xref: /linux/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.h (revision f9d48a88eba547d87d8ea7bfd25abd1836cdf06a)
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 /* gfx12 */
115 #define AMDGPU_PTE_PRT_GFX12		(1ULL << 56)
116 #define AMDGPU_PTE_PRT_FLAG(adev)	\
117 	((amdgpu_ip_version((adev), GC_HWIP, 0) >= IP_VERSION(12, 0, 0)) ? AMDGPU_PTE_PRT_GFX12 : AMDGPU_PTE_PRT)
118 
119 #define AMDGPU_PTE_MTYPE_GFX12(a)	((uint64_t)(a) << 54)
120 #define AMDGPU_PTE_MTYPE_GFX12_MASK	AMDGPU_PTE_MTYPE_GFX12(3ULL)
121 
122 #define AMDGPU_PTE_IS_PTE		(1ULL << 63)
123 
124 /* PDE Block Fragment Size for gfx v12 */
125 #define AMDGPU_PDE_BFS_GFX12(a)		((uint64_t)((a) & 0x1fULL) << 58)
126 #define AMDGPU_PDE_BFS_FLAG(adev, a)	\
127 	((amdgpu_ip_version((adev), GC_HWIP, 0) >= IP_VERSION(12, 0, 0)) ? AMDGPU_PDE_BFS_GFX12(a) : AMDGPU_PDE_BFS(a))
128 /* PDE is handled as PTE for gfx v12 */
129 #define AMDGPU_PDE_PTE_GFX12		(1ULL << 63)
130 #define AMDGPU_PDE_PTE_FLAG(adev)	\
131 	((amdgpu_ip_version((adev), GC_HWIP, 0) >= IP_VERSION(12, 0, 0)) ? AMDGPU_PDE_PTE_GFX12 : AMDGPU_PDE_PTE)
132 
133 /* How to program VM fault handling */
134 #define AMDGPU_VM_FAULT_STOP_NEVER	0
135 #define AMDGPU_VM_FAULT_STOP_FIRST	1
136 #define AMDGPU_VM_FAULT_STOP_ALWAYS	2
137 
138 /* How much VRAM be reserved for page tables */
139 #define AMDGPU_VM_RESERVED_VRAM		(8ULL << 20)
140 
141 /*
142  * max number of VMHUB
143  * layout: max 8 GFXHUB + 4 MMHUB0 + 1 MMHUB1
144  */
145 #define AMDGPU_MAX_VMHUBS			13
146 #define AMDGPU_GFXHUB_START			0
147 #define AMDGPU_MMHUB0_START			8
148 #define AMDGPU_MMHUB1_START			12
149 #define AMDGPU_GFXHUB(x)			(AMDGPU_GFXHUB_START + (x))
150 #define AMDGPU_MMHUB0(x)			(AMDGPU_MMHUB0_START + (x))
151 #define AMDGPU_MMHUB1(x)			(AMDGPU_MMHUB1_START + (x))
152 
153 #define AMDGPU_IS_GFXHUB(x) ((x) >= AMDGPU_GFXHUB_START && (x) < AMDGPU_MMHUB0_START)
154 #define AMDGPU_IS_MMHUB0(x) ((x) >= AMDGPU_MMHUB0_START && (x) < AMDGPU_MMHUB1_START)
155 #define AMDGPU_IS_MMHUB1(x) ((x) >= AMDGPU_MMHUB1_START && (x) < AMDGPU_MAX_VMHUBS)
156 
157 /* Reserve space at top/bottom of address space for kernel use */
158 #define AMDGPU_VA_RESERVED_CSA_SIZE		(2ULL << 20)
159 #define AMDGPU_VA_RESERVED_CSA_START(adev)	(((adev)->vm_manager.max_pfn \
160 						  << AMDGPU_GPU_PAGE_SHIFT)  \
161 						 - AMDGPU_VA_RESERVED_CSA_SIZE)
162 #define AMDGPU_VA_RESERVED_SEQ64_SIZE		(2ULL << 20)
163 #define AMDGPU_VA_RESERVED_SEQ64_START(adev)	(AMDGPU_VA_RESERVED_CSA_START(adev) \
164 						 - AMDGPU_VA_RESERVED_SEQ64_SIZE)
165 #define AMDGPU_VA_RESERVED_TRAP_SIZE		(2ULL << 12)
166 #define AMDGPU_VA_RESERVED_TRAP_START(adev)	(AMDGPU_VA_RESERVED_SEQ64_START(adev) \
167 						 - AMDGPU_VA_RESERVED_TRAP_SIZE)
168 #define AMDGPU_VA_RESERVED_BOTTOM		(1ULL << 16)
169 #define AMDGPU_VA_RESERVED_TOP			(AMDGPU_VA_RESERVED_TRAP_SIZE + \
170 						 AMDGPU_VA_RESERVED_SEQ64_SIZE + \
