xref: /linux/drivers/gpu/drm/xe/xe_bo.h (revision ca220141fa8ebae09765a242076b2b77338106b0)
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_ggtt.h"
13 #include "xe_macros.h"
14 #include "xe_validation.h"
15 #include "xe_vm_types.h"
16 #include "xe_vm.h"
17 #include "xe_vram_types.h"
18 
19 #define XE_DEFAULT_GTT_SIZE_MB          3072ULL /* 3GB by default */
20 
21 #define XE_BO_FLAG_USER		BIT(0)
22 /* The bits below need to be contiguous, or things break */
23 #define XE_BO_FLAG_SYSTEM		BIT(1)
24 #define XE_BO_FLAG_VRAM0		BIT(2)
25 #define XE_BO_FLAG_VRAM1		BIT(3)
26 #define XE_BO_FLAG_VRAM_MASK		(XE_BO_FLAG_VRAM0 | XE_BO_FLAG_VRAM1)
27 /* -- */
28 #define XE_BO_FLAG_STOLEN		BIT(4)
29 #define XE_BO_FLAG_VRAM(vram)		(XE_BO_FLAG_VRAM0 << ((vram)->id))
30 #define XE_BO_FLAG_VRAM_IF_DGFX(tile)	(IS_DGFX(tile_to_xe(tile)) ? \
31 					 XE_BO_FLAG_VRAM((tile)->mem.vram) : \
32 					 XE_BO_FLAG_SYSTEM)
33 #define XE_BO_FLAG_GGTT			BIT(5)
34 #define XE_BO_FLAG_IGNORE_MIN_PAGE_SIZE BIT(6)
35 #define XE_BO_FLAG_PINNED		BIT(7)
36 #define XE_BO_FLAG_NO_RESV_EVICT	BIT(8)
37 #define XE_BO_FLAG_DEFER_BACKING	BIT(9)
38 #define XE_BO_FLAG_SCANOUT		BIT(10)
39 #define XE_BO_FLAG_FIXED_PLACEMENT	BIT(11)
40 #define XE_BO_FLAG_PAGETABLE		BIT(12)
41 #define XE_BO_FLAG_NEEDS_CPU_ACCESS	BIT(13)
42 #define XE_BO_FLAG_NEEDS_UC		BIT(14)
43 #define XE_BO_FLAG_NEEDS_64K		BIT(15)
44 #define XE_BO_FLAG_NEEDS_2M		BIT(16)
45 #define XE_BO_FLAG_GGTT_INVALIDATE	BIT(17)
46 #define XE_BO_FLAG_PINNED_NORESTORE	BIT(18)
47 #define XE_BO_FLAG_PINNED_LATE_RESTORE	BIT(19)
48 #define XE_BO_FLAG_GGTT0		BIT(20)
49 #define XE_BO_FLAG_GGTT1		BIT(21)
50 #define XE_BO_FLAG_GGTT2		BIT(22)
51 #define XE_BO_FLAG_GGTT3		BIT(23)
52 #define XE_BO_FLAG_CPU_ADDR_MIRROR	BIT(24)
53 #define XE_BO_FLAG_FORCE_USER_VRAM	BIT(25)
54 #define XE_BO_FLAG_NO_COMPRESSION	BIT(26)
55 
56 /* this one is trigger internally only */
57 #define XE_BO_FLAG_INTERNAL_TEST	BIT(30)
58 #define XE_BO_FLAG_INTERNAL_64K		BIT(31)
59 
60 #define XE_BO_FLAG_GGTT_ALL		(XE_BO_FLAG_GGTT0 | \
61 					 XE_BO_FLAG_GGTT1 | \
62 					 XE_BO_FLAG_GGTT2 | \
63 					 XE_BO_FLAG_GGTT3)
64 
65 #define XE_BO_FLAG_GGTTx(tile) \
66 	(XE_BO_FLAG_GGTT0 << (tile)->id)
67 
68 #define XE_PTE_SHIFT			12
69 #define XE_PAGE_SIZE			(1 << XE_PTE_SHIFT)
70 #define XE_PTE_MASK			(XE_PAGE_SIZE - 1)
71 #define XE_PDE_SHIFT			(XE_PTE_SHIFT - 3)
72 #define XE_PDES				(1 << XE_PDE_SHIFT)
73 #define XE_PDE_MASK			(XE_PDES - 1)
74 
75 #define XE_64K_PTE_SHIFT		16
76 #define XE_64K_PAGE_SIZE		(1 << XE_64K_PTE_SHIFT)
77 #define XE_64K_PTE_MASK			(XE_64K_PAGE_SIZE - 1)
78 #define XE_64K_PDE_MASK			(XE_PDE_MASK >> 4)
79 
80 #define XE_PL_SYSTEM		TTM_PL_SYSTEM
81 #define XE_PL_TT		TTM_PL_TT
82 #define XE_PL_VRAM0		TTM_PL_VRAM
83 #define XE_PL_VRAM1		(XE_PL_VRAM0 + 1)
84 #define XE_PL_STOLEN		(TTM_NUM_MEM_TYPES - 1)
