xref: /linux/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.h (revision 4002a6c55e99046b4a09ae255d38d3620b31fb1d)
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_SHIFT(mtype)	((uint64_t)(mtype) << 48)
112 #define AMDGPU_PTE_MTYPE_NV10_MASK     AMDGPU_PTE_MTYPE_NV10_SHIFT(7ULL)
113 #define AMDGPU_PTE_MTYPE_NV10(flags, mtype)			\
114 	(((uint64_t)(flags) & (~AMDGPU_PTE_MTYPE_NV10_MASK)) |	\
115 	  AMDGPU_PTE_MTYPE_NV10_SHIFT(mtype))
116 
117 /* gfx12 */
118 #define AMDGPU_PTE_PRT_GFX12		(1ULL << 56)
119 #define AMDGPU_PTE_PRT_FLAG(adev)	\
120 	((amdgpu_ip_version((adev), GC_HWIP, 0) >= IP_VERSION(12, 0, 0)) ? AMDGPU_PTE_PRT_GFX12 : AMDGPU_PTE_PRT)
121 
122 #define AMDGPU_PTE_MTYPE_GFX12_SHIFT(mtype)	((uint64_t)(mtype) << 54)
123 #define AMDGPU_PTE_MTYPE_GFX12_MASK	AMDGPU_PTE_MTYPE_GFX12_SHIFT(3ULL)
124 #define AMDGPU_PTE_MTYPE_GFX12(flags, mtype)				\
125 	(((uint64_t)(flags) & (~AMDGPU_PTE_MTYPE_GFX12_MASK)) |	\
126 	  AMDGPU_PTE_MTYPE_GFX12_SHIFT(mtype))
127 
128 #define AMDGPU_PTE_IS_PTE		(1ULL << 63)
129 
130 /* PDE Block Fragment Size for gfx v12 */
131 #define AMDGPU_PDE_BFS_GFX12(a)		((uint64_t)((a) & 0x1fULL) << 58)
132 #define AMDGPU_PDE_BFS_FLAG(adev, a)	\
133 	((amdgpu_ip_version((adev), GC_HWIP, 0) >= IP_VERSION(12, 0, 0)) ? AMDGPU_PDE_BFS_GFX12(a) : AMDGPU_PDE_BFS(a))
134 /* PDE is handled as PTE for gfx v12 */
135 #define AMDGPU_PDE_PTE_GFX12		(1ULL << 63)
136 #define AMDGPU_PDE_PTE_FLAG(adev)	\
137 	((amdgpu_ip_version((adev), GC_HWIP, 0) >= IP_VERSION(12, 0, 0)) ? AMDGPU_PDE_PTE_GFX12 : AMDGPU_PDE_PTE)
138 
139 /* How to program VM fault handling */
140 #define AMDGPU_VM_FAULT_STOP_NEVER	0
141 #define AMDGPU_VM_FAULT_STOP_FIRST	1
142 #define AMDGPU_VM_FAULT_STOP_ALWAYS	2
143 
144 /* How much VRAM be reserved for page tables */
145 #define AMDGPU_VM_RESERVED_VRAM		(8ULL << 20)
146 
147 /*
148  * max number of VMHUB
149  * layout: max 8 GFXHUB + 4 MMHUB0 + 1 MMHUB1
150  */
151 #define AMDGPU_MAX_VMHUBS			13
152 #define AMDGPU_GFXHUB_START			0
153 #define AMDGPU_MMHUB0_START			8
154 #define AMDGPU_MMHUB1_START			12
155 #define AMDGPU_GFXHUB(x)			(AMDGPU_GFXHUB_START + (x))
156 #define AMDGPU_MMHUB0(x)			(AMDGPU_MMHUB0_START + (x))
157 #define AMDGPU_MMHUB1(x)			(AMDGPU_MMHUB1_START + (x))
158 
159 #define AMDGPU_IS_GFXHUB(x) ((x) >= AMDGPU_GFXHUB_START && (x) < AMDGPU_MMHUB0_START)
160 #define AMDGPU_IS_MMHUB0(x) ((x) >= AMDGPU_MMHUB0_START && (x) < AMDGPU_MMHUB1_START)
161 #define AMDGPU_IS_MMHUB1(x) ((x) >= AMDGPU_MMHUB1_START && (x) < AMDGPU_MAX_VMHUBS)
162 
163 /* Reserve space at top/bottom of address space for kernel use */
164 #define AMDGPU_VA_RESERVED_CSA_SIZE		(2ULL << 20)
165 #define AMDGPU_VA_RESERVED_CSA_START(adev)	(((adev)->vm_manager.max_pfn \
166 						  << AMDGPU_GPU_PAGE_SHIFT)  \
167 						 - AMDGPU_VA_RESERVED_CSA_SIZE)
