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
29 #include <linux/dma-fence-array.h>
30 #include <linux/interval_tree_generic.h>
31 #include <linux/idr.h>
32 #include <linux/dma-buf.h>
33
34 #include <drm/amdgpu_drm.h>
35 #include <drm/drm_drv.h>
36 #include "amdgpu.h"
37 #include "amdgpu_trace.h"
38 #include "amdgpu_amdkfd.h"
39 #include "amdgpu_gmc.h"
40 #include "amdgpu_xgmi.h"
41 #include "amdgpu_dma_buf.h"
42 #include "amdgpu_res_cursor.h"
43 #include "kfd_svm.h"
44
45 /**
46 * DOC: GPUVM
47 *
48 * GPUVM is similar to the legacy gart on older asics, however
49 * rather than there being a single global gart table
50 * for the entire GPU, there are multiple VM page tables active
51 * at any given time. The VM page tables can contain a mix
52 * vram pages and system memory pages and system memory pages
53 * can be mapped as snooped (cached system pages) or unsnooped
54 * (uncached system pages).
55 * Each VM has an ID associated with it and there is a page table
56 * associated with each VMID. When execting a command buffer,
57 * the kernel tells the the ring what VMID to use for that command
58 * buffer. VMIDs are allocated dynamically as commands are submitted.
59 * The userspace drivers maintain their own address space and the kernel
60 * sets up their pages tables accordingly when they submit their
61 * command buffers and a VMID is assigned.
62 * Cayman/Trinity support up to 8 active VMs at any given time;
63 * SI supports 16.
64 */
65
66 #define START(node) ((node)->start)
67 #define LAST(node) ((node)->last)
68
69 INTERVAL_TREE_DEFINE(struct amdgpu_bo_va_mapping, rb, uint64_t, __subtree_last,
70 START, LAST, static, amdgpu_vm_it)
71
72 #undef START
73 #undef LAST
74
75 /**
76 * struct amdgpu_prt_cb - Helper to disable partial resident texture feature from a fence callback
77 */
78 struct amdgpu_prt_cb {
79
80 /**
81 * @adev: amdgpu device
82 */
83 struct amdgpu_device *adev;
84
85 /**
86 * @cb: callback
87 */
88 struct dma_fence_cb cb;
89 };
90
91 /**
92 * amdgpu_vm_set_pasid - manage pasid and vm ptr mapping
93 *
94 * @adev: amdgpu_device pointer
95 * @vm: amdgpu_vm pointer
96 * @pasid: the pasid the VM is using on this GPU
97 *
98 * Set the pasid this VM is using on this GPU, can also be used to remove the
99 * pasid by passing in zero.
100 *
101 */
amdgpu_vm_set_pasid(struct amdgpu_device * adev,struct amdgpu_vm * vm,u32 pasid)102 int amdgpu_vm_set_pasid(struct amdgpu_device *adev, struct amdgpu_vm *vm,
103 u32 pasid)
104 {
105 int r;
106
107 if (vm->pasid == pasid)
108 return 0;
109
110 if (vm->pasid) {
111 r = xa_err(xa_erase_irq(&adev->vm_manager.pasids, vm->pasid));
112 if (r < 0)
113 return r;
114
115 vm->pasid = 0;
116 }
117
118 if (pasid) {
119 r = xa_err(xa_store_irq(&adev->vm_manager.pasids, pasid, vm,
120 GFP_KERNEL));
121 if (r < 0)
122 return r;
123
124 vm->pasid = pasid;
125 }
126
127
128 return 0;
129 }
130
131 /*
132 * vm eviction_lock can be taken in MMU notifiers. Make sure no reclaim-FS
133 * happens while holding this lock anywhere to prevent deadlocks when
134 * an MMU notifier runs in reclaim-FS context.
135 */
amdgpu_vm_eviction_lock(struct amdgpu_vm * vm)136 static inline void amdgpu_vm_eviction_lock(struct amdgpu_vm *vm)
137 {
138 mutex_lock(&vm->eviction_lock);
139 vm->saved_flags = memalloc_noreclaim_save();
140 }
141
amdgpu_vm_eviction_trylock(struct amdgpu_vm * vm)142 static inline int amdgpu_vm_eviction_trylock(struct amdgpu_vm *vm)
143 {
144 if (mutex_trylock(&vm->eviction_lock)) {
145 vm->saved_flags = memalloc_noreclaim_save();
146 return 1;
147 }
148 return 0;
149 }
150
amdgpu_vm_eviction_unlock(struct amdgpu_vm * vm)151 static inline void amdgpu_vm_eviction_unlock(struct amdgpu_vm *vm)
152 {
153 memalloc_noreclaim_restore(vm->saved_flags);
154 mutex_unlock(&vm->eviction_lock);
155 }
156
157 /**
158 * amdgpu_vm_level_shift - return the addr shift for each level
159 *
160 * @adev: amdgpu_device pointer
161 * @level: VMPT level
162 *
163 * Returns:
164 * The number of bits the pfn needs to be right shifted for a level.
165 */
amdgpu_vm_level_shift(struct amdgpu_device * adev,unsigned level)166 static unsigned amdgpu_vm_level_shift(struct amdgpu_device *adev,
167 unsigned level)
168 {
169 switch (level) {
170 case AMDGPU_VM_PDB2:
171 case AMDGPU_VM_PDB1:
172 case AMDGPU_VM_PDB0:
173 return 9 * (AMDGPU_VM_PDB0 - level) +
174 adev->vm_manager.block_size;
175 case AMDGPU_VM_PTB:
176 return 0;
177 default:
178 return ~0;
179 }
180 }
181
182 /**
183 * amdgpu_vm_num_entries - return the number of entries in a PD/PT
184 *
185 * @adev: amdgpu_device pointer
186 * @level: VMPT level
187 *
188 * Returns:
189 * The number of entries in a page directory or page table.
190 */
amdgpu_vm_num_entries(struct amdgpu_device * adev,unsigned level)191 static unsigned amdgpu_vm_num_entries(struct amdgpu_device *adev,
192 unsigned level)
193 {
194 unsigned shift = amdgpu_vm_level_shift(adev,
195 adev->vm_manager.root_level);
196
197 if (level == adev->vm_manager.root_level)
198 /* For the root directory */
199 return round_up(adev->vm_manager.max_pfn, 1ULL << shift)
200 >> shift;
201 else if (level != AMDGPU_VM_PTB)
202 /* Everything in between */
203 return 512;
204 else
205 /* For the page tables on the leaves */
206 return AMDGPU_VM_PTE_COUNT(adev);
207 }
208
209 /**
210 * amdgpu_vm_num_ats_entries - return the number of ATS entries in the root PD
211 *
212 * @adev: amdgpu_device pointer
213 *
214 * Returns:
215 * The number of entries in the root page directory which needs the ATS setting.
216 */
amdgpu_vm_num_ats_entries(struct amdgpu_device * adev)217 static unsigned amdgpu_vm_num_ats_entries(struct amdgpu_device *adev)
218 {
219 unsigned shift;
220
221 shift = amdgpu_vm_level_shift(adev, adev->vm_manager.root_level);
222 return AMDGPU_GMC_HOLE_START >> (shift + AMDGPU_GPU_PAGE_SHIFT);
223 }
224
225 /**
226 * amdgpu_vm_entries_mask - the mask to get the entry number of a PD/PT
227 *
228 * @adev: amdgpu_device pointer
229 * @level: VMPT level
230 *
231 * Returns:
232 * The mask to extract the entry number of a PD/PT from an address.
233 */
amdgpu_vm_entries_mask(struct amdgpu_device * adev,unsigned int level)234 static uint32_t amdgpu_vm_entries_mask(struct amdgpu_device *adev,
235 unsigned int level)
236 {
237 if (level <= adev->vm_manager.root_level)
238 return 0xffffffff;
239 else if (level != AMDGPU_VM_PTB)
240 return 0x1ff;
241 else
242 return AMDGPU_VM_PTE_COUNT(adev) - 1;
243 }
244
245 /**
246 * amdgpu_vm_bo_size - returns the size of the BOs in bytes
247 *
248 * @adev: amdgpu_device pointer
249 * @level: VMPT level
250 *
251 * Returns:
252 * The size of the BO for a page directory or page table in bytes.
253 */
amdgpu_vm_bo_size(struct amdgpu_device * adev,unsigned level)254 static unsigned amdgpu_vm_bo_size(struct amdgpu_device *adev, unsigned level)
255 {
256 return AMDGPU_GPU_PAGE_ALIGN(amdgpu_vm_num_entries(adev, level) * 8);
257 }
258
259 /**
260 * amdgpu_vm_bo_evicted - vm_bo is evicted
261 *
262 * @vm_bo: vm_bo which is evicted
263 *
264 * State for PDs/PTs and per VM BOs which are not at the location they should
265 * be.
266 */
amdgpu_vm_bo_evicted(struct amdgpu_vm_bo_base * vm_bo)267 static void amdgpu_vm_bo_evicted(struct amdgpu_vm_bo_base *vm_bo)
268 {
269 struct amdgpu_vm *vm = vm_bo->vm;
270 struct amdgpu_bo *bo = vm_bo->bo;
271
272 vm_bo->moved = true;
273 if (bo->tbo.type == ttm_bo_type_kernel)
274 list_move(&vm_bo->vm_status, &vm->evicted);
275 else
276 list_move_tail(&vm_bo->vm_status, &vm->evicted);
277 }
278 /**
279 * amdgpu_vm_bo_moved - vm_bo is moved
280 *
281 * @vm_bo: vm_bo which is moved
282 *
283 * State for per VM BOs which are moved, but that change is not yet reflected
284 * in the page tables.
285 */
amdgpu_vm_bo_moved(struct amdgpu_vm_bo_base * vm_bo)286 static void amdgpu_vm_bo_moved(struct amdgpu_vm_bo_base *vm_bo)
287 {
288 list_move(&vm_bo->vm_status, &vm_bo->vm->moved);
289 }
290
291 /**
292 * amdgpu_vm_bo_idle - vm_bo is idle
293 *
294 * @vm_bo: vm_bo which is now idle
295 *
296 * State for PDs/PTs and per VM BOs which have gone through the state machine
297 * and are now idle.
298 */
amdgpu_vm_bo_idle(struct amdgpu_vm_bo_base * vm_bo)299 static void amdgpu_vm_bo_idle(struct amdgpu_vm_bo_base *vm_bo)
300 {
301 list_move(&vm_bo->vm_status, &vm_bo->vm->idle);
302 vm_bo->moved = false;
303 }
304
305 /**
306 * amdgpu_vm_bo_invalidated - vm_bo is invalidated
307 *
308 * @vm_bo: vm_bo which is now invalidated
309 *
310 * State for normal BOs which are invalidated and that change not yet reflected
311 * in the PTs.
312 */
amdgpu_vm_bo_invalidated(struct amdgpu_vm_bo_base * vm_bo)313 static void amdgpu_vm_bo_invalidated(struct amdgpu_vm_bo_base *vm_bo)
314 {
315 spin_lock(&vm_bo->vm->invalidated_lock);
316 list_move(&vm_bo->vm_status, &vm_bo->vm->invalidated);
317 spin_unlock(&vm_bo->vm->invalidated_lock);
318 }
319
320 /**
321 * amdgpu_vm_bo_relocated - vm_bo is reloacted
322 *
323 * @vm_bo: vm_bo which is relocated
324 *
325 * State for PDs/PTs which needs to update their parent PD.
326 * For the root PD, just move to idle state.
327 */
amdgpu_vm_bo_relocated(struct amdgpu_vm_bo_base * vm_bo)328 static void amdgpu_vm_bo_relocated(struct amdgpu_vm_bo_base *vm_bo)
329 {
330 if (vm_bo->bo->parent)
331 list_move(&vm_bo->vm_status, &vm_bo->vm->relocated);
332 else
333 amdgpu_vm_bo_idle(vm_bo);
334 }
335
336 /**
337 * amdgpu_vm_bo_done - vm_bo is done
338 *
339 * @vm_bo: vm_bo which is now done
340 *
341 * State for normal BOs which are invalidated and that change has been updated
342 * in the PTs.
343 */
amdgpu_vm_bo_done(struct amdgpu_vm_bo_base * vm_bo)344 static void amdgpu_vm_bo_done(struct amdgpu_vm_bo_base *vm_bo)
345 {
346 spin_lock(&vm_bo->vm->invalidated_lock);
347 list_move(&vm_bo->vm_status, &vm_bo->vm->done);
348 spin_unlock(&vm_bo->vm->invalidated_lock);
349 }
350
351 /**
352 * amdgpu_vm_bo_base_init - Adds bo to the list of bos associated with the vm
353 *
354 * @base: base structure for tracking BO usage in a VM
355 * @vm: vm to which bo is to be added
356 * @bo: amdgpu buffer object
357 *
358 * Initialize a bo_va_base structure and add it to the appropriate lists
359 *
360 */
amdgpu_vm_bo_base_init(struct amdgpu_vm_bo_base * base,struct amdgpu_vm * vm,struct amdgpu_bo * bo)361 static void amdgpu_vm_bo_base_init(struct amdgpu_vm_bo_base *base,
362 struct amdgpu_vm *vm,
363 struct amdgpu_bo *bo)
364 {
365 base->vm = vm;
366 base->bo = bo;
367 base->next = NULL;
368 INIT_LIST_HEAD(&base->vm_status);
369
370 if (!bo)
371 return;
372 base->next = bo->vm_bo;
373 bo->vm_bo = base;
374
375 if (bo->tbo.base.resv != vm->root.bo->tbo.base.resv)
376 return;
377
378 vm->bulk_moveable = false;
379 if (bo->tbo.type == ttm_bo_type_kernel && bo->parent)
380 amdgpu_vm_bo_relocated(base);
381 else
382 amdgpu_vm_bo_idle(base);
383
384 if (bo->preferred_domains &
385 amdgpu_mem_type_to_domain(bo->tbo.resource->mem_type))
386 return;
387
388 /*
389 * we checked all the prerequisites, but it looks like this per vm bo
390 * is currently evicted. add the bo to the evicted list to make sure it
391 * is validated on next vm use to avoid fault.
392 * */
393 amdgpu_vm_bo_evicted(base);
394 }
395
396 /**
397 * amdgpu_vm_pt_parent - get the parent page directory
398 *
399 * @pt: child page table
400 *
401 * Helper to get the parent entry for the child page table. NULL if we are at
402 * the root page directory.
403 */
amdgpu_vm_pt_parent(struct amdgpu_vm_bo_base * pt)404 static struct amdgpu_vm_bo_base *amdgpu_vm_pt_parent(struct amdgpu_vm_bo_base *pt)
405 {
406 struct amdgpu_bo *parent = pt->bo->parent;
407
408 if (!parent)
409 return NULL;
410
411 return parent->vm_bo;
412 }
413
414 /*
415 * amdgpu_vm_pt_cursor - state for for_each_amdgpu_vm_pt
416 */
417 struct amdgpu_vm_pt_cursor {
418 uint64_t pfn;
419 struct amdgpu_vm_bo_base *parent;
420 struct amdgpu_vm_bo_base *entry;
421 unsigned level;
422 };
423
424 /**
425 * amdgpu_vm_pt_start - start PD/PT walk
426 *
427 * @adev: amdgpu_device pointer
428 * @vm: amdgpu_vm structure
429 * @start: start address of the walk
430 * @cursor: state to initialize
431 *
432 * Initialize a amdgpu_vm_pt_cursor to start a walk.
433 */
amdgpu_vm_pt_start(struct amdgpu_device * adev,struct amdgpu_vm * vm,uint64_t start,struct amdgpu_vm_pt_cursor * cursor)434 static void amdgpu_vm_pt_start(struct amdgpu_device *adev,
435 struct amdgpu_vm *vm, uint64_t start,
436 struct amdgpu_vm_pt_cursor *cursor)
437 {
438 cursor->pfn = start;
439 cursor->parent = NULL;
440 cursor->entry = &vm->root;
441 cursor->level = adev->vm_manager.root_level;
442 }
443
444 /**
445 * amdgpu_vm_pt_descendant - go to child node
446 *
447 * @adev: amdgpu_device pointer
448 * @cursor: current state
449 *
450 * Walk to the child node of the current node.
451 * Returns:
452 * True if the walk was possible, false otherwise.
453 */
amdgpu_vm_pt_descendant(struct amdgpu_device * adev,struct amdgpu_vm_pt_cursor * cursor)454 static bool amdgpu_vm_pt_descendant(struct amdgpu_device *adev,
455 struct amdgpu_vm_pt_cursor *cursor)
456 {
457 unsigned mask, shift, idx;
458
459 if ((cursor->level == AMDGPU_VM_PTB) || !cursor->entry ||
460 !cursor->entry->bo)
461 return false;
462
463 mask = amdgpu_vm_entries_mask(adev, cursor->level);
464 shift = amdgpu_vm_level_shift(adev, cursor->level);
465
466 ++cursor->level;
467 idx = (cursor->pfn >> shift) & mask;
468 cursor->parent = cursor->entry;
469 cursor->entry = &to_amdgpu_bo_vm(cursor->entry->bo)->entries[idx];
470 return true;
471 }
472
473 /**
474 * amdgpu_vm_pt_sibling - go to sibling node
475 *
476 * @adev: amdgpu_device pointer
477 * @cursor: current state
478 *
479 * Walk to the sibling node of the current node.
480 * Returns:
481 * True if the walk was possible, false otherwise.
