xref: /linux/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.h (revision c4a060372590848d886c24b8215139cab122c0ed)
1 /*
2  * Copyright 2016 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  * Authors: Christian König
23  */
24 #ifndef __AMDGPU_VM_H__
25 #define __AMDGPU_VM_H__
26 
27 #include <linux/idr.h>
28 #include <linux/kfifo.h>
29 #include <linux/rbtree.h>
30 #include <drm/gpu_scheduler.h>
31 #include <drm/drm_file.h>
32 #include <drm/ttm/ttm_bo.h>
33 #include <linux/sched/mm.h>
34 
35 #include "amdgpu_sync.h"
36 #include "amdgpu_ring.h"
37 #include "amdgpu_ids.h"
38 
39 struct drm_exec;
40 
41 struct amdgpu_bo_va;
42 struct amdgpu_job;
43 struct amdgpu_bo_list_entry;
44 struct amdgpu_bo_vm;
45 struct amdgpu_mem_stats;
46 
47 /*
48  * GPUVM handling
49  */
50 
51 /* Maximum number of PTEs the hardware can write with one command */
52 #define AMDGPU_VM_MAX_UPDATE_SIZE	0x3FFFF
53 
54 /* number of entries in page table */
55 #define AMDGPU_VM_PTE_COUNT(adev) (1 << (adev)->vm_manager.block_size)
56 
57 #define AMDGPU_PTE_VALID	(1ULL << 0)
58 #define AMDGPU_PTE_SYSTEM	(1ULL << 1)
59 #define AMDGPU_PTE_SNOOPED	(1ULL << 2)
60 
61 /* RV+ */
62 #define AMDGPU_PTE_TMZ		(1ULL << 3)
63 
64 /* VI only */
65 #define AMDGPU_PTE_EXECUTABLE	(1ULL << 4)
66 
67 #define AMDGPU_PTE_READABLE	(1ULL << 5)
68 #define AMDGPU_PTE_WRITEABLE	(1ULL << 6)
69 
70 #define AMDGPU_PTE_FRAG(x)	((x & 0x1fULL) << 7)
71 
72 /* TILED for VEGA10, reserved for older ASICs  */
73 #define AMDGPU_PTE_PRT		(1ULL << 51)
74 
75 /* PDE is handled as PTE for VEGA10 */
76 #define AMDGPU_PDE_PTE		(1ULL << 54)
77 
78 #define AMDGPU_PTE_LOG          (1ULL << 55)
79 
80 /* PTE is handled as PDE for VEGA10 (Translate Further) */
81 #define AMDGPU_PTE_TF		(1ULL << 56)
82 
83 /* MALL noalloc for sienna_cichlid, reserved for older ASICs  */
84 #define AMDGPU_PTE_NOALLOC	(1ULL << 58)
85 
86 /* PDE Block Fragment Size for VEGA10 */
87 #define AMDGPU_PDE_BFS(a)	((uint64_t)a << 59)
88 
89 /* Flag combination to set no-retry with TF disabled */
90 #define AMDGPU_VM_NORETRY_FLAGS	(AMDGPU_PTE_EXECUTABLE | AMDGPU_PDE_PTE | \
91 				AMDGPU_PTE_TF)
92 
93 /* Flag combination to set no-retry with TF enabled */
94 #define AMDGPU_VM_NORETRY_FLAGS_TF (AMDGPU_PTE_VALID | AMDGPU_PTE_SYSTEM | \
95 				   AMDGPU_PTE_PRT)
96 /* For GFX9 */
97 #define AMDGPU_PTE_MTYPE_VG10(a)	((uint64_t)(a) << 57)
98 #define AMDGPU_PTE_MTYPE_VG10_MASK	AMDGPU_PTE_MTYPE_VG10(3ULL)
99 
100 #define AMDGPU_MTYPE_NC 0
101 #define AMDGPU_MTYPE_CC 2
102 
103 #define AMDGPU_PTE_DEFAULT_ATC  (AMDGPU_PTE_SYSTEM      \
