xref: /linux/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.h (revision c7062be3380cb20c8b1c4a935a13f1848ead0719)
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 #include "amdgpu_ttm.h"
39 
40 struct drm_exec;
41 
42 struct amdgpu_bo_va;
43 struct amdgpu_job;
44 struct amdgpu_bo_list_entry;
45 struct amdgpu_bo_vm;
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_SHIFT(mtype)	((uint64_t)(mtype) << 57)
98 #define AMDGPU_PTE_MTYPE_VG10_MASK	AMDGPU_PTE_MTYPE_VG10_SHIFT(3ULL)
99 #define AMDGPU_PTE_MTYPE_VG10(flags, mtype)			\
100 	(((uint64_t)(flags) & (~AMDGPU_PTE_MTYPE_VG10_MASK)) |	\
101 	  AMDGPU_PTE_MTYPE_VG10_SHIFT(mtype))
102 
103 #define AMDGPU_MTYPE_NC 0
104 #define AMDGPU_MTYPE_CC 2
105 
106 #define AMDGPU_PTE_DEFAULT_ATC  (AMDGPU_PTE_SYSTEM      \
107                                 | AMDGPU_PTE_SNOOPED    \
108                                 | AMDGPU_PTE_EXECUTABLE \
109                                 | AMDGPU_PTE_READABLE   \
110                                 | AMDGPU_PTE_WRITEABLE  \
111                                 | AMDGPU_PTE_MTYPE_VG10(AMDGPU_MTYPE_CC))
112 
113 /* gfx10 */
114 #define AMDGPU_PTE_MTYPE_NV10_SHIFT(mtype)	((uint64_t)(mtype) << 48)
115 #define AMDGPU_PTE_MTYPE_NV10_MASK     AMDGPU_PTE_MTYPE_NV10_SHIFT(7ULL)
116 #define AMDGPU_PTE_MTYPE_NV10(flags, mtype)			\
117 	(((uint64_t)(flags) & (~AMDGPU_PTE_MTYPE_NV10_MASK)) |	\
118 	  AMDGPU_PTE_MTYPE_NV10_SHIFT(mtype))
119 
120 /* gfx12 */
121 #define AMDGPU_PTE_PRT_GFX12		(1ULL << 56)
122 #define AMDGPU_PTE_PRT_FLAG(adev)	\
123 	((amdgpu_ip_version((adev), GC_HWIP, 0) >= IP_VERSION(12, 0, 0)) ? AMDGPU_PTE_PRT_GFX12 : AMDGPU_PTE_PRT)
124 
125 #define AMDGPU_PTE_MTYPE_GFX12_SHIFT(mtype)	((uint64_t)(mtype) << 54)
126 #define AMDGPU_PTE_MTYPE_GFX12_MASK	AMDGPU_PTE_MTYPE_GFX12_SHIFT(3ULL)
127 #define AMDGPU_PTE_MTYPE_GFX12(flags, mtype)				\
128 	(((uint64_t)(flags) & (~AMDGPU_PTE_MTYPE_GFX12_MASK)) |	\
129 	  AMDGPU_PTE_MTYPE_GFX12_SHIFT(mtype))
130 
131 #define AMDGPU_PTE_DCC			(1ULL << 58)
132 #define AMDGPU_PTE_BUS_ATOMICS		(1ULL << 59)
133 #define AMDGPU_PTE_IS_PTE		(1ULL << 63)
134 
135 /* PDE Block Fragment Size for gfx v12 */
136 #define AMDGPU_PDE_BFS_GFX12(a)		((uint64_t)((a) & 0x1fULL) << 58)
137 #define AMDGPU_PDE_BFS_FLAG(adev, a)	\
138 	((amdgpu_ip_version((adev), GC_HWIP, 0) >= IP_VERSION(12, 0, 0)) ? AMDGPU_PDE_BFS_GFX12(a) : AMDGPU_PDE_BFS(a))
139 /* PDE is handled as PTE for gfx v12 */
140 #define AMDGPU_PDE_PTE_GFX12		(1ULL << 63)
141 #define AMDGPU_PDE_PTE_FLAG(adev)	\
142 	((amdgpu_ip_version((adev), GC_HWIP, 0) >= IP_VERSION(12, 0, 0)) ? AMDGPU_PDE_PTE_GFX12 : AMDGPU_PDE_PTE)
143 
144 /* How to program VM fault handling */
145 #define AMDGPU_VM_FAULT_STOP_NEVER	0
146 #define AMDGPU_VM_FAULT_STOP_FIRST	1
147 #define AMDGPU_VM_FAULT_STOP_ALWAYS	2
148 
149 /* How much VRAM be reserved for page tables */
150 #define AMDGPU_VM_RESERVED_VRAM		(8ULL << 20)
151 
152 /*
153  * max number of VMHUB
154  * layout: max 8 GFXHUB + 4 MMHUB0 + 1 MMHUB1
155  */
156 #define AMDGPU_MAX_VMHUBS			13
157 #define AMDGPU_GFXHUB_START			0
158 #define AMDGPU_MMHUB0_START			8
159 #define AMDGPU_MMHUB1_START			12
160 #define AMDGPU_GFXHUB(x)			(AMDGPU_GFXHUB_START + (x))
161 #define AMDGPU_MMHUB0(x)			(AMDGPU_MMHUB0_START + (x))
