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 	KFD_MEM_ATT_SG		/* Tag to DMA map SG BOs */
52 };
53 
54 struct kfd_mem_attachment {
55 	struct list_head list;
56 	enum kfd_mem_attachment_type type;
57 	bool is_mapped;
58 	struct amdgpu_bo_va *bo_va;
59 	struct amdgpu_device *adev;
60 	uint64_t va;
61 	uint64_t pte_flags;
62 };
63 
64 struct kgd_mem {
65 	struct mutex lock;
66 	struct amdgpu_bo *bo;
67 	struct dma_buf *dmabuf;
68 	struct list_head attachments;
69 	/* protected by amdkfd_process_info.lock */
70 	struct ttm_validate_buffer validate_list;
71 	struct ttm_validate_buffer resv_list;
72 	uint32_t domain;
73 	unsigned int mapped_to_gpu_memory;
74 	uint64_t va;
75 
76 	uint32_t alloc_flags;
77 
78 	atomic_t invalid;
79 	struct amdkfd_process_info *process_info;
80 
81 	struct amdgpu_sync sync;
82 
83 	bool aql_queue;
84 	bool is_imported;
85 };
86 
87 /* KFD Memory Eviction */
88 struct amdgpu_amdkfd_fence {
89 	struct dma_fence base;
90 	struct mm_struct *mm;
91 	spinlock_t lock;
92 	char timeline_name[TASK_COMM_LEN];
93 	struct svm_range_bo *svm_bo;
94 };
95 
96 struct amdgpu_kfd_dev {
97 	struct kfd_dev *dev;
98 	uint64_t vram_used;
99 	uint64_t vram_used_aligned;
100 	bool init_complete;
101 	struct work_struct reset_work;
102 };
103 
104 enum kgd_engine_type {
105 	KGD_ENGINE_PFP = 1,
106 	KGD_ENGINE_ME,
107 	KGD_ENGINE_CE,
108 	KGD_ENGINE_MEC1,
109 	KGD_ENGINE_MEC2,
110 	KGD_ENGINE_RLC,
111 	KGD_ENGINE_SDMA1,
112 	KGD_ENGINE_SDMA2,
113 	KGD_ENGINE_MAX
114 };
115 
116 
117 struct amdkfd_process_info {
118 	/* List head of all VMs that belong to a KFD process */
119 	struct list_head vm_list_head;
120 	/* List head for all KFD BOs that belong to a KFD process. */
121 	struct list_head kfd_bo_list;
122 	/* List of userptr BOs that are valid or invalid */
123 	struct list_head userptr_valid_list;
124 	struct list_head userptr_inval_list;
125 	/* Lock to protect kfd_bo_list */
126 	struct mutex lock;
127 
128 	/* Number of VMs */
129 	unsigned int n_vms;
130 	/* Eviction Fence */
131 	struct amdgpu_amdkfd_fence *eviction_fence;
132 
133 	/* MMU-notifier related fields */
134 	atomic_t evicted_bos;
135 	struct delayed_work restore_userptr_work;
136 	struct pid *pid;
137 	bool block_mmu_notifications;
138 };
139 
140 int amdgpu_amdkfd_init(void);
141 void amdgpu_amdkfd_fini(void);
142 
143 void amdgpu_amdkfd_suspend(struct amdgpu_device *adev, bool run_pm);
144 int amdgpu_amdkfd_resume_iommu(struct amdgpu_device *adev);
145 int amdgpu_amdkfd_resume(struct amdgpu_device *adev, bool run_pm);
146 void amdgpu_amdkfd_interrupt(struct amdgpu_device *adev,
147 			const void *ih_ring_entry);
148 void amdgpu_amdkfd_device_probe(struct amdgpu_device *adev);
149 void amdgpu_amdkfd_device_init(struct amdgpu_device *adev);
150 void amdgpu_amdkfd_device_fini_sw(struct amdgpu_device *adev);
151 int amdgpu_amdkfd_submit_ib(struct amdgpu_device *adev,
152 				enum kgd_engine_type engine,