171 						 AMDGPU_VA_RESERVED_CSA_SIZE)
172 
173 /* See vm_update_mode */
174 #define AMDGPU_VM_USE_CPU_FOR_GFX (1 << 0)
175 #define AMDGPU_VM_USE_CPU_FOR_COMPUTE (1 << 1)
176 
177 /* VMPT level enumerate, and the hiberachy is:
178  * PDB2->PDB1->PDB0->PTB
179  */
180 enum amdgpu_vm_level {
181 	AMDGPU_VM_PDB2,
182 	AMDGPU_VM_PDB1,
183 	AMDGPU_VM_PDB0,
184 	AMDGPU_VM_PTB
185 };
186 
187 /* base structure for tracking BO usage in a VM */
188 struct amdgpu_vm_bo_base {
189 	/* constant after initialization */
190 	struct amdgpu_vm		*vm;
191 	struct amdgpu_bo		*bo;
192 
193 	/* protected by bo being reserved */
194 	struct amdgpu_vm_bo_base	*next;
195 
196 	/* protected by spinlock */
197 	struct list_head		vm_status;
198 
199 	/* protected by the BO being reserved */
200 	bool				moved;
201 };
202 
203 /* provided by hw blocks that can write ptes, e.g., sdma */
204 struct amdgpu_vm_pte_funcs {
205 	/* number of dw to reserve per operation */
206 	unsigned	copy_pte_num_dw;
207 
208 	/* copy pte entries from GART */
209 	void (*copy_pte)(struct amdgpu_ib *ib,
210 			 uint64_t pe, uint64_t src,
211 			 unsigned count);
212 
213 	/* write pte one entry at a time with addr mapping */
214 	void (*write_pte)(struct amdgpu_ib *ib, uint64_t pe,
215 			  uint64_t value, unsigned count,
216 			  uint32_t incr);
217 	/* for linear pte/pde updates without addr mapping */
218 	void (*set_pte_pde)(struct amdgpu_ib *ib,
219 			    uint64_t pe,
220 			    uint64_t addr, unsigned count,
221 			    uint32_t incr, uint64_t flags);
222 };
223 
224 struct amdgpu_task_info {
225 	char		process_name[TASK_COMM_LEN];
226 	char		task_name[TASK_COMM_LEN];
227 	pid_t		pid;
228 	pid_t		tgid;
229 	struct kref	refcount;
230 };
231 
232 /**
233  * struct amdgpu_vm_update_params
234  *
235  * Encapsulate some VM table update parameters to reduce
236  * the number of function parameters
237  *
238  */
239 struct amdgpu_vm_update_params {
240 
241 	/**
242 	 * @adev: amdgpu device we do this update for
243 	 */
244 	struct amdgpu_device *adev;
245 
246 	/**
247 	 * @vm: optional amdgpu_vm we do this update for
248 	 */
249 	struct amdgpu_vm *vm;
250 
251 	/**
252 	 * @immediate: if changes should be made immediately
253 	 */
254 	bool immediate;
255 
256 	/**
257 	 * @unlocked: true if the root BO is not locked
258 	 */
259 	bool unlocked;
260 
261 	/**
262 	 * @pages_addr:
263 	 *
264 	 * DMA addresses to use for mapping
265 	 */
266 	dma_addr_t *pages_addr;
267 
268 	/**
269 	 * @job: job to used for hw submission
270 	 */
271 	struct amdgpu_job *job;
272 
273 	/**
274 	 * @num_dw_left: number of dw left for the IB
275 	 */
276 	unsigned int num_dw_left;
277 
278 	/**
279 	 * @needs_flush: true whenever we need to invalidate the TLB
280 	 */
281 	bool needs_flush;
282 
283 	/**
284 	 * @allow_override: true for memory that is not uncached: allows MTYPE
285 	 * to be overridden for NUMA local memory.
