Linux Audio

Check our new training course

Loading...
v6.8
   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_relocated - vm_bo is reloacted
 238 *
 239 * @vm_bo: vm_bo which is relocated
 240 *
 241 * State for PDs/PTs which needs to update their parent PD.
 242 * For the root PD, just move to idle state.
 243 */
 244static void amdgpu_vm_bo_relocated(struct amdgpu_vm_bo_base *vm_bo)
 245{
 246	if (vm_bo->bo->parent) {
 247		spin_lock(&vm_bo->vm->status_lock);
 248		list_move(&vm_bo->vm_status, &vm_bo->vm->relocated);
 249		spin_unlock(&vm_bo->vm->status_lock);
 250	} else {
 251		amdgpu_vm_bo_idle(vm_bo);
 252	}
 253}
 254
 255/**
 256 * amdgpu_vm_bo_done - vm_bo is done
 257 *
 258 * @vm_bo: vm_bo which is now done
 259 *
 260 * State for normal BOs which are invalidated and that change has been updated
 261 * in the PTs.
 262 */
 263static void amdgpu_vm_bo_done(struct amdgpu_vm_bo_base *vm_bo)
 264{
 265	spin_lock(&vm_bo->vm->status_lock);
 266	list_move(&vm_bo->vm_status, &vm_bo->vm->done);
 267	spin_unlock(&vm_bo->vm->status_lock);
 268}
 269
 270/**
 271 * amdgpu_vm_bo_reset_state_machine - reset the vm_bo state machine
 272 * @vm: the VM which state machine to reset
 273 *
 274 * Move all vm_bo object in the VM into a state where they will be updated
 275 * again during validation.
 276 */
 277static void amdgpu_vm_bo_reset_state_machine(struct amdgpu_vm *vm)
 278{
 279	struct amdgpu_vm_bo_base *vm_bo, *tmp;
 280
 281	spin_lock(&vm->status_lock);
 282	list_splice_init(&vm->done, &vm->invalidated);
 283	list_for_each_entry(vm_bo, &vm->invalidated, vm_status)
 284		vm_bo->moved = true;
 285	list_for_each_entry_safe(vm_bo, tmp, &vm->idle, vm_status) {
 286		struct amdgpu_bo *bo = vm_bo->bo;
 287
 288		vm_bo->moved = true;
 289		if (!bo || bo->tbo.type != ttm_bo_type_kernel)
 290			list_move(&vm_bo->vm_status, &vm_bo->vm->moved);
 291		else if (bo->parent)
 292			list_move(&vm_bo->vm_status, &vm_bo->vm->relocated);
 293	}
 294	spin_unlock(&vm->status_lock);
 295}
 296
 297/**
 298 * amdgpu_vm_bo_base_init - Adds bo to the list of bos associated with the vm
 299 *
 300 * @base: base structure for tracking BO usage in a VM
 301 * @vm: vm to which bo is to be added
 302 * @bo: amdgpu buffer object
 303 *
 304 * Initialize a bo_va_base structure and add it to the appropriate lists
 305 *
 306 */
 307void amdgpu_vm_bo_base_init(struct amdgpu_vm_bo_base *base,
 308			    struct amdgpu_vm *vm, struct amdgpu_bo *bo)
 309{
 310	base->vm = vm;
 311	base->bo = bo;
 312	base->next = NULL;
 313	INIT_LIST_HEAD(&base->vm_status);
 314
 315	if (!bo)
 316		return;
 317	base->next = bo->vm_bo;
 318	bo->vm_bo = base;
 319
 320	if (bo->tbo.base.resv != vm->root.bo->tbo.base.resv)
 321		return;
 322
 323	dma_resv_assert_held(vm->root.bo->tbo.base.resv);
 324
 325	ttm_bo_set_bulk_move(&bo->tbo, &vm->lru_bulk_move);
 326	if (bo->tbo.type == ttm_bo_type_kernel && bo->parent)
 327		amdgpu_vm_bo_relocated(base);
 328	else
 329		amdgpu_vm_bo_idle(base);
 330
 331	if (bo->preferred_domains &
 332	    amdgpu_mem_type_to_domain(bo->tbo.resource->mem_type))
 333		return;
 334
 335	/*
 336	 * we checked all the prerequisites, but it looks like this per vm bo
 337	 * is currently evicted. add the bo to the evicted list to make sure it
 338	 * is validated on next vm use to avoid fault.
 339	 * */
 340	amdgpu_vm_bo_evicted(base);
 341}
 342
 343/**
 344 * amdgpu_vm_lock_pd - lock PD in drm_exec
 345 *
 346 * @vm: vm providing the BOs
 347 * @exec: drm execution context
 348 * @num_fences: number of extra fences to reserve
 349 *
 350 * Lock the VM root PD in the DRM execution context.
 
 351 */
 352int amdgpu_vm_lock_pd(struct amdgpu_vm *vm, struct drm_exec *exec,
 353		      unsigned int num_fences)
 354{
 355	/* We need at least two fences for the VM PD/PT updates */
 356	return drm_exec_prepare_obj(exec, &vm->root.bo->tbo.base,
 357				    2 + num_fences);
 
 
 
