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 (!amdgpu_vm_is_bo_always_valid(vm, bo))
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 += lower_32_bits(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 uint64_t new_vm_generation = amdgpu_vm_generation(adev, vm);
467 struct amdgpu_vm_bo_base *bo_base;
468 struct amdgpu_bo *bo;
469 int r;
470
471 if (vm->generation != new_vm_generation) {
472 vm->generation = new_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
489 r = validate(param, bo);
490 if (r)
491 return r;
492
493 if (bo->tbo.type != ttm_bo_type_kernel) {
494 amdgpu_vm_bo_moved(bo_base);
495 } else {
496 vm->update_funcs->map_table(to_amdgpu_bo_vm(bo));
497 amdgpu_vm_bo_relocated(bo_base);
498 }
499 spin_lock(&vm->status_lock);
500 }
501 while (ticket && !list_empty(&vm->evicted_user)) {
502 bo_base = list_first_entry(&vm->evicted_user,
503 struct amdgpu_vm_bo_base,
504 vm_status);
505 spin_unlock(&vm->status_lock);
506
507 bo = bo_base->bo;
508
509 if (dma_resv_locking_ctx(bo->tbo.base.resv) != ticket) {
510 struct amdgpu_task_info *ti = amdgpu_vm_get_task_info_vm(vm);
511
512 pr_warn_ratelimited("Evicted user BO is not reserved\n");
513 if (ti) {
514 pr_warn_ratelimited("pid %d\n", ti->pid);
515 amdgpu_vm_put_task_info(ti);
516 }
517
518 return -EINVAL;
519 }
520
521 r = validate(param, bo);
522 if (r)
523 return r;
524
525 amdgpu_vm_bo_invalidated(bo_base);
526
527 spin_lock(&vm->status_lock);
528 }
529 spin_unlock(&vm->status_lock);
530
531 amdgpu_vm_eviction_lock(vm);
532 vm->evicting = false;
533 amdgpu_vm_eviction_unlock(vm);
534
535 return 0;
536}
537
538/**
539 * amdgpu_vm_ready - check VM is ready for updates
540 *
541 * @vm: VM to check
542 *
543 * Check if all VM PDs/PTs are ready for updates
544 *
545 * Returns:
546 * True if VM is not evicting.
547 */
548bool amdgpu_vm_ready(struct amdgpu_vm *vm)
549{
550 bool empty;
551 bool ret;
552
553 amdgpu_vm_eviction_lock(vm);
554 ret = !vm->evicting;
555 amdgpu_vm_eviction_unlock(vm);
556
557 spin_lock(&vm->status_lock);
558 empty = list_empty(&vm->evicted);
559 spin_unlock(&vm->status_lock);
560
561 return ret && empty;
562}
563
564/**
565 * amdgpu_vm_check_compute_bug - check whether asic has compute vm bug
566 *
567 * @adev: amdgpu_device pointer
568 */
569void amdgpu_vm_check_compute_bug(struct amdgpu_device *adev)
570{
571 const struct amdgpu_ip_block *ip_block;
572 bool has_compute_vm_bug;
573 struct amdgpu_ring *ring;
574 int i;
575
576 has_compute_vm_bug = false;
577
578 ip_block = amdgpu_device_ip_get_ip_block(adev, AMD_IP_BLOCK_TYPE_GFX);
579 if (ip_block) {
580 /* Compute has a VM bug for GFX version < 7.
581 Compute has a VM bug for GFX 8 MEC firmware version < 673.*/
582 if (ip_block->version->major <= 7)
583 has_compute_vm_bug = true;
584 else if (ip_block->version->major == 8)
585 if (adev->gfx.mec_fw_version < 673)
586 has_compute_vm_bug = true;
587 }
588
589 for (i = 0; i < adev->num_rings; i++) {
590 ring = adev->rings[i];
591 if (ring->funcs->type == AMDGPU_RING_TYPE_COMPUTE)
592 /* only compute rings */
593 ring->has_compute_vm_bug = has_compute_vm_bug;
594 else
595 ring->has_compute_vm_bug = false;
596 }
597}
598
599/**
600 * amdgpu_vm_need_pipeline_sync - Check if pipe sync is needed for job.
601 *
602 * @ring: ring on which the job will be submitted
603 * @job: job to submit
604 *
605 * Returns:
606 * True if sync is needed.
607 */
608bool amdgpu_vm_need_pipeline_sync(struct amdgpu_ring *ring,
609 struct amdgpu_job *job)
610{
611 struct amdgpu_device *adev = ring->adev;
612 unsigned vmhub = ring->vm_hub;
613 struct amdgpu_vmid_mgr *id_mgr = &adev->vm_manager.id_mgr[vmhub];
614
615 if (job->vmid == 0)
616 return false;
617
618 if (job->vm_needs_flush || ring->has_compute_vm_bug)
619 return true;
620
621 if (ring->funcs->emit_gds_switch && job->gds_switch_needed)
622 return true;
623
624 if (amdgpu_vmid_had_gpu_reset(adev, &id_mgr->ids[job->vmid]))
625 return true;
626
627 return false;
628}
629
630/**
631 * amdgpu_vm_flush - hardware flush the vm
632 *
633 * @ring: ring to use for flush
634 * @job: related job
635 * @need_pipe_sync: is pipe sync needed
636 *
637 * Emit a VM flush when it is necessary.
638 *
639 * Returns:
640 * 0 on success, errno otherwise.
641 */
642int amdgpu_vm_flush(struct amdgpu_ring *ring, struct amdgpu_job *job,
643 bool need_pipe_sync)
644{
645 struct amdgpu_device *adev = ring->adev;
646 unsigned vmhub = ring->vm_hub;
647 struct amdgpu_vmid_mgr *id_mgr = &adev->vm_manager.id_mgr[vmhub];
648 struct amdgpu_vmid *id = &id_mgr->ids[job->vmid];
649 bool spm_update_needed = job->spm_update_needed;
650 bool gds_switch_needed = ring->funcs->emit_gds_switch &&
651 job->gds_switch_needed;
652 bool vm_flush_needed = job->vm_needs_flush;
653 struct dma_fence *fence = NULL;
654 bool pasid_mapping_needed = false;
655 unsigned int patch;
656 int r;
657
658 if (amdgpu_vmid_had_gpu_reset(adev, id)) {
659 gds_switch_needed = true;
660 vm_flush_needed = true;
661 pasid_mapping_needed = true;
662 spm_update_needed = true;
663 }
664
665 mutex_lock(&id_mgr->lock);
666 if (id->pasid != job->pasid || !id->pasid_mapping ||
667 !dma_fence_is_signaled(id->pasid_mapping))
668 pasid_mapping_needed = true;
669 mutex_unlock(&id_mgr->lock);
670
671 gds_switch_needed &= !!ring->funcs->emit_gds_switch;
672 vm_flush_needed &= !!ring->funcs->emit_vm_flush &&
673 job->vm_pd_addr != AMDGPU_BO_INVALID_OFFSET;
674 pasid_mapping_needed &= adev->gmc.gmc_funcs->emit_pasid_mapping &&
675 ring->funcs->emit_wreg;
676
677 if (!vm_flush_needed && !gds_switch_needed && !need_pipe_sync &&
678 !(job->enforce_isolation && !job->vmid))
679 return 0;
680
681 amdgpu_ring_ib_begin(ring);
682 if (ring->funcs->init_cond_exec)
683 patch = amdgpu_ring_init_cond_exec(ring,
684 ring->cond_exe_gpu_addr);
685
686 if (need_pipe_sync)
687 amdgpu_ring_emit_pipeline_sync(ring);
688
689 if (adev->gfx.enable_cleaner_shader &&
690 ring->funcs->emit_cleaner_shader &&
691 job->enforce_isolation)
692 ring->funcs->emit_cleaner_shader(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
745 amdgpu_ring_ib_end(ring);
746 return 0;
747}
748
749/**
750 * amdgpu_vm_bo_find - find the bo_va for a specific vm & bo
751 *
752 * @vm: requested vm
753 * @bo: requested buffer object
754 *
755 * Find @bo inside the requested vm.
756 * Search inside the @bos vm list for the requested vm
757 * Returns the found bo_va or NULL if none is found
758 *
759 * Object has to be reserved!
760 *
761 * Returns:
762 * Found bo_va or NULL.
763 */
764struct amdgpu_bo_va *amdgpu_vm_bo_find(struct amdgpu_vm *vm,
765 struct amdgpu_bo *bo)
766{
767 struct amdgpu_vm_bo_base *base;
768
769 for (base = bo->vm_bo; base; base = base->next) {
770 if (base->vm != vm)
771 continue;
772
773 return container_of(base, struct amdgpu_bo_va, base);
774 }
775 return NULL;
776}
777
778/**
779 * amdgpu_vm_map_gart - Resolve gart mapping of addr
780 *
781 * @pages_addr: optional DMA address to use for lookup
782 * @addr: the unmapped addr
783 *
784 * Look up the physical address of the page that the pte resolves
785 * to.
786 *
787 * Returns:
788 * The pointer for the page table entry.
789 */
790uint64_t amdgpu_vm_map_gart(const dma_addr_t *pages_addr, uint64_t addr)
791{
792 uint64_t result;
793
794 /* page table offset */
795 result = pages_addr[addr >> PAGE_SHIFT];
796
797 /* in case cpu page size != gpu page size*/
798 result |= addr & (~PAGE_MASK);
799
800 result &= 0xFFFFFFFFFFFFF000ULL;
801
802 return result;
803}
804
805/**
806 * amdgpu_vm_update_pdes - make sure that all directories are valid
807 *
808 * @adev: amdgpu_device pointer
809 * @vm: requested vm
810 * @immediate: submit immediately to the paging queue
811 *
812 * Makes sure all directories are up to date.
813 *
814 * Returns:
815 * 0 for success, error for failure.
816 */
817int amdgpu_vm_update_pdes(struct amdgpu_device *adev,
818 struct amdgpu_vm *vm, bool immediate)
819{
820 struct amdgpu_vm_update_params params;
821 struct amdgpu_vm_bo_base *entry;
822 bool flush_tlb_needed = false;
823 LIST_HEAD(relocated);
824 int r, idx;
825
826 spin_lock(&vm->status_lock);
827 list_splice_init(&vm->relocated, &relocated);
828 spin_unlock(&vm->status_lock);
829
830 if (list_empty(&relocated))
831 return 0;
832
833 if (!drm_dev_enter(adev_to_drm(adev), &idx))
834 return -ENODEV;
835
836 memset(¶ms, 0, sizeof(params));
837 params.adev = adev;
838 params.vm = vm;
839 params.immediate = immediate;
840
841 r = vm->update_funcs->prepare(¶ms, NULL);
842 if (r)
843 goto error;
844
845 list_for_each_entry(entry, &relocated, vm_status) {
846 /* vm_flush_needed after updating moved PDEs */
847 flush_tlb_needed |= entry->moved;
848
849 r = amdgpu_vm_pde_update(¶ms, entry);
850 if (r)
851 goto error;
852 }
853
854 r = vm->update_funcs->commit(¶ms, &vm->last_update);
855 if (r)
856 goto error;
857
858 if (flush_tlb_needed)
859 atomic64_inc(&vm->tlb_seq);
860
861 while (!list_empty(&relocated)) {
862 entry = list_first_entry(&relocated, struct amdgpu_vm_bo_base,
863 vm_status);
864 amdgpu_vm_bo_idle(entry);
865 }
866
867error:
868 drm_dev_exit(idx);
869 return r;
870}
871
872/**
873 * amdgpu_vm_tlb_seq_cb - make sure to increment tlb sequence
874 * @fence: unused
875 * @cb: the callback structure
876 *
877 * Increments the tlb sequence to make sure that future CS execute a VM flush.
878 */
879static void amdgpu_vm_tlb_seq_cb(struct dma_fence *fence,
880 struct dma_fence_cb *cb)
881{
882 struct amdgpu_vm_tlb_seq_struct *tlb_cb;
883
884 tlb_cb = container_of(cb, typeof(*tlb_cb), cb);
885 atomic64_inc(&tlb_cb->vm->tlb_seq);
886 kfree(tlb_cb);
887}
888
889/**
890 * amdgpu_vm_tlb_flush - prepare TLB flush
891 *
892 * @params: parameters for update
893 * @fence: input fence to sync TLB flush with
894 * @tlb_cb: the callback structure
895 *
896 * Increments the tlb sequence to make sure that future CS execute a VM flush.
897 */
898static void
899amdgpu_vm_tlb_flush(struct amdgpu_vm_update_params *params,
900 struct dma_fence **fence,
901 struct amdgpu_vm_tlb_seq_struct *tlb_cb)
902{
903 struct amdgpu_vm *vm = params->vm;
904
905 tlb_cb->vm = vm;
906 if (!fence || !*fence) {
907 amdgpu_vm_tlb_seq_cb(NULL, &tlb_cb->cb);
908 return;
909 }
910
911 if (!dma_fence_add_callback(*fence, &tlb_cb->cb,
912 amdgpu_vm_tlb_seq_cb)) {
913 dma_fence_put(vm->last_tlb_flush);
914 vm->last_tlb_flush = dma_fence_get(*fence);
915 } else {
916 amdgpu_vm_tlb_seq_cb(NULL, &tlb_cb->cb);
917 }
918
919 /* Prepare a TLB flush fence to be attached to PTs */
920 if (!params->unlocked && vm->is_compute_context) {
921 amdgpu_vm_tlb_fence_create(params->adev, vm, fence);
922
923 /* Makes sure no PD/PT is freed before the flush */
924 dma_resv_add_fence(vm->root.bo->tbo.base.resv, *fence,
925 DMA_RESV_USAGE_BOOKKEEP);
926 }
927}
928
929/**
930 * amdgpu_vm_update_range - update a range in the vm page table
931 *
932 * @adev: amdgpu_device pointer to use for commands
933 * @vm: the VM to update the range
934 * @immediate: immediate submission in a page fault
935 * @unlocked: unlocked invalidation during MM callback
936 * @flush_tlb: trigger tlb invalidation after update completed
937 * @allow_override: change MTYPE for local NUMA nodes
938 * @sync: fences we need to sync to
939 * @start: start of mapped range
940 * @last: last mapped entry
941 * @flags: flags for the entries
942 * @offset: offset into nodes and pages_addr
943 * @vram_base: base for vram mappings
944 * @res: ttm_resource to map
945 * @pages_addr: DMA addresses to use for mapping
946 * @fence: optional resulting fence
947 *
948 * Fill in the page table entries between @start and @last.
949 *
950 * Returns:
951 * 0 for success, negative erro code for failure.
952 */
953int amdgpu_vm_update_range(struct amdgpu_device *adev, struct amdgpu_vm *vm,
954 bool immediate, bool unlocked, bool flush_tlb,
955 bool allow_override, struct amdgpu_sync *sync,
956 uint64_t start, uint64_t last, uint64_t flags,
957 uint64_t offset, uint64_t vram_base,
958 struct ttm_resource *res, dma_addr_t *pages_addr,
959 struct dma_fence **fence)
960{
961 struct amdgpu_vm_tlb_seq_struct *tlb_cb;
962 struct amdgpu_vm_update_params params;
963 struct amdgpu_res_cursor cursor;
964 int r, idx;
965
966 if (!drm_dev_enter(adev_to_drm(adev), &idx))
967 return -ENODEV;
968
969 tlb_cb = kmalloc(sizeof(*tlb_cb), GFP_KERNEL);
970 if (!tlb_cb) {
971 drm_dev_exit(idx);
972 return -ENOMEM;
973 }
974
975 /* Vega20+XGMI where PTEs get inadvertently cached in L2 texture cache,
976 * heavy-weight flush TLB unconditionally.
977 */
978 flush_tlb |= adev->gmc.xgmi.num_physical_nodes &&
979 amdgpu_ip_version(adev, GC_HWIP, 0) == IP_VERSION(9, 4, 0);
980
981 /*
982 * On GFX8 and older any 8 PTE block with a valid bit set enters the TLB
983 */
984 flush_tlb |= amdgpu_ip_version(adev, GC_HWIP, 0) < IP_VERSION(9, 0, 0);
985
986 memset(¶ms, 0, sizeof(params));
987 params.adev = adev;
988 params.vm = vm;
989 params.immediate = immediate;
990 params.pages_addr = pages_addr;
991 params.unlocked = unlocked;
992 params.needs_flush = flush_tlb;
993 params.allow_override = allow_override;
994 INIT_LIST_HEAD(¶ms.tlb_flush_waitlist);
995
996 amdgpu_vm_eviction_lock(vm);
997 if (vm->evicting) {
998 r = -EBUSY;
999 goto error_free;
1000 }
1001
1002 if (!unlocked && !dma_fence_is_signaled(vm->last_unlocked)) {
1003 struct dma_fence *tmp = dma_fence_get_stub();
1004
1005 amdgpu_bo_fence(vm->root.bo, vm->last_unlocked, true);
1006 swap(vm->last_unlocked, tmp);
1007 dma_fence_put(tmp);
1008 }
1009
1010 r = vm->update_funcs->prepare(¶ms, sync);
1011 if (r)
1012 goto error_free;
1013
1014 amdgpu_res_first(pages_addr ? NULL : res, offset,
1015 (last - start + 1) * AMDGPU_GPU_PAGE_SIZE, &cursor);
1016 while (cursor.remaining) {
1017 uint64_t tmp, num_entries, addr;
1018
1019 num_entries = cursor.size >> AMDGPU_GPU_PAGE_SHIFT;
1020 if (pages_addr) {
1021 bool contiguous = true;
1022
1023 if (num_entries > AMDGPU_GPU_PAGES_IN_CPU_PAGE) {
1024 uint64_t pfn = cursor.start >> PAGE_SHIFT;
1025 uint64_t count;
1026
1027 contiguous = pages_addr[pfn + 1] ==
1028 pages_addr[pfn] + PAGE_SIZE;
1029
1030 tmp = num_entries /
1031 AMDGPU_GPU_PAGES_IN_CPU_PAGE;
1032 for (count = 2; count < tmp; ++count) {
1033 uint64_t idx = pfn + count;
1034
1035 if (contiguous != (pages_addr[idx] ==
1036 pages_addr[idx - 1] + PAGE_SIZE))
1037 break;
1038 }
1039 if (!contiguous)
1040 count--;
1041 num_entries = count *
1042 AMDGPU_GPU_PAGES_IN_CPU_PAGE;
1043 }
1044
1045 if (!contiguous) {
1046 addr = cursor.start;
1047 params.pages_addr = pages_addr;
1048 } else {
1049 addr = pages_addr[cursor.start >> PAGE_SHIFT];
1050 params.pages_addr = NULL;
1051 }
1052
1053 } else if (flags & (AMDGPU_PTE_VALID | AMDGPU_PTE_PRT_FLAG(adev))) {
1054 addr = vram_base + cursor.start;
1055 } else {
1056 addr = 0;
1057 }
1058
1059 tmp = start + num_entries;
1060 r = amdgpu_vm_ptes_update(¶ms, start, tmp, addr, flags);
1061 if (r)
1062 goto error_free;
1063
1064 amdgpu_res_next(&cursor, num_entries * AMDGPU_GPU_PAGE_SIZE);
1065 start = tmp;
1066 }
1067
1068 r = vm->update_funcs->commit(¶ms, fence);
1069 if (r)
1070 goto error_free;
1071
1072 if (params.needs_flush) {
1073 amdgpu_vm_tlb_flush(¶ms, fence, tlb_cb);
1074 tlb_cb = NULL;
1075 }
1076
1077 amdgpu_vm_pt_free_list(adev, ¶ms);
1078
1079error_free:
1080 kfree(tlb_cb);
1081 amdgpu_vm_eviction_unlock(vm);
1082 drm_dev_exit(idx);
1083 return r;
1084}
1085
1086static void amdgpu_vm_bo_get_memory(struct amdgpu_bo_va *bo_va,
1087 struct amdgpu_mem_stats *stats,
1088 unsigned int size)
1089{
1090 struct amdgpu_vm *vm = bo_va->base.vm;
1091 struct amdgpu_bo *bo = bo_va->base.bo;
1092
1093 if (!bo)
1094 return;
1095
1096 /*
1097 * For now ignore BOs which are currently locked and potentially
1098 * changing their location.
1099 */
1100 if (!amdgpu_vm_is_bo_always_valid(vm, bo) &&
1101 !dma_resv_trylock(bo->tbo.base.resv))
1102 return;
1103
1104 amdgpu_bo_get_memory(bo, stats, size);
1105 if (!amdgpu_vm_is_bo_always_valid(vm, bo))
1106 dma_resv_unlock(bo->tbo.base.resv);
1107}
1108
1109void amdgpu_vm_get_memory(struct amdgpu_vm *vm,
1110 struct amdgpu_mem_stats *stats,
1111 unsigned int size)
1112{
1113 struct amdgpu_bo_va *bo_va, *tmp;
1114
1115 spin_lock(&vm->status_lock);
1116 list_for_each_entry_safe(bo_va, tmp, &vm->idle, base.vm_status)
1117 amdgpu_vm_bo_get_memory(bo_va, stats, size);
1118
1119 list_for_each_entry_safe(bo_va, tmp, &vm->evicted, base.vm_status)
1120 amdgpu_vm_bo_get_memory(bo_va, stats, size);
1121
1122 list_for_each_entry_safe(bo_va, tmp, &vm->relocated, base.vm_status)
1123 amdgpu_vm_bo_get_memory(bo_va, stats, size);
1124
1125 list_for_each_entry_safe(bo_va, tmp, &vm->moved, base.vm_status)
1126 amdgpu_vm_bo_get_memory(bo_va, stats, size);
1127
1128 list_for_each_entry_safe(bo_va, tmp, &vm->invalidated, base.vm_status)
1129 amdgpu_vm_bo_get_memory(bo_va, stats, size);
1130
1131 list_for_each_entry_safe(bo_va, tmp, &vm->done, base.vm_status)
1132 amdgpu_vm_bo_get_memory(bo_va, stats, size);
1133 spin_unlock(&vm->status_lock);
1134}
1135
1136/**
1137 * amdgpu_vm_bo_update - update all BO mappings in the vm page table
1138 *
1139 * @adev: amdgpu_device pointer
1140 * @bo_va: requested BO and VM object
1141 * @clear: if true clear the entries
1142 *
1143 * Fill in the page table entries for @bo_va.
1144 *
1145 * Returns:
1146 * 0 for success, -EINVAL for failure.
1147 */
1148int amdgpu_vm_bo_update(struct amdgpu_device *adev, struct amdgpu_bo_va *bo_va,
1149 bool clear)
1150{
1151 struct amdgpu_bo *bo = bo_va->base.bo;
1152 struct amdgpu_vm *vm = bo_va->base.vm;
1153 struct amdgpu_bo_va_mapping *mapping;
1154 struct dma_fence **last_update;
1155 dma_addr_t *pages_addr = NULL;
1156 struct ttm_resource *mem;
1157 struct amdgpu_sync sync;
1158 bool flush_tlb = clear;
1159 uint64_t vram_base;
1160 uint64_t flags;
1161 bool uncached;
1162 int r;
1163
1164 amdgpu_sync_create(&sync);
1165 if (clear) {
1166 mem = NULL;
1167
1168 /* Implicitly sync to command submissions in the same VM before
1169 * unmapping.