153 				uint32_t vmid, uint64_t gpu_addr,
154 				uint32_t *ib_cmd, uint32_t ib_len);
155 void amdgpu_amdkfd_set_compute_idle(struct amdgpu_device *adev, bool idle);
156 bool amdgpu_amdkfd_have_atomics_support(struct amdgpu_device *adev);
157 int amdgpu_amdkfd_flush_gpu_tlb_vmid(struct amdgpu_device *adev,
158 				uint16_t vmid);
159 int amdgpu_amdkfd_flush_gpu_tlb_pasid(struct amdgpu_device *adev,
160 				uint16_t pasid, enum TLB_FLUSH_TYPE flush_type);
161 
162 bool amdgpu_amdkfd_is_kfd_vmid(struct amdgpu_device *adev, u32 vmid);
163 
164 int amdgpu_amdkfd_pre_reset(struct amdgpu_device *adev);
165 
166 int amdgpu_amdkfd_post_reset(struct amdgpu_device *adev);
167 
168 void amdgpu_amdkfd_gpu_reset(struct amdgpu_device *adev);
169 
170 int amdgpu_queue_mask_bit_to_set_resource_bit(struct amdgpu_device *adev,
171 					int queue_bit);
172 
173 struct amdgpu_amdkfd_fence *amdgpu_amdkfd_fence_create(u64 context,
174 				struct mm_struct *mm,
175 				struct svm_range_bo *svm_bo);
176 #if defined(CONFIG_DEBUG_FS)
177 int kfd_debugfs_kfd_mem_limits(struct seq_file *m, void *data);
178 #endif
179 #if IS_ENABLED(CONFIG_HSA_AMD)
180 bool amdkfd_fence_check_mm(struct dma_fence *f, struct mm_struct *mm);
181 struct amdgpu_amdkfd_fence *to_amdgpu_amdkfd_fence(struct dma_fence *f);
182 int amdgpu_amdkfd_remove_fence_on_pt_pd_bos(struct amdgpu_bo *bo);
183 int amdgpu_amdkfd_evict_userptr(struct kgd_mem *mem, struct mm_struct *mm);
184 #else
185 static inline
amdkfd_fence_check_mm(struct dma_fence * f,struct mm_struct * mm)186 bool amdkfd_fence_check_mm(struct dma_fence *f, struct mm_struct *mm)
187 {
188 	return false;
189 }
190 
191 static inline
to_amdgpu_amdkfd_fence(struct dma_fence * f)192 struct amdgpu_amdkfd_fence *to_amdgpu_amdkfd_fence(struct dma_fence *f)
193 {
194 	return NULL;
195 }
196 
197 static inline
amdgpu_amdkfd_remove_fence_on_pt_pd_bos(struct amdgpu_bo * bo)198 int amdgpu_amdkfd_remove_fence_on_pt_pd_bos(struct amdgpu_bo *bo)
199 {
200 	return 0;
201 }
202 
203 static inline
amdgpu_amdkfd_evict_userptr(struct kgd_mem * mem,struct mm_struct * mm)204 int amdgpu_amdkfd_evict_userptr(struct kgd_mem *mem, struct mm_struct *mm)
205 {
206 	return 0;
207 }
208 #endif
209 /* Shared API */
210 int amdgpu_amdkfd_alloc_gtt_mem(struct amdgpu_device *adev, size_t size,
211 				void **mem_obj, uint64_t *gpu_addr,
212 				void **cpu_ptr, bool mqd_gfx9);
213 void amdgpu_amdkfd_free_gtt_mem(struct amdgpu_device *adev, void *mem_obj);
214 int amdgpu_amdkfd_alloc_gws(struct amdgpu_device *adev, size_t size,
215 				void **mem_obj);
216 void amdgpu_amdkfd_free_gws(struct amdgpu_device *adev, void *mem_obj);
217 int amdgpu_amdkfd_add_gws_to_process(void *info, void *gws, struct kgd_mem **mem);
218 int amdgpu_amdkfd_remove_gws_from_process(void *info, void *mem);
219 uint32_t amdgpu_amdkfd_get_fw_version(struct amdgpu_device *adev,
220 				      enum kgd_engine_type type);
221 void amdgpu_amdkfd_get_local_mem_info(struct amdgpu_device *adev,
