void i915_ggtt_insert_entries(struct i915_address_space *vm,
2627				     struct i915_vma *vma,
2628				     enum i915_cache_level cache_level,
2629				     u32 unused)
2630{
2631	unsigned int flags = (cache_level == I915_CACHE_NONE) ?
2632		AGP_USER_MEMORY : AGP_USER_CACHED_MEMORY;
2633
2634	intel_gtt_insert_sg_entries(vma->pages, vma->node.start >> PAGE_SHIFT,
2635				    flags);
2636}
2637
2638static void i915_ggtt_clear_range(struct i915_address_space *vm,
2639				  u64 start, u64 length)
2640{
2641	intel_gtt_clear_range(start >> PAGE_SHIFT, length >> PAGE_SHIFT);
2642}
2643
2644static int ggtt_bind_vma(struct i915_vma *vma,
2645			 enum i915_cache_level cache_level,
2646			 u32 flags)
2647{
2648	struct drm_i915_private *i915 = vma->vm->i915;
2649	struct drm_i915_gem_object *obj = vma->obj;
2650	u32 pte_flags;
2651
2652	/* Currently applicable only to VLV */
2653	pte_flags = 0;
2654	if (obj->gt_ro)
2655		pte_flags |= PTE_READ_ONLY;
2656
2657	intel_runtime_pm_get(i915);
2658	vma->vm->insert_entries(vma->vm, vma, cache_level, pte_flags);
2659	intel_runtime_pm_put(i915);
2660
2661	vma->page_sizes.gtt = I915_GTT_PAGE_SIZE;
2662
2663	/*
2664	 * Without aliasing PPGTT there's no difference between
2665	 * GLOBAL/LOCAL_BIND, it's all the same ptes. Hence unconditionally
2666	 * upgrade to both bound if we bind either to avoid double-binding.
2667	 */
2668	vma->flags |= I915_VMA_GLOBAL_BIND | I915_VMA_LOCAL_BIND;
2669
2670	return 0;
2671}
2672
2673static void ggtt_unbind_vma(struct i915_vma *vma)
2674{
2675	struct drm_i915_private *i915 = vma->vm->i915;
2676
2677	intel_runtime_pm_get(i915);
2678	vma->vm->clear_range(vma->vm, vma->node.start, vma->size);
2679	intel_runtime_pm_put(i915);
2680}
2681
2682static int aliasing_gtt_bind_vma(struct i915_vma *vma,
2683				 enum i915_cache_level cache_level,
2684				 u32 flags)
2685{
2686	struct drm_i915_private *i915 = vma->vm->i915;
2687	u32 pte_flags;
2688	int ret;
2689
2690	/* Currently applicable only to VLV */
2691	pte_flags = 0;
2692	if (vma->obj->gt_ro)
2693		pte_flags |= PTE_READ_ONLY;
2694
2695	if (flags & I915_VMA_LOCAL_BIND) {
2696		struct i915_hw_ppgtt *appgtt = i915->mm.aliasing_ppgtt;
2697
2698		if (!(vma->flags & I915_VMA_LOCAL_BIND) &&
2699		    appgtt->base.allocate_va_range) {
2700			ret = appgtt->base.allocate_va_range(&appgtt->base,
2701							     vma->node.start,
2702							     vma->size);
2703			if (ret)
2704				return ret;
2705		}
2706
2707		appgtt->base.insert_entries(&appgtt->base, vma, cache_level,
2708					    pte_flags);
2709	}
2710
2711	if (flags & I915_VMA_GLOBAL_BIND) {
2712		intel_runtime_pm_get(i915);
2713		vma->vm->insert_entries(vma->vm, vma, cache_level, pte_flags);
2714		intel_runtime_pm_put(i915);
2715	}
2716
2717	return 0;
2718}
2719
2720static void aliasing_gtt_unbind_vma(struct i915_vma *vma)
2721{
2722	struct drm_i915_private *i915 = vma->vm->i915;
2723
2724	if (vma->flags & I915_VMA_GLOBAL_BIND) {
2725		intel_runtime_pm_get(i915);
2726		vma->vm->clear_range(vma->vm, vma->node.start, vma->size);
2727		intel_runtime_pm_put(i915);
2728	}
2729
2730	if (vma->flags & I915_VMA_LOCAL_BIND) {
2731		struct i915_address_space *vm = &i915->mm.aliasing_ppgtt->base;
2732
2733		vm->clear_range(vm, vma->node.start, vma->size);
2734	}
2735}
2736
2737void i915_gem_gtt_finish_pages(struct drm_i915_gem_object *obj,
2738			       struct sg_table *pages)
2739{
2740	struct drm_i915_private *dev_priv = to_i915(obj->base.dev);
2741	struct device *kdev = &dev_priv->drm.pdev->dev;
2742	struct i915_ggtt *ggtt = &dev_priv->ggtt;
2743
2744	if (unlikely(ggtt->do_idle_maps)) {
2745		if (i915_gem_wait_for_idle(dev_priv, 0)) {
2746			DRM_ERROR("Failed to wait for idle; VT'd may hang.\n");
2747			/* Wait a bit, in hopes it avoids the hang */
2748			udelay(10);
2749		}
2750	}
2751
2752	dma_unmap_sg(kdev, pages->sgl, pages->nents, PCI_DMA_BIDIRECTIONAL);
2753}
2754
2755static int ggtt_set_pages(struct i915_vma *vma)
2756{
2757	int ret;
2758
2759	GEM_BUG_ON(vma->pages);
2760
2761	ret = i915_get_ggtt_vma_pages(vma);
2762	if (ret)
2763		return ret;
2764
2765	vma->page_sizes = vma->obj->mm.page_sizes;
2766
2767	return 0;
2768}
2769
2770static void i915_gtt_color_adjust(const struct drm_mm_node *node,
2771				  unsigned long color,
2772				  u64 *start,
2773				  u64 *end)
2774{
2775	if (node->allocated && node->color != color)
2776		*start += I915_GTT_PAGE_SIZE;
2777
2778	/* Also leave a space between the unallocated reserved node after the
2779	 * GTT and any objects within the GTT, i.e. we use the color adjustment
2780	 * to insert a guard page to prevent prefetches crossing over the
2781	 * GTT boundary.
2782	 */
2783	node = list_next_entry(node, node_list);
2784	if (node->color != color)
2785		*end -= I915_GTT_PAGE_SIZE;
2786}
2787
2788int i915_gem_init_aliasing_ppgtt(struct drm_i915_private *i915)
2789{
2790	struct i915_ggtt *ggtt = &i915->ggtt;
2791	struct i915_hw_ppgtt *ppgtt;
2792	int err;
2793
2794	ppgtt = i915_ppgtt_create(i915, ERR_PTR(-EPERM), "[alias]");
2795	if (IS_ERR(ppgtt))
2796		return PTR_ERR(ppgtt);
2797
2798	if (WARN_ON(ppgtt->base.total < ggtt->base.total)) {
2799		err = -ENODEV;
2800		goto err_ppgtt;
2801	}
2802
2803	if (ppgtt->base.allocate_va_range) {
2804		/* Note we only pre-allocate as far as the end of the global
2805		 * GTT. On 48b / 4-level page-tables, the difference is very,
2806		 * very significant! We have to preallocate as GVT/vgpu does
2807		 * not like the page directory disappearing.
