Loading...
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 <drm/ttm/ttm_tt.h>
37#include <drm/drm_exec.h>
38#include "amdgpu.h"
39#include "amdgpu_trace.h"
40#include "amdgpu_amdkfd.h"
41#include "amdgpu_gmc.h"
42#include "amdgpu_xgmi.h"
43#include "amdgpu_dma_buf.h"
44#include "amdgpu_res_cursor.h"
45#include "kfd_svm.h"
46
47/**
48 * DOC: GPUVM
49 *
50 * GPUVM is the MMU functionality provided on the GPU.
51 * GPUVM is similar to the legacy GART on older asics, however
52 * rather than there being a single global GART table
53 * for the entire GPU, there can be multiple GPUVM page tables active
54 * at any given time. The GPUVM page tables can contain a mix
55 * VRAM pages and system pages (both memory and MMIO) and system pages
56 * can be mapped as snooped (cached system pages) or unsnooped
57 * (uncached system pages).
58 *
59 * Each active GPUVM has an ID associated with it and there is a page table
60 * linked with each VMID. When executing a command buffer,
61 * the kernel tells the engine what VMID to use for that command
62 * buffer. VMIDs are allocated dynamically as commands are submitted.
63 * The userspace drivers maintain their own address space and the kernel
64 * sets up their pages tables accordingly when they submit their
65 * command buffers and a VMID is assigned.
66 * The hardware supports up to 16 active GPUVMs at any given time.
67 *
68 * Each GPUVM is represented by a 1-2 or 1-5 level page table, depending
69 * on the ASIC family. GPUVM supports RWX attributes on each page as well
70 * as other features such as encryption and caching attributes.
71 *
72 * VMID 0 is special. It is the GPUVM used for the kernel driver. In
73 * addition to an aperture managed by a page table, VMID 0 also has
74 * several other apertures. There is an aperture for direct access to VRAM
75 * and there is a legacy AGP aperture which just forwards accesses directly
76 * to the matching system physical addresses (or IOVAs when an IOMMU is
77 * present). These apertures provide direct access to these memories without
78 * incurring the overhead of a page table. VMID 0 is used by the kernel
79 * driver for tasks like memory management.
80 *
81 * GPU clients (i.e., engines on the GPU) use GPUVM VMIDs to access memory.
82 * For user applications, each application can have their own unique GPUVM
83 * address space. The application manages the address space and the kernel
84 * driver manages the GPUVM page tables for each process. If an GPU client
85 * accesses an invalid page, it will generate a GPU page fault, similar to
86 * accessing an invalid page on a CPU.
87 */
88
89#define START(node) ((node)->start)
90#define LAST(node) ((node)->last)
91
92INTERVAL_TREE_DEFINE(struct amdgpu_bo_va_mapping, rb, uint64_t, __subtree_last,
93 START, LAST, static, amdgpu_vm_it)
94
95#undef START
96#undef LAST
97
98/**
99 * struct amdgpu_prt_cb - Helper to disable partial resident texture feature from a fence callback
100 */
101struct amdgpu_prt_cb {
102
103 /**
104 * @adev: amdgpu device
105 */
106 struct amdgpu_device *adev;
107
108 /**
109 * @cb: callback
110 */
111 struct dma_fence_cb cb;
112};
113
114/**
115 * struct amdgpu_vm_tlb_seq_struct - Helper to increment the TLB flush sequence
116 */
117struct amdgpu_vm_tlb_seq_struct {
118 /**
119 * @vm: pointer to the amdgpu_vm structure to set the fence sequence on
120 */
121 struct amdgpu_vm *vm;
122
123 /**
124 * @cb: callback
125 */
126 struct dma_fence_cb cb;
127};
128
129/**
130 * amdgpu_vm_set_pasid - manage pasid and vm ptr mapping
131 *
132 * @adev: amdgpu_device pointer
133 * @vm: amdgpu_vm pointer
134 * @pasid: the pasid the VM is using on this GPU
135 *
136 * Set the pasid this VM is using on this GPU, can also be used to remove the
137 * pasid by passing in zero.
138 *
139 */
140int amdgpu_vm_set_pasid(struct amdgpu_device *adev, struct amdgpu_vm *vm,
141 u32 pasid)
142{
143 int r;
144
145 if (vm->pasid == pasid)
146 return 0;
147
148 if (vm->pasid) {
149 r = xa_err(xa_erase_irq(&adev->vm_manager.pasids, vm->pasid));
150 if (r < 0)
151 return r;
152
153 vm->pasid = 0;
154 }
155
156 if (pasid) {
157 r = xa_err(xa_store_irq(&adev->vm_manager.pasids, pasid, vm,
158 GFP_KERNEL));
159 if (r < 0)
160 return r;
161
162 vm->pasid = pasid;
163 }
164
165
166 return 0;
167}
168
169/**
170 * amdgpu_vm_bo_evicted - vm_bo is evicted
171 *
172 * @vm_bo: vm_bo which is evicted
173 *
174 * State for PDs/PTs and per VM BOs which are not at the location they should
175 * be.
176 */
177static void amdgpu_vm_bo_evicted(struct amdgpu_vm_bo_base *vm_bo)
178{
179 struct amdgpu_vm *vm = vm_bo->vm;
180 struct amdgpu_bo *bo = vm_bo->bo;
181
182 vm_bo->moved = true;
183 spin_lock(&vm_bo->vm->status_lock);
184 if (bo->tbo.type == ttm_bo_type_kernel)
185 list_move(&vm_bo->vm_status, &vm->evicted);
186 else
187 list_move_tail(&vm_bo->vm_status, &vm->evicted);
188 spin_unlock(&vm_bo->vm->status_lock);
189}
190/**
191 * amdgpu_vm_bo_moved - vm_bo is moved
192 *
193 * @vm_bo: vm_bo which is moved
194 *
195 * State for per VM BOs which are moved, but that change is not yet reflected
196 * in the page tables.
197 */
198static void amdgpu_vm_bo_moved(struct amdgpu_vm_bo_base *vm_bo)
199{
200 spin_lock(&vm_bo->vm->status_lock);
201 list_move(&vm_bo->vm_status, &vm_bo->vm->moved);
202 spin_unlock(&vm_bo->vm->status_lock);
203}
204
205/**
206 * amdgpu_vm_bo_idle - vm_bo is idle
207 *
208 * @vm_bo: vm_bo which is now idle
209 *
210 * State for PDs/PTs and per VM BOs which have gone through the state machine
211 * and are now idle.
212 */
213static void amdgpu_vm_bo_idle(struct amdgpu_vm_bo_base *vm_bo)
214{
215 spin_lock(&vm_bo->vm->status_lock);
216 list_move(&vm_bo->vm_status, &vm_bo->vm->idle);
217 spin_unlock(&vm_bo->vm->status_lock);
218 vm_bo->moved = false;
219}
220
221/**
222 * amdgpu_vm_bo_invalidated - vm_bo is invalidated
223 *
224 * @vm_bo: vm_bo which is now invalidated
225 *
226 * State for normal BOs which are invalidated and that change not yet reflected
227 * in the PTs.
228 */
229static void amdgpu_vm_bo_invalidated(struct amdgpu_vm_bo_base *vm_bo)
230{
231 spin_lock(&vm_bo->vm->status_lock);
232 list_move(&vm_bo->vm_status, &vm_bo->vm->invalidated);
233 spin_unlock(&vm_bo->vm->status_lock);
234}
235
236/**
237 * amdgpu_vm_bo_evicted_user - vm_bo is evicted
238 *
239 * @vm_bo: vm_bo which is evicted
240 *
241 * State for BOs used by user mode queues which are not at the location they
242 * should be.
243 */
244static void amdgpu_vm_bo_evicted_user(struct amdgpu_vm_bo_base *vm_bo)
245{
246 vm_bo->moved = true;
247 spin_lock(&vm_bo->vm->status_lock);
248 list_move(&vm_bo->vm_status, &vm_bo->vm->evicted_user);
249 spin_unlock(&vm_bo->vm->status_lock);
250}
251
252/**
253 * amdgpu_vm_bo_relocated - vm_bo is reloacted
254 *
255 * @vm_bo: vm_bo which is relocated
256 *
257 * State for PDs/PTs which needs to update their parent PD.
258 * For the root PD, just move to idle state.
259 */
260static void amdgpu_vm_bo_relocated(struct amdgpu_vm_bo_base *vm_bo)
261{
262 if (vm_bo->bo->parent) {
263 spin_lock(&vm_bo->vm->status_lock);
264 list_move(&vm_bo->vm_status, &vm_bo->vm->relocated);
265 spin_unlock(&vm_bo->vm->status_lock);
266 } else {
267 amdgpu_vm_bo_idle(vm_bo);
268 }
269}
270
271/**
272 * amdgpu_vm_bo_done - vm_bo is done
273 *
274 * @vm_bo: vm_bo which is now done
275 *
276 * State for normal BOs which are invalidated and that change has been updated
277 * in the PTs.
278 */
279static void amdgpu_vm_bo_done(struct amdgpu_vm_bo_base *vm_bo)
280{
281 spin_lock(&vm_bo->vm->status_lock);
282 list_move(&vm_bo->vm_status, &vm_bo->vm->done);
283 spin_unlock(&vm_bo->vm->status_lock);
284}
285
286/**
287 * amdgpu_vm_bo_reset_state_machine - reset the vm_bo state machine
288 * @vm: the VM which state machine to reset
289 *
290 * Move all vm_bo object in the VM into a state where they will be updated
291 * again during validation.
292 */
293static void amdgpu_vm_bo_reset_state_machine(struct amdgpu_vm *vm)
294{
295 struct amdgpu_vm_bo_base *vm_bo, *tmp;
296
297 spin_lock(&vm->status_lock);
298 list_splice_init(&vm->done, &vm->invalidated);
299 list_for_each_entry(vm_bo, &vm->invalidated, vm_status)
300 vm_bo->moved = true;
301 list_for_each_entry_safe(vm_bo, tmp, &vm->idle, vm_status) {
302 struct amdgpu_bo *bo = vm_bo->bo;
303
304 vm_bo->moved = true;
305 if (!bo || bo->tbo.type != ttm_bo_type_kernel)
306 list_move(&vm_bo->vm_status, &vm_bo->vm->moved);
307 else if (bo->parent)
308 list_move(&vm_bo->vm_status, &vm_bo->vm->relocated);
309 }
310 spin_unlock(&vm->status_lock);
311}
312
313/**
314 * amdgpu_vm_bo_base_init - Adds bo to the list of bos associated with the vm
315 *
316 * @base: base structure for tracking BO usage in a VM
317 * @vm: vm to which bo is to be added
318 * @bo: amdgpu buffer object
319 *
320 * Initialize a bo_va_base structure and add it to the appropriate lists
321 *
322 */
323void amdgpu_vm_bo_base_init(struct amdgpu_vm_bo_base *base,
324 struct amdgpu_vm *vm, struct amdgpu_bo *bo)
325{
326 base->vm = vm;
327 base->bo = bo;
328 base->next = NULL;
329 INIT_LIST_HEAD(&base->vm_status);
330
331 if (!bo)
332 return;
333 base->next = bo->vm_bo;
334 bo->vm_bo = base;
335
336 if (bo->tbo.base.resv != vm->root.bo->tbo.base.resv)
337 return;
338
339 dma_resv_assert_held(vm->root.bo->tbo.base.resv);
340
341 ttm_bo_set_bulk_move(&bo->tbo, &vm->lru_bulk_move);
342 if (bo->tbo.type == ttm_bo_type_kernel && bo->parent)
343 amdgpu_vm_bo_relocated(base);
344 else
345 amdgpu_vm_bo_idle(base);
346
347 if (bo->preferred_domains &
348 amdgpu_mem_type_to_domain(bo->tbo.resource->mem_type))
349 return;
350
351 /*
352 * we checked all the prerequisites, but it looks like this per vm bo
353 * is currently evicted. add the bo to the evicted list to make sure it
354 * is validated on next vm use to avoid fault.
355 * */
356 amdgpu_vm_bo_evicted(base);
357}
358
359/**
360 * amdgpu_vm_lock_pd - lock PD in drm_exec
361 *
362 * @vm: vm providing the BOs
363 * @exec: drm execution context
364 * @num_fences: number of extra fences to reserve
365 *
366 * Lock the VM root PD in the DRM execution context.
367 */
368int amdgpu_vm_lock_pd(struct amdgpu_vm *vm, struct drm_exec *exec,
369 unsigned int num_fences)
370{
371 /* We need at least two fences for the VM PD/PT updates */
372 return drm_exec_prepare_obj(exec, &vm->root.bo->tbo.base,
373 2 + num_fences);
374}
375
376/**
377 * amdgpu_vm_move_to_lru_tail - move all BOs to the end of LRU
378 *
379 * @adev: amdgpu device pointer
380 * @vm: vm providing the BOs
381 *
382 * Move all BOs to the end of LRU and remember their positions to put them
383 * together.
384 */
385void amdgpu_vm_move_to_lru_tail(struct amdgpu_device *adev,
386 struct amdgpu_vm *vm)
387{
388 spin_lock(&adev->mman.bdev.lru_lock);
389 ttm_lru_bulk_move_tail(&vm->lru_bulk_move);
390 spin_unlock(&adev->mman.bdev.lru_lock);
391}
392
393/* Create scheduler entities for page table updates */
394static int amdgpu_vm_init_entities(struct amdgpu_device *adev,
395 struct amdgpu_vm *vm)
396{
397 int r;
398
399 r = drm_sched_entity_init(&vm->immediate, DRM_SCHED_PRIORITY_NORMAL,
400 adev->vm_manager.vm_pte_scheds,
401 adev->vm_manager.vm_pte_num_scheds, NULL);
402 if (r)
403 goto error;
404
405 return drm_sched_entity_init(&vm->delayed, DRM_SCHED_PRIORITY_NORMAL,
406 adev->vm_manager.vm_pte_scheds,
407 adev->vm_manager.vm_pte_num_scheds, NULL);
408
409error:
410 drm_sched_entity_destroy(&vm->immediate);
411 return r;
412}
413
414/* Destroy the entities for page table updates again */
415static void amdgpu_vm_fini_entities(struct amdgpu_vm *vm)
416{
417 drm_sched_entity_destroy(&vm->immediate);
418 drm_sched_entity_destroy(&vm->delayed);
419}
420
421/**
422 * amdgpu_vm_generation - return the page table re-generation counter
423 * @adev: the amdgpu_device
424 * @vm: optional VM to check, might be NULL
425 *
426 * Returns a page table re-generation token to allow checking if submissions
427 * are still valid to use this VM. The VM parameter might be NULL in which case
428 * just the VRAM lost counter will be used.
429 */
430uint64_t amdgpu_vm_generation(struct amdgpu_device *adev, struct amdgpu_vm *vm)
431{
432 uint64_t result = (u64)atomic_read(&adev->vram_lost_counter) << 32;
433
434 if (!vm)
435 return result;
436
437 result += vm->generation;
438 /* Add one if the page tables will be re-generated on next CS */
439 if (drm_sched_entity_error(&vm->delayed))
440 ++result;
441
442 return result;
443}
444
445/**
446 * amdgpu_vm_validate - validate evicted BOs tracked in the VM
447 *
448 * @adev: amdgpu device pointer
449 * @vm: vm providing the BOs
450 * @ticket: optional reservation ticket used to reserve the VM
451 * @validate: callback to do the validation
452 * @param: parameter for the validation callback
453 *
454 * Validate the page table BOs and per-VM BOs on command submission if
455 * necessary. If a ticket is given, also try to validate evicted user queue
456 * BOs. They must already be reserved with the given ticket.
457 *
458 * Returns:
459 * Validation result.
460 */
461int amdgpu_vm_validate(struct amdgpu_device *adev, struct amdgpu_vm *vm,
462 struct ww_acquire_ctx *ticket,
463 int (*validate)(void *p, struct amdgpu_bo *bo),
464 void *param)
465{
466 struct amdgpu_vm_bo_base *bo_base;
467 struct amdgpu_bo *shadow;
468 struct amdgpu_bo *bo;
469 int r;
470
471 if (drm_sched_entity_error(&vm->delayed)) {
472 ++vm->generation;
473 amdgpu_vm_bo_reset_state_machine(vm);
474 amdgpu_vm_fini_entities(vm);
475 r = amdgpu_vm_init_entities(adev, vm);
476 if (r)
477 return r;
478 }
479
480 spin_lock(&vm->status_lock);
481 while (!list_empty(&vm->evicted)) {
482 bo_base = list_first_entry(&vm->evicted,
483 struct amdgpu_vm_bo_base,
484 vm_status);
485 spin_unlock(&vm->status_lock);
486
487 bo = bo_base->bo;
488 shadow = amdgpu_bo_shadowed(bo);
489
490 r = validate(param, bo);
491 if (r)
492 return r;
493 if (shadow) {
494 r = validate(param, shadow);
495 if (r)
496 return r;
497 }
498
499 if (bo->tbo.type != ttm_bo_type_kernel) {
500 amdgpu_vm_bo_moved(bo_base);
501 } else {
502 vm->update_funcs->map_table(to_amdgpu_bo_vm(bo));
503 amdgpu_vm_bo_relocated(bo_base);
504 }
505 spin_lock(&vm->status_lock);
506 }
507 while (ticket && !list_empty(&vm->evicted_user)) {
508 bo_base = list_first_entry(&vm->evicted_user,
509 struct amdgpu_vm_bo_base,
510 vm_status);
511 spin_unlock(&vm->status_lock);
512
513 bo = bo_base->bo;
514
515 if (dma_resv_locking_ctx(bo->tbo.base.resv) != ticket) {
516 struct amdgpu_task_info *ti = amdgpu_vm_get_task_info_vm(vm);
517
518 pr_warn_ratelimited("Evicted user BO is not reserved\n");
519 if (ti) {
520 pr_warn_ratelimited("pid %d\n", ti->pid);
521 amdgpu_vm_put_task_info(ti);
522 }
523
524 return -EINVAL;
525 }
526
527 r = validate(param, bo);
528 if (r)
529 return r;
530
531 amdgpu_vm_bo_invalidated(bo_base);
532
533 spin_lock(&vm->status_lock);
534 }
535 spin_unlock(&vm->status_lock);
536
537 amdgpu_vm_eviction_lock(vm);
538 vm->evicting = false;
539 amdgpu_vm_eviction_unlock(vm);
540
541 return 0;
542}
543
544/**
545 * amdgpu_vm_ready - check VM is ready for updates
546 *
547 * @vm: VM to check
548 *
549 * Check if all VM PDs/PTs are ready for updates
550 *
551 * Returns:
552 * True if VM is not evicting.
553 */
554bool amdgpu_vm_ready(struct amdgpu_vm *vm)
555{
556 bool empty;
557 bool ret;
558
559 amdgpu_vm_eviction_lock(vm);
560 ret = !vm->evicting;
561 amdgpu_vm_eviction_unlock(vm);
562
563 spin_lock(&vm->status_lock);
564 empty = list_empty(&vm->evicted);
565 spin_unlock(&vm->status_lock);
566
567 return ret && empty;
568}
569
570/**
571 * amdgpu_vm_check_compute_bug - check whether asic has compute vm bug
572 *
573 * @adev: amdgpu_device pointer
574 */
575void amdgpu_vm_check_compute_bug(struct amdgpu_device *adev)
576{
577 const struct amdgpu_ip_block *ip_block;
578 bool has_compute_vm_bug;
579 struct amdgpu_ring *ring;
580 int i;
581
582 has_compute_vm_bug = false;
583
584 ip_block = amdgpu_device_ip_get_ip_block(adev, AMD_IP_BLOCK_TYPE_GFX);
585 if (ip_block) {
586 /* Compute has a VM bug for GFX version < 7.
587 Compute has a VM bug for GFX 8 MEC firmware version < 673.*/
588 if (ip_block->version->major <= 7)
589 has_compute_vm_bug = true;
590 else if (ip_block->version->major == 8)
591 if (adev->gfx.mec_fw_version < 673)
592 has_compute_vm_bug = true;
593 }
594
595 for (i = 0; i < adev->num_rings; i++) {
596 ring = adev->rings[i];
597 if (ring->funcs->type == AMDGPU_RING_TYPE_COMPUTE)
598 /* only compute rings */
599 ring->has_compute_vm_bug = has_compute_vm_bug;
600 else
601 ring->has_compute_vm_bug = false;
602 }
603}
604
605/**
606 * amdgpu_vm_need_pipeline_sync - Check if pipe sync is needed for job.
607 *
608 * @ring: ring on which the job will be submitted
609 * @job: job to submit
610 *
611 * Returns:
612 * True if sync is needed.
613 */
614bool amdgpu_vm_need_pipeline_sync(struct amdgpu_ring *ring,
615 struct amdgpu_job *job)
616{
617 struct amdgpu_device *adev = ring->adev;
618 unsigned vmhub = ring->vm_hub;
619 struct amdgpu_vmid_mgr *id_mgr = &adev->vm_manager.id_mgr[vmhub];
620
621 if (job->vmid == 0)
622 return false;
623
624 if (job->vm_needs_flush || ring->has_compute_vm_bug)
625 return true;
626
627 if (ring->funcs->emit_gds_switch && job->gds_switch_needed)
628 return true;
629
630 if (amdgpu_vmid_had_gpu_reset(adev, &id_mgr->ids[job->vmid]))
631 return true;
632
633 return false;
634}
635
636/**
637 * amdgpu_vm_flush - hardware flush the vm
638 *
639 * @ring: ring to use for flush
640 * @job: related job
641 * @need_pipe_sync: is pipe sync needed
642 *
643 * Emit a VM flush when it is necessary.
644 *
645 * Returns:
646 * 0 on success, errno otherwise.
647 */
648int amdgpu_vm_flush(struct amdgpu_ring *ring, struct amdgpu_job *job,
649 bool need_pipe_sync)
650{
651 struct amdgpu_device *adev = ring->adev;
652 unsigned vmhub = ring->vm_hub;
653 struct amdgpu_vmid_mgr *id_mgr = &adev->vm_manager.id_mgr[vmhub];
654 struct amdgpu_vmid *id = &id_mgr->ids[job->vmid];
655 bool spm_update_needed = job->spm_update_needed;
656 bool gds_switch_needed = ring->funcs->emit_gds_switch &&
657 job->gds_switch_needed;
658 bool vm_flush_needed = job->vm_needs_flush;
659 struct dma_fence *fence = NULL;
660 bool pasid_mapping_needed = false;
661 unsigned int patch;
662 int r;
663
664 if (amdgpu_vmid_had_gpu_reset(adev, id)) {
665 gds_switch_needed = true;
666 vm_flush_needed = true;
667 pasid_mapping_needed = true;
668 spm_update_needed = true;
669 }
670
671 mutex_lock(&id_mgr->lock);
672 if (id->pasid != job->pasid || !id->pasid_mapping ||
673 !dma_fence_is_signaled(id->pasid_mapping))
674 pasid_mapping_needed = true;
675 mutex_unlock(&id_mgr->lock);
676
677 gds_switch_needed &= !!ring->funcs->emit_gds_switch;
678 vm_flush_needed &= !!ring->funcs->emit_vm_flush &&
679 job->vm_pd_addr != AMDGPU_BO_INVALID_OFFSET;
680 pasid_mapping_needed &= adev->gmc.gmc_funcs->emit_pasid_mapping &&
681 ring->funcs->emit_wreg;
682
683 if (!vm_flush_needed && !gds_switch_needed && !need_pipe_sync)
684 return 0;
685
686 amdgpu_ring_ib_begin(ring);
687 if (ring->funcs->init_cond_exec)
688 patch = amdgpu_ring_init_cond_exec(ring,
689 ring->cond_exe_gpu_addr);
690
691 if (need_pipe_sync)
692 amdgpu_ring_emit_pipeline_sync(ring);
693
694 if (vm_flush_needed) {
695 trace_amdgpu_vm_flush(ring, job->vmid, job->vm_pd_addr);
696 amdgpu_ring_emit_vm_flush(ring, job->vmid, job->vm_pd_addr);
697 }
698
699 if (pasid_mapping_needed)
700 amdgpu_gmc_emit_pasid_mapping(ring, job->vmid, job->pasid);
701
702 if (spm_update_needed && adev->gfx.rlc.funcs->update_spm_vmid)
703 adev->gfx.rlc.funcs->update_spm_vmid(adev, ring, job->vmid);
704
705 if (!ring->is_mes_queue && ring->funcs->emit_gds_switch &&
706 gds_switch_needed) {
707 amdgpu_ring_emit_gds_switch(ring, job->vmid, job->gds_base,
708 job->gds_size, job->gws_base,
709 job->gws_size, job->oa_base,
710 job->oa_size);
711 }
712
713 if (vm_flush_needed || pasid_mapping_needed) {
714 r = amdgpu_fence_emit(ring, &fence, NULL, 0);
715 if (r)
716 return r;
717 }
718
719 if (vm_flush_needed) {
720 mutex_lock(&id_mgr->lock);
721 dma_fence_put(id->last_flush);
722 id->last_flush = dma_fence_get(fence);
723 id->current_gpu_reset_count =
724 atomic_read(&adev->gpu_reset_counter);
725 mutex_unlock(&id_mgr->lock);
726 }
727
728 if (pasid_mapping_needed) {
729 mutex_lock(&id_mgr->lock);
730 id->pasid = job->pasid;
731 dma_fence_put(id->pasid_mapping);
732 id->pasid_mapping = dma_fence_get(fence);
733 mutex_unlock(&id_mgr->lock);
734 }
735 dma_fence_put(fence);
736
737 amdgpu_ring_patch_cond_exec(ring, patch);
738
739 /* the double SWITCH_BUFFER here *cannot* be skipped by COND_EXEC */
740 if (ring->funcs->emit_switch_buffer) {
741 amdgpu_ring_emit_switch_buffer(ring);
742 amdgpu_ring_emit_switch_buffer(ring);
743 }
744 amdgpu_ring_ib_end(ring);
745 return 0;
746}
747
748/**
749 * amdgpu_vm_bo_find - find the bo_va for a specific vm & bo
750 *
751 * @vm: requested vm
752 * @bo: requested buffer object
753 *
754 * Find @bo inside the requested vm.
755 * Search inside the @bos vm list for the requested vm
756 * Returns the found bo_va or NULL if none is found
757 *
758 * Object has to be reserved!
759 *
760 * Returns:
761 * Found bo_va or NULL.
762 */
763struct amdgpu_bo_va *amdgpu_vm_bo_find(struct amdgpu_vm *vm,
764 struct amdgpu_bo *bo)
765{
766 struct amdgpu_vm_bo_base *base;
767
768 for (base = bo->vm_bo; base; base = base->next) {
769 if (base->vm != vm)
770 continue;
771
772 return container_of(base, struct amdgpu_bo_va, base);
773 }
774 return NULL;
775}
776
777/**
778 * amdgpu_vm_map_gart - Resolve gart mapping of addr
779 *
780 * @pages_addr: optional DMA address to use for lookup
781 * @addr: the unmapped addr
782 *
783 * Look up the physical address of the page that the pte resolves
784 * to.
785 *
786 * Returns:
787 * The pointer for the page table entry.
788 */
789uint64_t amdgpu_vm_map_gart(const dma_addr_t *pages_addr, uint64_t addr)
790{
791 uint64_t result;
792
793 /* page table offset */
794 result = pages_addr[addr >> PAGE_SHIFT];
795
796 /* in case cpu page size != gpu page size*/
797 result |= addr & (~PAGE_MASK);
798
799 result &= 0xFFFFFFFFFFFFF000ULL;
800
801 return result;
802}
803
804/**
805 * amdgpu_vm_update_pdes - make sure that all directories are valid
806 *
807 * @adev: amdgpu_device pointer
808 * @vm: requested vm
809 * @immediate: submit immediately to the paging queue
810 *
811 * Makes sure all directories are up to date.
812 *
813 * Returns:
814 * 0 for success, error for failure.
