Linux Audio

Check our new training course

Yocto / OpenEmbedded training

Feb 10-13, 2025
Register
Loading...
v5.14.15
   1/*
   2 * Copyright 2008 Advanced Micro Devices, Inc.
   3 * Copyright 2008 Red Hat Inc.
   4 * Copyright 2009 Jerome Glisse.
   5 *
   6 * Permission is hereby granted, free of charge, to any person obtaining a
   7 * copy of this software and associated documentation files (the "Software"),
   8 * to deal in the Software without restriction, including without limitation
   9 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
  10 * and/or sell copies of the Software, and to permit persons to whom the
  11 * Software is furnished to do so, subject to the following conditions:
  12 *
  13 * The above copyright notice and this permission notice shall be included in
  14 * all copies or substantial portions of the Software.
  15 *
  16 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  17 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  18 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
  19 * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR
  20 * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
  21 * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
  22 * OTHER DEALINGS IN THE SOFTWARE.
  23 *
  24 * Authors: Dave Airlie
  25 *          Alex Deucher
  26 *          Jerome Glisse
  27 */
  28
  29#include <linux/dma-fence-array.h>
  30#include <linux/interval_tree_generic.h>
  31#include <linux/idr.h>
  32#include <linux/dma-buf.h>
  33
  34#include <drm/amdgpu_drm.h>
  35#include <drm/drm_drv.h>
  36#include "amdgpu.h"
  37#include "amdgpu_trace.h"
  38#include "amdgpu_amdkfd.h"
  39#include "amdgpu_gmc.h"
  40#include "amdgpu_xgmi.h"
  41#include "amdgpu_dma_buf.h"
  42#include "amdgpu_res_cursor.h"
  43#include "kfd_svm.h"
  44
  45/**
  46 * DOC: GPUVM
  47 *
  48 * GPUVM is similar to the legacy gart on older asics, however
  49 * rather than there being a single global gart table
  50 * for the entire GPU, there are multiple VM page tables active
  51 * at any given time.  The VM page tables can contain a mix
  52 * vram pages and system memory pages and system memory pages
  53 * can be mapped as snooped (cached system pages) or unsnooped
  54 * (uncached system pages).
  55 * Each VM has an ID associated with it and there is a page table
  56 * associated with each VMID.  When execting a command buffer,
  57 * the kernel tells the the ring what VMID to use for that command
  58 * buffer.  VMIDs are allocated dynamically as commands are submitted.
  59 * The userspace drivers maintain their own address space and the kernel
  60 * sets up their pages tables accordingly when they submit their
  61 * command buffers and a VMID is assigned.
  62 * Cayman/Trinity support up to 8 active VMs at any given time;
  63 * SI supports 16.
  64 */
  65
  66#define START(node) ((node)->start)
  67#define LAST(node) ((node)->last)
  68
  69INTERVAL_TREE_DEFINE(struct amdgpu_bo_va_mapping, rb, uint64_t, __subtree_last,
  70		     START, LAST, static, amdgpu_vm_it)
  71
  72#undef START
  73#undef LAST
  74
  75/**
  76 * struct amdgpu_prt_cb - Helper to disable partial resident texture feature from a fence callback
  77 */
  78struct amdgpu_prt_cb {
  79
  80	/**
  81	 * @adev: amdgpu device
  82	 */
  83	struct amdgpu_device *adev;
  84
  85	/**
  86	 * @cb: callback
  87	 */
  88	struct dma_fence_cb cb;
  89};
  90
  91/*
  92 * vm eviction_lock can be taken in MMU notifiers. Make sure no reclaim-FS
  93 * happens while holding this lock anywhere to prevent deadlocks when
  94 * an MMU notifier runs in reclaim-FS context.
  95 */
  96static inline void amdgpu_vm_eviction_lock(struct amdgpu_vm *vm)
  97{
  98	mutex_lock(&vm->eviction_lock);
  99	vm->saved_flags = memalloc_noreclaim_save();
 100}
 101
 102static inline int amdgpu_vm_eviction_trylock(struct amdgpu_vm *vm)
 103{
 104	if (mutex_trylock(&vm->eviction_lock)) {
 105		vm->saved_flags = memalloc_noreclaim_save();
 106		return 1;
 107	}
 108	return 0;
 109}
 110
 111static inline void amdgpu_vm_eviction_unlock(struct amdgpu_vm *vm)
 112{
 113	memalloc_noreclaim_restore(vm->saved_flags);
 114	mutex_unlock(&vm->eviction_lock);
 115}
 116
 117/**
 118 * amdgpu_vm_level_shift - return the addr shift for each level
 119 *
 120 * @adev: amdgpu_device pointer
 121 * @level: VMPT level
 122 *
 123 * Returns:
 124 * The number of bits the pfn needs to be right shifted for a level.
 125 */
 126static unsigned amdgpu_vm_level_shift(struct amdgpu_device *adev,
 127				      unsigned level)
 128{
 
 
 129	switch (level) {
 130	case AMDGPU_VM_PDB2:
 131	case AMDGPU_VM_PDB1:
 132	case AMDGPU_VM_PDB0:
 133		return 9 * (AMDGPU_VM_PDB0 - level) +
 134			adev->vm_manager.block_size;
 
 135	case AMDGPU_VM_PTB:
 136		return 0;
 
 137	default:
 138		return ~0;
 139	}
 
 
 140}
 141
 142/**
 143 * amdgpu_vm_num_entries - return the number of entries in a PD/PT
 144 *
 145 * @adev: amdgpu_device pointer
 146 * @level: VMPT level
 147 *
 148 * Returns:
 149 * The number of entries in a page directory or page table.
 150 */
 151static unsigned amdgpu_vm_num_entries(struct amdgpu_device *adev,
 152				      unsigned level)
 153{
 154	unsigned shift = amdgpu_vm_level_shift(adev,
 155					       adev->vm_manager.root_level);
 156
 157	if (level == adev->vm_manager.root_level)
 158		/* For the root directory */
 159		return round_up(adev->vm_manager.max_pfn, 1ULL << shift)
 160			>> shift;
 161	else if (level != AMDGPU_VM_PTB)
 162		/* Everything in between */
 163		return 512;
 164	else
 165		/* For the page tables on the leaves */
 166		return AMDGPU_VM_PTE_COUNT(adev);
 167}
 168
 169/**
 170 * amdgpu_vm_num_ats_entries - return the number of ATS entries in the root PD
 171 *
 172 * @adev: amdgpu_device pointer
 173 *
 174 * Returns:
 175 * The number of entries in the root page directory which needs the ATS setting.
 176 */
 177static unsigned amdgpu_vm_num_ats_entries(struct amdgpu_device *adev)
 178{
 179	unsigned shift;
 180
 181	shift = amdgpu_vm_level_shift(adev, adev->vm_manager.root_level);
 182	return AMDGPU_GMC_HOLE_START >> (shift + AMDGPU_GPU_PAGE_SHIFT);
 183}
 184
 185/**
 186 * amdgpu_vm_entries_mask - the mask to get the entry number of a PD/PT
 187 *
 188 * @adev: amdgpu_device pointer
 189 * @level: VMPT level
 190 *
 191 * Returns:
 192 * The mask to extract the entry number of a PD/PT from an address.
 193 */
 194static uint32_t amdgpu_vm_entries_mask(struct amdgpu_device *adev,
 195				       unsigned int level)
 196{
 197	if (level <= adev->vm_manager.root_level)
 198		return 0xffffffff;
 199	else if (level != AMDGPU_VM_PTB)
 200		return 0x1ff;
 201	else
 202		return AMDGPU_VM_PTE_COUNT(adev) - 1;
 203}
 204
 205/**
 206 * amdgpu_vm_bo_size - returns the size of the BOs in bytes
 207 *
 208 * @adev: amdgpu_device pointer
 209 * @level: VMPT level
 210 *
 211 * Returns:
 212 * The size of the BO for a page directory or page table in bytes.
 213 */
 214static unsigned amdgpu_vm_bo_size(struct amdgpu_device *adev, unsigned level)
 215{
 216	return AMDGPU_GPU_PAGE_ALIGN(amdgpu_vm_num_entries(adev, level) * 8);
 217}
 218
 219/**
 220 * amdgpu_vm_bo_evicted - vm_bo is evicted
 221 *
 222 * @vm_bo: vm_bo which is evicted
 223 *
 224 * State for PDs/PTs and per VM BOs which are not at the location they should
 225 * be.
 226 */
 227static void amdgpu_vm_bo_evicted(struct amdgpu_vm_bo_base *vm_bo)
 228{
 229	struct amdgpu_vm *vm = vm_bo->vm;
 230	struct amdgpu_bo *bo = vm_bo->bo;
 231
 232	vm_bo->moved = true;
 233	if (bo->tbo.type == ttm_bo_type_kernel)
 234		list_move(&vm_bo->vm_status, &vm->evicted);
 235	else
 236		list_move_tail(&vm_bo->vm_status, &vm->evicted);
 237}
 
 
 
 
 
 
 
 
 
 
 
 
 
 238/**
 239 * amdgpu_vm_bo_moved - vm_bo is moved
 240 *
 241 * @vm_bo: vm_bo which is moved
 242 *
 243 * State for per VM BOs which are moved, but that change is not yet reflected
 244 * in the page tables.
 245 */
 246static void amdgpu_vm_bo_moved(struct amdgpu_vm_bo_base *vm_bo)
 247{
 248	list_move(&vm_bo->vm_status, &vm_bo->vm->moved);
 249}
 250
 251/**
 252 * amdgpu_vm_bo_idle - vm_bo is idle
 253 *
 254 * @vm_bo: vm_bo which is now idle
 255 *
 256 * State for PDs/PTs and per VM BOs which have gone through the state machine
 257 * and are now idle.
 258 */
 259static void amdgpu_vm_bo_idle(struct amdgpu_vm_bo_base *vm_bo)
 260{
 261	list_move(&vm_bo->vm_status, &vm_bo->vm->idle);
 262	vm_bo->moved = false;
 263}
 264
 265/**
 266 * amdgpu_vm_bo_invalidated - vm_bo is invalidated
 267 *
 268 * @vm_bo: vm_bo which is now invalidated
 269 *
 270 * State for normal BOs which are invalidated and that change not yet reflected
 271 * in the PTs.
 272 */
 273static void amdgpu_vm_bo_invalidated(struct amdgpu_vm_bo_base *vm_bo)
 274{
 275	spin_lock(&vm_bo->vm->invalidated_lock);
 276	list_move(&vm_bo->vm_status, &vm_bo->vm->invalidated);
 277	spin_unlock(&vm_bo->vm->invalidated_lock);
 278}
 279
 280/**
 281 * amdgpu_vm_bo_relocated - vm_bo is reloacted
 282 *
 283 * @vm_bo: vm_bo which is relocated
 284 *
 285 * State for PDs/PTs which needs to update their parent PD.
 286 * For the root PD, just move to idle state.
 287 */
 288static void amdgpu_vm_bo_relocated(struct amdgpu_vm_bo_base *vm_bo)
 289{
 290	if (vm_bo->bo->parent)
 291		list_move(&vm_bo->vm_status, &vm_bo->vm->relocated);
 292	else
 293		amdgpu_vm_bo_idle(vm_bo);
 294}
 295
 296/**
 297 * amdgpu_vm_bo_done - vm_bo is done
 298 *
 299 * @vm_bo: vm_bo which is now done
 300 *
 301 * State for normal BOs which are invalidated and that change has been updated
 302 * in the PTs.
 303 */
 304static void amdgpu_vm_bo_done(struct amdgpu_vm_bo_base *vm_bo)
 305{
 306	spin_lock(&vm_bo->vm->invalidated_lock);
 307	list_move(&vm_bo->vm_status, &vm_bo->vm->done);
 308	spin_unlock(&vm_bo->vm->invalidated_lock);
 309}
 310
 311/**
 312 * amdgpu_vm_bo_base_init - Adds bo to the list of bos associated with the vm
 313 *
 314 * @base: base structure for tracking BO usage in a VM
 315 * @vm: vm to which bo is to be added
 316 * @bo: amdgpu buffer object
 317 *
 318 * Initialize a bo_va_base structure and add it to the appropriate lists
 319 *
 320 */
 321static void amdgpu_vm_bo_base_init(struct amdgpu_vm_bo_base *base,
 322				   struct amdgpu_vm *vm,
 323				   struct amdgpu_bo *bo)
 324{
 325	base->vm = vm;
 326	base->bo = bo;
 327	base->next = NULL;
 328	INIT_LIST_HEAD(&base->vm_status);
 329
 330	if (!bo)
 331		return;
 332	base->next = bo->vm_bo;
 333	bo->vm_bo = base;
 334
 335	if (bo->tbo.base.resv != vm->root.bo->tbo.base.resv)
 336		return;
 337
 338	vm->bulk_moveable = false;
 339	if (bo->tbo.type == ttm_bo_type_kernel && bo->parent)
 340		amdgpu_vm_bo_relocated(base);
 341	else
 342		amdgpu_vm_bo_idle(base);
 343
 344	if (bo->preferred_domains &
 345	    amdgpu_mem_type_to_domain(bo->tbo.resource->mem_type))
 346		return;
 347
 348	/*
 349	 * we checked all the prerequisites, but it looks like this per vm bo
 350	 * is currently evicted. add the bo to the evicted list to make sure it
 351	 * is validated on next vm use to avoid fault.
 352	 * */
 353	amdgpu_vm_bo_evicted(base);
 354}
 355
 356/**
 357 * amdgpu_vm_pt_parent - get the parent page directory
 358 *
 359 * @pt: child page table
 360 *
 361 * Helper to get the parent entry for the child page table. NULL if we are at
 362 * the root page directory.
 363 */
 364static struct amdgpu_vm_bo_base *amdgpu_vm_pt_parent(struct amdgpu_vm_bo_base *pt)
 365{
 366	struct amdgpu_bo *parent = pt->bo->parent;
 367
 368	if (!parent)
 369		return NULL;
 370
 371	return parent->vm_bo;
 372}
 373
 374/*
 375 * amdgpu_vm_pt_cursor - state for for_each_amdgpu_vm_pt
 376 */
 377struct amdgpu_vm_pt_cursor {
 378	uint64_t pfn;
 379	struct amdgpu_vm_bo_base *parent;
 380	struct amdgpu_vm_bo_base *entry;
 381	unsigned level;
 382};
 383
 384/**
 385 * amdgpu_vm_pt_start - start PD/PT walk
 386 *
 387 * @adev: amdgpu_device pointer
 388 * @vm: amdgpu_vm structure
 389 * @start: start address of the walk
 390 * @cursor: state to initialize
 391 *
 392 * Initialize a amdgpu_vm_pt_cursor to start a walk.
 393 */
 394static void amdgpu_vm_pt_start(struct amdgpu_device *adev,
 395			       struct amdgpu_vm *vm, uint64_t start,
 396			       struct amdgpu_vm_pt_cursor *cursor)
 397{
 398	cursor->pfn = start;
 399	cursor->parent = NULL;
 400	cursor->entry = &vm->root;
 401	cursor->level = adev->vm_manager.root_level;
 402}
 403
 404/**
 405 * amdgpu_vm_pt_descendant - go to child node
 406 *
 407 * @adev: amdgpu_device pointer
 408 * @cursor: current state
 409 *
 410 * Walk to the child node of the current node.
 411 * Returns:
 412 * True if the walk was possible, false otherwise.
 413 */
 414static bool amdgpu_vm_pt_descendant(struct amdgpu_device *adev,
 415				    struct amdgpu_vm_pt_cursor *cursor)
 416{
 417	unsigned mask, shift, idx;
 418
 419	if ((cursor->level == AMDGPU_VM_PTB) || !cursor->entry ||
 420	    !cursor->entry->bo)
 421		return false;
 422
 
