Loading...
Note: File does not exist in v3.1.
1/*
2 * Copyright 2008 Advanced Micro Devices, Inc.
3 * Copyright 2008 Red Hat Inc.
4 * Copyright 2009 Jerome Glisse.
5 *
6 * Permission is hereby granted, free of charge, to any person obtaining a
7 * copy of this software and associated documentation files (the "Software"),
8 * to deal in the Software without restriction, including without limitation
9 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
10 * and/or sell copies of the Software, and to permit persons to whom the
11 * Software is furnished to do so, subject to the following conditions:
12 *
13 * The above copyright notice and this permission notice shall be included in
14 * all copies or substantial portions of the Software.
15 *
16 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
17 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
18 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
19 * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR
20 * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
21 * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
22 * OTHER DEALINGS IN THE SOFTWARE.
23 *
24 * Authors: Dave Airlie
25 * Alex Deucher
26 * Jerome Glisse
27 */
28#include <linux/dma-fence-array.h>
29#include <linux/interval_tree_generic.h>
30#include <linux/idr.h>
31
32#include <drm/amdgpu_drm.h>
33#include "amdgpu.h"
34#include "amdgpu_trace.h"
35#include "amdgpu_amdkfd.h"
36#include "amdgpu_gmc.h"
37#include "amdgpu_xgmi.h"
38
39/**
40 * DOC: GPUVM
41 *
42 * GPUVM is similar to the legacy gart on older asics, however
43 * rather than there being a single global gart table
44 * for the entire GPU, there are multiple VM page tables active
45 * at any given time. The VM page tables can contain a mix
46 * vram pages and system memory pages and system memory pages
47 * can be mapped as snooped (cached system pages) or unsnooped
48 * (uncached system pages).
49 * Each VM has an ID associated with it and there is a page table
50 * associated with each VMID. When execting a command buffer,
51 * the kernel tells the the ring what VMID to use for that command
52 * buffer. VMIDs are allocated dynamically as commands are submitted.
53 * The userspace drivers maintain their own address space and the kernel
54 * sets up their pages tables accordingly when they submit their
55 * command buffers and a VMID is assigned.
56 * Cayman/Trinity support up to 8 active VMs at any given time;
57 * SI supports 16.
58 */
59
60#define START(node) ((node)->start)
61#define LAST(node) ((node)->last)
62
63INTERVAL_TREE_DEFINE(struct amdgpu_bo_va_mapping, rb, uint64_t, __subtree_last,
64 START, LAST, static, amdgpu_vm_it)
65
66#undef START
67#undef LAST
68
69/**
70 * struct amdgpu_prt_cb - Helper to disable partial resident texture feature from a fence callback
71 */
72struct amdgpu_prt_cb {
73
74 /**
75 * @adev: amdgpu device
76 */
77 struct amdgpu_device *adev;
78
79 /**
80 * @cb: callback
81 */
82 struct dma_fence_cb cb;
83};
84
85/**
86 * amdgpu_vm_level_shift - return the addr shift for each level
87 *
88 * @adev: amdgpu_device pointer
89 * @level: VMPT level
90 *
91 * Returns:
92 * The number of bits the pfn needs to be right shifted for a level.
93 */
94static unsigned amdgpu_vm_level_shift(struct amdgpu_device *adev,
95 unsigned level)
96{
97 unsigned shift = 0xff;
98
99 switch (level) {
100 case AMDGPU_VM_PDB2:
101 case AMDGPU_VM_PDB1:
102 case AMDGPU_VM_PDB0:
103 shift = 9 * (AMDGPU_VM_PDB0 - level) +
104 adev->vm_manager.block_size;
105 break;
106 case AMDGPU_VM_PTB:
107 shift = 0;
108 break;
109 default:
110 dev_err(adev->dev, "the level%d isn't supported.\n", level);
111 }
112
113 return shift;
114}
115
116/**
117 * amdgpu_vm_num_entries - return the number of entries in a PD/PT
118 *
119 * @adev: amdgpu_device pointer
120 * @level: VMPT level
121 *
122 * Returns:
123 * The number of entries in a page directory or page table.
124 */
125static unsigned amdgpu_vm_num_entries(struct amdgpu_device *adev,
126 unsigned level)
127{
128 unsigned shift = amdgpu_vm_level_shift(adev,
129 adev->vm_manager.root_level);
130
131 if (level == adev->vm_manager.root_level)
132 /* For the root directory */
133 return round_up(adev->vm_manager.max_pfn, 1ULL << shift) >> shift;
134 else if (level != AMDGPU_VM_PTB)
135 /* Everything in between */
136 return 512;
137 else
138 /* For the page tables on the leaves */
139 return AMDGPU_VM_PTE_COUNT(adev);
140}
141
142/**
143 * amdgpu_vm_num_ats_entries - return the number of ATS entries in the root PD
144 *
145 * @adev: amdgpu_device pointer
146 *
147 * Returns:
148 * The number of entries in the root page directory which needs the ATS setting.
149 */
150static unsigned amdgpu_vm_num_ats_entries(struct amdgpu_device *adev)
151{
152 unsigned shift;
153
154 shift = amdgpu_vm_level_shift(adev, adev->vm_manager.root_level);
155 return AMDGPU_GMC_HOLE_START >> (shift + AMDGPU_GPU_PAGE_SHIFT);
156}
157
158/**
159 * amdgpu_vm_entries_mask - the mask to get the entry number of a PD/PT
160 *
161 * @adev: amdgpu_device pointer
162 * @level: VMPT level
163 *
164 * Returns:
165 * The mask to extract the entry number of a PD/PT from an address.
166 */
167static uint32_t amdgpu_vm_entries_mask(struct amdgpu_device *adev,
168 unsigned int level)
169{
170 if (level <= adev->vm_manager.root_level)
171 return 0xffffffff;
172 else if (level != AMDGPU_VM_PTB)
173 return 0x1ff;
174 else
175 return AMDGPU_VM_PTE_COUNT(adev) - 1;
176}
177
178/**
179 * amdgpu_vm_bo_size - returns the size of the BOs in bytes
180 *
181 * @adev: amdgpu_device pointer
182 * @level: VMPT level
183 *
184 * Returns:
185 * The size of the BO for a page directory or page table in bytes.
186 */
187static unsigned amdgpu_vm_bo_size(struct amdgpu_device *adev, unsigned level)
188{
189 return AMDGPU_GPU_PAGE_ALIGN(amdgpu_vm_num_entries(adev, level) * 8);
190}
191
192/**
193 * amdgpu_vm_bo_evicted - vm_bo is evicted
194 *
195 * @vm_bo: vm_bo which is evicted
196 *
197 * State for PDs/PTs and per VM BOs which are not at the location they should
198 * be.
199 */
200static void amdgpu_vm_bo_evicted(struct amdgpu_vm_bo_base *vm_bo)
201{
202 struct amdgpu_vm *vm = vm_bo->vm;
203 struct amdgpu_bo *bo = vm_bo->bo;
204
205 vm_bo->moved = true;
206 if (bo->tbo.type == ttm_bo_type_kernel)
207 list_move(&vm_bo->vm_status, &vm->evicted);
208 else
209 list_move_tail(&vm_bo->vm_status, &vm->evicted);
210}
211
212/**
213 * amdgpu_vm_bo_relocated - vm_bo is reloacted
214 *
215 * @vm_bo: vm_bo which is relocated
216 *
217 * State for PDs/PTs which needs to update their parent PD.
218 */
219static void amdgpu_vm_bo_relocated(struct amdgpu_vm_bo_base *vm_bo)
220{
221 list_move(&vm_bo->vm_status, &vm_bo->vm->relocated);
222}
223
224/**
225 * amdgpu_vm_bo_moved - vm_bo is moved
226 *
227 * @vm_bo: vm_bo which is moved
228 *
229 * State for per VM BOs which are moved, but that change is not yet reflected
230 * in the page tables.
231 */
232static void amdgpu_vm_bo_moved(struct amdgpu_vm_bo_base *vm_bo)
233{
234 list_move(&vm_bo->vm_status, &vm_bo->vm->moved);
235}
236
237/**
238 * amdgpu_vm_bo_idle - vm_bo is idle
239 *
240 * @vm_bo: vm_bo which is now idle
241 *
242 * State for PDs/PTs and per VM BOs which have gone through the state machine
243 * and are now idle.
244 */
245static void amdgpu_vm_bo_idle(struct amdgpu_vm_bo_base *vm_bo)
246{
247 list_move(&vm_bo->vm_status, &vm_bo->vm->idle);
248 vm_bo->moved = false;
249}
250
251/**
252 * amdgpu_vm_bo_invalidated - vm_bo is invalidated
253 *
254 * @vm_bo: vm_bo which is now invalidated
255 *
256 * State for normal BOs which are invalidated and that change not yet reflected
257 * in the PTs.
258 */
259static void amdgpu_vm_bo_invalidated(struct amdgpu_vm_bo_base *vm_bo)
260{
261 spin_lock(&vm_bo->vm->invalidated_lock);
262 list_move(&vm_bo->vm_status, &vm_bo->vm->invalidated);
263 spin_unlock(&vm_bo->vm->invalidated_lock);
264}
265
266/**
267 * amdgpu_vm_bo_done - vm_bo is done
268 *
269 * @vm_bo: vm_bo which is now done
270 *
271 * State for normal BOs which are invalidated and that change has been updated
272 * in the PTs.
273 */
274static void amdgpu_vm_bo_done(struct amdgpu_vm_bo_base *vm_bo)
275{
276 spin_lock(&vm_bo->vm->invalidated_lock);
277 list_del_init(&vm_bo->vm_status);
278 spin_unlock(&vm_bo->vm->invalidated_lock);
279}
280
281/**
282 * amdgpu_vm_bo_base_init - Adds bo to the list of bos associated with the vm
283 *
284 * @base: base structure for tracking BO usage in a VM
285 * @vm: vm to which bo is to be added
286 * @bo: amdgpu buffer object
287 *
288 * Initialize a bo_va_base structure and add it to the appropriate lists
289 *
290 */
291static void amdgpu_vm_bo_base_init(struct amdgpu_vm_bo_base *base,
292 struct amdgpu_vm *vm,
293 struct amdgpu_bo *bo)
294{
295 base->vm = vm;
296 base->bo = bo;
297 base->next = NULL;
298 INIT_LIST_HEAD(&base->vm_status);
299
300 if (!bo)
301 return;
302 base->next = bo->vm_bo;
303 bo->vm_bo = base;
304
305 if (bo->tbo.base.resv != vm->root.base.bo->tbo.base.resv)
306 return;
307
308 vm->bulk_moveable = false;
309 if (bo->tbo.type == ttm_bo_type_kernel && bo->parent)
310 amdgpu_vm_bo_relocated(base);
311 else
312 amdgpu_vm_bo_idle(base);
313
314 if (bo->preferred_domains &
315 amdgpu_mem_type_to_domain(bo->tbo.mem.mem_type))
316 return;
317
318 /*
319 * we checked all the prerequisites, but it looks like this per vm bo
320 * is currently evicted. add the bo to the evicted list to make sure it
321 * is validated on next vm use to avoid fault.
322 * */
323 amdgpu_vm_bo_evicted(base);
324}
325
326/**
327 * amdgpu_vm_pt_parent - get the parent page directory
328 *
329 * @pt: child page table
330 *
331 * Helper to get the parent entry for the child page table. NULL if we are at
332 * the root page directory.
333 */
334static struct amdgpu_vm_pt *amdgpu_vm_pt_parent(struct amdgpu_vm_pt *pt)
335{
336 struct amdgpu_bo *parent = pt->base.bo->parent;
337
338 if (!parent)
339 return NULL;
340
341 return container_of(parent->vm_bo, struct amdgpu_vm_pt, base);
342}
343
344/**
345 * amdgpu_vm_pt_cursor - state for for_each_amdgpu_vm_pt
346 */
347struct amdgpu_vm_pt_cursor {
348 uint64_t pfn;
349 struct amdgpu_vm_pt *parent;
350 struct amdgpu_vm_pt *entry;
351 unsigned level;
352};
353
354/**
355 * amdgpu_vm_pt_start - start PD/PT walk
356 *
357 * @adev: amdgpu_device pointer
358 * @vm: amdgpu_vm structure
359 * @start: start address of the walk
360 * @cursor: state to initialize
361 *
362 * Initialize a amdgpu_vm_pt_cursor to start a walk.
363 */
364static void amdgpu_vm_pt_start(struct amdgpu_device *adev,
365 struct amdgpu_vm *vm, uint64_t start,
366 struct amdgpu_vm_pt_cursor *cursor)
367{
368 cursor->pfn = start;
369 cursor->parent = NULL;
370 cursor->entry = &vm->root;
371 cursor->level = adev->vm_manager.root_level;
372}
373
374/**
375 * amdgpu_vm_pt_descendant - go to child node
376 *
377 * @adev: amdgpu_device pointer
378 * @cursor: current state
379 *
380 * Walk to the child node of the current node.
381 * Returns:
382 * True if the walk was possible, false otherwise.
383 */
384static bool amdgpu_vm_pt_descendant(struct amdgpu_device *adev,
385 struct amdgpu_vm_pt_cursor *cursor)
386{
387 unsigned mask, shift, idx;
388
389 if (!cursor->entry->entries)
390 return false;
391
392 BUG_ON(!cursor->entry->base.bo);
393 mask = amdgpu_vm_entries_mask(adev, cursor->level);
394 shift = amdgpu_vm_level_shift(adev, cursor->level);
395
396 ++cursor->level;
397 idx = (cursor->pfn >> shift) & mask;
398 cursor->parent = cursor->entry;
399 cursor->entry = &cursor->entry->entries[idx];
400 return true;
401}
402
403/**
404 * amdgpu_vm_pt_sibling - go to sibling node
405 *
406 * @adev: amdgpu_device pointer
407 * @cursor: current state
408 *
409 * Walk to the sibling node of the current node.
410 * Returns:
411 * True if the walk was possible, false otherwise.
412 */
413static bool amdgpu_vm_pt_sibling(struct amdgpu_device *adev,
414 struct amdgpu_vm_pt_cursor *cursor)
415{
416 unsigned shift, num_entries;
417
418 /* Root doesn't have a sibling */
419 if (!cursor->parent)
420 return false;
421
422 /* Go to our parents and see if we got a sibling */
423 shift = amdgpu_vm_level_shift(adev, cursor->level - 1);
424 num_entries = amdgpu_vm_num_entries(adev, cursor->level - 1);
425
426 if (cursor->entry == &cursor->parent->entries[num_entries - 1])
427 return false;
428
429 cursor->pfn += 1ULL << shift;
430 cursor->pfn &= ~((1ULL << shift) - 1);
431 ++cursor->entry;
432 return true;
433}
434
435/**
436 * amdgpu_vm_pt_ancestor - go to parent node
437 *
438 * @cursor: current state
439 *
440 * Walk to the parent node of the current node.