2808		 */
2809		err = ppgtt->base.allocate_va_range(&ppgtt->base,
2810						    0, ggtt->base.total);
2811		if (err)
2812			goto err_ppgtt;
2813	}
2814
2815	i915->mm.aliasing_ppgtt = ppgtt;
2816
2817	GEM_BUG_ON(ggtt->base.bind_vma != ggtt_bind_vma);
2818	ggtt->base.bind_vma = aliasing_gtt_bind_vma;
2819
2820	GEM_BUG_ON(ggtt->base.unbind_vma != ggtt_unbind_vma);
2821	ggtt->base.unbind_vma = aliasing_gtt_unbind_vma;
2822
2823	return 0;
2824
2825err_ppgtt:
2826	i915_ppgtt_put(ppgtt);
2827	return err;
2828}
2829
2830void i915_gem_fini_aliasing_ppgtt(struct drm_i915_private *i915)
2831{
2832	struct i915_ggtt *ggtt = &i915->ggtt;
2833	struct i915_hw_ppgtt *ppgtt;
2834
2835	ppgtt = fetch_and_zero(&i915->mm.aliasing_ppgtt);
2836	if (!ppgtt)
2837		return;
2838
2839	i915_ppgtt_put(ppgtt);
2840
2841	ggtt->base.bind_vma = ggtt_bind_vma;
2842	ggtt->base.unbind_vma = ggtt_unbind_vma;
2843}
2844
2845int i915_gem_init_ggtt(struct drm_i915_private *dev_priv)
2846{
2847	/* Let GEM Manage all of the aperture.
2848	 *
2849	 * However, leave one page at the end still bound to the scratch page.
2850	 * There are a number of places where the hardware apparently prefetches
2851	 * past the end of the object, and we've seen multiple hangs with the
2852	 * GPU head pointer stuck in a batchbuffer bound at the last page of the
2853	 * aperture.  One page should be enough to keep any prefetching inside
2854	 * of the aperture.
2855	 */
2856	struct i915_ggtt *ggtt = &dev_priv->ggtt;
2857	unsigned long hole_start, hole_end;
2858	struct drm_mm_node *entry;
2859	int ret;
2860
2861	ret = intel_vgt_balloon(dev_priv);
2862	if (ret)
2863		return ret;
2864
2865	/* Reserve a mappable slot for our lockless error capture */
2866	ret = drm_mm_insert_node_in_range(&ggtt->base.mm, &ggtt->error_capture,
2867					  PAGE_SIZE, 0, I915_COLOR_UNEVICTABLE,
2868					  0, ggtt->mappable_end,
2869					  DRM_MM_INSERT_LOW);
2870	if (ret)
2871		return ret;
2872
2873	/* Clear any non-preallocated blocks */
2874	drm_mm_for_each_hole(entry, &ggtt->base.mm, hole_start, hole_end) {
2875		DRM_DEBUG_KMS("clearing unused GTT space: [%lx, %lx]\n",
2876			      hole_start, hole_end);
2877		ggtt->base.clear_range(&ggtt->base, hole_start,
2878				       hole_end - hole_start);
2879	}
2880
2881	/* And finally clear the reserved guard page */
2882	ggtt->base.clear_range(&ggtt->base,
2883			       ggtt->base.total - PAGE_SIZE, PAGE_SIZE);
2884
2885	if (USES_PPGTT(dev_priv) && !USES_FULL_PPGTT(dev_priv)) {
2886		ret = i915_gem_init_aliasing_ppgtt(dev_priv);
2887		if (ret)
2888			goto err;
2889	}
2890
2891	return 0;
2892
2893err:
2894	drm_mm_remove_node(&ggtt->error_capture);
2895	return ret;
2896}
2897
2898/**
2899 * i915_ggtt_cleanup_hw - Clean up GGTT hardware initialization
2900 * @dev_priv: i915 device
2901 */
2902void i915_ggtt_cleanup_hw(struct drm_i915_private *dev_priv)
2903{
2904	struct i915_ggtt *ggtt = &dev_priv->ggtt;
2905	struct i915_vma *vma, *vn;
2906	struct pagevec *pvec;
2907
2908	ggtt->base.closed = true;
2909
2910	mutex_lock(&dev_priv->drm.struct_mutex);
2911	GEM_BUG_ON(!list_empty(&ggtt->base.active_list));
2912	list_for_each_entry_safe(vma, vn, &ggtt->base.inactive_list, vm_link)
2913		WARN_ON(i915_vma_unbind(vma));
2914	mutex_unlock(&dev_priv->drm.struct_mutex);
2915
2916	i915_gem_cleanup_stolen(&dev_priv->drm);
2917
2918	mutex_lock(&dev_priv->drm.struct_mutex);
2919	i915_gem_fini_aliasing_ppgtt(dev_priv);
2920
2921	if (drm_mm_node_allocated(&ggtt->error_capture))
2922		drm_mm_remove_node(&ggtt->error_capture);
2923
2924	if (drm_mm_initialized(&ggtt->base.mm)) {
2925		intel_vgt_deballoon(dev_priv);
2926		i915_address_space_fini(&ggtt->base);
2927	}
2928
2929	ggtt->base.cleanup(&ggtt->base);
2930
2931	pvec = &dev_priv->mm.wc_stash;
2932	if (pvec->nr) {
2933		set_pages_array_wb(pvec->pages, pvec->nr);
2934		__pagevec_release(pvec);
2935	}
2936
2937	mutex_unlock(&dev_priv->drm.struct_mutex);
2938
2939	arch_phys_wc_del(ggtt->mtrr);
2940	io_mapping_fini(&ggtt->iomap);
2941}
2942
2943static unsigned int gen6_get_total_gtt_size(u16 snb_gmch_ctl)
2944{
2945	snb_gmch_ctl >>= SNB_GMCH_GGMS_SHIFT;
2946	snb_gmch_ctl &= SNB_GMCH_GGMS_MASK;
2947	return snb_gmch_ctl << 20;
2948}
2949
2950static unsigned int gen8_get_total_gtt_size(u16 bdw_gmch_ctl)
2951{
2952	bdw_gmch_ctl >>= BDW_GMCH_GGMS_SHIFT;
2953	bdw_gmch_ctl &= BDW_GMCH_GGMS_MASK;
2954	if (bdw_gmch_ctl)
2955		bdw_gmch_ctl = 1 << bdw_gmch_ctl;
2956
2957#ifdef CONFIG_X86_32
2958	/* Limit 32b platforms to a 2GB GGTT: 4 << 20 / pte size * PAGE_SIZE */
2959	if (bdw_gmch_ctl > 4)
2960		bdw_gmch_ctl = 4;
2961#endif
2962
2963	return bdw_gmch_ctl << 20;
2964}
2965
2966static unsigned int chv_get_total_gtt_size(u16 gmch_ctrl)
2967{
2968	gmch_ctrl >>= SNB_GMCH_GGMS_SHIFT;
2969	gmch_ctrl &= SNB_GMCH_GGMS_MASK;
2970
2971	if (gmch_ctrl)
2972		return 1 << (20 + gmch_ctrl);
2973
2974	return 0;
2975}
2976
2977static int ggtt_probe_common(struct i915_ggtt *ggtt, u64 size)
2978{
2979	struct drm_i915_private *dev_priv = ggtt->base.i915;
2980	struct pci_dev *pdev = dev_priv->drm.pdev;
2981	phys_addr_t phys_addr;
2982	int ret;
2983
2984	/* For Modern GENs the PTEs and register space are split in the BAR */
2985	phys_addr = pci_resource_start(pdev, 0) + pci_resource_len(pdev, 0) / 2;
2986
2987	/*
2988	 * On BXT+/CNL+ writes larger than 64 bit to the GTT pagetable range
2989	 * will be dropped. For WC mappings in general we have 64 byte burst
2990	 * writes when the WC buffer is flushed, so we can't use it, but have to
2991	 * resort to an uncached mapping. The WC issue is easily caught by the
2992	 * readback check when writing GTT PTE entries.