222 				      struct kfd_local_mem_info *mem_info);
223 uint64_t amdgpu_amdkfd_get_gpu_clock_counter(struct amdgpu_device *adev);
224 
225 uint32_t amdgpu_amdkfd_get_max_engine_clock_in_mhz(struct amdgpu_device *adev);
226 void amdgpu_amdkfd_get_cu_info(struct amdgpu_device *adev,
227 			       struct kfd_cu_info *cu_info);
228 int amdgpu_amdkfd_get_dmabuf_info(struct amdgpu_device *adev, int dma_buf_fd,
229 				  struct amdgpu_device **dmabuf_adev,
230 				  uint64_t *bo_size, void *metadata_buffer,
231 				  size_t buffer_size, uint32_t *metadata_size,
232 				  uint32_t *flags);
233 uint8_t amdgpu_amdkfd_get_xgmi_hops_count(struct amdgpu_device *dst,
234 					  struct amdgpu_device *src);
235 int amdgpu_amdkfd_get_xgmi_bandwidth_mbytes(struct amdgpu_device *dst,
236 					    struct amdgpu_device *src,
237 					    bool is_min);
238 int amdgpu_amdkfd_get_pcie_bandwidth_mbytes(struct amdgpu_device *adev, bool is_min);
239 
240 /* Read user wptr from a specified user address space with page fault
241  * disabled. The memory must be pinned and mapped to the hardware when
242  * this is called in hqd_load functions, so it should never fault in
243  * the first place. This resolves a circular lock dependency involving
244  * four locks, including the DQM lock and mmap_lock.
245  */
246 #define read_user_wptr(mmptr, wptr, dst)				\
247 	({								\
248 		bool valid = false;					\
249 		if ((mmptr) && (wptr)) {				\
250 			pagefault_disable();				\
251 			if ((mmptr) == current->mm) {			\
252 				valid = !get_user((dst), (wptr));	\
253 			} else if (current->flags & PF_KTHREAD) {	\
254 				kthread_use_mm(mmptr);			\
255 				valid = !get_user((dst), (wptr));	\
256 				kthread_unuse_mm(mmptr);		\
257 			}						\
258 			pagefault_enable();				\
259 		}							\
260 		valid;							\
261 	})
262 
263 /* GPUVM API */
264 #define drm_priv_to_vm(drm_priv)					\
265 	(&((struct amdgpu_fpriv *)					\
266 		((struct drm_file *)(drm_priv))->driver_priv)->vm)
267 
268 int amdgpu_amdkfd_gpuvm_set_vm_pasid(struct amdgpu_device *adev,
269 				     struct file *filp, u32 pasid);
270 int amdgpu_amdkfd_gpuvm_acquire_process_vm(struct amdgpu_device *adev,
271 					struct file *filp,
272 					void **process_info,
273 					struct dma_fence **ef);
274 void amdgpu_amdkfd_gpuvm_release_process_vm(struct amdgpu_device *adev,
275 					void *drm_priv);
276 uint64_t amdgpu_amdkfd_gpuvm_get_process_page_dir(void *drm_priv);
277 size_t amdgpu_amdkfd_get_available_memory(struct amdgpu_device *adev);
278 int amdgpu_amdkfd_gpuvm_alloc_memory_of_gpu(
279 		struct amdgpu_device *adev, uint64_t va, uint64_t size,
280 		void *drm_priv, struct kgd_mem **mem,
281 		uint64_t *offset, uint32_t flags, bool criu_resume);
282 int amdgpu_amdkfd_gpuvm_free_memory_of_gpu(
283 		struct amdgpu_device *adev, struct kgd_mem *mem, void *drm_priv,
284 		uint64_t *size);
285 int amdgpu_amdkfd_gpuvm_map_memory_to_gpu(struct amdgpu_device *adev,
