xref: /linux/include/drm/drm_gpusvm.h (revision 55a42f78ffd386e01a5404419f8c5ded7db70a21)
1 /* SPDX-License-Identifier: GPL-2.0-only OR MIT */
2 /*
3  * Copyright © 2024 Intel Corporation
4  */
5 
6 #ifndef __DRM_GPUSVM_H__
7 #define __DRM_GPUSVM_H__
8 
9 #include <linux/kref.h>
10 #include <linux/interval_tree.h>
11 #include <linux/mmu_notifier.h>
12 
13 struct dev_pagemap_ops;
14 struct drm_device;
15 struct drm_gpusvm;
16 struct drm_gpusvm_notifier;
17 struct drm_gpusvm_ops;
18 struct drm_gpusvm_range;
19 struct drm_pagemap;
20 struct drm_pagemap_addr;
21 
22 /**
23  * struct drm_gpusvm_ops - Operations structure for GPU SVM
24  *
25  * This structure defines the operations for GPU Shared Virtual Memory (SVM).
26  * These operations are provided by the GPU driver to manage SVM ranges and
27  * notifiers.
28  */
29 struct drm_gpusvm_ops {
30 	/**
31 	 * @notifier_alloc: Allocate a GPU SVM notifier (optional)
32 	 *
33 	 * Allocate a GPU SVM notifier.
34 	 *
35 	 * Return: Pointer to the allocated GPU SVM notifier on success, NULL on failure.
36 	 */
37 	struct drm_gpusvm_notifier *(*notifier_alloc)(void);
38 
39 	/**
40 	 * @notifier_free: Free a GPU SVM notifier (optional)
41 	 * @notifier: Pointer to the GPU SVM notifier to be freed
42 	 *
43 	 * Free a GPU SVM notifier.
44 	 */
45 	void (*notifier_free)(struct drm_gpusvm_notifier *notifier);
46 
47 	/**
48 	 * @range_alloc: Allocate a GPU SVM range (optional)
49 	 * @gpusvm: Pointer to the GPU SVM
50 	 *
51 	 * Allocate a GPU SVM range.
52 	 *
53 	 * Return: Pointer to the allocated GPU SVM range on success, NULL on failure.
54 	 */
55 	struct drm_gpusvm_range *(*range_alloc)(struct drm_gpusvm *gpusvm);
56 
57 	/**
58 	 * @range_free: Free a GPU SVM range (optional)
59 	 * @range: Pointer to the GPU SVM range to be freed
60 	 *
61 	 * Free a GPU SVM range.
62 	 */
63 	void (*range_free)(struct drm_gpusvm_range *range);
64 
65 	/**
66 	 * @invalidate: Invalidate GPU SVM notifier (required)
67 	 * @gpusvm: Pointer to the GPU SVM
68 	 * @notifier: Pointer to the GPU SVM notifier
69 	 * @mmu_range: Pointer to the mmu_notifier_range structure
70 	 *
71 	 * Invalidate the GPU page tables. It can safely walk the notifier range
72 	 * RB tree/list in this function. Called while holding the notifier lock.
73 	 */
74 	void (*invalidate)(struct drm_gpusvm *gpusvm,
75 			   struct drm_gpusvm_notifier *notifier,
76 			   const struct mmu_notifier_range *mmu_range);
77 };
78 
79 /**
80  * struct drm_gpusvm_notifier - Structure representing a GPU SVM notifier
81  *
82  * @gpusvm: Pointer to the GPU SVM structure
83  * @notifier: MMU interval notifier
84  * @itree: Interval tree node for the notifier (inserted in GPU SVM)
85  * @entry: List entry to fast interval tree traversal
86  * @root: Cached root node of the RB tree containing ranges
87  * @range_list: List head containing of ranges in the same order they appear in
88  *              interval tree. This is useful to keep iterating ranges while
89  *              doing modifications to RB tree.
90  * @flags: Flags for notifier
91  * @flags.removed: Flag indicating whether the MMU interval notifier has been
92  *                 removed
93  *
94  * This structure represents a GPU SVM notifier.
