Linux Audio

Check our new training course

Loading...
v6.13.7
  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
 39struct drm_exec;
 40
 41struct amdgpu_bo_va;
 42struct amdgpu_job;
 43struct amdgpu_bo_list_entry;
 44struct amdgpu_bo_vm;
 45
 46/*
 47 * GPUVM handling
 48 */
 49
 
 
 
 50/* Maximum number of PTEs the hardware can write with one command */
 51#define AMDGPU_VM_MAX_UPDATE_SIZE	0x3FFFF
 52
 53/* number of entries in page table */
 54#define AMDGPU_VM_PTE_COUNT(adev) (1 << (adev)->vm_manager.block_size)
 55
 56#define AMDGPU_PTE_VALID	(1ULL << 0)
 57#define AMDGPU_PTE_SYSTEM	(1ULL << 1)
 58#define AMDGPU_PTE_SNOOPED	(1ULL << 2)
 59
 60/* RV+ */
 61#define AMDGPU_PTE_TMZ		(1ULL << 3)
 62
 63/* VI only */
 64#define AMDGPU_PTE_EXECUTABLE	(1ULL << 4)
 
 65
 66#define AMDGPU_PTE_READABLE	(1ULL << 5)
 67#define AMDGPU_PTE_WRITEABLE	(1ULL << 6)
 68
 69#define AMDGPU_PTE_FRAG(x)	((x & 0x1fULL) << 7)
 
