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 /* max vmids dedicated for process */
123 #define AMDGPU_VM_MAX_RESERVED_VMID	1
124 
125 /* See vm_update_mode */
126 #define AMDGPU_VM_USE_CPU_FOR_GFX (1 << 0)
127 #define AMDGPU_VM_USE_CPU_FOR_COMPUTE (1 << 1)
128 
129 /* VMPT level enumerate, and the hiberachy is:
130  * PDB2->PDB1->PDB0->PTB
131  */
132 enum amdgpu_vm_level {
133 	AMDGPU_VM_PDB2,
134 	AMDGPU_VM_PDB1,
135 	AMDGPU_VM_PDB0,
136 	AMDGPU_VM_PTB
137 };
138 
139 /* base structure for tracking BO usage in a VM */
140 struct amdgpu_vm_bo_base {
141 	/* constant after initialization */
142 	struct amdgpu_vm		*vm;
143 	struct amdgpu_bo		*bo;
144 
145 	/* protected by bo being reserved */
146 	struct amdgpu_vm_bo_base	*next;
147 
148 	/* protected by spinlock */
149 	struct list_head		vm_status;
150 
151 	/* protected by the BO being reserved */
152 	bool				moved;
153 };
154 
155 /* provided by hw blocks that can write ptes, e.g., sdma */
156 struct amdgpu_vm_pte_funcs {
157 	/* number of dw to reserve per operation */
158 	unsigned	copy_pte_num_dw;
159 
160 	/* copy pte entries from GART */
161 	void (*copy_pte)(struct amdgpu_ib *ib,
162 			 uint64_t pe, uint64_t src,
163 			 unsigned count);
164 
165 	/* write pte one entry at a time with addr mapping */
166 	void (*write_pte)(struct amdgpu_ib *ib, uint64_t pe,
167 			  uint64_t value, unsigned count,
168 			  uint32_t incr);
169 	/* for linear pte/pde updates without addr mapping */
170 	void (*set_pte_pde)(struct amdgpu_ib *ib,
171 			    uint64_t pe,
172 			    uint64_t addr, unsigned count,
173 			    uint32_t incr, uint64_t flags);
174 };
175 
176 struct amdgpu_task_info {
177 	char	process_name[TASK_COMM_LEN];
178 	char	task_name[TASK_COMM_LEN];
179 	pid_t	pid;
180 	pid_t	tgid;
181 };
182 
183 /**
184  * struct amdgpu_vm_update_params
185  *
186  * Encapsulate some VM table update parameters to reduce
187  * the number of function parameters
188  *
189  */
190 struct amdgpu_vm_update_params {
191 
192 	/**
193 	 * @adev: amdgpu device we do this update for
194 	 */
195 	struct amdgpu_device *adev;
196 
197 	/**
198 	 * @vm: optional amdgpu_vm we do this update for
199 	 */
200 	struct amdgpu_vm *vm;
201 
202 	/**
203 	 * @immediate: if changes should be made immediately
204 	 */
205 	bool immediate;
206 
207 	/**
208 	 * @unlocked: true if the root BO is not locked
209 	 */
210 	bool unlocked;
211 
212 	/**
213 	 * @pages_addr:
214 	 *
215 	 * DMA addresses to use for mapping
216 	 */
217 	dma_addr_t *pages_addr;
218 
219 	/**
220 	 * @job: job to used for hw submission
221 	 */
222 	struct amdgpu_job *job;
223 
224 	/**
225 	 * @num_dw_left: number of dw left for the IB
226 	 */
227 	unsigned int num_dw_left;
228 
229 	/**
230 	 * @table_freed: return true if page table is freed when updating
231 	 */
232 	bool table_freed;
233 };
234 
235 struct amdgpu_vm_update_funcs {
236 	int (*map_table)(struct amdgpu_bo_vm *bo);
237 	int (*prepare)(struct amdgpu_vm_update_params *p, struct dma_resv *resv,
238 		       enum amdgpu_sync_mode sync_mode);
239 	int (*update)(struct amdgpu_vm_update_params *p,
240 		      struct amdgpu_bo_vm *bo, uint64_t pe, uint64_t addr,
241 		      unsigned count, uint32_t incr, uint64_t flags);
242 	int (*commit)(struct amdgpu_vm_update_params *p,
243 		      struct dma_fence **fence);
244 };
245 
246 struct amdgpu_vm {
247 	/* tree of virtual addresses mapped */
248 	struct rb_root_cached	va;
249 
250 	/* Lock to prevent eviction while we are updating page tables
251 	 * use vm_eviction_lock/unlock(vm)