104                                 | AMDGPU_PTE_SNOOPED    \
105                                 | AMDGPU_PTE_EXECUTABLE \
106                                 | AMDGPU_PTE_READABLE   \
107                                 | AMDGPU_PTE_WRITEABLE  \
108                                 | AMDGPU_PTE_MTYPE_VG10(AMDGPU_MTYPE_CC))
109 
110 /* gfx10 */
111 #define AMDGPU_PTE_MTYPE_NV10(a)       ((uint64_t)(a) << 48)
112 #define AMDGPU_PTE_MTYPE_NV10_MASK     AMDGPU_PTE_MTYPE_NV10(7ULL)
113 
114 /* How to program VM fault handling */
115 #define AMDGPU_VM_FAULT_STOP_NEVER	0
116 #define AMDGPU_VM_FAULT_STOP_FIRST	1
117 #define AMDGPU_VM_FAULT_STOP_ALWAYS	2
118 
119 /* How much VRAM be reserved for page tables */
120 #define AMDGPU_VM_RESERVED_VRAM		(8ULL << 20)
121 
122 /*
123  * max number of VMHUB
124  * layout: max 8 GFXHUB + 4 MMHUB0 + 1 MMHUB1
125  */
126 #define AMDGPU_MAX_VMHUBS			13
127 #define AMDGPU_GFXHUB_START			0
128 #define AMDGPU_MMHUB0_START			8
129 #define AMDGPU_MMHUB1_START			12
130 #define AMDGPU_GFXHUB(x)			(AMDGPU_GFXHUB_START + (x))
131 #define AMDGPU_MMHUB0(x)			(AMDGPU_MMHUB0_START + (x))
132 #define AMDGPU_MMHUB1(x)			(AMDGPU_MMHUB1_START + (x))
133 
134 #define AMDGPU_IS_GFXHUB(x) ((x) >= AMDGPU_GFXHUB_START && (x) < AMDGPU_MMHUB0_START)
135 #define AMDGPU_IS_MMHUB0(x) ((x) >= AMDGPU_MMHUB0_START && (x) < AMDGPU_MMHUB1_START)
136 #define AMDGPU_IS_MMHUB1(x) ((x) >= AMDGPU_MMHUB1_START && (x) < AMDGPU_MAX_VMHUBS)
137 
138 /* Reserve space at top/bottom of address space for kernel use */
139 #define AMDGPU_VA_RESERVED_CSA_SIZE		(2ULL << 20)
140 #define AMDGPU_VA_RESERVED_CSA_START(adev)	(((adev)->vm_manager.max_pfn \
141 						  << AMDGPU_GPU_PAGE_SHIFT)  \
142 						 - AMDGPU_VA_RESERVED_CSA_SIZE)
143 #define AMDGPU_VA_RESERVED_SEQ64_SIZE		(2ULL << 20)
144 #define AMDGPU_VA_RESERVED_SEQ64_START(adev)	(AMDGPU_VA_RESERVED_CSA_START(adev) \
145 						 - AMDGPU_VA_RESERVED_SEQ64_SIZE)
146 #define AMDGPU_VA_RESERVED_TRAP_SIZE		(2ULL << 12)
147 #define AMDGPU_VA_RESERVED_TRAP_START(adev)	(AMDGPU_VA_RESERVED_SEQ64_START(adev) \
148 						 - AMDGPU_VA_RESERVED_TRAP_SIZE)
149 #define AMDGPU_VA_RESERVED_BOTTOM		(1ULL << 16)
150 #define AMDGPU_VA_RESERVED_TOP			(AMDGPU_VA_RESERVED_TRAP_SIZE + \
151 						 AMDGPU_VA_RESERVED_SEQ64_SIZE + \
152 						 AMDGPU_VA_RESERVED_CSA_SIZE)
153 
154 /* See vm_update_mode */
155 #define AMDGPU_VM_USE_CPU_FOR_GFX (1 << 0)
156 #define AMDGPU_VM_USE_CPU_FOR_COMPUTE (1 << 1)
157 
158 /* VMPT level enumerate, and the hiberachy is:
159  * PDB2->PDB1->PDB0->PTB
160  */
161 enum amdgpu_vm_level {
162 	AMDGPU_VM_PDB2,