2993	 */
2994	if (IS_GEN9_LP(dev_priv) || INTEL_GEN(dev_priv) >= 10)
2995		ggtt->gsm = ioremap_nocache(phys_addr, size);
2996	else
2997		ggtt->gsm = ioremap_wc(phys_addr, size);
2998	if (!ggtt->gsm) {
2999		DRM_ERROR("Failed to map the ggtt page table\n");
3000		return -ENOMEM;
3001	}
3002
3003	ret = setup_scratch_page(&ggtt->base, GFP_DMA32);
3004	if (ret) {
3005		DRM_ERROR("Scratch setup failed\n");
3006		/* iounmap will also get called at remove, but meh */
3007		iounmap(ggtt->gsm);
3008		return ret;
3009	}
3010
3011	return 0;
3012}
3013
3014static struct intel_ppat_entry *
3015__alloc_ppat_entry(struct intel_ppat *ppat, unsigned int index, u8 value)
3016{
3017	struct intel_ppat_entry *entry = &ppat->entries[index];
3018
3019	GEM_BUG_ON(index >= ppat->max_entries);
3020	GEM_BUG_ON(test_bit(index, ppat->used));
3021
3022	entry->ppat = ppat;
3023	entry->value = value;
3024	kref_init(&entry->ref);
3025	set_bit(index, ppat->used);
3026	set_bit(index, ppat->dirty);
3027
3028	return entry;
3029}
3030
3031static void __free_ppat_entry(struct intel_ppat_entry *entry)
3032{
3033	struct intel_ppat *ppat = entry->ppat;
3034	unsigned int index = entry - ppat->entries;
3035
3036	GEM_BUG_ON(index >= ppat->max_entries);
3037	GEM_BUG_ON(!test_bit(index, ppat->used));
3038
3039	entry->value = ppat->clear_value;
3040	clear_bit(index, ppat->used);
3041	set_bit(index, ppat->dirty);
3042}
3043
3044/**
3045 * intel_ppat_get - get a usable PPAT entry
3046 * @i915: i915 device instance
3047 * @value: the PPAT value required by the caller
3048 *
3049 * The function tries to search if there is an existing PPAT entry which
3050 * matches with the required value. If perfectly matched, the existing PPAT
3051 * entry will be used. If only partially matched, it will try to check if
3052 * there is any available PPAT index. If yes, it will allocate a new PPAT
3053 * index for the required entry and update the HW. If not, the partially
3054 * matched entry will be used.
3055 */
3056const struct intel_ppat_entry *
3057intel_ppat_get(struct drm_i915_private *i915, u8 value)
3058{
3059	struct intel_ppat *ppat = &i915->ppat;
3060	struct intel_ppat_entry *entry = NULL;
3061	unsigned int scanned, best_score;
3062	int i;
3063
3064	GEM_BUG_ON(!ppat->max_entries);
3065
3066	scanned = best_score = 0;
3067	for_each_set_bit(i, ppat->used, ppat->max_entries) {
3068		unsigned int score;
3069
3070		score = ppat->match(ppat->entries[i].value, value);
3071		if (score > best_score) {
3072			entry = &ppat->entries[i];
3073			if (score == INTEL_PPAT_PERFECT_MATCH) {
3074				kref_get(&entry->ref);
3075				return entry;
3076			}
3077			best_score = score;
3078		}
3079		scanned++;
3080	}
3081
3082	if (scanned == ppat->max_entries) {
3083		if (!entry)
3084			return ERR_PTR(-ENOSPC);
3085
3086		kref_get(&entry->ref);
3087		return entry;
3088	}
3089
3090	i = find_first_zero_bit(ppat->used, ppat->max_entries);
3091	entry = __alloc_ppat_entry(ppat, i, value);
3092	ppat->update_hw(i915);
3093	return entry;
3094}
3095
3096static void release_ppat(struct kref *kref)
3097{
3098	struct intel_ppat_entry *entry =
3099		container_of(kref, struct intel_ppat_entry, ref);
3100	struct drm_i915_private *i915 = entry->ppat->i915;
3101
3102	__free_ppat_entry(entry);
3103	entry->ppat->update_hw(i915);
3104}
3105
3106/**
3107 * intel_ppat_put - put back the PPAT entry got from intel_ppat_get()
3108 * @entry: an intel PPAT entry
3109 *
3110 * Put back the PPAT entry got from intel_ppat_get(). If the PPAT index of the
3111 * entry is dynamically allocated, its reference count will be decreased. Once
3112 * the reference count becomes into zero, the PPAT index becomes free again.
3113 */
3114void intel_ppat_put(const struct intel_ppat_entry *entry)
3115{
3116	struct intel_ppat *ppat = entry->ppat;
3117	unsigned int index = entry - ppat->entries;
3118
3119	GEM_BUG_ON(!ppat->max_entries);
3120
3121	kref_put(&ppat->entries[index].ref, release_ppat);
3122}
3123
3124static void cnl_private_pat_update_hw(struct drm_i915_private *dev_priv)
3125{
3126	struct intel_ppat *ppat = &dev_priv->ppat;
3127	int i;
3128
3129	for_each_set_bit(i, ppat->dirty, ppat->max_entries) {
3130		I915_WRITE(GEN10_PAT_INDEX(i), ppat->entries[i].value);
3131		clear_bit(i, ppat->dirty);
3132	}
3133}
3134
3135static void bdw_private_pat_update_hw(struct drm_i915_private *dev_priv)
3136{
3137	struct intel_ppat *ppat = &dev_priv->ppat;
3138	u64 pat = 0;
3139	int i;
3140
3141	for (i = 0; i < ppat->max_entries; i++)
3142		pat |= GEN8_PPAT(i, ppat->entries[i].value);
3143
3144	bitmap_clear(ppat->dirty, 0, ppat->max_entries);
3145
3146	I915_WRITE(GEN8_PRIVATE_PAT_LO, lower_32_bits(pat));
3147	I915_WRITE(GEN8_PRIVATE_PAT_HI, upper_32_bits(pat));
3148}
3149
3150static unsigned int bdw_private_pat_match(u8 src, u8 dst)
3151{
3152	unsigned int score = 0;
3153	enum {
3154		AGE_MATCH = BIT(0),
3155		TC_MATCH = BIT(1),
3156		CA_MATCH = BIT(2),
3157	};
3158
3159	/* Cache attribute has to be matched. */
3160	if (GEN8_PPAT_GET_CA(src) != GEN8_PPAT_GET_CA(dst))
3161		return 0;
3162
3163	score |= CA_MATCH;
3164
3165	if (GEN8_PPAT_GET_TC(src) == GEN8_PPAT_GET_TC(dst))
3166		score |= TC_MATCH;
3167
3168	if (GEN8_PPAT_GET_AGE(src) == GEN8_PPAT_GET_AGE(dst))
3169		score |= AGE_MATCH;
3170
3171	if (score == (AGE_MATCH | TC_MATCH | CA_MATCH))
3172		return INTEL_PPAT_PERFECT_MATCH;
3173
3174	return score;
3175}
3176
3177static unsigned int chv_private_pat_match(u8 src, u8 dst)
3178{
3179	return (CHV_PPAT_GET_SNOOP(src) == CHV_PPAT_GET_SNOOP(dst)) ?
