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/list.h>
29 #include <linux/types.h>
30 #include <linux/mm.h>
31 #include <linux/kthread.h>
32 #include <linux/workqueue.h>
33 #include <linux/mmu_notifier.h>
34 #include <linux/memremap.h>
35 #include <kgd_kfd_interface.h>
36 #include "amdgpu_sync.h"
37 #include "amdgpu_vm.h"
38 #include "amdgpu_xcp.h"
39 
40 extern uint64_t amdgpu_amdkfd_total_mem_size;
41 
42 enum TLB_FLUSH_TYPE {
43 	TLB_FLUSH_LEGACY = 0,
44 	TLB_FLUSH_LIGHTWEIGHT,
45 	TLB_FLUSH_HEAVYWEIGHT
46 };
47 
48 struct amdgpu_device;
49 
50 enum kfd_mem_attachment_type {
51 	KFD_MEM_ATT_SHARED,	/* Share kgd_mem->bo or another attachment's */
52 	KFD_MEM_ATT_USERPTR,	/* SG bo to DMA map pages from a userptr bo */
53 	KFD_MEM_ATT_DMABUF,	/* DMAbuf to DMA map TTM BOs */
54 	KFD_MEM_ATT_SG		/* Tag to DMA map SG BOs */
55 };
56 
57 struct kfd_mem_attachment {
58 	struct list_head list;
59 	enum kfd_mem_attachment_type type;
60 	bool is_mapped;
61 	struct amdgpu_bo_va *bo_va;
62 	struct amdgpu_device *adev;
63 	uint64_t va;
64 	uint64_t pte_flags;
65 };
66 
67 struct kgd_mem {
68 	struct mutex lock;
69 	struct amdgpu_bo *bo;
70 	struct dma_buf *dmabuf;
71 	struct hmm_range *range;
72 	struct list_head attachments;
73 	/* protected by amdkfd_process_info.lock */
74 	struct list_head validate_list;
75 	uint32_t domain;
76 	unsigned int mapped_to_gpu_memory;
77 	uint64_t va;
78 
79 	uint32_t alloc_flags;
80 
81 	uint32_t invalid;
82 	struct amdkfd_process_info *process_info;
83 
84 	struct amdgpu_sync sync;
85 
86 	bool aql_queue;
87 	bool is_imported;
88 };
89 
90 /* KFD Memory Eviction */
91 struct amdgpu_amdkfd_fence {
92 	struct dma_fence base;
93 	struct mm_struct *mm;
94 	spinlock_t lock;
95 	char timeline_name[TASK_COMM_LEN];
96 	struct svm_range_bo *svm_bo;
97 };
98 
99 struct amdgpu_kfd_dev {
100 	struct kfd_dev *dev;
101 	int64_t vram_used[MAX_XCP];
102 	uint64_t vram_used_aligned[MAX_XCP];
103 	bool init_complete;
104 	struct work_struct reset_work;
105 
106 	/* HMM page migration MEMORY_DEVICE_PRIVATE mapping */
107 	struct dev_pagemap pgmap;
108 };
109 
110 enum kgd_engine_type {
111 	KGD_ENGINE_PFP = 1,
112 	KGD_ENGINE_ME,
113 	KGD_ENGINE_CE,
114 	KGD_ENGINE_MEC1,
115 	KGD_ENGINE_MEC2,
116 	KGD_ENGINE_RLC,
117 	KGD_ENGINE_SDMA1,
118 	KGD_ENGINE_SDMA2,
119 	KGD_ENGINE_MAX
120 };
121 
122 
123 struct amdkfd_process_info {
124 	/* List head of all VMs that belong to a KFD process */
125 	struct list_head vm_list_head;
126 	/* List head for all KFD BOs that belong to a KFD process. */
127 	struct list_head kfd_bo_list;
128 	/* List of userptr BOs that are valid or invalid */
129 	struct list_head userptr_valid_list;
130 	struct list_head userptr_inval_list;
131 	/* Lock to protect kfd_bo_list */
132 	struct mutex lock;
133 
134 	/* Number of VMs */
135 	unsigned int n_vms;