163 	AMDGPU_VM_PDB1,
164 	AMDGPU_VM_PDB0,
165 	AMDGPU_VM_PTB
166 };
167 
168 /* base structure for tracking BO usage in a VM */
169 struct amdgpu_vm_bo_base {
170 	/* constant after initialization */
171 	struct amdgpu_vm		*vm;
172 	struct amdgpu_bo		*bo;
173 
174 	/* protected by bo being reserved */
175 	struct amdgpu_vm_bo_base	*next;
176 
177 	/* protected by spinlock */
178 	struct list_head		vm_status;
179 
180 	/* protected by the BO being reserved */
181 	bool				moved;
182 };
183 
184 /* provided by hw blocks that can write ptes, e.g., sdma */
185 struct amdgpu_vm_pte_funcs {
186 	/* number of dw to reserve per operation */
187 	unsigned	copy_pte_num_dw;
188 
189 	/* copy pte entries from GART */
190 	void (*copy_pte)(struct amdgpu_ib *ib,
191 			 uint64_t pe, uint64_t src,
192 			 unsigned count);
193 
194 	/* write pte one entry at a time with addr mapping */
195 	void (*write_pte)(struct amdgpu_ib *ib, uint64_t pe,
196 			  uint64_t value, unsigned count,
197 			  uint32_t incr);
198 	/* for linear pte/pde updates without addr mapping */
199 	void (*set_pte_pde)(struct amdgpu_ib *ib,
200 			    uint64_t pe,
201 			    uint64_t addr, unsigned count,
202 			    uint32_t incr, uint64_t flags);
203 };
204 
205 struct amdgpu_task_info {
206 	char	process_name[TASK_COMM_LEN];
207 	char	task_name[TASK_COMM_LEN];
208 	pid_t	pid;
209 	pid_t	tgid;
210 };
211 
212 /**
213  * struct amdgpu_vm_update_params
214  *
215  * Encapsulate some VM table update parameters to reduce
216  * the number of function parameters
217  *
218  */
219 struct amdgpu_vm_update_params {
220 
221 	/**
222 	 * @adev: amdgpu device we do this update for
223 	 */
224 	struct amdgpu_device *adev;
225 
226 	/**
227 	 * @vm: optional amdgpu_vm we do this update for
228 	 */
229 	struct amdgpu_vm *vm;
230 
231 	/**
232 	 * @immediate: if changes should be made immediately
233 	 */
234 	bool immediate;
235 
236 	/**
237 	 * @unlocked: true if the root BO is not locked
238 	 */
239 	bool unlocked;
240 
241 	/**
242 	 * @pages_addr:
243 	 *
244 	 * DMA addresses to use for mapping
245 	 */
246 	dma_addr_t *pages_addr;
247 
248 	/**
249 	 * @job: job to used for hw submission
250 	 */
251 	struct amdgpu_job *job;
252 
253 	/**
254 	 * @num_dw_left: number of dw left for the IB
255 	 */
256 	unsigned int num_dw_left;
257 
258 	/**
259 	 * @table_freed: return true if page table is freed when updating
260 	 */
261 	bool table_freed;
262 
263 	/**
264 	 * @allow_override: true for memory that is not uncached: allows MTYPE
265 	 * to be overridden for NUMA local memory.
