xref: /linux/drivers/gpu/drm/amd/amdkfd/kfd_svm.h (revision 570f7e331f5febb30f1384817463c7e42b65ca7d)
1 /* SPDX-License-Identifier: GPL-2.0 OR MIT */
2 /*
3  * Copyright 2020-2021 Advanced Micro Devices, Inc.
4  *
5  * Permission is hereby granted, free of charge, to any person obtaining a
6  * copy of this software and associated documentation files (the "Software"),
7  * to deal in the Software without restriction, including without limitation
8  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
9  * and/or sell copies of the Software, and to permit persons to whom the
10  * Software is furnished to do so, subject to the following conditions:
11  *
12  * The above copyright notice and this permission notice shall be included in
13  * all copies or substantial portions of the Software.
14  *
15  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
18  * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR
19  * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
20  * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
21  * OTHER DEALINGS IN THE SOFTWARE.
22  *
23  */
24 
25 #ifndef KFD_SVM_H_
26 #define KFD_SVM_H_
27 
28 #if IS_ENABLED(CONFIG_HSA_AMD_SVM)
29 
30 #include <linux/rwsem.h>
31 #include <linux/list.h>
32 #include <linux/mutex.h>
33 #include <linux/sched/mm.h>
34 #include "amdgpu.h"
35 #include "kfd_priv.h"
36 
37 #define SVM_RANGE_VRAM_DOMAIN (1UL << 0)
38 #define SVM_ADEV_PGMAP_OWNER(adev)\
39 			((adev)->hive ? (void *)(adev)->hive : (void *)(adev))
40 
41 struct svm_range_bo {
42 	struct amdgpu_bo		bo;
43 	struct kref			kref;
44 	struct list_head		range_list; /* all svm ranges shared this bo */
45 	spinlock_t			list_lock;
46 	struct mm_struct		*mm;
47 	uint32_t			evicting;
48 	struct work_struct		release_work;
49 	struct kfd_node			*node;
50 };
51 
52 enum svm_work_list_ops {
53 	SVM_OP_NULL,
54 	SVM_OP_UNMAP_RANGE,
55 	SVM_OP_UPDATE_RANGE_NOTIFIER,
56 	SVM_OP_UPDATE_RANGE_NOTIFIER_AND_MAP,
57 	SVM_OP_ADD_RANGE,
58 	SVM_OP_ADD_RANGE_AND_MAP
59 };
60 
61 struct svm_work_list_item {
62 	enum svm_work_list_ops op;
63 	struct mm_struct *mm;
64 };
65 
66 /**
67  * struct svm_range - shared virtual memory range
68  *
69  * @svms:       list of svm ranges, structure defined in kfd_process
70  * @migrate_mutex: to serialize range migration, validation and mapping update
71  * @start:      range start address in pages
72  * @last:       range last address in pages
73  * @it_node:    node [start, last] stored in interval tree, start, last are page
74  *              aligned, page size is (last - start + 1)
75  * @list:       link list node, used to scan all ranges of svms
76  * @update_list:link list node used to add to update_list
77  * @mapping:    bo_va mapping structure to create and update GPU page table
78  * @npages:     number of pages
79  * @vram_pages: vram pages number in this svm_range
80  * @dma_addr:   dma mapping address on each GPU for system memory physical page
81  * @ttm_res:    vram ttm resource map
82  * @offset:     range start offset within mm_nodes
83  * @svm_bo:     struct to manage splited amdgpu_bo
84  * @svm_bo_list:link list node, to scan all ranges which share same svm_bo
85  * @lock:       protect prange start, last, child_list, svm_bo_list
86  * @saved_flags:save/restore current PF_MEMALLOC flags
87  * @flags:      flags defined as KFD_IOCTL_SVM_FLAG_*
88  * @perferred_loc: perferred location, 0 for CPU, or GPU id
89  * @perfetch_loc: last prefetch location, 0 for CPU, or GPU id
90  * @actual_loc: this svm_range location. 0: all pages are from sys ram;
91  *              GPU id: this svm_range may include vram pages from GPU with
92  *              id actual_loc.
