2 * Copyright 2016 Advanced Micro Devices, Inc.
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:
11 * The above copyright notice and this permission notice shall be included in
12 * all copies or substantial portions of the Software.
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.
22 * Authors: Christian König
24 #ifndef __AMDGPU_VM_H__
25 #define __AMDGPU_VM_H__
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>
35 #include "amdgpu_sync.h"
36 #include "amdgpu_ring.h"
37 #include "amdgpu_ids.h"
41 struct amdgpu_bo_list_entry;
43 struct amdgpu_mem_stats;
49 /* Maximum number of PTEs the hardware can write with one command */
50 #define AMDGPU_VM_MAX_UPDATE_SIZE 0x3FFFF
52 /* number of entries in page table */
53 #define AMDGPU_VM_PTE_COUNT(adev) (1 << (adev)->vm_manager.block_size)
55 #define AMDGPU_PTE_VALID (1ULL << 0)
56 #define AMDGPU_PTE_SYSTEM (1ULL << 1)
57 #define AMDGPU_PTE_SNOOPED (1ULL << 2)
60 #define AMDGPU_PTE_TMZ (1ULL << 3)
63 #define AMDGPU_PTE_EXECUTABLE (1ULL << 4)
65 #define AMDGPU_PTE_READABLE (1ULL << 5)
66 #define AMDGPU_PTE_WRITEABLE (1ULL << 6)
68 #define AMDGPU_PTE_FRAG(x) ((x & 0x1fULL) << 7)
70 /* TILED for VEGA10, reserved for older ASICs */
71 #define AMDGPU_PTE_PRT (1ULL << 51)
73 /* PDE is handled as PTE for VEGA10 */
74 #define AMDGPU_PDE_PTE (1ULL << 54)
76 #define AMDGPU_PTE_LOG (1ULL << 55)
78 /* PTE is handled as PDE for VEGA10 (Translate Further) */
79 #define AMDGPU_PTE_TF (1ULL << 56)
81 /* MALL noalloc for sienna_cichlid, reserved for older ASICs */
82 #define AMDGPU_PTE_NOALLOC (1ULL << 58)
84 /* PDE Block Fragment Size for VEGA10 */
85 #define AMDGPU_PDE_BFS(a) ((uint64_t)a << 59)
87 /* Flag combination to set no-retry with TF disabled */
88 #define AMDGPU_VM_NORETRY_FLAGS (AMDGPU_PTE_EXECUTABLE | AMDGPU_PDE_PTE | \
91 /* Flag combination to set no-retry with TF enabled */
92 #define AMDGPU_VM_NORETRY_FLAGS_TF (AMDGPU_PTE_VALID | AMDGPU_PTE_SYSTEM | \
95 #define AMDGPU_PTE_MTYPE_VG10(a) ((uint64_t)(a) << 57)
96 #define AMDGPU_PTE_MTYPE_VG10_MASK AMDGPU_PTE_MTYPE_VG10(3ULL)
98 #define AMDGPU_MTYPE_NC 0
99 #define AMDGPU_MTYPE_CC 2
101 #define AMDGPU_PTE_DEFAULT_ATC (AMDGPU_PTE_SYSTEM \
102 | AMDGPU_PTE_SNOOPED \
103 | AMDGPU_PTE_EXECUTABLE \
104 | AMDGPU_PTE_READABLE \
105 | AMDGPU_PTE_WRITEABLE \
106 | AMDGPU_PTE_MTYPE_VG10(AMDGPU_MTYPE_CC))
109 #define AMDGPU_PTE_MTYPE_NV10(a) ((uint64_t)(a) << 48)
110 #define AMDGPU_PTE_MTYPE_NV10_MASK AMDGPU_PTE_MTYPE_NV10(7ULL)
112 /* How to program VM fault handling */
113 #define AMDGPU_VM_FAULT_STOP_NEVER 0
114 #define AMDGPU_VM_FAULT_STOP_FIRST 1
115 #define AMDGPU_VM_FAULT_STOP_ALWAYS 2
117 /* Reserve 4MB VRAM for page tables */
118 #define AMDGPU_VM_RESERVED_VRAM (8ULL << 20)
121 * max number of VMHUB
122 * layout: max 8 GFXHUB + 4 MMHUB0 + 1 MMHUB1
124 #define AMDGPU_MAX_VMHUBS 13
125 #define AMDGPU_GFXHUB(x) (x)
126 #define AMDGPU_MMHUB0(x) (8 + x)
127 #define AMDGPU_MMHUB1(x) (8 + 4 + x)