168 #define AMDGPU_VA_RESERVED_SEQ64_SIZE		(2ULL << 20)
169 #define AMDGPU_VA_RESERVED_SEQ64_START(adev)	(AMDGPU_VA_RESERVED_CSA_START(adev) \
170 						 - AMDGPU_VA_RESERVED_SEQ64_SIZE)
171 #define AMDGPU_VA_RESERVED_TRAP_SIZE		(2ULL << 12)
172 #define AMDGPU_VA_RESERVED_TRAP_START(adev)	(AMDGPU_VA_RESERVED_SEQ64_START(adev) \
173 						 - AMDGPU_VA_RESERVED_TRAP_SIZE)
174 #define AMDGPU_VA_RESERVED_BOTTOM		(1ULL << 16)
175 #define AMDGPU_VA_RESERVED_TOP			(AMDGPU_VA_RESERVED_TRAP_SIZE + \
176 						 AMDGPU_VA_RESERVED_SEQ64_SIZE + \
177 						 AMDGPU_VA_RESERVED_CSA_SIZE)
178 
179 /* See vm_update_mode */
180 #define AMDGPU_VM_USE_CPU_FOR_GFX (1 << 0)
181 #define AMDGPU_VM_USE_CPU_FOR_COMPUTE (1 << 1)
182 
183 /* VMPT level enumerate, and the hiberachy is:
184  * PDB2->PDB1->PDB0->PTB
185  */
186 enum amdgpu_vm_level {
187 	AMDGPU_VM_PDB2,
188 	AMDGPU_VM_PDB1,
189 	AMDGPU_VM_PDB0,
190 	AMDGPU_VM_PTB
191 };
192 
193 /* base structure for tracking BO usage in a VM */
194 struct amdgpu_vm_bo_base {
195 	/* constant after initialization */
196 	struct amdgpu_vm		*vm;
197 	struct amdgpu_bo		*bo;
198 
199 	/* protected by bo being reserved */
200 	struct amdgpu_vm_bo_base	*next;
201 
202 	/* protected by spinlock */
203 	struct list_head		vm_status;
204 
205 	/* protected by the BO being reserved */
206 	bool				moved;
207 };
208 
209 /* provided by hw blocks that can write ptes, e.g., sdma */
210 struct amdgpu_vm_pte_funcs {
211 	/* number of dw to reserve per operation */
212 	unsigned	copy_pte_num_dw;
213 
214 	/* copy pte entries from GART */
215 	void (*copy_pte)(struct amdgpu_ib *ib,
216 			 uint64_t pe, uint64_t src,
217 			 unsigned count);
218 
219 	/* write pte one entry at a time with addr mapping */
220 	void (*write_pte)(struct amdgpu_ib *ib, uint64_t pe,
221 			  uint64_t value, unsigned count,
222 			  uint32_t incr);
223 	/* for linear pte/pde updates without addr mapping */
224 	void (*set_pte_pde)(struct amdgpu_ib *ib,
225 			    uint64_t pe,
226 			    uint64_t addr, unsigned count,
227 			    uint32_t incr, uint64_t flags);
228 };
229 
230 struct amdgpu_task_info {
231 	char		process_name[TASK_COMM_LEN];
232 	char		task_name[TASK_COMM_LEN];
233 	pid_t		pid;
234 	pid_t		tgid;
235 	struct kref	refcount;
236 };
237 
238 /**
239  * struct amdgpu_vm_update_params
240  *
241  * Encapsulate some VM table update parameters to reduce
242  * the number of function parameters
243  *
244  */
245 struct amdgpu_vm_update_params {
246 
247 	/**
248 	 * @adev: amdgpu device we do this update for
249 	 */
250 	struct amdgpu_device *adev;
251 
252 	/**
253 	 * @vm: optional amdgpu_vm we do this update for
254 	 */
255 	struct amdgpu_vm *vm;
256 
257 	/**
258 	 * @immediate: if changes should be made immediately
259 	 */
260 	bool immediate;
261 