266 	 */
267 	bool allow_override;
268 };
269 
270 struct amdgpu_vm_update_funcs {
271 	int (*map_table)(struct amdgpu_bo_vm *bo);
272 	int (*prepare)(struct amdgpu_vm_update_params *p, struct dma_resv *resv,
273 		       enum amdgpu_sync_mode sync_mode);
274 	int (*update)(struct amdgpu_vm_update_params *p,
275 		      struct amdgpu_bo_vm *bo, uint64_t pe, uint64_t addr,
276 		      unsigned count, uint32_t incr, uint64_t flags);
277 	int (*commit)(struct amdgpu_vm_update_params *p,
278 		      struct dma_fence **fence);
279 };
280 
281 struct amdgpu_vm_fault_info {
282 	/* fault address */
283 	uint64_t	addr;
284 	/* fault status register */
285 	uint32_t	status;
286 	/* which vmhub? gfxhub, mmhub, etc. */
287 	unsigned int	vmhub;
288 };
289 
290 struct amdgpu_vm {
291 	/* tree of virtual addresses mapped */
292 	struct rb_root_cached	va;
293 
294 	/* Lock to prevent eviction while we are updating page tables
295 	 * use vm_eviction_lock/unlock(vm)
296 	 */
297 	struct mutex		eviction_lock;
298 	bool			evicting;
299 	unsigned int		saved_flags;
300 
301 	/* Lock to protect vm_bo add/del/move on all lists of vm */
302 	spinlock_t		status_lock;
303 
304 	/* Per-VM and PT BOs who needs a validation */
305 	struct list_head	evicted;
306 
307 	/* BOs for user mode queues that need a validation */
308 	struct list_head	evicted_user;
309 
310 	/* PT BOs which relocated and their parent need an update */
311 	struct list_head	relocated;
312 
313 	/* per VM BOs moved, but not yet updated in the PT */
314 	struct list_head	moved;
315 
316 	/* All BOs of this VM not currently in the state machine */
317 	struct list_head	idle;
318 
319 	/* regular invalidated BOs, but not yet updated in the PT */
320 	struct list_head	invalidated;
321 
322 	/* BO mappings freed, but not yet updated in the PT */
323 	struct list_head	freed;
324 
325 	/* BOs which are invalidated, has been updated in the PTs */
326 	struct list_head        done;
327 
328 	/* PT BOs scheduled to free and fill with zero if vm_resv is not hold */
329 	struct list_head	pt_freed;
330 	struct work_struct	pt_free_work;
331 
332 	/* contains the page directory */
333 	struct amdgpu_vm_bo_base     root;
334 	struct dma_fence	*last_update;
335 
336 	/* Scheduler entities for page table updates */
337 	struct drm_sched_entity	immediate;
338 	struct drm_sched_entity	delayed;
339 
340 	/* Last finished delayed update */
341 	atomic64_t		tlb_seq;
342 	struct dma_fence	*last_tlb_flush;
343 	atomic64_t		kfd_last_flushed_seq;
344 
345 	/* How many times we had to re-generate the page tables */
346 	uint64_t		generation;
347 
348 	/* Last unlocked submission to the scheduler entities */
349 	struct dma_fence	*last_unlocked;
350 
351 	unsigned int		pasid;
352 	bool			reserved_vmid[AMDGPU_MAX_VMHUBS];
353 
354 	/* Flag to indicate if VM tables are updated by CPU or GPU (SDMA) */
355 	bool					use_cpu_for_update;
356 
357 	/* Functions to use for VM table updates */
358 	const struct amdgpu_vm_update_funcs	*update_funcs;
359 
360 	/* Flag to indicate ATS support from PTE for GFX9 */
361 	bool			pte_support_ats;
362 
363 	/* Up to 128 pending retry page faults */
364 	DECLARE_KFIFO(faults, u64, 128);
365 
366 	/* Points to the KFD process VM info */
367 	struct amdkfd_process_info *process_info;
368 
369 	/* List node in amdkfd_process_info.vm_list_head */
370 	struct list_head	vm_list_node;
371 
372 	/* Valid while the PD is reserved or fenced */
373 	uint64_t		pd_phys_addr;
374 
375 	/* Some basic info about the task */
376 	struct amdgpu_task_info task_info;
377 
378 	/* Store positions of group of BOs */
379 	struct ttm_lru_bulk_move lru_bulk_move;
380 	/* Flag to indicate if VM is used for compute */
381 	bool			is_compute_context;
382 
383 	/* Memory partition number, -1 means any partition */
384 	int8_t			mem_id;
385 
386 	/* cached fault info */
387 	struct amdgpu_vm_fault_info fault_info;
388 };
389 
390 struct amdgpu_vm_manager {
391 	/* Handling of VMIDs */
392 	struct amdgpu_vmid_mgr			id_mgr[AMDGPU_MAX_VMHUBS];
393 	unsigned int				first_kfd_vmid;
394 	bool					concurrent_flush;
395 
396 	/* Handling of VM fences */
397 	u64					fence_context;
398 	unsigned				seqno[AMDGPU_MAX_RINGS];
399 
400 	uint64_t				max_pfn;
401 	uint32_t				num_level;
402 	uint32_t				block_size;
403 	uint32_t				fragment_size;
404 	enum amdgpu_vm_level			root_level;
405 	/* vram base address for page table entry  */
406 	u64					vram_base_offset;
407 	/* vm pte handling */
408 	const struct amdgpu_vm_pte_funcs	*vm_pte_funcs;
409 	struct drm_gpu_scheduler		*vm_pte_scheds[AMDGPU_MAX_RINGS];
410 	unsigned				vm_pte_num_scheds;
411 	struct amdgpu_ring			*page_fault;
412 
413 	/* partial resident texture handling */
414 	spinlock_t				prt_lock;
415 	atomic_t				num_prt_users;
416 
417 	/* controls how VM page tables are updated for Graphics and Compute.