441 * Returns:
442 * True if the walk was possible, false otherwise.
443 */
444static bool amdgpu_vm_pt_ancestor(struct amdgpu_vm_pt_cursor *cursor)
445{
446 if (!cursor->parent)
447 return false;
448
449 --cursor->level;
450 cursor->entry = cursor->parent;
451 cursor->parent = amdgpu_vm_pt_parent(cursor->parent);
452 return true;
453}
454
455/**
456 * amdgpu_vm_pt_next - get next PD/PT in hieratchy
457 *
458 * @adev: amdgpu_device pointer
459 * @cursor: current state
460 *
461 * Walk the PD/PT tree to the next node.
462 */
463static void amdgpu_vm_pt_next(struct amdgpu_device *adev,
464 struct amdgpu_vm_pt_cursor *cursor)
465{
466 /* First try a newborn child */
467 if (amdgpu_vm_pt_descendant(adev, cursor))
468 return;
469
470 /* If that didn't worked try to find a sibling */
471 while (!amdgpu_vm_pt_sibling(adev, cursor)) {
472 /* No sibling, go to our parents and grandparents */
473 if (!amdgpu_vm_pt_ancestor(cursor)) {
474 cursor->pfn = ~0ll;
475 return;
476 }
477 }
478}
479
480/**
481 * amdgpu_vm_pt_first_dfs - start a deep first search
482 *
483 * @adev: amdgpu_device structure
484 * @vm: amdgpu_vm structure
485 * @cursor: state to initialize
486 *
487 * Starts a deep first traversal of the PD/PT tree.
488 */
489static void amdgpu_vm_pt_first_dfs(struct amdgpu_device *adev,
490 struct amdgpu_vm *vm,
491 struct amdgpu_vm_pt_cursor *start,
492 struct amdgpu_vm_pt_cursor *cursor)
493{
494 if (start)
495 *cursor = *start;
496 else
497 amdgpu_vm_pt_start(adev, vm, 0, cursor);
498 while (amdgpu_vm_pt_descendant(adev, cursor));
499}
500
501/**
502 * amdgpu_vm_pt_continue_dfs - check if the deep first search should continue
503 *
504 * @start: starting point for the search
505 * @entry: current entry
506 *
507 * Returns:
508 * True when the search should continue, false otherwise.
509 */
510static bool amdgpu_vm_pt_continue_dfs(struct amdgpu_vm_pt_cursor *start,
511 struct amdgpu_vm_pt *entry)
512{
513 return entry && (!start || entry != start->entry);
514}
515
516/**
517 * amdgpu_vm_pt_next_dfs - get the next node for a deep first search
518 *
519 * @adev: amdgpu_device structure
520 * @cursor: current state
521 *
522 * Move the cursor to the next node in a deep first search.
523 */
524static void amdgpu_vm_pt_next_dfs(struct amdgpu_device *adev,
525 struct amdgpu_vm_pt_cursor *cursor)
526{
527 if (!cursor->entry)
528 return;
529
530 if (!cursor->parent)
531 cursor->entry = NULL;
532 else if (amdgpu_vm_pt_sibling(adev, cursor))
533 while (amdgpu_vm_pt_descendant(adev, cursor));
534 else
535 amdgpu_vm_pt_ancestor(cursor);
536}
537
538/**
539 * for_each_amdgpu_vm_pt_dfs_safe - safe deep first search of all PDs/PTs
540 */
541#define for_each_amdgpu_vm_pt_dfs_safe(adev, vm, start, cursor, entry) \
542 for (amdgpu_vm_pt_first_dfs((adev), (vm), (start), &(cursor)), \
543 (entry) = (cursor).entry, amdgpu_vm_pt_next_dfs((adev), &(cursor));\
544 amdgpu_vm_pt_continue_dfs((start), (entry)); \
545 (entry) = (cursor).entry, amdgpu_vm_pt_next_dfs((adev), &(cursor)))
546
547/**
548 * amdgpu_vm_get_pd_bo - add the VM PD to a validation list
549 *
550 * @vm: vm providing the BOs
551 * @validated: head of validation list
552 * @entry: entry to add
553 *
554 * Add the page directory to the list of BOs to
555 * validate for command submission.
556 */
557void amdgpu_vm_get_pd_bo(struct amdgpu_vm *vm,
558 struct list_head *validated,
559 struct amdgpu_bo_list_entry *entry)
560{
561 entry->priority = 0;
562 entry->tv.bo = &vm->root.base.bo->tbo;
563 /* One for the VM updates, one for TTM and one for the CS job */
564 entry->tv.num_shared = 3;
565 entry->user_pages = NULL;
566 list_add(&entry->tv.head, validated);
567}
568
569void amdgpu_vm_del_from_lru_notify(struct ttm_buffer_object *bo)
570{
571 struct amdgpu_bo *abo;
572 struct amdgpu_vm_bo_base *bo_base;
573
574 if (!amdgpu_bo_is_amdgpu_bo(bo))
575 return;
576
577 if (bo->mem.placement & TTM_PL_FLAG_NO_EVICT)
578 return;
579
580 abo = ttm_to_amdgpu_bo(bo);
581 if (!abo->parent)
582 return;
583 for (bo_base = abo->vm_bo; bo_base; bo_base = bo_base->next) {
584 struct amdgpu_vm *vm = bo_base->vm;
585
586 if (abo->tbo.base.resv == vm->root.base.bo->tbo.base.resv)
587 vm->bulk_moveable = false;
588 }
589
590}
591/**
592 * amdgpu_vm_move_to_lru_tail - move all BOs to the end of LRU
593 *
594 * @adev: amdgpu device pointer
595 * @vm: vm providing the BOs
596 *
597 * Move all BOs to the end of LRU and remember their positions to put them
598 * together.
599 */
600void amdgpu_vm_move_to_lru_tail(struct amdgpu_device *adev,
601 struct amdgpu_vm *vm)
602{
603 struct ttm_bo_global *glob = adev->mman.bdev.glob;
604 struct amdgpu_vm_bo_base *bo_base;
605
606 if (vm->bulk_moveable) {
607 spin_lock(&glob->lru_lock);
608 ttm_bo_bulk_move_lru_tail(&vm->lru_bulk_move);
609 spin_unlock(&glob->lru_lock);
610 return;
611 }
612
613 memset(&vm->lru_bulk_move, 0, sizeof(vm->lru_bulk_move));
614
615 spin_lock(&glob->lru_lock);
616 list_for_each_entry(bo_base, &vm->idle, vm_status) {
617 struct amdgpu_bo *bo = bo_base->bo;
618
619 if (!bo->parent)
620 continue;
621
622 ttm_bo_move_to_lru_tail(&bo->tbo, &vm->lru_bulk_move);
623 if (bo->shadow)
624 ttm_bo_move_to_lru_tail(&bo->shadow->tbo,
625 &vm->lru_bulk_move);
626 }
627 spin_unlock(&glob->lru_lock);
628
629 vm->bulk_moveable = true;
630}
631
632/**
633 * amdgpu_vm_validate_pt_bos - validate the page table BOs
634 *
635 * @adev: amdgpu device pointer
636 * @vm: vm providing the BOs
637 * @validate: callback to do the validation
638 * @param: parameter for the validation callback
639 *
640 * Validate the page table BOs on command submission if neccessary.
641 *
642 * Returns:
643 * Validation result.
644 */
645int amdgpu_vm_validate_pt_bos(struct amdgpu_device *adev, struct amdgpu_vm *vm,
646 int (*validate)(void *p, struct amdgpu_bo *bo),
647 void *param)
648{
649 struct amdgpu_vm_bo_base *bo_base, *tmp;
650 int r = 0;
651
652 vm->bulk_moveable &= list_empty(&vm->evicted);
653
654 list_for_each_entry_safe(bo_base, tmp, &vm->evicted, vm_status) {
655 struct amdgpu_bo *bo = bo_base->bo;
656
657 r = validate(param, bo);
658 if (r)
659 break;
660
661 if (bo->tbo.type != ttm_bo_type_kernel) {
662 amdgpu_vm_bo_moved(bo_base);
663 } else {
664 vm->update_funcs->map_table(bo);
665 if (bo->parent)
666 amdgpu_vm_bo_relocated(bo_base);
667 else
668 amdgpu_vm_bo_idle(bo_base);
669 }
670 }
671
672 return r;
673}
674
675/**
676 * amdgpu_vm_ready - check VM is ready for updates
677 *
678 * @vm: VM to check
679 *
680 * Check if all VM PDs/PTs are ready for updates
681 *
682 * Returns:
683 * True if eviction list is empty.
684 */
685bool amdgpu_vm_ready(struct amdgpu_vm *vm)
686{
687 return list_empty(&vm->evicted);
688}
689
690/**
691 * amdgpu_vm_clear_bo - initially clear the PDs/PTs
692 *
693 * @adev: amdgpu_device pointer
694 * @vm: VM to clear BO from
695 * @bo: BO to clear
696 *
697 * Root PD needs to be reserved when calling this.
698 *
699 * Returns:
700 * 0 on success, errno otherwise.
701 */
702static int amdgpu_vm_clear_bo(struct amdgpu_device *adev,
703 struct amdgpu_vm *vm,
704 struct amdgpu_bo *bo)
705{
706 struct ttm_operation_ctx ctx = { true, false };
707 unsigned level = adev->vm_manager.root_level;
708 struct amdgpu_vm_update_params params;
709 struct amdgpu_bo *ancestor = bo;
710 unsigned entries, ats_entries;
711 uint64_t addr;
712 int r;
713
714 /* Figure out our place in the hierarchy */
715 if (ancestor->parent) {
716 ++level;
717 while (ancestor->parent->parent) {
718 ++level;
719 ancestor = ancestor->parent;
720 }
721 }
722
723 entries = amdgpu_bo_size(bo) / 8;
724 if (!vm->pte_support_ats) {
725 ats_entries = 0;
726
727 } else if (!bo->parent) {
728 ats_entries = amdgpu_vm_num_ats_entries(adev);
729 ats_entries = min(ats_entries, entries);
730 entries -= ats_entries;
731
732 } else {
733 struct amdgpu_vm_pt *pt;
734
735 pt = container_of(ancestor->vm_bo, struct amdgpu_vm_pt, base);
736 ats_entries = amdgpu_vm_num_ats_entries(adev);
737 if ((pt - vm->root.entries) >= ats_entries) {
738 ats_entries = 0;
739 } else {
740 ats_entries = entries;
741 entries = 0;
742 }
743 }
744
745 r = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
746 if (r)
747 return r;
748
749 if (bo->shadow) {
750 r = ttm_bo_validate(&bo->shadow->tbo, &bo->shadow->placement,
751 &ctx);
752 if (r)
753 return r;
754 }
755
756 r = vm->update_funcs->map_table(bo);
757 if (r)
758 return r;
759
760 memset(¶ms, 0, sizeof(params));
761 params.adev = adev;
762 params.vm = vm;
763
764 r = vm->update_funcs->prepare(¶ms, AMDGPU_FENCE_OWNER_KFD, NULL);
765 if (r)
766 return r;
767
768 addr = 0;
769 if (ats_entries) {
770 uint64_t value = 0, flags;
771
772 flags = AMDGPU_PTE_DEFAULT_ATC;
773 if (level != AMDGPU_VM_PTB) {
774 /* Handle leaf PDEs as PTEs */
775 flags |= AMDGPU_PDE_PTE;
776 amdgpu_gmc_get_vm_pde(adev, level, &value, &flags);
777 }
778
779 r = vm->update_funcs->update(¶ms, bo, addr, 0, ats_entries,
780 value, flags);
781 if (r)
782 return r;
783
784 addr += ats_entries * 8;
785 }
786
787 if (entries) {
788 uint64_t value = 0, flags = 0;
789
790 if (adev->asic_type >= CHIP_VEGA10) {
791 if (level != AMDGPU_VM_PTB) {
792 /* Handle leaf PDEs as PTEs */
793 flags |= AMDGPU_PDE_PTE;
794 amdgpu_gmc_get_vm_pde(adev, level,
795 &value, &flags);
796 } else {
797 /* Workaround for fault priority problem on GMC9 */
798 flags = AMDGPU_PTE_EXECUTABLE;
799 }
800 }
801
802 r = vm->update_funcs->update(¶ms, bo, addr, 0, entries,
803 value, flags);
804 if (r)
805 return r;
806 }
807
808 return vm->update_funcs->commit(¶ms, NULL);
809}
810
811/**
812 * amdgpu_vm_bo_param - fill in parameters for PD/PT allocation
813 *
814 * @adev: amdgpu_device pointer
815 * @vm: requesting vm
816 * @bp: resulting BO allocation parameters
817 */
818static void amdgpu_vm_bo_param(struct amdgpu_device *adev, struct amdgpu_vm *vm,
819 int level, struct amdgpu_bo_param *bp)
820{
821 memset(bp, 0, sizeof(*bp));
822
823 bp->size = amdgpu_vm_bo_size(adev, level);
824 bp->byte_align = AMDGPU_GPU_PAGE_SIZE;
825 bp->domain = AMDGPU_GEM_DOMAIN_VRAM;
826 bp->domain = amdgpu_bo_get_preferred_pin_domain(adev, bp->domain);
827 bp->flags = AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS |
828 AMDGPU_GEM_CREATE_CPU_GTT_USWC;
829 if (vm->use_cpu_for_update)
830 bp->flags |= AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED;
831 else if (!vm->root.base.bo || vm->root.base.bo->shadow)
832 bp->flags |= AMDGPU_GEM_CREATE_SHADOW;
833 bp->type = ttm_bo_type_kernel;
834 if (vm->root.base.bo)
835 bp->resv = vm->root.base.bo->tbo.base.resv;
836}
837
838/**
839 * amdgpu_vm_alloc_pts - Allocate a specific page table
840 *
841 * @adev: amdgpu_device pointer
842 * @vm: VM to allocate page tables for
843 * @cursor: Which page table to allocate
844 *
845 * Make sure a specific page table or directory is allocated.
846 *
847 * Returns:
848 * 1 if page table needed to be allocated, 0 if page table was already
849 * allocated, negative errno if an error occurred.
