xref: /linux/drivers/gpu/drm/xe/xe_bo.h (revision ddc73c46)
1 /* SPDX-License-Identifier: MIT */
2 /*
3  * Copyright © 2021 Intel Corporation
4  */
5 
6 #ifndef _XE_BO_H_
7 #define _XE_BO_H_
8 
9 #include <drm/ttm/ttm_tt.h>
10 
11 #include "xe_bo_types.h"
12 #include "xe_macros.h"
13 #include "xe_vm_types.h"
14 #include "xe_vm.h"
15 
16 #define XE_DEFAULT_GTT_SIZE_MB          3072ULL /* 3GB by default */
17 
18 #define XE_BO_FLAG_USER		BIT(0)
19 /* The bits below need to be contiguous, or things break */
20 #define XE_BO_FLAG_SYSTEM		BIT(1)
21 #define XE_BO_FLAG_VRAM0		BIT(2)
22 #define XE_BO_FLAG_VRAM1		BIT(3)
23 #define XE_BO_FLAG_VRAM_MASK		(XE_BO_FLAG_VRAM0 | XE_BO_FLAG_VRAM1)
24 /* -- */
25 #define XE_BO_FLAG_STOLEN		BIT(4)
26 #define XE_BO_FLAG_VRAM_IF_DGFX(tile)	(IS_DGFX(tile_to_xe(tile)) ? \
27 					 XE_BO_FLAG_VRAM0 << (tile)->id : \
28 					 XE_BO_FLAG_SYSTEM)
29 #define XE_BO_FLAG_GGTT			BIT(5)
30 #define XE_BO_FLAG_IGNORE_MIN_PAGE_SIZE BIT(6)
31 #define XE_BO_FLAG_PINNED		BIT(7)
32 #define XE_BO_FLAG_NO_RESV_EVICT	BIT(8)
33 #define XE_BO_FLAG_DEFER_BACKING	BIT(9)
34 #define XE_BO_FLAG_SCANOUT		BIT(10)
35 #define XE_BO_FLAG_FIXED_PLACEMENT	BIT(11)
36 #define XE_BO_FLAG_PAGETABLE		BIT(12)
37 #define XE_BO_FLAG_NEEDS_CPU_ACCESS	BIT(13)
38 #define XE_BO_FLAG_NEEDS_UC		BIT(14)
39 #define XE_BO_FLAG_NEEDS_64K		BIT(15)
40 #define XE_BO_FLAG_NEEDS_2M		BIT(16)
41 #define XE_BO_FLAG_GGTT_INVALIDATE	BIT(17)
42 /* this one is trigger internally only */
43 #define XE_BO_FLAG_INTERNAL_TEST	BIT(30)
44 #define XE_BO_FLAG_INTERNAL_64K		BIT(31)
45 
46 #define XE_PTE_SHIFT			12
47 #define XE_PAGE_SIZE			(1 << XE_PTE_SHIFT)
48 #define XE_PTE_MASK			(XE_PAGE_SIZE - 1)
49 #define XE_PDE_SHIFT			(XE_PTE_SHIFT - 3)
50 #define XE_PDES				(1 << XE_PDE_SHIFT)
51 #define XE_PDE_MASK			(XE_PDES - 1)
52 
53 #define XE_64K_PTE_SHIFT		16
54 #define XE_64K_PAGE_SIZE		(1 << XE_64K_PTE_SHIFT)
55 #define XE_64K_PTE_MASK			(XE_64K_PAGE_SIZE - 1)
56 #define XE_64K_PDE_MASK			(XE_PDE_MASK >> 4)
57 
58 #define XE_PL_SYSTEM		TTM_PL_SYSTEM
59 #define XE_PL_TT		TTM_PL_TT
60 #define XE_PL_VRAM0		TTM_PL_VRAM
61 #define XE_PL_VRAM1		(XE_PL_VRAM0 + 1)
62 #define XE_PL_STOLEN		(TTM_NUM_MEM_TYPES - 1)
63 
64 #define XE_BO_PROPS_INVALID	(-1)
65 
66 struct sg_table;
67 
68 struct xe_bo *xe_bo_alloc(void);
69 void xe_bo_free(struct xe_bo *bo);
70 
71 struct xe_bo *___xe_bo_create_locked(struct xe_device *xe, struct xe_bo *bo,
72 				     struct xe_tile *tile, struct dma_resv *resv,
73 				     struct ttm_lru_bulk_move *bulk, size_t size,
74 				     u16 cpu_caching, enum ttm_bo_type type,
75 				     u32 flags);
76 struct xe_bo *
77 xe_bo_create_locked_range(struct xe_device *xe,
78 			  struct xe_tile *tile, struct xe_vm *vm,
79 			  size_t size, u64 start, u64 end,
80 			  enum ttm_bo_type type, u32 flags);
81 struct xe_bo *xe_bo_create_locked(struct xe_device *xe, struct xe_tile *tile,
82 				  struct xe_vm *vm, size_t size,
83 				  enum ttm_bo_type type, u32 flags);