3180		INTEL_PPAT_PERFECT_MATCH : 0;
3181}
3182
3183static void cnl_setup_private_ppat(struct intel_ppat *ppat)
3184{
3185	ppat->max_entries = 8;
3186	ppat->update_hw = cnl_private_pat_update_hw;
3187	ppat->match = bdw_private_pat_match;
3188	ppat->clear_value = GEN8_PPAT_WB | GEN8_PPAT_LLCELLC | GEN8_PPAT_AGE(3);
3189
3190	__alloc_ppat_entry(ppat, 0, GEN8_PPAT_WB | GEN8_PPAT_LLC);
3191	__alloc_ppat_entry(ppat, 1, GEN8_PPAT_WC | GEN8_PPAT_LLCELLC);
3192	__alloc_ppat_entry(ppat, 2, GEN8_PPAT_WT | GEN8_PPAT_LLCELLC);
3193	__alloc_ppat_entry(ppat, 3, GEN8_PPAT_UC);
3194	__alloc_ppat_entry(ppat, 4, GEN8_PPAT_WB | GEN8_PPAT_LLCELLC | GEN8_PPAT_AGE(0));
3195	__alloc_ppat_entry(ppat, 5, GEN8_PPAT_WB | GEN8_PPAT_LLCELLC | GEN8_PPAT_AGE(1));
3196	__alloc_ppat_entry(ppat, 6, GEN8_PPAT_WB | GEN8_PPAT_LLCELLC | GEN8_PPAT_AGE(2));
3197	__alloc_ppat_entry(ppat, 7, GEN8_PPAT_WB | GEN8_PPAT_LLCELLC | GEN8_PPAT_AGE(3));
3198}
3199
3200/* The GGTT and PPGTT need a private PPAT setup in order to handle cacheability
3201 * bits. When using advanced contexts each context stores its own PAT, but
3202 * writing this data shouldn't be harmful even in those cases. */
3203static void bdw_setup_private_ppat(struct intel_ppat *ppat)
3204{
3205	ppat->max_entries = 8;
3206	ppat->update_hw = bdw_private_pat_update_hw;
3207	ppat->match = bdw_private_pat_match;
3208	ppat->clear_value = GEN8_PPAT_WB | GEN8_PPAT_LLCELLC | GEN8_PPAT_AGE(3);
3209
3210	if (!USES_PPGTT(ppat->i915)) {
3211		/* Spec: "For GGTT, there is NO pat_sel[2:0] from the entry,
3212		 * so RTL will always use the value corresponding to
3213		 * pat_sel = 000".
3214		 * So let's disable cache for GGTT to avoid screen corruptions.
3215		 * MOCS still can be used though.
3216		 * - System agent ggtt writes (i.e. cpu gtt mmaps) already work
3217		 * before this patch, i.e. the same uncached + snooping access
3218		 * like on gen6/7 seems to be in effect.
3219		 * - So this just fixes blitter/render access. Again it looks
3220		 * like it's not just uncached access, but uncached + snooping.
3221		 * So we can still hold onto all our assumptions wrt cpu
3222		 * clflushing on LLC machines.
3223		 */
3224		__alloc_ppat_entry(ppat, 0, GEN8_PPAT_UC);
3225		return;
3226	}
3227
3228	__alloc_ppat_entry(ppat, 0, GEN8_PPAT_WB | GEN8_PPAT_LLC);      /* for normal objects, no eLLC */
3229	__alloc_ppat_entry(ppat, 1, GEN8_PPAT_WC | GEN8_PPAT_LLCELLC);  /* for something pointing to ptes? */
3230	__alloc_ppat_entry(ppat, 2, GEN8_PPAT_WT | GEN8_PPAT_LLCELLC);  /* for scanout with eLLC */
3231	__alloc_ppat_entry(ppat, 3, GEN8_PPAT_UC);                      /* Uncached objects, mostly for scanout */
3232	__alloc_ppat_entry(ppat, 4, GEN8_PPAT_WB | GEN8_PPAT_LLCELLC | GEN8_PPAT_AGE(0));
3233	__alloc_ppat_entry(ppat, 5, GEN8_PPAT_WB | GEN8_PPAT_LLCELLC | GEN8_PPAT_AGE(1));
3234	__alloc_ppat_entry(ppat, 6, GEN8_PPAT_WB | GEN8_PPAT_LLCELLC | GEN8_PPAT_AGE(2));
3235	__alloc_ppat_entry(ppat, 7, GEN8_PPAT_WB | GEN8_PPAT_LLCELLC | GEN8_PPAT_AGE(3));
3236}
3237
3238static void chv_setup_private_ppat(struct intel_ppat *ppat)
3239{
3240	ppat->max_entries = 8;
3241	ppat->update_hw = bdw_private_pat_update_hw;
3242	ppat->match = chv_private_pat_match;
3243	ppat->clear_value = CHV_PPAT_SNOOP;
3244
3245	/*
3246	 * Map WB on BDW to snooped on CHV.
3247	 *
3248	 * Only the snoop bit has meaning for CHV, the rest is
3249	 * ignored.
3250	 *
3251	 * The hardware will never snoop for certain types of accesses:
3252	 * - CPU GTT (GMADR->GGTT->no snoop->memory)
3253	 * - PPGTT page tables
3254	 * - some other special cycles
3255	 *
3256	 * As with BDW, we also need to consider the following for GT accesses:
3257	 * "For GGTT, there is NO pat_sel[2:0] from the entry,
3258	 * so RTL will always use the value corresponding to
3259	 * pat_sel = 000".
3260	 * Which means we must set the snoop bit in PAT entry 0
3261	 * in order to keep the global status page working.
3262	 */
3263
3264	__alloc_ppat_entry(ppat, 0, CHV_PPAT_SNOOP);
3265	__alloc_ppat_entry(ppat, 1, 0);
3266	__alloc_ppat_entry(ppat, 2, 0);
3267	__alloc_ppat_entry(ppat, 3, 0);
3268	__alloc_ppat_entry(ppat, 4, CHV_PPAT_SNOOP);
3269	__alloc_ppat_entry(ppat, 5, CHV_PPAT_SNOOP);
3270	__alloc_ppat_entry(ppat, 6, CHV_PPAT_SNOOP);
3271	__alloc_ppat_entry(ppat, 7, CHV_PPAT_SNOOP);
3272}
3273
3274static void gen6_gmch_remove(struct i915_address_space *vm)
3275{
3276	struct i915_ggtt *ggtt = i915_vm_to_ggtt(vm);
3277
3278	iounmap(ggtt->gsm);
3279	cleanup_scratch_page(vm);
3280}
3281
3282static void setup_private_pat(struct drm_i915_private *dev_priv)
3283{
3284	struct intel_ppat *ppat = &dev_priv->ppat;
3285	int i;
3286
3287	ppat->i915 = dev_priv;
3288
3289	if (INTEL_GEN(dev_priv) >= 10)
3290		cnl_setup_private_ppat(ppat);
3291	else if (IS_CHERRYVIEW(dev_priv) || IS_GEN9_LP(dev_priv))
3292		chv_setup_private_ppat(ppat);
3293	else
3294		bdw_setup_private_ppat(ppat);
3295
3296	GEM_BUG_ON(ppat->max_entries > INTEL_MAX_PPAT_ENTRIES);
3297
3298	for_each_clear_bit(i, ppat->used, ppat->max_entries) {
3299		ppat->entries[i].value = ppat->clear_value;
3300		ppat->entries[i].ppat = ppat;
3301		set_bit(i, ppat->dirty);
3302	}
3303
3304	ppat->update_hw(dev_priv);
3305}
3306
3307static int gen8_gmch_probe(struct i915_ggtt *ggtt)
3308{
3309	struct drm_i915_private *dev_priv = ggtt->base.i915;
3310	struct pci_dev *pdev = dev_priv->drm.pdev;
3311	unsigned int size;
3312	u16 snb_gmch_ctl;
3313	int err;
3314
3315	/* TODO: We're not aware of mappable constraints on gen8 yet */
3316	ggtt->gmadr =
3317		(struct resource) DEFINE_RES_MEM(pci_resource_start(pdev, 2),
3318						 pci_resource_len(pdev, 2));
3319	ggtt->mappable_end = resource_size(&ggtt->gmadr);
3320