136 	/* Eviction Fence */
137 	struct amdgpu_amdkfd_fence *eviction_fence;
138 
139 	/* MMU-notifier related fields */
140 	struct mutex notifier_lock;
141 	uint32_t evicted_bos;
142 	struct delayed_work restore_userptr_work;
143 	struct pid *pid;
144 	bool block_mmu_notifications;
145 };
146 
147 int amdgpu_amdkfd_init(void);
148 void amdgpu_amdkfd_fini(void);
149 
150 void amdgpu_amdkfd_suspend(struct amdgpu_device *adev, bool run_pm);
151 int amdgpu_amdkfd_resume(struct amdgpu_device *adev, bool run_pm);
152 void amdgpu_amdkfd_interrupt(struct amdgpu_device *adev,
153 			const void *ih_ring_entry);
154 void amdgpu_amdkfd_device_probe(struct amdgpu_device *adev);
155 void amdgpu_amdkfd_device_init(struct amdgpu_device *adev);
156 void amdgpu_amdkfd_device_fini_sw(struct amdgpu_device *adev);
157 int amdgpu_amdkfd_check_and_lock_kfd(struct amdgpu_device *adev);
158 void amdgpu_amdkfd_unlock_kfd(struct amdgpu_device *adev);
159 int amdgpu_amdkfd_submit_ib(struct amdgpu_device *adev,
160 				enum kgd_engine_type engine,
161 				uint32_t vmid, uint64_t gpu_addr,
162 				uint32_t *ib_cmd, uint32_t ib_len);
163 void amdgpu_amdkfd_set_compute_idle(struct amdgpu_device *adev, bool idle);
164 bool amdgpu_amdkfd_have_atomics_support(struct amdgpu_device *adev);
165 int amdgpu_amdkfd_flush_gpu_tlb_vmid(struct amdgpu_device *adev,
166 				uint16_t vmid);
167 int amdgpu_amdkfd_flush_gpu_tlb_pasid(struct amdgpu_device *adev,
168 				uint16_t pasid, enum TLB_FLUSH_TYPE flush_type,
169 				uint32_t inst);
170 
171 bool amdgpu_amdkfd_is_kfd_vmid(struct amdgpu_device *adev, u32 vmid);
172 
173 int amdgpu_amdkfd_pre_reset(struct amdgpu_device *adev);
174 
175 int amdgpu_amdkfd_post_reset(struct amdgpu_device *adev);
176 
177 void amdgpu_amdkfd_gpu_reset(struct amdgpu_device *adev);
178 
179 int amdgpu_queue_mask_bit_to_set_resource_bit(struct amdgpu_device *adev,
180 					int queue_bit);
181 
182 struct amdgpu_amdkfd_fence *amdgpu_amdkfd_fence_create(u64 context,
183 				struct mm_struct *mm,
184 				struct svm_range_bo *svm_bo);
185 #if defined(CONFIG_DEBUG_FS)
186 int kfd_debugfs_kfd_mem_limits(struct seq_file *m, void *data);
187 #endif
188 #if IS_ENABLED(CONFIG_HSA_AMD)
189 bool amdkfd_fence_check_mm(struct dma_fence *f, struct mm_struct *mm);
190 struct amdgpu_amdkfd_fence *to_amdgpu_amdkfd_fence(struct dma_fence *f);
191 int amdgpu_amdkfd_remove_fence_on_pt_pd_bos(struct amdgpu_bo *bo);
192 int amdgpu_amdkfd_evict_userptr(struct mmu_interval_notifier *mni,
193 				unsigned long cur_seq, struct kgd_mem *mem);
194 #else
195 static inline
amdkfd_fence_check_mm(struct dma_fence * f,struct mm_struct * mm)196 bool amdkfd_fence_check_mm(struct dma_fence *f, struct mm_struct *mm)
197 {
198 	return false;
199 }
200 
201 static inline
to_amdgpu_amdkfd_fence(struct dma_fence * f)202 struct amdgpu_amdkfd_fence *to_amdgpu_amdkfd_fence(struct dma_fence *f)
