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)
89 #define AMDGPU_PTE_MTYPE_VG10(a) ((uint64_t)(a) << 57)
90 #define AMDGPU_PTE_MTYPE_VG10_MASK AMDGPU_PTE_MTYPE_VG10(3ULL)
92 #define AMDGPU_MTYPE_NC 0
93 #define AMDGPU_MTYPE_CC 2
95 #define AMDGPU_PTE_DEFAULT_ATC (AMDGPU_PTE_SYSTEM \
96 | AMDGPU_PTE_SNOOPED \
97 | AMDGPU_PTE_EXECUTABLE \
98 | AMDGPU_PTE_READABLE \
99 | AMDGPU_PTE_WRITEABLE \
100 | AMDGPU_PTE_MTYPE_VG10(AMDGPU_MTYPE_CC))
103 #define AMDGPU_PTE_MTYPE_NV10(a) ((uint64_t)(a) << 48)
104 #define AMDGPU_PTE_MTYPE_NV10_MASK AMDGPU_PTE_MTYPE_NV10(7ULL)
106 /* How to program VM fault handling */
107 #define AMDGPU_VM_FAULT_STOP_NEVER 0
108 #define AMDGPU_VM_FAULT_STOP_FIRST 1
109 #define AMDGPU_VM_FAULT_STOP_ALWAYS 2
111 /* Reserve 4MB VRAM for page tables */
112 #define AMDGPU_VM_RESERVED_VRAM (8ULL << 20)
115 * max number of VMHUB
116 * layout: max 8 GFXHUB + 4 MMHUB0 + 1 MMHUB1
118 #define AMDGPU_MAX_VMHUBS 13
119 #define AMDGPU_GFXHUB(x) (x)
120 #define AMDGPU_MMHUB0(x) (8 + x)
121 #define AMDGPU_MMHUB1(x) (8 + 4 + x)
123 /* Reserve 2MB at top/bottom of address space for kernel use */
124 #define AMDGPU_VA_RESERVED_SIZE (2ULL << 20)
126 /* See vm_update_mode */
127 #define AMDGPU_VM_USE_CPU_FOR_GFX (1 << 0)
128 #define AMDGPU_VM_USE_CPU_FOR_COMPUTE (1 << 1)
130 /* VMPT level enumerate, and the hiberachy is:
131 * PDB2->PDB1->PDB0->PTB
133 enum amdgpu_vm_level {
140 /* base structure for tracking BO usage in a VM */
141 struct amdgpu_vm_bo_base {
142 /* constant after initialization */
143 struct amdgpu_vm *vm;
144 struct amdgpu_bo *bo;
146 /* protected by bo being reserved */
147 struct amdgpu_vm_bo_base *next;
149 /* protected by spinlock */
150 struct list_head vm_status;
152 /* protected by the BO being reserved */
156 /* provided by hw blocks that can write ptes, e.g., sdma */
157 struct amdgpu_vm_pte_funcs {
158 /* number of dw to reserve per operation */
159 unsigned copy_pte_num_dw;
161 /* copy pte entries from GART */
162 void (*copy_pte)(struct amdgpu_ib *ib,
163 uint64_t pe, uint64_t src,
166 /* write pte one entry at a time with addr mapping */
167 void (*write_pte)(struct amdgpu_ib *ib, uint64_t pe,
168 uint64_t value, unsigned count,
170 /* for linear pte/pde updates without addr mapping */
171 void (*set_pte_pde)(struct amdgpu_ib *ib,
173 uint64_t addr, unsigned count,
174 uint32_t incr, uint64_t flags);
177 struct amdgpu_task_info {
178 char process_name[TASK_COMM_LEN];
179 char task_name[TASK_COMM_LEN];
185 * struct amdgpu_vm_update_params
187 * Encapsulate some VM table update parameters to reduce
188 * the number of function parameters
191 struct amdgpu_vm_update_params {
194 * @adev: amdgpu device we do this update for
196 struct amdgpu_device *adev;
199 * @vm: optional amdgpu_vm we do this update for
201 struct amdgpu_vm *vm;
204 * @immediate: if changes should be made immediately
209 * @unlocked: true if the root BO is not locked
216 * DMA addresses to use for mapping
218 dma_addr_t *pages_addr;
221 * @job: job to used for hw submission
223 struct amdgpu_job *job;
226 * @num_dw_left: number of dw left for the IB
228 unsigned int num_dw_left;
231 * @table_freed: return true if page table is freed when updating
236 struct amdgpu_vm_update_funcs {
