xref: /linux/drivers/gpu/drm/xe/xe_bo.h (revision d7b4e3287ca3a7baf66efd9158498e551a9550da)
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 /**
17  * xe_vm_assert_held(vm) - Assert that the vm's reservation object is held.
18  * @vm: The vm
19  */
20 #define xe_vm_assert_held(vm) dma_resv_assert_held(xe_vm_resv(vm))
21 
22 
23 
24 #define XE_DEFAULT_GTT_SIZE_MB          3072ULL /* 3GB by default */
25 
26 #define XE_BO_CREATE_USER_BIT		BIT(0)
27 /* The bits below need to be contiguous, or things break */
28 #define XE_BO_CREATE_SYSTEM_BIT		BIT(1)
29 #define XE_BO_CREATE_VRAM0_BIT		BIT(2)
30 #define XE_BO_CREATE_VRAM1_BIT		BIT(3)
31 #define XE_BO_CREATE_VRAM_MASK		(XE_BO_CREATE_VRAM0_BIT | \
32 					 XE_BO_CREATE_VRAM1_BIT)
33 /* -- */
34 #define XE_BO_CREATE_STOLEN_BIT		BIT(4)
35 #define XE_BO_CREATE_VRAM_IF_DGFX(tile) \
36 	(IS_DGFX(tile_to_xe(tile)) ? XE_BO_CREATE_VRAM0_BIT << (tile)->id : \
37 	 XE_BO_CREATE_SYSTEM_BIT)
38 #define XE_BO_CREATE_GGTT_BIT		BIT(5)
39 #define XE_BO_CREATE_IGNORE_MIN_PAGE_SIZE_BIT BIT(6)
40 #define XE_BO_CREATE_PINNED_BIT		BIT(7)
41 #define XE_BO_CREATE_NO_RESV_EVICT	BIT(8)
42 #define XE_BO_DEFER_BACKING		BIT(9)
43 #define XE_BO_SCANOUT_BIT		BIT(10)
44 #define XE_BO_FIXED_PLACEMENT_BIT	BIT(11)
45 #define XE_BO_PAGETABLE			BIT(12)
46 #define XE_BO_NEEDS_CPU_ACCESS		BIT(13)
47 /* this one is trigger internally only */
48 #define XE_BO_INTERNAL_TEST		BIT(30)
49 #define XE_BO_INTERNAL_64K		BIT(31)
50 
51 #define XELPG_PPGTT_PTE_PAT3		BIT_ULL(62)
52 #define XE2_PPGTT_PTE_PAT4		BIT_ULL(61)
53 #define XE_PPGTT_PDE_PDPE_PAT2		BIT_ULL(12)
54 #define XE_PPGTT_PTE_PAT2		BIT_ULL(7)
55 #define XE_PPGTT_PTE_PAT1		BIT_ULL(4)
56 #define XE_PPGTT_PTE_PAT0		BIT_ULL(3)
57 
58 #define XE_PTE_SHIFT			12
59 #define XE_PAGE_SIZE			(1 << XE_PTE_SHIFT)
60 #define XE_PTE_MASK			(XE_PAGE_SIZE - 1)
61 #define XE_PDE_SHIFT			(XE_PTE_SHIFT - 3)
62 #define XE_PDES				(1 << XE_PDE_SHIFT)
63 #define XE_PDE_MASK			(XE_PDES - 1)
64 
65 #define XE_64K_PTE_SHIFT		16
66 #define XE_64K_PAGE_SIZE		(1 << XE_64K_PTE_SHIFT)
67 #define XE_64K_PTE_MASK			(XE_64K_PAGE_SIZE - 1)
68 #define XE_64K_PDE_MASK			(XE_PDE_MASK >> 4)
69 
70 #define XE_PDE_PS_2M			BIT_ULL(7)
71 #define XE_PDPE_PS_1G			BIT_ULL(7)
72 #define XE_PDE_IPS_64K			BIT_ULL(11)
73 
74 #define XE_GGTT_PTE_DM			BIT_ULL(1)
75 #define XE_USM_PPGTT_PTE_AE		BIT_ULL(10)
76 #define XE_PPGTT_PTE_DM			BIT_ULL(11)
77 #define XE_PDE_64K			BIT_ULL(6)
78 #define XE_PTE_PS64			BIT_ULL(8)
79 #define XE_PTE_NULL			BIT_ULL(9)
80 
81 #define XE_PAGE_PRESENT			BIT_ULL(0)
82 #define XE_PAGE_RW			BIT_ULL(1)
83 
84 #define XE_PL_SYSTEM		TTM_PL_SYSTEM
85 #define XE_PL_TT		TTM_PL_TT
86 #define XE_PL_VRAM0		TTM_PL_VRAM
87 #define XE_PL_VRAM1		(XE_PL_VRAM0 + 1)
88 #define XE_PL_STOLEN		(TTM_NUM_MEM_TYPES - 1)
89 
90 #define XE_BO_PROPS_INVALID	(-1)
91 
92 struct sg_table;
93 
94 struct xe_bo *xe_bo_alloc(void);
95 void xe_bo_free(struct xe_bo *bo);
96 
97 struct xe_bo *___xe_bo_create_locked(struct xe_device *xe, struct xe_bo *bo,