 423	mask = amdgpu_vm_entries_mask(adev, cursor->level);
 424	shift = amdgpu_vm_level_shift(adev, cursor->level);
 425
 426	++cursor->level;
 427	idx = (cursor->pfn >> shift) & mask;
 428	cursor->parent = cursor->entry;
 429	cursor->entry = &to_amdgpu_bo_vm(cursor->entry->bo)->entries[idx];
 430	return true;
 431}
 432
 433/**
 434 * amdgpu_vm_pt_sibling - go to sibling node
 435 *
 436 * @adev: amdgpu_device pointer
 437 * @cursor: current state
 438 *
 439 * Walk to the sibling node of the current node.
 440 * Returns:
 441 * True if the walk was possible, false otherwise.
 442 */
 443static bool amdgpu_vm_pt_sibling(struct amdgpu_device *adev,
 444				 struct amdgpu_vm_pt_cursor *cursor)
 445{
 446	unsigned shift, num_entries;
 447
 448	/* Root doesn't have a sibling */
 449	if (!cursor->parent)
 450		return false;
 451
 452	/* Go to our parents and see if we got a sibling */
 453	shift = amdgpu_vm_level_shift(adev, cursor->level - 1);
 454	num_entries = amdgpu_vm_num_entries(adev, cursor->level - 1);
 455
 456	if (cursor->entry == &to_amdgpu_bo_vm(cursor->parent->bo)->entries[num_entries - 1])
 457		return false;
 458
 459	cursor->pfn += 1ULL << shift;
 460	cursor->pfn &= ~((1ULL << shift) - 1);
 461	++cursor->entry;
 462	return true;
 463}
 464
 465/**
 466 * amdgpu_vm_pt_ancestor - go to parent node
 467 *
 468 * @cursor: current state
 469 *
 470 * Walk to the parent node of the current node.
 471 * Returns:
 472 * True if the walk was possible, false otherwise.
 473 */
 474static bool amdgpu_vm_pt_ancestor(struct amdgpu_vm_pt_cursor *cursor)
 475{
 476	if (!cursor->parent)
 477		return false;
 478
 479	--cursor->level;
 480	cursor->entry = cursor->parent;
 481	cursor->parent = amdgpu_vm_pt_parent(cursor->parent);
 482	return true;
 483}
 484
 485/**
 486 * amdgpu_vm_pt_next - get next PD/PT in hieratchy
 487 *
 488 * @adev: amdgpu_device pointer
 489 * @cursor: current state
 490 *
 491 * Walk the PD/PT tree to the next node.
 492 */
 493static void amdgpu_vm_pt_next(struct amdgpu_device *adev,
 494			      struct amdgpu_vm_pt_cursor *cursor)
 495{
 496	/* First try a newborn child */
 497	if (amdgpu_vm_pt_descendant(adev, cursor))
 498		return;
 499
 500	/* If that didn't worked try to find a sibling */
 501	while (!amdgpu_vm_pt_sibling(adev, cursor)) {
 502		/* No sibling, go to our parents and grandparents */
 503		if (!amdgpu_vm_pt_ancestor(cursor)) {
 504			cursor->pfn = ~0ll;
 505			return;
 506		}
 507	}
 508}
 509
 510/**
 511 * amdgpu_vm_pt_first_dfs - start a deep first search
 512 *
 513 * @adev: amdgpu_device structure
 514 * @vm: amdgpu_vm structure
 515 * @start: optional cursor to start with
 516 * @cursor: state to initialize
 517 *
 518 * Starts a deep first traversal of the PD/PT tree.
 519 */
 520static void amdgpu_vm_pt_first_dfs(struct amdgpu_device *adev,
 521				   struct amdgpu_vm *vm,
 522				   struct amdgpu_vm_pt_cursor *start,
 523				   struct amdgpu_vm_pt_cursor *cursor)
 524{
 525	if (start)
 526		*cursor = *start;
 527	else
 528		amdgpu_vm_pt_start(adev, vm, 0, cursor);
 529	while (amdgpu_vm_pt_descendant(adev, cursor));
 530}
 531
 532/**
 533 * amdgpu_vm_pt_continue_dfs - check if the deep first search should continue
 534 *
 535 * @start: starting point for the search
 536 * @entry: current entry
 537 *
 538 * Returns:
 539 * True when the search should continue, false otherwise.
 540 */
 541static bool amdgpu_vm_pt_continue_dfs(struct amdgpu_vm_pt_cursor *start,
 542				      struct amdgpu_vm_bo_base *entry)
 543{
 544	return entry && (!start || entry != start->entry);
 545}
 546
 547/**
 548 * amdgpu_vm_pt_next_dfs - get the next node for a deep first search
 549 *
 550 * @adev: amdgpu_device structure
 551 * @cursor: current state
 552 *
 553 * Move the cursor to the next node in a deep first search.
 554 */
 555static void amdgpu_vm_pt_next_dfs(struct amdgpu_device *adev,
 556				  struct amdgpu_vm_pt_cursor *cursor)
 557{
 558	if (!cursor->entry)
 559		return;
 560
 561	if (!cursor->parent)
 562		cursor->entry = NULL;
 563	else if (amdgpu_vm_pt_sibling(adev, cursor))
 564		while (amdgpu_vm_pt_descendant(adev, cursor));
 565	else
 566		amdgpu_vm_pt_ancestor(cursor);
 567}
 568
 569/*
 570 * for_each_amdgpu_vm_pt_dfs_safe - safe deep first search of all PDs/PTs
 571 */
 572#define for_each_amdgpu_vm_pt_dfs_safe(adev, vm, start, cursor, entry)		\
 573	for (amdgpu_vm_pt_first_dfs((adev), (vm), (start), &(cursor)),		\
 574	     (entry) = (cursor).entry, amdgpu_vm_pt_next_dfs((adev), &(cursor));\
 575	     amdgpu_vm_pt_continue_dfs((start), (entry));			\
 576	     (entry) = (cursor).entry, amdgpu_vm_pt_next_dfs((adev), &(cursor)))
 577
 578/**
 579 * amdgpu_vm_get_pd_bo - add the VM PD to a validation list
 580 *
 581 * @vm: vm providing the BOs
 582 * @validated: head of validation list
 583 * @entry: entry to add
 584 *
 585 * Add the page directory to the list of BOs to
 586 * validate for command submission.
 587 */
 588void amdgpu_vm_get_pd_bo(struct amdgpu_vm *vm,
 589			 struct list_head *validated,
 590			 struct amdgpu_bo_list_entry *entry)
 591{
 592	entry->priority = 0;
 593	entry->tv.bo = &vm->root.bo->tbo;
 594	/* Two for VM updates, one for TTM and one for the CS job */
 595	entry->tv.num_shared = 4;
 596	entry->user_pages = NULL;
 597	list_add(&entry->tv.head, validated);
 598}
 599
 600/**
 601 * amdgpu_vm_del_from_lru_notify - update bulk_moveable flag
 602 *
 603 * @bo: BO which was removed from the LRU
 604 *
 605 * Make sure the bulk_moveable flag is updated when a BO is removed from the
 606 * LRU.
 607 */
 608void amdgpu_vm_del_from_lru_notify(struct ttm_buffer_object *bo)
 609{
 610	struct amdgpu_bo *abo;
 611	struct amdgpu_vm_bo_base *bo_base;
 612
 613	if (!amdgpu_bo_is_amdgpu_bo(bo))
 614		return;
 615
 616	if (bo->pin_count)
 617		return;
 618
 619	abo = ttm_to_amdgpu_bo(bo);
 620	if (!abo->parent)
 621		return;
 622	for (bo_base = abo->vm_bo; bo_base; bo_base = bo_base->next) {
 623		struct amdgpu_vm *vm = bo_base->vm;
 624
 625		if (abo->tbo.base.resv == vm->root.bo->tbo.base.resv)
 626			vm->bulk_moveable = false;
 627	}
 628
 629}
 630/**
 631 * amdgpu_vm_move_to_lru_tail - move all BOs to the end of LRU
 632 *
 633 * @adev: amdgpu device pointer
 634 * @vm: vm providing the BOs
 635 *
 636 * Move all BOs to the end of LRU and remember their positions to put them
 637 * together.
 638 */
 639void amdgpu_vm_move_to_lru_tail(struct amdgpu_device *adev,
 640				struct amdgpu_vm *vm)
 641{
 
 642	struct amdgpu_vm_bo_base *bo_base;
 643
 644	if (vm->bulk_moveable) {
 645		spin_lock(&adev->mman.bdev.lru_lock);
 646		ttm_bo_bulk_move_lru_tail(&vm->lru_bulk_move);
 647		spin_unlock(&adev->mman.bdev.lru_lock);
 648		return;
 649	}
 650
 651	memset(&vm->lru_bulk_move, 0, sizeof(vm->lru_bulk_move));
 652
 653	spin_lock(&adev->mman.bdev.lru_lock);
 654	list_for_each_entry(bo_base, &vm->idle, vm_status) {
 655		struct amdgpu_bo *bo = bo_base->bo;
 656		struct amdgpu_bo *shadow = amdgpu_bo_shadowed(bo);
 657
 658		if (!bo->parent)
 659			continue;
 660
 661		ttm_bo_move_to_lru_tail(&bo->tbo, bo->tbo.resource,
 662					&vm->lru_bulk_move);
 663		if (shadow)
 664			ttm_bo_move_to_lru_tail(&shadow->tbo,
 665						shadow->tbo.resource,
 666						&vm->lru_bulk_move);
 667	}
 668	spin_unlock(&adev->mman.bdev.lru_lock);
 669
 670	vm->bulk_moveable = true;
 671}
 672
 673/**
 674 * amdgpu_vm_validate_pt_bos - validate the page table BOs
 675 *
 676 * @adev: amdgpu device pointer
 677 * @vm: vm providing the BOs
 678 * @validate: callback to do the validation
 679 * @param: parameter for the validation callback
 680 *
 681 * Validate the page table BOs on command submission if neccessary.
 682 *
 683 * Returns:
 684 * Validation result.
 685 */
 686int amdgpu_vm_validate_pt_bos(struct amdgpu_device *adev, struct amdgpu_vm *vm,
 687			      int (*validate)(void *p, struct amdgpu_bo *bo),
 688			      void *param)
 689{
 690	struct amdgpu_vm_bo_base *bo_base, *tmp;
 691	int r;
 692
 693	vm->bulk_moveable &= list_empty(&vm->evicted);
 694
 695	list_for_each_entry_safe(bo_base, tmp, &vm->evicted, vm_status) {
 696		struct amdgpu_bo *bo = bo_base->bo;
 697		struct amdgpu_bo *shadow = amdgpu_bo_shadowed(bo);
 698
 699		r = validate(param, bo);
 700		if (r)
 701			return r;
 702		if (shadow) {
 703			r = validate(param, shadow);
 704			if (r)
 705				return r;
 706		}
 707
 708		if (bo->tbo.type != ttm_bo_type_kernel) {
 709			amdgpu_vm_bo_moved(bo_base);
 710		} else {
 711			vm->update_funcs->map_table(to_amdgpu_bo_vm(bo));
 712			amdgpu_vm_bo_relocated(bo_base);
 
 
 
 713		}
 714	}
 715
 716	amdgpu_vm_eviction_lock(vm);
 717	vm->evicting = false;
 718	amdgpu_vm_eviction_unlock(vm);
 719
 720	return 0;
 721}
 722
 723/**
 724 * amdgpu_vm_ready - check VM is ready for updates
 725 *
 726 * @vm: VM to check
 727 *
 728 * Check if all VM PDs/PTs are ready for updates
 729 *
 730 * Returns:
 731 * True if eviction list is empty.
 732 */
 733bool amdgpu_vm_ready(struct amdgpu_vm *vm)
 734{
 735	return list_empty(&vm->evicted);
 736}
 737
 738/**
 739 * amdgpu_vm_clear_bo - initially clear the PDs/PTs
 740 *
 741 * @adev: amdgpu_device pointer
 742 * @vm: VM to clear BO from
 743 * @vmbo: BO to clear
 744 * @immediate: use an immediate update
 745 *
 746 * Root PD needs to be reserved when calling this.
 747 *
 748 * Returns:
 749 * 0 on success, errno otherwise.
 750 */
 751static int amdgpu_vm_clear_bo(struct amdgpu_device *adev,
 752			      struct amdgpu_vm *vm,
 753			      struct amdgpu_bo_vm *vmbo,
 754			      bool immediate)
 755{
 756	struct ttm_operation_ctx ctx = { true, false };
 757	unsigned level = adev->vm_manager.root_level;
 758	struct amdgpu_vm_update_params params;
 759	struct amdgpu_bo *ancestor = &vmbo->bo;
 760	struct amdgpu_bo *bo = &vmbo->bo;
 761	unsigned entries, ats_entries;
 762	uint64_t addr;
 763	int r;
 764
 765	/* Figure out our place in the hierarchy */
 766	if (ancestor->parent) {
 767		++level;
 768		while (ancestor->parent->parent) {
 769			++level;
 770			ancestor = ancestor->parent;
 771		}
 772	}
 773
 774	entries = amdgpu_bo_size(bo) / 8;
 775	if (!vm->pte_support_ats) {
 776		ats_entries = 0;
 777
 778	} else if (!bo->parent) {
 779		ats_entries = amdgpu_vm_num_ats_entries(adev);
 780		ats_entries = min(ats_entries, entries);
 781		entries -= ats_entries;
 782
 783	} else {
 784		struct amdgpu_vm_bo_base *pt;
 785
 786		pt = ancestor->vm_bo;
 787		ats_entries = amdgpu_vm_num_ats_entries(adev);
 788		if ((pt - to_amdgpu_bo_vm(vm->root.bo)->entries) >= ats_entries) {
 789			ats_entries = 0;
 790		} else {
 791			ats_entries = entries;
 792			entries = 0;
 793		}
 794	}
 795
 796	r = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
 797	if (r)
 798		return r;
 799
 800	if (vmbo->shadow) {
 801		struct amdgpu_bo *shadow = vmbo->shadow;
 802
 803		r = ttm_bo_validate(&shadow->tbo, &shadow->placement, &ctx);
 804		if (r)
 805			return r;
 806	}
 807
 808	r = vm->update_funcs->map_table(vmbo);
 809	if (r)
 810		return r;
 811
 812	memset(&params, 0, sizeof(params));
 813	params.adev = adev;
 814	params.vm = vm;
 815	params.immediate = immediate;
 816
 817	r = vm->update_funcs->prepare(&params, NULL, AMDGPU_SYNC_EXPLICIT);
 818	if (r)
 819		return r;
 820
 821	addr = 0;
 822	if (ats_entries) {
 823		uint64_t value = 0, flags;
 824
 825		flags = AMDGPU_PTE_DEFAULT_ATC;
 826		if (level != AMDGPU_VM_PTB) {
 827			/* Handle leaf PDEs as PTEs */
 828			flags |= AMDGPU_PDE_PTE;
 829			amdgpu_gmc_get_vm_pde(adev, level, &value, &flags);
 830		}
 831
 832		r = vm->update_funcs->update(&params, vmbo, addr, 0, ats_entries,
 833					     value, flags);
 834		if (r)
 835			return r;
 836
 837		addr += ats_entries * 8;
 838	}
 839
 840	if (entries) {
 841		uint64_t value = 0, flags = 0;
 842
 843		if (adev->asic_type >= CHIP_VEGA10) {
 844			if (level != AMDGPU_VM_PTB) {
 845				/* Handle leaf PDEs as PTEs */
 846				flags |= AMDGPU_PDE_PTE;
 847				amdgpu_gmc_get_vm_pde(adev, level,
 848						      &value, &flags);
 849			} else {
 850				/* Workaround for fault priority problem on GMC9 */
 851				flags = AMDGPU_PTE_EXECUTABLE;
 852			}
 853		}
 854
 855		r = vm->update_funcs->update(&params, vmbo, addr, 0, entries,
 856					     value, flags);
 857		if (r)
 858			return r;
 859	}
 860
 861	return vm->update_funcs->commit(&params, NULL);
 862}
 863
 864/**
 865 * amdgpu_vm_pt_create - create bo for PD/PT
 866 *
 867 * @adev: amdgpu_device pointer
 868 * @vm: requesting vm
 869 * @level: the page table level
 870 * @immediate: use a immediate update
 871 * @vmbo: pointer to the buffer object pointer
 872 */
 873static int amdgpu_vm_pt_create(struct amdgpu_device *adev,
 874			       struct amdgpu_vm *vm,
 875			       int level, bool immediate,
 876			       struct amdgpu_bo_vm **vmbo)
 877{
 878	struct amdgpu_bo_param bp;
 879	struct amdgpu_bo *bo;
 880	struct dma_resv *resv;
 881	unsigned int num_entries;
 882	int r;
 883
 884	memset(&bp, 0, sizeof(bp));
 885
 886	bp.size = amdgpu_vm_bo_size(adev, level);
 887	bp.byte_align = AMDGPU_GPU_PAGE_SIZE;
 888	bp.domain = AMDGPU_GEM_DOMAIN_VRAM;
 889	bp.domain = amdgpu_bo_get_preferred_pin_domain(adev, bp.domain);
 890	bp.flags = AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS |
 891		AMDGPU_GEM_CREATE_CPU_GTT_USWC;
 892
 893	if (level < AMDGPU_VM_PTB)
 894		num_entries = amdgpu_vm_num_entries(adev, level);
 895	else
 896		num_entries = 0;
 897
 898	bp.bo_ptr_size = struct_size((*vmbo), entries, num_entries);
 899
 900	if (vm->use_cpu_for_update)
 901		bp.flags |= AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED;
 902
 903	bp.type = ttm_bo_type_kernel;
 904	bp.no_wait_gpu = immediate;
 905	if (vm->root.bo)
 906		bp.resv = vm->root.bo->tbo.base.resv;
 907
 908	r = amdgpu_bo_create_vm(adev, &bp, vmbo);
 909	if (r)
 910		return r;
 911
 912	bo = &(*vmbo)->bo;
 913	if (vm->is_compute_context || (adev->flags & AMD_IS_APU)) {
 914		(*vmbo)->shadow = NULL;
 915		return 0;
 916	}
 917
 918	if (!bp.resv)
 919		WARN_ON(dma_resv_lock(bo->tbo.base.resv,
 920				      NULL));
 921	resv = bp.resv;
 922	memset(&bp, 0, sizeof(bp));
 923	bp.size = amdgpu_vm_bo_size(adev, level);
 924	bp.domain = AMDGPU_GEM_DOMAIN_GTT;
 925	bp.flags = AMDGPU_GEM_CREATE_CPU_GTT_USWC;
 926	bp.type = ttm_bo_type_kernel;
 927	bp.resv = bo->tbo.base.resv;
 928	bp.bo_ptr_size = sizeof(struct amdgpu_bo);
 929
 930	r = amdgpu_bo_create(adev, &bp, &(*vmbo)->shadow);
 931
 932	if (!resv)
 933		dma_resv_unlock(bo->tbo.base.resv);
 934
 935	if (r) {
 936		amdgpu_bo_unref(&bo);
 937		return r;
 938	}
 939
 940	(*vmbo)->shadow->parent = amdgpu_bo_ref(bo);
 941	amdgpu_bo_add_to_shadow_list(*vmbo);
 942
 943	return 0;
 944}
 945
 946/**
 947 * amdgpu_vm_alloc_pts - Allocate a specific page table
 948 *
 949 * @adev: amdgpu_device pointer
 950 * @vm: VM to allocate page tables for
 951 * @cursor: Which page table to allocate
 952 * @immediate: use an immediate update
 953 *
 954 * Make sure a specific page table or directory is allocated.
 955 *
 956 * Returns:
 957 * 1 if page table needed to be allocated, 0 if page table was already
 958 * allocated, negative errno if an error occurred.
 959 */
 960static int amdgpu_vm_alloc_pts(struct amdgpu_device *adev,
 961			       struct amdgpu_vm *vm,
 962			       struct amdgpu_vm_pt_cursor *cursor,
 963			       bool immediate)
 964{
 965	struct amdgpu_vm_bo_base *entry = cursor->entry;
 966	struct amdgpu_bo *pt_bo;
 967	struct amdgpu_bo_vm *pt;
 968	int r;
 969
 970	if (entry->bo)
 