162 #define AMDGPU_MMHUB1(x)			(AMDGPU_MMHUB1_START + (x))
163 
164 #define AMDGPU_IS_GFXHUB(x) ((x) >= AMDGPU_GFXHUB_START && (x) < AMDGPU_MMHUB0_START)
165 #define AMDGPU_IS_MMHUB0(x) ((x) >= AMDGPU_MMHUB0_START && (x) < AMDGPU_MMHUB1_START)
166 #define AMDGPU_IS_MMHUB1(x) ((x) >= AMDGPU_MMHUB1_START && (x) < AMDGPU_MAX_VMHUBS)
167 
168 /* Reserve space at top/bottom of address space for kernel use */
169 #define AMDGPU_VA_RESERVED_CSA_SIZE		(2ULL << 20)
170 #define AMDGPU_VA_RESERVED_CSA_START(adev)	(((adev)->vm_manager.max_pfn \
171 						  << AMDGPU_GPU_PAGE_SHIFT)  \
172 						 - AMDGPU_VA_RESERVED_CSA_SIZE)
173 #define AMDGPU_VA_RESERVED_SEQ64_SIZE		(2ULL << 20)
174 #define AMDGPU_VA_RESERVED_SEQ64_START(adev)	(AMDGPU_VA_RESERVED_CSA_START(adev) \
175 						 - AMDGPU_VA_RESERVED_SEQ64_SIZE)
176 #define AMDGPU_VA_RESERVED_TRAP_SIZE		(2ULL << 12)
177 #define AMDGPU_VA_RESERVED_TRAP_START(adev)	(AMDGPU_VA_RESERVED_SEQ64_START(adev) \
178 						 - AMDGPU_VA_RESERVED_TRAP_SIZE)
179 #define AMDGPU_VA_RESERVED_BOTTOM		(1ULL << 16)
180 #define AMDGPU_VA_RESERVED_TOP			(AMDGPU_VA_RESERVED_TRAP_SIZE + \
181 						 AMDGPU_VA_RESERVED_SEQ64_SIZE + \
182 						 AMDGPU_VA_RESERVED_CSA_SIZE)
183 
184 /* See vm_update_mode */
185 #define AMDGPU_VM_USE_CPU_FOR_GFX (1 << 0)
186 #define AMDGPU_VM_USE_CPU_FOR_COMPUTE (1 << 1)
187 
188 /* VMPT level enumerate, and the hiberachy is:
189  * PDB3->PDB2->PDB1->PDB0->PTB
190  */
191 enum amdgpu_vm_level {
192 	AMDGPU_VM_PDB3,
193 	AMDGPU_VM_PDB2,
194 	AMDGPU_VM_PDB1,
195 	AMDGPU_VM_PDB0,
196 	AMDGPU_VM_PTB
197 };
198 
199 /* base structure for tracking BO usage in a VM */
200 struct amdgpu_vm_bo_base {
201 	/* constant after initialization */
202 	struct amdgpu_vm		*vm;
203 	struct amdgpu_bo		*bo;
204 
205 	/* protected by bo being reserved */
206 	struct amdgpu_vm_bo_base	*next;
207 
208 	/* protected by vm status_lock */
209 	struct list_head		vm_status;
210 
211 	/* if the bo is counted as shared in mem stats
212 	 * protected by vm status_lock */
213 	bool				shared;
214 
215 	/* protected by the BO being reserved */
216 	bool				moved;
217 };
218 
219 /* provided by hw blocks that can write ptes, e.g., sdma */
220 struct amdgpu_vm_pte_funcs {
221 	/* number of dw to reserve per operation */
222 	unsigned	copy_pte_num_dw;
223 
224 	/* copy pte entries from GART */
225 	void (*copy_pte)(struct amdgpu_ib *ib,
226 			 uint64_t pe, uint64_t src,
227 			 unsigned count);
228 
229 	/* write pte one entry at a time with addr mapping */
230 	void (*write_pte)(struct amdgpu_ib *ib, uint64_t pe,
231 			  uint64_t value, unsigned count,
232 			  uint32_t incr);
233 	/* for linear pte/pde updates without addr mapping */
234 	void (*set_pte_pde)(struct amdgpu_ib *ib,
235 			    uint64_t pe,
236 			    uint64_t addr, unsigned count,
237 			    uint32_t incr, uint64_t flags);
238 };
239 
240 struct amdgpu_task_info {
241 	struct drm_wedge_task_info task;
242 	char		process_name[TASK_COMM_LEN];
243 	pid_t		tgid;
244 	struct kref	refcount;
245 };
246 
247 /**
248  * struct amdgpu_vm_update_params
249  *
250  * Encapsulate some VM table update parameters to reduce
251  * the number of function parameters
252  *
253  */
254 struct amdgpu_vm_update_params {
255 