98 				     struct xe_tile *tile, struct dma_resv *resv,
99 				     struct ttm_lru_bulk_move *bulk, size_t size,
100 				     u16 cpu_caching, enum ttm_bo_type type,
101 				     u32 flags);
102 struct xe_bo *
103 xe_bo_create_locked_range(struct xe_device *xe,
104 			  struct xe_tile *tile, struct xe_vm *vm,
105 			  size_t size, u64 start, u64 end,
106 			  enum ttm_bo_type type, u32 flags);
107 struct xe_bo *xe_bo_create_locked(struct xe_device *xe, struct xe_tile *tile,
108 				  struct xe_vm *vm, size_t size,
109 				  enum ttm_bo_type type, u32 flags);
110 struct xe_bo *xe_bo_create(struct xe_device *xe, struct xe_tile *tile,
111 			   struct xe_vm *vm, size_t size,
112 			   enum ttm_bo_type type, u32 flags);
113 struct xe_bo *xe_bo_create_user(struct xe_device *xe, struct xe_tile *tile,
114 				struct xe_vm *vm, size_t size,
115 				u16 cpu_caching,
116 				enum ttm_bo_type type,
117 				u32 flags);
118 struct xe_bo *xe_bo_create_pin_map(struct xe_device *xe, struct xe_tile *tile,
119 				   struct xe_vm *vm, size_t size,
120 				   enum ttm_bo_type type, u32 flags);
121 struct xe_bo *xe_bo_create_pin_map_at(struct xe_device *xe, struct xe_tile *tile,
122 				      struct xe_vm *vm, size_t size, u64 offset,
123 				      enum ttm_bo_type type, u32 flags);
124 struct xe_bo *xe_bo_create_from_data(struct xe_device *xe, struct xe_tile *tile,
125 				     const void *data, size_t size,
126 				     enum ttm_bo_type type, u32 flags);
127 struct xe_bo *xe_managed_bo_create_pin_map(struct xe_device *xe, struct xe_tile *tile,
128 					   size_t size, u32 flags);
129 struct xe_bo *xe_managed_bo_create_from_data(struct xe_device *xe, struct xe_tile *tile,
130 					     const void *data, size_t size, u32 flags);
131 
132 int xe_bo_placement_for_flags(struct xe_device *xe, struct xe_bo *bo,
133 			      u32 bo_flags);
134 
135 static inline struct xe_bo *ttm_to_xe_bo(const struct ttm_buffer_object *bo)
136 {
137 	return container_of(bo, struct xe_bo, ttm);
138 }
139 
140 static inline struct xe_bo *gem_to_xe_bo(const struct drm_gem_object *obj)
141 {
142 	return container_of(obj, struct xe_bo, ttm.base);
143 }
144 
145 #define xe_bo_device(bo) ttm_to_xe_device((bo)->ttm.bdev)
146 
147 static inline struct xe_bo *xe_bo_get(struct xe_bo *bo)
148 {
149 	if (bo)
150 		drm_gem_object_get(&bo->ttm.base);
151 
152 	return bo;
153 }
154 
155 static inline void xe_bo_put(struct xe_bo *bo)
156 {
157 	if (bo)
158 		drm_gem_object_put(&bo->ttm.base);
159 }
160 
161 static inline void __xe_bo_unset_bulk_move(struct xe_bo *bo)
162 {
163 	if (bo)
164 		ttm_bo_set_bulk_move(&bo->ttm, NULL);
165 }
166 
167 static inline void xe_bo_assert_held(struct xe_bo *bo)
168 {
169 	if (bo)
170 		dma_resv_assert_held((bo)->ttm.base.resv);
171 }
172 
173 int xe_bo_lock(struct xe_bo *bo, bool intr);
174 
175 void xe_bo_unlock(struct xe_bo *bo);
176 
177 static inline void xe_bo_unlock_vm_held(struct xe_bo *bo)
178 {
179 	if (bo) {