203 {
204 	return NULL;
205 }
206 
207 static inline
amdgpu_amdkfd_remove_fence_on_pt_pd_bos(struct amdgpu_bo * bo)208 int amdgpu_amdkfd_remove_fence_on_pt_pd_bos(struct amdgpu_bo *bo)
209 {
210 	return 0;
211 }
212 
213 static inline
amdgpu_amdkfd_evict_userptr(struct mmu_interval_notifier * mni,unsigned long cur_seq,struct kgd_mem * mem)214 int amdgpu_amdkfd_evict_userptr(struct mmu_interval_notifier *mni,
215 				unsigned long cur_seq, struct kgd_mem *mem)
216 {
217 	return 0;
218 }
219 #endif
220 /* Shared API */
221 int amdgpu_amdkfd_alloc_gtt_mem(struct amdgpu_device *adev, size_t size,
222 				void **mem_obj, uint64_t *gpu_addr,
223 				void **cpu_ptr, bool mqd_gfx9);
224 void amdgpu_amdkfd_free_gtt_mem(struct amdgpu_device *adev, void **mem_obj);
225 int amdgpu_amdkfd_alloc_gws(struct amdgpu_device *adev, size_t size,
226 				void **mem_obj);
227 void amdgpu_amdkfd_free_gws(struct amdgpu_device *adev, void *mem_obj);
228 int amdgpu_amdkfd_add_gws_to_process(void *info, void *gws, struct kgd_mem **mem);
229 int amdgpu_amdkfd_remove_gws_from_process(void *info, void *mem);
230 uint32_t amdgpu_amdkfd_get_fw_version(struct amdgpu_device *adev,
231 				      enum kgd_engine_type type);
232 void amdgpu_amdkfd_get_local_mem_info(struct amdgpu_device *adev,
233 				      struct kfd_local_mem_info *mem_info,
234 				      struct amdgpu_xcp *xcp);
235 uint64_t amdgpu_amdkfd_get_gpu_clock_counter(struct amdgpu_device *adev);
236 
237 uint32_t amdgpu_amdkfd_get_max_engine_clock_in_mhz(struct amdgpu_device *adev);
238 void amdgpu_amdkfd_get_cu_info(struct amdgpu_device *adev,
239 			       struct kfd_cu_info *cu_info);
240 int amdgpu_amdkfd_get_dmabuf_info(struct amdgpu_device *adev, int dma_buf_fd,
241 				  struct amdgpu_device **dmabuf_adev,
242 				  uint64_t *bo_size, void *metadata_buffer,
243 				  size_t buffer_size, uint32_t *metadata_size,
244 				  uint32_t *flags, int8_t *xcp_id);
245 uint8_t amdgpu_amdkfd_get_xgmi_hops_count(struct amdgpu_device *dst,
246 					  struct amdgpu_device *src);
247 int amdgpu_amdkfd_get_xgmi_bandwidth_mbytes(struct amdgpu_device *dst,
248 					    struct amdgpu_device *src,
249 					    bool is_min);
250 int amdgpu_amdkfd_get_pcie_bandwidth_mbytes(struct amdgpu_device *adev, bool is_min);
251 int amdgpu_amdkfd_send_close_event_drain_irq(struct amdgpu_device *adev,
252 					uint32_t *payload);
253 int amdgpu_amdkfd_unmap_hiq(struct amdgpu_device *adev, u32 doorbell_off,
254 				u32 inst);
255 
256 /* Read user wptr from a specified user address space with page fault
257  * disabled. The memory must be pinned and mapped to the hardware when
258  * this is called in hqd_load functions, so it should never fault in
259  * the first place. This resolves a circular lock dependency involving
260  * four locks, including the DQM lock and mmap_lock.
