xref: /linux/include/drm/drm_gpusvm.h (revision 1260ed77798502de9c98020040d2995008de10cc)
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_gpusvm_devmem;
20 struct drm_pagemap;
21 struct drm_pagemap_device_addr;
22 
23 /**
24  * struct drm_gpusvm_devmem_ops - Operations structure for GPU SVM device memory
25  *
26  * This structure defines the operations for GPU Shared Virtual Memory (SVM)
27  * device memory. These operations are provided by the GPU driver to manage device memory
28  * allocations and perform operations such as migration between device memory and system
29  * RAM.
30  */
31 struct drm_gpusvm_devmem_ops {
32 	/**
33 	 * @devmem_release: Release device memory allocation (optional)
34 	 * @devmem_allocation: device memory allocation
35 	 *
36 	 * Release device memory allocation and drop a reference to device
37 	 * memory allocation.
38 	 */
39 	void (*devmem_release)(struct drm_gpusvm_devmem *devmem_allocation);
40 
41 	/**
42 	 * @populate_devmem_pfn: Populate device memory PFN (required for migration)
43 	 * @devmem_allocation: device memory allocation
44 	 * @npages: Number of pages to populate
45 	 * @pfn: Array of page frame numbers to populate
46 	 *
47 	 * Populate device memory page frame numbers (PFN).
48 	 *
49 	 * Return: 0 on success, a negative error code on failure.
50 	 */
51 	int (*populate_devmem_pfn)(struct drm_gpusvm_devmem *devmem_allocation,
52 				   unsigned long npages, unsigned long *pfn);
53 
54 	/**
55 	 * @copy_to_devmem: Copy to device memory (required for migration)
56 	 * @pages: Pointer to array of device memory pages (destination)
57 	 * @dma_addr: Pointer to array of DMA addresses (source)
58 	 * @npages: Number of pages to copy
59 	 *
60 	 * Copy pages to device memory.
61 	 *
62 	 * Return: 0 on success, a negative error code on failure.
63 	 */
64 	int (*copy_to_devmem)(struct page **pages,
65 			      dma_addr_t *dma_addr,
66 			      unsigned long npages);
67 
68 	/**
69 	 * @copy_to_ram: Copy to system RAM (required for migration)
70 	 * @pages: Pointer to array of device memory pages (source)
71 	 * @dma_addr: Pointer to array of DMA addresses (destination)
72 	 * @npages: Number of pages to copy
73 	 *
74 	 * Copy pages to system RAM.
75 	 *
76 	 * Return: 0 on success, a negative error code on failure.
77 	 */
78 	int (*copy_to_ram)(struct page **pages,
79 			   dma_addr_t *dma_addr,
80 			   unsigned long npages);
81 };
82 
83 /**
84  * struct drm_gpusvm_devmem - Structure representing a GPU SVM device memory allocation
85  *
86  * @dev: Pointer to the device structure which device memory allocation belongs to
87  * @mm: Pointer to the mm_struct for the address space
88  * @detached: device memory allocations is detached from device pages
89  * @ops: Pointer to the operations structure for GPU SVM device memory
90  * @dpagemap: The struct drm_pagemap of the pages this allocation belongs to.
91  * @size: Size of device memory allocation
92  */
93 struct drm_gpusvm_devmem {
94 	struct device *dev;
95 	struct mm_struct *mm;
96 	struct completion detached;
97 	const struct drm_gpusvm_devmem_ops *ops;
98 	struct drm_pagemap *dpagemap;
99 	size_t size;
100 };
101 
102 /**
103  * struct drm_gpusvm_ops - Operations structure for GPU SVM
104  *
105  * This structure defines the operations for GPU Shared Virtual Memory (SVM).
106  * These operations are provided by the GPU driver to manage SVM ranges and
107  * notifiers.
108  */
109 struct drm_gpusvm_ops {
110 	/**
111 	 * @notifier_alloc: Allocate a GPU SVM notifier (optional)
112 	 *
113 	 * Allocate a GPU SVM notifier.
114 	 *
115 	 * Return: Pointer to the allocated GPU SVM notifier on success, NULL on failure.