85 
86 #define XE_BO_PROPS_INVALID	(-1)
87 
88 #define XE_PCI_BARRIER_MMAP_OFFSET	(0x50 << XE_PTE_SHIFT)
89 
90 struct sg_table;
91 
92 struct xe_bo *xe_bo_alloc(void);
93 void xe_bo_free(struct xe_bo *bo);
94 
95 struct xe_bo *xe_bo_init_locked(struct xe_device *xe, struct xe_bo *bo,
96 				struct xe_tile *tile, struct dma_resv *resv,
97 				struct ttm_lru_bulk_move *bulk, size_t size,
98 				u16 cpu_caching, enum ttm_bo_type type,
99 				u32 flags, struct drm_exec *exec);
100 struct xe_bo *xe_bo_create_locked(struct xe_device *xe, struct xe_tile *tile,
101 				  struct xe_vm *vm, size_t size,
102 				  enum ttm_bo_type type, u32 flags,
103 				  struct drm_exec *exec);
104 struct xe_bo *xe_bo_create_user(struct xe_device *xe, struct xe_vm *vm, size_t size,
105 				u16 cpu_caching, u32 flags, struct drm_exec *exec);
106 struct xe_bo *xe_bo_create_pin_map(struct xe_device *xe, struct xe_tile *tile,
107 				   struct xe_vm *vm, size_t size,
108 				   enum ttm_bo_type type, u32 flags,
109 				   struct drm_exec *exec);
110 struct xe_bo *xe_bo_create_pin_map_novm(struct xe_device *xe, struct xe_tile *tile,
111 					size_t size, enum ttm_bo_type type, u32 flags,
112 					bool intr);
113 struct xe_bo *xe_bo_create_pin_range_novm(struct xe_device *xe, struct xe_tile *tile,
114 					  size_t size, u64 start, u64 end,
115 					  enum ttm_bo_type type, u32 flags);
116 struct xe_bo *
117 xe_bo_create_pin_map_at_novm(struct xe_device *xe, struct xe_tile *tile,
118 			     size_t size, u64 offset, enum ttm_bo_type type,
119 			     u32 flags, u64 alignment, bool intr);
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 void xe_managed_bo_unpin_map_no_vm(struct xe_bo *bo);
123 struct xe_bo *xe_managed_bo_create_from_data(struct xe_device *xe, struct xe_tile *tile,
124 					     const void *data, size_t size, u32 flags);
125 int xe_managed_bo_reinit_in_vram(struct xe_device *xe, struct xe_tile *tile, struct xe_bo **src);
126 
127 int xe_bo_placement_for_flags(struct xe_device *xe, struct xe_bo *bo,
128 			      u32 bo_flags, enum ttm_bo_type type);
129 
130 static inline struct xe_bo *ttm_to_xe_bo(const struct ttm_buffer_object *bo)
131 {
132 	return container_of(bo, struct xe_bo, ttm);
133 }
134 
135 static inline struct xe_bo *gem_to_xe_bo(const struct drm_gem_object *obj)
136 {
137 	return container_of(obj, struct xe_bo, ttm.base);
138 }
139 
140 #define xe_bo_device(bo) ttm_to_xe_device((bo)->ttm.bdev)
141 
142 static inline struct xe_bo *xe_bo_get(struct xe_bo *bo)
143 {
144 	if (bo)
145 		drm_gem_object_get(&bo->ttm.base);
146 
147 	return bo;
148 }
149 
150 void xe_bo_put(struct xe_bo *bo);
151 
152 /*
153  * xe_bo_get_unless_zero() - Conditionally obtain a GEM object refcount on an
154  * xe bo
155  * @bo: The bo for which we want to obtain a refcount.