261  */
262 #define read_user_wptr(mmptr, wptr, dst)				\
263 	({								\
264 		bool valid = false;					\
265 		if ((mmptr) && (wptr)) {				\
266 			pagefault_disable();				\
267 			if ((mmptr) == current->mm) {			\
268 				valid = !get_user((dst), (wptr));	\
269 			} else if (current->flags & PF_KTHREAD) {	\
270 				kthread_use_mm(mmptr);			\
271 				valid = !get_user((dst), (wptr));	\
272 				kthread_unuse_mm(mmptr);		\
273 			}						\
274 			pagefault_enable();				\
275 		}							\
276 		valid;							\
277 	})
278 
279 /* GPUVM API */
280 #define drm_priv_to_vm(drm_priv)					\
281 	(&((struct amdgpu_fpriv *)					\
282 		((struct drm_file *)(drm_priv))->driver_priv)->vm)
283 
284 int amdgpu_amdkfd_gpuvm_set_vm_pasid(struct amdgpu_device *adev,
285 				     struct amdgpu_vm *avm, u32 pasid);
286 int amdgpu_amdkfd_gpuvm_acquire_process_vm(struct amdgpu_device *adev,
287 					struct amdgpu_vm *avm,
288 					void **process_info,
289 					struct dma_fence **ef);
290 void amdgpu_amdkfd_gpuvm_release_process_vm(struct amdgpu_device *adev,
291 					void *drm_priv);
292 uint64_t amdgpu_amdkfd_gpuvm_get_process_page_dir(void *drm_priv);
293 size_t amdgpu_amdkfd_get_available_memory(struct amdgpu_device *adev,
294 					uint8_t xcp_id);
295 int amdgpu_amdkfd_gpuvm_alloc_memory_of_gpu(
296 		struct amdgpu_device *adev, uint64_t va, uint64_t size,
297 		void *drm_priv, struct kgd_mem **mem,
298 		uint64_t *offset, uint32_t flags, bool criu_resume);
299 int amdgpu_amdkfd_gpuvm_free_memory_of_gpu(
300 		struct amdgpu_device *adev, struct kgd_mem *mem, void *drm_priv,
301 		uint64_t *size);
302 int amdgpu_amdkfd_gpuvm_map_memory_to_gpu(struct amdgpu_device *adev,
303 					  struct kgd_mem *mem, void *drm_priv);
304 int amdgpu_amdkfd_gpuvm_unmap_memory_from_gpu(
305 		struct amdgpu_device *adev, struct kgd_mem *mem, void *drm_priv);
306 int amdgpu_amdkfd_gpuvm_dmaunmap_mem(struct kgd_mem *mem, void *drm_priv);
307 int amdgpu_amdkfd_gpuvm_sync_memory(
308 		struct amdgpu_device *adev, struct kgd_mem *mem, bool intr);
309 int amdgpu_amdkfd_gpuvm_map_gtt_bo_to_kernel(struct kgd_mem *mem,
310 					     void **kptr, uint64_t *size);
311 void amdgpu_amdkfd_gpuvm_unmap_gtt_bo_from_kernel(struct kgd_mem *mem);
312 
313 int amdgpu_amdkfd_map_gtt_bo_to_gart(struct amdgpu_device *adev, struct amdgpu_bo *bo);
314 
315 int amdgpu_amdkfd_gpuvm_restore_process_bos(void *process_info,
316 					    struct dma_fence **ef);
317 int amdgpu_amdkfd_gpuvm_get_vm_fault_info(struct amdgpu_device *adev,
318 					      struct kfd_vm_fault_info *info);
319 int amdgpu_amdkfd_gpuvm_import_dmabuf(struct amdgpu_device *adev,
320 				      struct dma_buf *dmabuf,
321 				      uint64_t va, void *drm_priv,
322 				      struct kgd_mem **mem, uint64_t *size,
323 				      uint64_t *mmap_offset);
