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 enum TLB_FLUSH_TYPE {
40 	TLB_FLUSH_LEGACY = 0,
41 	TLB_FLUSH_LIGHTWEIGHT,
42 	TLB_FLUSH_HEAVYWEIGHT
43 };
44 
45 struct amdgpu_device;
46 
47 enum kfd_mem_attachment_type {
48 	KFD_MEM_ATT_SHARED,	/* Share kgd_mem->bo or another attachment's */
49 	KFD_MEM_ATT_USERPTR,	/* SG bo to DMA map pages from a userptr bo */
50 	KFD_MEM_ATT_DMABUF,	/* DMAbuf to DMA map TTM BOs */
51 };
52 
53 struct kfd_mem_attachment {
54 	struct list_head list;
55 	enum kfd_mem_attachment_type type;
56 	bool is_mapped;
57 	struct amdgpu_bo_va *bo_va;
58 	struct amdgpu_device *adev;
59 	uint64_t va;
60 	uint64_t pte_flags;
61 };
62 
63 struct kgd_mem {
64 	struct mutex lock;
65 	struct amdgpu_bo *bo;
66 	struct dma_buf *dmabuf;
67 	struct list_head attachments;
68 	/* protected by amdkfd_process_info.lock */
69 	struct ttm_validate_buffer validate_list;
70 	struct ttm_validate_buffer resv_list;
71 	uint32_t domain;
72 	unsigned int mapped_to_gpu_memory;
73 	uint64_t va;
74 
75 	uint32_t alloc_flags;
76 
77 	atomic_t invalid;
78 	struct amdkfd_process_info *process_info;
79 
80 	struct amdgpu_sync sync;
81 
82 	bool aql_queue;
83 	bool is_imported;
84 };
85 
86 /* KFD Memory Eviction */
87 struct amdgpu_amdkfd_fence {
88 	struct dma_fence base;
89 	struct mm_struct *mm;
90 	spinlock_t lock;
91 	char timeline_name[TASK_COMM_LEN];
92 	struct svm_range_bo *svm_bo;
93 };
94 
95 struct amdgpu_kfd_dev {
96 	struct kfd_dev *dev;
97 	uint64_t vram_used;
98 	bool init_complete;
99 };
100 
101 enum kgd_engine_type {
102 	KGD_ENGINE_PFP = 1,
103 	KGD_ENGINE_ME,
104 	KGD_ENGINE_CE,
105 	KGD_ENGINE_MEC1,
106 	KGD_ENGINE_MEC2,
107 	KGD_ENGINE_RLC,
108 	KGD_ENGINE_SDMA1,
109 	KGD_ENGINE_SDMA2,
110 	KGD_ENGINE_MAX
111 };
112 
113 
114 struct amdkfd_process_info {
115 	/* List head of all VMs that belong to a KFD process */
116 	struct list_head vm_list_head;
117 	/* List head for all KFD BOs that belong to a KFD process. */
118 	struct list_head kfd_bo_list;
119 	/* List of userptr BOs that are valid or invalid */
120 	struct list_head userptr_valid_list;
121 	struct list_head userptr_inval_list;
122 	/* Lock to protect kfd_bo_list */
123 	struct mutex lock;
124 
125 	/* Number of VMs */
126 	unsigned int n_vms;
127 	/* Eviction Fence */
128 	struct amdgpu_amdkfd_fence *eviction_fence;
129 
130 	/* MMU-notifier related fields */
131 	atomic_t evicted_bos;
132 	struct delayed_work restore_userptr_work;
133 	struct pid *pid;
134 };
135 
136 int amdgpu_amdkfd_init(void);
137 void amdgpu_amdkfd_fini(void);
138 
139 void amdgpu_amdkfd_suspend(struct amdgpu_device *adev, bool run_pm);
140 int amdgpu_amdkfd_resume(struct amdgpu_device *adev, bool run_pm);
141 void amdgpu_amdkfd_interrupt(struct amdgpu_device *adev,
142 			const void *ih_ring_entry);
143 void amdgpu_amdkfd_device_probe(struct amdgpu_device *adev);
144 void amdgpu_amdkfd_device_init(struct amdgpu_device *adev);
145 void amdgpu_amdkfd_device_fini_sw(struct amdgpu_device *adev);
146 int amdgpu_amdkfd_submit_ib(struct kgd_dev *kgd, enum kgd_engine_type engine,
147 				uint32_t vmid, uint64_t gpu_addr,
148 				uint32_t *ib_cmd, uint32_t ib_len);
149 void amdgpu_amdkfd_set_compute_idle(struct kgd_dev *kgd, bool idle);
150 bool amdgpu_amdkfd_have_atomics_support(struct kgd_dev *kgd);
151 int amdgpu_amdkfd_flush_gpu_tlb_vmid(struct kgd_dev *kgd, uint16_t vmid);
152 int amdgpu_amdkfd_flush_gpu_tlb_pasid(struct kgd_dev *kgd, uint16_t pasid,
153 				      enum TLB_FLUSH_TYPE flush_type);
154 