3321	err = pci_set_dma_mask(pdev, DMA_BIT_MASK(39));
3322	if (!err)
3323		err = pci_set_consistent_dma_mask(pdev, DMA_BIT_MASK(39));
3324	if (err)
3325		DRM_ERROR("Can't set DMA mask/consistent mask (%d)\n", err);
3326
3327	pci_read_config_word(pdev, SNB_GMCH_CTRL, &snb_gmch_ctl);
3328
3329	if (INTEL_GEN(dev_priv) >= 9) {
3330		size = gen8_get_total_gtt_size(snb_gmch_ctl);
3331	} else if (IS_CHERRYVIEW(dev_priv)) {
3332		size = chv_get_total_gtt_size(snb_gmch_ctl);
3333	} else {
3334		size = gen8_get_total_gtt_size(snb_gmch_ctl);
3335	}
3336
3337	ggtt->base.total = (size / sizeof(gen8_pte_t)) << PAGE_SHIFT;
3338	ggtt->base.cleanup = gen6_gmch_remove;
3339	ggtt->base.bind_vma = ggtt_bind_vma;
3340	ggtt->base.unbind_vma = ggtt_unbind_vma;
3341	ggtt->base.set_pages = ggtt_set_pages;
3342	ggtt->base.clear_pages = clear_pages;
3343	ggtt->base.insert_page = gen8_ggtt_insert_page;
3344	ggtt->base.clear_range = nop_clear_range;
3345	if (!USES_FULL_PPGTT(dev_priv) || intel_scanout_needs_vtd_wa(dev_priv))
3346		ggtt->base.clear_range = gen8_ggtt_clear_range;
3347
3348	ggtt->base.insert_entries = gen8_ggtt_insert_entries;
3349
3350	/* Serialize GTT updates with aperture access on BXT if VT-d is on. */
3351	if (intel_ggtt_update_needs_vtd_wa(dev_priv)) {
3352		ggtt->base.insert_entries = bxt_vtd_ggtt_insert_entries__BKL;
3353		ggtt->base.insert_page    = bxt_vtd_ggtt_insert_page__BKL;
3354		if (ggtt->base.clear_range != nop_clear_range)
3355			ggtt->base.clear_range = bxt_vtd_ggtt_clear_range__BKL;
3356	}
3357
3358	ggtt->invalidate = gen6_ggtt_invalidate;
3359
3360	setup_private_pat(dev_priv);
3361
3362	return ggtt_probe_common(ggtt, size);
3363}
3364
3365static int gen6_gmch_probe(struct i915_ggtt *ggtt)
3366{
3367	struct drm_i915_private *dev_priv = ggtt->base.i915;
3368	struct pci_dev *pdev = dev_priv->drm.pdev;
3369	unsigned int size;
3370	u16 snb_gmch_ctl;
3371	int err;
3372
3373	ggtt->gmadr =
3374		(struct resource) DEFINE_RES_MEM(pci_resource_start(pdev, 2),
3375						 pci_resource_len(pdev, 2));
3376	ggtt->mappable_end = resource_size(&ggtt->gmadr);
3377
3378	/* 64/512MB is the current min/max we actually know of, but this is just
3379	 * a coarse sanity check.
3380	 */
3381	if (ggtt->mappable_end < (64<<20) || ggtt->mappable_end > (512<<20)) {
3382		DRM_ERROR("Unknown GMADR size (%pa)\n", &ggtt->mappable_end);
3383		return -ENXIO;
3384	}
3385
3386	err = pci_set_dma_mask(pdev, DMA_BIT_MASK(40));
3387	if (!err)
3388		err = pci_set_consistent_dma_mask(pdev, DMA_BIT_MASK(40));
3389	if (err)
3390		DRM_ERROR("Can't set DMA mask/consistent mask (%d)\n", err);
3391	pci_read_config_word(pdev, SNB_GMCH_CTRL, &snb_gmch_ctl);
3392
3393	size = gen6_get_total_gtt_size(snb_gmch_ctl);
3394	ggtt->base.total = (size / sizeof(gen6_pte_t)) << PAGE_SHIFT;
3395
3396	ggtt->base.clear_range = gen6_ggtt_clear_range;
3397	ggtt->base.insert_page = gen6_ggtt_insert_page;
3398	ggtt->base.insert_entries = gen6_ggtt_insert_entries;
3399	ggtt->base.bind_vma = ggtt_bind_vma;
3400	ggtt->base.unbind_vma = ggtt_unbind_vma;
3401	ggtt->base.set_pages = ggtt_set_pages;
3402	ggtt->base.clear_pages = clear_pages;
3403	ggtt->base.cleanup = gen6_gmch_remove;
3404
3405	ggtt->invalidate = gen6_ggtt_invalidate;
3406
3407	if (HAS_EDRAM(dev_priv))
3408		ggtt->base.pte_encode = iris_pte_encode;
3409	else if (IS_HASWELL(dev_priv))
3410		ggtt->base.pte_encode = hsw_pte_encode;
3411	else if (IS_VALLEYVIEW(dev_priv))
3412		ggtt->base.pte_encode = byt_pte_encode;
3413	else if (INTEL_GEN(dev_priv) >= 7)
3414		ggtt->base.pte_encode = ivb_pte_encode;
3415	else
3416		ggtt->base.pte_encode = snb_pte_encode;
3417
3418	return ggtt_probe_common(ggtt, size);
3419}
3420
3421static void i915_gmch_remove(struct i915_address_space *vm)
3422{
3423	intel_gmch_remove();
3424}
3425
3426static int i915_gmch_probe(struct i915_ggtt *ggtt)
3427{
3428	struct drm_i915_private *dev_priv = ggtt->base.i915;
3429	phys_addr_t gmadr_base;
3430	int ret;
3431
3432	ret = intel_gmch_probe(dev_priv->bridge_dev, dev_priv->drm.pdev, NULL);
3433	if (!ret) {
3434		DRM_ERROR("failed to set up gmch\n");
3435		return -EIO;
3436	}
3437
3438	intel_gtt_get(&ggtt->base.total,
3439		      &gmadr_base,
3440		      &ggtt->mappable_end);
3441
3442	ggtt->gmadr =
3443		(struct resource) DEFINE_RES_MEM(gmadr_base,
3444						 ggtt->mappable_end);
3445
3446	ggtt->do_idle_maps = needs_idle_maps(dev_priv);
3447	ggtt->base.insert_page = i915_ggtt_insert_page;
3448	ggtt->base.insert_entries = i915_ggtt_insert_entries;
3449	ggtt->base.clear_range = i915_ggtt_clear_range;
3450	ggtt->base.bind_vma = ggtt_bind_vma;
3451	ggtt->base.unbind_vma = ggtt_unbind_vma;
3452	ggtt->base.set_pages = ggtt_set_pages;
3453	ggtt->base.clear_pages = clear_pages;
3454	ggtt->base.cleanup = i915_gmch_remove;
3455
3456	ggtt->invalidate = gmch_ggtt_invalidate;
3457
3458	if (unlikely(ggtt->do_idle_maps))
3459		DRM_INFO("applying Ironlake quirks for intel_iommu\n");
3460
3461	return 0;
3462}
3463
3464/**
3465 * i915_ggtt_probe_hw - Probe GGTT hardware location
3466 * @dev_priv: i915 device
3467 */
3468int i915_ggtt_probe_hw(struct drm_i915_private *dev_priv)
3469{
3470	struct i915_ggtt *ggtt = &dev_priv->ggtt;
3471	int ret;
3472
3473	ggtt->base.i915 = dev_priv;
3474	ggtt->base.dma = &dev_priv->drm.pdev->dev;
3475
3476	if (INTEL_GEN(dev_priv) <= 5)
3477		ret = i915_gmch_probe(ggtt);
3478	else if (INTEL_GEN(dev_priv) < 8)
3479		ret = gen6_gmch_probe(ggtt);
3480	else
3481		ret = gen8_gmch_probe(ggtt);
3482	if (ret)
3483		return ret;
3484
3485	/* Trim the GGTT to fit the GuC mappable upper range (when enabled).
3486	 * This is easier than doing range restriction on the fly, as we
3487	 * currently don't have any bits spare to pass in this upper
3488	 * restriction!
