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_driver.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;
47 /* Maximum number of PTEs the hardware can write with one command */
48 #define AMDGPU_VM_MAX_UPDATE_SIZE 0x3FFFF
50 /* number of entries in page table */
51 #define AMDGPU_VM_PTE_COUNT(adev) (1 << (adev)->vm_manager.block_size)
53 #define AMDGPU_PTE_VALID (1ULL << 0)
54 #define AMDGPU_PTE_SYSTEM (1ULL << 1)
55 #define AMDGPU_PTE_SNOOPED (1ULL << 2)
58 #define AMDGPU_PTE_EXECUTABLE (1ULL << 4)
60 #define AMDGPU_PTE_READABLE (1ULL << 5)
61 #define AMDGPU_PTE_WRITEABLE (1ULL << 6)
63 #define AMDGPU_PTE_FRAG(x) ((x & 0x1fULL) << 7)
65 /* TILED for VEGA10, reserved for older ASICs */
66 #define AMDGPU_PTE_PRT (1ULL << 51)
68 /* PDE is handled as PTE for VEGA10 */
69 #define AMDGPU_PDE_PTE (1ULL << 54)
71 #define AMDGPU_PTE_LOG (1ULL << 55)
73 /* PTE is handled as PDE for VEGA10 (Translate Further) */
74 #define AMDGPU_PTE_TF (1ULL << 56)
76 /* PDE Block Fragment Size for VEGA10 */
77 #define AMDGPU_PDE_BFS(a) ((uint64_t)a << 59)
81 #define AMDGPU_PTE_MTYPE_VG10(a) ((uint64_t)(a) << 57)
82 #define AMDGPU_PTE_MTYPE_VG10_MASK AMDGPU_PTE_MTYPE_VG10(3ULL)
84 #define AMDGPU_MTYPE_NC 0
85 #define AMDGPU_MTYPE_CC 2
87 #define AMDGPU_PTE_DEFAULT_ATC (AMDGPU_PTE_SYSTEM \
88 | AMDGPU_PTE_SNOOPED \
89 | AMDGPU_PTE_EXECUTABLE \
90 | AMDGPU_PTE_READABLE \
91 | AMDGPU_PTE_WRITEABLE \
92 | AMDGPU_PTE_MTYPE_VG10(AMDGPU_MTYPE_CC))
95 #define AMDGPU_PTE_MTYPE_NV10(a) ((uint64_t)(a) << 48)
96 #define AMDGPU_PTE_MTYPE_NV10_MASK AMDGPU_PTE_MTYPE_NV10(7ULL)
98 /* How to programm VM fault handling */
99 #define AMDGPU_VM_FAULT_STOP_NEVER 0
100 #define AMDGPU_VM_FAULT_STOP_FIRST 1
101 #define AMDGPU_VM_FAULT_STOP_ALWAYS 2
103 /* Reserve 4MB VRAM for page tables */
104 #define AMDGPU_VM_RESERVED_VRAM (4ULL << 20)
106 /* max number of VMHUB */
107 #define AMDGPU_MAX_VMHUBS 3
108 #define AMDGPU_GFXHUB_0 0
109 #define AMDGPU_MMHUB_0 1
110 #define AMDGPU_MMHUB_1 2
112 /* hardcode that limit for now */
113 #define AMDGPU_VA_RESERVED_SIZE (1ULL << 20)
115 /* max vmids dedicated for process */
116 #define AMDGPU_VM_MAX_RESERVED_VMID 1
118 #define AMDGPU_VM_CONTEXT_GFX 0
119 #define AMDGPU_VM_CONTEXT_COMPUTE 1
121 /* See vm_update_mode */
122 #define AMDGPU_VM_USE_CPU_FOR_GFX (1 << 0)
123 #define AMDGPU_VM_USE_CPU_FOR_COMPUTE (1 << 1)
125 /* VMPT level enumerate, and the hiberachy is:
126 * PDB2->PDB1->PDB0->PTB
128 enum amdgpu_vm_level {
135 /* base structure for tracking BO usage in a VM */
136 struct amdgpu_vm_bo_base {
137 /* constant after initialization */
138 struct amdgpu_vm *vm;
139 struct amdgpu_bo *bo;
141 /* protected by bo being reserved */
142 struct amdgpu_vm_bo_base *next;
144 /* protected by spinlock */
145 struct list_head vm_status;
147 /* protected by the BO being reserved */
151 struct amdgpu_vm_pt {
152 struct amdgpu_vm_bo_base base;
154 /* array of page tables, one for each directory entry */
