xref: /linux/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.h (revision 3a39d672e7f48b8d6b91a09afa4b55352773b4b5)
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_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_IS_PTE		(1ULL << 63)
133 
134 /* PDE Block Fragment Size for gfx v12 */
135 #define AMDGPU_PDE_BFS_GFX12(a)		((uint64_t)((a) & 0x1fULL) << 58)
136 #define AMDGPU_PDE_BFS_FLAG(adev, a)	\
137 	((amdgpu_ip_version((adev), GC_HWIP, 0) >= IP_VERSION(12, 0, 0)) ? AMDGPU_PDE_BFS_GFX12(a) : AMDGPU_PDE_BFS(a))
138 /* PDE is handled as PTE for gfx v12 */
139 #define AMDGPU_PDE_PTE_GFX12		(1ULL << 63)
140 #define AMDGPU_PDE_PTE_FLAG(adev)	\
141 	((amdgpu_ip_version((adev), GC_HWIP, 0) >= IP_VERSION(12, 0, 0)) ? AMDGPU_PDE_PTE_GFX12 : AMDGPU_PDE_PTE)
142 
143 /* How to program VM fault handling */
144 #define AMDGPU_VM_FAULT_STOP_NEVER	0
145 #define AMDGPU_VM_FAULT_STOP_FIRST	1
146 #define AMDGPU_VM_FAULT_STOP_ALWAYS	2
147 
148 /* How much VRAM be reserved for page tables */
149 #define AMDGPU_VM_RESERVED_VRAM		(8ULL << 20)
150 
151 /*
152  * max number of VMHUB
153  * layout: max 8 GFXHUB + 4 MMHUB0 + 1 MMHUB1
154  */
155 #define AMDGPU_MAX_VMHUBS			13
156 #define AMDGPU_GFXHUB_START			0
157 #define AMDGPU_MMHUB0_START			8
158 #define AMDGPU_MMHUB1_START			12
159 #define AMDGPU_GFXHUB(x)			(AMDGPU_GFXHUB_START + (x))
160 #define AMDGPU_MMHUB0(x)			(AMDGPU_MMHUB0_START + (x))
161 #define AMDGPU_MMHUB1(x)			(AMDGPU_MMHUB1_START + (x))
162 
163 #define AMDGPU_IS_GFXHUB(x) ((x) >= AMDGPU_GFXHUB_START && (x) < AMDGPU_MMHUB0_START)
164 #define AMDGPU_IS_MMHUB0(x) ((x) >= AMDGPU_MMHUB0_START && (x) < AMDGPU_MMHUB1_START)
165 #define AMDGPU_IS_MMHUB1(x) ((x) >= AMDGPU_MMHUB1_START && (x) < AMDGPU_MAX_VMHUBS)
166 
167 /* Reserve space at top/bottom of address space for kernel use */
168 #define AMDGPU_VA_RESERVED_CSA_SIZE		(2ULL << 20)
169 #define AMDGPU_VA_RESERVED_CSA_START(adev)	(((adev)->vm_manager.max_pfn \
170 						  << AMDGPU_GPU_PAGE_SHIFT)  \
171 						 - AMDGPU_VA_RESERVED_CSA_SIZE)
172 #define AMDGPU_VA_RESERVED_SEQ64_SIZE		(2ULL << 20)
173 #define AMDGPU_VA_RESERVED_SEQ64_START(adev)	(AMDGPU_VA_RESERVED_CSA_START(adev) \
174 						 - AMDGPU_VA_RESERVED_SEQ64_SIZE)
175 #define AMDGPU_VA_RESERVED_TRAP_SIZE		(2ULL << 12)
176 #define AMDGPU_VA_RESERVED_TRAP_START(adev)	(AMDGPU_VA_RESERVED_SEQ64_START(adev) \
177 						 - AMDGPU_VA_RESERVED_TRAP_SIZE)
178 #define AMDGPU_VA_RESERVED_BOTTOM		(1ULL << 16)
179 #define AMDGPU_VA_RESERVED_TOP			(AMDGPU_VA_RESERVED_TRAP_SIZE + \
180 						 AMDGPU_VA_RESERVED_SEQ64_SIZE + \
181 						 AMDGPU_VA_RESERVED_CSA_SIZE)
182 
183 /* See vm_update_mode */
184 #define AMDGPU_VM_USE_CPU_FOR_GFX (1 << 0)
185 #define AMDGPU_VM_USE_CPU_FOR_COMPUTE (1 << 1)