262 	/**
263 	 * @unlocked: true if the root BO is not locked
264 	 */
265 	bool unlocked;
266 
267 	/**
268 	 * @pages_addr:
269 	 *
270 	 * DMA addresses to use for mapping
271 	 */
272 	dma_addr_t *pages_addr;
273 
274 	/**
275 	 * @job: job to used for hw submission
276 	 */
277 	struct amdgpu_job *job;
278 
279 	/**
280 	 * @num_dw_left: number of dw left for the IB
281 	 */
282 	unsigned int num_dw_left;
283 
284 	/**
285 	 * @needs_flush: true whenever we need to invalidate the TLB
286 	 */
287 	bool needs_flush;
288 
289 	/**
290 	 * @allow_override: true for memory that is not uncached: allows MTYPE
291 	 * to be overridden for NUMA local memory.
292 	 */
293 	bool allow_override;
294 
295 	/**
296 	 * @tlb_flush_waitlist: temporary storage for BOs until tlb_flush
297 	 */
298 	struct list_head tlb_flush_waitlist;
299 };
300 
301 struct amdgpu_vm_update_funcs {
302 	int (*map_table)(struct amdgpu_bo_vm *bo);
303 	int (*prepare)(struct amdgpu_vm_update_params *p, struct dma_resv *resv,
304 		       enum amdgpu_sync_mode sync_mode);
305 	int (*update)(struct amdgpu_vm_update_params *p,
306 		      struct amdgpu_bo_vm *bo, uint64_t pe, uint64_t addr,
307 		      unsigned count, uint32_t incr, uint64_t flags);
308 	int (*commit)(struct amdgpu_vm_update_params *p,
309 		      struct dma_fence **fence);
310 };
311 
312 struct amdgpu_vm_fault_info {
313 	/* fault address */
314 	uint64_t	addr;
315 	/* fault status register */
316 	uint32_t	status;
317 	/* which vmhub? gfxhub, mmhub, etc. */
318 	unsigned int	vmhub;
319 };
320 
321 struct amdgpu_vm {
322 	/* tree of virtual addresses mapped */
323 	struct rb_root_cached	va;
324 
325 	/* Lock to prevent eviction while we are updating page tables
326 	 * use vm_eviction_lock/unlock(vm)
327 	 */
328 	struct mutex		eviction_lock;
329 	bool			evicting;
330 	unsigned int		saved_flags;
331 
332 	/* Lock to protect vm_bo add/del/move on all lists of vm */
333 	spinlock_t		status_lock;
334 
335 	/* Per-VM and PT BOs who needs a validation */
336 	struct list_head	evicted;
337 
338 	/* BOs for user mode queues that need a validation */
339 	struct list_head	evicted_user;
340 
341 	/* PT BOs which relocated and their parent need an update */
342 	struct list_head	relocated;
343 
344 	/* per VM BOs moved, but not yet updated in the PT */
345 	struct list_head	moved;
346 
347 	/* All BOs of this VM not currently in the state machine */
348 	struct list_head	idle;
349 
350 	/* regular invalidated BOs, but not yet updated in the PT */
351 	struct list_head	invalidated;
352 
353 	/* BO mappings freed, but not yet updated in the PT */
354 	struct list_head	freed;
355 
356 	/* BOs which are invalidated, has been updated in the PTs */
357 	struct list_head        done;
358 
359 	/* PT BOs scheduled to free and fill with zero if vm_resv is not hold */