 70
 71/* TILED for VEGA10, reserved for older ASICs  */
 72#define AMDGPU_PTE_PRT		(1ULL << 51)
 73
 74/* PDE is handled as PTE for VEGA10 */
 75#define AMDGPU_PDE_PTE		(1ULL << 54)
 76
 77#define AMDGPU_PTE_LOG          (1ULL << 55)
 78
 79/* PTE is handled as PDE for VEGA10 (Translate Further) */
 80#define AMDGPU_PTE_TF		(1ULL << 56)
 81
 82/* MALL noalloc for sienna_cichlid, reserved for older ASICs  */
 83#define AMDGPU_PTE_NOALLOC	(1ULL << 58)
 84
 85/* PDE Block Fragment Size for VEGA10 */
 86#define AMDGPU_PDE_BFS(a)	((uint64_t)a << 59)
 87
 88/* Flag combination to set no-retry with TF disabled */
 89#define AMDGPU_VM_NORETRY_FLAGS	(AMDGPU_PTE_EXECUTABLE | AMDGPU_PDE_PTE | \
 90				AMDGPU_PTE_TF)
 91
 92/* Flag combination to set no-retry with TF enabled */
 93#define AMDGPU_VM_NORETRY_FLAGS_TF (AMDGPU_PTE_VALID | AMDGPU_PTE_SYSTEM | \
 94				   AMDGPU_PTE_PRT)
 95/* For GFX9 */
 96#define AMDGPU_PTE_MTYPE_VG10_SHIFT(mtype)	((uint64_t)(mtype) << 57)
 97#define AMDGPU_PTE_MTYPE_VG10_MASK	AMDGPU_PTE_MTYPE_VG10_SHIFT(3ULL)
 98#define AMDGPU_PTE_MTYPE_VG10(flags, mtype)			\
 99	(((uint64_t)(flags) & (~AMDGPU_PTE_MTYPE_VG10_MASK)) |	\
100	  AMDGPU_PTE_MTYPE_VG10_SHIFT(mtype))
101
102#define AMDGPU_MTYPE_NC 0
103#define AMDGPU_MTYPE_CC 2
104
105#define AMDGPU_PTE_DEFAULT_ATC  (AMDGPU_PTE_SYSTEM      \
106                                | AMDGPU_PTE_SNOOPED    \
107                                | AMDGPU_PTE_EXECUTABLE \
108                                | AMDGPU_PTE_READABLE   \
109                                | AMDGPU_PTE_WRITEABLE  \
110                                | AMDGPU_PTE_MTYPE_VG10(AMDGPU_MTYPE_CC))
111
112/* gfx10 */
113#define AMDGPU_PTE_MTYPE_NV10_SHIFT(mtype)	((uint64_t)(mtype) << 48)
114#define AMDGPU_PTE_MTYPE_NV10_MASK     AMDGPU_PTE_MTYPE_NV10_SHIFT(7ULL)
115#define AMDGPU_PTE_MTYPE_NV10(flags, mtype)			\
116	(((uint64_t)(flags) & (~AMDGPU_PTE_MTYPE_NV10_MASK)) |	\
117	  AMDGPU_PTE_MTYPE_NV10_SHIFT(mtype))
118
119/* gfx12 */
120#define AMDGPU_PTE_PRT_GFX12		(1ULL << 56)
121#define AMDGPU_PTE_PRT_FLAG(adev)	\
122	((amdgpu_ip_version((adev), GC_HWIP, 0) >= IP_VERSION(12, 0, 0)) ? AMDGPU_PTE_PRT_GFX12 : AMDGPU_PTE_PRT)
123
124#define AMDGPU_PTE_MTYPE_GFX12_SHIFT(mtype)	((uint64_t)(mtype) << 54)
125#define AMDGPU_PTE_MTYPE_GFX12_MASK	AMDGPU_PTE_MTYPE_GFX12_SHIFT(3ULL)
126#define AMDGPU_PTE_MTYPE_GFX12(flags, mtype)				\
127	(((uint64_t)(flags) & (~AMDGPU_PTE_MTYPE_GFX12_MASK)) |	\
128	  AMDGPU_PTE_MTYPE_GFX12_SHIFT(mtype))
129
130#define AMDGPU_PTE_DCC			(1ULL << 58)
131#define AMDGPU_PTE_IS_PTE		(1ULL << 63)
132
133/* PDE Block Fragment Size for gfx v12 */
134#define AMDGPU_PDE_BFS_GFX12(a)		((uint64_t)((a) & 0x1fULL) << 58)
135#define AMDGPU_PDE_BFS_FLAG(adev, a)	\
136	((amdgpu_ip_version((adev), GC_HWIP, 0) >= IP_VERSION(12, 0, 0)) ? AMDGPU_PDE_BFS_GFX12(a) : AMDGPU_PDE_BFS(a))
137/* PDE is handled as PTE for gfx v12 */
138#define AMDGPU_PDE_PTE_GFX12		(1ULL << 63)
139#define AMDGPU_PDE_PTE_FLAG(adev)	\
140	((amdgpu_ip_version((adev), GC_HWIP, 0) >= IP_VERSION(12, 0, 0)) ? AMDGPU_PDE_PTE_GFX12 : AMDGPU_PDE_PTE)
141
142/* How to program VM fault handling */
143#define AMDGPU_VM_FAULT_STOP_NEVER	0
144#define AMDGPU_VM_FAULT_STOP_FIRST	1
145#define AMDGPU_VM_FAULT_STOP_ALWAYS	2
146
147/* How much VRAM be reserved for page tables */
148#define AMDGPU_VM_RESERVED_VRAM		(8ULL << 20)
149
150/*
151 * max number of VMHUB
152 * layout: max 8 GFXHUB + 4 MMHUB0 + 1 MMHUB1
153 */
154#define AMDGPU_MAX_VMHUBS			13
155#define AMDGPU_GFXHUB_START			0
156#define AMDGPU_MMHUB0_START			8
157#define AMDGPU_MMHUB1_START			12
158#define AMDGPU_GFXHUB(x)			(AMDGPU_GFXHUB_START + (x))
159#define AMDGPU_MMHUB0(x)			(AMDGPU_MMHUB0_START + (x))
160#define AMDGPU_MMHUB1(x)			(AMDGPU_MMHUB1_START + (x))
161
162#define AMDGPU_IS_GFXHUB(x) ((x) >= AMDGPU_GFXHUB_START && (x) < AMDGPU_MMHUB0_START)
163#define AMDGPU_IS_MMHUB0(x) ((x) >= AMDGPU_MMHUB0_START && (x) < AMDGPU_MMHUB1_START)
164#define AMDGPU_IS_MMHUB1(x) ((x) >= AMDGPU_MMHUB1_START && (x) < AMDGPU_MAX_VMHUBS)
165
166/* Reserve space at top/bottom of address space for kernel use */
167#define AMDGPU_VA_RESERVED_CSA_SIZE		(2ULL << 20)
168#define AMDGPU_VA_RESERVED_CSA_START(adev)	(((adev)->vm_manager.max_pfn \
169						  << AMDGPU_GPU_PAGE_SHIFT)  \
170						 - AMDGPU_VA_RESERVED_CSA_SIZE)
171#define AMDGPU_VA_RESERVED_SEQ64_SIZE		(2ULL << 20)
172#define AMDGPU_VA_RESERVED_SEQ64_START(adev)	(AMDGPU_VA_RESERVED_CSA_START(adev) \
173						 - AMDGPU_VA_RESERVED_SEQ64_SIZE)
174#define AMDGPU_VA_RESERVED_TRAP_SIZE		(2ULL << 12)
175#define AMDGPU_VA_RESERVED_TRAP_START(adev)	(AMDGPU_VA_RESERVED_SEQ64_START(adev) \
176						 - AMDGPU_VA_RESERVED_TRAP_SIZE)
177#define AMDGPU_VA_RESERVED_BOTTOM		(1ULL << 16)
178#define AMDGPU_VA_RESERVED_TOP			(AMDGPU_VA_RESERVED_TRAP_SIZE + \
179						 AMDGPU_VA_RESERVED_SEQ64_SIZE + \