237 int (*map_table)(struct amdgpu_bo_vm *bo);
238 int (*prepare)(struct amdgpu_vm_update_params *p, struct dma_resv *resv,
239 enum amdgpu_sync_mode sync_mode);
240 int (*update)(struct amdgpu_vm_update_params *p,
241 struct amdgpu_bo_vm *bo, uint64_t pe, uint64_t addr,
242 unsigned count, uint32_t incr, uint64_t flags);
243 int (*commit)(struct amdgpu_vm_update_params *p,
244 struct dma_fence **fence);
248 /* tree of virtual addresses mapped */
249 struct rb_root_cached va;
251 /* Lock to prevent eviction while we are updating page tables
252 * use vm_eviction_lock/unlock(vm)
254 struct mutex eviction_lock;
256 unsigned int saved_flags;
258 /* Lock to protect vm_bo add/del/move on all lists of vm */
259 spinlock_t status_lock;
261 /* BOs who needs a validation */
262 struct list_head evicted;
264 /* PT BOs which relocated and their parent need an update */
265 struct list_head relocated;
267 /* per VM BOs moved, but not yet updated in the PT */
268 struct list_head moved;
270 /* All BOs of this VM not currently in the state machine */
271 struct list_head idle;
273 /* regular invalidated BOs, but not yet updated in the PT */
274 struct list_head invalidated;
276 /* BO mappings freed, but not yet updated in the PT */
277 struct list_head freed;
279 /* BOs which are invalidated, has been updated in the PTs */
280 struct list_head done;
282 /* PT BOs scheduled to free and fill with zero if vm_resv is not hold */
283 struct list_head pt_freed;
284 struct work_struct pt_free_work;
286 /* contains the page directory */
287 struct amdgpu_vm_bo_base root;
288 struct dma_fence *last_update;
290 /* Scheduler entities for page table updates */
291 struct drm_sched_entity immediate;
292 struct drm_sched_entity delayed;
294 /* Last finished delayed update */
296 struct dma_fence *last_tlb_flush;
298 /* Last unlocked submission to the scheduler entities */
299 struct dma_fence *last_unlocked;
302 bool reserved_vmid[AMDGPU_MAX_VMHUBS];
304 /* Flag to indicate if VM tables are updated by CPU or GPU (SDMA) */
305 bool use_cpu_for_update;
307 /* Functions to use for VM table updates */
308 const struct amdgpu_vm_update_funcs *update_funcs;
310 /* Flag to indicate ATS support from PTE for GFX9 */
311 bool pte_support_ats;
313 /* Up to 128 pending retry page faults */
314 DECLARE_KFIFO(faults, u64, 128);
316 /* Points to the KFD process VM info */
317 struct amdkfd_process_info *process_info;
319 /* List node in amdkfd_process_info.vm_list_head */
320 struct list_head vm_list_node;
322 /* Valid while the PD is reserved or fenced */
323 uint64_t pd_phys_addr;
325 /* Some basic info about the task */
326 struct amdgpu_task_info task_info;
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;
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;
340 /* Handling of VM fences */
342 unsigned seqno[AMDGPU_MAX_RINGS];
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;
357 /* partial resident texture handling */
359 atomic_t num_prt_users;
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
367 /* PASID to VM mapping, will be used in interrupt context to
368 * look up VM of a page fault
370 struct xarray pasids;
373 struct amdgpu_bo_va_mapping;
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)))