360 	struct list_head	pt_freed;
361 	struct work_struct	pt_free_work;
362 
363 	/* contains the page directory */
364 	struct amdgpu_vm_bo_base     root;
365 	struct dma_fence	*last_update;
366 
367 	/* Scheduler entities for page table updates */
368 	struct drm_sched_entity	immediate;
369 	struct drm_sched_entity	delayed;
370 
371 	/* Last finished delayed update */
372 	atomic64_t		tlb_seq;
373 	struct dma_fence	*last_tlb_flush;
374 	atomic64_t		kfd_last_flushed_seq;
375 	uint64_t		tlb_fence_context;
376 
377 	/* How many times we had to re-generate the page tables */
378 	uint64_t		generation;
379 
380 	/* Last unlocked submission to the scheduler entities */
381 	struct dma_fence	*last_unlocked;
382 
383 	unsigned int		pasid;
384 	bool			reserved_vmid[AMDGPU_MAX_VMHUBS];
385 
386 	/* Flag to indicate if VM tables are updated by CPU or GPU (SDMA) */
387 	bool					use_cpu_for_update;
388 
389 	/* Functions to use for VM table updates */
390 	const struct amdgpu_vm_update_funcs	*update_funcs;
391 
392 	/* Up to 128 pending retry page faults */
393 	DECLARE_KFIFO(faults, u64, 128);
394 
395 	/* Points to the KFD process VM info */
396 	struct amdkfd_process_info *process_info;
397 
398 	/* List node in amdkfd_process_info.vm_list_head */
399 	struct list_head	vm_list_node;
400 
401 	/* Valid while the PD is reserved or fenced */
402 	uint64_t		pd_phys_addr;
403 
404 	/* Some basic info about the task */
405 	struct amdgpu_task_info *task_info;
406 
407 	/* Store positions of group of BOs */
408 	struct ttm_lru_bulk_move lru_bulk_move;
409 	/* Flag to indicate if VM is used for compute */
410 	bool			is_compute_context;
411 
412 	/* Memory partition number, -1 means any partition */
413 	int8_t			mem_id;
414 
415 	/* cached fault info */
416 	struct amdgpu_vm_fault_info fault_info;
417 };
418 
419 struct amdgpu_vm_manager {
420 	/* Handling of VMIDs */
421 	struct amdgpu_vmid_mgr			id_mgr[AMDGPU_MAX_VMHUBS];
422 	unsigned int				first_kfd_vmid;
423 	bool					concurrent_flush;
424 
425 	/* Handling of VM fences */
426 	u64					fence_context;
427 	unsigned				seqno[AMDGPU_MAX_RINGS];
428 
429 	uint64_t				max_pfn;
430 	uint32_t				num_level;
431 	uint32_t				block_size;
432 	uint32_t				fragment_size;
433 	enum amdgpu_vm_level			root_level;
434 	/* vram base address for page table entry  */
435 	u64					vram_base_offset;
436 	/* vm pte handling */
437 	const struct amdgpu_vm_pte_funcs	*vm_pte_funcs;
438 	struct drm_gpu_scheduler		*vm_pte_scheds[AMDGPU_MAX_RINGS];
439 	unsigned				vm_pte_num_scheds;
440 	struct amdgpu_ring			*page_fault;
441 
442 	/* partial resident texture handling */
443 	spinlock_t				prt_lock;
444 	atomic_t				num_prt_users;
445 
446 	/* controls how VM page tables are updated for Graphics and Compute.