815 */
816int amdgpu_vm_update_pdes(struct amdgpu_device *adev,
817 struct amdgpu_vm *vm, bool immediate)
818{
819 struct amdgpu_vm_update_params params;
820 struct amdgpu_vm_bo_base *entry;
821 bool flush_tlb_needed = false;
822 LIST_HEAD(relocated);
823 int r, idx;
824
825 spin_lock(&vm->status_lock);
826 list_splice_init(&vm->relocated, &relocated);
827 spin_unlock(&vm->status_lock);
828
829 if (list_empty(&relocated))
830 return 0;
831
832 if (!drm_dev_enter(adev_to_drm(adev), &idx))
833 return -ENODEV;
834
835 memset(¶ms, 0, sizeof(params));
836 params.adev = adev;
837 params.vm = vm;
838 params.immediate = immediate;
839
840 r = vm->update_funcs->prepare(¶ms, NULL, AMDGPU_SYNC_EXPLICIT);
841 if (r)
842 goto error;
843
844 list_for_each_entry(entry, &relocated, vm_status) {
845 /* vm_flush_needed after updating moved PDEs */
846 flush_tlb_needed |= entry->moved;
847
848 r = amdgpu_vm_pde_update(¶ms, entry);
849 if (r)
850 goto error;
851 }
852
853 r = vm->update_funcs->commit(¶ms, &vm->last_update);
854 if (r)
855 goto error;
856
857 if (flush_tlb_needed)
858 atomic64_inc(&vm->tlb_seq);
859
860 while (!list_empty(&relocated)) {
861 entry = list_first_entry(&relocated, struct amdgpu_vm_bo_base,
862 vm_status);
863 amdgpu_vm_bo_idle(entry);
864 }
865
866error:
867 drm_dev_exit(idx);
868 return r;
869}
870
871/**
872 * amdgpu_vm_tlb_seq_cb - make sure to increment tlb sequence
873 * @fence: unused
874 * @cb: the callback structure
875 *
876 * Increments the tlb sequence to make sure that future CS execute a VM flush.
877 */
878static void amdgpu_vm_tlb_seq_cb(struct dma_fence *fence,
879 struct dma_fence_cb *cb)
880{
881 struct amdgpu_vm_tlb_seq_struct *tlb_cb;
882
883 tlb_cb = container_of(cb, typeof(*tlb_cb), cb);
884 atomic64_inc(&tlb_cb->vm->tlb_seq);
885 kfree(tlb_cb);
886}
887
888/**
889 * amdgpu_vm_update_range - update a range in the vm page table
890 *
891 * @adev: amdgpu_device pointer to use for commands
892 * @vm: the VM to update the range
893 * @immediate: immediate submission in a page fault
894 * @unlocked: unlocked invalidation during MM callback
895 * @flush_tlb: trigger tlb invalidation after update completed
896 * @allow_override: change MTYPE for local NUMA nodes
897 * @resv: fences we need to sync to
898 * @start: start of mapped range
899 * @last: last mapped entry
900 * @flags: flags for the entries
901 * @offset: offset into nodes and pages_addr
902 * @vram_base: base for vram mappings
903 * @res: ttm_resource to map
904 * @pages_addr: DMA addresses to use for mapping
905 * @fence: optional resulting fence
906 *
907 * Fill in the page table entries between @start and @last.
908 *
909 * Returns:
910 * 0 for success, negative erro code for failure.
911 */
912int amdgpu_vm_update_range(struct amdgpu_device *adev, struct amdgpu_vm *vm,
913 bool immediate, bool unlocked, bool flush_tlb, bool allow_override,
914 struct dma_resv *resv, uint64_t start, uint64_t last,
915 uint64_t flags, uint64_t offset, uint64_t vram_base,
916 struct ttm_resource *res, dma_addr_t *pages_addr,
917 struct dma_fence **fence)
918{
919 struct amdgpu_vm_update_params params;
920 struct amdgpu_vm_tlb_seq_struct *tlb_cb;
921 struct amdgpu_res_cursor cursor;
922 enum amdgpu_sync_mode sync_mode;
923 int r, idx;
924
925 if (!drm_dev_enter(adev_to_drm(adev), &idx))
926 return -ENODEV;
927
928 tlb_cb = kmalloc(sizeof(*tlb_cb), GFP_KERNEL);
929 if (!tlb_cb) {
930 r = -ENOMEM;
931 goto error_unlock;
932 }
933
934 /* Vega20+XGMI where PTEs get inadvertently cached in L2 texture cache,
935 * heavy-weight flush TLB unconditionally.
936 */
937 flush_tlb |= adev->gmc.xgmi.num_physical_nodes &&
938 amdgpu_ip_version(adev, GC_HWIP, 0) == IP_VERSION(9, 4, 0);
939
940 /*
941 * On GFX8 and older any 8 PTE block with a valid bit set enters the TLB
942 */
943 flush_tlb |= amdgpu_ip_version(adev, GC_HWIP, 0) < IP_VERSION(9, 0, 0);
944
945 memset(¶ms, 0, sizeof(params));
946 params.adev = adev;
947 params.vm = vm;
948 params.immediate = immediate;
949 params.pages_addr = pages_addr;
950 params.unlocked = unlocked;
951 params.allow_override = allow_override;
952
953 /* Implicitly sync to command submissions in the same VM before
954 * unmapping. Sync to moving fences before mapping.
955 */
956 if (!(flags & AMDGPU_PTE_VALID))
957 sync_mode = AMDGPU_SYNC_EQ_OWNER;
958 else
959 sync_mode = AMDGPU_SYNC_EXPLICIT;
960
961 amdgpu_vm_eviction_lock(vm);
962 if (vm->evicting) {
963 r = -EBUSY;
964 goto error_free;
965 }
966
967 if (!unlocked && !dma_fence_is_signaled(vm->last_unlocked)) {
968 struct dma_fence *tmp = dma_fence_get_stub();
969
970 amdgpu_bo_fence(vm->root.bo, vm->last_unlocked, true);
971 swap(vm->last_unlocked, tmp);
972 dma_fence_put(tmp);
973 }
974
975 r = vm->update_funcs->prepare(¶ms, resv, sync_mode);
976 if (r)
977 goto error_free;
978
979 amdgpu_res_first(pages_addr ? NULL : res, offset,
980 (last - start + 1) * AMDGPU_GPU_PAGE_SIZE, &cursor);
981 while (cursor.remaining) {
982 uint64_t tmp, num_entries, addr;
983
984 num_entries = cursor.size >> AMDGPU_GPU_PAGE_SHIFT;
985 if (pages_addr) {
986 bool contiguous = true;
987
988 if (num_entries > AMDGPU_GPU_PAGES_IN_CPU_PAGE) {
989 uint64_t pfn = cursor.start >> PAGE_SHIFT;
990 uint64_t count;
991
992 contiguous = pages_addr[pfn + 1] ==
993 pages_addr[pfn] + PAGE_SIZE;
994
995 tmp = num_entries /
996 AMDGPU_GPU_PAGES_IN_CPU_PAGE;
997 for (count = 2; count < tmp; ++count) {
998 uint64_t idx = pfn + count;
999
1000 if (contiguous != (pages_addr[idx] ==
1001 pages_addr[idx - 1] + PAGE_SIZE))
1002 break;
1003 }
1004 if (!contiguous)
1005 count--;
1006 num_entries = count *
1007 AMDGPU_GPU_PAGES_IN_CPU_PAGE;
1008 }
1009
1010 if (!contiguous) {
1011 addr = cursor.start;
1012 params.pages_addr = pages_addr;
1013 } else {
1014 addr = pages_addr[cursor.start >> PAGE_SHIFT];
1015 params.pages_addr = NULL;
1016 }
1017
1018 } else if (flags & (AMDGPU_PTE_VALID | AMDGPU_PTE_PRT)) {
1019 addr = vram_base + cursor.start;
1020 } else {
1021 addr = 0;
1022 }
1023
1024 tmp = start + num_entries;
1025 r = amdgpu_vm_ptes_update(¶ms, start, tmp, addr, flags);
1026 if (r)
1027 goto error_free;
1028
1029 amdgpu_res_next(&cursor, num_entries * AMDGPU_GPU_PAGE_SIZE);
1030 start = tmp;
1031 }
1032
1033 r = vm->update_funcs->commit(¶ms, fence);
1034
1035 if (flush_tlb || params.table_freed) {
1036 tlb_cb->vm = vm;
1037 if (fence && *fence &&
1038 !dma_fence_add_callback(*fence, &tlb_cb->cb,
1039 amdgpu_vm_tlb_seq_cb)) {
1040 dma_fence_put(vm->last_tlb_flush);
1041 vm->last_tlb_flush = dma_fence_get(*fence);
1042 } else {
1043 amdgpu_vm_tlb_seq_cb(NULL, &tlb_cb->cb);
1044 }
1045 tlb_cb = NULL;
1046 }
1047
1048error_free:
1049 kfree(tlb_cb);
1050
1051error_unlock:
1052 amdgpu_vm_eviction_unlock(vm);
1053 drm_dev_exit(idx);
1054 return r;
1055}
1056
1057static void amdgpu_vm_bo_get_memory(struct amdgpu_bo_va *bo_va,
1058 struct amdgpu_mem_stats *stats)
1059{
1060 struct amdgpu_vm *vm = bo_va->base.vm;
1061 struct amdgpu_bo *bo = bo_va->base.bo;
1062
1063 if (!bo)
1064 return;
1065
1066 /*
1067 * For now ignore BOs which are currently locked and potentially
1068 * changing their location.
1069 */
1070 if (bo->tbo.base.resv != vm->root.bo->tbo.base.resv &&
1071 !dma_resv_trylock(bo->tbo.base.resv))
1072 return;
1073
1074 amdgpu_bo_get_memory(bo, stats);
1075 if (bo->tbo.base.resv != vm->root.bo->tbo.base.resv)
1076 dma_resv_unlock(bo->tbo.base.resv);
1077}
1078
1079void amdgpu_vm_get_memory(struct amdgpu_vm *vm,
1080 struct amdgpu_mem_stats *stats)
1081{
1082 struct amdgpu_bo_va *bo_va, *tmp;
1083
1084 spin_lock(&vm->status_lock);
1085 list_for_each_entry_safe(bo_va, tmp, &vm->idle, base.vm_status)
1086 amdgpu_vm_bo_get_memory(bo_va, stats);
1087
1088 list_for_each_entry_safe(bo_va, tmp, &vm->evicted, base.vm_status)
1089 amdgpu_vm_bo_get_memory(bo_va, stats);
1090
1091 list_for_each_entry_safe(bo_va, tmp, &vm->relocated, base.vm_status)
1092 amdgpu_vm_bo_get_memory(bo_va, stats);
1093
1094 list_for_each_entry_safe(bo_va, tmp, &vm->moved, base.vm_status)
1095 amdgpu_vm_bo_get_memory(bo_va, stats);
1096
1097 list_for_each_entry_safe(bo_va, tmp, &vm->invalidated, base.vm_status)
1098 amdgpu_vm_bo_get_memory(bo_va, stats);
1099
1100 list_for_each_entry_safe(bo_va, tmp, &vm->done, base.vm_status)
1101 amdgpu_vm_bo_get_memory(bo_va, stats);
1102 spin_unlock(&vm->status_lock);
1103}
1104
1105/**
1106 * amdgpu_vm_bo_update - update all BO mappings in the vm page table
1107 *
1108 * @adev: amdgpu_device pointer
1109 * @bo_va: requested BO and VM object
1110 * @clear: if true clear the entries
1111 *
1112 * Fill in the page table entries for @bo_va.
1113 *
1114 * Returns:
1115 * 0 for success, -EINVAL for failure.
1116 */
1117int amdgpu_vm_bo_update(struct amdgpu_device *adev, struct amdgpu_bo_va *bo_va,
1118 bool clear)
1119{
1120 struct amdgpu_bo *bo = bo_va->base.bo;
1121 struct amdgpu_vm *vm = bo_va->base.vm;
1122 struct amdgpu_bo_va_mapping *mapping;
1123 dma_addr_t *pages_addr = NULL;
1124 struct ttm_resource *mem;
1125 struct dma_fence **last_update;
1126 bool flush_tlb = clear;
1127 bool uncached;
1128 struct dma_resv *resv;
1129 uint64_t vram_base;
1130 uint64_t flags;
1131 int r;
1132
1133 if (clear || !bo) {
1134 mem = NULL;
1135 resv = vm->root.bo->tbo.base.resv;
1136 } else {
1137 struct drm_gem_object *obj = &bo->tbo.base;
1138
1139 resv = bo->tbo.base.resv;
1140 if (obj->import_attach && bo_va->is_xgmi) {
1141 struct dma_buf *dma_buf = obj->import_attach->dmabuf;
1142 struct drm_gem_object *gobj = dma_buf->priv;
1143 struct amdgpu_bo *abo = gem_to_amdgpu_bo(gobj);
1144
1145 if (abo->tbo.resource &&
1146 abo->tbo.resource->mem_type == TTM_PL_VRAM)
1147 bo = gem_to_amdgpu_bo(gobj);
1148 }
1149 mem = bo->tbo.resource;
1150 if (mem && (mem->mem_type == TTM_PL_TT ||
1151 mem->mem_type == AMDGPU_PL_PREEMPT))
1152 pages_addr = bo->tbo.ttm->dma_address;
1153 }
1154
1155 if (bo) {
1156 struct amdgpu_device *bo_adev;
1157
1158 flags = amdgpu_ttm_tt_pte_flags(adev, bo->tbo.ttm, mem);
1159
1160 if (amdgpu_bo_encrypted(bo))
1161 flags |= AMDGPU_PTE_TMZ;
1162
1163 bo_adev = amdgpu_ttm_adev(bo->tbo.bdev);
1164 vram_base = bo_adev->vm_manager.vram_base_offset;
1165 uncached = (bo->flags & AMDGPU_GEM_CREATE_UNCACHED) != 0;
1166 } else {
1167 flags = 0x0;
1168 vram_base = 0;
1169 uncached = false;
1170 }
1171
1172 if (clear || (bo && bo->tbo.base.resv ==
1173 vm->root.bo->tbo.base.resv))
1174 last_update = &vm->last_update;
1175 else
1176 last_update = &bo_va->last_pt_update;
1177
1178 if (!clear && bo_va->base.moved) {
1179 flush_tlb = true;
1180 list_splice_init(&bo_va->valids, &bo_va->invalids);
1181
1182 } else if (bo_va->cleared != clear) {
1183 list_splice_init(&bo_va->valids, &bo_va->invalids);
1184 }
1185
1186 list_for_each_entry(mapping, &bo_va->invalids, list) {
1187 uint64_t update_flags = flags;
1188
1189 /* normally,bo_va->flags only contians READABLE and WIRTEABLE bit go here
1190 * but in case of something, we filter the flags in first place
1191 */
1192 if (!(mapping->flags & AMDGPU_PTE_READABLE))
1193 update_flags &= ~AMDGPU_PTE_READABLE;
1194 if (!(mapping->flags & AMDGPU_PTE_WRITEABLE))
1195 update_flags &= ~AMDGPU_PTE_WRITEABLE;
1196
1197 /* Apply ASIC specific mapping flags */
1198 amdgpu_gmc_get_vm_pte(adev, mapping, &update_flags);
1199
1200 trace_amdgpu_vm_bo_update(mapping);
1201
1202 r = amdgpu_vm_update_range(adev, vm, false, false, flush_tlb,
1203 !uncached, resv, mapping->start, mapping->last,
1204 update_flags, mapping->offset,
1205 vram_base, mem, pages_addr,
1206 last_update);
1207 if (r)
1208 return r;
1209 }
1210
1211 /* If the BO is not in its preferred location add it back to
1212 * the evicted list so that it gets validated again on the
1213 * next command submission.
1214 */
1215 if (bo && bo->tbo.base.resv == vm->root.bo->tbo.base.resv) {
1216 uint32_t mem_type = bo->tbo.resource->mem_type;
1217
1218 if (!(bo->preferred_domains &
1219 amdgpu_mem_type_to_domain(mem_type)))
1220 amdgpu_vm_bo_evicted(&bo_va->base);
1221 else
1222 amdgpu_vm_bo_idle(&bo_va->base);
1223 } else {
1224 amdgpu_vm_bo_done(&bo_va->base);
1225 }
1226
1227 list_splice_init(&bo_va->invalids, &bo_va->valids);
1228 bo_va->cleared = clear;
1229 bo_va->base.moved = false;
1230
1231 if (trace_amdgpu_vm_bo_mapping_enabled()) {
1232 list_for_each_entry(mapping, &bo_va->valids, list)
1233 trace_amdgpu_vm_bo_mapping(mapping);
1234 }
1235
1236 return 0;
1237}
1238
1239/**
1240 * amdgpu_vm_update_prt_state - update the global PRT state
1241 *
1242 * @adev: amdgpu_device pointer
1243 */
1244static void amdgpu_vm_update_prt_state(struct amdgpu_device *adev)
1245{
1246 unsigned long flags;
1247 bool enable;
1248
1249 spin_lock_irqsave(&adev->vm_manager.prt_lock, flags);
1250 enable = !!atomic_read(&adev->vm_manager.num_prt_users);
1251 adev->gmc.gmc_funcs->set_prt(adev, enable);
1252 spin_unlock_irqrestore(&adev->vm_manager.prt_lock, flags);
1253}
1254
1255/**
1256 * amdgpu_vm_prt_get - add a PRT user
1257 *
1258 * @adev: amdgpu_device pointer
1259 */
1260static void amdgpu_vm_prt_get(struct amdgpu_device *adev)
1261{
1262 if (!adev->gmc.gmc_funcs->set_prt)
1263 return;
1264
1265 if (atomic_inc_return(&adev->vm_manager.num_prt_users) == 1)
1266 amdgpu_vm_update_prt_state(adev);
1267}
1268
1269/**
1270 * amdgpu_vm_prt_put - drop a PRT user
1271 *
1272 * @adev: amdgpu_device pointer
1273 */
1274static void amdgpu_vm_prt_put(struct amdgpu_device *adev)
1275{
1276 if (atomic_dec_return(&adev->vm_manager.num_prt_users) == 0)
1277 amdgpu_vm_update_prt_state(adev);
1278}
1279
1280/**
1281 * amdgpu_vm_prt_cb - callback for updating the PRT status
1282 *
1283 * @fence: fence for the callback
1284 * @_cb: the callback function
1285 */
1286static void amdgpu_vm_prt_cb(struct dma_fence *fence, struct dma_fence_cb *_cb)
1287{
1288 struct amdgpu_prt_cb *cb = container_of(_cb, struct amdgpu_prt_cb, cb);
1289
1290 amdgpu_vm_prt_put(cb->adev);
1291 kfree(cb);
1292}
1293
1294/**
1295 * amdgpu_vm_add_prt_cb - add callback for updating the PRT status
1296 *
1297 * @adev: amdgpu_device pointer
1298 * @fence: fence for the callback
1299 */
1300static void amdgpu_vm_add_prt_cb(struct amdgpu_device *adev,
1301 struct dma_fence *fence)
1302{
1303 struct amdgpu_prt_cb *cb;
1304
1305 if (!adev->gmc.gmc_funcs->set_prt)
1306 return;
1307
1308 cb = kmalloc(sizeof(struct amdgpu_prt_cb), GFP_KERNEL);
1309 if (!cb) {
1310 /* Last resort when we are OOM */
1311 if (fence)
1312 dma_fence_wait(fence, false);
1313
1314 amdgpu_vm_prt_put(adev);
1315 } else {
1316 cb->adev = adev;
1317 if (!fence || dma_fence_add_callback(fence, &cb->cb,
1318 amdgpu_vm_prt_cb))
1319 amdgpu_vm_prt_cb(fence, &cb->cb);
1320 }
1321}
1322
1323/**
1324 * amdgpu_vm_free_mapping - free a mapping
1325 *
1326 * @adev: amdgpu_device pointer
1327 * @vm: requested vm
1328 * @mapping: mapping to be freed
1329 * @fence: fence of the unmap operation
1330 *
1331 * Free a mapping and make sure we decrease the PRT usage count if applicable.
1332 */
1333static void amdgpu_vm_free_mapping(struct amdgpu_device *adev,
1334 struct amdgpu_vm *vm,
1335 struct amdgpu_bo_va_mapping *mapping,
1336 struct dma_fence *fence)
1337{
1338 if (mapping->flags & AMDGPU_PTE_PRT)
1339 amdgpu_vm_add_prt_cb(adev, fence);
1340 kfree(mapping);
1341}
1342
1343/**
1344 * amdgpu_vm_prt_fini - finish all prt mappings
1345 *
1346 * @adev: amdgpu_device pointer
1347 * @vm: requested vm
1348 *
1349 * Register a cleanup callback to disable PRT support after VM dies.
1350 */
1351static void amdgpu_vm_prt_fini(struct amdgpu_device *adev, struct amdgpu_vm *vm)
1352{
1353 struct dma_resv *resv = vm->root.bo->tbo.base.resv;
1354 struct dma_resv_iter cursor;
1355 struct dma_fence *fence;
1356
1357 dma_resv_for_each_fence(&cursor, resv, DMA_RESV_USAGE_BOOKKEEP, fence) {
1358 /* Add a callback for each fence in the reservation object */
1359 amdgpu_vm_prt_get(adev);
1360 amdgpu_vm_add_prt_cb(adev, fence);
1361 }
1362}
1363
1364/**
1365 * amdgpu_vm_clear_freed - clear freed BOs in the PT
1366 *
1367 * @adev: amdgpu_device pointer
1368 * @vm: requested vm
1369 * @fence: optional resulting fence (unchanged if no work needed to be done
1370 * or if an error occurred)
1371 *
1372 * Make sure all freed BOs are cleared in the PT.
1373 * PTs have to be reserved and mutex must be locked!
1374 *
1375 * Returns:
1376 * 0 for success.
1377 *
1378 */
1379int amdgpu_vm_clear_freed(struct amdgpu_device *adev,
1380 struct amdgpu_vm *vm,
1381 struct dma_fence **fence)
1382{
1383 struct dma_resv *resv = vm->root.bo->tbo.base.resv;
1384 struct amdgpu_bo_va_mapping *mapping;
1385 uint64_t init_pte_value = 0;
1386 struct dma_fence *f = NULL;
1387 int r;
1388
1389 while (!list_empty(&vm->freed)) {
1390 mapping = list_first_entry(&vm->freed,
1391 struct amdgpu_bo_va_mapping, list);
1392 list_del(&mapping->list);
1393
1394 r = amdgpu_vm_update_range(adev, vm, false, false, true, false,
1395 resv, mapping->start, mapping->last,
1396 init_pte_value, 0, 0, NULL, NULL,
1397 &f);
1398 amdgpu_vm_free_mapping(adev, vm, mapping, f);
1399 if (r) {
1400 dma_fence_put(f);
1401 return r;
1402 }
1403 }
1404
1405 if (fence && f) {
1406 dma_fence_put(*fence);
1407 *fence = f;
1408 } else {
1409 dma_fence_put(f);
1410 }
1411
1412 return 0;
1413
1414}
1415
1416/**
1417 * amdgpu_vm_handle_moved - handle moved BOs in the PT
1418 *
1419 * @adev: amdgpu_device pointer
1420 * @vm: requested vm
1421 * @ticket: optional reservation ticket used to reserve the VM
1422 *
1423 * Make sure all BOs which are moved are updated in the PTs.
1424 *
1425 * Returns:
1426 * 0 for success.
1427 *
1428 * PTs have to be reserved!
1429 */
1430int amdgpu_vm_handle_moved(struct amdgpu_device *adev,
1431 struct amdgpu_vm *vm,
1432 struct ww_acquire_ctx *ticket)
1433{
1434 struct amdgpu_bo_va *bo_va;
1435 struct dma_resv *resv;
1436 bool clear, unlock;
1437 int r;
1438
1439 spin_lock(&vm->status_lock);
1440 while (!list_empty(&vm->moved)) {
1441 bo_va = list_first_entry(&vm->moved, struct amdgpu_bo_va,
1442 base.vm_status);
1443 spin_unlock(&vm->status_lock);
1444
1445 /* Per VM BOs never need to bo cleared in the page tables */
1446 r = amdgpu_vm_bo_update(adev, bo_va, false);
1447 if (r)
1448 return r;
1449 spin_lock(&vm->status_lock);
1450 }
1451
1452 while (!list_empty(&vm->invalidated)) {
1453 bo_va = list_first_entry(&vm->invalidated, struct amdgpu_bo_va,
1454 base.vm_status);
1455 resv = bo_va->base.bo->tbo.base.resv;
1456 spin_unlock(&vm->status_lock);
1457
1458 /* Try to reserve the BO to avoid clearing its ptes */
1459 if (!adev->debug_vm && dma_resv_trylock(resv)) {
1460 clear = false;
1461 unlock = true;
1462 /* The caller is already holding the reservation lock */
1463 } else if (ticket && dma_resv_locking_ctx(resv) == ticket) {
1464 clear = false;
1465 unlock = false;
1466 /* Somebody else is using the BO right now */
1467 } else {
1468 clear = true;
1469 unlock = false;
1470 }
1471
1472 r = amdgpu_vm_bo_update(adev, bo_va, clear);
1473
1474 if (unlock)
1475 dma_resv_unlock(resv);
1476 if (r)
1477 return r;
1478
1479 /* Remember evicted DMABuf imports in compute VMs for later
1480 * validation
1481 */
1482 if (vm->is_compute_context &&
1483 bo_va->base.bo->tbo.base.import_attach &&
1484 (!bo_va->base.bo->tbo.resource ||
1485 bo_va->base.bo->tbo.resource->mem_type == TTM_PL_SYSTEM))
1486 amdgpu_vm_bo_evicted_user(&bo_va->base);
1487
1488 spin_lock(&vm->status_lock);
1489 }
1490 spin_unlock(&vm->status_lock);
1491
1492 return 0;
1493}
1494
1495/**
1496 * amdgpu_vm_flush_compute_tlb - Flush TLB on compute VM
1497 *
1498 * @adev: amdgpu_device pointer
1499 * @vm: requested vm
1500 * @flush_type: flush type
1501 * @xcc_mask: mask of XCCs that belong to the compute partition in need of a TLB flush.
1502 *
1503 * Flush TLB if needed for a compute VM.
1504 *
1505 * Returns:
1506 * 0 for success.
1507 */
1508int amdgpu_vm_flush_compute_tlb(struct amdgpu_device *adev,
1509 struct amdgpu_vm *vm,
1510 uint32_t flush_type,
1511 uint32_t xcc_mask)
1512{
1513 uint64_t tlb_seq = amdgpu_vm_tlb_seq(vm);
1514 bool all_hub = false;
1515 int xcc = 0, r = 0;
1516
1517 WARN_ON_ONCE(!vm->is_compute_context);
1518
1519 /*
1520 * It can be that we race and lose here, but that is extremely unlikely
1521 * and the worst thing which could happen is that we flush the changes
1522 * into the TLB once more which is harmless.
1523 */
1524 if (atomic64_xchg(&vm->kfd_last_flushed_seq, tlb_seq) == tlb_seq)
1525 return 0;
1526
1527 if (adev->family == AMDGPU_FAMILY_AI ||
1528 adev->family == AMDGPU_FAMILY_RV)
1529 all_hub = true;
1530
1531 for_each_inst(xcc, xcc_mask) {
1532 r = amdgpu_gmc_flush_gpu_tlb_pasid(adev, vm->pasid, flush_type,
1533 all_hub, xcc);
1534 if (r)
1535 break;
1536 }
1537 return r;
1538}
1539
1540/**
1541 * amdgpu_vm_bo_add - add a bo to a specific vm
1542 *
1543 * @adev: amdgpu_device pointer
1544 * @vm: requested vm
1545 * @bo: amdgpu buffer object
1546 *
1547 * Add @bo into the requested vm.
1548 * Add @bo to the list of bos associated with the vm
1549 *
1550 * Returns:
1551 * Newly added bo_va or NULL for failure
1552 *
1553 * Object has to be reserved!
1554 */
1555struct amdgpu_bo_va *amdgpu_vm_bo_add(struct amdgpu_device *adev,
1556 struct amdgpu_vm *vm,
1557 struct amdgpu_bo *bo)
1558{
1559 struct amdgpu_bo_va *bo_va;
1560
1561 bo_va = kzalloc(sizeof(struct amdgpu_bo_va), GFP_KERNEL);
1562 if (bo_va == NULL) {
1563 return NULL;
1564 }
1565 amdgpu_vm_bo_base_init(&bo_va->base, vm, bo);
1566
1567 bo_va->ref_count = 1;
1568 bo_va->last_pt_update = dma_fence_get_stub();
1569 INIT_LIST_HEAD(&bo_va->valids);
1570 INIT_LIST_HEAD(&bo_va->invalids);
1571
1572 if (!bo)
1573 return bo_va;
1574
1575 dma_resv_assert_held(bo->tbo.base.resv);
1576 if (amdgpu_dmabuf_is_xgmi_accessible(adev, bo)) {
1577 bo_va->is_xgmi = true;
1578 /* Power up XGMI if it can be potentially used */
1579 amdgpu_xgmi_set_pstate(adev, AMDGPU_XGMI_PSTATE_MAX_VEGA20);
1580 }
1581
1582 return bo_va;
1583}
1584
1585
1586/**
1587 * amdgpu_vm_bo_insert_map - insert a new mapping
1588 *
1589 * @adev: amdgpu_device pointer
1590 * @bo_va: bo_va to store the address
1591 * @mapping: the mapping to insert
1592 *
1593 * Insert a new mapping into all structures.