129 /* Reserve 2MB at top/bottom of address space for kernel use */
130 #define AMDGPU_VA_RESERVED_SIZE (2ULL << 20)
132 /* See vm_update_mode */
133 #define AMDGPU_VM_USE_CPU_FOR_GFX (1 << 0)
134 #define AMDGPU_VM_USE_CPU_FOR_COMPUTE (1 << 1)
136 /* VMPT level enumerate, and the hiberachy is:
137 * PDB2->PDB1->PDB0->PTB
139 enum amdgpu_vm_level {
146 /* base structure for tracking BO usage in a VM */
147 struct amdgpu_vm_bo_base {
148 /* constant after initialization */
149 struct amdgpu_vm *vm;
150 struct amdgpu_bo *bo;
152 /* protected by bo being reserved */
153 struct amdgpu_vm_bo_base *next;
155 /* protected by spinlock */
156 struct list_head vm_status;
158 /* protected by the BO being reserved */
162 /* provided by hw blocks that can write ptes, e.g., sdma */
163 struct amdgpu_vm_pte_funcs {
164 /* number of dw to reserve per operation */
165 unsigned copy_pte_num_dw;
167 /* copy pte entries from GART */
168 void (*copy_pte)(struct amdgpu_ib *ib,
169 uint64_t pe, uint64_t src,
172 /* write pte one entry at a time with addr mapping */
173 void (*write_pte)(struct amdgpu_ib *ib, uint64_t pe,
174 uint64_t value, unsigned count,
176 /* for linear pte/pde updates without addr mapping */
177 void (*set_pte_pde)(struct amdgpu_ib *ib,
179 uint64_t addr, unsigned count,
180 uint32_t incr, uint64_t flags);
183 struct amdgpu_task_info {
184 char process_name[TASK_COMM_LEN];
185 char task_name[TASK_COMM_LEN];
191 * struct amdgpu_vm_update_params
193 * Encapsulate some VM table update parameters to reduce
194 * the number of function parameters
197 struct amdgpu_vm_update_params {
200 * @adev: amdgpu device we do this update for
202 struct amdgpu_device *adev;
205 * @vm: optional amdgpu_vm we do this update for
207 struct amdgpu_vm *vm;
210 * @immediate: if changes should be made immediately
215 * @unlocked: true if the root BO is not locked
222 * DMA addresses to use for mapping
224 dma_addr_t *pages_addr;
227 * @job: job to used for hw submission
229 struct amdgpu_job *job;
232 * @num_dw_left: number of dw left for the IB
234 unsigned int num_dw_left;
237 * @table_freed: return true if page table is freed when updating
242 struct amdgpu_vm_update_funcs {
243 int (*map_table)(struct amdgpu_bo_vm *bo);
244 int (*prepare)(struct amdgpu_vm_update_params *p, struct dma_resv *resv,
245 enum amdgpu_sync_mode sync_mode);
246 int (*update)(struct amdgpu_vm_update_params *p,
247 struct amdgpu_bo_vm *bo, uint64_t pe, uint64_t addr,
248 unsigned count, uint32_t incr, uint64_t flags);
249 int (*commit)(struct amdgpu_vm_update_params *p,
250 struct dma_fence **fence);
254 /* tree of virtual addresses mapped */
255 struct rb_root_cached va;
257 /* Lock to prevent eviction while we are updating page tables
258 * use vm_eviction_lock/unlock(vm)
260 struct mutex eviction_lock;
262 unsigned int saved_flags;
264 /* Lock to protect vm_bo add/del/move on all lists of vm */
265 spinlock_t status_lock;