447 	 * BIT0[= 0] Graphics updated by SDMA [= 1] by CPU
448 	 * BIT1[= 0] Compute updated by SDMA [= 1] by CPU
449 	 */
450 	int					vm_update_mode;
451 
452 	/* PASID to VM mapping, will be used in interrupt context to
453 	 * look up VM of a page fault
454 	 */
455 	struct xarray				pasids;
456 	/* Global registration of recent page fault information */
457 	struct amdgpu_vm_fault_info	fault_info;
458 };
459 
460 struct amdgpu_bo_va_mapping;
461 
462 #define amdgpu_vm_copy_pte(adev, ib, pe, src, count) ((adev)->vm_manager.vm_pte_funcs->copy_pte((ib), (pe), (src), (count)))
463 #define amdgpu_vm_write_pte(adev, ib, pe, value, count, incr) ((adev)->vm_manager.vm_pte_funcs->write_pte((ib), (pe), (value), (count), (incr)))
464 #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)))
465 
466 extern const struct amdgpu_vm_update_funcs amdgpu_vm_cpu_funcs;
467 extern const struct amdgpu_vm_update_funcs amdgpu_vm_sdma_funcs;
468 
469 void amdgpu_vm_manager_init(struct amdgpu_device *adev);
470 void amdgpu_vm_manager_fini(struct amdgpu_device *adev);
471 
472 int amdgpu_vm_set_pasid(struct amdgpu_device *adev, struct amdgpu_vm *vm,
473 			u32 pasid);
474 
475 long amdgpu_vm_wait_idle(struct amdgpu_vm *vm, long timeout);
476 int amdgpu_vm_init(struct amdgpu_device *adev, struct amdgpu_vm *vm, int32_t xcp_id);
477 int amdgpu_vm_make_compute(struct amdgpu_device *adev, struct amdgpu_vm *vm);
478 void amdgpu_vm_release_compute(struct amdgpu_device *adev, struct amdgpu_vm *vm);
479 void amdgpu_vm_fini(struct amdgpu_device *adev, struct amdgpu_vm *vm);
480 int amdgpu_vm_lock_pd(struct amdgpu_vm *vm, struct drm_exec *exec,
481 		      unsigned int num_fences);
482 bool amdgpu_vm_ready(struct amdgpu_vm *vm);
483 uint64_t amdgpu_vm_generation(struct amdgpu_device *adev, struct amdgpu_vm *vm);
484 int amdgpu_vm_validate(struct amdgpu_device *adev, struct amdgpu_vm *vm,
485 		       struct ww_acquire_ctx *ticket,
486 		       int (*callback)(void *p, struct amdgpu_bo *bo),
487 		       void *param);
488 int amdgpu_vm_flush(struct amdgpu_ring *ring, struct amdgpu_job *job, bool need_pipe_sync);
489 int amdgpu_vm_update_pdes(struct amdgpu_device *adev,
490 			  struct amdgpu_vm *vm, bool immediate);
491 int amdgpu_vm_clear_freed(struct amdgpu_device *adev,
492 			  struct amdgpu_vm *vm,
493 			  struct dma_fence **fence);
494 int amdgpu_vm_handle_moved(struct amdgpu_device *adev,
495 			   struct amdgpu_vm *vm,
496 			   struct ww_acquire_ctx *ticket);
497 int amdgpu_vm_flush_compute_tlb(struct amdgpu_device *adev,
498 				struct amdgpu_vm *vm,
499 				uint32_t flush_type,
500 				uint32_t xcc_mask);
501 void amdgpu_vm_bo_base_init(struct amdgpu_vm_bo_base *base,
502 			    struct amdgpu_vm *vm, struct amdgpu_bo *bo);
503 int amdgpu_vm_update_range(struct amdgpu_device *adev, struct amdgpu_vm *vm,
504 			   bool immediate, bool unlocked, bool flush_tlb, bool allow_override,
505 			   struct dma_resv *resv, uint64_t start, uint64_t last,