252 	 */
253 	struct mutex		eviction_lock;
254 	bool			evicting;
255 	unsigned int		saved_flags;
256 
257 	/* Lock to protect vm_bo add/del/move on all lists of vm */
258 	spinlock_t		status_lock;
259 
260 	/* BOs who needs a validation */
261 	struct list_head	evicted;
262 
263 	/* PT BOs which relocated and their parent need an update */
264 	struct list_head	relocated;
265 
266 	/* per VM BOs moved, but not yet updated in the PT */
267 	struct list_head	moved;
268 
269 	/* All BOs of this VM not currently in the state machine */
270 	struct list_head	idle;
271 
272 	/* regular invalidated BOs, but not yet updated in the PT */
273 	struct list_head	invalidated;
274 
275 	/* BO mappings freed, but not yet updated in the PT */
276 	struct list_head	freed;
277 
278 	/* BOs which are invalidated, has been updated in the PTs */
279 	struct list_head        done;
280 
281 	/* PT BOs scheduled to free and fill with zero if vm_resv is not hold */
282 	struct list_head	pt_freed;
283 	struct work_struct	pt_free_work;
284 
285 	/* contains the page directory */
286 	struct amdgpu_vm_bo_base     root;
287 	struct dma_fence	*last_update;
288 
289 	/* Scheduler entities for page table updates */
290 	struct drm_sched_entity	immediate;
291 	struct drm_sched_entity	delayed;
292 
293 	/* Last finished delayed update */
294 	atomic64_t		tlb_seq;
295 	struct dma_fence	*last_tlb_flush;
296 
297 	/* Last unlocked submission to the scheduler entities */
298 	struct dma_fence	*last_unlocked;
299 
300 	unsigned int		pasid;
301 	/* dedicated to vm */
302 	struct amdgpu_vmid	*reserved_vmid[AMDGPU_MAX_VMHUBS];
303 
304 	/* Flag to indicate if VM tables are updated by CPU or GPU (SDMA) */
305 	bool					use_cpu_for_update;
306 
307 	/* Functions to use for VM table updates */
308 	const struct amdgpu_vm_update_funcs	*update_funcs;
309 
310 	/* Flag to indicate ATS support from PTE for GFX9 */
311 	bool			pte_support_ats;
312 
313 	/* Up to 128 pending retry page faults */
314 	DECLARE_KFIFO(faults, u64, 128);
315 
316 	/* Points to the KFD process VM info */
317 	struct amdkfd_process_info *process_info;
318 
319 	/* List node in amdkfd_process_info.vm_list_head */
320 	struct list_head	vm_list_node;
321 
322 	/* Valid while the PD is reserved or fenced */
323 	uint64_t		pd_phys_addr;
324 
325 	/* Some basic info about the task */
326 	struct amdgpu_task_info task_info;
327 
328 	/* Store positions of group of BOs */
329 	struct ttm_lru_bulk_move lru_bulk_move;
330 	/* Flag to indicate if VM is used for compute */
331 	bool			is_compute_context;
332 };
333 
334 struct amdgpu_vm_manager {
335 	/* Handling of VMIDs */
336 	struct amdgpu_vmid_mgr			id_mgr[AMDGPU_MAX_VMHUBS];
337 	unsigned int				first_kfd_vmid;
338 	bool					concurrent_flush;
339 
340 	/* Handling of VM fences */
341 	u64					fence_context;
342 	unsigned				seqno[AMDGPU_MAX_RINGS];
343 
344 	uint64_t				max_pfn;
345 	uint32_t				num_level;
346 	uint32_t				block_size;
347 	uint32_t				fragment_size;
348 	enum amdgpu_vm_level			root_level;
349 	/* vram base address for page table entry  */
350 	u64					vram_base_offset;
351 	/* vm pte handling */
352 	const struct amdgpu_vm_pte_funcs	*vm_pte_funcs;
353 	struct drm_gpu_scheduler		*vm_pte_scheds[AMDGPU_MAX_RINGS];
354 	unsigned				vm_pte_num_scheds;
355 	struct amdgpu_ring			*page_fault;
356 
357 	/* partial resident texture handling */
358 	spinlock_t				prt_lock;
359 	atomic_t				num_prt_users;
360 
361 	/* controls how VM page tables are updated for Graphics and Compute.