180						 AMDGPU_VA_RESERVED_CSA_SIZE)
181
182/* See vm_update_mode */
183#define AMDGPU_VM_USE_CPU_FOR_GFX (1 << 0)
184#define AMDGPU_VM_USE_CPU_FOR_COMPUTE (1 << 1)
185
186/* VMPT level enumerate, and the hiberachy is:
187 * PDB2->PDB1->PDB0->PTB
188 */
189enum amdgpu_vm_level {
190	AMDGPU_VM_PDB2,
191	AMDGPU_VM_PDB1,
192	AMDGPU_VM_PDB0,
193	AMDGPU_VM_PTB
194};
195
196/* base structure for tracking BO usage in a VM */
197struct amdgpu_vm_bo_base {
198	/* constant after initialization */
199	struct amdgpu_vm		*vm;
200	struct amdgpu_bo		*bo;
201
202	/* protected by bo being reserved */
203	struct amdgpu_vm_bo_base	*next;
204
205	/* protected by spinlock */
206	struct list_head		vm_status;
207
208	/* protected by the BO being reserved */
209	bool				moved;
210};
211
212/* provided by hw blocks that can write ptes, e.g., sdma */
213struct amdgpu_vm_pte_funcs {
214	/* number of dw to reserve per operation */
215	unsigned	copy_pte_num_dw;
216
217	/* copy pte entries from GART */
218	void (*copy_pte)(struct amdgpu_ib *ib,
219			 uint64_t pe, uint64_t src,
220			 unsigned count);
221
222	/* write pte one entry at a time with addr mapping */
223	void (*write_pte)(struct amdgpu_ib *ib, uint64_t pe,
224			  uint64_t value, unsigned count,
225			  uint32_t incr);
226	/* for linear pte/pde updates without addr mapping */
227	void (*set_pte_pde)(struct amdgpu_ib *ib,
228			    uint64_t pe,
229			    uint64_t addr, unsigned count,
230			    uint32_t incr, uint64_t flags);
231};
232
233struct amdgpu_task_info {
234	char		process_name[TASK_COMM_LEN];
235	char		task_name[TASK_COMM_LEN];
236	pid_t		pid;
237	pid_t		tgid;
238	struct kref	refcount;
239};
240
241/**
242 * struct amdgpu_vm_update_params
243 *
244 * Encapsulate some VM table update parameters to reduce
245 * the number of function parameters
246 *
247 */
248struct amdgpu_vm_update_params {
249
250	/**
251	 * @adev: amdgpu device we do this update for
252	 */
253	struct amdgpu_device *adev;
254
255	/**
256	 * @vm: optional amdgpu_vm we do this update for
257	 */
258	struct amdgpu_vm *vm;
259
260	/**
261	 * @immediate: if changes should be made immediately
262	 */
263	bool immediate;
264
265	/**
266	 * @unlocked: true if the root BO is not locked
267	 */
268	bool unlocked;
269
270	/**
271	 * @pages_addr:
272	 *
273	 * DMA addresses to use for mapping
274	 */
275	dma_addr_t *pages_addr;
276
277	/**
278	 * @job: job to used for hw submission
279	 */
280	struct amdgpu_job *job;
281
282	/**
283	 * @num_dw_left: number of dw left for the IB
284	 */
285	unsigned int num_dw_left;
286
287	/**
288	 * @needs_flush: true whenever we need to invalidate the TLB
289	 */
290	bool needs_flush;
291
292	/**
293	 * @allow_override: true for memory that is not uncached: allows MTYPE
294	 * to be overridden for NUMA local memory.
295	 */
296	bool allow_override;
297
298	/**
299	 * @tlb_flush_waitlist: temporary storage for BOs until tlb_flush
300	 */
301	struct list_head tlb_flush_waitlist;
302};
303
304struct amdgpu_vm_update_funcs {
305	int (*map_table)(struct amdgpu_bo_vm *bo);
306	int (*prepare)(struct amdgpu_vm_update_params *p,
307		       struct amdgpu_sync *sync);
308	int (*update)(struct amdgpu_vm_update_params *p,
309		      struct amdgpu_bo_vm *bo, uint64_t pe, uint64_t addr,
310		      unsigned count, uint32_t incr, uint64_t flags);
311	int (*commit)(struct amdgpu_vm_update_params *p,
312		      struct dma_fence **fence);
313};
314
315struct amdgpu_vm_fault_info {
316	/* fault address */
317	uint64_t	addr;
318	/* fault status register */
319	uint32_t	status;
320	/* which vmhub? gfxhub, mmhub, etc. */
321	unsigned int	vmhub;
322};
323
324struct amdgpu_mem_stats {
325	struct drm_memory_stats drm;
326
327	/* buffers that requested this placement */
328	uint64_t requested;
329	/* buffers that requested this placement
330	 * but are currently evicted */
331	uint64_t evicted;
332};
333
334struct amdgpu_vm {
335	/* tree of virtual addresses mapped */
336	struct rb_root_cached	va;
337
338	/* Lock to prevent eviction while we are updating page tables
339	 * use vm_eviction_lock/unlock(vm)
340	 */
341	struct mutex		eviction_lock;
342	bool			evicting;
343	unsigned int		saved_flags;
344
345	/* Lock to protect vm_bo add/del/move on all lists of vm */
346	spinlock_t		status_lock;
347
348	/* Per-VM and PT BOs who needs a validation */
349	struct list_head	evicted;
350
351	/* BOs for user mode queues that need a validation */
352	struct list_head	evicted_user;
353
354	/* PT BOs which relocated and their parent need an update */
355	struct list_head	relocated;
356
357	/* per VM BOs moved, but not yet updated in the PT */
358	struct list_head	moved;
359
360	/* All BOs of this VM not currently in the state machine */
361	struct list_head	idle;
362
363	/* regular invalidated BOs, but not yet updated in the PT */
364	struct list_head	invalidated;
365
 