155 bool amdgpu_amdkfd_is_kfd_vmid(struct amdgpu_device *adev, u32 vmid);
156 
157 int amdgpu_amdkfd_pre_reset(struct amdgpu_device *adev);
158 
159 int amdgpu_amdkfd_post_reset(struct amdgpu_device *adev);
160 
161 void amdgpu_amdkfd_gpu_reset(struct kgd_dev *kgd);
162 
163 int amdgpu_queue_mask_bit_to_set_resource_bit(struct amdgpu_device *adev,
164 					int queue_bit);
165 
166 struct amdgpu_amdkfd_fence *amdgpu_amdkfd_fence_create(u64 context,
167 				struct mm_struct *mm,
168 				struct svm_range_bo *svm_bo);
169 #if IS_ENABLED(CONFIG_HSA_AMD)
170 bool amdkfd_fence_check_mm(struct dma_fence *f, struct mm_struct *mm);
171 struct amdgpu_amdkfd_fence *to_amdgpu_amdkfd_fence(struct dma_fence *f);
172 int amdgpu_amdkfd_remove_fence_on_pt_pd_bos(struct amdgpu_bo *bo);
173 int amdgpu_amdkfd_evict_userptr(struct kgd_mem *mem, struct mm_struct *mm);
174 #else
175 static inline
176 bool amdkfd_fence_check_mm(struct dma_fence *f, struct mm_struct *mm)
177 {
178 	return false;
179 }
180 
181 static inline
182 struct amdgpu_amdkfd_fence *to_amdgpu_amdkfd_fence(struct dma_fence *f)
183 {
184 	return NULL;
185 }
186 
187 static inline
188 int amdgpu_amdkfd_remove_fence_on_pt_pd_bos(struct amdgpu_bo *bo)
189 {
190 	return 0;
191 }
192 
193 static inline
194 int amdgpu_amdkfd_evict_userptr(struct kgd_mem *mem, struct mm_struct *mm)
195 {
196 	return 0;
197 }
198 #endif
199 /* Shared API */
200 int amdgpu_amdkfd_alloc_gtt_mem(struct kgd_dev *kgd, size_t size,
201 				void **mem_obj, uint64_t *gpu_addr,
202 				void **cpu_ptr, bool mqd_gfx9);
203 void amdgpu_amdkfd_free_gtt_mem(struct kgd_dev *kgd, void *mem_obj);
204 int amdgpu_amdkfd_alloc_gws(struct kgd_dev *kgd, size_t size, void **mem_obj);
205 void amdgpu_amdkfd_free_gws(struct kgd_dev *kgd, void *mem_obj);
206 int amdgpu_amdkfd_add_gws_to_process(void *info, void *gws, struct kgd_mem **mem);
207 int amdgpu_amdkfd_remove_gws_from_process(void *info, void *mem);
208 uint32_t amdgpu_amdkfd_get_fw_version(struct kgd_dev *kgd,
209 				      enum kgd_engine_type type);
210 void amdgpu_amdkfd_get_local_mem_info(struct kgd_dev *kgd,
211 				      struct kfd_local_mem_info *mem_info);
212 uint64_t amdgpu_amdkfd_get_gpu_clock_counter(struct kgd_dev *kgd);
213 
214 uint32_t amdgpu_amdkfd_get_max_engine_clock_in_mhz(struct kgd_dev *kgd);
215 void amdgpu_amdkfd_get_cu_info(struct kgd_dev *kgd, struct kfd_cu_info *cu_info);
216 int amdgpu_amdkfd_get_dmabuf_info(struct kgd_dev *kgd, int dma_buf_fd,
217 				  struct kgd_dev **dmabuf_kgd,
218 				  uint64_t *bo_size, void *metadata_buffer,
219 				  size_t buffer_size, uint32_t *metadata_size,
220 				  uint32_t *flags);
221 uint64_t amdgpu_amdkfd_get_vram_usage(struct kgd_dev *kgd);
222 uint64_t amdgpu_amdkfd_get_hive_id(struct kgd_dev *kgd);
223 uint64_t amdgpu_amdkfd_get_unique_id(struct kgd_dev *kgd);
224 uint64_t amdgpu_amdkfd_get_mmio_remap_phys_addr(struct kgd_dev *kgd);
225 uint32_t amdgpu_amdkfd_get_num_gws(struct kgd_dev *kgd);
226 uint32_t amdgpu_amdkfd_get_asic_rev_id(struct kgd_dev *kgd);
227 int amdgpu_amdkfd_get_noretry(struct kgd_dev *kgd);
228 uint8_t amdgpu_amdkfd_get_xgmi_hops_count(struct kgd_dev *dst, struct kgd_dev *src);
229 
230 /* Read user wptr from a specified user address space with page fault
231  * disabled. The memory must be pinned and mapped to the hardware when
232  * this is called in hqd_load functions, so it should never fault in
233  * the first place. This resolves a circular lock dependency involving
234  * four locks, including the DQM lock and mmap_lock.