1594 */
1595static void amdgpu_vm_bo_insert_map(struct amdgpu_device *adev,
1596 struct amdgpu_bo_va *bo_va,
1597 struct amdgpu_bo_va_mapping *mapping)
1598{
1599 struct amdgpu_vm *vm = bo_va->base.vm;
1600 struct amdgpu_bo *bo = bo_va->base.bo;
1601
1602 mapping->bo_va = bo_va;
1603 list_add(&mapping->list, &bo_va->invalids);
1604 amdgpu_vm_it_insert(mapping, &vm->va);
1605
1606 if (mapping->flags & AMDGPU_PTE_PRT)
1607 amdgpu_vm_prt_get(adev);
1608
1609 if (bo && bo->tbo.base.resv == vm->root.bo->tbo.base.resv &&
1610 !bo_va->base.moved) {
1611 amdgpu_vm_bo_moved(&bo_va->base);
1612 }
1613 trace_amdgpu_vm_bo_map(bo_va, mapping);
1614}
1615
1616/* Validate operation parameters to prevent potential abuse */
1617static int amdgpu_vm_verify_parameters(struct amdgpu_device *adev,
1618 struct amdgpu_bo *bo,
1619 uint64_t saddr,
1620 uint64_t offset,
1621 uint64_t size)
1622{
1623 uint64_t tmp, lpfn;
1624
1625 if (saddr & AMDGPU_GPU_PAGE_MASK
1626 || offset & AMDGPU_GPU_PAGE_MASK
1627 || size & AMDGPU_GPU_PAGE_MASK)
1628 return -EINVAL;
1629
1630 if (check_add_overflow(saddr, size, &tmp)
1631 || check_add_overflow(offset, size, &tmp)
1632 || size == 0 /* which also leads to end < begin */)
1633 return -EINVAL;
1634
1635 /* make sure object fit at this offset */
1636 if (bo && offset + size > amdgpu_bo_size(bo))
1637 return -EINVAL;
1638
1639 /* Ensure last pfn not exceed max_pfn */
1640 lpfn = (saddr + size - 1) >> AMDGPU_GPU_PAGE_SHIFT;
1641 if (lpfn >= adev->vm_manager.max_pfn)
1642 return -EINVAL;
1643
1644 return 0;
1645}
1646
1647/**
1648 * amdgpu_vm_bo_map - map bo inside a vm
1649 *
1650 * @adev: amdgpu_device pointer
1651 * @bo_va: bo_va to store the address
1652 * @saddr: where to map the BO
1653 * @offset: requested offset in the BO
1654 * @size: BO size in bytes
1655 * @flags: attributes of pages (read/write/valid/etc.)
1656 *
1657 * Add a mapping of the BO at the specefied addr into the VM.
1658 *
1659 * Returns:
1660 * 0 for success, error for failure.
1661 *
1662 * Object has to be reserved and unreserved outside!
1663 */
1664int amdgpu_vm_bo_map(struct amdgpu_device *adev,
1665 struct amdgpu_bo_va *bo_va,
1666 uint64_t saddr, uint64_t offset,
1667 uint64_t size, uint64_t flags)
1668{
1669 struct amdgpu_bo_va_mapping *mapping, *tmp;
1670 struct amdgpu_bo *bo = bo_va->base.bo;
1671 struct amdgpu_vm *vm = bo_va->base.vm;
1672 uint64_t eaddr;
1673 int r;
1674
1675 r = amdgpu_vm_verify_parameters(adev, bo, saddr, offset, size);
1676 if (r)
1677 return r;
1678
1679 saddr /= AMDGPU_GPU_PAGE_SIZE;
1680 eaddr = saddr + (size - 1) / AMDGPU_GPU_PAGE_SIZE;
1681
1682 tmp = amdgpu_vm_it_iter_first(&vm->va, saddr, eaddr);
1683 if (tmp) {
1684 /* bo and tmp overlap, invalid addr */
1685 dev_err(adev->dev, "bo %p va 0x%010Lx-0x%010Lx conflict with "
1686 "0x%010Lx-0x%010Lx\n", bo, saddr, eaddr,
1687 tmp->start, tmp->last + 1);
1688 return -EINVAL;
1689 }
1690
1691 mapping = kmalloc(sizeof(*mapping), GFP_KERNEL);
1692 if (!mapping)
1693 return -ENOMEM;
1694
1695 mapping->start = saddr;
1696 mapping->last = eaddr;
1697 mapping->offset = offset;
1698 mapping->flags = flags;
1699
1700 amdgpu_vm_bo_insert_map(adev, bo_va, mapping);
1701
1702 return 0;
1703}
1704
1705/**
1706 * amdgpu_vm_bo_replace_map - map bo inside a vm, replacing existing mappings
1707 *
1708 * @adev: amdgpu_device pointer
1709 * @bo_va: bo_va to store the address
1710 * @saddr: where to map the BO
1711 * @offset: requested offset in the BO
1712 * @size: BO size in bytes
1713 * @flags: attributes of pages (read/write/valid/etc.)
1714 *
1715 * Add a mapping of the BO at the specefied addr into the VM. Replace existing
1716 * mappings as we do so.
1717 *
1718 * Returns:
1719 * 0 for success, error for failure.
1720 *
1721 * Object has to be reserved and unreserved outside!
1722 */
1723int amdgpu_vm_bo_replace_map(struct amdgpu_device *adev,
1724 struct amdgpu_bo_va *bo_va,
1725 uint64_t saddr, uint64_t offset,
1726 uint64_t size, uint64_t flags)
1727{
1728 struct amdgpu_bo_va_mapping *mapping;
1729 struct amdgpu_bo *bo = bo_va->base.bo;
1730 uint64_t eaddr;
1731 int r;
1732
1733 r = amdgpu_vm_verify_parameters(adev, bo, saddr, offset, size);
1734 if (r)
1735 return r;
1736
1737 /* Allocate all the needed memory */
1738 mapping = kmalloc(sizeof(*mapping), GFP_KERNEL);
1739 if (!mapping)
1740 return -ENOMEM;
1741
1742 r = amdgpu_vm_bo_clear_mappings(adev, bo_va->base.vm, saddr, size);
1743 if (r) {
1744 kfree(mapping);
1745 return r;
1746 }
1747
1748 saddr /= AMDGPU_GPU_PAGE_SIZE;
1749 eaddr = saddr + (size - 1) / AMDGPU_GPU_PAGE_SIZE;
1750
1751 mapping->start = saddr;
1752 mapping->last = eaddr;
1753 mapping->offset = offset;
1754 mapping->flags = flags;
1755
1756 amdgpu_vm_bo_insert_map(adev, bo_va, mapping);
1757
1758 return 0;
1759}
1760
1761/**
1762 * amdgpu_vm_bo_unmap - remove bo mapping from vm
1763 *
1764 * @adev: amdgpu_device pointer
1765 * @bo_va: bo_va to remove the address from
1766 * @saddr: where to the BO is mapped
1767 *
1768 * Remove a mapping of the BO at the specefied addr from the VM.
1769 *
1770 * Returns:
1771 * 0 for success, error for failure.
1772 *
1773 * Object has to be reserved and unreserved outside!
1774 */
1775int amdgpu_vm_bo_unmap(struct amdgpu_device *adev,
1776 struct amdgpu_bo_va *bo_va,
1777 uint64_t saddr)
1778{
1779 struct amdgpu_bo_va_mapping *mapping;
1780 struct amdgpu_vm *vm = bo_va->base.vm;
1781 bool valid = true;
1782
1783 saddr /= AMDGPU_GPU_PAGE_SIZE;
1784
1785 list_for_each_entry(mapping, &bo_va->valids, list) {
1786 if (mapping->start == saddr)
1787 break;
1788 }
1789
1790 if (&mapping->list == &bo_va->valids) {
1791 valid = false;
1792
1793 list_for_each_entry(mapping, &bo_va->invalids, list) {
1794 if (mapping->start == saddr)
1795 break;
1796 }
1797
1798 if (&mapping->list == &bo_va->invalids)
1799 return -ENOENT;
1800 }
1801
1802 list_del(&mapping->list);
1803 amdgpu_vm_it_remove(mapping, &vm->va);
1804 mapping->bo_va = NULL;
1805 trace_amdgpu_vm_bo_unmap(bo_va, mapping);
1806
1807 if (valid)
1808 list_add(&mapping->list, &vm->freed);
1809 else
1810 amdgpu_vm_free_mapping(adev, vm, mapping,
1811 bo_va->last_pt_update);
1812
1813 return 0;
1814}
1815
1816/**
1817 * amdgpu_vm_bo_clear_mappings - remove all mappings in a specific range
1818 *
1819 * @adev: amdgpu_device pointer
1820 * @vm: VM structure to use
1821 * @saddr: start of the range
1822 * @size: size of the range
1823 *
1824 * Remove all mappings in a range, split them as appropriate.
1825 *
1826 * Returns:
1827 * 0 for success, error for failure.
1828 */
1829int amdgpu_vm_bo_clear_mappings(struct amdgpu_device *adev,
1830 struct amdgpu_vm *vm,
1831 uint64_t saddr, uint64_t size)
1832{
1833 struct amdgpu_bo_va_mapping *before, *after, *tmp, *next;
1834 LIST_HEAD(removed);
1835 uint64_t eaddr;
1836 int r;
1837
1838 r = amdgpu_vm_verify_parameters(adev, NULL, saddr, 0, size);
1839 if (r)
1840 return r;
1841
1842 saddr /= AMDGPU_GPU_PAGE_SIZE;
1843 eaddr = saddr + (size - 1) / AMDGPU_GPU_PAGE_SIZE;
1844
1845 /* Allocate all the needed memory */
1846 before = kzalloc(sizeof(*before), GFP_KERNEL);
1847 if (!before)
1848 return -ENOMEM;
1849 INIT_LIST_HEAD(&before->list);
1850
1851 after = kzalloc(sizeof(*after), GFP_KERNEL);
1852 if (!after) {
1853 kfree(before);
1854 return -ENOMEM;
1855 }
1856 INIT_LIST_HEAD(&after->list);
1857
1858 /* Now gather all removed mappings */
1859 tmp = amdgpu_vm_it_iter_first(&vm->va, saddr, eaddr);
1860 while (tmp) {
1861 /* Remember mapping split at the start */
1862 if (tmp->start < saddr) {
1863 before->start = tmp->start;
1864 before->last = saddr - 1;
1865 before->offset = tmp->offset;
1866 before->flags = tmp->flags;
1867 before->bo_va = tmp->bo_va;
1868 list_add(&before->list, &tmp->bo_va->invalids);
1869 }
1870
1871 /* Remember mapping split at the end */
1872 if (tmp->last > eaddr) {
1873 after->start = eaddr + 1;
1874 after->last = tmp->last;
1875 after->offset = tmp->offset;
1876 after->offset += (after->start - tmp->start) << PAGE_SHIFT;
1877 after->flags = tmp->flags;
1878 after->bo_va = tmp->bo_va;
1879 list_add(&after->list, &tmp->bo_va->invalids);
1880 }
1881
1882 list_del(&tmp->list);
1883 list_add(&tmp->list, &removed);
1884
1885 tmp = amdgpu_vm_it_iter_next(tmp, saddr, eaddr);
1886 }
1887
1888 /* And free them up */
1889 list_for_each_entry_safe(tmp, next, &removed, list) {
1890 amdgpu_vm_it_remove(tmp, &vm->va);
1891 list_del(&tmp->list);
1892
1893 if (tmp->start < saddr)
1894 tmp->start = saddr;
1895 if (tmp->last > eaddr)
1896 tmp->last = eaddr;
1897
1898 tmp->bo_va = NULL;
1899 list_add(&tmp->list, &vm->freed);
1900 trace_amdgpu_vm_bo_unmap(NULL, tmp);
1901 }
1902
1903 /* Insert partial mapping before the range */
1904 if (!list_empty(&before->list)) {
1905 struct amdgpu_bo *bo = before->bo_va->base.bo;
1906
1907 amdgpu_vm_it_insert(before, &vm->va);
1908 if (before->flags & AMDGPU_PTE_PRT)
1909 amdgpu_vm_prt_get(adev);
1910
1911 if (bo && bo->tbo.base.resv == vm->root.bo->tbo.base.resv &&
1912 !before->bo_va->base.moved)
1913 amdgpu_vm_bo_moved(&before->bo_va->base);
1914 } else {
1915 kfree(before);
1916 }
1917
1918 /* Insert partial mapping after the range */
1919 if (!list_empty(&after->list)) {
1920 struct amdgpu_bo *bo = after->bo_va->base.bo;
1921
1922 amdgpu_vm_it_insert(after, &vm->va);
1923 if (after->flags & AMDGPU_PTE_PRT)
1924 amdgpu_vm_prt_get(adev);
1925
1926 if (bo && bo->tbo.base.resv == vm->root.bo->tbo.base.resv &&
1927 !after->bo_va->base.moved)
1928 amdgpu_vm_bo_moved(&after->bo_va->base);
1929 } else {
1930 kfree(after);
1931 }
1932
1933 return 0;
1934}
1935
1936/**
1937 * amdgpu_vm_bo_lookup_mapping - find mapping by address
1938 *
1939 * @vm: the requested VM
1940 * @addr: the address
1941 *
1942 * Find a mapping by it's address.
1943 *
1944 * Returns:
1945 * The amdgpu_bo_va_mapping matching for addr or NULL
1946 *
1947 */
1948struct amdgpu_bo_va_mapping *amdgpu_vm_bo_lookup_mapping(struct amdgpu_vm *vm,
1949 uint64_t addr)
1950{
1951 return amdgpu_vm_it_iter_first(&vm->va, addr, addr);
1952}
1953
1954/**
1955 * amdgpu_vm_bo_trace_cs - trace all reserved mappings
1956 *
1957 * @vm: the requested vm
1958 * @ticket: CS ticket
1959 *
1960 * Trace all mappings of BOs reserved during a command submission.
1961 */
1962void amdgpu_vm_bo_trace_cs(struct amdgpu_vm *vm, struct ww_acquire_ctx *ticket)
1963{
1964 struct amdgpu_bo_va_mapping *mapping;
1965
1966 if (!trace_amdgpu_vm_bo_cs_enabled())
1967 return;
1968
1969 for (mapping = amdgpu_vm_it_iter_first(&vm->va, 0, U64_MAX); mapping;
1970 mapping = amdgpu_vm_it_iter_next(mapping, 0, U64_MAX)) {
1971 if (mapping->bo_va && mapping->bo_va->base.bo) {
1972 struct amdgpu_bo *bo;
1973
1974 bo = mapping->bo_va->base.bo;
1975 if (dma_resv_locking_ctx(bo->tbo.base.resv) !=
1976 ticket)
1977 continue;
1978 }
1979
1980 trace_amdgpu_vm_bo_cs(mapping);
1981 }
1982}
1983
1984/**
1985 * amdgpu_vm_bo_del - remove a bo from a specific vm
1986 *
1987 * @adev: amdgpu_device pointer
1988 * @bo_va: requested bo_va
1989 *
1990 * Remove @bo_va->bo from the requested vm.
1991 *
1992 * Object have to be reserved!
1993 */
1994void amdgpu_vm_bo_del(struct amdgpu_device *adev,
1995 struct amdgpu_bo_va *bo_va)
1996{
1997 struct amdgpu_bo_va_mapping *mapping, *next;
1998 struct amdgpu_bo *bo = bo_va->base.bo;
1999 struct amdgpu_vm *vm = bo_va->base.vm;
2000 struct amdgpu_vm_bo_base **base;
2001
2002 dma_resv_assert_held(vm->root.bo->tbo.base.resv);
2003
2004 if (bo) {
2005 dma_resv_assert_held(bo->tbo.base.resv);
2006 if (bo->tbo.base.resv == vm->root.bo->tbo.base.resv)
2007 ttm_bo_set_bulk_move(&bo->tbo, NULL);
2008
2009 for (base = &bo_va->base.bo->vm_bo; *base;
2010 base = &(*base)->next) {
2011 if (*base != &bo_va->base)
2012 continue;
2013
2014 *base = bo_va->base.next;
2015 break;
2016 }
2017 }
2018
2019 spin_lock(&vm->status_lock);
2020 list_del(&bo_va->base.vm_status);
2021 spin_unlock(&vm->status_lock);
2022
2023 list_for_each_entry_safe(mapping, next, &bo_va->valids, list) {
2024 list_del(&mapping->list);
2025 amdgpu_vm_it_remove(mapping, &vm->va);
2026 mapping->bo_va = NULL;
2027 trace_amdgpu_vm_bo_unmap(bo_va, mapping);
2028 list_add(&mapping->list, &vm->freed);
2029 }
2030 list_for_each_entry_safe(mapping, next, &bo_va->invalids, list) {
2031 list_del(&mapping->list);
2032 amdgpu_vm_it_remove(mapping, &vm->va);
2033 amdgpu_vm_free_mapping(adev, vm, mapping,
2034 bo_va->last_pt_update);
2035 }
2036
2037 dma_fence_put(bo_va->last_pt_update);
2038
2039 if (bo && bo_va->is_xgmi)
2040 amdgpu_xgmi_set_pstate(adev, AMDGPU_XGMI_PSTATE_MIN);
2041
2042 kfree(bo_va);
2043}
2044
2045/**
2046 * amdgpu_vm_evictable - check if we can evict a VM
2047 *
2048 * @bo: A page table of the VM.
2049 *
2050 * Check if it is possible to evict a VM.
2051 */
2052bool amdgpu_vm_evictable(struct amdgpu_bo *bo)
2053{
2054 struct amdgpu_vm_bo_base *bo_base = bo->vm_bo;
2055
2056 /* Page tables of a destroyed VM can go away immediately */
2057 if (!bo_base || !bo_base->vm)
2058 return true;
2059
2060 /* Don't evict VM page tables while they are busy */
2061 if (!dma_resv_test_signaled(bo->tbo.base.resv, DMA_RESV_USAGE_BOOKKEEP))
2062 return false;
2063
2064 /* Try to block ongoing updates */
2065 if (!amdgpu_vm_eviction_trylock(bo_base->vm))
2066 return false;
2067
2068 /* Don't evict VM page tables while they are updated */
2069 if (!dma_fence_is_signaled(bo_base->vm->last_unlocked)) {
2070 amdgpu_vm_eviction_unlock(bo_base->vm);
2071 return false;
2072 }
2073
2074 bo_base->vm->evicting = true;
2075 amdgpu_vm_eviction_unlock(bo_base->vm);
2076 return true;
2077}
2078
2079/**
2080 * amdgpu_vm_bo_invalidate - mark the bo as invalid
2081 *
2082 * @adev: amdgpu_device pointer
2083 * @bo: amdgpu buffer object
2084 * @evicted: is the BO evicted
2085 *
2086 * Mark @bo as invalid.
2087 */
2088void amdgpu_vm_bo_invalidate(struct amdgpu_device *adev,
2089 struct amdgpu_bo *bo, bool evicted)
2090{
2091 struct amdgpu_vm_bo_base *bo_base;
2092
2093 /* shadow bo doesn't have bo base, its validation needs its parent */
2094 if (bo->parent && (amdgpu_bo_shadowed(bo->parent) == bo))
2095 bo = bo->parent;
2096
2097 for (bo_base = bo->vm_bo; bo_base; bo_base = bo_base->next) {
2098 struct amdgpu_vm *vm = bo_base->vm;
2099
2100 if (evicted && bo->tbo.base.resv == vm->root.bo->tbo.base.resv) {
2101 amdgpu_vm_bo_evicted(bo_base);
2102 continue;
2103 }
2104
2105 if (bo_base->moved)
2106 continue;
2107 bo_base->moved = true;
2108
2109 if (bo->tbo.type == ttm_bo_type_kernel)
2110 amdgpu_vm_bo_relocated(bo_base);
2111 else if (bo->tbo.base.resv == vm->root.bo->tbo.base.resv)
2112 amdgpu_vm_bo_moved(bo_base);
2113 else
2114 amdgpu_vm_bo_invalidated(bo_base);
2115 }
2116}
2117
2118/**
2119 * amdgpu_vm_get_block_size - calculate VM page table size as power of two
2120 *
2121 * @vm_size: VM size
2122 *
2123 * Returns:
2124 * VM page table as power of two
2125 */
2126static uint32_t amdgpu_vm_get_block_size(uint64_t vm_size)
2127{
2128 /* Total bits covered by PD + PTs */
2129 unsigned bits = ilog2(vm_size) + 18;
2130
2131 /* Make sure the PD is 4K in size up to 8GB address space.
2132 Above that split equal between PD and PTs */
2133 if (vm_size <= 8)
2134 return (bits - 9);
2135 else
2136 return ((bits + 3) / 2);
2137}
2138
2139/**
2140 * amdgpu_vm_adjust_size - adjust vm size, block size and fragment size
2141 *
2142 * @adev: amdgpu_device pointer
2143 * @min_vm_size: the minimum vm size in GB if it's set auto
2144 * @fragment_size_default: Default PTE fragment size
2145 * @max_level: max VMPT level
2146 * @max_bits: max address space size in bits
2147 *
2148 */
2149void amdgpu_vm_adjust_size(struct amdgpu_device *adev, uint32_t min_vm_size,
2150 uint32_t fragment_size_default, unsigned max_level,
2151 unsigned max_bits)
2152{
2153 unsigned int max_size = 1 << (max_bits - 30);
2154 unsigned int vm_size;
2155 uint64_t tmp;
2156
2157 /* adjust vm size first */
2158 if (amdgpu_vm_size != -1) {
2159 vm_size = amdgpu_vm_size;
2160 if (vm_size > max_size) {
2161 dev_warn(adev->dev, "VM size (%d) too large, max is %u GB\n",
2162 amdgpu_vm_size, max_size);
2163 vm_size = max_size;
2164 }
2165 } else {
2166 struct sysinfo si;
2167 unsigned int phys_ram_gb;
2168
2169 /* Optimal VM size depends on the amount of physical
2170 * RAM available. Underlying requirements and
2171 * assumptions:
2172 *
2173 * - Need to map system memory and VRAM from all GPUs
2174 * - VRAM from other GPUs not known here
2175 * - Assume VRAM <= system memory
2176 * - On GFX8 and older, VM space can be segmented for
2177 * different MTYPEs
2178 * - Need to allow room for fragmentation, guard pages etc.
2179 *
2180 * This adds up to a rough guess of system memory x3.
2181 * Round up to power of two to maximize the available
2182 * VM size with the given page table size.
2183 */
2184 si_meminfo(&si);
2185 phys_ram_gb = ((uint64_t)si.totalram * si.mem_unit +
2186 (1 << 30) - 1) >> 30;
2187 vm_size = roundup_pow_of_two(
2188 min(max(phys_ram_gb * 3, min_vm_size), max_size));
2189 }
2190
2191 adev->vm_manager.max_pfn = (uint64_t)vm_size << 18;
2192
2193 tmp = roundup_pow_of_two(adev->vm_manager.max_pfn);
2194 if (amdgpu_vm_block_size != -1)
2195 tmp >>= amdgpu_vm_block_size - 9;
2196 tmp = DIV_ROUND_UP(fls64(tmp) - 1, 9) - 1;
2197 adev->vm_manager.num_level = min_t(unsigned int, max_level, tmp);
2198 switch (adev->vm_manager.num_level) {
2199 case 3:
2200 adev->vm_manager.root_level = AMDGPU_VM_PDB2;
2201 break;
2202 case 2:
2203 adev->vm_manager.root_level = AMDGPU_VM_PDB1;
2204 break;
2205 case 1:
2206 adev->vm_manager.root_level = AMDGPU_VM_PDB0;
2207 break;
2208 default:
2209 dev_err(adev->dev, "VMPT only supports 2~4+1 levels\n");
2210 }
2211 /* block size depends on vm size and hw setup*/
2212 if (amdgpu_vm_block_size != -1)
2213 adev->vm_manager.block_size =
2214 min((unsigned)amdgpu_vm_block_size, max_bits
2215 - AMDGPU_GPU_PAGE_SHIFT
2216 - 9 * adev->vm_manager.num_level);
2217 else if (adev->vm_manager.num_level > 1)
2218 adev->vm_manager.block_size = 9;
2219 else
2220 adev->vm_manager.block_size = amdgpu_vm_get_block_size(tmp);
2221
2222 if (amdgpu_vm_fragment_size == -1)
2223 adev->vm_manager.fragment_size = fragment_size_default;
2224 else
2225 adev->vm_manager.fragment_size = amdgpu_vm_fragment_size;
2226
2227 DRM_INFO("vm size is %u GB, %u levels, block size is %u-bit, fragment size is %u-bit\n",
2228 vm_size, adev->vm_manager.num_level + 1,
2229 adev->vm_manager.block_size,
2230 adev->vm_manager.fragment_size);
2231}
2232
2233/**
2234 * amdgpu_vm_wait_idle - wait for the VM to become idle
2235 *
2236 * @vm: VM object to wait for
2237 * @timeout: timeout to wait for VM to become idle
2238 */
2239long amdgpu_vm_wait_idle(struct amdgpu_vm *vm, long timeout)
2240{
2241 timeout = dma_resv_wait_timeout(vm->root.bo->tbo.base.resv,
2242 DMA_RESV_USAGE_BOOKKEEP,
2243 true, timeout);
2244 if (timeout <= 0)
2245 return timeout;
2246
2247 return dma_fence_wait_timeout(vm->last_unlocked, true, timeout);
2248}
2249
2250static void amdgpu_vm_destroy_task_info(struct kref *kref)
2251{
2252 struct amdgpu_task_info *ti = container_of(kref, struct amdgpu_task_info, refcount);
2253
2254 kfree(ti);
2255}
2256
2257static inline struct amdgpu_vm *
2258amdgpu_vm_get_vm_from_pasid(struct amdgpu_device *adev, u32 pasid)
2259{
2260 struct amdgpu_vm *vm;
2261 unsigned long flags;
2262
2263 xa_lock_irqsave(&adev->vm_manager.pasids, flags);
2264 vm = xa_load(&adev->vm_manager.pasids, pasid);
2265 xa_unlock_irqrestore(&adev->vm_manager.pasids, flags);
2266
2267 return vm;
2268}
2269
2270/**
2271 * amdgpu_vm_put_task_info - reference down the vm task_info ptr
2272 *
2273 * @task_info: task_info struct under discussion.
2274 *
2275 * frees the vm task_info ptr at the last put
2276 */
2277void amdgpu_vm_put_task_info(struct amdgpu_task_info *task_info)
2278{
2279 kref_put(&task_info->refcount, amdgpu_vm_destroy_task_info);
2280}
2281
2282/**
2283 * amdgpu_vm_get_task_info_vm - Extracts task info for a vm.
2284 *
2285 * @vm: VM to get info from
2286 *
2287 * Returns the reference counted task_info structure, which must be
2288 * referenced down with amdgpu_vm_put_task_info.
2289 */
2290struct amdgpu_task_info *
2291amdgpu_vm_get_task_info_vm(struct amdgpu_vm *vm)
2292{
2293 struct amdgpu_task_info *ti = NULL;
2294
2295 if (vm) {
2296 ti = vm->task_info;
2297 kref_get(&vm->task_info->refcount);
2298 }
2299
2300 return ti;
2301}
2302
2303/**
2304 * amdgpu_vm_get_task_info_pasid - Extracts task info for a PASID.
2305 *
2306 * @adev: drm device pointer
2307 * @pasid: PASID identifier for VM
2308 *
2309 * Returns the reference counted task_info structure, which must be
2310 * referenced down with amdgpu_vm_put_task_info.
2311 */
2312struct amdgpu_task_info *
2313amdgpu_vm_get_task_info_pasid(struct amdgpu_device *adev, u32 pasid)
2314{
2315 return amdgpu_vm_get_task_info_vm(
2316 amdgpu_vm_get_vm_from_pasid(adev, pasid));
2317}
2318
2319static int amdgpu_vm_create_task_info(struct amdgpu_vm *vm)
2320{
2321 vm->task_info = kzalloc(sizeof(struct amdgpu_task_info), GFP_KERNEL);
2322 if (!vm->task_info)
2323 return -ENOMEM;
2324
2325 kref_init(&vm->task_info->refcount);
2326 return 0;
2327}
2328
2329/**
2330 * amdgpu_vm_set_task_info - Sets VMs task info.