186 
187 /* VMPT level enumerate, and the hiberachy is:
188  * PDB2->PDB1->PDB0->PTB
189  */
190 enum amdgpu_vm_level {
191 	AMDGPU_VM_PDB2,
192 	AMDGPU_VM_PDB1,
193 	AMDGPU_VM_PDB0,
194 	AMDGPU_VM_PTB
195 };
196 
197 /* base structure for tracking BO usage in a VM */
198 struct amdgpu_vm_bo_base {
199 	/* constant after initialization */
200 	struct amdgpu_vm		*vm;
201 	struct amdgpu_bo		*bo;
202 
203 	/* protected by bo being reserved */
204 	struct amdgpu_vm_bo_base	*next;
205 
206 	/* protected by spinlock */
207 	struct list_head		vm_status;
208 
209 	/* protected by the BO being reserved */
210 	bool				moved;
211 };
212 
213 /* provided by hw blocks that can write ptes, e.g., sdma */
214 struct amdgpu_vm_pte_funcs {
215 	/* number of dw to reserve per operation */
216 	unsigned	copy_pte_num_dw;
217 
218 	/* copy pte entries from GART */
219 	void (*copy_pte)(struct amdgpu_ib *ib,
220 			 uint64_t pe, uint64_t src,
221 			 unsigned count);
222 
223 	/* write pte one entry at a time with addr mapping */
224 	void (*write_pte)(struct amdgpu_ib *ib, uint64_t pe,
225 			  uint64_t value, unsigned count,
226 			  uint32_t incr);
227 	/* for linear pte/pde updates without addr mapping */
228 	void (*set_pte_pde)(struct amdgpu_ib *ib,
229 			    uint64_t pe,
230 			    uint64_t addr, unsigned count,
231 			    uint32_t incr, uint64_t flags);
232 };
233 
234 struct amdgpu_task_info {
235 	char		process_name[TASK_COMM_LEN];
236 	char		task_name[TASK_COMM_LEN];
237 	pid_t		pid;
238 	pid_t		tgid;
239 	struct kref	refcount;
240 };
241 
242 /**
243  * struct amdgpu_vm_update_params
244  *
245  * Encapsulate some VM table update parameters to reduce
246  * the number of function parameters
247  *
248  */
249 struct amdgpu_vm_update_params {
250 
251 	/**
252 	 * @adev: amdgpu device we do this update for
253 	 */
254 	struct amdgpu_device *adev;
255 
256 	/**
257 	 * @vm: optional amdgpu_vm we do this update for
258 	 */
259 	struct amdgpu_vm *vm;
260 
261 	/**
262 	 * @immediate: if changes should be made immediately
263 	 */
264 	bool immediate;
265 
266 	/**
267 	 * @unlocked: true if the root BO is not locked
268 	 */
269 	bool unlocked;
270 
271 	/**
272 	 * @pages_addr:
273 	 *
274 	 * DMA addresses to use for mapping
275 	 */
276 	dma_addr_t *pages_addr;
277 
278 	/**
279 	 * @job: job to used for hw submission
280 	 */
281 	struct amdgpu_job *job;
282 
283 	/**
284 	 * @num_dw_left: number of dw left for the IB
285 	 */
286 	unsigned int num_dw_left;
287 
288 	/**
289 	 * @needs_flush: true whenever we need to invalidate the TLB
290 	 */
291 	bool needs_flush;
292 
293 	/**
294 	 * @allow_override: true for memory that is not uncached: allows MTYPE
295 	 * to be overridden for NUMA local memory.