482 */
amdgpu_vm_pt_sibling(struct amdgpu_device * adev,struct amdgpu_vm_pt_cursor * cursor)483 static bool amdgpu_vm_pt_sibling(struct amdgpu_device *adev,
484 struct amdgpu_vm_pt_cursor *cursor)
485 {
486 unsigned shift, num_entries;
487
488 /* Root doesn't have a sibling */
489 if (!cursor->parent)
490 return false;
491
492 /* Go to our parents and see if we got a sibling */
493 shift = amdgpu_vm_level_shift(adev, cursor->level - 1);
494 num_entries = amdgpu_vm_num_entries(adev, cursor->level - 1);
495
496 if (cursor->entry == &to_amdgpu_bo_vm(cursor->parent->bo)->entries[num_entries - 1])
497 return false;
498
499 cursor->pfn += 1ULL << shift;
500 cursor->pfn &= ~((1ULL << shift) - 1);
501 ++cursor->entry;
502 return true;
503 }
504
505 /**
506 * amdgpu_vm_pt_ancestor - go to parent node
507 *
508 * @cursor: current state
509 *
510 * Walk to the parent node of the current node.
511 * Returns:
512 * True if the walk was possible, false otherwise.
513 */
amdgpu_vm_pt_ancestor(struct amdgpu_vm_pt_cursor * cursor)514 static bool amdgpu_vm_pt_ancestor(struct amdgpu_vm_pt_cursor *cursor)
515 {
516 if (!cursor->parent)
517 return false;
518
519 --cursor->level;
520 cursor->entry = cursor->parent;
521 cursor->parent = amdgpu_vm_pt_parent(cursor->parent);
522 return true;
523 }
524
525 /**
526 * amdgpu_vm_pt_next - get next PD/PT in hieratchy
527 *
528 * @adev: amdgpu_device pointer
529 * @cursor: current state
530 *
531 * Walk the PD/PT tree to the next node.
532 */
amdgpu_vm_pt_next(struct amdgpu_device * adev,struct amdgpu_vm_pt_cursor * cursor)533 static void amdgpu_vm_pt_next(struct amdgpu_device *adev,
534 struct amdgpu_vm_pt_cursor *cursor)
535 {
536 /* First try a newborn child */
537 if (amdgpu_vm_pt_descendant(adev, cursor))
538 return;
539
540 /* If that didn't worked try to find a sibling */
541 while (!amdgpu_vm_pt_sibling(adev, cursor)) {
542 /* No sibling, go to our parents and grandparents */
543 if (!amdgpu_vm_pt_ancestor(cursor)) {
544 cursor->pfn = ~0ll;
545 return;
546 }
547 }
548 }
549
550 /**
551 * amdgpu_vm_pt_first_dfs - start a deep first search
552 *
553 * @adev: amdgpu_device structure
554 * @vm: amdgpu_vm structure
555 * @start: optional cursor to start with
556 * @cursor: state to initialize
557 *
558 * Starts a deep first traversal of the PD/PT tree.
559 */
amdgpu_vm_pt_first_dfs(struct amdgpu_device * adev,struct amdgpu_vm * vm,struct amdgpu_vm_pt_cursor * start,struct amdgpu_vm_pt_cursor * cursor)560 static void amdgpu_vm_pt_first_dfs(struct amdgpu_device *adev,
561 struct amdgpu_vm *vm,
562 struct amdgpu_vm_pt_cursor *start,
563 struct amdgpu_vm_pt_cursor *cursor)
564 {
565 if (start)
566 *cursor = *start;
567 else
568 amdgpu_vm_pt_start(adev, vm, 0, cursor);
569 while (amdgpu_vm_pt_descendant(adev, cursor));
570 }
571
572 /**
573 * amdgpu_vm_pt_continue_dfs - check if the deep first search should continue
574 *
575 * @start: starting point for the search
576 * @entry: current entry
577 *
578 * Returns:
579 * True when the search should continue, false otherwise.
580 */
amdgpu_vm_pt_continue_dfs(struct amdgpu_vm_pt_cursor * start,struct amdgpu_vm_bo_base * entry)581 static bool amdgpu_vm_pt_continue_dfs(struct amdgpu_vm_pt_cursor *start,
582 struct amdgpu_vm_bo_base *entry)
583 {
584 return entry && (!start || entry != start->entry);
585 }
586
587 /**
588 * amdgpu_vm_pt_next_dfs - get the next node for a deep first search
589 *
590 * @adev: amdgpu_device structure
591 * @cursor: current state
592 *
593 * Move the cursor to the next node in a deep first search.
594 */
amdgpu_vm_pt_next_dfs(struct amdgpu_device * adev,struct amdgpu_vm_pt_cursor * cursor)595 static void amdgpu_vm_pt_next_dfs(struct amdgpu_device *adev,
596 struct amdgpu_vm_pt_cursor *cursor)
597 {
598 if (!cursor->entry)
599 return;
600
601 if (!cursor->parent)
602 cursor->entry = NULL;
603 else if (amdgpu_vm_pt_sibling(adev, cursor))
604 while (amdgpu_vm_pt_descendant(adev, cursor));
605 else
606 amdgpu_vm_pt_ancestor(cursor);
607 }
608
609 /*
610 * for_each_amdgpu_vm_pt_dfs_safe - safe deep first search of all PDs/PTs
611 */
612 #define for_each_amdgpu_vm_pt_dfs_safe(adev, vm, start, cursor, entry) \
613 for (amdgpu_vm_pt_first_dfs((adev), (vm), (start), &(cursor)), \
614 (entry) = (cursor).entry, amdgpu_vm_pt_next_dfs((adev), &(cursor));\
615 amdgpu_vm_pt_continue_dfs((start), (entry)); \
616 (entry) = (cursor).entry, amdgpu_vm_pt_next_dfs((adev), &(cursor)))
617
618 /**
619 * amdgpu_vm_get_pd_bo - add the VM PD to a validation list
620 *
621 * @vm: vm providing the BOs
622 * @validated: head of validation list
623 * @entry: entry to add
624 *
625 * Add the page directory to the list of BOs to
626 * validate for command submission.
627 */
amdgpu_vm_get_pd_bo(struct amdgpu_vm * vm,struct list_head * validated,struct amdgpu_bo_list_entry * entry)628 void amdgpu_vm_get_pd_bo(struct amdgpu_vm *vm,
629 struct list_head *validated,
630 struct amdgpu_bo_list_entry *entry)
631 {
632 entry->priority = 0;
633 entry->tv.bo = &vm->root.bo->tbo;
634 /* Two for VM updates, one for TTM and one for the CS job */
635 entry->tv.num_shared = 4;
636 entry->user_pages = NULL;
637 list_add(&entry->tv.head, validated);
638 }
639
640 /**
641 * amdgpu_vm_del_from_lru_notify - update bulk_moveable flag
642 *
643 * @bo: BO which was removed from the LRU
644 *
645 * Make sure the bulk_moveable flag is updated when a BO is removed from the
646 * LRU.
647 */
amdgpu_vm_del_from_lru_notify(struct ttm_buffer_object * bo)648 void amdgpu_vm_del_from_lru_notify(struct ttm_buffer_object *bo)
649 {
650 struct amdgpu_bo *abo;
651 struct amdgpu_vm_bo_base *bo_base;
652
653 if (!amdgpu_bo_is_amdgpu_bo(bo))
654 return;
655
656 if (bo->pin_count)
657 return;
658
659 abo = ttm_to_amdgpu_bo(bo);
660 if (!abo->parent)
661 return;
662 for (bo_base = abo->vm_bo; bo_base; bo_base = bo_base->next) {
663 struct amdgpu_vm *vm = bo_base->vm;
664
665 if (abo->tbo.base.resv == vm->root.bo->tbo.base.resv)
666 vm->bulk_moveable = false;
667 }
668
669 }
670 /**
671 * amdgpu_vm_move_to_lru_tail - move all BOs to the end of LRU
672 *
673 * @adev: amdgpu device pointer
674 * @vm: vm providing the BOs
675 *
676 * Move all BOs to the end of LRU and remember their positions to put them
677 * together.
678 */
amdgpu_vm_move_to_lru_tail(struct amdgpu_device * adev,struct amdgpu_vm * vm)679 void amdgpu_vm_move_to_lru_tail(struct amdgpu_device *adev,
680 struct amdgpu_vm *vm)
681 {
682 struct amdgpu_vm_bo_base *bo_base;
683
684 if (vm->bulk_moveable) {
685 spin_lock(&adev->mman.bdev.lru_lock);
686 ttm_bo_bulk_move_lru_tail(&vm->lru_bulk_move);
687 spin_unlock(&adev->mman.bdev.lru_lock);
688 return;
689 }
690
691 memset(&vm->lru_bulk_move, 0, sizeof(vm->lru_bulk_move));
692
693 spin_lock(&adev->mman.bdev.lru_lock);
694 list_for_each_entry(bo_base, &vm->idle, vm_status) {
695 struct amdgpu_bo *bo = bo_base->bo;
696 struct amdgpu_bo *shadow = amdgpu_bo_shadowed(bo);
697
698 if (!bo->parent)
699 continue;
700
701 ttm_bo_move_to_lru_tail(&bo->tbo, bo->tbo.resource,
702 &vm->lru_bulk_move);
703 if (shadow)
704 ttm_bo_move_to_lru_tail(&shadow->tbo,
705 shadow->tbo.resource,
706 &vm->lru_bulk_move);
707 }
708 spin_unlock(&adev->mman.bdev.lru_lock);
709
710 vm->bulk_moveable = true;
711 }
712
713 /**
714 * amdgpu_vm_validate_pt_bos - validate the page table BOs
715 *
716 * @adev: amdgpu device pointer
717 * @vm: vm providing the BOs
718 * @validate: callback to do the validation
719 * @param: parameter for the validation callback
720 *
721 * Validate the page table BOs on command submission if neccessary.
722 *
723 * Returns:
724 * Validation result.
725 */
amdgpu_vm_validate_pt_bos(struct amdgpu_device * adev,struct amdgpu_vm * vm,int (* validate)(void * p,struct amdgpu_bo * bo),void * param)726 int amdgpu_vm_validate_pt_bos(struct amdgpu_device *adev, struct amdgpu_vm *vm,
727 int (*validate)(void *p, struct amdgpu_bo *bo),
728 void *param)
729 {
730 struct amdgpu_vm_bo_base *bo_base, *tmp;
731 int r;
732
733 vm->bulk_moveable &= list_empty(&vm->evicted);
734
735 list_for_each_entry_safe(bo_base, tmp, &vm->evicted, vm_status) {
736 struct amdgpu_bo *bo = bo_base->bo;
737 struct amdgpu_bo *shadow = amdgpu_bo_shadowed(bo);
738
739 r = validate(param, bo);
740 if (r)
741 return r;
742 if (shadow) {
743 r = validate(param, shadow);
744 if (r)
745 return r;
746 }
747
748 if (bo->tbo.type != ttm_bo_type_kernel) {
749 amdgpu_vm_bo_moved(bo_base);
750 } else {
751 vm->update_funcs->map_table(to_amdgpu_bo_vm(bo));
752 amdgpu_vm_bo_relocated(bo_base);
753 }
754 }
755
756 amdgpu_vm_eviction_lock(vm);
757 vm->evicting = false;
758 amdgpu_vm_eviction_unlock(vm);
759
760 return 0;
761 }
762
763 /**
764 * amdgpu_vm_ready - check VM is ready for updates
765 *
766 * @vm: VM to check
767 *
768 * Check if all VM PDs/PTs are ready for updates
769 *
770 * Returns:
771 * True if eviction list is empty.
772 */
amdgpu_vm_ready(struct amdgpu_vm * vm)773 bool amdgpu_vm_ready(struct amdgpu_vm *vm)
774 {
775 return list_empty(&vm->evicted);
776 }
777
778 /**
779 * amdgpu_vm_clear_bo - initially clear the PDs/PTs
780 *
781 * @adev: amdgpu_device pointer
782 * @vm: VM to clear BO from
783 * @vmbo: BO to clear
784 * @immediate: use an immediate update
785 *
786 * Root PD needs to be reserved when calling this.
787 *
788 * Returns:
789 * 0 on success, errno otherwise.
790 */
amdgpu_vm_clear_bo(struct amdgpu_device * adev,struct amdgpu_vm * vm,struct amdgpu_bo_vm * vmbo,bool immediate)791 static int amdgpu_vm_clear_bo(struct amdgpu_device *adev,
792 struct amdgpu_vm *vm,
793 struct amdgpu_bo_vm *vmbo,
794 bool immediate)
795 {
796 struct ttm_operation_ctx ctx = { true, false };
797 unsigned level = adev->vm_manager.root_level;
798 struct amdgpu_vm_update_params params;
799 struct amdgpu_bo *ancestor = &vmbo->bo;
800 struct amdgpu_bo *bo = &vmbo->bo;
801 unsigned entries, ats_entries;
802 uint64_t addr;
803 int r;
804
805 /* Figure out our place in the hierarchy */
806 if (ancestor->parent) {
807 ++level;
808 while (ancestor->parent->parent) {
809 ++level;
810 ancestor = ancestor->parent;
811 }
812 }
813
814 entries = amdgpu_bo_size(bo) / 8;
815 if (!vm->pte_support_ats) {
816 ats_entries = 0;
817
818 } else if (!bo->parent) {
819 ats_entries = amdgpu_vm_num_ats_entries(adev);
820 ats_entries = min(ats_entries, entries);
821 entries -= ats_entries;
822
823 } else {
824 struct amdgpu_vm_bo_base *pt;
825
826 pt = ancestor->vm_bo;
827 ats_entries = amdgpu_vm_num_ats_entries(adev);
828 if ((pt - to_amdgpu_bo_vm(vm->root.bo)->entries) >= ats_entries) {
829 ats_entries = 0;
830 } else {
831 ats_entries = entries;
832 entries = 0;
833 }
834 }
835
836 r = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
837 if (r)
838 return r;
839
840 if (vmbo->shadow) {
841 struct amdgpu_bo *shadow = vmbo->shadow;
842
843 r = ttm_bo_validate(&shadow->tbo, &shadow->placement, &ctx);
844 if (r)
845 return r;
846 }
847
848 r = vm->update_funcs->map_table(vmbo);
849 if (r)
850 return r;
851
852 memset(¶ms, 0, sizeof(params));
853 params.adev = adev;
854 params.vm = vm;
855 params.immediate = immediate;
856
857 r = vm->update_funcs->prepare(¶ms, NULL, AMDGPU_SYNC_EXPLICIT);
858 if (r)
859 return r;
860
861 addr = 0;
862 if (ats_entries) {
863 uint64_t value = 0, flags;
864
865 flags = AMDGPU_PTE_DEFAULT_ATC;
866 if (level != AMDGPU_VM_PTB) {
867 /* Handle leaf PDEs as PTEs */
868 flags |= AMDGPU_PDE_PTE;
869 amdgpu_gmc_get_vm_pde(adev, level, &value, &flags);
870 }
871
872 r = vm->update_funcs->update(¶ms, vmbo, addr, 0, ats_entries,
873 value, flags);
874 if (r)
875 return r;
876
877 addr += ats_entries * 8;
878 }
879
880 if (entries) {
881 uint64_t value = 0, flags = 0;
882
883 if (adev->asic_type >= CHIP_VEGA10) {
884 if (level != AMDGPU_VM_PTB) {
885 /* Handle leaf PDEs as PTEs */
886 flags |= AMDGPU_PDE_PTE;
887 amdgpu_gmc_get_vm_pde(adev, level,
888 &value, &flags);
889 } else {
890 /* Workaround for fault priority problem on GMC9 */
891 flags = AMDGPU_PTE_EXECUTABLE;
892 }
893 }
894
895 r = vm->update_funcs->update(¶ms, vmbo, addr, 0, entries,
896 value, flags);
897 if (r)
898 return r;
899 }
900
901 return vm->update_funcs->commit(¶ms, NULL);
902 }
903
904 /**
905 * amdgpu_vm_pt_create - create bo for PD/PT
906 *
907 * @adev: amdgpu_device pointer
908 * @vm: requesting vm
909 * @level: the page table level
910 * @immediate: use a immediate update
911 * @vmbo: pointer to the buffer object pointer
912 */
amdgpu_vm_pt_create(struct amdgpu_device * adev,struct amdgpu_vm * vm,int level,bool immediate,struct amdgpu_bo_vm ** vmbo)913 static int amdgpu_vm_pt_create(struct amdgpu_device *adev,
914 struct amdgpu_vm *vm,
915 int level, bool immediate,
916 struct amdgpu_bo_vm **vmbo)
917 {
918 struct amdgpu_bo_param bp;
919 struct amdgpu_bo *bo;
920 struct dma_resv *resv;
921 unsigned int num_entries;
922 int r;
923
924 memset(&bp, 0, sizeof(bp));
925
926 bp.size = amdgpu_vm_bo_size(adev, level);
927 bp.byte_align = AMDGPU_GPU_PAGE_SIZE;
928 bp.domain = AMDGPU_GEM_DOMAIN_VRAM;
929 bp.domain = amdgpu_bo_get_preferred_domain(adev, bp.domain);
930 bp.flags = AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS |
931 AMDGPU_GEM_CREATE_CPU_GTT_USWC;
932
933 if (level < AMDGPU_VM_PTB)
934 num_entries = amdgpu_vm_num_entries(adev, level);
935 else
936 num_entries = 0;
937
938 bp.bo_ptr_size = struct_size((*vmbo), entries, num_entries);
939
940 if (vm->use_cpu_for_update)
941 bp.flags |= AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED;
942
943 bp.type = ttm_bo_type_kernel;
944 bp.no_wait_gpu = immediate;
945 if (vm->root.bo)
946 bp.resv = vm->root.bo->tbo.base.resv;
947
948 r = amdgpu_bo_create_vm(adev, &bp, vmbo);
949 if (r)
950 return r;
951
952 bo = &(*vmbo)->bo;
953 if (vm->is_compute_context || (adev->flags & AMD_IS_APU)) {
954 (*vmbo)->shadow = NULL;
955 return 0;
956 }
957
958 if (!bp.resv)
959 WARN_ON(dma_resv_lock(bo->tbo.base.resv,
960 NULL));
961 resv = bp.resv;
962 memset(&bp, 0, sizeof(bp));
963 bp.size = amdgpu_vm_bo_size(adev, level);
964 bp.domain = AMDGPU_GEM_DOMAIN_GTT;
965 bp.flags = AMDGPU_GEM_CREATE_CPU_GTT_USWC;
966 bp.type = ttm_bo_type_kernel;
967 bp.resv = bo->tbo.base.resv;
968 bp.bo_ptr_size = sizeof(struct amdgpu_bo);
969
970 r = amdgpu_bo_create(adev, &bp, &(*vmbo)->shadow);
971
972 if (!resv)
973 dma_resv_unlock(bo->tbo.base.resv);
974
975 if (r) {
976 amdgpu_bo_unref(&bo);
977 return r;
978 }
979
980 (*vmbo)->shadow->parent = amdgpu_bo_ref(bo);
981 amdgpu_bo_add_to_shadow_list(*vmbo);
982
983 return 0;
984 }
985
986 /**
987 * amdgpu_vm_alloc_pts - Allocate a specific page table
988 *
989 * @adev: amdgpu_device pointer
990 * @vm: VM to allocate page tables for
991 * @cursor: Which page table to allocate
992 * @immediate: use an immediate update
993 *
994 * Make sure a specific page table or directory is allocated.