2331 *
2332 * @vm: vm for which to set the info
2333 */
2334void amdgpu_vm_set_task_info(struct amdgpu_vm *vm)
2335{
2336 if (!vm->task_info)
2337 return;
2338
2339 if (vm->task_info->pid == current->pid)
2340 return;
2341
2342 vm->task_info->pid = current->pid;
2343 get_task_comm(vm->task_info->task_name, current);
2344
2345 if (current->group_leader->mm != current->mm)
2346 return;
2347
2348 vm->task_info->tgid = current->group_leader->pid;
2349 get_task_comm(vm->task_info->process_name, current->group_leader);
2350}
2351
2352/**
2353 * amdgpu_vm_init - initialize a vm instance
2354 *
2355 * @adev: amdgpu_device pointer
2356 * @vm: requested vm
2357 * @xcp_id: GPU partition selection id
2358 *
2359 * Init @vm fields.
2360 *
2361 * Returns:
2362 * 0 for success, error for failure.
2363 */
2364int amdgpu_vm_init(struct amdgpu_device *adev, struct amdgpu_vm *vm,
2365 int32_t xcp_id)
2366{
2367 struct amdgpu_bo *root_bo;
2368 struct amdgpu_bo_vm *root;
2369 int r, i;
2370
2371 vm->va = RB_ROOT_CACHED;
2372 for (i = 0; i < AMDGPU_MAX_VMHUBS; i++)
2373 vm->reserved_vmid[i] = NULL;
2374 INIT_LIST_HEAD(&vm->evicted);
2375 INIT_LIST_HEAD(&vm->evicted_user);
2376 INIT_LIST_HEAD(&vm->relocated);
2377 INIT_LIST_HEAD(&vm->moved);
2378 INIT_LIST_HEAD(&vm->idle);
2379 INIT_LIST_HEAD(&vm->invalidated);
2380 spin_lock_init(&vm->status_lock);
2381 INIT_LIST_HEAD(&vm->freed);
2382 INIT_LIST_HEAD(&vm->done);
2383 INIT_LIST_HEAD(&vm->pt_freed);
2384 INIT_WORK(&vm->pt_free_work, amdgpu_vm_pt_free_work);
2385 INIT_KFIFO(vm->faults);
2386
2387 r = amdgpu_vm_init_entities(adev, vm);
2388 if (r)
2389 return r;
2390
2391 vm->is_compute_context = false;
2392
2393 vm->use_cpu_for_update = !!(adev->vm_manager.vm_update_mode &
2394 AMDGPU_VM_USE_CPU_FOR_GFX);
2395
2396 DRM_DEBUG_DRIVER("VM update mode is %s\n",
2397 vm->use_cpu_for_update ? "CPU" : "SDMA");
2398 WARN_ONCE((vm->use_cpu_for_update &&
2399 !amdgpu_gmc_vram_full_visible(&adev->gmc)),
2400 "CPU update of VM recommended only for large BAR system\n");
2401
2402 if (vm->use_cpu_for_update)
2403 vm->update_funcs = &amdgpu_vm_cpu_funcs;
2404 else
2405 vm->update_funcs = &amdgpu_vm_sdma_funcs;
2406
2407 vm->last_update = dma_fence_get_stub();
2408 vm->last_unlocked = dma_fence_get_stub();
2409 vm->last_tlb_flush = dma_fence_get_stub();
2410 vm->generation = 0;
2411
2412 mutex_init(&vm->eviction_lock);
2413 vm->evicting = false;
2414
2415 r = amdgpu_vm_pt_create(adev, vm, adev->vm_manager.root_level,
2416 false, &root, xcp_id);
2417 if (r)
2418 goto error_free_delayed;
2419
2420 root_bo = amdgpu_bo_ref(&root->bo);
2421 r = amdgpu_bo_reserve(root_bo, true);
2422 if (r) {
2423 amdgpu_bo_unref(&root->shadow);
2424 amdgpu_bo_unref(&root_bo);
2425 goto error_free_delayed;
2426 }
2427
2428 amdgpu_vm_bo_base_init(&vm->root, vm, root_bo);
2429 r = dma_resv_reserve_fences(root_bo->tbo.base.resv, 1);
2430 if (r)
2431 goto error_free_root;
2432
2433 r = amdgpu_vm_pt_clear(adev, vm, root, false);
2434 if (r)
2435 goto error_free_root;
2436
2437 r = amdgpu_vm_create_task_info(vm);
2438 if (r)
2439 DRM_DEBUG("Failed to create task info for VM\n");
2440
2441 amdgpu_bo_unreserve(vm->root.bo);
2442 amdgpu_bo_unref(&root_bo);
2443
2444 return 0;
2445
2446error_free_root:
2447 amdgpu_vm_pt_free_root(adev, vm);
2448 amdgpu_bo_unreserve(vm->root.bo);
2449 amdgpu_bo_unref(&root_bo);
2450
2451error_free_delayed:
2452 dma_fence_put(vm->last_tlb_flush);
2453 dma_fence_put(vm->last_unlocked);
2454 amdgpu_vm_fini_entities(vm);
2455
2456 return r;
2457}
2458
2459/**
2460 * amdgpu_vm_make_compute - Turn a GFX VM into a compute VM
2461 *
2462 * @adev: amdgpu_device pointer
2463 * @vm: requested vm
2464 *
2465 * This only works on GFX VMs that don't have any BOs added and no
2466 * page tables allocated yet.
2467 *
2468 * Changes the following VM parameters:
2469 * - use_cpu_for_update
2470 * - pte_supports_ats
2471 *
2472 * Reinitializes the page directory to reflect the changed ATS
2473 * setting.
2474 *
2475 * Returns:
2476 * 0 for success, -errno for errors.
2477 */
2478int amdgpu_vm_make_compute(struct amdgpu_device *adev, struct amdgpu_vm *vm)
2479{
2480 int r;
2481
2482 r = amdgpu_bo_reserve(vm->root.bo, true);
2483 if (r)
2484 return r;
2485
2486 /* Update VM state */
2487 vm->use_cpu_for_update = !!(adev->vm_manager.vm_update_mode &
2488 AMDGPU_VM_USE_CPU_FOR_COMPUTE);
2489 DRM_DEBUG_DRIVER("VM update mode is %s\n",
2490 vm->use_cpu_for_update ? "CPU" : "SDMA");
2491 WARN_ONCE((vm->use_cpu_for_update &&
2492 !amdgpu_gmc_vram_full_visible(&adev->gmc)),
2493 "CPU update of VM recommended only for large BAR system\n");
2494
2495 if (vm->use_cpu_for_update) {
2496 /* Sync with last SDMA update/clear before switching to CPU */
2497 r = amdgpu_bo_sync_wait(vm->root.bo,
2498 AMDGPU_FENCE_OWNER_UNDEFINED, true);
2499 if (r)
2500 goto unreserve_bo;
2501
2502 vm->update_funcs = &amdgpu_vm_cpu_funcs;
2503 r = amdgpu_vm_pt_map_tables(adev, vm);
2504 if (r)
2505 goto unreserve_bo;
2506
2507 } else {
2508 vm->update_funcs = &amdgpu_vm_sdma_funcs;
2509 }
2510
2511 dma_fence_put(vm->last_update);
2512 vm->last_update = dma_fence_get_stub();
2513 vm->is_compute_context = true;
2514
2515 /* Free the shadow bo for compute VM */
2516 amdgpu_bo_unref(&to_amdgpu_bo_vm(vm->root.bo)->shadow);
2517
2518 goto unreserve_bo;
2519
2520unreserve_bo:
2521 amdgpu_bo_unreserve(vm->root.bo);
2522 return r;
2523}
2524
2525/**
2526 * amdgpu_vm_release_compute - release a compute vm
2527 * @adev: amdgpu_device pointer
2528 * @vm: a vm turned into compute vm by calling amdgpu_vm_make_compute
2529 *
2530 * This is a correspondant of amdgpu_vm_make_compute. It decouples compute
2531 * pasid from vm. Compute should stop use of vm after this call.
2532 */
2533void amdgpu_vm_release_compute(struct amdgpu_device *adev, struct amdgpu_vm *vm)
2534{
2535 amdgpu_vm_set_pasid(adev, vm, 0);
2536 vm->is_compute_context = false;
2537}
2538
2539/**
2540 * amdgpu_vm_fini - tear down a vm instance
2541 *
2542 * @adev: amdgpu_device pointer
2543 * @vm: requested vm
2544 *
2545 * Tear down @vm.
2546 * Unbind the VM and remove all bos from the vm bo list
2547 */
2548void amdgpu_vm_fini(struct amdgpu_device *adev, struct amdgpu_vm *vm)
2549{
2550 struct amdgpu_bo_va_mapping *mapping, *tmp;
2551 bool prt_fini_needed = !!adev->gmc.gmc_funcs->set_prt;
2552 struct amdgpu_bo *root;
2553 unsigned long flags;
2554 int i;
2555
2556 amdgpu_amdkfd_gpuvm_destroy_cb(adev, vm);
2557
2558 flush_work(&vm->pt_free_work);
2559
2560 root = amdgpu_bo_ref(vm->root.bo);
2561 amdgpu_bo_reserve(root, true);
2562 amdgpu_vm_put_task_info(vm->task_info);
2563 amdgpu_vm_set_pasid(adev, vm, 0);
2564 dma_fence_wait(vm->last_unlocked, false);
2565 dma_fence_put(vm->last_unlocked);
2566 dma_fence_wait(vm->last_tlb_flush, false);
2567 /* Make sure that all fence callbacks have completed */
2568 spin_lock_irqsave(vm->last_tlb_flush->lock, flags);
2569 spin_unlock_irqrestore(vm->last_tlb_flush->lock, flags);
2570 dma_fence_put(vm->last_tlb_flush);
2571
2572 list_for_each_entry_safe(mapping, tmp, &vm->freed, list) {
2573 if (mapping->flags & AMDGPU_PTE_PRT && prt_fini_needed) {
2574 amdgpu_vm_prt_fini(adev, vm);
2575 prt_fini_needed = false;
2576 }
2577
2578 list_del(&mapping->list);
2579 amdgpu_vm_free_mapping(adev, vm, mapping, NULL);
2580 }
2581
2582 amdgpu_vm_pt_free_root(adev, vm);
2583 amdgpu_bo_unreserve(root);
2584 amdgpu_bo_unref(&root);
2585 WARN_ON(vm->root.bo);
2586
2587 amdgpu_vm_fini_entities(vm);
2588
2589 if (!RB_EMPTY_ROOT(&vm->va.rb_root)) {
2590 dev_err(adev->dev, "still active bo inside vm\n");
2591 }
2592 rbtree_postorder_for_each_entry_safe(mapping, tmp,
2593 &vm->va.rb_root, rb) {
2594 /* Don't remove the mapping here, we don't want to trigger a
2595 * rebalance and the tree is about to be destroyed anyway.
2596 */
2597 list_del(&mapping->list);
2598 kfree(mapping);
2599 }
2600
2601 dma_fence_put(vm->last_update);
2602
2603 for (i = 0; i < AMDGPU_MAX_VMHUBS; i++) {
2604 if (vm->reserved_vmid[i]) {
2605 amdgpu_vmid_free_reserved(adev, i);
2606 vm->reserved_vmid[i] = false;
2607 }
2608 }
2609
2610}
2611
2612/**
2613 * amdgpu_vm_manager_init - init the VM manager
2614 *
2615 * @adev: amdgpu_device pointer
2616 *
2617 * Initialize the VM manager structures
2618 */
2619void amdgpu_vm_manager_init(struct amdgpu_device *adev)
2620{
2621 unsigned i;
2622
2623 /* Concurrent flushes are only possible starting with Vega10 and
2624 * are broken on Navi10 and Navi14.
2625 */
2626 adev->vm_manager.concurrent_flush = !(adev->asic_type < CHIP_VEGA10 ||
2627 adev->asic_type == CHIP_NAVI10 ||
2628 adev->asic_type == CHIP_NAVI14);
2629 amdgpu_vmid_mgr_init(adev);
2630
2631 adev->vm_manager.fence_context =
2632 dma_fence_context_alloc(AMDGPU_MAX_RINGS);
2633 for (i = 0; i < AMDGPU_MAX_RINGS; ++i)
2634 adev->vm_manager.seqno[i] = 0;
2635
2636 spin_lock_init(&adev->vm_manager.prt_lock);
2637 atomic_set(&adev->vm_manager.num_prt_users, 0);
2638
2639 /* If not overridden by the user, by default, only in large BAR systems
2640 * Compute VM tables will be updated by CPU
2641 */
2642#ifdef CONFIG_X86_64
2643 if (amdgpu_vm_update_mode == -1) {
2644 /* For asic with VF MMIO access protection
2645 * avoid using CPU for VM table updates
2646 */
2647 if (amdgpu_gmc_vram_full_visible(&adev->gmc) &&
2648 !amdgpu_sriov_vf_mmio_access_protection(adev))
2649 adev->vm_manager.vm_update_mode =
2650 AMDGPU_VM_USE_CPU_FOR_COMPUTE;
2651 else
2652 adev->vm_manager.vm_update_mode = 0;
2653 } else
2654 adev->vm_manager.vm_update_mode = amdgpu_vm_update_mode;
2655#else
2656 adev->vm_manager.vm_update_mode = 0;
2657#endif
2658
2659 xa_init_flags(&adev->vm_manager.pasids, XA_FLAGS_LOCK_IRQ);
2660}
2661
2662/**
2663 * amdgpu_vm_manager_fini - cleanup VM manager
2664 *
2665 * @adev: amdgpu_device pointer
2666 *
2667 * Cleanup the VM manager and free resources.
2668 */
2669void amdgpu_vm_manager_fini(struct amdgpu_device *adev)
2670{
2671 WARN_ON(!xa_empty(&adev->vm_manager.pasids));
2672 xa_destroy(&adev->vm_manager.pasids);
2673
2674 amdgpu_vmid_mgr_fini(adev);
2675}
2676
2677/**
2678 * amdgpu_vm_ioctl - Manages VMID reservation for vm hubs.
2679 *
2680 * @dev: drm device pointer
2681 * @data: drm_amdgpu_vm
2682 * @filp: drm file pointer
2683 *
2684 * Returns:
2685 * 0 for success, -errno for errors.
2686 */
2687int amdgpu_vm_ioctl(struct drm_device *dev, void *data, struct drm_file *filp)
2688{
2689 union drm_amdgpu_vm *args = data;
2690 struct amdgpu_device *adev = drm_to_adev(dev);
2691 struct amdgpu_fpriv *fpriv = filp->driver_priv;
2692
2693 /* No valid flags defined yet */
2694 if (args->in.flags)
2695 return -EINVAL;
2696
2697 switch (args->in.op) {
2698 case AMDGPU_VM_OP_RESERVE_VMID:
2699 /* We only have requirement to reserve vmid from gfxhub */
2700 if (!fpriv->vm.reserved_vmid[AMDGPU_GFXHUB(0)]) {
2701 amdgpu_vmid_alloc_reserved(adev, AMDGPU_GFXHUB(0));
2702 fpriv->vm.reserved_vmid[AMDGPU_GFXHUB(0)] = true;
2703 }
2704
2705 break;
2706 case AMDGPU_VM_OP_UNRESERVE_VMID:
2707 if (fpriv->vm.reserved_vmid[AMDGPU_GFXHUB(0)]) {
2708 amdgpu_vmid_free_reserved(adev, AMDGPU_GFXHUB(0));
2709 fpriv->vm.reserved_vmid[AMDGPU_GFXHUB(0)] = false;
2710 }
2711 break;
2712 default:
2713 return -EINVAL;
2714 }
2715
2716 return 0;
2717}
2718
2719/**
2720 * amdgpu_vm_handle_fault - graceful handling of VM faults.
2721 * @adev: amdgpu device pointer
2722 * @pasid: PASID of the VM
2723 * @vmid: VMID, only used for GFX 9.4.3.
2724 * @node_id: Node_id received in IH cookie. Only applicable for
2725 * GFX 9.4.3.
2726 * @addr: Address of the fault
2727 * @write_fault: true is write fault, false is read fault
2728 *
2729 * Try to gracefully handle a VM fault. Return true if the fault was handled and
2730 * shouldn't be reported any more.
2731 */
2732bool amdgpu_vm_handle_fault(struct amdgpu_device *adev, u32 pasid,
2733 u32 vmid, u32 node_id, uint64_t addr,
2734 bool write_fault)
2735{
2736 bool is_compute_context = false;
2737 struct amdgpu_bo *root;
2738 unsigned long irqflags;
2739 uint64_t value, flags;
2740 struct amdgpu_vm *vm;
2741 int r;
2742
2743 xa_lock_irqsave(&adev->vm_manager.pasids, irqflags);
2744 vm = xa_load(&adev->vm_manager.pasids, pasid);
2745 if (vm) {
2746 root = amdgpu_bo_ref(vm->root.bo);
2747 is_compute_context = vm->is_compute_context;
2748 } else {
2749 root = NULL;
2750 }
2751 xa_unlock_irqrestore(&adev->vm_manager.pasids, irqflags);
2752
2753 if (!root)
2754 return false;
2755
2756 addr /= AMDGPU_GPU_PAGE_SIZE;
2757
2758 if (is_compute_context && !svm_range_restore_pages(adev, pasid, vmid,
2759 node_id, addr, write_fault)) {
2760 amdgpu_bo_unref(&root);
2761 return true;
2762 }
2763
2764 r = amdgpu_bo_reserve(root, true);
2765 if (r)
2766 goto error_unref;
2767
2768 /* Double check that the VM still exists */
2769 xa_lock_irqsave(&adev->vm_manager.pasids, irqflags);
2770 vm = xa_load(&adev->vm_manager.pasids, pasid);
2771 if (vm && vm->root.bo != root)
2772 vm = NULL;
2773 xa_unlock_irqrestore(&adev->vm_manager.pasids, irqflags);
2774 if (!vm)
2775 goto error_unlock;
2776
2777 flags = AMDGPU_PTE_VALID | AMDGPU_PTE_SNOOPED |
2778 AMDGPU_PTE_SYSTEM;
2779
2780 if (is_compute_context) {
2781 /* Intentionally setting invalid PTE flag
2782 * combination to force a no-retry-fault
2783 */
2784 flags = AMDGPU_VM_NORETRY_FLAGS;
2785 value = 0;
2786 } else if (amdgpu_vm_fault_stop == AMDGPU_VM_FAULT_STOP_NEVER) {
2787 /* Redirect the access to the dummy page */
2788 value = adev->dummy_page_addr;
2789 flags |= AMDGPU_PTE_EXECUTABLE | AMDGPU_PTE_READABLE |
2790 AMDGPU_PTE_WRITEABLE;
2791
2792 } else {
2793 /* Let the hw retry silently on the PTE */
2794 value = 0;
2795 }
2796
2797 r = dma_resv_reserve_fences(root->tbo.base.resv, 1);
2798 if (r) {
2799 pr_debug("failed %d to reserve fence slot\n", r);
2800 goto error_unlock;
2801 }
2802
2803 r = amdgpu_vm_update_range(adev, vm, true, false, false, false,
2804 NULL, addr, addr, flags, value, 0, NULL, NULL, NULL);
2805 if (r)
2806 goto error_unlock;
2807
2808 r = amdgpu_vm_update_pdes(adev, vm, true);
2809
2810error_unlock:
2811 amdgpu_bo_unreserve(root);
2812 if (r < 0)
2813 DRM_ERROR("Can't handle page fault (%d)\n", r);
2814
2815error_unref:
2816 amdgpu_bo_unref(&root);
2817
2818 return false;
2819}
2820
2821#if defined(CONFIG_DEBUG_FS)
2822/**
2823 * amdgpu_debugfs_vm_bo_info - print BO info for the VM
2824 *
2825 * @vm: Requested VM for printing BO info
2826 * @m: debugfs file
2827 *
2828 * Print BO information in debugfs file for the VM
2829 */
2830void amdgpu_debugfs_vm_bo_info(struct amdgpu_vm *vm, struct seq_file *m)
2831{
2832 struct amdgpu_bo_va *bo_va, *tmp;
2833 u64 total_idle = 0;
2834 u64 total_evicted = 0;
2835 u64 total_relocated = 0;
2836 u64 total_moved = 0;
2837 u64 total_invalidated = 0;
2838 u64 total_done = 0;
2839 unsigned int total_idle_objs = 0;
2840 unsigned int total_evicted_objs = 0;
2841 unsigned int total_relocated_objs = 0;
2842 unsigned int total_moved_objs = 0;
2843 unsigned int total_invalidated_objs = 0;
2844 unsigned int total_done_objs = 0;
2845 unsigned int id = 0;
2846
2847 spin_lock(&vm->status_lock);
2848 seq_puts(m, "\tIdle BOs:\n");
2849 list_for_each_entry_safe(bo_va, tmp, &vm->idle, base.vm_status) {
2850 if (!bo_va->base.bo)
2851 continue;
2852 total_idle += amdgpu_bo_print_info(id++, bo_va->base.bo, m);
2853 }
2854 total_idle_objs = id;
2855 id = 0;
2856
2857 seq_puts(m, "\tEvicted BOs:\n");
2858 list_for_each_entry_safe(bo_va, tmp, &vm->evicted, base.vm_status) {
2859 if (!bo_va->base.bo)
2860 continue;
2861 total_evicted += amdgpu_bo_print_info(id++, bo_va->base.bo, m);
2862 }
2863 total_evicted_objs = id;
2864 id = 0;
2865
2866 seq_puts(m, "\tRelocated BOs:\n");
2867 list_for_each_entry_safe(bo_va, tmp, &vm->relocated, base.vm_status) {
2868 if (!bo_va->base.bo)
2869 continue;
2870 total_relocated += amdgpu_bo_print_info(id++, bo_va->base.bo, m);
2871 }
2872 total_relocated_objs = id;
2873 id = 0;
2874
2875 seq_puts(m, "\tMoved BOs:\n");
2876 list_for_each_entry_safe(bo_va, tmp, &vm->moved, base.vm_status) {
2877 if (!bo_va->base.bo)
2878 continue;
2879 total_moved += amdgpu_bo_print_info(id++, bo_va->base.bo, m);
2880 }
2881 total_moved_objs = id;
2882 id = 0;
2883
2884 seq_puts(m, "\tInvalidated BOs:\n");
2885 list_for_each_entry_safe(bo_va, tmp, &vm->invalidated, base.vm_status) {
2886 if (!bo_va->base.bo)
2887 continue;
2888 total_invalidated += amdgpu_bo_print_info(id++, bo_va->base.bo, m);
2889 }
2890 total_invalidated_objs = id;
2891 id = 0;
2892
2893 seq_puts(m, "\tDone BOs:\n");
2894 list_for_each_entry_safe(bo_va, tmp, &vm->done, base.vm_status) {
2895 if (!bo_va->base.bo)
2896 continue;
2897 total_done += amdgpu_bo_print_info(id++, bo_va->base.bo, m);
2898 }
2899 spin_unlock(&vm->status_lock);
2900 total_done_objs = id;
2901
2902 seq_printf(m, "\tTotal idle size: %12lld\tobjs:\t%d\n", total_idle,
2903 total_idle_objs);
2904 seq_printf(m, "\tTotal evicted size: %12lld\tobjs:\t%d\n", total_evicted,
2905 total_evicted_objs);
2906 seq_printf(m, "\tTotal relocated size: %12lld\tobjs:\t%d\n", total_relocated,
2907 total_relocated_objs);
2908 seq_printf(m, "\tTotal moved size: %12lld\tobjs:\t%d\n", total_moved,
2909 total_moved_objs);
2910 seq_printf(m, "\tTotal invalidated size: %12lld\tobjs:\t%d\n", total_invalidated,
2911 total_invalidated_objs);
2912 seq_printf(m, "\tTotal done size: %12lld\tobjs:\t%d\n", total_done,
2913 total_done_objs);
2914}
2915#endif
2916
2917/**
2918 * amdgpu_vm_update_fault_cache - update cached fault into.
2919 * @adev: amdgpu device pointer
2920 * @pasid: PASID of the VM
2921 * @addr: Address of the fault
2922 * @status: GPUVM fault status register
2923 * @vmhub: which vmhub got the fault
2924 *
2925 * Cache the fault info for later use by userspace in debugging.
2926 */
2927void amdgpu_vm_update_fault_cache(struct amdgpu_device *adev,
2928 unsigned int pasid,
2929 uint64_t addr,
2930 uint32_t status,
2931 unsigned int vmhub)
2932{
2933 struct amdgpu_vm *vm;
2934 unsigned long flags;
2935
2936 xa_lock_irqsave(&adev->vm_manager.pasids, flags);
2937
2938 vm = xa_load(&adev->vm_manager.pasids, pasid);
2939 /* Don't update the fault cache if status is 0. In the multiple
2940 * fault case, subsequent faults will return a 0 status which is
2941 * useless for userspace and replaces the useful fault status, so
2942 * only update if status is non-0.
2943 */
2944 if (vm && status) {
2945 vm->fault_info.addr = addr;
2946 vm->fault_info.status = status;
2947 if (AMDGPU_IS_GFXHUB(vmhub)) {
2948 vm->fault_info.vmhub = AMDGPU_VMHUB_TYPE_GFX;
2949 vm->fault_info.vmhub |=
2950 (vmhub - AMDGPU_GFXHUB_START) << AMDGPU_VMHUB_IDX_SHIFT;
2951 } else if (AMDGPU_IS_MMHUB0(vmhub)) {
2952 vm->fault_info.vmhub = AMDGPU_VMHUB_TYPE_MM0;
2953 vm->fault_info.vmhub |=
2954 (vmhub - AMDGPU_MMHUB0_START) << AMDGPU_VMHUB_IDX_SHIFT;
2955 } else if (AMDGPU_IS_MMHUB1(vmhub)) {
2956 vm->fault_info.vmhub = AMDGPU_VMHUB_TYPE_MM1;
2957 vm->fault_info.vmhub |=
2958 (vmhub - AMDGPU_MMHUB1_START) << AMDGPU_VMHUB_IDX_SHIFT;
2959 } else {
2960 WARN_ONCE(1, "Invalid vmhub %u\n", vmhub);
2961 }
2962 }
2963 xa_unlock_irqrestore(&adev->vm_manager.pasids, flags);
2964}
2965
1/*
2 * Copyright 2008 Advanced Micro Devices, Inc.
3 * Copyright 2008 Red Hat Inc.
4 * Copyright 2009 Jerome Glisse.
5 *
6 * Permission is hereby granted, free of charge, to any person obtaining a
7 * copy of this software and associated documentation files (the "Software"),
8 * to deal in the Software without restriction, including without limitation
9 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
10 * and/or sell copies of the Software, and to permit persons to whom the
11 * Software is furnished to do so, subject to the following conditions:
12 *
13 * The above copyright notice and this permission notice shall be included in
14 * all copies or substantial portions of the Software.
15 *
16 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
17 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
18 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
19 * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR
20 * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
21 * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
22 * OTHER DEALINGS IN THE SOFTWARE.
23 *
24 * Authors: Dave Airlie
25 * Alex Deucher
26 * Jerome Glisse
27 */
28#include <linux/dma-fence-array.h>
29#include <linux/interval_tree_generic.h>
30#include <linux/idr.h>
31
32#include <drm/amdgpu_drm.h>
33#include "amdgpu.h"
34#include "amdgpu_trace.h"
35#include "amdgpu_amdkfd.h"
36#include "amdgpu_gmc.h"
37#include "amdgpu_xgmi.h"
38
39/**
40 * DOC: GPUVM
41 *
42 * GPUVM is similar to the legacy gart on older asics, however
43 * rather than there being a single global gart table
44 * for the entire GPU, there are multiple VM page tables active
45 * at any given time. The VM page tables can contain a mix
46 * vram pages and system memory pages and system memory pages
47 * can be mapped as snooped (cached system pages) or unsnooped
48 * (uncached system pages).
49 * Each VM has an ID associated with it and there is a page table
50 * associated with each VMID. When execting a command buffer,
51 * the kernel tells the the ring what VMID to use for that command
52 * buffer. VMIDs are allocated dynamically as commands are submitted.
53 * The userspace drivers maintain their own address space and the kernel
54 * sets up their pages tables accordingly when they submit their
55 * command buffers and a VMID is assigned.