267 /* BOs who needs a validation */
268 struct list_head evicted;
270 /* PT BOs which relocated and their parent need an update */
271 struct list_head relocated;
273 /* per VM BOs moved, but not yet updated in the PT */
274 struct list_head moved;
276 /* All BOs of this VM not currently in the state machine */
277 struct list_head idle;
279 /* regular invalidated BOs, but not yet updated in the PT */
280 struct list_head invalidated;
282 /* BO mappings freed, but not yet updated in the PT */
283 struct list_head freed;
285 /* BOs which are invalidated, has been updated in the PTs */
286 struct list_head done;
288 /* PT BOs scheduled to free and fill with zero if vm_resv is not hold */
289 struct list_head pt_freed;
290 struct work_struct pt_free_work;
292 /* contains the page directory */
293 struct amdgpu_vm_bo_base root;
294 struct dma_fence *last_update;
296 /* Scheduler entities for page table updates */
297 struct drm_sched_entity immediate;
298 struct drm_sched_entity delayed;
300 /* Last finished delayed update */
302 struct dma_fence *last_tlb_flush;
304 /* How many times we had to re-generate the page tables */
307 /* Last unlocked submission to the scheduler entities */
308 struct dma_fence *last_unlocked;
311 bool reserved_vmid[AMDGPU_MAX_VMHUBS];
313 /* Flag to indicate if VM tables are updated by CPU or GPU (SDMA) */
314 bool use_cpu_for_update;
316 /* Functions to use for VM table updates */
317 const struct amdgpu_vm_update_funcs *update_funcs;
319 /* Flag to indicate ATS support from PTE for GFX9 */
320 bool pte_support_ats;
322 /* Up to 128 pending retry page faults */
323 DECLARE_KFIFO(faults, u64, 128);
325 /* Points to the KFD process VM info */
326 struct amdkfd_process_info *process_info;
328 /* List node in amdkfd_process_info.vm_list_head */
329 struct list_head vm_list_node;
331 /* Valid while the PD is reserved or fenced */
332 uint64_t pd_phys_addr;
334 /* Some basic info about the task */
335 struct amdgpu_task_info task_info;
337 /* Store positions of group of BOs */
338 struct ttm_lru_bulk_move lru_bulk_move;
339 /* Flag to indicate if VM is used for compute */
340 bool is_compute_context;
342 /* Memory partition number, -1 means any partition */
346 struct amdgpu_vm_manager {
347 /* Handling of VMIDs */
348 struct amdgpu_vmid_mgr id_mgr[AMDGPU_MAX_VMHUBS];
349 unsigned int first_kfd_vmid;
350 bool concurrent_flush;
352 /* Handling of VM fences */
354 unsigned seqno[AMDGPU_MAX_RINGS];
359 uint32_t fragment_size;
360 enum amdgpu_vm_level root_level;
361 /* vram base address for page table entry */
362 u64 vram_base_offset;
363 /* vm pte handling */
364 const struct amdgpu_vm_pte_funcs *vm_pte_funcs;
365 struct drm_gpu_scheduler *vm_pte_scheds[AMDGPU_MAX_RINGS];
366 unsigned vm_pte_num_scheds;
367 struct amdgpu_ring *page_fault;
369 /* partial resident texture handling */
371 atomic_t num_prt_users;
373 /* controls how VM page tables are updated for Graphics and Compute.
