1 /*
2  * Copyright 2014 Advanced Micro Devices, Inc.
3  *
4  * Permission is hereby granted, free of charge, to any person obtaining a
5  * copy of this software and associated documentation files (the "Software"),
6  * to deal in the Software without restriction, including without limitation
7  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
8  * and/or sell copies of the Software, and to permit persons to whom the
9  * Software is furnished to do so, subject to the following conditions:
10  *
11  * The above copyright notice and this permission notice shall be included in
12  * all copies or substantial portions of the Software.
13  *
14  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
15  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
16  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
17  * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR
18  * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
19  * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
20  * OTHER DEALINGS IN THE SOFTWARE.
21  */
22 
23 /* amdgpu_amdkfd.h defines the private interface between amdgpu and amdkfd. */
24 
25 #ifndef AMDGPU_AMDKFD_H_INCLUDED
26 #define AMDGPU_AMDKFD_H_INCLUDED
27 
28 #include <linux/types.h>
29 #include <linux/mm.h>
30 #include <linux/kthread.h>
31 #include <linux/workqueue.h>
32 #include <kgd_kfd_interface.h>
33 #include <drm/ttm/ttm_execbuf_util.h>
34 #include "amdgpu_sync.h"
35 #include "amdgpu_vm.h"
36 
37 extern uint64_t amdgpu_amdkfd_total_mem_size;
38 
39 struct amdgpu_device;
40 
41 struct kfd_bo_va_list {
42 	struct list_head bo_list;
43 	struct amdgpu_bo_va *bo_va;
44 	void *kgd_dev;
45 	bool is_mapped;
46 	uint64_t va;
47 	uint64_t pte_flags;
48 };
49 
50 struct kgd_mem {
51 	struct mutex lock;
52 	struct amdgpu_bo *bo;
53 	struct list_head bo_va_list;
54 	/* protected by amdkfd_process_info.lock */
55 	struct ttm_validate_buffer validate_list;
56 	struct ttm_validate_buffer resv_list;
57 	uint32_t domain;
58 	unsigned int mapped_to_gpu_memory;
59 	uint64_t va;
60 
61 	uint32_t alloc_flags;
62 
63 	atomic_t invalid;
64 	struct amdkfd_process_info *process_info;
65 
66 	struct amdgpu_sync sync;
67 
68 	bool aql_queue;
69 	bool is_imported;
70 };
71 
72 /* KFD Memory Eviction */
73 struct amdgpu_amdkfd_fence {
74 	struct dma_fence base;
75 	struct mm_struct *mm;
76 	spinlock_t lock;
77 	char timeline_name[TASK_COMM_LEN];
78 };
79 
80 struct amdgpu_kfd_dev {
81 	struct kfd_dev *dev;
82 	uint64_t vram_used;
83 	bool init_complete;
84 };
85 
86 enum kgd_engine_type {
87 	KGD_ENGINE_PFP = 1,
88 	KGD_ENGINE_ME,
89 	KGD_ENGINE_CE,
90 	KGD_ENGINE_MEC1,
91 	KGD_ENGINE_MEC2,
92 	KGD_ENGINE_RLC,
93 	KGD_ENGINE_SDMA1,
94 	KGD_ENGINE_SDMA2,
95 	KGD_ENGINE_MAX
96 };
97 
98 
99 struct amdkfd_process_info {
100 	/* List head of all VMs that belong to a KFD process */
101 	struct list_head vm_list_head;
102 	/* List head for all KFD BOs that belong to a KFD process. */
103 	struct list_head kfd_bo_list;
104 	/* List of userptr BOs that are valid or invalid */
105 	struct list_head userptr_valid_list;
106 	struct list_head userptr_inval_list;
107 	/* Lock to protect kfd_bo_list */
108 	struct mutex lock;
109 
110 	/* Number of VMs */
111 	unsigned int n_vms;
112 	/* Eviction Fence */