256 	/**
257 	 * @adev: amdgpu device we do this update for
258 	 */
259 	struct amdgpu_device *adev;
260 
261 	/**
262 	 * @vm: optional amdgpu_vm we do this update for
263 	 */
264 	struct amdgpu_vm *vm;
265 
266 	/**
267 	 * @immediate: if changes should be made immediately
268 	 */
269 	bool immediate;
270 
271 	/**
272 	 * @unlocked: true if the root BO is not locked
273 	 */
274 	bool unlocked;
275 
276 	/**
277 	 * @pages_addr:
278 	 *
279 	 * DMA addresses to use for mapping
280 	 */
281 	dma_addr_t *pages_addr;
282 
283 	/**
284 	 * @job: job to used for hw submission
285 	 */
286 	struct amdgpu_job *job;
287 
288 	/**
289 	 * @num_dw_left: number of dw left for the IB
290 	 */
291 	unsigned int num_dw_left;
292 
293 	/**
294 	 * @needs_flush: true whenever we need to invalidate the TLB
295 	 */
296 	bool needs_flush;
297 
298 	/**
299 	 * @allow_override: true for memory that is not uncached: allows MTYPE
300 	 * to be overridden for NUMA local memory.
301 	 */
302 	bool allow_override;
303 
304 	/**
305 	 * @tlb_flush_waitlist: temporary storage for BOs until tlb_flush
306 	 */
307 	struct list_head tlb_flush_waitlist;
308 };
309 
310 struct amdgpu_vm_update_funcs {
311 	int (*map_table)(struct amdgpu_bo_vm *bo);
312 	int (*prepare)(struct amdgpu_vm_update_params *p,
313 		       struct amdgpu_sync *sync, u64 k_job_id);
314 	int (*update)(struct amdgpu_vm_update_params *p,
315 		      struct amdgpu_bo_vm *bo, uint64_t pe, uint64_t addr,
316 		      unsigned count, uint32_t incr, uint64_t flags);
317 	int (*commit)(struct amdgpu_vm_update_params *p,
318 		      struct dma_fence **fence);
319 };
320 
321 struct amdgpu_vm_fault_info {
322 	/* fault address */
323 	uint64_t	addr;
324 	/* fault status register */
325 	uint32_t	status;
326 	/* which vmhub? gfxhub, mmhub, etc. */
327 	unsigned int	vmhub;
328 };
329 
330 struct amdgpu_mem_stats {
331 	struct drm_memory_stats drm;
332 
333 	/* buffers that requested this placement but are currently evicted */
334 	uint64_t evicted;
335 };
336 
337 struct amdgpu_vm {
338 	/* tree of virtual addresses mapped */
339 	struct rb_root_cached	va;
340 
341 	/* Lock to prevent eviction while we are updating page tables
342 	 * use vm_eviction_lock/unlock(vm)
343 	 */
344 	struct mutex		eviction_lock;
345 	bool			evicting;
346 	unsigned int		saved_flags;
347 
348 	/* Lock to protect vm_bo add/del/move on all lists of vm */
349 	spinlock_t		status_lock;
350 
351 	/* Memory statistics for this vm, protected by status_lock */
352 	struct amdgpu_mem_stats stats[__AMDGPU_PL_NUM];
353 
354 	/*
355 	 * The following lists contain amdgpu_vm_bo_base objects for either
356 	 * PDs, PTs or per VM BOs. The state transits are:
357 	 *
358 	 * evicted -> relocated (PDs, PTs) or moved (per VM BOs) -> idle
359 	 */
360 
361 	/* Per-VM and PT BOs who needs a validation */
362 	struct list_head	evicted;
363 
364 	/* PT BOs which relocated and their parent need an update */
365 	struct list_head	relocated;
366 
367 	/* per VM BOs moved, but not yet updated in the PT */
368 	struct list_head	moved;
369 
370 	/* All BOs of this VM not currently in the state machine */
371 	struct list_head	idle;
372 
373 	/*
374 	 * The following lists contain amdgpu_vm_bo_base objects for BOs which