155 struct amdgpu_vm_pt *entries;
158 /* provided by hw blocks that can write ptes, e.g., sdma */
159 struct amdgpu_vm_pte_funcs {
160 /* number of dw to reserve per operation */
161 unsigned copy_pte_num_dw;
163 /* copy pte entries from GART */
164 void (*copy_pte)(struct amdgpu_ib *ib,
165 uint64_t pe, uint64_t src,
168 /* write pte one entry at a time with addr mapping */
169 void (*write_pte)(struct amdgpu_ib *ib, uint64_t pe,
170 uint64_t value, unsigned count,
172 /* for linear pte/pde updates without addr mapping */
173 void (*set_pte_pde)(struct amdgpu_ib *ib,
175 uint64_t addr, unsigned count,
176 uint32_t incr, uint64_t flags);
179 struct amdgpu_task_info {
180 char process_name[TASK_COMM_LEN];
181 char task_name[TASK_COMM_LEN];
187 * struct amdgpu_vm_update_params
189 * Encapsulate some VM table update parameters to reduce
190 * the number of function parameters
193 struct amdgpu_vm_update_params {
196 * @adev: amdgpu device we do this update for
198 struct amdgpu_device *adev;
201 * @vm: optional amdgpu_vm we do this update for
203 struct amdgpu_vm *vm;
206 * @direct: if changes should be made directly
213 * DMA addresses to use for mapping
215 dma_addr_t *pages_addr;
218 * @job: job to used for hw submission
220 struct amdgpu_job *job;
223 * @num_dw_left: number of dw left for the IB
225 unsigned int num_dw_left;
228 struct amdgpu_vm_update_funcs {
229 int (*map_table)(struct amdgpu_bo *bo);
230 int (*prepare)(struct amdgpu_vm_update_params *p, struct dma_resv *resv,
231 enum amdgpu_sync_mode sync_mode);
232 int (*update)(struct amdgpu_vm_update_params *p,
233 struct amdgpu_bo *bo, uint64_t pe, uint64_t addr,
234 unsigned count, uint32_t incr, uint64_t flags);
235 int (*commit)(struct amdgpu_vm_update_params *p,
236 struct dma_fence **fence);
240 /* tree of virtual addresses mapped */
241 struct rb_root_cached va;
243 /* Lock to prevent eviction while we are updating page tables
244 * use vm_eviction_lock/unlock(vm)
246 struct mutex eviction_lock;
248 unsigned int saved_flags;
250 /* BOs who needs a validation */
251 struct list_head evicted;
253 /* PT BOs which relocated and their parent need an update */
254 struct list_head relocated;
256 /* per VM BOs moved, but not yet updated in the PT */
257 struct list_head moved;
259 /* All BOs of this VM not currently in the state machine */
260 struct list_head idle;
262 /* regular invalidated BOs, but not yet updated in the PT */
263 struct list_head invalidated;
264 spinlock_t invalidated_lock;
266 /* BO mappings freed, but not yet updated in the PT */
267 struct list_head freed;
269 /* contains the page directory */
270 struct amdgpu_vm_pt root;
271 struct dma_fence *last_update;
273 /* Scheduler entities for page table updates */
274 struct drm_sched_entity direct;
275 struct drm_sched_entity delayed;
277 /* Last submission to the scheduler entities */
278 struct dma_fence *last_direct;
281 /* dedicated to vm */