995 *
996 * Returns:
997 * 1 if page table needed to be allocated, 0 if page table was already
998 * allocated, negative errno if an error occurred.
999 */
amdgpu_vm_alloc_pts(struct amdgpu_device * adev,struct amdgpu_vm * vm,struct amdgpu_vm_pt_cursor * cursor,bool immediate)1000 static int amdgpu_vm_alloc_pts(struct amdgpu_device *adev,
1001 struct amdgpu_vm *vm,
1002 struct amdgpu_vm_pt_cursor *cursor,
1003 bool immediate)
1004 {
1005 struct amdgpu_vm_bo_base *entry = cursor->entry;
1006 struct amdgpu_bo *pt_bo;
1007 struct amdgpu_bo_vm *pt;
1008 int r;
1009
1010 if (entry->bo)
1011 return 0;
1012
1013 r = amdgpu_vm_pt_create(adev, vm, cursor->level, immediate, &pt);
1014 if (r)
1015 return r;
1016
1017 /* Keep a reference to the root directory to avoid
1018 * freeing them up in the wrong order.
1019 */
1020 pt_bo = &pt->bo;
1021 pt_bo->parent = amdgpu_bo_ref(cursor->parent->bo);
1022 amdgpu_vm_bo_base_init(entry, vm, pt_bo);
1023 r = amdgpu_vm_clear_bo(adev, vm, pt, immediate);
1024 if (r)
1025 goto error_free_pt;
1026
1027 return 0;
1028
1029 error_free_pt:
1030 amdgpu_bo_unref(&pt->shadow);
1031 amdgpu_bo_unref(&pt_bo);
1032 return r;
1033 }
1034
1035 /**
1036 * amdgpu_vm_free_table - fre one PD/PT
1037 *
1038 * @entry: PDE to free
1039 */
amdgpu_vm_free_table(struct amdgpu_vm_bo_base * entry)1040 static void amdgpu_vm_free_table(struct amdgpu_vm_bo_base *entry)
1041 {
1042 struct amdgpu_bo *shadow;
1043
1044 if (!entry->bo)
1045 return;
1046 shadow = amdgpu_bo_shadowed(entry->bo);
1047 entry->bo->vm_bo = NULL;
1048 list_del(&entry->vm_status);
1049 amdgpu_bo_unref(&shadow);
1050 amdgpu_bo_unref(&entry->bo);
1051 }
1052
1053 /**
1054 * amdgpu_vm_free_pts - free PD/PT levels
1055 *
1056 * @adev: amdgpu device structure
1057 * @vm: amdgpu vm structure
1058 * @start: optional cursor where to start freeing PDs/PTs
1059 *
1060 * Free the page directory or page table level and all sub levels.
1061 */
amdgpu_vm_free_pts(struct amdgpu_device * adev,struct amdgpu_vm * vm,struct amdgpu_vm_pt_cursor * start)1062 static void amdgpu_vm_free_pts(struct amdgpu_device *adev,
1063 struct amdgpu_vm *vm,
1064 struct amdgpu_vm_pt_cursor *start)
1065 {
1066 struct amdgpu_vm_pt_cursor cursor;
1067 struct amdgpu_vm_bo_base *entry;
1068
1069 vm->bulk_moveable = false;
1070
1071 for_each_amdgpu_vm_pt_dfs_safe(adev, vm, start, cursor, entry)
1072 amdgpu_vm_free_table(entry);
1073
1074 if (start)
1075 amdgpu_vm_free_table(start->entry);
1076 }
1077
1078 /**
1079 * amdgpu_vm_check_compute_bug - check whether asic has compute vm bug
1080 *
1081 * @adev: amdgpu_device pointer
1082 */
amdgpu_vm_check_compute_bug(struct amdgpu_device * adev)1083 void amdgpu_vm_check_compute_bug(struct amdgpu_device *adev)
1084 {
1085 const struct amdgpu_ip_block *ip_block;
1086 bool has_compute_vm_bug;
1087 struct amdgpu_ring *ring;
1088 int i;
1089
1090 has_compute_vm_bug = false;
1091
1092 ip_block = amdgpu_device_ip_get_ip_block(adev, AMD_IP_BLOCK_TYPE_GFX);
1093 if (ip_block) {
1094 /* Compute has a VM bug for GFX version < 7.
1095 Compute has a VM bug for GFX 8 MEC firmware version < 673.*/
1096 if (ip_block->version->major <= 7)
1097 has_compute_vm_bug = true;
1098 else if (ip_block->version->major == 8)
1099 if (adev->gfx.mec_fw_version < 673)
1100 has_compute_vm_bug = true;
1101 }
1102
1103 for (i = 0; i < adev->num_rings; i++) {
1104 ring = adev->rings[i];
1105 if (ring->funcs->type == AMDGPU_RING_TYPE_COMPUTE)
1106 /* only compute rings */
1107 ring->has_compute_vm_bug = has_compute_vm_bug;
1108 else
1109 ring->has_compute_vm_bug = false;
1110 }
1111 }
1112
1113 /**
1114 * amdgpu_vm_need_pipeline_sync - Check if pipe sync is needed for job.
1115 *
1116 * @ring: ring on which the job will be submitted
1117 * @job: job to submit
1118 *
1119 * Returns:
1120 * True if sync is needed.
1121 */
amdgpu_vm_need_pipeline_sync(struct amdgpu_ring * ring,struct amdgpu_job * job)1122 bool amdgpu_vm_need_pipeline_sync(struct amdgpu_ring *ring,
1123 struct amdgpu_job *job)
1124 {
1125 struct amdgpu_device *adev = ring->adev;
1126 unsigned vmhub = ring->funcs->vmhub;
1127 struct amdgpu_vmid_mgr *id_mgr = &adev->vm_manager.id_mgr[vmhub];
1128 struct amdgpu_vmid *id;
1129 bool gds_switch_needed;
1130 bool vm_flush_needed = job->vm_needs_flush || ring->has_compute_vm_bug;
1131
1132 if (job->vmid == 0)
1133 return false;
1134 id = &id_mgr->ids[job->vmid];
1135 gds_switch_needed = ring->funcs->emit_gds_switch && (
1136 id->gds_base != job->gds_base ||
1137 id->gds_size != job->gds_size ||
1138 id->gws_base != job->gws_base ||
1139 id->gws_size != job->gws_size ||
1140 id->oa_base != job->oa_base ||
1141 id->oa_size != job->oa_size);
1142
1143 if (amdgpu_vmid_had_gpu_reset(adev, id))
1144 return true;
1145
1146 return vm_flush_needed || gds_switch_needed;
1147 }
1148
1149 /**
1150 * amdgpu_vm_flush - hardware flush the vm
1151 *
1152 * @ring: ring to use for flush
1153 * @job: related job
1154 * @need_pipe_sync: is pipe sync needed
1155 *
1156 * Emit a VM flush when it is necessary.
1157 *
1158 * Returns:
1159 * 0 on success, errno otherwise.
1160 */
amdgpu_vm_flush(struct amdgpu_ring * ring,struct amdgpu_job * job,bool need_pipe_sync)1161 int amdgpu_vm_flush(struct amdgpu_ring *ring, struct amdgpu_job *job,
1162 bool need_pipe_sync)
1163 {
1164 struct amdgpu_device *adev = ring->adev;
1165 unsigned vmhub = ring->funcs->vmhub;
1166 struct amdgpu_vmid_mgr *id_mgr = &adev->vm_manager.id_mgr[vmhub];
1167 struct amdgpu_vmid *id = &id_mgr->ids[job->vmid];
1168 bool gds_switch_needed = ring->funcs->emit_gds_switch && (
1169 id->gds_base != job->gds_base ||
1170 id->gds_size != job->gds_size ||
1171 id->gws_base != job->gws_base ||
1172 id->gws_size != job->gws_size ||
1173 id->oa_base != job->oa_base ||
1174 id->oa_size != job->oa_size);
1175 bool vm_flush_needed = job->vm_needs_flush;
1176 struct dma_fence *fence = NULL;
1177 bool pasid_mapping_needed = false;
1178 unsigned patch_offset = 0;
1179 bool update_spm_vmid_needed = (job->vm && (job->vm->reserved_vmid[vmhub] != NULL));
1180 int r;
1181
1182 if (update_spm_vmid_needed && adev->gfx.rlc.funcs->update_spm_vmid)
1183 adev->gfx.rlc.funcs->update_spm_vmid(adev, job->vmid);
1184
1185 if (amdgpu_vmid_had_gpu_reset(adev, id)) {
1186 gds_switch_needed = true;
1187 vm_flush_needed = true;
1188 pasid_mapping_needed = true;
1189 }
1190
1191 mutex_lock(&id_mgr->lock);
1192 if (id->pasid != job->pasid || !id->pasid_mapping ||
1193 !dma_fence_is_signaled(id->pasid_mapping))
1194 pasid_mapping_needed = true;
1195 mutex_unlock(&id_mgr->lock);
1196
1197 gds_switch_needed &= !!ring->funcs->emit_gds_switch;
1198 vm_flush_needed &= !!ring->funcs->emit_vm_flush &&
1199 job->vm_pd_addr != AMDGPU_BO_INVALID_OFFSET;
1200 pasid_mapping_needed &= adev->gmc.gmc_funcs->emit_pasid_mapping &&
1201 ring->funcs->emit_wreg;
1202
1203 if (!vm_flush_needed && !gds_switch_needed && !need_pipe_sync)
1204 return 0;
1205
1206 if (ring->funcs->init_cond_exec)
1207 patch_offset = amdgpu_ring_init_cond_exec(ring);
1208
1209 if (need_pipe_sync)
1210 amdgpu_ring_emit_pipeline_sync(ring);
1211
1212 if (vm_flush_needed) {
1213 trace_amdgpu_vm_flush(ring, job->vmid, job->vm_pd_addr);
1214 amdgpu_ring_emit_vm_flush(ring, job->vmid, job->vm_pd_addr);
1215 }
1216
1217 if (pasid_mapping_needed)
1218 amdgpu_gmc_emit_pasid_mapping(ring, job->vmid, job->pasid);
1219
1220 if (vm_flush_needed || pasid_mapping_needed) {
1221 r = amdgpu_fence_emit(ring, &fence, NULL, 0);
1222 if (r)
1223 return r;
1224 }
1225
1226 if (vm_flush_needed) {
1227 mutex_lock(&id_mgr->lock);
1228 dma_fence_put(id->last_flush);
1229 id->last_flush = dma_fence_get(fence);
1230 id->current_gpu_reset_count =
1231 atomic_read(&adev->gpu_reset_counter);
1232 mutex_unlock(&id_mgr->lock);
1233 }
1234
1235 if (pasid_mapping_needed) {
1236 mutex_lock(&id_mgr->lock);
1237 id->pasid = job->pasid;
1238 dma_fence_put(id->pasid_mapping);
1239 id->pasid_mapping = dma_fence_get(fence);
1240 mutex_unlock(&id_mgr->lock);
1241 }
1242 dma_fence_put(fence);
1243
1244 if (ring->funcs->emit_gds_switch && gds_switch_needed) {
1245 id->gds_base = job->gds_base;
1246 id->gds_size = job->gds_size;
1247 id->gws_base = job->gws_base;
1248 id->gws_size = job->gws_size;
1249 id->oa_base = job->oa_base;
1250 id->oa_size = job->oa_size;
1251 amdgpu_ring_emit_gds_switch(ring, job->vmid, job->gds_base,
1252 job->gds_size, job->gws_base,
1253 job->gws_size, job->oa_base,
1254 job->oa_size);
1255 }
1256
1257 if (ring->funcs->patch_cond_exec)
1258 amdgpu_ring_patch_cond_exec(ring, patch_offset);
1259
1260 /* the double SWITCH_BUFFER here *cannot* be skipped by COND_EXEC */
1261 if (ring->funcs->emit_switch_buffer) {
1262 amdgpu_ring_emit_switch_buffer(ring);
1263 amdgpu_ring_emit_switch_buffer(ring);
1264 }
1265 return 0;
1266 }
1267
1268 /**
1269 * amdgpu_vm_bo_find - find the bo_va for a specific vm & bo
1270 *
1271 * @vm: requested vm
1272 * @bo: requested buffer object
1273 *
1274 * Find @bo inside the requested vm.
1275 * Search inside the @bos vm list for the requested vm
1276 * Returns the found bo_va or NULL if none is found
1277 *
1278 * Object has to be reserved!
1279 *
1280 * Returns:
1281 * Found bo_va or NULL.
1282 */
amdgpu_vm_bo_find(struct amdgpu_vm * vm,struct amdgpu_bo * bo)1283 struct amdgpu_bo_va *amdgpu_vm_bo_find(struct amdgpu_vm *vm,
1284 struct amdgpu_bo *bo)
1285 {
1286 struct amdgpu_vm_bo_base *base;
1287
1288 for (base = bo->vm_bo; base; base = base->next) {
1289 if (base->vm != vm)
1290 continue;
1291
1292 return container_of(base, struct amdgpu_bo_va, base);
1293 }
1294 return NULL;
1295 }
1296
1297 /**
1298 * amdgpu_vm_map_gart - Resolve gart mapping of addr
1299 *
1300 * @pages_addr: optional DMA address to use for lookup
1301 * @addr: the unmapped addr
1302 *
1303 * Look up the physical address of the page that the pte resolves
1304 * to.
1305 *
1306 * Returns:
1307 * The pointer for the page table entry.
1308 */
amdgpu_vm_map_gart(const dma_addr_t * pages_addr,uint64_t addr)1309 uint64_t amdgpu_vm_map_gart(const dma_addr_t *pages_addr, uint64_t addr)
1310 {
1311 uint64_t result;
1312
1313 /* page table offset */
1314 result = pages_addr[addr >> PAGE_SHIFT];
1315
1316 /* in case cpu page size != gpu page size*/
1317 result |= addr & (~PAGE_MASK);
1318
1319 result &= 0xFFFFFFFFFFFFF000ULL;
1320
1321 return result;
1322 }
1323
1324 /**
1325 * amdgpu_vm_update_pde - update a single level in the hierarchy
1326 *
1327 * @params: parameters for the update
1328 * @vm: requested vm
1329 * @entry: entry to update
1330 *
1331 * Makes sure the requested entry in parent is up to date.
1332 */
amdgpu_vm_update_pde(struct amdgpu_vm_update_params * params,struct amdgpu_vm * vm,struct amdgpu_vm_bo_base * entry)1333 static int amdgpu_vm_update_pde(struct amdgpu_vm_update_params *params,
1334 struct amdgpu_vm *vm,
1335 struct amdgpu_vm_bo_base *entry)
1336 {
1337 struct amdgpu_vm_bo_base *parent = amdgpu_vm_pt_parent(entry);
1338 struct amdgpu_bo *bo = parent->bo, *pbo;
1339 uint64_t pde, pt, flags;
1340 unsigned level;
1341
1342 for (level = 0, pbo = bo->parent; pbo; ++level)
1343 pbo = pbo->parent;
1344
1345 level += params->adev->vm_manager.root_level;
1346 amdgpu_gmc_get_pde_for_bo(entry->bo, level, &pt, &flags);
1347 pde = (entry - to_amdgpu_bo_vm(parent->bo)->entries) * 8;
1348 return vm->update_funcs->update(params, to_amdgpu_bo_vm(bo), pde, pt,
1349 1, 0, flags);
1350 }
1351
1352 /**
1353 * amdgpu_vm_invalidate_pds - mark all PDs as invalid
1354 *
1355 * @adev: amdgpu_device pointer
1356 * @vm: related vm
1357 *
1358 * Mark all PD level as invalid after an error.
1359 */
amdgpu_vm_invalidate_pds(struct amdgpu_device * adev,struct amdgpu_vm * vm)1360 static void amdgpu_vm_invalidate_pds(struct amdgpu_device *adev,
1361 struct amdgpu_vm *vm)
1362 {
1363 struct amdgpu_vm_pt_cursor cursor;
1364 struct amdgpu_vm_bo_base *entry;
1365
1366 for_each_amdgpu_vm_pt_dfs_safe(adev, vm, NULL, cursor, entry)
1367 if (entry->bo && !entry->moved)
1368 amdgpu_vm_bo_relocated(entry);
1369 }
1370
1371 /**
1372 * amdgpu_vm_update_pdes - make sure that all directories are valid
1373 *
1374 * @adev: amdgpu_device pointer
1375 * @vm: requested vm
1376 * @immediate: submit immediately to the paging queue
1377 *
1378 * Makes sure all directories are up to date.
1379 *
1380 * Returns:
1381 * 0 for success, error for failure.
