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_driver.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 amdgpu_bo_va;
40 struct amdgpu_job;
41 struct amdgpu_bo_list_entry;
42 struct amdgpu_bo_vm;
43 
44 /*
45  * GPUVM handling
46  */
47 
48 /* Maximum number of PTEs the hardware can write with one command */
49 #define AMDGPU_VM_MAX_UPDATE_SIZE	0x3FFFF
50 
51 /* number of entries in page table */
52 #define AMDGPU_VM_PTE_COUNT(adev) (1 << (adev)->vm_manager.block_size)
53 
54 #define AMDGPU_PTE_VALID	(1ULL << 0)
55 #define AMDGPU_PTE_SYSTEM	(1ULL << 1)
56 #define AMDGPU_PTE_SNOOPED	(1ULL << 2)
57 
58 /* RV+ */
59 #define AMDGPU_PTE_TMZ		(1ULL << 3)
60 
61 /* VI only */
62 #define AMDGPU_PTE_EXECUTABLE	(1ULL << 4)
63 
64 #define AMDGPU_PTE_READABLE	(1ULL << 5)
65 #define AMDGPU_PTE_WRITEABLE	(1ULL << 6)
66 
67 #define AMDGPU_PTE_FRAG(x)	((x & 0x1fULL) << 7)
68 
69 /* TILED for VEGA10, reserved for older ASICs  */
70 #define AMDGPU_PTE_PRT		(1ULL << 51)
71 
72 /* PDE is handled as PTE for VEGA10 */
73 #define AMDGPU_PDE_PTE		(1ULL << 54)
74 
75 #define AMDGPU_PTE_LOG          (1ULL << 55)
76 
77 /* PTE is handled as PDE for VEGA10 (Translate Further) */
78 #define AMDGPU_PTE_TF		(1ULL << 56)
79 
80 /* MALL noalloc for sienna_cichlid, reserved for older ASICs  */
81 #define AMDGPU_PTE_NOALLOC	(1ULL << 58)
82 
83 /* PDE Block Fragment Size for VEGA10 */
84 #define AMDGPU_PDE_BFS(a)	((uint64_t)a << 59)
85 
86 
87 /* For GFX9 */
88 #define AMDGPU_PTE_MTYPE_VG10(a)	((uint64_t)(a) << 57)
89 #define AMDGPU_PTE_MTYPE_VG10_MASK	AMDGPU_PTE_MTYPE_VG10(3ULL)
90 
91 #define AMDGPU_MTYPE_NC 0
92 #define AMDGPU_MTYPE_CC 2
93 
94 #define AMDGPU_PTE_DEFAULT_ATC  (AMDGPU_PTE_SYSTEM      \
95                                 | AMDGPU_PTE_SNOOPED    \
96                                 | AMDGPU_PTE_EXECUTABLE \
97                                 | AMDGPU_PTE_READABLE   \
98                                 | AMDGPU_PTE_WRITEABLE  \
99                                 | AMDGPU_PTE_MTYPE_VG10(AMDGPU_MTYPE_CC))
100 
101 /* gfx10 */
102 #define AMDGPU_PTE_MTYPE_NV10(a)       ((uint64_t)(a) << 48)
103 #define AMDGPU_PTE_MTYPE_NV10_MASK     AMDGPU_PTE_MTYPE_NV10(7ULL)
104 
105 /* How to program VM fault handling */
106 #define AMDGPU_VM_FAULT_STOP_NEVER	0
107 #define AMDGPU_VM_FAULT_STOP_FIRST	1
108 #define AMDGPU_VM_FAULT_STOP_ALWAYS	2
109 
110 /* Reserve 4MB VRAM for page tables */
111 #define AMDGPU_VM_RESERVED_VRAM		(8ULL << 20)
112 
113 /* max number of VMHUB */
114 #define AMDGPU_MAX_VMHUBS			3
115 #define AMDGPU_GFXHUB_0				0
116 #define AMDGPU_MMHUB_0				1
117 #define AMDGPU_MMHUB_1				2
118 
119 /* Reserve 2MB at top/bottom of address space for kernel use */
120 #define AMDGPU_VA_RESERVED_SIZE			(2ULL << 20)
121 
122 /* See vm_update_mode */
123 #define AMDGPU_VM_USE_CPU_FOR_GFX (1 << 0)
124 #define AMDGPU_VM_USE_CPU_FOR_COMPUTE (1 << 1)
125 