116 	 */
117 	struct drm_gpusvm_notifier *(*notifier_alloc)(void);
118 
119 	/**
120 	 * @notifier_free: Free a GPU SVM notifier (optional)
121 	 * @notifier: Pointer to the GPU SVM notifier to be freed
122 	 *
123 	 * Free a GPU SVM notifier.
124 	 */
125 	void (*notifier_free)(struct drm_gpusvm_notifier *notifier);
126 
127 	/**
128 	 * @range_alloc: Allocate a GPU SVM range (optional)
129 	 * @gpusvm: Pointer to the GPU SVM
130 	 *
131 	 * Allocate a GPU SVM range.
132 	 *
133 	 * Return: Pointer to the allocated GPU SVM range on success, NULL on failure.
134 	 */
135 	struct drm_gpusvm_range *(*range_alloc)(struct drm_gpusvm *gpusvm);
136 
137 	/**
138 	 * @range_free: Free a GPU SVM range (optional)
139 	 * @range: Pointer to the GPU SVM range to be freed
140 	 *
141 	 * Free a GPU SVM range.
142 	 */
143 	void (*range_free)(struct drm_gpusvm_range *range);
144 
145 	/**
146 	 * @invalidate: Invalidate GPU SVM notifier (required)
147 	 * @gpusvm: Pointer to the GPU SVM
148 	 * @notifier: Pointer to the GPU SVM notifier
149 	 * @mmu_range: Pointer to the mmu_notifier_range structure
150 	 *
151 	 * Invalidate the GPU page tables. It can safely walk the notifier range
152 	 * RB tree/list in this function. Called while holding the notifier lock.
153 	 */
154 	void (*invalidate)(struct drm_gpusvm *gpusvm,
155 			   struct drm_gpusvm_notifier *notifier,
156 			   const struct mmu_notifier_range *mmu_range);
157 };
158 
159 /**
160  * struct drm_gpusvm_notifier - Structure representing a GPU SVM notifier
161  *
162  * @gpusvm: Pointer to the GPU SVM structure
163  * @notifier: MMU interval notifier
164  * @itree: Interval tree node for the notifier (inserted in GPU SVM)
165  * @entry: List entry to fast interval tree traversal
166  * @root: Cached root node of the RB tree containing ranges
167  * @range_list: List head containing of ranges in the same order they appear in
168  *              interval tree. This is useful to keep iterating ranges while
169  *              doing modifications to RB tree.
170  * @flags: Flags for notifier
171  * @flags.removed: Flag indicating whether the MMU interval notifier has been
172  *                 removed
173  *
174  * This structure represents a GPU SVM notifier.
175  */
176 struct drm_gpusvm_notifier {
177 	struct drm_gpusvm *gpusvm;
178 	struct mmu_interval_notifier notifier;
179 	struct interval_tree_node itree;
180 	struct list_head entry;
181 	struct rb_root_cached root;
182 	struct list_head range_list;
183 	struct {
184 		u32 removed : 1;
185 	} flags;
186 };
187 
188 /**
189  * struct drm_gpusvm_range - Structure representing a GPU SVM range
190  *
191  * @gpusvm: Pointer to the GPU SVM structure
192  * @notifier: Pointer to the GPU SVM notifier
193  * @refcount: Reference count for the range
194  * @itree: Interval tree node for the range (inserted in GPU SVM notifier)
195  * @entry: List entry to fast interval tree traversal
196  * @notifier_seq: Notifier sequence number of the range's pages
197  * @dma_addr: Device address array
198  * @dpagemap: The struct drm_pagemap of the device pages we're dma-mapping.
199  *            Note this is assuming only one drm_pagemap per range is allowed.
200  * @flags: Flags for range
201  * @flags.migrate_devmem: Flag indicating whether the range can be migrated to device memory
202  * @flags.unmapped: Flag indicating if the range has been unmapped
203  * @flags.partial_unmap: Flag indicating if the range has been partially unmapped
204  * @flags.has_devmem_pages: Flag indicating if the range has devmem pages
205  * @flags.has_dma_mapping: Flag indicating if the range has a DMA mapping
206  *
207  * This structure represents a GPU SVM range used for tracking memory ranges
208  * mapped in a DRM device.