95  */
96 struct drm_gpusvm_notifier {
97 	struct drm_gpusvm *gpusvm;
98 	struct mmu_interval_notifier notifier;
99 	struct interval_tree_node itree;
100 	struct list_head entry;
101 	struct rb_root_cached root;
102 	struct list_head range_list;
103 	struct {
104 		u32 removed : 1;
105 	} flags;
106 };
107 
108 /**
109  * struct drm_gpusvm_pages_flags - Structure representing a GPU SVM pages flags
110  *
111  * @migrate_devmem: Flag indicating whether the pages can be migrated to device memory
112  * @unmapped: Flag indicating if the pages has been unmapped
113  * @partial_unmap: Flag indicating if the pages has been partially unmapped
114  * @has_devmem_pages: Flag indicating if the pages has devmem pages
115  * @has_dma_mapping: Flag indicating if the pages has a DMA mapping
116  * @__flags: Flags for pages in u16 form (used for READ_ONCE)
117  */
118 struct drm_gpusvm_pages_flags {
119 	union {
120 		struct {
121 			/* All flags below must be set upon creation */
122 			u16 migrate_devmem : 1;
123 			/* All flags below must be set / cleared under notifier lock */
124 			u16 unmapped : 1;
125 			u16 partial_unmap : 1;
126 			u16 has_devmem_pages : 1;
127 			u16 has_dma_mapping : 1;
128 		};
129 		u16 __flags;
130 	};
131 };
132 
133 /**
134  * struct drm_gpusvm_pages - Structure representing a GPU SVM mapped pages
135  *
136  * @dma_addr: Device address array
137  * @dpagemap: The struct drm_pagemap of the device pages we're dma-mapping.
138  *            Note this is assuming only one drm_pagemap per range is allowed.
139  * @notifier_seq: Notifier sequence number of the range's pages
140  * @flags: Flags for range
141  * @flags.migrate_devmem: Flag indicating whether the range can be migrated to device memory
142  * @flags.unmapped: Flag indicating if the range has been unmapped
143  * @flags.partial_unmap: Flag indicating if the range has been partially unmapped
144  * @flags.has_devmem_pages: Flag indicating if the range has devmem pages
145  * @flags.has_dma_mapping: Flag indicating if the range has a DMA mapping
146  */
147 struct drm_gpusvm_pages {
148 	struct drm_pagemap_addr *dma_addr;
149 	struct drm_pagemap *dpagemap;
150 	unsigned long notifier_seq;
151 	struct drm_gpusvm_pages_flags flags;
152 };
153 
154 /**
155  * struct drm_gpusvm_range - Structure representing a GPU SVM range
156  *
157  * @gpusvm: Pointer to the GPU SVM structure
158  * @notifier: Pointer to the GPU SVM notifier
159  * @refcount: Reference count for the range
160  * @itree: Interval tree node for the range (inserted in GPU SVM notifier)
161  * @entry: List entry to fast interval tree traversal
162  * @pages: The pages for this range.
163  *
164  * This structure represents a GPU SVM range used for tracking memory ranges
165  * mapped in a DRM device.
166  */
167 struct drm_gpusvm_range {
168 	struct drm_gpusvm *gpusvm;
169 	struct drm_gpusvm_notifier *notifier;
170 	struct kref refcount;
171 	struct interval_tree_node itree;
172 	struct list_head entry;
173 	struct drm_gpusvm_pages pages;
174 };
175 
176 /**
177  * struct drm_gpusvm - GPU SVM structure
178  *
179  * @name: Name of the GPU SVM
180  * @drm: Pointer to the DRM device structure
181  * @mm: Pointer to the mm_struct for the address space
182  * @device_private_page_owner: Device private pages owner
183  * @mm_start: Start address of GPU SVM
184  * @mm_range: Range of the GPU SVM
185  * @notifier_size: Size of individual notifiers
186  * @ops: Pointer to the operations structure for GPU SVM
187  * @chunk_sizes: Pointer to the array of chunk sizes used in range allocation.