324 int amdgpu_amdkfd_gpuvm_export_dmabuf(struct kgd_mem *mem,
325 				      struct dma_buf **dmabuf);
326 void amdgpu_amdkfd_debug_mem_fence(struct amdgpu_device *adev);
327 int amdgpu_amdkfd_get_tile_config(struct amdgpu_device *adev,
328 				struct tile_config *config);
329 void amdgpu_amdkfd_ras_poison_consumption_handler(struct amdgpu_device *adev,
330 				bool reset);
331 bool amdgpu_amdkfd_bo_mapped_to_dev(struct amdgpu_device *adev, struct kgd_mem *mem);
332 void amdgpu_amdkfd_block_mmu_notifications(void *p);
333 int amdgpu_amdkfd_criu_resume(void *p);
334 bool amdgpu_amdkfd_ras_query_utcl2_poison_status(struct amdgpu_device *adev);
335 int amdgpu_amdkfd_reserve_mem_limit(struct amdgpu_device *adev,
336 		uint64_t size, u32 alloc_flag, int8_t xcp_id);
337 void amdgpu_amdkfd_unreserve_mem_limit(struct amdgpu_device *adev,
338 		uint64_t size, u32 alloc_flag, int8_t xcp_id);
339 
340 u64 amdgpu_amdkfd_xcp_memory_size(struct amdgpu_device *adev, int xcp_id);
341 
342 #define KFD_XCP_MEM_ID(adev, xcp_id) \
343 		((adev)->xcp_mgr && (xcp_id) >= 0 ?\
344 		(adev)->xcp_mgr->xcp[(xcp_id)].mem_id : -1)
345 
346 #define KFD_XCP_MEMORY_SIZE(adev, xcp_id) amdgpu_amdkfd_xcp_memory_size((adev), (xcp_id))
347 
348 
349 #if IS_ENABLED(CONFIG_HSA_AMD)
350 void amdgpu_amdkfd_gpuvm_init_mem_limits(void);
351 void amdgpu_amdkfd_gpuvm_destroy_cb(struct amdgpu_device *adev,
352 				struct amdgpu_vm *vm);
353 
354 /**
355  * @amdgpu_amdkfd_release_notify() - Notify KFD when GEM object is released
356  *
357  * Allows KFD to release its resources associated with the GEM object.
358  */
359 void amdgpu_amdkfd_release_notify(struct amdgpu_bo *bo);
360 void amdgpu_amdkfd_reserve_system_mem(uint64_t size);
361 #else
362 static inline
amdgpu_amdkfd_gpuvm_init_mem_limits(void)363 void amdgpu_amdkfd_gpuvm_init_mem_limits(void)
364 {
365 }
366 
367 static inline
amdgpu_amdkfd_gpuvm_destroy_cb(struct amdgpu_device * adev,struct amdgpu_vm * vm)368 void amdgpu_amdkfd_gpuvm_destroy_cb(struct amdgpu_device *adev,
369 					struct amdgpu_vm *vm)
370 {
371 }
372 
373 static inline
amdgpu_amdkfd_release_notify(struct amdgpu_bo * bo)374 void amdgpu_amdkfd_release_notify(struct amdgpu_bo *bo)
375 {
376 }
377 #endif
378 
379 #if IS_ENABLED(CONFIG_HSA_AMD_SVM)
380 int kgd2kfd_init_zone_device(struct amdgpu_device *adev);
381 #else
382 static inline
kgd2kfd_init_zone_device(struct amdgpu_device * adev)383 int kgd2kfd_init_zone_device(struct amdgpu_device *adev)
384 {
385 	return 0;
386 }
387 #endif
388 
389 /* KGD2KFD callbacks */
390 int kgd2kfd_quiesce_mm(struct mm_struct *mm, uint32_t trigger);
391 int kgd2kfd_resume_mm(struct mm_struct *mm);
392 int kgd2kfd_schedule_evict_and_restore_process(struct mm_struct *mm,
393 						struct dma_fence *fence);
394 #if IS_ENABLED(CONFIG_HSA_AMD)
395 int kgd2kfd_init(void);
396 void kgd2kfd_exit(void);
