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 <linux/hmm.h>
35 #include "amdgpu.h"
36 #include "kfd_priv.h"
37 
38 #define SVM_RANGE_VRAM_DOMAIN (1UL << 0)
39 #define SVM_ADEV_PGMAP_OWNER(adev)\
40 			((adev)->hive ? (void *)(adev)->hive : (void *)(adev))
41 
42 struct svm_range_bo {
43 	struct amdgpu_bo		*bo;
44 	struct kref			kref;
45 	struct list_head		range_list; /* all svm ranges shared this bo */
46 	spinlock_t			list_lock;
47 	struct amdgpu_amdkfd_fence	*eviction_fence;
48 	struct work_struct		eviction_work;
49 	struct svm_range_list		*svms;
50 	uint32_t			evicting;
51 };
52 
53 enum svm_work_list_ops {
54 	SVM_OP_NULL,
55 	SVM_OP_UNMAP_RANGE,
56 	SVM_OP_UPDATE_RANGE_NOTIFIER,
57 	SVM_OP_UPDATE_RANGE_NOTIFIER_AND_MAP,
58 	SVM_OP_ADD_RANGE,
59 	SVM_OP_ADD_RANGE_AND_MAP
60 };
61 
62 struct svm_work_list_item {
63 	enum svm_work_list_ops op;
64 	struct mm_struct *mm;
65 };
66 
67 /**
68  * struct svm_range - shared virtual memory range
69  *
70  * @svms:       list of svm ranges, structure defined in kfd_process
71  * @migrate_mutex: to serialize range migration, validation and mapping update
72  * @start:      range start address in pages
73  * @last:       range last address in pages
74  * @it_node:    node [start, last] stored in interval tree, start, last are page
75  *              aligned, page size is (last - start + 1)
76  * @list:       link list node, used to scan all ranges of svms
77  * @update_list:link list node used to add to update_list
78  * @remove_list:link list node used to add to remove list
79  * @insert_list:link list node used to add to insert list
80  * @mapping:    bo_va mapping structure to create and update GPU page table
81  * @npages:     number of pages
82  * @dma_addr:   dma mapping address on each GPU for system memory physical page
83  * @ttm_res:    vram ttm resource map
84  * @offset:     range start offset within mm_nodes
85  * @svm_bo:     struct to manage splited amdgpu_bo
86  * @svm_bo_list:link list node, to scan all ranges which share same svm_bo
87  * @lock:       protect prange start, last, child_list, svm_bo_list
88  * @saved_flags:save/restore current PF_MEMALLOC flags
89  * @flags:      flags defined as KFD_IOCTL_SVM_FLAG_*
90  * @perferred_loc: perferred location, 0 for CPU, or GPU id
91  * @perfetch_loc: last prefetch location, 0 for CPU, or GPU id
92  * @actual_loc: the actual location, 0 for CPU, or GPU id
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  *
103  * Data structure for virtual memory range shared by CPU and GPUs, it can be
104  * allocated from system memory ram or device vram, and migrate from ram to vram
105  * or from vram to ram.
106  */
107 struct svm_range {
108 	struct svm_range_list		*svms;
109 	struct mutex			migrate_mutex;
110 	unsigned long			start;
111 	unsigned long			last;
112 	struct interval_tree_node	it_node;
113 	struct list_head		list;
114 	struct list_head		update_list;
115 	struct list_head		remove_list;
116 	struct list_head		insert_list;
117 	uint64_t			npages;
118 	dma_addr_t			*dma_addr[MAX_GPU_INSTANCE];
119 	struct ttm_resource		*ttm_res;
120 	uint64_t			offset;
121 	struct svm_range_bo		*svm_bo;
122 	struct list_head		svm_bo_list;
123 	struct mutex                    lock;
124 	unsigned int                    saved_flags;
125 	uint32_t			flags;
126 	uint32_t			preferred_loc;
127 	uint32_t			prefetch_loc;
128 	uint32_t			actual_loc;
129 	uint8_t				granularity;
130 	atomic_t			invalid;
131 	uint64_t			validate_timestamp;
132 	struct mmu_interval_notifier	notifier;
133 	struct svm_work_list_item	work_item;
134 	struct list_head		deferred_list;
135 	struct list_head		child_list;
136 	DECLARE_BITMAP(bitmap_access, MAX_GPU_INSTANCE);
137 	DECLARE_BITMAP(bitmap_aip, MAX_GPU_INSTANCE);
138 	bool				validated_once;