1382 */
amdgpu_vm_update_pdes(struct amdgpu_device * adev,struct amdgpu_vm * vm,bool immediate)1383 int amdgpu_vm_update_pdes(struct amdgpu_device *adev,
1384 struct amdgpu_vm *vm, bool immediate)
1385 {
1386 struct amdgpu_vm_update_params params;
1387 int r;
1388
1389 if (list_empty(&vm->relocated))
1390 return 0;
1391
1392 memset(¶ms, 0, sizeof(params));
1393 params.adev = adev;
1394 params.vm = vm;
1395 params.immediate = immediate;
1396
1397 r = vm->update_funcs->prepare(¶ms, NULL, AMDGPU_SYNC_EXPLICIT);
1398 if (r)
1399 return r;
1400
1401 while (!list_empty(&vm->relocated)) {
1402 struct amdgpu_vm_bo_base *entry;
1403
1404 entry = list_first_entry(&vm->relocated,
1405 struct amdgpu_vm_bo_base,
1406 vm_status);
1407 amdgpu_vm_bo_idle(entry);
1408
1409 r = amdgpu_vm_update_pde(¶ms, vm, entry);
1410 if (r)
1411 goto error;
1412 }
1413
1414 r = vm->update_funcs->commit(¶ms, &vm->last_update);
1415 if (r)
1416 goto error;
1417 return 0;
1418
1419 error:
1420 amdgpu_vm_invalidate_pds(adev, vm);
1421 return r;
1422 }
1423
1424 /*
1425 * amdgpu_vm_update_flags - figure out flags for PTE updates
1426 *
1427 * Make sure to set the right flags for the PTEs at the desired level.
1428 */
amdgpu_vm_update_flags(struct amdgpu_vm_update_params * params,struct amdgpu_bo_vm * pt,unsigned int level,uint64_t pe,uint64_t addr,unsigned int count,uint32_t incr,uint64_t flags)1429 static void amdgpu_vm_update_flags(struct amdgpu_vm_update_params *params,
1430 struct amdgpu_bo_vm *pt, unsigned int level,
1431 uint64_t pe, uint64_t addr,
1432 unsigned int count, uint32_t incr,
1433 uint64_t flags)
1434
1435 {
1436 if (level != AMDGPU_VM_PTB) {
1437 flags |= AMDGPU_PDE_PTE;
1438 amdgpu_gmc_get_vm_pde(params->adev, level, &addr, &flags);
1439
1440 } else if (params->adev->asic_type >= CHIP_VEGA10 &&
1441 !(flags & AMDGPU_PTE_VALID) &&
1442 !(flags & AMDGPU_PTE_PRT)) {
1443
1444 /* Workaround for fault priority problem on GMC9 */
1445 flags |= AMDGPU_PTE_EXECUTABLE;
1446 }
1447
1448 params->vm->update_funcs->update(params, pt, pe, addr, count, incr,
1449 flags);
1450 }
1451
1452 /**
1453 * amdgpu_vm_fragment - get fragment for PTEs
1454 *
1455 * @params: see amdgpu_vm_update_params definition
1456 * @start: first PTE to handle
1457 * @end: last PTE to handle
1458 * @flags: hw mapping flags
1459 * @frag: resulting fragment size
1460 * @frag_end: end of this fragment
1461 *
1462 * Returns the first possible fragment for the start and end address.
1463 */
amdgpu_vm_fragment(struct amdgpu_vm_update_params * params,uint64_t start,uint64_t end,uint64_t flags,unsigned int * frag,uint64_t * frag_end)1464 static void amdgpu_vm_fragment(struct amdgpu_vm_update_params *params,
1465 uint64_t start, uint64_t end, uint64_t flags,
1466 unsigned int *frag, uint64_t *frag_end)
1467 {
1468 /**
1469 * The MC L1 TLB supports variable sized pages, based on a fragment
1470 * field in the PTE. When this field is set to a non-zero value, page
1471 * granularity is increased from 4KB to (1 << (12 + frag)). The PTE
1472 * flags are considered valid for all PTEs within the fragment range
1473 * and corresponding mappings are assumed to be physically contiguous.
1474 *
1475 * The L1 TLB can store a single PTE for the whole fragment,
1476 * significantly increasing the space available for translation
1477 * caching. This leads to large improvements in throughput when the
1478 * TLB is under pressure.
1479 *
1480 * The L2 TLB distributes small and large fragments into two
1481 * asymmetric partitions. The large fragment cache is significantly
1482 * larger. Thus, we try to use large fragments wherever possible.
1483 * Userspace can support this by aligning virtual base address and
1484 * allocation size to the fragment size.
1485 *
1486 * Starting with Vega10 the fragment size only controls the L1. The L2
1487 * is now directly feed with small/huge/giant pages from the walker.
1488 */
1489 unsigned max_frag;
1490
1491 if (params->adev->asic_type < CHIP_VEGA10)
1492 max_frag = params->adev->vm_manager.fragment_size;
1493 else
1494 max_frag = 31;
1495
1496 /* system pages are non continuously */
1497 if (params->pages_addr) {
1498 *frag = 0;
1499 *frag_end = end;
1500 return;
1501 }
1502
1503 /* This intentionally wraps around if no bit is set */
1504 *frag = min((unsigned)ffs(start) - 1, (unsigned)fls64(end - start) - 1);
1505 if (*frag >= max_frag) {
1506 *frag = max_frag;
1507 *frag_end = end & ~((1ULL << max_frag) - 1);
1508 } else {
1509 *frag_end = start + (1 << *frag);
1510 }
1511 }
1512
1513 /**
1514 * amdgpu_vm_update_ptes - make sure that page tables are valid
1515 *
1516 * @params: see amdgpu_vm_update_params definition
1517 * @start: start of GPU address range
1518 * @end: end of GPU address range
1519 * @dst: destination address to map to, the next dst inside the function
1520 * @flags: mapping flags
1521 *
1522 * Update the page tables in the range @start - @end.
1523 *
1524 * Returns:
1525 * 0 for success, -EINVAL for failure.
1526 */
amdgpu_vm_update_ptes(struct amdgpu_vm_update_params * params,uint64_t start,uint64_t end,uint64_t dst,uint64_t flags)1527 static int amdgpu_vm_update_ptes(struct amdgpu_vm_update_params *params,
1528 uint64_t start, uint64_t end,
1529 uint64_t dst, uint64_t flags)
1530 {
1531 struct amdgpu_device *adev = params->adev;
1532 struct amdgpu_vm_pt_cursor cursor;
1533 uint64_t frag_start = start, frag_end;
1534 unsigned int frag;
1535 int r;
1536
1537 /* figure out the initial fragment */
1538 amdgpu_vm_fragment(params, frag_start, end, flags, &frag, &frag_end);
1539
1540 /* walk over the address space and update the PTs */
1541 amdgpu_vm_pt_start(adev, params->vm, start, &cursor);
1542 while (cursor.pfn < end) {
1543 unsigned shift, parent_shift, mask;
1544 uint64_t incr, entry_end, pe_start;
1545 struct amdgpu_bo *pt;
1546
1547 if (!params->unlocked) {
1548 /* make sure that the page tables covering the
1549 * address range are actually allocated
1550 */
1551 r = amdgpu_vm_alloc_pts(params->adev, params->vm,
1552 &cursor, params->immediate);
1553 if (r)
1554 return r;
1555 }
1556
1557 shift = amdgpu_vm_level_shift(adev, cursor.level);
1558 parent_shift = amdgpu_vm_level_shift(adev, cursor.level - 1);
1559 if (params->unlocked) {
1560 /* Unlocked updates are only allowed on the leaves */
1561 if (amdgpu_vm_pt_descendant(adev, &cursor))
1562 continue;
1563 } else if (adev->asic_type < CHIP_VEGA10 &&
1564 (flags & AMDGPU_PTE_VALID)) {
1565 /* No huge page support before GMC v9 */
1566 if (cursor.level != AMDGPU_VM_PTB) {
1567 if (!amdgpu_vm_pt_descendant(adev, &cursor))
1568 return -ENOENT;
1569 continue;
1570 }
1571 } else if (frag < shift) {
1572 /* We can't use this level when the fragment size is
1573 * smaller than the address shift. Go to the next
1574 * child entry and try again.
1575 */
1576 if (amdgpu_vm_pt_descendant(adev, &cursor))
1577 continue;
1578 } else if (frag >= parent_shift) {
1579 /* If the fragment size is even larger than the parent
1580 * shift we should go up one level and check it again.
1581 */
1582 if (!amdgpu_vm_pt_ancestor(&cursor))
1583 return -EINVAL;
1584 continue;
1585 }
1586
1587 pt = cursor.entry->bo;
1588 if (!pt) {
1589 /* We need all PDs and PTs for mapping something, */
1590 if (flags & AMDGPU_PTE_VALID)
1591 return -ENOENT;
1592
1593 /* but unmapping something can happen at a higher
1594 * level.
1595 */
1596 if (!amdgpu_vm_pt_ancestor(&cursor))
1597 return -EINVAL;
1598
1599 pt = cursor.entry->bo;
1600 shift = parent_shift;
1601 frag_end = max(frag_end, ALIGN(frag_start + 1,
1602 1ULL << shift));
1603 }
1604
1605 /* Looks good so far, calculate parameters for the update */
1606 incr = (uint64_t)AMDGPU_GPU_PAGE_SIZE << shift;
1607 mask = amdgpu_vm_entries_mask(adev, cursor.level);
1608 pe_start = ((cursor.pfn >> shift) & mask) * 8;
1609 entry_end = ((uint64_t)mask + 1) << shift;
1610 entry_end += cursor.pfn & ~(entry_end - 1);
1611 entry_end = min(entry_end, end);
1612
1613 do {
1614 struct amdgpu_vm *vm = params->vm;
1615 uint64_t upd_end = min(entry_end, frag_end);
1616 unsigned nptes = (upd_end - frag_start) >> shift;
1617 uint64_t upd_flags = flags | AMDGPU_PTE_FRAG(frag);
1618
1619 /* This can happen when we set higher level PDs to
1620 * silent to stop fault floods.
1621 */
1622 nptes = max(nptes, 1u);
1623
1624 trace_amdgpu_vm_update_ptes(params, frag_start, upd_end,
1625 nptes, dst, incr, upd_flags,
1626 vm->task_info.pid,
1627 vm->immediate.fence_context);
1628 amdgpu_vm_update_flags(params, to_amdgpu_bo_vm(pt),
1629 cursor.level, pe_start, dst,
1630 nptes, incr, upd_flags);
1631
1632 pe_start += nptes * 8;
1633 dst += nptes * incr;
1634
1635 frag_start = upd_end;
1636 if (frag_start >= frag_end) {
1637 /* figure out the next fragment */
1638 amdgpu_vm_fragment(params, frag_start, end,
1639 flags, &frag, &frag_end);
1640 if (frag < shift)
1641 break;
1642 }
1643 } while (frag_start < entry_end);
1644
1645 if (amdgpu_vm_pt_descendant(adev, &cursor)) {
1646 /* Free all child entries.
1647 * Update the tables with the flags and addresses and free up subsequent
1648 * tables in the case of huge pages or freed up areas.
1649 * This is the maximum you can free, because all other page tables are not
1650 * completely covered by the range and so potentially still in use.
1651 */
1652 while (cursor.pfn < frag_start) {
1653 /* Make sure previous mapping is freed */
1654 if (cursor.entry->bo) {
1655 params->table_freed = true;
1656 amdgpu_vm_free_pts(adev, params->vm, &cursor);
1657 }
1658 amdgpu_vm_pt_next(adev, &cursor);
1659 }
1660
1661 } else if (frag >= shift) {
1662 /* or just move on to the next on the same level. */
1663 amdgpu_vm_pt_next(adev, &cursor);
1664 }
1665 }
1666
1667 return 0;
1668 }
1669
1670 /**
1671 * amdgpu_vm_bo_update_mapping - update a mapping in the vm page table
1672 *
1673 * @adev: amdgpu_device pointer of the VM
1674 * @bo_adev: amdgpu_device pointer of the mapped BO
1675 * @vm: requested vm
1676 * @immediate: immediate submission in a page fault
1677 * @unlocked: unlocked invalidation during MM callback
1678 * @resv: fences we need to sync to
1679 * @start: start of mapped range
1680 * @last: last mapped entry
1681 * @flags: flags for the entries
1682 * @offset: offset into nodes and pages_addr
1683 * @res: ttm_resource to map
1684 * @pages_addr: DMA addresses to use for mapping
1685 * @fence: optional resulting fence
1686 * @table_freed: return true if page table is freed
1687 *
1688 * Fill in the page table entries between @start and @last.
1689 *
1690 * Returns:
1691 * 0 for success, -EINVAL for failure.
1692 */
amdgpu_vm_bo_update_mapping(struct amdgpu_device * adev,struct amdgpu_device * bo_adev,struct amdgpu_vm * vm,bool immediate,bool unlocked,struct dma_resv * resv,uint64_t start,uint64_t last,uint64_t flags,uint64_t offset,struct ttm_resource * res,dma_addr_t * pages_addr,struct dma_fence ** fence,bool * table_freed)1693 int amdgpu_vm_bo_update_mapping(struct amdgpu_device *adev,
1694 struct amdgpu_device *bo_adev,
1695 struct amdgpu_vm *vm, bool immediate,
1696 bool unlocked, struct dma_resv *resv,
1697 uint64_t start, uint64_t last,
1698 uint64_t flags, uint64_t offset,
1699 struct ttm_resource *res,
1700 dma_addr_t *pages_addr,
1701 struct dma_fence **fence,
1702 bool *table_freed)
1703 {
1704 struct amdgpu_vm_update_params params;
1705 struct amdgpu_res_cursor cursor;
1706 enum amdgpu_sync_mode sync_mode;
1707 int r, idx;
1708
1709 if (!drm_dev_enter(&adev->ddev, &idx))
1710 return -ENODEV;
1711
1712 memset(¶ms, 0, sizeof(params));
1713 params.adev = adev;
1714 params.vm = vm;
1715 params.immediate = immediate;
1716 params.pages_addr = pages_addr;
1717 params.unlocked = unlocked;
1718
1719 /* Implicitly sync to command submissions in the same VM before
1720 * unmapping. Sync to moving fences before mapping.
1721 */
1722 if (!(flags & AMDGPU_PTE_VALID))
1723 sync_mode = AMDGPU_SYNC_EQ_OWNER;
1724 else
1725 sync_mode = AMDGPU_SYNC_EXPLICIT;
1726
1727 amdgpu_vm_eviction_lock(vm);
1728 if (vm->evicting) {
1729 r = -EBUSY;
1730 goto error_unlock;
1731 }
1732
1733 if (!unlocked && !dma_fence_is_signaled(vm->last_unlocked)) {
1734 struct dma_fence *tmp = dma_fence_get_stub();
1735
1736 amdgpu_bo_fence(vm->root.bo, vm->last_unlocked, true);
1737 swap(vm->last_unlocked, tmp);
1738 dma_fence_put(tmp);
1739 }
1740
1741 r = vm->update_funcs->prepare(¶ms, resv, sync_mode);
1742 if (r)
1743 goto error_unlock;
1744
1745 amdgpu_res_first(pages_addr ? NULL : res, offset,
1746 (last - start + 1) * AMDGPU_GPU_PAGE_SIZE, &cursor);
1747 while (cursor.remaining) {
1748 uint64_t tmp, num_entries, addr;
1749
1750 num_entries = cursor.size >> AMDGPU_GPU_PAGE_SHIFT;
1751 if (pages_addr) {
1752 bool contiguous = true;
1753
1754 if (num_entries > AMDGPU_GPU_PAGES_IN_CPU_PAGE) {
1755 uint64_t pfn = cursor.start >> PAGE_SHIFT;
1756 uint64_t count;
1757
1758 contiguous = pages_addr[pfn + 1] ==
1759 pages_addr[pfn] + PAGE_SIZE;
1760
1761 tmp = num_entries /
1762 AMDGPU_GPU_PAGES_IN_CPU_PAGE;
1763 for (count = 2; count < tmp; ++count) {
1764 uint64_t idx = pfn + count;
1765
1766 if (contiguous != (pages_addr[idx] ==
1767 pages_addr[idx - 1] + PAGE_SIZE))
1768 break;
1769 }
1770 num_entries = count *
1771 AMDGPU_GPU_PAGES_IN_CPU_PAGE;
1772 }
1773
1774 if (!contiguous) {
1775 addr = cursor.start;
1776 params.pages_addr = pages_addr;
1777 } else {
1778 addr = pages_addr[cursor.start >> PAGE_SHIFT];
1779 params.pages_addr = NULL;
1780 }
1781
1782 } else if (flags & (AMDGPU_PTE_VALID | AMDGPU_PTE_PRT)) {
1783 addr = bo_adev->vm_manager.vram_base_offset +
1784 cursor.start;
1785 } else {
1786 addr = 0;
1787 }
1788
1789 tmp = start + num_entries;
1790 r = amdgpu_vm_update_ptes(¶ms, start, tmp, addr, flags);
1791 if (r)
1792 goto error_unlock;
1793
1794 amdgpu_res_next(&cursor, num_entries * AMDGPU_GPU_PAGE_SIZE);
1795 start = tmp;
1796 }
1797
1798 r = vm->update_funcs->commit(¶ms, fence);
1799
1800 if (table_freed)
1801 *table_freed = *table_freed || params.table_freed;
1802
1803 error_unlock:
1804 amdgpu_vm_eviction_unlock(vm);
1805 drm_dev_exit(idx);
1806 return r;
1807 }
1808
amdgpu_vm_get_memory(struct amdgpu_vm * vm,uint64_t * vram_mem,uint64_t * gtt_mem,uint64_t * cpu_mem)1809 void amdgpu_vm_get_memory(struct amdgpu_vm *vm, uint64_t *vram_mem,
1810 uint64_t *gtt_mem, uint64_t *cpu_mem)
1811 {
1812 struct amdgpu_bo_va *bo_va, *tmp;
1813
1814 list_for_each_entry_safe(bo_va, tmp, &vm->idle, base.vm_status) {
1815 if (!bo_va->base.bo)
1816 continue;
1817 amdgpu_bo_get_memory(bo_va->base.bo, vram_mem,
1818 gtt_mem, cpu_mem);
1819 }
1820 list_for_each_entry_safe(bo_va, tmp, &vm->evicted, base.vm_status) {
1821 if (!bo_va->base.bo)
1822 continue;
1823 amdgpu_bo_get_memory(bo_va->base.bo, vram_mem,
1824 gtt_mem, cpu_mem);
1825 }
1826 list_for_each_entry_safe(bo_va, tmp, &vm->relocated, base.vm_status) {
1827 if (!bo_va->base.bo)
1828 continue;
1829 amdgpu_bo_get_memory(bo_va->base.bo, vram_mem,
1830 gtt_mem, cpu_mem);
1831 }
1832 list_for_each_entry_safe(bo_va, tmp, &vm->moved, base.vm_status) {
1833 if (!bo_va->base.bo)
1834 continue;
1835 amdgpu_bo_get_memory(bo_va->base.bo, vram_mem,
1836 gtt_mem, cpu_mem);
1837 }
1838 spin_lock(&vm->invalidated_lock);
1839 list_for_each_entry_safe(bo_va, tmp, &vm->invalidated, base.vm_status) {
1840 if (!bo_va->base.bo)
1841 continue;
1842 amdgpu_bo_get_memory(bo_va->base.bo, vram_mem,
1843 gtt_mem, cpu_mem);
1844 }
1845 list_for_each_entry_safe(bo_va, tmp, &vm->done, base.vm_status) {
1846 if (!bo_va->base.bo)
1847 continue;
1848 amdgpu_bo_get_memory(bo_va->base.bo, vram_mem,
1849 gtt_mem, cpu_mem);
1850 }
1851 spin_unlock(&vm->invalidated_lock);
1852 }
1853 /**
1854 * amdgpu_vm_bo_update - update all BO mappings in the vm page table
1855 *
1856 * @adev: amdgpu_device pointer
1857 * @bo_va: requested BO and VM object
1858 * @clear: if true clear the entries
1859 * @table_freed: return true if page table is freed
1860 *
1861 * Fill in the page table entries for @bo_va.