397 struct kfd_dev *kgd2kfd_probe(struct amdgpu_device *adev, bool vf);
398 bool kgd2kfd_device_init(struct kfd_dev *kfd,
399 			 const struct kgd2kfd_shared_resources *gpu_resources);
400 void kgd2kfd_device_exit(struct kfd_dev *kfd);
401 void kgd2kfd_suspend(struct kfd_dev *kfd, bool run_pm);
402 int kgd2kfd_resume(struct kfd_dev *kfd, bool run_pm);
403 int kgd2kfd_pre_reset(struct kfd_dev *kfd);
404 int kgd2kfd_post_reset(struct kfd_dev *kfd);
405 void kgd2kfd_interrupt(struct kfd_dev *kfd, const void *ih_ring_entry);
406 void kgd2kfd_set_sram_ecc_flag(struct kfd_dev *kfd);
407 void kgd2kfd_smi_event_throttle(struct kfd_dev *kfd, uint64_t throttle_bitmask);
408 int kgd2kfd_check_and_lock_kfd(void);
409 void kgd2kfd_unlock_kfd(void);
410 #else
kgd2kfd_init(void)411 static inline int kgd2kfd_init(void)
412 {
413 	return -ENOENT;
414 }
415 
kgd2kfd_exit(void)416 static inline void kgd2kfd_exit(void)
417 {
418 }
419 
420 static inline
kgd2kfd_probe(struct amdgpu_device * adev,bool vf)421 struct kfd_dev *kgd2kfd_probe(struct amdgpu_device *adev, bool vf)
422 {
423 	return NULL;
424 }
425 
426 static inline
kgd2kfd_device_init(struct kfd_dev * kfd,const struct kgd2kfd_shared_resources * gpu_resources)427 bool kgd2kfd_device_init(struct kfd_dev *kfd,
428 				const struct kgd2kfd_shared_resources *gpu_resources)
429 {
430 	return false;
431 }
432 
kgd2kfd_device_exit(struct kfd_dev * kfd)433 static inline void kgd2kfd_device_exit(struct kfd_dev *kfd)
434 {
435 }
436 
kgd2kfd_suspend(struct kfd_dev * kfd,bool run_pm)437 static inline void kgd2kfd_suspend(struct kfd_dev *kfd, bool run_pm)
438 {
439 }
440 
kgd2kfd_resume(struct kfd_dev * kfd,bool run_pm)441 static inline int kgd2kfd_resume(struct kfd_dev *kfd, bool run_pm)
442 {
443 	return 0;
444 }
445 
kgd2kfd_pre_reset(struct kfd_dev * kfd)446 static inline int kgd2kfd_pre_reset(struct kfd_dev *kfd)
447 {
448 	return 0;
449 }
450 
kgd2kfd_post_reset(struct kfd_dev * kfd)451 static inline int kgd2kfd_post_reset(struct kfd_dev *kfd)
452 {
453 	return 0;
454 }
455 
456 static inline
kgd2kfd_interrupt(struct kfd_dev * kfd,const void * ih_ring_entry)457 void kgd2kfd_interrupt(struct kfd_dev *kfd, const void *ih_ring_entry)
458 {
459 }
460 
461 static inline
kgd2kfd_set_sram_ecc_flag(struct kfd_dev * kfd)462 void kgd2kfd_set_sram_ecc_flag(struct kfd_dev *kfd)
463 {
464 }
465 
466 static inline
kgd2kfd_smi_event_throttle(struct kfd_dev * kfd,uint64_t throttle_bitmask)467 void kgd2kfd_smi_event_throttle(struct kfd_dev *kfd, uint64_t throttle_bitmask)
468 {
469 }
470 
kgd2kfd_check_and_lock_kfd(void)471 static inline int kgd2kfd_check_and_lock_kfd(void)
472 {
473 	return 0;
474 }
475 
kgd2kfd_unlock_kfd(void)476 static inline void kgd2kfd_unlock_kfd(void)
477 {
478 }
479 #endif
480 #endif /* AMDGPU_AMDKFD_H_INCLUDED */
481