374 * BIT0[= 0] Graphics updated by SDMA [= 1] by CPU
375 * BIT1[= 0] Compute updated by SDMA [= 1] by CPU
379 /* PASID to VM mapping, will be used in interrupt context to
380 * look up VM of a page fault
382 struct xarray pasids;
385 struct amdgpu_bo_va_mapping;
387 #define amdgpu_vm_copy_pte(adev, ib, pe, src, count) ((adev)->vm_manager.vm_pte_funcs->copy_pte((ib), (pe), (src), (count)))
388 #define amdgpu_vm_write_pte(adev, ib, pe, value, count, incr) ((adev)->vm_manager.vm_pte_funcs->write_pte((ib), (pe), (value), (count), (incr)))
389 #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)))
391 extern const struct amdgpu_vm_update_funcs amdgpu_vm_cpu_funcs;
392 extern const struct amdgpu_vm_update_funcs amdgpu_vm_sdma_funcs;
394 void amdgpu_vm_manager_init(struct amdgpu_device *adev);
395 void amdgpu_vm_manager_fini(struct amdgpu_device *adev);
397 int amdgpu_vm_set_pasid(struct amdgpu_device *adev, struct amdgpu_vm *vm,
400 long amdgpu_vm_wait_idle(struct amdgpu_vm *vm, long timeout);
401 int amdgpu_vm_init(struct amdgpu_device *adev, struct amdgpu_vm *vm, int32_t xcp_id);
402 int amdgpu_vm_make_compute(struct amdgpu_device *adev, struct amdgpu_vm *vm);
403 void amdgpu_vm_release_compute(struct amdgpu_device *adev, struct amdgpu_vm *vm);
404 void amdgpu_vm_fini(struct amdgpu_device *adev, struct amdgpu_vm *vm);
405 void amdgpu_vm_get_pd_bo(struct amdgpu_vm *vm,
406 struct list_head *validated,
407 struct amdgpu_bo_list_entry *entry);
408 bool amdgpu_vm_ready(struct amdgpu_vm *vm);
409 uint64_t amdgpu_vm_generation(struct amdgpu_device *adev, struct amdgpu_vm *vm);
410 int amdgpu_vm_validate_pt_bos(struct amdgpu_device *adev, struct amdgpu_vm *vm,
411 int (*callback)(void *p, struct amdgpu_bo *bo),
413 int amdgpu_vm_flush(struct amdgpu_ring *ring, struct amdgpu_job *job, bool need_pipe_sync);
414 int amdgpu_vm_update_pdes(struct amdgpu_device *adev,
415 struct amdgpu_vm *vm, bool immediate);
416 int amdgpu_vm_clear_freed(struct amdgpu_device *adev,
417 struct amdgpu_vm *vm,
418 struct dma_fence **fence);
419 int amdgpu_vm_handle_moved(struct amdgpu_device *adev,
420 struct amdgpu_vm *vm);
421 void amdgpu_vm_bo_base_init(struct amdgpu_vm_bo_base *base,
422 struct amdgpu_vm *vm, struct amdgpu_bo *bo);
423 int amdgpu_vm_update_range(struct amdgpu_device *adev, struct amdgpu_vm *vm,
424 bool immediate, bool unlocked, bool flush_tlb,
425 struct dma_resv *resv, uint64_t start, uint64_t last,
426 uint64_t flags, uint64_t offset, uint64_t vram_base,
427 struct ttm_resource *res, dma_addr_t *pages_addr,
428 struct dma_fence **fence);
429 int amdgpu_vm_bo_update(struct amdgpu_device *adev,
430 struct amdgpu_bo_va *bo_va,
432 bool amdgpu_vm_evictable(struct amdgpu_bo *bo);
433 void amdgpu_vm_bo_invalidate(struct amdgpu_device *adev,
434 struct amdgpu_bo *bo, bool evicted);
435 uint64_t amdgpu_vm_map_gart(const dma_addr_t *pages_addr, uint64_t addr);
436 struct amdgpu_bo_va *amdgpu_vm_bo_find(struct amdgpu_vm *vm,
437 struct amdgpu_bo *bo);
438 struct amdgpu_bo_va *amdgpu_vm_bo_add(struct amdgpu_device *adev,
439 struct amdgpu_vm *vm,
440 struct amdgpu_bo *bo);
441 int amdgpu_vm_bo_map(struct amdgpu_device *adev,
442 struct amdgpu_bo_va *bo_va,
443 uint64_t addr, uint64_t offset,
444 uint64_t size, uint64_t flags);
445 int amdgpu_vm_bo_replace_map(struct amdgpu_device *adev,
446 struct amdgpu_bo_va *bo_va,
447 uint64_t addr, uint64_t offset,
448 uint64_t size, uint64_t flags);
449 int amdgpu_vm_bo_unmap(struct amdgpu_device *adev,
450 struct amdgpu_bo_va *bo_va,