1862 *
1863 * Returns:
1864 * 0 for success, -EINVAL for failure.
1865 */
amdgpu_vm_bo_update(struct amdgpu_device * adev,struct amdgpu_bo_va * bo_va,bool clear,bool * table_freed)1866 int amdgpu_vm_bo_update(struct amdgpu_device *adev, struct amdgpu_bo_va *bo_va,
1867 bool clear, bool *table_freed)
1868 {
1869 struct amdgpu_bo *bo = bo_va->base.bo;
1870 struct amdgpu_vm *vm = bo_va->base.vm;
1871 struct amdgpu_bo_va_mapping *mapping;
1872 dma_addr_t *pages_addr = NULL;
1873 struct ttm_resource *mem;
1874 struct dma_fence **last_update;
1875 struct dma_resv *resv;
1876 uint64_t flags;
1877 struct amdgpu_device *bo_adev = adev;
1878 int r;
1879
1880 if (clear || !bo) {
1881 mem = NULL;
1882 resv = vm->root.bo->tbo.base.resv;
1883 } else {
1884 struct drm_gem_object *obj = &bo->tbo.base;
1885
1886 resv = bo->tbo.base.resv;
1887 if (obj->import_attach && bo_va->is_xgmi) {
1888 struct dma_buf *dma_buf = obj->import_attach->dmabuf;
1889 struct drm_gem_object *gobj = dma_buf->priv;
1890 struct amdgpu_bo *abo = gem_to_amdgpu_bo(gobj);
1891
1892 if (abo->tbo.resource->mem_type == TTM_PL_VRAM)
1893 bo = gem_to_amdgpu_bo(gobj);
1894 }
1895 mem = bo->tbo.resource;
1896 if (mem->mem_type == TTM_PL_TT ||
1897 mem->mem_type == AMDGPU_PL_PREEMPT)
1898 pages_addr = bo->tbo.ttm->dma_address;
1899 }
1900
1901 if (bo) {
1902 flags = amdgpu_ttm_tt_pte_flags(adev, bo->tbo.ttm, mem);
1903
1904 if (amdgpu_bo_encrypted(bo))
1905 flags |= AMDGPU_PTE_TMZ;
1906
1907 bo_adev = amdgpu_ttm_adev(bo->tbo.bdev);
1908 } else {
1909 flags = 0x0;
1910 }
1911
1912 if (clear || (bo && bo->tbo.base.resv ==
1913 vm->root.bo->tbo.base.resv))
1914 last_update = &vm->last_update;
1915 else
1916 last_update = &bo_va->last_pt_update;
1917
1918 if (!clear && bo_va->base.moved) {
1919 bo_va->base.moved = false;
1920 list_splice_init(&bo_va->valids, &bo_va->invalids);
1921
1922 } else if (bo_va->cleared != clear) {
1923 list_splice_init(&bo_va->valids, &bo_va->invalids);
1924 }
1925
1926 list_for_each_entry(mapping, &bo_va->invalids, list) {
1927 uint64_t update_flags = flags;
1928
1929 /* normally,bo_va->flags only contians READABLE and WIRTEABLE bit go here
1930 * but in case of something, we filter the flags in first place
1931 */
1932 if (!(mapping->flags & AMDGPU_PTE_READABLE))
1933 update_flags &= ~AMDGPU_PTE_READABLE;
1934 if (!(mapping->flags & AMDGPU_PTE_WRITEABLE))
1935 update_flags &= ~AMDGPU_PTE_WRITEABLE;
1936
1937 /* Apply ASIC specific mapping flags */
1938 amdgpu_gmc_get_vm_pte(adev, mapping, &update_flags);
1939
1940 trace_amdgpu_vm_bo_update(mapping);
1941
1942 r = amdgpu_vm_bo_update_mapping(adev, bo_adev, vm, false, false,
1943 resv, mapping->start,
1944 mapping->last, update_flags,
1945 mapping->offset, mem,
1946 pages_addr, last_update, table_freed);
1947 if (r)
1948 return r;
1949 }
1950
1951 /* If the BO is not in its preferred location add it back to
1952 * the evicted list so that it gets validated again on the
1953 * next command submission.
1954 */
1955 if (bo && bo->tbo.base.resv == vm->root.bo->tbo.base.resv) {
1956 uint32_t mem_type = bo->tbo.resource->mem_type;
1957
1958 if (!(bo->preferred_domains &
1959 amdgpu_mem_type_to_domain(mem_type)))
1960 amdgpu_vm_bo_evicted(&bo_va->base);
1961 else
1962 amdgpu_vm_bo_idle(&bo_va->base);
1963 } else {
1964 amdgpu_vm_bo_done(&bo_va->base);
1965 }
1966
1967 list_splice_init(&bo_va->invalids, &bo_va->valids);
1968 bo_va->cleared = clear;
1969
1970 if (trace_amdgpu_vm_bo_mapping_enabled()) {
1971 list_for_each_entry(mapping, &bo_va->valids, list)
1972 trace_amdgpu_vm_bo_mapping(mapping);
1973 }
1974
1975 return 0;
1976 }
1977
1978 /**
1979 * amdgpu_vm_update_prt_state - update the global PRT state
1980 *
1981 * @adev: amdgpu_device pointer
1982 */
amdgpu_vm_update_prt_state(struct amdgpu_device * adev)1983 static void amdgpu_vm_update_prt_state(struct amdgpu_device *adev)
1984 {
1985 unsigned long flags;
1986 bool enable;
1987
1988 spin_lock_irqsave(&adev->vm_manager.prt_lock, flags);
1989 enable = !!atomic_read(&adev->vm_manager.num_prt_users);
1990 adev->gmc.gmc_funcs->set_prt(adev, enable);
1991 spin_unlock_irqrestore(&adev->vm_manager.prt_lock, flags);
1992 }
1993
1994 /**
1995 * amdgpu_vm_prt_get - add a PRT user
1996 *
1997 * @adev: amdgpu_device pointer
1998 */
amdgpu_vm_prt_get(struct amdgpu_device * adev)1999 static void amdgpu_vm_prt_get(struct amdgpu_device *adev)
2000 {
2001 if (!adev->gmc.gmc_funcs->set_prt)
2002 return;
2003
2004 if (atomic_inc_return(&adev->vm_manager.num_prt_users) == 1)
2005 amdgpu_vm_update_prt_state(adev);
2006 }
2007
2008 /**
2009 * amdgpu_vm_prt_put - drop a PRT user
2010 *
2011 * @adev: amdgpu_device pointer
2012 */
amdgpu_vm_prt_put(struct amdgpu_device * adev)2013 static void amdgpu_vm_prt_put(struct amdgpu_device *adev)
2014 {
2015 if (atomic_dec_return(&adev->vm_manager.num_prt_users) == 0)
2016 amdgpu_vm_update_prt_state(adev);
2017 }
2018
2019 /**
2020 * amdgpu_vm_prt_cb - callback for updating the PRT status
2021 *
2022 * @fence: fence for the callback
2023 * @_cb: the callback function
2024 */
amdgpu_vm_prt_cb(struct dma_fence * fence,struct dma_fence_cb * _cb)2025 static void amdgpu_vm_prt_cb(struct dma_fence *fence, struct dma_fence_cb *_cb)
2026 {
2027 struct amdgpu_prt_cb *cb = container_of(_cb, struct amdgpu_prt_cb, cb);
2028
2029 amdgpu_vm_prt_put(cb->adev);
2030 kfree(cb);
2031 }
2032
2033 /**
2034 * amdgpu_vm_add_prt_cb - add callback for updating the PRT status
2035 *
2036 * @adev: amdgpu_device pointer
2037 * @fence: fence for the callback
2038 */
amdgpu_vm_add_prt_cb(struct amdgpu_device * adev,struct dma_fence * fence)2039 static void amdgpu_vm_add_prt_cb(struct amdgpu_device *adev,
2040 struct dma_fence *fence)
2041 {
2042 struct amdgpu_prt_cb *cb;
2043
2044 if (!adev->gmc.gmc_funcs->set_prt)
2045 return;
2046
2047 cb = kmalloc(sizeof(struct amdgpu_prt_cb), GFP_KERNEL);
2048 if (!cb) {
2049 /* Last resort when we are OOM */
2050 if (fence)
2051 dma_fence_wait(fence, false);
2052
2053 amdgpu_vm_prt_put(adev);
2054 } else {
2055 cb->adev = adev;
2056 if (!fence || dma_fence_add_callback(fence, &cb->cb,
2057 amdgpu_vm_prt_cb))
2058 amdgpu_vm_prt_cb(fence, &cb->cb);
2059 }
2060 }
2061
2062 /**
2063 * amdgpu_vm_free_mapping - free a mapping
2064 *
2065 * @adev: amdgpu_device pointer
2066 * @vm: requested vm
2067 * @mapping: mapping to be freed
2068 * @fence: fence of the unmap operation
2069 *
2070 * Free a mapping and make sure we decrease the PRT usage count if applicable.
2071 */
amdgpu_vm_free_mapping(struct amdgpu_device * adev,struct amdgpu_vm * vm,struct amdgpu_bo_va_mapping * mapping,struct dma_fence * fence)2072 static void amdgpu_vm_free_mapping(struct amdgpu_device *adev,
2073 struct amdgpu_vm *vm,
2074 struct amdgpu_bo_va_mapping *mapping,
2075 struct dma_fence *fence)
2076 {
2077 if (mapping->flags & AMDGPU_PTE_PRT)
2078 amdgpu_vm_add_prt_cb(adev, fence);
2079 kfree(mapping);
2080 }
2081
2082 /**
2083 * amdgpu_vm_prt_fini - finish all prt mappings
2084 *
2085 * @adev: amdgpu_device pointer
2086 * @vm: requested vm
2087 *
2088 * Register a cleanup callback to disable PRT support after VM dies.
2089 */
amdgpu_vm_prt_fini(struct amdgpu_device * adev,struct amdgpu_vm * vm)2090 static void amdgpu_vm_prt_fini(struct amdgpu_device *adev, struct amdgpu_vm *vm)
2091 {
2092 struct dma_resv *resv = vm->root.bo->tbo.base.resv;
2093 struct dma_fence *excl, **shared;
2094 unsigned i, shared_count;
2095 int r;
2096
2097 r = dma_resv_get_fences(resv, &excl, &shared_count, &shared);
2098 if (r) {
2099 /* Not enough memory to grab the fence list, as last resort
2100 * block for all the fences to complete.
2101 */
2102 dma_resv_wait_timeout(resv, true, false,
2103 MAX_SCHEDULE_TIMEOUT);
2104 return;
2105 }
2106
2107 /* Add a callback for each fence in the reservation object */
2108 amdgpu_vm_prt_get(adev);
2109 amdgpu_vm_add_prt_cb(adev, excl);
2110
2111 for (i = 0; i < shared_count; ++i) {
2112 amdgpu_vm_prt_get(adev);
2113 amdgpu_vm_add_prt_cb(adev, shared[i]);
2114 }
2115
2116 kfree(shared);
2117 }
2118
2119 /**
2120 * amdgpu_vm_clear_freed - clear freed BOs in the PT
2121 *
2122 * @adev: amdgpu_device pointer
2123 * @vm: requested vm
2124 * @fence: optional resulting fence (unchanged if no work needed to be done
2125 * or if an error occurred)
2126 *
2127 * Make sure all freed BOs are cleared in the PT.
2128 * PTs have to be reserved and mutex must be locked!
2129 *
2130 * Returns:
2131 * 0 for success.
2132 *
2133 */
amdgpu_vm_clear_freed(struct amdgpu_device * adev,struct amdgpu_vm * vm,struct dma_fence ** fence)2134 int amdgpu_vm_clear_freed(struct amdgpu_device *adev,
2135 struct amdgpu_vm *vm,
2136 struct dma_fence **fence)
2137 {
2138 struct dma_resv *resv = vm->root.bo->tbo.base.resv;
2139 struct amdgpu_bo_va_mapping *mapping;
2140 uint64_t init_pte_value = 0;
2141 struct dma_fence *f = NULL;
2142 int r;
2143
2144 while (!list_empty(&vm->freed)) {
2145 mapping = list_first_entry(&vm->freed,
2146 struct amdgpu_bo_va_mapping, list);
2147 list_del(&mapping->list);
2148
2149 if (vm->pte_support_ats &&
2150 mapping->start < AMDGPU_GMC_HOLE_START)
2151 init_pte_value = AMDGPU_PTE_DEFAULT_ATC;
2152
2153 r = amdgpu_vm_bo_update_mapping(adev, adev, vm, false, false,
2154 resv, mapping->start,
2155 mapping->last, init_pte_value,
2156 0, NULL, NULL, &f, NULL);
2157 amdgpu_vm_free_mapping(adev, vm, mapping, f);
2158 if (r) {
2159 dma_fence_put(f);
2160 return r;
2161 }
2162 }
2163
2164 if (fence && f) {
2165 dma_fence_put(*fence);
2166 *fence = f;
2167 } else {
2168 dma_fence_put(f);
2169 }
2170
2171 return 0;
2172
2173 }
2174
2175 /**
2176 * amdgpu_vm_handle_moved - handle moved BOs in the PT
2177 *
2178 * @adev: amdgpu_device pointer
2179 * @vm: requested vm
2180 *
2181 * Make sure all BOs which are moved are updated in the PTs.
2182 *
2183 * Returns:
2184 * 0 for success.
2185 *
2186 * PTs have to be reserved!
2187 */
amdgpu_vm_handle_moved(struct amdgpu_device * adev,struct amdgpu_vm * vm)2188 int amdgpu_vm_handle_moved(struct amdgpu_device *adev,
2189 struct amdgpu_vm *vm)
2190 {
2191 struct amdgpu_bo_va *bo_va, *tmp;
2192 struct dma_resv *resv;
2193 bool clear;
2194 int r;
2195
2196 list_for_each_entry_safe(bo_va, tmp, &vm->moved, base.vm_status) {
2197 /* Per VM BOs never need to bo cleared in the page tables */
2198 r = amdgpu_vm_bo_update(adev, bo_va, false, NULL);
2199 if (r)
2200 return r;
2201 }
2202
2203 spin_lock(&vm->invalidated_lock);
2204 while (!list_empty(&vm->invalidated)) {
2205 bo_va = list_first_entry(&vm->invalidated, struct amdgpu_bo_va,
2206 base.vm_status);
2207 resv = bo_va->base.bo->tbo.base.resv;
2208 spin_unlock(&vm->invalidated_lock);
2209
2210 /* Try to reserve the BO to avoid clearing its ptes */
2211 if (!amdgpu_vm_debug && dma_resv_trylock(resv))
2212 clear = false;
2213 /* Somebody else is using the BO right now */
2214 else
2215 clear = true;
2216
2217 r = amdgpu_vm_bo_update(adev, bo_va, clear, NULL);
2218 if (r)
2219 return r;
2220
2221 if (!clear)
2222 dma_resv_unlock(resv);
2223 spin_lock(&vm->invalidated_lock);
2224 }
2225 spin_unlock(&vm->invalidated_lock);
2226
2227 return 0;
2228 }
2229
2230 /**
2231 * amdgpu_vm_bo_add - add a bo to a specific vm
2232 *
2233 * @adev: amdgpu_device pointer
2234 * @vm: requested vm
2235 * @bo: amdgpu buffer object
2236 *
2237 * Add @bo into the requested vm.
2238 * Add @bo to the list of bos associated with the vm
2239 *
2240 * Returns:
2241 * Newly added bo_va or NULL for failure
2242 *
2243 * Object has to be reserved!