 
 358}
 359
 360/**
 361 * amdgpu_vm_move_to_lru_tail - move all BOs to the end of LRU
 362 *
 363 * @adev: amdgpu device pointer
 364 * @vm: vm providing the BOs
 365 *
 366 * Move all BOs to the end of LRU and remember their positions to put them
 367 * together.
 368 */
 369void amdgpu_vm_move_to_lru_tail(struct amdgpu_device *adev,
 370				struct amdgpu_vm *vm)
 371{
 372	spin_lock(&adev->mman.bdev.lru_lock);
 373	ttm_lru_bulk_move_tail(&vm->lru_bulk_move);
 374	spin_unlock(&adev->mman.bdev.lru_lock);
 375}
 376
 377/* Create scheduler entities for page table updates */
 378static int amdgpu_vm_init_entities(struct amdgpu_device *adev,
 379				   struct amdgpu_vm *vm)
 380{
 381	int r;
 382
 383	r = drm_sched_entity_init(&vm->immediate, DRM_SCHED_PRIORITY_NORMAL,
 384				  adev->vm_manager.vm_pte_scheds,
 385				  adev->vm_manager.vm_pte_num_scheds, NULL);
 386	if (r)
 387		goto error;
 388
 389	return drm_sched_entity_init(&vm->delayed, DRM_SCHED_PRIORITY_NORMAL,
 390				     adev->vm_manager.vm_pte_scheds,
 391				     adev->vm_manager.vm_pte_num_scheds, NULL);
 392
 393error:
 394	drm_sched_entity_destroy(&vm->immediate);
 395	return r;
 396}
 397
 398/* Destroy the entities for page table updates again */
 399static void amdgpu_vm_fini_entities(struct amdgpu_vm *vm)
 400{
 401	drm_sched_entity_destroy(&vm->immediate);
 402	drm_sched_entity_destroy(&vm->delayed);
 403}
 404
 405/**
 406 * amdgpu_vm_generation - return the page table re-generation counter
 407 * @adev: the amdgpu_device
 408 * @vm: optional VM to check, might be NULL
 409 *
 410 * Returns a page table re-generation token to allow checking if submissions
 411 * are still valid to use this VM. The VM parameter might be NULL in which case
 412 * just the VRAM lost counter will be used.
 413 */
 414uint64_t amdgpu_vm_generation(struct amdgpu_device *adev, struct amdgpu_vm *vm)
 415{
 416	uint64_t result = (u64)atomic_read(&adev->vram_lost_counter) << 32;
 417
 418	if (!vm)
 419		return result;
 420
 421	result += vm->generation;
 422	/* Add one if the page tables will be re-generated on next CS */
 423	if (drm_sched_entity_error(&vm->delayed))
 424		++result;
 425
 426	return result;
 427}
 428
 429/**
 430 * amdgpu_vm_validate_pt_bos - validate the page table BOs
 431 *
 432 * @adev: amdgpu device pointer
 433 * @vm: vm providing the BOs
 434 * @validate: callback to do the validation
 435 * @param: parameter for the validation callback
 436 *
 437 * Validate the page table BOs on command submission if neccessary.
 438 *
 439 * Returns:
 440 * Validation result.
 441 */
 442int amdgpu_vm_validate_pt_bos(struct amdgpu_device *adev, struct amdgpu_vm *vm,
 443			      int (*validate)(void *p, struct amdgpu_bo *bo),
 444			      void *param)
 445{
 446	struct amdgpu_vm_bo_base *bo_base;
 447	struct amdgpu_bo *shadow;
 448	struct amdgpu_bo *bo;
 449	int r;
 450
 451	if (drm_sched_entity_error(&vm->delayed)) {
 452		++vm->generation;
 453		amdgpu_vm_bo_reset_state_machine(vm);
 454		amdgpu_vm_fini_entities(vm);
 455		r = amdgpu_vm_init_entities(adev, vm);
 456		if (r)
 457			return r;
 458	}
 459
 460	spin_lock(&vm->status_lock);
 461	while (!list_empty(&vm->evicted)) {
 462		bo_base = list_first_entry(&vm->evicted,
 463					   struct amdgpu_vm_bo_base,
 464					   vm_status);
 465		spin_unlock(&vm->status_lock);
 466
 467		bo = bo_base->bo;
 468		shadow = amdgpu_bo_shadowed(bo);
 469
 470		r = validate(param, bo);
 471		if (r)
 472			return r;
 473		if (shadow) {
 474			r = validate(param, shadow);
 475			if (r)
 476				return r;
 477		}
 478
 479		if (bo->tbo.type != ttm_bo_type_kernel) {
 480			amdgpu_vm_bo_moved(bo_base);
 481		} else {
 482			vm->update_funcs->map_table(to_amdgpu_bo_vm(bo));
 483			amdgpu_vm_bo_relocated(bo_base);
 484		}
 485		spin_lock(&vm->status_lock);
 486	}
 487	spin_unlock(&vm->status_lock);
 488
 489	amdgpu_vm_eviction_lock(vm);
 490	vm->evicting = false;
 491	amdgpu_vm_eviction_unlock(vm);
 492
 493	return 0;
 494}
 495
 496/**
 497 * amdgpu_vm_ready - check VM is ready for updates
 498 *
 499 * @vm: VM to check
 500 *
 501 * Check if all VM PDs/PTs are ready for updates
 502 *
 503 * Returns:
 504 * True if VM is not evicting.
 505 */
 506bool amdgpu_vm_ready(struct amdgpu_vm *vm)
 507{
 508	bool empty;
 509	bool ret;
 510
 511	amdgpu_vm_eviction_lock(vm);
 512	ret = !vm->evicting;
 513	amdgpu_vm_eviction_unlock(vm);
 514
 515	spin_lock(&vm->status_lock);
 516	empty = list_empty(&vm->evicted);
 517	spin_unlock(&vm->status_lock);
 518
 519	return ret && empty;
 520}
 521
 522/**
 523 * amdgpu_vm_check_compute_bug - check whether asic has compute vm bug
 524 *
 525 * @adev: amdgpu_device pointer
 526 */
 527void amdgpu_vm_check_compute_bug(struct amdgpu_device *adev)
 528{
 529	const struct amdgpu_ip_block *ip_block;
 530	bool has_compute_vm_bug;
 531	struct amdgpu_ring *ring;
 532	int i;
 533
 534	has_compute_vm_bug = false;
 535
 536	ip_block = amdgpu_device_ip_get_ip_block(adev, AMD_IP_BLOCK_TYPE_GFX);
 537	if (ip_block) {
 538		/* Compute has a VM bug for GFX version < 7.
 539		   Compute has a VM bug for GFX 8 MEC firmware version < 673.*/
 540		if (ip_block->version->major <= 7)
 541			has_compute_vm_bug = true;
 542		else if (ip_block->version->major == 8)
 543			if (adev->gfx.mec_fw_version < 673)
 544				has_compute_vm_bug = true;
 545	}
 546
 547	for (i = 0; i < adev->num_rings; i++) {
 548		ring = adev->rings[i];
 549		if (ring->funcs->type == AMDGPU_RING_TYPE_COMPUTE)
 550			/* only compute rings */
 551			ring->has_compute_vm_bug = has_compute_vm_bug;
 552		else
 553			ring->has_compute_vm_bug = false;
 554	}
 555}
 556
 557/**
 558 * amdgpu_vm_need_pipeline_sync - Check if pipe sync is needed for job.
 559 *
 560 * @ring: ring on which the job will be submitted
 561 * @job: job to submit
 562 *
 563 * Returns:
 564 * True if sync is needed.
 565 */
 566bool amdgpu_vm_need_pipeline_sync(struct amdgpu_ring *ring,
 567				  struct amdgpu_job *job)
 568{
 569	struct amdgpu_device *adev = ring->adev;
 570	unsigned vmhub = ring->vm_hub;
 571	struct amdgpu_vmid_mgr *id_mgr = &adev->vm_manager.id_mgr[vmhub];
 572
 573	if (job->vmid == 0)
 574		return false;
 575
 576	if (job->vm_needs_flush || ring->has_compute_vm_bug)
 577		return true;
 578
 579	if (ring->funcs->emit_gds_switch && job->gds_switch_needed)
 580		return true;
 581
 582	if (amdgpu_vmid_had_gpu_reset(adev, &id_mgr->ids[job->vmid]))
 583		return true;
 584
 585	return false;
 586}
 587
 588/**
 589 * amdgpu_vm_flush - hardware flush the vm
 590 *
 591 * @ring: ring to use for flush
 592 * @job:  related job
 593 * @need_pipe_sync: is pipe sync needed
 594 *
 595 * Emit a VM flush when it is necessary.
 596 *
 597 * Returns:
 598 * 0 on success, errno otherwise.
 599 */
 600int amdgpu_vm_flush(struct amdgpu_ring *ring, struct amdgpu_job *job,
 601		    bool need_pipe_sync)
 602{
 603	struct amdgpu_device *adev = ring->adev;
 604	unsigned vmhub = ring->vm_hub;
 605	struct amdgpu_vmid_mgr *id_mgr = &adev->vm_manager.id_mgr[vmhub];
 606	struct amdgpu_vmid *id = &id_mgr->ids[job->vmid];
 607	bool spm_update_needed = job->spm_update_needed;
 608	bool gds_switch_needed = ring->funcs->emit_gds_switch &&
 609		job->gds_switch_needed;
 610	bool vm_flush_needed = job->vm_needs_flush;
 611	struct dma_fence *fence = NULL;
 612	bool pasid_mapping_needed = false;
 613	unsigned patch_offset = 0;
 614	int r;
 615
 616	if (amdgpu_vmid_had_gpu_reset(adev, id)) {
 617		gds_switch_needed = true;
 618		vm_flush_needed = true;
 619		pasid_mapping_needed = true;
 620		spm_update_needed = true;
 621	}
 622
 623	mutex_lock(&id_mgr->lock);
 624	if (id->pasid != job->pasid || !id->pasid_mapping ||
 625	    !dma_fence_is_signaled(id->pasid_mapping))
 626		pasid_mapping_needed = true;
 627	mutex_unlock(&id_mgr->lock);
 628
 629	gds_switch_needed &= !!ring->funcs->emit_gds_switch;
 630	vm_flush_needed &= !!ring->funcs->emit_vm_flush  &&
 631			job->vm_pd_addr != AMDGPU_BO_INVALID_OFFSET;
 632	pasid_mapping_needed &= adev->gmc.gmc_funcs->emit_pasid_mapping &&
 633		ring->funcs->emit_wreg;
 634
 635	if (!vm_flush_needed && !gds_switch_needed && !need_pipe_sync)
 636		return 0;
 637
 638	amdgpu_ring_ib_begin(ring);
 639	if (ring->funcs->init_cond_exec)
 640		patch_offset = amdgpu_ring_init_cond_exec(ring);
 641
 642	if (need_pipe_sync)
 643		amdgpu_ring_emit_pipeline_sync(ring);
 644
 645	if (vm_flush_needed) {
 646		trace_amdgpu_vm_flush(ring, job->vmid, job->vm_pd_addr);
 647		amdgpu_ring_emit_vm_flush(ring, job->vmid, job->vm_pd_addr);
 648	}
 649
 650	if (pasid_mapping_needed)
 651		amdgpu_gmc_emit_pasid_mapping(ring, job->vmid, job->pasid);
 652
 653	if (spm_update_needed && adev->gfx.rlc.funcs->update_spm_vmid)
 654		adev->gfx.rlc.funcs->update_spm_vmid(adev, job->vmid);
 655
 656	if (!ring->is_mes_queue && ring->funcs->emit_gds_switch &&
 657	    gds_switch_needed) {
 658		amdgpu_ring_emit_gds_switch(ring, job->vmid, job->gds_base,
 659					    job->gds_size, job->gws_base,
 660					    job->gws_size, job->oa_base,
 661					    job->oa_size);
 662	}
 663
 664	if (vm_flush_needed || pasid_mapping_needed) {
 665		r = amdgpu_fence_emit(ring, &fence, NULL, 0);
 666		if (r)
 667			return r;
 668	}
 669
 670	if (vm_flush_needed) {
 671		mutex_lock(&id_mgr->lock);
 672		dma_fence_put(id->last_flush);
 673		id->last_flush = dma_fence_get(fence);
 674		id->current_gpu_reset_count =
 675			atomic_read(&adev->gpu_reset_counter);
 676		mutex_unlock(&id_mgr->lock);
 677	}
 678
 679	if (pasid_mapping_needed) {
 680		mutex_lock(&id_mgr->lock);
 681		id->pasid = job->pasid;
 682		dma_fence_put(id->pasid_mapping);
 683		id->pasid_mapping = dma_fence_get(fence);
 684		mutex_unlock(&id_mgr->lock);
 685	}
 686	dma_fence_put(fence);
 687
 688	if (ring->funcs->patch_cond_exec)
 689		amdgpu_ring_patch_cond_exec(ring, patch_offset);
 690
 691	/* the double SWITCH_BUFFER here *cannot* be skipped by COND_EXEC */
 692	if (ring->funcs->emit_switch_buffer) {
 693		amdgpu_ring_emit_switch_buffer(ring);
 694		amdgpu_ring_emit_switch_buffer(ring);
 695	}
 696	amdgpu_ring_ib_end(ring);
 697	return 0;
 698}
 699
 700/**
 701 * amdgpu_vm_bo_find - find the bo_va for a specific vm & bo
 702 *
 703 * @vm: requested vm
 704 * @bo: requested buffer object
 705 *
 706 * Find @bo inside the requested vm.
 707 * Search inside the @bos vm list for the requested vm
 708 * Returns the found bo_va or NULL if none is found
 709 *
 710 * Object has to be reserved!
 711 *
 712 * Returns:
 713 * Found bo_va or NULL.
 714 */
 715struct amdgpu_bo_va *amdgpu_vm_bo_find(struct amdgpu_vm *vm,
 716				       struct amdgpu_bo *bo)
 717{
 718	struct amdgpu_vm_bo_base *base;
 719
 720	for (base = bo->vm_bo; base; base = base->next) {
 721		if (base->vm != vm)
 722			continue;
 723
 724		return container_of(base, struct amdgpu_bo_va, base);
 725	}
 726	return NULL;
 727}
 728
 729/**
 730 * amdgpu_vm_map_gart - Resolve gart mapping of addr
 731 *
 732 * @pages_addr: optional DMA address to use for lookup
 733 * @addr: the unmapped addr
 734 *
 735 * Look up the physical address of the page that the pte resolves
 736 * to.
 737 *
 738 * Returns:
 739 * The pointer for the page table entry.
 740 */
 741uint64_t amdgpu_vm_map_gart(const dma_addr_t *pages_addr, uint64_t addr)
 742{
 743	uint64_t result;
 744
 745	/* page table offset */
 746	result = pages_addr[addr >> PAGE_SHIFT];
 747
 748	/* in case cpu page size != gpu page size*/
 749	result |= addr & (~PAGE_MASK);
 750
 751	result &= 0xFFFFFFFFFFFFF000ULL;
 752
 753	return result;
 754}
 755
 756/**
 757 * amdgpu_vm_update_pdes - make sure that all directories are valid
 758 *
 759 * @adev: amdgpu_device pointer
 760 * @vm: requested vm
 761 * @immediate: submit immediately to the paging queue
 762 *
 763 * Makes sure all directories are up to date.
 764 *
 765 * Returns:
 766 * 0 for success, error for failure.
 767 */
 768int amdgpu_vm_update_pdes(struct amdgpu_device *adev,
 769			  struct amdgpu_vm *vm, bool immediate)
 770{
 771	struct amdgpu_vm_update_params params;
 772	struct amdgpu_vm_bo_base *entry;
 773	bool flush_tlb_needed = false;
 774	LIST_HEAD(relocated);
 775	int r, idx;
 776
 777	spin_lock(&vm->status_lock);
 778	list_splice_init(&vm->relocated, &relocated);
 779	spin_unlock(&vm->status_lock);
 780
 781	if (list_empty(&relocated))
 782		return 0;
 783
 784	if (!drm_dev_enter(adev_to_drm(adev), &idx))
 785		return -ENODEV;
 786
 787	memset(&params, 0, sizeof(params));
 788	params.adev = adev;
 789	params.vm = vm;
 790	params.immediate = immediate;
 791
 792	r = vm->update_funcs->prepare(&params, NULL, AMDGPU_SYNC_EXPLICIT);
 793	if (r)
 794		goto error;
 795
 796	list_for_each_entry(entry, &relocated, vm_status) {
 797		/* vm_flush_needed after updating moved PDEs */
 798		flush_tlb_needed |= entry->moved;
 799
 800		r = amdgpu_vm_pde_update(&params, entry);
 801		if (r)
 802			goto error;
 803	}
 804
 805	r = vm->update_funcs->commit(&params, &vm->last_update);
 806	if (r)
 807		goto error;
 808
 809	if (flush_tlb_needed)
 810		atomic64_inc(&vm->tlb_seq);
 811
 812	while (!list_empty(&relocated)) {
 813		entry = list_first_entry(&relocated, struct amdgpu_vm_bo_base,
 814					 vm_status);
 815		amdgpu_vm_bo_idle(entry);
 816	}
 817
 818error:
 819	drm_dev_exit(idx);
 820	return r;
 821}
 822
 823/**
 824 * amdgpu_vm_tlb_seq_cb - make sure to increment tlb sequence
 825 * @fence: unused
 826 * @cb: the callback structure
 827 *
 828 * Increments the tlb sequence to make sure that future CS execute a VM flush.
 829 */
 830static void amdgpu_vm_tlb_seq_cb(struct dma_fence *fence,
 831				 struct dma_fence_cb *cb)
 832{
 833	struct amdgpu_vm_tlb_seq_struct *tlb_cb;
 834
 835	tlb_cb = container_of(cb, typeof(*tlb_cb), cb);
 836	atomic64_inc(&tlb_cb->vm->tlb_seq);
 837	kfree(tlb_cb);
 838}
 839
 840/**
 841 * amdgpu_vm_update_range - update a range in the vm page table
 842 *
 843 * @adev: amdgpu_device pointer to use for commands
 844 * @vm: the VM to update the range
 845 * @immediate: immediate submission in a page fault
 846 * @unlocked: unlocked invalidation during MM callback
 847 * @flush_tlb: trigger tlb invalidation after update completed
 848 * @allow_override: change MTYPE for local NUMA nodes
 849 * @resv: fences we need to sync to
 850 * @start: start of mapped range
 851 * @last: last mapped entry
 852 * @flags: flags for the entries
 853 * @offset: offset into nodes and pages_addr
 854 * @vram_base: base for vram mappings
 855 * @res: ttm_resource to map
 856 * @pages_addr: DMA addresses to use for mapping
 857 * @fence: optional resulting fence
 858 *
 859 * Fill in the page table entries between @start and @last.
 860 *
 861 * Returns:
 862 * 0 for success, negative erro code for failure.
 863 */
 864int amdgpu_vm_update_range(struct amdgpu_device *adev, struct amdgpu_vm *vm,
 865			   bool immediate, bool unlocked, bool flush_tlb, bool allow_override,
 866			   struct dma_resv *resv, uint64_t start, uint64_t last,
 867			   uint64_t flags, uint64_t offset, uint64_t vram_base,
 868			   struct ttm_resource *res, dma_addr_t *pages_addr,
 869			   struct dma_fence **fence)
 870{
 871	struct amdgpu_vm_update_params params;
 872	struct amdgpu_vm_tlb_seq_struct *tlb_cb;
 873	struct amdgpu_res_cursor cursor;
 874	enum amdgpu_sync_mode sync_mode;
 875	int r, idx;
 876
 877	if (!drm_dev_enter(adev_to_drm(adev), &idx))
 878		return -ENODEV;
 879
 880	tlb_cb = kmalloc(sizeof(*tlb_cb), GFP_KERNEL);
 881	if (!tlb_cb) {
 882		r = -ENOMEM;
 883		goto error_unlock;
 884	}
 885
 886	/* Vega20+XGMI where PTEs get inadvertently cached in L2 texture cache,
 887	 * heavy-weight flush TLB unconditionally.
 888	 */
 889	flush_tlb |= adev->gmc.xgmi.num_physical_nodes &&
 890		     amdgpu_ip_version(adev, GC_HWIP, 0) == IP_VERSION(9, 4, 0);
 891
 892	/*
 893	 * On GFX8 and older any 8 PTE block with a valid bit set enters the TLB
 894	 */
 895	flush_tlb |= amdgpu_ip_version(adev, GC_HWIP, 0) < IP_VERSION(9, 0, 0);
 896
 897	memset(&params, 0, sizeof(params));
 898	params.adev = adev;
 899	params.vm = vm;
 900	params.immediate = immediate;
 901	params.pages_addr = pages_addr;
 902	params.unlocked = unlocked;
 903	params.allow_override = allow_override;
 904
 905	/* Implicitly sync to command submissions in the same VM before
 906	 * unmapping. Sync to moving fences before mapping.
 907	 */
 908	if (!(flags & AMDGPU_PTE_VALID))
 909		sync_mode = AMDGPU_SYNC_EQ_OWNER;
 910	else
 911		sync_mode = AMDGPU_SYNC_EXPLICIT;
 912
 913	amdgpu_vm_eviction_lock(vm);
 914	if (vm->evicting) {
 915		r = -EBUSY;
 916		goto error_free;
 917	}
 918
 919	if (!unlocked && !dma_fence_is_signaled(vm->last_unlocked)) {
 920		struct dma_fence *tmp = dma_fence_get_stub();
 921
 922		amdgpu_bo_fence(vm->root.bo, vm->last_unlocked, true);
 923		swap(vm->last_unlocked, tmp);
 924		dma_fence_put(tmp);
 925	}
 926
 927	r = vm->update_funcs->prepare(&params, resv, sync_mode);
 928	if (r)
 929		goto error_free;
 930
 931	amdgpu_res_first(pages_addr ? NULL : res, offset,
 932			 (last - start + 1) * AMDGPU_GPU_PAGE_SIZE, &cursor);
 933	while (cursor.remaining) {
 934		uint64_t tmp, num_entries, addr;
 935
 936		num_entries = cursor.size >> AMDGPU_GPU_PAGE_SHIFT;
 937		if (pages_addr) {
 938			bool contiguous = true;
 939
 940			if (num_entries > AMDGPU_GPU_PAGES_IN_CPU_PAGE) {
 941				uint64_t pfn = cursor.start >> PAGE_SHIFT;
 942				uint64_t count;
 943
 944				contiguous = pages_addr[pfn + 1] ==
 945					pages_addr[pfn] + PAGE_SIZE;
 946
 947				tmp = num_entries /
 948					AMDGPU_GPU_PAGES_IN_CPU_PAGE;
 949				for (count = 2; count < tmp; ++count) {
 950					uint64_t idx = pfn + count;
 951
 952					if (contiguous != (pages_addr[idx] ==
 953					    pages_addr[idx - 1] + PAGE_SIZE))
 954						break;
 955				}
 956				if (!contiguous)
 957					count--;
 958				num_entries = count *
 959					AMDGPU_GPU_PAGES_IN_CPU_PAGE;
 960			}
 961
 962			if (!contiguous) {
 963				addr = cursor.start;
 964				params.pages_addr = pages_addr;
 965			} else {
 966				addr = pages_addr[cursor.start >> PAGE_SHIFT];
 967				params.pages_addr = NULL;
 968			}
 969
 970		} else if (flags & (AMDGPU_PTE_VALID | AMDGPU_PTE_PRT)) {
 971			addr = vram_base + cursor.start;
 972		} else {
 973			addr = 0;
 974		}
 975
 976		tmp = start + num_entries;
 977		r = amdgpu_vm_ptes_update(&params, start, tmp, addr, flags);
 978		if (r)
 979			goto error_free;
 980
 981		amdgpu_res_next(&cursor, num_entries * AMDGPU_GPU_PAGE_SIZE);
 982		start = tmp;
 983	}
 984
 985	r = vm->update_funcs->commit(&params, fence);
 986
 987	if (flush_tlb || params.table_freed) {
 988		tlb_cb->vm = vm;
 989		if (fence && *fence &&
 990		    !dma_fence_add_callback(*fence, &tlb_cb->cb,
 991					   amdgpu_vm_tlb_seq_cb)) {
 992			dma_fence_put(vm->last_tlb_flush);
 993			vm->last_tlb_flush = dma_fence_get(*fence);
 994		} else {
 995			amdgpu_vm_tlb_seq_cb(NULL, &tlb_cb->cb);
 996		}
 997		tlb_cb = NULL;
 998	}
 999
1000error_free:
1001	kfree(tlb_cb);
1002
1003error_unlock:
1004	amdgpu_vm_eviction_unlock(vm);
1005	drm_dev_exit(idx);
1006	return r;
1007}
1008
1009static void amdgpu_vm_bo_get_memory(struct amdgpu_bo_va *bo_va,
1010				    struct amdgpu_mem_stats *stats)
1011{
1012	struct amdgpu_vm *vm = bo_va->base.vm;
1013	struct amdgpu_bo *bo = bo_va->base.bo;
1014
1015	if (!bo)
1016		return;
1017
1018	/*
1019	 * For now ignore BOs which are currently locked and potentially
1020	 * changing their location.
1021	 */
1022	if (bo->tbo.base.resv != vm->root.bo->tbo.base.resv &&
1023	    !dma_resv_trylock(bo->tbo.base.resv))
1024		return;
1025
1026	amdgpu_bo_get_memory(bo, stats);
1027	if (bo->tbo.base.resv != vm->root.bo->tbo.base.resv)
1028	    dma_resv_unlock(bo->tbo.base.resv);
1029}
1030
1031void amdgpu_vm_get_memory(struct amdgpu_vm *vm,
1032			  struct amdgpu_mem_stats *stats)
1033{
1034	struct amdgpu_bo_va *bo_va, *tmp;
1035
1036	spin_lock(&vm->status_lock);
1037	list_for_each_entry_safe(bo_va, tmp, &vm->idle, base.vm_status)
1038		amdgpu_vm_bo_get_memory(bo_va, stats);
1039
1040	list_for_each_entry_safe(bo_va, tmp, &vm->evicted, base.vm_status)
1041		amdgpu_vm_bo_get_memory(bo_va, stats);
1042
1043	list_for_each_entry_safe(bo_va, tmp, &vm->relocated, base.vm_status)
1044		amdgpu_vm_bo_get_memory(bo_va, stats);
1045
1046	list_for_each_entry_safe(bo_va, tmp, &vm->moved, base.vm_status)
1047		amdgpu_vm_bo_get_memory(bo_va, stats);
1048
1049	list_for_each_entry_safe(bo_va, tmp, &vm->invalidated, base.vm_status)
1050		amdgpu_vm_bo_get_memory(bo_va, stats);
1051
1052	list_for_each_entry_safe(bo_va, tmp, &vm->done, base.vm_status)
1053		amdgpu_vm_bo_get_memory(bo_va, stats);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1054	spin_unlock(&vm->status_lock);
1055}
1056
1057/**
1058 * amdgpu_vm_bo_update - update all BO mappings in the vm page table
1059 *
1060 * @adev: amdgpu_device pointer
1061 * @bo_va: requested BO and VM object
1062 * @clear: if true clear the entries
1063 *
1064 * Fill in the page table entries for @bo_va.
1065 *
1066 * Returns:
1067 * 0 for success, -EINVAL for failure.
1068 */
1069int amdgpu_vm_bo_update(struct amdgpu_device *adev, struct amdgpu_bo_va *bo_va,
1070			bool clear)
1071{
1072	struct amdgpu_bo *bo = bo_va->base.bo;
1073	struct amdgpu_vm *vm = bo_va->base.vm;
1074	struct amdgpu_bo_va_mapping *mapping;
1075	dma_addr_t *pages_addr = NULL;
1076	struct ttm_resource *mem;
1077	struct dma_fence **last_update;
1078	bool flush_tlb = clear;
1079	bool uncached;
1080	struct dma_resv *resv;
1081	uint64_t vram_base;
1082	uint64_t flags;
1083	int r;
1084
1085	if (clear || !bo) {
1086		mem = NULL;
1087		resv = vm->root.bo->tbo.base.resv;
1088	} else {
1089		struct drm_gem_object *obj = &bo->tbo.base;
1090
1091		resv = bo->tbo.base.resv;
1092		if (obj->import_attach && bo_va->is_xgmi) {
1093			struct dma_buf *dma_buf = obj->import_attach->dmabuf;
1094			struct drm_gem_object *gobj = dma_buf->priv;
1095			struct amdgpu_bo *abo = gem_to_amdgpu_bo(gobj);
1096
1097			if (abo->tbo.resource &&
1098			    abo->tbo.resource->mem_type == TTM_PL_VRAM)
1099				bo = gem_to_amdgpu_bo(gobj);
1100		}
1101		mem = bo->tbo.resource;
1102		if (mem && (mem->mem_type == TTM_PL_TT ||
1103			    mem->mem_type == AMDGPU_PL_PREEMPT))
1104			pages_addr = bo->tbo.ttm->dma_address;
1105	}
1106
1107	if (bo) {
1108		struct amdgpu_device *bo_adev;
1109
1110		flags = amdgpu_ttm_tt_pte_flags(adev, bo->tbo.ttm, mem);
1111
1112		if (amdgpu_bo_encrypted(bo))
1113			flags |= AMDGPU_PTE_TMZ;
1114
1115		bo_adev = amdgpu_ttm_adev(bo->tbo.bdev);
1116		vram_base = bo_adev->vm_manager.vram_base_offset;
1117		uncached = (bo->flags & AMDGPU_GEM_CREATE_UNCACHED) != 0;
1118	} else {
1119		flags = 0x0;
1120		vram_base = 0;
1121		uncached = false;
1122	}
1123
1124	if (clear || (bo && bo->tbo.base.resv ==
1125		      vm->root.bo->tbo.base.resv))
1126		last_update = &vm->last_update;
1127	else
1128		last_update = &bo_va->last_pt_update;
1129
1130	if (!clear && bo_va->base.moved) {
1131		flush_tlb = true;
1132		list_splice_init(&bo_va->valids, &bo_va->invalids);
1133
1134	} else if (bo_va->cleared != clear) {
1135		list_splice_init(&bo_va->valids, &bo_va->invalids);
1136	}
1137
1138	list_for_each_entry(mapping, &bo_va->invalids, list) {
1139		uint64_t update_flags = flags;
1140
1141		/* normally,bo_va->flags only contians READABLE and WIRTEABLE bit go here
1142		 * but in case of something, we filter the flags in first place
1143		 */
1144		if (!(mapping->flags & AMDGPU_PTE_READABLE))
1145			update_flags &= ~AMDGPU_PTE_READABLE;
1146		if (!(mapping->flags & AMDGPU_PTE_WRITEABLE))
1147			update_flags &= ~AMDGPU_PTE_WRITEABLE;
1148
1149		/* Apply ASIC specific mapping flags */
1150		amdgpu_gmc_get_vm_pte(adev, mapping, &update_flags);
1151
1152		trace_amdgpu_vm_bo_update(mapping);
1153
1154		r = amdgpu_vm_update_range(adev, vm, false, false, flush_tlb,
1155					   !uncached, resv, mapping->start, mapping->last,
1156					   update_flags, mapping->offset,
1157					   vram_base, mem, pages_addr,
1158					   last_update);
1159		if (r)
1160			return r;
1161	}
1162
1163	/* If the BO is not in its preferred location add it back to
1164	 * the evicted list so that it gets validated again on the
1165	 * next command submission.
1166	 */
1167	if (bo && bo->tbo.base.resv == vm->root.bo->tbo.base.resv) {
1168		uint32_t mem_type = bo->tbo.resource->mem_type;
1169
1170		if (!(bo->preferred_domains &
1171		      amdgpu_mem_type_to_domain(mem_type)))
1172			amdgpu_vm_bo_evicted(&bo_va->base);
1173		else
1174			amdgpu_vm_bo_idle(&bo_va->base);
1175	} else {
1176		amdgpu_vm_bo_done(&bo_va->base);
1177	}
1178
1179	list_splice_init(&bo_va->invalids, &bo_va->valids);
1180	bo_va->cleared = clear;
1181	bo_va->base.moved = false;
1182
1183	if (trace_amdgpu_vm_bo_mapping_enabled()) {
1184		list_for_each_entry(mapping, &bo_va->valids, list)
1185			trace_amdgpu_vm_bo_mapping(mapping);
1186	}
1187
1188	return 0;
1189}
1190
1191/**
1192 * amdgpu_vm_update_prt_state - update the global PRT state
1193 *
1194 * @adev: amdgpu_device pointer
1195 */
1196static void amdgpu_vm_update_prt_state(struct amdgpu_device *adev)
1197{
1198	unsigned long flags;
1199	bool enable;
1200
1201	spin_lock_irqsave(&adev->vm_manager.prt_lock, flags);
1202	enable = !!atomic_read(&adev->vm_manager.num_prt_users);
1203	adev->gmc.gmc_funcs->set_prt(adev, enable);
1204	spin_unlock_irqrestore(&adev->vm_manager.prt_lock, flags);
1205}
1206
1207/**
1208 * amdgpu_vm_prt_get - add a PRT user
1209 *
1210 * @adev: amdgpu_device pointer
1211 */
1212static void amdgpu_vm_prt_get(struct amdgpu_device *adev)
1213{
1214	if (!adev->gmc.gmc_funcs->set_prt)
1215		return;
1216
1217	if (atomic_inc_return(&adev->vm_manager.num_prt_users) == 1)
1218		amdgpu_vm_update_prt_state(adev);
1219}
1220
1221/**
1222 * amdgpu_vm_prt_put - drop a PRT user
1223 *
1224 * @adev: amdgpu_device pointer
1225 */
1226static void amdgpu_vm_prt_put(struct amdgpu_device *adev)
1227{
1228	if (atomic_dec_return(&adev->vm_manager.num_prt_users) == 0)
1229		amdgpu_vm_update_prt_state(adev);
1230}
1231
1232/**
1233 * amdgpu_vm_prt_cb - callback for updating the PRT status
1234 *
1235 * @fence: fence for the callback
1236 * @_cb: the callback function
1237 */
1238static void amdgpu_vm_prt_cb(struct dma_fence *fence, struct dma_fence_cb *_cb)
1239{
1240	struct amdgpu_prt_cb *cb = container_of(_cb, struct amdgpu_prt_cb, cb);
1241
1242	amdgpu_vm_prt_put(cb->adev);
1243	kfree(cb);
1244}
1245
1246/**
1247 * amdgpu_vm_add_prt_cb - add callback for updating the PRT status
1248 *
1249 * @adev: amdgpu_device pointer
1250 * @fence: fence for the callback
1251 */
1252static void amdgpu_vm_add_prt_cb(struct amdgpu_device *adev,
1253				 struct dma_fence *fence)
1254{
1255	struct amdgpu_prt_cb *cb;
1256
1257	if (!adev->gmc.gmc_funcs->set_prt)
1258		return;
1259
1260	cb = kmalloc(sizeof(struct amdgpu_prt_cb), GFP_KERNEL);
1261	if (!cb) {
1262		/* Last resort when we are OOM */
1263		if (fence)
1264			dma_fence_wait(fence, false);
1265
1266		amdgpu_vm_prt_put(adev);
1267	} else {
1268		cb->adev = adev;
1269		if (!fence || dma_fence_add_callback(fence, &cb->cb,
1270						     amdgpu_vm_prt_cb))
1271			amdgpu_vm_prt_cb(fence, &cb->cb);
1272	}
1273}
1274
1275/**
1276 * amdgpu_vm_free_mapping - free a mapping
1277 *
1278 * @adev: amdgpu_device pointer
1279 * @vm: requested vm
1280 * @mapping: mapping to be freed
1281 * @fence: fence of the unmap operation
1282 *
1283 * Free a mapping and make sure we decrease the PRT usage count if applicable.
1284 */
1285static void amdgpu_vm_free_mapping(struct amdgpu_device *adev,
1286				   struct amdgpu_vm *vm,
1287				   struct amdgpu_bo_va_mapping *mapping,
1288				   struct dma_fence *fence)
1289{
1290	if (mapping->flags & AMDGPU_PTE_PRT)
1291		amdgpu_vm_add_prt_cb(adev, fence);
1292	kfree(mapping);
1293}
1294
1295/**
1296 * amdgpu_vm_prt_fini - finish all prt mappings
1297 *
1298 * @adev: amdgpu_device pointer
1299 * @vm: requested vm
1300 *
1301 * Register a cleanup callback to disable PRT support after VM dies.
1302 */
1303static void amdgpu_vm_prt_fini(struct amdgpu_device *adev, struct amdgpu_vm *vm)
1304{
1305	struct dma_resv *resv = vm->root.bo->tbo.base.resv;
1306	struct dma_resv_iter cursor;
1307	struct dma_fence *fence;
1308
1309	dma_resv_for_each_fence(&cursor, resv, DMA_RESV_USAGE_BOOKKEEP, fence) {
1310		/* Add a callback for each fence in the reservation object */
1311		amdgpu_vm_prt_get(adev);
1312		amdgpu_vm_add_prt_cb(adev, fence);
1313	}
1314}
1315
1316/**
1317 * amdgpu_vm_clear_freed - clear freed BOs in the PT
1318 *
1319 * @adev: amdgpu_device pointer
1320 * @vm: requested vm
1321 * @fence: optional resulting fence (unchanged if no work needed to be done
1322 * or if an error occurred)
1323 *
1324 * Make sure all freed BOs are cleared in the PT.
1325 * PTs have to be reserved and mutex must be locked!
1326 *
1327 * Returns:
1328 * 0 for success.
1329 *
1330 */
1331int amdgpu_vm_clear_freed(struct amdgpu_device *adev,
1332			  struct amdgpu_vm *vm,
1333			  struct dma_fence **fence)
1334{
1335	struct dma_resv *resv = vm->root.bo->tbo.base.resv;
1336	struct amdgpu_bo_va_mapping *mapping;
1337	uint64_t init_pte_value = 0;
1338	struct dma_fence *f = NULL;
1339	int r;
1340
1341	while (!list_empty(&vm->freed)) {
1342		mapping = list_first_entry(&vm->freed,
1343			struct amdgpu_bo_va_mapping, list);
1344		list_del(&mapping->list);
1345
1346		if (vm->pte_support_ats &&
1347		    mapping->start < AMDGPU_GMC_HOLE_START)
1348			init_pte_value = AMDGPU_PTE_DEFAULT_ATC;
1349
1350		r = amdgpu_vm_update_range(adev, vm, false, false, true, false,
1351					   resv, mapping->start, mapping->last,
1352					   init_pte_value, 0, 0, NULL, NULL,
1353					   &f);
1354		amdgpu_vm_free_mapping(adev, vm, mapping, f);
1355		if (r) {
1356			dma_fence_put(f);
1357			return r;
1358		}
1359	}
1360
1361	if (fence && f) {
1362		dma_fence_put(*fence);
1363		*fence = f;
1364	} else {
1365		dma_fence_put(f);
1366	}
1367
1368	return 0;
1369
1370}
1371
1372/**
1373 * amdgpu_vm_handle_moved - handle moved BOs in the PT
1374 *
1375 * @adev: amdgpu_device pointer
1376 * @vm: requested vm
1377 * @ticket: optional reservation ticket used to reserve the VM
1378 *
1379 * Make sure all BOs which are moved are updated in the PTs.
1380 *
1381 * Returns:
1382 * 0 for success.
1383 *
1384 * PTs have to be reserved!
1385 */
1386int amdgpu_vm_handle_moved(struct amdgpu_device *adev,
1387			   struct amdgpu_vm *vm,
1388			   struct ww_acquire_ctx *ticket)
1389{
1390	struct amdgpu_bo_va *bo_va;
1391	struct dma_resv *resv;
1392	bool clear, unlock;
1393	int r;
1394
1395	spin_lock(&vm->status_lock);
1396	while (!list_empty(&vm->moved)) {
1397		bo_va = list_first_entry(&vm->moved, struct amdgpu_bo_va,
1398					 base.vm_status);
1399		spin_unlock(&vm->status_lock);
1400
1401		/* Per VM BOs never need to bo cleared in the page tables */
1402		r = amdgpu_vm_bo_update(adev, bo_va, false);
1403		if (r)
1404			return r;
1405		spin_lock(&vm->status_lock);
1406	}
1407
1408	while (!list_empty(&vm->invalidated)) {
1409		bo_va = list_first_entry(&vm->invalidated, struct amdgpu_bo_va,
1410					 base.vm_status);
1411		resv = bo_va->base.bo->tbo.base.resv;
1412		spin_unlock(&vm->status_lock);
1413
1414		/* Try to reserve the BO to avoid clearing its ptes */
1415		if (!adev->debug_vm && dma_resv_trylock(resv)) {
1416			clear = false;
1417			unlock = true;
1418		/* The caller is already holding the reservation lock */
1419		} else if (ticket && dma_resv_locking_ctx(resv) == ticket) {
1420			clear = false;
1421			unlock = false;
1422		/* Somebody else is using the BO right now */
1423		} else {
1424			clear = true;
1425			unlock = false;
1426		}
1427
1428		r = amdgpu_vm_bo_update(adev, bo_va, clear);
1429		if (r)
1430			return r;
1431
1432		if (unlock)
1433			dma_resv_unlock(resv);
1434		spin_lock(&vm->status_lock);
1435	}
1436	spin_unlock(&vm->status_lock);
1437
1438	return 0;
1439}
1440
1441/**
1442 * amdgpu_vm_flush_compute_tlb - Flush TLB on compute VM
1443 *
1444 * @adev: amdgpu_device pointer
1445 * @vm: requested vm
1446 * @flush_type: flush type
1447 * @xcc_mask: mask of XCCs that belong to the compute partition in need of a TLB flush.
1448 *
1449 * Flush TLB if needed for a compute VM.
1450 *
1451 * Returns:
1452 * 0 for success.
1453 */
1454int amdgpu_vm_flush_compute_tlb(struct amdgpu_device *adev,
1455				struct amdgpu_vm *vm,
1456				uint32_t flush_type,
1457				uint32_t xcc_mask)
1458{
1459	uint64_t tlb_seq = amdgpu_vm_tlb_seq(vm);
1460	bool all_hub = false;
1461	int xcc = 0, r = 0;
1462
1463	WARN_ON_ONCE(!vm->is_compute_context);
1464
1465	/*
1466	 * It can be that we race and lose here, but that is extremely unlikely
1467	 * and the worst thing which could happen is that we flush the changes
1468	 * into the TLB once more which is harmless.
1469	 */
1470	if (atomic64_xchg(&vm->kfd_last_flushed_seq, tlb_seq) == tlb_seq)
1471		return 0;
1472
1473	if (adev->family == AMDGPU_FAMILY_AI ||
1474	    adev->family == AMDGPU_FAMILY_RV)
1475		all_hub = true;
1476
1477	for_each_inst(xcc, xcc_mask) {
1478		r = amdgpu_gmc_flush_gpu_tlb_pasid(adev, vm->pasid, flush_type,
1479						   all_hub, xcc);
1480		if (r)
1481			break;
1482	}
1483	return r;
1484}
1485
1486/**
1487 * amdgpu_vm_bo_add - add a bo to a specific vm
1488 *
1489 * @adev: amdgpu_device pointer
1490 * @vm: requested vm
1491 * @bo: amdgpu buffer object
1492 *
1493 * Add @bo into the requested vm.
1494 * Add @bo to the list of bos associated with the vm
1495 *
1496 * Returns:
1497 * Newly added bo_va or NULL for failure
1498 *
1499 * Object has to be reserved!
1500 */
1501struct amdgpu_bo_va *amdgpu_vm_bo_add(struct amdgpu_device *adev,
1502				      struct amdgpu_vm *vm,
1503				      struct amdgpu_bo *bo)
1504{
1505	struct amdgpu_bo_va *bo_va;
1506
1507	bo_va = kzalloc(sizeof(struct amdgpu_bo_va), GFP_KERNEL);
1508	if (bo_va == NULL) {
1509		return NULL;
1510	}
1511	amdgpu_vm_bo_base_init(&bo_va->base, vm, bo);
1512
1513	bo_va->ref_count = 1;
1514	bo_va->last_pt_update = dma_fence_get_stub();
1515	INIT_LIST_HEAD(&bo_va->valids);
1516	INIT_LIST_HEAD(&bo_va->invalids);
1517
1518	if (!bo)
1519		return bo_va;
1520
1521	dma_resv_assert_held(bo->tbo.base.resv);
1522	if (amdgpu_dmabuf_is_xgmi_accessible(adev, bo)) {
1523		bo_va->is_xgmi = true;
1524		/* Power up XGMI if it can be potentially used */
1525		amdgpu_xgmi_set_pstate(adev, AMDGPU_XGMI_PSTATE_MAX_VEGA20);
1526	}
1527
1528	return bo_va;
1529}
1530
1531
1532/**
1533 * amdgpu_vm_bo_insert_map - insert a new mapping
1534 *
1535 * @adev: amdgpu_device pointer
1536 * @bo_va: bo_va to store the address
1537 * @mapping: the mapping to insert
1538 *
1539 * Insert a new mapping into all structures.
1540 */
1541static void amdgpu_vm_bo_insert_map(struct amdgpu_device *adev,
1542				    struct amdgpu_bo_va *bo_va,
1543				    struct amdgpu_bo_va_mapping *mapping)
1544{
1545	struct amdgpu_vm *vm = bo_va->base.vm;
1546	struct amdgpu_bo *bo = bo_va->base.bo;
1547
1548	mapping->bo_va = bo_va;
1549	list_add(&mapping->list, &bo_va->invalids);
1550	amdgpu_vm_it_insert(mapping, &vm->va);
1551
1552	if (mapping->flags & AMDGPU_PTE_PRT)
1553		amdgpu_vm_prt_get(adev);
1554
1555	if (bo && bo->tbo.base.resv == vm->root.bo->tbo.base.resv &&
1556	    !bo_va->base.moved) {
1557		amdgpu_vm_bo_moved(&bo_va->base);
1558	}
1559	trace_amdgpu_vm_bo_map(bo_va, mapping);
1560}
1561
1562/**
1563 * amdgpu_vm_bo_map - map bo inside a vm
1564 *
1565 * @adev: amdgpu_device pointer
1566 * @bo_va: bo_va to store the address
1567 * @saddr: where to map the BO
1568 * @offset: requested offset in the BO
1569 * @size: BO size in bytes
1570 * @flags: attributes of pages (read/write/valid/etc.)
1571 *
1572 * Add a mapping of the BO at the specefied addr into the VM.
1573 *
1574 * Returns:
1575 * 0 for success, error for failure.
1576 *
1577 * Object has to be reserved and unreserved outside!
1578 */
1579int amdgpu_vm_bo_map(struct amdgpu_device *adev,
1580		     struct amdgpu_bo_va *bo_va,
1581		     uint64_t saddr, uint64_t offset,
1582		     uint64_t size, uint64_t flags)
1583{
1584	struct amdgpu_bo_va_mapping *mapping, *tmp;
1585	struct amdgpu_bo *bo = bo_va->base.bo;
1586	struct amdgpu_vm *vm = bo_va->base.vm;
1587	uint64_t eaddr;
1588
1589	/* validate the parameters */
1590	if (saddr & ~PAGE_MASK || offset & ~PAGE_MASK || size & ~PAGE_MASK)
1591		return -EINVAL;
1592	if (saddr + size <= saddr || offset + size <= offset)
1593		return -EINVAL;
1594
1595	/* make sure object fit at this offset */
1596	eaddr = saddr + size - 1;
1597	if ((bo && offset + size > amdgpu_bo_size(bo)) ||
 