56 * Cayman/Trinity support up to 8 active VMs at any given time;
57 * SI supports 16.
58 */
59
60#define START(node) ((node)->start)
61#define LAST(node) ((node)->last)
62
63INTERVAL_TREE_DEFINE(struct amdgpu_bo_va_mapping, rb, uint64_t, __subtree_last,
64 START, LAST, static, amdgpu_vm_it)
65
66#undef START
67#undef LAST
68
69/**
70 * struct amdgpu_prt_cb - Helper to disable partial resident texture feature from a fence callback
71 */
72struct amdgpu_prt_cb {
73
74 /**
75 * @adev: amdgpu device
76 */
77 struct amdgpu_device *adev;
78
79 /**
80 * @cb: callback
81 */
82 struct dma_fence_cb cb;
83};
84
85/**
86 * amdgpu_vm_level_shift - return the addr shift for each level
87 *
88 * @adev: amdgpu_device pointer
89 * @level: VMPT level
90 *
91 * Returns:
92 * The number of bits the pfn needs to be right shifted for a level.
93 */
94static unsigned amdgpu_vm_level_shift(struct amdgpu_device *adev,
95 unsigned level)
96{
97 unsigned shift = 0xff;
98
99 switch (level) {
100 case AMDGPU_VM_PDB2:
101 case AMDGPU_VM_PDB1:
102 case AMDGPU_VM_PDB0:
103 shift = 9 * (AMDGPU_VM_PDB0 - level) +
104 adev->vm_manager.block_size;
105 break;
106 case AMDGPU_VM_PTB:
107 shift = 0;
108 break;
109 default:
110 dev_err(adev->dev, "the level%d isn't supported.\n", level);
111 }
112
113 return shift;
114}
115
116/**
117 * amdgpu_vm_num_entries - return the number of entries in a PD/PT
118 *
119 * @adev: amdgpu_device pointer
120 * @level: VMPT level
121 *
122 * Returns:
123 * The number of entries in a page directory or page table.
124 */
125static unsigned amdgpu_vm_num_entries(struct amdgpu_device *adev,
126 unsigned level)
127{
128 unsigned shift = amdgpu_vm_level_shift(adev,
129 adev->vm_manager.root_level);
130
131 if (level == adev->vm_manager.root_level)
132 /* For the root directory */
133 return round_up(adev->vm_manager.max_pfn, 1ULL << shift) >> shift;
134 else if (level != AMDGPU_VM_PTB)
135 /* Everything in between */
136 return 512;
137 else
138 /* For the page tables on the leaves */
139 return AMDGPU_VM_PTE_COUNT(adev);
140}
141
142/**
143 * amdgpu_vm_num_ats_entries - return the number of ATS entries in the root PD
144 *
145 * @adev: amdgpu_device pointer
146 *
147 * Returns:
148 * The number of entries in the root page directory which needs the ATS setting.
149 */
150static unsigned amdgpu_vm_num_ats_entries(struct amdgpu_device *adev)
151{
152 unsigned shift;
153
154 shift = amdgpu_vm_level_shift(adev, adev->vm_manager.root_level);
155 return AMDGPU_GMC_HOLE_START >> (shift + AMDGPU_GPU_PAGE_SHIFT);
156}
157
158/**
159 * amdgpu_vm_entries_mask - the mask to get the entry number of a PD/PT
160 *
161 * @adev: amdgpu_device pointer
162 * @level: VMPT level
163 *
164 * Returns:
165 * The mask to extract the entry number of a PD/PT from an address.
166 */
167static uint32_t amdgpu_vm_entries_mask(struct amdgpu_device *adev,
168 unsigned int level)
169{
170 if (level <= adev->vm_manager.root_level)
171 return 0xffffffff;
172 else if (level != AMDGPU_VM_PTB)
173 return 0x1ff;
174 else
175 return AMDGPU_VM_PTE_COUNT(adev) - 1;
176}
177
178/**
179 * amdgpu_vm_bo_size - returns the size of the BOs in bytes
180 *
181 * @adev: amdgpu_device pointer
182 * @level: VMPT level
183 *
184 * Returns:
185 * The size of the BO for a page directory or page table in bytes.
186 */
187static unsigned amdgpu_vm_bo_size(struct amdgpu_device *adev, unsigned level)
188{
189 return AMDGPU_GPU_PAGE_ALIGN(amdgpu_vm_num_entries(adev, level) * 8);
190}
191
192/**
193 * amdgpu_vm_bo_evicted - vm_bo is evicted
194 *
195 * @vm_bo: vm_bo which is evicted
196 *
197 * State for PDs/PTs and per VM BOs which are not at the location they should
198 * be.
199 */
200static void amdgpu_vm_bo_evicted(struct amdgpu_vm_bo_base *vm_bo)
201{
202 struct amdgpu_vm *vm = vm_bo->vm;
203 struct amdgpu_bo *bo = vm_bo->bo;
204
205 vm_bo->moved = true;
206 if (bo->tbo.type == ttm_bo_type_kernel)
207 list_move(&vm_bo->vm_status, &vm->evicted);
208 else
209 list_move_tail(&vm_bo->vm_status, &vm->evicted);
210}
211
212/**
213 * amdgpu_vm_bo_relocated - vm_bo is reloacted
214 *
215 * @vm_bo: vm_bo which is relocated
216 *
217 * State for PDs/PTs which needs to update their parent PD.
218 */
219static void amdgpu_vm_bo_relocated(struct amdgpu_vm_bo_base *vm_bo)
220{
221 list_move(&vm_bo->vm_status, &vm_bo->vm->relocated);
222}
223
224/**
225 * amdgpu_vm_bo_moved - vm_bo is moved
226 *
227 * @vm_bo: vm_bo which is moved
228 *
229 * State for per VM BOs which are moved, but that change is not yet reflected
230 * in the page tables.
231 */
232static void amdgpu_vm_bo_moved(struct amdgpu_vm_bo_base *vm_bo)
233{
234 list_move(&vm_bo->vm_status, &vm_bo->vm->moved);
235}
236
237/**
238 * amdgpu_vm_bo_idle - vm_bo is idle
239 *
240 * @vm_bo: vm_bo which is now idle
241 *
242 * State for PDs/PTs and per VM BOs which have gone through the state machine
243 * and are now idle.
244 */
245static void amdgpu_vm_bo_idle(struct amdgpu_vm_bo_base *vm_bo)
246{
247 list_move(&vm_bo->vm_status, &vm_bo->vm->idle);
248 vm_bo->moved = false;
249}
250
251/**
252 * amdgpu_vm_bo_invalidated - vm_bo is invalidated
253 *
254 * @vm_bo: vm_bo which is now invalidated
255 *
256 * State for normal BOs which are invalidated and that change not yet reflected
257 * in the PTs.
258 */
259static void amdgpu_vm_bo_invalidated(struct amdgpu_vm_bo_base *vm_bo)
260{
261 spin_lock(&vm_bo->vm->invalidated_lock);
262 list_move(&vm_bo->vm_status, &vm_bo->vm->invalidated);
263 spin_unlock(&vm_bo->vm->invalidated_lock);
264}
265
266/**
267 * amdgpu_vm_bo_done - vm_bo is done
268 *
269 * @vm_bo: vm_bo which is now done
270 *
271 * State for normal BOs which are invalidated and that change has been updated
272 * in the PTs.
273 */
274static void amdgpu_vm_bo_done(struct amdgpu_vm_bo_base *vm_bo)
275{
276 spin_lock(&vm_bo->vm->invalidated_lock);
277 list_del_init(&vm_bo->vm_status);
278 spin_unlock(&vm_bo->vm->invalidated_lock);
279}
280
281/**
282 * amdgpu_vm_bo_base_init - Adds bo to the list of bos associated with the vm
283 *
284 * @base: base structure for tracking BO usage in a VM
285 * @vm: vm to which bo is to be added
286 * @bo: amdgpu buffer object
287 *
288 * Initialize a bo_va_base structure and add it to the appropriate lists
289 *
290 */
291static void amdgpu_vm_bo_base_init(struct amdgpu_vm_bo_base *base,
292 struct amdgpu_vm *vm,
293 struct amdgpu_bo *bo)
294{
295 base->vm = vm;
296 base->bo = bo;
297 base->next = NULL;
298 INIT_LIST_HEAD(&base->vm_status);
299
300 if (!bo)
301 return;
302 base->next = bo->vm_bo;
303 bo->vm_bo = base;
304
305 if (bo->tbo.base.resv != vm->root.base.bo->tbo.base.resv)
306 return;
307
308 vm->bulk_moveable = false;
309 if (bo->tbo.type == ttm_bo_type_kernel && bo->parent)
310 amdgpu_vm_bo_relocated(base);
311 else
312 amdgpu_vm_bo_idle(base);
313
314 if (bo->preferred_domains &
315 amdgpu_mem_type_to_domain(bo->tbo.mem.mem_type))
316 return;
317
318 /*
319 * we checked all the prerequisites, but it looks like this per vm bo
320 * is currently evicted. add the bo to the evicted list to make sure it
321 * is validated on next vm use to avoid fault.
322 * */
323 amdgpu_vm_bo_evicted(base);
324}
325
326/**
327 * amdgpu_vm_pt_parent - get the parent page directory
328 *
329 * @pt: child page table
330 *
331 * Helper to get the parent entry for the child page table. NULL if we are at
332 * the root page directory.
333 */
334static struct amdgpu_vm_pt *amdgpu_vm_pt_parent(struct amdgpu_vm_pt *pt)
335{
336 struct amdgpu_bo *parent = pt->base.bo->parent;
337
338 if (!parent)
339 return NULL;
340
341 return container_of(parent->vm_bo, struct amdgpu_vm_pt, base);
342}
343
344/**
345 * amdgpu_vm_pt_cursor - state for for_each_amdgpu_vm_pt
346 */
347struct amdgpu_vm_pt_cursor {
348 uint64_t pfn;
349 struct amdgpu_vm_pt *parent;
350 struct amdgpu_vm_pt *entry;
351 unsigned level;
352};
353
354/**
355 * amdgpu_vm_pt_start - start PD/PT walk
356 *
357 * @adev: amdgpu_device pointer
358 * @vm: amdgpu_vm structure
359 * @start: start address of the walk
360 * @cursor: state to initialize
361 *
362 * Initialize a amdgpu_vm_pt_cursor to start a walk.
363 */
364static void amdgpu_vm_pt_start(struct amdgpu_device *adev,
365 struct amdgpu_vm *vm, uint64_t start,
366 struct amdgpu_vm_pt_cursor *cursor)
367{
368 cursor->pfn = start;
369 cursor->parent = NULL;
370 cursor->entry = &vm->root;
371 cursor->level = adev->vm_manager.root_level;
372}
373
374/**
375 * amdgpu_vm_pt_descendant - go to child node
376 *
377 * @adev: amdgpu_device pointer
378 * @cursor: current state
379 *
380 * Walk to the child node of the current node.
381 * Returns:
382 * True if the walk was possible, false otherwise.
383 */
384static bool amdgpu_vm_pt_descendant(struct amdgpu_device *adev,
385 struct amdgpu_vm_pt_cursor *cursor)
386{
387 unsigned mask, shift, idx;
388
389 if (!cursor->entry->entries)
390 return false;
391
392 BUG_ON(!cursor->entry->base.bo);
393 mask = amdgpu_vm_entries_mask(adev, cursor->level);
394 shift = amdgpu_vm_level_shift(adev, cursor->level);
395
396 ++cursor->level;
397 idx = (cursor->pfn >> shift) & mask;
398 cursor->parent = cursor->entry;
399 cursor->entry = &cursor->entry->entries[idx];
400 return true;
401}
402
403/**
404 * amdgpu_vm_pt_sibling - go to sibling node
405 *
406 * @adev: amdgpu_device pointer
407 * @cursor: current state
408 *
409 * Walk to the sibling node of the current node.
410 * Returns:
411 * True if the walk was possible, false otherwise.
412 */
413static bool amdgpu_vm_pt_sibling(struct amdgpu_device *adev,
414 struct amdgpu_vm_pt_cursor *cursor)
415{
416 unsigned shift, num_entries;
417
418 /* Root doesn't have a sibling */
419 if (!cursor->parent)
420 return false;
421
422 /* Go to our parents and see if we got a sibling */
423 shift = amdgpu_vm_level_shift(adev, cursor->level - 1);
424 num_entries = amdgpu_vm_num_entries(adev, cursor->level - 1);
425
426 if (cursor->entry == &cursor->parent->entries[num_entries - 1])
427 return false;
428
429 cursor->pfn += 1ULL << shift;
430 cursor->pfn &= ~((1ULL << shift) - 1);
431 ++cursor->entry;
432 return true;
433}
434
435/**
436 * amdgpu_vm_pt_ancestor - go to parent node
437 *
438 * @cursor: current state
439 *
440 * Walk to the parent node of the current node.
441 * Returns:
442 * True if the walk was possible, false otherwise.
443 */
444static bool amdgpu_vm_pt_ancestor(struct amdgpu_vm_pt_cursor *cursor)
445{
446 if (!cursor->parent)
447 return false;
448
449 --cursor->level;
450 cursor->entry = cursor->parent;
451 cursor->parent = amdgpu_vm_pt_parent(cursor->parent);
452 return true;
453}
454
455/**
456 * amdgpu_vm_pt_next - get next PD/PT in hieratchy
457 *
458 * @adev: amdgpu_device pointer
459 * @cursor: current state
460 *
461 * Walk the PD/PT tree to the next node.
462 */
463static void amdgpu_vm_pt_next(struct amdgpu_device *adev,
464 struct amdgpu_vm_pt_cursor *cursor)
465{
466 /* First try a newborn child */
467 if (amdgpu_vm_pt_descendant(adev, cursor))
468 return;
469
470 /* If that didn't worked try to find a sibling */
471 while (!amdgpu_vm_pt_sibling(adev, cursor)) {
472 /* No sibling, go to our parents and grandparents */
473 if (!amdgpu_vm_pt_ancestor(cursor)) {
474 cursor->pfn = ~0ll;
475 return;
476 }
477 }
478}
479
480/**
481 * amdgpu_vm_pt_first_dfs - start a deep first search
482 *
483 * @adev: amdgpu_device structure
484 * @vm: amdgpu_vm structure
485 * @cursor: state to initialize
486 *
487 * Starts a deep first traversal of the PD/PT tree.
488 */
489static void amdgpu_vm_pt_first_dfs(struct amdgpu_device *adev,
490 struct amdgpu_vm *vm,
491 struct amdgpu_vm_pt_cursor *start,
492 struct amdgpu_vm_pt_cursor *cursor)
493{
494 if (start)
495 *cursor = *start;
496 else
497 amdgpu_vm_pt_start(adev, vm, 0, cursor);
498 while (amdgpu_vm_pt_descendant(adev, cursor));
499}
500
501/**
502 * amdgpu_vm_pt_continue_dfs - check if the deep first search should continue
503 *
504 * @start: starting point for the search
505 * @entry: current entry
506 *
507 * Returns:
508 * True when the search should continue, false otherwise.
509 */
510static bool amdgpu_vm_pt_continue_dfs(struct amdgpu_vm_pt_cursor *start,
511 struct amdgpu_vm_pt *entry)
512{
513 return entry && (!start || entry != start->entry);
514}
515
516/**
517 * amdgpu_vm_pt_next_dfs - get the next node for a deep first search
518 *
519 * @adev: amdgpu_device structure
520 * @cursor: current state
521 *
522 * Move the cursor to the next node in a deep first search.
523 */
524static void amdgpu_vm_pt_next_dfs(struct amdgpu_device *adev,
525 struct amdgpu_vm_pt_cursor *cursor)
526{
527 if (!cursor->entry)
528 return;
529
530 if (!cursor->parent)
531 cursor->entry = NULL;
532 else if (amdgpu_vm_pt_sibling(adev, cursor))
533 while (amdgpu_vm_pt_descendant(adev, cursor));
534 else
535 amdgpu_vm_pt_ancestor(cursor);
536}
537
538/**
539 * for_each_amdgpu_vm_pt_dfs_safe - safe deep first search of all PDs/PTs
540 */
541#define for_each_amdgpu_vm_pt_dfs_safe(adev, vm, start, cursor, entry) \
542 for (amdgpu_vm_pt_first_dfs((adev), (vm), (start), &(cursor)), \
543 (entry) = (cursor).entry, amdgpu_vm_pt_next_dfs((adev), &(cursor));\
544 amdgpu_vm_pt_continue_dfs((start), (entry)); \
545 (entry) = (cursor).entry, amdgpu_vm_pt_next_dfs((adev), &(cursor)))
546
547/**
548 * amdgpu_vm_get_pd_bo - add the VM PD to a validation list
549 *
550 * @vm: vm providing the BOs
551 * @validated: head of validation list
552 * @entry: entry to add
553 *
554 * Add the page directory to the list of BOs to
555 * validate for command submission.
556 */
557void amdgpu_vm_get_pd_bo(struct amdgpu_vm *vm,
558 struct list_head *validated,
559 struct amdgpu_bo_list_entry *entry)
560{
561 entry->priority = 0;
562 entry->tv.bo = &vm->root.base.bo->tbo;
563 /* One for the VM updates, one for TTM and one for the CS job */
564 entry->tv.num_shared = 3;
565 entry->user_pages = NULL;
566 list_add(&entry->tv.head, validated);
567}
568
569void amdgpu_vm_del_from_lru_notify(struct ttm_buffer_object *bo)
570{
571 struct amdgpu_bo *abo;
572 struct amdgpu_vm_bo_base *bo_base;
573
574 if (!amdgpu_bo_is_amdgpu_bo(bo))
575 return;
576
577 if (bo->mem.placement & TTM_PL_FLAG_NO_EVICT)
578 return;
579
580 abo = ttm_to_amdgpu_bo(bo);
581 if (!abo->parent)
582 return;
583 for (bo_base = abo->vm_bo; bo_base; bo_base = bo_base->next) {
584 struct amdgpu_vm *vm = bo_base->vm;
585
586 if (abo->tbo.base.resv == vm->root.base.bo->tbo.base.resv)
587 vm->bulk_moveable = false;
588 }
589
590}
591/**
592 * amdgpu_vm_move_to_lru_tail - move all BOs to the end of LRU
593 *
594 * @adev: amdgpu device pointer
595 * @vm: vm providing the BOs
596 *
597 * Move all BOs to the end of LRU and remember their positions to put them
598 * together.
599 */
600void amdgpu_vm_move_to_lru_tail(struct amdgpu_device *adev,
601 struct amdgpu_vm *vm)
602{
603 struct ttm_bo_global *glob = adev->mman.bdev.glob;
604 struct amdgpu_vm_bo_base *bo_base;
605
606 if (vm->bulk_moveable) {
607 spin_lock(&glob->lru_lock);
608 ttm_bo_bulk_move_lru_tail(&vm->lru_bulk_move);
609 spin_unlock(&glob->lru_lock);
610 return;
611 }
612
613 memset(&vm->lru_bulk_move, 0, sizeof(vm->lru_bulk_move));
614
615 spin_lock(&glob->lru_lock);
616 list_for_each_entry(bo_base, &vm->idle, vm_status) {
617 struct amdgpu_bo *bo = bo_base->bo;
618
619 if (!bo->parent)
620 continue;
621
622 ttm_bo_move_to_lru_tail(&bo->tbo, &vm->lru_bulk_move);
623 if (bo->shadow)
624 ttm_bo_move_to_lru_tail(&bo->shadow->tbo,
625 &vm->lru_bulk_move);
626 }
627 spin_unlock(&glob->lru_lock);
628
629 vm->bulk_moveable = true;
630}
631
632/**
633 * amdgpu_vm_validate_pt_bos - validate the page table BOs
634 *
635 * @adev: amdgpu device pointer
636 * @vm: vm providing the BOs
637 * @validate: callback to do the validation
638 * @param: parameter for the validation callback
639 *
640 * Validate the page table BOs on command submission if neccessary.
641 *
642 * Returns:
643 * Validation result.
644 */
645int amdgpu_vm_validate_pt_bos(struct amdgpu_device *adev, struct amdgpu_vm *vm,
646 int (*validate)(void *p, struct amdgpu_bo *bo),
647 void *param)
648{
649 struct amdgpu_vm_bo_base *bo_base, *tmp;
650 int r = 0;
651
652 vm->bulk_moveable &= list_empty(&vm->evicted);
653
654 list_for_each_entry_safe(bo_base, tmp, &vm->evicted, vm_status) {
655 struct amdgpu_bo *bo = bo_base->bo;
656
657 r = validate(param, bo);
658 if (r)
659 break;
660
661 if (bo->tbo.type != ttm_bo_type_kernel) {
662 amdgpu_vm_bo_moved(bo_base);
663 } else {
664 vm->update_funcs->map_table(bo);
665 if (bo->parent)
666 amdgpu_vm_bo_relocated(bo_base);
667 else
668 amdgpu_vm_bo_idle(bo_base);
669 }
670 }
671
672 return r;
673}
674
675/**
676 * amdgpu_vm_ready - check VM is ready for updates
677 *
678 * @vm: VM to check
679 *
680 * Check if all VM PDs/PTs are ready for updates
681 *
682 * Returns:
683 * True if eviction list is empty.
684 */
685bool amdgpu_vm_ready(struct amdgpu_vm *vm)
686{
687 return list_empty(&vm->evicted);
688}
689
690/**
691 * amdgpu_vm_clear_bo - initially clear the PDs/PTs
692 *
693 * @adev: amdgpu_device pointer
694 * @vm: VM to clear BO from
695 * @bo: BO to clear
696 *
697 * Root PD needs to be reserved when calling this.
698 *
699 * Returns:
700 * 0 on success, errno otherwise.
701 */
702static int amdgpu_vm_clear_bo(struct amdgpu_device *adev,
703 struct amdgpu_vm *vm,
704 struct amdgpu_bo *bo)
705{
706 struct ttm_operation_ctx ctx = { true, false };
707 unsigned level = adev->vm_manager.root_level;
708 struct amdgpu_vm_update_params params;
709 struct amdgpu_bo *ancestor = bo;
710 unsigned entries, ats_entries;
711 uint64_t addr;
712 int r;
713
714 /* Figure out our place in the hierarchy */
715 if (ancestor->parent) {
716 ++level;
717 while (ancestor->parent->parent) {
718 ++level;
719 ancestor = ancestor->parent;
720 }
721 }
722
723 entries = amdgpu_bo_size(bo) / 8;
724 if (!vm->pte_support_ats) {
725 ats_entries = 0;
726
727 } else if (!bo->parent) {
728 ats_entries = amdgpu_vm_num_ats_entries(adev);
729 ats_entries = min(ats_entries, entries);
730 entries -= ats_entries;
731
732 } else {
733 struct amdgpu_vm_pt *pt;
734
735 pt = container_of(ancestor->vm_bo, struct amdgpu_vm_pt, base);
736 ats_entries = amdgpu_vm_num_ats_entries(adev);
737 if ((pt - vm->root.entries) >= ats_entries) {
738 ats_entries = 0;
739 } else {
740 ats_entries = entries;
741 entries = 0;
742 }
743 }
744
745 r = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
746 if (r)
747 return r;
748
749 if (bo->shadow) {
750 r = ttm_bo_validate(&bo->shadow->tbo, &bo->shadow->placement,
751 &ctx);
752 if (r)
753 return r;
754 }
755
756 r = vm->update_funcs->map_table(bo);
757 if (r)
758 return r;
759
760 memset(¶ms, 0, sizeof(params));
761 params.adev = adev;
762 params.vm = vm;
763
764 r = vm->update_funcs->prepare(¶ms, AMDGPU_FENCE_OWNER_KFD, NULL);
765 if (r)
766 return r;
767
768 addr = 0;
769 if (ats_entries) {
770 uint64_t value = 0, flags;
771
772 flags = AMDGPU_PTE_DEFAULT_ATC;
773 if (level != AMDGPU_VM_PTB) {
774 /* Handle leaf PDEs as PTEs */
775 flags |= AMDGPU_PDE_PTE;
776 amdgpu_gmc_get_vm_pde(adev, level, &value, &flags);
777 }
778
779 r = vm->update_funcs->update(¶ms, bo, addr, 0, ats_entries,
780 value, flags);
781 if (r)
782 return r;
783
784 addr += ats_entries * 8;
785 }
786
787 if (entries) {
788 uint64_t value = 0, flags = 0;
789
790 if (adev->asic_type >= CHIP_VEGA10) {
791 if (level != AMDGPU_VM_PTB) {
792 /* Handle leaf PDEs as PTEs */
793 flags |= AMDGPU_PDE_PTE;
794 amdgpu_gmc_get_vm_pde(adev, level,
795 &value, &flags);
796 } else {
797 /* Workaround for fault priority problem on GMC9 */
798 flags = AMDGPU_PTE_EXECUTABLE;
799 }
800 }
801
802 r = vm->update_funcs->update(¶ms, bo, addr, 0, entries,
803 value, flags);
804 if (r)
805 return r;
806 }
807
808 return vm->update_funcs->commit(¶ms, NULL);
809}
810
811/**
812 * amdgpu_vm_bo_param - fill in parameters for PD/PT allocation
813 *
814 * @adev: amdgpu_device pointer
815 * @vm: requesting vm
816 * @bp: resulting BO allocation parameters
817 */
818static void amdgpu_vm_bo_param(struct amdgpu_device *adev, struct amdgpu_vm *vm,
819 int level, struct amdgpu_bo_param *bp)
820{
821 memset(bp, 0, sizeof(*bp));
822
823 bp->size = amdgpu_vm_bo_size(adev, level);
824 bp->byte_align = AMDGPU_GPU_PAGE_SIZE;
825 bp->domain = AMDGPU_GEM_DOMAIN_VRAM;
826 bp->domain = amdgpu_bo_get_preferred_pin_domain(adev, bp->domain);
827 bp->flags = AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS |
828 AMDGPU_GEM_CREATE_CPU_GTT_USWC;
829 if (vm->use_cpu_for_update)
830 bp->flags |= AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED;
831 else if (!vm->root.base.bo || vm->root.base.bo->shadow)
832 bp->flags |= AMDGPU_GEM_CREATE_SHADOW;
833 bp->type = ttm_bo_type_kernel;
834 if (vm->root.base.bo)
835 bp->resv = vm->root.base.bo->tbo.base.resv;
836}
837
838/**
839 * amdgpu_vm_alloc_pts - Allocate a specific page table
840 *
841 * @adev: amdgpu_device pointer
842 * @vm: VM to allocate page tables for
843 * @cursor: Which page table to allocate
844 *
845 * Make sure a specific page table or directory is allocated.
846 *
847 * Returns:
848 * 1 if page table needed to be allocated, 0 if page table was already
849 * allocated, negative errno if an error occurred.
850 */
851static int amdgpu_vm_alloc_pts(struct amdgpu_device *adev,
852 struct amdgpu_vm *vm,
853 struct amdgpu_vm_pt_cursor *cursor)
854{
855 struct amdgpu_vm_pt *entry = cursor->entry;
856 struct amdgpu_bo_param bp;
857 struct amdgpu_bo *pt;
858 int r;
859
860 if (cursor->level < AMDGPU_VM_PTB && !entry->entries) {
861 unsigned num_entries;
862
863 num_entries = amdgpu_vm_num_entries(adev, cursor->level);
864 entry->entries = kvmalloc_array(num_entries,
865 sizeof(*entry->entries),
866 GFP_KERNEL | __GFP_ZERO);
867 if (!entry->entries)
868 return -ENOMEM;
869 }
870
871 if (entry->base.bo)
872 return 0;
873
874 amdgpu_vm_bo_param(adev, vm, cursor->level, &bp);
875
876 r = amdgpu_bo_create(adev, &bp, &pt);
877 if (r)
878 return r;
879
880 /* Keep a reference to the root directory to avoid
881 * freeing them up in the wrong order.
882 */
883 pt->parent = amdgpu_bo_ref(cursor->parent->base.bo);
884 amdgpu_vm_bo_base_init(&entry->base, vm, pt);
885
886 r = amdgpu_vm_clear_bo(adev, vm, pt);
887 if (r)
888 goto error_free_pt;
889
890 return 0;
891
892error_free_pt:
893 amdgpu_bo_unref(&pt->shadow);
894 amdgpu_bo_unref(&pt);
895 return r;
896}
897
898/**
899 * amdgpu_vm_free_table - fre one PD/PT
900 *
901 * @entry: PDE to free
902 */
903static void amdgpu_vm_free_table(struct amdgpu_vm_pt *entry)
904{
905 if (entry->base.bo) {
906 entry->base.bo->vm_bo = NULL;
907 list_del(&entry->base.vm_status);
908 amdgpu_bo_unref(&entry->base.bo->shadow);
909 amdgpu_bo_unref(&entry->base.bo);
910 }
911 kvfree(entry->entries);
912 entry->entries = NULL;
913}
914
915/**
916 * amdgpu_vm_free_pts - free PD/PT levels
917 *
918 * @adev: amdgpu device structure
919 * @vm: amdgpu vm structure
920 * @start: optional cursor where to start freeing PDs/PTs
921 *
922 * Free the page directory or page table level and all sub levels.