84 struct xe_bo *xe_bo_create(struct xe_device *xe, struct xe_tile *tile,
85 			   struct xe_vm *vm, size_t size,
86 			   enum ttm_bo_type type, u32 flags);
87 struct xe_bo *xe_bo_create_user(struct xe_device *xe, struct xe_tile *tile,
88 				struct xe_vm *vm, size_t size,
89 				u16 cpu_caching,
90 				u32 flags);
91 struct xe_bo *xe_bo_create_pin_map(struct xe_device *xe, struct xe_tile *tile,
92 				   struct xe_vm *vm, size_t size,
93 				   enum ttm_bo_type type, u32 flags);
94 struct xe_bo *xe_bo_create_pin_map_at(struct xe_device *xe, struct xe_tile *tile,
95 				      struct xe_vm *vm, size_t size, u64 offset,
96 				      enum ttm_bo_type type, u32 flags);
97 struct xe_bo *xe_bo_create_from_data(struct xe_device *xe, struct xe_tile *tile,
98 				     const void *data, size_t size,
99 				     enum ttm_bo_type type, u32 flags);
100 struct xe_bo *xe_managed_bo_create_pin_map(struct xe_device *xe, struct xe_tile *tile,
101 					   size_t size, u32 flags);
102 struct xe_bo *xe_managed_bo_create_from_data(struct xe_device *xe, struct xe_tile *tile,
103 					     const void *data, size_t size, u32 flags);
104 int xe_managed_bo_reinit_in_vram(struct xe_device *xe, struct xe_tile *tile, struct xe_bo **src);
105 
106 int xe_bo_placement_for_flags(struct xe_device *xe, struct xe_bo *bo,
107 			      u32 bo_flags);
108 
ttm_to_xe_bo(const struct ttm_buffer_object * bo)109 static inline struct xe_bo *ttm_to_xe_bo(const struct ttm_buffer_object *bo)
110 {
111 	return container_of(bo, struct xe_bo, ttm);
112 }
113 
gem_to_xe_bo(const struct drm_gem_object * obj)114 static inline struct xe_bo *gem_to_xe_bo(const struct drm_gem_object *obj)
115 {
116 	return container_of(obj, struct xe_bo, ttm.base);
117 }
118 
119 #define xe_bo_device(bo) ttm_to_xe_device((bo)->ttm.bdev)
120 
xe_bo_get(struct xe_bo * bo)121 static inline struct xe_bo *xe_bo_get(struct xe_bo *bo)
122 {
123 	if (bo)
124 		drm_gem_object_get(&bo->ttm.base);
125 
126 	return bo;
127 }
128 
129 void xe_bo_put(struct xe_bo *bo);
130 
__xe_bo_unset_bulk_move(struct xe_bo * bo)131 static inline void __xe_bo_unset_bulk_move(struct xe_bo *bo)
132 {
133 	if (bo)
134 		ttm_bo_set_bulk_move(&bo->ttm, NULL);
135 }
136 
xe_bo_assert_held(struct xe_bo * bo)137 static inline void xe_bo_assert_held(struct xe_bo *bo)
138 {
139 	if (bo)
140 		dma_resv_assert_held((bo)->ttm.base.resv);
141 }
142 
143 int xe_bo_lock(struct xe_bo *bo, bool intr);
144 
145 void xe_bo_unlock(struct xe_bo *bo);
146 
xe_bo_unlock_vm_held(struct xe_bo * bo)147 static inline void xe_bo_unlock_vm_held(struct xe_bo *bo)
148 {
149 	if (bo) {
150 		XE_WARN_ON(bo->vm && bo->ttm.base.resv != xe_vm_resv(bo->vm));
151 		if (bo->vm)
152 			xe_vm_assert_held(bo->vm);
153 		else
154 			dma_resv_unlock(bo->ttm.base.resv);
155 	}
156 }
157 
158 int xe_bo_pin_external(struct xe_bo *bo);
159 int xe_bo_pin(struct xe_bo *bo);
160 void xe_bo_unpin_external(struct xe_bo *bo);
161 void xe_bo_unpin(struct xe_bo *bo);
162 int xe_bo_validate(struct xe_bo *bo, struct xe_vm *vm, bool allow_res_evict);