286 	 */
287 	bool allow_override;
288 
289 	/**
290 	 * @tlb_flush_waitlist: temporary storage for BOs until tlb_flush
291 	 */
292 	struct list_head tlb_flush_waitlist;
293 };
294 
295 struct amdgpu_vm_update_funcs {
296 	int (*map_table)(struct amdgpu_bo_vm *bo);
297 	int (*prepare)(struct amdgpu_vm_update_params *p, struct dma_resv *resv,
298 		       enum amdgpu_sync_mode sync_mode);
299 	int (*update)(struct amdgpu_vm_update_params *p,
300 		      struct amdgpu_bo_vm *bo, uint64_t pe, uint64_t addr,
301 		      unsigned count, uint32_t incr, uint64_t flags);
302 	int (*commit)(struct amdgpu_vm_update_params *p,
303 		      struct dma_fence **fence);
304 };
305 
306 struct amdgpu_vm_fault_info {
307 	/* fault address */
308 	uint64_t	addr;
309 	/* fault status register */
310 	uint32_t	status;
311 	/* which vmhub? gfxhub, mmhub, etc. */
312 	unsigned int	vmhub;
313 };
314 
315 struct amdgpu_vm {
316 	/* tree of virtual addresses mapped */
317 	struct rb_root_cached	va;
318 
319 	/* Lock to prevent eviction while we are updating page tables
320 	 * use vm_eviction_lock/unlock(vm)
321 	 */
322 	struct mutex		eviction_lock;
323 	bool			evicting;
324 	unsigned int		saved_flags;
325 
326 	/* Lock to protect vm_bo add/del/move on all lists of vm */
327 	spinlock_t		status_lock;
328 
329 	/* Per-VM and PT BOs who needs a validation */
330 	struct list_head	evicted;
331 
332 	/* BOs for user mode queues that need a validation */
333 	struct list_head	evicted_user;
334 
335 	/* PT BOs which relocated and their parent need an update */
336 	struct list_head	relocated;
337 
338 	/* per VM BOs moved, but not yet updated in the PT */
339 	struct list_head	moved;
340 
341 	/* All BOs of this VM not currently in the state machine */
342 	struct list_head	idle;
343 
344 	/* regular invalidated BOs, but not yet updated in the PT */
345 	struct list_head	invalidated;
346 
347 	/* BO mappings freed, but not yet updated in the PT */
348 	struct list_head	freed;
349 
350 	/* BOs which are invalidated, has been updated in the PTs */
351 	struct list_head        done;
352 
353 	/* PT BOs scheduled to free and fill with zero if vm_resv is not hold */
354 	struct list_head	pt_freed;
355 	struct work_struct	pt_free_work;
356 
357 	/* contains the page directory */
358 	struct amdgpu_vm_bo_base     root;
359 	struct dma_fence	*last_update;
360 
361 	/* Scheduler entities for page table updates */
362 	struct drm_sched_entity	immediate;
363 	struct drm_sched_entity	delayed;
364 
365 	/* Last finished delayed update */
366 	atomic64_t		tlb_seq;
367 	struct dma_fence	*last_tlb_flush;
368 	atomic64_t		kfd_last_flushed_seq;
369 	uint64_t		tlb_fence_context;
370 
371 	/* How many times we had to re-generate the page tables */
372 	uint64_t		generation;
373 
374 	/* Last unlocked submission to the scheduler entities */
375 	struct dma_fence	*last_unlocked;
376 
377 	unsigned int		pasid;
378 	bool			reserved_vmid[AMDGPU_MAX_VMHUBS];
379 
380 	/* Flag to indicate if VM tables are updated by CPU or GPU (SDMA) */
381 	bool					use_cpu_for_update;
382 
383 	/* Functions to use for VM table updates */
384 	const struct amdgpu_vm_update_funcs	*update_funcs;
385 
386 	/* Up to 128 pending retry page faults */
387 	DECLARE_KFIFO(faults, u64, 128);
388 
389 	/* Points to the KFD process VM info */
390 	struct amdkfd_process_info *process_info;
391 
392 	/* List node in amdkfd_process_info.vm_list_head */