923 */
924static void amdgpu_vm_free_pts(struct amdgpu_device *adev,
925 struct amdgpu_vm *vm,
926 struct amdgpu_vm_pt_cursor *start)
927{
928 struct amdgpu_vm_pt_cursor cursor;
929 struct amdgpu_vm_pt *entry;
930
931 vm->bulk_moveable = false;
932
933 for_each_amdgpu_vm_pt_dfs_safe(adev, vm, start, cursor, entry)
934 amdgpu_vm_free_table(entry);
935
936 if (start)
937 amdgpu_vm_free_table(start->entry);
938}
939
940/**
941 * amdgpu_vm_check_compute_bug - check whether asic has compute vm bug
942 *
943 * @adev: amdgpu_device pointer
944 */
945void amdgpu_vm_check_compute_bug(struct amdgpu_device *adev)
946{
947 const struct amdgpu_ip_block *ip_block;
948 bool has_compute_vm_bug;
949 struct amdgpu_ring *ring;
950 int i;
951
952 has_compute_vm_bug = false;
953
954 ip_block = amdgpu_device_ip_get_ip_block(adev, AMD_IP_BLOCK_TYPE_GFX);
955 if (ip_block) {
956 /* Compute has a VM bug for GFX version < 7.
957 Compute has a VM bug for GFX 8 MEC firmware version < 673.*/
958 if (ip_block->version->major <= 7)
959 has_compute_vm_bug = true;
960 else if (ip_block->version->major == 8)
961 if (adev->gfx.mec_fw_version < 673)
962 has_compute_vm_bug = true;
963 }
964
965 for (i = 0; i < adev->num_rings; i++) {
966 ring = adev->rings[i];
967 if (ring->funcs->type == AMDGPU_RING_TYPE_COMPUTE)
968 /* only compute rings */
969 ring->has_compute_vm_bug = has_compute_vm_bug;
970 else
971 ring->has_compute_vm_bug = false;
972 }
973}
974
975/**
976 * amdgpu_vm_need_pipeline_sync - Check if pipe sync is needed for job.
977 *
978 * @ring: ring on which the job will be submitted
979 * @job: job to submit
980 *
981 * Returns:
982 * True if sync is needed.
983 */
984bool amdgpu_vm_need_pipeline_sync(struct amdgpu_ring *ring,
985 struct amdgpu_job *job)
986{
987 struct amdgpu_device *adev = ring->adev;
988 unsigned vmhub = ring->funcs->vmhub;
989 struct amdgpu_vmid_mgr *id_mgr = &adev->vm_manager.id_mgr[vmhub];
990 struct amdgpu_vmid *id;
991 bool gds_switch_needed;
992 bool vm_flush_needed = job->vm_needs_flush || ring->has_compute_vm_bug;
993
994 if (job->vmid == 0)
995 return false;
996 id = &id_mgr->ids[job->vmid];
997 gds_switch_needed = ring->funcs->emit_gds_switch && (
998 id->gds_base != job->gds_base ||
999 id->gds_size != job->gds_size ||
1000 id->gws_base != job->gws_base ||
1001 id->gws_size != job->gws_size ||
1002 id->oa_base != job->oa_base ||
1003 id->oa_size != job->oa_size);
1004
1005 if (amdgpu_vmid_had_gpu_reset(adev, id))
1006 return true;
1007
1008 return vm_flush_needed || gds_switch_needed;
1009}
1010
1011/**
1012 * amdgpu_vm_flush - hardware flush the vm
1013 *
1014 * @ring: ring to use for flush
1015 * @job: related job
1016 * @need_pipe_sync: is pipe sync needed
1017 *
1018 * Emit a VM flush when it is necessary.
1019 *
1020 * Returns:
1021 * 0 on success, errno otherwise.
1022 */
1023int amdgpu_vm_flush(struct amdgpu_ring *ring, struct amdgpu_job *job, bool need_pipe_sync)
1024{
1025 struct amdgpu_device *adev = ring->adev;
1026 unsigned vmhub = ring->funcs->vmhub;
1027 struct amdgpu_vmid_mgr *id_mgr = &adev->vm_manager.id_mgr[vmhub];
1028 struct amdgpu_vmid *id = &id_mgr->ids[job->vmid];
1029 bool gds_switch_needed = ring->funcs->emit_gds_switch && (
1030 id->gds_base != job->gds_base ||
1031 id->gds_size != job->gds_size ||
1032 id->gws_base != job->gws_base ||
1033 id->gws_size != job->gws_size ||
1034 id->oa_base != job->oa_base ||
1035 id->oa_size != job->oa_size);
1036 bool vm_flush_needed = job->vm_needs_flush;
1037 bool pasid_mapping_needed = id->pasid != job->pasid ||
1038 !id->pasid_mapping ||
1039 !dma_fence_is_signaled(id->pasid_mapping);
1040 struct dma_fence *fence = NULL;
1041 unsigned patch_offset = 0;
1042 int r;
1043
1044 if (amdgpu_vmid_had_gpu_reset(adev, id)) {
1045 gds_switch_needed = true;
1046 vm_flush_needed = true;
1047 pasid_mapping_needed = true;
1048 }
1049
1050 gds_switch_needed &= !!ring->funcs->emit_gds_switch;
1051 vm_flush_needed &= !!ring->funcs->emit_vm_flush &&
1052 job->vm_pd_addr != AMDGPU_BO_INVALID_OFFSET;
1053 pasid_mapping_needed &= adev->gmc.gmc_funcs->emit_pasid_mapping &&
1054 ring->funcs->emit_wreg;
1055
1056 if (!vm_flush_needed && !gds_switch_needed && !need_pipe_sync)
1057 return 0;
1058
1059 if (ring->funcs->init_cond_exec)
1060 patch_offset = amdgpu_ring_init_cond_exec(ring);
1061
1062 if (need_pipe_sync)
1063 amdgpu_ring_emit_pipeline_sync(ring);
1064
1065 if (vm_flush_needed) {
1066 trace_amdgpu_vm_flush(ring, job->vmid, job->vm_pd_addr);
1067 amdgpu_ring_emit_vm_flush(ring, job->vmid, job->vm_pd_addr);
1068 }
1069
1070 if (pasid_mapping_needed)
1071 amdgpu_gmc_emit_pasid_mapping(ring, job->vmid, job->pasid);
1072
1073 if (vm_flush_needed || pasid_mapping_needed) {
1074 r = amdgpu_fence_emit(ring, &fence, 0);
1075 if (r)
1076 return r;
1077 }
1078
1079 if (vm_flush_needed) {
1080 mutex_lock(&id_mgr->lock);
1081 dma_fence_put(id->last_flush);
1082 id->last_flush = dma_fence_get(fence);
1083 id->current_gpu_reset_count =
1084 atomic_read(&adev->gpu_reset_counter);
1085 mutex_unlock(&id_mgr->lock);
1086 }
1087
1088 if (pasid_mapping_needed) {
1089 id->pasid = job->pasid;
1090 dma_fence_put(id->pasid_mapping);
1091 id->pasid_mapping = dma_fence_get(fence);
1092 }
1093 dma_fence_put(fence);
1094
1095 if (ring->funcs->emit_gds_switch && gds_switch_needed) {
1096 id->gds_base = job->gds_base;
1097 id->gds_size = job->gds_size;
1098 id->gws_base = job->gws_base;
1099 id->gws_size = job->gws_size;
1100 id->oa_base = job->oa_base;
1101 id->oa_size = job->oa_size;
1102 amdgpu_ring_emit_gds_switch(ring, job->vmid, job->gds_base,
1103 job->gds_size, job->gws_base,
1104 job->gws_size, job->oa_base,
1105 job->oa_size);
1106 }
1107
1108 if (ring->funcs->patch_cond_exec)
1109 amdgpu_ring_patch_cond_exec(ring, patch_offset);
1110
1111 /* the double SWITCH_BUFFER here *cannot* be skipped by COND_EXEC */
1112 if (ring->funcs->emit_switch_buffer) {
1113 amdgpu_ring_emit_switch_buffer(ring);
1114 amdgpu_ring_emit_switch_buffer(ring);
1115 }
1116 return 0;
1117}
1118
1119/**
1120 * amdgpu_vm_bo_find - find the bo_va for a specific vm & bo
1121 *
1122 * @vm: requested vm
1123 * @bo: requested buffer object
1124 *
1125 * Find @bo inside the requested vm.
1126 * Search inside the @bos vm list for the requested vm
1127 * Returns the found bo_va or NULL if none is found
1128 *
1129 * Object has to be reserved!
1130 *
1131 * Returns:
1132 * Found bo_va or NULL.
1133 */
1134struct amdgpu_bo_va *amdgpu_vm_bo_find(struct amdgpu_vm *vm,
1135 struct amdgpu_bo *bo)
1136{
1137 struct amdgpu_vm_bo_base *base;
1138
1139 for (base = bo->vm_bo; base; base = base->next) {
1140 if (base->vm != vm)
1141 continue;
1142
1143 return container_of(base, struct amdgpu_bo_va, base);
1144 }
1145 return NULL;
1146}
1147
1148/**
1149 * amdgpu_vm_map_gart - Resolve gart mapping of addr
1150 *
1151 * @pages_addr: optional DMA address to use for lookup
1152 * @addr: the unmapped addr
1153 *
1154 * Look up the physical address of the page that the pte resolves
1155 * to.
1156 *
1157 * Returns:
1158 * The pointer for the page table entry.
1159 */
1160uint64_t amdgpu_vm_map_gart(const dma_addr_t *pages_addr, uint64_t addr)
1161{
1162 uint64_t result;
1163
1164 /* page table offset */
1165 result = pages_addr[addr >> PAGE_SHIFT];
1166
1167 /* in case cpu page size != gpu page size*/
1168 result |= addr & (~PAGE_MASK);
1169
1170 result &= 0xFFFFFFFFFFFFF000ULL;
1171
1172 return result;
1173}
1174
1175/*
1176 * amdgpu_vm_update_pde - update a single level in the hierarchy
1177 *
1178 * @param: parameters for the update
1179 * @vm: requested vm
1180 * @entry: entry to update
1181 *
1182 * Makes sure the requested entry in parent is up to date.
1183 */
1184static int amdgpu_vm_update_pde(struct amdgpu_vm_update_params *params,
1185 struct amdgpu_vm *vm,
1186 struct amdgpu_vm_pt *entry)
1187{
1188 struct amdgpu_vm_pt *parent = amdgpu_vm_pt_parent(entry);
1189 struct amdgpu_bo *bo = parent->base.bo, *pbo;
1190 uint64_t pde, pt, flags;
1191 unsigned level;
1192
1193 for (level = 0, pbo = bo->parent; pbo; ++level)
1194 pbo = pbo->parent;
1195
1196 level += params->adev->vm_manager.root_level;
1197 amdgpu_gmc_get_pde_for_bo(entry->base.bo, level, &pt, &flags);
1198 pde = (entry - parent->entries) * 8;
1199 return vm->update_funcs->update(params, bo, pde, pt, 1, 0, flags);
1200}
1201
1202/*
1203 * amdgpu_vm_invalidate_pds - mark all PDs as invalid
1204 *
1205 * @adev: amdgpu_device pointer
1206 * @vm: related vm
1207 *
1208 * Mark all PD level as invalid after an error.
1209 */
1210static void amdgpu_vm_invalidate_pds(struct amdgpu_device *adev,
1211 struct amdgpu_vm *vm)
1212{
1213 struct amdgpu_vm_pt_cursor cursor;
1214 struct amdgpu_vm_pt *entry;
1215
1216 for_each_amdgpu_vm_pt_dfs_safe(adev, vm, NULL, cursor, entry)
1217 if (entry->base.bo && !entry->base.moved)
1218 amdgpu_vm_bo_relocated(&entry->base);
1219}
1220
1221/*
1222 * amdgpu_vm_update_directories - make sure that all directories are valid
1223 *
1224 * @adev: amdgpu_device pointer
1225 * @vm: requested vm
1226 *
1227 * Makes sure all directories are up to date.
1228 *
1229 * Returns:
1230 * 0 for success, error for failure.
1231 */
1232int amdgpu_vm_update_directories(struct amdgpu_device *adev,
1233 struct amdgpu_vm *vm)
1234{
1235 struct amdgpu_vm_update_params params;
1236 int r;
1237
1238 if (list_empty(&vm->relocated))
1239 return 0;
1240
1241 memset(¶ms, 0, sizeof(params));
1242 params.adev = adev;
1243 params.vm = vm;
1244
1245 r = vm->update_funcs->prepare(¶ms, AMDGPU_FENCE_OWNER_VM, NULL);
1246 if (r)
1247 return r;
1248
1249 while (!list_empty(&vm->relocated)) {
1250 struct amdgpu_vm_pt *entry;
1251
1252 entry = list_first_entry(&vm->relocated, struct amdgpu_vm_pt,
1253 base.vm_status);
1254 amdgpu_vm_bo_idle(&entry->base);
1255
1256 r = amdgpu_vm_update_pde(¶ms, vm, entry);
1257 if (r)
1258 goto error;
1259 }
1260
1261 r = vm->update_funcs->commit(¶ms, &vm->last_update);
1262 if (r)
1263 goto error;
1264 return 0;
1265
1266error:
1267 amdgpu_vm_invalidate_pds(adev, vm);
1268 return r;
1269}
1270
1271/**
1272 * amdgpu_vm_update_flags - figure out flags for PTE updates
1273 *
1274 * Make sure to set the right flags for the PTEs at the desired level.
1275 */
1276static void amdgpu_vm_update_flags(struct amdgpu_vm_update_params *params,
1277 struct amdgpu_bo *bo, unsigned level,
1278 uint64_t pe, uint64_t addr,
1279 unsigned count, uint32_t incr,
1280 uint64_t flags)
1281
1282{
1283 if (level != AMDGPU_VM_PTB) {
1284 flags |= AMDGPU_PDE_PTE;
1285 amdgpu_gmc_get_vm_pde(params->adev, level, &addr, &flags);
1286
1287 } else if (params->adev->asic_type >= CHIP_VEGA10 &&
1288 !(flags & AMDGPU_PTE_VALID) &&
1289 !(flags & AMDGPU_PTE_PRT)) {
1290
1291 /* Workaround for fault priority problem on GMC9 */
1292 flags |= AMDGPU_PTE_EXECUTABLE;
1293 }
1294
1295 params->vm->update_funcs->update(params, bo, pe, addr, count, incr,
1296 flags);
1297}
1298
1299/**
1300 * amdgpu_vm_fragment - get fragment for PTEs
1301 *
1302 * @params: see amdgpu_vm_update_params definition
1303 * @start: first PTE to handle
1304 * @end: last PTE to handle
1305 * @flags: hw mapping flags
1306 * @frag: resulting fragment size
1307 * @frag_end: end of this fragment
1308 *
1309 * Returns the first possible fragment for the start and end address.
1310 */
1311static void amdgpu_vm_fragment(struct amdgpu_vm_update_params *params,
1312 uint64_t start, uint64_t end, uint64_t flags,
1313 unsigned int *frag, uint64_t *frag_end)
1314{
1315 /**
1316 * The MC L1 TLB supports variable sized pages, based on a fragment
1317 * field in the PTE. When this field is set to a non-zero value, page
1318 * granularity is increased from 4KB to (1 << (12 + frag)). The PTE
1319 * flags are considered valid for all PTEs within the fragment range
1320 * and corresponding mappings are assumed to be physically contiguous.
1321 *
1322 * The L1 TLB can store a single PTE for the whole fragment,
1323 * significantly increasing the space available for translation
1324 * caching. This leads to large improvements in throughput when the
1325 * TLB is under pressure.
1326 *
1327 * The L2 TLB distributes small and large fragments into two
1328 * asymmetric partitions. The large fragment cache is significantly
1329 * larger. Thus, we try to use large fragments wherever possible.
1330 * Userspace can support this by aligning virtual base address and
1331 * allocation size to the fragment size.
1332 *
1333 * Starting with Vega10 the fragment size only controls the L1. The L2
1334 * is now directly feed with small/huge/giant pages from the walker.
1335 */
1336 unsigned max_frag;
1337
1338 if (params->adev->asic_type < CHIP_VEGA10)
1339 max_frag = params->adev->vm_manager.fragment_size;
1340 else
1341 max_frag = 31;
1342
1343 /* system pages are non continuously */
1344 if (params->pages_addr) {
1345 *frag = 0;
1346 *frag_end = end;
1347 return;
1348 }
1349
1350 /* This intentionally wraps around if no bit is set */
1351 *frag = min((unsigned)ffs(start) - 1, (unsigned)fls64(end - start) - 1);
1352 if (*frag >= max_frag) {
1353 *frag = max_frag;
1354 *frag_end = end & ~((1ULL << max_frag) - 1);
1355 } else {
1356 *frag_end = start + (1 << *frag);
1357 }
1358}
1359
1360/**
1361 * amdgpu_vm_update_ptes - make sure that page tables are valid
1362 *
1363 * @params: see amdgpu_vm_update_params definition
1364 * @start: start of GPU address range
1365 * @end: end of GPU address range
1366 * @dst: destination address to map to, the next dst inside the function
1367 * @flags: mapping flags
1368 *
1369 * Update the page tables in the range @start - @end.
1370 *
1371 * Returns:
1372 * 0 for success, -EINVAL for failure.
1373 */
1374static int amdgpu_vm_update_ptes(struct amdgpu_vm_update_params *params,
1375 uint64_t start, uint64_t end,
1376 uint64_t dst, uint64_t flags)
1377{
1378 struct amdgpu_device *adev = params->adev;
1379 struct amdgpu_vm_pt_cursor cursor;
1380 uint64_t frag_start = start, frag_end;
1381 unsigned int frag;
1382 int r;
1383
1384 /* figure out the initial fragment */
1385 amdgpu_vm_fragment(params, frag_start, end, flags, &frag, &frag_end);
1386
1387 /* walk over the address space and update the PTs */
1388 amdgpu_vm_pt_start(adev, params->vm, start, &cursor);
1389 while (cursor.pfn < end) {
1390 unsigned shift, parent_shift, mask;
1391 uint64_t incr, entry_end, pe_start;
1392 struct amdgpu_bo *pt;
1393
1394 r = amdgpu_vm_alloc_pts(params->adev, params->vm, &cursor);
1395 if (r)
1396 return r;
1397
1398 pt = cursor.entry->base.bo;
1399
1400 /* The root level can't be a huge page */
1401 if (cursor.level == adev->vm_manager.root_level) {
1402 if (!amdgpu_vm_pt_descendant(adev, &cursor))
1403 return -ENOENT;
1404 continue;
1405 }
1406
1407 shift = amdgpu_vm_level_shift(adev, cursor.level);
1408 parent_shift = amdgpu_vm_level_shift(adev, cursor.level - 1);
1409 if (adev->asic_type < CHIP_VEGA10 &&
1410 (flags & AMDGPU_PTE_VALID)) {
1411 /* No huge page support before GMC v9 */
1412 if (cursor.level != AMDGPU_VM_PTB) {
1413 if (!amdgpu_vm_pt_descendant(adev, &cursor))
1414 return -ENOENT;
1415 continue;
1416 }
1417 } else if (frag < shift) {
1418 /* We can't use this level when the fragment size is
1419 * smaller than the address shift. Go to the next
1420 * child entry and try again.
1421 */
1422 if (!amdgpu_vm_pt_descendant(adev, &cursor))
1423 return -ENOENT;
1424 continue;
1425 } else if (frag >= parent_shift &&
1426 cursor.level - 1 != adev->vm_manager.root_level) {
1427 /* If the fragment size is even larger than the parent
1428 * shift we should go up one level and check it again
1429 * unless one level up is the root level.
1430 */
1431 if (!amdgpu_vm_pt_ancestor(&cursor))
1432 return -ENOENT;
1433 continue;
1434 }
1435
1436 /* Looks good so far, calculate parameters for the update */
1437 incr = (uint64_t)AMDGPU_GPU_PAGE_SIZE << shift;
1438 mask = amdgpu_vm_entries_mask(adev, cursor.level);
1439 pe_start = ((cursor.pfn >> shift) & mask) * 8;
1440 entry_end = (uint64_t)(mask + 1) << shift;
1441 entry_end += cursor.pfn & ~(entry_end - 1);
1442 entry_end = min(entry_end, end);
1443
1444 do {
1445 uint64_t upd_end = min(entry_end, frag_end);
1446 unsigned nptes = (upd_end - frag_start) >> shift;
1447
1448 amdgpu_vm_update_flags(params, pt, cursor.level,
1449 pe_start, dst, nptes, incr,
1450 flags | AMDGPU_PTE_FRAG(frag));
1451
1452 pe_start += nptes * 8;
1453 dst += (uint64_t)nptes * AMDGPU_GPU_PAGE_SIZE << shift;
1454
1455 frag_start = upd_end;
1456 if (frag_start >= frag_end) {
1457 /* figure out the next fragment */
1458 amdgpu_vm_fragment(params, frag_start, end,
1459 flags, &frag, &frag_end);
1460 if (frag < shift)
1461 break;
1462 }
1463 } while (frag_start < entry_end);
1464
1465 if (amdgpu_vm_pt_descendant(adev, &cursor)) {
1466 /* Free all child entries */
1467 while (cursor.pfn < frag_start) {
1468 amdgpu_vm_free_pts(adev, params->vm, &cursor);
1469 amdgpu_vm_pt_next(adev, &cursor);
1470 }
1471
1472 } else if (frag >= shift) {
1473 /* or just move on to the next on the same level. */
1474 amdgpu_vm_pt_next(adev, &cursor);
1475 }
1476 }
1477
1478 return 0;
1479}
1480
1481/**
1482 * amdgpu_vm_bo_update_mapping - update a mapping in the vm page table
1483 *
1484 * @adev: amdgpu_device pointer
1485 * @exclusive: fence we need to sync to
1486 * @pages_addr: DMA addresses to use for mapping
1487 * @vm: requested vm
1488 * @start: start of mapped range
1489 * @last: last mapped entry
1490 * @flags: flags for the entries
1491 * @addr: addr to set the area to
1492 * @fence: optional resulting fence
1493 *
1494 * Fill in the page table entries between @start and @last.
1495 *
1496 * Returns:
1497 * 0 for success, -EINVAL for failure.
1498 */
1499static int amdgpu_vm_bo_update_mapping(struct amdgpu_device *adev,
1500 struct dma_fence *exclusive,
1501 dma_addr_t *pages_addr,
1502 struct amdgpu_vm *vm,
1503 uint64_t start, uint64_t last,
1504 uint64_t flags, uint64_t addr,
1505 struct dma_fence **fence)
1506{
1507 struct amdgpu_vm_update_params params;
1508 void *owner = AMDGPU_FENCE_OWNER_VM;
1509 int r;
1510
1511 memset(¶ms, 0, sizeof(params));
1512 params.adev = adev;
1513 params.vm = vm;
1514 params.pages_addr = pages_addr;
1515
1516 /* sync to everything except eviction fences on unmapping */
1517 if (!(flags & AMDGPU_PTE_VALID))
1518 owner = AMDGPU_FENCE_OWNER_KFD;
1519
1520 r = vm->update_funcs->prepare(¶ms, owner, exclusive);
1521 if (r)
1522 return r;
1523
1524 r = amdgpu_vm_update_ptes(¶ms, start, last + 1, addr, flags);
1525 if (r)
1526 return r;
1527
1528 return vm->update_funcs->commit(¶ms, fence);
1529}
1530
1531/**
1532 * amdgpu_vm_bo_split_mapping - split a mapping into smaller chunks
1533 *
1534 * @adev: amdgpu_device pointer
1535 * @exclusive: fence we need to sync to
1536 * @pages_addr: DMA addresses to use for mapping
1537 * @vm: requested vm
1538 * @mapping: mapped range and flags to use for the update
1539 * @flags: HW flags for the mapping
1540 * @bo_adev: amdgpu_device pointer that bo actually been allocated
1541 * @nodes: array of drm_mm_nodes with the MC addresses
1542 * @fence: optional resulting fence
1543 *
1544 * Split the mapping into smaller chunks so that each update fits
1545 * into a SDMA IB.
1546 *
1547 * Returns:
1548 * 0 for success, -EINVAL for failure.
1549 */
1550static int amdgpu_vm_bo_split_mapping(struct amdgpu_device *adev,
1551 struct dma_fence *exclusive,
1552 dma_addr_t *pages_addr,
1553 struct amdgpu_vm *vm,
1554 struct amdgpu_bo_va_mapping *mapping,
1555 uint64_t flags,
1556 struct amdgpu_device *bo_adev,
1557 struct drm_mm_node *nodes,
1558 struct dma_fence **fence)
1559{
1560 unsigned min_linear_pages = 1 << adev->vm_manager.fragment_size;
1561 uint64_t pfn, start = mapping->start;
1562 int r;
1563
1564 /* normally,bo_va->flags only contians READABLE and WIRTEABLE bit go here
1565 * but in case of something, we filter the flags in first place
1566 */
1567 if (!(mapping->flags & AMDGPU_PTE_READABLE))
1568 flags &= ~AMDGPU_PTE_READABLE;
1569 if (!(mapping->flags & AMDGPU_PTE_WRITEABLE))
1570 flags &= ~AMDGPU_PTE_WRITEABLE;
1571
1572 flags &= ~AMDGPU_PTE_EXECUTABLE;
1573 flags |= mapping->flags & AMDGPU_PTE_EXECUTABLE;
1574
1575 if (adev->asic_type >= CHIP_NAVI10) {
1576 flags &= ~AMDGPU_PTE_MTYPE_NV10_MASK;
1577 flags |= (mapping->flags & AMDGPU_PTE_MTYPE_NV10_MASK);
1578 } else {
1579 flags &= ~AMDGPU_PTE_MTYPE_VG10_MASK;
1580 flags |= (mapping->flags & AMDGPU_PTE_MTYPE_VG10_MASK);
1581 }
1582
1583 if ((mapping->flags & AMDGPU_PTE_PRT) &&
1584 (adev->asic_type >= CHIP_VEGA10)) {
1585 flags |= AMDGPU_PTE_PRT;
1586 if (adev->asic_type >= CHIP_NAVI10) {
1587 flags |= AMDGPU_PTE_SNOOPED;
1588 flags |= AMDGPU_PTE_LOG;
1589 flags |= AMDGPU_PTE_SYSTEM;
1590 }
1591 flags &= ~AMDGPU_PTE_VALID;
1592 }
1593
1594 trace_amdgpu_vm_bo_update(mapping);
1595
1596 pfn = mapping->offset >> PAGE_SHIFT;
1597 if (nodes) {
1598 while (pfn >= nodes->size) {
1599 pfn -= nodes->size;
1600 ++nodes;
1601 }
1602 }
1603
1604 do {
1605 dma_addr_t *dma_addr = NULL;
1606 uint64_t max_entries;
1607 uint64_t addr, last;
1608
1609 if (nodes) {
1610 addr = nodes->start << PAGE_SHIFT;
1611 max_entries = (nodes->size - pfn) *
1612 AMDGPU_GPU_PAGES_IN_CPU_PAGE;
1613 } else {
1614 addr = 0;
1615 max_entries = S64_MAX;
1616 }
1617
1618 if (pages_addr) {
1619 uint64_t count;
1620
1621 for (count = 1;
1622 count < max_entries / AMDGPU_GPU_PAGES_IN_CPU_PAGE;
1623 ++count) {
1624 uint64_t idx = pfn + count;
1625
1626 if (pages_addr[idx] !=
1627 (pages_addr[idx - 1] + PAGE_SIZE))
1628 break;
1629 }
1630
1631 if (count < min_linear_pages) {
1632 addr = pfn << PAGE_SHIFT;
1633 dma_addr = pages_addr;
1634 } else {
1635 addr = pages_addr[pfn];
1636 max_entries = count * AMDGPU_GPU_PAGES_IN_CPU_PAGE;
1637 }
1638
1639 } else if (flags & AMDGPU_PTE_VALID) {
1640 addr += bo_adev->vm_manager.vram_base_offset;
1641 addr += pfn << PAGE_SHIFT;
1642 }
1643
1644 last = min((uint64_t)mapping->last, start + max_entries - 1);
1645 r = amdgpu_vm_bo_update_mapping(adev, exclusive, dma_addr, vm,
1646 start, last, flags, addr,
1647 fence);
1648 if (r)
1649 return r;
1650
1651 pfn += (last - start + 1) / AMDGPU_GPU_PAGES_IN_CPU_PAGE;
1652 if (nodes && nodes->size == pfn) {
1653 pfn = 0;
1654 ++nodes;
1655 }
1656 start = last + 1;
1657
1658 } while (unlikely(start != mapping->last + 1));
1659
1660 return 0;
1661}
1662
1663/**
1664 * amdgpu_vm_bo_update - update all BO mappings in the vm page table
1665 *
1666 * @adev: amdgpu_device pointer
1667 * @bo_va: requested BO and VM object
1668 * @clear: if true clear the entries
1669 *
1670 * Fill in the page table entries for @bo_va.