506 			   uint64_t flags, uint64_t offset, uint64_t vram_base,
507 			   struct ttm_resource *res, dma_addr_t *pages_addr,
508 			   struct dma_fence **fence);
509 int amdgpu_vm_bo_update(struct amdgpu_device *adev,
510 			struct amdgpu_bo_va *bo_va,
511 			bool clear);
512 bool amdgpu_vm_evictable(struct amdgpu_bo *bo);
513 void amdgpu_vm_bo_invalidate(struct amdgpu_device *adev,
514 			     struct amdgpu_bo *bo, bool evicted);
515 uint64_t amdgpu_vm_map_gart(const dma_addr_t *pages_addr, uint64_t addr);
516 struct amdgpu_bo_va *amdgpu_vm_bo_find(struct amdgpu_vm *vm,
517 				       struct amdgpu_bo *bo);
518 struct amdgpu_bo_va *amdgpu_vm_bo_add(struct amdgpu_device *adev,
519 				      struct amdgpu_vm *vm,
520 				      struct amdgpu_bo *bo);
521 int amdgpu_vm_bo_map(struct amdgpu_device *adev,
522 		     struct amdgpu_bo_va *bo_va,
523 		     uint64_t addr, uint64_t offset,
524 		     uint64_t size, uint64_t flags);
525 int amdgpu_vm_bo_replace_map(struct amdgpu_device *adev,
526 			     struct amdgpu_bo_va *bo_va,
527 			     uint64_t addr, uint64_t offset,
528 			     uint64_t size, uint64_t flags);
529 int amdgpu_vm_bo_unmap(struct amdgpu_device *adev,
530 		       struct amdgpu_bo_va *bo_va,
531 		       uint64_t addr);
532 int amdgpu_vm_bo_clear_mappings(struct amdgpu_device *adev,
533 				struct amdgpu_vm *vm,
534 				uint64_t saddr, uint64_t size);
535 struct amdgpu_bo_va_mapping *amdgpu_vm_bo_lookup_mapping(struct amdgpu_vm *vm,
536 							 uint64_t addr);
537 void amdgpu_vm_bo_trace_cs(struct amdgpu_vm *vm, struct ww_acquire_ctx *ticket);
538 void amdgpu_vm_bo_del(struct amdgpu_device *adev,
539 		      struct amdgpu_bo_va *bo_va);
540 void amdgpu_vm_adjust_size(struct amdgpu_device *adev, uint32_t min_vm_size,
541 			   uint32_t fragment_size_default, unsigned max_level,
542 			   unsigned max_bits);
543 int amdgpu_vm_ioctl(struct drm_device *dev, void *data, struct drm_file *filp);
544 bool amdgpu_vm_need_pipeline_sync(struct amdgpu_ring *ring,
545 				  struct amdgpu_job *job);
546 void amdgpu_vm_check_compute_bug(struct amdgpu_device *adev);
547 
548 struct amdgpu_task_info *
549 amdgpu_vm_get_task_info_pasid(struct amdgpu_device *adev, u32 pasid);
550 
551 struct amdgpu_task_info *
552 amdgpu_vm_get_task_info_vm(struct amdgpu_vm *vm);
553 
554 void amdgpu_vm_put_task_info(struct amdgpu_task_info *task_info);
555 
556 bool amdgpu_vm_handle_fault(struct amdgpu_device *adev, u32 pasid,
557 			    u32 vmid, u32 node_id, uint64_t addr,
558 			    bool write_fault);
559 
560 void amdgpu_vm_set_task_info(struct amdgpu_vm *vm);
561 
562 void amdgpu_vm_move_to_lru_tail(struct amdgpu_device *adev,
563 				struct amdgpu_vm *vm);
564 void amdgpu_vm_get_memory(struct amdgpu_vm *vm,
565 			  struct amdgpu_mem_stats *stats);
566 
567 int amdgpu_vm_pt_clear(struct amdgpu_device *adev, struct amdgpu_vm *vm,
568 		       struct amdgpu_bo_vm *vmbo, bool immediate);