2244 */
amdgpu_vm_bo_add(struct amdgpu_device * adev,struct amdgpu_vm * vm,struct amdgpu_bo * bo)2245 struct amdgpu_bo_va *amdgpu_vm_bo_add(struct amdgpu_device *adev,
2246 struct amdgpu_vm *vm,
2247 struct amdgpu_bo *bo)
2248 {
2249 struct amdgpu_bo_va *bo_va;
2250
2251 bo_va = kzalloc(sizeof(struct amdgpu_bo_va), GFP_KERNEL);
2252 if (bo_va == NULL) {
2253 return NULL;
2254 }
2255 amdgpu_vm_bo_base_init(&bo_va->base, vm, bo);
2256
2257 bo_va->ref_count = 1;
2258 INIT_LIST_HEAD(&bo_va->valids);
2259 INIT_LIST_HEAD(&bo_va->invalids);
2260
2261 if (!bo)
2262 return bo_va;
2263
2264 if (amdgpu_dmabuf_is_xgmi_accessible(adev, bo)) {
2265 bo_va->is_xgmi = true;
2266 /* Power up XGMI if it can be potentially used */
2267 amdgpu_xgmi_set_pstate(adev, AMDGPU_XGMI_PSTATE_MAX_VEGA20);
2268 }
2269
2270 return bo_va;
2271 }
2272
2273
2274 /**
2275 * amdgpu_vm_bo_insert_map - insert a new mapping
2276 *
2277 * @adev: amdgpu_device pointer
2278 * @bo_va: bo_va to store the address
2279 * @mapping: the mapping to insert
2280 *
2281 * Insert a new mapping into all structures.
2282 */
amdgpu_vm_bo_insert_map(struct amdgpu_device * adev,struct amdgpu_bo_va * bo_va,struct amdgpu_bo_va_mapping * mapping)2283 static void amdgpu_vm_bo_insert_map(struct amdgpu_device *adev,
2284 struct amdgpu_bo_va *bo_va,
2285 struct amdgpu_bo_va_mapping *mapping)
2286 {
2287 struct amdgpu_vm *vm = bo_va->base.vm;
2288 struct amdgpu_bo *bo = bo_va->base.bo;
2289
2290 mapping->bo_va = bo_va;
2291 list_add(&mapping->list, &bo_va->invalids);
2292 amdgpu_vm_it_insert(mapping, &vm->va);
2293
2294 if (mapping->flags & AMDGPU_PTE_PRT)
2295 amdgpu_vm_prt_get(adev);
2296
2297 if (bo && bo->tbo.base.resv == vm->root.bo->tbo.base.resv &&
2298 !bo_va->base.moved) {
2299 list_move(&bo_va->base.vm_status, &vm->moved);
2300 }
2301 trace_amdgpu_vm_bo_map(bo_va, mapping);
2302 }
2303
2304 /**
2305 * amdgpu_vm_bo_map - map bo inside a vm
2306 *
2307 * @adev: amdgpu_device pointer
2308 * @bo_va: bo_va to store the address
2309 * @saddr: where to map the BO
2310 * @offset: requested offset in the BO
2311 * @size: BO size in bytes
2312 * @flags: attributes of pages (read/write/valid/etc.)
2313 *
2314 * Add a mapping of the BO at the specefied addr into the VM.
2315 *
2316 * Returns:
2317 * 0 for success, error for failure.
2318 *
2319 * Object has to be reserved and unreserved outside!
2320 */
amdgpu_vm_bo_map(struct amdgpu_device * adev,struct amdgpu_bo_va * bo_va,uint64_t saddr,uint64_t offset,uint64_t size,uint64_t flags)2321 int amdgpu_vm_bo_map(struct amdgpu_device *adev,
2322 struct amdgpu_bo_va *bo_va,
2323 uint64_t saddr, uint64_t offset,
2324 uint64_t size, uint64_t flags)
2325 {
2326 struct amdgpu_bo_va_mapping *mapping, *tmp;
2327 struct amdgpu_bo *bo = bo_va->base.bo;
2328 struct amdgpu_vm *vm = bo_va->base.vm;
2329 uint64_t eaddr;
2330
2331 /* validate the parameters */
2332 if (saddr & ~PAGE_MASK || offset & ~PAGE_MASK ||
2333 size == 0 || size & ~PAGE_MASK)
2334 return -EINVAL;
2335
2336 /* make sure object fit at this offset */
2337 eaddr = saddr + size - 1;
2338 if (saddr >= eaddr ||
2339 (bo && offset + size > amdgpu_bo_size(bo)) ||
2340 (eaddr >= adev->vm_manager.max_pfn << AMDGPU_GPU_PAGE_SHIFT))
2341 return -EINVAL;
2342
2343 saddr /= AMDGPU_GPU_PAGE_SIZE;
2344 eaddr /= AMDGPU_GPU_PAGE_SIZE;
2345
2346 tmp = amdgpu_vm_it_iter_first(&vm->va, saddr, eaddr);
2347 if (tmp) {
2348 /* bo and tmp overlap, invalid addr */
2349 dev_err(adev->dev, "bo %p va 0x%010Lx-0x%010Lx conflict with "
2350 "0x%010Lx-0x%010Lx\n", bo, saddr, eaddr,
2351 tmp->start, tmp->last + 1);
2352 return -EINVAL;
2353 }
2354
2355 mapping = kmalloc(sizeof(*mapping), GFP_KERNEL);
2356 if (!mapping)
2357 return -ENOMEM;
2358
2359 mapping->start = saddr;
2360 mapping->last = eaddr;
2361 mapping->offset = offset;
2362 mapping->flags = flags;
2363
2364 amdgpu_vm_bo_insert_map(adev, bo_va, mapping);
2365
2366 return 0;
2367 }
2368
2369 /**
2370 * amdgpu_vm_bo_replace_map - map bo inside a vm, replacing existing mappings
2371 *
2372 * @adev: amdgpu_device pointer
2373 * @bo_va: bo_va to store the address
2374 * @saddr: where to map the BO
2375 * @offset: requested offset in the BO
2376 * @size: BO size in bytes
2377 * @flags: attributes of pages (read/write/valid/etc.)
2378 *
2379 * Add a mapping of the BO at the specefied addr into the VM. Replace existing
2380 * mappings as we do so.
2381 *
2382 * Returns:
2383 * 0 for success, error for failure.
2384 *
2385 * Object has to be reserved and unreserved outside!
2386 */
amdgpu_vm_bo_replace_map(struct amdgpu_device * adev,struct amdgpu_bo_va * bo_va,uint64_t saddr,uint64_t offset,uint64_t size,uint64_t flags)2387 int amdgpu_vm_bo_replace_map(struct amdgpu_device *adev,
2388 struct amdgpu_bo_va *bo_va,
2389 uint64_t saddr, uint64_t offset,
2390 uint64_t size, uint64_t flags)
2391 {
2392 struct amdgpu_bo_va_mapping *mapping;
2393 struct amdgpu_bo *bo = bo_va->base.bo;
2394 uint64_t eaddr;
2395 int r;
2396
2397 /* validate the parameters */
2398 if (saddr & ~PAGE_MASK || offset & ~PAGE_MASK ||
2399 size == 0 || size & ~PAGE_MASK)
2400 return -EINVAL;
2401
2402 /* make sure object fit at this offset */
2403 eaddr = saddr + size - 1;
2404 if (saddr >= eaddr ||
2405 (bo && offset + size > amdgpu_bo_size(bo)) ||
2406 (eaddr >= adev->vm_manager.max_pfn << AMDGPU_GPU_PAGE_SHIFT))
2407 return -EINVAL;
2408
2409 /* Allocate all the needed memory */
2410 mapping = kmalloc(sizeof(*mapping), GFP_KERNEL);
2411 if (!mapping)
2412 return -ENOMEM;
2413
2414 r = amdgpu_vm_bo_clear_mappings(adev, bo_va->base.vm, saddr, size);
2415 if (r) {
2416 kfree(mapping);
2417 return r;
2418 }
2419
2420 saddr /= AMDGPU_GPU_PAGE_SIZE;
2421 eaddr /= AMDGPU_GPU_PAGE_SIZE;
2422
2423 mapping->start = saddr;
2424 mapping->last = eaddr;
2425 mapping->offset = offset;
2426 mapping->flags = flags;
2427
2428 amdgpu_vm_bo_insert_map(adev, bo_va, mapping);
2429
2430 return 0;
2431 }
2432
2433 /**
2434 * amdgpu_vm_bo_unmap - remove bo mapping from vm
2435 *
2436 * @adev: amdgpu_device pointer
2437 * @bo_va: bo_va to remove the address from
2438 * @saddr: where to the BO is mapped
2439 *
2440 * Remove a mapping of the BO at the specefied addr from the VM.
2441 *
2442 * Returns:
2443 * 0 for success, error for failure.
2444 *
2445 * Object has to be reserved and unreserved outside!
2446 */
amdgpu_vm_bo_unmap(struct amdgpu_device * adev,struct amdgpu_bo_va * bo_va,uint64_t saddr)2447 int amdgpu_vm_bo_unmap(struct amdgpu_device *adev,
2448 struct amdgpu_bo_va *bo_va,
2449 uint64_t saddr)
2450 {
2451 struct amdgpu_bo_va_mapping *mapping;
2452 struct amdgpu_vm *vm = bo_va->base.vm;
2453 bool valid = true;
2454
2455 saddr /= AMDGPU_GPU_PAGE_SIZE;
2456
2457 list_for_each_entry(mapping, &bo_va->valids, list) {
2458 if (mapping->start == saddr)
2459 break;
2460 }
2461
2462 if (&mapping->list == &bo_va->valids) {
2463 valid = false;
2464
2465 list_for_each_entry(mapping, &bo_va->invalids, list) {
2466 if (mapping->start == saddr)
2467 break;
2468 }
2469
2470 if (&mapping->list == &bo_va->invalids)
2471 return -ENOENT;
2472 }
2473
2474 list_del(&mapping->list);
2475 amdgpu_vm_it_remove(mapping, &vm->va);
2476 mapping->bo_va = NULL;
2477 trace_amdgpu_vm_bo_unmap(bo_va, mapping);
2478
2479 if (valid)
2480 list_add(&mapping->list, &vm->freed);
2481 else
2482 amdgpu_vm_free_mapping(adev, vm, mapping,
2483 bo_va->last_pt_update);
2484
2485 return 0;
2486 }
2487
2488 /**
2489 * amdgpu_vm_bo_clear_mappings - remove all mappings in a specific range
2490 *
2491 * @adev: amdgpu_device pointer
2492 * @vm: VM structure to use
2493 * @saddr: start of the range
2494 * @size: size of the range
2495 *
2496 * Remove all mappings in a range, split them as appropriate.
2497 *
2498 * Returns:
2499 * 0 for success, error for failure.
2500 */
amdgpu_vm_bo_clear_mappings(struct amdgpu_device * adev,struct amdgpu_vm * vm,uint64_t saddr,uint64_t size)2501 int amdgpu_vm_bo_clear_mappings(struct amdgpu_device *adev,
2502 struct amdgpu_vm *vm,
2503 uint64_t saddr, uint64_t size)
2504 {
2505 struct amdgpu_bo_va_mapping *before, *after, *tmp, *next;
2506 LIST_HEAD(removed);
2507 uint64_t eaddr;
2508
2509 eaddr = saddr + size - 1;
2510 saddr /= AMDGPU_GPU_PAGE_SIZE;
2511 eaddr /= AMDGPU_GPU_PAGE_SIZE;
2512
2513 /* Allocate all the needed memory */
2514 before = kzalloc(sizeof(*before), GFP_KERNEL);
2515 if (!before)
2516 return -ENOMEM;
2517 INIT_LIST_HEAD(&before->list);
2518
2519 after = kzalloc(sizeof(*after), GFP_KERNEL);
2520 if (!after) {
2521 kfree(before);
2522 return -ENOMEM;
2523 }
2524 INIT_LIST_HEAD(&after->list);
2525
2526 /* Now gather all removed mappings */
2527 tmp = amdgpu_vm_it_iter_first(&vm->va, saddr, eaddr);
2528 while (tmp) {
2529 /* Remember mapping split at the start */
2530 if (tmp->start < saddr) {
2531 before->start = tmp->start;
2532 before->last = saddr - 1;
2533 before->offset = tmp->offset;
2534 before->flags = tmp->flags;
2535 before->bo_va = tmp->bo_va;
2536 list_add(&before->list, &tmp->bo_va->invalids);
2537 }
2538
2539 /* Remember mapping split at the end */
2540 if (tmp->last > eaddr) {
2541 after->start = eaddr + 1;
2542 after->last = tmp->last;
2543 after->offset = tmp->offset;
2544 after->offset += (after->start - tmp->start) << PAGE_SHIFT;
2545 after->flags = tmp->flags;
2546 after->bo_va = tmp->bo_va;
2547 list_add(&after->list, &tmp->bo_va->invalids);
2548 }
2549
2550 list_del(&tmp->list);
2551 list_add(&tmp->list, &removed);
2552
2553 tmp = amdgpu_vm_it_iter_next(tmp, saddr, eaddr);
2554 }
2555
2556 /* And free them up */
2557 list_for_each_entry_safe(tmp, next, &removed, list) {
2558 amdgpu_vm_it_remove(tmp, &vm->va);
2559 list_del(&tmp->list);
2560
2561 if (tmp->start < saddr)
2562 tmp->start = saddr;
2563 if (tmp->last > eaddr)
2564 tmp->last = eaddr;
2565
2566 tmp->bo_va = NULL;
2567 list_add(&tmp->list, &vm->freed);
2568 trace_amdgpu_vm_bo_unmap(NULL, tmp);
2569 }
2570
2571 /* Insert partial mapping before the range */
2572 if (!list_empty(&before->list)) {
2573 amdgpu_vm_it_insert(before, &vm->va);
2574 if (before->flags & AMDGPU_PTE_PRT)
2575 amdgpu_vm_prt_get(adev);
2576 } else {
2577 kfree(before);
2578 }
2579
2580 /* Insert partial mapping after the range */
2581 if (!list_empty(&after->list)) {
2582 amdgpu_vm_it_insert(after, &vm->va);
2583 if (after->flags & AMDGPU_PTE_PRT)
2584 amdgpu_vm_prt_get(adev);
2585 } else {
2586 kfree(after);
2587 }
2588
2589 return 0;
2590 }
2591
2592 /**
2593 * amdgpu_vm_bo_lookup_mapping - find mapping by address
2594 *
2595 * @vm: the requested VM
2596 * @addr: the address
2597 *
2598 * Find a mapping by it's address.
2599 *
2600 * Returns:
2601 * The amdgpu_bo_va_mapping matching for addr or NULL
2602 *
2603 */
amdgpu_vm_bo_lookup_mapping(struct amdgpu_vm * vm,uint64_t addr)2604 struct amdgpu_bo_va_mapping *amdgpu_vm_bo_lookup_mapping(struct amdgpu_vm *vm,
2605 uint64_t addr)
2606 {
2607 return amdgpu_vm_it_iter_first(&vm->va, addr, addr);
2608 }
2609
2610 /**
2611 * amdgpu_vm_bo_trace_cs - trace all reserved mappings
2612 *
2613 * @vm: the requested vm
2614 * @ticket: CS ticket
2615 *
2616 * Trace all mappings of BOs reserved during a command submission.
2617 */
amdgpu_vm_bo_trace_cs(struct amdgpu_vm * vm,struct ww_acquire_ctx * ticket)2618 void amdgpu_vm_bo_trace_cs(struct amdgpu_vm *vm, struct ww_acquire_ctx *ticket)
2619 {
2620 struct amdgpu_bo_va_mapping *mapping;
2621
2622 if (!trace_amdgpu_vm_bo_cs_enabled())
2623 return;
2624
2625 for (mapping = amdgpu_vm_it_iter_first(&vm->va, 0, U64_MAX); mapping;
2626 mapping = amdgpu_vm_it_iter_next(mapping, 0, U64_MAX)) {
2627 if (mapping->bo_va && mapping->bo_va->base.bo) {
2628 struct amdgpu_bo *bo;
2629
2630 bo = mapping->bo_va->base.bo;
2631 if (dma_resv_locking_ctx(bo->tbo.base.resv) !=
2632 ticket)
2633 continue;
2634 }
2635
2636 trace_amdgpu_vm_bo_cs(mapping);
2637 }
2638 }
2639
2640 /**
2641 * amdgpu_vm_bo_rmv - remove a bo to a specific vm
2642 *
2643 * @adev: amdgpu_device pointer
2644 * @bo_va: requested bo_va
2645 *
2646 * Remove @bo_va->bo from the requested vm.
2647 *
2648 * Object have to be reserved!
2649 */
amdgpu_vm_bo_rmv(struct amdgpu_device * adev,struct amdgpu_bo_va * bo_va)2650 void amdgpu_vm_bo_rmv(struct amdgpu_device *adev,
2651 struct amdgpu_bo_va *bo_va)
2652 {
2653 struct amdgpu_bo_va_mapping *mapping, *next;
2654 struct amdgpu_bo *bo = bo_va->base.bo;
2655 struct amdgpu_vm *vm = bo_va->base.vm;
2656 struct amdgpu_vm_bo_base **base;
2657
2658 if (bo) {
2659 if (bo->tbo.base.resv == vm->root.bo->tbo.base.resv)
2660 vm->bulk_moveable = false;
2661
2662 for (base = &bo_va->base.bo->vm_bo; *base;
2663 base = &(*base)->next) {
2664 if (*base != &bo_va->base)
2665 continue;
2666
2667 *base = bo_va->base.next;
2668 break;
2669 }
2670 }
2671
2672 spin_lock(&vm->invalidated_lock);
2673 list_del(&bo_va->base.vm_status);
2674 spin_unlock(&vm->invalidated_lock);
2675
2676 list_for_each_entry_safe(mapping, next, &bo_va->valids, list) {
2677 list_del(&mapping->list);
2678 amdgpu_vm_it_remove(mapping, &vm->va);
2679 mapping->bo_va = NULL;
2680 trace_amdgpu_vm_bo_unmap(bo_va, mapping);
2681 list_add(&mapping->list, &vm->freed);
2682 }
2683 list_for_each_entry_safe(mapping, next, &bo_va->invalids, list) {
2684 list_del(&mapping->list);
2685 amdgpu_vm_it_remove(mapping, &vm->va);
2686 amdgpu_vm_free_mapping(adev, vm, mapping,
2687 bo_va->last_pt_update);
2688 }
2689
2690 dma_fence_put(bo_va->last_pt_update);
2691
2692 if (bo && bo_va->is_xgmi)
2693 amdgpu_xgmi_set_pstate(adev, AMDGPU_XGMI_PSTATE_MIN);
2694
2695 kfree(bo_va);
2696 }
2697
2698 /**
2699 * amdgpu_vm_evictable - check if we can evict a VM
2700 *
2701 * @bo: A page table of the VM.
