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 #define XE_BO_FLAG_GGTT0 BIT(18) 43 #define XE_BO_FLAG_GGTT1 BIT(19) 44 #define XE_BO_FLAG_GGTT2 BIT(20) 45 #define XE_BO_FLAG_GGTT3 BIT(21) 46 #define XE_BO_FLAG_GGTT_ALL (XE_BO_FLAG_GGTT0 | \ 47 XE_BO_FLAG_GGTT1 | \ 48 XE_BO_FLAG_GGTT2 | \ 49 XE_BO_FLAG_GGTT3) 50 51 /* this one is trigger internally only */ 52 #define XE_BO_FLAG_INTERNAL_TEST BIT(30) 53 #define XE_BO_FLAG_INTERNAL_64K BIT(31) 54 55 #define XE_BO_FLAG_GGTTx(tile) \ 56 (XE_BO_FLAG_GGTT0 << (tile)->id) 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_PL_SYSTEM TTM_PL_SYSTEM 71 #define XE_PL_TT TTM_PL_TT 72 #define XE_PL_VRAM0 TTM_PL_VRAM 73 #define XE_PL_VRAM1 (XE_PL_VRAM0 + 1) 74 #define XE_PL_STOLEN (TTM_NUM_MEM_TYPES - 1) 75 76 #define XE_BO_PROPS_INVALID (-1) 77 78 #define XE_PCI_BARRIER_MMAP_OFFSET (0x50 << XE_PTE_SHIFT) 79 80 struct sg_table; 81 82 struct xe_bo *xe_bo_alloc(void); 83 void xe_bo_free(struct xe_bo *bo); 84 85 struct xe_bo *___xe_bo_create_locked(struct xe_device *xe, struct xe_bo *bo, 86 struct xe_tile *tile, struct dma_resv *resv, 87 struct ttm_lru_bulk_move *bulk, size_t size, 88 u16 cpu_caching, enum ttm_bo_type type, 89 u32 flags); 90 struct xe_bo * 91 xe_bo_create_locked_range(struct xe_device *xe, 92 struct xe_tile *tile, struct xe_vm *vm, 93 size_t size, u64 start, u64 end, 94 enum ttm_bo_type type, u32 flags, u64 alignment); 95 struct xe_bo *xe_bo_create_locked(struct xe_device *xe, struct xe_tile *tile, 96 struct xe_vm *vm, size_t size, 97 enum ttm_bo_type type, u32 flags); 98 struct xe_bo *xe_bo_create(struct xe_device *xe, struct xe_tile *tile, 99 struct xe_vm *vm, size_t size, 100 enum ttm_bo_type type, u32 flags); 101 struct xe_bo *xe_bo_create_user(struct xe_device *xe, struct xe_tile *tile, 102 struct xe_vm *vm, size_t size, 103 u16 cpu_caching, 104 u32 flags); 105 struct xe_bo *xe_bo_create_pin_map(struct xe_device *xe, struct xe_tile *tile, 106 struct xe_vm *vm, size_t size, 107 enum ttm_bo_type type, u32 flags); 108 struct xe_bo *xe_bo_create_pin_map_at(struct xe_device *xe, struct xe_tile *tile, 109 struct xe_vm *vm, size_t size, u64 offset, 110 enum ttm_bo_type type, u32 flags); 111 struct xe_bo *xe_bo_create_pin_map_at_aligned(struct xe_device *xe, 112 struct xe_tile *tile, 113 struct xe_vm *vm, 114 size_t size, u64 offset, 115 enum ttm_bo_type type, u32 flags, 116 u64 alignment); 117 struct xe_bo *xe_bo_create_from_data(struct xe_device *xe, struct xe_tile *tile, 118 const void *data, size_t size, 119 enum ttm_bo_type type, u32 flags); 120 struct xe_bo *xe_managed_bo_create_pin_map(struct xe_device *xe, struct xe_tile *tile, 121 size_t size, u32 flags); 122 struct xe_bo *xe_managed_bo_create_from_data(struct xe_device *xe, struct xe_tile *tile, 123 const void *data, size_t size, u32 flags); 124 int xe_managed_bo_reinit_in_vram(struct xe_device *xe, struct xe_tile *tile, struct xe_bo **src); 125 126 int xe_bo_placement_for_flags(struct xe_device *xe, struct xe_bo *bo, 127 u32 bo_flags); 128 129 static inline struct xe_bo *ttm_to_xe_bo(const struct ttm_buffer_object *bo) 130 { 131 return container_of(bo, struct xe_bo, ttm); 132 } 133 134 static inline struct xe_bo *gem_to_xe_bo(const struct drm_gem_object *obj) 135 { 136 return container_of(obj, struct xe_bo, ttm.base); 137 } 138 139 #define xe_bo_device(bo) ttm_to_xe_device((bo)->ttm.bdev) 140 141 static inline struct xe_bo *xe_bo_get(struct xe_bo *bo) 142 { 143 if (bo) 144 drm_gem_object_get(&bo->ttm.base); 145 146 return bo; 147 } 148 149 void xe_bo_put(struct xe_bo *bo); 150 151 /* 152 * xe_bo_get_unless_zero() - Conditionally obtain a GEM object refcount on an 153 * xe bo 154 * @bo: The bo for which we want to obtain a refcount. 