3489	 */
3490	if (USES_GUC(dev_priv)) {
3491		ggtt->base.total = min_t(u64, ggtt->base.total, GUC_GGTT_TOP);
3492		ggtt->mappable_end = min_t(u64, ggtt->mappable_end, ggtt->base.total);
3493	}
3494
3495	if ((ggtt->base.total - 1) >> 32) {
3496		DRM_ERROR("We never expected a Global GTT with more than 32bits"
3497			  " of address space! Found %lldM!\n",
3498			  ggtt->base.total >> 20);
3499		ggtt->base.total = 1ULL << 32;
3500		ggtt->mappable_end = min_t(u64, ggtt->mappable_end, ggtt->base.total);
3501	}
3502
3503	if (ggtt->mappable_end > ggtt->base.total) {
3504		DRM_ERROR("mappable aperture extends past end of GGTT,"
3505			  " aperture=%pa, total=%llx\n",
3506			  &ggtt->mappable_end, ggtt->base.total);
3507		ggtt->mappable_end = ggtt->base.total;
3508	}
3509
3510	/* GMADR is the PCI mmio aperture into the global GTT. */
3511	DRM_DEBUG_DRIVER("GGTT size = %lluM\n", ggtt->base.total >> 20);
3512	DRM_DEBUG_DRIVER("GMADR size = %lluM\n", (u64)ggtt->mappable_end >> 20);
3513	DRM_DEBUG_DRIVER("DSM size = %lluM\n",
3514			 (u64)resource_size(&intel_graphics_stolen_res) >> 20);
3515	if (intel_vtd_active())
3516		DRM_INFO("VT-d active for gfx access\n");
3517
3518	return 0;
3519}
3520
3521/**
3522 * i915_ggtt_init_hw - Initialize GGTT hardware
3523 * @dev_priv: i915 device
3524 */
3525int i915_ggtt_init_hw(struct drm_i915_private *dev_priv)
3526{
3527	struct i915_ggtt *ggtt = &dev_priv->ggtt;
3528	int ret;
3529
3530	INIT_LIST_HEAD(&dev_priv->vm_list);
3531
3532	/* Note that we use page colouring to enforce a guard page at the
3533	 * end of the address space. This is required as the CS may prefetch
3534	 * beyond the end of the batch buffer, across the page boundary,
3535	 * and beyond the end of the GTT if we do not provide a guard.
3536	 */
3537	mutex_lock(&dev_priv->drm.struct_mutex);
3538	i915_address_space_init(&ggtt->base, dev_priv, "[global]");
3539	if (!HAS_LLC(dev_priv) && !USES_PPGTT(dev_priv))
3540		ggtt->base.mm.color_adjust = i915_gtt_color_adjust;
3541	mutex_unlock(&dev_priv->drm.struct_mutex);
3542
3543	if (!io_mapping_init_wc(&dev_priv->ggtt.iomap,
3544				dev_priv->ggtt.gmadr.start,
3545				dev_priv->ggtt.mappable_end)) {
3546		ret = -EIO;
3547		goto out_gtt_cleanup;
3548	}
3549
3550	ggtt->mtrr = arch_phys_wc_add(ggtt->gmadr.start, ggtt->mappable_end);
3551
3552	/*
3553	 * Initialise stolen early so that we may reserve preallocated
3554	 * objects for the BIOS to KMS transition.
3555	 */
3556	ret = i915_gem_init_stolen(dev_priv);
3557	if (ret)
3558		goto out_gtt_cleanup;
3559
3560	return 0;
3561
3562out_gtt_cleanup:
3563	ggtt->base.cleanup(&ggtt->base);
3564	return ret;
3565}
3566
3567int i915_ggtt_enable_hw(struct drm_i915_private *dev_priv)
3568{
3569	if (INTEL_GEN(dev_priv) < 6 && !intel_enable_gtt())
3570		return -EIO;
3571
3572	return 0;
3573}
3574
3575void i915_ggtt_enable_guc(struct drm_i915_private *i915)
3576{
3577	GEM_BUG_ON(i915->ggtt.invalidate != gen6_ggtt_invalidate);
3578
3579	i915->ggtt.invalidate = guc_ggtt_invalidate;
3580
3581	i915_ggtt_invalidate(i915);
3582}
3583
3584void i915_ggtt_disable_guc(struct drm_i915_private *i915)
3585{
3586	/* We should only be called after i915_ggtt_enable_guc() */
3587	GEM_BUG_ON(i915->ggtt.invalidate != guc_ggtt_invalidate);
3588
3589	i915->ggtt.invalidate = gen6_ggtt_invalidate;
3590
3591	i915_ggtt_invalidate(i915);
3592}
3593
3594void i915_gem_restore_gtt_mappings(struct drm_i915_private *dev_priv)
3595{
3596	struct i915_ggtt *ggtt = &dev_priv->ggtt;
3597	struct drm_i915_gem_object *obj, *on;
3598
3599	i915_check_and_clear_faults(dev_priv);
3600
3601	/* First fill our portion of the GTT with scratch pages */
3602	ggtt->base.clear_range(&ggtt->base, 0, ggtt->base.total);
3603
3604	ggtt->base.closed = true; /* skip rewriting PTE on VMA unbind */
3605
3606	/* clflush objects bound into the GGTT and rebind them. */
3607	list_for_each_entry_safe(obj, on, &dev_priv->mm.bound_list, mm.link) {
3608		bool ggtt_bound = false;
3609		struct i915_vma *vma;
3610
3611		for_each_ggtt_vma(vma, obj) {
3612			if (!i915_vma_unbind(vma))
3613				continue;
3614
3615			WARN_ON(i915_vma_bind(vma, obj->cache_level,
3616					      PIN_UPDATE));
3617			ggtt_bound = true;
3618		}
3619
3620		if (ggtt_bound)
3621			WARN_ON(i915_gem_object_set_to_gtt_domain(obj, false));
3622	}
3623
3624	ggtt->base.closed = false;
3625
3626	if (INTEL_GEN(dev_priv) >= 8) {
3627		struct intel_ppat *ppat = &dev_priv->ppat;
3628
3629		bitmap_set(ppat->dirty, 0, ppat->max_entries);
3630		dev_priv->ppat.update_hw(dev_priv);
3631		return;
3632	}
3633
3634	if (USES_PPGTT(dev_priv)) {
3635		struct i915_address_space *vm;
3636
3637		list_for_each_entry(vm, &dev_priv->vm_list, global_link) {
3638			struct i915_hw_ppgtt *ppgtt;
3639
3640			if (i915_is_ggtt(vm))
3641				ppgtt = dev_priv->mm.aliasing_ppgtt;
3642			else
3643				ppgtt = i915_vm_to_ppgtt(vm);
3644
3645			gen6_write_page_range(ppgtt, 0, ppgtt->base.total);
3646		}
3647	}
3648
3649	i915_ggtt_invalidate(dev_priv);
3650}
3651
3652static struct scatterlist *
3653rotate_pages(const dma_addr_t *in, unsigned int offset,
3654	     unsigned int width, unsigned int height,
3655	     unsigned int stride,
3656	     struct sg_table *st, struct scatterlist *sg)
3657{
3658	unsigned int column, row;
3659	unsigned int src_idx;
3660
3661	for (column = 0; column < width; column++) {
3662		src_idx = stride * (height - 1) + column;
3663		for (row = 0; row < height; row++) {
3664			st->nents++;
3665			/* We don't need the pages, but need to initialize
3666			 * the entries so the sg list can be happily traversed.
3667			 * The only thing we need are DMA addresses.