296 	 */
297 	bool allow_override;
298 
299 	/**
300 	 * @tlb_flush_waitlist: temporary storage for BOs until tlb_flush
301 	 */
302 	struct list_head tlb_flush_waitlist;
303 };
304 
305 struct amdgpu_vm_update_funcs {
306 	int (*map_table)(struct amdgpu_bo_vm *bo);
307 	int (*prepare)(struct amdgpu_vm_update_params *p,
308 		       struct amdgpu_sync *sync);
309 	int (*update)(struct amdgpu_vm_update_params *p,
310 		      struct amdgpu_bo_vm *bo, uint64_t pe, uint64_t addr,
311 		      unsigned count, uint32_t incr, uint64_t flags);
312 	int (*commit)(struct amdgpu_vm_update_params *p,
313 		      struct dma_fence **fence);
314 };
315 
316 struct amdgpu_vm_fault_info {
317 	/* fault address */
318 	uint64_t	addr;
319 	/* fault status register */
320 	uint32_t	status;
321 	/* which vmhub? gfxhub, mmhub, etc. */
322 	unsigned int	vmhub;
323 };
324 
325 struct amdgpu_vm {
326 	/* tree of virtual addresses mapped */
327 	struct rb_root_cached	va;
328 
329 	/* Lock to prevent eviction while we are updating page tables
330 	 * use vm_eviction_lock/unlock(vm)
331 	 */
332 	struct mutex		eviction_lock;
333 	bool			evicting;
334 	unsigned int		saved_flags;
335 
336 	/* Lock to protect vm_bo add/del/move on all lists of vm */
337 	spinlock_t		status_lock;
338 
339 	/* Per-VM and PT BOs who needs a validation */
340 	struct list_head	evicted;
341 
342 	/* BOs for user mode queues that need a validation */
343 	struct list_head	evicted_user;
344 
345 	/* PT BOs which relocated and their parent need an update */
346 	struct list_head	relocated;
347 
348 	/* per VM BOs moved, but not yet updated in the PT */
349 	struct list_head	moved;
350 
351 	/* All BOs of this VM not currently in the state machine */
352 	struct list_head	idle;
353 
354 	/* regular invalidated BOs, but not yet updated in the PT */
355 	struct list_head	invalidated;
356 
357 	/* BO mappings freed, but not yet updated in the PT */
358 	struct list_head	freed;
359 
360 	/* BOs which are invalidated, has been updated in the PTs */
361 	struct list_head        done;
362 
363 	/* PT BOs scheduled to free and fill with zero if vm_resv is not hold */
364 	struct list_head	pt_freed;
365 	struct work_struct	pt_free_work;
366 
367 	/* contains the page directory */
368 	struct amdgpu_vm_bo_base     root;
369 	struct dma_fence	*last_update;
370 
371 	/* Scheduler entities for page table updates */
372 	struct drm_sched_entity	immediate;
373 	struct drm_sched_entity	delayed;
374 
375 	/* Last finished delayed update */
376 	atomic64_t		tlb_seq;
377 	struct dma_fence	*last_tlb_flush;
378 	atomic64_t		kfd_last_flushed_seq;
379 	uint64_t		tlb_fence_context;
380 
381 	/* How many times we had to re-generate the page tables */
382 	uint64_t		generation;
383 
384 	/* Last unlocked submission to the scheduler entities */
385 	struct dma_fence	*last_unlocked;
386 
387 	unsigned int		pasid;
388 	bool			reserved_vmid[AMDGPU_MAX_VMHUBS];
389 
390 	/* Flag to indicate if VM tables are updated by CPU or GPU (SDMA) */
391 	bool					use_cpu_for_update;
392 
393 	/* Functions to use for VM table updates */
394 	const struct amdgpu_vm_update_funcs	*update_funcs;
395 
396 	/* Up to 128 pending retry page faults */
397 	DECLARE_KFIFO(faults, u64, 128);
398 
399 	/* Points to the KFD process VM info */
400 	struct amdkfd_process_info *process_info;
401 
402 	/* List node in amdkfd_process_info.vm_list_head */
403 	struct list_head	vm_list_node;
404 
405 	/* Valid while the PD is reserved or fenced */
406 	uint64_t		pd_phys_addr;
407 
408 	/* Some basic info about the task */
409 	struct amdgpu_task_info *task_info;
410 
411 	/* Store positions of group of BOs */
412 	struct ttm_lru_bulk_move lru_bulk_move;
413 	/* Flag to indicate if VM is used for compute */
414 	bool			is_compute_context;
415 
416 	/* Memory partition number, -1 means any partition */
417 	int8_t			mem_id;
418 
419 	/* cached fault info */
420 	struct amdgpu_vm_fault_info fault_info;
421 };
422 
423 struct amdgpu_vm_manager {
424 	/* Handling of VMIDs */
425 	struct amdgpu_vmid_mgr			id_mgr[AMDGPU_MAX_VMHUBS];
426 	unsigned int				first_kfd_vmid;
427 	bool					concurrent_flush;
428 
429 	/* Handling of VM fences */
430 	u64					fence_context;
431 	unsigned				seqno[AMDGPU_MAX_RINGS];
432 
433 	uint64_t				max_pfn;
434 	uint32_t				num_level;
435 	uint32_t				block_size;
436 	uint32_t				fragment_size;
437 	enum amdgpu_vm_level			root_level;
438 	/* vram base address for page table entry  */
439 	u64					vram_base_offset;
440 	/* vm pte handling */
441 	const struct amdgpu_vm_pte_funcs	*vm_pte_funcs;
442 	struct drm_gpu_scheduler		*vm_pte_scheds[AMDGPU_MAX_RINGS];
443 	unsigned				vm_pte_num_scheds;
444 	struct amdgpu_ring			*page_fault;
445 
446 	/* partial resident texture handling */
447 	spinlock_t				prt_lock;
448 	atomic_t				num_prt_users;
449 
450 	/* controls how VM page tables are updated for Graphics and Compute.