188  *               Entries should be powers of 2 in descending order.
189  * @num_chunks: Number of chunks
190  * @notifier_lock: Read-write semaphore for protecting notifier operations
191  * @root: Cached root node of the Red-Black tree containing GPU SVM notifiers
192  * @notifier_list: list head containing of notifiers in the same order they
193  *                 appear in interval tree. This is useful to keep iterating
194  *                 notifiers while doing modifications to RB tree.
195  *
196  * This structure represents a GPU SVM (Shared Virtual Memory) used for tracking
197  * memory ranges mapped in a DRM (Direct Rendering Manager) device.
198  *
199  * No reference counting is provided, as this is expected to be embedded in the
200  * driver VM structure along with the struct drm_gpuvm, which handles reference
201  * counting.
202  */
203 struct drm_gpusvm {
204 	const char *name;
205 	struct drm_device *drm;
206 	struct mm_struct *mm;
207 	void *device_private_page_owner;
208 	unsigned long mm_start;
209 	unsigned long mm_range;
210 	unsigned long notifier_size;
211 	const struct drm_gpusvm_ops *ops;
212 	const unsigned long *chunk_sizes;
213 	int num_chunks;
214 	struct rw_semaphore notifier_lock;
215 	struct rb_root_cached root;
216 	struct list_head notifier_list;
217 #ifdef CONFIG_LOCKDEP
218 	/**
219 	 * @lock_dep_map: Annotates drm_gpusvm_range_find_or_insert and
220 	 * drm_gpusvm_range_remove with a driver provided lock.
221 	 */
222 	struct lockdep_map *lock_dep_map;
223 #endif
224 };
225 
226 /**
227  * struct drm_gpusvm_ctx - DRM GPU SVM context
228  *
229  * @check_pages_threshold: Check CPU pages for present if chunk is less than or
230  *                         equal to threshold. If not present, reduce chunk
231  *                         size.
232  * @timeslice_ms: The timeslice MS which in minimum time a piece of memory
233  *		  remains with either exclusive GPU or CPU access.
234  * @in_notifier: entering from a MMU notifier
235  * @read_only: operating on read-only memory
236  * @devmem_possible: possible to use device memory
237  * @devmem_only: use only device memory
238  *
239  * Context that is DRM GPUSVM is operating in (i.e. user arguments).
240  */
241 struct drm_gpusvm_ctx {
242 	unsigned long check_pages_threshold;
243 	unsigned long timeslice_ms;
244 	unsigned int in_notifier :1;
245 	unsigned int read_only :1;
246 	unsigned int devmem_possible :1;
247 	unsigned int devmem_only :1;
248 };
249 
250 int drm_gpusvm_init(struct drm_gpusvm *gpusvm,
251 		    const char *name, struct drm_device *drm,
252 		    struct mm_struct *mm, void *device_private_page_owner,
253 		    unsigned long mm_start, unsigned long mm_range,
254 		    unsigned long notifier_size,
255 		    const struct drm_gpusvm_ops *ops,
256 		    const unsigned long *chunk_sizes, int num_chunks);
257 
258 void drm_gpusvm_fini(struct drm_gpusvm *gpusvm);
259 
260 void drm_gpusvm_free(struct drm_gpusvm *gpusvm);
261 
262 unsigned long
263 drm_gpusvm_find_vma_start(struct drm_gpusvm *gpusvm,
264 			  unsigned long start,
265 			  unsigned long end);
266 
267 struct drm_gpusvm_range *
268 drm_gpusvm_range_find_or_insert(struct drm_gpusvm *gpusvm,
269 				unsigned long fault_addr,