156  *
157  * There is a short window between where the bo's GEM object refcount reaches
158  * zero and where we put the final ttm_bo reference. Code in the eviction- and
159  * shrinking path should therefore attempt to grab a gem object reference before
160  * trying to use members outside of the base class ttm object. This function is
161  * intended for that purpose. On successful return, this function must be paired
162  * with an xe_bo_put().
163  *
164  * Return: @bo on success, NULL on failure.
165  */
166 static inline __must_check struct xe_bo *xe_bo_get_unless_zero(struct xe_bo *bo)
167 {
168 	if (!bo || !kref_get_unless_zero(&bo->ttm.base.refcount))
169 		return NULL;
170 
171 	return bo;
172 }
173 
174 static inline void __xe_bo_unset_bulk_move(struct xe_bo *bo)
175 {
176 	if (bo)
177 		ttm_bo_set_bulk_move(&bo->ttm, NULL);
178 }
179 
180 static inline void xe_bo_assert_held(struct xe_bo *bo)
181 {
182 	if (bo)
183 		dma_resv_assert_held((bo)->ttm.base.resv);
184 }
185 
186 int xe_bo_lock(struct xe_bo *bo, bool intr);
187 
188 void xe_bo_unlock(struct xe_bo *bo);
189 
190 static inline void xe_bo_unlock_vm_held(struct xe_bo *bo)
191 {
192 	if (bo) {
193 		XE_WARN_ON(bo->vm && bo->ttm.base.resv != xe_vm_resv(bo->vm));
194 		if (bo->vm)
195 			xe_vm_assert_held(bo->vm);
196 		else
197 			dma_resv_unlock(bo->ttm.base.resv);
198 	}
199 }
200 
201 int xe_bo_pin_external(struct xe_bo *bo, bool in_place, struct drm_exec *exec);
202 int xe_bo_pin(struct xe_bo *bo, struct drm_exec *exec);
203 void xe_bo_unpin_external(struct xe_bo *bo);
204 void xe_bo_unpin(struct xe_bo *bo);
205 int xe_bo_validate(struct xe_bo *bo, struct xe_vm *vm, bool allow_res_evict,
206 		   struct drm_exec *exec);
207 
208 static inline bool xe_bo_is_pinned(struct xe_bo *bo)
209 {
210 	return bo->ttm.pin_count;
211 }
212 
213 static inline bool xe_bo_is_protected(const struct xe_bo *bo)
214 {
215 	return bo->pxp_key_instance;
216 }
217 
218 static inline void xe_bo_unpin_map_no_vm(struct xe_bo *bo)
219 {
220 	if (likely(bo)) {
221 		xe_bo_lock(bo, false);
222 		xe_bo_unpin(bo);
223 		xe_bo_unlock(bo);
224 
225 		xe_bo_put(bo);
226 	}
227 }
228 
229 bool xe_bo_is_xe_bo(struct ttm_buffer_object *bo);
230 dma_addr_t __xe_bo_addr(struct xe_bo *bo, u64 offset, size_t page_size);
231 dma_addr_t xe_bo_addr(struct xe_bo *bo, u64 offset, size_t page_size);
232 
233 static inline dma_addr_t
234 xe_bo_main_addr(struct xe_bo *bo, size_t page_size)
235 {
236 	return xe_bo_addr(bo, 0, page_size);
237 }
238 
239 /**
240  * xe_bo_size() - Xe BO size
241  * @bo: The bo object.
242  *
243  * Simple helper to return Xe BO's size.