1598	    (eaddr >= adev->vm_manager.max_pfn << AMDGPU_GPU_PAGE_SHIFT))
1599		return -EINVAL;
1600
1601	saddr /= AMDGPU_GPU_PAGE_SIZE;
1602	eaddr /= AMDGPU_GPU_PAGE_SIZE;
1603
1604	tmp = amdgpu_vm_it_iter_first(&vm->va, saddr, eaddr);
1605	if (tmp) {
1606		/* bo and tmp overlap, invalid addr */
1607		dev_err(adev->dev, "bo %p va 0x%010Lx-0x%010Lx conflict with "
1608			"0x%010Lx-0x%010Lx\n", bo, saddr, eaddr,
1609			tmp->start, tmp->last + 1);
1610		return -EINVAL;
1611	}
1612
1613	mapping = kmalloc(sizeof(*mapping), GFP_KERNEL);
1614	if (!mapping)
1615		return -ENOMEM;
1616
1617	mapping->start = saddr;
1618	mapping->last = eaddr;
1619	mapping->offset = offset;
1620	mapping->flags = flags;
1621
1622	amdgpu_vm_bo_insert_map(adev, bo_va, mapping);
1623
1624	return 0;
1625}
1626
1627/**
1628 * amdgpu_vm_bo_replace_map - map bo inside a vm, replacing existing mappings
1629 *
1630 * @adev: amdgpu_device pointer
1631 * @bo_va: bo_va to store the address
1632 * @saddr: where to map the BO
1633 * @offset: requested offset in the BO
1634 * @size: BO size in bytes
1635 * @flags: attributes of pages (read/write/valid/etc.)
1636 *
1637 * Add a mapping of the BO at the specefied addr into the VM. Replace existing
1638 * mappings as we do so.
1639 *
1640 * Returns:
1641 * 0 for success, error for failure.
1642 *
1643 * Object has to be reserved and unreserved outside!
1644 */
1645int amdgpu_vm_bo_replace_map(struct amdgpu_device *adev,
1646			     struct amdgpu_bo_va *bo_va,
1647			     uint64_t saddr, uint64_t offset,
1648			     uint64_t size, uint64_t flags)
1649{
1650	struct amdgpu_bo_va_mapping *mapping;
1651	struct amdgpu_bo *bo = bo_va->base.bo;
1652	uint64_t eaddr;
1653	int r;
1654
1655	/* validate the parameters */
1656	if (saddr & ~PAGE_MASK || offset & ~PAGE_MASK || size & ~PAGE_MASK)
1657		return -EINVAL;
1658	if (saddr + size <= saddr || offset + size <= offset)
1659		return -EINVAL;
1660
1661	/* make sure object fit at this offset */
1662	eaddr = saddr + size - 1;
1663	if ((bo && offset + size > amdgpu_bo_size(bo)) ||
 
1664	    (eaddr >= adev->vm_manager.max_pfn << AMDGPU_GPU_PAGE_SHIFT))
1665		return -EINVAL;
1666
1667	/* Allocate all the needed memory */
1668	mapping = kmalloc(sizeof(*mapping), GFP_KERNEL);
1669	if (!mapping)
1670		return -ENOMEM;
1671
1672	r = amdgpu_vm_bo_clear_mappings(adev, bo_va->base.vm, saddr, size);
1673	if (r) {
1674		kfree(mapping);
1675		return r;
1676	}
1677
1678	saddr /= AMDGPU_GPU_PAGE_SIZE;
1679	eaddr /= AMDGPU_GPU_PAGE_SIZE;
1680
1681	mapping->start = saddr;
1682	mapping->last = eaddr;
1683	mapping->offset = offset;
1684	mapping->flags = flags;
1685
1686	amdgpu_vm_bo_insert_map(adev, bo_va, mapping);
1687
1688	return 0;
1689}
1690
1691/**
1692 * amdgpu_vm_bo_unmap - remove bo mapping from vm
1693 *
1694 * @adev: amdgpu_device pointer
1695 * @bo_va: bo_va to remove the address from
1696 * @saddr: where to the BO is mapped
1697 *
1698 * Remove a mapping of the BO at the specefied addr from the VM.
1699 *
1700 * Returns:
1701 * 0 for success, error for failure.
1702 *
1703 * Object has to be reserved and unreserved outside!
1704 */
1705int amdgpu_vm_bo_unmap(struct amdgpu_device *adev,
1706		       struct amdgpu_bo_va *bo_va,
1707		       uint64_t saddr)
1708{
1709	struct amdgpu_bo_va_mapping *mapping;
1710	struct amdgpu_vm *vm = bo_va->base.vm;
1711	bool valid = true;
1712
1713	saddr /= AMDGPU_GPU_PAGE_SIZE;
1714
1715	list_for_each_entry(mapping, &bo_va->valids, list) {
1716		if (mapping->start == saddr)
1717			break;
1718	}
1719
1720	if (&mapping->list == &bo_va->valids) {
1721		valid = false;
1722
1723		list_for_each_entry(mapping, &bo_va->invalids, list) {
1724			if (mapping->start == saddr)
1725				break;
1726		}
1727
1728		if (&mapping->list == &bo_va->invalids)
1729			return -ENOENT;
1730	}
1731
1732	list_del(&mapping->list);
1733	amdgpu_vm_it_remove(mapping, &vm->va);
1734	mapping->bo_va = NULL;
1735	trace_amdgpu_vm_bo_unmap(bo_va, mapping);
1736
1737	if (valid)
1738		list_add(&mapping->list, &vm->freed);
1739	else
1740		amdgpu_vm_free_mapping(adev, vm, mapping,
1741				       bo_va->last_pt_update);
1742
1743	return 0;
1744}
1745
1746/**
1747 * amdgpu_vm_bo_clear_mappings - remove all mappings in a specific range
1748 *
1749 * @adev: amdgpu_device pointer
1750 * @vm: VM structure to use
1751 * @saddr: start of the range
1752 * @size: size of the range
1753 *
1754 * Remove all mappings in a range, split them as appropriate.
1755 *
1756 * Returns:
1757 * 0 for success, error for failure.
1758 */
1759int amdgpu_vm_bo_clear_mappings(struct amdgpu_device *adev,
1760				struct amdgpu_vm *vm,
1761				uint64_t saddr, uint64_t size)
1762{
1763	struct amdgpu_bo_va_mapping *before, *after, *tmp, *next;
1764	LIST_HEAD(removed);
1765	uint64_t eaddr;
1766
1767	eaddr = saddr + size - 1;
1768	saddr /= AMDGPU_GPU_PAGE_SIZE;
1769	eaddr /= AMDGPU_GPU_PAGE_SIZE;
1770
1771	/* Allocate all the needed memory */
1772	before = kzalloc(sizeof(*before), GFP_KERNEL);
1773	if (!before)
1774		return -ENOMEM;
1775	INIT_LIST_HEAD(&before->list);
1776
1777	after = kzalloc(sizeof(*after), GFP_KERNEL);
1778	if (!after) {
1779		kfree(before);
1780		return -ENOMEM;
1781	}
1782	INIT_LIST_HEAD(&after->list);
1783
1784	/* Now gather all removed mappings */
1785	tmp = amdgpu_vm_it_iter_first(&vm->va, saddr, eaddr);
1786	while (tmp) {
1787		/* Remember mapping split at the start */
1788		if (tmp->start < saddr) {
1789			before->start = tmp->start;
1790			before->last = saddr - 1;
1791			before->offset = tmp->offset;
1792			before->flags = tmp->flags;
1793			before->bo_va = tmp->bo_va;
1794			list_add(&before->list, &tmp->bo_va->invalids);
1795		}
1796
1797		/* Remember mapping split at the end */
1798		if (tmp->last > eaddr) {
1799			after->start = eaddr + 1;
1800			after->last = tmp->last;
1801			after->offset = tmp->offset;
1802			after->offset += (after->start - tmp->start) << PAGE_SHIFT;
1803			after->flags = tmp->flags;
1804			after->bo_va = tmp->bo_va;
1805			list_add(&after->list, &tmp->bo_va->invalids);
1806		}
1807
1808		list_del(&tmp->list);
1809		list_add(&tmp->list, &removed);
1810
1811		tmp = amdgpu_vm_it_iter_next(tmp, saddr, eaddr);
1812	}
1813
1814	/* And free them up */
1815	list_for_each_entry_safe(tmp, next, &removed, list) {
1816		amdgpu_vm_it_remove(tmp, &vm->va);
1817		list_del(&tmp->list);
1818
1819		if (tmp->start < saddr)
1820		    tmp->start = saddr;
1821		if (tmp->last > eaddr)
1822		    tmp->last = eaddr;
1823
1824		tmp->bo_va = NULL;
1825		list_add(&tmp->list, &vm->freed);
1826		trace_amdgpu_vm_bo_unmap(NULL, tmp);
1827	}
1828
1829	/* Insert partial mapping before the range */
1830	if (!list_empty(&before->list)) {
1831		struct amdgpu_bo *bo = before->bo_va->base.bo;
1832
1833		amdgpu_vm_it_insert(before, &vm->va);
1834		if (before->flags & AMDGPU_PTE_PRT)
1835			amdgpu_vm_prt_get(adev);
1836
1837		if (bo && bo->tbo.base.resv == vm->root.bo->tbo.base.resv &&
1838		    !before->bo_va->base.moved)
1839			amdgpu_vm_bo_moved(&before->bo_va->base);
1840	} else {
1841		kfree(before);
1842	}
1843
1844	/* Insert partial mapping after the range */
1845	if (!list_empty(&after->list)) {
1846		struct amdgpu_bo *bo = after->bo_va->base.bo;
1847
1848		amdgpu_vm_it_insert(after, &vm->va);
1849		if (after->flags & AMDGPU_PTE_PRT)
1850			amdgpu_vm_prt_get(adev);
1851
1852		if (bo && bo->tbo.base.resv == vm->root.bo->tbo.base.resv &&
1853		    !after->bo_va->base.moved)
1854			amdgpu_vm_bo_moved(&after->bo_va->base);
1855	} else {
1856		kfree(after);
1857	}
1858
1859	return 0;
1860}
1861
1862/**
1863 * amdgpu_vm_bo_lookup_mapping - find mapping by address
1864 *
1865 * @vm: the requested VM
1866 * @addr: the address
1867 *
1868 * Find a mapping by it's address.
1869 *
1870 * Returns:
1871 * The amdgpu_bo_va_mapping matching for addr or NULL
1872 *
1873 */
1874struct amdgpu_bo_va_mapping *amdgpu_vm_bo_lookup_mapping(struct amdgpu_vm *vm,
1875							 uint64_t addr)
1876{
1877	return amdgpu_vm_it_iter_first(&vm->va, addr, addr);
1878}
1879
1880/**
1881 * amdgpu_vm_bo_trace_cs - trace all reserved mappings
1882 *
1883 * @vm: the requested vm
1884 * @ticket: CS ticket
1885 *
1886 * Trace all mappings of BOs reserved during a command submission.
1887 */
1888void amdgpu_vm_bo_trace_cs(struct amdgpu_vm *vm, struct ww_acquire_ctx *ticket)
1889{
1890	struct amdgpu_bo_va_mapping *mapping;
1891
1892	if (!trace_amdgpu_vm_bo_cs_enabled())
1893		return;
1894
1895	for (mapping = amdgpu_vm_it_iter_first(&vm->va, 0, U64_MAX); mapping;
1896	     mapping = amdgpu_vm_it_iter_next(mapping, 0, U64_MAX)) {
1897		if (mapping->bo_va && mapping->bo_va->base.bo) {
1898			struct amdgpu_bo *bo;
1899
1900			bo = mapping->bo_va->base.bo;
1901			if (dma_resv_locking_ctx(bo->tbo.base.resv) !=
1902			    ticket)
1903				continue;
1904		}
1905
1906		trace_amdgpu_vm_bo_cs(mapping);
1907	}
1908}
1909
1910/**
1911 * amdgpu_vm_bo_del - remove a bo from a specific vm
1912 *
1913 * @adev: amdgpu_device pointer
1914 * @bo_va: requested bo_va
1915 *
1916 * Remove @bo_va->bo from the requested vm.
1917 *
1918 * Object have to be reserved!
1919 */
1920void amdgpu_vm_bo_del(struct amdgpu_device *adev,
1921		      struct amdgpu_bo_va *bo_va)
1922{
1923	struct amdgpu_bo_va_mapping *mapping, *next;
1924	struct amdgpu_bo *bo = bo_va->base.bo;
1925	struct amdgpu_vm *vm = bo_va->base.vm;
1926	struct amdgpu_vm_bo_base **base;
1927
1928	dma_resv_assert_held(vm->root.bo->tbo.base.resv);
1929
1930	if (bo) {
1931		dma_resv_assert_held(bo->tbo.base.resv);
1932		if (bo->tbo.base.resv == vm->root.bo->tbo.base.resv)
1933			ttm_bo_set_bulk_move(&bo->tbo, NULL);
1934
1935		for (base = &bo_va->base.bo->vm_bo; *base;
1936		     base = &(*base)->next) {
1937			if (*base != &bo_va->base)
1938				continue;
1939
1940			*base = bo_va->base.next;
1941			break;
1942		}
1943	}
1944
1945	spin_lock(&vm->status_lock);
1946	list_del(&bo_va->base.vm_status);
1947	spin_unlock(&vm->status_lock);
1948
1949	list_for_each_entry_safe(mapping, next, &bo_va->valids, list) {
1950		list_del(&mapping->list);
1951		amdgpu_vm_it_remove(mapping, &vm->va);
1952		mapping->bo_va = NULL;
1953		trace_amdgpu_vm_bo_unmap(bo_va, mapping);
1954		list_add(&mapping->list, &vm->freed);
1955	}
1956	list_for_each_entry_safe(mapping, next, &bo_va->invalids, list) {
1957		list_del(&mapping->list);
1958		amdgpu_vm_it_remove(mapping, &vm->va);
1959		amdgpu_vm_free_mapping(adev, vm, mapping,
1960				       bo_va->last_pt_update);
1961	}
1962
1963	dma_fence_put(bo_va->last_pt_update);
1964
1965	if (bo && bo_va->is_xgmi)
1966		amdgpu_xgmi_set_pstate(adev, AMDGPU_XGMI_PSTATE_MIN);
1967
1968	kfree(bo_va);
1969}
1970
1971/**
1972 * amdgpu_vm_evictable - check if we can evict a VM
1973 *
1974 * @bo: A page table of the VM.
1975 *
1976 * Check if it is possible to evict a VM.
1977 */
1978bool amdgpu_vm_evictable(struct amdgpu_bo *bo)
1979{
1980	struct amdgpu_vm_bo_base *bo_base = bo->vm_bo;
1981
1982	/* Page tables of a destroyed VM can go away immediately */
1983	if (!bo_base || !bo_base->vm)
1984		return true;
1985
1986	/* Don't evict VM page tables while they are busy */
1987	if (!dma_resv_test_signaled(bo->tbo.base.resv, DMA_RESV_USAGE_BOOKKEEP))
1988		return false;
1989
1990	/* Try to block ongoing updates */
1991	if (!amdgpu_vm_eviction_trylock(bo_base->vm))
1992		return false;
1993
1994	/* Don't evict VM page tables while they are updated */
1995	if (!dma_fence_is_signaled(bo_base->vm->last_unlocked)) {
1996		amdgpu_vm_eviction_unlock(bo_base->vm);
1997		return false;
1998	}
1999
2000	bo_base->vm->evicting = true;
2001	amdgpu_vm_eviction_unlock(bo_base->vm);
2002	return true;
2003}
2004
2005/**
2006 * amdgpu_vm_bo_invalidate - mark the bo as invalid
2007 *
2008 * @adev: amdgpu_device pointer
2009 * @bo: amdgpu buffer object
2010 * @evicted: is the BO evicted
2011 *
2012 * Mark @bo as invalid.
2013 */
2014void amdgpu_vm_bo_invalidate(struct amdgpu_device *adev,
2015			     struct amdgpu_bo *bo, bool evicted)
2016{
2017	struct amdgpu_vm_bo_base *bo_base;
2018
2019	/* shadow bo doesn't have bo base, its validation needs its parent */
2020	if (bo->parent && (amdgpu_bo_shadowed(bo->parent) == bo))
2021		bo = bo->parent;
2022
2023	for (bo_base = bo->vm_bo; bo_base; bo_base = bo_base->next) {
2024		struct amdgpu_vm *vm = bo_base->vm;
2025
2026		if (evicted && bo->tbo.base.resv == vm->root.bo->tbo.base.resv) {
2027			amdgpu_vm_bo_evicted(bo_base);
2028			continue;
2029		}
2030
2031		if (bo_base->moved)
2032			continue;
2033		bo_base->moved = true;
2034
2035		if (bo->tbo.type == ttm_bo_type_kernel)
2036			amdgpu_vm_bo_relocated(bo_base);
2037		else if (bo->tbo.base.resv == vm->root.bo->tbo.base.resv)
2038			amdgpu_vm_bo_moved(bo_base);
2039		else
2040			amdgpu_vm_bo_invalidated(bo_base);
2041	}
2042}
2043
2044/**
2045 * amdgpu_vm_get_block_size - calculate VM page table size as power of two
2046 *
2047 * @vm_size: VM size
2048 *
2049 * Returns:
2050 * VM page table as power of two
2051 */
2052static uint32_t amdgpu_vm_get_block_size(uint64_t vm_size)
2053{
2054	/* Total bits covered by PD + PTs */
2055	unsigned bits = ilog2(vm_size) + 18;
2056
2057	/* Make sure the PD is 4K in size up to 8GB address space.
2058	   Above that split equal between PD and PTs */
2059	if (vm_size <= 8)
2060		return (bits - 9);
2061	else
2062		return ((bits + 3) / 2);
2063}
2064
2065/**
2066 * amdgpu_vm_adjust_size - adjust vm size, block size and fragment size
2067 *
2068 * @adev: amdgpu_device pointer
2069 * @min_vm_size: the minimum vm size in GB if it's set auto
2070 * @fragment_size_default: Default PTE fragment size
2071 * @max_level: max VMPT level
2072 * @max_bits: max address space size in bits
2073 *
2074 */
2075void amdgpu_vm_adjust_size(struct amdgpu_device *adev, uint32_t min_vm_size,
2076			   uint32_t fragment_size_default, unsigned max_level,
2077			   unsigned max_bits)
2078{
2079	unsigned int max_size = 1 << (max_bits - 30);
2080	unsigned int vm_size;
2081	uint64_t tmp;
2082
2083	/* adjust vm size first */
2084	if (amdgpu_vm_size != -1) {
2085		vm_size = amdgpu_vm_size;
2086		if (vm_size > max_size) {
2087			dev_warn(adev->dev, "VM size (%d) too large, max is %u GB\n",
2088				 amdgpu_vm_size, max_size);
2089			vm_size = max_size;
2090		}
2091	} else {
2092		struct sysinfo si;
2093		unsigned int phys_ram_gb;
2094
2095		/* Optimal VM size depends on the amount of physical
2096		 * RAM available. Underlying requirements and
2097		 * assumptions:
2098		 *
2099		 *  - Need to map system memory and VRAM from all GPUs
2100		 *     - VRAM from other GPUs not known here
2101		 *     - Assume VRAM <= system memory
2102		 *  - On GFX8 and older, VM space can be segmented for
2103		 *    different MTYPEs
2104		 *  - Need to allow room for fragmentation, guard pages etc.
2105		 *
2106		 * This adds up to a rough guess of system memory x3.
2107		 * Round up to power of two to maximize the available
2108		 * VM size with the given page table size.
2109		 */
2110		si_meminfo(&si);
2111		phys_ram_gb = ((uint64_t)si.totalram * si.mem_unit +
2112			       (1 << 30) - 1) >> 30;
2113		vm_size = roundup_pow_of_two(
2114			min(max(phys_ram_gb * 3, min_vm_size), max_size));
2115	}
2116
2117	adev->vm_manager.max_pfn = (uint64_t)vm_size << 18;
2118
2119	tmp = roundup_pow_of_two(adev->vm_manager.max_pfn);
2120	if (amdgpu_vm_block_size != -1)
2121		tmp >>= amdgpu_vm_block_size - 9;
2122	tmp = DIV_ROUND_UP(fls64(tmp) - 1, 9) - 1;
2123	adev->vm_manager.num_level = min_t(unsigned int, max_level, tmp);
2124	switch (adev->vm_manager.num_level) {
2125	case 3:
2126		adev->vm_manager.root_level = AMDGPU_VM_PDB2;
2127		break;
2128	case 2:
2129		adev->vm_manager.root_level = AMDGPU_VM_PDB1;
2130		break;
2131	case 1:
2132		adev->vm_manager.root_level = AMDGPU_VM_PDB0;
2133		break;
2134	default:
2135		dev_err(adev->dev, "VMPT only supports 2~4+1 levels\n");
2136	}
2137	/* block size depends on vm size and hw setup*/
2138	if (amdgpu_vm_block_size != -1)
2139		adev->vm_manager.block_size =
2140			min((unsigned)amdgpu_vm_block_size, max_bits
2141			    - AMDGPU_GPU_PAGE_SHIFT
2142			    - 9 * adev->vm_manager.num_level);
2143	else if (adev->vm_manager.num_level > 1)
2144		adev->vm_manager.block_size = 9;
2145	else
2146		adev->vm_manager.block_size = amdgpu_vm_get_block_size(tmp);
2147
2148	if (amdgpu_vm_fragment_size == -1)
2149		adev->vm_manager.fragment_size = fragment_size_default;
2150	else
2151		adev->vm_manager.fragment_size = amdgpu_vm_fragment_size;
2152
2153	DRM_INFO("vm size is %u GB, %u levels, block size is %u-bit, fragment size is %u-bit\n",
2154		 vm_size, adev->vm_manager.num_level + 1,
2155		 adev->vm_manager.block_size,
2156		 adev->vm_manager.fragment_size);
2157}
2158
2159/**
2160 * amdgpu_vm_wait_idle - wait for the VM to become idle
2161 *
2162 * @vm: VM object to wait for
2163 * @timeout: timeout to wait for VM to become idle
2164 */
2165long amdgpu_vm_wait_idle(struct amdgpu_vm *vm, long timeout)
2166{
2167	timeout = dma_resv_wait_timeout(vm->root.bo->tbo.base.resv,
2168					DMA_RESV_USAGE_BOOKKEEP,
2169					true, timeout);
2170	if (timeout <= 0)
2171		return timeout;
2172
2173	return dma_fence_wait_timeout(vm->last_unlocked, true, timeout);
2174}
2175
2176/**
2177 * amdgpu_vm_init - initialize a vm instance
2178 *
2179 * @adev: amdgpu_device pointer
2180 * @vm: requested vm
2181 * @xcp_id: GPU partition selection id
2182 *
2183 * Init @vm fields.
2184 *
2185 * Returns:
2186 * 0 for success, error for failure.
2187 */
2188int amdgpu_vm_init(struct amdgpu_device *adev, struct amdgpu_vm *vm,
2189		   int32_t xcp_id)
2190{
2191	struct amdgpu_bo *root_bo;
2192	struct amdgpu_bo_vm *root;
2193	int r, i;
2194
2195	vm->va = RB_ROOT_CACHED;
2196	for (i = 0; i < AMDGPU_MAX_VMHUBS; i++)
2197		vm->reserved_vmid[i] = NULL;
2198	INIT_LIST_HEAD(&vm->evicted);
2199	INIT_LIST_HEAD(&vm->relocated);
2200	INIT_LIST_HEAD(&vm->moved);
2201	INIT_LIST_HEAD(&vm->idle);
2202	INIT_LIST_HEAD(&vm->invalidated);
2203	spin_lock_init(&vm->status_lock);
2204	INIT_LIST_HEAD(&vm->freed);
2205	INIT_LIST_HEAD(&vm->done);
2206	INIT_LIST_HEAD(&vm->pt_freed);
2207	INIT_WORK(&vm->pt_free_work, amdgpu_vm_pt_free_work);
2208	INIT_KFIFO(vm->faults);
2209
2210	r = amdgpu_vm_init_entities(adev, vm);
 
 
 
2211	if (r)
2212		return r;
2213
 
 
 
 
 