1170 */
1171 r = amdgpu_sync_resv(adev, &sync, vm->root.bo->tbo.base.resv,
1172 AMDGPU_SYNC_EQ_OWNER, vm);
1173 if (r)
1174 goto error_free;
1175 if (bo) {
1176 r = amdgpu_sync_kfd(&sync, bo->tbo.base.resv);
1177 if (r)
1178 goto error_free;
1179 }
1180 } else if (!bo) {
1181 mem = NULL;
1182
1183 /* PRT map operations don't need to sync to anything. */
1184
1185 } else {
1186 struct drm_gem_object *obj = &bo->tbo.base;
1187
1188 if (obj->import_attach && bo_va->is_xgmi) {
1189 struct dma_buf *dma_buf = obj->import_attach->dmabuf;
1190 struct drm_gem_object *gobj = dma_buf->priv;
1191 struct amdgpu_bo *abo = gem_to_amdgpu_bo(gobj);
1192
1193 if (abo->tbo.resource &&
1194 abo->tbo.resource->mem_type == TTM_PL_VRAM)
1195 bo = gem_to_amdgpu_bo(gobj);
1196 }
1197 mem = bo->tbo.resource;
1198 if (mem && (mem->mem_type == TTM_PL_TT ||
1199 mem->mem_type == AMDGPU_PL_PREEMPT))
1200 pages_addr = bo->tbo.ttm->dma_address;
1201
1202 /* Implicitly sync to moving fences before mapping anything */
1203 r = amdgpu_sync_resv(adev, &sync, bo->tbo.base.resv,
1204 AMDGPU_SYNC_EXPLICIT, vm);
1205 if (r)
1206 goto error_free;
1207 }
1208
1209 if (bo) {
1210 struct amdgpu_device *bo_adev;
1211
1212 flags = amdgpu_ttm_tt_pte_flags(adev, bo->tbo.ttm, mem);
1213
1214 if (amdgpu_bo_encrypted(bo))
1215 flags |= AMDGPU_PTE_TMZ;
1216
1217 bo_adev = amdgpu_ttm_adev(bo->tbo.bdev);
1218 vram_base = bo_adev->vm_manager.vram_base_offset;
1219 uncached = (bo->flags & AMDGPU_GEM_CREATE_UNCACHED) != 0;
1220 } else {
1221 flags = 0x0;
1222 vram_base = 0;
1223 uncached = false;
1224 }
1225
1226 if (clear || amdgpu_vm_is_bo_always_valid(vm, bo))
1227 last_update = &vm->last_update;
1228 else
1229 last_update = &bo_va->last_pt_update;
1230
1231 if (!clear && bo_va->base.moved) {
1232 flush_tlb = true;
1233 list_splice_init(&bo_va->valids, &bo_va->invalids);
1234
1235 } else if (bo_va->cleared != clear) {
1236 list_splice_init(&bo_va->valids, &bo_va->invalids);
1237 }
1238
1239 list_for_each_entry(mapping, &bo_va->invalids, list) {
1240 uint64_t update_flags = flags;
1241
1242 /* normally,bo_va->flags only contians READABLE and WIRTEABLE bit go here
1243 * but in case of something, we filter the flags in first place
1244 */
1245 if (!(mapping->flags & AMDGPU_PTE_READABLE))
1246 update_flags &= ~AMDGPU_PTE_READABLE;
1247 if (!(mapping->flags & AMDGPU_PTE_WRITEABLE))
1248 update_flags &= ~AMDGPU_PTE_WRITEABLE;
1249
1250 /* Apply ASIC specific mapping flags */
1251 amdgpu_gmc_get_vm_pte(adev, mapping, &update_flags);
1252
1253 trace_amdgpu_vm_bo_update(mapping);
1254
1255 r = amdgpu_vm_update_range(adev, vm, false, false, flush_tlb,
1256 !uncached, &sync, mapping->start,
1257 mapping->last, update_flags,
1258 mapping->offset, vram_base, mem,
1259 pages_addr, last_update);
1260 if (r)
1261 goto error_free;
1262 }
1263
1264 /* If the BO is not in its preferred location add it back to
1265 * the evicted list so that it gets validated again on the
1266 * next command submission.
1267 */
1268 if (amdgpu_vm_is_bo_always_valid(vm, bo)) {
1269 if (bo->tbo.resource &&
1270 !(bo->preferred_domains &
1271 amdgpu_mem_type_to_domain(bo->tbo.resource->mem_type)))
1272 amdgpu_vm_bo_evicted(&bo_va->base);
1273 else
1274 amdgpu_vm_bo_idle(&bo_va->base);
1275 } else {
1276 amdgpu_vm_bo_done(&bo_va->base);
1277 }
1278
1279 list_splice_init(&bo_va->invalids, &bo_va->valids);
1280 bo_va->cleared = clear;
1281 bo_va->base.moved = false;
1282
1283 if (trace_amdgpu_vm_bo_mapping_enabled()) {
1284 list_for_each_entry(mapping, &bo_va->valids, list)
1285 trace_amdgpu_vm_bo_mapping(mapping);
1286 }
1287
1288error_free:
1289 amdgpu_sync_free(&sync);
1290 return r;
1291}
1292
1293/**
1294 * amdgpu_vm_update_prt_state - update the global PRT state
1295 *
1296 * @adev: amdgpu_device pointer
1297 */
1298static void amdgpu_vm_update_prt_state(struct amdgpu_device *adev)
1299{
1300 unsigned long flags;
1301 bool enable;
1302
1303 spin_lock_irqsave(&adev->vm_manager.prt_lock, flags);
1304 enable = !!atomic_read(&adev->vm_manager.num_prt_users);
1305 adev->gmc.gmc_funcs->set_prt(adev, enable);
1306 spin_unlock_irqrestore(&adev->vm_manager.prt_lock, flags);
1307}
1308
1309/**
1310 * amdgpu_vm_prt_get - add a PRT user
1311 *
1312 * @adev: amdgpu_device pointer
1313 */
1314static void amdgpu_vm_prt_get(struct amdgpu_device *adev)
1315{
1316 if (!adev->gmc.gmc_funcs->set_prt)
1317 return;
1318
1319 if (atomic_inc_return(&adev->vm_manager.num_prt_users) == 1)
1320 amdgpu_vm_update_prt_state(adev);
1321}
1322
1323/**
1324 * amdgpu_vm_prt_put - drop a PRT user
1325 *
1326 * @adev: amdgpu_device pointer
1327 */
1328static void amdgpu_vm_prt_put(struct amdgpu_device *adev)
1329{
1330 if (atomic_dec_return(&adev->vm_manager.num_prt_users) == 0)
1331 amdgpu_vm_update_prt_state(adev);
1332}
1333
1334/**
1335 * amdgpu_vm_prt_cb - callback for updating the PRT status
1336 *
1337 * @fence: fence for the callback
1338 * @_cb: the callback function
1339 */
1340static void amdgpu_vm_prt_cb(struct dma_fence *fence, struct dma_fence_cb *_cb)
1341{
1342 struct amdgpu_prt_cb *cb = container_of(_cb, struct amdgpu_prt_cb, cb);
1343
1344 amdgpu_vm_prt_put(cb->adev);
1345 kfree(cb);
1346}
1347
1348/**
1349 * amdgpu_vm_add_prt_cb - add callback for updating the PRT status
1350 *
1351 * @adev: amdgpu_device pointer
1352 * @fence: fence for the callback
1353 */
1354static void amdgpu_vm_add_prt_cb(struct amdgpu_device *adev,
1355 struct dma_fence *fence)
1356{
1357 struct amdgpu_prt_cb *cb;
1358
1359 if (!adev->gmc.gmc_funcs->set_prt)
1360 return;
1361
1362 cb = kmalloc(sizeof(struct amdgpu_prt_cb), GFP_KERNEL);
1363 if (!cb) {
1364 /* Last resort when we are OOM */
1365 if (fence)
1366 dma_fence_wait(fence, false);
1367
1368 amdgpu_vm_prt_put(adev);
1369 } else {
1370 cb->adev = adev;
1371 if (!fence || dma_fence_add_callback(fence, &cb->cb,
1372 amdgpu_vm_prt_cb))
1373 amdgpu_vm_prt_cb(fence, &cb->cb);
1374 }
1375}
1376
1377/**
1378 * amdgpu_vm_free_mapping - free a mapping
1379 *
1380 * @adev: amdgpu_device pointer
1381 * @vm: requested vm
1382 * @mapping: mapping to be freed
1383 * @fence: fence of the unmap operation
1384 *
1385 * Free a mapping and make sure we decrease the PRT usage count if applicable.
1386 */
1387static void amdgpu_vm_free_mapping(struct amdgpu_device *adev,
1388 struct amdgpu_vm *vm,
1389 struct amdgpu_bo_va_mapping *mapping,
1390 struct dma_fence *fence)
1391{
1392 if (mapping->flags & AMDGPU_PTE_PRT_FLAG(adev))
1393 amdgpu_vm_add_prt_cb(adev, fence);
1394 kfree(mapping);
1395}
1396
1397/**
1398 * amdgpu_vm_prt_fini - finish all prt mappings
1399 *
1400 * @adev: amdgpu_device pointer
1401 * @vm: requested vm
1402 *
1403 * Register a cleanup callback to disable PRT support after VM dies.
1404 */
1405static void amdgpu_vm_prt_fini(struct amdgpu_device *adev, struct amdgpu_vm *vm)
1406{
1407 struct dma_resv *resv = vm->root.bo->tbo.base.resv;
1408 struct dma_resv_iter cursor;
1409 struct dma_fence *fence;
1410
1411 dma_resv_for_each_fence(&cursor, resv, DMA_RESV_USAGE_BOOKKEEP, fence) {
1412 /* Add a callback for each fence in the reservation object */
1413 amdgpu_vm_prt_get(adev);
1414 amdgpu_vm_add_prt_cb(adev, fence);
1415 }
1416}
1417
1418/**
1419 * amdgpu_vm_clear_freed - clear freed BOs in the PT
1420 *
1421 * @adev: amdgpu_device pointer
1422 * @vm: requested vm
1423 * @fence: optional resulting fence (unchanged if no work needed to be done
1424 * or if an error occurred)
1425 *
1426 * Make sure all freed BOs are cleared in the PT.
1427 * PTs have to be reserved and mutex must be locked!
1428 *
1429 * Returns:
1430 * 0 for success.
1431 *
1432 */
1433int amdgpu_vm_clear_freed(struct amdgpu_device *adev,
1434 struct amdgpu_vm *vm,
1435 struct dma_fence **fence)
1436{
1437 struct amdgpu_bo_va_mapping *mapping;
1438 struct dma_fence *f = NULL;
1439 struct amdgpu_sync sync;
1440 int r;
1441
1442
1443 /*
1444 * Implicitly sync to command submissions in the same VM before
1445 * unmapping.
1446 */
1447 amdgpu_sync_create(&sync);
1448 r = amdgpu_sync_resv(adev, &sync, vm->root.bo->tbo.base.resv,
1449 AMDGPU_SYNC_EQ_OWNER, vm);
1450 if (r)
1451 goto error_free;
1452
1453 while (!list_empty(&vm->freed)) {
1454 mapping = list_first_entry(&vm->freed,
1455 struct amdgpu_bo_va_mapping, list);
1456 list_del(&mapping->list);
1457
1458 r = amdgpu_vm_update_range(adev, vm, false, false, true, false,
1459 &sync, mapping->start, mapping->last,
1460 0, 0, 0, NULL, NULL, &f);
1461 amdgpu_vm_free_mapping(adev, vm, mapping, f);
1462 if (r) {
1463 dma_fence_put(f);
1464 goto error_free;
1465 }
1466 }
1467
1468 if (fence && f) {
1469 dma_fence_put(*fence);
1470 *fence = f;
1471 } else {
1472 dma_fence_put(f);
1473 }
1474
1475error_free:
1476 amdgpu_sync_free(&sync);
1477 return r;
1478
1479}
1480
1481/**
1482 * amdgpu_vm_handle_moved - handle moved BOs in the PT
1483 *
1484 * @adev: amdgpu_device pointer
1485 * @vm: requested vm
1486 * @ticket: optional reservation ticket used to reserve the VM
1487 *
1488 * Make sure all BOs which are moved are updated in the PTs.
1489 *
1490 * Returns:
1491 * 0 for success.
1492 *
1493 * PTs have to be reserved!
1494 */
1495int amdgpu_vm_handle_moved(struct amdgpu_device *adev,
1496 struct amdgpu_vm *vm,
1497 struct ww_acquire_ctx *ticket)
1498{
1499 struct amdgpu_bo_va *bo_va;
1500 struct dma_resv *resv;
1501 bool clear, unlock;
1502 int r;
1503
1504 spin_lock(&vm->status_lock);
1505 while (!list_empty(&vm->moved)) {
1506 bo_va = list_first_entry(&vm->moved, struct amdgpu_bo_va,
1507 base.vm_status);
1508 spin_unlock(&vm->status_lock);
1509
1510 /* Per VM BOs never need to bo cleared in the page tables */
1511 r = amdgpu_vm_bo_update(adev, bo_va, false);
1512 if (r)
1513 return r;
1514 spin_lock(&vm->status_lock);
1515 }
1516
1517 while (!list_empty(&vm->invalidated)) {
1518 bo_va = list_first_entry(&vm->invalidated, struct amdgpu_bo_va,
1519 base.vm_status);
1520 resv = bo_va->base.bo->tbo.base.resv;
1521 spin_unlock(&vm->status_lock);
1522
1523 /* Try to reserve the BO to avoid clearing its ptes */
1524 if (!adev->debug_vm && dma_resv_trylock(resv)) {
1525 clear = false;
1526 unlock = true;
1527 /* The caller is already holding the reservation lock */
1528 } else if (ticket && dma_resv_locking_ctx(resv) == ticket) {
1529 clear = false;
1530 unlock = false;
1531 /* Somebody else is using the BO right now */
1532 } else {
1533 clear = true;
1534 unlock = false;
1535 }
1536
1537 r = amdgpu_vm_bo_update(adev, bo_va, clear);
1538
1539 if (unlock)
1540 dma_resv_unlock(resv);
1541 if (r)
1542 return r;
1543
1544 /* Remember evicted DMABuf imports in compute VMs for later
1545 * validation
1546 */
1547 if (vm->is_compute_context &&
1548 bo_va->base.bo->tbo.base.import_attach &&
1549 (!bo_va->base.bo->tbo.resource ||
1550 bo_va->base.bo->tbo.resource->mem_type == TTM_PL_SYSTEM))
1551 amdgpu_vm_bo_evicted_user(&bo_va->base);
1552
1553 spin_lock(&vm->status_lock);
1554 }
1555 spin_unlock(&vm->status_lock);
1556
1557 return 0;
1558}
1559
1560/**
1561 * amdgpu_vm_flush_compute_tlb - Flush TLB on compute VM
1562 *
1563 * @adev: amdgpu_device pointer
1564 * @vm: requested vm
1565 * @flush_type: flush type
1566 * @xcc_mask: mask of XCCs that belong to the compute partition in need of a TLB flush.
1567 *
1568 * Flush TLB if needed for a compute VM.
1569 *
1570 * Returns:
1571 * 0 for success.
1572 */
1573int amdgpu_vm_flush_compute_tlb(struct amdgpu_device *adev,
1574 struct amdgpu_vm *vm,
1575 uint32_t flush_type,
1576 uint32_t xcc_mask)
1577{
1578 uint64_t tlb_seq = amdgpu_vm_tlb_seq(vm);
1579 bool all_hub = false;
1580 int xcc = 0, r = 0;
1581
1582 WARN_ON_ONCE(!vm->is_compute_context);
1583
1584 /*
1585 * It can be that we race and lose here, but that is extremely unlikely
1586 * and the worst thing which could happen is that we flush the changes
1587 * into the TLB once more which is harmless.
1588 */
1589 if (atomic64_xchg(&vm->kfd_last_flushed_seq, tlb_seq) == tlb_seq)
1590 return 0;
1591
1592 if (adev->family == AMDGPU_FAMILY_AI ||
1593 adev->family == AMDGPU_FAMILY_RV)
1594 all_hub = true;
1595
1596 for_each_inst(xcc, xcc_mask) {
1597 r = amdgpu_gmc_flush_gpu_tlb_pasid(adev, vm->pasid, flush_type,
1598 all_hub, xcc);
1599 if (r)
1600 break;
1601 }
1602 return r;
1603}
1604
1605/**
1606 * amdgpu_vm_bo_add - add a bo to a specific vm
1607 *
1608 * @adev: amdgpu_device pointer
1609 * @vm: requested vm
1610 * @bo: amdgpu buffer object
1611 *
1612 * Add @bo into the requested vm.
1613 * Add @bo to the list of bos associated with the vm
1614 *
1615 * Returns:
1616 * Newly added bo_va or NULL for failure
1617 *
1618 * Object has to be reserved!
1619 */
1620struct amdgpu_bo_va *amdgpu_vm_bo_add(struct amdgpu_device *adev,
1621 struct amdgpu_vm *vm,
1622 struct amdgpu_bo *bo)
1623{
1624 struct amdgpu_bo_va *bo_va;
1625
1626 bo_va = kzalloc(sizeof(struct amdgpu_bo_va), GFP_KERNEL);
1627 if (bo_va == NULL) {
1628 return NULL;
1629 }
1630 amdgpu_vm_bo_base_init(&bo_va->base, vm, bo);
1631
1632 bo_va->ref_count = 1;
1633 bo_va->last_pt_update = dma_fence_get_stub();
1634 INIT_LIST_HEAD(&bo_va->valids);
1635 INIT_LIST_HEAD(&bo_va->invalids);
1636
1637 if (!bo)
1638 return bo_va;
1639
1640 dma_resv_assert_held(bo->tbo.base.resv);
1641 if (amdgpu_dmabuf_is_xgmi_accessible(adev, bo)) {
1642 bo_va->is_xgmi = true;
1643 /* Power up XGMI if it can be potentially used */
1644 amdgpu_xgmi_set_pstate(adev, AMDGPU_XGMI_PSTATE_MAX_VEGA20);
1645 }
1646
1647 return bo_va;
1648}
1649
1650
1651/**
1652 * amdgpu_vm_bo_insert_map - insert a new mapping
1653 *
1654 * @adev: amdgpu_device pointer
1655 * @bo_va: bo_va to store the address
1656 * @mapping: the mapping to insert
1657 *
1658 * Insert a new mapping into all structures.
1659 */
1660static void amdgpu_vm_bo_insert_map(struct amdgpu_device *adev,
1661 struct amdgpu_bo_va *bo_va,
1662 struct amdgpu_bo_va_mapping *mapping)
1663{
1664 struct amdgpu_vm *vm = bo_va->base.vm;
1665 struct amdgpu_bo *bo = bo_va->base.bo;
1666
1667 mapping->bo_va = bo_va;
1668 list_add(&mapping->list, &bo_va->invalids);
1669 amdgpu_vm_it_insert(mapping, &vm->va);
1670
1671 if (mapping->flags & AMDGPU_PTE_PRT_FLAG(adev))
1672 amdgpu_vm_prt_get(adev);
1673
1674 if (amdgpu_vm_is_bo_always_valid(vm, bo) && !bo_va->base.moved)
1675 amdgpu_vm_bo_moved(&bo_va->base);
1676
1677 trace_amdgpu_vm_bo_map(bo_va, mapping);
1678}
1679
1680/* Validate operation parameters to prevent potential abuse */
1681static int amdgpu_vm_verify_parameters(struct amdgpu_device *adev,
1682 struct amdgpu_bo *bo,
1683 uint64_t saddr,
1684 uint64_t offset,
1685 uint64_t size)
1686{
1687 uint64_t tmp, lpfn;
1688
1689 if (saddr & AMDGPU_GPU_PAGE_MASK
1690 || offset & AMDGPU_GPU_PAGE_MASK
1691 || size & AMDGPU_GPU_PAGE_MASK)
1692 return -EINVAL;
1693
1694 if (check_add_overflow(saddr, size, &tmp)
1695 || check_add_overflow(offset, size, &tmp)
1696 || size == 0 /* which also leads to end < begin */)
1697 return -EINVAL;
1698
1699 /* make sure object fit at this offset */
1700 if (bo && offset + size > amdgpu_bo_size(bo))
1701 return -EINVAL;
1702
1703 /* Ensure last pfn not exceed max_pfn */
1704 lpfn = (saddr + size - 1) >> AMDGPU_GPU_PAGE_SHIFT;
1705 if (lpfn >= adev->vm_manager.max_pfn)
1706 return -EINVAL;
1707
1708 return 0;
1709}
1710
1711/**
1712 * amdgpu_vm_bo_map - map bo inside a vm
1713 *
1714 * @adev: amdgpu_device pointer
1715 * @bo_va: bo_va to store the address
1716 * @saddr: where to map the BO
1717 * @offset: requested offset in the BO
1718 * @size: BO size in bytes
1719 * @flags: attributes of pages (read/write/valid/etc.)
1720 *
1721 * Add a mapping of the BO at the specefied addr into the VM.
1722 *
1723 * Returns:
1724 * 0 for success, error for failure.
1725 *
1726 * Object has to be reserved and unreserved outside!
1727 */
1728int amdgpu_vm_bo_map(struct amdgpu_device *adev,
1729 struct amdgpu_bo_va *bo_va,
1730 uint64_t saddr, uint64_t offset,
1731 uint64_t size, uint64_t flags)
1732{
1733 struct amdgpu_bo_va_mapping *mapping, *tmp;
1734 struct amdgpu_bo *bo = bo_va->base.bo;
1735 struct amdgpu_vm *vm = bo_va->base.vm;
1736 uint64_t eaddr;
1737 int r;
1738
1739 r = amdgpu_vm_verify_parameters(adev, bo, saddr, offset, size);
1740 if (r)
1741 return r;
1742
1743 saddr /= AMDGPU_GPU_PAGE_SIZE;
1744 eaddr = saddr + (size - 1) / AMDGPU_GPU_PAGE_SIZE;
1745
1746 tmp = amdgpu_vm_it_iter_first(&vm->va, saddr, eaddr);
1747 if (tmp) {
1748 /* bo and tmp overlap, invalid addr */
1749 dev_err(adev->dev, "bo %p va 0x%010Lx-0x%010Lx conflict with "
1750 "0x%010Lx-0x%010Lx\n", bo, saddr, eaddr,
1751 tmp->start, tmp->last + 1);
1752 return -EINVAL;
1753 }
1754
1755 mapping = kmalloc(sizeof(*mapping), GFP_KERNEL);
1756 if (!mapping)
1757 return -ENOMEM;
1758
1759 mapping->start = saddr;
1760 mapping->last = eaddr;
1761 mapping->offset = offset;
1762 mapping->flags = flags;
1763
1764 amdgpu_vm_bo_insert_map(adev, bo_va, mapping);
1765
1766 return 0;
1767}
1768
1769/**
1770 * amdgpu_vm_bo_replace_map - map bo inside a vm, replacing existing mappings
1771 *
1772 * @adev: amdgpu_device pointer
1773 * @bo_va: bo_va to store the address
1774 * @saddr: where to map the BO
1775 * @offset: requested offset in the BO
1776 * @size: BO size in bytes
1777 * @flags: attributes of pages (read/write/valid/etc.)
1778 *
1779 * Add a mapping of the BO at the specefied addr into the VM. Replace existing
1780 * mappings as we do so.
1781 *
1782 * Returns:
1783 * 0 for success, error for failure.
1784 *
1785 * Object has to be reserved and unreserved outside!
1786 */
1787int amdgpu_vm_bo_replace_map(struct amdgpu_device *adev,
1788 struct amdgpu_bo_va *bo_va,
1789 uint64_t saddr, uint64_t offset,
1790 uint64_t size, uint64_t flags)
1791{
1792 struct amdgpu_bo_va_mapping *mapping;
1793 struct amdgpu_bo *bo = bo_va->base.bo;
1794 uint64_t eaddr;
1795 int r;
1796
1797 r = amdgpu_vm_verify_parameters(adev, bo, saddr, offset, size);
1798 if (r)
1799 return r;
1800
1801 /* Allocate all the needed memory */
1802 mapping = kmalloc(sizeof(*mapping), GFP_KERNEL);
1803 if (!mapping)
1804 return -ENOMEM;
1805
1806 r = amdgpu_vm_bo_clear_mappings(adev, bo_va->base.vm, saddr, size);
1807 if (r) {
1808 kfree(mapping);
1809 return r;
1810 }
1811
1812 saddr /= AMDGPU_GPU_PAGE_SIZE;
1813 eaddr = saddr + (size - 1) / AMDGPU_GPU_PAGE_SIZE;
1814
1815 mapping->start = saddr;
1816 mapping->last = eaddr;
1817 mapping->offset = offset;
1818 mapping->flags = flags;
1819
1820 amdgpu_vm_bo_insert_map(adev, bo_va, mapping);
1821
1822 return 0;
1823}
1824
1825/**
1826 * amdgpu_vm_bo_unmap - remove bo mapping from vm
1827 *
1828 * @adev: amdgpu_device pointer
1829 * @bo_va: bo_va to remove the address from
1830 * @saddr: where to the BO is mapped
1831 *
1832 * Remove a mapping of the BO at the specefied addr from the VM.
1833 *
1834 * Returns:
1835 * 0 for success, error for failure.
1836 *
1837 * Object has to be reserved and unreserved outside!
1838 */
1839int amdgpu_vm_bo_unmap(struct amdgpu_device *adev,
1840 struct amdgpu_bo_va *bo_va,
1841 uint64_t saddr)
1842{
1843 struct amdgpu_bo_va_mapping *mapping;
1844 struct amdgpu_vm *vm = bo_va->base.vm;
1845 bool valid = true;
1846
1847 saddr /= AMDGPU_GPU_PAGE_SIZE;
1848
1849 list_for_each_entry(mapping, &bo_va->valids, list) {
1850 if (mapping->start == saddr)
1851 break;
1852 }
1853
1854 if (&mapping->list == &bo_va->valids) {
1855 valid = false;
1856
1857 list_for_each_entry(mapping, &bo_va->invalids, list) {
1858 if (mapping->start == saddr)
1859 break;
1860 }
1861
1862 if (&mapping->list == &bo_va->invalids)
1863 return -ENOENT;
1864 }
1865
1866 list_del(&mapping->list);
1867 amdgpu_vm_it_remove(mapping, &vm->va);
1868 mapping->bo_va = NULL;
1869 trace_amdgpu_vm_bo_unmap(bo_va, mapping);
1870
1871 if (valid)
1872 list_add(&mapping->list, &vm->freed);
1873 else
1874 amdgpu_vm_free_mapping(adev, vm, mapping,
1875 bo_va->last_pt_update);
1876
1877 return 0;
1878}
1879
1880/**
1881 * amdgpu_vm_bo_clear_mappings - remove all mappings in a specific range
1882 *
1883 * @adev: amdgpu_device pointer
1884 * @vm: VM structure to use
1885 * @saddr: start of the range
1886 * @size: size of the range
1887 *
1888 * Remove all mappings in a range, split them as appropriate.
