1 /*
2  * SPDX-License-Identifier: MIT
3  *
4  * Copyright © 2016 Intel Corporation
5  */
6 
7 #ifndef __I915_GEM_OBJECT_TYPES_H__
8 #define __I915_GEM_OBJECT_TYPES_H__
9 
10 #include <linux/mmu_notifier.h>
11 
12 #include <drm/drm_gem.h>
13 #include <uapi/drm/i915_drm.h>
14 
15 #include "i915_active.h"
16 #include "i915_selftest.h"
17 
18 struct drm_i915_gem_object;
19 struct intel_fronbuffer;
20 
21 /*
22  * struct i915_lut_handle tracks the fast lookups from handle to vma used
23  * for execbuf. Although we use a radixtree for that mapping, in order to
24  * remove them as the object or context is closed, we need a secondary list
25  * and a translation entry (i915_lut_handle).
26  */
27 struct i915_lut_handle {
28 	struct list_head obj_link;
29 	struct i915_gem_context *ctx;
30 	u32 handle;
31 };
32 
33 struct drm_i915_gem_object_ops {
34 	unsigned int flags;
35 #define I915_GEM_OBJECT_HAS_IOMEM	BIT(1)
36 #define I915_GEM_OBJECT_IS_SHRINKABLE	BIT(2)
37 #define I915_GEM_OBJECT_IS_PROXY	BIT(3)
38 #define I915_GEM_OBJECT_NO_MMAP		BIT(4)
39 
40 	/* Interface between the GEM object and its backing storage.
41 	 * get_pages() is called once prior to the use of the associated set
42 	 * of pages before to binding them into the GTT, and put_pages() is
43 	 * called after we no longer need them. As we expect there to be
44 	 * associated cost with migrating pages between the backing storage
45 	 * and making them available for the GPU (e.g. clflush), we may hold
46 	 * onto the pages after they are no longer referenced by the GPU
47 	 * in case they may be used again shortly (for example migrating the
48 	 * pages to a different memory domain within the GTT). put_pages()
49 	 * will therefore most likely be called when the object itself is
50 	 * being released or under memory pressure (where we attempt to
51 	 * reap pages for the shrinker).
52 	 */
53 	int (*get_pages)(struct drm_i915_gem_object *obj);
54 	void (*put_pages)(struct drm_i915_gem_object *obj,
55 			  struct sg_table *pages);
56 	void (*truncate)(struct drm_i915_gem_object *obj);
57 	void (*writeback)(struct drm_i915_gem_object *obj);
58 
59 	int (*pread)(struct drm_i915_gem_object *obj,
60 		     const struct drm_i915_gem_pread *arg);
61 	int (*pwrite)(struct drm_i915_gem_object *obj,
62 		      const struct drm_i915_gem_pwrite *arg);
63 
64 	int (*dmabuf_export)(struct drm_i915_gem_object *obj);
65 	void (*release)(struct drm_i915_gem_object *obj);
66 
67 	const char *name; /* friendly name for debug, e.g. lockdep classes */
68 };
69 
70 enum i915_map_type {
71 	I915_MAP_WB = 0,
72 	I915_MAP_WC,
73 #define I915_MAP_OVERRIDE BIT(31)
74 	I915_MAP_FORCE_WB = I915_MAP_WB | I915_MAP_OVERRIDE,
75 	I915_MAP_FORCE_WC = I915_MAP_WC | I915_MAP_OVERRIDE,
76 };
77 
78 enum i915_mmap_type {
79 	I915_MMAP_TYPE_GTT = 0,
80 	I915_MMAP_TYPE_WC,
81 	I915_MMAP_TYPE_WB,
82 	I915_MMAP_TYPE_UC,
83 };
84 
85 struct i915_mmap_offset {
86 	struct drm_vma_offset_node vma_node;
87 	struct drm_i915_gem_object *obj;
88 	enum i915_mmap_type mmap_type;
89 
90 	struct rb_node offset;
91 };
92 
93 struct i915_gem_object_page_iter {
94 	struct scatterlist *sg_pos;
95 	unsigned int sg_idx; /* in pages, but 32bit eek! */
96 
97 	struct radix_tree_root radix;
98 	struct mutex lock; /* protects this cache */
99 };
100 
101 struct drm_i915_gem_object {
102 	struct drm_gem_object base;
103 
104 	const struct drm_i915_gem_object_ops *ops;
105 
106 	struct {
107 		/**
108 		 * @vma.lock: protect the list/tree of vmas
109 		 */
110 		spinlock_t lock;
111 
112 		/**
113 		 * @vma.list: List of VMAs backed by this object
114 		 *
115 		 * The VMA on this list are ordered by type, all GGTT vma are
116 		 * placed at the head and all ppGTT vma are placed at the tail.