1671 *
1672 * Returns:
1673 * 0 for success, -EINVAL for failure.
1674 */
1675int amdgpu_vm_bo_update(struct amdgpu_device *adev,
1676 struct amdgpu_bo_va *bo_va,
1677 bool clear)
1678{
1679 struct amdgpu_bo *bo = bo_va->base.bo;
1680 struct amdgpu_vm *vm = bo_va->base.vm;
1681 struct amdgpu_bo_va_mapping *mapping;
1682 dma_addr_t *pages_addr = NULL;
1683 struct ttm_mem_reg *mem;
1684 struct drm_mm_node *nodes;
1685 struct dma_fence *exclusive, **last_update;
1686 uint64_t flags;
1687 struct amdgpu_device *bo_adev = adev;
1688 int r;
1689
1690 if (clear || !bo) {
1691 mem = NULL;
1692 nodes = NULL;
1693 exclusive = NULL;
1694 } else {
1695 struct ttm_dma_tt *ttm;
1696
1697 mem = &bo->tbo.mem;
1698 nodes = mem->mm_node;
1699 if (mem->mem_type == TTM_PL_TT) {
1700 ttm = container_of(bo->tbo.ttm, struct ttm_dma_tt, ttm);
1701 pages_addr = ttm->dma_address;
1702 }
1703 exclusive = dma_resv_get_excl(bo->tbo.base.resv);
1704 }
1705
1706 if (bo) {
1707 flags = amdgpu_ttm_tt_pte_flags(adev, bo->tbo.ttm, mem);
1708 bo_adev = amdgpu_ttm_adev(bo->tbo.bdev);
1709 } else {
1710 flags = 0x0;
1711 }
1712
1713 if (clear || (bo && bo->tbo.base.resv == vm->root.base.bo->tbo.base.resv))
1714 last_update = &vm->last_update;
1715 else
1716 last_update = &bo_va->last_pt_update;
1717
1718 if (!clear && bo_va->base.moved) {
1719 bo_va->base.moved = false;
1720 list_splice_init(&bo_va->valids, &bo_va->invalids);
1721
1722 } else if (bo_va->cleared != clear) {
1723 list_splice_init(&bo_va->valids, &bo_va->invalids);
1724 }
1725
1726 list_for_each_entry(mapping, &bo_va->invalids, list) {
1727 r = amdgpu_vm_bo_split_mapping(adev, exclusive, pages_addr, vm,
1728 mapping, flags, bo_adev, nodes,
1729 last_update);
1730 if (r)
1731 return r;
1732 }
1733
1734 if (vm->use_cpu_for_update) {
1735 /* Flush HDP */
1736 mb();
1737 amdgpu_asic_flush_hdp(adev, NULL);
1738 }
1739
1740 /* If the BO is not in its preferred location add it back to
1741 * the evicted list so that it gets validated again on the
1742 * next command submission.
1743 */
1744 if (bo && bo->tbo.base.resv == vm->root.base.bo->tbo.base.resv) {
1745 uint32_t mem_type = bo->tbo.mem.mem_type;
1746
1747 if (!(bo->preferred_domains & amdgpu_mem_type_to_domain(mem_type)))
1748 amdgpu_vm_bo_evicted(&bo_va->base);
1749 else
1750 amdgpu_vm_bo_idle(&bo_va->base);
1751 } else {
1752 amdgpu_vm_bo_done(&bo_va->base);
1753 }
1754
1755 list_splice_init(&bo_va->invalids, &bo_va->valids);
1756 bo_va->cleared = clear;
1757
1758 if (trace_amdgpu_vm_bo_mapping_enabled()) {
1759 list_for_each_entry(mapping, &bo_va->valids, list)
1760 trace_amdgpu_vm_bo_mapping(mapping);
1761 }
1762
1763 return 0;
1764}
1765
1766/**
1767 * amdgpu_vm_update_prt_state - update the global PRT state
1768 *
1769 * @adev: amdgpu_device pointer
1770 */
1771static void amdgpu_vm_update_prt_state(struct amdgpu_device *adev)
1772{
1773 unsigned long flags;
1774 bool enable;
1775
1776 spin_lock_irqsave(&adev->vm_manager.prt_lock, flags);
1777 enable = !!atomic_read(&adev->vm_manager.num_prt_users);
1778 adev->gmc.gmc_funcs->set_prt(adev, enable);
1779 spin_unlock_irqrestore(&adev->vm_manager.prt_lock, flags);
1780}
1781
1782/**
1783 * amdgpu_vm_prt_get - add a PRT user
1784 *
1785 * @adev: amdgpu_device pointer
1786 */
1787static void amdgpu_vm_prt_get(struct amdgpu_device *adev)
1788{
1789 if (!adev->gmc.gmc_funcs->set_prt)
1790 return;
1791
1792 if (atomic_inc_return(&adev->vm_manager.num_prt_users) == 1)
1793 amdgpu_vm_update_prt_state(adev);
1794}
1795
1796/**
1797 * amdgpu_vm_prt_put - drop a PRT user
1798 *
1799 * @adev: amdgpu_device pointer
1800 */
1801static void amdgpu_vm_prt_put(struct amdgpu_device *adev)
1802{
1803 if (atomic_dec_return(&adev->vm_manager.num_prt_users) == 0)
1804 amdgpu_vm_update_prt_state(adev);
1805}
1806
1807/**
1808 * amdgpu_vm_prt_cb - callback for updating the PRT status
1809 *
1810 * @fence: fence for the callback
1811 * @_cb: the callback function
1812 */
1813static void amdgpu_vm_prt_cb(struct dma_fence *fence, struct dma_fence_cb *_cb)
1814{
1815 struct amdgpu_prt_cb *cb = container_of(_cb, struct amdgpu_prt_cb, cb);
1816
1817 amdgpu_vm_prt_put(cb->adev);
1818 kfree(cb);
1819}
1820
1821/**
1822 * amdgpu_vm_add_prt_cb - add callback for updating the PRT status
1823 *
1824 * @adev: amdgpu_device pointer
1825 * @fence: fence for the callback
1826 */
1827static void amdgpu_vm_add_prt_cb(struct amdgpu_device *adev,
1828 struct dma_fence *fence)
1829{
1830 struct amdgpu_prt_cb *cb;
1831
1832 if (!adev->gmc.gmc_funcs->set_prt)
1833 return;
1834
1835 cb = kmalloc(sizeof(struct amdgpu_prt_cb), GFP_KERNEL);
1836 if (!cb) {
1837 /* Last resort when we are OOM */
1838 if (fence)
1839 dma_fence_wait(fence, false);
1840
1841 amdgpu_vm_prt_put(adev);
1842 } else {
1843 cb->adev = adev;
1844 if (!fence || dma_fence_add_callback(fence, &cb->cb,
1845 amdgpu_vm_prt_cb))
1846 amdgpu_vm_prt_cb(fence, &cb->cb);
1847 }
1848}
1849
1850/**
1851 * amdgpu_vm_free_mapping - free a mapping
1852 *
1853 * @adev: amdgpu_device pointer
1854 * @vm: requested vm
1855 * @mapping: mapping to be freed
1856 * @fence: fence of the unmap operation
1857 *
1858 * Free a mapping and make sure we decrease the PRT usage count if applicable.
1859 */
1860static void amdgpu_vm_free_mapping(struct amdgpu_device *adev,
1861 struct amdgpu_vm *vm,
1862 struct amdgpu_bo_va_mapping *mapping,
1863 struct dma_fence *fence)
1864{
1865 if (mapping->flags & AMDGPU_PTE_PRT)
1866 amdgpu_vm_add_prt_cb(adev, fence);
1867 kfree(mapping);
1868}
1869
1870/**
1871 * amdgpu_vm_prt_fini - finish all prt mappings
1872 *
1873 * @adev: amdgpu_device pointer
1874 * @vm: requested vm
1875 *
1876 * Register a cleanup callback to disable PRT support after VM dies.
1877 */
1878static void amdgpu_vm_prt_fini(struct amdgpu_device *adev, struct amdgpu_vm *vm)
1879{
1880 struct dma_resv *resv = vm->root.base.bo->tbo.base.resv;
1881 struct dma_fence *excl, **shared;
1882 unsigned i, shared_count;
1883 int r;
1884
1885 r = dma_resv_get_fences_rcu(resv, &excl,
1886 &shared_count, &shared);
1887 if (r) {
1888 /* Not enough memory to grab the fence list, as last resort
1889 * block for all the fences to complete.
1890 */
1891 dma_resv_wait_timeout_rcu(resv, true, false,
1892 MAX_SCHEDULE_TIMEOUT);
1893 return;
1894 }
1895
1896 /* Add a callback for each fence in the reservation object */
1897 amdgpu_vm_prt_get(adev);
1898 amdgpu_vm_add_prt_cb(adev, excl);
1899
1900 for (i = 0; i < shared_count; ++i) {
1901 amdgpu_vm_prt_get(adev);
1902 amdgpu_vm_add_prt_cb(adev, shared[i]);
1903 }
1904
1905 kfree(shared);
1906}
1907
1908/**
1909 * amdgpu_vm_clear_freed - clear freed BOs in the PT
1910 *
1911 * @adev: amdgpu_device pointer
1912 * @vm: requested vm
1913 * @fence: optional resulting fence (unchanged if no work needed to be done
1914 * or if an error occurred)
1915 *
1916 * Make sure all freed BOs are cleared in the PT.
1917 * PTs have to be reserved and mutex must be locked!
1918 *
1919 * Returns:
1920 * 0 for success.
1921 *
1922 */
1923int amdgpu_vm_clear_freed(struct amdgpu_device *adev,
1924 struct amdgpu_vm *vm,
1925 struct dma_fence **fence)
1926{
1927 struct amdgpu_bo_va_mapping *mapping;
1928 uint64_t init_pte_value = 0;
1929 struct dma_fence *f = NULL;
1930 int r;
1931
1932 while (!list_empty(&vm->freed)) {
1933 mapping = list_first_entry(&vm->freed,
1934 struct amdgpu_bo_va_mapping, list);
1935 list_del(&mapping->list);
1936
1937 if (vm->pte_support_ats &&
1938 mapping->start < AMDGPU_GMC_HOLE_START)
1939 init_pte_value = AMDGPU_PTE_DEFAULT_ATC;
1940
1941 r = amdgpu_vm_bo_update_mapping(adev, NULL, NULL, vm,
1942 mapping->start, mapping->last,
1943 init_pte_value, 0, &f);
1944 amdgpu_vm_free_mapping(adev, vm, mapping, f);
1945 if (r) {
1946 dma_fence_put(f);
1947 return r;
1948 }
1949 }
1950
1951 if (fence && f) {
1952 dma_fence_put(*fence);
1953 *fence = f;
1954 } else {
1955 dma_fence_put(f);
1956 }
1957
1958 return 0;
1959
1960}
1961
1962/**
1963 * amdgpu_vm_handle_moved - handle moved BOs in the PT
1964 *
1965 * @adev: amdgpu_device pointer
1966 * @vm: requested vm
1967 *
1968 * Make sure all BOs which are moved are updated in the PTs.
1969 *
1970 * Returns:
1971 * 0 for success.
1972 *
1973 * PTs have to be reserved!
1974 */
1975int amdgpu_vm_handle_moved(struct amdgpu_device *adev,
1976 struct amdgpu_vm *vm)
1977{
1978 struct amdgpu_bo_va *bo_va, *tmp;
1979 struct dma_resv *resv;
1980 bool clear;
1981 int r;
1982
1983 list_for_each_entry_safe(bo_va, tmp, &vm->moved, base.vm_status) {
1984 /* Per VM BOs never need to bo cleared in the page tables */
1985 r = amdgpu_vm_bo_update(adev, bo_va, false);
1986 if (r)
1987 return r;
1988 }
1989
1990 spin_lock(&vm->invalidated_lock);
1991 while (!list_empty(&vm->invalidated)) {
1992 bo_va = list_first_entry(&vm->invalidated, struct amdgpu_bo_va,
1993 base.vm_status);
1994 resv = bo_va->base.bo->tbo.base.resv;
1995 spin_unlock(&vm->invalidated_lock);
1996
1997 /* Try to reserve the BO to avoid clearing its ptes */
1998 if (!amdgpu_vm_debug && dma_resv_trylock(resv))
1999 clear = false;
2000 /* Somebody else is using the BO right now */
2001 else
2002 clear = true;
2003
2004 r = amdgpu_vm_bo_update(adev, bo_va, clear);
2005 if (r)
2006 return r;
2007
2008 if (!clear)
2009 dma_resv_unlock(resv);
2010 spin_lock(&vm->invalidated_lock);
2011 }
2012 spin_unlock(&vm->invalidated_lock);
2013
2014 return 0;
2015}
2016
2017/**
2018 * amdgpu_vm_bo_add - add a bo to a specific vm
2019 *
2020 * @adev: amdgpu_device pointer
2021 * @vm: requested vm
2022 * @bo: amdgpu buffer object
2023 *
2024 * Add @bo into the requested vm.
2025 * Add @bo to the list of bos associated with the vm
2026 *
2027 * Returns:
2028 * Newly added bo_va or NULL for failure
2029 *
2030 * Object has to be reserved!
2031 */
2032struct amdgpu_bo_va *amdgpu_vm_bo_add(struct amdgpu_device *adev,
2033 struct amdgpu_vm *vm,
2034 struct amdgpu_bo *bo)
2035{
2036 struct amdgpu_bo_va *bo_va;
2037
2038 bo_va = kzalloc(sizeof(struct amdgpu_bo_va), GFP_KERNEL);
2039 if (bo_va == NULL) {
2040 return NULL;
2041 }
2042 amdgpu_vm_bo_base_init(&bo_va->base, vm, bo);
2043
2044 bo_va->ref_count = 1;
2045 INIT_LIST_HEAD(&bo_va->valids);
2046 INIT_LIST_HEAD(&bo_va->invalids);
2047
2048 if (bo && amdgpu_xgmi_same_hive(adev, amdgpu_ttm_adev(bo->tbo.bdev)) &&
2049 (bo->preferred_domains & AMDGPU_GEM_DOMAIN_VRAM)) {
2050 bo_va->is_xgmi = true;
2051 mutex_lock(&adev->vm_manager.lock_pstate);
2052 /* Power up XGMI if it can be potentially used */
2053 if (++adev->vm_manager.xgmi_map_counter == 1)
2054 amdgpu_xgmi_set_pstate(adev, 1);
2055 mutex_unlock(&adev->vm_manager.lock_pstate);
2056 }
2057
2058 return bo_va;
2059}
2060
2061
2062/**
2063 * amdgpu_vm_bo_insert_mapping - insert a new mapping
2064 *
2065 * @adev: amdgpu_device pointer
2066 * @bo_va: bo_va to store the address
2067 * @mapping: the mapping to insert
2068 *
2069 * Insert a new mapping into all structures.
2070 */
2071static void amdgpu_vm_bo_insert_map(struct amdgpu_device *adev,
2072 struct amdgpu_bo_va *bo_va,
2073 struct amdgpu_bo_va_mapping *mapping)
2074{
2075 struct amdgpu_vm *vm = bo_va->base.vm;
2076 struct amdgpu_bo *bo = bo_va->base.bo;
2077
2078 mapping->bo_va = bo_va;
2079 list_add(&mapping->list, &bo_va->invalids);
2080 amdgpu_vm_it_insert(mapping, &vm->va);
2081
2082 if (mapping->flags & AMDGPU_PTE_PRT)
2083 amdgpu_vm_prt_get(adev);
2084
2085 if (bo && bo->tbo.base.resv == vm->root.base.bo->tbo.base.resv &&
2086 !bo_va->base.moved) {
2087 list_move(&bo_va->base.vm_status, &vm->moved);
2088 }
2089 trace_amdgpu_vm_bo_map(bo_va, mapping);
2090}
2091
2092/**
2093 * amdgpu_vm_bo_map - map bo inside a vm
2094 *
2095 * @adev: amdgpu_device pointer
2096 * @bo_va: bo_va to store the address
2097 * @saddr: where to map the BO
2098 * @offset: requested offset in the BO
2099 * @size: BO size in bytes
2100 * @flags: attributes of pages (read/write/valid/etc.)
2101 *
2102 * Add a mapping of the BO at the specefied addr into the VM.
2103 *
2104 * Returns:
2105 * 0 for success, error for failure.
2106 *
2107 * Object has to be reserved and unreserved outside!
2108 */
2109int amdgpu_vm_bo_map(struct amdgpu_device *adev,
2110 struct amdgpu_bo_va *bo_va,
2111 uint64_t saddr, uint64_t offset,
2112 uint64_t size, uint64_t flags)
2113{
2114 struct amdgpu_bo_va_mapping *mapping, *tmp;
2115 struct amdgpu_bo *bo = bo_va->base.bo;
2116 struct amdgpu_vm *vm = bo_va->base.vm;
2117 uint64_t eaddr;
2118
2119 /* validate the parameters */
2120 if (saddr & AMDGPU_GPU_PAGE_MASK || offset & AMDGPU_GPU_PAGE_MASK ||
2121 size == 0 || size & AMDGPU_GPU_PAGE_MASK)
2122 return -EINVAL;
2123
2124 /* make sure object fit at this offset */
2125 eaddr = saddr + size - 1;
2126 if (saddr >= eaddr ||
2127 (bo && offset + size > amdgpu_bo_size(bo)))
2128 return -EINVAL;
2129
2130 saddr /= AMDGPU_GPU_PAGE_SIZE;
2131 eaddr /= AMDGPU_GPU_PAGE_SIZE;
2132
2133 tmp = amdgpu_vm_it_iter_first(&vm->va, saddr, eaddr);
2134 if (tmp) {
2135 /* bo and tmp overlap, invalid addr */
2136 dev_err(adev->dev, "bo %p va 0x%010Lx-0x%010Lx conflict with "
2137 "0x%010Lx-0x%010Lx\n", bo, saddr, eaddr,
2138 tmp->start, tmp->last + 1);
2139 return -EINVAL;
2140 }
2141
2142 mapping = kmalloc(sizeof(*mapping), GFP_KERNEL);
2143 if (!mapping)
2144 return -ENOMEM;
2145
2146 mapping->start = saddr;
2147 mapping->last = eaddr;
2148 mapping->offset = offset;
2149 mapping->flags = flags;
2150
2151 amdgpu_vm_bo_insert_map(adev, bo_va, mapping);
2152
2153 return 0;
2154}
2155
2156/**
2157 * amdgpu_vm_bo_replace_map - map bo inside a vm, replacing existing mappings
2158 *
2159 * @adev: amdgpu_device pointer
2160 * @bo_va: bo_va to store the address
2161 * @saddr: where to map the BO
2162 * @offset: requested offset in the BO
2163 * @size: BO size in bytes
2164 * @flags: attributes of pages (read/write/valid/etc.)
2165 *
2166 * Add a mapping of the BO at the specefied addr into the VM. Replace existing
2167 * mappings as we do so.
2168 *
2169 * Returns:
2170 * 0 for success, error for failure.
2171 *
2172 * Object has to be reserved and unreserved outside!
2173 */
2174int amdgpu_vm_bo_replace_map(struct amdgpu_device *adev,
2175 struct amdgpu_bo_va *bo_va,
2176 uint64_t saddr, uint64_t offset,
2177 uint64_t size, uint64_t flags)
2178{
2179 struct amdgpu_bo_va_mapping *mapping;
2180 struct amdgpu_bo *bo = bo_va->base.bo;
2181 uint64_t eaddr;
2182 int r;
2183
2184 /* validate the parameters */
2185 if (saddr & AMDGPU_GPU_PAGE_MASK || offset & AMDGPU_GPU_PAGE_MASK ||
2186 size == 0 || size & AMDGPU_GPU_PAGE_MASK)
2187 return -EINVAL;
2188
2189 /* make sure object fit at this offset */
2190 eaddr = saddr + size - 1;
2191 if (saddr >= eaddr ||
2192 (bo && offset + size > amdgpu_bo_size(bo)))
2193 return -EINVAL;
2194
2195 /* Allocate all the needed memory */
2196 mapping = kmalloc(sizeof(*mapping), GFP_KERNEL);
2197 if (!mapping)
2198 return -ENOMEM;
2199
2200 r = amdgpu_vm_bo_clear_mappings(adev, bo_va->base.vm, saddr, size);
2201 if (r) {
2202 kfree(mapping);
2203 return r;
2204 }
2205
2206 saddr /= AMDGPU_GPU_PAGE_SIZE;
2207 eaddr /= AMDGPU_GPU_PAGE_SIZE;
2208
2209 mapping->start = saddr;
2210 mapping->last = eaddr;
2211 mapping->offset = offset;
2212 mapping->flags = flags;
2213
2214 amdgpu_vm_bo_insert_map(adev, bo_va, mapping);
2215
2216 return 0;
2217}
2218
2219/**
2220 * amdgpu_vm_bo_unmap - remove bo mapping from vm
2221 *
2222 * @adev: amdgpu_device pointer
2223 * @bo_va: bo_va to remove the address from
2224 * @saddr: where to the BO is mapped
2225 *
2226 * Remove a mapping of the BO at the specefied addr from the VM.
2227 *
2228 * Returns:
2229 * 0 for success, error for failure.
2230 *
2231 * Object has to be reserved and unreserved outside!
2232 */
2233int amdgpu_vm_bo_unmap(struct amdgpu_device *adev,
2234 struct amdgpu_bo_va *bo_va,
2235 uint64_t saddr)
2236{
2237 struct amdgpu_bo_va_mapping *mapping;
2238 struct amdgpu_vm *vm = bo_va->base.vm;
2239 bool valid = true;
2240
2241 saddr /= AMDGPU_GPU_PAGE_SIZE;
2242
2243 list_for_each_entry(mapping, &bo_va->valids, list) {
2244 if (mapping->start == saddr)
2245 break;
2246 }
2247
2248 if (&mapping->list == &bo_va->valids) {
2249 valid = false;
2250
2251 list_for_each_entry(mapping, &bo_va->invalids, list) {
2252 if (mapping->start == saddr)
2253 break;
2254 }
2255
2256 if (&mapping->list == &bo_va->invalids)
2257 return -ENOENT;
2258 }
2259
2260 list_del(&mapping->list);
2261 amdgpu_vm_it_remove(mapping, &vm->va);
2262 mapping->bo_va = NULL;
2263 trace_amdgpu_vm_bo_unmap(bo_va, mapping);
2264
2265 if (valid)
2266 list_add(&mapping->list, &vm->freed);
2267 else
2268 amdgpu_vm_free_mapping(adev, vm, mapping,
2269 bo_va->last_pt_update);
2270
2271 return 0;
2272}
2273
2274/**
2275 * amdgpu_vm_bo_clear_mappings - remove all mappings in a specific range
2276 *
2277 * @adev: amdgpu_device pointer
2278 * @vm: VM structure to use
2279 * @saddr: start of the range
2280 * @size: size of the range
2281 *
2282 * Remove all mappings in a range, split them as appropriate.
2283 *
2284 * Returns:
2285 * 0 for success, error for failure.
2286 */
2287int amdgpu_vm_bo_clear_mappings(struct amdgpu_device *adev,
2288 struct amdgpu_vm *vm,
2289 uint64_t saddr, uint64_t size)
2290{
2291 struct amdgpu_bo_va_mapping *before, *after, *tmp, *next;
2292 LIST_HEAD(removed);
2293 uint64_t eaddr;
2294
2295 eaddr = saddr + size - 1;
2296 saddr /= AMDGPU_GPU_PAGE_SIZE;
2297 eaddr /= AMDGPU_GPU_PAGE_SIZE;
2298
2299 /* Allocate all the needed memory */
2300 before = kzalloc(sizeof(*before), GFP_KERNEL);
2301 if (!before)
2302 return -ENOMEM;
2303 INIT_LIST_HEAD(&before->list);
2304
2305 after = kzalloc(sizeof(*after), GFP_KERNEL);
2306 if (!after) {
2307 kfree(before);
2308 return -ENOMEM;
2309 }
2310 INIT_LIST_HEAD(&after->list);
2311
2312 /* Now gather all removed mappings */
2313 tmp = amdgpu_vm_it_iter_first(&vm->va, saddr, eaddr);
2314 while (tmp) {
2315 /* Remember mapping split at the start */
2316 if (tmp->start < saddr) {
2317 before->start = tmp->start;
2318 before->last = saddr - 1;
2319 before->offset = tmp->offset;
2320 before->flags = tmp->flags;
2321 before->bo_va = tmp->bo_va;
2322 list_add(&before->list, &tmp->bo_va->invalids);
2323 }
2324
2325 /* Remember mapping split at the end */
2326 if (tmp->last > eaddr) {
2327 after->start = eaddr + 1;
2328 after->last = tmp->last;
2329 after->offset = tmp->offset;
2330 after->offset += after->start - tmp->start;
2331 after->flags = tmp->flags;
2332 after->bo_va = tmp->bo_va;
2333 list_add(&after->list, &tmp->bo_va->invalids);
2334 }
2335
2336 list_del(&tmp->list);
2337 list_add(&tmp->list, &removed);
2338
2339 tmp = amdgpu_vm_it_iter_next(tmp, saddr, eaddr);
2340 }
2341
2342 /* And free them up */
2343 list_for_each_entry_safe(tmp, next, &removed, list) {
2344 amdgpu_vm_it_remove(tmp, &vm->va);
2345 list_del(&tmp->list);
2346
2347 if (tmp->start < saddr)
2348 tmp->start = saddr;
2349 if (tmp->last > eaddr)
2350 tmp->last = eaddr;
2351
2352 tmp->bo_va = NULL;
2353 list_add(&tmp->list, &vm->freed);
2354 trace_amdgpu_vm_bo_unmap(NULL, tmp);
2355 }
2356
2357 /* Insert partial mapping before the range */
2358 if (!list_empty(&before->list)) {
2359 amdgpu_vm_it_insert(before, &vm->va);
2360 if (before->flags & AMDGPU_PTE_PRT)
2361 amdgpu_vm_prt_get(adev);
2362 } else {
2363 kfree(before);
2364 }
2365
2366 /* Insert partial mapping after the range */
2367 if (!list_empty(&after->list)) {
2368 amdgpu_vm_it_insert(after, &vm->va);
2369 if (after->flags & AMDGPU_PTE_PRT)
2370 amdgpu_vm_prt_get(adev);
2371 } else {
2372 kfree(after);
2373 }
2374
2375 return 0;
2376}
2377
2378/**
2379 * amdgpu_vm_bo_lookup_mapping - find mapping by address
2380 *
2381 * @vm: the requested VM
2382 * @addr: the address
2383 *
2384 * Find a mapping by it's address.
2385 *
2386 * Returns:
2387 * The amdgpu_bo_va_mapping matching for addr or NULL
2388 *
2389 */
2390struct amdgpu_bo_va_mapping *amdgpu_vm_bo_lookup_mapping(struct amdgpu_vm *vm,
2391 uint64_t addr)
2392{
2393 return amdgpu_vm_it_iter_first(&vm->va, addr, addr);
2394}
2395
2396/**
2397 * amdgpu_vm_bo_trace_cs - trace all reserved mappings
2398 *
2399 * @vm: the requested vm
2400 * @ticket: CS ticket
2401 *
2402 * Trace all mappings of BOs reserved during a command submission.