126 /* VMPT level enumerate, and the hiberachy is:
127  * PDB2->PDB1->PDB0->PTB
128  */
129 enum amdgpu_vm_level {
130 	AMDGPU_VM_PDB2,
131 	AMDGPU_VM_PDB1,
132 	AMDGPU_VM_PDB0,
133 	AMDGPU_VM_PTB
134 };
135 
136 /* base structure for tracking BO usage in a VM */
137 struct amdgpu_vm_bo_base {
138 	/* constant after initialization */
139 	struct amdgpu_vm		*vm;
140 	struct amdgpu_bo		*bo;
141 
142 	/* protected by bo being reserved */
143 	struct amdgpu_vm_bo_base	*next;
144 
145 	/* protected by spinlock */
146 	struct list_head		vm_status;
147 
148 	/* protected by the BO being reserved */
149 	bool				moved;
150 };
151 
152 /* provided by hw blocks that can write ptes, e.g., sdma */
153 struct amdgpu_vm_pte_funcs {
154 	/* number of dw to reserve per operation */
155 	unsigned	copy_pte_num_dw;
156 
157 	/* copy pte entries from GART */
158 	void (*copy_pte)(struct amdgpu_ib *ib,
159 			 uint64_t pe, uint64_t src,
160 			 unsigned count);
161 
162 	/* write pte one entry at a time with addr mapping */
163 	void (*write_pte)(struct amdgpu_ib *ib, uint64_t pe,
164 			  uint64_t value, unsigned count,
165 			  uint32_t incr);
166 	/* for linear pte/pde updates without addr mapping */
167 	void (*set_pte_pde)(struct amdgpu_ib *ib,
168 			    uint64_t pe,
169 			    uint64_t addr, unsigned count,
170 			    uint32_t incr, uint64_t flags);
171 };
172 
173 struct amdgpu_task_info {
174 	char	process_name[TASK_COMM_LEN];
175 	char	task_name[TASK_COMM_LEN];
176 	pid_t	pid;
177 	pid_t	tgid;
178 };
179 
180 /**
181  * struct amdgpu_vm_update_params
182  *
183  * Encapsulate some VM table update parameters to reduce
184  * the number of function parameters
185  *
186  */
187 struct amdgpu_vm_update_params {
188 
189 	/**
190 	 * @adev: amdgpu device we do this update for
191 	 */
192 	struct amdgpu_device *adev;
193 
194 	/**
195 	 * @vm: optional amdgpu_vm we do this update for
196 	 */
197 	struct amdgpu_vm *vm;
198 
199 	/**
200 	 * @immediate: if changes should be made immediately
201 	 */
202 	bool immediate;
203 
204 	/**
205 	 * @unlocked: true if the root BO is not locked
206 	 */
207 	bool unlocked;
208 
209 	/**
210 	 * @pages_addr:
211 	 *
212 	 * DMA addresses to use for mapping
213 	 */
214 	dma_addr_t *pages_addr;
215 
216 	/**
217 	 * @job: job to used for hw submission
218 	 */
219 	struct amdgpu_job *job;
220 
221 	/**
222 	 * @num_dw_left: number of dw left for the IB
223 	 */
224 	unsigned int num_dw_left;
225 
226 	/**
227 	 * @table_freed: return true if page table is freed when updating
228 	 */
229 	bool table_freed;
230 };
231 
232 struct amdgpu_vm_update_funcs {
233 	int (*map_table)(struct amdgpu_bo_vm *bo);
234 	int (*prepare)(struct amdgpu_vm_update_params *p, struct dma_resv *resv,
235 		       enum amdgpu_sync_mode sync_mode);
236 	int (*update)(struct amdgpu_vm_update_params *p,
237 		      struct amdgpu_bo_vm *bo, uint64_t pe, uint64_t addr,