180 		XE_WARN_ON(bo->vm && bo->ttm.base.resv != xe_vm_resv(bo->vm));
181 		if (bo->vm)
182 			xe_vm_assert_held(bo->vm);
183 		else
184 			dma_resv_unlock(bo->ttm.base.resv);
185 	}
186 }
187 
188 int xe_bo_pin_external(struct xe_bo *bo);
189 int xe_bo_pin(struct xe_bo *bo);
190 void xe_bo_unpin_external(struct xe_bo *bo);
191 void xe_bo_unpin(struct xe_bo *bo);
192 int xe_bo_validate(struct xe_bo *bo, struct xe_vm *vm, bool allow_res_evict);
193 
194 static inline bool xe_bo_is_pinned(struct xe_bo *bo)
195 {
196 	return bo->ttm.pin_count;
197 }
198 
199 static inline void xe_bo_unpin_map_no_vm(struct xe_bo *bo)
200 {
201 	if (likely(bo)) {
202 		xe_bo_lock(bo, false);
203 		xe_bo_unpin(bo);
204 		xe_bo_unlock(bo);
205 
206 		xe_bo_put(bo);
207 	}
208 }
209 
210 bool xe_bo_is_xe_bo(struct ttm_buffer_object *bo);
211 dma_addr_t __xe_bo_addr(struct xe_bo *bo, u64 offset, size_t page_size);
212 dma_addr_t xe_bo_addr(struct xe_bo *bo, u64 offset, size_t page_size);
213 
214 static inline dma_addr_t
215 xe_bo_main_addr(struct xe_bo *bo, size_t page_size)
216 {
217 	return xe_bo_addr(bo, 0, page_size);
218 }
219 
220 static inline u32
221 xe_bo_ggtt_addr(struct xe_bo *bo)
222 {
223 	XE_WARN_ON(bo->ggtt_node.size > bo->size);
224 	XE_WARN_ON(bo->ggtt_node.start + bo->ggtt_node.size > (1ull << 32));
225 	return bo->ggtt_node.start;
226 }
227 
228 int xe_bo_vmap(struct xe_bo *bo);
229 void xe_bo_vunmap(struct xe_bo *bo);
230 
231 bool mem_type_is_vram(u32 mem_type);
232 bool xe_bo_is_vram(struct xe_bo *bo);
233 bool xe_bo_is_stolen(struct xe_bo *bo);
234 bool xe_bo_is_stolen_devmem(struct xe_bo *bo);
235 uint64_t vram_region_gpu_offset(struct ttm_resource *res);
236 
237 bool xe_bo_can_migrate(struct xe_bo *bo, u32 mem_type);
238 
239 int xe_bo_migrate(struct xe_bo *bo, u32 mem_type);
240 int xe_bo_evict(struct xe_bo *bo, bool force_alloc);
241 
242 int xe_bo_evict_pinned(struct xe_bo *bo);
243 int xe_bo_restore_pinned(struct xe_bo *bo);
244 
245 extern struct ttm_device_funcs xe_ttm_funcs;
246 extern const char *const xe_mem_type_to_name[];
247 
248 int xe_gem_create_ioctl(struct drm_device *dev, void *data,
249 			struct drm_file *file);
250 int xe_gem_mmap_offset_ioctl(struct drm_device *dev, void *data,
251 			     struct drm_file *file);
252 int xe_bo_dumb_create(struct drm_file *file_priv,
253 		      struct drm_device *dev,
254 		      struct drm_mode_create_dumb *args);
255 
256 bool xe_bo_needs_ccs_pages(struct xe_bo *bo);
257 
258 static inline size_t xe_bo_ccs_pages_start(struct xe_bo *bo)
259 {
260 	return PAGE_ALIGN(bo->ttm.base.size);
261 }
262 
263 static inline bool xe_bo_has_pages(struct xe_bo *bo)
264 {
265 	if ((bo->ttm.ttm && ttm_tt_is_populated(bo->ttm.ttm)) ||
266 	    xe_bo_is_vram(bo))
267 		return true;
268 
269 	return false;
270 }
271 
272 void __xe_bo_release_dummy(struct kref *kref);
273 
274 /**
275  * xe_bo_put_deferred() - Put a buffer object with delayed final freeing
276  * @bo: The bo to put.