1889 *
1890 * Returns:
1891 * 0 for success, error for failure.
1892 */
1893int amdgpu_vm_bo_clear_mappings(struct amdgpu_device *adev,
1894 struct amdgpu_vm *vm,
1895 uint64_t saddr, uint64_t size)
1896{
1897 struct amdgpu_bo_va_mapping *before, *after, *tmp, *next;
1898 LIST_HEAD(removed);
1899 uint64_t eaddr;
1900 int r;
1901
1902 r = amdgpu_vm_verify_parameters(adev, NULL, saddr, 0, size);
1903 if (r)
1904 return r;
1905
1906 saddr /= AMDGPU_GPU_PAGE_SIZE;
1907 eaddr = saddr + (size - 1) / AMDGPU_GPU_PAGE_SIZE;
1908
1909 /* Allocate all the needed memory */
1910 before = kzalloc(sizeof(*before), GFP_KERNEL);
1911 if (!before)
1912 return -ENOMEM;
1913 INIT_LIST_HEAD(&before->list);
1914
1915 after = kzalloc(sizeof(*after), GFP_KERNEL);
1916 if (!after) {
1917 kfree(before);
1918 return -ENOMEM;
1919 }
1920 INIT_LIST_HEAD(&after->list);
1921
1922 /* Now gather all removed mappings */
1923 tmp = amdgpu_vm_it_iter_first(&vm->va, saddr, eaddr);
1924 while (tmp) {
1925 /* Remember mapping split at the start */
1926 if (tmp->start < saddr) {
1927 before->start = tmp->start;
1928 before->last = saddr - 1;
1929 before->offset = tmp->offset;
1930 before->flags = tmp->flags;
1931 before->bo_va = tmp->bo_va;
1932 list_add(&before->list, &tmp->bo_va->invalids);
1933 }
1934
1935 /* Remember mapping split at the end */
1936 if (tmp->last > eaddr) {
1937 after->start = eaddr + 1;
1938 after->last = tmp->last;
1939 after->offset = tmp->offset;
1940 after->offset += (after->start - tmp->start) << PAGE_SHIFT;
1941 after->flags = tmp->flags;
1942 after->bo_va = tmp->bo_va;
1943 list_add(&after->list, &tmp->bo_va->invalids);
1944 }
1945
1946 list_del(&tmp->list);
1947 list_add(&tmp->list, &removed);
1948
1949 tmp = amdgpu_vm_it_iter_next(tmp, saddr, eaddr);
1950 }
1951
1952 /* And free them up */
1953 list_for_each_entry_safe(tmp, next, &removed, list) {
1954 amdgpu_vm_it_remove(tmp, &vm->va);
1955 list_del(&tmp->list);
1956
1957 if (tmp->start < saddr)
1958 tmp->start = saddr;
1959 if (tmp->last > eaddr)
1960 tmp->last = eaddr;
1961
1962 tmp->bo_va = NULL;
1963 list_add(&tmp->list, &vm->freed);
1964 trace_amdgpu_vm_bo_unmap(NULL, tmp);
1965 }
1966
1967 /* Insert partial mapping before the range */
1968 if (!list_empty(&before->list)) {
1969 struct amdgpu_bo *bo = before->bo_va->base.bo;
1970
1971 amdgpu_vm_it_insert(before, &vm->va);
1972 if (before->flags & AMDGPU_PTE_PRT_FLAG(adev))
1973 amdgpu_vm_prt_get(adev);
1974
1975 if (amdgpu_vm_is_bo_always_valid(vm, bo) &&
1976 !before->bo_va->base.moved)
1977 amdgpu_vm_bo_moved(&before->bo_va->base);
1978 } else {
1979 kfree(before);
1980 }
1981
1982 /* Insert partial mapping after the range */
1983 if (!list_empty(&after->list)) {
1984 struct amdgpu_bo *bo = after->bo_va->base.bo;
1985
1986 amdgpu_vm_it_insert(after, &vm->va);
1987 if (after->flags & AMDGPU_PTE_PRT_FLAG(adev))
1988 amdgpu_vm_prt_get(adev);
1989
1990 if (amdgpu_vm_is_bo_always_valid(vm, bo) &&
1991 !after->bo_va->base.moved)
1992 amdgpu_vm_bo_moved(&after->bo_va->base);
1993 } else {
1994 kfree(after);
1995 }
1996
1997 return 0;
1998}
1999
2000/**
2001 * amdgpu_vm_bo_lookup_mapping - find mapping by address
2002 *
2003 * @vm: the requested VM
2004 * @addr: the address
2005 *
2006 * Find a mapping by it's address.
2007 *
2008 * Returns:
2009 * The amdgpu_bo_va_mapping matching for addr or NULL
2010 *
2011 */
2012struct amdgpu_bo_va_mapping *amdgpu_vm_bo_lookup_mapping(struct amdgpu_vm *vm,
2013 uint64_t addr)
2014{
2015 return amdgpu_vm_it_iter_first(&vm->va, addr, addr);
2016}
2017
2018/**
2019 * amdgpu_vm_bo_trace_cs - trace all reserved mappings
2020 *
2021 * @vm: the requested vm
2022 * @ticket: CS ticket
2023 *
2024 * Trace all mappings of BOs reserved during a command submission.
2025 */
2026void amdgpu_vm_bo_trace_cs(struct amdgpu_vm *vm, struct ww_acquire_ctx *ticket)
2027{
2028 struct amdgpu_bo_va_mapping *mapping;
2029
2030 if (!trace_amdgpu_vm_bo_cs_enabled())
2031 return;
2032
2033 for (mapping = amdgpu_vm_it_iter_first(&vm->va, 0, U64_MAX); mapping;
2034 mapping = amdgpu_vm_it_iter_next(mapping, 0, U64_MAX)) {
2035 if (mapping->bo_va && mapping->bo_va->base.bo) {
2036 struct amdgpu_bo *bo;
2037
2038 bo = mapping->bo_va->base.bo;
2039 if (dma_resv_locking_ctx(bo->tbo.base.resv) !=
2040 ticket)
2041 continue;
2042 }
2043
2044 trace_amdgpu_vm_bo_cs(mapping);
2045 }
2046}
2047
2048/**
2049 * amdgpu_vm_bo_del - remove a bo from a specific vm
2050 *
2051 * @adev: amdgpu_device pointer
2052 * @bo_va: requested bo_va
2053 *
2054 * Remove @bo_va->bo from the requested vm.
2055 *
2056 * Object have to be reserved!
2057 */
2058void amdgpu_vm_bo_del(struct amdgpu_device *adev,
2059 struct amdgpu_bo_va *bo_va)
2060{
2061 struct amdgpu_bo_va_mapping *mapping, *next;
2062 struct amdgpu_bo *bo = bo_va->base.bo;
2063 struct amdgpu_vm *vm = bo_va->base.vm;
2064 struct amdgpu_vm_bo_base **base;
2065
2066 dma_resv_assert_held(vm->root.bo->tbo.base.resv);
2067
2068 if (bo) {
2069 dma_resv_assert_held(bo->tbo.base.resv);
2070 if (amdgpu_vm_is_bo_always_valid(vm, bo))
2071 ttm_bo_set_bulk_move(&bo->tbo, NULL);
2072
2073 for (base = &bo_va->base.bo->vm_bo; *base;
2074 base = &(*base)->next) {
2075 if (*base != &bo_va->base)
2076 continue;
2077
2078 *base = bo_va->base.next;
2079 break;
2080 }
2081 }
2082
2083 spin_lock(&vm->status_lock);
2084 list_del(&bo_va->base.vm_status);
2085 spin_unlock(&vm->status_lock);
2086
2087 list_for_each_entry_safe(mapping, next, &bo_va->valids, list) {
2088 list_del(&mapping->list);
2089 amdgpu_vm_it_remove(mapping, &vm->va);
2090 mapping->bo_va = NULL;
2091 trace_amdgpu_vm_bo_unmap(bo_va, mapping);
2092 list_add(&mapping->list, &vm->freed);
2093 }
2094 list_for_each_entry_safe(mapping, next, &bo_va->invalids, list) {
2095 list_del(&mapping->list);
2096 amdgpu_vm_it_remove(mapping, &vm->va);
2097 amdgpu_vm_free_mapping(adev, vm, mapping,
2098 bo_va->last_pt_update);
2099 }
2100
2101 dma_fence_put(bo_va->last_pt_update);
2102
2103 if (bo && bo_va->is_xgmi)
2104 amdgpu_xgmi_set_pstate(adev, AMDGPU_XGMI_PSTATE_MIN);
2105
2106 kfree(bo_va);
2107}
2108
2109/**
2110 * amdgpu_vm_evictable - check if we can evict a VM
2111 *
2112 * @bo: A page table of the VM.
2113 *
2114 * Check if it is possible to evict a VM.
2115 */
2116bool amdgpu_vm_evictable(struct amdgpu_bo *bo)
2117{
2118 struct amdgpu_vm_bo_base *bo_base = bo->vm_bo;
2119
2120 /* Page tables of a destroyed VM can go away immediately */
2121 if (!bo_base || !bo_base->vm)
2122 return true;
2123
2124 /* Don't evict VM page tables while they are busy */
2125 if (!dma_resv_test_signaled(bo->tbo.base.resv, DMA_RESV_USAGE_BOOKKEEP))
2126 return false;
2127
2128 /* Try to block ongoing updates */
2129 if (!amdgpu_vm_eviction_trylock(bo_base->vm))
2130 return false;
2131
2132 /* Don't evict VM page tables while they are updated */
2133 if (!dma_fence_is_signaled(bo_base->vm->last_unlocked)) {
2134 amdgpu_vm_eviction_unlock(bo_base->vm);
2135 return false;
2136 }
2137
2138 bo_base->vm->evicting = true;
2139 amdgpu_vm_eviction_unlock(bo_base->vm);
2140 return true;
2141}
2142
2143/**
2144 * amdgpu_vm_bo_invalidate - mark the bo as invalid
2145 *
2146 * @adev: amdgpu_device pointer
2147 * @bo: amdgpu buffer object
2148 * @evicted: is the BO evicted
2149 *
2150 * Mark @bo as invalid.
2151 */
2152void amdgpu_vm_bo_invalidate(struct amdgpu_device *adev,
2153 struct amdgpu_bo *bo, bool evicted)
2154{
2155 struct amdgpu_vm_bo_base *bo_base;
2156
2157 for (bo_base = bo->vm_bo; bo_base; bo_base = bo_base->next) {
2158 struct amdgpu_vm *vm = bo_base->vm;
2159
2160 if (evicted && amdgpu_vm_is_bo_always_valid(vm, bo)) {
2161 amdgpu_vm_bo_evicted(bo_base);
2162 continue;
2163 }
2164
2165 if (bo_base->moved)
2166 continue;
2167 bo_base->moved = true;
2168
2169 if (bo->tbo.type == ttm_bo_type_kernel)
2170 amdgpu_vm_bo_relocated(bo_base);
2171 else if (amdgpu_vm_is_bo_always_valid(vm, bo))
2172 amdgpu_vm_bo_moved(bo_base);
2173 else
2174 amdgpu_vm_bo_invalidated(bo_base);
2175 }
2176}
2177
2178/**
2179 * amdgpu_vm_get_block_size - calculate VM page table size as power of two
2180 *
2181 * @vm_size: VM size
2182 *
2183 * Returns:
2184 * VM page table as power of two
2185 */
2186static uint32_t amdgpu_vm_get_block_size(uint64_t vm_size)
2187{
2188 /* Total bits covered by PD + PTs */
2189 unsigned bits = ilog2(vm_size) + 18;
2190
2191 /* Make sure the PD is 4K in size up to 8GB address space.
2192 Above that split equal between PD and PTs */
2193 if (vm_size <= 8)
2194 return (bits - 9);
2195 else
2196 return ((bits + 3) / 2);
2197}
2198
2199/**
2200 * amdgpu_vm_adjust_size - adjust vm size, block size and fragment size
2201 *
2202 * @adev: amdgpu_device pointer
2203 * @min_vm_size: the minimum vm size in GB if it's set auto
2204 * @fragment_size_default: Default PTE fragment size
2205 * @max_level: max VMPT level
2206 * @max_bits: max address space size in bits
2207 *
2208 */
2209void amdgpu_vm_adjust_size(struct amdgpu_device *adev, uint32_t min_vm_size,
2210 uint32_t fragment_size_default, unsigned max_level,
2211 unsigned max_bits)
2212{
2213 unsigned int max_size = 1 << (max_bits - 30);
2214 unsigned int vm_size;
2215 uint64_t tmp;
2216
2217 /* adjust vm size first */
2218 if (amdgpu_vm_size != -1) {
2219 vm_size = amdgpu_vm_size;
2220 if (vm_size > max_size) {
2221 dev_warn(adev->dev, "VM size (%d) too large, max is %u GB\n",
2222 amdgpu_vm_size, max_size);
2223 vm_size = max_size;
2224 }
2225 } else {
2226 struct sysinfo si;
2227 unsigned int phys_ram_gb;
2228
2229 /* Optimal VM size depends on the amount of physical
2230 * RAM available. Underlying requirements and
2231 * assumptions:
2232 *
2233 * - Need to map system memory and VRAM from all GPUs
2234 * - VRAM from other GPUs not known here
2235 * - Assume VRAM <= system memory
2236 * - On GFX8 and older, VM space can be segmented for
2237 * different MTYPEs
2238 * - Need to allow room for fragmentation, guard pages etc.
2239 *
2240 * This adds up to a rough guess of system memory x3.
2241 * Round up to power of two to maximize the available
2242 * VM size with the given page table size.
2243 */
2244 si_meminfo(&si);
2245 phys_ram_gb = ((uint64_t)si.totalram * si.mem_unit +
2246 (1 << 30) - 1) >> 30;
2247 vm_size = roundup_pow_of_two(
2248 clamp(phys_ram_gb * 3, min_vm_size, max_size));
2249 }
2250
2251 adev->vm_manager.max_pfn = (uint64_t)vm_size << 18;
2252
2253 tmp = roundup_pow_of_two(adev->vm_manager.max_pfn);
2254 if (amdgpu_vm_block_size != -1)
2255 tmp >>= amdgpu_vm_block_size - 9;
2256 tmp = DIV_ROUND_UP(fls64(tmp) - 1, 9) - 1;
2257 adev->vm_manager.num_level = min_t(unsigned int, max_level, tmp);
2258 switch (adev->vm_manager.num_level) {
2259 case 3:
2260 adev->vm_manager.root_level = AMDGPU_VM_PDB2;
2261 break;
2262 case 2:
2263 adev->vm_manager.root_level = AMDGPU_VM_PDB1;
2264 break;
2265 case 1:
2266 adev->vm_manager.root_level = AMDGPU_VM_PDB0;
2267 break;
2268 default:
2269 dev_err(adev->dev, "VMPT only supports 2~4+1 levels\n");
2270 }
2271 /* block size depends on vm size and hw setup*/
2272 if (amdgpu_vm_block_size != -1)
2273 adev->vm_manager.block_size =
2274 min((unsigned)amdgpu_vm_block_size, max_bits
2275 - AMDGPU_GPU_PAGE_SHIFT
2276 - 9 * adev->vm_manager.num_level);
2277 else if (adev->vm_manager.num_level > 1)
2278 adev->vm_manager.block_size = 9;
2279 else
2280 adev->vm_manager.block_size = amdgpu_vm_get_block_size(tmp);
2281
2282 if (amdgpu_vm_fragment_size == -1)
2283 adev->vm_manager.fragment_size = fragment_size_default;
2284 else
2285 adev->vm_manager.fragment_size = amdgpu_vm_fragment_size;
2286
2287 DRM_INFO("vm size is %u GB, %u levels, block size is %u-bit, fragment size is %u-bit\n",
2288 vm_size, adev->vm_manager.num_level + 1,
2289 adev->vm_manager.block_size,
2290 adev->vm_manager.fragment_size);
2291}
2292
2293/**
2294 * amdgpu_vm_wait_idle - wait for the VM to become idle
2295 *
2296 * @vm: VM object to wait for
2297 * @timeout: timeout to wait for VM to become idle
2298 */
2299long amdgpu_vm_wait_idle(struct amdgpu_vm *vm, long timeout)
2300{
2301 timeout = dma_resv_wait_timeout(vm->root.bo->tbo.base.resv,
2302 DMA_RESV_USAGE_BOOKKEEP,
2303 true, timeout);
2304 if (timeout <= 0)
2305 return timeout;
2306
2307 return dma_fence_wait_timeout(vm->last_unlocked, true, timeout);
2308}
2309
2310static void amdgpu_vm_destroy_task_info(struct kref *kref)
2311{
2312 struct amdgpu_task_info *ti = container_of(kref, struct amdgpu_task_info, refcount);
2313
2314 kfree(ti);
2315}
2316
2317static inline struct amdgpu_vm *
2318amdgpu_vm_get_vm_from_pasid(struct amdgpu_device *adev, u32 pasid)
2319{
2320 struct amdgpu_vm *vm;
2321 unsigned long flags;
2322
2323 xa_lock_irqsave(&adev->vm_manager.pasids, flags);
2324 vm = xa_load(&adev->vm_manager.pasids, pasid);
2325 xa_unlock_irqrestore(&adev->vm_manager.pasids, flags);
2326
2327 return vm;
2328}
2329
2330/**
2331 * amdgpu_vm_put_task_info - reference down the vm task_info ptr
2332 *
2333 * @task_info: task_info struct under discussion.
2334 *
2335 * frees the vm task_info ptr at the last put
2336 */
2337void amdgpu_vm_put_task_info(struct amdgpu_task_info *task_info)
2338{
2339 kref_put(&task_info->refcount, amdgpu_vm_destroy_task_info);
2340}
2341
2342/**
2343 * amdgpu_vm_get_task_info_vm - Extracts task info for a vm.
2344 *
2345 * @vm: VM to get info from
2346 *
2347 * Returns the reference counted task_info structure, which must be
2348 * referenced down with amdgpu_vm_put_task_info.
2349 */
2350struct amdgpu_task_info *
2351amdgpu_vm_get_task_info_vm(struct amdgpu_vm *vm)
2352{
2353 struct amdgpu_task_info *ti = NULL;
2354
2355 if (vm) {
2356 ti = vm->task_info;
2357 kref_get(&vm->task_info->refcount);
2358 }
2359
2360 return ti;
2361}
2362
2363/**
2364 * amdgpu_vm_get_task_info_pasid - Extracts task info for a PASID.
2365 *
2366 * @adev: drm device pointer
2367 * @pasid: PASID identifier for VM
2368 *
2369 * Returns the reference counted task_info structure, which must be
2370 * referenced down with amdgpu_vm_put_task_info.
2371 */
2372struct amdgpu_task_info *
2373amdgpu_vm_get_task_info_pasid(struct amdgpu_device *adev, u32 pasid)
2374{
2375 return amdgpu_vm_get_task_info_vm(
2376 amdgpu_vm_get_vm_from_pasid(adev, pasid));
2377}
2378
2379static int amdgpu_vm_create_task_info(struct amdgpu_vm *vm)
2380{
2381 vm->task_info = kzalloc(sizeof(struct amdgpu_task_info), GFP_KERNEL);
2382 if (!vm->task_info)
2383 return -ENOMEM;
2384
2385 kref_init(&vm->task_info->refcount);
2386 return 0;
2387}
2388
2389/**
2390 * amdgpu_vm_set_task_info - Sets VMs task info.
2391 *
2392 * @vm: vm for which to set the info
2393 */
2394void amdgpu_vm_set_task_info(struct amdgpu_vm *vm)
2395{
2396 if (!vm->task_info)
2397 return;
2398
2399 if (vm->task_info->pid == current->pid)
2400 return;
2401
2402 vm->task_info->pid = current->pid;
2403 get_task_comm(vm->task_info->task_name, current);
2404
2405 if (current->group_leader->mm != current->mm)
2406 return;
2407
2408 vm->task_info->tgid = current->group_leader->pid;
2409 get_task_comm(vm->task_info->process_name, current->group_leader);
2410}
2411
2412/**
2413 * amdgpu_vm_init - initialize a vm instance
2414 *
2415 * @adev: amdgpu_device pointer
2416 * @vm: requested vm
2417 * @xcp_id: GPU partition selection id
2418 *
2419 * Init @vm fields.
2420 *
2421 * Returns:
2422 * 0 for success, error for failure.
2423 */
2424int amdgpu_vm_init(struct amdgpu_device *adev, struct amdgpu_vm *vm,
2425 int32_t xcp_id)
2426{
2427 struct amdgpu_bo *root_bo;
2428 struct amdgpu_bo_vm *root;
2429 int r, i;
2430
2431 vm->va = RB_ROOT_CACHED;
2432 for (i = 0; i < AMDGPU_MAX_VMHUBS; i++)
2433 vm->reserved_vmid[i] = NULL;
2434 INIT_LIST_HEAD(&vm->evicted);
2435 INIT_LIST_HEAD(&vm->evicted_user);
2436 INIT_LIST_HEAD(&vm->relocated);
2437 INIT_LIST_HEAD(&vm->moved);
2438 INIT_LIST_HEAD(&vm->idle);
2439 INIT_LIST_HEAD(&vm->invalidated);
2440 spin_lock_init(&vm->status_lock);
2441 INIT_LIST_HEAD(&vm->freed);
2442 INIT_LIST_HEAD(&vm->done);
2443 INIT_LIST_HEAD(&vm->pt_freed);
2444 INIT_WORK(&vm->pt_free_work, amdgpu_vm_pt_free_work);
2445 INIT_KFIFO(vm->faults);
2446
2447 r = amdgpu_vm_init_entities(adev, vm);
2448 if (r)
2449 return r;
2450
2451 ttm_lru_bulk_move_init(&vm->lru_bulk_move);
2452
2453 vm->is_compute_context = false;
2454
2455 vm->use_cpu_for_update = !!(adev->vm_manager.vm_update_mode &
2456 AMDGPU_VM_USE_CPU_FOR_GFX);
2457
2458 DRM_DEBUG_DRIVER("VM update mode is %s\n",
2459 vm->use_cpu_for_update ? "CPU" : "SDMA");
2460 WARN_ONCE((vm->use_cpu_for_update &&
2461 !amdgpu_gmc_vram_full_visible(&adev->gmc)),
2462 "CPU update of VM recommended only for large BAR system\n");
2463
2464 if (vm->use_cpu_for_update)
2465 vm->update_funcs = &amdgpu_vm_cpu_funcs;
2466 else
2467 vm->update_funcs = &amdgpu_vm_sdma_funcs;
2468
2469 vm->last_update = dma_fence_get_stub();
2470 vm->last_unlocked = dma_fence_get_stub();
2471 vm->last_tlb_flush = dma_fence_get_stub();
2472 vm->generation = amdgpu_vm_generation(adev, NULL);
2473
2474 mutex_init(&vm->eviction_lock);
2475 vm->evicting = false;
2476 vm->tlb_fence_context = dma_fence_context_alloc(1);
2477
2478 r = amdgpu_vm_pt_create(adev, vm, adev->vm_manager.root_level,
2479 false, &root, xcp_id);
2480 if (r)
2481 goto error_free_delayed;
2482
2483 root_bo = amdgpu_bo_ref(&root->bo);
2484 r = amdgpu_bo_reserve(root_bo, true);
2485 if (r) {
2486 amdgpu_bo_unref(&root_bo);
2487 goto error_free_delayed;
2488 }
2489
2490 amdgpu_vm_bo_base_init(&vm->root, vm, root_bo);
2491 r = dma_resv_reserve_fences(root_bo->tbo.base.resv, 1);
2492 if (r)
2493 goto error_free_root;
2494
2495 r = amdgpu_vm_pt_clear(adev, vm, root, false);
2496 if (r)
2497 goto error_free_root;
2498
2499 r = amdgpu_vm_create_task_info(vm);
2500 if (r)
2501 DRM_DEBUG("Failed to create task info for VM\n");
2502
2503 amdgpu_bo_unreserve(vm->root.bo);
2504 amdgpu_bo_unref(&root_bo);
2505
2506 return 0;
2507
2508error_free_root:
2509 amdgpu_vm_pt_free_root(adev, vm);
2510 amdgpu_bo_unreserve(vm->root.bo);
2511 amdgpu_bo_unref(&root_bo);
2512
2513error_free_delayed:
2514 dma_fence_put(vm->last_tlb_flush);
2515 dma_fence_put(vm->last_unlocked);
2516 ttm_lru_bulk_move_fini(&adev->mman.bdev, &vm->lru_bulk_move);
2517 amdgpu_vm_fini_entities(vm);
2518
2519 return r;
2520}
2521
2522/**
2523 * amdgpu_vm_make_compute - Turn a GFX VM into a compute VM
2524 *
2525 * @adev: amdgpu_device pointer
2526 * @vm: requested vm
2527 *
2528 * This only works on GFX VMs that don't have any BOs added and no
2529 * page tables allocated yet.