238 		      unsigned count, uint32_t incr, uint64_t flags);
239 	int (*commit)(struct amdgpu_vm_update_params *p,
240 		      struct dma_fence **fence);
241 };
242 
243 struct amdgpu_vm {
244 	/* tree of virtual addresses mapped */
245 	struct rb_root_cached	va;
246 
247 	/* Lock to prevent eviction while we are updating page tables
248 	 * use vm_eviction_lock/unlock(vm)
249 	 */
250 	struct mutex		eviction_lock;
251 	bool			evicting;
252 	unsigned int		saved_flags;
253 
254 	/* Lock to protect vm_bo add/del/move on all lists of vm */
255 	spinlock_t		status_lock;
256 
257 	/* BOs who needs a validation */
258 	struct list_head	evicted;
259 
260 	/* PT BOs which relocated and their parent need an update */
261 	struct list_head	relocated;
262 
263 	/* per VM BOs moved, but not yet updated in the PT */
264 	struct list_head	moved;
265 
266 	/* All BOs of this VM not currently in the state machine */
267 	struct list_head	idle;
268 
269 	/* regular invalidated BOs, but not yet updated in the PT */
270 	struct list_head	invalidated;
271 
272 	/* BO mappings freed, but not yet updated in the PT */
273 	struct list_head	freed;
274 
275 	/* BOs which are invalidated, has been updated in the PTs */
276 	struct list_head        done;
277 
278 	/* PT BOs scheduled to free and fill with zero if vm_resv is not hold */
279 	struct list_head	pt_freed;
280 	struct work_struct	pt_free_work;
281 
282 	/* contains the page directory */
283 	struct amdgpu_vm_bo_base     root;
284 	struct dma_fence	*last_update;
285 
286 	/* Scheduler entities for page table updates */
287 	struct drm_sched_entity	immediate;
288 	struct drm_sched_entity	delayed;
289 
290 	/* Last finished delayed update */
291 	atomic64_t		tlb_seq;
292 	struct dma_fence	*last_tlb_flush;
293 
294 	/* Last unlocked submission to the scheduler entities */
295 	struct dma_fence	*last_unlocked;
296 
297 	unsigned int		pasid;
298 	bool			reserved_vmid[AMDGPU_MAX_VMHUBS];
299 
300 	/* Flag to indicate if VM tables are updated by CPU or GPU (SDMA) */
301 	bool					use_cpu_for_update;
302 
303 	/* Functions to use for VM table updates */
304 	const struct amdgpu_vm_update_funcs	*update_funcs;
305 
306 	/* Flag to indicate ATS support from PTE for GFX9 */
307 	bool			pte_support_ats;
308 
309 	/* Up to 128 pending retry page faults */
310 	DECLARE_KFIFO(faults, u64, 128);
311 
312 	/* Points to the KFD process VM info */
313 	struct amdkfd_process_info *process_info;
314 
315 	/* List node in amdkfd_process_info.vm_list_head */
316 	struct list_head	vm_list_node;
317 
318 	/* Valid while the PD is reserved or fenced */
319 	uint64_t		pd_phys_addr;
320 
321 	/* Some basic info about the task */
322 	struct amdgpu_task_info task_info;
323 
324 	/* Store positions of group of BOs */
325 	struct ttm_lru_bulk_move lru_bulk_move;
326 	/* Flag to indicate if VM is used for compute */
327 	bool			is_compute_context;