209  */
210 struct drm_gpusvm_range {
211 	struct drm_gpusvm *gpusvm;
212 	struct drm_gpusvm_notifier *notifier;
213 	struct kref refcount;
214 	struct interval_tree_node itree;
215 	struct list_head entry;
216 	unsigned long notifier_seq;
217 	struct drm_pagemap_device_addr *dma_addr;
218 	struct drm_pagemap *dpagemap;
219 	struct {
220 		/* All flags below must be set upon creation */
221 		u16 migrate_devmem : 1;
222 		/* All flags below must be set / cleared under notifier lock */
223 		u16 unmapped : 1;
224 		u16 partial_unmap : 1;
225 		u16 has_devmem_pages : 1;
226 		u16 has_dma_mapping : 1;
227 	} flags;
228 };
229 
230 /**
231  * struct drm_gpusvm - GPU SVM structure
232  *
233  * @name: Name of the GPU SVM
234  * @drm: Pointer to the DRM device structure
235  * @mm: Pointer to the mm_struct for the address space
236  * @device_private_page_owner: Device private pages owner
237  * @mm_start: Start address of GPU SVM
238  * @mm_range: Range of the GPU SVM
239  * @notifier_size: Size of individual notifiers
240  * @ops: Pointer to the operations structure for GPU SVM
241  * @chunk_sizes: Pointer to the array of chunk sizes used in range allocation.
242  *               Entries should be powers of 2 in descending order.
243  * @num_chunks: Number of chunks
244  * @notifier_lock: Read-write semaphore for protecting notifier operations
245  * @root: Cached root node of the Red-Black tree containing GPU SVM notifiers
246  * @notifier_list: list head containing of notifiers in the same order they
247  *                 appear in interval tree. This is useful to keep iterating
248  *                 notifiers while doing modifications to RB tree.
249  *
250  * This structure represents a GPU SVM (Shared Virtual Memory) used for tracking
251  * memory ranges mapped in a DRM (Direct Rendering Manager) device.
252  *
253  * No reference counting is provided, as this is expected to be embedded in the
254  * driver VM structure along with the struct drm_gpuvm, which handles reference
255  * counting.
256  */
257 struct drm_gpusvm {
258 	const char *name;
259 	struct drm_device *drm;
260 	struct mm_struct *mm;
261 	void *device_private_page_owner;
262 	unsigned long mm_start;
263 	unsigned long mm_range;
264 	unsigned long notifier_size;
265 	const struct drm_gpusvm_ops *ops;
266 	const unsigned long *chunk_sizes;
267 	int num_chunks;
268 	struct rw_semaphore notifier_lock;
269 	struct rb_root_cached root;
270 	struct list_head notifier_list;
271 #ifdef CONFIG_LOCKDEP
272 	/**
273 	 * @lock_dep_map: Annotates drm_gpusvm_range_find_or_insert and
274 	 * drm_gpusvm_range_remove with a driver provided lock.
275 	 */
276 	struct lockdep_map *lock_dep_map;
277 #endif
278 };
279 
280 /**
281  * struct drm_gpusvm_ctx - DRM GPU SVM context
282  *
283  * @check_pages_threshold: Check CPU pages for present if chunk is less than or
284  *                         equal to threshold. If not present, reduce chunk
285  *                         size.
286  * @in_notifier: entering from a MMU notifier
287  * @read_only: operating on read-only memory
288  * @devmem_possible: possible to use device memory
289  *
290  * Context that is DRM GPUSVM is operating in (i.e. user arguments).