235  */
236 #define read_user_wptr(mmptr, wptr, dst)				\
237 	({								\
238 		bool valid = false;					\
239 		if ((mmptr) && (wptr)) {				\
240 			pagefault_disable();				\
241 			if ((mmptr) == current->mm) {			\
242 				valid = !get_user((dst), (wptr));	\
243 			} else if (current->flags & PF_KTHREAD) {	\
244 				kthread_use_mm(mmptr);			\
245 				valid = !get_user((dst), (wptr));	\
246 				kthread_unuse_mm(mmptr);		\
247 			}						\
248 			pagefault_enable();				\
249 		}							\
250 		valid;							\
251 	})
252 
253 /* GPUVM API */
254 #define drm_priv_to_vm(drm_priv)					\
255 	(&((struct amdgpu_fpriv *)					\
256 		((struct drm_file *)(drm_priv))->driver_priv)->vm)
257 
258 int amdgpu_amdkfd_gpuvm_acquire_process_vm(struct kgd_dev *kgd,
259 					struct file *filp, u32 pasid,
260 					void **process_info,
261 					struct dma_fence **ef);
262 void amdgpu_amdkfd_gpuvm_release_process_vm(struct kgd_dev *kgd, void *drm_priv);
263 uint64_t amdgpu_amdkfd_gpuvm_get_process_page_dir(void *drm_priv);
264 int amdgpu_amdkfd_gpuvm_alloc_memory_of_gpu(
265 		struct kgd_dev *kgd, uint64_t va, uint64_t size,
266 		void *drm_priv, struct kgd_mem **mem,
267 		uint64_t *offset, uint32_t flags);
268 int amdgpu_amdkfd_gpuvm_free_memory_of_gpu(
269 		struct kgd_dev *kgd, struct kgd_mem *mem, void *drm_priv,
270 		uint64_t *size);
271 int amdgpu_amdkfd_gpuvm_map_memory_to_gpu(
272 		struct kgd_dev *kgd, struct kgd_mem *mem, void *drm_priv);
273 int amdgpu_amdkfd_gpuvm_unmap_memory_from_gpu(
274 		struct kgd_dev *kgd, struct kgd_mem *mem, void *drm_priv);
275 int amdgpu_amdkfd_gpuvm_sync_memory(
276 		struct kgd_dev *kgd, struct kgd_mem *mem, bool intr);
277 int amdgpu_amdkfd_gpuvm_map_gtt_bo_to_kernel(struct kgd_dev *kgd,
278 		struct kgd_mem *mem, void **kptr, uint64_t *size);
279 int amdgpu_amdkfd_gpuvm_restore_process_bos(void *process_info,
280 					    struct dma_fence **ef);
281 int amdgpu_amdkfd_gpuvm_get_vm_fault_info(struct kgd_dev *kgd,
282 					      struct kfd_vm_fault_info *info);
283 int amdgpu_amdkfd_gpuvm_import_dmabuf(struct kgd_dev *kgd,
284 				      struct dma_buf *dmabuf,
285 				      uint64_t va, void *drm_priv,
286 				      struct kgd_mem **mem, uint64_t *size,
287 				      uint64_t *mmap_offset);
288 int amdgpu_amdkfd_get_tile_config(struct kgd_dev *kgd,
289 				struct tile_config *config);
290 #if IS_ENABLED(CONFIG_HSA_AMD)
291 void amdgpu_amdkfd_gpuvm_init_mem_limits(void);
292 void amdgpu_amdkfd_gpuvm_destroy_cb(struct amdgpu_device *adev,
293 				struct amdgpu_vm *vm);
294 void amdgpu_amdkfd_unreserve_memory_limit(struct amdgpu_bo *bo);
295 void amdgpu_amdkfd_reserve_system_mem(uint64_t size);
296 #else
297 static inline