2530 *
2531 * Changes the following VM parameters:
2532 * - use_cpu_for_update
2533 * - pte_supports_ats
2534 *
2535 * Reinitializes the page directory to reflect the changed ATS
2536 * setting.
2537 *
2538 * Returns:
2539 * 0 for success, -errno for errors.
2540 */
2541int amdgpu_vm_make_compute(struct amdgpu_device *adev, struct amdgpu_vm *vm)
2542{
2543 int r;
2544
2545 r = amdgpu_bo_reserve(vm->root.bo, true);
2546 if (r)
2547 return r;
2548
2549 /* Update VM state */
2550 vm->use_cpu_for_update = !!(adev->vm_manager.vm_update_mode &
2551 AMDGPU_VM_USE_CPU_FOR_COMPUTE);
2552 DRM_DEBUG_DRIVER("VM update mode is %s\n",
2553 vm->use_cpu_for_update ? "CPU" : "SDMA");
2554 WARN_ONCE((vm->use_cpu_for_update &&
2555 !amdgpu_gmc_vram_full_visible(&adev->gmc)),
2556 "CPU update of VM recommended only for large BAR system\n");
2557
2558 if (vm->use_cpu_for_update) {
2559 /* Sync with last SDMA update/clear before switching to CPU */
2560 r = amdgpu_bo_sync_wait(vm->root.bo,
2561 AMDGPU_FENCE_OWNER_UNDEFINED, true);
2562 if (r)
2563 goto unreserve_bo;
2564
2565 vm->update_funcs = &amdgpu_vm_cpu_funcs;
2566 r = amdgpu_vm_pt_map_tables(adev, vm);
2567 if (r)
2568 goto unreserve_bo;
2569
2570 } else {
2571 vm->update_funcs = &amdgpu_vm_sdma_funcs;
2572 }
2573
2574 dma_fence_put(vm->last_update);
2575 vm->last_update = dma_fence_get_stub();
2576 vm->is_compute_context = true;
2577
2578unreserve_bo:
2579 amdgpu_bo_unreserve(vm->root.bo);
2580 return r;
2581}
2582
2583/**
2584 * amdgpu_vm_release_compute - release a compute vm
2585 * @adev: amdgpu_device pointer
2586 * @vm: a vm turned into compute vm by calling amdgpu_vm_make_compute
2587 *
2588 * This is a correspondant of amdgpu_vm_make_compute. It decouples compute
2589 * pasid from vm. Compute should stop use of vm after this call.
2590 */
2591void amdgpu_vm_release_compute(struct amdgpu_device *adev, struct amdgpu_vm *vm)
2592{
2593 amdgpu_vm_set_pasid(adev, vm, 0);
2594 vm->is_compute_context = false;
2595}
2596
2597/**
2598 * amdgpu_vm_fini - tear down a vm instance
2599 *
2600 * @adev: amdgpu_device pointer
2601 * @vm: requested vm
2602 *
2603 * Tear down @vm.
2604 * Unbind the VM and remove all bos from the vm bo list
2605 */
2606void amdgpu_vm_fini(struct amdgpu_device *adev, struct amdgpu_vm *vm)
2607{
2608 struct amdgpu_bo_va_mapping *mapping, *tmp;
2609 bool prt_fini_needed = !!adev->gmc.gmc_funcs->set_prt;
2610 struct amdgpu_bo *root;
2611 unsigned long flags;
2612 int i;
2613
2614 amdgpu_amdkfd_gpuvm_destroy_cb(adev, vm);
2615
2616 flush_work(&vm->pt_free_work);
2617
2618 root = amdgpu_bo_ref(vm->root.bo);
2619 amdgpu_bo_reserve(root, true);
2620 amdgpu_vm_put_task_info(vm->task_info);
2621 amdgpu_vm_set_pasid(adev, vm, 0);
2622 dma_fence_wait(vm->last_unlocked, false);
2623 dma_fence_put(vm->last_unlocked);
2624 dma_fence_wait(vm->last_tlb_flush, false);
2625 /* Make sure that all fence callbacks have completed */
2626 spin_lock_irqsave(vm->last_tlb_flush->lock, flags);
2627 spin_unlock_irqrestore(vm->last_tlb_flush->lock, flags);
2628 dma_fence_put(vm->last_tlb_flush);
2629
2630 list_for_each_entry_safe(mapping, tmp, &vm->freed, list) {
2631 if (mapping->flags & AMDGPU_PTE_PRT_FLAG(adev) && prt_fini_needed) {
2632 amdgpu_vm_prt_fini(adev, vm);
2633 prt_fini_needed = false;
2634 }
2635
2636 list_del(&mapping->list);
2637 amdgpu_vm_free_mapping(adev, vm, mapping, NULL);
2638 }
2639
2640 amdgpu_vm_pt_free_root(adev, vm);
2641 amdgpu_bo_unreserve(root);
2642 amdgpu_bo_unref(&root);
2643 WARN_ON(vm->root.bo);
2644
2645 amdgpu_vm_fini_entities(vm);
2646
2647 if (!RB_EMPTY_ROOT(&vm->va.rb_root)) {
2648 dev_err(adev->dev, "still active bo inside vm\n");
2649 }
2650 rbtree_postorder_for_each_entry_safe(mapping, tmp,
2651 &vm->va.rb_root, rb) {
2652 /* Don't remove the mapping here, we don't want to trigger a
2653 * rebalance and the tree is about to be destroyed anyway.
2654 */
2655 list_del(&mapping->list);
2656 kfree(mapping);
2657 }
2658
2659 dma_fence_put(vm->last_update);
2660
2661 for (i = 0; i < AMDGPU_MAX_VMHUBS; i++) {
2662 if (vm->reserved_vmid[i]) {
2663 amdgpu_vmid_free_reserved(adev, i);
2664 vm->reserved_vmid[i] = false;
2665 }
2666 }
2667
2668 ttm_lru_bulk_move_fini(&adev->mman.bdev, &vm->lru_bulk_move);
2669}
2670
2671/**
2672 * amdgpu_vm_manager_init - init the VM manager
2673 *
2674 * @adev: amdgpu_device pointer
2675 *
2676 * Initialize the VM manager structures
2677 */
2678void amdgpu_vm_manager_init(struct amdgpu_device *adev)
2679{
2680 unsigned i;
2681
2682 /* Concurrent flushes are only possible starting with Vega10 and
2683 * are broken on Navi10 and Navi14.
2684 */
2685 adev->vm_manager.concurrent_flush = !(adev->asic_type < CHIP_VEGA10 ||
2686 adev->asic_type == CHIP_NAVI10 ||
2687 adev->asic_type == CHIP_NAVI14);
2688 amdgpu_vmid_mgr_init(adev);
2689
2690 adev->vm_manager.fence_context =
2691 dma_fence_context_alloc(AMDGPU_MAX_RINGS);
2692 for (i = 0; i < AMDGPU_MAX_RINGS; ++i)
2693 adev->vm_manager.seqno[i] = 0;
2694
2695 spin_lock_init(&adev->vm_manager.prt_lock);
2696 atomic_set(&adev->vm_manager.num_prt_users, 0);
2697
2698 /* If not overridden by the user, by default, only in large BAR systems
2699 * Compute VM tables will be updated by CPU
2700 */
2701#ifdef CONFIG_X86_64
2702 if (amdgpu_vm_update_mode == -1) {
2703 /* For asic with VF MMIO access protection
2704 * avoid using CPU for VM table updates
2705 */
2706 if (amdgpu_gmc_vram_full_visible(&adev->gmc) &&
2707 !amdgpu_sriov_vf_mmio_access_protection(adev))
2708 adev->vm_manager.vm_update_mode =
2709 AMDGPU_VM_USE_CPU_FOR_COMPUTE;
2710 else
2711 adev->vm_manager.vm_update_mode = 0;
2712 } else
2713 adev->vm_manager.vm_update_mode = amdgpu_vm_update_mode;
2714#else
2715 adev->vm_manager.vm_update_mode = 0;
2716#endif
2717
2718 xa_init_flags(&adev->vm_manager.pasids, XA_FLAGS_LOCK_IRQ);
2719}
2720
2721/**
2722 * amdgpu_vm_manager_fini - cleanup VM manager
2723 *
2724 * @adev: amdgpu_device pointer
2725 *
2726 * Cleanup the VM manager and free resources.
2727 */
2728void amdgpu_vm_manager_fini(struct amdgpu_device *adev)
2729{
2730 WARN_ON(!xa_empty(&adev->vm_manager.pasids));
2731 xa_destroy(&adev->vm_manager.pasids);
2732
2733 amdgpu_vmid_mgr_fini(adev);
2734}
2735
2736/**
2737 * amdgpu_vm_ioctl - Manages VMID reservation for vm hubs.
2738 *
2739 * @dev: drm device pointer
2740 * @data: drm_amdgpu_vm
2741 * @filp: drm file pointer
2742 *
2743 * Returns:
2744 * 0 for success, -errno for errors.
2745 */
2746int amdgpu_vm_ioctl(struct drm_device *dev, void *data, struct drm_file *filp)
2747{
2748 union drm_amdgpu_vm *args = data;
2749 struct amdgpu_device *adev = drm_to_adev(dev);
2750 struct amdgpu_fpriv *fpriv = filp->driver_priv;
2751
2752 /* No valid flags defined yet */
2753 if (args->in.flags)
2754 return -EINVAL;
2755
2756 switch (args->in.op) {
2757 case AMDGPU_VM_OP_RESERVE_VMID:
2758 /* We only have requirement to reserve vmid from gfxhub */
2759 if (!fpriv->vm.reserved_vmid[AMDGPU_GFXHUB(0)]) {
2760 amdgpu_vmid_alloc_reserved(adev, AMDGPU_GFXHUB(0));
2761 fpriv->vm.reserved_vmid[AMDGPU_GFXHUB(0)] = true;
2762 }
2763
2764 break;
2765 case AMDGPU_VM_OP_UNRESERVE_VMID:
2766 if (fpriv->vm.reserved_vmid[AMDGPU_GFXHUB(0)]) {
2767 amdgpu_vmid_free_reserved(adev, AMDGPU_GFXHUB(0));
2768 fpriv->vm.reserved_vmid[AMDGPU_GFXHUB(0)] = false;
2769 }
2770 break;
2771 default:
2772 return -EINVAL;
2773 }
2774
2775 return 0;
2776}
2777
2778/**
2779 * amdgpu_vm_handle_fault - graceful handling of VM faults.
2780 * @adev: amdgpu device pointer
2781 * @pasid: PASID of the VM
2782 * @ts: Timestamp of the fault
2783 * @vmid: VMID, only used for GFX 9.4.3.
2784 * @node_id: Node_id received in IH cookie. Only applicable for
2785 * GFX 9.4.3.
2786 * @addr: Address of the fault
2787 * @write_fault: true is write fault, false is read fault
2788 *
2789 * Try to gracefully handle a VM fault. Return true if the fault was handled and
2790 * shouldn't be reported any more.
2791 */
2792bool amdgpu_vm_handle_fault(struct amdgpu_device *adev, u32 pasid,
2793 u32 vmid, u32 node_id, uint64_t addr, uint64_t ts,
2794 bool write_fault)
2795{
2796 bool is_compute_context = false;
2797 struct amdgpu_bo *root;
2798 unsigned long irqflags;
2799 uint64_t value, flags;
2800 struct amdgpu_vm *vm;
2801 int r;
2802
2803 xa_lock_irqsave(&adev->vm_manager.pasids, irqflags);
2804 vm = xa_load(&adev->vm_manager.pasids, pasid);
2805 if (vm) {
2806 root = amdgpu_bo_ref(vm->root.bo);
2807 is_compute_context = vm->is_compute_context;
2808 } else {
2809 root = NULL;
2810 }
2811 xa_unlock_irqrestore(&adev->vm_manager.pasids, irqflags);
2812
2813 if (!root)
2814 return false;
2815
2816 addr /= AMDGPU_GPU_PAGE_SIZE;
2817
2818 if (is_compute_context && !svm_range_restore_pages(adev, pasid, vmid,
2819 node_id, addr, ts, write_fault)) {
2820 amdgpu_bo_unref(&root);
2821 return true;
2822 }
2823
2824 r = amdgpu_bo_reserve(root, true);
2825 if (r)
2826 goto error_unref;
2827
2828 /* Double check that the VM still exists */
2829 xa_lock_irqsave(&adev->vm_manager.pasids, irqflags);
2830 vm = xa_load(&adev->vm_manager.pasids, pasid);
2831 if (vm && vm->root.bo != root)
2832 vm = NULL;
2833 xa_unlock_irqrestore(&adev->vm_manager.pasids, irqflags);
2834 if (!vm)
2835 goto error_unlock;
2836
2837 flags = AMDGPU_PTE_VALID | AMDGPU_PTE_SNOOPED |
2838 AMDGPU_PTE_SYSTEM;
2839
2840 if (is_compute_context) {
2841 /* Intentionally setting invalid PTE flag
2842 * combination to force a no-retry-fault
2843 */
2844 flags = AMDGPU_VM_NORETRY_FLAGS;
2845 value = 0;
2846 } else if (amdgpu_vm_fault_stop == AMDGPU_VM_FAULT_STOP_NEVER) {
2847 /* Redirect the access to the dummy page */
2848 value = adev->dummy_page_addr;
2849 flags |= AMDGPU_PTE_EXECUTABLE | AMDGPU_PTE_READABLE |
2850 AMDGPU_PTE_WRITEABLE;
2851
2852 } else {
2853 /* Let the hw retry silently on the PTE */
2854 value = 0;
2855 }
2856
2857 r = dma_resv_reserve_fences(root->tbo.base.resv, 1);
2858 if (r) {
2859 pr_debug("failed %d to reserve fence slot\n", r);
2860 goto error_unlock;
2861 }
2862
2863 r = amdgpu_vm_update_range(adev, vm, true, false, false, false,
2864 NULL, addr, addr, flags, value, 0, NULL, NULL, NULL);
2865 if (r)
2866 goto error_unlock;
2867
2868 r = amdgpu_vm_update_pdes(adev, vm, true);
2869
2870error_unlock:
2871 amdgpu_bo_unreserve(root);
2872 if (r < 0)
2873 DRM_ERROR("Can't handle page fault (%d)\n", r);
2874
2875error_unref:
2876 amdgpu_bo_unref(&root);
2877
2878 return false;
2879}
2880
2881#if defined(CONFIG_DEBUG_FS)
2882/**
2883 * amdgpu_debugfs_vm_bo_info - print BO info for the VM
2884 *
2885 * @vm: Requested VM for printing BO info
2886 * @m: debugfs file
2887 *
2888 * Print BO information in debugfs file for the VM
2889 */
2890void amdgpu_debugfs_vm_bo_info(struct amdgpu_vm *vm, struct seq_file *m)
2891{
2892 struct amdgpu_bo_va *bo_va, *tmp;
2893 u64 total_idle = 0;
2894 u64 total_evicted = 0;
2895 u64 total_relocated = 0;
2896 u64 total_moved = 0;
2897 u64 total_invalidated = 0;
2898 u64 total_done = 0;
2899 unsigned int total_idle_objs = 0;
2900 unsigned int total_evicted_objs = 0;
2901 unsigned int total_relocated_objs = 0;
2902 unsigned int total_moved_objs = 0;
2903 unsigned int total_invalidated_objs = 0;
2904 unsigned int total_done_objs = 0;
2905 unsigned int id = 0;
2906
2907 spin_lock(&vm->status_lock);
2908 seq_puts(m, "\tIdle BOs:\n");
2909 list_for_each_entry_safe(bo_va, tmp, &vm->idle, base.vm_status) {
2910 if (!bo_va->base.bo)
2911 continue;
2912 total_idle += amdgpu_bo_print_info(id++, bo_va->base.bo, m);
2913 }
2914 total_idle_objs = id;
2915 id = 0;
2916
2917 seq_puts(m, "\tEvicted BOs:\n");
2918 list_for_each_entry_safe(bo_va, tmp, &vm->evicted, base.vm_status) {
2919 if (!bo_va->base.bo)
2920 continue;
2921 total_evicted += amdgpu_bo_print_info(id++, bo_va->base.bo, m);
2922 }
2923 total_evicted_objs = id;
2924 id = 0;
2925
2926 seq_puts(m, "\tRelocated BOs:\n");
2927 list_for_each_entry_safe(bo_va, tmp, &vm->relocated, base.vm_status) {
2928 if (!bo_va->base.bo)
2929 continue;
2930 total_relocated += amdgpu_bo_print_info(id++, bo_va->base.bo, m);
2931 }
2932 total_relocated_objs = id;
2933 id = 0;
2934
2935 seq_puts(m, "\tMoved BOs:\n");
2936 list_for_each_entry_safe(bo_va, tmp, &vm->moved, base.vm_status) {
2937 if (!bo_va->base.bo)
2938 continue;
2939 total_moved += amdgpu_bo_print_info(id++, bo_va->base.bo, m);
2940 }
2941 total_moved_objs = id;
2942 id = 0;
2943
2944 seq_puts(m, "\tInvalidated BOs:\n");
2945 list_for_each_entry_safe(bo_va, tmp, &vm->invalidated, base.vm_status) {
2946 if (!bo_va->base.bo)
2947 continue;
2948 total_invalidated += amdgpu_bo_print_info(id++, bo_va->base.bo, m);
2949 }
2950 total_invalidated_objs = id;
2951 id = 0;
2952
2953 seq_puts(m, "\tDone BOs:\n");
2954 list_for_each_entry_safe(bo_va, tmp, &vm->done, base.vm_status) {
2955 if (!bo_va->base.bo)
2956 continue;
2957 total_done += amdgpu_bo_print_info(id++, bo_va->base.bo, m);
2958 }
2959 spin_unlock(&vm->status_lock);
2960 total_done_objs = id;
2961
2962 seq_printf(m, "\tTotal idle size: %12lld\tobjs:\t%d\n", total_idle,
2963 total_idle_objs);
2964 seq_printf(m, "\tTotal evicted size: %12lld\tobjs:\t%d\n", total_evicted,
2965 total_evicted_objs);
2966 seq_printf(m, "\tTotal relocated size: %12lld\tobjs:\t%d\n", total_relocated,
2967 total_relocated_objs);
2968 seq_printf(m, "\tTotal moved size: %12lld\tobjs:\t%d\n", total_moved,
2969 total_moved_objs);
2970 seq_printf(m, "\tTotal invalidated size: %12lld\tobjs:\t%d\n", total_invalidated,
2971 total_invalidated_objs);
2972 seq_printf(m, "\tTotal done size: %12lld\tobjs:\t%d\n", total_done,
2973 total_done_objs);
2974}
2975#endif
2976
2977/**
2978 * amdgpu_vm_update_fault_cache - update cached fault into.
2979 * @adev: amdgpu device pointer
2980 * @pasid: PASID of the VM
2981 * @addr: Address of the fault
2982 * @status: GPUVM fault status register
2983 * @vmhub: which vmhub got the fault
2984 *
2985 * Cache the fault info for later use by userspace in debugging.
2986 */
2987void amdgpu_vm_update_fault_cache(struct amdgpu_device *adev,
2988 unsigned int pasid,
2989 uint64_t addr,
2990 uint32_t status,
2991 unsigned int vmhub)
2992{
2993 struct amdgpu_vm *vm;
2994 unsigned long flags;
2995
2996 xa_lock_irqsave(&adev->vm_manager.pasids, flags);
2997
2998 vm = xa_load(&adev->vm_manager.pasids, pasid);
2999 /* Don't update the fault cache if status is 0. In the multiple
3000 * fault case, subsequent faults will return a 0 status which is
3001 * useless for userspace and replaces the useful fault status, so
3002 * only update if status is non-0.
3003 */
3004 if (vm && status) {
3005 vm->fault_info.addr = addr;
3006 vm->fault_info.status = status;
3007 /*
3008 * Update the fault information globally for later usage
3009 * when vm could be stale or freed.
3010 */
3011 adev->vm_manager.fault_info.addr = addr;
3012 adev->vm_manager.fault_info.vmhub = vmhub;
3013 adev->vm_manager.fault_info.status = status;
3014
3015 if (AMDGPU_IS_GFXHUB(vmhub)) {
3016 vm->fault_info.vmhub = AMDGPU_VMHUB_TYPE_GFX;
3017 vm->fault_info.vmhub |=
3018 (vmhub - AMDGPU_GFXHUB_START) << AMDGPU_VMHUB_IDX_SHIFT;
3019 } else if (AMDGPU_IS_MMHUB0(vmhub)) {
3020 vm->fault_info.vmhub = AMDGPU_VMHUB_TYPE_MM0;
3021 vm->fault_info.vmhub |=
3022 (vmhub - AMDGPU_MMHUB0_START) << AMDGPU_VMHUB_IDX_SHIFT;
3023 } else if (AMDGPU_IS_MMHUB1(vmhub)) {
3024 vm->fault_info.vmhub = AMDGPU_VMHUB_TYPE_MM1;
3025 vm->fault_info.vmhub |=
3026 (vmhub - AMDGPU_MMHUB1_START) << AMDGPU_VMHUB_IDX_SHIFT;
3027 } else {
3028 WARN_ONCE(1, "Invalid vmhub %u\n", vmhub);
3029 }
3030 }
3031 xa_unlock_irqrestore(&adev->vm_manager.pasids, flags);
3032}
3033
3034/**
3035 * amdgpu_vm_is_bo_always_valid - check if the BO is VM always valid
3036 *
3037 * @vm: VM to test against.
3038 * @bo: BO to be tested.
3039 *
3040 * Returns true if the BO shares the dma_resv object with the root PD and is
3041 * always guaranteed to be valid inside the VM.
3042 */
3043bool amdgpu_vm_is_bo_always_valid(struct amdgpu_vm *vm, struct amdgpu_bo *bo)
3044{
3045 return bo && bo->tbo.base.resv == vm->root.bo->tbo.base.resv;
3046}
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#include <drm/drmP.h>
32#include <drm/amdgpu_drm.h>
33#include "amdgpu.h"
34#include "amdgpu_trace.h"
35#include "amdgpu_amdkfd.h"
36
37/*
38 * GPUVM
39 * GPUVM is similar to the legacy gart on older asics, however
40 * rather than there being a single global gart table
41 * for the entire GPU, there are multiple VM page tables active
42 * at any given time. The VM page tables can contain a mix
43 * vram pages and system memory pages and system memory pages
44 * can be mapped as snooped (cached system pages) or unsnooped
45 * (uncached system pages).
46 * Each VM has an ID associated with it and there is a page table
47 * associated with each VMID. When execting a command buffer,
48 * the kernel tells the the ring what VMID to use for that command
49 * buffer. VMIDs are allocated dynamically as commands are submitted.
50 * The userspace drivers maintain their own address space and the kernel
51 * sets up their pages tables accordingly when they submit their
52 * command buffers and a VMID is assigned.
53 * Cayman/Trinity support up to 8 active VMs at any given time;
54 * SI supports 16.
55 */
56
57#define START(node) ((node)->start)
58#define LAST(node) ((node)->last)
59
60INTERVAL_TREE_DEFINE(struct amdgpu_bo_va_mapping, rb, uint64_t, __subtree_last,
61 START, LAST, static, amdgpu_vm_it)
62
63#undef START
64#undef LAST
65
66/* Local structure. Encapsulate some VM table update parameters to reduce
67 * the number of function parameters
68 */
69struct amdgpu_pte_update_params {
70 /* amdgpu device we do this update for */
71 struct amdgpu_device *adev;
72 /* optional amdgpu_vm we do this update for */
73 struct amdgpu_vm *vm;
74 /* address where to copy page table entries from */
75 uint64_t src;
76 /* indirect buffer to fill with commands */
77 struct amdgpu_ib *ib;
78 /* Function which actually does the update */
79 void (*func)(struct amdgpu_pte_update_params *params,
80 struct amdgpu_bo *bo, uint64_t pe,
81 uint64_t addr, unsigned count, uint32_t incr,
82 uint64_t flags);
83 /* The next two are used during VM update by CPU
84 * DMA addresses to use for mapping
85 * Kernel pointer of PD/PT BO that needs to be updated
86 */
87 dma_addr_t *pages_addr;
88 void *kptr;
89};
90
91/* Helper to disable partial resident texture feature from a fence callback */
92struct amdgpu_prt_cb {
93 struct amdgpu_device *adev;
94 struct dma_fence_cb cb;
95};
96
97/**
98 * amdgpu_vm_level_shift - return the addr shift for each level
99 *
100 * @adev: amdgpu_device pointer
101 *
102 * Returns the number of bits the pfn needs to be right shifted for a level.