375 	 * have their own dma_resv object and not depend on the root PD. Their
376 	 * state transits are:
377 	 *
378 	 * evicted_user or invalidated -> done
379 	 */
380 
381 	/* BOs for user mode queues that need a validation */
382 	struct list_head	evicted_user;
383 
384 	/* regular invalidated BOs, but not yet updated in the PT */
385 	struct list_head	invalidated;
386 
387 	/* BOs which are invalidated, has been updated in the PTs */
388 	struct list_head        done;
389 
390 	/*
391 	 * This list contains amdgpu_bo_va_mapping objects which have been freed
392 	 * but not updated in the PTs
393 	 */
394 	struct list_head	freed;
395 
396 	/* contains the page directory */
397 	struct amdgpu_vm_bo_base     root;
398 	struct dma_fence	*last_update;
399 
400 	/* Scheduler entities for page table updates */
401 	struct drm_sched_entity	immediate;
402 	struct drm_sched_entity	delayed;
403 
404 	/* Last finished delayed update */
405 	atomic64_t		tlb_seq;
406 	struct dma_fence	*last_tlb_flush;
407 	atomic64_t		kfd_last_flushed_seq;
408 	uint64_t		tlb_fence_context;
409 
410 	/* How many times we had to re-generate the page tables */
411 	uint64_t		generation;
412 
413 	/* Last unlocked submission to the scheduler entities */
414 	struct dma_fence	*last_unlocked;
415 
416 	unsigned int		pasid;
417 	struct amdgpu_vmid	*reserved_vmid[AMDGPU_MAX_VMHUBS];
418 
419 	/* Flag to indicate if VM tables are updated by CPU or GPU (SDMA) */
420 	bool					use_cpu_for_update;
421 
422 	/* Functions to use for VM table updates */
423 	const struct amdgpu_vm_update_funcs	*update_funcs;
424 
425 	/* Up to 128 pending retry page faults */
426 	DECLARE_KFIFO(faults, u64, 128);
427 
428 	/* Points to the KFD process VM info */
429 	struct amdkfd_process_info *process_info;
430 
431 	/* List node in amdkfd_process_info.vm_list_head */
432 	struct list_head	vm_list_node;
433 
434 	/* Valid while the PD is reserved or fenced */
435 	uint64_t		pd_phys_addr;
436 
437 	/* Some basic info about the task */
438 	struct amdgpu_task_info *task_info;
439 
440 	/* Store positions of group of BOs */
441 	struct ttm_lru_bulk_move lru_bulk_move;
442 	/* Flag to indicate if VM is used for compute */
443 	bool			is_compute_context;
444 
445 	/* Memory partition number, -1 means any partition */
446 	int8_t			mem_id;
447 
448 	/* cached fault info */
449 	struct amdgpu_vm_fault_info fault_info;
450 };
451 
452 struct amdgpu_vm_manager {
453 	/* Handling of VMIDs */
454 	struct amdgpu_vmid_mgr			id_mgr[AMDGPU_MAX_VMHUBS];
455 	unsigned int				first_kfd_vmid;
456 	bool					concurrent_flush;
457 
458 	uint64_t				max_pfn;
459 	uint32_t				num_level;
460 	uint32_t				block_size;
461 	uint32_t				fragment_size;
462 	enum amdgpu_vm_level			root_level;
463 	/* vram base address for page table entry  */
464 	u64					vram_base_offset;
465 	/* vm pte handling */
466 	const struct amdgpu_vm_pte_funcs	*vm_pte_funcs;
467 	struct drm_gpu_scheduler		*vm_pte_scheds[AMDGPU_MAX_RINGS];
468 	unsigned				vm_pte_num_scheds;
469 	struct amdgpu_ring			*page_fault;
470 
471 	/* partial resident texture handling */
472 	spinlock_t				prt_lock;
473 	atomic_t				num_prt_users;
474 
475 	/* controls how VM page tables are updated for Graphics and Compute.