244  *
245  * Return: Xe BO's size
246  */
247 static inline size_t xe_bo_size(struct xe_bo *bo)
248 {
249 	return bo->ttm.base.size;
250 }
251 
252 static inline u32
253 __xe_bo_ggtt_addr(struct xe_bo *bo, u8 tile_id)
254 {
255 	struct xe_ggtt_node *ggtt_node = bo->ggtt_node[tile_id];
256 	u64 offset;
257 
258 	if (XE_WARN_ON(!ggtt_node))
259 		return 0;
260 
261 	offset = xe_ggtt_node_addr(ggtt_node);
262 	XE_WARN_ON(offset + xe_bo_size(bo) > (1ull << 32));
263 	return offset;
264 }
265 
266 static inline u32
267 xe_bo_ggtt_addr(struct xe_bo *bo)
268 {
269 	xe_assert(xe_bo_device(bo), bo->tile);
270 
271 	return __xe_bo_ggtt_addr(bo, bo->tile->id);
272 }
273 
274 int xe_bo_vmap(struct xe_bo *bo);
275 void xe_bo_vunmap(struct xe_bo *bo);
276 int xe_bo_read(struct xe_bo *bo, u64 offset, void *dst, int size);
277 
278 bool mem_type_is_vram(u32 mem_type);
279 bool xe_bo_is_vram(struct xe_bo *bo);
280 bool xe_bo_is_visible_vram(struct xe_bo *bo);
281 bool xe_bo_is_stolen(struct xe_bo *bo);
282 bool xe_bo_is_stolen_devmem(struct xe_bo *bo);
283 bool xe_bo_is_vm_bound(struct xe_bo *bo);
284 bool xe_bo_has_single_placement(struct xe_bo *bo);
285 uint64_t vram_region_gpu_offset(struct ttm_resource *res);
286 
287 bool xe_bo_can_migrate(struct xe_bo *bo, u32 mem_type);
288 
289 int xe_bo_migrate(struct xe_bo *bo, u32 mem_type, struct ttm_operation_ctx *ctc,
290 		  struct drm_exec *exec);
291 int xe_bo_evict(struct xe_bo *bo, struct drm_exec *exec);
292 
293 int xe_bo_evict_pinned(struct xe_bo *bo);
294 int xe_bo_notifier_prepare_pinned(struct xe_bo *bo);
295 int xe_bo_notifier_unprepare_pinned(struct xe_bo *bo);
296 int xe_bo_restore_pinned(struct xe_bo *bo);
297 
298 int xe_bo_dma_unmap_pinned(struct xe_bo *bo);
299 
300 extern const struct ttm_device_funcs xe_ttm_funcs;
301 extern const char *const xe_mem_type_to_name[];
302 
303 int xe_gem_create_ioctl(struct drm_device *dev, void *data,
304 			struct drm_file *file);
305 int xe_gem_mmap_offset_ioctl(struct drm_device *dev, void *data,
306 			     struct drm_file *file);
307 void xe_bo_runtime_pm_release_mmap_offset(struct xe_bo *bo);
308 
309 int xe_bo_dumb_create(struct drm_file *file_priv,
310 		      struct drm_device *dev,
311 		      struct drm_mode_create_dumb *args);
312 
313 bool xe_bo_needs_ccs_pages(struct xe_bo *bo);
314 
315 static inline size_t xe_bo_ccs_pages_start(struct xe_bo *bo)
316 {
317 	return PAGE_ALIGN(xe_bo_size(bo));
318 }
319 
320 /**
321  * xe_bo_has_valid_ccs_bb - Check if CCS's BBs were setup for the BO.
322  * @bo: the &xe_bo to check
323  *
324  * The CCS's BBs should only be setup by the driver VF, but it is safe
325  * to call this function also by non-VF driver.
326  *
327  * Return: true iff the CCS's BBs are setup, false otherwise.
328  */
329 static inline bool xe_bo_has_valid_ccs_bb(struct xe_bo *bo)
330 {
331 	return bo->bb_ccs[XE_SRIOV_VF_CCS_READ_CTX] &&
332 	       bo->bb_ccs[XE_SRIOV_VF_CCS_WRITE_CTX];
333 }
334 
335 static inline bool xe_bo_has_pages(struct xe_bo *bo)
336 {
337 	if ((bo->ttm.ttm && ttm_tt_is_populated(bo->ttm.ttm)) ||
338 	    xe_bo_is_vram(bo))
339 		return true;
340 
341 	return false;
342 }
343 
344 void __xe_bo_release_dummy(struct kref *kref);
345 
346 /**
347  * xe_bo_put_deferred() - Put a buffer object with delayed final freeing
348  * @bo: The bo to put.
349  * @deferred: List to which to add the buffer object if we cannot put, or
350  * NULL if the function is to put unconditionally.