2702 *
2703 * Check if it is possible to evict a VM.
2704 */
amdgpu_vm_evictable(struct amdgpu_bo * bo)2705 bool amdgpu_vm_evictable(struct amdgpu_bo *bo)
2706 {
2707 struct amdgpu_vm_bo_base *bo_base = bo->vm_bo;
2708
2709 /* Page tables of a destroyed VM can go away immediately */
2710 if (!bo_base || !bo_base->vm)
2711 return true;
2712
2713 /* Don't evict VM page tables while they are busy */
2714 if (!dma_resv_test_signaled(bo->tbo.base.resv, true))
2715 return false;
2716
2717 /* Try to block ongoing updates */
2718 if (!amdgpu_vm_eviction_trylock(bo_base->vm))
2719 return false;
2720
2721 /* Don't evict VM page tables while they are updated */
2722 if (!dma_fence_is_signaled(bo_base->vm->last_unlocked)) {
2723 amdgpu_vm_eviction_unlock(bo_base->vm);
2724 return false;
2725 }
2726
2727 bo_base->vm->evicting = true;
2728 amdgpu_vm_eviction_unlock(bo_base->vm);
2729 return true;
2730 }
2731
2732 /**
2733 * amdgpu_vm_bo_invalidate - mark the bo as invalid
2734 *
2735 * @adev: amdgpu_device pointer
2736 * @bo: amdgpu buffer object
2737 * @evicted: is the BO evicted
2738 *
2739 * Mark @bo as invalid.
2740 */
amdgpu_vm_bo_invalidate(struct amdgpu_device * adev,struct amdgpu_bo * bo,bool evicted)2741 void amdgpu_vm_bo_invalidate(struct amdgpu_device *adev,
2742 struct amdgpu_bo *bo, bool evicted)
2743 {
2744 struct amdgpu_vm_bo_base *bo_base;
2745
2746 /* shadow bo doesn't have bo base, its validation needs its parent */
2747 if (bo->parent && (amdgpu_bo_shadowed(bo->parent) == bo))
2748 bo = bo->parent;
2749
2750 for (bo_base = bo->vm_bo; bo_base; bo_base = bo_base->next) {
2751 struct amdgpu_vm *vm = bo_base->vm;
2752
2753 if (evicted && bo->tbo.base.resv == vm->root.bo->tbo.base.resv) {
2754 amdgpu_vm_bo_evicted(bo_base);
2755 continue;
2756 }
2757
2758 if (bo_base->moved)
2759 continue;
2760 bo_base->moved = true;
2761
2762 if (bo->tbo.type == ttm_bo_type_kernel)
2763 amdgpu_vm_bo_relocated(bo_base);
2764 else if (bo->tbo.base.resv == vm->root.bo->tbo.base.resv)
2765 amdgpu_vm_bo_moved(bo_base);
2766 else
2767 amdgpu_vm_bo_invalidated(bo_base);
2768 }
2769 }
2770
2771 /**
2772 * amdgpu_vm_get_block_size - calculate VM page table size as power of two
2773 *
2774 * @vm_size: VM size
2775 *
2776 * Returns:
2777 * VM page table as power of two
2778 */
amdgpu_vm_get_block_size(uint64_t vm_size)2779 static uint32_t amdgpu_vm_get_block_size(uint64_t vm_size)
2780 {
2781 /* Total bits covered by PD + PTs */
2782 unsigned bits = ilog2(vm_size) + 18;
2783
2784 /* Make sure the PD is 4K in size up to 8GB address space.
2785 Above that split equal between PD and PTs */
2786 if (vm_size <= 8)
2787 return (bits - 9);
2788 else
2789 return ((bits + 3) / 2);
2790 }
2791
2792 /**
2793 * amdgpu_vm_adjust_size - adjust vm size, block size and fragment size
2794 *
2795 * @adev: amdgpu_device pointer
2796 * @min_vm_size: the minimum vm size in GB if it's set auto
2797 * @fragment_size_default: Default PTE fragment size
2798 * @max_level: max VMPT level
2799 * @max_bits: max address space size in bits
2800 *
2801 */
amdgpu_vm_adjust_size(struct amdgpu_device * adev,uint32_t min_vm_size,uint32_t fragment_size_default,unsigned max_level,unsigned max_bits)2802 void amdgpu_vm_adjust_size(struct amdgpu_device *adev, uint32_t min_vm_size,
2803 uint32_t fragment_size_default, unsigned max_level,
2804 unsigned max_bits)
2805 {
2806 unsigned int max_size = 1 << (max_bits - 30);
2807 unsigned int vm_size;
2808 uint64_t tmp;
2809
2810 /* adjust vm size first */
2811 if (amdgpu_vm_size != -1) {
2812 vm_size = amdgpu_vm_size;
2813 if (vm_size > max_size) {
2814 dev_warn(adev->dev, "VM size (%d) too large, max is %u GB\n",
2815 amdgpu_vm_size, max_size);
2816 vm_size = max_size;
2817 }
2818 } else {
2819 struct sysinfo si;
2820 unsigned int phys_ram_gb;
2821
2822 /* Optimal VM size depends on the amount of physical
2823 * RAM available. Underlying requirements and
2824 * assumptions:
2825 *
2826 * - Need to map system memory and VRAM from all GPUs
2827 * - VRAM from other GPUs not known here
2828 * - Assume VRAM <= system memory
2829 * - On GFX8 and older, VM space can be segmented for
2830 * different MTYPEs
2831 * - Need to allow room for fragmentation, guard pages etc.
2832 *
2833 * This adds up to a rough guess of system memory x3.
2834 * Round up to power of two to maximize the available
2835 * VM size with the given page table size.
2836 */
2837 si_meminfo(&si);
2838 phys_ram_gb = ((uint64_t)si.totalram * si.mem_unit +
2839 (1 << 30) - 1) >> 30;
2840 vm_size = roundup_pow_of_two(
2841 min(max(phys_ram_gb * 3, min_vm_size), max_size));
2842 }
2843
2844 adev->vm_manager.max_pfn = (uint64_t)vm_size << 18;
2845
2846 tmp = roundup_pow_of_two(adev->vm_manager.max_pfn);
2847 if (amdgpu_vm_block_size != -1)
2848 tmp >>= amdgpu_vm_block_size - 9;
2849 tmp = DIV_ROUND_UP(fls64(tmp) - 1, 9) - 1;
2850 adev->vm_manager.num_level = min(max_level, (unsigned)tmp);
2851 switch (adev->vm_manager.num_level) {
2852 case 3:
2853 adev->vm_manager.root_level = AMDGPU_VM_PDB2;
2854 break;
2855 case 2:
2856 adev->vm_manager.root_level = AMDGPU_VM_PDB1;
2857 break;
2858 case 1:
2859 adev->vm_manager.root_level = AMDGPU_VM_PDB0;
2860 break;
2861 default:
2862 dev_err(adev->dev, "VMPT only supports 2~4+1 levels\n");
2863 }
2864 /* block size depends on vm size and hw setup*/
2865 if (amdgpu_vm_block_size != -1)
2866 adev->vm_manager.block_size =
2867 min((unsigned)amdgpu_vm_block_size, max_bits
2868 - AMDGPU_GPU_PAGE_SHIFT
2869 - 9 * adev->vm_manager.num_level);
2870 else if (adev->vm_manager.num_level > 1)
2871 adev->vm_manager.block_size = 9;
2872 else
2873 adev->vm_manager.block_size = amdgpu_vm_get_block_size(tmp);
2874
2875 if (amdgpu_vm_fragment_size == -1)
2876 adev->vm_manager.fragment_size = fragment_size_default;
2877 else
2878 adev->vm_manager.fragment_size = amdgpu_vm_fragment_size;
2879
2880 DRM_INFO("vm size is %u GB, %u levels, block size is %u-bit, fragment size is %u-bit\n",
2881 vm_size, adev->vm_manager.num_level + 1,
2882 adev->vm_manager.block_size,
2883 adev->vm_manager.fragment_size);
2884 }
2885
2886 /**
2887 * amdgpu_vm_wait_idle - wait for the VM to become idle
2888 *
2889 * @vm: VM object to wait for
2890 * @timeout: timeout to wait for VM to become idle
2891 */
amdgpu_vm_wait_idle(struct amdgpu_vm * vm,long timeout)2892 long amdgpu_vm_wait_idle(struct amdgpu_vm *vm, long timeout)
2893 {
2894 timeout = dma_resv_wait_timeout(vm->root.bo->tbo.base.resv, true,
2895 true, timeout);
2896 if (timeout <= 0)
2897 return timeout;
2898
2899 return dma_fence_wait_timeout(vm->last_unlocked, true, timeout);
2900 }
2901
2902 /**
2903 * amdgpu_vm_init - initialize a vm instance
2904 *
2905 * @adev: amdgpu_device pointer
2906 * @vm: requested vm
2907 *
2908 * Init @vm fields.
2909 *
2910 * Returns:
2911 * 0 for success, error for failure.
2912 */
amdgpu_vm_init(struct amdgpu_device * adev,struct amdgpu_vm * vm)2913 int amdgpu_vm_init(struct amdgpu_device *adev, struct amdgpu_vm *vm)
2914 {
2915 struct amdgpu_bo *root_bo;
2916 struct amdgpu_bo_vm *root;
2917 int r, i;
2918
2919 vm->va = RB_ROOT_CACHED;
2920 for (i = 0; i < AMDGPU_MAX_VMHUBS; i++)
2921 vm->reserved_vmid[i] = NULL;
2922 INIT_LIST_HEAD(&vm->evicted);
2923 INIT_LIST_HEAD(&vm->relocated);
2924 INIT_LIST_HEAD(&vm->moved);
2925 INIT_LIST_HEAD(&vm->idle);
2926 INIT_LIST_HEAD(&vm->invalidated);
2927 spin_lock_init(&vm->invalidated_lock);
2928 INIT_LIST_HEAD(&vm->freed);
2929 INIT_LIST_HEAD(&vm->done);
2930
2931 /* create scheduler entities for page table updates */
2932 r = drm_sched_entity_init(&vm->immediate, DRM_SCHED_PRIORITY_NORMAL,
2933 adev->vm_manager.vm_pte_scheds,
2934 adev->vm_manager.vm_pte_num_scheds, NULL);
2935 if (r)
2936 return r;
2937
2938 r = drm_sched_entity_init(&vm->delayed, DRM_SCHED_PRIORITY_NORMAL,
2939 adev->vm_manager.vm_pte_scheds,
2940 adev->vm_manager.vm_pte_num_scheds, NULL);
2941 if (r)
2942 goto error_free_immediate;
2943
2944 vm->pte_support_ats = false;
2945 vm->is_compute_context = false;
2946
2947 vm->use_cpu_for_update = !!(adev->vm_manager.vm_update_mode &
2948 AMDGPU_VM_USE_CPU_FOR_GFX);
2949
2950 DRM_DEBUG_DRIVER("VM update mode is %s\n",
2951 vm->use_cpu_for_update ? "CPU" : "SDMA");
2952 WARN_ONCE((vm->use_cpu_for_update &&
2953 !amdgpu_gmc_vram_full_visible(&adev->gmc)),
2954 "CPU update of VM recommended only for large BAR system\n");
2955
2956 if (vm->use_cpu_for_update)
2957 vm->update_funcs = &amdgpu_vm_cpu_funcs;
2958 else
2959 vm->update_funcs = &amdgpu_vm_sdma_funcs;
2960 vm->last_update = NULL;
2961 vm->last_unlocked = dma_fence_get_stub();
2962
2963 mutex_init(&vm->eviction_lock);
2964 vm->evicting = false;
2965
2966 r = amdgpu_vm_pt_create(adev, vm, adev->vm_manager.root_level,
2967 false, &root);
2968 if (r)
2969 goto error_free_delayed;
2970 root_bo = &root->bo;
2971 r = amdgpu_bo_reserve(root_bo, true);
2972 if (r)
2973 goto error_free_root;
2974
2975 r = dma_resv_reserve_shared(root_bo->tbo.base.resv, 1);
2976 if (r)
2977 goto error_unreserve;
2978
2979 amdgpu_vm_bo_base_init(&vm->root, vm, root_bo);
2980
2981 r = amdgpu_vm_clear_bo(adev, vm, root, false);
2982 if (r)
2983 goto error_unreserve;
2984
2985 amdgpu_bo_unreserve(vm->root.bo);
2986
2987 INIT_KFIFO(vm->faults);
2988
2989 return 0;
2990
2991 error_unreserve:
2992 amdgpu_bo_unreserve(vm->root.bo);
2993
2994 error_free_root:
2995 amdgpu_bo_unref(&root->shadow);
2996 amdgpu_bo_unref(&root_bo);
2997 vm->root.bo = NULL;
2998
2999 error_free_delayed:
3000 dma_fence_put(vm->last_unlocked);
3001 drm_sched_entity_destroy(&vm->delayed);
3002
3003 error_free_immediate:
3004 drm_sched_entity_destroy(&vm->immediate);
3005
3006 return r;
3007 }
3008
3009 /**
3010 * amdgpu_vm_check_clean_reserved - check if a VM is clean
3011 *
3012 * @adev: amdgpu_device pointer
3013 * @vm: the VM to check
3014 *
3015 * check all entries of the root PD, if any subsequent PDs are allocated,
3016 * it means there are page table creating and filling, and is no a clean
3017 * VM
3018 *
3019 * Returns:
3020 * 0 if this VM is clean
3021 */
amdgpu_vm_check_clean_reserved(struct amdgpu_device * adev,struct amdgpu_vm * vm)3022 static int amdgpu_vm_check_clean_reserved(struct amdgpu_device *adev,
3023 struct amdgpu_vm *vm)
3024 {
3025 enum amdgpu_vm_level root = adev->vm_manager.root_level;
3026 unsigned int entries = amdgpu_vm_num_entries(adev, root);
3027 unsigned int i = 0;
3028
3029 for (i = 0; i < entries; i++) {
3030 if (to_amdgpu_bo_vm(vm->root.bo)->entries[i].bo)
3031 return -EINVAL;
3032 }
3033
3034 return 0;
3035 }
3036
3037 /**
3038 * amdgpu_vm_make_compute - Turn a GFX VM into a compute VM
3039 *
3040 * @adev: amdgpu_device pointer
3041 * @vm: requested vm
3042 *
3043 * This only works on GFX VMs that don't have any BOs added and no
3044 * page tables allocated yet.
3045 *
3046 * Changes the following VM parameters:
3047 * - use_cpu_for_update
3048 * - pte_supports_ats
3049 *
3050 * Reinitializes the page directory to reflect the changed ATS
3051 * setting.
3052 *
3053 * Returns:
3054 * 0 for success, -errno for errors.
3055 */
amdgpu_vm_make_compute(struct amdgpu_device * adev,struct amdgpu_vm * vm)3056 int amdgpu_vm_make_compute(struct amdgpu_device *adev, struct amdgpu_vm *vm)
3057 {
3058 bool pte_support_ats = (adev->asic_type == CHIP_RAVEN);
3059 int r;
3060
3061 r = amdgpu_bo_reserve(vm->root.bo, true);
3062 if (r)
3063 return r;
3064
3065 /* Sanity checks */
3066 r = amdgpu_vm_check_clean_reserved(adev, vm);
3067 if (r)
3068 goto unreserve_bo;
3069
3070 /* Check if PD needs to be reinitialized and do it before
3071 * changing any other state, in case it fails.
3072 */
3073 if (pte_support_ats != vm->pte_support_ats) {
3074 vm->pte_support_ats = pte_support_ats;
3075 r = amdgpu_vm_clear_bo(adev, vm,
3076 to_amdgpu_bo_vm(vm->root.bo),
3077 false);
3078 if (r)
3079 goto unreserve_bo;
3080 }
3081
3082 /* Update VM state */
3083 vm->use_cpu_for_update = !!(adev->vm_manager.vm_update_mode &
3084 AMDGPU_VM_USE_CPU_FOR_COMPUTE);
3085 DRM_DEBUG_DRIVER("VM update mode is %s\n",
3086 vm->use_cpu_for_update ? "CPU" : "SDMA");
3087 WARN_ONCE((vm->use_cpu_for_update &&
3088 !amdgpu_gmc_vram_full_visible(&adev->gmc)),
3089 "CPU update of VM recommended only for large BAR system\n");
3090
3091 if (vm->use_cpu_for_update) {
3092 /* Sync with last SDMA update/clear before switching to CPU */
3093 r = amdgpu_bo_sync_wait(vm->root.bo,
3094 AMDGPU_FENCE_OWNER_UNDEFINED, true);
3095 if (r)
3096 goto unreserve_bo;
3097
3098 vm->update_funcs = &amdgpu_vm_cpu_funcs;
3099 } else {
3100 vm->update_funcs = &amdgpu_vm_sdma_funcs;
3101 }
3102 dma_fence_put(vm->last_update);
3103 vm->last_update = NULL;
3104 vm->is_compute_context = true;
3105
3106 /* Free the shadow bo for compute VM */
3107 amdgpu_bo_unref(&to_amdgpu_bo_vm(vm->root.bo)->shadow);
3108
3109 goto unreserve_bo;
3110
3111 unreserve_bo:
3112 amdgpu_bo_unreserve(vm->root.bo);
3113 return r;
3114 }
3115
3116 /**
3117 * amdgpu_vm_release_compute - release a compute vm
3118 * @adev: amdgpu_device pointer
3119 * @vm: a vm turned into compute vm by calling amdgpu_vm_make_compute
3120 *
3121 * This is a correspondant of amdgpu_vm_make_compute. It decouples compute
3122 * pasid from vm. Compute should stop use of vm after this call.