418 	 * BIT0[= 0] Graphics updated by SDMA [= 1] by CPU
419 	 * BIT1[= 0] Compute updated by SDMA [= 1] by CPU
420 	 */
421 	int					vm_update_mode;
422 
423 	/* PASID to VM mapping, will be used in interrupt context to
424 	 * look up VM of a page fault
425 	 */
426 	struct xarray				pasids;
427 };
428 
429 struct amdgpu_bo_va_mapping;
430 
431 #define amdgpu_vm_copy_pte(adev, ib, pe, src, count) ((adev)->vm_manager.vm_pte_funcs->copy_pte((ib), (pe), (src), (count)))
432 #define amdgpu_vm_write_pte(adev, ib, pe, value, count, incr) ((adev)->vm_manager.vm_pte_funcs->write_pte((ib), (pe), (value), (count), (incr)))
433 #define amdgpu_vm_set_pte_pde(adev, ib, pe, addr, count, incr, flags) ((adev)->vm_manager.vm_pte_funcs->set_pte_pde((ib), (pe), (addr), (count), (incr), (flags)))
434 
435 extern const struct amdgpu_vm_update_funcs amdgpu_vm_cpu_funcs;
436 extern const struct amdgpu_vm_update_funcs amdgpu_vm_sdma_funcs;
437 
438 void amdgpu_vm_manager_init(struct amdgpu_device *adev);
439 void amdgpu_vm_manager_fini(struct amdgpu_device *adev);
440 
441 int amdgpu_vm_set_pasid(struct amdgpu_device *adev, struct amdgpu_vm *vm,
442 			u32 pasid);
443 
444 long amdgpu_vm_wait_idle(struct amdgpu_vm *vm, long timeout);
445 int amdgpu_vm_init(struct amdgpu_device *adev, struct amdgpu_vm *vm, int32_t xcp_id);
446 int amdgpu_vm_make_compute(struct amdgpu_device *adev, struct amdgpu_vm *vm);
447 void amdgpu_vm_release_compute(struct amdgpu_device *adev, struct amdgpu_vm *vm);
448 void amdgpu_vm_fini(struct amdgpu_device *adev, struct amdgpu_vm *vm);
449 int amdgpu_vm_lock_pd(struct amdgpu_vm *vm, struct drm_exec *exec,
450 		      unsigned int num_fences);
451 bool amdgpu_vm_ready(struct amdgpu_vm *vm);
452 uint64_t amdgpu_vm_generation(struct amdgpu_device *adev, struct amdgpu_vm *vm);
453 int amdgpu_vm_validate(struct amdgpu_device *adev, struct amdgpu_vm *vm,
454 		       struct ww_acquire_ctx *ticket,
455 		       int (*callback)(void *p, struct amdgpu_bo *bo),
456 		       void *param);
457 int amdgpu_vm_flush(struct amdgpu_ring *ring, struct amdgpu_job *job, bool need_pipe_sync);
458 int amdgpu_vm_update_pdes(struct amdgpu_device *adev,
459 			  struct amdgpu_vm *vm, bool immediate);
460 int amdgpu_vm_clear_freed(struct amdgpu_device *adev,
461 			  struct amdgpu_vm *vm,
462 			  struct dma_fence **fence);
463 int amdgpu_vm_handle_moved(struct amdgpu_device *adev,
464 			   struct amdgpu_vm *vm,
465 			   struct ww_acquire_ctx *ticket);
466 int amdgpu_vm_flush_compute_tlb(struct amdgpu_device *adev,
467 				struct amdgpu_vm *vm,
468 				uint32_t flush_type,
469 				uint32_t xcc_mask);
470 void amdgpu_vm_bo_base_init(struct amdgpu_vm_bo_base *base,
471 			    struct amdgpu_vm *vm, struct amdgpu_bo *bo);
472 int amdgpu_vm_update_range(struct amdgpu_device *adev, struct amdgpu_vm *vm,