850 */
851static int amdgpu_vm_alloc_pts(struct amdgpu_device *adev,
852 struct amdgpu_vm *vm,
853 struct amdgpu_vm_pt_cursor *cursor)
854{
855 struct amdgpu_vm_pt *entry = cursor->entry;
856 struct amdgpu_bo_param bp;
857 struct amdgpu_bo *pt;
858 int r;
859
860 if (cursor->level < AMDGPU_VM_PTB && !entry->entries) {
861 unsigned num_entries;
862
863 num_entries = amdgpu_vm_num_entries(adev, cursor->level);
864 entry->entries = kvmalloc_array(num_entries,
865 sizeof(*entry->entries),
866 GFP_KERNEL | __GFP_ZERO);
867 if (!entry->entries)
868 return -ENOMEM;
869 }
870
871 if (entry->base.bo)
872 return 0;
873
874 amdgpu_vm_bo_param(adev, vm, cursor->level, &bp);
875
876 r = amdgpu_bo_create(adev, &bp, &pt);
877 if (r)
878 return r;
879
880 /* Keep a reference to the root directory to avoid
881 * freeing them up in the wrong order.
882 */
883 pt->parent = amdgpu_bo_ref(cursor->parent->base.bo);
884 amdgpu_vm_bo_base_init(&entry->base, vm, pt);
885
886 r = amdgpu_vm_clear_bo(adev, vm, pt);
887 if (r)
888 goto error_free_pt;
889
890 return 0;
891
892error_free_pt:
893 amdgpu_bo_unref(&pt->shadow);
894 amdgpu_bo_unref(&pt);
895 return r;
896}
897
898/**
899 * amdgpu_vm_free_table - fre one PD/PT
900 *
901 * @entry: PDE to free
902 */
903static void amdgpu_vm_free_table(struct amdgpu_vm_pt *entry)
904{
905 if (entry->base.bo) {
906 entry->base.bo->vm_bo = NULL;
907 list_del(&entry->base.vm_status);
908 amdgpu_bo_unref(&entry->base.bo->shadow);
909 amdgpu_bo_unref(&entry->base.bo);
910 }
911 kvfree(entry->entries);
912 entry->entries = NULL;
913}
914
915/**
916 * amdgpu_vm_free_pts - free PD/PT levels
917 *
918 * @adev: amdgpu device structure
919 * @vm: amdgpu vm structure
920 * @start: optional cursor where to start freeing PDs/PTs
921 *
922 * Free the page directory or page table level and all sub levels.
923 */
924static void amdgpu_vm_free_pts(struct amdgpu_device *adev,
925 struct amdgpu_vm *vm,
926 struct amdgpu_vm_pt_cursor *start)
927{
928 struct amdgpu_vm_pt_cursor cursor;
929 struct amdgpu_vm_pt *entry;
930
931 vm->bulk_moveable = false;
932
933 for_each_amdgpu_vm_pt_dfs_safe(adev, vm, start, cursor, entry)
934 amdgpu_vm_free_table(entry);
935
936 if (start)
937 amdgpu_vm_free_table(start->entry);
938}
939
940/**
941 * amdgpu_vm_check_compute_bug - check whether asic has compute vm bug
942 *
943 * @adev: amdgpu_device pointer
944 */
945void amdgpu_vm_check_compute_bug(struct amdgpu_device *adev)
946{
947 const struct amdgpu_ip_block *ip_block;
948 bool has_compute_vm_bug;
949 struct amdgpu_ring *ring;
950 int i;
951
952 has_compute_vm_bug = false;
953
954 ip_block = amdgpu_device_ip_get_ip_block(adev, AMD_IP_BLOCK_TYPE_GFX);
955 if (ip_block) {
956 /* Compute has a VM bug for GFX version < 7.
957 Compute has a VM bug for GFX 8 MEC firmware version < 673.*/
958 if (ip_block->version->major <= 7)
959 has_compute_vm_bug = true;
960 else if (ip_block->version->major == 8)
961 if (adev->gfx.mec_fw_version < 673)
962 has_compute_vm_bug = true;
963 }
964
965 for (i = 0; i < adev->num_rings; i++) {
966 ring = adev->rings[i];
967 if (ring->funcs->type == AMDGPU_RING_TYPE_COMPUTE)
968 /* only compute rings */
969 ring->has_compute_vm_bug = has_compute_vm_bug;
970 else
971 ring->has_compute_vm_bug = false;
972 }
973}
974
975/**
976 * amdgpu_vm_need_pipeline_sync - Check if pipe sync is needed for job.
977 *
978 * @ring: ring on which the job will be submitted
979 * @job: job to submit
980 *
981 * Returns:
982 * True if sync is needed.
983 */
984bool amdgpu_vm_need_pipeline_sync(struct amdgpu_ring *ring,
985 struct amdgpu_job *job)
986{
987 struct amdgpu_device *adev = ring->adev;
988 unsigned vmhub = ring->funcs->vmhub;
989 struct amdgpu_vmid_mgr *id_mgr = &adev->vm_manager.id_mgr[vmhub];
990 struct amdgpu_vmid *id;
991 bool gds_switch_needed;
992 bool vm_flush_needed = job->vm_needs_flush || ring->has_compute_vm_bug;
993
994 if (job->vmid == 0)
995 return false;
996 id = &id_mgr->ids[job->vmid];
997 gds_switch_needed = ring->funcs->emit_gds_switch && (
998 id->gds_base != job->gds_base ||
999 id->gds_size != job->gds_size ||
1000 id->gws_base != job->gws_base ||
1001 id->gws_size != job->gws_size ||
1002 id->oa_base != job->oa_base ||
1003 id->oa_size != job->oa_size);
1004
1005 if (amdgpu_vmid_had_gpu_reset(adev, id))
1006 return true;
1007
1008 return vm_flush_needed || gds_switch_needed;
1009}
1010
1011/**
1012 * amdgpu_vm_flush - hardware flush the vm
1013 *
1014 * @ring: ring to use for flush
1015 * @job: related job
1016 * @need_pipe_sync: is pipe sync needed
1017 *
1018 * Emit a VM flush when it is necessary.
1019 *
1020 * Returns:
1021 * 0 on success, errno otherwise.
1022 */
1023int amdgpu_vm_flush(struct amdgpu_ring *ring, struct amdgpu_job *job, bool need_pipe_sync)
1024{
1025 struct amdgpu_device *adev = ring->adev;
1026 unsigned vmhub = ring->funcs->vmhub;
1027 struct amdgpu_vmid_mgr *id_mgr = &adev->vm_manager.id_mgr[vmhub];
1028 struct amdgpu_vmid *id = &id_mgr->ids[job->vmid];
1029 bool gds_switch_needed = ring->funcs->emit_gds_switch && (
1030 id->gds_base != job->gds_base ||
1031 id->gds_size != job->gds_size ||
1032 id->gws_base != job->gws_base ||
1033 id->gws_size != job->gws_size ||
1034 id->oa_base != job->oa_base ||
1035 id->oa_size != job->oa_size);
1036 bool vm_flush_needed = job->vm_needs_flush;
1037 bool pasid_mapping_needed = id->pasid != job->pasid ||
1038 !id->pasid_mapping ||
1039 !dma_fence_is_signaled(id->pasid_mapping);
1040 struct dma_fence *fence = NULL;
1041 unsigned patch_offset = 0;
1042 int r;
1043
1044 if (amdgpu_vmid_had_gpu_reset(adev, id)) {
1045 gds_switch_needed = true;
1046 vm_flush_needed = true;
1047 pasid_mapping_needed = true;
1048 }
1049
1050 gds_switch_needed &= !!ring->funcs->emit_gds_switch;
1051 vm_flush_needed &= !!ring->funcs->emit_vm_flush &&
1052 job->vm_pd_addr != AMDGPU_BO_INVALID_OFFSET;
1053 pasid_mapping_needed &= adev->gmc.gmc_funcs->emit_pasid_mapping &&
1054 ring->funcs->emit_wreg;
1055
1056 if (!vm_flush_needed && !gds_switch_needed && !need_pipe_sync)
1057 return 0;
1058
1059 if (ring->funcs->init_cond_exec)
1060 patch_offset = amdgpu_ring_init_cond_exec(ring);
1061
1062 if (need_pipe_sync)
1063 amdgpu_ring_emit_pipeline_sync(ring);
1064
1065 if (vm_flush_needed) {
1066 trace_amdgpu_vm_flush(ring, job->vmid, job->vm_pd_addr);
1067 amdgpu_ring_emit_vm_flush(ring, job->vmid, job->vm_pd_addr);
1068 }
1069
1070 if (pasid_mapping_needed)
1071 amdgpu_gmc_emit_pasid_mapping(ring, job->vmid, job->pasid);
1072
1073 if (vm_flush_needed || pasid_mapping_needed) {
1074 r = amdgpu_fence_emit(ring, &fence, 0);
1075 if (r)
1076 return r;
1077 }
1078
1079 if (vm_flush_needed) {
1080 mutex_lock(&id_mgr->lock);
1081 dma_fence_put(id->last_flush);
1082 id->last_flush = dma_fence_get(fence);
1083 id->current_gpu_reset_count =
1084 atomic_read(&adev->gpu_reset_counter);
1085 mutex_unlock(&id_mgr->lock);
1086 }
1087
1088 if (pasid_mapping_needed) {
1089 id->pasid = job->pasid;
1090 dma_fence_put(id->pasid_mapping);
1091 id->pasid_mapping = dma_fence_get(fence);
1092 }
1093 dma_fence_put(fence);
1094
1095 if (ring->funcs->emit_gds_switch && gds_switch_needed) {
1096 id->gds_base = job->gds_base;
1097 id->gds_size = job->gds_size;
1098 id->gws_base = job->gws_base;
1099 id->gws_size = job->gws_size;
1100 id->oa_base = job->oa_base;
1101 id->oa_size = job->oa_size;
1102 amdgpu_ring_emit_gds_switch(ring, job->vmid, job->gds_base,
1103 job->gds_size, job->gws_base,
1104 job->gws_size, job->oa_base,
1105 job->oa_size);
1106 }
1107
1108 if (ring->funcs->patch_cond_exec)
1109 amdgpu_ring_patch_cond_exec(ring, patch_offset);
1110
1111 /* the double SWITCH_BUFFER here *cannot* be skipped by COND_EXEC */
1112 if (ring->funcs->emit_switch_buffer) {
1113 amdgpu_ring_emit_switch_buffer(ring);
1114 amdgpu_ring_emit_switch_buffer(ring);
1115 }
1116 return 0;
1117}
1118
1119/**
1120 * amdgpu_vm_bo_find - find the bo_va for a specific vm & bo
1121 *
1122 * @vm: requested vm
1123 * @bo: requested buffer object
1124 *
1125 * Find @bo inside the requested vm.
1126 * Search inside the @bos vm list for the requested vm
1127 * Returns the found bo_va or NULL if none is found
1128 *
1129 * Object has to be reserved!
1130 *
1131 * Returns:
1132 * Found bo_va or NULL.
1133 */
1134struct amdgpu_bo_va *amdgpu_vm_bo_find(struct amdgpu_vm *vm,
1135 struct amdgpu_bo *bo)
1136{
1137 struct amdgpu_vm_bo_base *base;
1138
1139 for (base = bo->vm_bo; base; base = base->next) {
1140 if (base->vm != vm)
1141 continue;
1142
1143 return container_of(base, struct amdgpu_bo_va, base);
1144 }
1145 return NULL;
1146}
1147
1148/**
1149 * amdgpu_vm_map_gart - Resolve gart mapping of addr
1150 *
1151 * @pages_addr: optional DMA address to use for lookup
1152 * @addr: the unmapped addr
1153 *
1154 * Look up the physical address of the page that the pte resolves
1155 * to.
1156 *
1157 * Returns:
1158 * The pointer for the page table entry.
1159 */
1160uint64_t amdgpu_vm_map_gart(const dma_addr_t *pages_addr, uint64_t addr)
1161{
1162 uint64_t result;
1163
1164 /* page table offset */
1165 result = pages_addr[addr >> PAGE_SHIFT];
1166
1167 /* in case cpu page size != gpu page size*/
1168 result |= addr & (~PAGE_MASK);
1169
1170 result &= 0xFFFFFFFFFFFFF000ULL;
1171
1172 return result;
1173}
1174
1175/*
1176 * amdgpu_vm_update_pde - update a single level in the hierarchy
1177 *
1178 * @param: parameters for the update
1179 * @vm: requested vm
1180 * @entry: entry to update
1181 *
1182 * Makes sure the requested entry in parent is up to date.
1183 */
1184static int amdgpu_vm_update_pde(struct amdgpu_vm_update_params *params,
1185 struct amdgpu_vm *vm,
1186 struct amdgpu_vm_pt *entry)
1187{
1188 struct amdgpu_vm_pt *parent = amdgpu_vm_pt_parent(entry);
1189 struct amdgpu_bo *bo = parent->base.bo, *pbo;
1190 uint64_t pde, pt, flags;
1191 unsigned level;
1192
1193 for (level = 0, pbo = bo->parent; pbo; ++level)
1194 pbo = pbo->parent;
1195
1196 level += params->adev->vm_manager.root_level;
1197 amdgpu_gmc_get_pde_for_bo(entry->base.bo, level, &pt, &flags);
1198 pde = (entry - parent->entries) * 8;
1199 return vm->update_funcs->update(params, bo, pde, pt, 1, 0, flags);
1200}
1201
1202/*
1203 * amdgpu_vm_invalidate_pds - mark all PDs as invalid
1204 *
1205 * @adev: amdgpu_device pointer
1206 * @vm: related vm
1207 *
1208 * Mark all PD level as invalid after an error.
1209 */
1210static void amdgpu_vm_invalidate_pds(struct amdgpu_device *adev,
1211 struct amdgpu_vm *vm)
1212{
1213 struct amdgpu_vm_pt_cursor cursor;
1214 struct amdgpu_vm_pt *entry;
1215
1216 for_each_amdgpu_vm_pt_dfs_safe(adev, vm, NULL, cursor, entry)
1217 if (entry->base.bo && !entry->base.moved)
1218 amdgpu_vm_bo_relocated(&entry->base);
1219}
1220
1221/*
1222 * amdgpu_vm_update_directories - make sure that all directories are valid
1223 *
1224 * @adev: amdgpu_device pointer
1225 * @vm: requested vm
1226 *
1227 * Makes sure all directories are up to date.
1228 *
1229 * Returns:
1230 * 0 for success, error for failure.