 
 
 
 
 
 
 
 
 
 
 971		return 0;
 972
 973	r = amdgpu_vm_pt_create(adev, vm, cursor->level, immediate, &pt);
 
 
 974	if (r)
 975		return r;
 976
 977	/* Keep a reference to the root directory to avoid
 978	 * freeing them up in the wrong order.
 979	 */
 980	pt_bo = &pt->bo;
 981	pt_bo->parent = amdgpu_bo_ref(cursor->parent->bo);
 982	amdgpu_vm_bo_base_init(entry, vm, pt_bo);
 983	r = amdgpu_vm_clear_bo(adev, vm, pt, immediate);
 984	if (r)
 985		goto error_free_pt;
 986
 987	return 0;
 988
 989error_free_pt:
 990	amdgpu_bo_unref(&pt->shadow);
 991	amdgpu_bo_unref(&pt_bo);
 992	return r;
 993}
 994
 995/**
 996 * amdgpu_vm_free_table - fre one PD/PT
 997 *
 998 * @entry: PDE to free
 999 */
1000static void amdgpu_vm_free_table(struct amdgpu_vm_bo_base *entry)
1001{
1002	struct amdgpu_bo *shadow;
1003
1004	if (!entry->bo)
1005		return;
1006	shadow = amdgpu_bo_shadowed(entry->bo);
1007	entry->bo->vm_bo = NULL;
1008	list_del(&entry->vm_status);
1009	amdgpu_bo_unref(&shadow);
1010	amdgpu_bo_unref(&entry->bo);
1011}
1012
1013/**
1014 * amdgpu_vm_free_pts - free PD/PT levels
1015 *
1016 * @adev: amdgpu device structure
1017 * @vm: amdgpu vm structure
1018 * @start: optional cursor where to start freeing PDs/PTs
1019 *
1020 * Free the page directory or page table level and all sub levels.
1021 */
1022static void amdgpu_vm_free_pts(struct amdgpu_device *adev,
1023			       struct amdgpu_vm *vm,
1024			       struct amdgpu_vm_pt_cursor *start)
1025{
1026	struct amdgpu_vm_pt_cursor cursor;
1027	struct amdgpu_vm_bo_base *entry;
1028
1029	vm->bulk_moveable = false;
1030
1031	for_each_amdgpu_vm_pt_dfs_safe(adev, vm, start, cursor, entry)
1032		amdgpu_vm_free_table(entry);
1033
1034	if (start)
1035		amdgpu_vm_free_table(start->entry);
1036}
1037
1038/**
1039 * amdgpu_vm_check_compute_bug - check whether asic has compute vm bug
1040 *
1041 * @adev: amdgpu_device pointer
1042 */
1043void amdgpu_vm_check_compute_bug(struct amdgpu_device *adev)
1044{
1045	const struct amdgpu_ip_block *ip_block;
1046	bool has_compute_vm_bug;
1047	struct amdgpu_ring *ring;
1048	int i;
1049
1050	has_compute_vm_bug = false;
1051
1052	ip_block = amdgpu_device_ip_get_ip_block(adev, AMD_IP_BLOCK_TYPE_GFX);
1053	if (ip_block) {
1054		/* Compute has a VM bug for GFX version < 7.
1055		   Compute has a VM bug for GFX 8 MEC firmware version < 673.*/
1056		if (ip_block->version->major <= 7)
1057			has_compute_vm_bug = true;
1058		else if (ip_block->version->major == 8)
1059			if (adev->gfx.mec_fw_version < 673)
1060				has_compute_vm_bug = true;
1061	}
1062
1063	for (i = 0; i < adev->num_rings; i++) {
1064		ring = adev->rings[i];
1065		if (ring->funcs->type == AMDGPU_RING_TYPE_COMPUTE)
1066			/* only compute rings */
1067			ring->has_compute_vm_bug = has_compute_vm_bug;
1068		else
1069			ring->has_compute_vm_bug = false;
1070	}
1071}
1072
1073/**
1074 * amdgpu_vm_need_pipeline_sync - Check if pipe sync is needed for job.
1075 *
1076 * @ring: ring on which the job will be submitted
1077 * @job: job to submit
1078 *
1079 * Returns:
1080 * True if sync is needed.
1081 */
1082bool amdgpu_vm_need_pipeline_sync(struct amdgpu_ring *ring,
1083				  struct amdgpu_job *job)
1084{
1085	struct amdgpu_device *adev = ring->adev;
1086	unsigned vmhub = ring->funcs->vmhub;
1087	struct amdgpu_vmid_mgr *id_mgr = &adev->vm_manager.id_mgr[vmhub];
1088	struct amdgpu_vmid *id;
1089	bool gds_switch_needed;
1090	bool vm_flush_needed = job->vm_needs_flush || ring->has_compute_vm_bug;
1091
1092	if (job->vmid == 0)
1093		return false;
1094	id = &id_mgr->ids[job->vmid];
1095	gds_switch_needed = ring->funcs->emit_gds_switch && (
1096		id->gds_base != job->gds_base ||
1097		id->gds_size != job->gds_size ||
1098		id->gws_base != job->gws_base ||
1099		id->gws_size != job->gws_size ||
1100		id->oa_base != job->oa_base ||
1101		id->oa_size != job->oa_size);
1102
1103	if (amdgpu_vmid_had_gpu_reset(adev, id))
1104		return true;
1105
1106	return vm_flush_needed || gds_switch_needed;
1107}
1108
1109/**
1110 * amdgpu_vm_flush - hardware flush the vm
1111 *
1112 * @ring: ring to use for flush
1113 * @job:  related job
1114 * @need_pipe_sync: is pipe sync needed
1115 *
1116 * Emit a VM flush when it is necessary.
1117 *
1118 * Returns:
1119 * 0 on success, errno otherwise.
1120 */
1121int amdgpu_vm_flush(struct amdgpu_ring *ring, struct amdgpu_job *job,
1122		    bool need_pipe_sync)
1123{
1124	struct amdgpu_device *adev = ring->adev;
1125	unsigned vmhub = ring->funcs->vmhub;
1126	struct amdgpu_vmid_mgr *id_mgr = &adev->vm_manager.id_mgr[vmhub];
1127	struct amdgpu_vmid *id = &id_mgr->ids[job->vmid];
1128	bool gds_switch_needed = ring->funcs->emit_gds_switch && (
1129		id->gds_base != job->gds_base ||
1130		id->gds_size != job->gds_size ||
1131		id->gws_base != job->gws_base ||
1132		id->gws_size != job->gws_size ||
1133		id->oa_base != job->oa_base ||
1134		id->oa_size != job->oa_size);
1135	bool vm_flush_needed = job->vm_needs_flush;
 
 
 
1136	struct dma_fence *fence = NULL;
1137	bool pasid_mapping_needed = false;
1138	unsigned patch_offset = 0;
1139	bool update_spm_vmid_needed = (job->vm && (job->vm->reserved_vmid[vmhub] != NULL));
1140	int r;
1141
1142	if (update_spm_vmid_needed && adev->gfx.rlc.funcs->update_spm_vmid)
1143		adev->gfx.rlc.funcs->update_spm_vmid(adev, job->vmid);
1144
1145	if (amdgpu_vmid_had_gpu_reset(adev, id)) {
1146		gds_switch_needed = true;
1147		vm_flush_needed = true;
1148		pasid_mapping_needed = true;
1149	}
1150
1151	mutex_lock(&id_mgr->lock);
1152	if (id->pasid != job->pasid || !id->pasid_mapping ||
1153	    !dma_fence_is_signaled(id->pasid_mapping))
1154		pasid_mapping_needed = true;
1155	mutex_unlock(&id_mgr->lock);
1156
1157	gds_switch_needed &= !!ring->funcs->emit_gds_switch;
1158	vm_flush_needed &= !!ring->funcs->emit_vm_flush  &&
1159			job->vm_pd_addr != AMDGPU_BO_INVALID_OFFSET;
1160	pasid_mapping_needed &= adev->gmc.gmc_funcs->emit_pasid_mapping &&
1161		ring->funcs->emit_wreg;
1162
1163	if (!vm_flush_needed && !gds_switch_needed && !need_pipe_sync)
1164		return 0;
1165
1166	if (ring->funcs->init_cond_exec)
1167		patch_offset = amdgpu_ring_init_cond_exec(ring);
1168
1169	if (need_pipe_sync)
1170		amdgpu_ring_emit_pipeline_sync(ring);
1171
1172	if (vm_flush_needed) {
1173		trace_amdgpu_vm_flush(ring, job->vmid, job->vm_pd_addr);
1174		amdgpu_ring_emit_vm_flush(ring, job->vmid, job->vm_pd_addr);
1175	}
1176
1177	if (pasid_mapping_needed)
1178		amdgpu_gmc_emit_pasid_mapping(ring, job->vmid, job->pasid);
1179
1180	if (vm_flush_needed || pasid_mapping_needed) {
1181		r = amdgpu_fence_emit(ring, &fence, 0);
1182		if (r)
1183			return r;
1184	}
1185
1186	if (vm_flush_needed) {
1187		mutex_lock(&id_mgr->lock);
1188		dma_fence_put(id->last_flush);
1189		id->last_flush = dma_fence_get(fence);
1190		id->current_gpu_reset_count =
1191			atomic_read(&adev->gpu_reset_counter);
1192		mutex_unlock(&id_mgr->lock);
1193	}
1194
1195	if (pasid_mapping_needed) {
1196		mutex_lock(&id_mgr->lock);
1197		id->pasid = job->pasid;
1198		dma_fence_put(id->pasid_mapping);
1199		id->pasid_mapping = dma_fence_get(fence);
1200		mutex_unlock(&id_mgr->lock);
1201	}
1202	dma_fence_put(fence);
1203
1204	if (ring->funcs->emit_gds_switch && gds_switch_needed) {
1205		id->gds_base = job->gds_base;
1206		id->gds_size = job->gds_size;
1207		id->gws_base = job->gws_base;
1208		id->gws_size = job->gws_size;
1209		id->oa_base = job->oa_base;
1210		id->oa_size = job->oa_size;
1211		amdgpu_ring_emit_gds_switch(ring, job->vmid, job->gds_base,
1212					    job->gds_size, job->gws_base,
1213					    job->gws_size, job->oa_base,
1214					    job->oa_size);
1215	}
1216
1217	if (ring->funcs->patch_cond_exec)
1218		amdgpu_ring_patch_cond_exec(ring, patch_offset);
1219
1220	/* the double SWITCH_BUFFER here *cannot* be skipped by COND_EXEC */
1221	if (ring->funcs->emit_switch_buffer) {
1222		amdgpu_ring_emit_switch_buffer(ring);
1223		amdgpu_ring_emit_switch_buffer(ring);
1224	}
1225	return 0;
1226}
1227
1228/**
1229 * amdgpu_vm_bo_find - find the bo_va for a specific vm & bo
1230 *
1231 * @vm: requested vm
1232 * @bo: requested buffer object
1233 *
1234 * Find @bo inside the requested vm.
1235 * Search inside the @bos vm list for the requested vm
1236 * Returns the found bo_va or NULL if none is found
1237 *
1238 * Object has to be reserved!
1239 *
1240 * Returns:
1241 * Found bo_va or NULL.
1242 */
1243struct amdgpu_bo_va *amdgpu_vm_bo_find(struct amdgpu_vm *vm,
1244				       struct amdgpu_bo *bo)
1245{
1246	struct amdgpu_vm_bo_base *base;
1247
1248	for (base = bo->vm_bo; base; base = base->next) {
1249		if (base->vm != vm)
1250			continue;
1251
1252		return container_of(base, struct amdgpu_bo_va, base);
1253	}
1254	return NULL;
1255}
1256
1257/**
1258 * amdgpu_vm_map_gart - Resolve gart mapping of addr
1259 *
1260 * @pages_addr: optional DMA address to use for lookup
1261 * @addr: the unmapped addr
1262 *
1263 * Look up the physical address of the page that the pte resolves
1264 * to.
1265 *
1266 * Returns:
1267 * The pointer for the page table entry.
1268 */
1269uint64_t amdgpu_vm_map_gart(const dma_addr_t *pages_addr, uint64_t addr)
1270{
1271	uint64_t result;
1272
1273	/* page table offset */
1274	result = pages_addr[addr >> PAGE_SHIFT];
1275
1276	/* in case cpu page size != gpu page size*/
1277	result |= addr & (~PAGE_MASK);
1278
1279	result &= 0xFFFFFFFFFFFFF000ULL;
1280
1281	return result;
1282}
1283
1284/**
1285 * amdgpu_vm_update_pde - update a single level in the hierarchy
1286 *
1287 * @params: parameters for the update
1288 * @vm: requested vm
1289 * @entry: entry to update
1290 *
1291 * Makes sure the requested entry in parent is up to date.
1292 */
1293static int amdgpu_vm_update_pde(struct amdgpu_vm_update_params *params,
1294				struct amdgpu_vm *vm,
1295				struct amdgpu_vm_bo_base *entry)
1296{
1297	struct amdgpu_vm_bo_base *parent = amdgpu_vm_pt_parent(entry);
1298	struct amdgpu_bo *bo = parent->bo, *pbo;
1299	uint64_t pde, pt, flags;
1300	unsigned level;
1301
1302	for (level = 0, pbo = bo->parent; pbo; ++level)
1303		pbo = pbo->parent;
1304
1305	level += params->adev->vm_manager.root_level;
1306	amdgpu_gmc_get_pde_for_bo(entry->bo, level, &pt, &flags);
1307	pde = (entry - to_amdgpu_bo_vm(parent->bo)->entries) * 8;
1308	return vm->update_funcs->update(params, to_amdgpu_bo_vm(bo), pde, pt,
1309					1, 0, flags);
1310}
1311
1312/**
1313 * amdgpu_vm_invalidate_pds - mark all PDs as invalid
1314 *
1315 * @adev: amdgpu_device pointer
1316 * @vm: related vm
1317 *
1318 * Mark all PD level as invalid after an error.
1319 */
1320static void amdgpu_vm_invalidate_pds(struct amdgpu_device *adev,
1321				     struct amdgpu_vm *vm)
1322{
1323	struct amdgpu_vm_pt_cursor cursor;
1324	struct amdgpu_vm_bo_base *entry;
1325
1326	for_each_amdgpu_vm_pt_dfs_safe(adev, vm, NULL, cursor, entry)
1327		if (entry->bo && !entry->moved)
1328			amdgpu_vm_bo_relocated(entry);
1329}
1330
1331/**
1332 * amdgpu_vm_update_pdes - make sure that all directories are valid
1333 *
1334 * @adev: amdgpu_device pointer
1335 * @vm: requested vm
1336 * @immediate: submit immediately to the paging queue
1337 *
1338 * Makes sure all directories are up to date.
1339 *
1340 * Returns:
1341 * 0 for success, error for failure.
1342 */
1343int amdgpu_vm_update_pdes(struct amdgpu_device *adev,
1344			  struct amdgpu_vm *vm, bool immediate)
1345{
1346	struct amdgpu_vm_update_params params;
1347	int r;
1348
1349	if (list_empty(&vm->relocated))
1350		return 0;
1351
1352	memset(&params, 0, sizeof(params));
1353	params.adev = adev;
1354	params.vm = vm;
1355	params.immediate = immediate;
1356
1357	r = vm->update_funcs->prepare(&params, NULL, AMDGPU_SYNC_EXPLICIT);
1358	if (r)
1359		return r;
1360
1361	while (!list_empty(&vm->relocated)) {
1362		struct amdgpu_vm_bo_base *entry;
1363
1364		entry = list_first_entry(&vm->relocated,
1365					 struct amdgpu_vm_bo_base,
1366					 vm_status);
1367		amdgpu_vm_bo_idle(entry);
1368
1369		r = amdgpu_vm_update_pde(&params, vm, entry);
1370		if (r)
1371			goto error;
1372	}
1373
1374	r = vm->update_funcs->commit(&params, &vm->last_update);
1375	if (r)
1376		goto error;
1377	return 0;
1378
1379error:
1380	amdgpu_vm_invalidate_pds(adev, vm);
1381	return r;
1382}
1383
1384/*
1385 * amdgpu_vm_update_flags - figure out flags for PTE updates
1386 *
1387 * Make sure to set the right flags for the PTEs at the desired level.
1388 */
1389static void amdgpu_vm_update_flags(struct amdgpu_vm_update_params *params,
1390				   struct amdgpu_bo_vm *pt, unsigned int level,
1391				   uint64_t pe, uint64_t addr,
1392				   unsigned int count, uint32_t incr,
1393				   uint64_t flags)
1394
1395{
1396	if (level != AMDGPU_VM_PTB) {
1397		flags |= AMDGPU_PDE_PTE;
1398		amdgpu_gmc_get_vm_pde(params->adev, level, &addr, &flags);
1399
1400	} else if (params->adev->asic_type >= CHIP_VEGA10 &&
1401		   !(flags & AMDGPU_PTE_VALID) &&
1402		   !(flags & AMDGPU_PTE_PRT)) {
1403
1404		/* Workaround for fault priority problem on GMC9 */
1405		flags |= AMDGPU_PTE_EXECUTABLE;
1406	}
1407
1408	params->vm->update_funcs->update(params, pt, pe, addr, count, incr,
1409					 flags);
1410}
1411
1412/**
1413 * amdgpu_vm_fragment - get fragment for PTEs
1414 *
1415 * @params: see amdgpu_vm_update_params definition
1416 * @start: first PTE to handle
1417 * @end: last PTE to handle
1418 * @flags: hw mapping flags
1419 * @frag: resulting fragment size
1420 * @frag_end: end of this fragment
1421 *
1422 * Returns the first possible fragment for the start and end address.
1423 */
1424static void amdgpu_vm_fragment(struct amdgpu_vm_update_params *params,
1425			       uint64_t start, uint64_t end, uint64_t flags,
1426			       unsigned int *frag, uint64_t *frag_end)
1427{
1428	/**
1429	 * The MC L1 TLB supports variable sized pages, based on a fragment
1430	 * field in the PTE. When this field is set to a non-zero value, page
1431	 * granularity is increased from 4KB to (1 << (12 + frag)). The PTE
1432	 * flags are considered valid for all PTEs within the fragment range
1433	 * and corresponding mappings are assumed to be physically contiguous.
1434	 *
1435	 * The L1 TLB can store a single PTE for the whole fragment,
1436	 * significantly increasing the space available for translation
1437	 * caching. This leads to large improvements in throughput when the
1438	 * TLB is under pressure.
1439	 *
1440	 * The L2 TLB distributes small and large fragments into two
1441	 * asymmetric partitions. The large fragment cache is significantly
1442	 * larger. Thus, we try to use large fragments wherever possible.
1443	 * Userspace can support this by aligning virtual base address and
1444	 * allocation size to the fragment size.
1445	 *
1446	 * Starting with Vega10 the fragment size only controls the L1. The L2
1447	 * is now directly feed with small/huge/giant pages from the walker.
1448	 */
1449	unsigned max_frag;
1450
1451	if (params->adev->asic_type < CHIP_VEGA10)
1452		max_frag = params->adev->vm_manager.fragment_size;
1453	else
1454		max_frag = 31;
1455
1456	/* system pages are non continuously */
1457	if (params->pages_addr) {
1458		*frag = 0;
1459		*frag_end = end;
1460		return;
1461	}
1462
1463	/* This intentionally wraps around if no bit is set */
1464	*frag = min((unsigned)ffs(start) - 1, (unsigned)fls64(end - start) - 1);
1465	if (*frag >= max_frag) {
1466		*frag = max_frag;
1467		*frag_end = end & ~((1ULL << max_frag) - 1);
1468	} else {
1469		*frag_end = start + (1 << *frag);
1470	}
1471}
1472
1473/**
1474 * amdgpu_vm_update_ptes - make sure that page tables are valid
1475 *
1476 * @params: see amdgpu_vm_update_params definition
1477 * @start: start of GPU address range
1478 * @end: end of GPU address range
1479 * @dst: destination address to map to, the next dst inside the function
1480 * @flags: mapping flags
1481 *
1482 * Update the page tables in the range @start - @end.
1483 *
1484 * Returns:
1485 * 0 for success, -EINVAL for failure.
1486 */
1487static int amdgpu_vm_update_ptes(struct amdgpu_vm_update_params *params,
1488				 uint64_t start, uint64_t end,
1489				 uint64_t dst, uint64_t flags)
1490{
1491	struct amdgpu_device *adev = params->adev;
1492	struct amdgpu_vm_pt_cursor cursor;
1493	uint64_t frag_start = start, frag_end;
1494	unsigned int frag;
1495	int r;
1496
1497	/* figure out the initial fragment */
1498	amdgpu_vm_fragment(params, frag_start, end, flags, &frag, &frag_end);
1499
1500	/* walk over the address space and update the PTs */
1501	amdgpu_vm_pt_start(adev, params->vm, start, &cursor);
1502	while (cursor.pfn < end) {
1503		unsigned shift, parent_shift, mask;
1504		uint64_t incr, entry_end, pe_start;
1505		struct amdgpu_bo *pt;
1506
1507		if (!params->unlocked) {
1508			/* make sure that the page tables covering the
1509			 * address range are actually allocated
1510			 */
1511			r = amdgpu_vm_alloc_pts(params->adev, params->vm,
1512						&cursor, params->immediate);
1513			if (r)
1514				return r;
 