 
 
366	/* BO mappings freed, but not yet updated in the PT */
367	struct list_head	freed;
368
369	/* BOs which are invalidated, has been updated in the PTs */
370	struct list_head        done;
371
372	/* PT BOs scheduled to free and fill with zero if vm_resv is not hold */
373	struct list_head	pt_freed;
374	struct work_struct	pt_free_work;
375
376	/* contains the page directory */
377	struct amdgpu_vm_bo_base     root;
378	struct dma_fence	*last_update;
379
380	/* Scheduler entities for page table updates */
381	struct drm_sched_entity	immediate;
382	struct drm_sched_entity	delayed;
383
384	/* Last finished delayed update */
385	atomic64_t		tlb_seq;
386	struct dma_fence	*last_tlb_flush;
387	atomic64_t		kfd_last_flushed_seq;
388	uint64_t		tlb_fence_context;
389
390	/* How many times we had to re-generate the page tables */
391	uint64_t		generation;
392
393	/* Last unlocked submission to the scheduler entities */
394	struct dma_fence	*last_unlocked;
395
396	unsigned int		pasid;
397	bool			reserved_vmid[AMDGPU_MAX_VMHUBS];
398
399	/* Flag to indicate if VM tables are updated by CPU or GPU (SDMA) */
400	bool					use_cpu_for_update;
401
402	/* Functions to use for VM table updates */
403	const struct amdgpu_vm_update_funcs	*update_funcs;
404
405	/* Up to 128 pending retry page faults */
406	DECLARE_KFIFO(faults, u64, 128);
407
408	/* Points to the KFD process VM info */
409	struct amdkfd_process_info *process_info;
410
411	/* List node in amdkfd_process_info.vm_list_head */
412	struct list_head	vm_list_node;
413
414	/* Valid while the PD is reserved or fenced */
415	uint64_t		pd_phys_addr;
416
417	/* Some basic info about the task */
418	struct amdgpu_task_info *task_info;
419
420	/* Store positions of group of BOs */
421	struct ttm_lru_bulk_move lru_bulk_move;
422	/* Flag to indicate if VM is used for compute */
423	bool			is_compute_context;
424
425	/* Memory partition number, -1 means any partition */
426	int8_t			mem_id;
427
428	/* cached fault info */
429	struct amdgpu_vm_fault_info fault_info;
430};
431
432struct amdgpu_vm_manager {
433	/* Handling of VMIDs */
434	struct amdgpu_vmid_mgr			id_mgr[AMDGPU_MAX_VMHUBS];
435	unsigned int				first_kfd_vmid;
436	bool					concurrent_flush;
 
437
438	/* Handling of VM fences */
439	u64					fence_context;
440	unsigned				seqno[AMDGPU_MAX_RINGS];
441
442	uint64_t				max_pfn;
443	uint32_t				num_level;
444	uint32_t				block_size;
445	uint32_t				fragment_size;
446	enum amdgpu_vm_level			root_level;
447	/* vram base address for page table entry  */
448	u64					vram_base_offset;
 