451 	 * BIT0[= 0] Graphics updated by SDMA [= 1] by CPU
452 	 * BIT1[= 0] Compute updated by SDMA [= 1] by CPU
453 	 */
454 	int					vm_update_mode;
455 
456 	/* PASID to VM mapping, will be used in interrupt context to
457 	 * look up VM of a page fault
458 	 */
459 	struct xarray				pasids;
460 	/* Global registration of recent page fault information */
461 	struct amdgpu_vm_fault_info	fault_info;
462 };
463 
464 struct amdgpu_bo_va_mapping;
465 
466 #define amdgpu_vm_copy_pte(adev, ib, pe, src, count) ((adev)->vm_manager.vm_pte_funcs->copy_pte((ib), (pe), (src), (count)))
467 #define amdgpu_vm_write_pte(adev, ib, pe, value, count, incr) ((adev)->vm_manager.vm_pte_funcs->write_pte((ib), (pe), (value), (count), (incr)))
468 #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)))
469 
470 extern const struct amdgpu_vm_update_funcs amdgpu_vm_cpu_funcs;
471 extern const struct amdgpu_vm_update_funcs amdgpu_vm_sdma_funcs;
472 
473 void amdgpu_vm_manager_init(struct amdgpu_device *adev);
474 void amdgpu_vm_manager_fini(struct amdgpu_device *adev);
475 
476 int amdgpu_vm_set_pasid(struct amdgpu_device *adev, struct amdgpu_vm *vm,
477 			u32 pasid);
478 
479 long amdgpu_vm_wait_idle(struct amdgpu_vm *vm, long timeout);
480 int amdgpu_vm_init(struct amdgpu_device *adev, struct amdgpu_vm *vm, int32_t xcp_id);
481 int amdgpu_vm_make_compute(struct amdgpu_device *adev, struct amdgpu_vm *vm);
482 void amdgpu_vm_release_compute(struct amdgpu_device *adev, struct amdgpu_vm *vm);
483 void amdgpu_vm_fini(struct amdgpu_device *adev, struct amdgpu_vm *vm);
484 int amdgpu_vm_lock_pd(struct amdgpu_vm *vm, struct drm_exec *exec,
485 		      unsigned int num_fences);
486 bool amdgpu_vm_ready(struct amdgpu_vm *vm);
487 uint64_t amdgpu_vm_generation(struct amdgpu_device *adev, struct amdgpu_vm *vm);
488 int amdgpu_vm_validate(struct amdgpu_device *adev, struct amdgpu_vm *vm,
489 		       struct ww_acquire_ctx *ticket,
490 		       int (*callback)(void *p, struct amdgpu_bo *bo),
491 		       void *param);
492 int amdgpu_vm_flush(struct amdgpu_ring *ring, struct amdgpu_job *job, bool need_pipe_sync);
493 int amdgpu_vm_update_pdes(struct amdgpu_device *adev,
494 			  struct amdgpu_vm *vm, bool immediate);
495 int amdgpu_vm_clear_freed(struct amdgpu_device *adev,
496 			  struct amdgpu_vm *vm,
497 			  struct dma_fence **fence);
498 int amdgpu_vm_handle_moved(struct amdgpu_device *adev,
499 			   struct amdgpu_vm *vm,
500 			   struct ww_acquire_ctx *ticket);
501 int amdgpu_vm_flush_compute_tlb(struct amdgpu_device *adev,
502 				struct amdgpu_vm *vm,
503 				uint32_t flush_type,
504 				uint32_t xcc_mask);
505 void amdgpu_vm_bo_base_init(struct amdgpu_vm_bo_base *base,
506 			    struct amdgpu_vm *vm, struct amdgpu_bo *bo);
507 int amdgpu_vm_update_range(struct amdgpu_device *adev, struct amdgpu_vm *vm,
508 			   bool immediate, bool unlocked, bool flush_tlb,
509 			   bool allow_override, struct amdgpu_sync *sync,
510 			   uint64_t start, uint64_t last, uint64_t flags,