 
 
1515		}
1516
1517		shift = amdgpu_vm_level_shift(adev, cursor.level);
1518		parent_shift = amdgpu_vm_level_shift(adev, cursor.level - 1);
1519		if (params->unlocked) {
1520			/* Unlocked updates are only allowed on the leaves */
1521			if (amdgpu_vm_pt_descendant(adev, &cursor))
1522				continue;
1523		} else if (adev->asic_type < CHIP_VEGA10 &&
1524			   (flags & AMDGPU_PTE_VALID)) {
1525			/* No huge page support before GMC v9 */
1526			if (cursor.level != AMDGPU_VM_PTB) {
1527				if (!amdgpu_vm_pt_descendant(adev, &cursor))
1528					return -ENOENT;
1529				continue;
1530			}
1531		} else if (frag < shift) {
1532			/* We can't use this level when the fragment size is
1533			 * smaller than the address shift. Go to the next
1534			 * child entry and try again.
1535			 */
1536			if (amdgpu_vm_pt_descendant(adev, &cursor))
1537				continue;
1538		} else if (frag >= parent_shift) {
 
 
1539			/* If the fragment size is even larger than the parent
1540			 * shift we should go up one level and check it again.
 
1541			 */
1542			if (!amdgpu_vm_pt_ancestor(&cursor))
1543				return -EINVAL;
1544			continue;
1545		}
1546
1547		pt = cursor.entry->bo;
1548		if (!pt) {
1549			/* We need all PDs and PTs for mapping something, */
1550			if (flags & AMDGPU_PTE_VALID)
1551				return -ENOENT;
1552
1553			/* but unmapping something can happen at a higher
1554			 * level.
1555			 */
1556			if (!amdgpu_vm_pt_ancestor(&cursor))
1557				return -EINVAL;
1558
1559			pt = cursor.entry->bo;
1560			shift = parent_shift;
1561			frag_end = max(frag_end, ALIGN(frag_start + 1,
1562				   1ULL << shift));
1563		}
1564
1565		/* Looks good so far, calculate parameters for the update */
1566		incr = (uint64_t)AMDGPU_GPU_PAGE_SIZE << shift;
1567		mask = amdgpu_vm_entries_mask(adev, cursor.level);
1568		pe_start = ((cursor.pfn >> shift) & mask) * 8;
1569		entry_end = ((uint64_t)mask + 1) << shift;
1570		entry_end += cursor.pfn & ~(entry_end - 1);
1571		entry_end = min(entry_end, end);
1572
1573		do {
1574			struct amdgpu_vm *vm = params->vm;
1575			uint64_t upd_end = min(entry_end, frag_end);
1576			unsigned nptes = (upd_end - frag_start) >> shift;
1577			uint64_t upd_flags = flags | AMDGPU_PTE_FRAG(frag);
1578
1579			/* This can happen when we set higher level PDs to
1580			 * silent to stop fault floods.
1581			 */
1582			nptes = max(nptes, 1u);
1583
1584			trace_amdgpu_vm_update_ptes(params, frag_start, upd_end,
1585						    nptes, dst, incr, upd_flags,
1586						    vm->task_info.pid,
1587						    vm->immediate.fence_context);
1588			amdgpu_vm_update_flags(params, to_amdgpu_bo_vm(pt),
1589					       cursor.level, pe_start, dst,
1590					       nptes, incr, upd_flags);
1591
1592			pe_start += nptes * 8;
1593			dst += nptes * incr;
1594
1595			frag_start = upd_end;
1596			if (frag_start >= frag_end) {
1597				/* figure out the next fragment */
1598				amdgpu_vm_fragment(params, frag_start, end,
1599						   flags, &frag, &frag_end);
1600				if (frag < shift)
1601					break;
1602			}
1603		} while (frag_start < entry_end);
1604
1605		if (amdgpu_vm_pt_descendant(adev, &cursor)) {
1606			/* Free all child entries.
1607			 * Update the tables with the flags and addresses and free up subsequent
1608			 * tables in the case of huge pages or freed up areas.
1609			 * This is the maximum you can free, because all other page tables are not
1610			 * completely covered by the range and so potentially still in use.
1611			 */
1612			while (cursor.pfn < frag_start) {
1613				/* Make sure previous mapping is freed */
1614				if (cursor.entry->bo) {
1615					params->table_freed = true;
1616					amdgpu_vm_free_pts(adev, params->vm, &cursor);
1617				}
1618				amdgpu_vm_pt_next(adev, &cursor);
1619			}
1620
1621		} else if (frag >= shift) {
1622			/* or just move on to the next on the same level. */
1623			amdgpu_vm_pt_next(adev, &cursor);
1624		}
1625	}
1626
1627	return 0;
1628}
1629
1630/**
1631 * amdgpu_vm_bo_update_mapping - update a mapping in the vm page table
1632 *
1633 * @adev: amdgpu_device pointer of the VM
1634 * @bo_adev: amdgpu_device pointer of the mapped BO
 
1635 * @vm: requested vm
1636 * @immediate: immediate submission in a page fault
1637 * @unlocked: unlocked invalidation during MM callback
1638 * @resv: fences we need to sync to
1639 * @start: start of mapped range
1640 * @last: last mapped entry
1641 * @flags: flags for the entries
1642 * @offset: offset into nodes and pages_addr
1643 * @res: ttm_resource to map
1644 * @pages_addr: DMA addresses to use for mapping
1645 * @fence: optional resulting fence
1646 * @table_freed: return true if page table is freed
1647 *
1648 * Fill in the page table entries between @start and @last.
1649 *
1650 * Returns:
1651 * 0 for success, -EINVAL for failure.
1652 */
1653int amdgpu_vm_bo_update_mapping(struct amdgpu_device *adev,
1654				struct amdgpu_device *bo_adev,
1655				struct amdgpu_vm *vm, bool immediate,
1656				bool unlocked, struct dma_resv *resv,
1657				uint64_t start, uint64_t last,
1658				uint64_t flags, uint64_t offset,
1659				struct ttm_resource *res,
1660				dma_addr_t *pages_addr,
1661				struct dma_fence **fence,
1662				bool *table_freed)
1663{
1664	struct amdgpu_vm_update_params params;
1665	struct amdgpu_res_cursor cursor;
1666	enum amdgpu_sync_mode sync_mode;
1667	int r, idx;
1668
1669	if (!drm_dev_enter(&adev->ddev, &idx))
1670		return -ENODEV;
1671
1672	memset(&params, 0, sizeof(params));
1673	params.adev = adev;
1674	params.vm = vm;
1675	params.immediate = immediate;
1676	params.pages_addr = pages_addr;
1677	params.unlocked = unlocked;
1678
1679	/* Implicitly sync to command submissions in the same VM before
1680	 * unmapping. Sync to moving fences before mapping.
1681	 */
1682	if (!(flags & AMDGPU_PTE_VALID))
1683		sync_mode = AMDGPU_SYNC_EQ_OWNER;
1684	else
1685		sync_mode = AMDGPU_SYNC_EXPLICIT;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1686
1687	amdgpu_vm_eviction_lock(vm);
1688	if (vm->evicting) {
1689		r = -EBUSY;
1690		goto error_unlock;
 
 
 
 
 
1691	}
1692
1693	if (!unlocked && !dma_fence_is_signaled(vm->last_unlocked)) {
1694		struct dma_fence *tmp = dma_fence_get_stub();
1695
1696		amdgpu_bo_fence(vm->root.bo, vm->last_unlocked, true);
1697		swap(vm->last_unlocked, tmp);
1698		dma_fence_put(tmp);
 
 
 
1699	}
1700
1701	r = vm->update_funcs->prepare(&params, resv, sync_mode);
1702	if (r)
1703		goto error_unlock;
 
1704
1705	amdgpu_res_first(pages_addr ? NULL : res, offset,
1706			 (last - start + 1) * AMDGPU_GPU_PAGE_SIZE, &cursor);
1707	while (cursor.remaining) {
1708		uint64_t tmp, num_entries, addr;
 
 
 
 
1709
1710		num_entries = cursor.size >> AMDGPU_GPU_PAGE_SHIFT;
1711		if (pages_addr) {
1712			bool contiguous = true;
 
 
 
 
 
1713
1714			if (num_entries > AMDGPU_GPU_PAGES_IN_CPU_PAGE) {
1715				uint64_t pfn = cursor.start >> PAGE_SHIFT;
1716				uint64_t count;
1717
1718				contiguous = pages_addr[pfn + 1] ==
1719					pages_addr[pfn] + PAGE_SIZE;
1720
1721				tmp = num_entries /
1722					AMDGPU_GPU_PAGES_IN_CPU_PAGE;
1723				for (count = 2; count < tmp; ++count) {
1724					uint64_t idx = pfn + count;
1725
1726					if (contiguous != (pages_addr[idx] ==
1727					    pages_addr[idx - 1] + PAGE_SIZE))
1728						break;
1729				}
1730				num_entries = count *
1731					AMDGPU_GPU_PAGES_IN_CPU_PAGE;
1732			}
1733
1734			if (!contiguous) {
1735				addr = cursor.start;
1736				params.pages_addr = pages_addr;
1737			} else {
1738				addr = pages_addr[cursor.start >> PAGE_SHIFT];
1739				params.pages_addr = NULL;
1740			}
1741
1742		} else if (flags & (AMDGPU_PTE_VALID | AMDGPU_PTE_PRT)) {
1743			addr = bo_adev->vm_manager.vram_base_offset +
1744				cursor.start;
1745		} else {
1746			addr = 0;
1747		}
1748
1749		tmp = start + num_entries;
1750		r = amdgpu_vm_update_ptes(&params, start, tmp, addr, flags);
 
 
1751		if (r)
1752			goto error_unlock;
1753
1754		amdgpu_res_next(&cursor, num_entries * AMDGPU_GPU_PAGE_SIZE);
1755		start = tmp;
1756	}
1757
1758	r = vm->update_funcs->commit(&params, fence);
 
 
 
 
 
1759
1760	if (table_freed)
1761		*table_freed = params.table_freed;
1762
1763error_unlock:
1764	amdgpu_vm_eviction_unlock(vm);
1765	drm_dev_exit(idx);
1766	return r;
1767}
1768
1769void amdgpu_vm_get_memory(struct amdgpu_vm *vm, uint64_t *vram_mem,
1770				uint64_t *gtt_mem, uint64_t *cpu_mem)
1771{
1772	struct amdgpu_bo_va *bo_va, *tmp;
1773
1774	list_for_each_entry_safe(bo_va, tmp, &vm->idle, base.vm_status) {
1775		if (!bo_va->base.bo)
1776			continue;
1777		amdgpu_bo_get_memory(bo_va->base.bo, vram_mem,
1778				gtt_mem, cpu_mem);
1779	}
1780	list_for_each_entry_safe(bo_va, tmp, &vm->evicted, base.vm_status) {
1781		if (!bo_va->base.bo)
1782			continue;
1783		amdgpu_bo_get_memory(bo_va->base.bo, vram_mem,
1784				gtt_mem, cpu_mem);
1785	}
1786	list_for_each_entry_safe(bo_va, tmp, &vm->relocated, base.vm_status) {
1787		if (!bo_va->base.bo)
1788			continue;
1789		amdgpu_bo_get_memory(bo_va->base.bo, vram_mem,
1790				gtt_mem, cpu_mem);
1791	}
1792	list_for_each_entry_safe(bo_va, tmp, &vm->moved, base.vm_status) {
1793		if (!bo_va->base.bo)
1794			continue;
1795		amdgpu_bo_get_memory(bo_va->base.bo, vram_mem,
1796				gtt_mem, cpu_mem);
1797	}
1798	spin_lock(&vm->invalidated_lock);
1799	list_for_each_entry_safe(bo_va, tmp, &vm->invalidated, base.vm_status) {
1800		if (!bo_va->base.bo)
1801			continue;
1802		amdgpu_bo_get_memory(bo_va->base.bo, vram_mem,
1803				gtt_mem, cpu_mem);
1804	}
1805	list_for_each_entry_safe(bo_va, tmp, &vm->done, base.vm_status) {
1806		if (!bo_va->base.bo)
1807			continue;
1808		amdgpu_bo_get_memory(bo_va->base.bo, vram_mem,
1809				gtt_mem, cpu_mem);
1810	}
1811	spin_unlock(&vm->invalidated_lock);
1812}
1813/**
1814 * amdgpu_vm_bo_update - update all BO mappings in the vm page table
1815 *
1816 * @adev: amdgpu_device pointer
1817 * @bo_va: requested BO and VM object
1818 * @clear: if true clear the entries
1819 *
1820 * Fill in the page table entries for @bo_va.
1821 *
1822 * Returns:
1823 * 0 for success, -EINVAL for failure.
1824 */
1825int amdgpu_vm_bo_update(struct amdgpu_device *adev, struct amdgpu_bo_va *bo_va,
 