1231 */
1232int amdgpu_vm_update_directories(struct amdgpu_device *adev,
1233 struct amdgpu_vm *vm)
1234{
1235 struct amdgpu_vm_update_params params;
1236 int r;
1237
1238 if (list_empty(&vm->relocated))
1239 return 0;
1240
1241 memset(¶ms, 0, sizeof(params));
1242 params.adev = adev;
1243 params.vm = vm;
1244
1245 r = vm->update_funcs->prepare(¶ms, AMDGPU_FENCE_OWNER_VM, NULL);
1246 if (r)
1247 return r;
1248
1249 while (!list_empty(&vm->relocated)) {
1250 struct amdgpu_vm_pt *entry;
1251
1252 entry = list_first_entry(&vm->relocated, struct amdgpu_vm_pt,
1253 base.vm_status);
1254 amdgpu_vm_bo_idle(&entry->base);
1255
1256 r = amdgpu_vm_update_pde(¶ms, vm, entry);
1257 if (r)
1258 goto error;
1259 }
1260
1261 r = vm->update_funcs->commit(¶ms, &vm->last_update);
1262 if (r)
1263 goto error;
1264 return 0;
1265
1266error:
1267 amdgpu_vm_invalidate_pds(adev, vm);
1268 return r;
1269}
1270
1271/**
1272 * amdgpu_vm_update_flags - figure out flags for PTE updates
1273 *
1274 * Make sure to set the right flags for the PTEs at the desired level.
1275 */
1276static void amdgpu_vm_update_flags(struct amdgpu_vm_update_params *params,
1277 struct amdgpu_bo *bo, unsigned level,
1278 uint64_t pe, uint64_t addr,
1279 unsigned count, uint32_t incr,
1280 uint64_t flags)
1281
1282{
1283 if (level != AMDGPU_VM_PTB) {
1284 flags |= AMDGPU_PDE_PTE;
1285 amdgpu_gmc_get_vm_pde(params->adev, level, &addr, &flags);
1286
1287 } else if (params->adev->asic_type >= CHIP_VEGA10 &&
1288 !(flags & AMDGPU_PTE_VALID) &&
1289 !(flags & AMDGPU_PTE_PRT)) {
1290
1291 /* Workaround for fault priority problem on GMC9 */
1292 flags |= AMDGPU_PTE_EXECUTABLE;
1293 }
1294
1295 params->vm->update_funcs->update(params, bo, pe, addr, count, incr,
1296 flags);
1297}
1298
1299/**
1300 * amdgpu_vm_fragment - get fragment for PTEs
1301 *
1302 * @params: see amdgpu_vm_update_params definition
1303 * @start: first PTE to handle
1304 * @end: last PTE to handle
1305 * @flags: hw mapping flags
1306 * @frag: resulting fragment size
1307 * @frag_end: end of this fragment
1308 *
1309 * Returns the first possible fragment for the start and end address.
1310 */
1311static void amdgpu_vm_fragment(struct amdgpu_vm_update_params *params,
1312 uint64_t start, uint64_t end, uint64_t flags,
1313 unsigned int *frag, uint64_t *frag_end)
1314{
1315 /**
1316 * The MC L1 TLB supports variable sized pages, based on a fragment
1317 * field in the PTE. When this field is set to a non-zero value, page
1318 * granularity is increased from 4KB to (1 << (12 + frag)). The PTE
1319 * flags are considered valid for all PTEs within the fragment range
1320 * and corresponding mappings are assumed to be physically contiguous.
1321 *
1322 * The L1 TLB can store a single PTE for the whole fragment,
1323 * significantly increasing the space available for translation
1324 * caching. This leads to large improvements in throughput when the
1325 * TLB is under pressure.
1326 *
1327 * The L2 TLB distributes small and large fragments into two
1328 * asymmetric partitions. The large fragment cache is significantly
1329 * larger. Thus, we try to use large fragments wherever possible.
1330 * Userspace can support this by aligning virtual base address and
1331 * allocation size to the fragment size.
1332 *
1333 * Starting with Vega10 the fragment size only controls the L1. The L2
1334 * is now directly feed with small/huge/giant pages from the walker.
1335 */
1336 unsigned max_frag;
1337
1338 if (params->adev->asic_type < CHIP_VEGA10)
1339 max_frag = params->adev->vm_manager.fragment_size;
1340 else
1341 max_frag = 31;
1342
1343 /* system pages are non continuously */
1344 if (params->pages_addr) {
1345 *frag = 0;
1346 *frag_end = end;
1347 return;
1348 }
1349
1350 /* This intentionally wraps around if no bit is set */
1351 *frag = min((unsigned)ffs(start) - 1, (unsigned)fls64(end - start) - 1);
1352 if (*frag >= max_frag) {
1353 *frag = max_frag;
1354 *frag_end = end & ~((1ULL << max_frag) - 1);
1355 } else {
1356 *frag_end = start + (1 << *frag);
1357 }
1358}
1359
1360/**
1361 * amdgpu_vm_update_ptes - make sure that page tables are valid
1362 *
1363 * @params: see amdgpu_vm_update_params definition
1364 * @start: start of GPU address range
1365 * @end: end of GPU address range
1366 * @dst: destination address to map to, the next dst inside the function
1367 * @flags: mapping flags
1368 *
1369 * Update the page tables in the range @start - @end.
1370 *
1371 * Returns:
1372 * 0 for success, -EINVAL for failure.
1373 */
1374static int amdgpu_vm_update_ptes(struct amdgpu_vm_update_params *params,
1375 uint64_t start, uint64_t end,
1376 uint64_t dst, uint64_t flags)
1377{
1378 struct amdgpu_device *adev = params->adev;
1379 struct amdgpu_vm_pt_cursor cursor;
1380 uint64_t frag_start = start, frag_end;
1381 unsigned int frag;
1382 int r;
1383
1384 /* figure out the initial fragment */
1385 amdgpu_vm_fragment(params, frag_start, end, flags, &frag, &frag_end);
1386
1387 /* walk over the address space and update the PTs */
1388 amdgpu_vm_pt_start(adev, params->vm, start, &cursor);
1389 while (cursor.pfn < end) {
1390 unsigned shift, parent_shift, mask;
1391 uint64_t incr, entry_end, pe_start;
1392 struct amdgpu_bo *pt;
1393
1394 r = amdgpu_vm_alloc_pts(params->adev, params->vm, &cursor);
1395 if (r)
1396 return r;
1397
1398 pt = cursor.entry->base.bo;
1399
1400 /* The root level can't be a huge page */
1401 if (cursor.level == adev->vm_manager.root_level) {
1402 if (!amdgpu_vm_pt_descendant(adev, &cursor))
1403 return -ENOENT;
1404 continue;
1405 }
1406
1407 shift = amdgpu_vm_level_shift(adev, cursor.level);
1408 parent_shift = amdgpu_vm_level_shift(adev, cursor.level - 1);
1409 if (adev->asic_type < CHIP_VEGA10 &&
1410 (flags & AMDGPU_PTE_VALID)) {
1411 /* No huge page support before GMC v9 */
1412 if (cursor.level != AMDGPU_VM_PTB) {
1413 if (!amdgpu_vm_pt_descendant(adev, &cursor))
1414 return -ENOENT;
1415 continue;
1416 }
1417 } else if (frag < shift) {
1418 /* We can't use this level when the fragment size is
1419 * smaller than the address shift. Go to the next
1420 * child entry and try again.
1421 */
1422 if (!amdgpu_vm_pt_descendant(adev, &cursor))
1423 return -ENOENT;
1424 continue;
1425 } else if (frag >= parent_shift &&
1426 cursor.level - 1 != adev->vm_manager.root_level) {
1427 /* If the fragment size is even larger than the parent
1428 * shift we should go up one level and check it again
1429 * unless one level up is the root level.
1430 */
1431 if (!amdgpu_vm_pt_ancestor(&cursor))
1432 return -ENOENT;
1433 continue;
1434 }
1435
1436 /* Looks good so far, calculate parameters for the update */
1437 incr = (uint64_t)AMDGPU_GPU_PAGE_SIZE << shift;
1438 mask = amdgpu_vm_entries_mask(adev, cursor.level);
1439 pe_start = ((cursor.pfn >> shift) & mask) * 8;
1440 entry_end = (uint64_t)(mask + 1) << shift;
1441 entry_end += cursor.pfn & ~(entry_end - 1);
1442 entry_end = min(entry_end, end);
1443
1444 do {
1445 uint64_t upd_end = min(entry_end, frag_end);
1446 unsigned nptes = (upd_end - frag_start) >> shift;
1447
1448 amdgpu_vm_update_flags(params, pt, cursor.level,
1449 pe_start, dst, nptes, incr,
1450 flags | AMDGPU_PTE_FRAG(frag));
1451
1452 pe_start += nptes * 8;
1453 dst += (uint64_t)nptes * AMDGPU_GPU_PAGE_SIZE << shift;
1454
1455 frag_start = upd_end;
1456 if (frag_start >= frag_end) {
1457 /* figure out the next fragment */
1458 amdgpu_vm_fragment(params, frag_start, end,
1459 flags, &frag, &frag_end);
1460 if (frag < shift)
1461 break;
1462 }
1463 } while (frag_start < entry_end);
1464
1465 if (amdgpu_vm_pt_descendant(adev, &cursor)) {
1466 /* Free all child entries */
1467 while (cursor.pfn < frag_start) {
1468 amdgpu_vm_free_pts(adev, params->vm, &cursor);
1469 amdgpu_vm_pt_next(adev, &cursor);
1470 }
1471
1472 } else if (frag >= shift) {
1473 /* or just move on to the next on the same level. */
1474 amdgpu_vm_pt_next(adev, &cursor);
1475 }
1476 }
1477
1478 return 0;
1479}
1480
1481/**
1482 * amdgpu_vm_bo_update_mapping - update a mapping in the vm page table
1483 *
1484 * @adev: amdgpu_device pointer
1485 * @exclusive: fence we need to sync to
1486 * @pages_addr: DMA addresses to use for mapping
1487 * @vm: requested vm
1488 * @start: start of mapped range
1489 * @last: last mapped entry
1490 * @flags: flags for the entries
1491 * @addr: addr to set the area to
1492 * @fence: optional resulting fence
1493 *
1494 * Fill in the page table entries between @start and @last.
1495 *
1496 * Returns:
1497 * 0 for success, -EINVAL for failure.
1498 */
1499static int amdgpu_vm_bo_update_mapping(struct amdgpu_device *adev,
1500 struct dma_fence *exclusive,
1501 dma_addr_t *pages_addr,
1502 struct amdgpu_vm *vm,
1503 uint64_t start, uint64_t last,
1504 uint64_t flags, uint64_t addr,
1505 struct dma_fence **fence)
1506{
1507 struct amdgpu_vm_update_params params;
1508 void *owner = AMDGPU_FENCE_OWNER_VM;
1509 int r;
1510
1511 memset(¶ms, 0, sizeof(params));
1512 params.adev = adev;
1513 params.vm = vm;
1514 params.pages_addr = pages_addr;
1515
1516 /* sync to everything except eviction fences on unmapping */
1517 if (!(flags & AMDGPU_PTE_VALID))
1518 owner = AMDGPU_FENCE_OWNER_KFD;
1519
1520 r = vm->update_funcs->prepare(¶ms, owner, exclusive);
1521 if (r)
1522 return r;
1523
1524 r = amdgpu_vm_update_ptes(¶ms, start, last + 1, addr, flags);
1525 if (r)
1526 return r;
1527
1528 return vm->update_funcs->commit(¶ms, fence);
1529}
1530
1531/**
1532 * amdgpu_vm_bo_split_mapping - split a mapping into smaller chunks
1533 *
1534 * @adev: amdgpu_device pointer
1535 * @exclusive: fence we need to sync to
1536 * @pages_addr: DMA addresses to use for mapping
1537 * @vm: requested vm
1538 * @mapping: mapped range and flags to use for the update
1539 * @flags: HW flags for the mapping
1540 * @bo_adev: amdgpu_device pointer that bo actually been allocated
1541 * @nodes: array of drm_mm_nodes with the MC addresses
1542 * @fence: optional resulting fence
1543 *
1544 * Split the mapping into smaller chunks so that each update fits
1545 * into a SDMA IB.
1546 *
1547 * Returns:
1548 * 0 for success, -EINVAL for failure.
1549 */
1550static int amdgpu_vm_bo_split_mapping(struct amdgpu_device *adev,
1551 struct dma_fence *exclusive,
1552 dma_addr_t *pages_addr,
1553 struct amdgpu_vm *vm,
1554 struct amdgpu_bo_va_mapping *mapping,
1555 uint64_t flags,
1556 struct amdgpu_device *bo_adev,
1557 struct drm_mm_node *nodes,
1558 struct dma_fence **fence)
1559{
1560 unsigned min_linear_pages = 1 << adev->vm_manager.fragment_size;
1561 uint64_t pfn, start = mapping->start;
1562 int r;
1563
1564 /* normally,bo_va->flags only contians READABLE and WIRTEABLE bit go here
1565 * but in case of something, we filter the flags in first place
1566 */
1567 if (!(mapping->flags & AMDGPU_PTE_READABLE))
1568 flags &= ~AMDGPU_PTE_READABLE;
1569 if (!(mapping->flags & AMDGPU_PTE_WRITEABLE))
1570 flags &= ~AMDGPU_PTE_WRITEABLE;
1571
1572 flags &= ~AMDGPU_PTE_EXECUTABLE;
1573 flags |= mapping->flags & AMDGPU_PTE_EXECUTABLE;
1574
1575 if (adev->asic_type >= CHIP_NAVI10) {
1576 flags &= ~AMDGPU_PTE_MTYPE_NV10_MASK;
1577 flags |= (mapping->flags & AMDGPU_PTE_MTYPE_NV10_MASK);
1578 } else {
1579 flags &= ~AMDGPU_PTE_MTYPE_VG10_MASK;
1580 flags |= (mapping->flags & AMDGPU_PTE_MTYPE_VG10_MASK);
1581 }
1582
1583 if ((mapping->flags & AMDGPU_PTE_PRT) &&
1584 (adev->asic_type >= CHIP_VEGA10)) {
1585 flags |= AMDGPU_PTE_PRT;
1586 if (adev->asic_type >= CHIP_NAVI10) {
1587 flags |= AMDGPU_PTE_SNOOPED;
1588 flags |= AMDGPU_PTE_LOG;
1589 flags |= AMDGPU_PTE_SYSTEM;
1590 }
1591 flags &= ~AMDGPU_PTE_VALID;
1592 }
1593
1594 trace_amdgpu_vm_bo_update(mapping);
1595
1596 pfn = mapping->offset >> PAGE_SHIFT;
1597 if (nodes) {
1598 while (pfn >= nodes->size) {
1599 pfn -= nodes->size;
1600 ++nodes;
1601 }
1602 }
1603
1604 do {
1605 dma_addr_t *dma_addr = NULL;
1606 uint64_t max_entries;
1607 uint64_t addr, last;
1608
1609 if (nodes) {
1610 addr = nodes->start << PAGE_SHIFT;
1611 max_entries = (nodes->size - pfn) *
1612 AMDGPU_GPU_PAGES_IN_CPU_PAGE;
1613 } else {
1614 addr = 0;
1615 max_entries = S64_MAX;
1616 }
1617
1618 if (pages_addr) {
1619 uint64_t count;
1620
1621 for (count = 1;
1622 count < max_entries / AMDGPU_GPU_PAGES_IN_CPU_PAGE;
1623 ++count) {
1624 uint64_t idx = pfn + count;
1625
1626 if (pages_addr[idx] !=
1627 (pages_addr[idx - 1] + PAGE_SIZE))
1628 break;
1629 }
1630
1631 if (count < min_linear_pages) {
1632 addr = pfn << PAGE_SHIFT;
1633 dma_addr = pages_addr;
1634 } else {
1635 addr = pages_addr[pfn];
1636 max_entries = count * AMDGPU_GPU_PAGES_IN_CPU_PAGE;
1637 }
1638
1639 } else if (flags & AMDGPU_PTE_VALID) {
1640 addr += bo_adev->vm_manager.vram_base_offset;
1641 addr += pfn << PAGE_SHIFT;
1642 }
1643
1644 last = min((uint64_t)mapping->last, start + max_entries - 1);
1645 r = amdgpu_vm_bo_update_mapping(adev, exclusive, dma_addr, vm,
1646 start, last, flags, addr,
1647 fence);
1648 if (r)
1649 return r;
1650
1651 pfn += (last - start + 1) / AMDGPU_GPU_PAGES_IN_CPU_PAGE;
1652 if (nodes && nodes->size == pfn) {
1653 pfn = 0;
1654 ++nodes;
1655 }
1656 start = last + 1;
1657
1658 } while (unlikely(start != mapping->last + 1));
1659
1660 return 0;
1661}
1662
1663/**
1664 * amdgpu_vm_bo_update - update all BO mappings in the vm page table
1665 *
1666 * @adev: amdgpu_device pointer
1667 * @bo_va: requested BO and VM object
1668 * @clear: if true clear the entries
1669 *
1670 * Fill in the page table entries for @bo_va.