511 			   uint64_t offset, uint64_t vram_base,
512 			   struct ttm_resource *res, dma_addr_t *pages_addr,
513 			   struct dma_fence **fence);
514 int amdgpu_vm_bo_update(struct amdgpu_device *adev,
515 			struct amdgpu_bo_va *bo_va,
516 			bool clear);
517 bool amdgpu_vm_evictable(struct amdgpu_bo *bo);
518 void amdgpu_vm_bo_invalidate(struct amdgpu_device *adev,
519 			     struct amdgpu_bo *bo, bool evicted);
520 uint64_t amdgpu_vm_map_gart(const dma_addr_t *pages_addr, uint64_t addr);
521 struct amdgpu_bo_va *amdgpu_vm_bo_find(struct amdgpu_vm *vm,
522 				       struct amdgpu_bo *bo);
523 struct amdgpu_bo_va *amdgpu_vm_bo_add(struct amdgpu_device *adev,
524 				      struct amdgpu_vm *vm,
525 				      struct amdgpu_bo *bo);
526 int amdgpu_vm_bo_map(struct amdgpu_device *adev,
527 		     struct amdgpu_bo_va *bo_va,
528 		     uint64_t addr, uint64_t offset,
529 		     uint64_t size, uint64_t flags);
530 int amdgpu_vm_bo_replace_map(struct amdgpu_device *adev,
531 			     struct amdgpu_bo_va *bo_va,
532 			     uint64_t addr, uint64_t offset,
533 			     uint64_t size, uint64_t flags);
534 int amdgpu_vm_bo_unmap(struct amdgpu_device *adev,
535 		       struct amdgpu_bo_va *bo_va,
536 		       uint64_t addr);
537 int amdgpu_vm_bo_clear_mappings(struct amdgpu_device *adev,
538 				struct amdgpu_vm *vm,
539 				uint64_t saddr, uint64_t size);
540 struct amdgpu_bo_va_mapping *amdgpu_vm_bo_lookup_mapping(struct amdgpu_vm *vm,
541 							 uint64_t addr);
542 void amdgpu_vm_bo_trace_cs(struct amdgpu_vm *vm, struct ww_acquire_ctx *ticket);
543 void amdgpu_vm_bo_del(struct amdgpu_device *adev,
544 		      struct amdgpu_bo_va *bo_va);
545 void amdgpu_vm_adjust_size(struct amdgpu_device *adev, uint32_t min_vm_size,
546 			   uint32_t fragment_size_default, unsigned max_level,
547 			   unsigned max_bits);
548 int amdgpu_vm_ioctl(struct drm_device *dev, void *data, struct drm_file *filp);
549 bool amdgpu_vm_need_pipeline_sync(struct amdgpu_ring *ring,
550 				  struct amdgpu_job *job);
551 void amdgpu_vm_check_compute_bug(struct amdgpu_device *adev);
552 
553 struct amdgpu_task_info *
554 amdgpu_vm_get_task_info_pasid(struct amdgpu_device *adev, u32 pasid);
555 
556 struct amdgpu_task_info *
557 amdgpu_vm_get_task_info_vm(struct amdgpu_vm *vm);
558 
559 void amdgpu_vm_put_task_info(struct amdgpu_task_info *task_info);
560 
561 bool amdgpu_vm_handle_fault(struct amdgpu_device *adev, u32 pasid,
562 			    u32 vmid, u32 node_id, uint64_t addr, uint64_t ts,
563 			    bool write_fault);
564 
565 void amdgpu_vm_set_task_info(struct amdgpu_vm *vm);
566 
567 void amdgpu_vm_move_to_lru_tail(struct amdgpu_device *adev,
568 				struct amdgpu_vm *vm);
569 void amdgpu_vm_get_memory(struct amdgpu_vm *vm,
570 			  struct amdgpu_mem_stats *stats);
571 
572 int amdgpu_vm_pt_clear(struct amdgpu_device *adev, struct amdgpu_vm *vm,
573 		       struct amdgpu_bo_vm *vmbo, bool immediate);
574 int amdgpu_vm_pt_create(struct amdgpu_device *adev, struct amdgpu_vm *vm,
575 			int level, bool immediate, struct amdgpu_bo_vm **vmbo,