1826			bool clear)
1827{
1828	struct amdgpu_bo *bo = bo_va->base.bo;
1829	struct amdgpu_vm *vm = bo_va->base.vm;
1830	struct amdgpu_bo_va_mapping *mapping;
1831	dma_addr_t *pages_addr = NULL;
1832	struct ttm_resource *mem;
1833	struct dma_fence **last_update;
1834	struct dma_resv *resv;
1835	uint64_t flags;
1836	struct amdgpu_device *bo_adev = adev;
1837	int r;
1838
1839	if (clear || !bo) {
1840		mem = NULL;
1841		resv = vm->root.bo->tbo.base.resv;
 
1842	} else {
1843		struct drm_gem_object *obj = &bo->tbo.base;
1844
1845		resv = bo->tbo.base.resv;
1846		if (obj->import_attach && bo_va->is_xgmi) {
1847			struct dma_buf *dma_buf = obj->import_attach->dmabuf;
1848			struct drm_gem_object *gobj = dma_buf->priv;
1849			struct amdgpu_bo *abo = gem_to_amdgpu_bo(gobj);
1850
1851			if (abo->tbo.resource->mem_type == TTM_PL_VRAM)
1852				bo = gem_to_amdgpu_bo(gobj);
1853		}
1854		mem = bo->tbo.resource;
1855		if (mem->mem_type == TTM_PL_TT ||
1856		    mem->mem_type == AMDGPU_PL_PREEMPT)
1857			pages_addr = bo->tbo.ttm->dma_address;
1858	}
1859
1860	if (bo) {
1861		flags = amdgpu_ttm_tt_pte_flags(adev, bo->tbo.ttm, mem);
1862
1863		if (amdgpu_bo_encrypted(bo))
1864			flags |= AMDGPU_PTE_TMZ;
1865
1866		bo_adev = amdgpu_ttm_adev(bo->tbo.bdev);
1867	} else {
1868		flags = 0x0;
1869	}
1870
1871	if (clear || (bo && bo->tbo.base.resv ==
1872		      vm->root.bo->tbo.base.resv))
1873		last_update = &vm->last_update;
1874	else
1875		last_update = &bo_va->last_pt_update;
1876
1877	if (!clear && bo_va->base.moved) {
1878		bo_va->base.moved = false;
1879		list_splice_init(&bo_va->valids, &bo_va->invalids);
1880
1881	} else if (bo_va->cleared != clear) {
1882		list_splice_init(&bo_va->valids, &bo_va->invalids);
1883	}
1884
1885	list_for_each_entry(mapping, &bo_va->invalids, list) {
1886		uint64_t update_flags = flags;
1887
1888		/* normally,bo_va->flags only contians READABLE and WIRTEABLE bit go here
1889		 * but in case of something, we filter the flags in first place
1890		 */
1891		if (!(mapping->flags & AMDGPU_PTE_READABLE))
1892			update_flags &= ~AMDGPU_PTE_READABLE;
1893		if (!(mapping->flags & AMDGPU_PTE_WRITEABLE))
1894			update_flags &= ~AMDGPU_PTE_WRITEABLE;
1895
1896		/* Apply ASIC specific mapping flags */
1897		amdgpu_gmc_get_vm_pte(adev, mapping, &update_flags);
1898
1899		trace_amdgpu_vm_bo_update(mapping);
1900
1901		r = amdgpu_vm_bo_update_mapping(adev, bo_adev, vm, false, false,
1902						resv, mapping->start,
1903						mapping->last, update_flags,
1904						mapping->offset, mem,
1905						pages_addr, last_update, NULL);
1906		if (r)
1907			return r;
1908	}
1909
 
 
 
 
 
 
1910	/* If the BO is not in its preferred location add it back to
1911	 * the evicted list so that it gets validated again on the
1912	 * next command submission.
1913	 */
1914	if (bo && bo->tbo.base.resv == vm->root.bo->tbo.base.resv) {
1915		uint32_t mem_type = bo->tbo.resource->mem_type;
1916
1917		if (!(bo->preferred_domains &
1918		      amdgpu_mem_type_to_domain(mem_type)))
1919			amdgpu_vm_bo_evicted(&bo_va->base);
1920		else
1921			amdgpu_vm_bo_idle(&bo_va->base);
1922	} else {
1923		amdgpu_vm_bo_done(&bo_va->base);
1924	}
1925
1926	list_splice_init(&bo_va->invalids, &bo_va->valids);
1927	bo_va->cleared = clear;
1928
1929	if (trace_amdgpu_vm_bo_mapping_enabled()) {
1930		list_for_each_entry(mapping, &bo_va->valids, list)
1931			trace_amdgpu_vm_bo_mapping(mapping);
1932	}
1933
1934	return 0;
1935}
1936
1937/**
1938 * amdgpu_vm_update_prt_state - update the global PRT state
1939 *
1940 * @adev: amdgpu_device pointer
1941 */
1942static void amdgpu_vm_update_prt_state(struct amdgpu_device *adev)
1943{
1944	unsigned long flags;
1945	bool enable;
1946
1947	spin_lock_irqsave(&adev->vm_manager.prt_lock, flags);
1948	enable = !!atomic_read(&adev->vm_manager.num_prt_users);
1949	adev->gmc.gmc_funcs->set_prt(adev, enable);
1950	spin_unlock_irqrestore(&adev->vm_manager.prt_lock, flags);
1951}
1952
1953/**
1954 * amdgpu_vm_prt_get - add a PRT user
1955 *
1956 * @adev: amdgpu_device pointer
1957 */
1958static void amdgpu_vm_prt_get(struct amdgpu_device *adev)
1959{
1960	if (!adev->gmc.gmc_funcs->set_prt)
1961		return;
1962
1963	if (atomic_inc_return(&adev->vm_manager.num_prt_users) == 1)
1964		amdgpu_vm_update_prt_state(adev);
1965}
1966
1967/**
1968 * amdgpu_vm_prt_put - drop a PRT user
1969 *
1970 * @adev: amdgpu_device pointer
1971 */
1972static void amdgpu_vm_prt_put(struct amdgpu_device *adev)
1973{
1974	if (atomic_dec_return(&adev->vm_manager.num_prt_users) == 0)
1975		amdgpu_vm_update_prt_state(adev);
1976}
1977
1978/**
1979 * amdgpu_vm_prt_cb - callback for updating the PRT status
1980 *
1981 * @fence: fence for the callback
1982 * @_cb: the callback function
1983 */
1984static void amdgpu_vm_prt_cb(struct dma_fence *fence, struct dma_fence_cb *_cb)
1985{
1986	struct amdgpu_prt_cb *cb = container_of(_cb, struct amdgpu_prt_cb, cb);
1987
1988	amdgpu_vm_prt_put(cb->adev);
1989	kfree(cb);
1990}
1991
1992/**
1993 * amdgpu_vm_add_prt_cb - add callback for updating the PRT status
1994 *
1995 * @adev: amdgpu_device pointer
1996 * @fence: fence for the callback
1997 */
1998static void amdgpu_vm_add_prt_cb(struct amdgpu_device *adev,
1999				 struct dma_fence *fence)
2000{
2001	struct amdgpu_prt_cb *cb;
2002
2003	if (!adev->gmc.gmc_funcs->set_prt)
2004		return;
2005
2006	cb = kmalloc(sizeof(struct amdgpu_prt_cb), GFP_KERNEL);
2007	if (!cb) {
2008		/* Last resort when we are OOM */
2009		if (fence)
2010			dma_fence_wait(fence, false);
2011
2012		amdgpu_vm_prt_put(adev);
2013	} else {
2014		cb->adev = adev;
2015		if (!fence || dma_fence_add_callback(fence, &cb->cb,
2016						     amdgpu_vm_prt_cb))
2017			amdgpu_vm_prt_cb(fence, &cb->cb);
2018	}
2019}
2020
2021/**
2022 * amdgpu_vm_free_mapping - free a mapping
2023 *
2024 * @adev: amdgpu_device pointer
2025 * @vm: requested vm
2026 * @mapping: mapping to be freed
2027 * @fence: fence of the unmap operation
2028 *
2029 * Free a mapping and make sure we decrease the PRT usage count if applicable.
2030 */
2031static void amdgpu_vm_free_mapping(struct amdgpu_device *adev,
2032				   struct amdgpu_vm *vm,
2033				   struct amdgpu_bo_va_mapping *mapping,
2034				   struct dma_fence *fence)
2035{
2036	if (mapping->flags & AMDGPU_PTE_PRT)
2037		amdgpu_vm_add_prt_cb(adev, fence);
2038	kfree(mapping);
2039}
2040
2041/**
2042 * amdgpu_vm_prt_fini - finish all prt mappings
2043 *
2044 * @adev: amdgpu_device pointer
2045 * @vm: requested vm
2046 *
2047 * Register a cleanup callback to disable PRT support after VM dies.
2048 */
2049static void amdgpu_vm_prt_fini(struct amdgpu_device *adev, struct amdgpu_vm *vm)
2050{
2051	struct dma_resv *resv = vm->root.bo->tbo.base.resv;
2052	struct dma_fence *excl, **shared;
2053	unsigned i, shared_count;
2054	int r;
2055
2056	r = dma_resv_get_fences(resv, &excl, &shared_count, &shared);
 
2057	if (r) {
2058		/* Not enough memory to grab the fence list, as last resort
2059		 * block for all the fences to complete.
2060		 */
2061		dma_resv_wait_timeout(resv, true, false,
2062						    MAX_SCHEDULE_TIMEOUT);
2063		return;
2064	}
2065
2066	/* Add a callback for each fence in the reservation object */
2067	amdgpu_vm_prt_get(adev);
2068	amdgpu_vm_add_prt_cb(adev, excl);
2069
2070	for (i = 0; i < shared_count; ++i) {
2071		amdgpu_vm_prt_get(adev);
2072		amdgpu_vm_add_prt_cb(adev, shared[i]);
2073	}
2074
2075	kfree(shared);
2076}
2077
2078/**
2079 * amdgpu_vm_clear_freed - clear freed BOs in the PT
2080 *
2081 * @adev: amdgpu_device pointer
2082 * @vm: requested vm
2083 * @fence: optional resulting fence (unchanged if no work needed to be done
2084 * or if an error occurred)
2085 *
2086 * Make sure all freed BOs are cleared in the PT.
2087 * PTs have to be reserved and mutex must be locked!
2088 *
2089 * Returns:
2090 * 0 for success.
2091 *
2092 */
2093int amdgpu_vm_clear_freed(struct amdgpu_device *adev,
2094			  struct amdgpu_vm *vm,
2095			  struct dma_fence **fence)
2096{
2097	struct dma_resv *resv = vm->root.bo->tbo.base.resv;
2098	struct amdgpu_bo_va_mapping *mapping;
2099	uint64_t init_pte_value = 0;
2100	struct dma_fence *f = NULL;
2101	int r;
2102
2103	while (!list_empty(&vm->freed)) {
2104		mapping = list_first_entry(&vm->freed,
2105			struct amdgpu_bo_va_mapping, list);
2106		list_del(&mapping->list);
2107
2108		if (vm->pte_support_ats &&
2109		    mapping->start < AMDGPU_GMC_HOLE_START)
2110			init_pte_value = AMDGPU_PTE_DEFAULT_ATC;
2111
2112		r = amdgpu_vm_bo_update_mapping(adev, adev, vm, false, false,
2113						resv, mapping->start,
2114						mapping->last, init_pte_value,
2115						0, NULL, NULL, &f, NULL);
2116		amdgpu_vm_free_mapping(adev, vm, mapping, f);
2117		if (r) {
2118			dma_fence_put(f);
2119			return r;
2120		}
2121	}
2122
2123	if (fence && f) {
2124		dma_fence_put(*fence);
2125		*fence = f;
2126	} else {
2127		dma_fence_put(f);
2128	}
2129
2130	return 0;
2131
2132}
2133
2134/**
2135 * amdgpu_vm_handle_moved - handle moved BOs in the PT
2136 *
2137 * @adev: amdgpu_device pointer
2138 * @vm: requested vm
2139 *
2140 * Make sure all BOs which are moved are updated in the PTs.
2141 *
2142 * Returns:
2143 * 0 for success.
2144 *
2145 * PTs have to be reserved!
2146 */
2147int amdgpu_vm_handle_moved(struct amdgpu_device *adev,
2148			   struct amdgpu_vm *vm)
2149{
2150	struct amdgpu_bo_va *bo_va, *tmp;
2151	struct dma_resv *resv;
2152	bool clear;
2153	int r;
2154
2155	list_for_each_entry_safe(bo_va, tmp, &vm->moved, base.vm_status) {
2156		/* Per VM BOs never need to bo cleared in the page tables */
2157		r = amdgpu_vm_bo_update(adev, bo_va, false);
2158		if (r)
2159			return r;
2160	}
2161
2162	spin_lock(&vm->invalidated_lock);
2163	while (!list_empty(&vm->invalidated)) {
2164		bo_va = list_first_entry(&vm->invalidated, struct amdgpu_bo_va,
2165					 base.vm_status);
2166		resv = bo_va->base.bo->tbo.base.resv;
2167		spin_unlock(&vm->invalidated_lock);
2168
2169		/* Try to reserve the BO to avoid clearing its ptes */
2170		if (!amdgpu_vm_debug && dma_resv_trylock(resv))
2171			clear = false;
2172		/* Somebody else is using the BO right now */
2173		else
2174			clear = true;
2175
2176		r = amdgpu_vm_bo_update(adev, bo_va, clear);
2177		if (r)
2178			return r;
2179
2180		if (!clear)
2181			dma_resv_unlock(resv);
2182		spin_lock(&vm->invalidated_lock);
2183	}
2184	spin_unlock(&vm->invalidated_lock);
2185
2186	return 0;
2187}
2188
2189/**
2190 * amdgpu_vm_bo_add - add a bo to a specific vm
2191 *
2192 * @adev: amdgpu_device pointer
2193 * @vm: requested vm
2194 * @bo: amdgpu buffer object
2195 *
2196 * Add @bo into the requested vm.
2197 * Add @bo to the list of bos associated with the vm
2198 *
2199 * Returns:
2200 * Newly added bo_va or NULL for failure
2201 *
2202 * Object has to be reserved!
2203 */
2204struct amdgpu_bo_va *amdgpu_vm_bo_add(struct amdgpu_device *adev,
2205				      struct amdgpu_vm *vm,
2206				      struct amdgpu_bo *bo)
2207{
2208	struct amdgpu_bo_va *bo_va;
2209
2210	bo_va = kzalloc(sizeof(struct amdgpu_bo_va), GFP_KERNEL);
2211	if (bo_va == NULL) {
2212		return NULL;
2213	}
2214	amdgpu_vm_bo_base_init(&bo_va->base, vm, bo);
2215
2216	bo_va->ref_count = 1;
2217	INIT_LIST_HEAD(&bo_va->valids);
2218	INIT_LIST_HEAD(&bo_va->invalids);
2219
2220	if (!bo)
2221		return bo_va;
2222
2223	if (amdgpu_dmabuf_is_xgmi_accessible(adev, bo)) {
2224		bo_va->is_xgmi = true;
 
2225		/* Power up XGMI if it can be potentially used */
2226		amdgpu_xgmi_set_pstate(adev, AMDGPU_XGMI_PSTATE_MAX_VEGA20);
 