452 int amdgpu_vm_bo_clear_mappings(struct amdgpu_device *adev,
453 struct amdgpu_vm *vm,
454 uint64_t saddr, uint64_t size);
455 struct amdgpu_bo_va_mapping *amdgpu_vm_bo_lookup_mapping(struct amdgpu_vm *vm,
457 void amdgpu_vm_bo_trace_cs(struct amdgpu_vm *vm, struct ww_acquire_ctx *ticket);
458 void amdgpu_vm_bo_del(struct amdgpu_device *adev,
459 struct amdgpu_bo_va *bo_va);
460 void amdgpu_vm_adjust_size(struct amdgpu_device *adev, uint32_t min_vm_size,
461 uint32_t fragment_size_default, unsigned max_level,
463 int amdgpu_vm_ioctl(struct drm_device *dev, void *data, struct drm_file *filp);
464 bool amdgpu_vm_need_pipeline_sync(struct amdgpu_ring *ring,
465 struct amdgpu_job *job);
466 void amdgpu_vm_check_compute_bug(struct amdgpu_device *adev);
468 void amdgpu_vm_get_task_info(struct amdgpu_device *adev, u32 pasid,
469 struct amdgpu_task_info *task_info);
470 bool amdgpu_vm_handle_fault(struct amdgpu_device *adev, u32 pasid,
471 u32 vmid, u32 node_id, uint64_t addr,
474 void amdgpu_vm_set_task_info(struct amdgpu_vm *vm);
476 void amdgpu_vm_move_to_lru_tail(struct amdgpu_device *adev,
477 struct amdgpu_vm *vm);
478 void amdgpu_vm_get_memory(struct amdgpu_vm *vm,
479 struct amdgpu_mem_stats *stats);
481 int amdgpu_vm_pt_clear(struct amdgpu_device *adev, struct amdgpu_vm *vm,
482 struct amdgpu_bo_vm *vmbo, bool immediate);
483 int amdgpu_vm_pt_create(struct amdgpu_device *adev, struct amdgpu_vm *vm,
484 int level, bool immediate, struct amdgpu_bo_vm **vmbo,
486 void amdgpu_vm_pt_free_root(struct amdgpu_device *adev, struct amdgpu_vm *vm);
487 bool amdgpu_vm_pt_is_root_clean(struct amdgpu_device *adev,
488 struct amdgpu_vm *vm);
490 int amdgpu_vm_pde_update(struct amdgpu_vm_update_params *params,
491 struct amdgpu_vm_bo_base *entry);
492 int amdgpu_vm_ptes_update(struct amdgpu_vm_update_params *params,
493 uint64_t start, uint64_t end,
494 uint64_t dst, uint64_t flags);
495 void amdgpu_vm_pt_free_work(struct work_struct *work);
497 #if defined(CONFIG_DEBUG_FS)
498 void amdgpu_debugfs_vm_bo_info(struct amdgpu_vm *vm, struct seq_file *m);
501 int amdgpu_vm_pt_map_tables(struct amdgpu_device *adev, struct amdgpu_vm *vm);
504 * amdgpu_vm_tlb_seq - return tlb flush sequence number
505 * @vm: the amdgpu_vm structure to query
507 * Returns the tlb flush sequence number which indicates that the VM TLBs needs
508 * to be invalidated whenever the sequence number change.
510 static inline uint64_t amdgpu_vm_tlb_seq(struct amdgpu_vm *vm)
516 * Workaround to stop racing between the fence signaling and handling
517 * the cb. The lock is static after initially setting it up, just make
518 * sure that the dma_fence structure isn't freed up.
521 lock = vm->last_tlb_flush->lock;
524 spin_lock_irqsave(lock, flags);
525 spin_unlock_irqrestore(lock, flags);
527 return atomic64_read(&vm->tlb_seq);
531 * vm eviction_lock can be taken in MMU notifiers. Make sure no reclaim-FS
532 * happens while holding this lock anywhere to prevent deadlocks when
533 * an MMU notifier runs in reclaim-FS context.
535 static inline void amdgpu_vm_eviction_lock(struct amdgpu_vm *vm)
537 mutex_lock(&vm->eviction_lock);
538 vm->saved_flags = memalloc_noreclaim_save();
541 static inline bool amdgpu_vm_eviction_trylock(struct amdgpu_vm *vm)
543 if (mutex_trylock(&vm->eviction_lock)) {
544 vm->saved_flags = memalloc_noreclaim_save();
550 static inline void amdgpu_vm_eviction_unlock(struct amdgpu_vm *vm)
552 memalloc_noreclaim_restore(vm->saved_flags);
553 mutex_unlock(&vm->eviction_lock);