282 struct amdgpu_vmid *reserved_vmid[AMDGPU_MAX_VMHUBS];
284 /* Flag to indicate if VM tables are updated by CPU or GPU (SDMA) */
285 bool use_cpu_for_update;
287 /* Functions to use for VM table updates */
288 const struct amdgpu_vm_update_funcs *update_funcs;
290 /* Flag to indicate ATS support from PTE for GFX9 */
291 bool pte_support_ats;
293 /* Up to 128 pending retry page faults */
294 DECLARE_KFIFO(faults, u64, 128);
296 /* Points to the KFD process VM info */
297 struct amdkfd_process_info *process_info;
299 /* List node in amdkfd_process_info.vm_list_head */
300 struct list_head vm_list_node;
302 /* Valid while the PD is reserved or fenced */
303 uint64_t pd_phys_addr;
305 /* Some basic info about the task */
306 struct amdgpu_task_info task_info;
308 /* Store positions of group of BOs */
309 struct ttm_lru_bulk_move lru_bulk_move;
310 /* mark whether can do the bulk move */
312 /* Flag to indicate if VM is used for compute */
313 bool is_compute_context;
316 struct amdgpu_vm_manager {
317 /* Handling of VMIDs */
318 struct amdgpu_vmid_mgr id_mgr[AMDGPU_MAX_VMHUBS];
320 /* Handling of VM fences */
322 unsigned seqno[AMDGPU_MAX_RINGS];
327 uint32_t fragment_size;
328 enum amdgpu_vm_level root_level;
329 /* vram base address for page table entry */
330 u64 vram_base_offset;
331 /* vm pte handling */
332 const struct amdgpu_vm_pte_funcs *vm_pte_funcs;
333 struct drm_gpu_scheduler *vm_pte_scheds[AMDGPU_MAX_RINGS];
334 unsigned vm_pte_num_scheds;
335 struct amdgpu_ring *page_fault;
337 /* partial resident texture handling */
339 atomic_t num_prt_users;
341 /* controls how VM page tables are updated for Graphics and Compute.
342 * BIT0[= 0] Graphics updated by SDMA [= 1] by CPU
343 * BIT1[= 0] Compute updated by SDMA [= 1] by CPU
347 /* PASID to VM mapping, will be used in interrupt context to
348 * look up VM of a page fault
350 struct idr pasid_idr;
351 spinlock_t pasid_lock;
353 /* counter of mapped memory through xgmi */
354 uint32_t xgmi_map_counter;
355 struct mutex lock_pstate;
358 #define amdgpu_vm_copy_pte(adev, ib, pe, src, count) ((adev)->vm_manager.vm_pte_funcs->copy_pte((ib), (pe), (src), (count)))
359 #define amdgpu_vm_write_pte(adev, ib, pe, value, count, incr) ((adev)->vm_manager.vm_pte_funcs->write_pte((ib), (pe), (value), (count), (incr)))
360 #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)))
362 extern const struct amdgpu_vm_update_funcs amdgpu_vm_cpu_funcs;
363 extern const struct amdgpu_vm_update_funcs amdgpu_vm_sdma_funcs;
365 void amdgpu_vm_manager_init(struct amdgpu_device *adev);
366 void amdgpu_vm_manager_fini(struct amdgpu_device *adev);
368 long amdgpu_vm_wait_idle(struct amdgpu_vm *vm, long timeout);
369 int amdgpu_vm_init(struct amdgpu_device *adev, struct amdgpu_vm *vm,
370 int vm_context, unsigned int pasid);
371 int amdgpu_vm_make_compute(struct amdgpu_device *adev, struct amdgpu_vm *vm, unsigned int pasid);
372 void amdgpu_vm_release_compute(struct amdgpu_device *adev, struct amdgpu_vm *vm);