476 	 * BIT0[= 0] Graphics updated by SDMA [= 1] by CPU
477 	 * BIT1[= 0] Compute updated by SDMA [= 1] by CPU
478 	 */
479 	int					vm_update_mode;
480 
481 	/* PASID to VM mapping, will be used in interrupt context to
482 	 * look up VM of a page fault
483 	 */
484 	struct xarray				pasids;
485 	/* Global registration of recent page fault information */
486 	struct amdgpu_vm_fault_info	fault_info;
487 };
488 
489 struct amdgpu_bo_va_mapping;
490 
491 #define amdgpu_vm_copy_pte(adev, ib, pe, src, count) ((adev)->vm_manager.vm_pte_funcs->copy_pte((ib), (pe), (src), (count)))
492 #define amdgpu_vm_write_pte(adev, ib, pe, value, count, incr) ((adev)->vm_manager.vm_pte_funcs->write_pte((ib), (pe), (value), (count), (incr)))
493 #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)))
494 
495 extern const struct amdgpu_vm_update_funcs amdgpu_vm_cpu_funcs;
496 extern const struct amdgpu_vm_update_funcs amdgpu_vm_sdma_funcs;
497 
498 void amdgpu_vm_manager_init(struct amdgpu_device *adev);
499 void amdgpu_vm_manager_fini(struct amdgpu_device *adev);
500 
501 long amdgpu_vm_wait_idle(struct amdgpu_vm *vm, long timeout);
502 int amdgpu_vm_init(struct amdgpu_device *adev, struct amdgpu_vm *vm, int32_t xcp_id, uint32_t pasid);
503 int amdgpu_vm_make_compute(struct amdgpu_device *adev, struct amdgpu_vm *vm);
504 void amdgpu_vm_fini(struct amdgpu_device *adev, struct amdgpu_vm *vm);
505 int amdgpu_vm_lock_pd(struct amdgpu_vm *vm, struct drm_exec *exec,
506 		      unsigned int num_fences);
507 int amdgpu_vm_lock_done_list(struct amdgpu_vm *vm, struct drm_exec *exec,
508 			     unsigned int num_fences);
509 bool amdgpu_vm_ready(struct amdgpu_vm *vm);
510 uint64_t amdgpu_vm_generation(struct amdgpu_device *adev, struct amdgpu_vm *vm);
511 int amdgpu_vm_validate(struct amdgpu_device *adev, struct amdgpu_vm *vm,
512 		       struct ww_acquire_ctx *ticket,
513 		       int (*callback)(void *p, struct amdgpu_bo *bo),
514 		       void *param);
515 int amdgpu_vm_flush(struct amdgpu_ring *ring, struct amdgpu_job *job, bool need_pipe_sync);
516 int amdgpu_vm_update_pdes(struct amdgpu_device *adev,
517 			  struct amdgpu_vm *vm, bool immediate);
518 int amdgpu_vm_clear_freed(struct amdgpu_device *adev,
519 			  struct amdgpu_vm *vm,
520 			  struct dma_fence **fence);
521 int amdgpu_vm_handle_moved(struct amdgpu_device *adev,
522 			   struct amdgpu_vm *vm,
523 			   struct ww_acquire_ctx *ticket);
524 int amdgpu_vm_flush_compute_tlb(struct amdgpu_device *adev,
525 				struct amdgpu_vm *vm,
526 				uint32_t flush_type,
527 				uint32_t xcc_mask);
528 void amdgpu_vm_bo_base_init(struct amdgpu_vm_bo_base *base,
529 			    struct amdgpu_vm *vm, struct amdgpu_bo *bo);
530 int amdgpu_vm_update_range(struct amdgpu_device *adev, struct amdgpu_vm *vm,
531 			   bool immediate, bool unlocked, bool flush_tlb,
532 			   bool allow_override, struct amdgpu_sync *sync,
533 			   uint64_t start, uint64_t last, uint64_t flags,
534 			   uint64_t offset, uint64_t vram_base,
535 			   struct ttm_resource *res, dma_addr_t *pages_addr,
536 			   struct dma_fence **fence);
537 int amdgpu_vm_bo_update(struct amdgpu_device *adev,