3668			 */
3669			sg_set_page(sg, NULL, PAGE_SIZE, 0);
3670			sg_dma_address(sg) = in[offset + src_idx];
3671			sg_dma_len(sg) = PAGE_SIZE;
3672			sg = sg_next(sg);
3673			src_idx -= stride;
3674		}
3675	}
3676
3677	return sg;
3678}
3679
3680static noinline struct sg_table *
3681intel_rotate_pages(struct intel_rotation_info *rot_info,
3682		   struct drm_i915_gem_object *obj)
3683{
3684	const unsigned long n_pages = obj->base.size / PAGE_SIZE;
3685	unsigned int size = intel_rotation_info_size(rot_info);
3686	struct sgt_iter sgt_iter;
3687	dma_addr_t dma_addr;
3688	unsigned long i;
3689	dma_addr_t *page_addr_list;
3690	struct sg_table *st;
3691	struct scatterlist *sg;
3692	int ret = -ENOMEM;
3693
3694	/* Allocate a temporary list of source pages for random access. */
3695	page_addr_list = kvmalloc_array(n_pages,
3696					sizeof(dma_addr_t),
3697					GFP_KERNEL);
3698	if (!page_addr_list)
3699		return ERR_PTR(ret);
3700
3701	/* Allocate target SG list. */
3702	st = kmalloc(sizeof(*st), GFP_KERNEL);
3703	if (!st)
3704		goto err_st_alloc;
3705
3706	ret = sg_alloc_table(st, size, GFP_KERNEL);
3707	if (ret)
3708		goto err_sg_alloc;
3709
3710	/* Populate source page list from the object. */
3711	i = 0;
3712	for_each_sgt_dma(dma_addr, sgt_iter, obj->mm.pages)
3713		page_addr_list[i++] = dma_addr;
3714
3715	GEM_BUG_ON(i != n_pages);
3716	st->nents = 0;
3717	sg = st->sgl;
3718
3719	for (i = 0 ; i < ARRAY_SIZE(rot_info->plane); i++) {
3720		sg = rotate_pages(page_addr_list, rot_info->plane[i].offset,
3721				  rot_info->plane[i].width, rot_info->plane[i].height,
3722				  rot_info->plane[i].stride, st, sg);
3723	}
3724
3725	kvfree(page_addr_list);
3726
3727	return st;
3728
3729err_sg_alloc:
3730	kfree(st);
3731err_st_alloc:
3732	kvfree(page_addr_list);
3733
3734	DRM_DEBUG_DRIVER("Failed to create rotated mapping for object size %zu! (%ux%u tiles, %u pages)\n",
3735			 obj->base.size, rot_info->plane[0].width, rot_info->plane[0].height, size);
3736
3737	return ERR_PTR(ret);
3738}
3739
3740static noinline struct sg_table *
3741intel_partial_pages(const struct i915_ggtt_view *view,
3742		    struct drm_i915_gem_object *obj)
3743{
3744	struct sg_table *st;
3745	struct scatterlist *sg, *iter;
3746	unsigned int count = view->partial.size;
3747	unsigned int offset;
3748	int ret = -ENOMEM;
3749
3750	st = kmalloc(sizeof(*st), GFP_KERNEL);
3751	if (!st)
3752		goto err_st_alloc;
3753
3754	ret = sg_alloc_table(st, count, GFP_KERNEL);
3755	if (ret)
3756		goto err_sg_alloc;
3757
3758	iter = i915_gem_object_get_sg(obj, view->partial.offset, &offset);
3759	GEM_BUG_ON(!iter);
3760
3761	sg = st->sgl;
3762	st->nents = 0;
3763	do {
3764		unsigned int len;
3765
3766		len = min(iter->length - (offset << PAGE_SHIFT),
3767			  count << PAGE_SHIFT);
3768		sg_set_page(sg, NULL, len, 0);
3769		sg_dma_address(sg) =
3770			sg_dma_address(iter) + (offset << PAGE_SHIFT);
3771		sg_dma_len(sg) = len;
3772
3773		st->nents++;
3774		count -= len >> PAGE_SHIFT;
3775		if (count == 0) {
3776			sg_mark_end(sg);
3777			return st;
3778		}
3779
3780		sg = __sg_next(sg);
3781		iter = __sg_next(iter);
3782		offset = 0;
3783	} while (1);
3784
3785err_sg_alloc:
3786	kfree(st);
3787err_st_alloc:
3788	return ERR_PTR(ret);
3789}
3790
3791static int
3792i915_get_ggtt_vma_pages(struct i915_vma *vma)
3793{
3794	int ret;
3795
3796	/* The vma->pages are only valid within the lifespan of the borrowed
3797	 * obj->mm.pages. When the obj->mm.pages sg_table is regenerated, so
3798	 * must be the vma->pages. A simple rule is that vma->pages must only
3799	 * be accessed when the obj->mm.pages are pinned.
3800	 */
3801	GEM_BUG_ON(!i915_gem_object_has_pinned_pages(vma->obj));
3802
3803	switch (vma->ggtt_view.type) {
3804	default:
3805		GEM_BUG_ON(vma->ggtt_view.type);
3806		/* fall through */
3807	case I915_GGTT_VIEW_NORMAL:
3808		vma->pages = vma->obj->mm.pages;
3809		return 0;
3810
3811	case I915_GGTT_VIEW_ROTATED:
3812		vma->pages =
3813			intel_rotate_pages(&vma->ggtt_view.rotated, vma->obj);
3814		break;
3815
3816	case I915_GGTT_VIEW_PARTIAL:
3817		vma->pages = intel_partial_pages(&vma->ggtt_view, vma->obj);
3818		break;
3819	}
3820
3821	ret = 0;
3822	if (unlikely(IS_ERR(vma->pages))) {
3823		ret = PTR_ERR(vma->pages);
3824		vma->pages = NULL;
3825		DRM_ERROR("Failed to get pages for VMA view type %u (%d)!\n",
3826			  vma->ggtt_view.type, ret);
3827	}
3828	return ret;
3829}
3830
3831/**
3832 * i915_gem_gtt_reserve - reserve a node in an address_space (GTT)
3833 * @vm: the &struct i915_address_space
3834 * @node: the &struct drm_mm_node (typically i915_vma.mode)
3835 * @size: how much space to allocate inside the GTT,
3836 *        must be #I915_GTT_PAGE_SIZE aligned
3837 * @offset: where to insert inside the GTT,
3838 *          must be #I915_GTT_MIN_ALIGNMENT aligned, and the node
3839 *          (@offset + @size) must fit within the address space
3840 * @color: color to apply to node, if this node is not from a VMA,
3841 *         color must be #I915_COLOR_UNEVICTABLE
3842 * @flags: control search and eviction behaviour
3843 *
3844 * i915_gem_gtt_reserve() tries to insert the @node at the exact @offset inside
3845 * the address space (using @size and @color). If the @node does not fit, it
3846 * tries to evict any overlapping nodes from the GTT, including any
3847 * neighbouring nodes if the colors do not match (to ensure guard pages between
3848 * differing domains). See i915_gem_evict_for_node() for the gory details
3849 * on the eviction algorithm. #PIN_NONBLOCK may used to prevent waiting on
3850 * evicting active overlapping objects, and any overlapping node that is pinned
3851 * or marked as unevictable will also result in failure.
3852 *
3853 * Returns: 0 on success, -ENOSPC if no suitable hole is found, -EINTR if
3854 * asked to wait for eviction and interrupted.