1671 *
1672 * Returns:
1673 * 0 for success, -EINVAL for failure.
1674 */
1675int amdgpu_vm_bo_update(struct amdgpu_device *adev,
1676 struct amdgpu_bo_va *bo_va,
1677 bool clear)
1678{
1679 struct amdgpu_bo *bo = bo_va->base.bo;
1680 struct amdgpu_vm *vm = bo_va->base.vm;
1681 struct amdgpu_bo_va_mapping *mapping;
1682 dma_addr_t *pages_addr = NULL;
1683 struct ttm_mem_reg *mem;
1684 struct drm_mm_node *nodes;
1685 struct dma_fence *exclusive, **last_update;
1686 uint64_t flags;
1687 struct amdgpu_device *bo_adev = adev;
1688 int r;
1689
1690 if (clear || !bo) {
1691 mem = NULL;
1692 nodes = NULL;
1693 exclusive = NULL;
1694 } else {
1695 struct ttm_dma_tt *ttm;
1696
1697 mem = &bo->tbo.mem;
1698 nodes = mem->mm_node;
1699 if (mem->mem_type == TTM_PL_TT) {
1700 ttm = container_of(bo->tbo.ttm, struct ttm_dma_tt, ttm);
1701 pages_addr = ttm->dma_address;
1702 }
1703 exclusive = dma_resv_get_excl(bo->tbo.base.resv);
1704 }
1705
1706 if (bo) {
1707 flags = amdgpu_ttm_tt_pte_flags(adev, bo->tbo.ttm, mem);
1708 bo_adev = amdgpu_ttm_adev(bo->tbo.bdev);
1709 } else {
1710 flags = 0x0;
1711 }
1712
1713 if (clear || (bo && bo->tbo.base.resv == vm->root.base.bo->tbo.base.resv))
1714 last_update = &vm->last_update;
1715 else
1716 last_update = &bo_va->last_pt_update;
1717
1718 if (!clear && bo_va->base.moved) {
1719 bo_va->base.moved = false;
1720 list_splice_init(&bo_va->valids, &bo_va->invalids);
1721
1722 } else if (bo_va->cleared != clear) {
1723 list_splice_init(&bo_va->valids, &bo_va->invalids);
1724 }
1725
1726 list_for_each_entry(mapping, &bo_va->invalids, list) {
1727 r = amdgpu_vm_bo_split_mapping(adev, exclusive, pages_addr, vm,
1728 mapping, flags, bo_adev, nodes,
1729 last_update);
1730 if (r)
1731 return r;
1732 }
1733
1734 if (vm->use_cpu_for_update) {
1735 /* Flush HDP */
1736 mb();
1737 amdgpu_asic_flush_hdp(adev, NULL);
1738 }
1739
1740 /* If the BO is not in its preferred location add it back to
1741 * the evicted list so that it gets validated again on the
1742 * next command submission.
1743 */
1744 if (bo && bo->tbo.base.resv == vm->root.base.bo->tbo.base.resv) {
1745 uint32_t mem_type = bo->tbo.mem.mem_type;
1746
1747 if (!(bo->preferred_domains & amdgpu_mem_type_to_domain(mem_type)))
1748 amdgpu_vm_bo_evicted(&bo_va->base);
1749 else
1750 amdgpu_vm_bo_idle(&bo_va->base);
1751 } else {
1752 amdgpu_vm_bo_done(&bo_va->base);
1753 }
1754
1755 list_splice_init(&bo_va->invalids, &bo_va->valids);
1756 bo_va->cleared = clear;
1757
1758 if (trace_amdgpu_vm_bo_mapping_enabled()) {
1759 list_for_each_entry(mapping, &bo_va->valids, list)
1760 trace_amdgpu_vm_bo_mapping(mapping);
1761 }
1762
1763 return 0;
1764}
1765
1766/**
1767 * amdgpu_vm_update_prt_state - update the global PRT state
1768 *
1769 * @adev: amdgpu_device pointer
1770 */
1771static void amdgpu_vm_update_prt_state(struct amdgpu_device *adev)
1772{
1773 unsigned long flags;
1774 bool enable;
1775
1776 spin_lock_irqsave(&adev->vm_manager.prt_lock, flags);
1777 enable = !!atomic_read(&adev->vm_manager.num_prt_users);
1778 adev->gmc.gmc_funcs->set_prt(adev, enable);
1779 spin_unlock_irqrestore(&adev->vm_manager.prt_lock, flags);
1780}
1781
1782/**
1783 * amdgpu_vm_prt_get - add a PRT user
1784 *
1785 * @adev: amdgpu_device pointer
1786 */
1787static void amdgpu_vm_prt_get(struct amdgpu_device *adev)
1788{
1789 if (!adev->gmc.gmc_funcs->set_prt)
1790 return;
1791
1792 if (atomic_inc_return(&adev->vm_manager.num_prt_users) == 1)
1793 amdgpu_vm_update_prt_state(adev);
1794}
1795
1796/**
1797 * amdgpu_vm_prt_put - drop a PRT user
1798 *
1799 * @adev: amdgpu_device pointer
1800 */
1801static void amdgpu_vm_prt_put(struct amdgpu_device *adev)
1802{
1803 if (atomic_dec_return(&adev->vm_manager.num_prt_users) == 0)
1804 amdgpu_vm_update_prt_state(adev);
1805}
1806
1807/**
1808 * amdgpu_vm_prt_cb - callback for updating the PRT status
1809 *
1810 * @fence: fence for the callback
1811 * @_cb: the callback function
1812 */
1813static void amdgpu_vm_prt_cb(struct dma_fence *fence, struct dma_fence_cb *_cb)
1814{
1815 struct amdgpu_prt_cb *cb = container_of(_cb, struct amdgpu_prt_cb, cb);
1816
1817 amdgpu_vm_prt_put(cb->adev);
1818 kfree(cb);
1819}
1820
1821/**
1822 * amdgpu_vm_add_prt_cb - add callback for updating the PRT status
1823 *
1824 * @adev: amdgpu_device pointer
1825 * @fence: fence for the callback
1826 */
1827static void amdgpu_vm_add_prt_cb(struct amdgpu_device *adev,
1828 struct dma_fence *fence)
1829{
1830 struct amdgpu_prt_cb *cb;
1831
1832 if (!adev->gmc.gmc_funcs->set_prt)
1833 return;
1834
1835 cb = kmalloc(sizeof(struct amdgpu_prt_cb), GFP_KERNEL);
1836 if (!cb) {
1837 /* Last resort when we are OOM */
1838 if (fence)
1839 dma_fence_wait(fence, false);
1840
1841 amdgpu_vm_prt_put(adev);
1842 } else {
1843 cb->adev = adev;
1844 if (!fence || dma_fence_add_callback(fence, &cb->cb,
1845 amdgpu_vm_prt_cb))
1846 amdgpu_vm_prt_cb(fence, &cb->cb);
1847 }
1848}
1849
1850/**
1851 * amdgpu_vm_free_mapping - free a mapping
1852 *
1853 * @adev: amdgpu_device pointer
1854 * @vm: requested vm
1855 * @mapping: mapping to be freed
1856 * @fence: fence of the unmap operation
1857 *
1858 * Free a mapping and make sure we decrease the PRT usage count if applicable.
1859 */
1860static void amdgpu_vm_free_mapping(struct amdgpu_device *adev,
1861 struct amdgpu_vm *vm,
1862 struct amdgpu_bo_va_mapping *mapping,
1863 struct dma_fence *fence)
1864{
1865 if (mapping->flags & AMDGPU_PTE_PRT)
1866 amdgpu_vm_add_prt_cb(adev, fence);
1867 kfree(mapping);
1868}
1869
1870/**
1871 * amdgpu_vm_prt_fini - finish all prt mappings
1872 *
1873 * @adev: amdgpu_device pointer
1874 * @vm: requested vm
1875 *
1876 * Register a cleanup callback to disable PRT support after VM dies.
1877 */
1878static void amdgpu_vm_prt_fini(struct amdgpu_device *adev, struct amdgpu_vm *vm)
1879{
1880 struct dma_resv *resv = vm->root.base.bo->tbo.base.resv;
1881 struct dma_fence *excl, **shared;
1882 unsigned i, shared_count;
1883 int r;
1884
1885 r = dma_resv_get_fences_rcu(resv, &excl,
1886 &shared_count, &shared);
1887 if (r) {
1888 /* Not enough memory to grab the fence list, as last resort
1889 * block for all the fences to complete.
1890 */
1891 dma_resv_wait_timeout_rcu(resv, true, false,
1892 MAX_SCHEDULE_TIMEOUT);
1893 return;
1894 }
1895
1896 /* Add a callback for each fence in the reservation object */
1897 amdgpu_vm_prt_get(adev);
1898 amdgpu_vm_add_prt_cb(adev, excl);
1899
1900 for (i = 0; i < shared_count; ++i) {
1901 amdgpu_vm_prt_get(adev);
1902 amdgpu_vm_add_prt_cb(adev, shared[i]);
1903 }
1904
1905 kfree(shared);
1906}
1907
1908/**
1909 * amdgpu_vm_clear_freed - clear freed BOs in the PT
1910 *
1911 * @adev: amdgpu_device pointer
1912 * @vm: requested vm
1913 * @fence: optional resulting fence (unchanged if no work needed to be done
1914 * or if an error occurred)
1915 *
1916 * Make sure all freed BOs are cleared in the PT.
1917 * PTs have to be reserved and mutex must be locked!
1918 *
1919 * Returns:
1920 * 0 for success.
1921 *
1922 */
1923int amdgpu_vm_clear_freed(struct amdgpu_device *adev,
1924 struct amdgpu_vm *vm,
1925 struct dma_fence **fence)
1926{
1927 struct amdgpu_bo_va_mapping *mapping;
1928 uint64_t init_pte_value = 0;
1929 struct dma_fence *f = NULL;
1930 int r;
1931
1932 while (!list_empty(&vm->freed)) {
1933 mapping = list_first_entry(&vm->freed,
1934 struct amdgpu_bo_va_mapping, list);
1935 list_del(&mapping->list);
1936
1937 if (vm->pte_support_ats &&
1938 mapping->start < AMDGPU_GMC_HOLE_START)
1939 init_pte_value = AMDGPU_PTE_DEFAULT_ATC;
1940
1941 r = amdgpu_vm_bo_update_mapping(adev, NULL, NULL, vm,
1942 mapping->start, mapping->last,
1943 init_pte_value, 0, &f);
1944 amdgpu_vm_free_mapping(adev, vm, mapping, f);
1945 if (r) {
1946 dma_fence_put(f);
1947 return r;
1948 }
1949 }
1950
1951 if (fence && f) {
1952 dma_fence_put(*fence);
1953 *fence = f;
1954 } else {
1955 dma_fence_put(f);
1956 }
1957
1958 return 0;
1959
1960}
1961
1962/**
1963 * amdgpu_vm_handle_moved - handle moved BOs in the PT
1964 *
1965 * @adev: amdgpu_device pointer
1966 * @vm: requested vm
1967 *
1968 * Make sure all BOs which are moved are updated in the PTs.
1969 *
1970 * Returns:
1971 * 0 for success.
1972 *
1973 * PTs have to be reserved!
1974 */
1975int amdgpu_vm_handle_moved(struct amdgpu_device *adev,
1976 struct amdgpu_vm *vm)
1977{
1978 struct amdgpu_bo_va *bo_va, *tmp;
1979 struct dma_resv *resv;
1980 bool clear;
1981 int r;
1982
1983 list_for_each_entry_safe(bo_va, tmp, &vm->moved, base.vm_status) {
1984 /* Per VM BOs never need to bo cleared in the page tables */
1985 r = amdgpu_vm_bo_update(adev, bo_va, false);
1986 if (r)
1987 return r;
1988 }
1989
1990 spin_lock(&vm->invalidated_lock);
1991 while (!list_empty(&vm->invalidated)) {
1992 bo_va = list_first_entry(&vm->invalidated, struct amdgpu_bo_va,
1993 base.vm_status);
1994 resv = bo_va->base.bo->tbo.base.resv;
1995 spin_unlock(&vm->invalidated_lock);
1996
1997 /* Try to reserve the BO to avoid clearing its ptes */
1998 if (!amdgpu_vm_debug && dma_resv_trylock(resv))
1999 clear = false;
2000 /* Somebody else is using the BO right now */
2001 else
2002 clear = true;
2003
2004 r = amdgpu_vm_bo_update(adev, bo_va, clear);
2005 if (r)
2006 return r;
2007
2008 if (!clear)
2009 dma_resv_unlock(resv);
2010 spin_lock(&vm->invalidated_lock);
2011 }
2012 spin_unlock(&vm->invalidated_lock);
2013
2014 return 0;
2015}
2016
2017/**
2018 * amdgpu_vm_bo_add - add a bo to a specific vm
2019 *
2020 * @adev: amdgpu_device pointer
2021 * @vm: requested vm
2022 * @bo: amdgpu buffer object
2023 *
2024 * Add @bo into the requested vm.