538 			struct amdgpu_bo_va *bo_va,
539 			bool clear);
540 bool amdgpu_vm_evictable(struct amdgpu_bo *bo);
541 void amdgpu_vm_bo_invalidate(struct amdgpu_bo *bo, bool evicted);
542 void amdgpu_vm_update_stats(struct amdgpu_vm_bo_base *base,
543 			    struct ttm_resource *new_res, int sign);
544 void amdgpu_vm_bo_update_shared(struct amdgpu_bo *bo);
545 void amdgpu_vm_bo_move(struct amdgpu_bo *bo, struct ttm_resource *new_mem,
546 		       bool evicted);
547 uint64_t amdgpu_vm_map_gart(const dma_addr_t *pages_addr, uint64_t addr);
548 struct amdgpu_bo_va *amdgpu_vm_bo_find(struct amdgpu_vm *vm,
549 				       struct amdgpu_bo *bo);
550 struct amdgpu_bo_va *amdgpu_vm_bo_add(struct amdgpu_device *adev,
551 				      struct amdgpu_vm *vm,
552 				      struct amdgpu_bo *bo);
553 int amdgpu_vm_bo_map(struct amdgpu_device *adev,
554 		     struct amdgpu_bo_va *bo_va,
555 		     uint64_t addr, uint64_t offset,
556 		     uint64_t size, uint32_t flags);
557 int amdgpu_vm_bo_replace_map(struct amdgpu_device *adev,
558 			     struct amdgpu_bo_va *bo_va,
559 			     uint64_t addr, uint64_t offset,
560 			     uint64_t size, uint32_t flags);
561 int amdgpu_vm_bo_unmap(struct amdgpu_device *adev,
562 		       struct amdgpu_bo_va *bo_va,
563 		       uint64_t addr);
564 int amdgpu_vm_bo_clear_mappings(struct amdgpu_device *adev,
565 				struct amdgpu_vm *vm,
566 				uint64_t saddr, uint64_t size);
567 struct amdgpu_bo_va_mapping *amdgpu_vm_bo_lookup_mapping(struct amdgpu_vm *vm,
568 							 uint64_t addr);
569 void amdgpu_vm_bo_trace_cs(struct amdgpu_vm *vm, struct ww_acquire_ctx *ticket);
570 void amdgpu_vm_bo_del(struct amdgpu_device *adev,
571 		      struct amdgpu_bo_va *bo_va);
572 void amdgpu_vm_adjust_size(struct amdgpu_device *adev, uint32_t min_vm_size,
573 			   uint32_t fragment_size_default, unsigned max_level,
574 			   unsigned max_bits);
575 int amdgpu_vm_ioctl(struct drm_device *dev, void *data, struct drm_file *filp);
576 bool amdgpu_vm_need_pipeline_sync(struct amdgpu_ring *ring,
577 				  struct amdgpu_job *job);
578 void amdgpu_vm_check_compute_bug(struct amdgpu_device *adev);
579 
580 struct amdgpu_task_info *
581 amdgpu_vm_get_task_info_pasid(struct amdgpu_device *adev, u32 pasid);
582 
583 struct amdgpu_task_info *
584 amdgpu_vm_get_task_info_vm(struct amdgpu_vm *vm);
585 
586 void amdgpu_vm_put_task_info(struct amdgpu_task_info *task_info);
587 
588 bool amdgpu_vm_handle_fault(struct amdgpu_device *adev, u32 pasid,
589 			    u32 vmid, u32 node_id, uint64_t addr, uint64_t ts,
590 			    bool write_fault);
591 
592 void amdgpu_vm_set_task_info(struct amdgpu_vm *vm);
593 
594 void amdgpu_vm_move_to_lru_tail(struct amdgpu_device *adev,
595 				struct amdgpu_vm *vm);
596 void amdgpu_vm_get_memory(struct amdgpu_vm *vm,
597 			  struct amdgpu_mem_stats stats[__AMDGPU_PL_NUM]);
598 
599 int amdgpu_vm_pt_clear(struct amdgpu_device *adev, struct amdgpu_vm *vm,
600 		       struct amdgpu_bo_vm *vmbo, bool immediate);
601 int amdgpu_vm_pt_create(struct amdgpu_device *adev, struct amdgpu_vm *vm,
602 			int level, bool immediate, struct amdgpu_bo_vm **vmbo,
603 			int32_t xcp_id);