93  * @granularity:migration granularity, log2 num pages
94  * @invalid:    not 0 means cpu page table is invalidated
95  * @validate_timestamp: system timestamp when range is validated
96  * @notifier:   register mmu interval notifier
97  * @work_item:  deferred work item information
98  * @deferred_list: list header used to add range to deferred list
99  * @child_list: list header for split ranges which are not added to svms yet
100  * @bitmap_access: index bitmap of GPUs which can access the range
101  * @bitmap_aip: index bitmap of GPUs which can access the range in place
102  * @bitmap_needs_unmap: index bitmap of GPUs which currently set NO_ACCESS
103  * @bitmap_mapped: index bitmap of GPUs which currently have the range mapped
104  * @mapping_done: true if range_validate_and_map complete successfully
105  *
106  * Data structure for virtual memory range shared by CPU and GPUs, it can be
107  * allocated from system memory ram or device vram, and migrate from ram to vram
108  * or from vram to ram.
109  */
110 struct svm_range {
111 	struct svm_range_list		*svms;
112 	struct mutex			migrate_mutex;
113 	unsigned long			start;
114 	unsigned long			last;
115 	struct interval_tree_node	it_node;
116 	struct list_head		list;
117 	struct list_head		update_list;
118 	uint64_t			npages;
119 	uint64_t			vram_pages;
120 	dma_addr_t			*dma_addr[MAX_GPU_INSTANCE];
121 	struct ttm_resource		*ttm_res;
122 	uint64_t			offset;
123 	struct svm_range_bo		*svm_bo;
124 	struct list_head		svm_bo_list;
125 	struct mutex                    lock;
126 	unsigned int                    saved_flags;
127 	uint32_t			flags;
128 	uint32_t			preferred_loc;
129 	uint32_t			prefetch_loc;
130 	uint32_t			actual_loc;
131 	uint8_t				granularity;
132 	atomic_t			invalid;
133 	ktime_t				validate_timestamp;
134 	struct mmu_interval_notifier	notifier;
135 	struct svm_work_list_item	work_item;
136 	struct list_head		deferred_list;
137 	struct list_head		child_list;
138 	DECLARE_BITMAP(bitmap_access, MAX_GPU_INSTANCE);
139 	DECLARE_BITMAP(bitmap_aip, MAX_GPU_INSTANCE);
140 	DECLARE_BITMAP(bitmap_needs_unmap, MAX_GPU_INSTANCE);
141 	DECLARE_BITMAP(bitmap_mapped, MAX_GPU_INSTANCE);
142 	bool				mapping_done;
143 	atomic_t			queue_refcount;
144 };
145 
146 static inline void svm_range_lock(struct svm_range *prange)
147 {
148 	mutex_lock(&prange->lock);
149 	prange->saved_flags = memalloc_noreclaim_save();
150 
151 }
152 static inline void svm_range_unlock(struct svm_range *prange)
153 {
154 	memalloc_noreclaim_restore(prange->saved_flags);
155 	mutex_unlock(&prange->lock);
156 }
157 
158 static inline struct svm_range_bo *svm_range_bo_ref(struct svm_range_bo *svm_bo)
159 {
160 	if (svm_bo)
161 		kref_get(&svm_bo->kref);
162 
163 	return svm_bo;
164 }
165 
166 int svm_range_list_init(struct kfd_process *p);
167 void svm_range_list_fini(struct kfd_process *p);
168 int svm_ioctl(struct kfd_process *p, enum kfd_ioctl_svm_op op, uint64_t start,
169 	      uint64_t size, uint32_t nattrs,
170 	      struct kfd_ioctl_svm_attribute *attrs);
171 struct svm_range *svm_range_from_addr(struct svm_range_list *svms,
172 				      unsigned long addr,
173 				      struct svm_range **parent);
174 struct kfd_node *svm_range_get_node_by_id(struct svm_range *prange,
175 					  uint32_t gpu_id);
176 void svm_range_bo_destroy(struct ttm_buffer_object *tbo);
177 int svm_range_vram_node_new(struct kfd_node *node, struct svm_range *prange,
178 			    bool clear);
179 void svm_range_vram_node_free(struct svm_range *prange);