393 	struct list_head	vm_list_node;
394 
395 	/* Valid while the PD is reserved or fenced */
396 	uint64_t		pd_phys_addr;
397 
398 	/* Some basic info about the task */
399 	struct amdgpu_task_info *task_info;
400 
401 	/* Store positions of group of BOs */
402 	struct ttm_lru_bulk_move lru_bulk_move;
403 	/* Flag to indicate if VM is used for compute */
404 	bool			is_compute_context;
405 
406 	/* Memory partition number, -1 means any partition */
407 	int8_t			mem_id;
408 
409 	/* cached fault info */
410 	struct amdgpu_vm_fault_info fault_info;
411 };
412 
413 struct amdgpu_vm_manager {
414 	/* Handling of VMIDs */
415 	struct amdgpu_vmid_mgr			id_mgr[AMDGPU_MAX_VMHUBS];
416 	unsigned int				first_kfd_vmid;
417 	bool					concurrent_flush;
418 
419 	/* Handling of VM fences */
420 	u64					fence_context;
421 	unsigned				seqno[AMDGPU_MAX_RINGS];
422 
423 	uint64_t				max_pfn;
424 	uint32_t				num_level;
425 	uint32_t				block_size;
426 	uint32_t				fragment_size;
427 	enum amdgpu_vm_level			root_level;
428 	/* vram base address for page table entry  */
429 	u64					vram_base_offset;
430 	/* vm pte handling */
431 	const struct amdgpu_vm_pte_funcs	*vm_pte_funcs;
432 	struct drm_gpu_scheduler		*vm_pte_scheds[AMDGPU_MAX_RINGS];
433 	unsigned				vm_pte_num_scheds;
434 	struct amdgpu_ring			*page_fault;
435 
436 	/* partial resident texture handling */
437 	spinlock_t				prt_lock;
438 	atomic_t				num_prt_users;
439 
440 	/* controls how VM page tables are updated for Graphics and Compute.
441 	 * BIT0[= 0] Graphics updated by SDMA [= 1] by CPU
442 	 * BIT1[= 0] Compute updated by SDMA [= 1] by CPU
443 	 */
444 	int					vm_update_mode;
445 
446 	/* PASID to VM mapping, will be used in interrupt context to
447 	 * look up VM of a page fault
448 	 */
449 	struct xarray				pasids;
450 	/* Global registration of recent page fault information */
451 	struct amdgpu_vm_fault_info	fault_info;
452 };
453 
454 struct amdgpu_bo_va_mapping;
455 
456 #define amdgpu_vm_copy_pte(adev, ib, pe, src, count) ((adev)->vm_manager.vm_pte_funcs->copy_pte((ib), (pe), (src), (count)))
457 #define amdgpu_vm_write_pte(adev, ib, pe, value, count, incr) ((adev)->vm_manager.vm_pte_funcs->write_pte((ib), (pe), (value), (count), (incr)))
458 #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)))
459 
460 extern const struct amdgpu_vm_update_funcs amdgpu_vm_cpu_funcs;
461 extern const struct amdgpu_vm_update_funcs amdgpu_vm_sdma_funcs;
462 
463 void amdgpu_vm_manager_init(struct amdgpu_device *adev);
464 void amdgpu_vm_manager_fini(struct amdgpu_device *adev);
465 
466 int amdgpu_vm_set_pasid(struct amdgpu_device *adev, struct amdgpu_vm *vm,
467 			u32 pasid);
468 
469 long amdgpu_vm_wait_idle(struct amdgpu_vm *vm, long timeout);
470 int amdgpu_vm_init(struct amdgpu_device *adev, struct amdgpu_vm *vm, int32_t xcp_id);
471 int amdgpu_vm_make_compute(struct amdgpu_device *adev, struct amdgpu_vm *vm);
472 void amdgpu_vm_release_compute(struct amdgpu_device *adev, struct amdgpu_vm *vm);
473 void amdgpu_vm_fini(struct amdgpu_device *adev, struct amdgpu_vm *vm);
474 int amdgpu_vm_lock_pd(struct amdgpu_vm *vm, struct drm_exec *exec,
475 		      unsigned int num_fences);
476 bool amdgpu_vm_ready(struct amdgpu_vm *vm);
477 uint64_t amdgpu_vm_generation(struct amdgpu_device *adev, struct amdgpu_vm *vm);
478 int amdgpu_vm_validate(struct amdgpu_device *adev, struct amdgpu_vm *vm,
479 		       struct ww_acquire_ctx *ticket,