576 			int32_t xcp_id);
577 void amdgpu_vm_pt_free_root(struct amdgpu_device *adev, struct amdgpu_vm *vm);
578 
579 int amdgpu_vm_pde_update(struct amdgpu_vm_update_params *params,
580 			 struct amdgpu_vm_bo_base *entry);
581 int amdgpu_vm_ptes_update(struct amdgpu_vm_update_params *params,
582 			  uint64_t start, uint64_t end,
583 			  uint64_t dst, uint64_t flags);
584 void amdgpu_vm_pt_free_work(struct work_struct *work);
585 void amdgpu_vm_pt_free_list(struct amdgpu_device *adev,
586 			    struct amdgpu_vm_update_params *params);
587 
588 #if defined(CONFIG_DEBUG_FS)
589 void amdgpu_debugfs_vm_bo_info(struct amdgpu_vm *vm, struct seq_file *m);
590 #endif
591 
592 int amdgpu_vm_pt_map_tables(struct amdgpu_device *adev, struct amdgpu_vm *vm);
593 
594 bool amdgpu_vm_is_bo_always_valid(struct amdgpu_vm *vm, struct amdgpu_bo *bo);
595 
596 /**
597  * amdgpu_vm_tlb_seq - return tlb flush sequence number
598  * @vm: the amdgpu_vm structure to query
599  *
600  * Returns the tlb flush sequence number which indicates that the VM TLBs needs
601  * to be invalidated whenever the sequence number change.
602  */
amdgpu_vm_tlb_seq(struct amdgpu_vm * vm)603 static inline uint64_t amdgpu_vm_tlb_seq(struct amdgpu_vm *vm)
604 {
605 	unsigned long flags;
606 	spinlock_t *lock;
607 
608 	/*
609 	 * Workaround to stop racing between the fence signaling and handling
610 	 * the cb. The lock is static after initially setting it up, just make
611 	 * sure that the dma_fence structure isn't freed up.
612 	 */
613 	rcu_read_lock();
614 	lock = vm->last_tlb_flush->lock;
615 	rcu_read_unlock();
616 
617 	spin_lock_irqsave(lock, flags);
618 	spin_unlock_irqrestore(lock, flags);
619 
620 	return atomic64_read(&vm->tlb_seq);
621 }
622 
623 /*
624  * vm eviction_lock can be taken in MMU notifiers. Make sure no reclaim-FS
625  * happens while holding this lock anywhere to prevent deadlocks when
626  * an MMU notifier runs in reclaim-FS context.
627  */
amdgpu_vm_eviction_lock(struct amdgpu_vm * vm)628 static inline void amdgpu_vm_eviction_lock(struct amdgpu_vm *vm)
629 {
630 	mutex_lock(&vm->eviction_lock);
631 	vm->saved_flags = memalloc_noreclaim_save();
632 }
633 
amdgpu_vm_eviction_trylock(struct amdgpu_vm * vm)634 static inline bool amdgpu_vm_eviction_trylock(struct amdgpu_vm *vm)
635 {
636 	if (mutex_trylock(&vm->eviction_lock)) {
637 		vm->saved_flags = memalloc_noreclaim_save();
638 		return true;
639 	}
640 	return false;
641 }
642 
amdgpu_vm_eviction_unlock(struct amdgpu_vm * vm)643 static inline void amdgpu_vm_eviction_unlock(struct amdgpu_vm *vm)
644 {
645 	memalloc_noreclaim_restore(vm->saved_flags);
646 	mutex_unlock(&vm->eviction_lock);
647 }
648 
649 void amdgpu_vm_update_fault_cache(struct amdgpu_device *adev,
650 				  unsigned int pasid,
651 				  uint64_t addr,
652 				  uint32_t status,
653 				  unsigned int vmhub);
654 void amdgpu_vm_tlb_fence_create(struct amdgpu_device *adev,
655 				 struct amdgpu_vm *vm,
656 				 struct dma_fence **fence);
657 
658 #endif
659