 
2214	vm->pte_support_ats = false;
2215	vm->is_compute_context = false;
2216
2217	vm->use_cpu_for_update = !!(adev->vm_manager.vm_update_mode &
2218				    AMDGPU_VM_USE_CPU_FOR_GFX);
2219
2220	DRM_DEBUG_DRIVER("VM update mode is %s\n",
2221			 vm->use_cpu_for_update ? "CPU" : "SDMA");
2222	WARN_ONCE((vm->use_cpu_for_update &&
2223		   !amdgpu_gmc_vram_full_visible(&adev->gmc)),
2224		  "CPU update of VM recommended only for large BAR system\n");
2225
2226	if (vm->use_cpu_for_update)
2227		vm->update_funcs = &amdgpu_vm_cpu_funcs;
2228	else
2229		vm->update_funcs = &amdgpu_vm_sdma_funcs;
2230
2231	vm->last_update = dma_fence_get_stub();
2232	vm->last_unlocked = dma_fence_get_stub();
2233	vm->last_tlb_flush = dma_fence_get_stub();
2234	vm->generation = 0;
2235
2236	mutex_init(&vm->eviction_lock);
2237	vm->evicting = false;
2238
2239	r = amdgpu_vm_pt_create(adev, vm, adev->vm_manager.root_level,
2240				false, &root, xcp_id);
2241	if (r)
2242		goto error_free_delayed;
2243
2244	root_bo = amdgpu_bo_ref(&root->bo);
2245	r = amdgpu_bo_reserve(root_bo, true);
2246	if (r) {
2247		amdgpu_bo_unref(&root->shadow);
2248		amdgpu_bo_unref(&root_bo);
2249		goto error_free_delayed;
2250	}
2251
2252	amdgpu_vm_bo_base_init(&vm->root, vm, root_bo);
2253	r = dma_resv_reserve_fences(root_bo->tbo.base.resv, 1);
2254	if (r)
2255		goto error_free_root;
 
 
2256
2257	r = amdgpu_vm_pt_clear(adev, vm, root, false);
2258	if (r)
2259		goto error_free_root;
2260
2261	amdgpu_bo_unreserve(vm->root.bo);
2262	amdgpu_bo_unref(&root_bo);
 
2263
2264	return 0;
2265
2266error_free_root:
2267	amdgpu_vm_pt_free_root(adev, vm);
2268	amdgpu_bo_unreserve(vm->root.bo);
 
 
 
2269	amdgpu_bo_unref(&root_bo);
 
2270
2271error_free_delayed:
2272	dma_fence_put(vm->last_tlb_flush);
2273	dma_fence_put(vm->last_unlocked);
2274	amdgpu_vm_fini_entities(vm);
 
 
 
2275
2276	return r;
2277}
2278
2279/**
2280 * amdgpu_vm_make_compute - Turn a GFX VM into a compute VM
2281 *
2282 * @adev: amdgpu_device pointer
2283 * @vm: requested vm
2284 *
2285 * This only works on GFX VMs that don't have any BOs added and no
2286 * page tables allocated yet.
2287 *
2288 * Changes the following VM parameters:
2289 * - use_cpu_for_update
2290 * - pte_supports_ats
2291 *
2292 * Reinitializes the page directory to reflect the changed ATS
2293 * setting.
2294 *
2295 * Returns:
2296 * 0 for success, -errno for errors.
2297 */
2298int amdgpu_vm_make_compute(struct amdgpu_device *adev, struct amdgpu_vm *vm)
2299{
2300	bool pte_support_ats = (adev->asic_type == CHIP_RAVEN);
2301	int r;
2302
2303	r = amdgpu_bo_reserve(vm->root.bo, true);
2304	if (r)
2305		return r;
2306
 
 
 
 
 
 
2307	/* Check if PD needs to be reinitialized and do it before
2308	 * changing any other state, in case it fails.
2309	 */
2310	if (pte_support_ats != vm->pte_support_ats) {
2311		/* Sanity checks */
2312		if (!amdgpu_vm_pt_is_root_clean(adev, vm)) {
2313			r = -EINVAL;
2314			goto unreserve_bo;
2315		}
2316
2317		vm->pte_support_ats = pte_support_ats;
2318		r = amdgpu_vm_pt_clear(adev, vm, to_amdgpu_bo_vm(vm->root.bo),
2319				       false);
2320		if (r)
2321			goto unreserve_bo;
2322	}
2323
2324	/* Update VM state */
2325	vm->use_cpu_for_update = !!(adev->vm_manager.vm_update_mode &
2326				    AMDGPU_VM_USE_CPU_FOR_COMPUTE);
2327	DRM_DEBUG_DRIVER("VM update mode is %s\n",
2328			 vm->use_cpu_for_update ? "CPU" : "SDMA");
2329	WARN_ONCE((vm->use_cpu_for_update &&
2330		   !amdgpu_gmc_vram_full_visible(&adev->gmc)),
2331		  "CPU update of VM recommended only for large BAR system\n");
2332
2333	if (vm->use_cpu_for_update) {
2334		/* Sync with last SDMA update/clear before switching to CPU */
2335		r = amdgpu_bo_sync_wait(vm->root.bo,
2336					AMDGPU_FENCE_OWNER_UNDEFINED, true);
2337		if (r)
2338			goto unreserve_bo;
2339
2340		vm->update_funcs = &amdgpu_vm_cpu_funcs;
2341		r = amdgpu_vm_pt_map_tables(adev, vm);
2342		if (r)
2343			goto unreserve_bo;
2344
2345	} else {
2346		vm->update_funcs = &amdgpu_vm_sdma_funcs;
2347	}
 
 
 
 
 
 
 
2348
2349	dma_fence_put(vm->last_update);
2350	vm->last_update = dma_fence_get_stub();
2351	vm->is_compute_context = true;
2352
2353	/* Free the shadow bo for compute VM */
2354	amdgpu_bo_unref(&to_amdgpu_bo_vm(vm->root.bo)->shadow);
2355
2356	goto unreserve_bo;
2357
2358unreserve_bo:
2359	amdgpu_bo_unreserve(vm->root.bo);
2360	return r;
2361}
2362
2363/**
2364 * amdgpu_vm_release_compute - release a compute vm
2365 * @adev: amdgpu_device pointer
2366 * @vm: a vm turned into compute vm by calling amdgpu_vm_make_compute
2367 *
2368 * This is a correspondant of amdgpu_vm_make_compute. It decouples compute
2369 * pasid from vm. Compute should stop use of vm after this call.
2370 */
2371void amdgpu_vm_release_compute(struct amdgpu_device *adev, struct amdgpu_vm *vm)
2372{
2373	amdgpu_vm_set_pasid(adev, vm, 0);
2374	vm->is_compute_context = false;
2375}
2376
2377/**
2378 * amdgpu_vm_fini - tear down a vm instance
2379 *
2380 * @adev: amdgpu_device pointer
2381 * @vm: requested vm
2382 *
2383 * Tear down @vm.
2384 * Unbind the VM and remove all bos from the vm bo list
2385 */
2386void amdgpu_vm_fini(struct amdgpu_device *adev, struct amdgpu_vm *vm)
2387{
2388	struct amdgpu_bo_va_mapping *mapping, *tmp;
2389	bool prt_fini_needed = !!adev->gmc.gmc_funcs->set_prt;
2390	struct amdgpu_bo *root;
2391	unsigned long flags;
2392	int i;
2393
2394	amdgpu_amdkfd_gpuvm_destroy_cb(adev, vm);
2395
2396	flush_work(&vm->pt_free_work);
2397
2398	root = amdgpu_bo_ref(vm->root.bo);
2399	amdgpu_bo_reserve(root, true);
2400	amdgpu_vm_set_pasid(adev, vm, 0);
2401	dma_fence_wait(vm->last_unlocked, false);
2402	dma_fence_put(vm->last_unlocked);
2403	dma_fence_wait(vm->last_tlb_flush, false);
2404	/* Make sure that all fence callbacks have completed */
2405	spin_lock_irqsave(vm->last_tlb_flush->lock, flags);
2406	spin_unlock_irqrestore(vm->last_tlb_flush->lock, flags);
2407	dma_fence_put(vm->last_tlb_flush);
2408
2409	list_for_each_entry_safe(mapping, tmp, &vm->freed, list) {
2410		if (mapping->flags & AMDGPU_PTE_PRT && prt_fini_needed) {
2411			amdgpu_vm_prt_fini(adev, vm);
2412			prt_fini_needed = false;
2413		}
2414
2415		list_del(&mapping->list);
2416		amdgpu_vm_free_mapping(adev, vm, mapping, NULL);
2417	}
2418
2419	amdgpu_vm_pt_free_root(adev, vm);
2420	amdgpu_bo_unreserve(root);
2421	amdgpu_bo_unref(&root);
2422	WARN_ON(vm->root.bo);
2423
2424	amdgpu_vm_fini_entities(vm);
 
2425
2426	if (!RB_EMPTY_ROOT(&vm->va.rb_root)) {
2427		dev_err(adev->dev, "still active bo inside vm\n");
2428	}
2429	rbtree_postorder_for_each_entry_safe(mapping, tmp,
2430					     &vm->va.rb_root, rb) {
2431		/* Don't remove the mapping here, we don't want to trigger a
2432		 * rebalance and the tree is about to be destroyed anyway.
2433		 */
2434		list_del(&mapping->list);
2435		kfree(mapping);
2436	}
2437
2438	dma_fence_put(vm->last_update);
2439
2440	for (i = 0; i < AMDGPU_MAX_VMHUBS; i++) {
2441		if (vm->reserved_vmid[i]) {
2442			amdgpu_vmid_free_reserved(adev, i);
2443			vm->reserved_vmid[i] = false;
2444		}
2445	}
2446
2447}
2448
2449/**
2450 * amdgpu_vm_manager_init - init the VM manager
2451 *
2452 * @adev: amdgpu_device pointer
2453 *
2454 * Initialize the VM manager structures
2455 */
2456void amdgpu_vm_manager_init(struct amdgpu_device *adev)
2457{
2458	unsigned i;
2459
2460	/* Concurrent flushes are only possible starting with Vega10 and
2461	 * are broken on Navi10 and Navi14.
2462	 */
2463	adev->vm_manager.concurrent_flush = !(adev->asic_type < CHIP_VEGA10 ||
2464					      adev->asic_type == CHIP_NAVI10 ||
2465					      adev->asic_type == CHIP_NAVI14);
2466	amdgpu_vmid_mgr_init(adev);
2467
2468	adev->vm_manager.fence_context =
2469		dma_fence_context_alloc(AMDGPU_MAX_RINGS);
2470	for (i = 0; i < AMDGPU_MAX_RINGS; ++i)
2471		adev->vm_manager.seqno[i] = 0;
2472
2473	spin_lock_init(&adev->vm_manager.prt_lock);
2474	atomic_set(&adev->vm_manager.num_prt_users, 0);
2475
2476	/* If not overridden by the user, by default, only in large BAR systems
2477	 * Compute VM tables will be updated by CPU
2478	 */
2479#ifdef CONFIG_X86_64
2480	if (amdgpu_vm_update_mode == -1) {
2481		/* For asic with VF MMIO access protection
2482		 * avoid using CPU for VM table updates
2483		 */
2484		if (amdgpu_gmc_vram_full_visible(&adev->gmc) &&
2485		    !amdgpu_sriov_vf_mmio_access_protection(adev))
2486			adev->vm_manager.vm_update_mode =
2487				AMDGPU_VM_USE_CPU_FOR_COMPUTE;
2488		else
2489			adev->vm_manager.vm_update_mode = 0;
2490	} else
2491		adev->vm_manager.vm_update_mode = amdgpu_vm_update_mode;
2492#else
2493	adev->vm_manager.vm_update_mode = 0;
2494#endif
2495
2496	xa_init_flags(&adev->vm_manager.pasids, XA_FLAGS_LOCK_IRQ);
2497}
2498
2499/**
2500 * amdgpu_vm_manager_fini - cleanup VM manager
2501 *
2502 * @adev: amdgpu_device pointer
2503 *
2504 * Cleanup the VM manager and free resources.
2505 */
2506void amdgpu_vm_manager_fini(struct amdgpu_device *adev)
2507{
2508	WARN_ON(!xa_empty(&adev->vm_manager.pasids));
2509	xa_destroy(&adev->vm_manager.pasids);
2510
2511	amdgpu_vmid_mgr_fini(adev);
2512}
2513
2514/**
2515 * amdgpu_vm_ioctl - Manages VMID reservation for vm hubs.
2516 *
2517 * @dev: drm device pointer
2518 * @data: drm_amdgpu_vm
2519 * @filp: drm file pointer
2520 *
2521 * Returns:
2522 * 0 for success, -errno for errors.
2523 */
2524int amdgpu_vm_ioctl(struct drm_device *dev, void *data, struct drm_file *filp)
2525{
2526	union drm_amdgpu_vm *args = data;
2527	struct amdgpu_device *adev = drm_to_adev(dev);
2528	struct amdgpu_fpriv *fpriv = filp->driver_priv;
2529
2530	/* No valid flags defined yet */
2531	if (args->in.flags)
2532		return -EINVAL;
2533
2534	switch (args->in.op) {
2535	case AMDGPU_VM_OP_RESERVE_VMID:
2536		/* We only have requirement to reserve vmid from gfxhub */
2537		if (!fpriv->vm.reserved_vmid[AMDGPU_GFXHUB(0)]) {
2538			amdgpu_vmid_alloc_reserved(adev, AMDGPU_GFXHUB(0));
2539			fpriv->vm.reserved_vmid[AMDGPU_GFXHUB(0)] = true;
2540		}
2541
2542		break;
2543	case AMDGPU_VM_OP_UNRESERVE_VMID:
2544		if (fpriv->vm.reserved_vmid[AMDGPU_GFXHUB(0)]) {
2545			amdgpu_vmid_free_reserved(adev, AMDGPU_GFXHUB(0));
2546			fpriv->vm.reserved_vmid[AMDGPU_GFXHUB(0)] = false;
2547		}
2548		break;
2549	default:
2550		return -EINVAL;
2551	}
2552
2553	return 0;
2554}
2555
2556/**
2557 * amdgpu_vm_get_task_info - Extracts task info for a PASID.
2558 *
2559 * @adev: drm device pointer
2560 * @pasid: PASID identifier for VM
2561 * @task_info: task_info to fill.
2562 */
2563void amdgpu_vm_get_task_info(struct amdgpu_device *adev, u32 pasid,
2564			 struct amdgpu_task_info *task_info)
2565{
2566	struct amdgpu_vm *vm;
2567	unsigned long flags;
2568
2569	xa_lock_irqsave(&adev->vm_manager.pasids, flags);
2570
2571	vm = xa_load(&adev->vm_manager.pasids, pasid);
2572	if (vm)
2573		*task_info = vm->task_info;
2574
2575	xa_unlock_irqrestore(&adev->vm_manager.pasids, flags);
2576}
2577
2578/**
2579 * amdgpu_vm_set_task_info - Sets VMs task info.
2580 *
2581 * @vm: vm for which to set the info
2582 */
2583void amdgpu_vm_set_task_info(struct amdgpu_vm *vm)
2584{
2585	if (vm->task_info.pid)
2586		return;
2587
2588	vm->task_info.pid = current->pid;
2589	get_task_comm(vm->task_info.task_name, current);
2590
2591	if (current->group_leader->mm != current->mm)
2592		return;
2593
2594	vm->task_info.tgid = current->group_leader->pid;
2595	get_task_comm(vm->task_info.process_name, current->group_leader);
2596}
2597
2598/**
2599 * amdgpu_vm_handle_fault - graceful handling of VM faults.
2600 * @adev: amdgpu device pointer
2601 * @pasid: PASID of the VM
2602 * @vmid: VMID, only used for GFX 9.4.3.
2603 * @node_id: Node_id received in IH cookie. Only applicable for
2604 *           GFX 9.4.3.
2605 * @addr: Address of the fault
2606 * @write_fault: true is write fault, false is read fault
2607 *
2608 * Try to gracefully handle a VM fault. Return true if the fault was handled and
2609 * shouldn't be reported any more.
2610 */
2611bool amdgpu_vm_handle_fault(struct amdgpu_device *adev, u32 pasid,
2612			    u32 vmid, u32 node_id, uint64_t addr,
2613			    bool write_fault)
2614{
2615	bool is_compute_context = false;
2616	struct amdgpu_bo *root;
2617	unsigned long irqflags;
2618	uint64_t value, flags;
2619	struct amdgpu_vm *vm;
2620	int r;
2621
2622	xa_lock_irqsave(&adev->vm_manager.pasids, irqflags);
2623	vm = xa_load(&adev->vm_manager.pasids, pasid);
2624	if (vm) {
2625		root = amdgpu_bo_ref(vm->root.bo);
2626		is_compute_context = vm->is_compute_context;
2627	} else {
2628		root = NULL;
2629	}
2630	xa_unlock_irqrestore(&adev->vm_manager.pasids, irqflags);
2631
2632	if (!root)
2633		return false;
2634
2635	addr /= AMDGPU_GPU_PAGE_SIZE;
2636
2637	if (is_compute_context && !svm_range_restore_pages(adev, pasid, vmid,
2638	    node_id, addr, write_fault)) {
2639		amdgpu_bo_unref(&root);
2640		return true;
2641	}
2642
2643	r = amdgpu_bo_reserve(root, true);
2644	if (r)
2645		goto error_unref;
2646
2647	/* Double check that the VM still exists */
2648	xa_lock_irqsave(&adev->vm_manager.pasids, irqflags);
2649	vm = xa_load(&adev->vm_manager.pasids, pasid);
2650	if (vm && vm->root.bo != root)
2651		vm = NULL;
2652	xa_unlock_irqrestore(&adev->vm_manager.pasids, irqflags);
2653	if (!vm)
2654		goto error_unlock;
2655
2656	flags = AMDGPU_PTE_VALID | AMDGPU_PTE_SNOOPED |
2657		AMDGPU_PTE_SYSTEM;
2658
2659	if (is_compute_context) {
2660		/* Intentionally setting invalid PTE flag
2661		 * combination to force a no-retry-fault
2662		 */
2663		flags = AMDGPU_VM_NORETRY_FLAGS;
2664		value = 0;
2665	} else if (amdgpu_vm_fault_stop == AMDGPU_VM_FAULT_STOP_NEVER) {
2666		/* Redirect the access to the dummy page */
2667		value = adev->dummy_page_addr;
2668		flags |= AMDGPU_PTE_EXECUTABLE | AMDGPU_PTE_READABLE |
2669			AMDGPU_PTE_WRITEABLE;
2670
2671	} else {
2672		/* Let the hw retry silently on the PTE */
2673		value = 0;
2674	}
2675
2676	r = dma_resv_reserve_fences(root->tbo.base.resv, 1);
2677	if (r) {
2678		pr_debug("failed %d to reserve fence slot\n", r);
2679		goto error_unlock;
2680	}
2681
2682	r = amdgpu_vm_update_range(adev, vm, true, false, false, false,
2683				   NULL, addr, addr, flags, value, 0, NULL, NULL, NULL);
2684	if (r)
2685		goto error_unlock;
2686
2687	r = amdgpu_vm_update_pdes(adev, vm, true);
2688
2689error_unlock:
2690	amdgpu_bo_unreserve(root);
2691	if (r < 0)
2692		DRM_ERROR("Can't handle page fault (%d)\n", r);
2693
2694error_unref:
2695	amdgpu_bo_unref(&root);
2696
2697	return false;
2698}
2699
2700#if defined(CONFIG_DEBUG_FS)
2701/**
2702 * amdgpu_debugfs_vm_bo_info  - print BO info for the VM
2703 *
2704 * @vm: Requested VM for printing BO info
2705 * @m: debugfs file
2706 *
2707 * Print BO information in debugfs file for the VM
2708 */
2709void amdgpu_debugfs_vm_bo_info(struct amdgpu_vm *vm, struct seq_file *m)
2710{
2711	struct amdgpu_bo_va *bo_va, *tmp;
2712	u64 total_idle = 0;
2713	u64 total_evicted = 0;
2714	u64 total_relocated = 0;
2715	u64 total_moved = 0;
2716	u64 total_invalidated = 0;
2717	u64 total_done = 0;
2718	unsigned int total_idle_objs = 0;
2719	unsigned int total_evicted_objs = 0;
2720	unsigned int total_relocated_objs = 0;
2721	unsigned int total_moved_objs = 0;
2722	unsigned int total_invalidated_objs = 0;
2723	unsigned int total_done_objs = 0;
2724	unsigned int id = 0;
2725
2726	spin_lock(&vm->status_lock);
2727	seq_puts(m, "\tIdle BOs:\n");
2728	list_for_each_entry_safe(bo_va, tmp, &vm->idle, base.vm_status) {
2729		if (!bo_va->base.bo)
2730			continue;
2731		total_idle += amdgpu_bo_print_info(id++, bo_va->base.bo, m);
2732	}
2733	total_idle_objs = id;
2734	id = 0;
2735
2736	seq_puts(m, "\tEvicted BOs:\n");
2737	list_for_each_entry_safe(bo_va, tmp, &vm->evicted, base.vm_status) {
2738		if (!bo_va->base.bo)
2739			continue;
2740		total_evicted += amdgpu_bo_print_info(id++, bo_va->base.bo, m);
2741	}
2742	total_evicted_objs = id;
2743	id = 0;
2744
2745	seq_puts(m, "\tRelocated BOs:\n");
2746	list_for_each_entry_safe(bo_va, tmp, &vm->relocated, base.vm_status) {
2747		if (!bo_va->base.bo)
2748			continue;
2749		total_relocated += amdgpu_bo_print_info(id++, bo_va->base.bo, m);
2750	}
2751	total_relocated_objs = id;
2752	id = 0;
2753
2754	seq_puts(m, "\tMoved BOs:\n");
2755	list_for_each_entry_safe(bo_va, tmp, &vm->moved, base.vm_status) {
2756		if (!bo_va->base.bo)
2757			continue;
2758		total_moved += amdgpu_bo_print_info(id++, bo_va->base.bo, m);
2759	}
2760	total_moved_objs = id;
2761	id = 0;
2762
2763	seq_puts(m, "\tInvalidated BOs:\n");
2764	list_for_each_entry_safe(bo_va, tmp, &vm->invalidated, base.vm_status) {
2765		if (!bo_va->base.bo)
2766			continue;
2767		total_invalidated += amdgpu_bo_print_info(id++,	bo_va->base.bo, m);
2768	}
2769	total_invalidated_objs = id;
2770	id = 0;
2771
2772	seq_puts(m, "\tDone BOs:\n");
2773	list_for_each_entry_safe(bo_va, tmp, &vm->done, base.vm_status) {
2774		if (!bo_va->base.bo)
2775			continue;
2776		total_done += amdgpu_bo_print_info(id++, bo_va->base.bo, m);
2777	}
2778	spin_unlock(&vm->status_lock);
2779	total_done_objs = id;
2780
2781	seq_printf(m, "\tTotal idle size:        %12lld\tobjs:\t%d\n", total_idle,
2782		   total_idle_objs);
2783	seq_printf(m, "\tTotal evicted size:     %12lld\tobjs:\t%d\n", total_evicted,
2784		   total_evicted_objs);
2785	seq_printf(m, "\tTotal relocated size:   %12lld\tobjs:\t%d\n", total_relocated,
2786		   total_relocated_objs);
2787	seq_printf(m, "\tTotal moved size:       %12lld\tobjs:\t%d\n", total_moved,
2788		   total_moved_objs);
2789	seq_printf(m, "\tTotal invalidated size: %12lld\tobjs:\t%d\n", total_invalidated,
2790		   total_invalidated_objs);
2791	seq_printf(m, "\tTotal done size:        %12lld\tobjs:\t%d\n", total_done,
2792		   total_done_objs);
2793}
2794#endif
2795
2796/**
2797 * amdgpu_vm_update_fault_cache - update cached fault into.
2798 * @adev: amdgpu device pointer
2799 * @pasid: PASID of the VM
2800 * @addr: Address of the fault
2801 * @status: GPUVM fault status register
2802 * @vmhub: which vmhub got the fault
2803 *
2804 * Cache the fault info for later use by userspace in debugging.
2805 */
2806void amdgpu_vm_update_fault_cache(struct amdgpu_device *adev,
2807				  unsigned int pasid,
2808				  uint64_t addr,
2809				  uint32_t status,
2810				  unsigned int vmhub)
2811{
2812	struct amdgpu_vm *vm;
2813	unsigned long flags;
2814
2815	xa_lock_irqsave(&adev->vm_manager.pasids, flags);
2816
2817	vm = xa_load(&adev->vm_manager.pasids, pasid);
2818	/* Don't update the fault cache if status is 0.  In the multiple
2819	 * fault case, subsequent faults will return a 0 status which is
2820	 * useless for userspace and replaces the useful fault status, so
2821	 * only update if status is non-0.
2822	 */
2823	if (vm && status) {
2824		vm->fault_info.addr = addr;
2825		vm->fault_info.status = status;
2826		if (AMDGPU_IS_GFXHUB(vmhub)) {
2827			vm->fault_info.vmhub = AMDGPU_VMHUB_TYPE_GFX;
2828			vm->fault_info.vmhub |=
2829				(vmhub - AMDGPU_GFXHUB_START) << AMDGPU_VMHUB_IDX_SHIFT;
2830		} else if (AMDGPU_IS_MMHUB0(vmhub)) {
2831			vm->fault_info.vmhub = AMDGPU_VMHUB_TYPE_MM0;
2832			vm->fault_info.vmhub |=
2833				(vmhub - AMDGPU_MMHUB0_START) << AMDGPU_VMHUB_IDX_SHIFT;
2834		} else if (AMDGPU_IS_MMHUB1(vmhub)) {
2835			vm->fault_info.vmhub = AMDGPU_VMHUB_TYPE_MM1;
2836			vm->fault_info.vmhub |=
2837				(vmhub - AMDGPU_MMHUB1_START) << AMDGPU_VMHUB_IDX_SHIFT;
2838		} else {
2839			WARN_ONCE(1, "Invalid vmhub %u\n", vmhub);
2840		}
2841	}
2842	xa_unlock_irqrestore(&adev->vm_manager.pasids, flags);
2843}
2844
v6.2
   1/*
   2 * Copyright 2008 Advanced Micro Devices, Inc.
   3 * Copyright 2008 Red Hat Inc.
   4 * Copyright 2009 Jerome Glisse.
   5 *
   6 * Permission is hereby granted, free of charge, to any person obtaining a
   7 * copy of this software and associated documentation files (the "Software"),
   8 * to deal in the Software without restriction, including without limitation
   9 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
  10 * and/or sell copies of the Software, and to permit persons to whom the
  11 * Software is furnished to do so, subject to the following conditions:
  12 *
  13 * The above copyright notice and this permission notice shall be included in
  14 * all copies or substantial portions of the Software.
  15 *
  16 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  17 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  18 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
  19 * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR
  20 * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
  21 * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
  22 * OTHER DEALINGS IN THE SOFTWARE.
  23 *
  24 * Authors: Dave Airlie
  25 *          Alex Deucher
  26 *          Jerome Glisse
  27 */
  28
  29#include <linux/dma-fence-array.h>
  30#include <linux/interval_tree_generic.h>
  31#include <linux/idr.h>
  32#include <linux/dma-buf.h>
  33
  34#include <drm/amdgpu_drm.h>
  35#include <drm/drm_drv.h>
 