2025 * Add @bo to the list of bos associated with the vm
2026 *
2027 * Returns:
2028 * Newly added bo_va or NULL for failure
2029 *
2030 * Object has to be reserved!
2031 */
2032struct amdgpu_bo_va *amdgpu_vm_bo_add(struct amdgpu_device *adev,
2033 struct amdgpu_vm *vm,
2034 struct amdgpu_bo *bo)
2035{
2036 struct amdgpu_bo_va *bo_va;
2037
2038 bo_va = kzalloc(sizeof(struct amdgpu_bo_va), GFP_KERNEL);
2039 if (bo_va == NULL) {
2040 return NULL;
2041 }
2042 amdgpu_vm_bo_base_init(&bo_va->base, vm, bo);
2043
2044 bo_va->ref_count = 1;
2045 INIT_LIST_HEAD(&bo_va->valids);
2046 INIT_LIST_HEAD(&bo_va->invalids);
2047
2048 if (bo && amdgpu_xgmi_same_hive(adev, amdgpu_ttm_adev(bo->tbo.bdev)) &&
2049 (bo->preferred_domains & AMDGPU_GEM_DOMAIN_VRAM)) {
2050 bo_va->is_xgmi = true;
2051 mutex_lock(&adev->vm_manager.lock_pstate);
2052 /* Power up XGMI if it can be potentially used */
2053 if (++adev->vm_manager.xgmi_map_counter == 1)
2054 amdgpu_xgmi_set_pstate(adev, 1);
2055 mutex_unlock(&adev->vm_manager.lock_pstate);
2056 }
2057
2058 return bo_va;
2059}
2060
2061
2062/**
2063 * amdgpu_vm_bo_insert_mapping - insert a new mapping
2064 *
2065 * @adev: amdgpu_device pointer
2066 * @bo_va: bo_va to store the address
2067 * @mapping: the mapping to insert
2068 *
2069 * Insert a new mapping into all structures.
2070 */
2071static void amdgpu_vm_bo_insert_map(struct amdgpu_device *adev,
2072 struct amdgpu_bo_va *bo_va,
2073 struct amdgpu_bo_va_mapping *mapping)
2074{
2075 struct amdgpu_vm *vm = bo_va->base.vm;
2076 struct amdgpu_bo *bo = bo_va->base.bo;
2077
2078 mapping->bo_va = bo_va;
2079 list_add(&mapping->list, &bo_va->invalids);
2080 amdgpu_vm_it_insert(mapping, &vm->va);
2081
2082 if (mapping->flags & AMDGPU_PTE_PRT)
2083 amdgpu_vm_prt_get(adev);
2084
2085 if (bo && bo->tbo.base.resv == vm->root.base.bo->tbo.base.resv &&
2086 !bo_va->base.moved) {
2087 list_move(&bo_va->base.vm_status, &vm->moved);
2088 }
2089 trace_amdgpu_vm_bo_map(bo_va, mapping);
2090}
2091
2092/**
2093 * amdgpu_vm_bo_map - map bo inside a vm
2094 *
2095 * @adev: amdgpu_device pointer
2096 * @bo_va: bo_va to store the address
2097 * @saddr: where to map the BO
2098 * @offset: requested offset in the BO
2099 * @size: BO size in bytes
2100 * @flags: attributes of pages (read/write/valid/etc.)
2101 *
2102 * Add a mapping of the BO at the specefied addr into the VM.
2103 *
2104 * Returns:
2105 * 0 for success, error for failure.
2106 *
2107 * Object has to be reserved and unreserved outside!
2108 */
2109int amdgpu_vm_bo_map(struct amdgpu_device *adev,
2110 struct amdgpu_bo_va *bo_va,
2111 uint64_t saddr, uint64_t offset,
2112 uint64_t size, uint64_t flags)
2113{
2114 struct amdgpu_bo_va_mapping *mapping, *tmp;
2115 struct amdgpu_bo *bo = bo_va->base.bo;
2116 struct amdgpu_vm *vm = bo_va->base.vm;
2117 uint64_t eaddr;
2118
2119 /* validate the parameters */
2120 if (saddr & AMDGPU_GPU_PAGE_MASK || offset & AMDGPU_GPU_PAGE_MASK ||
2121 size == 0 || size & AMDGPU_GPU_PAGE_MASK)
2122 return -EINVAL;
2123
2124 /* make sure object fit at this offset */
2125 eaddr = saddr + size - 1;
2126 if (saddr >= eaddr ||
2127 (bo && offset + size > amdgpu_bo_size(bo)))
2128 return -EINVAL;
2129
2130 saddr /= AMDGPU_GPU_PAGE_SIZE;
2131 eaddr /= AMDGPU_GPU_PAGE_SIZE;
2132
2133 tmp = amdgpu_vm_it_iter_first(&vm->va, saddr, eaddr);
2134 if (tmp) {
2135 /* bo and tmp overlap, invalid addr */
2136 dev_err(adev->dev, "bo %p va 0x%010Lx-0x%010Lx conflict with "
2137 "0x%010Lx-0x%010Lx\n", bo, saddr, eaddr,
2138 tmp->start, tmp->last + 1);
2139 return -EINVAL;
2140 }
2141
2142 mapping = kmalloc(sizeof(*mapping), GFP_KERNEL);
2143 if (!mapping)
2144 return -ENOMEM;
2145
2146 mapping->start = saddr;
2147 mapping->last = eaddr;
2148 mapping->offset = offset;
2149 mapping->flags = flags;
2150
2151 amdgpu_vm_bo_insert_map(adev, bo_va, mapping);
2152
2153 return 0;
2154}
2155
2156/**
2157 * amdgpu_vm_bo_replace_map - map bo inside a vm, replacing existing mappings
2158 *
2159 * @adev: amdgpu_device pointer
2160 * @bo_va: bo_va to store the address
2161 * @saddr: where to map the BO
2162 * @offset: requested offset in the BO
2163 * @size: BO size in bytes
2164 * @flags: attributes of pages (read/write/valid/etc.)
2165 *
2166 * Add a mapping of the BO at the specefied addr into the VM. Replace existing
2167 * mappings as we do so.
2168 *
2169 * Returns:
2170 * 0 for success, error for failure.
2171 *
2172 * Object has to be reserved and unreserved outside!
2173 */
2174int amdgpu_vm_bo_replace_map(struct amdgpu_device *adev,
2175 struct amdgpu_bo_va *bo_va,
2176 uint64_t saddr, uint64_t offset,
2177 uint64_t size, uint64_t flags)
2178{
2179 struct amdgpu_bo_va_mapping *mapping;
2180 struct amdgpu_bo *bo = bo_va->base.bo;
2181 uint64_t eaddr;
2182 int r;
2183
2184 /* validate the parameters */
2185 if (saddr & AMDGPU_GPU_PAGE_MASK || offset & AMDGPU_GPU_PAGE_MASK ||
2186 size == 0 || size & AMDGPU_GPU_PAGE_MASK)
2187 return -EINVAL;
2188
2189 /* make sure object fit at this offset */
2190 eaddr = saddr + size - 1;
2191 if (saddr >= eaddr ||
2192 (bo && offset + size > amdgpu_bo_size(bo)))
2193 return -EINVAL;
2194
2195 /* Allocate all the needed memory */
2196 mapping = kmalloc(sizeof(*mapping), GFP_KERNEL);
2197 if (!mapping)
2198 return -ENOMEM;
2199
2200 r = amdgpu_vm_bo_clear_mappings(adev, bo_va->base.vm, saddr, size);
2201 if (r) {
2202 kfree(mapping);
2203 return r;
2204 }
2205
2206 saddr /= AMDGPU_GPU_PAGE_SIZE;
2207 eaddr /= AMDGPU_GPU_PAGE_SIZE;
2208
2209 mapping->start = saddr;
2210 mapping->last = eaddr;
2211 mapping->offset = offset;
2212 mapping->flags = flags;
2213
2214 amdgpu_vm_bo_insert_map(adev, bo_va, mapping);
2215
2216 return 0;
2217}
2218
2219/**
2220 * amdgpu_vm_bo_unmap - remove bo mapping from vm
2221 *
2222 * @adev: amdgpu_device pointer
2223 * @bo_va: bo_va to remove the address from
2224 * @saddr: where to the BO is mapped
2225 *
2226 * Remove a mapping of the BO at the specefied addr from the VM.
2227 *
2228 * Returns:
2229 * 0 for success, error for failure.
2230 *
2231 * Object has to be reserved and unreserved outside!
2232 */
2233int amdgpu_vm_bo_unmap(struct amdgpu_device *adev,
2234 struct amdgpu_bo_va *bo_va,
2235 uint64_t saddr)
2236{
2237 struct amdgpu_bo_va_mapping *mapping;
2238 struct amdgpu_vm *vm = bo_va->base.vm;
2239 bool valid = true;
2240
2241 saddr /= AMDGPU_GPU_PAGE_SIZE;
2242
2243 list_for_each_entry(mapping, &bo_va->valids, list) {
2244 if (mapping->start == saddr)
2245 break;
2246 }
2247
2248 if (&mapping->list == &bo_va->valids) {
2249 valid = false;
2250
2251 list_for_each_entry(mapping, &bo_va->invalids, list) {
2252 if (mapping->start == saddr)
2253 break;
2254 }
2255
2256 if (&mapping->list == &bo_va->invalids)
2257 return -ENOENT;
2258 }
2259
2260 list_del(&mapping->list);
2261 amdgpu_vm_it_remove(mapping, &vm->va);
2262 mapping->bo_va = NULL;
2263 trace_amdgpu_vm_bo_unmap(bo_va, mapping);
2264
2265 if (valid)
2266 list_add(&mapping->list, &vm->freed);
2267 else
2268 amdgpu_vm_free_mapping(adev, vm, mapping,
2269 bo_va->last_pt_update);
2270
2271 return 0;
2272}
2273
2274/**
2275 * amdgpu_vm_bo_clear_mappings - remove all mappings in a specific range
2276 *
2277 * @adev: amdgpu_device pointer
2278 * @vm: VM structure to use
2279 * @saddr: start of the range
2280 * @size: size of the range
2281 *
2282 * Remove all mappings in a range, split them as appropriate.
2283 *
2284 * Returns:
2285 * 0 for success, error for failure.
2286 */
2287int amdgpu_vm_bo_clear_mappings(struct amdgpu_device *adev,
2288 struct amdgpu_vm *vm,
2289 uint64_t saddr, uint64_t size)
2290{
2291 struct amdgpu_bo_va_mapping *before, *after, *tmp, *next;
2292 LIST_HEAD(removed);
2293 uint64_t eaddr;
2294
2295 eaddr = saddr + size - 1;
2296 saddr /= AMDGPU_GPU_PAGE_SIZE;
2297 eaddr /= AMDGPU_GPU_PAGE_SIZE;
2298
2299 /* Allocate all the needed memory */
2300 before = kzalloc(sizeof(*before), GFP_KERNEL);
2301 if (!before)
2302 return -ENOMEM;
2303 INIT_LIST_HEAD(&before->list);
2304
2305 after = kzalloc(sizeof(*after), GFP_KERNEL);
2306 if (!after) {
2307 kfree(before);
2308 return -ENOMEM;
2309 }
2310 INIT_LIST_HEAD(&after->list);
2311
2312 /* Now gather all removed mappings */
2313 tmp = amdgpu_vm_it_iter_first(&vm->va, saddr, eaddr);
2314 while (tmp) {
2315 /* Remember mapping split at the start */
2316 if (tmp->start < saddr) {
2317 before->start = tmp->start;
2318 before->last = saddr - 1;
2319 before->offset = tmp->offset;
2320 before->flags = tmp->flags;
2321 before->bo_va = tmp->bo_va;
2322 list_add(&before->list, &tmp->bo_va->invalids);
2323 }
2324
2325 /* Remember mapping split at the end */
2326 if (tmp->last > eaddr) {
2327 after->start = eaddr + 1;
2328 after->last = tmp->last;
2329 after->offset = tmp->offset;
2330 after->offset += after->start - tmp->start;
2331 after->flags = tmp->flags;
2332 after->bo_va = tmp->bo_va;
2333 list_add(&after->list, &tmp->bo_va->invalids);
2334 }
2335
2336 list_del(&tmp->list);
2337 list_add(&tmp->list, &removed);
2338
2339 tmp = amdgpu_vm_it_iter_next(tmp, saddr, eaddr);
2340 }
2341
2342 /* And free them up */
2343 list_for_each_entry_safe(tmp, next, &removed, list) {
2344 amdgpu_vm_it_remove(tmp, &vm->va);
2345 list_del(&tmp->list);
2346
2347 if (tmp->start < saddr)
2348 tmp->start = saddr;
2349 if (tmp->last > eaddr)
2350 tmp->last = eaddr;
2351
2352 tmp->bo_va = NULL;
2353 list_add(&tmp->list, &vm->freed);
2354 trace_amdgpu_vm_bo_unmap(NULL, tmp);
2355 }
2356
2357 /* Insert partial mapping before the range */
2358 if (!list_empty(&before->list)) {
2359 amdgpu_vm_it_insert(before, &vm->va);
2360 if (before->flags & AMDGPU_PTE_PRT)
2361 amdgpu_vm_prt_get(adev);
2362 } else {
2363 kfree(before);
2364 }
2365
2366 /* Insert partial mapping after the range */
2367 if (!list_empty(&after->list)) {
2368 amdgpu_vm_it_insert(after, &vm->va);
2369 if (after->flags & AMDGPU_PTE_PRT)
2370 amdgpu_vm_prt_get(adev);
2371 } else {
2372 kfree(after);
2373 }
2374
2375 return 0;
2376}
2377
2378/**
2379 * amdgpu_vm_bo_lookup_mapping - find mapping by address
2380 *
2381 * @vm: the requested VM
2382 * @addr: the address
2383 *
2384 * Find a mapping by it's address.
2385 *
2386 * Returns:
2387 * The amdgpu_bo_va_mapping matching for addr or NULL
2388 *
2389 */
2390struct amdgpu_bo_va_mapping *amdgpu_vm_bo_lookup_mapping(struct amdgpu_vm *vm,
2391 uint64_t addr)
2392{
2393 return amdgpu_vm_it_iter_first(&vm->va, addr, addr);
2394}
2395
2396/**
2397 * amdgpu_vm_bo_trace_cs - trace all reserved mappings
2398 *
2399 * @vm: the requested vm
2400 * @ticket: CS ticket
2401 *
2402 * Trace all mappings of BOs reserved during a command submission.