379 extern const struct amdgpu_vm_update_funcs amdgpu_vm_cpu_funcs;
380 extern const struct amdgpu_vm_update_funcs amdgpu_vm_sdma_funcs;
382 void amdgpu_vm_manager_init(struct amdgpu_device *adev);
383 void amdgpu_vm_manager_fini(struct amdgpu_device *adev);
385 int amdgpu_vm_set_pasid(struct amdgpu_device *adev, struct amdgpu_vm *vm,
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),
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,
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,
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,
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,
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);
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 u32 client_id, u32 node_id, uint64_t addr,
461 void amdgpu_vm_set_task_info(struct amdgpu_vm *vm);
463 void amdgpu_vm_move_to_lru_tail(struct amdgpu_device *adev,
464 struct amdgpu_vm *vm);
465 void amdgpu_vm_get_memory(struct amdgpu_vm *vm,
466 struct amdgpu_mem_stats *stats);
468 int amdgpu_vm_pt_clear(struct amdgpu_device *adev, struct amdgpu_vm *vm,
469 struct amdgpu_bo_vm *vmbo, bool immediate);
470 int amdgpu_vm_pt_create(struct amdgpu_device *adev, struct amdgpu_vm *vm,
471 int level, bool immediate, struct amdgpu_bo_vm **vmbo);
472 void amdgpu_vm_pt_free_root(struct amdgpu_device *adev, struct amdgpu_vm *vm);
473 bool amdgpu_vm_pt_is_root_clean(struct amdgpu_device *adev,
474 struct amdgpu_vm *vm);
476 int amdgpu_vm_pde_update(struct amdgpu_vm_update_params *params,
477 struct amdgpu_vm_bo_base *entry);
478 int amdgpu_vm_ptes_update(struct amdgpu_vm_update_params *params,
479 uint64_t start, uint64_t end,
480 uint64_t dst, uint64_t flags);
481 void amdgpu_vm_pt_free_work(struct work_struct *work);
483 #if defined(CONFIG_DEBUG_FS)
484 void amdgpu_debugfs_vm_bo_info(struct amdgpu_vm *vm, struct seq_file *m);
488 * amdgpu_vm_tlb_seq - return tlb flush sequence number
489 * @vm: the amdgpu_vm structure to query
491 * Returns the tlb flush sequence number which indicates that the VM TLBs needs
492 * to be invalidated whenever the sequence number change.
494 static inline uint64_t amdgpu_vm_tlb_seq(struct amdgpu_vm *vm)
500 * Workaround to stop racing between the fence signaling and handling
501 * the cb. The lock is static after initially setting it up, just make
502 * sure that the dma_fence structure isn't freed up.
505 lock = vm->last_tlb_flush->lock;
508 spin_lock_irqsave(lock, flags);
509 spin_unlock_irqrestore(lock, flags);
511 return atomic64_read(&vm->tlb_seq);
515 * vm eviction_lock can be taken in MMU notifiers. Make sure no reclaim-FS
516 * happens while holding this lock anywhere to prevent deadlocks when
517 * an MMU notifier runs in reclaim-FS context.
519 static inline void amdgpu_vm_eviction_lock(struct amdgpu_vm *vm)
521 mutex_lock(&vm->eviction_lock);
522 vm->saved_flags = memalloc_noreclaim_save();
525 static inline bool amdgpu_vm_eviction_trylock(struct amdgpu_vm *vm)
527 if (mutex_trylock(&vm->eviction_lock)) {
528 vm->saved_flags = memalloc_noreclaim_save();
534 static inline void amdgpu_vm_eviction_unlock(struct amdgpu_vm *vm)
536 memalloc_noreclaim_restore(vm->saved_flags);
537 mutex_unlock(&vm->eviction_lock);