3123 */
amdgpu_vm_release_compute(struct amdgpu_device * adev,struct amdgpu_vm * vm)3124 void amdgpu_vm_release_compute(struct amdgpu_device *adev, struct amdgpu_vm *vm)
3125 {
3126 amdgpu_vm_set_pasid(adev, vm, 0);
3127 vm->is_compute_context = false;
3128 }
3129
3130 /**
3131 * amdgpu_vm_fini - tear down a vm instance
3132 *
3133 * @adev: amdgpu_device pointer
3134 * @vm: requested vm
3135 *
3136 * Tear down @vm.
3137 * Unbind the VM and remove all bos from the vm bo list
3138 */
amdgpu_vm_fini(struct amdgpu_device * adev,struct amdgpu_vm * vm)3139 void amdgpu_vm_fini(struct amdgpu_device *adev, struct amdgpu_vm *vm)
3140 {
3141 struct amdgpu_bo_va_mapping *mapping, *tmp;
3142 bool prt_fini_needed = !!adev->gmc.gmc_funcs->set_prt;
3143 struct amdgpu_bo *root;
3144 int i;
3145
3146 amdgpu_amdkfd_gpuvm_destroy_cb(adev, vm);
3147
3148 root = amdgpu_bo_ref(vm->root.bo);
3149 amdgpu_bo_reserve(root, true);
3150 amdgpu_vm_set_pasid(adev, vm, 0);
3151 dma_fence_wait(vm->last_unlocked, false);
3152 dma_fence_put(vm->last_unlocked);
3153
3154 list_for_each_entry_safe(mapping, tmp, &vm->freed, list) {
3155 if (mapping->flags & AMDGPU_PTE_PRT && prt_fini_needed) {
3156 amdgpu_vm_prt_fini(adev, vm);
3157 prt_fini_needed = false;
3158 }
3159
3160 list_del(&mapping->list);
3161 amdgpu_vm_free_mapping(adev, vm, mapping, NULL);
3162 }
3163
3164 amdgpu_vm_free_pts(adev, vm, NULL);
3165 amdgpu_bo_unreserve(root);
3166 amdgpu_bo_unref(&root);
3167 WARN_ON(vm->root.bo);
3168
3169 drm_sched_entity_destroy(&vm->immediate);
3170 drm_sched_entity_destroy(&vm->delayed);
3171
3172 if (!RB_EMPTY_ROOT(&vm->va.rb_root)) {
3173 dev_err(adev->dev, "still active bo inside vm\n");
3174 }
3175 rbtree_postorder_for_each_entry_safe(mapping, tmp,
3176 &vm->va.rb_root, rb) {
3177 /* Don't remove the mapping here, we don't want to trigger a
3178 * rebalance and the tree is about to be destroyed anyway.
3179 */
3180 list_del(&mapping->list);
3181 kfree(mapping);
3182 }
3183
3184 dma_fence_put(vm->last_update);
3185 for (i = 0; i < AMDGPU_MAX_VMHUBS; i++)
3186 amdgpu_vmid_free_reserved(adev, vm, i);
3187 }
3188
3189 /**
3190 * amdgpu_vm_manager_init - init the VM manager
3191 *
3192 * @adev: amdgpu_device pointer
3193 *
3194 * Initialize the VM manager structures
3195 */
amdgpu_vm_manager_init(struct amdgpu_device * adev)3196 void amdgpu_vm_manager_init(struct amdgpu_device *adev)
3197 {
3198 unsigned i;
3199
3200 /* Concurrent flushes are only possible starting with Vega10 and
3201 * are broken on Navi10 and Navi14.
3202 */
3203 adev->vm_manager.concurrent_flush = !(adev->asic_type < CHIP_VEGA10 ||
3204 adev->asic_type == CHIP_NAVI10 ||
3205 adev->asic_type == CHIP_NAVI14);
3206 amdgpu_vmid_mgr_init(adev);
3207
3208 adev->vm_manager.fence_context =
3209 dma_fence_context_alloc(AMDGPU_MAX_RINGS);
3210 for (i = 0; i < AMDGPU_MAX_RINGS; ++i)
3211 adev->vm_manager.seqno[i] = 0;
3212
3213 spin_lock_init(&adev->vm_manager.prt_lock);
3214 atomic_set(&adev->vm_manager.num_prt_users, 0);
3215
3216 /* If not overridden by the user, by default, only in large BAR systems
3217 * Compute VM tables will be updated by CPU
3218 */
3219 #ifdef CONFIG_X86_64
3220 if (amdgpu_vm_update_mode == -1) {
3221 if (amdgpu_gmc_vram_full_visible(&adev->gmc))
3222 adev->vm_manager.vm_update_mode =
3223 AMDGPU_VM_USE_CPU_FOR_COMPUTE;
3224 else
3225 adev->vm_manager.vm_update_mode = 0;
3226 } else
3227 adev->vm_manager.vm_update_mode = amdgpu_vm_update_mode;
3228 #else
3229 adev->vm_manager.vm_update_mode = 0;
3230 #endif
3231
3232 xa_init_flags(&adev->vm_manager.pasids, XA_FLAGS_LOCK_IRQ);
3233 }
3234
3235 /**
3236 * amdgpu_vm_manager_fini - cleanup VM manager
3237 *
3238 * @adev: amdgpu_device pointer
3239 *
3240 * Cleanup the VM manager and free resources.
3241 */
amdgpu_vm_manager_fini(struct amdgpu_device * adev)3242 void amdgpu_vm_manager_fini(struct amdgpu_device *adev)
3243 {
3244 WARN_ON(!xa_empty(&adev->vm_manager.pasids));
3245 xa_destroy(&adev->vm_manager.pasids);
3246
3247 amdgpu_vmid_mgr_fini(adev);
3248 }
3249
3250 /**
3251 * amdgpu_vm_ioctl - Manages VMID reservation for vm hubs.
3252 *
3253 * @dev: drm device pointer
3254 * @data: drm_amdgpu_vm
3255 * @filp: drm file pointer
3256 *
3257 * Returns:
3258 * 0 for success, -errno for errors.
3259 */
amdgpu_vm_ioctl(struct drm_device * dev,void * data,struct drm_file * filp)3260 int amdgpu_vm_ioctl(struct drm_device *dev, void *data, struct drm_file *filp)
3261 {
3262 union drm_amdgpu_vm *args = data;
3263 struct amdgpu_device *adev = drm_to_adev(dev);
3264 struct amdgpu_fpriv *fpriv = filp->driver_priv;
3265 long timeout = msecs_to_jiffies(2000);
3266 int r;
3267
3268 switch (args->in.op) {
3269 case AMDGPU_VM_OP_RESERVE_VMID:
3270 /* We only have requirement to reserve vmid from gfxhub */
3271 r = amdgpu_vmid_alloc_reserved(adev, &fpriv->vm,
3272 AMDGPU_GFXHUB_0);
3273 if (r)
3274 return r;
3275 break;
3276 case AMDGPU_VM_OP_UNRESERVE_VMID:
3277 if (amdgpu_sriov_runtime(adev))
3278 timeout = 8 * timeout;
3279
3280 /* Wait vm idle to make sure the vmid set in SPM_VMID is
3281 * not referenced anymore.
3282 */
3283 r = amdgpu_bo_reserve(fpriv->vm.root.bo, true);
3284 if (r)
3285 return r;
3286
3287 r = amdgpu_vm_wait_idle(&fpriv->vm, timeout);
3288 if (r < 0)
3289 return r;
3290
3291 amdgpu_bo_unreserve(fpriv->vm.root.bo);
3292 amdgpu_vmid_free_reserved(adev, &fpriv->vm, AMDGPU_GFXHUB_0);
3293 break;
3294 default:
3295 return -EINVAL;
3296 }
3297
3298 return 0;
3299 }
3300
3301 /**
3302 * amdgpu_vm_get_task_info - Extracts task info for a PASID.
3303 *
3304 * @adev: drm device pointer
3305 * @pasid: PASID identifier for VM
3306 * @task_info: task_info to fill.
3307 */
amdgpu_vm_get_task_info(struct amdgpu_device * adev,u32 pasid,struct amdgpu_task_info * task_info)3308 void amdgpu_vm_get_task_info(struct amdgpu_device *adev, u32 pasid,
3309 struct amdgpu_task_info *task_info)
3310 {
3311 struct amdgpu_vm *vm;
3312 unsigned long flags;
3313
3314 xa_lock_irqsave(&adev->vm_manager.pasids, flags);
3315
3316 vm = xa_load(&adev->vm_manager.pasids, pasid);
3317 if (vm)
3318 *task_info = vm->task_info;
3319
3320 xa_unlock_irqrestore(&adev->vm_manager.pasids, flags);
3321 }
3322
3323 /**
3324 * amdgpu_vm_set_task_info - Sets VMs task info.
3325 *
3326 * @vm: vm for which to set the info
3327 */
amdgpu_vm_set_task_info(struct amdgpu_vm * vm)3328 void amdgpu_vm_set_task_info(struct amdgpu_vm *vm)
3329 {
3330 if (vm->task_info.pid)
3331 return;
3332
3333 vm->task_info.pid = current->pid;
3334 get_task_comm(vm->task_info.task_name, current);
3335
3336 if (current->group_leader->mm != current->mm)
3337 return;
3338
3339 vm->task_info.tgid = current->group_leader->pid;
3340 get_task_comm(vm->task_info.process_name, current->group_leader);
3341 }
3342
3343 /**
3344 * amdgpu_vm_handle_fault - graceful handling of VM faults.
3345 * @adev: amdgpu device pointer
3346 * @pasid: PASID of the VM
3347 * @addr: Address of the fault
3348 * @write_fault: true is write fault, false is read fault
3349 *
3350 * Try to gracefully handle a VM fault. Return true if the fault was handled and
3351 * shouldn't be reported any more.
3352 */
amdgpu_vm_handle_fault(struct amdgpu_device * adev,u32 pasid,uint64_t addr,bool write_fault)3353 bool amdgpu_vm_handle_fault(struct amdgpu_device *adev, u32 pasid,
3354 uint64_t addr, bool write_fault)
3355 {
3356 bool is_compute_context = false;
3357 struct amdgpu_bo *root;
3358 unsigned long irqflags;
3359 uint64_t value, flags;
3360 struct amdgpu_vm *vm;
3361 int r;
3362
3363 xa_lock_irqsave(&adev->vm_manager.pasids, irqflags);
3364 vm = xa_load(&adev->vm_manager.pasids, pasid);
3365 if (vm) {
3366 root = amdgpu_bo_ref(vm->root.bo);
3367 is_compute_context = vm->is_compute_context;
3368 } else {
3369 root = NULL;
3370 }
3371 xa_unlock_irqrestore(&adev->vm_manager.pasids, irqflags);
3372
3373 if (!root)
3374 return false;
3375
3376 addr /= AMDGPU_GPU_PAGE_SIZE;
3377
3378 if (is_compute_context &&
3379 !svm_range_restore_pages(adev, pasid, addr, write_fault)) {
3380 amdgpu_bo_unref(&root);
3381 return true;
3382 }
3383
3384 r = amdgpu_bo_reserve(root, true);
3385 if (r)
3386 goto error_unref;
3387
3388 /* Double check that the VM still exists */
3389 xa_lock_irqsave(&adev->vm_manager.pasids, irqflags);
3390 vm = xa_load(&adev->vm_manager.pasids, pasid);
3391 if (vm && vm->root.bo != root)
3392 vm = NULL;
3393 xa_unlock_irqrestore(&adev->vm_manager.pasids, irqflags);
3394 if (!vm)
3395 goto error_unlock;
3396
3397 flags = AMDGPU_PTE_VALID | AMDGPU_PTE_SNOOPED |
3398 AMDGPU_PTE_SYSTEM;
3399
3400 if (is_compute_context) {
3401 /* Intentionally setting invalid PTE flag
3402 * combination to force a no-retry-fault
3403 */
3404 flags = AMDGPU_PTE_EXECUTABLE | AMDGPU_PDE_PTE |
3405 AMDGPU_PTE_TF;
3406 value = 0;
3407 } else if (amdgpu_vm_fault_stop == AMDGPU_VM_FAULT_STOP_NEVER) {
3408 /* Redirect the access to the dummy page */
3409 value = adev->dummy_page_addr;
3410 flags |= AMDGPU_PTE_EXECUTABLE | AMDGPU_PTE_READABLE |
3411 AMDGPU_PTE_WRITEABLE;
3412
3413 } else {
3414 /* Let the hw retry silently on the PTE */
3415 value = 0;
3416 }
3417
3418 r = dma_resv_reserve_shared(root->tbo.base.resv, 1);
3419 if (r) {
3420 pr_debug("failed %d to reserve fence slot\n", r);
3421 goto error_unlock;
3422 }
3423
3424 r = amdgpu_vm_bo_update_mapping(adev, adev, vm, true, false, NULL, addr,
3425 addr, flags, value, NULL, NULL, NULL,
3426 NULL);
3427 if (r)
3428 goto error_unlock;
3429
3430 r = amdgpu_vm_update_pdes(adev, vm, true);
3431
3432 error_unlock:
3433 amdgpu_bo_unreserve(root);
3434 if (r < 0)
3435 DRM_ERROR("Can't handle page fault (%d)\n", r);
3436
3437 error_unref:
3438 amdgpu_bo_unref(&root);
3439
3440 return false;
3441 }
3442
3443 #if defined(CONFIG_DEBUG_FS)
3444 /**
3445 * amdgpu_debugfs_vm_bo_info - print BO info for the VM
3446 *
3447 * @vm: Requested VM for printing BO info
3448 * @m: debugfs file
3449 *
3450 * Print BO information in debugfs file for the VM
3451 */
amdgpu_debugfs_vm_bo_info(struct amdgpu_vm * vm,struct seq_file * m)3452 void amdgpu_debugfs_vm_bo_info(struct amdgpu_vm *vm, struct seq_file *m)
3453 {
3454 struct amdgpu_bo_va *bo_va, *tmp;
3455 u64 total_idle = 0;
3456 u64 total_evicted = 0;
3457 u64 total_relocated = 0;
3458 u64 total_moved = 0;
3459 u64 total_invalidated = 0;
3460 u64 total_done = 0;
3461 unsigned int total_idle_objs = 0;
3462 unsigned int total_evicted_objs = 0;
3463 unsigned int total_relocated_objs = 0;
3464 unsigned int total_moved_objs = 0;
3465 unsigned int total_invalidated_objs = 0;
3466 unsigned int total_done_objs = 0;
3467 unsigned int id = 0;
3468
3469 seq_puts(m, "\tIdle BOs:\n");
3470 list_for_each_entry_safe(bo_va, tmp, &vm->idle, base.vm_status) {
3471 if (!bo_va->base.bo)
3472 continue;
3473 total_idle += amdgpu_bo_print_info(id++, bo_va->base.bo, m);
3474 }
3475 total_idle_objs = id;
3476 id = 0;
3477
3478 seq_puts(m, "\tEvicted BOs:\n");
3479 list_for_each_entry_safe(bo_va, tmp, &vm->evicted, base.vm_status) {
3480 if (!bo_va->base.bo)
3481 continue;
3482 total_evicted += amdgpu_bo_print_info(id++, bo_va->base.bo, m);
3483 }
3484 total_evicted_objs = id;
3485 id = 0;
3486
3487 seq_puts(m, "\tRelocated BOs:\n");
3488 list_for_each_entry_safe(bo_va, tmp, &vm->relocated, base.vm_status) {
3489 if (!bo_va->base.bo)
3490 continue;
3491 total_relocated += amdgpu_bo_print_info(id++, bo_va->base.bo, m);
3492 }
3493 total_relocated_objs = id;
3494 id = 0;
3495
3496 seq_puts(m, "\tMoved BOs:\n");
3497 list_for_each_entry_safe(bo_va, tmp, &vm->moved, base.vm_status) {
3498 if (!bo_va->base.bo)
3499 continue;
3500 total_moved += amdgpu_bo_print_info(id++, bo_va->base.bo, m);
3501 }
3502 total_moved_objs = id;
3503 id = 0;
3504
3505 seq_puts(m, "\tInvalidated BOs:\n");
3506 spin_lock(&vm->invalidated_lock);
3507 list_for_each_entry_safe(bo_va, tmp, &vm->invalidated, base.vm_status) {
3508 if (!bo_va->base.bo)
3509 continue;
3510 total_invalidated += amdgpu_bo_print_info(id++, bo_va->base.bo, m);
3511 }
3512 total_invalidated_objs = id;
3513 id = 0;
3514
3515 seq_puts(m, "\tDone BOs:\n");
3516 list_for_each_entry_safe(bo_va, tmp, &vm->done, base.vm_status) {
3517 if (!bo_va->base.bo)
3518 continue;
3519 total_done += amdgpu_bo_print_info(id++, bo_va->base.bo, m);
3520 }
3521 spin_unlock(&vm->invalidated_lock);
3522 total_done_objs = id;
3523
3524 seq_printf(m, "\tTotal idle size: %12lld\tobjs:\t%d\n", total_idle,
3525 total_idle_objs);
3526 seq_printf(m, "\tTotal evicted size: %12lld\tobjs:\t%d\n", total_evicted,
3527 total_evicted_objs);
3528 seq_printf(m, "\tTotal relocated size: %12lld\tobjs:\t%d\n", total_relocated,
3529 total_relocated_objs);
3530 seq_printf(m, "\tTotal moved size: %12lld\tobjs:\t%d\n", total_moved,
3531 total_moved_objs);
3532 seq_printf(m, "\tTotal invalidated size: %12lld\tobjs:\t%d\n", total_invalidated,
3533 total_invalidated_objs);
3534 seq_printf(m, "\tTotal done size: %12lld\tobjs:\t%d\n", total_done,
3535 total_done_objs);
3536 }
3537 #endif
3538