103 */
104static unsigned amdgpu_vm_level_shift(struct amdgpu_device *adev,
105 unsigned level)
106{
107 unsigned shift = 0xff;
108
109 switch (level) {
110 case AMDGPU_VM_PDB2:
111 case AMDGPU_VM_PDB1:
112 case AMDGPU_VM_PDB0:
113 shift = 9 * (AMDGPU_VM_PDB0 - level) +
114 adev->vm_manager.block_size;
115 break;
116 case AMDGPU_VM_PTB:
117 shift = 0;
118 break;
119 default:
120 dev_err(adev->dev, "the level%d isn't supported.\n", level);
121 }
122
123 return shift;
124}
125
126/**
127 * amdgpu_vm_num_entries - return the number of entries in a PD/PT
128 *
129 * @adev: amdgpu_device pointer
130 *
131 * Calculate the number of entries in a page directory or page table.
132 */
133static unsigned amdgpu_vm_num_entries(struct amdgpu_device *adev,
134 unsigned level)
135{
136 unsigned shift = amdgpu_vm_level_shift(adev,
137 adev->vm_manager.root_level);
138
139 if (level == adev->vm_manager.root_level)
140 /* For the root directory */
141 return round_up(adev->vm_manager.max_pfn, 1 << shift) >> shift;
142 else if (level != AMDGPU_VM_PTB)
143 /* Everything in between */
144 return 512;
145 else
146 /* For the page tables on the leaves */
147 return AMDGPU_VM_PTE_COUNT(adev);
148}
149
150/**
151 * amdgpu_vm_bo_size - returns the size of the BOs in bytes
152 *
153 * @adev: amdgpu_device pointer
154 *
155 * Calculate the size of the BO for a page directory or page table in bytes.
156 */
157static unsigned amdgpu_vm_bo_size(struct amdgpu_device *adev, unsigned level)
158{
159 return AMDGPU_GPU_PAGE_ALIGN(amdgpu_vm_num_entries(adev, level) * 8);
160}
161
162/**
163 * amdgpu_vm_get_pd_bo - add the VM PD to a validation list
164 *
165 * @vm: vm providing the BOs
166 * @validated: head of validation list
167 * @entry: entry to add
168 *
169 * Add the page directory to the list of BOs to
170 * validate for command submission.
171 */
172void amdgpu_vm_get_pd_bo(struct amdgpu_vm *vm,
173 struct list_head *validated,
174 struct amdgpu_bo_list_entry *entry)
175{
176 entry->robj = vm->root.base.bo;
177 entry->priority = 0;
178 entry->tv.bo = &entry->robj->tbo;
179 entry->tv.shared = true;
180 entry->user_pages = NULL;
181 list_add(&entry->tv.head, validated);
182}
183
184/**
185 * amdgpu_vm_validate_pt_bos - validate the page table BOs
186 *
187 * @adev: amdgpu device pointer
188 * @vm: vm providing the BOs
189 * @validate: callback to do the validation
190 * @param: parameter for the validation callback
191 *
192 * Validate the page table BOs on command submission if neccessary.
193 */
194int amdgpu_vm_validate_pt_bos(struct amdgpu_device *adev, struct amdgpu_vm *vm,
195 int (*validate)(void *p, struct amdgpu_bo *bo),
196 void *param)
197{
198 struct ttm_bo_global *glob = adev->mman.bdev.glob;
199 int r;
200
201 spin_lock(&vm->status_lock);
202 while (!list_empty(&vm->evicted)) {
203 struct amdgpu_vm_bo_base *bo_base;
204 struct amdgpu_bo *bo;
205
206 bo_base = list_first_entry(&vm->evicted,
207 struct amdgpu_vm_bo_base,
208 vm_status);
209 spin_unlock(&vm->status_lock);
210
211 bo = bo_base->bo;
212 BUG_ON(!bo);
213 if (bo->parent) {
214 r = validate(param, bo);
215 if (r)
216 return r;
217
218 spin_lock(&glob->lru_lock);
219 ttm_bo_move_to_lru_tail(&bo->tbo);
220 if (bo->shadow)
221 ttm_bo_move_to_lru_tail(&bo->shadow->tbo);
222 spin_unlock(&glob->lru_lock);
223 }
224
225 if (bo->tbo.type == ttm_bo_type_kernel &&
226 vm->use_cpu_for_update) {
227 r = amdgpu_bo_kmap(bo, NULL);
228 if (r)
229 return r;
230 }
231
232 spin_lock(&vm->status_lock);
233 if (bo->tbo.type != ttm_bo_type_kernel)
234 list_move(&bo_base->vm_status, &vm->moved);
235 else
236 list_move(&bo_base->vm_status, &vm->relocated);
237 }
238 spin_unlock(&vm->status_lock);
239
240 return 0;
241}
242
243/**
244 * amdgpu_vm_ready - check VM is ready for updates
245 *
246 * @vm: VM to check
247 *
248 * Check if all VM PDs/PTs are ready for updates
249 */
250bool amdgpu_vm_ready(struct amdgpu_vm *vm)
251{
252 bool ready;
253
254 spin_lock(&vm->status_lock);
255 ready = list_empty(&vm->evicted);
256 spin_unlock(&vm->status_lock);
257
258 return ready;
259}
260
261/**
262 * amdgpu_vm_clear_bo - initially clear the PDs/PTs
263 *
264 * @adev: amdgpu_device pointer
265 * @bo: BO to clear
266 * @level: level this BO is at
267 *
268 * Root PD needs to be reserved when calling this.
269 */
270static int amdgpu_vm_clear_bo(struct amdgpu_device *adev,
271 struct amdgpu_vm *vm, struct amdgpu_bo *bo,
272 unsigned level, bool pte_support_ats)
273{
274 struct ttm_operation_ctx ctx = { true, false };
275 struct dma_fence *fence = NULL;
276 unsigned entries, ats_entries;
277 struct amdgpu_ring *ring;
278 struct amdgpu_job *job;
279 uint64_t addr;
280 int r;
281
282 addr = amdgpu_bo_gpu_offset(bo);
283 entries = amdgpu_bo_size(bo) / 8;
284
285 if (pte_support_ats) {
286 if (level == adev->vm_manager.root_level) {
287 ats_entries = amdgpu_vm_level_shift(adev, level);
288 ats_entries += AMDGPU_GPU_PAGE_SHIFT;
289 ats_entries = AMDGPU_VA_HOLE_START >> ats_entries;
290 ats_entries = min(ats_entries, entries);
291 entries -= ats_entries;
292 } else {
293 ats_entries = entries;
294 entries = 0;
295 }
296 } else {
297 ats_entries = 0;
298 }
299
300 ring = container_of(vm->entity.sched, struct amdgpu_ring, sched);
301
302 r = reservation_object_reserve_shared(bo->tbo.resv);
303 if (r)
304 return r;
305
306 r = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
307 if (r)
308 goto error;
309
310 r = amdgpu_job_alloc_with_ib(adev, 64, &job);
311 if (r)
312 goto error;
313
314 if (ats_entries) {
315 uint64_t ats_value;
316
317 ats_value = AMDGPU_PTE_DEFAULT_ATC;
318 if (level != AMDGPU_VM_PTB)
319 ats_value |= AMDGPU_PDE_PTE;
320
321 amdgpu_vm_set_pte_pde(adev, &job->ibs[0], addr, 0,
322 ats_entries, 0, ats_value);
323 addr += ats_entries * 8;
324 }
325
326 if (entries)
327 amdgpu_vm_set_pte_pde(adev, &job->ibs[0], addr, 0,
328 entries, 0, 0);
329
330 amdgpu_ring_pad_ib(ring, &job->ibs[0]);
331
332 WARN_ON(job->ibs[0].length_dw > 64);
333 r = amdgpu_sync_resv(adev, &job->sync, bo->tbo.resv,
334 AMDGPU_FENCE_OWNER_UNDEFINED, false);
335 if (r)
336 goto error_free;
337
338 r = amdgpu_job_submit(job, ring, &vm->entity,
339 AMDGPU_FENCE_OWNER_UNDEFINED, &fence);
340 if (r)
341 goto error_free;
342
343 amdgpu_bo_fence(bo, fence, true);
344 dma_fence_put(fence);
345
346 if (bo->shadow)
347 return amdgpu_vm_clear_bo(adev, vm, bo->shadow,
348 level, pte_support_ats);
349
350 return 0;
351
352error_free:
353 amdgpu_job_free(job);
354
355error:
356 return r;
357}
358
359/**
360 * amdgpu_vm_alloc_levels - allocate the PD/PT levels
361 *
362 * @adev: amdgpu_device pointer
363 * @vm: requested vm
364 * @saddr: start of the address range
365 * @eaddr: end of the address range
366 *
367 * Make sure the page directories and page tables are allocated
368 */
369static int amdgpu_vm_alloc_levels(struct amdgpu_device *adev,
370 struct amdgpu_vm *vm,
371 struct amdgpu_vm_pt *parent,
372 uint64_t saddr, uint64_t eaddr,
373 unsigned level, bool ats)
374{
375 unsigned shift = amdgpu_vm_level_shift(adev, level);
376 unsigned pt_idx, from, to;
377 u64 flags;
378 int r;
379
380 if (!parent->entries) {
381 unsigned num_entries = amdgpu_vm_num_entries(adev, level);
382
383 parent->entries = kvmalloc_array(num_entries,
384 sizeof(struct amdgpu_vm_pt),
385 GFP_KERNEL | __GFP_ZERO);
386 if (!parent->entries)
387 return -ENOMEM;
388 memset(parent->entries, 0 , sizeof(struct amdgpu_vm_pt));
389 }
390
391 from = saddr >> shift;
392 to = eaddr >> shift;
393 if (from >= amdgpu_vm_num_entries(adev, level) ||
394 to >= amdgpu_vm_num_entries(adev, level))
395 return -EINVAL;
396
397 ++level;
398 saddr = saddr & ((1 << shift) - 1);
399 eaddr = eaddr & ((1 << shift) - 1);
400
401 flags = AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS;
402 if (vm->use_cpu_for_update)
403 flags |= AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED;
404 else
405 flags |= (AMDGPU_GEM_CREATE_NO_CPU_ACCESS |
406 AMDGPU_GEM_CREATE_SHADOW);
407
408 /* walk over the address space and allocate the page tables */
409 for (pt_idx = from; pt_idx <= to; ++pt_idx) {
410 struct reservation_object *resv = vm->root.base.bo->tbo.resv;
411 struct amdgpu_vm_pt *entry = &parent->entries[pt_idx];
412 struct amdgpu_bo *pt;
413
414 if (!entry->base.bo) {
415 r = amdgpu_bo_create(adev,
416 amdgpu_vm_bo_size(adev, level),
417 AMDGPU_GPU_PAGE_SIZE,
418 AMDGPU_GEM_DOMAIN_VRAM, flags,
419 ttm_bo_type_kernel, resv, &pt);
420 if (r)
421 return r;
422
423 r = amdgpu_vm_clear_bo(adev, vm, pt, level, ats);
424 if (r) {
425 amdgpu_bo_unref(&pt->shadow);
426 amdgpu_bo_unref(&pt);
427 return r;
428 }
429
430 if (vm->use_cpu_for_update) {
431 r = amdgpu_bo_kmap(pt, NULL);
432 if (r) {
433 amdgpu_bo_unref(&pt->shadow);
434 amdgpu_bo_unref(&pt);
435 return r;
436 }
437 }
438
439 /* Keep a reference to the root directory to avoid
440 * freeing them up in the wrong order.
441 */
442 pt->parent = amdgpu_bo_ref(parent->base.bo);
443
444 entry->base.vm = vm;
445 entry->base.bo = pt;
446 list_add_tail(&entry->base.bo_list, &pt->va);
447 spin_lock(&vm->status_lock);
448 list_add(&entry->base.vm_status, &vm->relocated);
449 spin_unlock(&vm->status_lock);
450 }
451
452 if (level < AMDGPU_VM_PTB) {
453 uint64_t sub_saddr = (pt_idx == from) ? saddr : 0;
454 uint64_t sub_eaddr = (pt_idx == to) ? eaddr :
455 ((1 << shift) - 1);
456 r = amdgpu_vm_alloc_levels(adev, vm, entry, sub_saddr,
457 sub_eaddr, level, ats);
458 if (r)
459 return r;
460 }
461 }
462
463 return 0;
464}
465
466/**
467 * amdgpu_vm_alloc_pts - Allocate page tables.
468 *
469 * @adev: amdgpu_device pointer
470 * @vm: VM to allocate page tables for
471 * @saddr: Start address which needs to be allocated
472 * @size: Size from start address we need.
473 *
474 * Make sure the page tables are allocated.
475 */
476int amdgpu_vm_alloc_pts(struct amdgpu_device *adev,
477 struct amdgpu_vm *vm,
478 uint64_t saddr, uint64_t size)
479{
480 uint64_t eaddr;
481 bool ats = false;
482
483 /* validate the parameters */
484 if (saddr & AMDGPU_GPU_PAGE_MASK || size & AMDGPU_GPU_PAGE_MASK)
485 return -EINVAL;
486
487 eaddr = saddr + size - 1;
488
489 if (vm->pte_support_ats)
490 ats = saddr < AMDGPU_VA_HOLE_START;
491
492 saddr /= AMDGPU_GPU_PAGE_SIZE;
493 eaddr /= AMDGPU_GPU_PAGE_SIZE;
494
495 if (eaddr >= adev->vm_manager.max_pfn) {
496 dev_err(adev->dev, "va above limit (0x%08llX >= 0x%08llX)\n",
497 eaddr, adev->vm_manager.max_pfn);
498 return -EINVAL;
499 }
500
501 return amdgpu_vm_alloc_levels(adev, vm, &vm->root, saddr, eaddr,
502 adev->vm_manager.root_level, ats);
503}
504
505/**
506 * amdgpu_vm_check_compute_bug - check whether asic has compute vm bug
507 *
508 * @adev: amdgpu_device pointer
509 */
510void amdgpu_vm_check_compute_bug(struct amdgpu_device *adev)
511{
512 const struct amdgpu_ip_block *ip_block;
513 bool has_compute_vm_bug;
514 struct amdgpu_ring *ring;
515 int i;
516
517 has_compute_vm_bug = false;
518
519 ip_block = amdgpu_device_ip_get_ip_block(adev, AMD_IP_BLOCK_TYPE_GFX);
520 if (ip_block) {
521 /* Compute has a VM bug for GFX version < 7.
522 Compute has a VM bug for GFX 8 MEC firmware version < 673.*/
523 if (ip_block->version->major <= 7)
524 has_compute_vm_bug = true;
525 else if (ip_block->version->major == 8)
526 if (adev->gfx.mec_fw_version < 673)
527 has_compute_vm_bug = true;
528 }
529
530 for (i = 0; i < adev->num_rings; i++) {
531 ring = adev->rings[i];
532 if (ring->funcs->type == AMDGPU_RING_TYPE_COMPUTE)
533 /* only compute rings */
534 ring->has_compute_vm_bug = has_compute_vm_bug;
535 else
536 ring->has_compute_vm_bug = false;
537 }
538}
539
540bool amdgpu_vm_need_pipeline_sync(struct amdgpu_ring *ring,
541 struct amdgpu_job *job)
542{
543 struct amdgpu_device *adev = ring->adev;
544 unsigned vmhub = ring->funcs->vmhub;
545 struct amdgpu_vmid_mgr *id_mgr = &adev->vm_manager.id_mgr[vmhub];
546 struct amdgpu_vmid *id;
547 bool gds_switch_needed;
548 bool vm_flush_needed = job->vm_needs_flush || ring->has_compute_vm_bug;
549
550 if (job->vmid == 0)
551 return false;
552 id = &id_mgr->ids[job->vmid];
553 gds_switch_needed = ring->funcs->emit_gds_switch && (
554 id->gds_base != job->gds_base ||
555 id->gds_size != job->gds_size ||
556 id->gws_base != job->gws_base ||
557 id->gws_size != job->gws_size ||
558 id->oa_base != job->oa_base ||
559 id->oa_size != job->oa_size);
560
561 if (amdgpu_vmid_had_gpu_reset(adev, id))
562 return true;
563
564 return vm_flush_needed || gds_switch_needed;
565}
566
567static bool amdgpu_vm_is_large_bar(struct amdgpu_device *adev)
568{
569 return (adev->gmc.real_vram_size == adev->gmc.visible_vram_size);
570}
571
572/**
573 * amdgpu_vm_flush - hardware flush the vm
574 *
575 * @ring: ring to use for flush
576 * @vmid: vmid number to use
577 * @pd_addr: address of the page directory
578 *
579 * Emit a VM flush when it is necessary.
580 */
581int amdgpu_vm_flush(struct amdgpu_ring *ring, struct amdgpu_job *job, bool need_pipe_sync)
582{
583 struct amdgpu_device *adev = ring->adev;
584 unsigned vmhub = ring->funcs->vmhub;
585 struct amdgpu_vmid_mgr *id_mgr = &adev->vm_manager.id_mgr[vmhub];
586 struct amdgpu_vmid *id = &id_mgr->ids[job->vmid];
587 bool gds_switch_needed = ring->funcs->emit_gds_switch && (
588 id->gds_base != job->gds_base ||
589 id->gds_size != job->gds_size ||
590 id->gws_base != job->gws_base ||
591 id->gws_size != job->gws_size ||
592 id->oa_base != job->oa_base ||
593 id->oa_size != job->oa_size);
594 bool vm_flush_needed = job->vm_needs_flush;
595 bool pasid_mapping_needed = id->pasid != job->pasid ||
596 !id->pasid_mapping ||
597 !dma_fence_is_signaled(id->pasid_mapping);
598 struct dma_fence *fence = NULL;
599 unsigned patch_offset = 0;
600 int r;
601
602 if (amdgpu_vmid_had_gpu_reset(adev, id)) {
603 gds_switch_needed = true;
604 vm_flush_needed = true;
605 pasid_mapping_needed = true;
606 }
607
608 gds_switch_needed &= !!ring->funcs->emit_gds_switch;
609 vm_flush_needed &= !!ring->funcs->emit_vm_flush;
610 pasid_mapping_needed &= adev->gmc.gmc_funcs->emit_pasid_mapping &&
611 ring->funcs->emit_wreg;
612
613 if (!vm_flush_needed && !gds_switch_needed && !need_pipe_sync)
614 return 0;
615
616 if (ring->funcs->init_cond_exec)
617 patch_offset = amdgpu_ring_init_cond_exec(ring);
618
619 if (need_pipe_sync)
620 amdgpu_ring_emit_pipeline_sync(ring);
621
622 if (vm_flush_needed) {
623 trace_amdgpu_vm_flush(ring, job->vmid, job->vm_pd_addr);
624 amdgpu_ring_emit_vm_flush(ring, job->vmid, job->vm_pd_addr);
625 }
626
627 if (pasid_mapping_needed)
628 amdgpu_gmc_emit_pasid_mapping(ring, job->vmid, job->pasid);
629
630 if (vm_flush_needed || pasid_mapping_needed) {
631 r = amdgpu_fence_emit(ring, &fence);
632 if (r)
633 return r;
634 }
635
636 if (vm_flush_needed) {
637 mutex_lock(&id_mgr->lock);
638 dma_fence_put(id->last_flush);
639 id->last_flush = dma_fence_get(fence);
640 id->current_gpu_reset_count =
641 atomic_read(&adev->gpu_reset_counter);
642 mutex_unlock(&id_mgr->lock);
643 }
644
645 if (pasid_mapping_needed) {
646 id->pasid = job->pasid;
647 dma_fence_put(id->pasid_mapping);
648 id->pasid_mapping = dma_fence_get(fence);
649 }
650 dma_fence_put(fence);
651
652 if (ring->funcs->emit_gds_switch && gds_switch_needed) {
653 id->gds_base = job->gds_base;
654 id->gds_size = job->gds_size;
655 id->gws_base = job->gws_base;
656 id->gws_size = job->gws_size;
657 id->oa_base = job->oa_base;
658 id->oa_size = job->oa_size;
659 amdgpu_ring_emit_gds_switch(ring, job->vmid, job->gds_base,
660 job->gds_size, job->gws_base,
661 job->gws_size, job->oa_base,
662 job->oa_size);
663 }
664
665 if (ring->funcs->patch_cond_exec)
666 amdgpu_ring_patch_cond_exec(ring, patch_offset);
667
668 /* the double SWITCH_BUFFER here *cannot* be skipped by COND_EXEC */
669 if (ring->funcs->emit_switch_buffer) {
670 amdgpu_ring_emit_switch_buffer(ring);
671 amdgpu_ring_emit_switch_buffer(ring);
672 }
673 return 0;
674}
675
676/**
677 * amdgpu_vm_bo_find - find the bo_va for a specific vm & bo
678 *
679 * @vm: requested vm
680 * @bo: requested buffer object
681 *
682 * Find @bo inside the requested vm.
683 * Search inside the @bos vm list for the requested vm
684 * Returns the found bo_va or NULL if none is found
685 *
686 * Object has to be reserved!
687 */
688struct amdgpu_bo_va *amdgpu_vm_bo_find(struct amdgpu_vm *vm,
689 struct amdgpu_bo *bo)
690{
691 struct amdgpu_bo_va *bo_va;
692
693 list_for_each_entry(bo_va, &bo->va, base.bo_list) {
694 if (bo_va->base.vm == vm) {
695 return bo_va;
696 }
697 }
698 return NULL;
699}
700
701/**
702 * amdgpu_vm_do_set_ptes - helper to call the right asic function
703 *
704 * @params: see amdgpu_pte_update_params definition
705 * @bo: PD/PT to update
706 * @pe: addr of the page entry
707 * @addr: dst addr to write into pe
708 * @count: number of page entries to update
709 * @incr: increase next addr by incr bytes
710 * @flags: hw access flags
711 *
712 * Traces the parameters and calls the right asic functions
713 * to setup the page table using the DMA.
714 */
715static void amdgpu_vm_do_set_ptes(struct amdgpu_pte_update_params *params,
716 struct amdgpu_bo *bo,
717 uint64_t pe, uint64_t addr,
718 unsigned count, uint32_t incr,
719 uint64_t flags)
720{
721 pe += amdgpu_bo_gpu_offset(bo);
722 trace_amdgpu_vm_set_ptes(pe, addr, count, incr, flags);
723
724 if (count < 3) {
725 amdgpu_vm_write_pte(params->adev, params->ib, pe,
726 addr | flags, count, incr);
727
728 } else {
729 amdgpu_vm_set_pte_pde(params->adev, params->ib, pe, addr,
730 count, incr, flags);
731 }
732}
733
734/**
735 * amdgpu_vm_do_copy_ptes - copy the PTEs from the GART
736 *
737 * @params: see amdgpu_pte_update_params definition
738 * @bo: PD/PT to update
739 * @pe: addr of the page entry
740 * @addr: dst addr to write into pe
741 * @count: number of page entries to update
742 * @incr: increase next addr by incr bytes
743 * @flags: hw access flags
744 *
745 * Traces the parameters and calls the DMA function to copy the PTEs.
746 */
747static void amdgpu_vm_do_copy_ptes(struct amdgpu_pte_update_params *params,
748 struct amdgpu_bo *bo,
749 uint64_t pe, uint64_t addr,
750 unsigned count, uint32_t incr,
751 uint64_t flags)
752{
753 uint64_t src = (params->src + (addr >> 12) * 8);
754
755 pe += amdgpu_bo_gpu_offset(bo);
756 trace_amdgpu_vm_copy_ptes(pe, src, count);
757
758 amdgpu_vm_copy_pte(params->adev, params->ib, pe, src, count);
759}
760
761/**
762 * amdgpu_vm_map_gart - Resolve gart mapping of addr
763 *
764 * @pages_addr: optional DMA address to use for lookup
765 * @addr: the unmapped addr
766 *
767 * Look up the physical address of the page that the pte resolves
768 * to and return the pointer for the page table entry.
769 */
770static uint64_t amdgpu_vm_map_gart(const dma_addr_t *pages_addr, uint64_t addr)
771{
772 uint64_t result;
773
774 /* page table offset */
775 result = pages_addr[addr >> PAGE_SHIFT];
776
777 /* in case cpu page size != gpu page size*/
778 result |= addr & (~PAGE_MASK);
779
780 result &= 0xFFFFFFFFFFFFF000ULL;
781
782 return result;
783}
784
785/**
786 * amdgpu_vm_cpu_set_ptes - helper to update page tables via CPU
787 *
788 * @params: see amdgpu_pte_update_params definition
789 * @bo: PD/PT to update
790 * @pe: kmap addr of the page entry
791 * @addr: dst addr to write into pe
792 * @count: number of page entries to update
793 * @incr: increase next addr by incr bytes
794 * @flags: hw access flags
795 *
796 * Write count number of PT/PD entries directly.