473 			   bool immediate, bool unlocked, bool flush_tlb, bool allow_override,
474 			   struct dma_resv *resv, uint64_t start, uint64_t last,
475 			   uint64_t flags, uint64_t offset, uint64_t vram_base,
476 			   struct ttm_resource *res, dma_addr_t *pages_addr,
477 			   struct dma_fence **fence);
478 int amdgpu_vm_bo_update(struct amdgpu_device *adev,
479 			struct amdgpu_bo_va *bo_va,
480 			bool clear);
481 bool amdgpu_vm_evictable(struct amdgpu_bo *bo);
482 void amdgpu_vm_bo_invalidate(struct amdgpu_device *adev,
483 			     struct amdgpu_bo *bo, bool evicted);
484 uint64_t amdgpu_vm_map_gart(const dma_addr_t *pages_addr, uint64_t addr);
485 struct amdgpu_bo_va *amdgpu_vm_bo_find(struct amdgpu_vm *vm,
486 				       struct amdgpu_bo *bo);
487 struct amdgpu_bo_va *amdgpu_vm_bo_add(struct amdgpu_device *adev,
488 				      struct amdgpu_vm *vm,
489 				      struct amdgpu_bo *bo);
490 int amdgpu_vm_bo_map(struct amdgpu_device *adev,
491 		     struct amdgpu_bo_va *bo_va,
492 		     uint64_t addr, uint64_t offset,
493 		     uint64_t size, uint64_t flags);
494 int amdgpu_vm_bo_replace_map(struct amdgpu_device *adev,
495 			     struct amdgpu_bo_va *bo_va,
496 			     uint64_t addr, uint64_t offset,
497 			     uint64_t size, uint64_t flags);
498 int amdgpu_vm_bo_unmap(struct amdgpu_device *adev,
499 		       struct amdgpu_bo_va *bo_va,
500 		       uint64_t addr);
501 int amdgpu_vm_bo_clear_mappings(struct amdgpu_device *adev,
502 				struct amdgpu_vm *vm,
503 				uint64_t saddr, uint64_t size);
504 struct amdgpu_bo_va_mapping *amdgpu_vm_bo_lookup_mapping(struct amdgpu_vm *vm,
505 							 uint64_t addr);
506 void amdgpu_vm_bo_trace_cs(struct amdgpu_vm *vm, struct ww_acquire_ctx *ticket);
507 void amdgpu_vm_bo_del(struct amdgpu_device *adev,
508 		      struct amdgpu_bo_va *bo_va);
509 void amdgpu_vm_adjust_size(struct amdgpu_device *adev, uint32_t min_vm_size,
510 			   uint32_t fragment_size_default, unsigned max_level,
511 			   unsigned max_bits);
512 int amdgpu_vm_ioctl(struct drm_device *dev, void *data, struct drm_file *filp);
513 bool amdgpu_vm_need_pipeline_sync(struct amdgpu_ring *ring,
514 				  struct amdgpu_job *job);
515 void amdgpu_vm_check_compute_bug(struct amdgpu_device *adev);
516 
517 void amdgpu_vm_get_task_info(struct amdgpu_device *adev, u32 pasid,
518 			     struct amdgpu_task_info *task_info);
519 bool amdgpu_vm_handle_fault(struct amdgpu_device *adev, u32 pasid,
520 			    u32 vmid, u32 node_id, uint64_t addr,
521 			    bool write_fault);
522 
523 void amdgpu_vm_set_task_info(struct amdgpu_vm *vm);
524 
525 void amdgpu_vm_move_to_lru_tail(struct amdgpu_device *adev,
526 				struct amdgpu_vm *vm);
527 void amdgpu_vm_get_memory(struct amdgpu_vm *vm,
528 			  struct amdgpu_mem_stats *stats);
529 
530 int amdgpu_vm_pt_clear(struct amdgpu_device *adev, struct amdgpu_vm *vm,