 
2227	}
2228
2229	return bo_va;
2230}
2231
2232
2233/**
2234 * amdgpu_vm_bo_insert_map - insert a new mapping
2235 *
2236 * @adev: amdgpu_device pointer
2237 * @bo_va: bo_va to store the address
2238 * @mapping: the mapping to insert
2239 *
2240 * Insert a new mapping into all structures.
2241 */
2242static void amdgpu_vm_bo_insert_map(struct amdgpu_device *adev,
2243				    struct amdgpu_bo_va *bo_va,
2244				    struct amdgpu_bo_va_mapping *mapping)
2245{
2246	struct amdgpu_vm *vm = bo_va->base.vm;
2247	struct amdgpu_bo *bo = bo_va->base.bo;
2248
2249	mapping->bo_va = bo_va;
2250	list_add(&mapping->list, &bo_va->invalids);
2251	amdgpu_vm_it_insert(mapping, &vm->va);
2252
2253	if (mapping->flags & AMDGPU_PTE_PRT)
2254		amdgpu_vm_prt_get(adev);
2255
2256	if (bo && bo->tbo.base.resv == vm->root.bo->tbo.base.resv &&
2257	    !bo_va->base.moved) {
2258		list_move(&bo_va->base.vm_status, &vm->moved);
2259	}
2260	trace_amdgpu_vm_bo_map(bo_va, mapping);
2261}
2262
2263/**
2264 * amdgpu_vm_bo_map - map bo inside a vm
2265 *
2266 * @adev: amdgpu_device pointer
2267 * @bo_va: bo_va to store the address
2268 * @saddr: where to map the BO
2269 * @offset: requested offset in the BO
2270 * @size: BO size in bytes
2271 * @flags: attributes of pages (read/write/valid/etc.)
2272 *
2273 * Add a mapping of the BO at the specefied addr into the VM.
2274 *
2275 * Returns:
2276 * 0 for success, error for failure.
2277 *
2278 * Object has to be reserved and unreserved outside!
2279 */
2280int amdgpu_vm_bo_map(struct amdgpu_device *adev,
2281		     struct amdgpu_bo_va *bo_va,
2282		     uint64_t saddr, uint64_t offset,
2283		     uint64_t size, uint64_t flags)
2284{
2285	struct amdgpu_bo_va_mapping *mapping, *tmp;
2286	struct amdgpu_bo *bo = bo_va->base.bo;
2287	struct amdgpu_vm *vm = bo_va->base.vm;
2288	uint64_t eaddr;
2289
2290	/* validate the parameters */
2291	if (saddr & ~PAGE_MASK || offset & ~PAGE_MASK ||
2292	    size == 0 || size & ~PAGE_MASK)
2293		return -EINVAL;
2294
2295	/* make sure object fit at this offset */
2296	eaddr = saddr + size - 1;
2297	if (saddr >= eaddr ||
2298	    (bo && offset + size > amdgpu_bo_size(bo)) ||
2299	    (eaddr >= adev->vm_manager.max_pfn << AMDGPU_GPU_PAGE_SHIFT))
2300		return -EINVAL;
2301
2302	saddr /= AMDGPU_GPU_PAGE_SIZE;
2303	eaddr /= AMDGPU_GPU_PAGE_SIZE;
2304
2305	tmp = amdgpu_vm_it_iter_first(&vm->va, saddr, eaddr);
2306	if (tmp) {
2307		/* bo and tmp overlap, invalid addr */
2308		dev_err(adev->dev, "bo %p va 0x%010Lx-0x%010Lx conflict with "
2309			"0x%010Lx-0x%010Lx\n", bo, saddr, eaddr,
2310			tmp->start, tmp->last + 1);
2311		return -EINVAL;
2312	}
2313
2314	mapping = kmalloc(sizeof(*mapping), GFP_KERNEL);
2315	if (!mapping)
2316		return -ENOMEM;
2317
2318	mapping->start = saddr;
2319	mapping->last = eaddr;
2320	mapping->offset = offset;
2321	mapping->flags = flags;
2322
2323	amdgpu_vm_bo_insert_map(adev, bo_va, mapping);
2324
2325	return 0;
2326}
2327
2328/**
2329 * amdgpu_vm_bo_replace_map - map bo inside a vm, replacing existing mappings
2330 *
2331 * @adev: amdgpu_device pointer
2332 * @bo_va: bo_va to store the address
2333 * @saddr: where to map the BO
2334 * @offset: requested offset in the BO
2335 * @size: BO size in bytes
2336 * @flags: attributes of pages (read/write/valid/etc.)
2337 *
2338 * Add a mapping of the BO at the specefied addr into the VM. Replace existing
2339 * mappings as we do so.
2340 *
2341 * Returns:
2342 * 0 for success, error for failure.
2343 *
2344 * Object has to be reserved and unreserved outside!
2345 */
2346int amdgpu_vm_bo_replace_map(struct amdgpu_device *adev,
2347			     struct amdgpu_bo_va *bo_va,
2348			     uint64_t saddr, uint64_t offset,
2349			     uint64_t size, uint64_t flags)
2350{
2351	struct amdgpu_bo_va_mapping *mapping;
2352	struct amdgpu_bo *bo = bo_va->base.bo;
2353	uint64_t eaddr;
2354	int r;
2355
2356	/* validate the parameters */
2357	if (saddr & ~PAGE_MASK || offset & ~PAGE_MASK ||
2358	    size == 0 || size & ~PAGE_MASK)
2359		return -EINVAL;
2360
2361	/* make sure object fit at this offset */
2362	eaddr = saddr + size - 1;
2363	if (saddr >= eaddr ||
2364	    (bo && offset + size > amdgpu_bo_size(bo)) ||
2365	    (eaddr >= adev->vm_manager.max_pfn << AMDGPU_GPU_PAGE_SHIFT))
2366		return -EINVAL;
2367
2368	/* Allocate all the needed memory */
2369	mapping = kmalloc(sizeof(*mapping), GFP_KERNEL);
2370	if (!mapping)
2371		return -ENOMEM;
2372
2373	r = amdgpu_vm_bo_clear_mappings(adev, bo_va->base.vm, saddr, size);
2374	if (r) {
2375		kfree(mapping);
2376		return r;
2377	}
2378
2379	saddr /= AMDGPU_GPU_PAGE_SIZE;
2380	eaddr /= AMDGPU_GPU_PAGE_SIZE;
2381
2382	mapping->start = saddr;
2383	mapping->last = eaddr;
2384	mapping->offset = offset;
2385	mapping->flags = flags;
2386
2387	amdgpu_vm_bo_insert_map(adev, bo_va, mapping);
2388
2389	return 0;
2390}
2391
2392/**
2393 * amdgpu_vm_bo_unmap - remove bo mapping from vm
2394 *
2395 * @adev: amdgpu_device pointer
2396 * @bo_va: bo_va to remove the address from
2397 * @saddr: where to the BO is mapped
2398 *
2399 * Remove a mapping of the BO at the specefied addr from the VM.
2400 *
2401 * Returns:
2402 * 0 for success, error for failure.
2403 *
2404 * Object has to be reserved and unreserved outside!
2405 */
2406int amdgpu_vm_bo_unmap(struct amdgpu_device *adev,
2407		       struct amdgpu_bo_va *bo_va,
2408		       uint64_t saddr)
2409{
2410	struct amdgpu_bo_va_mapping *mapping;
2411	struct amdgpu_vm *vm = bo_va->base.vm;
2412	bool valid = true;
2413
2414	saddr /= AMDGPU_GPU_PAGE_SIZE;
2415
2416	list_for_each_entry(mapping, &bo_va->valids, list) {
2417		if (mapping->start == saddr)
2418			break;
2419	}
2420
2421	if (&mapping->list == &bo_va->valids) {
2422		valid = false;
2423
2424		list_for_each_entry(mapping, &bo_va->invalids, list) {
2425			if (mapping->start == saddr)
2426				break;
2427		}
2428
2429		if (&mapping->list == &bo_va->invalids)
2430			return -ENOENT;
2431	}
2432
2433	list_del(&mapping->list);
2434	amdgpu_vm_it_remove(mapping, &vm->va);
2435	mapping->bo_va = NULL;
2436	trace_amdgpu_vm_bo_unmap(bo_va, mapping);
2437
2438	if (valid)
2439		list_add(&mapping->list, &vm->freed);
2440	else
2441		amdgpu_vm_free_mapping(adev, vm, mapping,
2442				       bo_va->last_pt_update);
2443
2444	return 0;
2445}
2446
2447/**
2448 * amdgpu_vm_bo_clear_mappings - remove all mappings in a specific range
2449 *
2450 * @adev: amdgpu_device pointer
2451 * @vm: VM structure to use
2452 * @saddr: start of the range
2453 * @size: size of the range
2454 *
2455 * Remove all mappings in a range, split them as appropriate.
2456 *
2457 * Returns:
2458 * 0 for success, error for failure.
2459 */
2460int amdgpu_vm_bo_clear_mappings(struct amdgpu_device *adev,
2461				struct amdgpu_vm *vm,
2462				uint64_t saddr, uint64_t size)
2463{
2464	struct amdgpu_bo_va_mapping *before, *after, *tmp, *next;
2465	LIST_HEAD(removed);
2466	uint64_t eaddr;
2467
2468	eaddr = saddr + size - 1;
2469	saddr /= AMDGPU_GPU_PAGE_SIZE;
2470	eaddr /= AMDGPU_GPU_PAGE_SIZE;
2471
2472	/* Allocate all the needed memory */
2473	before = kzalloc(sizeof(*before), GFP_KERNEL);
2474	if (!before)
2475		return -ENOMEM;
2476	INIT_LIST_HEAD(&before->list);
2477
2478	after = kzalloc(sizeof(*after), GFP_KERNEL);
2479	if (!after) {
2480		kfree(before);
2481		return -ENOMEM;
2482	}
2483	INIT_LIST_HEAD(&after->list);
2484
2485	/* Now gather all removed mappings */
2486	tmp = amdgpu_vm_it_iter_first(&vm->va, saddr, eaddr);
2487	while (tmp) {
2488		/* Remember mapping split at the start */
2489		if (tmp->start < saddr) {
2490			before->start = tmp->start;
2491			before->last = saddr - 1;
2492			before->offset = tmp->offset;
2493			before->flags = tmp->flags;
2494			before->bo_va = tmp->bo_va;
2495			list_add(&before->list, &tmp->bo_va->invalids);
2496		}
2497
2498		/* Remember mapping split at the end */
2499		if (tmp->last > eaddr) {
2500			after->start = eaddr + 1;
2501			after->last = tmp->last;
2502			after->offset = tmp->offset;
2503			after->offset += (after->start - tmp->start) << PAGE_SHIFT;
2504			after->flags = tmp->flags;
2505			after->bo_va = tmp->bo_va;
2506			list_add(&after->list, &tmp->bo_va->invalids);
2507		}
2508
2509		list_del(&tmp->list);
2510		list_add(&tmp->list, &removed);
2511
2512		tmp = amdgpu_vm_it_iter_next(tmp, saddr, eaddr);
2513	}
2514
2515	/* And free them up */
2516	list_for_each_entry_safe(tmp, next, &removed, list) {
2517		amdgpu_vm_it_remove(tmp, &vm->va);
2518		list_del(&tmp->list);
2519
2520		if (tmp->start < saddr)
2521		    tmp->start = saddr;
2522		if (tmp->last > eaddr)
2523		    tmp->last = eaddr;
2524
2525		tmp->bo_va = NULL;
2526		list_add(&tmp->list, &vm->freed);
2527		trace_amdgpu_vm_bo_unmap(NULL, tmp);
2528	}
2529
2530	/* Insert partial mapping before the range */
2531	if (!list_empty(&before->list)) {
2532		amdgpu_vm_it_insert(before, &vm->va);
2533		if (before->flags & AMDGPU_PTE_PRT)
2534			amdgpu_vm_prt_get(adev);
2535	} else {
2536		kfree(before);
2537	}
2538
2539	/* Insert partial mapping after the range */
2540	if (!list_empty(&after->list)) {
2541		amdgpu_vm_it_insert(after, &vm->va);
2542		if (after->flags & AMDGPU_PTE_PRT)
2543			amdgpu_vm_prt_get(adev);
2544	} else {
2545		kfree(after);
2546	}
2547
2548	return 0;
2549}
2550
2551/**
2552 * amdgpu_vm_bo_lookup_mapping - find mapping by address
2553 *
2554 * @vm: the requested VM
2555 * @addr: the address
2556 *
2557 * Find a mapping by it's address.
2558 *
2559 * Returns:
2560 * The amdgpu_bo_va_mapping matching for addr or NULL
2561 *
2562 */
2563struct amdgpu_bo_va_mapping *amdgpu_vm_bo_lookup_mapping(struct amdgpu_vm *vm,
2564							 uint64_t addr)
2565{
2566	return amdgpu_vm_it_iter_first(&vm->va, addr, addr);
2567}
2568
2569/**
2570 * amdgpu_vm_bo_trace_cs - trace all reserved mappings
2571 *
2572 * @vm: the requested vm
2573 * @ticket: CS ticket
2574 *
2575 * Trace all mappings of BOs reserved during a command submission.
2576 */
2577void amdgpu_vm_bo_trace_cs(struct amdgpu_vm *vm, struct ww_acquire_ctx *ticket)
2578{
2579	struct amdgpu_bo_va_mapping *mapping;
2580
2581	if (!trace_amdgpu_vm_bo_cs_enabled())
2582		return;
2583
2584	for (mapping = amdgpu_vm_it_iter_first(&vm->va, 0, U64_MAX); mapping;
2585	     mapping = amdgpu_vm_it_iter_next(mapping, 0, U64_MAX)) {
2586		if (mapping->bo_va && mapping->bo_va->base.bo) {
2587			struct amdgpu_bo *bo;
2588
2589			bo = mapping->bo_va->base.bo;
2590			if (dma_resv_locking_ctx(bo->tbo.base.resv) !=
2591			    ticket)
2592				continue;
2593		}
2594
2595		trace_amdgpu_vm_bo_cs(mapping);
2596	}
2597}
2598
2599/**
2600 * amdgpu_vm_bo_rmv - remove a bo to a specific vm
2601 *
2602 * @adev: amdgpu_device pointer
2603 * @bo_va: requested bo_va
2604 *
2605 * Remove @bo_va->bo from the requested vm.
2606 *
2607 * Object have to be reserved!
2608 */
2609void amdgpu_vm_bo_rmv(struct amdgpu_device *adev,
2610		      struct amdgpu_bo_va *bo_va)
2611{
2612	struct amdgpu_bo_va_mapping *mapping, *next;
2613	struct amdgpu_bo *bo = bo_va->base.bo;
2614	struct amdgpu_vm *vm = bo_va->base.vm;
2615	struct amdgpu_vm_bo_base **base;
2616
2617	if (bo) {
2618		if (bo->tbo.base.resv == vm->root.bo->tbo.base.resv)
2619			vm->bulk_moveable = false;
2620
2621		for (base = &bo_va->base.bo->vm_bo; *base;
2622		     base = &(*base)->next) {
2623			if (*base != &bo_va->base)
2624				continue;
2625
2626			*base = bo_va->base.next;
2627			break;
2628		}
2629	}
2630
2631	spin_lock(&vm->invalidated_lock);
2632	list_del(&bo_va->base.vm_status);
2633	spin_unlock(&vm->invalidated_lock);
2634
2635	list_for_each_entry_safe(mapping, next, &bo_va->valids, list) {
2636		list_del(&mapping->list);
2637		amdgpu_vm_it_remove(mapping, &vm->va);
2638		mapping->bo_va = NULL;
2639		trace_amdgpu_vm_bo_unmap(bo_va, mapping);
2640		list_add(&mapping->list, &vm->freed);
2641	}
2642	list_for_each_entry_safe(mapping, next, &bo_va->invalids, list) {
2643		list_del(&mapping->list);
2644		amdgpu_vm_it_remove(mapping, &vm->va);
2645		amdgpu_vm_free_mapping(adev, vm, mapping,
2646				       bo_va->last_pt_update);
2647	}
2648
2649	dma_fence_put(bo_va->last_pt_update);
2650
2651	if (bo && bo_va->is_xgmi)
2652		amdgpu_xgmi_set_pstate(adev, AMDGPU_XGMI_PSTATE_MIN);
2653
2654	kfree(bo_va);
2655}
2656
2657/**
2658 * amdgpu_vm_evictable - check if we can evict a VM
2659 *
2660 * @bo: A page table of the VM.
2661 *
2662 * Check if it is possible to evict a VM.
2663 */
2664bool amdgpu_vm_evictable(struct amdgpu_bo *bo)
2665{
2666	struct amdgpu_vm_bo_base *bo_base = bo->vm_bo;
2667
2668	/* Page tables of a destroyed VM can go away immediately */
2669	if (!bo_base || !bo_base->vm)
2670		return true;
2671
2672	/* Don't evict VM page tables while they are busy */
2673	if (!dma_resv_test_signaled(bo->tbo.base.resv, true))
2674		return false;
2675
2676	/* Try to block ongoing updates */
2677	if (!amdgpu_vm_eviction_trylock(bo_base->vm))
2678		return false;
2679
2680	/* Don't evict VM page tables while they are updated */
2681	if (!dma_fence_is_signaled(bo_base->vm->last_unlocked)) {
2682		amdgpu_vm_eviction_unlock(bo_base->vm);
2683		return false;
2684	}
2685
2686	bo_base->vm->evicting = true;
2687	amdgpu_vm_eviction_unlock(bo_base->vm);
2688	return true;
2689}
2690
2691/**
2692 * amdgpu_vm_bo_invalidate - mark the bo as invalid
2693 *
2694 * @adev: amdgpu_device pointer
2695 * @bo: amdgpu buffer object
2696 * @evicted: is the BO evicted
2697 *
2698 * Mark @bo as invalid.
2699 */
2700void amdgpu_vm_bo_invalidate(struct amdgpu_device *adev,
2701			     struct amdgpu_bo *bo, bool evicted)
2702{
2703	struct amdgpu_vm_bo_base *bo_base;
2704
2705	/* shadow bo doesn't have bo base, its validation needs its parent */
2706	if (bo->parent && (amdgpu_bo_shadowed(bo->parent) == bo))
2707		bo = bo->parent;
2708
2709	for (bo_base = bo->vm_bo; bo_base; bo_base = bo_base->next) {
2710		struct amdgpu_vm *vm = bo_base->vm;
2711
2712		if (evicted && bo->tbo.base.resv == vm->root.bo->tbo.base.resv) {
2713			amdgpu_vm_bo_evicted(bo_base);
2714			continue;
2715		}
2716
2717		if (bo_base->moved)
2718			continue;
2719		bo_base->moved = true;
2720
2721		if (bo->tbo.type == ttm_bo_type_kernel)
2722			amdgpu_vm_bo_relocated(bo_base);
2723		else if (bo->tbo.base.resv == vm->root.bo->tbo.base.resv)
2724			amdgpu_vm_bo_moved(bo_base);
2725		else
2726			amdgpu_vm_bo_invalidated(bo_base);
2727	}
2728}
2729
2730/**
2731 * amdgpu_vm_get_block_size - calculate VM page table size as power of two
2732 *
2733 * @vm_size: VM size
2734 *
2735 * Returns:
2736 * VM page table as power of two
2737 */
2738static uint32_t amdgpu_vm_get_block_size(uint64_t vm_size)
2739{
2740	/* Total bits covered by PD + PTs */
2741	unsigned bits = ilog2(vm_size) + 18;
2742
2743	/* Make sure the PD is 4K in size up to 8GB address space.
2744	   Above that split equal between PD and PTs */
2745	if (vm_size <= 8)
2746		return (bits - 9);
2747	else
2748		return ((bits + 3) / 2);
2749}
2750
2751/**
2752 * amdgpu_vm_adjust_size - adjust vm size, block size and fragment size
2753 *
2754 * @adev: amdgpu_device pointer
2755 * @min_vm_size: the minimum vm size in GB if it's set auto
2756 * @fragment_size_default: Default PTE fragment size
2757 * @max_level: max VMPT level
2758 * @max_bits: max address space size in bits
2759 *
2760 */
2761void amdgpu_vm_adjust_size(struct amdgpu_device *adev, uint32_t min_vm_size,
2762			   uint32_t fragment_size_default, unsigned max_level,
2763			   unsigned max_bits)
2764{
2765	unsigned int max_size = 1 << (max_bits - 30);
2766	unsigned int vm_size;
2767	uint64_t tmp;
2768
2769	/* adjust vm size first */
2770	if (amdgpu_vm_size != -1) {
2771		vm_size = amdgpu_vm_size;
2772		if (vm_size > max_size) {
2773			dev_warn(adev->dev, "VM size (%d) too large, max is %u GB\n",
2774				 amdgpu_vm_size, max_size);
2775			vm_size = max_size;
2776		}
2777	} else {
2778		struct sysinfo si;
2779		unsigned int phys_ram_gb;
2780
2781		/* Optimal VM size depends on the amount of physical
2782		 * RAM available. Underlying requirements and
2783		 * assumptions:
2784		 *
2785		 *  - Need to map system memory and VRAM from all GPUs
2786		 *     - VRAM from other GPUs not known here
2787		 *     - Assume VRAM <= system memory
2788		 *  - On GFX8 and older, VM space can be segmented for
2789		 *    different MTYPEs
2790		 *  - Need to allow room for fragmentation, guard pages etc.
2791		 *
2792		 * This adds up to a rough guess of system memory x3.
2793		 * Round up to power of two to maximize the available
2794		 * VM size with the given page table size.
2795		 */
2796		si_meminfo(&si);
2797		phys_ram_gb = ((uint64_t)si.totalram * si.mem_unit +
2798			       (1 << 30) - 1) >> 30;
2799		vm_size = roundup_pow_of_two(
2800			min(max(phys_ram_gb * 3, min_vm_size), max_size));
2801	}
2802
2803	adev->vm_manager.max_pfn = (uint64_t)vm_size << 18;
2804
2805	tmp = roundup_pow_of_two(adev->vm_manager.max_pfn);
2806	if (amdgpu_vm_block_size != -1)
2807		tmp >>= amdgpu_vm_block_size - 9;
2808	tmp = DIV_ROUND_UP(fls64(tmp) - 1, 9) - 1;
2809	adev->vm_manager.num_level = min(max_level, (unsigned)tmp);
2810	switch (adev->vm_manager.num_level) {
2811	case 3:
2812		adev->vm_manager.root_level = AMDGPU_VM_PDB2;
2813		break;
2814	case 2:
2815		adev->vm_manager.root_level = AMDGPU_VM_PDB1;
2816		break;
2817	case 1:
2818		adev->vm_manager.root_level = AMDGPU_VM_PDB0;
2819		break;
2820	default:
2821		dev_err(adev->dev, "VMPT only supports 2~4+1 levels\n");
2822	}
2823	/* block size depends on vm size and hw setup*/
2824	if (amdgpu_vm_block_size != -1)
2825		adev->vm_manager.block_size =
2826			min((unsigned)amdgpu_vm_block_size, max_bits
2827			    - AMDGPU_GPU_PAGE_SHIFT
2828			    - 9 * adev->vm_manager.num_level);
2829	else if (adev->vm_manager.num_level > 1)
2830		adev->vm_manager.block_size = 9;
2831	else
2832		adev->vm_manager.block_size = amdgpu_vm_get_block_size(tmp);
2833
2834	if (amdgpu_vm_fragment_size == -1)
2835		adev->vm_manager.fragment_size = fragment_size_default;
2836	else
2837		adev->vm_manager.fragment_size = amdgpu_vm_fragment_size;
2838
2839	DRM_INFO("vm size is %u GB, %u levels, block size is %u-bit, fragment size is %u-bit\n",
2840		 vm_size, adev->vm_manager.num_level + 1,
2841		 adev->vm_manager.block_size,
2842		 adev->vm_manager.fragment_size);
2843}
2844
2845/**
2846 * amdgpu_vm_wait_idle - wait for the VM to become idle
2847 *
2848 * @vm: VM object to wait for
2849 * @timeout: timeout to wait for VM to become idle
2850 */
2851long amdgpu_vm_wait_idle(struct amdgpu_vm *vm, long timeout)
2852{
2853	timeout = dma_resv_wait_timeout(vm->root.bo->tbo.base.resv, true,
2854					true, timeout);
2855	if (timeout <= 0)
2856		return timeout;
2857
2858	return dma_fence_wait_timeout(vm->last_unlocked, true, timeout);
2859}
2860
2861/**
2862 * amdgpu_vm_init - initialize a vm instance
2863 *
2864 * @adev: amdgpu_device pointer
2865 * @vm: requested vm
 