797 */
798static void amdgpu_vm_cpu_set_ptes(struct amdgpu_pte_update_params *params,
799 struct amdgpu_bo *bo,
800 uint64_t pe, uint64_t addr,
801 unsigned count, uint32_t incr,
802 uint64_t flags)
803{
804 unsigned int i;
805 uint64_t value;
806
807 pe += (unsigned long)amdgpu_bo_kptr(bo);
808
809 trace_amdgpu_vm_set_ptes(pe, addr, count, incr, flags);
810
811 for (i = 0; i < count; i++) {
812 value = params->pages_addr ?
813 amdgpu_vm_map_gart(params->pages_addr, addr) :
814 addr;
815 amdgpu_gmc_set_pte_pde(params->adev, (void *)(uintptr_t)pe,
816 i, value, flags);
817 addr += incr;
818 }
819}
820
821static int amdgpu_vm_wait_pd(struct amdgpu_device *adev, struct amdgpu_vm *vm,
822 void *owner)
823{
824 struct amdgpu_sync sync;
825 int r;
826
827 amdgpu_sync_create(&sync);
828 amdgpu_sync_resv(adev, &sync, vm->root.base.bo->tbo.resv, owner, false);
829 r = amdgpu_sync_wait(&sync, true);
830 amdgpu_sync_free(&sync);
831
832 return r;
833}
834
835/*
836 * amdgpu_vm_update_pde - update a single level in the hierarchy
837 *
838 * @param: parameters for the update
839 * @vm: requested vm
840 * @parent: parent directory
841 * @entry: entry to update
842 *
843 * Makes sure the requested entry in parent is up to date.
844 */
845static void amdgpu_vm_update_pde(struct amdgpu_pte_update_params *params,
846 struct amdgpu_vm *vm,
847 struct amdgpu_vm_pt *parent,
848 struct amdgpu_vm_pt *entry)
849{
850 struct amdgpu_bo *bo = parent->base.bo, *pbo;
851 uint64_t pde, pt, flags;
852 unsigned level;
853
854 /* Don't update huge pages here */
855 if (entry->huge)
856 return;
857
858 for (level = 0, pbo = bo->parent; pbo; ++level)
859 pbo = pbo->parent;
860
861 level += params->adev->vm_manager.root_level;
862 pt = amdgpu_bo_gpu_offset(entry->base.bo);
863 flags = AMDGPU_PTE_VALID;
864 amdgpu_gmc_get_vm_pde(params->adev, level, &pt, &flags);
865 pde = (entry - parent->entries) * 8;
866 if (bo->shadow)
867 params->func(params, bo->shadow, pde, pt, 1, 0, flags);
868 params->func(params, bo, pde, pt, 1, 0, flags);
869}
870
871/*
872 * amdgpu_vm_invalidate_level - mark all PD levels as invalid
873 *
874 * @parent: parent PD
875 *
876 * Mark all PD level as invalid after an error.
877 */
878static void amdgpu_vm_invalidate_level(struct amdgpu_device *adev,
879 struct amdgpu_vm *vm,
880 struct amdgpu_vm_pt *parent,
881 unsigned level)
882{
883 unsigned pt_idx, num_entries;
884
885 /*
886 * Recurse into the subdirectories. This recursion is harmless because
887 * we only have a maximum of 5 layers.
888 */
889 num_entries = amdgpu_vm_num_entries(adev, level);
890 for (pt_idx = 0; pt_idx < num_entries; ++pt_idx) {
891 struct amdgpu_vm_pt *entry = &parent->entries[pt_idx];
892
893 if (!entry->base.bo)
894 continue;
895
896 spin_lock(&vm->status_lock);
897 if (list_empty(&entry->base.vm_status))
898 list_add(&entry->base.vm_status, &vm->relocated);
899 spin_unlock(&vm->status_lock);
900 amdgpu_vm_invalidate_level(adev, vm, entry, level + 1);
901 }
902}
903
904/*
905 * amdgpu_vm_update_directories - make sure that all directories are valid
906 *
907 * @adev: amdgpu_device pointer
908 * @vm: requested vm
909 *
910 * Makes sure all directories are up to date.
911 * Returns 0 for success, error for failure.
912 */
913int amdgpu_vm_update_directories(struct amdgpu_device *adev,
914 struct amdgpu_vm *vm)
915{
916 struct amdgpu_pte_update_params params;
917 struct amdgpu_job *job;
918 unsigned ndw = 0;
919 int r = 0;
920
921 if (list_empty(&vm->relocated))
922 return 0;
923
924restart:
925 memset(¶ms, 0, sizeof(params));
926 params.adev = adev;
927
928 if (vm->use_cpu_for_update) {
929 r = amdgpu_vm_wait_pd(adev, vm, AMDGPU_FENCE_OWNER_VM);
930 if (unlikely(r))
931 return r;
932
933 params.func = amdgpu_vm_cpu_set_ptes;
934 } else {
935 ndw = 512 * 8;
936 r = amdgpu_job_alloc_with_ib(adev, ndw * 4, &job);
937 if (r)
938 return r;
939
940 params.ib = &job->ibs[0];
941 params.func = amdgpu_vm_do_set_ptes;
942 }
943
944 spin_lock(&vm->status_lock);
945 while (!list_empty(&vm->relocated)) {
946 struct amdgpu_vm_bo_base *bo_base, *parent;
947 struct amdgpu_vm_pt *pt, *entry;
948 struct amdgpu_bo *bo;
949
950 bo_base = list_first_entry(&vm->relocated,
951 struct amdgpu_vm_bo_base,
952 vm_status);
953 list_del_init(&bo_base->vm_status);
954 spin_unlock(&vm->status_lock);
955
956 bo = bo_base->bo->parent;
957 if (!bo) {
958 spin_lock(&vm->status_lock);
959 continue;
960 }
961
962 parent = list_first_entry(&bo->va, struct amdgpu_vm_bo_base,
963 bo_list);
964 pt = container_of(parent, struct amdgpu_vm_pt, base);
965 entry = container_of(bo_base, struct amdgpu_vm_pt, base);
966
967 amdgpu_vm_update_pde(¶ms, vm, pt, entry);
968
969 spin_lock(&vm->status_lock);
970 if (!vm->use_cpu_for_update &&
971 (ndw - params.ib->length_dw) < 32)
972 break;
973 }
974 spin_unlock(&vm->status_lock);
975
976 if (vm->use_cpu_for_update) {
977 /* Flush HDP */
978 mb();
979 amdgpu_asic_flush_hdp(adev, NULL);
980 } else if (params.ib->length_dw == 0) {
981 amdgpu_job_free(job);
982 } else {
983 struct amdgpu_bo *root = vm->root.base.bo;
984 struct amdgpu_ring *ring;
985 struct dma_fence *fence;
986
987 ring = container_of(vm->entity.sched, struct amdgpu_ring,
988 sched);
989
990 amdgpu_ring_pad_ib(ring, params.ib);
991 amdgpu_sync_resv(adev, &job->sync, root->tbo.resv,
992 AMDGPU_FENCE_OWNER_VM, false);
993 WARN_ON(params.ib->length_dw > ndw);
994 r = amdgpu_job_submit(job, ring, &vm->entity,
995 AMDGPU_FENCE_OWNER_VM, &fence);
996 if (r)
997 goto error;
998
999 amdgpu_bo_fence(root, fence, true);
1000 dma_fence_put(vm->last_update);
1001 vm->last_update = fence;
1002 }
1003
1004 if (!list_empty(&vm->relocated))
1005 goto restart;
1006
1007 return 0;
1008
1009error:
1010 amdgpu_vm_invalidate_level(adev, vm, &vm->root,
1011 adev->vm_manager.root_level);
1012 amdgpu_job_free(job);
1013 return r;
1014}
1015
1016/**
1017 * amdgpu_vm_find_entry - find the entry for an address
1018 *
1019 * @p: see amdgpu_pte_update_params definition
1020 * @addr: virtual address in question
1021 * @entry: resulting entry or NULL
1022 * @parent: parent entry
1023 *
1024 * Find the vm_pt entry and it's parent for the given address.
1025 */
1026void amdgpu_vm_get_entry(struct amdgpu_pte_update_params *p, uint64_t addr,
1027 struct amdgpu_vm_pt **entry,
1028 struct amdgpu_vm_pt **parent)
1029{
1030 unsigned level = p->adev->vm_manager.root_level;
1031
1032 *parent = NULL;
1033 *entry = &p->vm->root;
1034 while ((*entry)->entries) {
1035 unsigned shift = amdgpu_vm_level_shift(p->adev, level++);
1036
1037 *parent = *entry;
1038 *entry = &(*entry)->entries[addr >> shift];
1039 addr &= (1ULL << shift) - 1;
1040 }
1041
1042 if (level != AMDGPU_VM_PTB)
1043 *entry = NULL;
1044}
1045
1046/**
1047 * amdgpu_vm_handle_huge_pages - handle updating the PD with huge pages
1048 *
1049 * @p: see amdgpu_pte_update_params definition
1050 * @entry: vm_pt entry to check
1051 * @parent: parent entry
1052 * @nptes: number of PTEs updated with this operation
1053 * @dst: destination address where the PTEs should point to
1054 * @flags: access flags fro the PTEs
1055 *
1056 * Check if we can update the PD with a huge page.
1057 */
1058static void amdgpu_vm_handle_huge_pages(struct amdgpu_pte_update_params *p,
1059 struct amdgpu_vm_pt *entry,
1060 struct amdgpu_vm_pt *parent,
1061 unsigned nptes, uint64_t dst,
1062 uint64_t flags)
1063{
1064 uint64_t pde;
1065
1066 /* In the case of a mixed PT the PDE must point to it*/
1067 if (p->adev->asic_type >= CHIP_VEGA10 && !p->src &&
1068 nptes == AMDGPU_VM_PTE_COUNT(p->adev)) {
1069 /* Set the huge page flag to stop scanning at this PDE */
1070 flags |= AMDGPU_PDE_PTE;
1071 }
1072
1073 if (!(flags & AMDGPU_PDE_PTE)) {
1074 if (entry->huge) {
1075 /* Add the entry to the relocated list to update it. */
1076 entry->huge = false;
1077 spin_lock(&p->vm->status_lock);
1078 list_move(&entry->base.vm_status, &p->vm->relocated);
1079 spin_unlock(&p->vm->status_lock);
1080 }
1081 return;
1082 }
1083
1084 entry->huge = true;
1085 amdgpu_gmc_get_vm_pde(p->adev, AMDGPU_VM_PDB0, &dst, &flags);
1086
1087 pde = (entry - parent->entries) * 8;
1088 if (parent->base.bo->shadow)
1089 p->func(p, parent->base.bo->shadow, pde, dst, 1, 0, flags);
1090 p->func(p, parent->base.bo, pde, dst, 1, 0, flags);
1091}
1092
1093/**
1094 * amdgpu_vm_update_ptes - make sure that page tables are valid
1095 *
1096 * @params: see amdgpu_pte_update_params definition
1097 * @vm: requested vm
1098 * @start: start of GPU address range
1099 * @end: end of GPU address range
1100 * @dst: destination address to map to, the next dst inside the function
1101 * @flags: mapping flags
1102 *
1103 * Update the page tables in the range @start - @end.
1104 * Returns 0 for success, -EINVAL for failure.
1105 */
1106static int amdgpu_vm_update_ptes(struct amdgpu_pte_update_params *params,
1107 uint64_t start, uint64_t end,
1108 uint64_t dst, uint64_t flags)
1109{
1110 struct amdgpu_device *adev = params->adev;
1111 const uint64_t mask = AMDGPU_VM_PTE_COUNT(adev) - 1;
1112
1113 uint64_t addr, pe_start;
1114 struct amdgpu_bo *pt;
1115 unsigned nptes;
1116
1117 /* walk over the address space and update the page tables */
1118 for (addr = start; addr < end; addr += nptes,
1119 dst += nptes * AMDGPU_GPU_PAGE_SIZE) {
1120 struct amdgpu_vm_pt *entry, *parent;
1121
1122 amdgpu_vm_get_entry(params, addr, &entry, &parent);
1123 if (!entry)
1124 return -ENOENT;
1125
1126 if ((addr & ~mask) == (end & ~mask))
1127 nptes = end - addr;
1128 else
1129 nptes = AMDGPU_VM_PTE_COUNT(adev) - (addr & mask);
1130
1131 amdgpu_vm_handle_huge_pages(params, entry, parent,
1132 nptes, dst, flags);
1133 /* We don't need to update PTEs for huge pages */
1134 if (entry->huge)
1135 continue;
1136
1137 pt = entry->base.bo;
1138 pe_start = (addr & mask) * 8;
1139 if (pt->shadow)
1140 params->func(params, pt->shadow, pe_start, dst, nptes,
1141 AMDGPU_GPU_PAGE_SIZE, flags);
1142 params->func(params, pt, pe_start, dst, nptes,
1143 AMDGPU_GPU_PAGE_SIZE, flags);
1144 }
1145
1146 return 0;
1147}
1148
1149/*
1150 * amdgpu_vm_frag_ptes - add fragment information to PTEs
1151 *
1152 * @params: see amdgpu_pte_update_params definition
1153 * @vm: requested vm
1154 * @start: first PTE to handle
1155 * @end: last PTE to handle
1156 * @dst: addr those PTEs should point to
1157 * @flags: hw mapping flags
1158 * Returns 0 for success, -EINVAL for failure.
1159 */
1160static int amdgpu_vm_frag_ptes(struct amdgpu_pte_update_params *params,
1161 uint64_t start, uint64_t end,
1162 uint64_t dst, uint64_t flags)
1163{
1164 /**
1165 * The MC L1 TLB supports variable sized pages, based on a fragment
1166 * field in the PTE. When this field is set to a non-zero value, page
1167 * granularity is increased from 4KB to (1 << (12 + frag)). The PTE
1168 * flags are considered valid for all PTEs within the fragment range
1169 * and corresponding mappings are assumed to be physically contiguous.
1170 *
1171 * The L1 TLB can store a single PTE for the whole fragment,
1172 * significantly increasing the space available for translation
1173 * caching. This leads to large improvements in throughput when the
1174 * TLB is under pressure.
1175 *
1176 * The L2 TLB distributes small and large fragments into two
1177 * asymmetric partitions. The large fragment cache is significantly
1178 * larger. Thus, we try to use large fragments wherever possible.
1179 * Userspace can support this by aligning virtual base address and
1180 * allocation size to the fragment size.
1181 */
1182 unsigned max_frag = params->adev->vm_manager.fragment_size;
1183 int r;
1184
1185 /* system pages are non continuously */
1186 if (params->src || !(flags & AMDGPU_PTE_VALID))
1187 return amdgpu_vm_update_ptes(params, start, end, dst, flags);
1188
1189 while (start != end) {
1190 uint64_t frag_flags, frag_end;
1191 unsigned frag;
1192
1193 /* This intentionally wraps around if no bit is set */
1194 frag = min((unsigned)ffs(start) - 1,
1195 (unsigned)fls64(end - start) - 1);
1196 if (frag >= max_frag) {
1197 frag_flags = AMDGPU_PTE_FRAG(max_frag);
1198 frag_end = end & ~((1ULL << max_frag) - 1);
1199 } else {
1200 frag_flags = AMDGPU_PTE_FRAG(frag);
1201 frag_end = start + (1 << frag);
1202 }
1203
1204 r = amdgpu_vm_update_ptes(params, start, frag_end, dst,
1205 flags | frag_flags);
1206 if (r)
1207 return r;
1208
1209 dst += (frag_end - start) * AMDGPU_GPU_PAGE_SIZE;
1210 start = frag_end;
1211 }
1212
1213 return 0;
1214}
1215
1216/**
1217 * amdgpu_vm_bo_update_mapping - update a mapping in the vm page table
1218 *
1219 * @adev: amdgpu_device pointer
1220 * @exclusive: fence we need to sync to
1221 * @pages_addr: DMA addresses to use for mapping
1222 * @vm: requested vm
1223 * @start: start of mapped range
1224 * @last: last mapped entry
1225 * @flags: flags for the entries
1226 * @addr: addr to set the area to
1227 * @fence: optional resulting fence
1228 *
1229 * Fill in the page table entries between @start and @last.
1230 * Returns 0 for success, -EINVAL for failure.
1231 */
1232static int amdgpu_vm_bo_update_mapping(struct amdgpu_device *adev,
1233 struct dma_fence *exclusive,
1234 dma_addr_t *pages_addr,
1235 struct amdgpu_vm *vm,
1236 uint64_t start, uint64_t last,
1237 uint64_t flags, uint64_t addr,
1238 struct dma_fence **fence)
1239{
1240 struct amdgpu_ring *ring;
1241 void *owner = AMDGPU_FENCE_OWNER_VM;
1242 unsigned nptes, ncmds, ndw;
1243 struct amdgpu_job *job;
1244 struct amdgpu_pte_update_params params;
1245 struct dma_fence *f = NULL;
1246 int r;
1247
1248 memset(¶ms, 0, sizeof(params));
1249 params.adev = adev;
1250 params.vm = vm;
1251
1252 /* sync to everything on unmapping */
1253 if (!(flags & AMDGPU_PTE_VALID))
1254 owner = AMDGPU_FENCE_OWNER_UNDEFINED;
1255
1256 if (vm->use_cpu_for_update) {
1257 /* params.src is used as flag to indicate system Memory */
1258 if (pages_addr)
1259 params.src = ~0;
1260
1261 /* Wait for PT BOs to be free. PTs share the same resv. object
1262 * as the root PD BO
1263 */
1264 r = amdgpu_vm_wait_pd(adev, vm, owner);
1265 if (unlikely(r))
1266 return r;
1267
1268 params.func = amdgpu_vm_cpu_set_ptes;
1269 params.pages_addr = pages_addr;
1270 return amdgpu_vm_frag_ptes(¶ms, start, last + 1,
1271 addr, flags);
1272 }
1273
1274 ring = container_of(vm->entity.sched, struct amdgpu_ring, sched);
1275
1276 nptes = last - start + 1;
1277
1278 /*
1279 * reserve space for two commands every (1 << BLOCK_SIZE)
1280 * entries or 2k dwords (whatever is smaller)
1281 *
1282 * The second command is for the shadow pagetables.
1283 */
1284 if (vm->root.base.bo->shadow)
1285 ncmds = ((nptes >> min(adev->vm_manager.block_size, 11u)) + 1) * 2;
1286 else
1287 ncmds = ((nptes >> min(adev->vm_manager.block_size, 11u)) + 1);
1288
1289 /* padding, etc. */
1290 ndw = 64;
1291
1292 if (pages_addr) {
1293 /* copy commands needed */
1294 ndw += ncmds * adev->vm_manager.vm_pte_funcs->copy_pte_num_dw;
1295
1296 /* and also PTEs */
1297 ndw += nptes * 2;
1298
1299 params.func = amdgpu_vm_do_copy_ptes;
1300
1301 } else {
1302 /* set page commands needed */
1303 ndw += ncmds * 10;
1304
1305 /* extra commands for begin/end fragments */
1306 ndw += 2 * 10 * adev->vm_manager.fragment_size;
1307
1308 params.func = amdgpu_vm_do_set_ptes;
1309 }
1310
1311 r = amdgpu_job_alloc_with_ib(adev, ndw * 4, &job);
1312 if (r)
1313 return r;
1314
1315 params.ib = &job->ibs[0];
1316
1317 if (pages_addr) {
1318 uint64_t *pte;
1319 unsigned i;
1320
1321 /* Put the PTEs at the end of the IB. */
1322 i = ndw - nptes * 2;
1323 pte= (uint64_t *)&(job->ibs->ptr[i]);
1324 params.src = job->ibs->gpu_addr + i * 4;
1325
1326 for (i = 0; i < nptes; ++i) {
1327 pte[i] = amdgpu_vm_map_gart(pages_addr, addr + i *
1328 AMDGPU_GPU_PAGE_SIZE);
1329 pte[i] |= flags;
1330 }
1331 addr = 0;
1332 }
1333
1334 r = amdgpu_sync_fence(adev, &job->sync, exclusive, false);
1335 if (r)
1336 goto error_free;
1337
1338 r = amdgpu_sync_resv(adev, &job->sync, vm->root.base.bo->tbo.resv,
1339 owner, false);
1340 if (r)
1341 goto error_free;
1342
1343 r = reservation_object_reserve_shared(vm->root.base.bo->tbo.resv);
1344 if (r)
1345 goto error_free;
1346
1347 r = amdgpu_vm_frag_ptes(¶ms, start, last + 1, addr, flags);
1348 if (r)
1349 goto error_free;
1350
1351 amdgpu_ring_pad_ib(ring, params.ib);
1352 WARN_ON(params.ib->length_dw > ndw);
1353 r = amdgpu_job_submit(job, ring, &vm->entity,
1354 AMDGPU_FENCE_OWNER_VM, &f);
1355 if (r)
1356 goto error_free;
1357
1358 amdgpu_bo_fence(vm->root.base.bo, f, true);
1359 dma_fence_put(*fence);
1360 *fence = f;
1361 return 0;
1362
1363error_free:
1364 amdgpu_job_free(job);
1365 return r;
1366}
1367
1368/**
1369 * amdgpu_vm_bo_split_mapping - split a mapping into smaller chunks
1370 *
1371 * @adev: amdgpu_device pointer
1372 * @exclusive: fence we need to sync to
1373 * @pages_addr: DMA addresses to use for mapping
1374 * @vm: requested vm
1375 * @mapping: mapped range and flags to use for the update
1376 * @flags: HW flags for the mapping
1377 * @nodes: array of drm_mm_nodes with the MC addresses
1378 * @fence: optional resulting fence
1379 *
1380 * Split the mapping into smaller chunks so that each update fits
1381 * into a SDMA IB.
1382 * Returns 0 for success, -EINVAL for failure.
1383 */
1384static int amdgpu_vm_bo_split_mapping(struct amdgpu_device *adev,
1385 struct dma_fence *exclusive,
1386 dma_addr_t *pages_addr,
1387 struct amdgpu_vm *vm,
1388 struct amdgpu_bo_va_mapping *mapping,
1389 uint64_t flags,
1390 struct drm_mm_node *nodes,
1391 struct dma_fence **fence)
1392{
1393 unsigned min_linear_pages = 1 << adev->vm_manager.fragment_size;
1394 uint64_t pfn, start = mapping->start;
1395 int r;
1396
1397 /* normally,bo_va->flags only contians READABLE and WIRTEABLE bit go here
1398 * but in case of something, we filter the flags in first place
1399 */
1400 if (!(mapping->flags & AMDGPU_PTE_READABLE))
1401 flags &= ~AMDGPU_PTE_READABLE;
1402 if (!(mapping->flags & AMDGPU_PTE_WRITEABLE))
1403 flags &= ~AMDGPU_PTE_WRITEABLE;
1404
1405 flags &= ~AMDGPU_PTE_EXECUTABLE;
1406 flags |= mapping->flags & AMDGPU_PTE_EXECUTABLE;
1407
1408 flags &= ~AMDGPU_PTE_MTYPE_MASK;
1409 flags |= (mapping->flags & AMDGPU_PTE_MTYPE_MASK);
1410
1411 if ((mapping->flags & AMDGPU_PTE_PRT) &&
1412 (adev->asic_type >= CHIP_VEGA10)) {
1413 flags |= AMDGPU_PTE_PRT;
1414 flags &= ~AMDGPU_PTE_VALID;
1415 }
1416
1417 trace_amdgpu_vm_bo_update(mapping);
1418
1419 pfn = mapping->offset >> PAGE_SHIFT;
1420 if (nodes) {
1421 while (pfn >= nodes->size) {
1422 pfn -= nodes->size;
1423 ++nodes;
1424 }
1425 }
1426
1427 do {
1428 dma_addr_t *dma_addr = NULL;
1429 uint64_t max_entries;
1430 uint64_t addr, last;
1431
1432 if (nodes) {
1433 addr = nodes->start << PAGE_SHIFT;
1434 max_entries = (nodes->size - pfn) *
1435 (PAGE_SIZE / AMDGPU_GPU_PAGE_SIZE);
1436 } else {
1437 addr = 0;
1438 max_entries = S64_MAX;
1439 }
1440
1441 if (pages_addr) {
1442 uint64_t count;
1443
1444 max_entries = min(max_entries, 16ull * 1024ull);
1445 for (count = 1; count < max_entries; ++count) {
1446 uint64_t idx = pfn + count;
1447
1448 if (pages_addr[idx] !=
1449 (pages_addr[idx - 1] + PAGE_SIZE))
1450 break;
1451 }
1452
1453 if (count < min_linear_pages) {
1454 addr = pfn << PAGE_SHIFT;
1455 dma_addr = pages_addr;
1456 } else {
1457 addr = pages_addr[pfn];
1458 max_entries = count;
1459 }
1460
1461 } else if (flags & AMDGPU_PTE_VALID) {
1462 addr += adev->vm_manager.vram_base_offset;
1463 addr += pfn << PAGE_SHIFT;
1464 }
1465
1466 last = min((uint64_t)mapping->last, start + max_entries - 1);
1467 r = amdgpu_vm_bo_update_mapping(adev, exclusive, dma_addr, vm,
1468 start, last, flags, addr,
1469 fence);
1470 if (r)
1471 return r;
1472
1473 pfn += last - start + 1;
1474 if (nodes && nodes->size == pfn) {
1475 pfn = 0;
1476 ++nodes;
1477 }
1478 start = last + 1;
1479
1480 } while (unlikely(start != mapping->last + 1));
1481
1482 return 0;
1483}
1484
1485/**
1486 * amdgpu_vm_bo_update - update all BO mappings in the vm page table
1487 *
1488 * @adev: amdgpu_device pointer
1489 * @bo_va: requested BO and VM object
1490 * @clear: if true clear the entries
1491 *
1492 * Fill in the page table entries for @bo_va.