328 };
329 
330 struct amdgpu_vm_manager {
331 	/* Handling of VMIDs */
332 	struct amdgpu_vmid_mgr			id_mgr[AMDGPU_MAX_VMHUBS];
333 	unsigned int				first_kfd_vmid;
334 	bool					concurrent_flush;
335 
336 	/* Handling of VM fences */
337 	u64					fence_context;
338 	unsigned				seqno[AMDGPU_MAX_RINGS];
339 
340 	uint64_t				max_pfn;
341 	uint32_t				num_level;
342 	uint32_t				block_size;
343 	uint32_t				fragment_size;
344 	enum amdgpu_vm_level			root_level;
345 	/* vram base address for page table entry  */
346 	u64					vram_base_offset;
347 	/* vm pte handling */
348 	const struct amdgpu_vm_pte_funcs	*vm_pte_funcs;
349 	struct drm_gpu_scheduler		*vm_pte_scheds[AMDGPU_MAX_RINGS];
350 	unsigned				vm_pte_num_scheds;
351 	struct amdgpu_ring			*page_fault;
352 
353 	/* partial resident texture handling */
354 	spinlock_t				prt_lock;
355 	atomic_t				num_prt_users;
356 
357 	/* controls how VM page tables are updated for Graphics and Compute.
358 	 * BIT0[= 0] Graphics updated by SDMA [= 1] by CPU
359 	 * BIT1[= 0] Compute updated by SDMA [= 1] by CPU
360 	 */
361 	int					vm_update_mode;
362 
363 	/* PASID to VM mapping, will be used in interrupt context to
364 	 * look up VM of a page fault
365 	 */
366 	struct xarray				pasids;
367 };
368 
369 struct amdgpu_bo_va_mapping;
370 
371 #define amdgpu_vm_copy_pte(adev, ib, pe, src, count) ((adev)->vm_manager.vm_pte_funcs->copy_pte((ib), (pe), (src), (count)))
372 #define amdgpu_vm_write_pte(adev, ib, pe, value, count, incr) ((adev)->vm_manager.vm_pte_funcs->write_pte((ib), (pe), (value), (count), (incr)))
373 #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)))
374 
375 extern const struct amdgpu_vm_update_funcs amdgpu_vm_cpu_funcs;
376 extern const struct amdgpu_vm_update_funcs amdgpu_vm_sdma_funcs;
377 
378 void amdgpu_vm_manager_init(struct amdgpu_device *adev);
379 void amdgpu_vm_manager_fini(struct amdgpu_device *adev);
380 
381 int amdgpu_vm_set_pasid(struct amdgpu_device *adev, struct amdgpu_vm *vm,
382 			u32 pasid);
383 
384 long amdgpu_vm_wait_idle(struct amdgpu_vm *vm, long timeout);
385 int amdgpu_vm_init(struct amdgpu_device *adev, struct amdgpu_vm *vm);
386 int amdgpu_vm_make_compute(struct amdgpu_device *adev, struct amdgpu_vm *vm);
387 void amdgpu_vm_release_compute(struct amdgpu_device *adev, struct amdgpu_vm *vm);
388 void amdgpu_vm_fini(struct amdgpu_device *adev, struct amdgpu_vm *vm);
389 void amdgpu_vm_get_pd_bo(struct amdgpu_vm *vm,
390 			 struct list_head *validated,
391 			 struct amdgpu_bo_list_entry *entry);
392 bool amdgpu_vm_ready(struct amdgpu_vm *vm);
393 int amdgpu_vm_validate_pt_bos(struct amdgpu_device *adev, struct amdgpu_vm *vm,
394 			      int (*callback)(void *p, struct amdgpu_bo *bo),
395 			      void *param);
396 int amdgpu_vm_flush(struct amdgpu_ring *ring, struct amdgpu_job *job, bool need_pipe_sync);
397 int amdgpu_vm_update_pdes(struct amdgpu_device *adev,