 
449	/* vm pte handling */
450	const struct amdgpu_vm_pte_funcs	*vm_pte_funcs;
451	struct drm_gpu_scheduler		*vm_pte_scheds[AMDGPU_MAX_RINGS];
452	unsigned				vm_pte_num_scheds;
453	struct amdgpu_ring			*page_fault;
454
455	/* partial resident texture handling */
456	spinlock_t				prt_lock;
457	atomic_t				num_prt_users;
458
459	/* controls how VM page tables are updated for Graphics and Compute.
460	 * BIT0[= 0] Graphics updated by SDMA [= 1] by CPU
461	 * BIT1[= 0] Compute updated by SDMA [= 1] by CPU
462	 */
463	int					vm_update_mode;
464
465	/* PASID to VM mapping, will be used in interrupt context to
466	 * look up VM of a page fault
467	 */
468	struct xarray				pasids;
469	/* Global registration of recent page fault information */
470	struct amdgpu_vm_fault_info	fault_info;
471};
472
473struct amdgpu_bo_va_mapping;
474
475#define amdgpu_vm_copy_pte(adev, ib, pe, src, count) ((adev)->vm_manager.vm_pte_funcs->copy_pte((ib), (pe), (src), (count)))
476#define amdgpu_vm_write_pte(adev, ib, pe, value, count, incr) ((adev)->vm_manager.vm_pte_funcs->write_pte((ib), (pe), (value), (count), (incr)))
477#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)))
478
479extern const struct amdgpu_vm_update_funcs amdgpu_vm_cpu_funcs;
480extern const struct amdgpu_vm_update_funcs amdgpu_vm_sdma_funcs;
481
482void amdgpu_vm_manager_init(struct amdgpu_device *adev);
483void amdgpu_vm_manager_fini(struct amdgpu_device *adev);
484
485int amdgpu_vm_set_pasid(struct amdgpu_device *adev, struct amdgpu_vm *vm,
486			u32 pasid);
487
488long amdgpu_vm_wait_idle(struct amdgpu_vm *vm, long timeout);
489int amdgpu_vm_init(struct amdgpu_device *adev, struct amdgpu_vm *vm, int32_t xcp_id);
490int amdgpu_vm_make_compute(struct amdgpu_device *adev, struct amdgpu_vm *vm);
491void amdgpu_vm_release_compute(struct amdgpu_device *adev, struct amdgpu_vm *vm);
492void amdgpu_vm_fini(struct amdgpu_device *adev, struct amdgpu_vm *vm);
493int amdgpu_vm_lock_pd(struct amdgpu_vm *vm, struct drm_exec *exec,
494		      unsigned int num_fences);
495bool amdgpu_vm_ready(struct amdgpu_vm *vm);
496uint64_t amdgpu_vm_generation(struct amdgpu_device *adev, struct amdgpu_vm *vm);
497int amdgpu_vm_validate(struct amdgpu_device *adev, struct amdgpu_vm *vm,
498		       struct ww_acquire_ctx *ticket,
499		       int (*callback)(void *p, struct amdgpu_bo *bo),
500		       void *param);
501int amdgpu_vm_flush(struct amdgpu_ring *ring, struct amdgpu_job *job, bool need_pipe_sync);
502int amdgpu_vm_update_pdes(struct amdgpu_device *adev,
503			  struct amdgpu_vm *vm, bool immediate);
 
 
 