 
  36#include "amdgpu.h"
  37#include "amdgpu_trace.h"
  38#include "amdgpu_amdkfd.h"
  39#include "amdgpu_gmc.h"
  40#include "amdgpu_xgmi.h"
  41#include "amdgpu_dma_buf.h"
  42#include "amdgpu_res_cursor.h"
  43#include "kfd_svm.h"
  44
  45/**
  46 * DOC: GPUVM
  47 *
  48 * GPUVM is the MMU functionality provided on the GPU.
  49 * GPUVM is similar to the legacy GART on older asics, however
  50 * rather than there being a single global GART table
  51 * for the entire GPU, there can be multiple GPUVM page tables active
  52 * at any given time.  The GPUVM page tables can contain a mix
  53 * VRAM pages and system pages (both memory and MMIO) and system pages
  54 * can be mapped as snooped (cached system pages) or unsnooped
  55 * (uncached system pages).
  56 *
  57 * Each active GPUVM has an ID associated with it and there is a page table
  58 * linked with each VMID.  When executing a command buffer,
  59 * the kernel tells the engine what VMID to use for that command
  60 * buffer.  VMIDs are allocated dynamically as commands are submitted.
  61 * The userspace drivers maintain their own address space and the kernel
  62 * sets up their pages tables accordingly when they submit their
  63 * command buffers and a VMID is assigned.
  64 * The hardware supports up to 16 active GPUVMs at any given time.
  65 *
  66 * Each GPUVM is represented by a 1-2 or 1-5 level page table, depending
  67 * on the ASIC family.  GPUVM supports RWX attributes on each page as well
  68 * as other features such as encryption and caching attributes.
  69 *
  70 * VMID 0 is special.  It is the GPUVM used for the kernel driver.  In
  71 * addition to an aperture managed by a page table, VMID 0 also has
  72 * several other apertures.  There is an aperture for direct access to VRAM
  73 * and there is a legacy AGP aperture which just forwards accesses directly
  74 * to the matching system physical addresses (or IOVAs when an IOMMU is
  75 * present).  These apertures provide direct access to these memories without
  76 * incurring the overhead of a page table.  VMID 0 is used by the kernel
  77 * driver for tasks like memory management.
  78 *
  79 * GPU clients (i.e., engines on the GPU) use GPUVM VMIDs to access memory.
  80 * For user applications, each application can have their own unique GPUVM
  81 * address space.  The application manages the address space and the kernel
  82 * driver manages the GPUVM page tables for each process.  If an GPU client
  83 * accesses an invalid page, it will generate a GPU page fault, similar to
  84 * accessing an invalid page on a CPU.
  85 */
  86
  87#define START(node) ((node)->start)
  88#define LAST(node) ((node)->last)
  89
  90INTERVAL_TREE_DEFINE(struct amdgpu_bo_va_mapping, rb, uint64_t, __subtree_last,
  91		     START, LAST, static, amdgpu_vm_it)
  92
  93#undef START
  94#undef LAST
  95
  96/**
  97 * struct amdgpu_prt_cb - Helper to disable partial resident texture feature from a fence callback
  98 */
  99struct amdgpu_prt_cb {
 100
 101	/**
 102	 * @adev: amdgpu device
 103	 */
 104	struct amdgpu_device *adev;
 105
 106	/**
 107	 * @cb: callback
 108	 */
 109	struct dma_fence_cb cb;
 110};
 111
 112/**
 113 * struct amdgpu_vm_tlb_seq_cb - Helper to increment the TLB flush sequence
 114 */
 115struct amdgpu_vm_tlb_seq_cb {
 116	/**
 117	 * @vm: pointer to the amdgpu_vm structure to set the fence sequence on
 118	 */
 119	struct amdgpu_vm *vm;
 120
 121	/**
 122	 * @cb: callback
 123	 */
 124	struct dma_fence_cb cb;
 125};
 126
 127/**
 128 * amdgpu_vm_set_pasid - manage pasid and vm ptr mapping
 129 *
 130 * @adev: amdgpu_device pointer
 131 * @vm: amdgpu_vm pointer
 132 * @pasid: the pasid the VM is using on this GPU
 133 *
 134 * Set the pasid this VM is using on this GPU, can also be used to remove the
 135 * pasid by passing in zero.
 136 *
 137 */
 138int amdgpu_vm_set_pasid(struct amdgpu_device *adev, struct amdgpu_vm *vm,
 139			u32 pasid)
 140{
 141	int r;
 142
 143	if (vm->pasid == pasid)
 144		return 0;
 145
 146	if (vm->pasid) {
 147		r = xa_err(xa_erase_irq(&adev->vm_manager.pasids, vm->pasid));
 148		if (r < 0)
 149			return r;
 150
 151		vm->pasid = 0;
 152	}
 153
 154	if (pasid) {
 155		r = xa_err(xa_store_irq(&adev->vm_manager.pasids, pasid, vm,
 156					GFP_KERNEL));
 157		if (r < 0)
 158			return r;
 159
 160		vm->pasid = pasid;
 161	}
 162
 163
 164	return 0;
 165}
 166
 167/**
 168 * amdgpu_vm_bo_evicted - vm_bo is evicted
 169 *
 170 * @vm_bo: vm_bo which is evicted
 171 *
 172 * State for PDs/PTs and per VM BOs which are not at the location they should
 173 * be.
 174 */
 175static void amdgpu_vm_bo_evicted(struct amdgpu_vm_bo_base *vm_bo)
 176{
 177	struct amdgpu_vm *vm = vm_bo->vm;
 178	struct amdgpu_bo *bo = vm_bo->bo;
 179
 180	vm_bo->moved = true;
 181	spin_lock(&vm_bo->vm->status_lock);
 182	if (bo->tbo.type == ttm_bo_type_kernel)
 183		list_move(&vm_bo->vm_status, &vm->evicted);
 184	else
 185		list_move_tail(&vm_bo->vm_status, &vm->evicted);
 186	spin_unlock(&vm_bo->vm->status_lock);
 187}
 188/**
 189 * amdgpu_vm_bo_moved - vm_bo is moved
 190 *
 191 * @vm_bo: vm_bo which is moved
 192 *
 193 * State for per VM BOs which are moved, but that change is not yet reflected
 194 * in the page tables.
 195 */
 196static void amdgpu_vm_bo_moved(struct amdgpu_vm_bo_base *vm_bo)
 197{
 198	spin_lock(&vm_bo->vm->status_lock);
 199	list_move(&vm_bo->vm_status, &vm_bo->vm->moved);
 200	spin_unlock(&vm_bo->vm->status_lock);
 201}
 202
 203/**
 204 * amdgpu_vm_bo_idle - vm_bo is idle
 205 *
 206 * @vm_bo: vm_bo which is now idle
 207 *
 208 * State for PDs/PTs and per VM BOs which have gone through the state machine
 209 * and are now idle.
 210 */
 211static void amdgpu_vm_bo_idle(struct amdgpu_vm_bo_base *vm_bo)
 212{
 213	spin_lock(&vm_bo->vm->status_lock);
 214	list_move(&vm_bo->vm_status, &vm_bo->vm->idle);
 215	spin_unlock(&vm_bo->vm->status_lock);
 216	vm_bo->moved = false;
 217}
 218
 219/**
 220 * amdgpu_vm_bo_invalidated - vm_bo is invalidated
 221 *
 222 * @vm_bo: vm_bo which is now invalidated
 223 *
 224 * State for normal BOs which are invalidated and that change not yet reflected
 225 * in the PTs.
 226 */
 227static void amdgpu_vm_bo_invalidated(struct amdgpu_vm_bo_base *vm_bo)
 228{
 229	spin_lock(&vm_bo->vm->status_lock);
 230	list_move(&vm_bo->vm_status, &vm_bo->vm->invalidated);
 231	spin_unlock(&vm_bo->vm->status_lock);
 232}
 233
 234/**
 235 * amdgpu_vm_bo_relocated - vm_bo is reloacted
 236 *
 237 * @vm_bo: vm_bo which is relocated
 238 *
 239 * State for PDs/PTs which needs to update their parent PD.
 240 * For the root PD, just move to idle state.
 241 */
 242static void amdgpu_vm_bo_relocated(struct amdgpu_vm_bo_base *vm_bo)
 243{
 244	if (vm_bo->bo->parent) {
 245		spin_lock(&vm_bo->vm->status_lock);
 246		list_move(&vm_bo->vm_status, &vm_bo->vm->relocated);
 247		spin_unlock(&vm_bo->vm->status_lock);
 248	} else {
 249		amdgpu_vm_bo_idle(vm_bo);
 250	}
 251}
 252
 253/**
 254 * amdgpu_vm_bo_done - vm_bo is done
 255 *
 256 * @vm_bo: vm_bo which is now done
 257 *
 258 * State for normal BOs which are invalidated and that change has been updated
 259 * in the PTs.
 260 */
 261static void amdgpu_vm_bo_done(struct amdgpu_vm_bo_base *vm_bo)
 262{
 263	spin_lock(&vm_bo->vm->status_lock);
 264	list_move(&vm_bo->vm_status, &vm_bo->vm->done);
 265	spin_unlock(&vm_bo->vm->status_lock);
 266}
 267
 268/**
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 269 * amdgpu_vm_bo_base_init - Adds bo to the list of bos associated with the vm
 270 *
 271 * @base: base structure for tracking BO usage in a VM
 272 * @vm: vm to which bo is to be added
 273 * @bo: amdgpu buffer object
 274 *
 275 * Initialize a bo_va_base structure and add it to the appropriate lists
 276 *
 277 */
 278void amdgpu_vm_bo_base_init(struct amdgpu_vm_bo_base *base,
 279			    struct amdgpu_vm *vm, struct amdgpu_bo *bo)
 280{
 281	base->vm = vm;
 282	base->bo = bo;
 283	base->next = NULL;
 284	INIT_LIST_HEAD(&base->vm_status);
 285
 286	if (!bo)
 287		return;
 288	base->next = bo->vm_bo;
 289	bo->vm_bo = base;
 290
 291	if (bo->tbo.base.resv != vm->root.bo->tbo.base.resv)
 292		return;
 293
 294	dma_resv_assert_held(vm->root.bo->tbo.base.resv);
 295
 296	ttm_bo_set_bulk_move(&bo->tbo, &vm->lru_bulk_move);
 297	if (bo->tbo.type == ttm_bo_type_kernel && bo->parent)
 298		amdgpu_vm_bo_relocated(base);
 299	else
 300		amdgpu_vm_bo_idle(base);
 301
 302	if (bo->preferred_domains &
 303	    amdgpu_mem_type_to_domain(bo->tbo.resource->mem_type))
 304		return;
 305
 306	/*
 307	 * we checked all the prerequisites, but it looks like this per vm bo
 308	 * is currently evicted. add the bo to the evicted list to make sure it
 309	 * is validated on next vm use to avoid fault.
 310	 * */
 311	amdgpu_vm_bo_evicted(base);
 312}
 313
 314/**
 315 * amdgpu_vm_get_pd_bo - add the VM PD to a validation list
 316 *
 317 * @vm: vm providing the BOs
 318 * @validated: head of validation list
 319 * @entry: entry to add
 320 *
 321 * Add the page directory to the list of BOs to
 322 * validate for command submission.
 323 */
 324void amdgpu_vm_get_pd_bo(struct amdgpu_vm *vm,
 325			 struct list_head *validated,
 326			 struct amdgpu_bo_list_entry *entry)
 327{
 328	entry->priority = 0;
 329	entry->tv.bo = &vm->root.bo->tbo;
 330	/* Two for VM updates, one for TTM and one for the CS job */
 331	entry->tv.num_shared = 4;
 332	entry->user_pages = NULL;
 333	list_add(&entry->tv.head, validated);
 334}
 335
 336/**
 337 * amdgpu_vm_move_to_lru_tail - move all BOs to the end of LRU
 338 *
 339 * @adev: amdgpu device pointer
 340 * @vm: vm providing the BOs
 341 *
 342 * Move all BOs to the end of LRU and remember their positions to put them
 343 * together.
 344 */
 345void amdgpu_vm_move_to_lru_tail(struct amdgpu_device *adev,
 346				struct amdgpu_vm *vm)
 347{
 348	spin_lock(&adev->mman.bdev.lru_lock);
 349	ttm_lru_bulk_move_tail(&vm->lru_bulk_move);
 350	spin_unlock(&adev->mman.bdev.lru_lock);
 351}
 352
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 353/**
 354 * amdgpu_vm_validate_pt_bos - validate the page table BOs
 355 *
 356 * @adev: amdgpu device pointer
 357 * @vm: vm providing the BOs
 358 * @validate: callback to do the validation
 359 * @param: parameter for the validation callback
 360 *
 361 * Validate the page table BOs on command submission if neccessary.
 362 *
 363 * Returns:
 364 * Validation result.
 365 */
 366int amdgpu_vm_validate_pt_bos(struct amdgpu_device *adev, struct amdgpu_vm *vm,
 367			      int (*validate)(void *p, struct amdgpu_bo *bo),
 368			      void *param)
 369{
 370	struct amdgpu_vm_bo_base *bo_base;
 371	struct amdgpu_bo *shadow;
 372	struct amdgpu_bo *bo;
 373	int r;
 374
 
 
 
 
 
 
 
 
 
 375	spin_lock(&vm->status_lock);
 376	while (!list_empty(&vm->evicted)) {
 377		bo_base = list_first_entry(&vm->evicted,
 378					   struct amdgpu_vm_bo_base,
 379					   vm_status);
 380		spin_unlock(&vm->status_lock);
 381
 382		bo = bo_base->bo;
 383		shadow = amdgpu_bo_shadowed(bo);
 384
 385		r = validate(param, bo);
 386		if (r)
 387			return r;
 388		if (shadow) {
 389			r = validate(param, shadow);
 390			if (r)
 391				return r;
 392		}
 393
 394		if (bo->tbo.type != ttm_bo_type_kernel) {
 395			amdgpu_vm_bo_moved(bo_base);
 396		} else {
 397			vm->update_funcs->map_table(to_amdgpu_bo_vm(bo));
 398			amdgpu_vm_bo_relocated(bo_base);
 399		}
 400		spin_lock(&vm->status_lock);
 401	}
 402	spin_unlock(&vm->status_lock);
 403
 404	amdgpu_vm_eviction_lock(vm);
 405	vm->evicting = false;
 406	amdgpu_vm_eviction_unlock(vm);
 407
 408	return 0;
 409}
 410
 411/**
 412 * amdgpu_vm_ready - check VM is ready for updates
 413 *
 414 * @vm: VM to check
 415 *
 416 * Check if all VM PDs/PTs are ready for updates
 417 *
 418 * Returns:
 419 * True if VM is not evicting.
 420 */
 421bool amdgpu_vm_ready(struct amdgpu_vm *vm)
 422{
 423	bool empty;
 424	bool ret;
 425
 426	amdgpu_vm_eviction_lock(vm);
 427	ret = !vm->evicting;
 428	amdgpu_vm_eviction_unlock(vm);
 429
 430	spin_lock(&vm->status_lock);
 431	empty = list_empty(&vm->evicted);
 432	spin_unlock(&vm->status_lock);
 433
 434	return ret && empty;
 435}
 436
 437/**
 438 * amdgpu_vm_check_compute_bug - check whether asic has compute vm bug
 439 *
 440 * @adev: amdgpu_device pointer
 441 */
 442void amdgpu_vm_check_compute_bug(struct amdgpu_device *adev)
 443{
 444	const struct amdgpu_ip_block *ip_block;
 445	bool has_compute_vm_bug;
 446	struct amdgpu_ring *ring;
 447	int i;
 448
 449	has_compute_vm_bug = false;
 450
 451	ip_block = amdgpu_device_ip_get_ip_block(adev, AMD_IP_BLOCK_TYPE_GFX);
 452	if (ip_block) {
 453		/* Compute has a VM bug for GFX version < 7.
 454		   Compute has a VM bug for GFX 8 MEC firmware version < 673.*/
 455		if (ip_block->version->major <= 7)
 456			has_compute_vm_bug = true;
 457		else if (ip_block->version->major == 8)
 458			if (adev->gfx.mec_fw_version < 673)
 459				has_compute_vm_bug = true;
 460	}
 461
 462	for (i = 0; i < adev->num_rings; i++) {
 463		ring = adev->rings[i];
 464		if (ring->funcs->type == AMDGPU_RING_TYPE_COMPUTE)
 465			/* only compute rings */
 466			ring->has_compute_vm_bug = has_compute_vm_bug;
 467		else
 468			ring->has_compute_vm_bug = false;
 469	}
 470}
 471
 472/**
 473 * amdgpu_vm_need_pipeline_sync - Check if pipe sync is needed for job.
 474 *
 475 * @ring: ring on which the job will be submitted
 476 * @job: job to submit
 477 *
 478 * Returns:
 479 * True if sync is needed.
 480 */
 481bool amdgpu_vm_need_pipeline_sync(struct amdgpu_ring *ring,
 482				  struct amdgpu_job *job)
 483{
 484	struct amdgpu_device *adev = ring->adev;
 485	unsigned vmhub = ring->funcs->vmhub;
 486	struct amdgpu_vmid_mgr *id_mgr = &adev->vm_manager.id_mgr[vmhub];
 487
 488	if (job->vmid == 0)
 489		return false;
 490
 491	if (job->vm_needs_flush || ring->has_compute_vm_bug)
 492		return true;
 493
 494	if (ring->funcs->emit_gds_switch && job->gds_switch_needed)
 495		return true;
 496
 497	if (amdgpu_vmid_had_gpu_reset(adev, &id_mgr->ids[job->vmid]))
 498		return true;
 499
 500	return false;
 501}
 502
 503/**
 504 * amdgpu_vm_flush - hardware flush the vm
 505 *
 506 * @ring: ring to use for flush
 507 * @job:  related job
 508 * @need_pipe_sync: is pipe sync needed
 509 *
 510 * Emit a VM flush when it is necessary.
 511 *
 512 * Returns:
 513 * 0 on success, errno otherwise.
 514 */
 515int amdgpu_vm_flush(struct amdgpu_ring *ring, struct amdgpu_job *job,
 516		    bool need_pipe_sync)
 517{
 518	struct amdgpu_device *adev = ring->adev;
 519	unsigned vmhub = ring->funcs->vmhub;
 520	struct amdgpu_vmid_mgr *id_mgr = &adev->vm_manager.id_mgr[vmhub];
 521	struct amdgpu_vmid *id = &id_mgr->ids[job->vmid];
 522	bool spm_update_needed = job->spm_update_needed;
 523	bool gds_switch_needed = ring->funcs->emit_gds_switch &&
 524		job->gds_switch_needed;
 525	bool vm_flush_needed = job->vm_needs_flush;
 526	struct dma_fence *fence = NULL;
 527	bool pasid_mapping_needed = false;
 528	unsigned patch_offset = 0;
 529	int r;
 530
 531	if (amdgpu_vmid_had_gpu_reset(adev, id)) {
 532		gds_switch_needed = true;
 533		vm_flush_needed = true;
 534		pasid_mapping_needed = true;
 535		spm_update_needed = true;
 536	}
 537
 538	mutex_lock(&id_mgr->lock);
 539	if (id->pasid != job->pasid || !id->pasid_mapping ||
 540	    !dma_fence_is_signaled(id->pasid_mapping))
 541		pasid_mapping_needed = true;
 542	mutex_unlock(&id_mgr->lock);
 543
 544	gds_switch_needed &= !!ring->funcs->emit_gds_switch;
 545	vm_flush_needed &= !!ring->funcs->emit_vm_flush  &&
 546			job->vm_pd_addr != AMDGPU_BO_INVALID_OFFSET;
 547	pasid_mapping_needed &= adev->gmc.gmc_funcs->emit_pasid_mapping &&
 548		ring->funcs->emit_wreg;
 549
 550	if (!vm_flush_needed && !gds_switch_needed && !need_pipe_sync)
 551		return 0;
 552
 553	amdgpu_ring_ib_begin(ring);
 554	if (ring->funcs->init_cond_exec)
 555		patch_offset = amdgpu_ring_init_cond_exec(ring);
 556
 557	if (need_pipe_sync)
 558		amdgpu_ring_emit_pipeline_sync(ring);
 559
 560	if (vm_flush_needed) {
 561		trace_amdgpu_vm_flush(ring, job->vmid, job->vm_pd_addr);
 562		amdgpu_ring_emit_vm_flush(ring, job->vmid, job->vm_pd_addr);
 563	}
 564
 565	if (pasid_mapping_needed)
 566		amdgpu_gmc_emit_pasid_mapping(ring, job->vmid, job->pasid);
 567
 568	if (spm_update_needed && adev->gfx.rlc.funcs->update_spm_vmid)
 569		adev->gfx.rlc.funcs->update_spm_vmid(adev, job->vmid);
 570
 571	if (!ring->is_mes_queue && ring->funcs->emit_gds_switch &&
 572	    gds_switch_needed) {
 573		amdgpu_ring_emit_gds_switch(ring, job->vmid, job->gds_base,
 574					    job->gds_size, job->gws_base,
 575					    job->gws_size, job->oa_base,
 576					    job->oa_size);
 577	}
 578
 579	if (vm_flush_needed || pasid_mapping_needed) {
 580		r = amdgpu_fence_emit(ring, &fence, NULL, 0);
 581		if (r)
 582			return r;
 583	}
 584
 585	if (vm_flush_needed) {
 586		mutex_lock(&id_mgr->lock);
 587		dma_fence_put(id->last_flush);
 588		id->last_flush = dma_fence_get(fence);
 589		id->current_gpu_reset_count =
 590			atomic_read(&adev->gpu_reset_counter);
 591		mutex_unlock(&id_mgr->lock);
 592	}
 593
 594	if (pasid_mapping_needed) {
 595		mutex_lock(&id_mgr->lock);
 596		id->pasid = job->pasid;
 597		dma_fence_put(id->pasid_mapping);
 598		id->pasid_mapping = dma_fence_get(fence);
 599		mutex_unlock(&id_mgr->lock);
 600	}
 601	dma_fence_put(fence);
 602
 603	if (ring->funcs->patch_cond_exec)
 604		amdgpu_ring_patch_cond_exec(ring, patch_offset);
 605
 606	/* the double SWITCH_BUFFER here *cannot* be skipped by COND_EXEC */
 607	if (ring->funcs->emit_switch_buffer) {
 608		amdgpu_ring_emit_switch_buffer(ring);
 609		amdgpu_ring_emit_switch_buffer(ring);
 610	}
 611	amdgpu_ring_ib_end(ring);
 612	return 0;
 613}
 614
 615/**
 616 * amdgpu_vm_bo_find - find the bo_va for a specific vm & bo
 617 *
 618 * @vm: requested vm
 619 * @bo: requested buffer object
 620 *
 621 * Find @bo inside the requested vm.
 622 * Search inside the @bos vm list for the requested vm
 623 * Returns the found bo_va or NULL if none is found
 624 *
 625 * Object has to be reserved!
 626 *
 627 * Returns:
 628 * Found bo_va or NULL.
 629 */
 630struct amdgpu_bo_va *amdgpu_vm_bo_find(struct amdgpu_vm *vm,
 631				       struct amdgpu_bo *bo)
 632{
 633	struct amdgpu_vm_bo_base *base;
 634
 635	for (base = bo->vm_bo; base; base = base->next) {
 636		if (base->vm != vm)
 637			continue;
 638
 639		return container_of(base, struct amdgpu_bo_va, base);
 640	}
 641	return NULL;
 642}
 643
 644/**
 645 * amdgpu_vm_map_gart - Resolve gart mapping of addr
 646 *
 647 * @pages_addr: optional DMA address to use for lookup
 648 * @addr: the unmapped addr
 649 *
 650 * Look up the physical address of the page that the pte resolves
 651 * to.
 652 *
 653 * Returns:
 654 * The pointer for the page table entry.
 655 */
 656uint64_t amdgpu_vm_map_gart(const dma_addr_t *pages_addr, uint64_t addr)
 657{
 658	uint64_t result;
 659
 660	/* page table offset */
 661	result = pages_addr[addr >> PAGE_SHIFT];
 662
 663	/* in case cpu page size != gpu page size*/
 664	result |= addr & (~PAGE_MASK);
 665
 666	result &= 0xFFFFFFFFFFFFF000ULL;
 667
 668	return result;
 669}
 670
 671/**
 672 * amdgpu_vm_update_pdes - make sure that all directories are valid
 673 *
 674 * @adev: amdgpu_device pointer
 675 * @vm: requested vm
 676 * @immediate: submit immediately to the paging queue
 677 *
 678 * Makes sure all directories are up to date.
 679 *
 680 * Returns:
 681 * 0 for success, error for failure.
 682 */
 683int amdgpu_vm_update_pdes(struct amdgpu_device *adev,
 684			  struct amdgpu_vm *vm, bool immediate)
 685{
 686	struct amdgpu_vm_update_params params;
 687	struct amdgpu_vm_bo_base *entry;
 688	bool flush_tlb_needed = false;
 689	LIST_HEAD(relocated);
 690	int r, idx;
 691
 692	spin_lock(&vm->status_lock);
 693	list_splice_init(&vm->relocated, &relocated);
 694	spin_unlock(&vm->status_lock);
 695
 696	if (list_empty(&relocated))
 697		return 0;
 698
 699	if (!drm_dev_enter(adev_to_drm(adev), &idx))
 700		return -ENODEV;
 701
 702	memset(&params, 0, sizeof(params));
 703	params.adev = adev;
 704	params.vm = vm;
 705	params.immediate = immediate;
 706
 707	r = vm->update_funcs->prepare(&params, NULL, AMDGPU_SYNC_EXPLICIT);
 708	if (r)
 709		goto error;
 710
 711	list_for_each_entry(entry, &relocated, vm_status) {
 712		/* vm_flush_needed after updating moved PDEs */
 713		flush_tlb_needed |= entry->moved;
 714
 715		r = amdgpu_vm_pde_update(&params, entry);
 716		if (r)
 717			goto error;
 718	}
 719
 720	r = vm->update_funcs->commit(&params, &vm->last_update);
 721	if (r)
 722		goto error;
 723
 724	if (flush_tlb_needed)
 725		atomic64_inc(&vm->tlb_seq);
 726
 727	while (!list_empty(&relocated)) {
 728		entry = list_first_entry(&relocated, struct amdgpu_vm_bo_base,
 729					 vm_status);
 730		amdgpu_vm_bo_idle(entry);
 731	}
 732
 733error:
 734	drm_dev_exit(idx);
 735	return r;
 736}
 737
 738/**
 739 * amdgpu_vm_tlb_seq_cb - make sure to increment tlb sequence
 740 * @fence: unused
 741 * @cb: the callback structure
 742 *
 743 * Increments the tlb sequence to make sure that future CS execute a VM flush.
 744 */
 745static void amdgpu_vm_tlb_seq_cb(struct dma_fence *fence,
 746				 struct dma_fence_cb *cb)
 747{
 748	struct amdgpu_vm_tlb_seq_cb *tlb_cb;
 749
 750	tlb_cb = container_of(cb, typeof(*tlb_cb), cb);
 751	atomic64_inc(&tlb_cb->vm->tlb_seq);
 752	kfree(tlb_cb);
 753}
 754
 755/**
 756 * amdgpu_vm_update_range - update a range in the vm page table
 757 *
 758 * @adev: amdgpu_device pointer to use for commands
 759 * @vm: the VM to update the range
 760 * @immediate: immediate submission in a page fault
 761 * @unlocked: unlocked invalidation during MM callback
 762 * @flush_tlb: trigger tlb invalidation after update completed
 