291  */
292 struct drm_gpusvm_ctx {
293 	unsigned long check_pages_threshold;
294 	unsigned int in_notifier :1;
295 	unsigned int read_only :1;
296 	unsigned int devmem_possible :1;
297 };
298 
299 int drm_gpusvm_init(struct drm_gpusvm *gpusvm,
300 		    const char *name, struct drm_device *drm,
301 		    struct mm_struct *mm, void *device_private_page_owner,
302 		    unsigned long mm_start, unsigned long mm_range,
303 		    unsigned long notifier_size,
304 		    const struct drm_gpusvm_ops *ops,
305 		    const unsigned long *chunk_sizes, int num_chunks);
306 
307 void drm_gpusvm_fini(struct drm_gpusvm *gpusvm);
308 
309 void drm_gpusvm_free(struct drm_gpusvm *gpusvm);
310 
311 struct drm_gpusvm_range *
312 drm_gpusvm_range_find_or_insert(struct drm_gpusvm *gpusvm,
313 				unsigned long fault_addr,
314 				unsigned long gpuva_start,
315 				unsigned long gpuva_end,
316 				const struct drm_gpusvm_ctx *ctx);
317 
318 void drm_gpusvm_range_remove(struct drm_gpusvm *gpusvm,
319 			     struct drm_gpusvm_range *range);
320 
321 int drm_gpusvm_range_evict(struct drm_gpusvm *gpusvm,
322 			   struct drm_gpusvm_range *range);
323 
324 struct drm_gpusvm_range *
325 drm_gpusvm_range_get(struct drm_gpusvm_range *range);
326 
327 void drm_gpusvm_range_put(struct drm_gpusvm_range *range);
328 
329 bool drm_gpusvm_range_pages_valid(struct drm_gpusvm *gpusvm,
330 				  struct drm_gpusvm_range *range);
331 
332 int drm_gpusvm_range_get_pages(struct drm_gpusvm *gpusvm,
333 			       struct drm_gpusvm_range *range,
334 			       const struct drm_gpusvm_ctx *ctx);
335 
336 void drm_gpusvm_range_unmap_pages(struct drm_gpusvm *gpusvm,
337 				  struct drm_gpusvm_range *range,
338 				  const struct drm_gpusvm_ctx *ctx);
339 
340 int drm_gpusvm_migrate_to_devmem(struct drm_gpusvm *gpusvm,
341 				 struct drm_gpusvm_range *range,
342 				 struct drm_gpusvm_devmem *devmem_allocation,
343 				 const struct drm_gpusvm_ctx *ctx);
344 
345 int drm_gpusvm_evict_to_ram(struct drm_gpusvm_devmem *devmem_allocation);
346 
347 const struct dev_pagemap_ops *drm_gpusvm_pagemap_ops_get(void);
348 
349 bool drm_gpusvm_has_mapping(struct drm_gpusvm *gpusvm, unsigned long start,
350 			    unsigned long end);
351 
352 struct drm_gpusvm_range *
353 drm_gpusvm_range_find(struct drm_gpusvm_notifier *notifier, unsigned long start,
354 		      unsigned long end);
355 
356 void drm_gpusvm_range_set_unmapped(struct drm_gpusvm_range *range,
357 				   const struct mmu_notifier_range *mmu_range);
358 
359 void drm_gpusvm_devmem_init(struct drm_gpusvm_devmem *devmem_allocation,
360 			    struct device *dev, struct mm_struct *mm,
361 			    const struct drm_gpusvm_devmem_ops *ops,
362 			    struct drm_pagemap *dpagemap, size_t size);
363 
364 #ifdef CONFIG_LOCKDEP
365 /**
366  * drm_gpusvm_driver_set_lock() - Set the lock protecting accesses to GPU SVM
367  * @gpusvm: Pointer to the GPU SVM structure.
368  * @lock: the lock used to protect the gpuva list. The locking primitive
369  * must contain a dep_map field.
370  *
371  * Call this to annotate drm_gpusvm_range_find_or_insert and
372  * drm_gpusvm_range_remove.
373  */
374 #define drm_gpusvm_driver_set_lock(gpusvm, lock) \
375 	do { \
376 		if (!WARN((gpusvm)->lock_dep_map, \
377 			  "GPUSVM range lock should be set only once."))\
378 			(gpusvm)->lock_dep_map = &(lock)->dep_map;	\
379 	} while (0)
380 #else
381 #define drm_gpusvm_driver_set_lock(gpusvm, lock) do {} while (0)
382 #endif
383 
384 /**
385  * drm_gpusvm_notifier_lock() - Lock GPU SVM notifier
386  * @gpusvm__: Pointer to the GPU SVM structure.
387  *
388  * Abstract client usage GPU SVM notifier lock, take lock
389  */
390 #define drm_gpusvm_notifier_lock(gpusvm__)	\
391 	down_read(&(gpusvm__)->notifier_lock)
392 
393 /**
394  * drm_gpusvm_notifier_unlock() - Unlock GPU SVM notifier
395  * @gpusvm__: Pointer to the GPU SVM structure.