3855 */
3856int i915_gem_gtt_reserve(struct i915_address_space *vm,
3857			 struct drm_mm_node *node,
3858			 u64 size, u64 offset, unsigned long color,
3859			 unsigned int flags)
3860{
3861	int err;
3862
3863	GEM_BUG_ON(!size);
3864	GEM_BUG_ON(!IS_ALIGNED(size, I915_GTT_PAGE_SIZE));
3865	GEM_BUG_ON(!IS_ALIGNED(offset, I915_GTT_MIN_ALIGNMENT));
3866	GEM_BUG_ON(range_overflows(offset, size, vm->total));
3867	GEM_BUG_ON(vm == &vm->i915->mm.aliasing_ppgtt->base);
3868	GEM_BUG_ON(drm_mm_node_allocated(node));
3869
3870	node->size = size;
3871	node->start = offset;
3872	node->color = color;
3873
3874	err = drm_mm_reserve_node(&vm->mm, node);
3875	if (err != -ENOSPC)
3876		return err;
3877
3878	if (flags & PIN_NOEVICT)
3879		return -ENOSPC;
3880
3881	err = i915_gem_evict_for_node(vm, node, flags);
3882	if (err == 0)
3883		err = drm_mm_reserve_node(&vm->mm, node);
3884
3885	return err;
3886}
3887
3888static u64 random_offset(u64 start, u64 end, u64 len, u64 align)
3889{
3890	u64 range, addr;
3891
3892	GEM_BUG_ON(range_overflows(start, len, end));
3893	GEM_BUG_ON(round_up(start, align) > round_down(end - len, align));
3894
3895	range = round_down(end - len, align) - round_up(start, align);
3896	if (range) {
3897		if (sizeof(unsigned long) == sizeof(u64)) {
3898			addr = get_random_long();
3899		} else {
3900			addr = get_random_int();
3901			if (range > U32_MAX) {
3902				addr <<= 32;
3903				addr |= get_random_int();
3904			}
3905		}
3906		div64_u64_rem(addr, range, &addr);
3907		start += addr;
3908	}
3909
3910	return round_up(start, align);
3911}
3912
3913/**
3914 * i915_gem_gtt_insert - insert a node into an address_space (GTT)
3915 * @vm: the &struct i915_address_space
3916 * @node: the &struct drm_mm_node (typically i915_vma.node)
3917 * @size: how much space to allocate inside the GTT,
3918 *        must be #I915_GTT_PAGE_SIZE aligned
3919 * @alignment: required alignment of starting offset, may be 0 but
3920 *             if specified, this must be a power-of-two and at least
3921 *             #I915_GTT_MIN_ALIGNMENT
3922 * @color: color to apply to node
3923 * @start: start of any range restriction inside GTT (0 for all),
3924 *         must be #I915_GTT_PAGE_SIZE aligned
3925 * @end: end of any range restriction inside GTT (U64_MAX for all),
3926 *       must be #I915_GTT_PAGE_SIZE aligned if not U64_MAX
3927 * @flags: control search and eviction behaviour
3928 *
3929 * i915_gem_gtt_insert() first searches for an available hole into which
3930 * is can insert the node. The hole address is aligned to @alignment and
3931 * its @size must then fit entirely within the [@start, @end] bounds. The
3932 * nodes on either side of the hole must match @color, or else a guard page
3933 * will be inserted between the two nodes (or the node evicted). If no
3934 * suitable hole is found, first a victim is randomly selected and tested
3935 * for eviction, otherwise then the LRU list of objects within the GTT
3936 * is scanned to find the first set of replacement nodes to create the hole.
3937 * Those old overlapping nodes are evicted from the GTT (and so must be
3938 * rebound before any future use). Any node that is currently pinned cannot
3939 * be evicted (see i915_vma_pin()). Similar if the node's VMA is currently
3940 * active and #PIN_NONBLOCK is specified, that node is also skipped when
3941 * searching for an eviction candidate. See i915_gem_evict_something() for
3942 * the gory details on the eviction algorithm.
3943 *
3944 * Returns: 0 on success, -ENOSPC if no suitable hole is found, -EINTR if
3945 * asked to wait for eviction and interrupted.
3946 */
3947int i915_gem_gtt_insert(struct i915_address_space *vm,
3948			struct drm_mm_node *node,
3949			u64 size, u64 alignment, unsigned long color,
3950			u64 start, u64 end, unsigned int flags)
3951{
3952	enum drm_mm_insert_mode mode;
3953	u64 offset;
3954	int err;
3955
3956	lockdep_assert_held(&vm->i915->drm.struct_mutex);
3957	GEM_BUG_ON(!size);
3958	GEM_BUG_ON(!IS_ALIGNED(size, I915_GTT_PAGE_SIZE));
3959	GEM_BUG_ON(alignment && !is_power_of_2(alignment));
3960	GEM_BUG_ON(alignment && !IS_ALIGNED(alignment, I915_GTT_MIN_ALIGNMENT));
3961	GEM_BUG_ON(start >= end);
3962	GEM_BUG_ON(start > 0  && !IS_ALIGNED(start, I915_GTT_PAGE_SIZE));
3963	GEM_BUG_ON(end < U64_MAX && !IS_ALIGNED(end, I915_GTT_PAGE_SIZE));
3964	GEM_BUG_ON(vm == &vm->i915->mm.aliasing_ppgtt->base);
3965	GEM_BUG_ON(drm_mm_node_allocated(node));
3966
3967	if (unlikely(range_overflows(start, size, end)))
3968		return -ENOSPC;
3969
3970	if (unlikely(round_up(start, alignment) > round_down(end - size, alignment)))
3971		return -ENOSPC;
3972
3973	mode = DRM_MM_INSERT_BEST;
3974	if (flags & PIN_HIGH)
3975		mode = DRM_MM_INSERT_HIGH;
3976	if (flags & PIN_MAPPABLE)
3977		mode = DRM_MM_INSERT_LOW;
3978
3979	/* We only allocate in PAGE_SIZE/GTT_PAGE_SIZE (4096) chunks,
3980	 * so we know that we always have a minimum alignment of 4096.
3981	 * The drm_mm range manager is optimised to return results
3982	 * with zero alignment, so where possible use the optimal
3983	 * path.
3984	 */
3985	BUILD_BUG_ON(I915_GTT_MIN_ALIGNMENT > I915_GTT_PAGE_SIZE);
3986	if (alignment <= I915_GTT_MIN_ALIGNMENT)
3987		alignment = 0;
3988
3989	err = drm_mm_insert_node_in_range(&vm->mm, node,
3990					  size, alignment, color,
3991					  start, end, mode);
3992	if (err != -ENOSPC)
3993		return err;
3994
3995	if (flags & PIN_NOEVICT)
3996		return -ENOSPC;
3997
3998	/* No free space, pick a slot at random.
3999	 *
4000	 * There is a pathological case here using a GTT shared between
4001	 * mmap and GPU (i.e. ggtt/aliasing_ppgtt but not full-ppgtt):
4002	 *
4003	 *    |<-- 256 MiB aperture -->||<-- 1792 MiB unmappable -->|
4004	 *         (64k objects)             (448k objects)
4005	 *
4006	 * Now imagine that the eviction LRU is ordered top-down (just because
4007	 * pathology meets real life), and that we need to evict an object to
4008	 * make room inside the aperture. The eviction scan then has to walk
4009	 * the 448k list before it finds one within range. And now imagine that
4010	 * it has to search for a new hole between every byte inside the memcpy,
4011	 * for several simultaneous clients.
4012	 *
4013	 * On a full-ppgtt system, if we have run out of available space, there
4014	 * will be lots and lots of objects in the eviction list! Again,
4015	 * searching that LRU list may be slow if we are also applying any
4016	 * range restrictions (e.g. restriction to low 4GiB) and so, for
4017	 * simplicity and similarilty between different GTT, try the single
4018	 * random replacement first.
4019	 */
4020	offset = random_offset(start, end,
4021			       size, alignment ?: I915_GTT_MIN_ALIGNMENT);
4022	err = i915_gem_gtt_reserve(vm, node, size, offset, color, flags);
4023	if (err != -ENOSPC)
4024		return err;
4025
4026	/* Randomly selected placement is pinned, do a search */
4027	err = i915_gem_evict_something(vm, size, alignment, color,
4028				       start, end, flags);
4029	if (err)
4030		return err;
4031
4032	return drm_mm_insert_node_in_range(&vm->mm, node,
4033					   size, alignment, color,
4034					   start, end, DRM_MM_INSERT_EVICT);
4035}
4036
4037#if IS_ENABLED(CONFIG_DRM_I915_SELFTEST)
4038#include "selftests/mock_gtt.c"
4039#include "selftests/i915_gem_gtt.c"
4040#endif