 763 * @resv: fences we need to sync to
 764 * @start: start of mapped range
 765 * @last: last mapped entry
 766 * @flags: flags for the entries
 767 * @offset: offset into nodes and pages_addr
 768 * @vram_base: base for vram mappings
 769 * @res: ttm_resource to map
 770 * @pages_addr: DMA addresses to use for mapping
 771 * @fence: optional resulting fence
 772 *
 773 * Fill in the page table entries between @start and @last.
 774 *
 775 * Returns:
 776 * 0 for success, negative erro code for failure.
 777 */
 778int amdgpu_vm_update_range(struct amdgpu_device *adev, struct amdgpu_vm *vm,
 779			   bool immediate, bool unlocked, bool flush_tlb,
 780			   struct dma_resv *resv, uint64_t start, uint64_t last,
 781			   uint64_t flags, uint64_t offset, uint64_t vram_base,
 782			   struct ttm_resource *res, dma_addr_t *pages_addr,
 783			   struct dma_fence **fence)
 784{
 785	struct amdgpu_vm_update_params params;
 786	struct amdgpu_vm_tlb_seq_cb *tlb_cb;
 787	struct amdgpu_res_cursor cursor;
 788	enum amdgpu_sync_mode sync_mode;
 789	int r, idx;
 790
 791	if (!drm_dev_enter(adev_to_drm(adev), &idx))
 792		return -ENODEV;
 793
 794	tlb_cb = kmalloc(sizeof(*tlb_cb), GFP_KERNEL);
 795	if (!tlb_cb) {
 796		r = -ENOMEM;
 797		goto error_unlock;
 798	}
 799
 800	/* Vega20+XGMI where PTEs get inadvertently cached in L2 texture cache,
 801	 * heavy-weight flush TLB unconditionally.
 802	 */
 803	flush_tlb |= adev->gmc.xgmi.num_physical_nodes &&
 804		     adev->ip_versions[GC_HWIP][0] == IP_VERSION(9, 4, 0);
 805
 806	/*
 807	 * On GFX8 and older any 8 PTE block with a valid bit set enters the TLB
 808	 */
 809	flush_tlb |= adev->ip_versions[GC_HWIP][0] < IP_VERSION(9, 0, 0);
 810
 811	memset(&params, 0, sizeof(params));
 812	params.adev = adev;
 813	params.vm = vm;
 814	params.immediate = immediate;
 815	params.pages_addr = pages_addr;
 816	params.unlocked = unlocked;
 
 817
 818	/* Implicitly sync to command submissions in the same VM before
 819	 * unmapping. Sync to moving fences before mapping.
 820	 */
 821	if (!(flags & AMDGPU_PTE_VALID))
 822		sync_mode = AMDGPU_SYNC_EQ_OWNER;
 823	else
 824		sync_mode = AMDGPU_SYNC_EXPLICIT;
 825
 826	amdgpu_vm_eviction_lock(vm);
 827	if (vm->evicting) {
 828		r = -EBUSY;
 829		goto error_free;
 830	}
 831
 832	if (!unlocked && !dma_fence_is_signaled(vm->last_unlocked)) {
 833		struct dma_fence *tmp = dma_fence_get_stub();
 834
 835		amdgpu_bo_fence(vm->root.bo, vm->last_unlocked, true);
 836		swap(vm->last_unlocked, tmp);
 837		dma_fence_put(tmp);
 838	}
 839
 840	r = vm->update_funcs->prepare(&params, resv, sync_mode);
 841	if (r)
 842		goto error_free;
 843
 844	amdgpu_res_first(pages_addr ? NULL : res, offset,
 845			 (last - start + 1) * AMDGPU_GPU_PAGE_SIZE, &cursor);
 846	while (cursor.remaining) {
 847		uint64_t tmp, num_entries, addr;
 848
 849		num_entries = cursor.size >> AMDGPU_GPU_PAGE_SHIFT;
 850		if (pages_addr) {
 851			bool contiguous = true;
 852
 853			if (num_entries > AMDGPU_GPU_PAGES_IN_CPU_PAGE) {
 854				uint64_t pfn = cursor.start >> PAGE_SHIFT;
 855				uint64_t count;
 856
 857				contiguous = pages_addr[pfn + 1] ==
 858					pages_addr[pfn] + PAGE_SIZE;
 859
 860				tmp = num_entries /
 861					AMDGPU_GPU_PAGES_IN_CPU_PAGE;
 862				for (count = 2; count < tmp; ++count) {
 863					uint64_t idx = pfn + count;
 864
 865					if (contiguous != (pages_addr[idx] ==
 866					    pages_addr[idx - 1] + PAGE_SIZE))
 867						break;
 868				}
 
 
 869				num_entries = count *
 870					AMDGPU_GPU_PAGES_IN_CPU_PAGE;
 871			}
 872
 873			if (!contiguous) {
 874				addr = cursor.start;
 875				params.pages_addr = pages_addr;
 876			} else {
 877				addr = pages_addr[cursor.start >> PAGE_SHIFT];
 878				params.pages_addr = NULL;
 879			}
 880
 881		} else if (flags & (AMDGPU_PTE_VALID | AMDGPU_PTE_PRT)) {
 882			addr = vram_base + cursor.start;
 883		} else {
 884			addr = 0;
 885		}
 886
 887		tmp = start + num_entries;
 888		r = amdgpu_vm_ptes_update(&params, start, tmp, addr, flags);
 889		if (r)
 890			goto error_free;
 891
 892		amdgpu_res_next(&cursor, num_entries * AMDGPU_GPU_PAGE_SIZE);
 893		start = tmp;
 894	}
 895
 896	r = vm->update_funcs->commit(&params, fence);
 897
 898	if (flush_tlb || params.table_freed) {
 899		tlb_cb->vm = vm;
 900		if (fence && *fence &&
 901		    !dma_fence_add_callback(*fence, &tlb_cb->cb,
 902					   amdgpu_vm_tlb_seq_cb)) {
 903			dma_fence_put(vm->last_tlb_flush);
 904			vm->last_tlb_flush = dma_fence_get(*fence);
 905		} else {
 906			amdgpu_vm_tlb_seq_cb(NULL, &tlb_cb->cb);
 907		}
 908		tlb_cb = NULL;
 909	}
 910
 911error_free:
 912	kfree(tlb_cb);
 913
 914error_unlock:
 915	amdgpu_vm_eviction_unlock(vm);
 916	drm_dev_exit(idx);
 917	return r;
 918}
 919
 920void amdgpu_vm_get_memory(struct amdgpu_vm *vm, uint64_t *vram_mem,
 921				uint64_t *gtt_mem, uint64_t *cpu_mem)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 922{
 923	struct amdgpu_bo_va *bo_va, *tmp;
 924
 925	spin_lock(&vm->status_lock);
 926	list_for_each_entry_safe(bo_va, tmp, &vm->idle, base.vm_status) {
 927		if (!bo_va->base.bo)
 928			continue;
 929		amdgpu_bo_get_memory(bo_va->base.bo, vram_mem,
 930				gtt_mem, cpu_mem);
 931	}
 932	list_for_each_entry_safe(bo_va, tmp, &vm->evicted, base.vm_status) {
 933		if (!bo_va->base.bo)
 934			continue;
 935		amdgpu_bo_get_memory(bo_va->base.bo, vram_mem,
 936				gtt_mem, cpu_mem);
 937	}
 938	list_for_each_entry_safe(bo_va, tmp, &vm->relocated, base.vm_status) {
 939		if (!bo_va->base.bo)
 940			continue;
 941		amdgpu_bo_get_memory(bo_va->base.bo, vram_mem,
 942				gtt_mem, cpu_mem);
 943	}
 944	list_for_each_entry_safe(bo_va, tmp, &vm->moved, base.vm_status) {
 945		if (!bo_va->base.bo)
 946			continue;
 947		amdgpu_bo_get_memory(bo_va->base.bo, vram_mem,
 948				gtt_mem, cpu_mem);
 949	}
 950	list_for_each_entry_safe(bo_va, tmp, &vm->invalidated, base.vm_status) {
 951		if (!bo_va->base.bo)
 952			continue;
 953		amdgpu_bo_get_memory(bo_va->base.bo, vram_mem,
 954				gtt_mem, cpu_mem);
 955	}
 956	list_for_each_entry_safe(bo_va, tmp, &vm->done, base.vm_status) {
 957		if (!bo_va->base.bo)
 958			continue;
 959		amdgpu_bo_get_memory(bo_va->base.bo, vram_mem,
 960				gtt_mem, cpu_mem);
 961	}
 962	spin_unlock(&vm->status_lock);
 963}
 
 964/**
 965 * amdgpu_vm_bo_update - update all BO mappings in the vm page table
 966 *
 967 * @adev: amdgpu_device pointer
 968 * @bo_va: requested BO and VM object
 969 * @clear: if true clear the entries
 970 *
 971 * Fill in the page table entries for @bo_va.
 972 *
 973 * Returns:
 974 * 0 for success, -EINVAL for failure.
 975 */
 976int amdgpu_vm_bo_update(struct amdgpu_device *adev, struct amdgpu_bo_va *bo_va,
 977			bool clear)
 978{
 979	struct amdgpu_bo *bo = bo_va->base.bo;
 980	struct amdgpu_vm *vm = bo_va->base.vm;
 981	struct amdgpu_bo_va_mapping *mapping;
 982	dma_addr_t *pages_addr = NULL;
 983	struct ttm_resource *mem;
 984	struct dma_fence **last_update;
 985	bool flush_tlb = clear;
 
 986	struct dma_resv *resv;
 987	uint64_t vram_base;
 988	uint64_t flags;
 989	int r;
 990
 991	if (clear || !bo) {
 992		mem = NULL;
 993		resv = vm->root.bo->tbo.base.resv;
 994	} else {
 995		struct drm_gem_object *obj = &bo->tbo.base;
 996
 997		resv = bo->tbo.base.resv;
 998		if (obj->import_attach && bo_va->is_xgmi) {
 999			struct dma_buf *dma_buf = obj->import_attach->dmabuf;
1000			struct drm_gem_object *gobj = dma_buf->priv;
1001			struct amdgpu_bo *abo = gem_to_amdgpu_bo(gobj);
1002
1003			if (abo->tbo.resource->mem_type == TTM_PL_VRAM)
 
1004				bo = gem_to_amdgpu_bo(gobj);
1005		}
1006		mem = bo->tbo.resource;
1007		if (mem->mem_type == TTM_PL_TT ||
1008		    mem->mem_type == AMDGPU_PL_PREEMPT)
1009			pages_addr = bo->tbo.ttm->dma_address;
1010	}
1011
1012	if (bo) {
1013		struct amdgpu_device *bo_adev;
1014
1015		flags = amdgpu_ttm_tt_pte_flags(adev, bo->tbo.ttm, mem);
1016
1017		if (amdgpu_bo_encrypted(bo))
1018			flags |= AMDGPU_PTE_TMZ;
1019
1020		bo_adev = amdgpu_ttm_adev(bo->tbo.bdev);
1021		vram_base = bo_adev->vm_manager.vram_base_offset;
 
1022	} else {
1023		flags = 0x0;
1024		vram_base = 0;
 
1025	}
1026
1027	if (clear || (bo && bo->tbo.base.resv ==
1028		      vm->root.bo->tbo.base.resv))
1029		last_update = &vm->last_update;
1030	else
1031		last_update = &bo_va->last_pt_update;
1032
1033	if (!clear && bo_va->base.moved) {
1034		flush_tlb = true;
1035		list_splice_init(&bo_va->valids, &bo_va->invalids);
1036
1037	} else if (bo_va->cleared != clear) {
1038		list_splice_init(&bo_va->valids, &bo_va->invalids);
1039	}
1040
1041	list_for_each_entry(mapping, &bo_va->invalids, list) {
1042		uint64_t update_flags = flags;
1043
1044		/* normally,bo_va->flags only contians READABLE and WIRTEABLE bit go here
1045		 * but in case of something, we filter the flags in first place
1046		 */
1047		if (!(mapping->flags & AMDGPU_PTE_READABLE))
1048			update_flags &= ~AMDGPU_PTE_READABLE;
1049		if (!(mapping->flags & AMDGPU_PTE_WRITEABLE))
1050			update_flags &= ~AMDGPU_PTE_WRITEABLE;
1051
1052		/* Apply ASIC specific mapping flags */
1053		amdgpu_gmc_get_vm_pte(adev, mapping, &update_flags);
1054
1055		trace_amdgpu_vm_bo_update(mapping);
1056
1057		r = amdgpu_vm_update_range(adev, vm, false, false, flush_tlb,
1058					   resv, mapping->start, mapping->last,
1059					   update_flags, mapping->offset,
1060					   vram_base, mem, pages_addr,
1061					   last_update);
1062		if (r)
1063			return r;
1064	}
1065
1066	/* If the BO is not in its preferred location add it back to
1067	 * the evicted list so that it gets validated again on the
1068	 * next command submission.
1069	 */
1070	if (bo && bo->tbo.base.resv == vm->root.bo->tbo.base.resv) {
1071		uint32_t mem_type = bo->tbo.resource->mem_type;
1072
1073		if (!(bo->preferred_domains &
1074		      amdgpu_mem_type_to_domain(mem_type)))
1075			amdgpu_vm_bo_evicted(&bo_va->base);
1076		else
1077			amdgpu_vm_bo_idle(&bo_va->base);
1078	} else {
1079		amdgpu_vm_bo_done(&bo_va->base);
1080	}
1081
1082	list_splice_init(&bo_va->invalids, &bo_va->valids);
1083	bo_va->cleared = clear;
1084	bo_va->base.moved = false;
1085
1086	if (trace_amdgpu_vm_bo_mapping_enabled()) {
1087		list_for_each_entry(mapping, &bo_va->valids, list)
1088			trace_amdgpu_vm_bo_mapping(mapping);
1089	}
1090
1091	return 0;
1092}
1093
1094/**
1095 * amdgpu_vm_update_prt_state - update the global PRT state
1096 *
1097 * @adev: amdgpu_device pointer
1098 */
1099static void amdgpu_vm_update_prt_state(struct amdgpu_device *adev)
1100{
1101	unsigned long flags;
1102	bool enable;
1103
1104	spin_lock_irqsave(&adev->vm_manager.prt_lock, flags);
1105	enable = !!atomic_read(&adev->vm_manager.num_prt_users);
1106	adev->gmc.gmc_funcs->set_prt(adev, enable);
1107	spin_unlock_irqrestore(&adev->vm_manager.prt_lock, flags);
1108}
1109
1110/**
1111 * amdgpu_vm_prt_get - add a PRT user
1112 *
1113 * @adev: amdgpu_device pointer
1114 */
1115static void amdgpu_vm_prt_get(struct amdgpu_device *adev)
1116{
1117	if (!adev->gmc.gmc_funcs->set_prt)
1118		return;
1119
1120	if (atomic_inc_return(&adev->vm_manager.num_prt_users) == 1)
1121		amdgpu_vm_update_prt_state(adev);
1122}
1123
1124/**
1125 * amdgpu_vm_prt_put - drop a PRT user
1126 *
1127 * @adev: amdgpu_device pointer
1128 */
1129static void amdgpu_vm_prt_put(struct amdgpu_device *adev)
1130{
1131	if (atomic_dec_return(&adev->vm_manager.num_prt_users) == 0)
1132		amdgpu_vm_update_prt_state(adev);
1133}
1134
1135/**
1136 * amdgpu_vm_prt_cb - callback for updating the PRT status
1137 *
1138 * @fence: fence for the callback
1139 * @_cb: the callback function
1140 */
1141static void amdgpu_vm_prt_cb(struct dma_fence *fence, struct dma_fence_cb *_cb)
1142{
1143	struct amdgpu_prt_cb *cb = container_of(_cb, struct amdgpu_prt_cb, cb);
1144
1145	amdgpu_vm_prt_put(cb->adev);
1146	kfree(cb);
1147}
1148
1149/**
1150 * amdgpu_vm_add_prt_cb - add callback for updating the PRT status
1151 *
1152 * @adev: amdgpu_device pointer
1153 * @fence: fence for the callback
1154 */
1155static void amdgpu_vm_add_prt_cb(struct amdgpu_device *adev,
1156				 struct dma_fence *fence)
1157{
1158	struct amdgpu_prt_cb *cb;
1159
1160	if (!adev->gmc.gmc_funcs->set_prt)
1161		return;
1162
1163	cb = kmalloc(sizeof(struct amdgpu_prt_cb), GFP_KERNEL);
1164	if (!cb) {
1165		/* Last resort when we are OOM */
1166		if (fence)
1167			dma_fence_wait(fence, false);
1168
1169		amdgpu_vm_prt_put(adev);
1170	} else {
1171		cb->adev = adev;
1172		if (!fence || dma_fence_add_callback(fence, &cb->cb,
1173						     amdgpu_vm_prt_cb))
1174			amdgpu_vm_prt_cb(fence, &cb->cb);
1175	}
1176}
1177
1178/**
1179 * amdgpu_vm_free_mapping - free a mapping
1180 *
1181 * @adev: amdgpu_device pointer
1182 * @vm: requested vm
1183 * @mapping: mapping to be freed
1184 * @fence: fence of the unmap operation
1185 *
1186 * Free a mapping and make sure we decrease the PRT usage count if applicable.
1187 */
1188static void amdgpu_vm_free_mapping(struct amdgpu_device *adev,
1189				   struct amdgpu_vm *vm,
1190				   struct amdgpu_bo_va_mapping *mapping,
1191				   struct dma_fence *fence)
1192{
1193	if (mapping->flags & AMDGPU_PTE_PRT)
1194		amdgpu_vm_add_prt_cb(adev, fence);
1195	kfree(mapping);
1196}
1197
1198/**
1199 * amdgpu_vm_prt_fini - finish all prt mappings
1200 *
1201 * @adev: amdgpu_device pointer
1202 * @vm: requested vm
1203 *
1204 * Register a cleanup callback to disable PRT support after VM dies.
1205 */
1206static void amdgpu_vm_prt_fini(struct amdgpu_device *adev, struct amdgpu_vm *vm)
1207{
1208	struct dma_resv *resv = vm->root.bo->tbo.base.resv;
1209	struct dma_resv_iter cursor;
1210	struct dma_fence *fence;
1211
1212	dma_resv_for_each_fence(&cursor, resv, DMA_RESV_USAGE_BOOKKEEP, fence) {
1213		/* Add a callback for each fence in the reservation object */
1214		amdgpu_vm_prt_get(adev);
1215		amdgpu_vm_add_prt_cb(adev, fence);
1216	}
1217}
1218
1219/**
1220 * amdgpu_vm_clear_freed - clear freed BOs in the PT
1221 *
1222 * @adev: amdgpu_device pointer
1223 * @vm: requested vm
1224 * @fence: optional resulting fence (unchanged if no work needed to be done
1225 * or if an error occurred)
1226 *
1227 * Make sure all freed BOs are cleared in the PT.
1228 * PTs have to be reserved and mutex must be locked!
1229 *
1230 * Returns:
1231 * 0 for success.
1232 *
1233 */
1234int amdgpu_vm_clear_freed(struct amdgpu_device *adev,
1235			  struct amdgpu_vm *vm,
1236			  struct dma_fence **fence)
1237{
1238	struct dma_resv *resv = vm->root.bo->tbo.base.resv;
1239	struct amdgpu_bo_va_mapping *mapping;
1240	uint64_t init_pte_value = 0;
1241	struct dma_fence *f = NULL;
1242	int r;
1243
1244	while (!list_empty(&vm->freed)) {
1245		mapping = list_first_entry(&vm->freed,
1246			struct amdgpu_bo_va_mapping, list);
1247		list_del(&mapping->list);
1248
1249		if (vm->pte_support_ats &&
1250		    mapping->start < AMDGPU_GMC_HOLE_START)
1251			init_pte_value = AMDGPU_PTE_DEFAULT_ATC;
1252
1253		r = amdgpu_vm_update_range(adev, vm, false, false, true, resv,
1254					   mapping->start, mapping->last,
1255					   init_pte_value, 0, 0, NULL, NULL,
1256					   &f);
1257		amdgpu_vm_free_mapping(adev, vm, mapping, f);
1258		if (r) {
1259			dma_fence_put(f);
1260			return r;
1261		}
1262	}
1263
1264	if (fence && f) {
1265		dma_fence_put(*fence);
1266		*fence = f;
1267	} else {
1268		dma_fence_put(f);
1269	}
1270
1271	return 0;
1272
1273}
1274
1275/**
1276 * amdgpu_vm_handle_moved - handle moved BOs in the PT
1277 *
1278 * @adev: amdgpu_device pointer
1279 * @vm: requested vm
 
1280 *
1281 * Make sure all BOs which are moved are updated in the PTs.
1282 *
1283 * Returns:
1284 * 0 for success.
1285 *
1286 * PTs have to be reserved!
1287 */
1288int amdgpu_vm_handle_moved(struct amdgpu_device *adev,
1289			   struct amdgpu_vm *vm)
 
1290{
1291	struct amdgpu_bo_va *bo_va;
1292	struct dma_resv *resv;
1293	bool clear;
1294	int r;
1295
1296	spin_lock(&vm->status_lock);
1297	while (!list_empty(&vm->moved)) {
1298		bo_va = list_first_entry(&vm->moved, struct amdgpu_bo_va,
1299					 base.vm_status);
1300		spin_unlock(&vm->status_lock);
1301
1302		/* Per VM BOs never need to bo cleared in the page tables */
1303		r = amdgpu_vm_bo_update(adev, bo_va, false);
1304		if (r)
1305			return r;
1306		spin_lock(&vm->status_lock);
1307	}
1308
1309	while (!list_empty(&vm->invalidated)) {
1310		bo_va = list_first_entry(&vm->invalidated, struct amdgpu_bo_va,
1311					 base.vm_status);
1312		resv = bo_va->base.bo->tbo.base.resv;
1313		spin_unlock(&vm->status_lock);
1314
1315		/* Try to reserve the BO to avoid clearing its ptes */
1316		if (!amdgpu_vm_debug && dma_resv_trylock(resv))
1317			clear = false;
 
 
 
 
 
1318		/* Somebody else is using the BO right now */
1319		else
1320			clear = true;
 