163 
xe_bo_is_pinned(struct xe_bo * bo)164 static inline bool xe_bo_is_pinned(struct xe_bo *bo)
165 {
166 	return bo->ttm.pin_count;
167 }
168 
xe_bo_unpin_map_no_vm(struct xe_bo * bo)169 static inline void xe_bo_unpin_map_no_vm(struct xe_bo *bo)
170 {
171 	if (likely(bo)) {
172 		xe_bo_lock(bo, false);
173 		xe_bo_unpin(bo);
174 		xe_bo_unlock(bo);
175 
176 		xe_bo_put(bo);
177 	}
178 }
179 
180 bool xe_bo_is_xe_bo(struct ttm_buffer_object *bo);
181 dma_addr_t __xe_bo_addr(struct xe_bo *bo, u64 offset, size_t page_size);
182 dma_addr_t xe_bo_addr(struct xe_bo *bo, u64 offset, size_t page_size);
183 
184 static inline dma_addr_t
xe_bo_main_addr(struct xe_bo * bo,size_t page_size)185 xe_bo_main_addr(struct xe_bo *bo, size_t page_size)
186 {
187 	return xe_bo_addr(bo, 0, page_size);
188 }
189 
190 static inline u32
xe_bo_ggtt_addr(struct xe_bo * bo)191 xe_bo_ggtt_addr(struct xe_bo *bo)
192 {
193 	if (XE_WARN_ON(!bo->ggtt_node))
194 		return 0;
195 
196 	XE_WARN_ON(bo->ggtt_node->base.size > bo->size);
197 	XE_WARN_ON(bo->ggtt_node->base.start + bo->ggtt_node->base.size > (1ull << 32));
198 	return bo->ggtt_node->base.start;
199 }
200 
201 int xe_bo_vmap(struct xe_bo *bo);
202 void xe_bo_vunmap(struct xe_bo *bo);
203 
204 bool mem_type_is_vram(u32 mem_type);
205 bool xe_bo_is_vram(struct xe_bo *bo);
206 bool xe_bo_is_stolen(struct xe_bo *bo);
207 bool xe_bo_is_stolen_devmem(struct xe_bo *bo);
208 bool xe_bo_has_single_placement(struct xe_bo *bo);
209 uint64_t vram_region_gpu_offset(struct ttm_resource *res);
210 
211 bool xe_bo_can_migrate(struct xe_bo *bo, u32 mem_type);
212 
213 int xe_bo_migrate(struct xe_bo *bo, u32 mem_type);
214 int xe_bo_evict(struct xe_bo *bo, bool force_alloc);
215 
216 int xe_bo_evict_pinned(struct xe_bo *bo);
217 int xe_bo_restore_pinned(struct xe_bo *bo);
218 
219 extern const struct ttm_device_funcs xe_ttm_funcs;
220 extern const char *const xe_mem_type_to_name[];
221 
222 int xe_gem_create_ioctl(struct drm_device *dev, void *data,
223 			struct drm_file *file);
224 int xe_gem_mmap_offset_ioctl(struct drm_device *dev, void *data,
225 			     struct drm_file *file);
226 void xe_bo_runtime_pm_release_mmap_offset(struct xe_bo *bo);
227 
228 int xe_bo_dumb_create(struct drm_file *file_priv,
229 		      struct drm_device *dev,
230 		      struct drm_mode_create_dumb *args);
231 
232 bool xe_bo_needs_ccs_pages(struct xe_bo *bo);
233 
xe_bo_ccs_pages_start(struct xe_bo * bo)234 static inline size_t xe_bo_ccs_pages_start(struct xe_bo *bo)
235 {
236 	return PAGE_ALIGN(bo->ttm.base.size);
237 }
238 
xe_bo_has_pages(struct xe_bo * bo)239 static inline bool xe_bo_has_pages(struct xe_bo *bo)
240 {
241 	if ((bo->ttm.ttm && ttm_tt_is_populated(bo->ttm.ttm)) ||
242 	    xe_bo_is_vram(bo))
243 		return true;
244 
245 	return false;
246 }
247 
248 void __xe_bo_release_dummy(struct kref *kref);
249 
250 /**
251  * xe_bo_put_deferred() - Put a buffer object with delayed final freeing
252  * @bo: The bo to put.