480 		       int (*callback)(void *p, struct amdgpu_bo *bo),
481 		       void *param);
482 int amdgpu_vm_flush(struct amdgpu_ring *ring, struct amdgpu_job *job, bool need_pipe_sync);
483 int amdgpu_vm_update_pdes(struct amdgpu_device *adev,
484 			  struct amdgpu_vm *vm, bool immediate);
485 int amdgpu_vm_clear_freed(struct amdgpu_device *adev,
486 			  struct amdgpu_vm *vm,
487 			  struct dma_fence **fence);
488 int amdgpu_vm_handle_moved(struct amdgpu_device *adev,
489 			   struct amdgpu_vm *vm,
490 			   struct ww_acquire_ctx *ticket);
491 int amdgpu_vm_flush_compute_tlb(struct amdgpu_device *adev,
492 				struct amdgpu_vm *vm,
493 				uint32_t flush_type,
494 				uint32_t xcc_mask);
495 void amdgpu_vm_bo_base_init(struct amdgpu_vm_bo_base *base,
496 			    struct amdgpu_vm *vm, struct amdgpu_bo *bo);
497 int amdgpu_vm_update_range(struct amdgpu_device *adev, struct amdgpu_vm *vm,
498 			   bool immediate, bool unlocked, bool flush_tlb, bool allow_override,
499 			   struct dma_resv *resv, uint64_t start, uint64_t last,
500 			   uint64_t flags, uint64_t offset, uint64_t vram_base,
501 			   struct ttm_resource *res, dma_addr_t *pages_addr,
502 			   struct dma_fence **fence);
503 int amdgpu_vm_bo_update(struct amdgpu_device *adev,
504 			struct amdgpu_bo_va *bo_va,
505 			bool clear);
506 bool amdgpu_vm_evictable(struct amdgpu_bo *bo);
507 void amdgpu_vm_bo_invalidate(struct amdgpu_device *adev,
508 			     struct amdgpu_bo *bo, bool evicted);
509 uint64_t amdgpu_vm_map_gart(const dma_addr_t *pages_addr, uint64_t addr);
510 struct amdgpu_bo_va *amdgpu_vm_bo_find(struct amdgpu_vm *vm,
511 				       struct amdgpu_bo *bo);
512 struct amdgpu_bo_va *amdgpu_vm_bo_add(struct amdgpu_device *adev,
513 				      struct amdgpu_vm *vm,
514 				      struct amdgpu_bo *bo);
515 int amdgpu_vm_bo_map(struct amdgpu_device *adev,
516 		     struct amdgpu_bo_va *bo_va,
517 		     uint64_t addr, uint64_t offset,
518 		     uint64_t size, uint64_t flags);
519 int amdgpu_vm_bo_replace_map(struct amdgpu_device *adev,
520 			     struct amdgpu_bo_va *bo_va,
521 			     uint64_t addr, uint64_t offset,
522 			     uint64_t size, uint64_t flags);
523 int amdgpu_vm_bo_unmap(struct amdgpu_device *adev,
524 		       struct amdgpu_bo_va *bo_va,
525 		       uint64_t addr);
526 int amdgpu_vm_bo_clear_mappings(struct amdgpu_device *adev,
527 				struct amdgpu_vm *vm,
528 				uint64_t saddr, uint64_t size);
529 struct amdgpu_bo_va_mapping *amdgpu_vm_bo_lookup_mapping(struct amdgpu_vm *vm,
530 							 uint64_t addr);
531 void amdgpu_vm_bo_trace_cs(struct amdgpu_vm *vm, struct ww_acquire_ctx *ticket);
532 void amdgpu_vm_bo_del(struct amdgpu_device *adev,
533 		      struct amdgpu_bo_va *bo_va);
534 void amdgpu_vm_adjust_size(struct amdgpu_device *adev, uint32_t min_vm_size,
535 			   uint32_t fragment_size_default, unsigned max_level,
536 			   unsigned max_bits);
537 int amdgpu_vm_ioctl(struct drm_device *dev, void *data, struct drm_file *filp);
538 bool amdgpu_vm_need_pipeline_sync(struct amdgpu_ring *ring,
539 				  struct amdgpu_job *job);
540 void amdgpu_vm_check_compute_bug(struct amdgpu_device *adev);
541 
542 struct amdgpu_task_info *
543 amdgpu_vm_get_task_info_pasid(struct amdgpu_device *adev, u32 pasid);
544 
545 struct amdgpu_task_info *
546 amdgpu_vm_get_task_info_vm(struct amdgpu_vm *vm);
547 
548 void amdgpu_vm_put_task_info(struct amdgpu_task_info *task_info);
549 