531 		       struct amdgpu_bo_vm *vmbo, bool immediate);
532 int amdgpu_vm_pt_create(struct amdgpu_device *adev, struct amdgpu_vm *vm,
533 			int level, bool immediate, struct amdgpu_bo_vm **vmbo,
534 			int32_t xcp_id);
535 void amdgpu_vm_pt_free_root(struct amdgpu_device *adev, struct amdgpu_vm *vm);
536 bool amdgpu_vm_pt_is_root_clean(struct amdgpu_device *adev,
537 				struct amdgpu_vm *vm);
538 
539 int amdgpu_vm_pde_update(struct amdgpu_vm_update_params *params,
540 			 struct amdgpu_vm_bo_base *entry);
541 int amdgpu_vm_ptes_update(struct amdgpu_vm_update_params *params,
542 			  uint64_t start, uint64_t end,
543 			  uint64_t dst, uint64_t flags);
544 void amdgpu_vm_pt_free_work(struct work_struct *work);
545 
546 #if defined(CONFIG_DEBUG_FS)
547 void amdgpu_debugfs_vm_bo_info(struct amdgpu_vm *vm, struct seq_file *m);
548 #endif
549 
550 int amdgpu_vm_pt_map_tables(struct amdgpu_device *adev, struct amdgpu_vm *vm);
551 
552 /**
553  * amdgpu_vm_tlb_seq - return tlb flush sequence number
554  * @vm: the amdgpu_vm structure to query
555  *
556  * Returns the tlb flush sequence number which indicates that the VM TLBs needs
557  * to be invalidated whenever the sequence number change.
558  */
559 static inline uint64_t amdgpu_vm_tlb_seq(struct amdgpu_vm *vm)
560 {
561 	unsigned long flags;
562 	spinlock_t *lock;
563 
564 	/*
565 	 * Workaround to stop racing between the fence signaling and handling
566 	 * the cb. The lock is static after initially setting it up, just make
567 	 * sure that the dma_fence structure isn't freed up.
568 	 */
569 	rcu_read_lock();
570 	lock = vm->last_tlb_flush->lock;
571 	rcu_read_unlock();
572 
573 	spin_lock_irqsave(lock, flags);
574 	spin_unlock_irqrestore(lock, flags);
575 
576 	return atomic64_read(&vm->tlb_seq);
577 }
578 
579 /*
580  * vm eviction_lock can be taken in MMU notifiers. Make sure no reclaim-FS
581  * happens while holding this lock anywhere to prevent deadlocks when
582  * an MMU notifier runs in reclaim-FS context.
583  */
584 static inline void amdgpu_vm_eviction_lock(struct amdgpu_vm *vm)
585 {
586 	mutex_lock(&vm->eviction_lock);
587 	vm->saved_flags = memalloc_noreclaim_save();
588 }
589 
590 static inline bool amdgpu_vm_eviction_trylock(struct amdgpu_vm *vm)
591 {
592 	if (mutex_trylock(&vm->eviction_lock)) {
593 		vm->saved_flags = memalloc_noreclaim_save();
594 		return true;
595 	}
596 	return false;
597 }
598 
599 static inline void amdgpu_vm_eviction_unlock(struct amdgpu_vm *vm)
600 {
601 	memalloc_noreclaim_restore(vm->saved_flags);
602 	mutex_unlock(&vm->eviction_lock);
603 }
604 
605 void amdgpu_vm_update_fault_cache(struct amdgpu_device *adev,
606 				  unsigned int pasid,
607 				  uint64_t addr,
608 				  uint32_t status,
609 				  unsigned int vmhub);
610 
611 #endif
612