1493 * Returns 0 for success, -EINVAL for failure.
1494 */
1495int amdgpu_vm_bo_update(struct amdgpu_device *adev,
1496 struct amdgpu_bo_va *bo_va,
1497 bool clear)
1498{
1499 struct amdgpu_bo *bo = bo_va->base.bo;
1500 struct amdgpu_vm *vm = bo_va->base.vm;
1501 struct amdgpu_bo_va_mapping *mapping;
1502 dma_addr_t *pages_addr = NULL;
1503 struct ttm_mem_reg *mem;
1504 struct drm_mm_node *nodes;
1505 struct dma_fence *exclusive, **last_update;
1506 uint64_t flags;
1507 int r;
1508
1509 if (clear || !bo_va->base.bo) {
1510 mem = NULL;
1511 nodes = NULL;
1512 exclusive = NULL;
1513 } else {
1514 struct ttm_dma_tt *ttm;
1515
1516 mem = &bo_va->base.bo->tbo.mem;
1517 nodes = mem->mm_node;
1518 if (mem->mem_type == TTM_PL_TT) {
1519 ttm = container_of(bo_va->base.bo->tbo.ttm,
1520 struct ttm_dma_tt, ttm);
1521 pages_addr = ttm->dma_address;
1522 }
1523 exclusive = reservation_object_get_excl(bo->tbo.resv);
1524 }
1525
1526 if (bo)
1527 flags = amdgpu_ttm_tt_pte_flags(adev, bo->tbo.ttm, mem);
1528 else
1529 flags = 0x0;
1530
1531 if (clear || (bo && bo->tbo.resv == vm->root.base.bo->tbo.resv))
1532 last_update = &vm->last_update;
1533 else
1534 last_update = &bo_va->last_pt_update;
1535
1536 if (!clear && bo_va->base.moved) {
1537 bo_va->base.moved = false;
1538 list_splice_init(&bo_va->valids, &bo_va->invalids);
1539
1540 } else if (bo_va->cleared != clear) {
1541 list_splice_init(&bo_va->valids, &bo_va->invalids);
1542 }
1543
1544 list_for_each_entry(mapping, &bo_va->invalids, list) {
1545 r = amdgpu_vm_bo_split_mapping(adev, exclusive, pages_addr, vm,
1546 mapping, flags, nodes,
1547 last_update);
1548 if (r)
1549 return r;
1550 }
1551
1552 if (vm->use_cpu_for_update) {
1553 /* Flush HDP */
1554 mb();
1555 amdgpu_asic_flush_hdp(adev, NULL);
1556 }
1557
1558 spin_lock(&vm->status_lock);
1559 list_del_init(&bo_va->base.vm_status);
1560 spin_unlock(&vm->status_lock);
1561
1562 list_splice_init(&bo_va->invalids, &bo_va->valids);
1563 bo_va->cleared = clear;
1564
1565 if (trace_amdgpu_vm_bo_mapping_enabled()) {
1566 list_for_each_entry(mapping, &bo_va->valids, list)
1567 trace_amdgpu_vm_bo_mapping(mapping);
1568 }
1569
1570 return 0;
1571}
1572
1573/**
1574 * amdgpu_vm_update_prt_state - update the global PRT state
1575 */
1576static void amdgpu_vm_update_prt_state(struct amdgpu_device *adev)
1577{
1578 unsigned long flags;
1579 bool enable;
1580
1581 spin_lock_irqsave(&adev->vm_manager.prt_lock, flags);
1582 enable = !!atomic_read(&adev->vm_manager.num_prt_users);
1583 adev->gmc.gmc_funcs->set_prt(adev, enable);
1584 spin_unlock_irqrestore(&adev->vm_manager.prt_lock, flags);
1585}
1586
1587/**
1588 * amdgpu_vm_prt_get - add a PRT user
1589 */
1590static void amdgpu_vm_prt_get(struct amdgpu_device *adev)
1591{
1592 if (!adev->gmc.gmc_funcs->set_prt)
1593 return;
1594
1595 if (atomic_inc_return(&adev->vm_manager.num_prt_users) == 1)
1596 amdgpu_vm_update_prt_state(adev);
1597}
1598
1599/**
1600 * amdgpu_vm_prt_put - drop a PRT user
1601 */
1602static void amdgpu_vm_prt_put(struct amdgpu_device *adev)
1603{
1604 if (atomic_dec_return(&adev->vm_manager.num_prt_users) == 0)
1605 amdgpu_vm_update_prt_state(adev);
1606}
1607
1608/**
1609 * amdgpu_vm_prt_cb - callback for updating the PRT status
1610 */
1611static void amdgpu_vm_prt_cb(struct dma_fence *fence, struct dma_fence_cb *_cb)
1612{
1613 struct amdgpu_prt_cb *cb = container_of(_cb, struct amdgpu_prt_cb, cb);
1614
1615 amdgpu_vm_prt_put(cb->adev);
1616 kfree(cb);
1617}
1618
1619/**
1620 * amdgpu_vm_add_prt_cb - add callback for updating the PRT status
1621 */
1622static void amdgpu_vm_add_prt_cb(struct amdgpu_device *adev,
1623 struct dma_fence *fence)
1624{
1625 struct amdgpu_prt_cb *cb;
1626
1627 if (!adev->gmc.gmc_funcs->set_prt)
1628 return;
1629
1630 cb = kmalloc(sizeof(struct amdgpu_prt_cb), GFP_KERNEL);
1631 if (!cb) {
1632 /* Last resort when we are OOM */
1633 if (fence)
1634 dma_fence_wait(fence, false);
1635
1636 amdgpu_vm_prt_put(adev);
1637 } else {
1638 cb->adev = adev;
1639 if (!fence || dma_fence_add_callback(fence, &cb->cb,
1640 amdgpu_vm_prt_cb))
1641 amdgpu_vm_prt_cb(fence, &cb->cb);
1642 }
1643}
1644
1645/**
1646 * amdgpu_vm_free_mapping - free a mapping
1647 *
1648 * @adev: amdgpu_device pointer
1649 * @vm: requested vm
1650 * @mapping: mapping to be freed
1651 * @fence: fence of the unmap operation
1652 *
1653 * Free a mapping and make sure we decrease the PRT usage count if applicable.
1654 */
1655static void amdgpu_vm_free_mapping(struct amdgpu_device *adev,
1656 struct amdgpu_vm *vm,
1657 struct amdgpu_bo_va_mapping *mapping,
1658 struct dma_fence *fence)
1659{
1660 if (mapping->flags & AMDGPU_PTE_PRT)
1661 amdgpu_vm_add_prt_cb(adev, fence);
1662 kfree(mapping);
1663}
1664
1665/**
1666 * amdgpu_vm_prt_fini - finish all prt mappings
1667 *
1668 * @adev: amdgpu_device pointer
1669 * @vm: requested vm
1670 *
1671 * Register a cleanup callback to disable PRT support after VM dies.
1672 */
1673static void amdgpu_vm_prt_fini(struct amdgpu_device *adev, struct amdgpu_vm *vm)
1674{
1675 struct reservation_object *resv = vm->root.base.bo->tbo.resv;
1676 struct dma_fence *excl, **shared;
1677 unsigned i, shared_count;
1678 int r;
1679
1680 r = reservation_object_get_fences_rcu(resv, &excl,
1681 &shared_count, &shared);
1682 if (r) {
1683 /* Not enough memory to grab the fence list, as last resort
1684 * block for all the fences to complete.
1685 */
1686 reservation_object_wait_timeout_rcu(resv, true, false,
1687 MAX_SCHEDULE_TIMEOUT);
1688 return;
1689 }
1690
1691 /* Add a callback for each fence in the reservation object */
1692 amdgpu_vm_prt_get(adev);
1693 amdgpu_vm_add_prt_cb(adev, excl);
1694
1695 for (i = 0; i < shared_count; ++i) {
1696 amdgpu_vm_prt_get(adev);
1697 amdgpu_vm_add_prt_cb(adev, shared[i]);
1698 }
1699
1700 kfree(shared);
1701}
1702
1703/**
1704 * amdgpu_vm_clear_freed - clear freed BOs in the PT
1705 *
1706 * @adev: amdgpu_device pointer
1707 * @vm: requested vm
1708 * @fence: optional resulting fence (unchanged if no work needed to be done
1709 * or if an error occurred)
1710 *
1711 * Make sure all freed BOs are cleared in the PT.
1712 * Returns 0 for success.
1713 *
1714 * PTs have to be reserved and mutex must be locked!
1715 */
1716int amdgpu_vm_clear_freed(struct amdgpu_device *adev,
1717 struct amdgpu_vm *vm,
1718 struct dma_fence **fence)
1719{
1720 struct amdgpu_bo_va_mapping *mapping;
1721 uint64_t init_pte_value = 0;
1722 struct dma_fence *f = NULL;
1723 int r;
1724
1725 while (!list_empty(&vm->freed)) {
1726 mapping = list_first_entry(&vm->freed,
1727 struct amdgpu_bo_va_mapping, list);
1728 list_del(&mapping->list);
1729
1730 if (vm->pte_support_ats && mapping->start < AMDGPU_VA_HOLE_START)
1731 init_pte_value = AMDGPU_PTE_DEFAULT_ATC;
1732
1733 r = amdgpu_vm_bo_update_mapping(adev, NULL, NULL, vm,
1734 mapping->start, mapping->last,
1735 init_pte_value, 0, &f);
1736 amdgpu_vm_free_mapping(adev, vm, mapping, f);
1737 if (r) {
1738 dma_fence_put(f);
1739 return r;
1740 }
1741 }
1742
1743 if (fence && f) {
1744 dma_fence_put(*fence);
1745 *fence = f;
1746 } else {
1747 dma_fence_put(f);
1748 }
1749
1750 return 0;
1751
1752}
1753
1754/**
1755 * amdgpu_vm_handle_moved - handle moved BOs in the PT
1756 *
1757 * @adev: amdgpu_device pointer
1758 * @vm: requested vm
1759 * @sync: sync object to add fences to
1760 *
1761 * Make sure all BOs which are moved are updated in the PTs.
1762 * Returns 0 for success.
1763 *
1764 * PTs have to be reserved!
1765 */
1766int amdgpu_vm_handle_moved(struct amdgpu_device *adev,
1767 struct amdgpu_vm *vm)
1768{
1769 bool clear;
1770 int r = 0;
1771
1772 spin_lock(&vm->status_lock);
1773 while (!list_empty(&vm->moved)) {
1774 struct amdgpu_bo_va *bo_va;
1775 struct reservation_object *resv;
1776
1777 bo_va = list_first_entry(&vm->moved,
1778 struct amdgpu_bo_va, base.vm_status);
1779 spin_unlock(&vm->status_lock);
1780
1781 resv = bo_va->base.bo->tbo.resv;
1782
1783 /* Per VM BOs never need to bo cleared in the page tables */
1784 if (resv == vm->root.base.bo->tbo.resv)
1785 clear = false;
1786 /* Try to reserve the BO to avoid clearing its ptes */
1787 else if (!amdgpu_vm_debug && reservation_object_trylock(resv))
1788 clear = false;
1789 /* Somebody else is using the BO right now */
1790 else
1791 clear = true;
1792
1793 r = amdgpu_vm_bo_update(adev, bo_va, clear);
1794 if (r)
1795 return r;
1796
1797 if (!clear && resv != vm->root.base.bo->tbo.resv)
1798 reservation_object_unlock(resv);
1799
1800 spin_lock(&vm->status_lock);
1801 }
1802 spin_unlock(&vm->status_lock);
1803
1804 return r;
1805}
1806
1807/**
1808 * amdgpu_vm_bo_add - add a bo to a specific vm
1809 *
1810 * @adev: amdgpu_device pointer
1811 * @vm: requested vm
1812 * @bo: amdgpu buffer object
1813 *
1814 * Add @bo into the requested vm.
1815 * Add @bo to the list of bos associated with the vm
1816 * Returns newly added bo_va or NULL for failure
1817 *
1818 * Object has to be reserved!
1819 */
1820struct amdgpu_bo_va *amdgpu_vm_bo_add(struct amdgpu_device *adev,
1821 struct amdgpu_vm *vm,
1822 struct amdgpu_bo *bo)
1823{
1824 struct amdgpu_bo_va *bo_va;
1825
1826 bo_va = kzalloc(sizeof(struct amdgpu_bo_va), GFP_KERNEL);
1827 if (bo_va == NULL) {
1828 return NULL;
1829 }
1830 bo_va->base.vm = vm;
1831 bo_va->base.bo = bo;
1832 INIT_LIST_HEAD(&bo_va->base.bo_list);
1833 INIT_LIST_HEAD(&bo_va->base.vm_status);
1834
1835 bo_va->ref_count = 1;
1836 INIT_LIST_HEAD(&bo_va->valids);
1837 INIT_LIST_HEAD(&bo_va->invalids);
1838
1839 if (!bo)
1840 return bo_va;
1841
1842 list_add_tail(&bo_va->base.bo_list, &bo->va);
1843
1844 if (bo->tbo.resv != vm->root.base.bo->tbo.resv)
1845 return bo_va;
1846
1847 if (bo->preferred_domains &
1848 amdgpu_mem_type_to_domain(bo->tbo.mem.mem_type))
1849 return bo_va;
1850
1851 /*
1852 * We checked all the prerequisites, but it looks like this per VM BO
1853 * is currently evicted. add the BO to the evicted list to make sure it
1854 * is validated on next VM use to avoid fault.
1855 * */
1856 spin_lock(&vm->status_lock);
1857 list_move_tail(&bo_va->base.vm_status, &vm->evicted);
1858 spin_unlock(&vm->status_lock);
1859
1860 return bo_va;
1861}
1862
1863
1864/**
1865 * amdgpu_vm_bo_insert_mapping - insert a new mapping
1866 *
1867 * @adev: amdgpu_device pointer
1868 * @bo_va: bo_va to store the address
1869 * @mapping: the mapping to insert
1870 *
1871 * Insert a new mapping into all structures.
1872 */
1873static void amdgpu_vm_bo_insert_map(struct amdgpu_device *adev,
1874 struct amdgpu_bo_va *bo_va,
1875 struct amdgpu_bo_va_mapping *mapping)
1876{
1877 struct amdgpu_vm *vm = bo_va->base.vm;
1878 struct amdgpu_bo *bo = bo_va->base.bo;
1879
1880 mapping->bo_va = bo_va;
1881 list_add(&mapping->list, &bo_va->invalids);
1882 amdgpu_vm_it_insert(mapping, &vm->va);
1883
1884 if (mapping->flags & AMDGPU_PTE_PRT)
1885 amdgpu_vm_prt_get(adev);
1886
1887 if (bo && bo->tbo.resv == vm->root.base.bo->tbo.resv) {
1888 spin_lock(&vm->status_lock);
1889 if (list_empty(&bo_va->base.vm_status))
1890 list_add(&bo_va->base.vm_status, &vm->moved);
1891 spin_unlock(&vm->status_lock);
1892 }
1893 trace_amdgpu_vm_bo_map(bo_va, mapping);
1894}
1895
1896/**
1897 * amdgpu_vm_bo_map - map bo inside a vm
1898 *
1899 * @adev: amdgpu_device pointer
1900 * @bo_va: bo_va to store the address
1901 * @saddr: where to map the BO
1902 * @offset: requested offset in the BO
1903 * @flags: attributes of pages (read/write/valid/etc.)
1904 *
1905 * Add a mapping of the BO at the specefied addr into the VM.
1906 * Returns 0 for success, error for failure.
1907 *
1908 * Object has to be reserved and unreserved outside!
1909 */
1910int amdgpu_vm_bo_map(struct amdgpu_device *adev,
1911 struct amdgpu_bo_va *bo_va,
1912 uint64_t saddr, uint64_t offset,
1913 uint64_t size, uint64_t flags)
1914{
1915 struct amdgpu_bo_va_mapping *mapping, *tmp;
1916 struct amdgpu_bo *bo = bo_va->base.bo;
1917 struct amdgpu_vm *vm = bo_va->base.vm;
1918 uint64_t eaddr;
1919
1920 /* validate the parameters */
1921 if (saddr & AMDGPU_GPU_PAGE_MASK || offset & AMDGPU_GPU_PAGE_MASK ||
1922 size == 0 || size & AMDGPU_GPU_PAGE_MASK)
1923 return -EINVAL;
1924
1925 /* make sure object fit at this offset */
1926 eaddr = saddr + size - 1;
1927 if (saddr >= eaddr ||
1928 (bo && offset + size > amdgpu_bo_size(bo)))
1929 return -EINVAL;
1930
1931 saddr /= AMDGPU_GPU_PAGE_SIZE;
1932 eaddr /= AMDGPU_GPU_PAGE_SIZE;
1933
1934 tmp = amdgpu_vm_it_iter_first(&vm->va, saddr, eaddr);
1935 if (tmp) {
1936 /* bo and tmp overlap, invalid addr */
1937 dev_err(adev->dev, "bo %p va 0x%010Lx-0x%010Lx conflict with "
1938 "0x%010Lx-0x%010Lx\n", bo, saddr, eaddr,
1939 tmp->start, tmp->last + 1);
1940 return -EINVAL;
1941 }
1942
1943 mapping = kmalloc(sizeof(*mapping), GFP_KERNEL);
1944 if (!mapping)
1945 return -ENOMEM;
1946
1947 mapping->start = saddr;
1948 mapping->last = eaddr;
1949 mapping->offset = offset;
1950 mapping->flags = flags;
1951
1952 amdgpu_vm_bo_insert_map(adev, bo_va, mapping);
1953
1954 return 0;
1955}
1956
1957/**
1958 * amdgpu_vm_bo_replace_map - map bo inside a vm, replacing existing mappings
1959 *
1960 * @adev: amdgpu_device pointer
1961 * @bo_va: bo_va to store the address
1962 * @saddr: where to map the BO
1963 * @offset: requested offset in the BO
1964 * @flags: attributes of pages (read/write/valid/etc.)
1965 *
1966 * Add a mapping of the BO at the specefied addr into the VM. Replace existing
1967 * mappings as we do so.
1968 * Returns 0 for success, error for failure.
1969 *
1970 * Object has to be reserved and unreserved outside!
1971 */
1972int amdgpu_vm_bo_replace_map(struct amdgpu_device *adev,
1973 struct amdgpu_bo_va *bo_va,
1974 uint64_t saddr, uint64_t offset,
1975 uint64_t size, uint64_t flags)
1976{
1977 struct amdgpu_bo_va_mapping *mapping;
1978 struct amdgpu_bo *bo = bo_va->base.bo;
1979 uint64_t eaddr;
1980 int r;
1981
1982 /* validate the parameters */
1983 if (saddr & AMDGPU_GPU_PAGE_MASK || offset & AMDGPU_GPU_PAGE_MASK ||
1984 size == 0 || size & AMDGPU_GPU_PAGE_MASK)
1985 return -EINVAL;
1986
1987 /* make sure object fit at this offset */
1988 eaddr = saddr + size - 1;
1989 if (saddr >= eaddr ||
1990 (bo && offset + size > amdgpu_bo_size(bo)))
1991 return -EINVAL;
1992
1993 /* Allocate all the needed memory */
1994 mapping = kmalloc(sizeof(*mapping), GFP_KERNEL);
1995 if (!mapping)
1996 return -ENOMEM;
1997
1998 r = amdgpu_vm_bo_clear_mappings(adev, bo_va->base.vm, saddr, size);
1999 if (r) {
2000 kfree(mapping);
2001 return r;
2002 }
2003
2004 saddr /= AMDGPU_GPU_PAGE_SIZE;
2005 eaddr /= AMDGPU_GPU_PAGE_SIZE;
2006
2007 mapping->start = saddr;
2008 mapping->last = eaddr;
2009 mapping->offset = offset;
2010 mapping->flags = flags;
2011
2012 amdgpu_vm_bo_insert_map(adev, bo_va, mapping);
2013
2014 return 0;
2015}
2016
2017/**
2018 * amdgpu_vm_bo_unmap - remove bo mapping from vm
2019 *
2020 * @adev: amdgpu_device pointer
2021 * @bo_va: bo_va to remove the address from
2022 * @saddr: where to the BO is mapped
2023 *
2024 * Remove a mapping of the BO at the specefied addr from the VM.
2025 * Returns 0 for success, error for failure.
2026 *
2027 * Object has to be reserved and unreserved outside!
2028 */
2029int amdgpu_vm_bo_unmap(struct amdgpu_device *adev,
2030 struct amdgpu_bo_va *bo_va,
2031 uint64_t saddr)
2032{
2033 struct amdgpu_bo_va_mapping *mapping;
2034 struct amdgpu_vm *vm = bo_va->base.vm;
2035 bool valid = true;
2036
2037 saddr /= AMDGPU_GPU_PAGE_SIZE;
2038
2039 list_for_each_entry(mapping, &bo_va->valids, list) {
2040 if (mapping->start == saddr)
2041 break;
2042 }
2043
2044 if (&mapping->list == &bo_va->valids) {
2045 valid = false;
2046
2047 list_for_each_entry(mapping, &bo_va->invalids, list) {
2048 if (mapping->start == saddr)
2049 break;
2050 }
2051
2052 if (&mapping->list == &bo_va->invalids)
2053 return -ENOENT;
2054 }
2055
2056 list_del(&mapping->list);
2057 amdgpu_vm_it_remove(mapping, &vm->va);
2058 mapping->bo_va = NULL;
2059 trace_amdgpu_vm_bo_unmap(bo_va, mapping);
2060
2061 if (valid)
2062 list_add(&mapping->list, &vm->freed);
2063 else
2064 amdgpu_vm_free_mapping(adev, vm, mapping,
2065 bo_va->last_pt_update);
2066
2067 return 0;
2068}
2069
2070/**
2071 * amdgpu_vm_bo_clear_mappings - remove all mappings in a specific range
2072 *
2073 * @adev: amdgpu_device pointer
2074 * @vm: VM structure to use
2075 * @saddr: start of the range
2076 * @size: size of the range
2077 *
2078 * Remove all mappings in a range, split them as appropriate.
2079 * Returns 0 for success, error for failure.
2080 */
2081int amdgpu_vm_bo_clear_mappings(struct amdgpu_device *adev,
2082 struct amdgpu_vm *vm,
2083 uint64_t saddr, uint64_t size)
2084{
2085 struct amdgpu_bo_va_mapping *before, *after, *tmp, *next;
2086 LIST_HEAD(removed);
2087 uint64_t eaddr;
2088
2089 eaddr = saddr + size - 1;
2090 saddr /= AMDGPU_GPU_PAGE_SIZE;
2091 eaddr /= AMDGPU_GPU_PAGE_SIZE;
2092
2093 /* Allocate all the needed memory */
2094 before = kzalloc(sizeof(*before), GFP_KERNEL);
2095 if (!before)
2096 return -ENOMEM;
2097 INIT_LIST_HEAD(&before->list);
2098
2099 after = kzalloc(sizeof(*after), GFP_KERNEL);
2100 if (!after) {
2101 kfree(before);
2102 return -ENOMEM;
2103 }
2104 INIT_LIST_HEAD(&after->list);
2105
2106 /* Now gather all removed mappings */
2107 tmp = amdgpu_vm_it_iter_first(&vm->va, saddr, eaddr);
2108 while (tmp) {
2109 /* Remember mapping split at the start */
2110 if (tmp->start < saddr) {
2111 before->start = tmp->start;
2112 before->last = saddr - 1;
2113 before->offset = tmp->offset;
2114 before->flags = tmp->flags;
2115 list_add(&before->list, &tmp->list);
2116 }
2117
2118 /* Remember mapping split at the end */
2119 if (tmp->last > eaddr) {
2120 after->start = eaddr + 1;
2121 after->last = tmp->last;
2122 after->offset = tmp->offset;
2123 after->offset += after->start - tmp->start;
2124 after->flags = tmp->flags;
2125 list_add(&after->list, &tmp->list);
2126 }
2127
2128 list_del(&tmp->list);
2129 list_add(&tmp->list, &removed);
2130
2131 tmp = amdgpu_vm_it_iter_next(tmp, saddr, eaddr);
2132 }
2133
2134 /* And free them up */
2135 list_for_each_entry_safe(tmp, next, &removed, list) {
2136 amdgpu_vm_it_remove(tmp, &vm->va);
2137 list_del(&tmp->list);
2138
2139 if (tmp->start < saddr)
2140 tmp->start = saddr;
2141 if (tmp->last > eaddr)
2142 tmp->last = eaddr;
2143
2144 tmp->bo_va = NULL;
2145 list_add(&tmp->list, &vm->freed);
2146 trace_amdgpu_vm_bo_unmap(NULL, tmp);
2147 }
2148
2149 /* Insert partial mapping before the range */
2150 if (!list_empty(&before->list)) {
2151 amdgpu_vm_it_insert(before, &vm->va);
2152 if (before->flags & AMDGPU_PTE_PRT)
2153 amdgpu_vm_prt_get(adev);
2154 } else {
2155 kfree(before);
2156 }
2157
2158 /* Insert partial mapping after the range */
2159 if (!list_empty(&after->list)) {
2160 amdgpu_vm_it_insert(after, &vm->va);
2161 if (after->flags & AMDGPU_PTE_PRT)
2162 amdgpu_vm_prt_get(adev);
2163 } else {
2164 kfree(after);
2165 }
2166
2167 return 0;
2168}
2169
2170/**
2171 * amdgpu_vm_bo_lookup_mapping - find mapping by address
2172 *
2173 * @vm: the requested VM
2174 *
2175 * Find a mapping by it's address.