362 	 * BIT0[= 0] Graphics updated by SDMA [= 1] by CPU
363 	 * BIT1[= 0] Compute updated by SDMA [= 1] by CPU
364 	 */
365 	int					vm_update_mode;
366 
367 	/* PASID to VM mapping, will be used in interrupt context to
368 	 * look up VM of a page fault
369 	 */
370 	struct xarray				pasids;
371 };
372 
373 struct amdgpu_bo_va_mapping;
374 
375 #define amdgpu_vm_copy_pte(adev, ib, pe, src, count) ((adev)->vm_manager.vm_pte_funcs->copy_pte((ib), (pe), (src), (count)))
376 #define amdgpu_vm_write_pte(adev, ib, pe, value, count, incr) ((adev)->vm_manager.vm_pte_funcs->write_pte((ib), (pe), (value), (count), (incr)))
377 #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)))
378 
379 extern const struct amdgpu_vm_update_funcs amdgpu_vm_cpu_funcs;
380 extern const struct amdgpu_vm_update_funcs amdgpu_vm_sdma_funcs;
381 
382 void amdgpu_vm_manager_init(struct amdgpu_device *adev);
383 void amdgpu_vm_manager_fini(struct amdgpu_device *adev);
384 
385 int amdgpu_vm_set_pasid(struct amdgpu_device *adev, struct amdgpu_vm *vm,
386 			u32 pasid);
387 
388 long amdgpu_vm_wait_idle(struct amdgpu_vm *vm, long timeout);
389 int amdgpu_vm_init(struct amdgpu_device *adev, struct amdgpu_vm *vm);
390 int amdgpu_vm_make_compute(struct amdgpu_device *adev, struct amdgpu_vm *vm);
391 void amdgpu_vm_release_compute(struct amdgpu_device *adev, struct amdgpu_vm *vm);
392 void amdgpu_vm_fini(struct amdgpu_device *adev, struct amdgpu_vm *vm);
393 void amdgpu_vm_get_pd_bo(struct amdgpu_vm *vm,
394 			 struct list_head *validated,
395 			 struct amdgpu_bo_list_entry *entry);
396 bool amdgpu_vm_ready(struct amdgpu_vm *vm);
397 int amdgpu_vm_validate_pt_bos(struct amdgpu_device *adev, struct amdgpu_vm *vm,
398 			      int (*callback)(void *p, struct amdgpu_bo *bo),
399 			      void *param);
400 int amdgpu_vm_flush(struct amdgpu_ring *ring, struct amdgpu_job *job, bool need_pipe_sync);
401 int amdgpu_vm_update_pdes(struct amdgpu_device *adev,
402 			  struct amdgpu_vm *vm, bool immediate);
403 int amdgpu_vm_clear_freed(struct amdgpu_device *adev,
404 			  struct amdgpu_vm *vm,
405 			  struct dma_fence **fence);
406 int amdgpu_vm_handle_moved(struct amdgpu_device *adev,
407 			   struct amdgpu_vm *vm);
408 void amdgpu_vm_bo_base_init(struct amdgpu_vm_bo_base *base,
409 			    struct amdgpu_vm *vm, struct amdgpu_bo *bo);
410 int amdgpu_vm_update_range(struct amdgpu_device *adev, struct amdgpu_vm *vm,
411 			   bool immediate, bool unlocked, bool flush_tlb,
412 			   struct dma_resv *resv, uint64_t start, uint64_t last,
413 			   uint64_t flags, uint64_t offset, uint64_t vram_base,
414 			   struct ttm_resource *res, dma_addr_t *pages_addr,
415 			   struct dma_fence **fence);
416 int amdgpu_vm_bo_update(struct amdgpu_device *adev,
417 			struct amdgpu_bo_va *bo_va,
418 			bool clear);
419 bool amdgpu_vm_evictable(struct amdgpu_bo *bo);
420 void amdgpu_vm_bo_invalidate(struct amdgpu_device *adev,
421 			     struct amdgpu_bo *bo, bool evicted);
422 uint64_t amdgpu_vm_map_gart(const dma_addr_t *pages_addr, uint64_t addr);
423 struct amdgpu_bo_va *amdgpu_vm_bo_find(struct amdgpu_vm *vm,
424 				       struct amdgpu_bo *bo);
425 struct amdgpu_bo_va *amdgpu_vm_bo_add(struct amdgpu_device *adev,
426 				      struct amdgpu_vm *vm,
427 				      struct amdgpu_bo *bo);
428 int amdgpu_vm_bo_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_replace_map(struct amdgpu_device *adev,
433 			     struct amdgpu_bo_va *bo_va,
434 			     uint64_t addr, uint64_t offset,
435 			     uint64_t size, uint64_t flags);
436 int amdgpu_vm_bo_unmap(struct amdgpu_device *adev,
437 		       struct amdgpu_bo_va *bo_va,
438 		       uint64_t addr);
439 int amdgpu_vm_bo_clear_mappings(struct amdgpu_device *adev,
440 				struct amdgpu_vm *vm,
441 				uint64_t saddr, uint64_t size);