2403 */
2404void amdgpu_vm_bo_trace_cs(struct amdgpu_vm *vm, struct ww_acquire_ctx *ticket)
2405{
2406 struct amdgpu_bo_va_mapping *mapping;
2407
2408 if (!trace_amdgpu_vm_bo_cs_enabled())
2409 return;
2410
2411 for (mapping = amdgpu_vm_it_iter_first(&vm->va, 0, U64_MAX); mapping;
2412 mapping = amdgpu_vm_it_iter_next(mapping, 0, U64_MAX)) {
2413 if (mapping->bo_va && mapping->bo_va->base.bo) {
2414 struct amdgpu_bo *bo;
2415
2416 bo = mapping->bo_va->base.bo;
2417 if (dma_resv_locking_ctx(bo->tbo.base.resv) !=
2418 ticket)
2419 continue;
2420 }
2421
2422 trace_amdgpu_vm_bo_cs(mapping);
2423 }
2424}
2425
2426/**
2427 * amdgpu_vm_bo_rmv - remove a bo to a specific vm
2428 *
2429 * @adev: amdgpu_device pointer
2430 * @bo_va: requested bo_va
2431 *
2432 * Remove @bo_va->bo from the requested vm.
2433 *
2434 * Object have to be reserved!
2435 */
2436void amdgpu_vm_bo_rmv(struct amdgpu_device *adev,
2437 struct amdgpu_bo_va *bo_va)
2438{
2439 struct amdgpu_bo_va_mapping *mapping, *next;
2440 struct amdgpu_bo *bo = bo_va->base.bo;
2441 struct amdgpu_vm *vm = bo_va->base.vm;
2442 struct amdgpu_vm_bo_base **base;
2443
2444 if (bo) {
2445 if (bo->tbo.base.resv == vm->root.base.bo->tbo.base.resv)
2446 vm->bulk_moveable = false;
2447
2448 for (base = &bo_va->base.bo->vm_bo; *base;
2449 base = &(*base)->next) {
2450 if (*base != &bo_va->base)
2451 continue;
2452
2453 *base = bo_va->base.next;
2454 break;
2455 }
2456 }
2457
2458 spin_lock(&vm->invalidated_lock);
2459 list_del(&bo_va->base.vm_status);
2460 spin_unlock(&vm->invalidated_lock);
2461
2462 list_for_each_entry_safe(mapping, next, &bo_va->valids, list) {
2463 list_del(&mapping->list);
2464 amdgpu_vm_it_remove(mapping, &vm->va);
2465 mapping->bo_va = NULL;
2466 trace_amdgpu_vm_bo_unmap(bo_va, mapping);
2467 list_add(&mapping->list, &vm->freed);
2468 }
2469 list_for_each_entry_safe(mapping, next, &bo_va->invalids, list) {
2470 list_del(&mapping->list);
2471 amdgpu_vm_it_remove(mapping, &vm->va);
2472 amdgpu_vm_free_mapping(adev, vm, mapping,
2473 bo_va->last_pt_update);
2474 }
2475
2476 dma_fence_put(bo_va->last_pt_update);
2477
2478 if (bo && bo_va->is_xgmi) {
2479 mutex_lock(&adev->vm_manager.lock_pstate);
2480 if (--adev->vm_manager.xgmi_map_counter == 0)
2481 amdgpu_xgmi_set_pstate(adev, 0);
2482 mutex_unlock(&adev->vm_manager.lock_pstate);
2483 }
2484
2485 kfree(bo_va);
2486}
2487
2488/**
2489 * amdgpu_vm_bo_invalidate - mark the bo as invalid
2490 *
2491 * @adev: amdgpu_device pointer
2492 * @bo: amdgpu buffer object
2493 * @evicted: is the BO evicted
2494 *
2495 * Mark @bo as invalid.
2496 */
2497void amdgpu_vm_bo_invalidate(struct amdgpu_device *adev,
2498 struct amdgpu_bo *bo, bool evicted)
2499{
2500 struct amdgpu_vm_bo_base *bo_base;
2501
2502 /* shadow bo doesn't have bo base, its validation needs its parent */
2503 if (bo->parent && bo->parent->shadow == bo)
2504 bo = bo->parent;
2505
2506 for (bo_base = bo->vm_bo; bo_base; bo_base = bo_base->next) {
2507 struct amdgpu_vm *vm = bo_base->vm;
2508
2509 if (evicted && bo->tbo.base.resv == vm->root.base.bo->tbo.base.resv) {
2510 amdgpu_vm_bo_evicted(bo_base);
2511 continue;
2512 }
2513
2514 if (bo_base->moved)
2515 continue;
2516 bo_base->moved = true;
2517
2518 if (bo->tbo.type == ttm_bo_type_kernel)
2519 amdgpu_vm_bo_relocated(bo_base);
2520 else if (bo->tbo.base.resv == vm->root.base.bo->tbo.base.resv)
2521 amdgpu_vm_bo_moved(bo_base);
2522 else
2523 amdgpu_vm_bo_invalidated(bo_base);
2524 }
2525}
2526
2527/**
2528 * amdgpu_vm_get_block_size - calculate VM page table size as power of two
2529 *
2530 * @vm_size: VM size
2531 *
2532 * Returns:
2533 * VM page table as power of two
2534 */
2535static uint32_t amdgpu_vm_get_block_size(uint64_t vm_size)
2536{
2537 /* Total bits covered by PD + PTs */
2538 unsigned bits = ilog2(vm_size) + 18;
2539
2540 /* Make sure the PD is 4K in size up to 8GB address space.
2541 Above that split equal between PD and PTs */
2542 if (vm_size <= 8)
2543 return (bits - 9);
2544 else
2545 return ((bits + 3) / 2);
2546}
2547
2548/**
2549 * amdgpu_vm_adjust_size - adjust vm size, block size and fragment size
2550 *
2551 * @adev: amdgpu_device pointer
2552 * @min_vm_size: the minimum vm size in GB if it's set auto
2553 * @fragment_size_default: Default PTE fragment size
2554 * @max_level: max VMPT level
2555 * @max_bits: max address space size in bits
2556 *
2557 */
2558void amdgpu_vm_adjust_size(struct amdgpu_device *adev, uint32_t min_vm_size,
2559 uint32_t fragment_size_default, unsigned max_level,
2560 unsigned max_bits)
2561{
2562 unsigned int max_size = 1 << (max_bits - 30);
2563 unsigned int vm_size;
2564 uint64_t tmp;
2565
2566 /* adjust vm size first */
2567 if (amdgpu_vm_size != -1) {
2568 vm_size = amdgpu_vm_size;
2569 if (vm_size > max_size) {
2570 dev_warn(adev->dev, "VM size (%d) too large, max is %u GB\n",
2571 amdgpu_vm_size, max_size);
2572 vm_size = max_size;
2573 }
2574 } else {
2575 struct sysinfo si;
2576 unsigned int phys_ram_gb;
2577
2578 /* Optimal VM size depends on the amount of physical
2579 * RAM available. Underlying requirements and
2580 * assumptions:
2581 *
2582 * - Need to map system memory and VRAM from all GPUs
2583 * - VRAM from other GPUs not known here
2584 * - Assume VRAM <= system memory
2585 * - On GFX8 and older, VM space can be segmented for
2586 * different MTYPEs
2587 * - Need to allow room for fragmentation, guard pages etc.
2588 *
2589 * This adds up to a rough guess of system memory x3.
2590 * Round up to power of two to maximize the available
2591 * VM size with the given page table size.
2592 */
2593 si_meminfo(&si);
2594 phys_ram_gb = ((uint64_t)si.totalram * si.mem_unit +
2595 (1 << 30) - 1) >> 30;
2596 vm_size = roundup_pow_of_two(
2597 min(max(phys_ram_gb * 3, min_vm_size), max_size));
2598 }
2599
2600 adev->vm_manager.max_pfn = (uint64_t)vm_size << 18;
2601
2602 tmp = roundup_pow_of_two(adev->vm_manager.max_pfn);
2603 if (amdgpu_vm_block_size != -1)
2604 tmp >>= amdgpu_vm_block_size - 9;
2605 tmp = DIV_ROUND_UP(fls64(tmp) - 1, 9) - 1;
2606 adev->vm_manager.num_level = min(max_level, (unsigned)tmp);
2607 switch (adev->vm_manager.num_level) {
2608 case 3:
2609 adev->vm_manager.root_level = AMDGPU_VM_PDB2;
2610 break;
2611 case 2:
2612 adev->vm_manager.root_level = AMDGPU_VM_PDB1;
2613 break;
2614 case 1:
2615 adev->vm_manager.root_level = AMDGPU_VM_PDB0;
2616 break;
2617 default:
2618 dev_err(adev->dev, "VMPT only supports 2~4+1 levels\n");
2619 }
2620 /* block size depends on vm size and hw setup*/
2621 if (amdgpu_vm_block_size != -1)
2622 adev->vm_manager.block_size =
2623 min((unsigned)amdgpu_vm_block_size, max_bits
2624 - AMDGPU_GPU_PAGE_SHIFT
2625 - 9 * adev->vm_manager.num_level);
2626 else if (adev->vm_manager.num_level > 1)
2627 adev->vm_manager.block_size = 9;
2628 else
2629 adev->vm_manager.block_size = amdgpu_vm_get_block_size(tmp);
2630
2631 if (amdgpu_vm_fragment_size == -1)
2632 adev->vm_manager.fragment_size = fragment_size_default;
2633 else
2634 adev->vm_manager.fragment_size = amdgpu_vm_fragment_size;
2635
2636 DRM_INFO("vm size is %u GB, %u levels, block size is %u-bit, fragment size is %u-bit\n",
2637 vm_size, adev->vm_manager.num_level + 1,
2638 adev->vm_manager.block_size,
2639 adev->vm_manager.fragment_size);
2640}
2641
2642/**
2643 * amdgpu_vm_wait_idle - wait for the VM to become idle
2644 *
2645 * @vm: VM object to wait for
2646 * @timeout: timeout to wait for VM to become idle
2647 */
2648long amdgpu_vm_wait_idle(struct amdgpu_vm *vm, long timeout)
2649{
2650 return dma_resv_wait_timeout_rcu(vm->root.base.bo->tbo.base.resv,
2651 true, true, timeout);
2652}
2653
2654/**
2655 * amdgpu_vm_init - initialize a vm instance
2656 *
2657 * @adev: amdgpu_device pointer
2658 * @vm: requested vm
2659 * @vm_context: Indicates if it GFX or Compute context
2660 * @pasid: Process address space identifier
2661 *
2662 * Init @vm fields.
2663 *
2664 * Returns:
2665 * 0 for success, error for failure.
2666 */
2667int amdgpu_vm_init(struct amdgpu_device *adev, struct amdgpu_vm *vm,
2668 int vm_context, unsigned int pasid)
2669{
2670 struct amdgpu_bo_param bp;
2671 struct amdgpu_bo *root;
2672 int r, i;
2673
2674 vm->va = RB_ROOT_CACHED;
2675 for (i = 0; i < AMDGPU_MAX_VMHUBS; i++)
2676 vm->reserved_vmid[i] = NULL;
2677 INIT_LIST_HEAD(&vm->evicted);
2678 INIT_LIST_HEAD(&vm->relocated);
2679 INIT_LIST_HEAD(&vm->moved);
2680 INIT_LIST_HEAD(&vm->idle);
2681 INIT_LIST_HEAD(&vm->invalidated);
2682 spin_lock_init(&vm->invalidated_lock);
2683 INIT_LIST_HEAD(&vm->freed);
2684
2685 /* create scheduler entity for page table updates */
2686 r = drm_sched_entity_init(&vm->entity, adev->vm_manager.vm_pte_rqs,
2687 adev->vm_manager.vm_pte_num_rqs, NULL);
2688 if (r)
2689 return r;
2690
2691 vm->pte_support_ats = false;
2692
2693 if (vm_context == AMDGPU_VM_CONTEXT_COMPUTE) {
2694 vm->use_cpu_for_update = !!(adev->vm_manager.vm_update_mode &
2695 AMDGPU_VM_USE_CPU_FOR_COMPUTE);
2696
2697 if (adev->asic_type == CHIP_RAVEN)
2698 vm->pte_support_ats = true;
2699 } else {
2700 vm->use_cpu_for_update = !!(adev->vm_manager.vm_update_mode &
2701 AMDGPU_VM_USE_CPU_FOR_GFX);
2702 }
2703 DRM_DEBUG_DRIVER("VM update mode is %s\n",
2704 vm->use_cpu_for_update ? "CPU" : "SDMA");
2705 WARN_ONCE((vm->use_cpu_for_update && !amdgpu_gmc_vram_full_visible(&adev->gmc)),
2706 "CPU update of VM recommended only for large BAR system\n");
2707
2708 if (vm->use_cpu_for_update)
2709 vm->update_funcs = &amdgpu_vm_cpu_funcs;
2710 else
2711 vm->update_funcs = &amdgpu_vm_sdma_funcs;
2712 vm->last_update = NULL;
2713
2714 amdgpu_vm_bo_param(adev, vm, adev->vm_manager.root_level, &bp);
2715 if (vm_context == AMDGPU_VM_CONTEXT_COMPUTE)
2716 bp.flags &= ~AMDGPU_GEM_CREATE_SHADOW;
2717 r = amdgpu_bo_create(adev, &bp, &root);
2718 if (r)
2719 goto error_free_sched_entity;
2720
2721 r = amdgpu_bo_reserve(root, true);
2722 if (r)
2723 goto error_free_root;
2724
2725 r = dma_resv_reserve_shared(root->tbo.base.resv, 1);
2726 if (r)
2727 goto error_unreserve;
2728
2729 amdgpu_vm_bo_base_init(&vm->root.base, vm, root);
2730
2731 r = amdgpu_vm_clear_bo(adev, vm, root);
2732 if (r)
2733 goto error_unreserve;
2734
2735 amdgpu_bo_unreserve(vm->root.base.bo);
2736
2737 if (pasid) {
2738 unsigned long flags;
2739
2740 spin_lock_irqsave(&adev->vm_manager.pasid_lock, flags);
2741 r = idr_alloc(&adev->vm_manager.pasid_idr, vm, pasid, pasid + 1,
2742 GFP_ATOMIC);
2743 spin_unlock_irqrestore(&adev->vm_manager.pasid_lock, flags);
2744 if (r < 0)
2745 goto error_free_root;
2746
2747 vm->pasid = pasid;
2748 }
2749
2750 INIT_KFIFO(vm->faults);
2751
2752 return 0;
2753
2754error_unreserve:
2755 amdgpu_bo_unreserve(vm->root.base.bo);
2756
2757error_free_root:
2758 amdgpu_bo_unref(&vm->root.base.bo->shadow);
2759 amdgpu_bo_unref(&vm->root.base.bo);
2760 vm->root.base.bo = NULL;
2761
2762error_free_sched_entity:
2763 drm_sched_entity_destroy(&vm->entity);
2764
2765 return r;
2766}
2767
2768/**
2769 * amdgpu_vm_check_clean_reserved - check if a VM is clean
2770 *
2771 * @adev: amdgpu_device pointer
2772 * @vm: the VM to check
2773 *
2774 * check all entries of the root PD, if any subsequent PDs are allocated,
2775 * it means there are page table creating and filling, and is no a clean
2776 * VM
2777 *
2778 * Returns:
2779 * 0 if this VM is clean
2780 */
2781static int amdgpu_vm_check_clean_reserved(struct amdgpu_device *adev,
2782 struct amdgpu_vm *vm)
2783{
2784 enum amdgpu_vm_level root = adev->vm_manager.root_level;
2785 unsigned int entries = amdgpu_vm_num_entries(adev, root);
2786 unsigned int i = 0;
2787
2788 if (!(vm->root.entries))
2789 return 0;
2790
2791 for (i = 0; i < entries; i++) {
2792 if (vm->root.entries[i].base.bo)
2793 return -EINVAL;
2794 }
2795
2796 return 0;
2797}
2798
2799/**
2800 * amdgpu_vm_make_compute - Turn a GFX VM into a compute VM
2801 *
2802 * @adev: amdgpu_device pointer
2803 * @vm: requested vm
2804 *
2805 * This only works on GFX VMs that don't have any BOs added and no
2806 * page tables allocated yet.