351  *
352  * Since the final freeing of an object includes both sleeping and (!)
353  * memory allocation in the dma_resv individualization, it's not ok
354  * to put an object from atomic context nor from within a held lock
355  * tainted by reclaim. In such situations we want to defer the final
356  * freeing until we've exited the restricting context, or in the worst
357  * case to a workqueue.
358  * This function either puts the object if possible without the refcount
359  * reaching zero, or adds it to the @deferred list if that was not possible.
360  * The caller needs to follow up with a call to xe_bo_put_commit() to actually
361  * put the bo iff this function returns true. It's safe to always
362  * follow up with a call to xe_bo_put_commit().
363  * TODO: It's TTM that is the villain here. Perhaps TTM should add an
364  * interface like this.
365  *
366  * Return: true if @bo was the first object put on the @freed list,
367  * false otherwise.
368  */
369 static inline bool
370 xe_bo_put_deferred(struct xe_bo *bo, struct llist_head *deferred)
371 {
372 	if (!deferred) {
373 		xe_bo_put(bo);
374 		return false;
375 	}
376 
377 	if (!kref_put(&bo->ttm.base.refcount, __xe_bo_release_dummy))
378 		return false;
379 
380 	return llist_add(&bo->freed, deferred);
381 }
382 
383 void xe_bo_put_commit(struct llist_head *deferred);
384 
385 /**
386  * xe_bo_put_async() - Put BO async
387  * @bo: The bo to put.
388  *
389  * Put BO async, the final put is deferred to a worker to exit an IRQ context.
390  */
391 static inline void
392 xe_bo_put_async(struct xe_bo *bo)
393 {
394 	struct xe_bo_dev *bo_device = &xe_bo_device(bo)->bo_device;
395 
396 	if (xe_bo_put_deferred(bo, &bo_device->async_list))
397 		schedule_work(&bo_device->async_free);
398 }
399 
400 void xe_bo_dev_init(struct xe_bo_dev *bo_device);
401 
402 void xe_bo_dev_fini(struct xe_bo_dev *bo_device);
403 
404 struct sg_table *xe_bo_sg(struct xe_bo *bo);
405 
406 /*
407  * xe_sg_segment_size() - Provides upper limit for sg segment size.
408  * @dev: device pointer
409  *
410  * Returns the maximum segment size for the 'struct scatterlist'
411  * elements.
412  */
413 static inline unsigned int xe_sg_segment_size(struct device *dev)
414 {
415 	struct scatterlist __maybe_unused sg;
416 	size_t max = BIT_ULL(sizeof(sg.length) * 8) - 1;
417 
418 	max = min_t(size_t, max, dma_max_mapping_size(dev));
419 
420 	/*
421 	 * The iommu_dma_map_sg() function ensures iova allocation doesn't
422 	 * cross dma segment boundary. It does so by padding some sg elements.
423 	 * This can cause overflow, ending up with sg->length being set to 0.
424 	 * Avoid this by ensuring maximum segment size is half of 'max'
425 	 * rounded down to PAGE_SIZE.
426 	 */
427 	return round_down(max / 2, PAGE_SIZE);
428 }
429 
430 /**
431  * struct xe_bo_shrink_flags - flags governing the shrink behaviour.
432  * @purge: Only purging allowed. Don't shrink if bo not purgeable.
433  * @writeback: Attempt to immediately move content to swap.
434  */
435 struct xe_bo_shrink_flags {
436 	u32 purge : 1;
437 	u32 writeback : 1;
438 };
439 
440 long xe_bo_shrink(struct ttm_operation_ctx *ctx, struct ttm_buffer_object *bo,
441 		  const struct xe_bo_shrink_flags flags,
442 		  unsigned long *scanned);
443 
444 /**
445  * xe_bo_is_mem_type - Whether the bo currently resides in the given
446  * TTM memory type
447  * @bo: The bo to check.
448  * @mem_type: The TTM memory type.
449  *
450  * Return: true iff the bo resides in @mem_type, false otherwise.
451  */
452 static inline bool xe_bo_is_mem_type(struct xe_bo *bo, u32 mem_type)
453 {
454 	xe_bo_assert_held(bo);
455 	return bo->ttm.resource->mem_type == mem_type;
456 }
457 #endif
458