442 struct amdgpu_bo_va_mapping *amdgpu_vm_bo_lookup_mapping(struct amdgpu_vm *vm,
443 							 uint64_t addr);
444 void amdgpu_vm_bo_trace_cs(struct amdgpu_vm *vm, struct ww_acquire_ctx *ticket);
445 void amdgpu_vm_bo_del(struct amdgpu_device *adev,
446 		      struct amdgpu_bo_va *bo_va);
447 void amdgpu_vm_adjust_size(struct amdgpu_device *adev, uint32_t min_vm_size,
448 			   uint32_t fragment_size_default, unsigned max_level,
449 			   unsigned max_bits);
450 int amdgpu_vm_ioctl(struct drm_device *dev, void *data, struct drm_file *filp);
451 bool amdgpu_vm_need_pipeline_sync(struct amdgpu_ring *ring,
452 				  struct amdgpu_job *job);
453 void amdgpu_vm_check_compute_bug(struct amdgpu_device *adev);
454 
455 void amdgpu_vm_get_task_info(struct amdgpu_device *adev, u32 pasid,
456 			     struct amdgpu_task_info *task_info);
457 bool amdgpu_vm_handle_fault(struct amdgpu_device *adev, u32 pasid,
458 			    uint64_t addr, bool write_fault);
459 
460 void amdgpu_vm_set_task_info(struct amdgpu_vm *vm);
461 
462 void amdgpu_vm_move_to_lru_tail(struct amdgpu_device *adev,
463 				struct amdgpu_vm *vm);
464 void amdgpu_vm_get_memory(struct amdgpu_vm *vm, uint64_t *vram_mem,
465 				uint64_t *gtt_mem, uint64_t *cpu_mem);
466 
467 int amdgpu_vm_pt_clear(struct amdgpu_device *adev, struct amdgpu_vm *vm,
468 		       struct amdgpu_bo_vm *vmbo, bool immediate);
469 int amdgpu_vm_pt_create(struct amdgpu_device *adev, struct amdgpu_vm *vm,
470 			int level, bool immediate, struct amdgpu_bo_vm **vmbo);
471 void amdgpu_vm_pt_free_root(struct amdgpu_device *adev, struct amdgpu_vm *vm);
472 bool amdgpu_vm_pt_is_root_clean(struct amdgpu_device *adev,
473 				struct amdgpu_vm *vm);
474 
475 int amdgpu_vm_pde_update(struct amdgpu_vm_update_params *params,
476 			 struct amdgpu_vm_bo_base *entry);
477 int amdgpu_vm_ptes_update(struct amdgpu_vm_update_params *params,
478 			  uint64_t start, uint64_t end,
479 			  uint64_t dst, uint64_t flags);
480 void amdgpu_vm_pt_free_work(struct work_struct *work);
481 
482 #if defined(CONFIG_DEBUG_FS)
483 void amdgpu_debugfs_vm_bo_info(struct amdgpu_vm *vm, struct seq_file *m);
484 #endif
485 
486 /**
487  * amdgpu_vm_tlb_seq - return tlb flush sequence number
488  * @vm: the amdgpu_vm structure to query
489  *
490  * Returns the tlb flush sequence number which indicates that the VM TLBs needs
491  * to be invalidated whenever the sequence number change.
492  */
493 static inline uint64_t amdgpu_vm_tlb_seq(struct amdgpu_vm *vm)
494 {
495 	unsigned long flags;
496 	spinlock_t *lock;
497 
498 	/*
499 	 * Workaround to stop racing between the fence signaling and handling
500 	 * the cb. The lock is static after initially setting it up, just make
501 	 * sure that the dma_fence structure isn't freed up.
502 	 */
503 	rcu_read_lock();
504 	lock = vm->last_tlb_flush->lock;
505 	rcu_read_unlock();
506 
507 	spin_lock_irqsave(lock, flags);
508 	spin_unlock_irqrestore(lock, flags);
509 
510 	return atomic64_read(&vm->tlb_seq);
511 }
512 
513 /*
514  * vm eviction_lock can be taken in MMU notifiers. Make sure no reclaim-FS
515  * happens while holding this lock anywhere to prevent deadlocks when
516  * an MMU notifier runs in reclaim-FS context.
517  */
518 static inline void amdgpu_vm_eviction_lock(struct amdgpu_vm *vm)
519 {
520 	mutex_lock(&vm->eviction_lock);
521 	vm->saved_flags = memalloc_noreclaim_save();
522 }
523 
524 static inline bool amdgpu_vm_eviction_trylock(struct amdgpu_vm *vm)
525 {
526 	if (mutex_trylock(&vm->eviction_lock)) {
527 		vm->saved_flags = memalloc_noreclaim_save();
528 		return true;
529 	}
530 	return false;
531 }
532 
533 static inline void amdgpu_vm_eviction_unlock(struct amdgpu_vm *vm)
534 {
535 	memalloc_noreclaim_restore(vm->saved_flags);
536 	mutex_unlock(&vm->eviction_lock);
537 }
538 
539 #endif
540