604 void amdgpu_vm_pt_free_root(struct amdgpu_device *adev, struct amdgpu_vm *vm);
605 
606 int amdgpu_vm_pde_update(struct amdgpu_vm_update_params *params,
607 			 struct amdgpu_vm_bo_base *entry);
608 int amdgpu_vm_ptes_update(struct amdgpu_vm_update_params *params,
609 			  uint64_t start, uint64_t end,
610 			  uint64_t dst, uint64_t flags);
611 void amdgpu_vm_pt_free_work(struct work_struct *work);
612 void amdgpu_vm_pt_free_list(struct amdgpu_device *adev,
613 			    struct amdgpu_vm_update_params *params);
614 
615 #if defined(CONFIG_DEBUG_FS)
616 void amdgpu_debugfs_vm_bo_info(struct amdgpu_vm *vm, struct seq_file *m);
617 #endif
618 
619 int amdgpu_vm_pt_map_tables(struct amdgpu_device *adev, struct amdgpu_vm *vm);
620 
621 bool amdgpu_vm_is_bo_always_valid(struct amdgpu_vm *vm, struct amdgpu_bo *bo);
622 
623 /**
624  * amdgpu_vm_tlb_seq - return tlb flush sequence number
625  * @vm: the amdgpu_vm structure to query
626  *
627  * Returns the tlb flush sequence number which indicates that the VM TLBs needs
628  * to be invalidated whenever the sequence number change.
629  */
630 static inline uint64_t amdgpu_vm_tlb_seq(struct amdgpu_vm *vm)
631 {
632 	unsigned long flags;
633 	spinlock_t *lock;
634 
635 	/*
636 	 * Workaround to stop racing between the fence signaling and handling
637 	 * the cb. The lock is static after initially setting it up, just make
638 	 * sure that the dma_fence structure isn't freed up.
639 	 */
640 	rcu_read_lock();
641 	lock = vm->last_tlb_flush->lock;
642 	rcu_read_unlock();
643 
644 	spin_lock_irqsave(lock, flags);
645 	spin_unlock_irqrestore(lock, flags);
646 
647 	return atomic64_read(&vm->tlb_seq);
648 }
649 
650 /*
651  * vm eviction_lock can be taken in MMU notifiers. Make sure no reclaim-FS
652  * happens while holding this lock anywhere to prevent deadlocks when
653  * an MMU notifier runs in reclaim-FS context.
654  */
655 static inline void amdgpu_vm_eviction_lock(struct amdgpu_vm *vm)
656 {
657 	mutex_lock(&vm->eviction_lock);
658 	vm->saved_flags = memalloc_noreclaim_save();
659 }
660 
661 static inline bool amdgpu_vm_eviction_trylock(struct amdgpu_vm *vm)
662 {
663 	if (mutex_trylock(&vm->eviction_lock)) {
664 		vm->saved_flags = memalloc_noreclaim_save();
665 		return true;
666 	}
667 	return false;
668 }
669 
670 static inline void amdgpu_vm_eviction_unlock(struct amdgpu_vm *vm)
671 {
672 	memalloc_noreclaim_restore(vm->saved_flags);
673 	mutex_unlock(&vm->eviction_lock);
674 }
675 
676 void amdgpu_vm_update_fault_cache(struct amdgpu_device *adev,
677 				  unsigned int pasid,
678 				  uint64_t addr,
679 				  uint32_t status,
680 				  unsigned int vmhub);
681 void amdgpu_vm_tlb_fence_create(struct amdgpu_device *adev,
682 				 struct amdgpu_vm *vm,
683 				 struct dma_fence **fence);
684 
685 void amdgpu_vm_print_task_info(struct amdgpu_device *adev,
686 			       struct amdgpu_task_info *task_info);
687 
688 #define amdgpu_vm_bo_va_for_each_valid_mapping(bo_va, mapping) \
689 		list_for_each_entry(mapping, &(bo_va)->valids, list)
690 #define amdgpu_vm_bo_va_for_each_invalid_mapping(bo_va, mapping) \
691 		list_for_each_entry(mapping, &(bo_va)->invalids, list)
692 
693 #endif
694