550 bool amdgpu_vm_handle_fault(struct amdgpu_device *adev, u32 pasid,
551 			    u32 vmid, u32 node_id, uint64_t addr,
552 			    bool write_fault);
553 
554 void amdgpu_vm_set_task_info(struct amdgpu_vm *vm);
555 
556 void amdgpu_vm_move_to_lru_tail(struct amdgpu_device *adev,
557 				struct amdgpu_vm *vm);
558 void amdgpu_vm_get_memory(struct amdgpu_vm *vm,
559 			  struct amdgpu_mem_stats *stats);
560 
561 int amdgpu_vm_pt_clear(struct amdgpu_device *adev, struct amdgpu_vm *vm,
562 		       struct amdgpu_bo_vm *vmbo, bool immediate);
563 int amdgpu_vm_pt_create(struct amdgpu_device *adev, struct amdgpu_vm *vm,
564 			int level, bool immediate, struct amdgpu_bo_vm **vmbo,
565 			int32_t xcp_id);
566 void amdgpu_vm_pt_free_root(struct amdgpu_device *adev, struct amdgpu_vm *vm);
567 
568 int amdgpu_vm_pde_update(struct amdgpu_vm_update_params *params,
569 			 struct amdgpu_vm_bo_base *entry);
570 int amdgpu_vm_ptes_update(struct amdgpu_vm_update_params *params,
571 			  uint64_t start, uint64_t end,
572 			  uint64_t dst, uint64_t flags);
573 void amdgpu_vm_pt_free_work(struct work_struct *work);
574 void amdgpu_vm_pt_free_list(struct amdgpu_device *adev,
575 			    struct amdgpu_vm_update_params *params);
576 
577 #if defined(CONFIG_DEBUG_FS)
578 void amdgpu_debugfs_vm_bo_info(struct amdgpu_vm *vm, struct seq_file *m);
579 #endif
580 
581 int amdgpu_vm_pt_map_tables(struct amdgpu_device *adev, struct amdgpu_vm *vm);
582 
583 /**
584  * amdgpu_vm_tlb_seq - return tlb flush sequence number
585  * @vm: the amdgpu_vm structure to query
586  *
587  * Returns the tlb flush sequence number which indicates that the VM TLBs needs
588  * to be invalidated whenever the sequence number change.
589  */
590 static inline uint64_t amdgpu_vm_tlb_seq(struct amdgpu_vm *vm)
591 {
592 	unsigned long flags;
593 	spinlock_t *lock;
594 
595 	/*
596 	 * Workaround to stop racing between the fence signaling and handling
597 	 * the cb. The lock is static after initially setting it up, just make
598 	 * sure that the dma_fence structure isn't freed up.
599 	 */
600 	rcu_read_lock();
601 	lock = vm->last_tlb_flush->lock;
602 	rcu_read_unlock();
603 
604 	spin_lock_irqsave(lock, flags);
605 	spin_unlock_irqrestore(lock, flags);
606 
607 	return atomic64_read(&vm->tlb_seq);
608 }
609 
610 /*
611  * vm eviction_lock can be taken in MMU notifiers. Make sure no reclaim-FS
612  * happens while holding this lock anywhere to prevent deadlocks when
613  * an MMU notifier runs in reclaim-FS context.
614  */
615 static inline void amdgpu_vm_eviction_lock(struct amdgpu_vm *vm)
616 {
617 	mutex_lock(&vm->eviction_lock);
618 	vm->saved_flags = memalloc_noreclaim_save();
619 }
620 
621 static inline bool amdgpu_vm_eviction_trylock(struct amdgpu_vm *vm)
622 {
623 	if (mutex_trylock(&vm->eviction_lock)) {
624 		vm->saved_flags = memalloc_noreclaim_save();
625 		return true;
626 	}
627 	return false;
628 }
629 
630 static inline void amdgpu_vm_eviction_unlock(struct amdgpu_vm *vm)
631 {
632 	memalloc_noreclaim_restore(vm->saved_flags);
633 	mutex_unlock(&vm->eviction_lock);
634 }
635 
636 void amdgpu_vm_update_fault_cache(struct amdgpu_device *adev,
637 				  unsigned int pasid,
638 				  uint64_t addr,
639 				  uint32_t status,
640 				  unsigned int vmhub);
641 void amdgpu_vm_tlb_fence_create(struct amdgpu_device *adev,
642 				 struct amdgpu_vm *vm,
643 				 struct dma_fence **fence);
644 
645 #endif
646