 
504int amdgpu_vm_clear_freed(struct amdgpu_device *adev,
505			  struct amdgpu_vm *vm,
506			  struct dma_fence **fence);
507int amdgpu_vm_handle_moved(struct amdgpu_device *adev,
508			   struct amdgpu_vm *vm,
509			   struct ww_acquire_ctx *ticket);
510int amdgpu_vm_flush_compute_tlb(struct amdgpu_device *adev,
511				struct amdgpu_vm *vm,
512				uint32_t flush_type,
513				uint32_t xcc_mask);
514void amdgpu_vm_bo_base_init(struct amdgpu_vm_bo_base *base,
515			    struct amdgpu_vm *vm, struct amdgpu_bo *bo);
516int amdgpu_vm_update_range(struct amdgpu_device *adev, struct amdgpu_vm *vm,
517			   bool immediate, bool unlocked, bool flush_tlb,
518			   bool allow_override, struct amdgpu_sync *sync,
519			   uint64_t start, uint64_t last, uint64_t flags,
520			   uint64_t offset, uint64_t vram_base,
521			   struct ttm_resource *res, dma_addr_t *pages_addr,
522			   struct dma_fence **fence);
523int amdgpu_vm_bo_update(struct amdgpu_device *adev,
524			struct amdgpu_bo_va *bo_va,
525			bool clear);
526bool amdgpu_vm_evictable(struct amdgpu_bo *bo);
527void amdgpu_vm_bo_invalidate(struct amdgpu_device *adev,
528			     struct amdgpu_bo *bo, bool evicted);
529uint64_t amdgpu_vm_map_gart(const dma_addr_t *pages_addr, uint64_t addr);
530struct amdgpu_bo_va *amdgpu_vm_bo_find(struct amdgpu_vm *vm,
531				       struct amdgpu_bo *bo);
532struct amdgpu_bo_va *amdgpu_vm_bo_add(struct amdgpu_device *adev,
533				      struct amdgpu_vm *vm,
534				      struct amdgpu_bo *bo);
535int amdgpu_vm_bo_map(struct amdgpu_device *adev,
536		     struct amdgpu_bo_va *bo_va,
537		     uint64_t addr, uint64_t offset,
538		     uint64_t size, uint64_t flags);
539int amdgpu_vm_bo_replace_map(struct amdgpu_device *adev,
540			     struct amdgpu_bo_va *bo_va,
541			     uint64_t addr, uint64_t offset,
542			     uint64_t size, uint64_t flags);
543int amdgpu_vm_bo_unmap(struct amdgpu_device *adev,
544		       struct amdgpu_bo_va *bo_va,
545		       uint64_t addr);
546int amdgpu_vm_bo_clear_mappings(struct amdgpu_device *adev,
547				struct amdgpu_vm *vm,
548				uint64_t saddr, uint64_t size);
549struct amdgpu_bo_va_mapping *amdgpu_vm_bo_lookup_mapping(struct amdgpu_vm *vm,
550							 uint64_t addr);
551void amdgpu_vm_bo_trace_cs(struct amdgpu_vm *vm, struct ww_acquire_ctx *ticket);
552void amdgpu_vm_bo_del(struct amdgpu_device *adev,
553		      struct amdgpu_bo_va *bo_va);
554void amdgpu_vm_adjust_size(struct amdgpu_device *adev, uint32_t min_vm_size,
555			   uint32_t fragment_size_default, unsigned max_level,
556			   unsigned max_bits);
557int amdgpu_vm_ioctl(struct drm_device *dev, void *data, struct drm_file *filp);
558bool amdgpu_vm_need_pipeline_sync(struct amdgpu_ring *ring,
559				  struct amdgpu_job *job);
560void amdgpu_vm_check_compute_bug(struct amdgpu_device *adev);
561
562struct amdgpu_task_info *
563amdgpu_vm_get_task_info_pasid(struct amdgpu_device *adev, u32 pasid);
564
565struct amdgpu_task_info *
566amdgpu_vm_get_task_info_vm(struct amdgpu_vm *vm);
567
568void amdgpu_vm_put_task_info(struct amdgpu_task_info *task_info);
569
570bool amdgpu_vm_handle_fault(struct amdgpu_device *adev, u32 pasid,
571			    u32 vmid, u32 node_id, uint64_t addr, uint64_t ts,
572			    bool write_fault);
573
574void amdgpu_vm_set_task_info(struct amdgpu_vm *vm);
575
576void amdgpu_vm_move_to_lru_tail(struct amdgpu_device *adev,
577				struct amdgpu_vm *vm);
578void amdgpu_vm_get_memory(struct amdgpu_vm *vm,
579			  struct amdgpu_mem_stats *stats,
580			  unsigned int size);
581
582int amdgpu_vm_pt_clear(struct amdgpu_device *adev, struct amdgpu_vm *vm,
583		       struct amdgpu_bo_vm *vmbo, bool immediate);
584int amdgpu_vm_pt_create(struct amdgpu_device *adev, struct amdgpu_vm *vm,
585			int level, bool immediate, struct amdgpu_bo_vm **vmbo,
586			int32_t xcp_id);
587void amdgpu_vm_pt_free_root(struct amdgpu_device *adev, struct amdgpu_vm *vm);
588
589int amdgpu_vm_pde_update(struct amdgpu_vm_update_params *params,
590			 struct amdgpu_vm_bo_base *entry);
591int amdgpu_vm_ptes_update(struct amdgpu_vm_update_params *params,
592			  uint64_t start, uint64_t end,
593			  uint64_t dst, uint64_t flags);
594void amdgpu_vm_pt_free_work(struct work_struct *work);
595void amdgpu_vm_pt_free_list(struct amdgpu_device *adev,
596			    struct amdgpu_vm_update_params *params);
597
598#if defined(CONFIG_DEBUG_FS)
599void amdgpu_debugfs_vm_bo_info(struct amdgpu_vm *vm, struct seq_file *m);
600#endif
601
602int amdgpu_vm_pt_map_tables(struct amdgpu_device *adev, struct amdgpu_vm *vm);
603
604bool amdgpu_vm_is_bo_always_valid(struct amdgpu_vm *vm, struct amdgpu_bo *bo);
605
606/**
607 * amdgpu_vm_tlb_seq - return tlb flush sequence number
608 * @vm: the amdgpu_vm structure to query
609 *
610 * Returns the tlb flush sequence number which indicates that the VM TLBs needs
611 * to be invalidated whenever the sequence number change.
612 */
613static inline uint64_t amdgpu_vm_tlb_seq(struct amdgpu_vm *vm)
614{
615	unsigned long flags;
616	spinlock_t *lock;
617
618	/*
619	 * Workaround to stop racing between the fence signaling and handling
620	 * the cb. The lock is static after initially setting it up, just make
621	 * sure that the dma_fence structure isn't freed up.
622	 */
623	rcu_read_lock();
624	lock = vm->last_tlb_flush->lock;
625	rcu_read_unlock();
626
627	spin_lock_irqsave(lock, flags);
628	spin_unlock_irqrestore(lock, flags);
629
630	return atomic64_read(&vm->tlb_seq);
631}
632
633/*
634 * vm eviction_lock can be taken in MMU notifiers. Make sure no reclaim-FS
635 * happens while holding this lock anywhere to prevent deadlocks when
636 * an MMU notifier runs in reclaim-FS context.
637 */
638static inline void amdgpu_vm_eviction_lock(struct amdgpu_vm *vm)
639{
640	mutex_lock(&vm->eviction_lock);
641	vm->saved_flags = memalloc_noreclaim_save();
642}
643
644static inline bool amdgpu_vm_eviction_trylock(struct amdgpu_vm *vm)
645{
646	if (mutex_trylock(&vm->eviction_lock)) {
647		vm->saved_flags = memalloc_noreclaim_save();
648		return true;
649	}
650	return false;
651}
652
653static inline void amdgpu_vm_eviction_unlock(struct amdgpu_vm *vm)
654{
655	memalloc_noreclaim_restore(vm->saved_flags);
656	mutex_unlock(&vm->eviction_lock);
657}
658
659void amdgpu_vm_update_fault_cache(struct amdgpu_device *adev,
660				  unsigned int pasid,
661				  uint64_t addr,
662				  uint32_t status,
663				  unsigned int vmhub);
664void amdgpu_vm_tlb_fence_create(struct amdgpu_device *adev,
665				 struct amdgpu_vm *vm,
666				 struct dma_fence **fence);
667
668#endif
v4.10.11
  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/rbtree.h>
 