277  * @deferred: List to which to add the buffer object if we cannot put, or
278  * NULL if the function is to put unconditionally.
279  *
280  * Since the final freeing of an object includes both sleeping and (!)
281  * memory allocation in the dma_resv individualization, it's not ok
282  * to put an object from atomic context nor from within a held lock
283  * tainted by reclaim. In such situations we want to defer the final
284  * freeing until we've exited the restricting context, or in the worst
285  * case to a workqueue.
286  * This function either puts the object if possible without the refcount
287  * reaching zero, or adds it to the @deferred list if that was not possible.
288  * The caller needs to follow up with a call to xe_bo_put_commit() to actually
289  * put the bo iff this function returns true. It's safe to always
290  * follow up with a call to xe_bo_put_commit().
291  * TODO: It's TTM that is the villain here. Perhaps TTM should add an
292  * interface like this.
293  *
294  * Return: true if @bo was the first object put on the @freed list,
295  * false otherwise.
296  */
297 static inline bool
298 xe_bo_put_deferred(struct xe_bo *bo, struct llist_head *deferred)
299 {
300 	if (!deferred) {
301 		xe_bo_put(bo);
302 		return false;
303 	}
304 
305 	if (!kref_put(&bo->ttm.base.refcount, __xe_bo_release_dummy))
306 		return false;
307 
308 	return llist_add(&bo->freed, deferred);
309 }
310 
311 void xe_bo_put_commit(struct llist_head *deferred);
312 
313 struct sg_table *xe_bo_sg(struct xe_bo *bo);
314 
315 /*
316  * xe_sg_segment_size() - Provides upper limit for sg segment size.
317  * @dev: device pointer
318  *
319  * Returns the maximum segment size for the 'struct scatterlist'
320  * elements.
321  */
322 static inline unsigned int xe_sg_segment_size(struct device *dev)
323 {
324 	struct scatterlist __maybe_unused sg;
325 	size_t max = BIT_ULL(sizeof(sg.length) * 8) - 1;
326 
327 	max = min_t(size_t, max, dma_max_mapping_size(dev));
328 
329 	/*
330 	 * The iommu_dma_map_sg() function ensures iova allocation doesn't
331 	 * cross dma segment boundary. It does so by padding some sg elements.
332 	 * This can cause overflow, ending up with sg->length being set to 0.
333 	 * Avoid this by ensuring maximum segment size is half of 'max'
334 	 * rounded down to PAGE_SIZE.
335 	 */
336 	return round_down(max / 2, PAGE_SIZE);
337 }
338 
339 #define i915_gem_object_flush_if_display(obj)		((void)(obj))
340 
341 #if IS_ENABLED(CONFIG_DRM_XE_KUNIT_TEST)
342 /**
343  * xe_bo_is_mem_type - Whether the bo currently resides in the given
344  * TTM memory type
345  * @bo: The bo to check.
346  * @mem_type: The TTM memory type.
347  *
348  * Return: true iff the bo resides in @mem_type, false otherwise.
349  */
350 static inline bool xe_bo_is_mem_type(struct xe_bo *bo, u32 mem_type)
351 {
352 	xe_bo_assert_held(bo);
353 	return bo->ttm.resource->mem_type == mem_type;
354 }
355 #endif
356 #endif
357