113 	struct amdgpu_amdkfd_fence *eviction_fence;
114 
115 	/* MMU-notifier related fields */
116 	atomic_t evicted_bos;
117 	struct delayed_work restore_userptr_work;
118 	struct pid *pid;
119 };
120 
121 int amdgpu_amdkfd_init(void);
122 void amdgpu_amdkfd_fini(void);
123 
124 void amdgpu_amdkfd_suspend(struct amdgpu_device *adev, bool run_pm);
125 int amdgpu_amdkfd_resume(struct amdgpu_device *adev, bool run_pm);
126 void amdgpu_amdkfd_interrupt(struct amdgpu_device *adev,
127 			const void *ih_ring_entry);
128 void amdgpu_amdkfd_device_probe(struct amdgpu_device *adev);
129 void amdgpu_amdkfd_device_init(struct amdgpu_device *adev);
130 void amdgpu_amdkfd_device_fini(struct amdgpu_device *adev);
131 int amdgpu_amdkfd_submit_ib(struct kgd_dev *kgd, enum kgd_engine_type engine,
132 				uint32_t vmid, uint64_t gpu_addr,
133 				uint32_t *ib_cmd, uint32_t ib_len);
134 void amdgpu_amdkfd_set_compute_idle(struct kgd_dev *kgd, bool idle);
135 bool amdgpu_amdkfd_have_atomics_support(struct kgd_dev *kgd);
136 int amdgpu_amdkfd_flush_gpu_tlb_vmid(struct kgd_dev *kgd, uint16_t vmid);
137 int amdgpu_amdkfd_flush_gpu_tlb_pasid(struct kgd_dev *kgd, uint16_t pasid);
138 
139 bool amdgpu_amdkfd_is_kfd_vmid(struct amdgpu_device *adev, u32 vmid);
140 
141 int amdgpu_amdkfd_pre_reset(struct amdgpu_device *adev);
142 
143 int amdgpu_amdkfd_post_reset(struct amdgpu_device *adev);
144 
145 void amdgpu_amdkfd_gpu_reset(struct kgd_dev *kgd);
146 
147 int amdgpu_queue_mask_bit_to_set_resource_bit(struct amdgpu_device *adev,
148 					int queue_bit);
149 
150 struct amdgpu_amdkfd_fence *amdgpu_amdkfd_fence_create(u64 context,
151 								struct mm_struct *mm);
152 #if IS_ENABLED(CONFIG_HSA_AMD)
153 bool amdkfd_fence_check_mm(struct dma_fence *f, struct mm_struct *mm);
154 struct amdgpu_amdkfd_fence *to_amdgpu_amdkfd_fence(struct dma_fence *f);
155 int amdgpu_amdkfd_remove_fence_on_pt_pd_bos(struct amdgpu_bo *bo);
156 int amdgpu_amdkfd_evict_userptr(struct kgd_mem *mem, struct mm_struct *mm);
157 #else
158 static inline
159 bool amdkfd_fence_check_mm(struct dma_fence *f, struct mm_struct *mm)
160 {
161 	return false;
162 }
163 
164 static inline
165 struct amdgpu_amdkfd_fence *to_amdgpu_amdkfd_fence(struct dma_fence *f)
166 {
167 	return NULL;
168 }
169 
170 static inline
171 int amdgpu_amdkfd_remove_fence_on_pt_pd_bos(struct amdgpu_bo *bo)
172 {
173 	return 0;
174 }
175 
176 static inline
177 int amdgpu_amdkfd_evict_userptr(struct kgd_mem *mem, struct mm_struct *mm)
178 {
179 	return 0;
180 }
181 #endif
182 /* Shared API */
183 int amdgpu_amdkfd_alloc_gtt_mem(struct kgd_dev *kgd, size_t size,
184 				void **mem_obj, uint64_t *gpu_addr,
185 				void **cpu_ptr, bool mqd_gfx9);
186 void amdgpu_amdkfd_free_gtt_mem(struct kgd_dev *kgd, void *mem_obj);
187 int amdgpu_amdkfd_alloc_gws(struct kgd_dev *kgd, size_t size, void **mem_obj);
188 void amdgpu_amdkfd_free_gws(struct kgd_dev *kgd, void *mem_obj);
189 int amdgpu_amdkfd_add_gws_to_process(void *info, void *gws, struct kgd_mem **mem);
190 int amdgpu_amdkfd_remove_gws_from_process(void *info, void *mem);
191 uint32_t amdgpu_amdkfd_get_fw_version(struct kgd_dev *kgd,
192 				      enum kgd_engine_type type);
193 void amdgpu_amdkfd_get_local_mem_info(struct kgd_dev *kgd,
194 				      struct kfd_local_mem_info *mem_info);