2866 * @pasid: Process address space identifier
2867 *
2868 * Init @vm fields.
2869 *
2870 * Returns:
2871 * 0 for success, error for failure.
2872 */
2873int amdgpu_vm_init(struct amdgpu_device *adev, struct amdgpu_vm *vm, u32 pasid)
 
2874{
2875	struct amdgpu_bo *root_bo;
2876	struct amdgpu_bo_vm *root;
2877	int r, i;
2878
2879	vm->va = RB_ROOT_CACHED;
2880	for (i = 0; i < AMDGPU_MAX_VMHUBS; i++)
2881		vm->reserved_vmid[i] = NULL;
2882	INIT_LIST_HEAD(&vm->evicted);
2883	INIT_LIST_HEAD(&vm->relocated);
2884	INIT_LIST_HEAD(&vm->moved);
2885	INIT_LIST_HEAD(&vm->idle);
2886	INIT_LIST_HEAD(&vm->invalidated);
2887	spin_lock_init(&vm->invalidated_lock);
2888	INIT_LIST_HEAD(&vm->freed);
2889	INIT_LIST_HEAD(&vm->done);
2890
2891	/* create scheduler entities for page table updates */
2892	r = drm_sched_entity_init(&vm->immediate, DRM_SCHED_PRIORITY_NORMAL,
2893				  adev->vm_manager.vm_pte_scheds,
2894				  adev->vm_manager.vm_pte_num_scheds, NULL);
2895	if (r)
2896		return r;
2897
2898	r = drm_sched_entity_init(&vm->delayed, DRM_SCHED_PRIORITY_NORMAL,
2899				  adev->vm_manager.vm_pte_scheds,
2900				  adev->vm_manager.vm_pte_num_scheds, NULL);
2901	if (r)
2902		goto error_free_immediate;
2903
2904	vm->pte_support_ats = false;
2905	vm->is_compute_context = false;
2906
2907	vm->use_cpu_for_update = !!(adev->vm_manager.vm_update_mode &
2908				    AMDGPU_VM_USE_CPU_FOR_GFX);
 
2909
 
 
 
 
 
 
2910	DRM_DEBUG_DRIVER("VM update mode is %s\n",
2911			 vm->use_cpu_for_update ? "CPU" : "SDMA");
2912	WARN_ONCE((vm->use_cpu_for_update &&
2913		   !amdgpu_gmc_vram_full_visible(&adev->gmc)),
2914		  "CPU update of VM recommended only for large BAR system\n");
2915
2916	if (vm->use_cpu_for_update)
2917		vm->update_funcs = &amdgpu_vm_cpu_funcs;
2918	else
2919		vm->update_funcs = &amdgpu_vm_sdma_funcs;
2920	vm->last_update = NULL;
2921	vm->last_unlocked = dma_fence_get_stub();
2922
2923	mutex_init(&vm->eviction_lock);
2924	vm->evicting = false;
2925
2926	r = amdgpu_vm_pt_create(adev, vm, adev->vm_manager.root_level,
2927				false, &root);
2928	if (r)
2929		goto error_free_delayed;
2930	root_bo = &root->bo;
2931	r = amdgpu_bo_reserve(root_bo, true);
2932	if (r)
2933		goto error_free_root;
2934
2935	r = dma_resv_reserve_shared(root_bo->tbo.base.resv, 1);
2936	if (r)
2937		goto error_unreserve;
2938
2939	amdgpu_vm_bo_base_init(&vm->root, vm, root_bo);
2940
2941	r = amdgpu_vm_clear_bo(adev, vm, root, false);
2942	if (r)
2943		goto error_unreserve;
2944
2945	amdgpu_bo_unreserve(vm->root.bo);
2946
2947	if (pasid) {
2948		unsigned long flags;
2949
2950		spin_lock_irqsave(&adev->vm_manager.pasid_lock, flags);
2951		r = idr_alloc(&adev->vm_manager.pasid_idr, vm, pasid, pasid + 1,
2952			      GFP_ATOMIC);
2953		spin_unlock_irqrestore(&adev->vm_manager.pasid_lock, flags);
2954		if (r < 0)
2955			goto error_free_root;
2956
2957		vm->pasid = pasid;
2958	}
2959
2960	INIT_KFIFO(vm->faults);
2961
2962	return 0;
2963
2964error_unreserve:
2965	amdgpu_bo_unreserve(vm->root.bo);
2966
2967error_free_root:
2968	amdgpu_bo_unref(&root->shadow);
2969	amdgpu_bo_unref(&root_bo);
2970	vm->root.bo = NULL;
2971
2972error_free_delayed:
2973	dma_fence_put(vm->last_unlocked);
2974	drm_sched_entity_destroy(&vm->delayed);
2975
2976error_free_immediate:
2977	drm_sched_entity_destroy(&vm->immediate);
2978
2979	return r;
2980}
2981
2982/**
2983 * amdgpu_vm_check_clean_reserved - check if a VM is clean
2984 *
2985 * @adev: amdgpu_device pointer
2986 * @vm: the VM to check
2987 *
2988 * check all entries of the root PD, if any subsequent PDs are allocated,
2989 * it means there are page table creating and filling, and is no a clean
2990 * VM
2991 *
2992 * Returns:
2993 *	0 if this VM is clean
2994 */
2995static int amdgpu_vm_check_clean_reserved(struct amdgpu_device *adev,
2996					  struct amdgpu_vm *vm)
2997{
2998	enum amdgpu_vm_level root = adev->vm_manager.root_level;
2999	unsigned int entries = amdgpu_vm_num_entries(adev, root);
3000	unsigned int i = 0;
3001
 
 
 
3002	for (i = 0; i < entries; i++) {
3003		if (to_amdgpu_bo_vm(vm->root.bo)->entries[i].bo)
3004			return -EINVAL;
3005	}
3006
3007	return 0;
3008}
3009
3010/**
3011 * amdgpu_vm_make_compute - Turn a GFX VM into a compute VM
3012 *
3013 * @adev: amdgpu_device pointer
3014 * @vm: requested vm
3015 * @pasid: pasid to use
3016 *
3017 * This only works on GFX VMs that don't have any BOs added and no
3018 * page tables allocated yet.
3019 *
3020 * Changes the following VM parameters:
3021 * - use_cpu_for_update
3022 * - pte_supports_ats
3023 * - pasid (old PASID is released, because compute manages its own PASIDs)
3024 *
3025 * Reinitializes the page directory to reflect the changed ATS
3026 * setting.
3027 *
3028 * Returns:
3029 * 0 for success, -errno for errors.
3030 */
3031int amdgpu_vm_make_compute(struct amdgpu_device *adev, struct amdgpu_vm *vm,
3032			   u32 pasid)
3033{
3034	bool pte_support_ats = (adev->asic_type == CHIP_RAVEN);
3035	int r;
3036
3037	r = amdgpu_bo_reserve(vm->root.bo, true);
3038	if (r)
3039		return r;
3040
3041	/* Sanity checks */
3042	r = amdgpu_vm_check_clean_reserved(adev, vm);
3043	if (r)
3044		goto unreserve_bo;
3045
3046	if (pasid) {
3047		unsigned long flags;
3048
3049		spin_lock_irqsave(&adev->vm_manager.pasid_lock, flags);
3050		r = idr_alloc(&adev->vm_manager.pasid_idr, vm, pasid, pasid + 1,
3051			      GFP_ATOMIC);
3052		spin_unlock_irqrestore(&adev->vm_manager.pasid_lock, flags);
3053
3054		if (r == -ENOSPC)
3055			goto unreserve_bo;
3056		r = 0;
3057	}
3058
3059	/* Check if PD needs to be reinitialized and do it before
3060	 * changing any other state, in case it fails.
3061	 */
3062	if (pte_support_ats != vm->pte_support_ats) {
3063		vm->pte_support_ats = pte_support_ats;
3064		r = amdgpu_vm_clear_bo(adev, vm,
3065				       to_amdgpu_bo_vm(vm->root.bo),
3066				       false);
3067		if (r)
3068			goto free_idr;
3069	}
3070
3071	/* Update VM state */
3072	vm->use_cpu_for_update = !!(adev->vm_manager.vm_update_mode &
3073				    AMDGPU_VM_USE_CPU_FOR_COMPUTE);
3074	DRM_DEBUG_DRIVER("VM update mode is %s\n",
3075			 vm->use_cpu_for_update ? "CPU" : "SDMA");
3076	WARN_ONCE((vm->use_cpu_for_update &&
3077		   !amdgpu_gmc_vram_full_visible(&adev->gmc)),
3078		  "CPU update of VM recommended only for large BAR system\n");
3079
3080	if (vm->use_cpu_for_update) {
3081		/* Sync with last SDMA update/clear before switching to CPU */
3082		r = amdgpu_bo_sync_wait(vm->root.bo,
3083					AMDGPU_FENCE_OWNER_UNDEFINED, true);
3084		if (r)
3085			goto free_idr;
3086
3087		vm->update_funcs = &amdgpu_vm_cpu_funcs;
3088	} else {
3089		vm->update_funcs = &amdgpu_vm_sdma_funcs;
3090	}
3091	dma_fence_put(vm->last_update);
3092	vm->last_update = NULL;
3093	vm->is_compute_context = true;
3094
3095	if (vm->pasid) {
3096		unsigned long flags;
3097
3098		spin_lock_irqsave(&adev->vm_manager.pasid_lock, flags);
3099		idr_remove(&adev->vm_manager.pasid_idr, vm->pasid);
3100		spin_unlock_irqrestore(&adev->vm_manager.pasid_lock, flags);
3101
3102		/* Free the original amdgpu allocated pasid
3103		 * Will be replaced with kfd allocated pasid
3104		 */
3105		amdgpu_pasid_free(vm->pasid);
3106		vm->pasid = 0;
3107	}
3108
3109	/* Free the shadow bo for compute VM */
3110	amdgpu_bo_unref(&to_amdgpu_bo_vm(vm->root.bo)->shadow);
3111
3112	if (pasid)
3113		vm->pasid = pasid;
3114
3115	goto unreserve_bo;
3116
3117free_idr:
3118	if (pasid) {
3119		unsigned long flags;
3120
3121		spin_lock_irqsave(&adev->vm_manager.pasid_lock, flags);
3122		idr_remove(&adev->vm_manager.pasid_idr, pasid);
3123		spin_unlock_irqrestore(&adev->vm_manager.pasid_lock, flags);
3124	}
3125unreserve_bo:
3126	amdgpu_bo_unreserve(vm->root.bo);
3127	return r;
3128}
3129
3130/**
3131 * amdgpu_vm_release_compute - release a compute vm
3132 * @adev: amdgpu_device pointer
3133 * @vm: a vm turned into compute vm by calling amdgpu_vm_make_compute
3134 *
3135 * This is a correspondant of amdgpu_vm_make_compute. It decouples compute
3136 * pasid from vm. Compute should stop use of vm after this call.
3137 */
3138void amdgpu_vm_release_compute(struct amdgpu_device *adev, struct amdgpu_vm *vm)
3139{
3140	if (vm->pasid) {
3141		unsigned long flags;
3142
3143		spin_lock_irqsave(&adev->vm_manager.pasid_lock, flags);
3144		idr_remove(&adev->vm_manager.pasid_idr, vm->pasid);
3145		spin_unlock_irqrestore(&adev->vm_manager.pasid_lock, flags);
3146	}
3147	vm->pasid = 0;
3148	vm->is_compute_context = false;
3149}
3150
3151/**
3152 * amdgpu_vm_fini - tear down a vm instance
3153 *
3154 * @adev: amdgpu_device pointer
3155 * @vm: requested vm
3156 *
3157 * Tear down @vm.
3158 * Unbind the VM and remove all bos from the vm bo list
3159 */
3160void amdgpu_vm_fini(struct amdgpu_device *adev, struct amdgpu_vm *vm)
3161{
3162	struct amdgpu_bo_va_mapping *mapping, *tmp;
3163	bool prt_fini_needed = !!adev->gmc.gmc_funcs->set_prt;
3164	struct amdgpu_bo *root;
3165	int i;
3166
3167	amdgpu_amdkfd_gpuvm_destroy_cb(adev, vm);
3168
3169	root = amdgpu_bo_ref(vm->root.bo);
3170	amdgpu_bo_reserve(root, true);
3171	if (vm->pasid) {
3172		unsigned long flags;
3173
3174		spin_lock_irqsave(&adev->vm_manager.pasid_lock, flags);
3175		idr_remove(&adev->vm_manager.pasid_idr, vm->pasid);
3176		spin_unlock_irqrestore(&adev->vm_manager.pasid_lock, flags);
3177		vm->pasid = 0;
3178	}
3179
3180	dma_fence_wait(vm->last_unlocked, false);
3181	dma_fence_put(vm->last_unlocked);
3182
3183	list_for_each_entry_safe(mapping, tmp, &vm->freed, list) {
3184		if (mapping->flags & AMDGPU_PTE_PRT && prt_fini_needed) {
3185			amdgpu_vm_prt_fini(adev, vm);
3186			prt_fini_needed = false;
3187		}
3188
3189		list_del(&mapping->list);
3190		amdgpu_vm_free_mapping(adev, vm, mapping, NULL);
3191	}
3192
3193	amdgpu_vm_free_pts(adev, vm, NULL);
3194	amdgpu_bo_unreserve(root);
3195	amdgpu_bo_unref(&root);
3196	WARN_ON(vm->root.bo);
3197
3198	drm_sched_entity_destroy(&vm->immediate);
3199	drm_sched_entity_destroy(&vm->delayed);
3200
3201	if (!RB_EMPTY_ROOT(&vm->va.rb_root)) {
3202		dev_err(adev->dev, "still active bo inside vm\n");
3203	}
3204	rbtree_postorder_for_each_entry_safe(mapping, tmp,
3205					     &vm->va.rb_root, rb) {
3206		/* Don't remove the mapping here, we don't want to trigger a
3207		 * rebalance and the tree is about to be destroyed anyway.
3208		 */
3209		list_del(&mapping->list);
3210		kfree(mapping);
3211	}
 