 
 
 
 28
 29#include "gpu_scheduler.h"
 30#include "amdgpu_sync.h"
 31#include "amdgpu_ring.h"
 
 
 
 32
 33struct amdgpu_bo_va;
 34struct amdgpu_job;
 35struct amdgpu_bo_list_entry;
 
 36
 37/*
 38 * GPUVM handling
 39 */
 40
 41/* maximum number of VMIDs */
 42#define AMDGPU_NUM_VM	16
 43
 44/* Maximum number of PTEs the hardware can write with one command */
 45#define AMDGPU_VM_MAX_UPDATE_SIZE	0x3FFFF
 46
 47/* number of entries in page table */
 48#define AMDGPU_VM_PTE_COUNT (1 << amdgpu_vm_block_size)
 49
 50/* PTBs (Page Table Blocks) need to be aligned to 32K */
 51#define AMDGPU_VM_PTB_ALIGN_SIZE   32768
 
 52
 53/* LOG2 number of continuous pages for the fragment field */
 54#define AMDGPU_LOG2_PAGES_PER_FRAG 4
 55
 56#define AMDGPU_PTE_VALID	(1 << 0)
 57#define AMDGPU_PTE_SYSTEM	(1 << 1)
 58#define AMDGPU_PTE_SNOOPED	(1 << 2)
 59
 60/* VI only */
 61#define AMDGPU_PTE_EXECUTABLE	(1 << 4)
 62
 63#define AMDGPU_PTE_READABLE	(1 << 5)
 64#define AMDGPU_PTE_WRITEABLE	(1 << 6)
 65
 66#define AMDGPU_PTE_FRAG(x)	((x & 0x1f) << 7)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 67
 68/* How to programm VM fault handling */
 69#define AMDGPU_VM_FAULT_STOP_NEVER	0
 70#define AMDGPU_VM_FAULT_STOP_FIRST	1
 71#define AMDGPU_VM_FAULT_STOP_ALWAYS	2
 72
 73struct amdgpu_vm_pt {
 74	struct amdgpu_bo	*bo;
 75	uint64_t		addr;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 76};
 77
 78struct amdgpu_vm {
 79	/* tree of virtual addresses mapped */
 80	struct rb_root		va;
 
 
 
 
 
 
 
 81
 82	/* protecting invalidated */
 83	spinlock_t		status_lock;
 84
 85	/* BOs moved, but not yet updated in the PT */
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 86	struct list_head	invalidated;
 87
 88	/* BOs cleared in the PT because of a move */
 89	struct list_head	cleared;
 90
 91	/* BO mappings freed, but not yet updated in the PT */
 92	struct list_head	freed;
 93
 
 
 
 
 
 
 