195 uint64_t amdgpu_amdkfd_get_gpu_clock_counter(struct kgd_dev *kgd);
196 
197 uint32_t amdgpu_amdkfd_get_max_engine_clock_in_mhz(struct kgd_dev *kgd);
198 void amdgpu_amdkfd_get_cu_info(struct kgd_dev *kgd, struct kfd_cu_info *cu_info);
199 int amdgpu_amdkfd_get_dmabuf_info(struct kgd_dev *kgd, int dma_buf_fd,
200 				  struct kgd_dev **dmabuf_kgd,
201 				  uint64_t *bo_size, void *metadata_buffer,
202 				  size_t buffer_size, uint32_t *metadata_size,
203 				  uint32_t *flags);
204 uint64_t amdgpu_amdkfd_get_vram_usage(struct kgd_dev *kgd);
205 uint64_t amdgpu_amdkfd_get_hive_id(struct kgd_dev *kgd);
206 uint64_t amdgpu_amdkfd_get_unique_id(struct kgd_dev *kgd);
207 uint64_t amdgpu_amdkfd_get_mmio_remap_phys_addr(struct kgd_dev *kgd);
208 uint32_t amdgpu_amdkfd_get_num_gws(struct kgd_dev *kgd);
209 uint32_t amdgpu_amdkfd_get_asic_rev_id(struct kgd_dev *kgd);
210 int amdgpu_amdkfd_get_noretry(struct kgd_dev *kgd);
211 uint8_t amdgpu_amdkfd_get_xgmi_hops_count(struct kgd_dev *dst, struct kgd_dev *src);
212 
213 /* Read user wptr from a specified user address space with page fault
214  * disabled. The memory must be pinned and mapped to the hardware when
215  * this is called in hqd_load functions, so it should never fault in
216  * the first place. This resolves a circular lock dependency involving
217  * four locks, including the DQM lock and mmap_lock.
218  */
219 #define read_user_wptr(mmptr, wptr, dst)				\
220 	({								\
221 		bool valid = false;					\
222 		if ((mmptr) && (wptr)) {				\
223 			pagefault_disable();				\
224 			if ((mmptr) == current->mm) {			\
225 				valid = !get_user((dst), (wptr));	\
226 			} else if (current->flags & PF_KTHREAD) {	\
227 				kthread_use_mm(mmptr);			\
228 				valid = !get_user((dst), (wptr));	\
229 				kthread_unuse_mm(mmptr);		\
230 			}						\
231 			pagefault_enable();				\
232 		}							\
233 		valid;							\
234 	})
235 
236 /* GPUVM API */
237 int amdgpu_amdkfd_gpuvm_acquire_process_vm(struct kgd_dev *kgd,
238 					struct file *filp, u32 pasid,
239 					void **vm, void **process_info,
240 					struct dma_fence **ef);
241 void amdgpu_amdkfd_gpuvm_release_process_vm(struct kgd_dev *kgd, void *vm);
242 uint64_t amdgpu_amdkfd_gpuvm_get_process_page_dir(void *vm);
243 int amdgpu_amdkfd_gpuvm_alloc_memory_of_gpu(
244 		struct kgd_dev *kgd, uint64_t va, uint64_t size,
245 		void *vm, struct kgd_mem **mem,
246 		uint64_t *offset, uint32_t flags);
247 int amdgpu_amdkfd_gpuvm_free_memory_of_gpu(
248 		struct kgd_dev *kgd, struct kgd_mem *mem, uint64_t *size);
249 int amdgpu_amdkfd_gpuvm_map_memory_to_gpu(
250 		struct kgd_dev *kgd, struct kgd_mem *mem, void *vm);
251 int amdgpu_amdkfd_gpuvm_unmap_memory_from_gpu(
252 		struct kgd_dev *kgd, struct kgd_mem *mem, void *vm);
253 int amdgpu_amdkfd_gpuvm_sync_memory(
254 		struct kgd_dev *kgd, struct kgd_mem *mem, bool intr);
255 int amdgpu_amdkfd_gpuvm_map_gtt_bo_to_kernel(struct kgd_dev *kgd,
256 		struct kgd_mem *mem, void **kptr, uint64_t *size);
257 int amdgpu_amdkfd_gpuvm_restore_process_bos(void *process_info,
258 					    struct dma_fence **ef);
259 int amdgpu_amdkfd_gpuvm_get_vm_fault_info(struct kgd_dev *kgd,
260 					      struct kfd_vm_fault_info *info);
261 int amdgpu_amdkfd_gpuvm_import_dmabuf(struct kgd_dev *kgd,