2403 */
2404void amdgpu_vm_bo_trace_cs(struct amdgpu_vm *vm, struct ww_acquire_ctx *ticket)
2405{
2406 struct amdgpu_bo_va_mapping *mapping;
2407
2408 if (!trace_amdgpu_vm_bo_cs_enabled())
2409 return;
2410
2411 for (mapping = amdgpu_vm_it_iter_first(&vm->va, 0, U64_MAX); mapping;
2412 mapping = amdgpu_vm_it_iter_next(mapping, 0, U64_MAX)) {
2413 if (mapping->bo_va && mapping->bo_va->base.bo) {
2414 struct amdgpu_bo *bo;
2415
2416 bo = mapping->bo_va->base.bo;
2417 if (dma_resv_locking_ctx(bo->tbo.base.resv) !=
2418 ticket)
2419 continue;
2420 }
2421
2422 trace_amdgpu_vm_bo_cs(mapping);
2423 }
2424}
2425
2426/**
2427 * amdgpu_vm_bo_rmv - remove a bo to a specific vm
2428 *
2429 * @adev: amdgpu_device pointer
2430 * @bo_va: requested bo_va
2431 *
2432 * Remove @bo_va->bo from the requested vm.
2433 *
2434 * Object have to be reserved!
2435 */
2436void amdgpu_vm_bo_rmv(struct amdgpu_device *adev,
2437 struct amdgpu_bo_va *bo_va)
2438{
2439 struct amdgpu_bo_va_mapping *mapping, *next;
2440 struct amdgpu_bo *bo = bo_va->base.bo;
2441 struct amdgpu_vm *vm = bo_va->base.vm;
2442 struct amdgpu_vm_bo_base **base;
2443
2444 if (bo) {
2445 if (bo->tbo.base.resv == vm->root.base.bo->tbo.base.resv)
2446 vm->bulk_moveable = false;
2447
2448 for (base = &bo_va->base.bo->vm_bo; *base;
2449 base = &(*base)->next) {
2450 if (*base != &bo_va->base)
2451 continue;
2452
2453 *base = bo_va->base.next;
2454 break;
2455 }
2456 }
2457
2458 spin_lock(&vm->invalidated_lock);
2459 list_del(&bo_va->base.vm_status);
2460 spin_unlock(&vm->invalidated_lock);
2461
2462 list_for_each_entry_safe(mapping, next, &bo_va->valids, list) {
2463 list_del(&mapping->list);
2464 amdgpu_vm_it_remove(mapping, &vm->va);
2465 mapping->bo_va = NULL;
2466 trace_amdgpu_vm_bo_unmap(bo_va, mapping);
2467 list_add(&mapping->list, &vm->freed);
2468 }
2469 list_for_each_entry_safe(mapping, next, &bo_va->invalids, list) {
2470 list_del(&mapping->list);
2471 amdgpu_vm_it_remove(mapping, &vm->va);
2472 amdgpu_vm_free_mapping(adev, vm, mapping,
2473 bo_va->last_pt_update);
2474 }
2475
2476 dma_fence_put(bo_va->last_pt_update);
2477
2478 if (bo && bo_va->is_xgmi) {
2479 mutex_lock(&adev->vm_manager.lock_pstate);
2480 if (--adev->vm_manager.xgmi_map_counter == 0)
2481 amdgpu_xgmi_set_pstate(adev, 0);
2482 mutex_unlock(&adev->vm_manager.lock_pstate);
2483 }
2484
2485 kfree(bo_va);
2486}
2487
2488/**
2489 * amdgpu_vm_bo_invalidate - mark the bo as invalid
2490 *
2491 * @adev: amdgpu_device pointer
2492 * @bo: amdgpu buffer object
2493 * @evicted: is the BO evicted
2494 *
2495 * Mark @bo as invalid.
2496 */
2497void amdgpu_vm_bo_invalidate(struct amdgpu_device *adev,
2498 struct amdgpu_bo *bo, bool evicted)
2499{
2500 struct amdgpu_vm_bo_base *bo_base;
2501
2502 /* shadow bo doesn't have bo base, its validation needs its parent */
2503 if (bo->parent && bo->parent->shadow == bo)
2504 bo = bo->parent;
2505
2506 for (bo_base = bo->vm_bo; bo_base; bo_base = bo_base->next) {
2507 struct amdgpu_vm *vm = bo_base->vm;
2508
2509 if (evicted && bo->tbo.base.resv == vm->root.base.bo->tbo.base.resv) {
2510 amdgpu_vm_bo_evicted(bo_base);
2511 continue;
2512 }
2513
2514 if (bo_base->moved)
2515 continue;
2516 bo_base->moved = true;
2517
2518 if (bo->tbo.type == ttm_bo_type_kernel)
2519 amdgpu_vm_bo_relocated(bo_base);
2520 else if (bo->tbo.base.resv == vm->root.base.bo->tbo.base.resv)
2521 amdgpu_vm_bo_moved(bo_base);
2522 else
2523 amdgpu_vm_bo_invalidated(bo_base);
2524 }
2525}
2526
2527/**
2528 * amdgpu_vm_get_block_size - calculate VM page table size as power of two
2529 *
2530 * @vm_size: VM size
2531 *
2532 * Returns:
2533 * VM page table as power of two
2534 */
2535static uint32_t amdgpu_vm_get_block_size(uint64_t vm_size)
2536{
2537 /* Total bits covered by PD + PTs */
2538 unsigned bits = ilog2(vm_size) + 18;
2539
2540 /* Make sure the PD is 4K in size up to 8GB address space.
2541 Above that split equal between PD and PTs */
2542 if (vm_size <= 8)
2543 return (bits - 9);
2544 else
2545 return ((bits + 3) / 2);
2546}
2547
2548/**
2549 * amdgpu_vm_adjust_size - adjust vm size, block size and fragment size
2550 *
2551 * @adev: amdgpu_device pointer
2552 * @min_vm_size: the minimum vm size in GB if it's set auto
2553 * @fragment_size_default: Default PTE fragment size
2554 * @max_level: max VMPT level
2555 * @max_bits: max address space size in bits
2556 *
2557 */
2558void amdgpu_vm_adjust_size(struct amdgpu_device *adev, uint32_t min_vm_size,
2559 uint32_t fragment_size_default, unsigned max_level,
2560 unsigned max_bits)
2561{
2562 unsigned int max_size = 1 << (max_bits - 30);
2563 unsigned int vm_size;
2564 uint64_t tmp;
2565
2566 /* adjust vm size first */
2567 if (amdgpu_vm_size != -1) {
2568 vm_size = amdgpu_vm_size;
2569 if (vm_size > max_size) {
2570 dev_warn(adev->dev, "VM size (%d) too large, max is %u GB\n",
2571 amdgpu_vm_size, max_size);
2572 vm_size = max_size;
2573 }
2574 } else {
2575 struct sysinfo si;
2576 unsigned int phys_ram_gb;
2577
2578 /* Optimal VM size depends on the amount of physical
2579 * RAM available. Underlying requirements and
2580 * assumptions:
2581 *
2582 * - Need to map system memory and VRAM from all GPUs
2583 * - VRAM from other GPUs not known here
2584 * - Assume VRAM <= system memory
2585 * - On GFX8 and older, VM space can be segmented for
2586 * different MTYPEs
2587 * - Need to allow room for fragmentation, guard pages etc.
2588 *
2589 * This adds up to a rough guess of system memory x3.
2590 * Round up to power of two to maximize the available
2591 * VM size with the given page table size.
2592 */
2593 si_meminfo(&si);
2594 phys_ram_gb = ((uint64_t)si.totalram * si.mem_unit +
2595 (1 << 30) - 1) >> 30;
2596 vm_size = roundup_pow_of_two(
2597 min(max(phys_ram_gb * 3, min_vm_size), max_size));
2598 }
2599
2600 adev->vm_manager.max_pfn = (uint64_t)vm_size << 18;
2601
2602 tmp = roundup_pow_of_two(adev->vm_manager.max_pfn);
2603 if (amdgpu_vm_block_size != -1)
2604 tmp >>= amdgpu_vm_block_size - 9;
2605 tmp = DIV_ROUND_UP(fls64(tmp) - 1, 9) - 1;
2606 adev->vm_manager.num_level = min(max_level, (unsigned)tmp);
2607 switch (adev->vm_manager.num_level) {
2608 case 3:
2609 adev->vm_manager.root_level = AMDGPU_VM_PDB2;
2610 break;
2611 case 2:
2612 adev->vm_manager.root_level = AMDGPU_VM_PDB1;
2613 break;
2614 case 1:
2615 adev->vm_manager.root_level = AMDGPU_VM_PDB0;
2616 break;
2617 default:
2618 dev_err(adev->dev, "VMPT only supports 2~4+1 levels\n");
2619 }
2620 /* block size depends on vm size and hw setup*/
2621 if (amdgpu_vm_block_size != -1)
2622 adev->vm_manager.block_size =
2623 min((unsigned)amdgpu_vm_block_size, max_bits
2624 - AMDGPU_GPU_PAGE_SHIFT
2625 - 9 * adev->vm_manager.num_level);
2626 else if (adev->vm_manager.num_level > 1)
2627 adev->vm_manager.block_size = 9;
2628 else
2629 adev->vm_manager.block_size = amdgpu_vm_get_block_size(tmp);
2630
2631 if (amdgpu_vm_fragment_size == -1)
2632 adev->vm_manager.fragment_size = fragment_size_default;
2633 else
2634 adev->vm_manager.fragment_size = amdgpu_vm_fragment_size;
2635
2636 DRM_INFO("vm size is %u GB, %u levels, block size is %u-bit, fragment size is %u-bit\n",
2637 vm_size, adev->vm_manager.num_level + 1,
2638 adev->vm_manager.block_size,
2639 adev->vm_manager.fragment_size);
2640}
2641
2642/**
2643 * amdgpu_vm_wait_idle - wait for the VM to become idle
2644 *
2645 * @vm: VM object to wait for
2646 * @timeout: timeout to wait for VM to become idle
2647 */
2648long amdgpu_vm_wait_idle(struct amdgpu_vm *vm, long timeout)
2649{
2650 return dma_resv_wait_timeout_rcu(vm->root.base.bo->tbo.base.resv,
2651 true, true, timeout);
2652}
2653
2654/**
2655 * amdgpu_vm_init - initialize a vm instance
2656 *
2657 * @adev: amdgpu_device pointer
2658 * @vm: requested vm
2659 * @vm_context: Indicates if it GFX or Compute context
2660 * @pasid: Process address space identifier
2661 *
2662 * Init @vm fields.
2663 *
2664 * Returns:
2665 * 0 for success, error for failure.
2666 */
2667int amdgpu_vm_init(struct amdgpu_device *adev, struct amdgpu_vm *vm,
2668 int vm_context, unsigned int pasid)
2669{
2670 struct amdgpu_bo_param bp;
2671 struct amdgpu_bo *root;
2672 int r, i;
2673
2674 vm->va = RB_ROOT_CACHED;
2675 for (i = 0; i < AMDGPU_MAX_VMHUBS; i++)
2676 vm->reserved_vmid[i] = NULL;
2677 INIT_LIST_HEAD(&vm->evicted);
2678 INIT_LIST_HEAD(&vm->relocated);
2679 INIT_LIST_HEAD(&vm->moved);
2680 INIT_LIST_HEAD(&vm->idle);
2681 INIT_LIST_HEAD(&vm->invalidated);
2682 spin_lock_init(&vm->invalidated_lock);
2683 INIT_LIST_HEAD(&vm->freed);
2684
2685 /* create scheduler entity for page table updates */
2686 r = drm_sched_entity_init(&vm->entity, adev->vm_manager.vm_pte_rqs,
2687 adev->vm_manager.vm_pte_num_rqs, NULL);
2688 if (r)
2689 return r;
2690
2691 vm->pte_support_ats = false;
2692
2693 if (vm_context == AMDGPU_VM_CONTEXT_COMPUTE) {
2694 vm->use_cpu_for_update = !!(adev->vm_manager.vm_update_mode &
2695 AMDGPU_VM_USE_CPU_FOR_COMPUTE);
2696
2697 if (adev->asic_type == CHIP_RAVEN)
2698 vm->pte_support_ats = true;
2699 } else {
2700 vm->use_cpu_for_update = !!(adev->vm_manager.vm_update_mode &
2701 AMDGPU_VM_USE_CPU_FOR_GFX);
2702 }
2703 DRM_DEBUG_DRIVER("VM update mode is %s\n",
2704 vm->use_cpu_for_update ? "CPU" : "SDMA");
2705 WARN_ONCE((vm->use_cpu_for_update && !amdgpu_gmc_vram_full_visible(&adev->gmc)),
2706 "CPU update of VM recommended only for large BAR system\n");
2707
2708 if (vm->use_cpu_for_update)
2709 vm->update_funcs = &amdgpu_vm_cpu_funcs;
2710 else
2711 vm->update_funcs = &amdgpu_vm_sdma_funcs;
2712 vm->last_update = NULL;
2713
2714 amdgpu_vm_bo_param(adev, vm, adev->vm_manager.root_level, &bp);
2715 if (vm_context == AMDGPU_VM_CONTEXT_COMPUTE)
2716 bp.flags &= ~AMDGPU_GEM_CREATE_SHADOW;
2717 r = amdgpu_bo_create(adev, &bp, &root);
2718 if (r)
2719 goto error_free_sched_entity;
2720
2721 r = amdgpu_bo_reserve(root, true);
2722 if (r)
2723 goto error_free_root;
2724
2725 r = dma_resv_reserve_shared(root->tbo.base.resv, 1);
2726 if (r)
2727 goto error_unreserve;
2728
2729 amdgpu_vm_bo_base_init(&vm->root.base, vm, root);
2730
2731 r = amdgpu_vm_clear_bo(adev, vm, root);
2732 if (r)
2733 goto error_unreserve;
2734
2735 amdgpu_bo_unreserve(vm->root.base.bo);
2736
2737 if (pasid) {
2738 unsigned long flags;
2739
2740 spin_lock_irqsave(&adev->vm_manager.pasid_lock, flags);
2741 r = idr_alloc(&adev->vm_manager.pasid_idr, vm, pasid, pasid + 1,
2742 GFP_ATOMIC);
2743 spin_unlock_irqrestore(&adev->vm_manager.pasid_lock, flags);
2744 if (r < 0)
2745 goto error_free_root;
2746
2747 vm->pasid = pasid;
2748 }
2749
2750 INIT_KFIFO(vm->faults);
2751
2752 return 0;
2753
2754error_unreserve:
2755 amdgpu_bo_unreserve(vm->root.base.bo);
2756
2757error_free_root:
2758 amdgpu_bo_unref(&vm->root.base.bo->shadow);
2759 amdgpu_bo_unref(&vm->root.base.bo);
2760 vm->root.base.bo = NULL;
2761
2762error_free_sched_entity:
2763 drm_sched_entity_destroy(&vm->entity);
2764
2765 return r;
2766}
2767
2768/**
2769 * amdgpu_vm_check_clean_reserved - check if a VM is clean
2770 *
2771 * @adev: amdgpu_device pointer
2772 * @vm: the VM to check
2773 *
2774 * check all entries of the root PD, if any subsequent PDs are allocated,
2775 * it means there are page table creating and filling, and is no a clean
2776 * VM
2777 *
2778 * Returns:
2779 * 0 if this VM is clean
2780 */
2781static int amdgpu_vm_check_clean_reserved(struct amdgpu_device *adev,
2782 struct amdgpu_vm *vm)
2783{
2784 enum amdgpu_vm_level root = adev->vm_manager.root_level;
2785 unsigned int entries = amdgpu_vm_num_entries(adev, root);
2786 unsigned int i = 0;
2787
2788 if (!(vm->root.entries))
2789 return 0;
2790
2791 for (i = 0; i < entries; i++) {
2792 if (vm->root.entries[i].base.bo)
2793 return -EINVAL;
2794 }
2795
2796 return 0;
2797}
2798
2799/**
2800 * amdgpu_vm_make_compute - Turn a GFX VM into a compute VM
2801 *
2802 * @adev: amdgpu_device pointer
2803 * @vm: requested vm
2804 *
2805 * This only works on GFX VMs that don't have any BOs added and no
2806 * page tables allocated yet.
2807 *
2808 * Changes the following VM parameters:
2809 * - use_cpu_for_update
2810 * - pte_supports_ats
2811 * - pasid (old PASID is released, because compute manages its own PASIDs)
2812 *
2813 * Reinitializes the page directory to reflect the changed ATS
2814 * setting.
2815 *
2816 * Returns:
2817 * 0 for success, -errno for errors.
2818 */
2819int amdgpu_vm_make_compute(struct amdgpu_device *adev, struct amdgpu_vm *vm, unsigned int pasid)
2820{
2821 bool pte_support_ats = (adev->asic_type == CHIP_RAVEN);
2822 int r;
2823
2824 r = amdgpu_bo_reserve(vm->root.base.bo, true);
2825 if (r)
2826 return r;
2827
2828 /* Sanity checks */
2829 r = amdgpu_vm_check_clean_reserved(adev, vm);
2830 if (r)
2831 goto unreserve_bo;
2832
2833 if (pasid) {
2834 unsigned long flags;
2835
2836 spin_lock_irqsave(&adev->vm_manager.pasid_lock, flags);
2837 r = idr_alloc(&adev->vm_manager.pasid_idr, vm, pasid, pasid + 1,
2838 GFP_ATOMIC);
2839 spin_unlock_irqrestore(&adev->vm_manager.pasid_lock, flags);
2840
2841 if (r == -ENOSPC)
2842 goto unreserve_bo;
2843 r = 0;
2844 }
2845
2846 /* Check if PD needs to be reinitialized and do it before
2847 * changing any other state, in case it fails.
2848 */
2849 if (pte_support_ats != vm->pte_support_ats) {
2850 vm->pte_support_ats = pte_support_ats;
2851 r = amdgpu_vm_clear_bo(adev, vm, vm->root.base.bo);
2852 if (r)
2853 goto free_idr;
2854 }
2855
2856 /* Update VM state */
2857 vm->use_cpu_for_update = !!(adev->vm_manager.vm_update_mode &
2858 AMDGPU_VM_USE_CPU_FOR_COMPUTE);
2859 DRM_DEBUG_DRIVER("VM update mode is %s\n",
2860 vm->use_cpu_for_update ? "CPU" : "SDMA");
2861 WARN_ONCE((vm->use_cpu_for_update && !amdgpu_gmc_vram_full_visible(&adev->gmc)),
2862 "CPU update of VM recommended only for large BAR system\n");
2863
2864 if (vm->use_cpu_for_update)
2865 vm->update_funcs = &amdgpu_vm_cpu_funcs;
2866 else
2867 vm->update_funcs = &amdgpu_vm_sdma_funcs;
2868 dma_fence_put(vm->last_update);
2869 vm->last_update = NULL;
2870
2871 if (vm->pasid) {
2872 unsigned long flags;
2873
2874 spin_lock_irqsave(&adev->vm_manager.pasid_lock, flags);
2875 idr_remove(&adev->vm_manager.pasid_idr, vm->pasid);
2876 spin_unlock_irqrestore(&adev->vm_manager.pasid_lock, flags);
2877
2878 /* Free the original amdgpu allocated pasid
2879 * Will be replaced with kfd allocated pasid
2880 */
2881 amdgpu_pasid_free(vm->pasid);
2882 vm->pasid = 0;
2883 }
2884
2885 /* Free the shadow bo for compute VM */
2886 amdgpu_bo_unref(&vm->root.base.bo->shadow);
2887
2888 if (pasid)
2889 vm->pasid = pasid;
2890
2891 goto unreserve_bo;
2892
2893free_idr:
2894 if (pasid) {
2895 unsigned long flags;
2896
2897 spin_lock_irqsave(&adev->vm_manager.pasid_lock, flags);
2898 idr_remove(&adev->vm_manager.pasid_idr, pasid);
2899 spin_unlock_irqrestore(&adev->vm_manager.pasid_lock, flags);
2900 }
2901unreserve_bo:
2902 amdgpu_bo_unreserve(vm->root.base.bo);
2903 return r;
2904}
2905
2906/**
2907 * amdgpu_vm_release_compute - release a compute vm
2908 * @adev: amdgpu_device pointer
2909 * @vm: a vm turned into compute vm by calling amdgpu_vm_make_compute
2910 *
2911 * This is a correspondant of amdgpu_vm_make_compute. It decouples compute
2912 * pasid from vm. Compute should stop use of vm after this call.
2913 */
2914void amdgpu_vm_release_compute(struct amdgpu_device *adev, struct amdgpu_vm *vm)
2915{
2916 if (vm->pasid) {
2917 unsigned long flags;
2918
2919 spin_lock_irqsave(&adev->vm_manager.pasid_lock, flags);
2920 idr_remove(&adev->vm_manager.pasid_idr, vm->pasid);
2921 spin_unlock_irqrestore(&adev->vm_manager.pasid_lock, flags);
2922 }
2923 vm->pasid = 0;
2924}
2925
2926/**
2927 * amdgpu_vm_fini - tear down a vm instance
2928 *
2929 * @adev: amdgpu_device pointer
2930 * @vm: requested vm
2931 *
2932 * Tear down @vm.
2933 * Unbind the VM and remove all bos from the vm bo list
2934 */
2935void amdgpu_vm_fini(struct amdgpu_device *adev, struct amdgpu_vm *vm)
2936{
2937 struct amdgpu_bo_va_mapping *mapping, *tmp;
2938 bool prt_fini_needed = !!adev->gmc.gmc_funcs->set_prt;
2939 struct amdgpu_bo *root;
2940 int i, r;
2941
2942 amdgpu_amdkfd_gpuvm_destroy_cb(adev, vm);
2943
2944 if (vm->pasid) {
2945 unsigned long flags;
2946
2947 spin_lock_irqsave(&adev->vm_manager.pasid_lock, flags);
2948 idr_remove(&adev->vm_manager.pasid_idr, vm->pasid);
2949 spin_unlock_irqrestore(&adev->vm_manager.pasid_lock, flags);
2950 }
2951
2952 drm_sched_entity_destroy(&vm->entity);
2953
2954 if (!RB_EMPTY_ROOT(&vm->va.rb_root)) {
2955 dev_err(adev->dev, "still active bo inside vm\n");
2956 }
2957 rbtree_postorder_for_each_entry_safe(mapping, tmp,
2958 &vm->va.rb_root, rb) {
2959 /* Don't remove the mapping here, we don't want to trigger a
2960 * rebalance and the tree is about to be destroyed anyway.
2961 */
2962 list_del(&mapping->list);
2963 kfree(mapping);
2964 }
2965 list_for_each_entry_safe(mapping, tmp, &vm->freed, list) {
2966 if (mapping->flags & AMDGPU_PTE_PRT && prt_fini_needed) {
2967 amdgpu_vm_prt_fini(adev, vm);
2968 prt_fini_needed = false;
2969 }
2970
2971 list_del(&mapping->list);
2972 amdgpu_vm_free_mapping(adev, vm, mapping, NULL);
2973 }
2974
2975 root = amdgpu_bo_ref(vm->root.base.bo);
2976 r = amdgpu_bo_reserve(root, true);
2977 if (r) {
2978 dev_err(adev->dev, "Leaking page tables because BO reservation failed\n");
2979 } else {
2980 amdgpu_vm_free_pts(adev, vm, NULL);
2981 amdgpu_bo_unreserve(root);
2982 }
2983 amdgpu_bo_unref(&root);
2984 WARN_ON(vm->root.base.bo);
2985 dma_fence_put(vm->last_update);
2986 for (i = 0; i < AMDGPU_MAX_VMHUBS; i++)
2987 amdgpu_vmid_free_reserved(adev, vm, i);
2988}
2989
2990/**
2991 * amdgpu_vm_manager_init - init the VM manager
2992 *
2993 * @adev: amdgpu_device pointer
2994 *
2995 * Initialize the VM manager structures
2996 */
2997void amdgpu_vm_manager_init(struct amdgpu_device *adev)
2998{
2999 unsigned i;
3000
3001 amdgpu_vmid_mgr_init(adev);
3002
3003 adev->vm_manager.fence_context =
3004 dma_fence_context_alloc(AMDGPU_MAX_RINGS);
3005 for (i = 0; i < AMDGPU_MAX_RINGS; ++i)
3006 adev->vm_manager.seqno[i] = 0;
3007
3008 spin_lock_init(&adev->vm_manager.prt_lock);
3009 atomic_set(&adev->vm_manager.num_prt_users, 0);
3010
3011 /* If not overridden by the user, by default, only in large BAR systems
3012 * Compute VM tables will be updated by CPU
3013 */
3014#ifdef CONFIG_X86_64
3015 if (amdgpu_vm_update_mode == -1) {
3016 if (amdgpu_gmc_vram_full_visible(&adev->gmc))
3017 adev->vm_manager.vm_update_mode =
3018 AMDGPU_VM_USE_CPU_FOR_COMPUTE;
3019 else
3020 adev->vm_manager.vm_update_mode = 0;
3021 } else
3022 adev->vm_manager.vm_update_mode = amdgpu_vm_update_mode;
3023#else
3024 adev->vm_manager.vm_update_mode = 0;
3025#endif
3026
3027 idr_init(&adev->vm_manager.pasid_idr);
3028 spin_lock_init(&adev->vm_manager.pasid_lock);
3029
3030 adev->vm_manager.xgmi_map_counter = 0;
3031 mutex_init(&adev->vm_manager.lock_pstate);
3032}
3033
3034/**
3035 * amdgpu_vm_manager_fini - cleanup VM manager
3036 *
3037 * @adev: amdgpu_device pointer
3038 *
3039 * Cleanup the VM manager and free resources.
3040 */
3041void amdgpu_vm_manager_fini(struct amdgpu_device *adev)
3042{
3043 WARN_ON(!idr_is_empty(&adev->vm_manager.pasid_idr));
3044 idr_destroy(&adev->vm_manager.pasid_idr);
3045
3046 amdgpu_vmid_mgr_fini(adev);
3047}
3048
3049/**
3050 * amdgpu_vm_ioctl - Manages VMID reservation for vm hubs.
3051 *
3052 * @dev: drm device pointer
3053 * @data: drm_amdgpu_vm
3054 * @filp: drm file pointer
3055 *
3056 * Returns:
3057 * 0 for success, -errno for errors.
3058 */
3059int amdgpu_vm_ioctl(struct drm_device *dev, void *data, struct drm_file *filp)
3060{
3061 union drm_amdgpu_vm *args = data;
3062 struct amdgpu_device *adev = dev->dev_private;
3063 struct amdgpu_fpriv *fpriv = filp->driver_priv;
3064 int r;
3065
3066 switch (args->in.op) {
3067 case AMDGPU_VM_OP_RESERVE_VMID:
3068 /* current, we only have requirement to reserve vmid from gfxhub */
3069 r = amdgpu_vmid_alloc_reserved(adev, &fpriv->vm, AMDGPU_GFXHUB_0);
3070 if (r)
3071 return r;
3072 break;
3073 case AMDGPU_VM_OP_UNRESERVE_VMID:
3074 amdgpu_vmid_free_reserved(adev, &fpriv->vm, AMDGPU_GFXHUB_0);
3075 break;
3076 default:
3077 return -EINVAL;
3078 }
3079
3080 return 0;
3081}
3082
3083/**
3084 * amdgpu_vm_get_task_info - Extracts task info for a PASID.
3085 *
3086 * @adev: drm device pointer
3087 * @pasid: PASID identifier for VM
3088 * @task_info: task_info to fill.
3089 */
3090void amdgpu_vm_get_task_info(struct amdgpu_device *adev, unsigned int pasid,
3091 struct amdgpu_task_info *task_info)
3092{
3093 struct amdgpu_vm *vm;
3094 unsigned long flags;
3095
3096 spin_lock_irqsave(&adev->vm_manager.pasid_lock, flags);
3097
3098 vm = idr_find(&adev->vm_manager.pasid_idr, pasid);
3099 if (vm)
3100 *task_info = vm->task_info;
3101
3102 spin_unlock_irqrestore(&adev->vm_manager.pasid_lock, flags);
3103}
3104
3105/**
3106 * amdgpu_vm_set_task_info - Sets VMs task info.
3107 *
3108 * @vm: vm for which to set the info
3109 */
3110void amdgpu_vm_set_task_info(struct amdgpu_vm *vm)
3111{
3112 if (!vm->task_info.pid) {
3113 vm->task_info.pid = current->pid;
3114 get_task_comm(vm->task_info.task_name, current);
3115
3116 if (current->group_leader->mm == current->mm) {
3117 vm->task_info.tgid = current->group_leader->pid;
3118 get_task_comm(vm->task_info.process_name, current->group_leader);
3119 }
3120 }
3121}