117 		 * The different types of GGTT vma are unordered between
118 		 * themselves, use the @vma.tree (which has a defined order
119 		 * between all VMA) to quickly find an exact match.
120 		 */
121 		struct list_head list;
122 
123 		/**
124 		 * @vma.tree: Ordered tree of VMAs backed by this object
125 		 *
126 		 * All VMA created for this object are placed in the @vma.tree
127 		 * for fast retrieval via a binary search in
128 		 * i915_vma_instance(). They are also added to @vma.list for
129 		 * easy iteration.
130 		 */
131 		struct rb_root tree;
132 	} vma;
133 
134 	/**
135 	 * @lut_list: List of vma lookup entries in use for this object.
136 	 *
137 	 * If this object is closed, we need to remove all of its VMA from
138 	 * the fast lookup index in associated contexts; @lut_list provides
139 	 * this translation from object to context->handles_vma.
140 	 */
141 	struct list_head lut_list;
142 	spinlock_t lut_lock; /* guards lut_list */
143 
144 	/**
145 	 * @obj_link: Link into @i915_gem_ww_ctx.obj_list
146 	 *
147 	 * When we lock this object through i915_gem_object_lock() with a
148 	 * context, we add it to the list to ensure we can unlock everything
149 	 * when i915_gem_ww_ctx_backoff() or i915_gem_ww_ctx_fini() are called.
150 	 */
151 	struct list_head obj_link;
152 
153 	union {
154 		struct rcu_head rcu;
155 		struct llist_node freed;
156 	};
157 
158 	/**
159 	 * Whether the object is currently in the GGTT mmap.
160 	 */
161 	unsigned int userfault_count;
162 	struct list_head userfault_link;
163 
164 	struct {
165 		spinlock_t lock; /* Protects access to mmo offsets */
166 		struct rb_root offsets;
167 	} mmo;
168 
169 	I915_SELFTEST_DECLARE(struct list_head st_link);
170 
171 	unsigned long flags;
172 #define I915_BO_ALLOC_CONTIGUOUS BIT(0)
173 #define I915_BO_ALLOC_VOLATILE   BIT(1)
174 #define I915_BO_ALLOC_STRUCT_PAGE BIT(2)
175 #define I915_BO_ALLOC_FLAGS (I915_BO_ALLOC_CONTIGUOUS | \
176 			     I915_BO_ALLOC_VOLATILE | \
177 			     I915_BO_ALLOC_STRUCT_PAGE)
178 #define I915_BO_READONLY         BIT(3)
179 #define I915_TILING_QUIRK_BIT    4 /* unknown swizzling; do not release! */
180 
181 	/*
182 	 * Is the object to be mapped as read-only to the GPU
183 	 * Only honoured if hardware has relevant pte bit
184 	 */
185 	unsigned int cache_level:3;
186 	unsigned int cache_coherent:2;
187 #define I915_BO_CACHE_COHERENT_FOR_READ BIT(0)
188 #define I915_BO_CACHE_COHERENT_FOR_WRITE BIT(1)
189 	unsigned int cache_dirty:1;
190 
191 	/**
192 	 * @read_domains: Read memory domains.
193 	 *
194 	 * These monitor which caches contain read/write data related to the
195 	 * object. When transitioning from one set of domains to another,
196 	 * the driver is called to ensure that caches are suitably flushed and
197 	 * invalidated.