 94	/* contains the page directory */
 95	struct amdgpu_bo	*page_directory;
 96	unsigned		max_pde_used;
 97	struct dma_fence		*page_directory_fence;
 98	uint64_t		last_eviction_counter;
 99
100	/* array of page tables, one for each page directory entry */
101	struct amdgpu_vm_pt	*page_tables;
102
103	/* for id and flush management per ring */
104	struct amdgpu_vm_id	*ids[AMDGPU_MAX_RINGS];
105
106	/* protecting freed */
107	spinlock_t		freed_lock;
108
109	/* Scheduler entity for page table updates */
110	struct amd_sched_entity	entity;
111
112	/* client id */
113	u64                     client_id;
114};
115
116struct amdgpu_vm_id {
117	struct list_head	list;
118	struct dma_fence		*first;
119	struct amdgpu_sync	active;
120	struct dma_fence		*last_flush;
121	atomic64_t		owner;
122
123	uint64_t		pd_gpu_addr;
124	/* last flushed PD/PT update */
125	struct dma_fence		*flushed_updates;
126
127	uint32_t                current_gpu_reset_count;
128
129	uint32_t		gds_base;
130	uint32_t		gds_size;
131	uint32_t		gws_base;
132	uint32_t		gws_size;
133	uint32_t		oa_base;
134	uint32_t		oa_size;
 
 
 
 
 
 
 
 
 
 
 
 
 
135};
136
137struct amdgpu_vm_manager {
138	/* Handling of VMIDs */
139	struct mutex				lock;
140	unsigned				num_ids;
141	struct list_head			ids_lru;
142	struct amdgpu_vm_id			ids[AMDGPU_NUM_VM];
143
144	/* Handling of VM fences */
145	u64					fence_context;
146	unsigned				seqno[AMDGPU_MAX_RINGS];
147
148	uint32_t				max_pfn;
 
 
 
 
149	/* vram base address for page table entry  */
150	u64					vram_base_offset;
151	/* is vm enabled? */
152	bool					enabled;
153	/* vm pte handling */
154	const struct amdgpu_vm_pte_funcs        *vm_pte_funcs;
155	struct amdgpu_ring                      *vm_pte_rings[AMDGPU_MAX_RINGS];
156	unsigned				vm_pte_num_rings;
157	atomic_t				vm_pte_next_ring;
158	/* client id counter */
159	atomic64_t				client_counter;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
160};
161
 
 
 
 
 
 
 
 
 
162void amdgpu_vm_manager_init(struct amdgpu_device *adev);
163void amdgpu_vm_manager_fini(struct amdgpu_device *adev);
164int amdgpu_vm_init(struct amdgpu_device *adev, struct amdgpu_vm *vm);
 
 
 
 
 
 
 
165void amdgpu_vm_fini(struct amdgpu_device *adev, struct amdgpu_vm *vm);
166void amdgpu_vm_get_pd_bo(struct amdgpu_vm *vm,
167			 struct list_head *validated,
168			 struct amdgpu_bo_list_entry *entry);
169int amdgpu_vm_validate_pt_bos(struct amdgpu_device *adev, struct amdgpu_vm *vm,
170			      int (*callback)(void *p, struct amdgpu_bo *bo),
171			      void *param);
172void amdgpu_vm_move_pt_bos_in_lru(struct amdgpu_device *adev,
173				  struct amdgpu_vm *vm);
174int amdgpu_vm_grab_id(struct amdgpu_vm *vm, struct amdgpu_ring *ring,
175		      struct amdgpu_sync *sync, struct dma_fence *fence,
176		      struct amdgpu_job *job);
177int amdgpu_vm_flush(struct amdgpu_ring *ring, struct amdgpu_job *job);
178void amdgpu_vm_reset_id(struct amdgpu_device *adev, unsigned vm_id);
179int amdgpu_vm_update_page_directory(struct amdgpu_device *adev,
180				    struct amdgpu_vm *vm);
181int amdgpu_vm_clear_freed(struct amdgpu_device *adev,
182			  struct amdgpu_vm *vm);
183int amdgpu_vm_clear_invalids(struct amdgpu_device *adev, struct amdgpu_vm *vm,
184			     struct amdgpu_sync *sync);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
185int amdgpu_vm_bo_update(struct amdgpu_device *adev,
186			struct amdgpu_bo_va *bo_va,
187			bool clear);
 
188void amdgpu_vm_bo_invalidate(struct amdgpu_device *adev,
189			     struct amdgpu_bo *bo);
 
190struct amdgpu_bo_va *amdgpu_vm_bo_find(struct amdgpu_vm *vm,
191				       struct amdgpu_bo *bo);
192struct amdgpu_bo_va *amdgpu_vm_bo_add(struct amdgpu_device *adev,
193				      struct amdgpu_vm *vm,
194				      struct amdgpu_bo *bo);
195int amdgpu_vm_bo_map(struct amdgpu_device *adev,
196		     struct amdgpu_bo_va *bo_va,
197		     uint64_t addr, uint64_t offset,
198		     uint64_t size, uint32_t flags);
 
 
 
 
199int amdgpu_vm_bo_unmap(struct amdgpu_device *adev,
200		       struct amdgpu_bo_va *bo_va,
201		       uint64_t addr);
202void amdgpu_vm_bo_rmv(struct amdgpu_device *adev,
 
 
 
 
 
 
203		      struct amdgpu_bo_va *bo_va);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
204
205#endif