262 				      struct dma_buf *dmabuf,
263 				      uint64_t va, void *vm,
264 				      struct kgd_mem **mem, uint64_t *size,
265 				      uint64_t *mmap_offset);
266 int amdgpu_amdkfd_get_tile_config(struct kgd_dev *kgd,
267 				struct tile_config *config);
268 #if IS_ENABLED(CONFIG_HSA_AMD)
269 void amdgpu_amdkfd_gpuvm_init_mem_limits(void);
270 void amdgpu_amdkfd_gpuvm_destroy_cb(struct amdgpu_device *adev,
271 				struct amdgpu_vm *vm);
272 void amdgpu_amdkfd_unreserve_memory_limit(struct amdgpu_bo *bo);
273 #else
274 static inline
275 void amdgpu_amdkfd_gpuvm_init_mem_limits(void)
276 {
277 }
278 
279 static inline
280 void amdgpu_amdkfd_gpuvm_destroy_cb(struct amdgpu_device *adev,
281 					struct amdgpu_vm *vm)
282 {
283 }
284 
285 static inline
286 void amdgpu_amdkfd_unreserve_memory_limit(struct amdgpu_bo *bo)
287 {
288 }
289 #endif
290 /* KGD2KFD callbacks */
291 int kgd2kfd_quiesce_mm(struct mm_struct *mm);
292 int kgd2kfd_resume_mm(struct mm_struct *mm);
293 int kgd2kfd_schedule_evict_and_restore_process(struct mm_struct *mm,
294 						struct dma_fence *fence);
295 #if IS_ENABLED(CONFIG_HSA_AMD)
296 int kgd2kfd_init(void);
297 void kgd2kfd_exit(void);
298 struct kfd_dev *kgd2kfd_probe(struct kgd_dev *kgd, struct pci_dev *pdev,
299 			      unsigned int asic_type, bool vf);
300 bool kgd2kfd_device_init(struct kfd_dev *kfd,
301 			 struct drm_device *ddev,
302 			 const struct kgd2kfd_shared_resources *gpu_resources);
303 void kgd2kfd_device_exit(struct kfd_dev *kfd);
304 void kgd2kfd_suspend(struct kfd_dev *kfd, bool run_pm);
305 int kgd2kfd_resume(struct kfd_dev *kfd, bool run_pm);
306 int kgd2kfd_pre_reset(struct kfd_dev *kfd);
307 int kgd2kfd_post_reset(struct kfd_dev *kfd);
308 void kgd2kfd_interrupt(struct kfd_dev *kfd, const void *ih_ring_entry);
309 void kgd2kfd_set_sram_ecc_flag(struct kfd_dev *kfd);
310 void kgd2kfd_smi_event_throttle(struct kfd_dev *kfd, uint32_t throttle_bitmask);
311 #else
312 static inline int kgd2kfd_init(void)
313 {
314 	return -ENOENT;
315 }
316 
317 static inline void kgd2kfd_exit(void)
318 {
319 }
320 
321 static inline
322 struct kfd_dev *kgd2kfd_probe(struct kgd_dev *kgd, struct pci_dev *pdev,
323 					unsigned int asic_type, bool vf)
324 {
325 	return NULL;
326 }
327 
328 static inline
329 bool kgd2kfd_device_init(struct kfd_dev *kfd, struct drm_device *ddev,
330 				const struct kgd2kfd_shared_resources *gpu_resources)
331 {
332 	return false;
333 }
334 
335 static inline void kgd2kfd_device_exit(struct kfd_dev *kfd)
336 {
337 }
338 
339 static inline void kgd2kfd_suspend(struct kfd_dev *kfd, bool run_pm)
340 {
341 }
342 
343 static inline int kgd2kfd_resume(struct kfd_dev *kfd, bool run_pm)
344 {
345 	return 0;
346 }
347 
348 static inline int kgd2kfd_pre_reset(struct kfd_dev *kfd)
349 {
350 	return 0;
351 }
352 
353 static inline int kgd2kfd_post_reset(struct kfd_dev *kfd)
354 {
355 	return 0;
356 }
357 
358 static inline
359 void kgd2kfd_interrupt(struct kfd_dev *kfd, const void *ih_ring_entry)
360 {
361 }
362 
363 static inline
364 void kgd2kfd_set_sram_ecc_flag(struct kfd_dev *kfd)
365 {
366 }
367 
368 static inline
369 void kgd2kfd_smi_event_throttle(struct kfd_dev *kfd, uint32_t throttle_bitmask)
370 {
371 }
372 #endif
373 #endif /* AMDGPU_AMDKFD_H_INCLUDED */
374