 
1321
1322		r = amdgpu_vm_bo_update(adev, bo_va, clear);
1323		if (r)
1324			return r;
1325
1326		if (!clear)
1327			dma_resv_unlock(resv);
1328		spin_lock(&vm->status_lock);
1329	}
1330	spin_unlock(&vm->status_lock);
1331
1332	return 0;
1333}
1334
1335/**
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1336 * amdgpu_vm_bo_add - add a bo to a specific vm
1337 *
1338 * @adev: amdgpu_device pointer
1339 * @vm: requested vm
1340 * @bo: amdgpu buffer object
1341 *
1342 * Add @bo into the requested vm.
1343 * Add @bo to the list of bos associated with the vm
1344 *
1345 * Returns:
1346 * Newly added bo_va or NULL for failure
1347 *
1348 * Object has to be reserved!
1349 */
1350struct amdgpu_bo_va *amdgpu_vm_bo_add(struct amdgpu_device *adev,
1351				      struct amdgpu_vm *vm,
1352				      struct amdgpu_bo *bo)
1353{
1354	struct amdgpu_bo_va *bo_va;
1355
1356	bo_va = kzalloc(sizeof(struct amdgpu_bo_va), GFP_KERNEL);
1357	if (bo_va == NULL) {
1358		return NULL;
1359	}
1360	amdgpu_vm_bo_base_init(&bo_va->base, vm, bo);
1361
1362	bo_va->ref_count = 1;
 
1363	INIT_LIST_HEAD(&bo_va->valids);
1364	INIT_LIST_HEAD(&bo_va->invalids);
1365
1366	if (!bo)
1367		return bo_va;
1368
1369	dma_resv_assert_held(bo->tbo.base.resv);
1370	if (amdgpu_dmabuf_is_xgmi_accessible(adev, bo)) {
1371		bo_va->is_xgmi = true;
1372		/* Power up XGMI if it can be potentially used */
1373		amdgpu_xgmi_set_pstate(adev, AMDGPU_XGMI_PSTATE_MAX_VEGA20);
1374	}
1375
1376	return bo_va;
1377}
1378
1379
1380/**
1381 * amdgpu_vm_bo_insert_map - insert a new mapping
1382 *
1383 * @adev: amdgpu_device pointer
1384 * @bo_va: bo_va to store the address
1385 * @mapping: the mapping to insert
1386 *
1387 * Insert a new mapping into all structures.
1388 */
1389static void amdgpu_vm_bo_insert_map(struct amdgpu_device *adev,
1390				    struct amdgpu_bo_va *bo_va,
1391				    struct amdgpu_bo_va_mapping *mapping)
1392{
1393	struct amdgpu_vm *vm = bo_va->base.vm;
1394	struct amdgpu_bo *bo = bo_va->base.bo;
1395
1396	mapping->bo_va = bo_va;
1397	list_add(&mapping->list, &bo_va->invalids);
1398	amdgpu_vm_it_insert(mapping, &vm->va);
1399
1400	if (mapping->flags & AMDGPU_PTE_PRT)
1401		amdgpu_vm_prt_get(adev);
1402
1403	if (bo && bo->tbo.base.resv == vm->root.bo->tbo.base.resv &&
1404	    !bo_va->base.moved) {
1405		amdgpu_vm_bo_moved(&bo_va->base);
1406	}
1407	trace_amdgpu_vm_bo_map(bo_va, mapping);
1408}
1409
1410/**
1411 * amdgpu_vm_bo_map - map bo inside a vm
1412 *
1413 * @adev: amdgpu_device pointer
1414 * @bo_va: bo_va to store the address
1415 * @saddr: where to map the BO
1416 * @offset: requested offset in the BO
1417 * @size: BO size in bytes
1418 * @flags: attributes of pages (read/write/valid/etc.)
1419 *
1420 * Add a mapping of the BO at the specefied addr into the VM.
1421 *
1422 * Returns:
1423 * 0 for success, error for failure.
1424 *
1425 * Object has to be reserved and unreserved outside!
1426 */
1427int amdgpu_vm_bo_map(struct amdgpu_device *adev,
1428		     struct amdgpu_bo_va *bo_va,
1429		     uint64_t saddr, uint64_t offset,
1430		     uint64_t size, uint64_t flags)
1431{
1432	struct amdgpu_bo_va_mapping *mapping, *tmp;
1433	struct amdgpu_bo *bo = bo_va->base.bo;
1434	struct amdgpu_vm *vm = bo_va->base.vm;
1435	uint64_t eaddr;
1436
1437	/* validate the parameters */
1438	if (saddr & ~PAGE_MASK || offset & ~PAGE_MASK ||
1439	    size == 0 || size & ~PAGE_MASK)
 
1440		return -EINVAL;
1441
1442	/* make sure object fit at this offset */
1443	eaddr = saddr + size - 1;
1444	if (saddr >= eaddr ||
1445	    (bo && offset + size > amdgpu_bo_size(bo)) ||
1446	    (eaddr >= adev->vm_manager.max_pfn << AMDGPU_GPU_PAGE_SHIFT))
1447		return -EINVAL;
1448
1449	saddr /= AMDGPU_GPU_PAGE_SIZE;
1450	eaddr /= AMDGPU_GPU_PAGE_SIZE;
1451
1452	tmp = amdgpu_vm_it_iter_first(&vm->va, saddr, eaddr);
1453	if (tmp) {
1454		/* bo and tmp overlap, invalid addr */
1455		dev_err(adev->dev, "bo %p va 0x%010Lx-0x%010Lx conflict with "
1456			"0x%010Lx-0x%010Lx\n", bo, saddr, eaddr,
1457			tmp->start, tmp->last + 1);
1458		return -EINVAL;
1459	}
1460
1461	mapping = kmalloc(sizeof(*mapping), GFP_KERNEL);
1462	if (!mapping)
1463		return -ENOMEM;
1464
1465	mapping->start = saddr;
1466	mapping->last = eaddr;
1467	mapping->offset = offset;
1468	mapping->flags = flags;
1469
1470	amdgpu_vm_bo_insert_map(adev, bo_va, mapping);
1471
1472	return 0;
1473}
1474
1475/**
1476 * amdgpu_vm_bo_replace_map - map bo inside a vm, replacing existing mappings
1477 *
1478 * @adev: amdgpu_device pointer
1479 * @bo_va: bo_va to store the address
1480 * @saddr: where to map the BO
1481 * @offset: requested offset in the BO
1482 * @size: BO size in bytes
1483 * @flags: attributes of pages (read/write/valid/etc.)
1484 *
1485 * Add a mapping of the BO at the specefied addr into the VM. Replace existing
1486 * mappings as we do so.
1487 *
1488 * Returns:
1489 * 0 for success, error for failure.
1490 *
1491 * Object has to be reserved and unreserved outside!
1492 */
1493int amdgpu_vm_bo_replace_map(struct amdgpu_device *adev,
1494			     struct amdgpu_bo_va *bo_va,
1495			     uint64_t saddr, uint64_t offset,
1496			     uint64_t size, uint64_t flags)
1497{
1498	struct amdgpu_bo_va_mapping *mapping;
1499	struct amdgpu_bo *bo = bo_va->base.bo;
1500	uint64_t eaddr;
1501	int r;
1502
1503	/* validate the parameters */
1504	if (saddr & ~PAGE_MASK || offset & ~PAGE_MASK ||
1505	    size == 0 || size & ~PAGE_MASK)
 
1506		return -EINVAL;
1507
1508	/* make sure object fit at this offset */
1509	eaddr = saddr + size - 1;
1510	if (saddr >= eaddr ||
1511	    (bo && offset + size > amdgpu_bo_size(bo)) ||
1512	    (eaddr >= adev->vm_manager.max_pfn << AMDGPU_GPU_PAGE_SHIFT))
1513		return -EINVAL;
1514
1515	/* Allocate all the needed memory */
1516	mapping = kmalloc(sizeof(*mapping), GFP_KERNEL);
1517	if (!mapping)
1518		return -ENOMEM;
1519
1520	r = amdgpu_vm_bo_clear_mappings(adev, bo_va->base.vm, saddr, size);
1521	if (r) {
1522		kfree(mapping);
1523		return r;
1524	}
1525
1526	saddr /= AMDGPU_GPU_PAGE_SIZE;
1527	eaddr /= AMDGPU_GPU_PAGE_SIZE;
1528
1529	mapping->start = saddr;
1530	mapping->last = eaddr;
1531	mapping->offset = offset;
1532	mapping->flags = flags;
1533
1534	amdgpu_vm_bo_insert_map(adev, bo_va, mapping);
1535
1536	return 0;
1537}
1538
1539/**
1540 * amdgpu_vm_bo_unmap - remove bo mapping from vm
1541 *
1542 * @adev: amdgpu_device pointer
1543 * @bo_va: bo_va to remove the address from
1544 * @saddr: where to the BO is mapped
1545 *
1546 * Remove a mapping of the BO at the specefied addr from the VM.
1547 *
1548 * Returns:
1549 * 0 for success, error for failure.
1550 *
1551 * Object has to be reserved and unreserved outside!
1552 */
1553int amdgpu_vm_bo_unmap(struct amdgpu_device *adev,
1554		       struct amdgpu_bo_va *bo_va,
1555		       uint64_t saddr)
1556{
1557	struct amdgpu_bo_va_mapping *mapping;
1558	struct amdgpu_vm *vm = bo_va->base.vm;
1559	bool valid = true;
1560
1561	saddr /= AMDGPU_GPU_PAGE_SIZE;
1562
1563	list_for_each_entry(mapping, &bo_va->valids, list) {
1564		if (mapping->start == saddr)
1565			break;
1566	}
1567
1568	if (&mapping->list == &bo_va->valids) {
1569		valid = false;
1570
1571		list_for_each_entry(mapping, &bo_va->invalids, list) {
1572			if (mapping->start == saddr)
1573				break;
1574		}
1575
1576		if (&mapping->list == &bo_va->invalids)
1577			return -ENOENT;
1578	}
1579
1580	list_del(&mapping->list);
1581	amdgpu_vm_it_remove(mapping, &vm->va);
1582	mapping->bo_va = NULL;
1583	trace_amdgpu_vm_bo_unmap(bo_va, mapping);
1584
1585	if (valid)
1586		list_add(&mapping->list, &vm->freed);
1587	else
1588		amdgpu_vm_free_mapping(adev, vm, mapping,
1589				       bo_va->last_pt_update);
1590
1591	return 0;
1592}
1593
1594/**
1595 * amdgpu_vm_bo_clear_mappings - remove all mappings in a specific range
1596 *
1597 * @adev: amdgpu_device pointer
1598 * @vm: VM structure to use
1599 * @saddr: start of the range
1600 * @size: size of the range
1601 *
1602 * Remove all mappings in a range, split them as appropriate.
1603 *
1604 * Returns:
1605 * 0 for success, error for failure.
1606 */
1607int amdgpu_vm_bo_clear_mappings(struct amdgpu_device *adev,
1608				struct amdgpu_vm *vm,
1609				uint64_t saddr, uint64_t size)
1610{
1611	struct amdgpu_bo_va_mapping *before, *after, *tmp, *next;
1612	LIST_HEAD(removed);
1613	uint64_t eaddr;
1614
1615	eaddr = saddr + size - 1;
1616	saddr /= AMDGPU_GPU_PAGE_SIZE;
1617	eaddr /= AMDGPU_GPU_PAGE_SIZE;
1618
1619	/* Allocate all the needed memory */
1620	before = kzalloc(sizeof(*before), GFP_KERNEL);
1621	if (!before)
1622		return -ENOMEM;
1623	INIT_LIST_HEAD(&before->list);
1624
1625	after = kzalloc(sizeof(*after), GFP_KERNEL);
1626	if (!after) {
1627		kfree(before);
1628		return -ENOMEM;
1629	}
1630	INIT_LIST_HEAD(&after->list);
1631
1632	/* Now gather all removed mappings */
1633	tmp = amdgpu_vm_it_iter_first(&vm->va, saddr, eaddr);
1634	while (tmp) {
1635		/* Remember mapping split at the start */
1636		if (tmp->start < saddr) {
1637			before->start = tmp->start;
1638			before->last = saddr - 1;
1639			before->offset = tmp->offset;
1640			before->flags = tmp->flags;
1641			before->bo_va = tmp->bo_va;
1642			list_add(&before->list, &tmp->bo_va->invalids);
1643		}
1644
1645		/* Remember mapping split at the end */
1646		if (tmp->last > eaddr) {
1647			after->start = eaddr + 1;
1648			after->last = tmp->last;
1649			after->offset = tmp->offset;
1650			after->offset += (after->start - tmp->start) << PAGE_SHIFT;
1651			after->flags = tmp->flags;
1652			after->bo_va = tmp->bo_va;
1653			list_add(&after->list, &tmp->bo_va->invalids);
1654		}
1655
1656		list_del(&tmp->list);
1657		list_add(&tmp->list, &removed);
1658
1659		tmp = amdgpu_vm_it_iter_next(tmp, saddr, eaddr);
1660	}
1661
1662	/* And free them up */
1663	list_for_each_entry_safe(tmp, next, &removed, list) {
1664		amdgpu_vm_it_remove(tmp, &vm->va);
1665		list_del(&tmp->list);
1666
1667		if (tmp->start < saddr)
1668		    tmp->start = saddr;
1669		if (tmp->last > eaddr)
1670		    tmp->last = eaddr;
1671
1672		tmp->bo_va = NULL;
1673		list_add(&tmp->list, &vm->freed);
1674		trace_amdgpu_vm_bo_unmap(NULL, tmp);
1675	}
1676
1677	/* Insert partial mapping before the range */
1678	if (!list_empty(&before->list)) {
 
 
1679		amdgpu_vm_it_insert(before, &vm->va);
1680		if (before->flags & AMDGPU_PTE_PRT)
1681			amdgpu_vm_prt_get(adev);
 
 
 
 
1682	} else {
1683		kfree(before);
1684	}
1685
1686	/* Insert partial mapping after the range */
1687	if (!list_empty(&after->list)) {
 
 
1688		amdgpu_vm_it_insert(after, &vm->va);
1689		if (after->flags & AMDGPU_PTE_PRT)
1690			amdgpu_vm_prt_get(adev);
 
 
 
 
1691	} else {
1692		kfree(after);
1693	}
1694
1695	return 0;
1696}
1697
1698/**
1699 * amdgpu_vm_bo_lookup_mapping - find mapping by address
1700 *
1701 * @vm: the requested VM
1702 * @addr: the address
1703 *
1704 * Find a mapping by it's address.
1705 *
1706 * Returns:
1707 * The amdgpu_bo_va_mapping matching for addr or NULL
1708 *
1709 */
1710struct amdgpu_bo_va_mapping *amdgpu_vm_bo_lookup_mapping(struct amdgpu_vm *vm,
1711							 uint64_t addr)
1712{
1713	return amdgpu_vm_it_iter_first(&vm->va, addr, addr);
1714}
1715
1716/**
1717 * amdgpu_vm_bo_trace_cs - trace all reserved mappings
1718 *
1719 * @vm: the requested vm
1720 * @ticket: CS ticket
1721 *
1722 * Trace all mappings of BOs reserved during a command submission.
1723 */
1724void amdgpu_vm_bo_trace_cs(struct amdgpu_vm *vm, struct ww_acquire_ctx *ticket)
1725{
1726	struct amdgpu_bo_va_mapping *mapping;
1727
1728	if (!trace_amdgpu_vm_bo_cs_enabled())
1729		return;
1730
1731	for (mapping = amdgpu_vm_it_iter_first(&vm->va, 0, U64_MAX); mapping;
1732	     mapping = amdgpu_vm_it_iter_next(mapping, 0, U64_MAX)) {
1733		if (mapping->bo_va && mapping->bo_va->base.bo) {
1734			struct amdgpu_bo *bo;
1735
1736			bo = mapping->bo_va->base.bo;
1737			if (dma_resv_locking_ctx(bo->tbo.base.resv) !=
1738			    ticket)
1739				continue;
1740		}
1741
1742		trace_amdgpu_vm_bo_cs(mapping);
1743	}
1744}
1745
1746/**
1747 * amdgpu_vm_bo_del - remove a bo from a specific vm
1748 *
1749 * @adev: amdgpu_device pointer
1750 * @bo_va: requested bo_va
1751 *
1752 * Remove @bo_va->bo from the requested vm.
1753 *
1754 * Object have to be reserved!
1755 */
1756void amdgpu_vm_bo_del(struct amdgpu_device *adev,
1757		      struct amdgpu_bo_va *bo_va)
1758{
1759	struct amdgpu_bo_va_mapping *mapping, *next;
1760	struct amdgpu_bo *bo = bo_va->base.bo;
1761	struct amdgpu_vm *vm = bo_va->base.vm;
1762	struct amdgpu_vm_bo_base **base;
1763
1764	dma_resv_assert_held(vm->root.bo->tbo.base.resv);
1765
1766	if (bo) {
1767		dma_resv_assert_held(bo->tbo.base.resv);
1768		if (bo->tbo.base.resv == vm->root.bo->tbo.base.resv)
1769			ttm_bo_set_bulk_move(&bo->tbo, NULL);
1770
1771		for (base = &bo_va->base.bo->vm_bo; *base;
1772		     base = &(*base)->next) {
1773			if (*base != &bo_va->base)
1774				continue;
1775
1776			*base = bo_va->base.next;
1777			break;
1778		}
1779	}
1780
1781	spin_lock(&vm->status_lock);
1782	list_del(&bo_va->base.vm_status);
1783	spin_unlock(&vm->status_lock);
1784
1785	list_for_each_entry_safe(mapping, next, &bo_va->valids, list) {
1786		list_del(&mapping->list);
1787		amdgpu_vm_it_remove(mapping, &vm->va);
1788		mapping->bo_va = NULL;
1789		trace_amdgpu_vm_bo_unmap(bo_va, mapping);
1790		list_add(&mapping->list, &vm->freed);
1791	}
1792	list_for_each_entry_safe(mapping, next, &bo_va->invalids, list) {
1793		list_del(&mapping->list);
1794		amdgpu_vm_it_remove(mapping, &vm->va);
1795		amdgpu_vm_free_mapping(adev, vm, mapping,
1796				       bo_va->last_pt_update);
1797	}
1798
1799	dma_fence_put(bo_va->last_pt_update);
1800
1801	if (bo && bo_va->is_xgmi)
1802		amdgpu_xgmi_set_pstate(adev, AMDGPU_XGMI_PSTATE_MIN);
1803
1804	kfree(bo_va);
1805}
1806
1807/**
1808 * amdgpu_vm_evictable - check if we can evict a VM
1809 *
1810 * @bo: A page table of the VM.
1811 *
1812 * Check if it is possible to evict a VM.
1813 */
1814bool amdgpu_vm_evictable(struct amdgpu_bo *bo)
1815{
1816	struct amdgpu_vm_bo_base *bo_base = bo->vm_bo;
1817
1818	/* Page tables of a destroyed VM can go away immediately */
1819	if (!bo_base || !bo_base->vm)
1820		return true;
1821
1822	/* Don't evict VM page tables while they are busy */
1823	if (!dma_resv_test_signaled(bo->tbo.base.resv, DMA_RESV_USAGE_BOOKKEEP))
1824		return false;
1825
1826	/* Try to block ongoing updates */
1827	if (!amdgpu_vm_eviction_trylock(bo_base->vm))
1828		return false;
1829
1830	/* Don't evict VM page tables while they are updated */
1831	if (!dma_fence_is_signaled(bo_base->vm->last_unlocked)) {
1832		amdgpu_vm_eviction_unlock(bo_base->vm);
1833		return false;
1834	}
1835
1836	bo_base->vm->evicting = true;
1837	amdgpu_vm_eviction_unlock(bo_base->vm);
1838	return true;
1839}
1840
1841/**
1842 * amdgpu_vm_bo_invalidate - mark the bo as invalid
1843 *
1844 * @adev: amdgpu_device pointer
1845 * @bo: amdgpu buffer object
1846 * @evicted: is the BO evicted
1847 *
1848 * Mark @bo as invalid.
1849 */
1850void amdgpu_vm_bo_invalidate(struct amdgpu_device *adev,
1851			     struct amdgpu_bo *bo, bool evicted)
1852{
1853	struct amdgpu_vm_bo_base *bo_base;
1854
1855	/* shadow bo doesn't have bo base, its validation needs its parent */
1856	if (bo->parent && (amdgpu_bo_shadowed(bo->parent) == bo))
1857		bo = bo->parent;
1858
1859	for (bo_base = bo->vm_bo; bo_base; bo_base = bo_base->next) {
1860		struct amdgpu_vm *vm = bo_base->vm;
1861
1862		if (evicted && bo->tbo.base.resv == vm->root.bo->tbo.base.resv) {
1863			amdgpu_vm_bo_evicted(bo_base);
1864			continue;
1865		}
1866
1867		if (bo_base->moved)
1868			continue;
1869		bo_base->moved = true;
1870
1871		if (bo->tbo.type == ttm_bo_type_kernel)
1872			amdgpu_vm_bo_relocated(bo_base);
1873		else if (bo->tbo.base.resv == vm->root.bo->tbo.base.resv)
1874			amdgpu_vm_bo_moved(bo_base);
1875		else
1876			amdgpu_vm_bo_invalidated(bo_base);
1877	}
1878}
1879
1880/**
1881 * amdgpu_vm_get_block_size - calculate VM page table size as power of two
1882 *
1883 * @vm_size: VM size
1884 *
1885 * Returns:
1886 * VM page table as power of two
1887 */
1888static uint32_t amdgpu_vm_get_block_size(uint64_t vm_size)
1889{
1890	/* Total bits covered by PD + PTs */
1891	unsigned bits = ilog2(vm_size) + 18;
1892
1893	/* Make sure the PD is 4K in size up to 8GB address space.
1894	   Above that split equal between PD and PTs */
1895	if (vm_size <= 8)
1896		return (bits - 9);
1897	else
1898		return ((bits + 3) / 2);
1899}
1900
1901/**
1902 * amdgpu_vm_adjust_size - adjust vm size, block size and fragment size
1903 *
1904 * @adev: amdgpu_device pointer
1905 * @min_vm_size: the minimum vm size in GB if it's set auto
1906 * @fragment_size_default: Default PTE fragment size
1907 * @max_level: max VMPT level
1908 * @max_bits: max address space size in bits
1909 *
1910 */
1911void amdgpu_vm_adjust_size(struct amdgpu_device *adev, uint32_t min_vm_size,
1912			   uint32_t fragment_size_default, unsigned max_level,
1913			   unsigned max_bits)
1914{
1915	unsigned int max_size = 1 << (max_bits - 30);
1916	unsigned int vm_size;
1917	uint64_t tmp;
1918
1919	/* adjust vm size first */
1920	if (amdgpu_vm_size != -1) {
1921		vm_size = amdgpu_vm_size;
1922		if (vm_size > max_size) {
1923			dev_warn(adev->dev, "VM size (%d) too large, max is %u GB\n",
1924				 amdgpu_vm_size, max_size);
1925			vm_size = max_size;
1926		}
1927	} else {
1928		struct sysinfo si;
1929		unsigned int phys_ram_gb;
1930
1931		/* Optimal VM size depends on the amount of physical
1932		 * RAM available. Underlying requirements and
1933		 * assumptions:
1934		 *
1935		 *  - Need to map system memory and VRAM from all GPUs
1936		 *     - VRAM from other GPUs not known here
1937		 *     - Assume VRAM <= system memory
1938		 *  - On GFX8 and older, VM space can be segmented for
1939		 *    different MTYPEs
1940		 *  - Need to allow room for fragmentation, guard pages etc.
1941		 *
1942		 * This adds up to a rough guess of system memory x3.
1943		 * Round up to power of two to maximize the available
1944		 * VM size with the given page table size.
1945		 */
1946		si_meminfo(&si);
1947		phys_ram_gb = ((uint64_t)si.totalram * si.mem_unit +
1948			       (1 << 30) - 1) >> 30;
1949		vm_size = roundup_pow_of_two(
1950			min(max(phys_ram_gb * 3, min_vm_size), max_size));
1951	}
1952
1953	adev->vm_manager.max_pfn = (uint64_t)vm_size << 18;
1954
1955	tmp = roundup_pow_of_two(adev->vm_manager.max_pfn);
1956	if (amdgpu_vm_block_size != -1)
1957		tmp >>= amdgpu_vm_block_size - 9;
1958	tmp = DIV_ROUND_UP(fls64(tmp) - 1, 9) - 1;
1959	adev->vm_manager.num_level = min(max_level, (unsigned)tmp);
1960	switch (adev->vm_manager.num_level) {
1961	case 3:
1962		adev->vm_manager.root_level = AMDGPU_VM_PDB2;
1963		break;
1964	case 2:
1965		adev->vm_manager.root_level = AMDGPU_VM_PDB1;
1966		break;
1967	case 1:
1968		adev->vm_manager.root_level = AMDGPU_VM_PDB0;
1969		break;
1970	default:
1971		dev_err(adev->dev, "VMPT only supports 2~4+1 levels\n");
1972	}
1973	/* block size depends on vm size and hw setup*/
1974	if (amdgpu_vm_block_size != -1)
1975		adev->vm_manager.block_size =
1976			min((unsigned)amdgpu_vm_block_size, max_bits
1977			    - AMDGPU_GPU_PAGE_SHIFT
1978			    - 9 * adev->vm_manager.num_level);
1979	else if (adev->vm_manager.num_level > 1)
1980		adev->vm_manager.block_size = 9;
1981	else
1982		adev->vm_manager.block_size = amdgpu_vm_get_block_size(tmp);
1983
1984	if (amdgpu_vm_fragment_size == -1)
1985		adev->vm_manager.fragment_size = fragment_size_default;
1986	else
1987		adev->vm_manager.fragment_size = amdgpu_vm_fragment_size;
1988
1989	DRM_INFO("vm size is %u GB, %u levels, block size is %u-bit, fragment size is %u-bit\n",
1990		 vm_size, adev->vm_manager.num_level + 1,
1991		 adev->vm_manager.block_size,
1992		 adev->vm_manager.fragment_size);
1993}
1994
1995/**
1996 * amdgpu_vm_wait_idle - wait for the VM to become idle
1997 *
1998 * @vm: VM object to wait for
1999 * @timeout: timeout to wait for VM to become idle
2000 */
2001long amdgpu_vm_wait_idle(struct amdgpu_vm *vm, long timeout)
2002{
2003	timeout = dma_resv_wait_timeout(vm->root.bo->tbo.base.resv,
2004					DMA_RESV_USAGE_BOOKKEEP,
2005					true, timeout);
2006	if (timeout <= 0)
2007		return timeout;
2008
2009	return dma_fence_wait_timeout(vm->last_unlocked, true, timeout);
2010}
2011
2012/**
2013 * amdgpu_vm_init - initialize a vm instance
2014 *
2015 * @adev: amdgpu_device pointer
2016 * @vm: requested vm
 