569 int amdgpu_vm_pt_create(struct amdgpu_device *adev, struct amdgpu_vm *vm,
570 			int level, bool immediate, struct amdgpu_bo_vm **vmbo,
571 			int32_t xcp_id);
572 void amdgpu_vm_pt_free_root(struct amdgpu_device *adev, struct amdgpu_vm *vm);
573 
574 int amdgpu_vm_pde_update(struct amdgpu_vm_update_params *params,
575 			 struct amdgpu_vm_bo_base *entry);
576 int amdgpu_vm_ptes_update(struct amdgpu_vm_update_params *params,
577 			  uint64_t start, uint64_t end,
578 			  uint64_t dst, uint64_t flags);
579 void amdgpu_vm_pt_free_work(struct work_struct *work);
580 void amdgpu_vm_pt_free_list(struct amdgpu_device *adev,
581 			    struct amdgpu_vm_update_params *params);
582 
583 #if defined(CONFIG_DEBUG_FS)
584 void amdgpu_debugfs_vm_bo_info(struct amdgpu_vm *vm, struct seq_file *m);
585 #endif
586 
587 int amdgpu_vm_pt_map_tables(struct amdgpu_device *adev, struct amdgpu_vm *vm);
588 
589 bool amdgpu_vm_is_bo_always_valid(struct amdgpu_vm *vm, struct amdgpu_bo *bo);
590 
591 /**
592  * amdgpu_vm_tlb_seq - return tlb flush sequence number
593  * @vm: the amdgpu_vm structure to query
594  *
595  * Returns the tlb flush sequence number which indicates that the VM TLBs needs
596  * to be invalidated whenever the sequence number change.
597  */
598 static inline uint64_t amdgpu_vm_tlb_seq(struct amdgpu_vm *vm)
599 {
600 	unsigned long flags;
601 	spinlock_t *lock;
602 
603 	/*
604 	 * Workaround to stop racing between the fence signaling and handling
605 	 * the cb. The lock is static after initially setting it up, just make
606 	 * sure that the dma_fence structure isn't freed up.
607 	 */
608 	rcu_read_lock();
609 	lock = vm->last_tlb_flush->lock;
610 	rcu_read_unlock();
611 
612 	spin_lock_irqsave(lock, flags);
613 	spin_unlock_irqrestore(lock, flags);
614 
615 	return atomic64_read(&vm->tlb_seq);
616 }
617 
618 /*
619  * vm eviction_lock can be taken in MMU notifiers. Make sure no reclaim-FS
620  * happens while holding this lock anywhere to prevent deadlocks when
621  * an MMU notifier runs in reclaim-FS context.
622  */
623 static inline void amdgpu_vm_eviction_lock(struct amdgpu_vm *vm)
624 {
625 	mutex_lock(&vm->eviction_lock);
626 	vm->saved_flags = memalloc_noreclaim_save();
627 }
628 
629 static inline bool amdgpu_vm_eviction_trylock(struct amdgpu_vm *vm)
630 {
631 	if (mutex_trylock(&vm->eviction_lock)) {
632 		vm->saved_flags = memalloc_noreclaim_save();
633 		return true;
634 	}
635 	return false;
636 }
637 
638 static inline void amdgpu_vm_eviction_unlock(struct amdgpu_vm *vm)
639 {
640 	memalloc_noreclaim_restore(vm->saved_flags);
641 	mutex_unlock(&vm->eviction_lock);
642 }
643 
644 void amdgpu_vm_update_fault_cache(struct amdgpu_device *adev,
645 				  unsigned int pasid,
646 				  uint64_t addr,
647 				  uint32_t status,
648 				  unsigned int vmhub);
649 void amdgpu_vm_tlb_fence_create(struct amdgpu_device *adev,
650 				 struct amdgpu_vm *vm,
651 				 struct dma_fence **fence);
652 
653 #endif
654