398 			  struct amdgpu_vm *vm, bool immediate);
399 int amdgpu_vm_clear_freed(struct amdgpu_device *adev,
400 			  struct amdgpu_vm *vm,
401 			  struct dma_fence **fence);
402 int amdgpu_vm_handle_moved(struct amdgpu_device *adev,
403 			   struct amdgpu_vm *vm);
404 void amdgpu_vm_bo_base_init(struct amdgpu_vm_bo_base *base,
405 			    struct amdgpu_vm *vm, struct amdgpu_bo *bo);
406 int amdgpu_vm_update_range(struct amdgpu_device *adev, struct amdgpu_vm *vm,
407 			   bool immediate, bool unlocked, bool flush_tlb,
408 			   struct dma_resv *resv, uint64_t start, uint64_t last,
409 			   uint64_t flags, uint64_t offset, uint64_t vram_base,
410 			   struct ttm_resource *res, dma_addr_t *pages_addr,
411 			   struct dma_fence **fence);
412 int amdgpu_vm_bo_update(struct amdgpu_device *adev,
413 			struct amdgpu_bo_va *bo_va,
414 			bool clear);
415 bool amdgpu_vm_evictable(struct amdgpu_bo *bo);
416 void amdgpu_vm_bo_invalidate(struct amdgpu_device *adev,
417 			     struct amdgpu_bo *bo, bool evicted);
418 uint64_t amdgpu_vm_map_gart(const dma_addr_t *pages_addr, uint64_t addr);
419 struct amdgpu_bo_va *amdgpu_vm_bo_find(struct amdgpu_vm *vm,
420 				       struct amdgpu_bo *bo);
421 struct amdgpu_bo_va *amdgpu_vm_bo_add(struct amdgpu_device *adev,
422 				      struct amdgpu_vm *vm,
423 				      struct amdgpu_bo *bo);
424 int amdgpu_vm_bo_map(struct amdgpu_device *adev,
425 		     struct amdgpu_bo_va *bo_va,
426 		     uint64_t addr, uint64_t offset,
427 		     uint64_t size, uint64_t flags);
428 int amdgpu_vm_bo_replace_map(struct amdgpu_device *adev,
429 			     struct amdgpu_bo_va *bo_va,
430 			     uint64_t addr, uint64_t offset,
431 			     uint64_t size, uint64_t flags);
432 int amdgpu_vm_bo_unmap(struct amdgpu_device *adev,
433 		       struct amdgpu_bo_va *bo_va,
434 		       uint64_t addr);
435 int amdgpu_vm_bo_clear_mappings(struct amdgpu_device *adev,
436 				struct amdgpu_vm *vm,
437 				uint64_t saddr, uint64_t size);
438 struct amdgpu_bo_va_mapping *amdgpu_vm_bo_lookup_mapping(struct amdgpu_vm *vm,
439 							 uint64_t addr);
440 void amdgpu_vm_bo_trace_cs(struct amdgpu_vm *vm, struct ww_acquire_ctx *ticket);
441 void amdgpu_vm_bo_del(struct amdgpu_device *adev,
442 		      struct amdgpu_bo_va *bo_va);
443 void amdgpu_vm_adjust_size(struct amdgpu_device *adev, uint32_t min_vm_size,
444 			   uint32_t fragment_size_default, unsigned max_level,
445 			   unsigned max_bits);
446 int amdgpu_vm_ioctl(struct drm_device *dev, void *data, struct drm_file *filp);
447 bool amdgpu_vm_need_pipeline_sync(struct amdgpu_ring *ring,
448 				  struct amdgpu_job *job);
449 void amdgpu_vm_check_compute_bug(struct amdgpu_device *adev);
450 
451 void amdgpu_vm_get_task_info(struct amdgpu_device *adev, u32 pasid,
452 			     struct amdgpu_task_info *task_info);
453 bool amdgpu_vm_handle_fault(struct amdgpu_device *adev, u32 pasid,