2807 *
2808 * Changes the following VM parameters:
2809 * - use_cpu_for_update
2810 * - pte_supports_ats
2811 * - pasid (old PASID is released, because compute manages its own PASIDs)
2812 *
2813 * Reinitializes the page directory to reflect the changed ATS
2814 * setting.
2815 *
2816 * Returns:
2817 * 0 for success, -errno for errors.
2818 */
2819int amdgpu_vm_make_compute(struct amdgpu_device *adev, struct amdgpu_vm *vm, unsigned int pasid)
2820{
2821 bool pte_support_ats = (adev->asic_type == CHIP_RAVEN);
2822 int r;
2823
2824 r = amdgpu_bo_reserve(vm->root.base.bo, true);
2825 if (r)
2826 return r;
2827
2828 /* Sanity checks */
2829 r = amdgpu_vm_check_clean_reserved(adev, vm);
2830 if (r)
2831 goto unreserve_bo;
2832
2833 if (pasid) {
2834 unsigned long flags;
2835
2836 spin_lock_irqsave(&adev->vm_manager.pasid_lock, flags);
2837 r = idr_alloc(&adev->vm_manager.pasid_idr, vm, pasid, pasid + 1,
2838 GFP_ATOMIC);
2839 spin_unlock_irqrestore(&adev->vm_manager.pasid_lock, flags);
2840
2841 if (r == -ENOSPC)
2842 goto unreserve_bo;
2843 r = 0;
2844 }
2845
2846 /* Check if PD needs to be reinitialized and do it before
2847 * changing any other state, in case it fails.
2848 */
2849 if (pte_support_ats != vm->pte_support_ats) {
2850 vm->pte_support_ats = pte_support_ats;
2851 r = amdgpu_vm_clear_bo(adev, vm, vm->root.base.bo);
2852 if (r)
2853 goto free_idr;
2854 }
2855
2856 /* Update VM state */
2857 vm->use_cpu_for_update = !!(adev->vm_manager.vm_update_mode &
2858 AMDGPU_VM_USE_CPU_FOR_COMPUTE);
2859 DRM_DEBUG_DRIVER("VM update mode is %s\n",
2860 vm->use_cpu_for_update ? "CPU" : "SDMA");
2861 WARN_ONCE((vm->use_cpu_for_update && !amdgpu_gmc_vram_full_visible(&adev->gmc)),
2862 "CPU update of VM recommended only for large BAR system\n");
2863
2864 if (vm->use_cpu_for_update)
2865 vm->update_funcs = &amdgpu_vm_cpu_funcs;
2866 else
2867 vm->update_funcs = &amdgpu_vm_sdma_funcs;
2868 dma_fence_put(vm->last_update);
2869 vm->last_update = NULL;
2870
2871 if (vm->pasid) {
2872 unsigned long flags;
2873
2874 spin_lock_irqsave(&adev->vm_manager.pasid_lock, flags);
2875 idr_remove(&adev->vm_manager.pasid_idr, vm->pasid);
2876 spin_unlock_irqrestore(&adev->vm_manager.pasid_lock, flags);
2877
2878 /* Free the original amdgpu allocated pasid
2879 * Will be replaced with kfd allocated pasid
2880 */
2881 amdgpu_pasid_free(vm->pasid);
2882 vm->pasid = 0;
2883 }
2884
2885 /* Free the shadow bo for compute VM */
2886 amdgpu_bo_unref(&vm->root.base.bo->shadow);
2887
2888 if (pasid)
2889 vm->pasid = pasid;
2890
2891 goto unreserve_bo;
2892
2893free_idr:
2894 if (pasid) {
2895 unsigned long flags;
2896
2897 spin_lock_irqsave(&adev->vm_manager.pasid_lock, flags);
2898 idr_remove(&adev->vm_manager.pasid_idr, pasid);
2899 spin_unlock_irqrestore(&adev->vm_manager.pasid_lock, flags);
2900 }
2901unreserve_bo:
2902 amdgpu_bo_unreserve(vm->root.base.bo);
2903 return r;
2904}
2905
2906/**
2907 * amdgpu_vm_release_compute - release a compute vm
2908 * @adev: amdgpu_device pointer
2909 * @vm: a vm turned into compute vm by calling amdgpu_vm_make_compute
2910 *
2911 * This is a correspondant of amdgpu_vm_make_compute. It decouples compute
2912 * pasid from vm. Compute should stop use of vm after this call.
2913 */
2914void amdgpu_vm_release_compute(struct amdgpu_device *adev, struct amdgpu_vm *vm)
2915{
2916 if (vm->pasid) {
2917 unsigned long flags;
2918
2919 spin_lock_irqsave(&adev->vm_manager.pasid_lock, flags);
2920 idr_remove(&adev->vm_manager.pasid_idr, vm->pasid);
2921 spin_unlock_irqrestore(&adev->vm_manager.pasid_lock, flags);
2922 }
2923 vm->pasid = 0;
2924}
2925
2926/**
2927 * amdgpu_vm_fini - tear down a vm instance
2928 *
2929 * @adev: amdgpu_device pointer
2930 * @vm: requested vm
2931 *
2932 * Tear down @vm.
2933 * Unbind the VM and remove all bos from the vm bo list
2934 */
2935void amdgpu_vm_fini(struct amdgpu_device *adev, struct amdgpu_vm *vm)
2936{
2937 struct amdgpu_bo_va_mapping *mapping, *tmp;
2938 bool prt_fini_needed = !!adev->gmc.gmc_funcs->set_prt;
2939 struct amdgpu_bo *root;
2940 int i, r;
2941
2942 amdgpu_amdkfd_gpuvm_destroy_cb(adev, vm);
2943
2944 if (vm->pasid) {
2945 unsigned long flags;
2946
2947 spin_lock_irqsave(&adev->vm_manager.pasid_lock, flags);
2948 idr_remove(&adev->vm_manager.pasid_idr, vm->pasid);
2949 spin_unlock_irqrestore(&adev->vm_manager.pasid_lock, flags);
2950 }
2951
2952 drm_sched_entity_destroy(&vm->entity);
2953
2954 if (!RB_EMPTY_ROOT(&vm->va.rb_root)) {
2955 dev_err(adev->dev, "still active bo inside vm\n");
2956 }
2957 rbtree_postorder_for_each_entry_safe(mapping, tmp,
2958 &vm->va.rb_root, rb) {
2959 /* Don't remove the mapping here, we don't want to trigger a
2960 * rebalance and the tree is about to be destroyed anyway.
2961 */
2962 list_del(&mapping->list);
2963 kfree(mapping);
2964 }
2965 list_for_each_entry_safe(mapping, tmp, &vm->freed, list) {
2966 if (mapping->flags & AMDGPU_PTE_PRT && prt_fini_needed) {
2967 amdgpu_vm_prt_fini(adev, vm);
2968 prt_fini_needed = false;
2969 }
2970
2971 list_del(&mapping->list);
2972 amdgpu_vm_free_mapping(adev, vm, mapping, NULL);
2973 }
2974
2975 root = amdgpu_bo_ref(vm->root.base.bo);
2976 r = amdgpu_bo_reserve(root, true);
2977 if (r) {
2978 dev_err(adev->dev, "Leaking page tables because BO reservation failed\n");
2979 } else {
2980 amdgpu_vm_free_pts(adev, vm, NULL);
2981 amdgpu_bo_unreserve(root);
2982 }
2983 amdgpu_bo_unref(&root);
2984 WARN_ON(vm->root.base.bo);
2985 dma_fence_put(vm->last_update);
2986 for (i = 0; i < AMDGPU_MAX_VMHUBS; i++)
2987 amdgpu_vmid_free_reserved(adev, vm, i);
2988}
2989
2990/**
2991 * amdgpu_vm_manager_init - init the VM manager
2992 *
2993 * @adev: amdgpu_device pointer
2994 *
2995 * Initialize the VM manager structures
2996 */
2997void amdgpu_vm_manager_init(struct amdgpu_device *adev)
2998{
2999 unsigned i;
3000
3001 amdgpu_vmid_mgr_init(adev);
3002
3003 adev->vm_manager.fence_context =
3004 dma_fence_context_alloc(AMDGPU_MAX_RINGS);
3005 for (i = 0; i < AMDGPU_MAX_RINGS; ++i)
3006 adev->vm_manager.seqno[i] = 0;
3007
3008 spin_lock_init(&adev->vm_manager.prt_lock);
3009 atomic_set(&adev->vm_manager.num_prt_users, 0);
3010
3011 /* If not overridden by the user, by default, only in large BAR systems
3012 * Compute VM tables will be updated by CPU
3013 */
3014#ifdef CONFIG_X86_64
3015 if (amdgpu_vm_update_mode == -1) {
3016 if (amdgpu_gmc_vram_full_visible(&adev->gmc))
3017 adev->vm_manager.vm_update_mode =
3018 AMDGPU_VM_USE_CPU_FOR_COMPUTE;
3019 else
3020 adev->vm_manager.vm_update_mode = 0;
3021 } else
3022 adev->vm_manager.vm_update_mode = amdgpu_vm_update_mode;
3023#else
3024 adev->vm_manager.vm_update_mode = 0;
3025#endif
3026
3027 idr_init(&adev->vm_manager.pasid_idr);
3028 spin_lock_init(&adev->vm_manager.pasid_lock);
3029
3030 adev->vm_manager.xgmi_map_counter = 0;
3031 mutex_init(&adev->vm_manager.lock_pstate);
3032}
3033
3034/**
3035 * amdgpu_vm_manager_fini - cleanup VM manager
3036 *
3037 * @adev: amdgpu_device pointer
3038 *
3039 * Cleanup the VM manager and free resources.
3040 */
3041void amdgpu_vm_manager_fini(struct amdgpu_device *adev)
3042{
3043 WARN_ON(!idr_is_empty(&adev->vm_manager.pasid_idr));
3044 idr_destroy(&adev->vm_manager.pasid_idr);
3045
3046 amdgpu_vmid_mgr_fini(adev);
3047}
3048
3049/**
3050 * amdgpu_vm_ioctl - Manages VMID reservation for vm hubs.
3051 *
3052 * @dev: drm device pointer
3053 * @data: drm_amdgpu_vm
3054 * @filp: drm file pointer
3055 *
3056 * Returns:
3057 * 0 for success, -errno for errors.
3058 */
3059int amdgpu_vm_ioctl(struct drm_device *dev, void *data, struct drm_file *filp)
3060{
3061 union drm_amdgpu_vm *args = data;
3062 struct amdgpu_device *adev = dev->dev_private;
3063 struct amdgpu_fpriv *fpriv = filp->driver_priv;
3064 int r;
3065
3066 switch (args->in.op) {
3067 case AMDGPU_VM_OP_RESERVE_VMID:
3068 /* current, we only have requirement to reserve vmid from gfxhub */
3069 r = amdgpu_vmid_alloc_reserved(adev, &fpriv->vm, AMDGPU_GFXHUB_0);
3070 if (r)
3071 return r;
3072 break;
3073 case AMDGPU_VM_OP_UNRESERVE_VMID:
3074 amdgpu_vmid_free_reserved(adev, &fpriv->vm, AMDGPU_GFXHUB_0);
3075 break;
3076 default:
3077 return -EINVAL;
3078 }
3079
3080 return 0;
3081}
3082
3083/**
3084 * amdgpu_vm_get_task_info - Extracts task info for a PASID.
3085 *
3086 * @adev: drm device pointer
3087 * @pasid: PASID identifier for VM
3088 * @task_info: task_info to fill.
3089 */
3090void amdgpu_vm_get_task_info(struct amdgpu_device *adev, unsigned int pasid,
3091 struct amdgpu_task_info *task_info)
3092{
3093 struct amdgpu_vm *vm;
3094 unsigned long flags;
3095
3096 spin_lock_irqsave(&adev->vm_manager.pasid_lock, flags);
3097
3098 vm = idr_find(&adev->vm_manager.pasid_idr, pasid);
3099 if (vm)
3100 *task_info = vm->task_info;
3101
3102 spin_unlock_irqrestore(&adev->vm_manager.pasid_lock, flags);
3103}
3104
3105/**
3106 * amdgpu_vm_set_task_info - Sets VMs task info.
3107 *
3108 * @vm: vm for which to set the info
3109 */
3110void amdgpu_vm_set_task_info(struct amdgpu_vm *vm)
3111{
3112 if (!vm->task_info.pid) {
3113 vm->task_info.pid = current->pid;
3114 get_task_comm(vm->task_info.task_name, current);
3115
3116 if (current->group_leader->mm == current->mm) {
3117 vm->task_info.tgid = current->group_leader->pid;
3118 get_task_comm(vm->task_info.process_name, current->group_leader);
3119 }
3120 }
3121}