270 				unsigned long gpuva_start,
271 				unsigned long gpuva_end,
272 				const struct drm_gpusvm_ctx *ctx);
273 
274 void drm_gpusvm_range_remove(struct drm_gpusvm *gpusvm,
275 			     struct drm_gpusvm_range *range);
276 
277 int drm_gpusvm_range_evict(struct drm_gpusvm *gpusvm,
278 			   struct drm_gpusvm_range *range);
279 
280 struct drm_gpusvm_range *
281 drm_gpusvm_range_get(struct drm_gpusvm_range *range);
282 
283 void drm_gpusvm_range_put(struct drm_gpusvm_range *range);
284 
285 bool drm_gpusvm_range_pages_valid(struct drm_gpusvm *gpusvm,
286 				  struct drm_gpusvm_range *range);
287 
288 int drm_gpusvm_range_get_pages(struct drm_gpusvm *gpusvm,
289 			       struct drm_gpusvm_range *range,
290 			       const struct drm_gpusvm_ctx *ctx);
291 
292 void drm_gpusvm_range_unmap_pages(struct drm_gpusvm *gpusvm,
293 				  struct drm_gpusvm_range *range,
294 				  const struct drm_gpusvm_ctx *ctx);
295 
296 bool drm_gpusvm_has_mapping(struct drm_gpusvm *gpusvm, unsigned long start,
297 			    unsigned long end);
298 
299 struct drm_gpusvm_notifier *
300 drm_gpusvm_notifier_find(struct drm_gpusvm *gpusvm, unsigned long start,
301 			 unsigned long end);
302 
303 struct drm_gpusvm_range *
304 drm_gpusvm_range_find(struct drm_gpusvm_notifier *notifier, unsigned long start,
305 		      unsigned long end);
306 
307 void drm_gpusvm_range_set_unmapped(struct drm_gpusvm_range *range,
308 				   const struct mmu_notifier_range *mmu_range);
309 
310 int drm_gpusvm_get_pages(struct drm_gpusvm *gpusvm,
311 			 struct drm_gpusvm_pages *svm_pages,
312 			 struct mm_struct *mm,
313 			 struct mmu_interval_notifier *notifier,
314 			 unsigned long pages_start, unsigned long pages_end,
315 			 const struct drm_gpusvm_ctx *ctx);
316 
317 void drm_gpusvm_unmap_pages(struct drm_gpusvm *gpusvm,
318 			    struct drm_gpusvm_pages *svm_pages,
319 			    unsigned long npages,
320 			    const struct drm_gpusvm_ctx *ctx);
321 
322 void drm_gpusvm_free_pages(struct drm_gpusvm *gpusvm,
323 			   struct drm_gpusvm_pages *svm_pages,
324 			   unsigned long npages);
325 
326 #ifdef CONFIG_LOCKDEP
327 /**
328  * drm_gpusvm_driver_set_lock() - Set the lock protecting accesses to GPU SVM
329  * @gpusvm: Pointer to the GPU SVM structure.
330  * @lock: the lock used to protect the gpuva list. The locking primitive
331  * must contain a dep_map field.
332  *
333  * Call this to annotate drm_gpusvm_range_find_or_insert and
334  * drm_gpusvm_range_remove.
335  */
336 #define drm_gpusvm_driver_set_lock(gpusvm, lock) \
337 	do { \
338 		if (!WARN((gpusvm)->lock_dep_map, \
339 			  "GPUSVM range lock should be set only once."))\
340 			(gpusvm)->lock_dep_map = &(lock)->dep_map;	\
341 	} while (0)
342 #else
343 #define drm_gpusvm_driver_set_lock(gpusvm, lock) do {} while (0)
344 #endif
345 
346 /**
347  * drm_gpusvm_notifier_lock() - Lock GPU SVM notifier
348  * @gpusvm__: Pointer to the GPU SVM structure.
349  *
350  * Abstract client usage GPU SVM notifier lock, take lock
351  */
352 #define drm_gpusvm_notifier_lock(gpusvm__)	\
353 	down_read(&(gpusvm__)->notifier_lock)
354 
355 /**
356  * drm_gpusvm_notifier_unlock() - Unlock GPU SVM notifier
357  * @gpusvm__: Pointer to the GPU SVM structure.