286 					  struct kgd_mem *mem, void *drm_priv);
287 int amdgpu_amdkfd_gpuvm_unmap_memory_from_gpu(
288 		struct amdgpu_device *adev, struct kgd_mem *mem, void *drm_priv);
289 int amdgpu_amdkfd_gpuvm_sync_memory(
290 		struct amdgpu_device *adev, struct kgd_mem *mem, bool intr);
291 int amdgpu_amdkfd_gpuvm_map_gtt_bo_to_kernel(struct kgd_mem *mem,
292 					     void **kptr, uint64_t *size);
293 void amdgpu_amdkfd_gpuvm_unmap_gtt_bo_from_kernel(struct kgd_mem *mem);
294 
295 int amdgpu_amdkfd_map_gtt_bo_to_gart(struct amdgpu_device *adev, struct amdgpu_bo *bo);
296 
297 int amdgpu_amdkfd_gpuvm_restore_process_bos(void *process_info,
298 					    struct dma_fence **ef);
299 int amdgpu_amdkfd_gpuvm_get_vm_fault_info(struct amdgpu_device *adev,
300 					      struct kfd_vm_fault_info *info);
301 int amdgpu_amdkfd_gpuvm_import_dmabuf(struct amdgpu_device *adev,
302 				      struct dma_buf *dmabuf,
303 				      uint64_t va, void *drm_priv,
304 				      struct kgd_mem **mem, uint64_t *size,
305 				      uint64_t *mmap_offset);
306 int amdgpu_amdkfd_get_tile_config(struct amdgpu_device *adev,
307 				struct tile_config *config);
308 void amdgpu_amdkfd_ras_poison_consumption_handler(struct amdgpu_device *adev,
309 				bool reset);
310 bool amdgpu_amdkfd_bo_mapped_to_dev(struct amdgpu_device *adev, struct kgd_mem *mem);
311 void amdgpu_amdkfd_block_mmu_notifications(void *p);
312 int amdgpu_amdkfd_criu_resume(void *p);
313 bool amdgpu_amdkfd_ras_query_utcl2_poison_status(struct amdgpu_device *adev);
314 int amdgpu_amdkfd_reserve_mem_limit(struct amdgpu_device *adev,
315 		uint64_t size, u32 alloc_flag);
316 void amdgpu_amdkfd_unreserve_mem_limit(struct amdgpu_device *adev,
317 		uint64_t size, u32 alloc_flag);
318 
319 #if IS_ENABLED(CONFIG_HSA_AMD)
320 void amdgpu_amdkfd_gpuvm_init_mem_limits(void);
321 void amdgpu_amdkfd_gpuvm_destroy_cb(struct amdgpu_device *adev,
322 				struct amdgpu_vm *vm);
323 
324 /**
325  * @amdgpu_amdkfd_release_notify() - Notify KFD when GEM object is released
326  *
327  * Allows KFD to release its resources associated with the GEM object.
328  */
329 void amdgpu_amdkfd_release_notify(struct amdgpu_bo *bo);
330 void amdgpu_amdkfd_reserve_system_mem(uint64_t size);
331 #else
332 static inline
amdgpu_amdkfd_gpuvm_init_mem_limits(void)333 void amdgpu_amdkfd_gpuvm_init_mem_limits(void)
334 {
335 }
336 
337 static inline
amdgpu_amdkfd_gpuvm_destroy_cb(struct amdgpu_device * adev,struct amdgpu_vm * vm)338 void amdgpu_amdkfd_gpuvm_destroy_cb(struct amdgpu_device *adev,
339 					struct amdgpu_vm *vm)
340 {
341 }
342 
343 static inline
amdgpu_amdkfd_release_notify(struct amdgpu_bo * bo)344 void amdgpu_amdkfd_release_notify(struct amdgpu_bo *bo)
345 {
346 }
347 #endif
348 /* KGD2KFD callbacks */
349 int kgd2kfd_quiesce_mm(struct mm_struct *mm, uint32_t trigger);
350 int kgd2kfd_resume_mm(struct mm_struct *mm);
351 int kgd2kfd_schedule_evict_and_restore_process(struct mm_struct *mm,