180 int svm_range_restore_pages(struct amdgpu_device *adev, unsigned int pasid,
181 			    uint32_t vmid, uint32_t node_id, uint64_t addr, uint64_t ts,
182 			    bool write_fault);
183 int svm_range_evict_svm_bo(struct amdgpu_bo *bo);
184 
185 void svm_range_add_list_work(struct svm_range_list *svms,
186 			     struct svm_range *prange, struct mm_struct *mm,
187 			     enum svm_work_list_ops op);
188 void schedule_deferred_list_work(struct svm_range_list *svms);
189 void svm_range_dma_unmap_dev(struct device *dev, dma_addr_t *dma_addr,
190 			 unsigned long offset, unsigned long npages);
191 void svm_range_dma_unmap(struct svm_range *prange);
192 void svm_range_get_info(struct kfd_process *p, uint32_t *num_svm_ranges,
193 			uint64_t *svm_priv_data_size);
194 int kfd_criu_checkpoint_svm(struct kfd_process *p,
195 			    uint8_t __user *user_priv_data,
196 			    uint64_t *priv_offset);
197 int kfd_criu_restore_svm(struct kfd_process *p,
198 			 uint8_t __user *user_priv_ptr,
199 			 uint64_t *priv_data_offset,
200 			 uint64_t max_priv_data_size);
201 int kfd_criu_resume_svm(struct kfd_process *p);
202 struct kfd_process_device *
203 svm_range_get_pdd_by_node(struct svm_range *prange, struct kfd_node *node);
204 void svm_range_list_lock_and_flush_work(struct svm_range_list *svms, struct mm_struct *mm);
205 
206 /* SVM API and HMM page migration work together, device memory type
207  * is initialized to not 0 when page migration register device memory.
208  */
209 #define KFD_IS_SVM_API_SUPPORTED(adev) ((adev)->kfd.pgmap.type != 0 ||\
210 					((adev)->apu_prefer_gtt))
211 
212 void svm_range_bo_unref_async(struct svm_range_bo *svm_bo);
213 
214 void svm_range_set_max_pages(struct amdgpu_device *adev);
215 int svm_range_switch_xnack_reserve_mem(struct kfd_process *p, bool xnack_enabled);
216 
217 #else
218 
219 struct kfd_process;
220 
221 static inline int svm_range_list_init(struct kfd_process *p)
222 {
223 	return 0;
224 }
225 static inline void svm_range_list_fini(struct kfd_process *p)
226 {
227 	/* empty */
228 }
229 
230 static inline int svm_range_restore_pages(struct amdgpu_device *adev,
231 					  unsigned int pasid,
232 					  uint32_t client_id, uint32_t node_id,
233 					  uint64_t addr, uint64_t ts, bool write_fault)
234 {
235 	return -EFAULT;
236 }
237 
238 static inline void svm_range_bo_destroy(struct ttm_buffer_object *tbo)
239 {
240 }
241 
242 static inline int svm_range_evict_svm_bo(struct amdgpu_bo *bo)
243 {
244 	return 0;
245 }
246 
247 static inline void svm_range_get_info(struct kfd_process *p,
248 				      uint32_t *num_svm_ranges,
249 				      uint64_t *svm_priv_data_size)
250 {
251 	*num_svm_ranges = 0;
252 	*svm_priv_data_size = 0;
253 }
254 
255 static inline int kfd_criu_checkpoint_svm(struct kfd_process *p,
256 					  uint8_t __user *user_priv_data,
257 					  uint64_t *priv_offset)
258 {
259 	return 0;
260 }
261 
262 static inline int kfd_criu_restore_svm(struct kfd_process *p,
263 				       uint8_t __user *user_priv_ptr,
264 				       uint64_t *priv_data_offset,
265 				       uint64_t max_priv_data_size)
266 {
267 	return -EINVAL;
268 }
269 
270 static inline int kfd_criu_resume_svm(struct kfd_process *p)
271 {
272 	return 0;
273 }
274 
275 static inline void svm_range_set_max_pages(struct amdgpu_device *adev)
276 {
277 }
278 
279 #define KFD_IS_SVM_API_SUPPORTED(dev) false
280 
281 #endif /* IS_ENABLED(CONFIG_HSA_AMD_SVM) */
282 
283 #endif /* KFD_SVM_H_ */
284