2017 *
2018 * Init @vm fields.
2019 *
2020 * Returns:
2021 * 0 for success, error for failure.
2022 */
2023int amdgpu_vm_init(struct amdgpu_device *adev, struct amdgpu_vm *vm)
 
2024{
2025	struct amdgpu_bo *root_bo;
2026	struct amdgpu_bo_vm *root;
2027	int r, i;
2028
2029	vm->va = RB_ROOT_CACHED;
2030	for (i = 0; i < AMDGPU_MAX_VMHUBS; i++)
2031		vm->reserved_vmid[i] = NULL;
2032	INIT_LIST_HEAD(&vm->evicted);
2033	INIT_LIST_HEAD(&vm->relocated);
2034	INIT_LIST_HEAD(&vm->moved);
2035	INIT_LIST_HEAD(&vm->idle);
2036	INIT_LIST_HEAD(&vm->invalidated);
2037	spin_lock_init(&vm->status_lock);
2038	INIT_LIST_HEAD(&vm->freed);
2039	INIT_LIST_HEAD(&vm->done);
2040	INIT_LIST_HEAD(&vm->pt_freed);
2041	INIT_WORK(&vm->pt_free_work, amdgpu_vm_pt_free_work);
 
2042
2043	/* create scheduler entities for page table updates */
2044	r = drm_sched_entity_init(&vm->immediate, DRM_SCHED_PRIORITY_NORMAL,
2045				  adev->vm_manager.vm_pte_scheds,
2046				  adev->vm_manager.vm_pte_num_scheds, NULL);
2047	if (r)
2048		return r;
2049
2050	r = drm_sched_entity_init(&vm->delayed, DRM_SCHED_PRIORITY_NORMAL,
2051				  adev->vm_manager.vm_pte_scheds,
2052				  adev->vm_manager.vm_pte_num_scheds, NULL);
2053	if (r)
2054		goto error_free_immediate;
2055
2056	vm->pte_support_ats = false;
2057	vm->is_compute_context = false;
2058
2059	vm->use_cpu_for_update = !!(adev->vm_manager.vm_update_mode &
2060				    AMDGPU_VM_USE_CPU_FOR_GFX);
2061
2062	DRM_DEBUG_DRIVER("VM update mode is %s\n",
2063			 vm->use_cpu_for_update ? "CPU" : "SDMA");
2064	WARN_ONCE((vm->use_cpu_for_update &&
2065		   !amdgpu_gmc_vram_full_visible(&adev->gmc)),
2066		  "CPU update of VM recommended only for large BAR system\n");
2067
2068	if (vm->use_cpu_for_update)
2069		vm->update_funcs = &amdgpu_vm_cpu_funcs;
2070	else
2071		vm->update_funcs = &amdgpu_vm_sdma_funcs;
2072	vm->last_update = NULL;
 
2073	vm->last_unlocked = dma_fence_get_stub();
2074	vm->last_tlb_flush = dma_fence_get_stub();
 
2075
2076	mutex_init(&vm->eviction_lock);
2077	vm->evicting = false;
2078
2079	r = amdgpu_vm_pt_create(adev, vm, adev->vm_manager.root_level,
2080				false, &root);
2081	if (r)
2082		goto error_free_delayed;
2083	root_bo = &root->bo;
 
2084	r = amdgpu_bo_reserve(root_bo, true);
2085	if (r)
2086		goto error_free_root;
 
 
 
2087
 
2088	r = dma_resv_reserve_fences(root_bo->tbo.base.resv, 1);
2089	if (r)
2090		goto error_unreserve;
2091
2092	amdgpu_vm_bo_base_init(&vm->root, vm, root_bo);
2093
2094	r = amdgpu_vm_pt_clear(adev, vm, root, false);
2095	if (r)
2096		goto error_unreserve;
2097
2098	amdgpu_bo_unreserve(vm->root.bo);
2099
2100	INIT_KFIFO(vm->faults);
2101
2102	return 0;
2103
2104error_unreserve:
 
2105	amdgpu_bo_unreserve(vm->root.bo);
2106
2107error_free_root:
2108	amdgpu_bo_unref(&root->shadow);
2109	amdgpu_bo_unref(&root_bo);
2110	vm->root.bo = NULL;
2111
2112error_free_delayed:
2113	dma_fence_put(vm->last_tlb_flush);
2114	dma_fence_put(vm->last_unlocked);
2115	drm_sched_entity_destroy(&vm->delayed);
2116
2117error_free_immediate:
2118	drm_sched_entity_destroy(&vm->immediate);
2119
2120	return r;
2121}
2122
2123/**
2124 * amdgpu_vm_make_compute - Turn a GFX VM into a compute VM
2125 *
2126 * @adev: amdgpu_device pointer
2127 * @vm: requested vm
2128 *
2129 * This only works on GFX VMs that don't have any BOs added and no
2130 * page tables allocated yet.
2131 *
2132 * Changes the following VM parameters:
2133 * - use_cpu_for_update
2134 * - pte_supports_ats
2135 *
2136 * Reinitializes the page directory to reflect the changed ATS
2137 * setting.
2138 *
2139 * Returns:
2140 * 0 for success, -errno for errors.
2141 */
2142int amdgpu_vm_make_compute(struct amdgpu_device *adev, struct amdgpu_vm *vm)
2143{
2144	bool pte_support_ats = (adev->asic_type == CHIP_RAVEN);
2145	int r;
2146
2147	r = amdgpu_bo_reserve(vm->root.bo, true);
2148	if (r)
2149		return r;
2150
2151	/* Sanity checks */
2152	if (!amdgpu_vm_pt_is_root_clean(adev, vm)) {
2153		r = -EINVAL;
2154		goto unreserve_bo;
2155	}
2156
2157	/* Check if PD needs to be reinitialized and do it before
2158	 * changing any other state, in case it fails.
2159	 */
2160	if (pte_support_ats != vm->pte_support_ats) {
 
 
 
 
 
 
2161		vm->pte_support_ats = pte_support_ats;
2162		r = amdgpu_vm_pt_clear(adev, vm, to_amdgpu_bo_vm(vm->root.bo),
2163				       false);
2164		if (r)
2165			goto unreserve_bo;
2166	}
2167
2168	/* Update VM state */
2169	vm->use_cpu_for_update = !!(adev->vm_manager.vm_update_mode &
2170				    AMDGPU_VM_USE_CPU_FOR_COMPUTE);
2171	DRM_DEBUG_DRIVER("VM update mode is %s\n",
2172			 vm->use_cpu_for_update ? "CPU" : "SDMA");
2173	WARN_ONCE((vm->use_cpu_for_update &&
2174		   !amdgpu_gmc_vram_full_visible(&adev->gmc)),
2175		  "CPU update of VM recommended only for large BAR system\n");
2176
2177	if (vm->use_cpu_for_update) {
2178		/* Sync with last SDMA update/clear before switching to CPU */
2179		r = amdgpu_bo_sync_wait(vm->root.bo,
2180					AMDGPU_FENCE_OWNER_UNDEFINED, true);
2181		if (r)
2182			goto unreserve_bo;
2183
2184		vm->update_funcs = &amdgpu_vm_cpu_funcs;
 
 
 
 
2185	} else {
2186		vm->update_funcs = &amdgpu_vm_sdma_funcs;
2187	}
2188	/*
2189	 * Make sure root PD gets mapped. As vm_update_mode could be changed
2190	 * when turning a GFX VM into a compute VM.
2191	 */
2192	r = vm->update_funcs->map_table(to_amdgpu_bo_vm(vm->root.bo));
2193	if (r)
2194		goto unreserve_bo;
2195
2196	dma_fence_put(vm->last_update);
2197	vm->last_update = NULL;
2198	vm->is_compute_context = true;
2199
2200	/* Free the shadow bo for compute VM */
2201	amdgpu_bo_unref(&to_amdgpu_bo_vm(vm->root.bo)->shadow);
2202
2203	goto unreserve_bo;
2204
2205unreserve_bo:
2206	amdgpu_bo_unreserve(vm->root.bo);
2207	return r;
2208}
2209
2210/**
2211 * amdgpu_vm_release_compute - release a compute vm
2212 * @adev: amdgpu_device pointer
2213 * @vm: a vm turned into compute vm by calling amdgpu_vm_make_compute
2214 *
2215 * This is a correspondant of amdgpu_vm_make_compute. It decouples compute
2216 * pasid from vm. Compute should stop use of vm after this call.
2217 */
2218void amdgpu_vm_release_compute(struct amdgpu_device *adev, struct amdgpu_vm *vm)
2219{
2220	amdgpu_vm_set_pasid(adev, vm, 0);
2221	vm->is_compute_context = false;
2222}
2223
2224/**
2225 * amdgpu_vm_fini - tear down a vm instance
2226 *
2227 * @adev: amdgpu_device pointer
2228 * @vm: requested vm
2229 *
2230 * Tear down @vm.
2231 * Unbind the VM and remove all bos from the vm bo list
2232 */
2233void amdgpu_vm_fini(struct amdgpu_device *adev, struct amdgpu_vm *vm)
2234{
2235	struct amdgpu_bo_va_mapping *mapping, *tmp;
2236	bool prt_fini_needed = !!adev->gmc.gmc_funcs->set_prt;
2237	struct amdgpu_bo *root;
2238	unsigned long flags;
2239	int i;
2240
2241	amdgpu_amdkfd_gpuvm_destroy_cb(adev, vm);
2242
2243	flush_work(&vm->pt_free_work);
2244
2245	root = amdgpu_bo_ref(vm->root.bo);
2246	amdgpu_bo_reserve(root, true);
2247	amdgpu_vm_set_pasid(adev, vm, 0);
2248	dma_fence_wait(vm->last_unlocked, false);
2249	dma_fence_put(vm->last_unlocked);
2250	dma_fence_wait(vm->last_tlb_flush, false);
2251	/* Make sure that all fence callbacks have completed */
2252	spin_lock_irqsave(vm->last_tlb_flush->lock, flags);
2253	spin_unlock_irqrestore(vm->last_tlb_flush->lock, flags);
2254	dma_fence_put(vm->last_tlb_flush);
2255
2256	list_for_each_entry_safe(mapping, tmp, &vm->freed, list) {
2257		if (mapping->flags & AMDGPU_PTE_PRT && prt_fini_needed) {
2258			amdgpu_vm_prt_fini(adev, vm);
2259			prt_fini_needed = false;
2260		}
2261
2262		list_del(&mapping->list);
2263		amdgpu_vm_free_mapping(adev, vm, mapping, NULL);
2264	}
2265
2266	amdgpu_vm_pt_free_root(adev, vm);
2267	amdgpu_bo_unreserve(root);
2268	amdgpu_bo_unref(&root);
2269	WARN_ON(vm->root.bo);
2270
2271	drm_sched_entity_destroy(&vm->immediate);
2272	drm_sched_entity_destroy(&vm->delayed);
2273
2274	if (!RB_EMPTY_ROOT(&vm->va.rb_root)) {
2275		dev_err(adev->dev, "still active bo inside vm\n");
2276	}
2277	rbtree_postorder_for_each_entry_safe(mapping, tmp,
2278					     &vm->va.rb_root, rb) {
2279		/* Don't remove the mapping here, we don't want to trigger a
2280		 * rebalance and the tree is about to be destroyed anyway.
2281		 */
2282		list_del(&mapping->list);
2283		kfree(mapping);
2284	}
2285
2286	dma_fence_put(vm->last_update);
2287	for (i = 0; i < AMDGPU_MAX_VMHUBS; i++)
2288		amdgpu_vmid_free_reserved(adev, vm, i);
 
 
 
 
 
 
2289}
2290
2291/**
2292 * amdgpu_vm_manager_init - init the VM manager
2293 *
2294 * @adev: amdgpu_device pointer
2295 *
2296 * Initialize the VM manager structures
2297 */
2298void amdgpu_vm_manager_init(struct amdgpu_device *adev)
2299{
2300	unsigned i;
2301
2302	/* Concurrent flushes are only possible starting with Vega10 and
2303	 * are broken on Navi10 and Navi14.
2304	 */
2305	adev->vm_manager.concurrent_flush = !(adev->asic_type < CHIP_VEGA10 ||
2306					      adev->asic_type == CHIP_NAVI10 ||
2307					      adev->asic_type == CHIP_NAVI14);
2308	amdgpu_vmid_mgr_init(adev);
2309
2310	adev->vm_manager.fence_context =
2311		dma_fence_context_alloc(AMDGPU_MAX_RINGS);
2312	for (i = 0; i < AMDGPU_MAX_RINGS; ++i)
2313		adev->vm_manager.seqno[i] = 0;
2314
2315	spin_lock_init(&adev->vm_manager.prt_lock);
2316	atomic_set(&adev->vm_manager.num_prt_users, 0);
2317
2318	/* If not overridden by the user, by default, only in large BAR systems
2319	 * Compute VM tables will be updated by CPU
2320	 */
2321#ifdef CONFIG_X86_64
2322	if (amdgpu_vm_update_mode == -1) {
2323		/* For asic with VF MMIO access protection
2324		 * avoid using CPU for VM table updates
2325		 */
2326		if (amdgpu_gmc_vram_full_visible(&adev->gmc) &&
2327		    !amdgpu_sriov_vf_mmio_access_protection(adev))
2328			adev->vm_manager.vm_update_mode =
2329				AMDGPU_VM_USE_CPU_FOR_COMPUTE;
2330		else
2331			adev->vm_manager.vm_update_mode = 0;
2332	} else
2333		adev->vm_manager.vm_update_mode = amdgpu_vm_update_mode;
2334#else
2335	adev->vm_manager.vm_update_mode = 0;
2336#endif
2337
2338	xa_init_flags(&adev->vm_manager.pasids, XA_FLAGS_LOCK_IRQ);
2339}
2340
2341/**
2342 * amdgpu_vm_manager_fini - cleanup VM manager
2343 *
2344 * @adev: amdgpu_device pointer
2345 *
2346 * Cleanup the VM manager and free resources.
2347 */
2348void amdgpu_vm_manager_fini(struct amdgpu_device *adev)
2349{
2350	WARN_ON(!xa_empty(&adev->vm_manager.pasids));
2351	xa_destroy(&adev->vm_manager.pasids);
2352
2353	amdgpu_vmid_mgr_fini(adev);
2354}
2355
2356/**
2357 * amdgpu_vm_ioctl - Manages VMID reservation for vm hubs.
2358 *
2359 * @dev: drm device pointer
2360 * @data: drm_amdgpu_vm
2361 * @filp: drm file pointer
2362 *
2363 * Returns:
2364 * 0 for success, -errno for errors.
2365 */
2366int amdgpu_vm_ioctl(struct drm_device *dev, void *data, struct drm_file *filp)
2367{
2368	union drm_amdgpu_vm *args = data;
2369	struct amdgpu_device *adev = drm_to_adev(dev);
2370	struct amdgpu_fpriv *fpriv = filp->driver_priv;
2371	int r;
 
 
 
2372
2373	switch (args->in.op) {
2374	case AMDGPU_VM_OP_RESERVE_VMID:
2375		/* We only have requirement to reserve vmid from gfxhub */
2376		r = amdgpu_vmid_alloc_reserved(adev, &fpriv->vm,
2377					       AMDGPU_GFXHUB_0);
2378		if (r)
2379			return r;
 
2380		break;
2381	case AMDGPU_VM_OP_UNRESERVE_VMID:
2382		amdgpu_vmid_free_reserved(adev, &fpriv->vm, AMDGPU_GFXHUB_0);
 
 
 
2383		break;
2384	default:
2385		return -EINVAL;
2386	}
2387
2388	return 0;
2389}
2390
2391/**
2392 * amdgpu_vm_get_task_info - Extracts task info for a PASID.
2393 *
2394 * @adev: drm device pointer
2395 * @pasid: PASID identifier for VM
2396 * @task_info: task_info to fill.
2397 */
2398void amdgpu_vm_get_task_info(struct amdgpu_device *adev, u32 pasid,
2399			 struct amdgpu_task_info *task_info)
2400{
2401	struct amdgpu_vm *vm;
2402	unsigned long flags;
2403
2404	xa_lock_irqsave(&adev->vm_manager.pasids, flags);
2405
2406	vm = xa_load(&adev->vm_manager.pasids, pasid);
2407	if (vm)
2408		*task_info = vm->task_info;
2409
2410	xa_unlock_irqrestore(&adev->vm_manager.pasids, flags);
2411}
2412
2413/**
2414 * amdgpu_vm_set_task_info - Sets VMs task info.
2415 *
2416 * @vm: vm for which to set the info
2417 */
2418void amdgpu_vm_set_task_info(struct amdgpu_vm *vm)
2419{
2420	if (vm->task_info.pid)
2421		return;
2422
2423	vm->task_info.pid = current->pid;
2424	get_task_comm(vm->task_info.task_name, current);
2425
2426	if (current->group_leader->mm != current->mm)
2427		return;
2428
2429	vm->task_info.tgid = current->group_leader->pid;
2430	get_task_comm(vm->task_info.process_name, current->group_leader);
2431}
2432
2433/**
2434 * amdgpu_vm_handle_fault - graceful handling of VM faults.
2435 * @adev: amdgpu device pointer
2436 * @pasid: PASID of the VM
 
 
 
2437 * @addr: Address of the fault
2438 * @write_fault: true is write fault, false is read fault
2439 *
2440 * Try to gracefully handle a VM fault. Return true if the fault was handled and
2441 * shouldn't be reported any more.
2442 */
2443bool amdgpu_vm_handle_fault(struct amdgpu_device *adev, u32 pasid,
2444			    uint64_t addr, bool write_fault)
 
2445{
2446	bool is_compute_context = false;
2447	struct amdgpu_bo *root;
2448	unsigned long irqflags;
2449	uint64_t value, flags;
2450	struct amdgpu_vm *vm;
2451	int r;
2452
2453	xa_lock_irqsave(&adev->vm_manager.pasids, irqflags);
2454	vm = xa_load(&adev->vm_manager.pasids, pasid);
2455	if (vm) {
2456		root = amdgpu_bo_ref(vm->root.bo);
2457		is_compute_context = vm->is_compute_context;
2458	} else {
2459		root = NULL;
2460	}
2461	xa_unlock_irqrestore(&adev->vm_manager.pasids, irqflags);
2462
2463	if (!root)
2464		return false;
2465
2466	addr /= AMDGPU_GPU_PAGE_SIZE;
2467
2468	if (is_compute_context &&
2469	    !svm_range_restore_pages(adev, pasid, addr, write_fault)) {
2470		amdgpu_bo_unref(&root);
2471		return true;
2472	}
2473
2474	r = amdgpu_bo_reserve(root, true);
2475	if (r)
2476		goto error_unref;
2477
2478	/* Double check that the VM still exists */
2479	xa_lock_irqsave(&adev->vm_manager.pasids, irqflags);
2480	vm = xa_load(&adev->vm_manager.pasids, pasid);
2481	if (vm && vm->root.bo != root)
2482		vm = NULL;
2483	xa_unlock_irqrestore(&adev->vm_manager.pasids, irqflags);
2484	if (!vm)
2485		goto error_unlock;
2486
2487	flags = AMDGPU_PTE_VALID | AMDGPU_PTE_SNOOPED |
2488		AMDGPU_PTE_SYSTEM;
2489
2490	if (is_compute_context) {
2491		/* Intentionally setting invalid PTE flag
2492		 * combination to force a no-retry-fault
2493		 */
2494		flags = AMDGPU_PTE_SNOOPED | AMDGPU_PTE_PRT;
2495		value = 0;
2496	} else if (amdgpu_vm_fault_stop == AMDGPU_VM_FAULT_STOP_NEVER) {
2497		/* Redirect the access to the dummy page */
2498		value = adev->dummy_page_addr;
2499		flags |= AMDGPU_PTE_EXECUTABLE | AMDGPU_PTE_READABLE |
2500			AMDGPU_PTE_WRITEABLE;
2501
2502	} else {
2503		/* Let the hw retry silently on the PTE */
2504		value = 0;
2505	}
2506
2507	r = dma_resv_reserve_fences(root->tbo.base.resv, 1);
2508	if (r) {
2509		pr_debug("failed %d to reserve fence slot\n", r);
2510		goto error_unlock;
2511	}
2512
2513	r = amdgpu_vm_update_range(adev, vm, true, false, false, NULL, addr,
2514				   addr, flags, value, 0, NULL, NULL, NULL);
2515	if (r)
2516		goto error_unlock;
2517
2518	r = amdgpu_vm_update_pdes(adev, vm, true);
2519
2520error_unlock:
2521	amdgpu_bo_unreserve(root);
2522	if (r < 0)
2523		DRM_ERROR("Can't handle page fault (%d)\n", r);
2524
2525error_unref:
2526	amdgpu_bo_unref(&root);
2527
2528	return false;
2529}
2530
2531#if defined(CONFIG_DEBUG_FS)
2532/**
2533 * amdgpu_debugfs_vm_bo_info  - print BO info for the VM
2534 *
2535 * @vm: Requested VM for printing BO info
2536 * @m: debugfs file
2537 *
2538 * Print BO information in debugfs file for the VM
2539 */
2540void amdgpu_debugfs_vm_bo_info(struct amdgpu_vm *vm, struct seq_file *m)
2541{
2542	struct amdgpu_bo_va *bo_va, *tmp;
2543	u64 total_idle = 0;
2544	u64 total_evicted = 0;
2545	u64 total_relocated = 0;
2546	u64 total_moved = 0;
2547	u64 total_invalidated = 0;
2548	u64 total_done = 0;
2549	unsigned int total_idle_objs = 0;
2550	unsigned int total_evicted_objs = 0;
2551	unsigned int total_relocated_objs = 0;
2552	unsigned int total_moved_objs = 0;
2553	unsigned int total_invalidated_objs = 0;
2554	unsigned int total_done_objs = 0;
2555	unsigned int id = 0;
2556
2557	spin_lock(&vm->status_lock);
2558	seq_puts(m, "\tIdle BOs:\n");
2559	list_for_each_entry_safe(bo_va, tmp, &vm->idle, base.vm_status) {
2560		if (!bo_va->base.bo)
2561			continue;
2562		total_idle += amdgpu_bo_print_info(id++, bo_va->base.bo, m);
2563	}
2564	total_idle_objs = id;
2565	id = 0;
2566
2567	seq_puts(m, "\tEvicted BOs:\n");
2568	list_for_each_entry_safe(bo_va, tmp, &vm->evicted, base.vm_status) {
2569		if (!bo_va->base.bo)
2570			continue;
2571		total_evicted += amdgpu_bo_print_info(id++, bo_va->base.bo, m);
2572	}
2573	total_evicted_objs = id;
2574	id = 0;
2575
2576	seq_puts(m, "\tRelocated BOs:\n");
2577	list_for_each_entry_safe(bo_va, tmp, &vm->relocated, base.vm_status) {
2578		if (!bo_va->base.bo)
2579			continue;
2580		total_relocated += amdgpu_bo_print_info(id++, bo_va->base.bo, m);
2581	}
2582	total_relocated_objs = id;
2583	id = 0;
2584
2585	seq_puts(m, "\tMoved BOs:\n");
2586	list_for_each_entry_safe(bo_va, tmp, &vm->moved, base.vm_status) {
2587		if (!bo_va->base.bo)
2588			continue;
2589		total_moved += amdgpu_bo_print_info(id++, bo_va->base.bo, m);
2590	}
2591	total_moved_objs = id;
2592	id = 0;
2593
2594	seq_puts(m, "\tInvalidated BOs:\n");
2595	list_for_each_entry_safe(bo_va, tmp, &vm->invalidated, base.vm_status) {
2596		if (!bo_va->base.bo)
2597			continue;
2598		total_invalidated += amdgpu_bo_print_info(id++,	bo_va->base.bo, m);
2599	}
2600	total_invalidated_objs = id;
2601	id = 0;
2602
2603	seq_puts(m, "\tDone BOs:\n");
2604	list_for_each_entry_safe(bo_va, tmp, &vm->done, base.vm_status) {
2605		if (!bo_va->base.bo)
2606			continue;
2607		total_done += amdgpu_bo_print_info(id++, bo_va->base.bo, m);
2608	}
2609	spin_unlock(&vm->status_lock);
2610	total_done_objs = id;
2611
2612	seq_printf(m, "\tTotal idle size:        %12lld\tobjs:\t%d\n", total_idle,
2613		   total_idle_objs);
2614	seq_printf(m, "\tTotal evicted size:     %12lld\tobjs:\t%d\n", total_evicted,
2615		   total_evicted_objs);
2616	seq_printf(m, "\tTotal relocated size:   %12lld\tobjs:\t%d\n", total_relocated,
2617		   total_relocated_objs);
2618	seq_printf(m, "\tTotal moved size:       %12lld\tobjs:\t%d\n", total_moved,
2619		   total_moved_objs);
2620	seq_printf(m, "\tTotal invalidated size: %12lld\tobjs:\t%d\n", total_invalidated,
2621		   total_invalidated_objs);
2622	seq_printf(m, "\tTotal done size:        %12lld\tobjs:\t%d\n", total_done,
2623		   total_done_objs);
2624}
2625#endif