155 * 156 * There is a short window between where the bo's GEM object refcount reaches 157 * zero and where we put the final ttm_bo reference. Code in the eviction- and 158 * shrinking path should therefore attempt to grab a gem object reference before 159 * trying to use members outside of the base class ttm object. This function is 160 * intended for that purpose. On successful return, this function must be paired 161 * with an xe_bo_put(). 162 * 163 * Return: @bo on success, NULL on failure. 164 */ 165 static inline __must_check struct xe_bo *xe_bo_get_unless_zero(struct xe_bo *bo) 166 { 167 if (!bo || !kref_get_unless_zero(&bo->ttm.base.refcount)) 168 return NULL; 169 170 return bo; 171 } 172 173 static inline void __xe_bo_unset_bulk_move(struct xe_bo *bo) 174 { 175 if (bo) 176 ttm_bo_set_bulk_move(&bo->ttm, NULL); 177 } 178 179 static inline void xe_bo_assert_held(struct xe_bo *bo) 180 { 181 if (bo) 182 dma_resv_assert_held((bo)->ttm.base.resv); 183 } 184 185 int xe_bo_lock(struct xe_bo *bo, bool intr); 186 187 void xe_bo_unlock(struct xe_bo *bo); 188 189 static inline void xe_bo_unlock_vm_held(struct xe_bo *bo) 190 { 191 if (bo) { 192 XE_WARN_ON(bo->vm && bo->ttm.base.resv != xe_vm_resv(bo->vm)); 193 if (bo->vm) 194 xe_vm_assert_held(bo->vm); 195 else 196 dma_resv_unlock(bo->ttm.base.resv); 197 } 198 } 199 200 int xe_bo_pin_external(struct xe_bo *bo); 201 int xe_bo_pin(struct xe_bo *bo); 202 void xe_bo_unpin_external(struct xe_bo *bo); 203 void xe_bo_unpin(struct xe_bo *bo); 204 int xe_bo_validate(struct xe_bo *bo, struct xe_vm *vm, bool allow_res_evict); 205 206 static inline bool xe_bo_is_pinned(struct xe_bo *bo) 207 { 208 return bo->ttm.pin_count; 209 } 210 211 static inline bool xe_bo_is_protected(const struct xe_bo *bo) 212 { 213 return bo->pxp_key_instance; 214 } 215 216 static inline void xe_bo_unpin_map_no_vm(struct xe_bo *bo) 217 { 218 if (likely(bo)) { 219 xe_bo_lock(bo, false); 220 xe_bo_unpin(bo); 221 xe_bo_unlock(bo); 222 223 xe_bo_put(bo); 224 } 225 } 226 227 bool xe_bo_is_xe_bo(struct ttm_buffer_object *bo); 228 dma_addr_t __xe_bo_addr(struct xe_bo *bo, u64 offset, size_t page_size); 229 dma_addr_t xe_bo_addr(struct xe_bo *bo, u64 offset, size_t page_size); 230 231 static inline dma_addr_t 232 xe_bo_main_addr(struct xe_bo *bo, size_t page_size) 233 { 234 return xe_bo_addr(bo, 0, page_size); 235 } 236 237 static inline u32 238 __xe_bo_ggtt_addr(struct xe_bo *bo, u8 tile_id) 239 { 240 struct xe_ggtt_node *ggtt_node = bo->ggtt_node[tile_id]; 241 242 if (XE_WARN_ON(!ggtt_node)) 243 return 0; 244 245 XE_WARN_ON(ggtt_node->base.size > bo->size); 246 XE_WARN_ON(ggtt_node->base.start + ggtt_node->base.size > (1ull << 32)); 247 return ggtt_node->base.start; 248 } 249 250 static inline u32 251 xe_bo_ggtt_addr(struct xe_bo *bo) 252 { 253 xe_assert(xe_bo_device(bo), bo->tile); 254 255 return __xe_bo_ggtt_addr(bo, bo->tile->id); 256 } 257 258 int xe_bo_vmap(struct xe_bo *bo); 259 void xe_bo_vunmap(struct xe_bo *bo); 260 int xe_bo_read(struct xe_bo *bo, u64 offset, void *dst, int size); 261 262 bool mem_type_is_vram(u32 mem_type); 263 bool xe_bo_is_vram(struct xe_bo *bo); 264 bool xe_bo_is_stolen(struct xe_bo *bo); 265 bool xe_bo_is_stolen_devmem(struct xe_bo *bo); 266 bool xe_bo_is_vm_bound(struct xe_bo *bo); 267 bool xe_bo_has_single_placement(struct xe_bo *bo); 268 uint64_t vram_region_gpu_offset(struct ttm_resource *res); 269 270 bool xe_bo_can_migrate(struct xe_bo *bo, u32 mem_type); 271 272 int xe_bo_migrate(struct xe_bo *bo, u32 mem_type); 273 int xe_bo_evict(struct xe_bo *bo, bool force_alloc); 274 275 int xe_bo_evict_pinned(struct xe_bo *bo); 276 int xe_bo_restore_pinned(struct xe_bo *bo); 277 278 extern const struct ttm_device_funcs xe_ttm_funcs; 279 extern const char *const xe_mem_type_to_name[]; 280 281 int xe_gem_create_ioctl(struct drm_device *dev, void *data, 282 struct drm_file *file); 283 int xe_gem_mmap_offset_ioctl(struct drm_device *dev, void *data, 284 struct drm_file *file); 285 void xe_bo_runtime_pm_release_mmap_offset(struct xe_bo *bo); 286 287 int xe_bo_dumb_create(struct drm_file *file_priv, 288 struct drm_device *dev, 289 struct drm_mode_create_dumb *args); 290 291 bool xe_bo_needs_ccs_pages(struct xe_bo *bo); 292 293 static inline size_t xe_bo_ccs_pages_start(struct xe_bo *bo) 294 { 295 return PAGE_ALIGN(bo->ttm.base.size); 296 } 297 298 static inline bool xe_bo_has_pages(struct xe_bo *bo) 299 { 300 if ((bo->ttm.ttm && ttm_tt_is_populated(bo->ttm.ttm)) || 301 xe_bo_is_vram(bo)) 302 return true; 303 304 return false; 305 } 306 307 void __xe_bo_release_dummy(struct kref *kref); 308 309 /** 310 * xe_bo_put_deferred() - Put a buffer object with delayed final freeing 311 * @bo: The bo to put. 312 * @deferred: List to which to add the buffer object if we cannot put, or 313 * NULL if the function is to put unconditionally. 314 * 315 * Since the final freeing of an object includes both sleeping and (!) 316 * memory allocation in the dma_resv individualization, it's not ok 317 * to put an object from atomic context nor from within a held lock 318 * tainted by reclaim. In such situations we want to defer the final 319 * freeing until we've exited the restricting context, or in the worst 320 * case to a workqueue. 321 * This function either puts the object if possible without the refcount 322 * reaching zero, or adds it to the @deferred list if that was not possible. 323 * The caller needs to follow up with a call to xe_bo_put_commit() to actually 324 * put the bo iff this function returns true. It's safe to always 325 * follow up with a call to xe_bo_put_commit(). 326 * TODO: It's TTM that is the villain here. Perhaps TTM should add an 327 * interface like this. 328 * 329 * Return: true if @bo was the first object put on the @freed list, 330 * false otherwise. 331 */ 332 static inline bool 333 xe_bo_put_deferred(struct xe_bo *bo, struct llist_head *deferred) 334 { 335 if (!deferred) { 336 xe_bo_put(bo); 337 return false; 338 } 339 340 if (!kref_put(&bo->ttm.base.refcount, __xe_bo_release_dummy)) 341 return false; 342 343 return llist_add(&bo->freed, deferred); 344 } 345 346 void xe_bo_put_commit(struct llist_head *deferred); 347 348 struct sg_table *xe_bo_sg(struct xe_bo *bo); 349 350 /* 351 * xe_sg_segment_size() - Provides upper limit for sg segment size. 352 * @dev: device pointer 353 * 354 * Returns the maximum segment size for the 'struct scatterlist' 355 * elements. 356 */ 357 static inline unsigned int xe_sg_segment_size(struct device *dev) 358 { 359 struct scatterlist __maybe_unused sg; 360 size_t max = BIT_ULL(sizeof(sg.length) * 8) - 1; 361 362 max = min_t(size_t, max, dma_max_mapping_size(dev)); 363 364 /* 365 * The iommu_dma_map_sg() function ensures iova allocation doesn't 366 * cross dma segment boundary. It does so by padding some sg elements. 367 * This can cause overflow, ending up with sg->length being set to 0. 368 * Avoid this by ensuring maximum segment size is half of 'max' 369 * rounded down to PAGE_SIZE. 370 */ 371 return round_down(max / 2, PAGE_SIZE); 372 } 373 374 /** 375 * struct xe_bo_shrink_flags - flags governing the shrink behaviour. 376 * @purge: Only purging allowed. Don't shrink if bo not purgeable. 377 * @writeback: Attempt to immediately move content to swap. 378 */ 379 struct xe_bo_shrink_flags { 380 u32 purge : 1; 381 u32 writeback : 1; 382 }; 383 384 long xe_bo_shrink(struct ttm_operation_ctx *ctx, struct ttm_buffer_object *bo, 385 const struct xe_bo_shrink_flags flags, 386 unsigned long *scanned); 387 388 #if IS_ENABLED(CONFIG_DRM_XE_KUNIT_TEST) 389 /** 390 * xe_bo_is_mem_type - Whether the bo currently resides in the given 391 * TTM memory type 392 * @bo: The bo to check. 393 * @mem_type: The TTM memory type. 394 * 395 * Return: true iff the bo resides in @mem_type, false otherwise. 396 */ 397 static inline bool xe_bo_is_mem_type(struct xe_bo *bo, u32 mem_type) 398 { 399 xe_bo_assert_held(bo); 400 return bo->ttm.resource->mem_type == mem_type; 401 } 402 #endif 403 #endif 404