139 };
140 
svm_range_lock(struct svm_range * prange)141 static inline void svm_range_lock(struct svm_range *prange)
142 {
143 	mutex_lock(&prange->lock);
144 	prange->saved_flags = memalloc_noreclaim_save();
145 
146 }
svm_range_unlock(struct svm_range * prange)147 static inline void svm_range_unlock(struct svm_range *prange)
148 {
149 	memalloc_noreclaim_restore(prange->saved_flags);
150 	mutex_unlock(&prange->lock);
151 }
152 
svm_range_bo_ref(struct svm_range_bo * svm_bo)153 static inline struct svm_range_bo *svm_range_bo_ref(struct svm_range_bo *svm_bo)
154 {
155 	if (svm_bo)
156 		kref_get(&svm_bo->kref);
157 
158 	return svm_bo;
159 }
160 
161 int svm_range_list_init(struct kfd_process *p);
162 void svm_range_list_fini(struct kfd_process *p);
163 int svm_ioctl(struct kfd_process *p, enum kfd_ioctl_svm_op op, uint64_t start,
164 	      uint64_t size, uint32_t nattrs,
165 	      struct kfd_ioctl_svm_attribute *attrs);
166 struct svm_range *svm_range_from_addr(struct svm_range_list *svms,
167 				      unsigned long addr,
168 				      struct svm_range **parent);
169 struct amdgpu_device *svm_range_get_adev_by_id(struct svm_range *prange,
170 					       uint32_t id);
171 int svm_range_vram_node_new(struct amdgpu_device *adev,
172 			    struct svm_range *prange, bool clear);
173 void svm_range_vram_node_free(struct svm_range *prange);
174 int svm_range_split_by_granularity(struct kfd_process *p, struct mm_struct *mm,
175 			       unsigned long addr, struct svm_range *parent,
176 			       struct svm_range *prange);
177 int svm_range_restore_pages(struct amdgpu_device *adev,
178 			    unsigned int pasid, uint64_t addr, bool write_fault);
179 int svm_range_schedule_evict_svm_bo(struct amdgpu_amdkfd_fence *fence);
180 void svm_range_add_list_work(struct svm_range_list *svms,
181 			     struct svm_range *prange, struct mm_struct *mm,
182 			     enum svm_work_list_ops op);
183 void schedule_deferred_list_work(struct svm_range_list *svms);
184 void svm_range_dma_unmap(struct device *dev, dma_addr_t *dma_addr,
185 			 unsigned long offset, unsigned long npages);
186 void svm_range_free_dma_mappings(struct svm_range *prange);
187 void svm_range_prefault(struct svm_range *prange, struct mm_struct *mm,
188 			void *owner);
189 struct kfd_process_device *
190 svm_range_get_pdd_by_adev(struct svm_range *prange, struct amdgpu_device *adev);
191 void svm_range_list_lock_and_flush_work(struct svm_range_list *svms, struct mm_struct *mm);
192 
193 /* SVM API and HMM page migration work together, device memory type
194  * is initialized to not 0 when page migration register device memory.
195  */
196 #define KFD_IS_SVM_API_SUPPORTED(dev) ((dev)->pgmap.type != 0)
197 
198 void svm_range_bo_unref(struct svm_range_bo *svm_bo);
199 #else
200 
201 struct kfd_process;
202 
svm_range_list_init(struct kfd_process * p)203 static inline int svm_range_list_init(struct kfd_process *p)
204 {
205 	return 0;
206 }
svm_range_list_fini(struct kfd_process * p)207 static inline void svm_range_list_fini(struct kfd_process *p)
208 {
209 	/* empty */
210 }
211 
svm_range_restore_pages(struct amdgpu_device * adev,unsigned int pasid,uint64_t addr,bool write_fault)212 static inline int svm_range_restore_pages(struct amdgpu_device *adev,
213 					  unsigned int pasid, uint64_t addr,
214 					  bool write_fault)
215 {
216 	return -EFAULT;
217 }
218 
svm_range_schedule_evict_svm_bo(struct amdgpu_amdkfd_fence * fence)219 static inline int svm_range_schedule_evict_svm_bo(
220 		struct amdgpu_amdkfd_fence *fence)
221 {
222 	WARN_ONCE(1, "SVM eviction fence triggered, but SVM is disabled");
223 	return -EINVAL;
224 }
225 
226 #define KFD_IS_SVM_API_SUPPORTED(dev) false
227 
228 #endif /* IS_ENABLED(CONFIG_HSA_AMD_SVM) */
229 
230 #endif /* KFD_SVM_H_ */
231