253  * @deferred: List to which to add the buffer object if we cannot put, or
254  * NULL if the function is to put unconditionally.
255  *
256  * Since the final freeing of an object includes both sleeping and (!)
257  * memory allocation in the dma_resv individualization, it's not ok
258  * to put an object from atomic context nor from within a held lock
259  * tainted by reclaim. In such situations we want to defer the final
260  * freeing until we've exited the restricting context, or in the worst
261  * case to a workqueue.
262  * This function either puts the object if possible without the refcount
263  * reaching zero, or adds it to the @deferred list if that was not possible.
264  * The caller needs to follow up with a call to xe_bo_put_commit() to actually
265  * put the bo iff this function returns true. It's safe to always
266  * follow up with a call to xe_bo_put_commit().
267  * TODO: It's TTM that is the villain here. Perhaps TTM should add an
268  * interface like this.
269  *
270  * Return: true if @bo was the first object put on the @freed list,
271  * false otherwise.
272  */
273 static inline bool
xe_bo_put_deferred(struct xe_bo * bo,struct llist_head * deferred)274 xe_bo_put_deferred(struct xe_bo *bo, struct llist_head *deferred)
275 {
276 	if (!deferred) {
277 		xe_bo_put(bo);
278 		return false;
279 	}
280 
281 	if (!kref_put(&bo->ttm.base.refcount, __xe_bo_release_dummy))
282 		return false;
283 
284 	return llist_add(&bo->freed, deferred);
285 }
286 
287 void xe_bo_put_commit(struct llist_head *deferred);
288 
289 struct sg_table *xe_bo_sg(struct xe_bo *bo);
290 
291 /*
292  * xe_sg_segment_size() - Provides upper limit for sg segment size.
293  * @dev: device pointer
294  *
295  * Returns the maximum segment size for the 'struct scatterlist'
296  * elements.
297  */
xe_sg_segment_size(struct device * dev)298 static inline unsigned int xe_sg_segment_size(struct device *dev)
299 {
300 	struct scatterlist __maybe_unused sg;
301 	size_t max = BIT_ULL(sizeof(sg.length) * 8) - 1;
302 
303 	max = min_t(size_t, max, dma_max_mapping_size(dev));
304 
305 	/*
306 	 * The iommu_dma_map_sg() function ensures iova allocation doesn't
307 	 * cross dma segment boundary. It does so by padding some sg elements.
308 	 * This can cause overflow, ending up with sg->length being set to 0.
309 	 * Avoid this by ensuring maximum segment size is half of 'max'
310 	 * rounded down to PAGE_SIZE.
311 	 */
312 	return round_down(max / 2, PAGE_SIZE);
313 }
314 
315 #define i915_gem_object_flush_if_display(obj)		((void)(obj))
316 
317 #if IS_ENABLED(CONFIG_DRM_XE_KUNIT_TEST)
318 /**
319  * xe_bo_is_mem_type - Whether the bo currently resides in the given
320  * TTM memory type
321  * @bo: The bo to check.
322  * @mem_type: The TTM memory type.
323  *
324  * Return: true iff the bo resides in @mem_type, false otherwise.
325  */
xe_bo_is_mem_type(struct xe_bo * bo,u32 mem_type)326 static inline bool xe_bo_is_mem_type(struct xe_bo *bo, u32 mem_type)
327 {
328 	xe_bo_assert_held(bo);
329 	return bo->ttm.resource->mem_type == mem_type;
330 }
331 #endif
332 #endif
333