198 	 */
199 	u16 read_domains;
200 
201 	/**
202 	 * @write_domain: Corresponding unique write memory domain.
203 	 */
204 	u16 write_domain;
205 
206 	struct intel_frontbuffer __rcu *frontbuffer;
207 
208 	/** Current tiling stride for the object, if it's tiled. */
209 	unsigned int tiling_and_stride;
210 #define FENCE_MINIMUM_STRIDE 128 /* See i915_tiling_ok() */
211 #define TILING_MASK (FENCE_MINIMUM_STRIDE - 1)
212 #define STRIDE_MASK (~TILING_MASK)
213 
214 	struct {
215 		/*
216 		 * Protects the pages and their use. Do not use directly, but
217 		 * instead go through the pin/unpin interfaces.
218 		 */
219 		atomic_t pages_pin_count;
220 		atomic_t shrink_pin;
221 
222 		/**
223 		 * Memory region for this object.
224 		 */
225 		struct intel_memory_region *region;
226 		/**
227 		 * List of memory region blocks allocated for this object.
228 		 */
229 		struct list_head blocks;
230 		/**
231 		 * Element within memory_region->objects or region->purgeable
232 		 * if the object is marked as DONTNEED. Access is protected by
233 		 * region->obj_lock.
234 		 */
235 		struct list_head region_link;
236 
237 		struct sg_table *pages;
238 		void *mapping;
239 
240 		struct i915_page_sizes {
241 			/**
242 			 * The sg mask of the pages sg_table. i.e the mask of
243 			 * of the lengths for each sg entry.
244 			 */
245 			unsigned int phys;
246 
247 			/**
248 			 * The gtt page sizes we are allowed to use given the
249 			 * sg mask and the supported page sizes. This will
250 			 * express the smallest unit we can use for the whole
251 			 * object, as well as the larger sizes we may be able
252 			 * to use opportunistically.
253 			 */
254 			unsigned int sg;
255 
256 			/**
257 			 * The actual gtt page size usage. Since we can have
258 			 * multiple vma associated with this object we need to
259 			 * prevent any trampling of state, hence a copy of this
260 			 * struct also lives in each vma, therefore the gtt
261 			 * value here should only be read/write through the vma.
262 			 */
263 			unsigned int gtt;
264 		} page_sizes;
265 
266 		I915_SELFTEST_DECLARE(unsigned int page_mask);
267 
268 		struct i915_gem_object_page_iter get_page;
269 		struct i915_gem_object_page_iter get_dma_page;
270 
271 		/**
272 		 * Element within i915->mm.unbound_list or i915->mm.bound_list,
273 		 * locked by i915->mm.obj_lock.
274 		 */
275 		struct list_head link;
276 
277 		/**
278 		 * Advice: are the backing pages purgeable?
279 		 */
280 		unsigned int madv:2;
281 
282 		/**
283 		 * This is set if the object has been written to since the
284 		 * pages were last acquired.
285 		 */
286 		bool dirty:1;
287 	} mm;
288 
289 	/** Record of address bit 17 of each page at last unbind. */
290 	unsigned long *bit_17;
291 
292 	union {
293 #ifdef CONFIG_MMU_NOTIFIER
294 		struct i915_gem_userptr {
295 			uintptr_t ptr;
296 			unsigned long notifier_seq;
297 
298 			struct mmu_interval_notifier notifier;
299 			struct page **pvec;
300 			int page_ref;
301 		} userptr;
302 #endif
303 
304 		struct drm_mm_node *stolen;
305 
306 		unsigned long scratch;
307 		u64 encode;
308 
309 		void *gvt_info;
310 	};
311 };
312 
313 static inline struct drm_i915_gem_object *
to_intel_bo(struct drm_gem_object * gem)314 to_intel_bo(struct drm_gem_object *gem)
315 {
316 	/* Assert that to_intel_bo(NULL) == NULL */
317 	BUILD_BUG_ON(offsetof(struct drm_i915_gem_object, base));
318 
319 	return container_of(gem, struct drm_i915_gem_object, base);
320 }
321 
322 #endif
323