373 void amdgpu_vm_fini(struct amdgpu_device *adev, struct amdgpu_vm *vm);
374 void amdgpu_vm_get_pd_bo(struct amdgpu_vm *vm,
375 struct list_head *validated,
376 struct amdgpu_bo_list_entry *entry);
377 bool amdgpu_vm_ready(struct amdgpu_vm *vm);
378 int amdgpu_vm_validate_pt_bos(struct amdgpu_device *adev, struct amdgpu_vm *vm,
379 int (*callback)(void *p, struct amdgpu_bo *bo),
381 int amdgpu_vm_flush(struct amdgpu_ring *ring, struct amdgpu_job *job, bool need_pipe_sync);
382 int amdgpu_vm_update_pdes(struct amdgpu_device *adev,
383 struct amdgpu_vm *vm, bool direct);
384 int amdgpu_vm_clear_freed(struct amdgpu_device *adev,
385 struct amdgpu_vm *vm,
386 struct dma_fence **fence);
387 int amdgpu_vm_handle_moved(struct amdgpu_device *adev,
388 struct amdgpu_vm *vm);
389 int amdgpu_vm_bo_update(struct amdgpu_device *adev,
390 struct amdgpu_bo_va *bo_va,
392 bool amdgpu_vm_evictable(struct amdgpu_bo *bo);
393 void amdgpu_vm_bo_invalidate(struct amdgpu_device *adev,
394 struct amdgpu_bo *bo, bool evicted);
395 uint64_t amdgpu_vm_map_gart(const dma_addr_t *pages_addr, uint64_t addr);
396 struct amdgpu_bo_va *amdgpu_vm_bo_find(struct amdgpu_vm *vm,
397 struct amdgpu_bo *bo);
398 struct amdgpu_bo_va *amdgpu_vm_bo_add(struct amdgpu_device *adev,
399 struct amdgpu_vm *vm,
400 struct amdgpu_bo *bo);
401 int amdgpu_vm_bo_map(struct amdgpu_device *adev,
402 struct amdgpu_bo_va *bo_va,
403 uint64_t addr, uint64_t offset,
404 uint64_t size, uint64_t flags);
405 int amdgpu_vm_bo_replace_map(struct amdgpu_device *adev,
406 struct amdgpu_bo_va *bo_va,
407 uint64_t addr, uint64_t offset,
408 uint64_t size, uint64_t flags);
409 int amdgpu_vm_bo_unmap(struct amdgpu_device *adev,
410 struct amdgpu_bo_va *bo_va,
412 int amdgpu_vm_bo_clear_mappings(struct amdgpu_device *adev,
413 struct amdgpu_vm *vm,
414 uint64_t saddr, uint64_t size);
415 struct amdgpu_bo_va_mapping *amdgpu_vm_bo_lookup_mapping(struct amdgpu_vm *vm,
417 void amdgpu_vm_bo_trace_cs(struct amdgpu_vm *vm, struct ww_acquire_ctx *ticket);
418 void amdgpu_vm_bo_rmv(struct amdgpu_device *adev,
419 struct amdgpu_bo_va *bo_va);
420 void amdgpu_vm_adjust_size(struct amdgpu_device *adev, uint32_t min_vm_size,
421 uint32_t fragment_size_default, unsigned max_level,
423 int amdgpu_vm_ioctl(struct drm_device *dev, void *data, struct drm_file *filp);
424 bool amdgpu_vm_need_pipeline_sync(struct amdgpu_ring *ring,
425 struct amdgpu_job *job);
426 void amdgpu_vm_check_compute_bug(struct amdgpu_device *adev);
428 void amdgpu_vm_get_task_info(struct amdgpu_device *adev, unsigned int pasid,
429 struct amdgpu_task_info *task_info);
430 bool amdgpu_vm_handle_fault(struct amdgpu_device *adev, unsigned int pasid,
433 void amdgpu_vm_set_task_info(struct amdgpu_vm *vm);
435 void amdgpu_vm_move_to_lru_tail(struct amdgpu_device *adev,
436 struct amdgpu_vm *vm);
437 void amdgpu_vm_del_from_lru_notify(struct ttm_buffer_object *bo);