2176 */
2177struct amdgpu_bo_va_mapping *amdgpu_vm_bo_lookup_mapping(struct amdgpu_vm *vm,
2178 uint64_t addr)
2179{
2180 return amdgpu_vm_it_iter_first(&vm->va, addr, addr);
2181}
2182
2183/**
2184 * amdgpu_vm_bo_rmv - remove a bo to a specific vm
2185 *
2186 * @adev: amdgpu_device pointer
2187 * @bo_va: requested bo_va
2188 *
2189 * Remove @bo_va->bo from the requested vm.
2190 *
2191 * Object have to be reserved!
2192 */
2193void amdgpu_vm_bo_rmv(struct amdgpu_device *adev,
2194 struct amdgpu_bo_va *bo_va)
2195{
2196 struct amdgpu_bo_va_mapping *mapping, *next;
2197 struct amdgpu_vm *vm = bo_va->base.vm;
2198
2199 list_del(&bo_va->base.bo_list);
2200
2201 spin_lock(&vm->status_lock);
2202 list_del(&bo_va->base.vm_status);
2203 spin_unlock(&vm->status_lock);
2204
2205 list_for_each_entry_safe(mapping, next, &bo_va->valids, list) {
2206 list_del(&mapping->list);
2207 amdgpu_vm_it_remove(mapping, &vm->va);
2208 mapping->bo_va = NULL;
2209 trace_amdgpu_vm_bo_unmap(bo_va, mapping);
2210 list_add(&mapping->list, &vm->freed);
2211 }
2212 list_for_each_entry_safe(mapping, next, &bo_va->invalids, list) {
2213 list_del(&mapping->list);
2214 amdgpu_vm_it_remove(mapping, &vm->va);
2215 amdgpu_vm_free_mapping(adev, vm, mapping,
2216 bo_va->last_pt_update);
2217 }
2218
2219 dma_fence_put(bo_va->last_pt_update);
2220 kfree(bo_va);
2221}
2222
2223/**
2224 * amdgpu_vm_bo_invalidate - mark the bo as invalid
2225 *
2226 * @adev: amdgpu_device pointer
2227 * @vm: requested vm
2228 * @bo: amdgpu buffer object
2229 *
2230 * Mark @bo as invalid.
2231 */
2232void amdgpu_vm_bo_invalidate(struct amdgpu_device *adev,
2233 struct amdgpu_bo *bo, bool evicted)
2234{
2235 struct amdgpu_vm_bo_base *bo_base;
2236
2237 list_for_each_entry(bo_base, &bo->va, bo_list) {
2238 struct amdgpu_vm *vm = bo_base->vm;
2239
2240 bo_base->moved = true;
2241 if (evicted && bo->tbo.resv == vm->root.base.bo->tbo.resv) {
2242 spin_lock(&bo_base->vm->status_lock);
2243 if (bo->tbo.type == ttm_bo_type_kernel)
2244 list_move(&bo_base->vm_status, &vm->evicted);
2245 else
2246 list_move_tail(&bo_base->vm_status,
2247 &vm->evicted);
2248 spin_unlock(&bo_base->vm->status_lock);
2249 continue;
2250 }
2251
2252 if (bo->tbo.type == ttm_bo_type_kernel) {
2253 spin_lock(&bo_base->vm->status_lock);
2254 if (list_empty(&bo_base->vm_status))
2255 list_add(&bo_base->vm_status, &vm->relocated);
2256 spin_unlock(&bo_base->vm->status_lock);
2257 continue;
2258 }
2259
2260 spin_lock(&bo_base->vm->status_lock);
2261 if (list_empty(&bo_base->vm_status))
2262 list_add(&bo_base->vm_status, &vm->moved);
2263 spin_unlock(&bo_base->vm->status_lock);
2264 }
2265}
2266
2267static uint32_t amdgpu_vm_get_block_size(uint64_t vm_size)
2268{
2269 /* Total bits covered by PD + PTs */
2270 unsigned bits = ilog2(vm_size) + 18;
2271
2272 /* Make sure the PD is 4K in size up to 8GB address space.
2273 Above that split equal between PD and PTs */
2274 if (vm_size <= 8)
2275 return (bits - 9);
2276 else
2277 return ((bits + 3) / 2);
2278}
2279
2280/**
2281 * amdgpu_vm_adjust_size - adjust vm size, block size and fragment size
2282 *
2283 * @adev: amdgpu_device pointer
2284 * @vm_size: the default vm size if it's set auto
2285 */
2286void amdgpu_vm_adjust_size(struct amdgpu_device *adev, uint32_t vm_size,
2287 uint32_t fragment_size_default, unsigned max_level,
2288 unsigned max_bits)
2289{
2290 uint64_t tmp;
2291
2292 /* adjust vm size first */
2293 if (amdgpu_vm_size != -1) {
2294 unsigned max_size = 1 << (max_bits - 30);
2295
2296 vm_size = amdgpu_vm_size;
2297 if (vm_size > max_size) {
2298 dev_warn(adev->dev, "VM size (%d) too large, max is %u GB\n",
2299 amdgpu_vm_size, max_size);
2300 vm_size = max_size;
2301 }
2302 }
2303
2304 adev->vm_manager.max_pfn = (uint64_t)vm_size << 18;
2305
2306 tmp = roundup_pow_of_two(adev->vm_manager.max_pfn);
2307 if (amdgpu_vm_block_size != -1)
2308 tmp >>= amdgpu_vm_block_size - 9;
2309 tmp = DIV_ROUND_UP(fls64(tmp) - 1, 9) - 1;
2310 adev->vm_manager.num_level = min(max_level, (unsigned)tmp);
2311 switch (adev->vm_manager.num_level) {
2312 case 3:
2313 adev->vm_manager.root_level = AMDGPU_VM_PDB2;
2314 break;
2315 case 2:
2316 adev->vm_manager.root_level = AMDGPU_VM_PDB1;
2317 break;
2318 case 1:
2319 adev->vm_manager.root_level = AMDGPU_VM_PDB0;
2320 break;
2321 default:
2322 dev_err(adev->dev, "VMPT only supports 2~4+1 levels\n");
2323 }
2324 /* block size depends on vm size and hw setup*/
2325 if (amdgpu_vm_block_size != -1)
2326 adev->vm_manager.block_size =
2327 min((unsigned)amdgpu_vm_block_size, max_bits
2328 - AMDGPU_GPU_PAGE_SHIFT
2329 - 9 * adev->vm_manager.num_level);
2330 else if (adev->vm_manager.num_level > 1)
2331 adev->vm_manager.block_size = 9;
2332 else
2333 adev->vm_manager.block_size = amdgpu_vm_get_block_size(tmp);
2334
2335 if (amdgpu_vm_fragment_size == -1)
2336 adev->vm_manager.fragment_size = fragment_size_default;
2337 else
2338 adev->vm_manager.fragment_size = amdgpu_vm_fragment_size;
2339
2340 DRM_INFO("vm size is %u GB, %u levels, block size is %u-bit, fragment size is %u-bit\n",
2341 vm_size, adev->vm_manager.num_level + 1,
2342 adev->vm_manager.block_size,
2343 adev->vm_manager.fragment_size);
2344}
2345
2346/**
2347 * amdgpu_vm_init - initialize a vm instance
2348 *
2349 * @adev: amdgpu_device pointer
2350 * @vm: requested vm
2351 * @vm_context: Indicates if it GFX or Compute context
2352 *
2353 * Init @vm fields.
2354 */
2355int amdgpu_vm_init(struct amdgpu_device *adev, struct amdgpu_vm *vm,
2356 int vm_context, unsigned int pasid)
2357{
2358 const unsigned align = min(AMDGPU_VM_PTB_ALIGN_SIZE,
2359 AMDGPU_VM_PTE_COUNT(adev) * 8);
2360 unsigned ring_instance;
2361 struct amdgpu_ring *ring;
2362 struct drm_sched_rq *rq;
2363 unsigned long size;
2364 uint64_t flags;
2365 int r, i;
2366
2367 vm->va = RB_ROOT_CACHED;
2368 for (i = 0; i < AMDGPU_MAX_VMHUBS; i++)
2369 vm->reserved_vmid[i] = NULL;
2370 spin_lock_init(&vm->status_lock);
2371 INIT_LIST_HEAD(&vm->evicted);
2372 INIT_LIST_HEAD(&vm->relocated);
2373 INIT_LIST_HEAD(&vm->moved);
2374 INIT_LIST_HEAD(&vm->freed);
2375
2376 /* create scheduler entity for page table updates */
2377
2378 ring_instance = atomic_inc_return(&adev->vm_manager.vm_pte_next_ring);
2379 ring_instance %= adev->vm_manager.vm_pte_num_rings;
2380 ring = adev->vm_manager.vm_pte_rings[ring_instance];
2381 rq = &ring->sched.sched_rq[DRM_SCHED_PRIORITY_KERNEL];
2382 r = drm_sched_entity_init(&ring->sched, &vm->entity,
2383 rq, amdgpu_sched_jobs, NULL);
2384 if (r)
2385 return r;
2386
2387 vm->pte_support_ats = false;
2388
2389 if (vm_context == AMDGPU_VM_CONTEXT_COMPUTE) {
2390 vm->use_cpu_for_update = !!(adev->vm_manager.vm_update_mode &
2391 AMDGPU_VM_USE_CPU_FOR_COMPUTE);
2392
2393 if (adev->asic_type == CHIP_RAVEN)
2394 vm->pte_support_ats = true;
2395 } else {
2396 vm->use_cpu_for_update = !!(adev->vm_manager.vm_update_mode &
2397 AMDGPU_VM_USE_CPU_FOR_GFX);
2398 }
2399 DRM_DEBUG_DRIVER("VM update mode is %s\n",
2400 vm->use_cpu_for_update ? "CPU" : "SDMA");
2401 WARN_ONCE((vm->use_cpu_for_update & !amdgpu_vm_is_large_bar(adev)),
2402 "CPU update of VM recommended only for large BAR system\n");
2403 vm->last_update = NULL;
2404
2405 flags = AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS;
2406 if (vm->use_cpu_for_update)
2407 flags |= AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED;
2408 else
2409 flags |= AMDGPU_GEM_CREATE_SHADOW;
2410
2411 size = amdgpu_vm_bo_size(adev, adev->vm_manager.root_level);
2412 r = amdgpu_bo_create(adev, size, align, AMDGPU_GEM_DOMAIN_VRAM, flags,
2413 ttm_bo_type_kernel, NULL, &vm->root.base.bo);
2414 if (r)
2415 goto error_free_sched_entity;
2416
2417 r = amdgpu_bo_reserve(vm->root.base.bo, true);
2418 if (r)
2419 goto error_free_root;
2420
2421 r = amdgpu_vm_clear_bo(adev, vm, vm->root.base.bo,
2422 adev->vm_manager.root_level,
2423 vm->pte_support_ats);
2424 if (r)
2425 goto error_unreserve;
2426
2427 vm->root.base.vm = vm;
2428 list_add_tail(&vm->root.base.bo_list, &vm->root.base.bo->va);
2429 list_add_tail(&vm->root.base.vm_status, &vm->evicted);
2430 amdgpu_bo_unreserve(vm->root.base.bo);
2431
2432 if (pasid) {
2433 unsigned long flags;
2434
2435 spin_lock_irqsave(&adev->vm_manager.pasid_lock, flags);
2436 r = idr_alloc(&adev->vm_manager.pasid_idr, vm, pasid, pasid + 1,
2437 GFP_ATOMIC);
2438 spin_unlock_irqrestore(&adev->vm_manager.pasid_lock, flags);
2439 if (r < 0)
2440 goto error_free_root;
2441
2442 vm->pasid = pasid;
2443 }
2444
2445 INIT_KFIFO(vm->faults);
2446 vm->fault_credit = 16;
2447
2448 return 0;
2449
2450error_unreserve:
2451 amdgpu_bo_unreserve(vm->root.base.bo);
2452
2453error_free_root:
2454 amdgpu_bo_unref(&vm->root.base.bo->shadow);
2455 amdgpu_bo_unref(&vm->root.base.bo);
2456 vm->root.base.bo = NULL;
2457
2458error_free_sched_entity:
2459 drm_sched_entity_fini(&ring->sched, &vm->entity);
2460
2461 return r;
2462}
2463
2464/**
2465 * amdgpu_vm_make_compute - Turn a GFX VM into a compute VM
2466 *
2467 * This only works on GFX VMs that don't have any BOs added and no
2468 * page tables allocated yet.
2469 *
2470 * Changes the following VM parameters:
2471 * - use_cpu_for_update
2472 * - pte_supports_ats
2473 * - pasid (old PASID is released, because compute manages its own PASIDs)
2474 *
2475 * Reinitializes the page directory to reflect the changed ATS
2476 * setting. May leave behind an unused shadow BO for the page
2477 * directory when switching from SDMA updates to CPU updates.
2478 *
2479 * Returns 0 for success, -errno for errors.
2480 */
2481int amdgpu_vm_make_compute(struct amdgpu_device *adev, struct amdgpu_vm *vm)
2482{
2483 bool pte_support_ats = (adev->asic_type == CHIP_RAVEN);
2484 int r;
2485
2486 r = amdgpu_bo_reserve(vm->root.base.bo, true);
2487 if (r)
2488 return r;
2489
2490 /* Sanity checks */
2491 if (!RB_EMPTY_ROOT(&vm->va.rb_root) || vm->root.entries) {
2492 r = -EINVAL;
2493 goto error;
2494 }
2495
2496 /* Check if PD needs to be reinitialized and do it before
2497 * changing any other state, in case it fails.
2498 */
2499 if (pte_support_ats != vm->pte_support_ats) {
2500 r = amdgpu_vm_clear_bo(adev, vm, vm->root.base.bo,
2501 adev->vm_manager.root_level,
2502 pte_support_ats);
2503 if (r)
2504 goto error;
2505 }
2506
2507 /* Update VM state */
2508 vm->use_cpu_for_update = !!(adev->vm_manager.vm_update_mode &
2509 AMDGPU_VM_USE_CPU_FOR_COMPUTE);
2510 vm->pte_support_ats = pte_support_ats;
2511 DRM_DEBUG_DRIVER("VM update mode is %s\n",
2512 vm->use_cpu_for_update ? "CPU" : "SDMA");
2513 WARN_ONCE((vm->use_cpu_for_update & !amdgpu_vm_is_large_bar(adev)),
2514 "CPU update of VM recommended only for large BAR system\n");
2515
2516 if (vm->pasid) {
2517 unsigned long flags;
2518
2519 spin_lock_irqsave(&adev->vm_manager.pasid_lock, flags);
2520 idr_remove(&adev->vm_manager.pasid_idr, vm->pasid);
2521 spin_unlock_irqrestore(&adev->vm_manager.pasid_lock, flags);
2522
2523 vm->pasid = 0;
2524 }
2525
2526error:
2527 amdgpu_bo_unreserve(vm->root.base.bo);
2528 return r;
2529}
2530
2531/**
2532 * amdgpu_vm_free_levels - free PD/PT levels
2533 *
2534 * @adev: amdgpu device structure
2535 * @parent: PD/PT starting level to free
2536 * @level: level of parent structure
2537 *
2538 * Free the page directory or page table level and all sub levels.
2539 */
2540static void amdgpu_vm_free_levels(struct amdgpu_device *adev,
2541 struct amdgpu_vm_pt *parent,
2542 unsigned level)
2543{
2544 unsigned i, num_entries = amdgpu_vm_num_entries(adev, level);
2545
2546 if (parent->base.bo) {
2547 list_del(&parent->base.bo_list);
2548 list_del(&parent->base.vm_status);
2549 amdgpu_bo_unref(&parent->base.bo->shadow);
2550 amdgpu_bo_unref(&parent->base.bo);
2551 }
2552
2553 if (parent->entries)
2554 for (i = 0; i < num_entries; i++)
2555 amdgpu_vm_free_levels(adev, &parent->entries[i],
2556 level + 1);
2557
2558 kvfree(parent->entries);
2559}
2560
2561/**
2562 * amdgpu_vm_fini - tear down a vm instance
2563 *
2564 * @adev: amdgpu_device pointer
2565 * @vm: requested vm
2566 *
2567 * Tear down @vm.
2568 * Unbind the VM and remove all bos from the vm bo list
2569 */
2570void amdgpu_vm_fini(struct amdgpu_device *adev, struct amdgpu_vm *vm)
2571{
2572 struct amdgpu_bo_va_mapping *mapping, *tmp;
2573 bool prt_fini_needed = !!adev->gmc.gmc_funcs->set_prt;
2574 struct amdgpu_bo *root;
2575 u64 fault;
2576 int i, r;
2577
2578 amdgpu_amdkfd_gpuvm_destroy_cb(adev, vm);
2579
2580 /* Clear pending page faults from IH when the VM is destroyed */
2581 while (kfifo_get(&vm->faults, &fault))
2582 amdgpu_ih_clear_fault(adev, fault);
2583
2584 if (vm->pasid) {
2585 unsigned long flags;
2586
2587 spin_lock_irqsave(&adev->vm_manager.pasid_lock, flags);
2588 idr_remove(&adev->vm_manager.pasid_idr, vm->pasid);
2589 spin_unlock_irqrestore(&adev->vm_manager.pasid_lock, flags);
2590 }
2591
2592 drm_sched_entity_fini(vm->entity.sched, &vm->entity);
2593
2594 if (!RB_EMPTY_ROOT(&vm->va.rb_root)) {
2595 dev_err(adev->dev, "still active bo inside vm\n");
2596 }
2597 rbtree_postorder_for_each_entry_safe(mapping, tmp,
2598 &vm->va.rb_root, rb) {
2599 list_del(&mapping->list);
2600 amdgpu_vm_it_remove(mapping, &vm->va);
2601 kfree(mapping);
2602 }
2603 list_for_each_entry_safe(mapping, tmp, &vm->freed, list) {
2604 if (mapping->flags & AMDGPU_PTE_PRT && prt_fini_needed) {
2605 amdgpu_vm_prt_fini(adev, vm);
2606 prt_fini_needed = false;
2607 }
2608
2609 list_del(&mapping->list);
2610 amdgpu_vm_free_mapping(adev, vm, mapping, NULL);
2611 }
2612
2613 root = amdgpu_bo_ref(vm->root.base.bo);
2614 r = amdgpu_bo_reserve(root, true);
2615 if (r) {
2616 dev_err(adev->dev, "Leaking page tables because BO reservation failed\n");
2617 } else {
2618 amdgpu_vm_free_levels(adev, &vm->root,
2619 adev->vm_manager.root_level);
2620 amdgpu_bo_unreserve(root);
2621 }
2622 amdgpu_bo_unref(&root);
2623 dma_fence_put(vm->last_update);
2624 for (i = 0; i < AMDGPU_MAX_VMHUBS; i++)
2625 amdgpu_vmid_free_reserved(adev, vm, i);
2626}
2627
2628/**
2629 * amdgpu_vm_pasid_fault_credit - Check fault credit for given PASID
2630 *
2631 * @adev: amdgpu_device pointer
2632 * @pasid: PASID do identify the VM
2633 *
2634 * This function is expected to be called in interrupt context. Returns
2635 * true if there was fault credit, false otherwise
2636 */
2637bool amdgpu_vm_pasid_fault_credit(struct amdgpu_device *adev,
2638 unsigned int pasid)
2639{
2640 struct amdgpu_vm *vm;
2641
2642 spin_lock(&adev->vm_manager.pasid_lock);
2643 vm = idr_find(&adev->vm_manager.pasid_idr, pasid);
2644 if (!vm) {
2645 /* VM not found, can't track fault credit */
2646 spin_unlock(&adev->vm_manager.pasid_lock);
2647 return true;
2648 }
2649
2650 /* No lock needed. only accessed by IRQ handler */
2651 if (!vm->fault_credit) {
2652 /* Too many faults in this VM */
2653 spin_unlock(&adev->vm_manager.pasid_lock);
2654 return false;
2655 }
2656
2657 vm->fault_credit--;
2658 spin_unlock(&adev->vm_manager.pasid_lock);
2659 return true;
2660}
2661
2662/**
2663 * amdgpu_vm_manager_init - init the VM manager
2664 *
2665 * @adev: amdgpu_device pointer
2666 *
2667 * Initialize the VM manager structures
2668 */
2669void amdgpu_vm_manager_init(struct amdgpu_device *adev)
2670{
2671 unsigned i;
2672
2673 amdgpu_vmid_mgr_init(adev);
2674
2675 adev->vm_manager.fence_context =
2676 dma_fence_context_alloc(AMDGPU_MAX_RINGS);
2677 for (i = 0; i < AMDGPU_MAX_RINGS; ++i)
2678 adev->vm_manager.seqno[i] = 0;
2679
2680 atomic_set(&adev->vm_manager.vm_pte_next_ring, 0);
2681 spin_lock_init(&adev->vm_manager.prt_lock);
2682 atomic_set(&adev->vm_manager.num_prt_users, 0);
2683
2684 /* If not overridden by the user, by default, only in large BAR systems
2685 * Compute VM tables will be updated by CPU
2686 */
2687#ifdef CONFIG_X86_64
2688 if (amdgpu_vm_update_mode == -1) {
2689 if (amdgpu_vm_is_large_bar(adev))
2690 adev->vm_manager.vm_update_mode =
2691 AMDGPU_VM_USE_CPU_FOR_COMPUTE;
2692 else
2693 adev->vm_manager.vm_update_mode = 0;
2694 } else
2695 adev->vm_manager.vm_update_mode = amdgpu_vm_update_mode;
2696#else
2697 adev->vm_manager.vm_update_mode = 0;
2698#endif
2699
2700 idr_init(&adev->vm_manager.pasid_idr);
2701 spin_lock_init(&adev->vm_manager.pasid_lock);
2702}
2703
2704/**
2705 * amdgpu_vm_manager_fini - cleanup VM manager
2706 *
2707 * @adev: amdgpu_device pointer
2708 *
2709 * Cleanup the VM manager and free resources.
2710 */
2711void amdgpu_vm_manager_fini(struct amdgpu_device *adev)
2712{
2713 WARN_ON(!idr_is_empty(&adev->vm_manager.pasid_idr));
2714 idr_destroy(&adev->vm_manager.pasid_idr);
2715
2716 amdgpu_vmid_mgr_fini(adev);
2717}
2718
2719int amdgpu_vm_ioctl(struct drm_device *dev, void *data, struct drm_file *filp)
2720{
2721 union drm_amdgpu_vm *args = data;
2722 struct amdgpu_device *adev = dev->dev_private;
2723 struct amdgpu_fpriv *fpriv = filp->driver_priv;
2724 int r;
2725
2726 switch (args->in.op) {
2727 case AMDGPU_VM_OP_RESERVE_VMID:
2728 /* current, we only have requirement to reserve vmid from gfxhub */
2729 r = amdgpu_vmid_alloc_reserved(adev, &fpriv->vm, AMDGPU_GFXHUB);
2730 if (r)
2731 return r;
2732 break;
2733 case AMDGPU_VM_OP_UNRESERVE_VMID:
2734 amdgpu_vmid_free_reserved(adev, &fpriv->vm, AMDGPU_GFXHUB);
2735 break;
2736 default:
2737 return -EINVAL;
2738 }
2739
2740 return 0;
2741}