358  *
359  * Abstract client usage GPU SVM notifier lock, drop lock
360  */
361 #define drm_gpusvm_notifier_unlock(gpusvm__)	\
362 	up_read(&(gpusvm__)->notifier_lock)
363 
364 /**
365  * drm_gpusvm_range_start() - GPU SVM range start address
366  * @range: Pointer to the GPU SVM range
367  *
368  * Return: GPU SVM range start address
369  */
370 static inline unsigned long
371 drm_gpusvm_range_start(struct drm_gpusvm_range *range)
372 {
373 	return range->itree.start;
374 }
375 
376 /**
377  * drm_gpusvm_range_end() - GPU SVM range end address
378  * @range: Pointer to the GPU SVM range
379  *
380  * Return: GPU SVM range end address
381  */
382 static inline unsigned long
383 drm_gpusvm_range_end(struct drm_gpusvm_range *range)
384 {
385 	return range->itree.last + 1;
386 }
387 
388 /**
389  * drm_gpusvm_range_size() - GPU SVM range size
390  * @range: Pointer to the GPU SVM range
391  *
392  * Return: GPU SVM range size
393  */
394 static inline unsigned long
395 drm_gpusvm_range_size(struct drm_gpusvm_range *range)
396 {
397 	return drm_gpusvm_range_end(range) - drm_gpusvm_range_start(range);
398 }
399 
400 /**
401  * drm_gpusvm_notifier_start() - GPU SVM notifier start address
402  * @notifier: Pointer to the GPU SVM notifier
403  *
404  * Return: GPU SVM notifier start address
405  */
406 static inline unsigned long
407 drm_gpusvm_notifier_start(struct drm_gpusvm_notifier *notifier)
408 {
409 	return notifier->itree.start;
410 }
411 
412 /**
413  * drm_gpusvm_notifier_end() - GPU SVM notifier end address
414  * @notifier: Pointer to the GPU SVM notifier
415  *
416  * Return: GPU SVM notifier end address
417  */
418 static inline unsigned long
419 drm_gpusvm_notifier_end(struct drm_gpusvm_notifier *notifier)
420 {
421 	return notifier->itree.last + 1;
422 }
423 
424 /**
425  * drm_gpusvm_notifier_size() - GPU SVM notifier size
426  * @notifier: Pointer to the GPU SVM notifier
427  *
428  * Return: GPU SVM notifier size
429  */
430 static inline unsigned long
431 drm_gpusvm_notifier_size(struct drm_gpusvm_notifier *notifier)
432 {
433 	return drm_gpusvm_notifier_end(notifier) -
434 		drm_gpusvm_notifier_start(notifier);
435 }
436 
437 /**
438  * __drm_gpusvm_range_next() - Get the next GPU SVM range in the list
439  * @range: a pointer to the current GPU SVM range
440  *
441  * Return: A pointer to the next drm_gpusvm_range if available, or NULL if the
442  *         current range is the last one or if the input range is NULL.
443  */
444 static inline struct drm_gpusvm_range *
445 __drm_gpusvm_range_next(struct drm_gpusvm_range *range)
446 {
447 	if (range && !list_is_last(&range->entry,
448 				   &range->notifier->range_list))
449 		return list_next_entry(range, entry);
450 
451 	return NULL;
452 }
453 
454 /**
455  * drm_gpusvm_for_each_range() - Iterate over GPU SVM ranges in a notifier
456  * @range__: Iterator variable for the ranges. If set, it indicates the start of
457  *	     the iterator. If NULL, call drm_gpusvm_range_find() to get the range.
458  * @notifier__: Pointer to the GPU SVM notifier
459  * @start__: Start address of the range
460  * @end__: End address of the range
461  *
462  * This macro is used to iterate over GPU SVM ranges in a notifier. It is safe
463  * to use while holding the driver SVM lock or the notifier lock.