454 			    uint64_t addr, bool write_fault);
455 
456 void amdgpu_vm_set_task_info(struct amdgpu_vm *vm);
457 
458 void amdgpu_vm_move_to_lru_tail(struct amdgpu_device *adev,
459 				struct amdgpu_vm *vm);
460 void amdgpu_vm_get_memory(struct amdgpu_vm *vm, uint64_t *vram_mem,
461 				uint64_t *gtt_mem, uint64_t *cpu_mem);
462 
463 int amdgpu_vm_pt_clear(struct amdgpu_device *adev, struct amdgpu_vm *vm,
464 		       struct amdgpu_bo_vm *vmbo, bool immediate);
465 int amdgpu_vm_pt_create(struct amdgpu_device *adev, struct amdgpu_vm *vm,
466 			int level, bool immediate, struct amdgpu_bo_vm **vmbo);
467 void amdgpu_vm_pt_free_root(struct amdgpu_device *adev, struct amdgpu_vm *vm);
468 bool amdgpu_vm_pt_is_root_clean(struct amdgpu_device *adev,
469 				struct amdgpu_vm *vm);
470 
471 int amdgpu_vm_pde_update(struct amdgpu_vm_update_params *params,
472 			 struct amdgpu_vm_bo_base *entry);
473 int amdgpu_vm_ptes_update(struct amdgpu_vm_update_params *params,
474 			  uint64_t start, uint64_t end,
475 			  uint64_t dst, uint64_t flags);
476 void amdgpu_vm_pt_free_work(struct work_struct *work);
477 
478 #if defined(CONFIG_DEBUG_FS)
479 void amdgpu_debugfs_vm_bo_info(struct amdgpu_vm *vm, struct seq_file *m);
480 #endif
481 
482 /**
483  * amdgpu_vm_tlb_seq - return tlb flush sequence number
484  * @vm: the amdgpu_vm structure to query
485  *
486  * Returns the tlb flush sequence number which indicates that the VM TLBs needs
487  * to be invalidated whenever the sequence number change.
488  */
489 static inline uint64_t amdgpu_vm_tlb_seq(struct amdgpu_vm *vm)
490 {
491 	unsigned long flags;
492 	spinlock_t *lock;
493 
494 	/*
495 	 * Workaround to stop racing between the fence signaling and handling
496 	 * the cb. The lock is static after initially setting it up, just make
497 	 * sure that the dma_fence structure isn't freed up.
498 	 */
499 	rcu_read_lock();
500 	lock = vm->last_tlb_flush->lock;
501 	rcu_read_unlock();
502 
503 	spin_lock_irqsave(lock, flags);
504 	spin_unlock_irqrestore(lock, flags);
505 
506 	return atomic64_read(&vm->tlb_seq);
507 }
508 
509 /*
510  * vm eviction_lock can be taken in MMU notifiers. Make sure no reclaim-FS
511  * happens while holding this lock anywhere to prevent deadlocks when
512  * an MMU notifier runs in reclaim-FS context.
513  */
514 static inline void amdgpu_vm_eviction_lock(struct amdgpu_vm *vm)
515 {
516 	mutex_lock(&vm->eviction_lock);
517 	vm->saved_flags = memalloc_noreclaim_save();
518 }
519 
520 static inline bool amdgpu_vm_eviction_trylock(struct amdgpu_vm *vm)
521 {
522 	if (mutex_trylock(&vm->eviction_lock)) {
523 		vm->saved_flags = memalloc_noreclaim_save();
524 		return true;
525 	}
526 	return false;
527 }
528 
529 static inline void amdgpu_vm_eviction_unlock(struct amdgpu_vm *vm)
530 {
531 	memalloc_noreclaim_restore(vm->saved_flags);
532 	mutex_unlock(&vm->eviction_lock);
533 }
534 
535 #endif
536