352 						struct dma_fence *fence);
353 #if IS_ENABLED(CONFIG_HSA_AMD)
354 int kgd2kfd_init(void);
355 void kgd2kfd_exit(void);
356 struct kfd_dev *kgd2kfd_probe(struct amdgpu_device *adev, bool vf);
357 bool kgd2kfd_device_init(struct kfd_dev *kfd,
358 			 struct drm_device *ddev,
359 			 const struct kgd2kfd_shared_resources *gpu_resources);
360 void kgd2kfd_device_exit(struct kfd_dev *kfd);
361 void kgd2kfd_suspend(struct kfd_dev *kfd, bool run_pm);
362 int kgd2kfd_resume_iommu(struct kfd_dev *kfd);
363 int kgd2kfd_resume(struct kfd_dev *kfd, bool run_pm);
364 int kgd2kfd_pre_reset(struct kfd_dev *kfd);
365 int kgd2kfd_post_reset(struct kfd_dev *kfd);
366 void kgd2kfd_interrupt(struct kfd_dev *kfd, const void *ih_ring_entry);
367 void kgd2kfd_set_sram_ecc_flag(struct kfd_dev *kfd);
368 void kgd2kfd_smi_event_throttle(struct kfd_dev *kfd, uint64_t throttle_bitmask);
369 #else
kgd2kfd_init(void)370 static inline int kgd2kfd_init(void)
371 {
372 	return -ENOENT;
373 }
374 
kgd2kfd_exit(void)375 static inline void kgd2kfd_exit(void)
376 {
377 }
378 
379 static inline
kgd2kfd_probe(struct amdgpu_device * adev,bool vf)380 struct kfd_dev *kgd2kfd_probe(struct amdgpu_device *adev, bool vf)
381 {
382 	return NULL;
383 }
384 
385 static inline
kgd2kfd_device_init(struct kfd_dev * kfd,struct drm_device * ddev,const struct kgd2kfd_shared_resources * gpu_resources)386 bool kgd2kfd_device_init(struct kfd_dev *kfd, struct drm_device *ddev,
387 				const struct kgd2kfd_shared_resources *gpu_resources)
388 {
389 	return false;
390 }
391 
kgd2kfd_device_exit(struct kfd_dev * kfd)392 static inline void kgd2kfd_device_exit(struct kfd_dev *kfd)
393 {
394 }
395 
kgd2kfd_suspend(struct kfd_dev * kfd,bool run_pm)396 static inline void kgd2kfd_suspend(struct kfd_dev *kfd, bool run_pm)
397 {
398 }
399 
kgd2kfd_resume_iommu(struct kfd_dev * kfd)400 static int __maybe_unused kgd2kfd_resume_iommu(struct kfd_dev *kfd)
401 {
402 	return 0;
403 }
404 
kgd2kfd_resume(struct kfd_dev * kfd,bool run_pm)405 static inline int kgd2kfd_resume(struct kfd_dev *kfd, bool run_pm)
406 {
407 	return 0;
408 }
409 
kgd2kfd_pre_reset(struct kfd_dev * kfd)410 static inline int kgd2kfd_pre_reset(struct kfd_dev *kfd)
411 {
412 	return 0;
413 }
414 
kgd2kfd_post_reset(struct kfd_dev * kfd)415 static inline int kgd2kfd_post_reset(struct kfd_dev *kfd)
416 {
417 	return 0;
418 }
419 
420 static inline
kgd2kfd_interrupt(struct kfd_dev * kfd,const void * ih_ring_entry)421 void kgd2kfd_interrupt(struct kfd_dev *kfd, const void *ih_ring_entry)
422 {
423 }
424 
425 static inline
kgd2kfd_set_sram_ecc_flag(struct kfd_dev * kfd)426 void kgd2kfd_set_sram_ecc_flag(struct kfd_dev *kfd)
427 {
428 }
429 
430 static inline
kgd2kfd_smi_event_throttle(struct kfd_dev * kfd,uint64_t throttle_bitmask)431 void kgd2kfd_smi_event_throttle(struct kfd_dev *kfd, uint64_t throttle_bitmask)
432 {
433 }
434 #endif
435 #endif /* AMDGPU_AMDKFD_H_INCLUDED */
436