396  *
397  * Abstract client usage GPU SVM notifier lock, drop lock
398  */
399 #define drm_gpusvm_notifier_unlock(gpusvm__)	\
400 	up_read(&(gpusvm__)->notifier_lock)
401 
402 /**
403  * drm_gpusvm_range_start() - GPU SVM range start address
404  * @range: Pointer to the GPU SVM range
405  *
406  * Return: GPU SVM range start address
407  */
408 static inline unsigned long
409 drm_gpusvm_range_start(struct drm_gpusvm_range *range)
410 {
411 	return range->itree.start;
412 }
413 
414 /**
415  * drm_gpusvm_range_end() - GPU SVM range end address
416  * @range: Pointer to the GPU SVM range
417  *
418  * Return: GPU SVM range end address
419  */
420 static inline unsigned long
421 drm_gpusvm_range_end(struct drm_gpusvm_range *range)
422 {
423 	return range->itree.last + 1;
424 }
425 
426 /**
427  * drm_gpusvm_range_size() - GPU SVM range size
428  * @range: Pointer to the GPU SVM range
429  *
430  * Return: GPU SVM range size
431  */
432 static inline unsigned long
433 drm_gpusvm_range_size(struct drm_gpusvm_range *range)
434 {
435 	return drm_gpusvm_range_end(range) - drm_gpusvm_range_start(range);
436 }
437 
438 /**
439  * drm_gpusvm_notifier_start() - GPU SVM notifier start address
440  * @notifier: Pointer to the GPU SVM notifier
441  *
442  * Return: GPU SVM notifier start address
443  */
444 static inline unsigned long
445 drm_gpusvm_notifier_start(struct drm_gpusvm_notifier *notifier)
446 {
447 	return notifier->itree.start;
448 }
449 
450 /**
451  * drm_gpusvm_notifier_end() - GPU SVM notifier end address
452  * @notifier: Pointer to the GPU SVM notifier
453  *
454  * Return: GPU SVM notifier end address
455  */
456 static inline unsigned long
457 drm_gpusvm_notifier_end(struct drm_gpusvm_notifier *notifier)
458 {
459 	return notifier->itree.last + 1;
460 }
461 
462 /**
463  * drm_gpusvm_notifier_size() - GPU SVM notifier size
464  * @notifier: Pointer to the GPU SVM notifier
465  *
466  * Return: GPU SVM notifier size
467  */
468 static inline unsigned long
469 drm_gpusvm_notifier_size(struct drm_gpusvm_notifier *notifier)
470 {
471 	return drm_gpusvm_notifier_end(notifier) -
472 		drm_gpusvm_notifier_start(notifier);
473 }
474 
475 /**
476  * __drm_gpusvm_range_next() - Get the next GPU SVM range in the list
477  * @range: a pointer to the current GPU SVM range
478  *
479  * Return: A pointer to the next drm_gpusvm_range if available, or NULL if the
480  *         current range is the last one or if the input range is NULL.
481  */
482 static inline struct drm_gpusvm_range *
483 __drm_gpusvm_range_next(struct drm_gpusvm_range *range)
484 {
485 	if (range && !list_is_last(&range->entry,
486 				   &range->notifier->range_list))
487 		return list_next_entry(range, entry);
488 
489 	return NULL;
490 }
491 
492 /**
493  * drm_gpusvm_for_each_range() - Iterate over GPU SVM ranges in a notifier
494  * @range__: Iterator variable for the ranges. If set, it indicates the start of
495  *	     the iterator. If NULL, call drm_gpusvm_range_find() to get the range.
496  * @notifier__: Pointer to the GPU SVM notifier
497  * @start__: Start address of the range
498  * @end__: End address of the range
499  *
500  * This macro is used to iterate over GPU SVM ranges in a notifier. It is safe
501  * to use while holding the driver SVM lock or the notifier lock.
502  */
503 #define drm_gpusvm_for_each_range(range__, notifier__, start__, end__)	\
504 	for ((range__) = (range__) ?:					\
505 	     drm_gpusvm_range_find((notifier__), (start__), (end__));	\
506 	     (range__) && (drm_gpusvm_range_start(range__) < (end__));	\
507 	     (range__) = __drm_gpusvm_range_next(range__))
508 
509 #endif /* __DRM_GPUSVM_H__ */
510