298 void amdgpu_amdkfd_gpuvm_init_mem_limits(void)
299 {
300 }
301 
302 static inline
303 void amdgpu_amdkfd_gpuvm_destroy_cb(struct amdgpu_device *adev,
304 					struct amdgpu_vm *vm)
305 {
306 }
307 
308 static inline
309 void amdgpu_amdkfd_unreserve_memory_limit(struct amdgpu_bo *bo)
310 {
311 }
312 #endif
313 /* KGD2KFD callbacks */
314 int kgd2kfd_quiesce_mm(struct mm_struct *mm);
315 int kgd2kfd_resume_mm(struct mm_struct *mm);
316 int kgd2kfd_schedule_evict_and_restore_process(struct mm_struct *mm,
317 						struct dma_fence *fence);
318 #if IS_ENABLED(CONFIG_HSA_AMD)
319 int kgd2kfd_init(void);
320 void kgd2kfd_exit(void);
321 struct kfd_dev *kgd2kfd_probe(struct kgd_dev *kgd, struct pci_dev *pdev,
322 			      unsigned int asic_type, bool vf);
323 bool kgd2kfd_device_init(struct kfd_dev *kfd,
324 			 struct drm_device *ddev,
325 			 const struct kgd2kfd_shared_resources *gpu_resources);
326 void kgd2kfd_device_exit(struct kfd_dev *kfd);
327 void kgd2kfd_suspend(struct kfd_dev *kfd, bool run_pm);
328 int kgd2kfd_resume(struct kfd_dev *kfd, bool run_pm);
329 int kgd2kfd_pre_reset(struct kfd_dev *kfd);
330 int kgd2kfd_post_reset(struct kfd_dev *kfd);
331 void kgd2kfd_interrupt(struct kfd_dev *kfd, const void *ih_ring_entry);
332 void kgd2kfd_set_sram_ecc_flag(struct kfd_dev *kfd);
333 void kgd2kfd_smi_event_throttle(struct kfd_dev *kfd, uint32_t throttle_bitmask);
334 #else
335 static inline int kgd2kfd_init(void)
336 {
337 	return -ENOENT;
338 }
339 
340 static inline void kgd2kfd_exit(void)
341 {
342 }
343 
344 static inline
345 struct kfd_dev *kgd2kfd_probe(struct kgd_dev *kgd, struct pci_dev *pdev,
346 					unsigned int asic_type, bool vf)
347 {
348 	return NULL;
349 }
350 
351 static inline
352 bool kgd2kfd_device_init(struct kfd_dev *kfd, struct drm_device *ddev,
353 				const struct kgd2kfd_shared_resources *gpu_resources)
354 {
355 	return false;
356 }
357 
358 static inline void kgd2kfd_device_exit(struct kfd_dev *kfd)
359 {
360 }
361 
362 static inline void kgd2kfd_suspend(struct kfd_dev *kfd, bool run_pm)
363 {
364 }
365 
366 static inline int kgd2kfd_resume(struct kfd_dev *kfd, bool run_pm)
367 {
368 	return 0;
369 }
370 
371 static inline int kgd2kfd_pre_reset(struct kfd_dev *kfd)
372 {
373 	return 0;
374 }
375 
376 static inline int kgd2kfd_post_reset(struct kfd_dev *kfd)
377 {
378 	return 0;
379 }
380 
381 static inline
382 void kgd2kfd_interrupt(struct kfd_dev *kfd, const void *ih_ring_entry)
383 {
384 }
385 
386 static inline
387 void kgd2kfd_set_sram_ecc_flag(struct kfd_dev *kfd)
388 {
389 }
390 
391 static inline
392 void kgd2kfd_smi_event_throttle(struct kfd_dev *kfd, uint32_t throttle_bitmask)
393 {
394 }
395 #endif
396 #endif /* AMDGPU_AMDKFD_H_INCLUDED */
397