464  */
465 #define drm_gpusvm_for_each_range(range__, notifier__, start__, end__)	\
466 	for ((range__) = (range__) ?:					\
467 	     drm_gpusvm_range_find((notifier__), (start__), (end__));	\
468 	     (range__) && (drm_gpusvm_range_start(range__) < (end__));	\
469 	     (range__) = __drm_gpusvm_range_next(range__))
470 
471 /**
472  * drm_gpusvm_for_each_range_safe() - Safely iterate over GPU SVM ranges in a notifier
473  * @range__: Iterator variable for the ranges
474  * @next__: Iterator variable for the ranges temporay storage
475  * @notifier__: Pointer to the GPU SVM notifier
476  * @start__: Start address of the range
477  * @end__: End address of the range
478  *
479  * This macro is used to iterate over GPU SVM ranges in a notifier while
480  * removing ranges from it.
481  */
482 #define drm_gpusvm_for_each_range_safe(range__, next__, notifier__, start__, end__)	\
483 	for ((range__) = drm_gpusvm_range_find((notifier__), (start__), (end__)),	\
484 	     (next__) = __drm_gpusvm_range_next(range__);				\
485 	     (range__) && (drm_gpusvm_range_start(range__) < (end__));			\
486 	     (range__) = (next__), (next__) = __drm_gpusvm_range_next(range__))
487 
488 /**
489  * __drm_gpusvm_notifier_next() - get the next drm_gpusvm_notifier in the list
490  * @notifier: a pointer to the current drm_gpusvm_notifier
491  *
492  * Return: A pointer to the next drm_gpusvm_notifier if available, or NULL if
493  *         the current notifier is the last one or if the input notifier is
494  *         NULL.
495  */
496 static inline struct drm_gpusvm_notifier *
497 __drm_gpusvm_notifier_next(struct drm_gpusvm_notifier *notifier)
498 {
499 	if (notifier && !list_is_last(&notifier->entry,
500 				      &notifier->gpusvm->notifier_list))
501 		return list_next_entry(notifier, entry);
502 
503 	return NULL;
504 }
505 
506 /**
507  * drm_gpusvm_for_each_notifier() - Iterate over GPU SVM notifiers in a gpusvm
508  * @notifier__: Iterator variable for the notifiers
509  * @gpusvm__: Pointer to the GPU SVM notifier
510  * @start__: Start address of the notifier
511  * @end__: End address of the notifier
512  *
513  * This macro is used to iterate over GPU SVM notifiers in a gpusvm.
514  */
515 #define drm_gpusvm_for_each_notifier(notifier__, gpusvm__, start__, end__)		\
516 	for ((notifier__) = drm_gpusvm_notifier_find((gpusvm__), (start__), (end__));	\
517 	     (notifier__) && (drm_gpusvm_notifier_start(notifier__) < (end__));		\
518 	     (notifier__) = __drm_gpusvm_notifier_next(notifier__))
519 
520 /**
521  * drm_gpusvm_for_each_notifier_safe() - Safely iterate over GPU SVM notifiers in a gpusvm
522  * @notifier__: Iterator variable for the notifiers
523  * @next__: Iterator variable for the notifiers temporay storage
524  * @gpusvm__: Pointer to the GPU SVM notifier
525  * @start__: Start address of the notifier
526  * @end__: End address of the notifier
527  *
528  * This macro is used to iterate over GPU SVM notifiers in a gpusvm while
529  * removing notifiers from it.
530  */
531 #define drm_gpusvm_for_each_notifier_safe(notifier__, next__, gpusvm__, start__, end__)	\
532 	for ((notifier__) = drm_gpusvm_notifier_find((gpusvm__), (start__), (end__)),	\
533 	     (next__) = __drm_gpusvm_notifier_next(notifier__);				\
534 	     (notifier__) && (drm_gpusvm_notifier_start(notifier__) < (end__));		\
535 	     (notifier__) = (next__), (next__) = __drm_gpusvm_notifier_next(notifier__))
536 
537 #endif /* __DRM_GPUSVM_H__ */
538