 
 
 
 
3212
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3213	dma_fence_put(vm->last_update);
3214	for (i = 0; i < AMDGPU_MAX_VMHUBS; i++)
3215		amdgpu_vmid_free_reserved(adev, vm, i);
3216}
3217
3218/**
3219 * amdgpu_vm_manager_init - init the VM manager
3220 *
3221 * @adev: amdgpu_device pointer
3222 *
3223 * Initialize the VM manager structures
3224 */
3225void amdgpu_vm_manager_init(struct amdgpu_device *adev)
3226{
3227	unsigned i;
3228
3229	/* Concurrent flushes are only possible starting with Vega10 and
3230	 * are broken on Navi10 and Navi14.
3231	 */
3232	adev->vm_manager.concurrent_flush = !(adev->asic_type < CHIP_VEGA10 ||
3233					      adev->asic_type == CHIP_NAVI10 ||
3234					      adev->asic_type == CHIP_NAVI14);
3235	amdgpu_vmid_mgr_init(adev);
3236
3237	adev->vm_manager.fence_context =
3238		dma_fence_context_alloc(AMDGPU_MAX_RINGS);
3239	for (i = 0; i < AMDGPU_MAX_RINGS; ++i)
3240		adev->vm_manager.seqno[i] = 0;
3241
3242	spin_lock_init(&adev->vm_manager.prt_lock);
3243	atomic_set(&adev->vm_manager.num_prt_users, 0);
3244
3245	/* If not overridden by the user, by default, only in large BAR systems
3246	 * Compute VM tables will be updated by CPU
3247	 */
3248#ifdef CONFIG_X86_64
3249	if (amdgpu_vm_update_mode == -1) {
3250		if (amdgpu_gmc_vram_full_visible(&adev->gmc))
3251			adev->vm_manager.vm_update_mode =
3252				AMDGPU_VM_USE_CPU_FOR_COMPUTE;
3253		else
3254			adev->vm_manager.vm_update_mode = 0;
3255	} else
3256		adev->vm_manager.vm_update_mode = amdgpu_vm_update_mode;
3257#else
3258	adev->vm_manager.vm_update_mode = 0;
3259#endif
3260
3261	idr_init(&adev->vm_manager.pasid_idr);
3262	spin_lock_init(&adev->vm_manager.pasid_lock);
 
 
 
3263}
3264
3265/**
3266 * amdgpu_vm_manager_fini - cleanup VM manager
3267 *
3268 * @adev: amdgpu_device pointer
3269 *
3270 * Cleanup the VM manager and free resources.
3271 */
3272void amdgpu_vm_manager_fini(struct amdgpu_device *adev)
3273{
3274	WARN_ON(!idr_is_empty(&adev->vm_manager.pasid_idr));
3275	idr_destroy(&adev->vm_manager.pasid_idr);
3276
3277	amdgpu_vmid_mgr_fini(adev);
3278}
3279
3280/**
3281 * amdgpu_vm_ioctl - Manages VMID reservation for vm hubs.
3282 *
3283 * @dev: drm device pointer
3284 * @data: drm_amdgpu_vm
3285 * @filp: drm file pointer
3286 *
3287 * Returns:
3288 * 0 for success, -errno for errors.
3289 */
3290int amdgpu_vm_ioctl(struct drm_device *dev, void *data, struct drm_file *filp)
3291{
3292	union drm_amdgpu_vm *args = data;
3293	struct amdgpu_device *adev = drm_to_adev(dev);
3294	struct amdgpu_fpriv *fpriv = filp->driver_priv;
3295	long timeout = msecs_to_jiffies(2000);
3296	int r;
3297
3298	switch (args->in.op) {
3299	case AMDGPU_VM_OP_RESERVE_VMID:
3300		/* We only have requirement to reserve vmid from gfxhub */
3301		r = amdgpu_vmid_alloc_reserved(adev, &fpriv->vm,
3302					       AMDGPU_GFXHUB_0);
3303		if (r)
3304			return r;
3305		break;
3306	case AMDGPU_VM_OP_UNRESERVE_VMID:
3307		if (amdgpu_sriov_runtime(adev))
3308			timeout = 8 * timeout;
3309
3310		/* Wait vm idle to make sure the vmid set in SPM_VMID is
3311		 * not referenced anymore.
3312		 */
3313		r = amdgpu_bo_reserve(fpriv->vm.root.bo, true);
3314		if (r)
3315			return r;
3316
3317		r = amdgpu_vm_wait_idle(&fpriv->vm, timeout);
3318		if (r < 0)
3319			return r;
3320
3321		amdgpu_bo_unreserve(fpriv->vm.root.bo);
3322		amdgpu_vmid_free_reserved(adev, &fpriv->vm, AMDGPU_GFXHUB_0);
3323		break;
3324	default:
3325		return -EINVAL;
3326	}
3327
3328	return 0;
3329}
3330
3331/**
3332 * amdgpu_vm_get_task_info - Extracts task info for a PASID.
3333 *
3334 * @adev: drm device pointer
3335 * @pasid: PASID identifier for VM
3336 * @task_info: task_info to fill.
3337 */
3338void amdgpu_vm_get_task_info(struct amdgpu_device *adev, u32 pasid,
3339			 struct amdgpu_task_info *task_info)
3340{
3341	struct amdgpu_vm *vm;
3342	unsigned long flags;
3343
3344	spin_lock_irqsave(&adev->vm_manager.pasid_lock, flags);
3345
3346	vm = idr_find(&adev->vm_manager.pasid_idr, pasid);
3347	if (vm)
3348		*task_info = vm->task_info;
3349
3350	spin_unlock_irqrestore(&adev->vm_manager.pasid_lock, flags);
3351}
3352
3353/**
3354 * amdgpu_vm_set_task_info - Sets VMs task info.
3355 *
3356 * @vm: vm for which to set the info
3357 */
3358void amdgpu_vm_set_task_info(struct amdgpu_vm *vm)
3359{
3360	if (vm->task_info.pid)
3361		return;
3362
3363	vm->task_info.pid = current->pid;
3364	get_task_comm(vm->task_info.task_name, current);
3365
3366	if (current->group_leader->mm != current->mm)
3367		return;
3368
3369	vm->task_info.tgid = current->group_leader->pid;
3370	get_task_comm(vm->task_info.process_name, current->group_leader);
3371}
3372
3373/**
3374 * amdgpu_vm_handle_fault - graceful handling of VM faults.
3375 * @adev: amdgpu device pointer
3376 * @pasid: PASID of the VM
3377 * @addr: Address of the fault
3378 *
3379 * Try to gracefully handle a VM fault. Return true if the fault was handled and
3380 * shouldn't be reported any more.
3381 */
3382bool amdgpu_vm_handle_fault(struct amdgpu_device *adev, u32 pasid,
3383			    uint64_t addr)
3384{
3385	bool is_compute_context = false;
3386	struct amdgpu_bo *root;
3387	unsigned long irqflags;
3388	uint64_t value, flags;
3389	struct amdgpu_vm *vm;
3390	int r;
3391
3392	spin_lock_irqsave(&adev->vm_manager.pasid_lock, irqflags);
3393	vm = idr_find(&adev->vm_manager.pasid_idr, pasid);
3394	if (vm) {
3395		root = amdgpu_bo_ref(vm->root.bo);
3396		is_compute_context = vm->is_compute_context;
3397	} else {
3398		root = NULL;
3399	}
3400	spin_unlock_irqrestore(&adev->vm_manager.pasid_lock, irqflags);
3401
3402	if (!root)
3403		return false;
3404
3405	addr /= AMDGPU_GPU_PAGE_SIZE;
3406
3407	if (is_compute_context &&
3408	    !svm_range_restore_pages(adev, pasid, addr)) {
3409		amdgpu_bo_unref(&root);
3410		return true;
3411	}
3412
3413	r = amdgpu_bo_reserve(root, true);
3414	if (r)
3415		goto error_unref;
3416
3417	/* Double check that the VM still exists */
3418	spin_lock_irqsave(&adev->vm_manager.pasid_lock, irqflags);
3419	vm = idr_find(&adev->vm_manager.pasid_idr, pasid);
3420	if (vm && vm->root.bo != root)
3421		vm = NULL;
3422	spin_unlock_irqrestore(&adev->vm_manager.pasid_lock, irqflags);
3423	if (!vm)
3424		goto error_unlock;
3425
3426	flags = AMDGPU_PTE_VALID | AMDGPU_PTE_SNOOPED |
3427		AMDGPU_PTE_SYSTEM;
3428
3429	if (is_compute_context) {
3430		/* Intentionally setting invalid PTE flag
3431		 * combination to force a no-retry-fault
3432		 */
3433		flags = AMDGPU_PTE_EXECUTABLE | AMDGPU_PDE_PTE |
3434			AMDGPU_PTE_TF;
3435		value = 0;
3436	} else if (amdgpu_vm_fault_stop == AMDGPU_VM_FAULT_STOP_NEVER) {
3437		/* Redirect the access to the dummy page */
3438		value = adev->dummy_page_addr;
3439		flags |= AMDGPU_PTE_EXECUTABLE | AMDGPU_PTE_READABLE |
3440			AMDGPU_PTE_WRITEABLE;
3441
3442	} else {
3443		/* Let the hw retry silently on the PTE */
3444		value = 0;
3445	}
3446
3447	r = dma_resv_reserve_shared(root->tbo.base.resv, 1);
3448	if (r) {
3449		pr_debug("failed %d to reserve fence slot\n", r);
3450		goto error_unlock;
3451	}
3452
3453	r = amdgpu_vm_bo_update_mapping(adev, adev, vm, true, false, NULL, addr,
3454					addr, flags, value, NULL, NULL, NULL,
3455					NULL);
3456	if (r)
3457		goto error_unlock;
3458
3459	r = amdgpu_vm_update_pdes(adev, vm, true);
3460
3461error_unlock:
3462	amdgpu_bo_unreserve(root);
3463	if (r < 0)
3464		DRM_ERROR("Can't handle page fault (%d)\n", r);
3465
3466error_unref:
3467	amdgpu_bo_unref(&root);
3468
3469	return false;
3470}
3471
3472#if defined(CONFIG_DEBUG_FS)
3473/**
3474 * amdgpu_debugfs_vm_bo_info  - print BO info for the VM
3475 *
3476 * @vm: Requested VM for printing BO info
3477 * @m: debugfs file
3478 *
3479 * Print BO information in debugfs file for the VM
3480 */
3481void amdgpu_debugfs_vm_bo_info(struct amdgpu_vm *vm, struct seq_file *m)
3482{
3483	struct amdgpu_bo_va *bo_va, *tmp;
3484	u64 total_idle = 0;
3485	u64 total_evicted = 0;
3486	u64 total_relocated = 0;
3487	u64 total_moved = 0;
3488	u64 total_invalidated = 0;
3489	u64 total_done = 0;
3490	unsigned int total_idle_objs = 0;
3491	unsigned int total_evicted_objs = 0;
3492	unsigned int total_relocated_objs = 0;
3493	unsigned int total_moved_objs = 0;
3494	unsigned int total_invalidated_objs = 0;
3495	unsigned int total_done_objs = 0;
3496	unsigned int id = 0;
3497
3498	seq_puts(m, "\tIdle BOs:\n");
3499	list_for_each_entry_safe(bo_va, tmp, &vm->idle, base.vm_status) {
3500		if (!bo_va->base.bo)
3501			continue;
3502		total_idle += amdgpu_bo_print_info(id++, bo_va->base.bo, m);
3503	}
3504	total_idle_objs = id;
3505	id = 0;
3506
3507	seq_puts(m, "\tEvicted BOs:\n");
3508	list_for_each_entry_safe(bo_va, tmp, &vm->evicted, base.vm_status) {
3509		if (!bo_va->base.bo)
3510			continue;
3511		total_evicted += amdgpu_bo_print_info(id++, bo_va->base.bo, m);
3512	}
3513	total_evicted_objs = id;
3514	id = 0;
3515
3516	seq_puts(m, "\tRelocated BOs:\n");
3517	list_for_each_entry_safe(bo_va, tmp, &vm->relocated, base.vm_status) {
3518		if (!bo_va->base.bo)
3519			continue;
3520		total_relocated += amdgpu_bo_print_info(id++, bo_va->base.bo, m);
3521	}
3522	total_relocated_objs = id;
3523	id = 0;
3524
3525	seq_puts(m, "\tMoved BOs:\n");
3526	list_for_each_entry_safe(bo_va, tmp, &vm->moved, base.vm_status) {
3527		if (!bo_va->base.bo)
3528			continue;
3529		total_moved += amdgpu_bo_print_info(id++, bo_va->base.bo, m);
3530	}
3531	total_moved_objs = id;
3532	id = 0;
3533
3534	seq_puts(m, "\tInvalidated BOs:\n");
3535	spin_lock(&vm->invalidated_lock);
3536	list_for_each_entry_safe(bo_va, tmp, &vm->invalidated, base.vm_status) {
3537		if (!bo_va->base.bo)
3538			continue;
3539		total_invalidated += amdgpu_bo_print_info(id++,	bo_va->base.bo, m);
3540	}
3541	total_invalidated_objs = id;
3542	id = 0;
3543
3544	seq_puts(m, "\tDone BOs:\n");
3545	list_for_each_entry_safe(bo_va, tmp, &vm->done, base.vm_status) {
3546		if (!bo_va->base.bo)
3547			continue;
3548		total_done += amdgpu_bo_print_info(id++, bo_va->base.bo, m);
3549	}
3550	spin_unlock(&vm->invalidated_lock);
3551	total_done_objs = id;
3552
3553	seq_printf(m, "\tTotal idle size:        %12lld\tobjs:\t%d\n", total_idle,
3554		   total_idle_objs);
3555	seq_printf(m, "\tTotal evicted size:     %12lld\tobjs:\t%d\n", total_evicted,
3556		   total_evicted_objs);
3557	seq_printf(m, "\tTotal relocated size:   %12lld\tobjs:\t%d\n", total_relocated,
3558		   total_relocated_objs);
3559	seq_printf(m, "\tTotal moved size:       %12lld\tobjs:\t%d\n", total_moved,
3560		   total_moved_objs);
3561	seq_printf(m, "\tTotal invalidated size: %12lld\tobjs:\t%d\n", total_invalidated,
3562		   total_invalidated_objs);
3563	seq_printf(m, "\tTotal done size:        %12lld\tobjs:\t%d\n", total_done,
3564		   total_done_objs);
3565}
3566#endif
v5.4
   1/*
   2 * Copyright 2008 Advanced Micro Devices, Inc.
   3 * Copyright 2008 Red Hat Inc.
   4 * Copyright 2009 Jerome Glisse.
   5 *
   6 * Permission is hereby granted, free of charge, to any person obtaining a
   7 * copy of this software and associated documentation files (the "Software"),
   8 * to deal in the Software without restriction, including without limitation
   9 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
  10 * and/or sell copies of the Software, and to permit persons to whom the
  11 * Software is furnished to do so, subject to the following conditions:
  12 *
  13 * The above copyright notice and this permission notice shall be included in
  14 * all copies or substantial portions of the Software.
  15 *
  16 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  17 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  18 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
  19 * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR
  20 * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
  21 * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
  22 * OTHER DEALINGS IN THE SOFTWARE.
  23 *
  24 * Authors: Dave Airlie
  25 *          Alex Deucher
  26 *          Jerome Glisse
  27 */
 
  28#include <linux/dma-fence-array.h>
  29#include <linux/interval_tree_generic.h>
  30#include <linux/idr.h>
 
  31
  32#include <drm/amdgpu_drm.h>
 
  33#include "amdgpu.h"
  34#include "amdgpu_trace.h"
  35#include "amdgpu_amdkfd.h"
  36#include "amdgpu_gmc.h"
  37#include "amdgpu_xgmi.h"
 
 
 
  38
  39/**
  40 * DOC: GPUVM
  41 *
  42 * GPUVM is similar to the legacy gart on older asics, however
  43 * rather than there being a single global gart table
  44 * for the entire GPU, there are multiple VM page tables active
  45 * at any given time.  The VM page tables can contain a mix
  46 * vram pages and system memory pages and system memory pages
  47 * can be mapped as snooped (cached system pages) or unsnooped
  48 * (uncached system pages).
  49 * Each VM has an ID associated with it and there is a page table
  50 * associated with each VMID.  When execting a command buffer,
  51 * the kernel tells the the ring what VMID to use for that command
  52 * buffer.  VMIDs are allocated dynamically as commands are submitted.
  53 * The userspace drivers maintain their own address space and the kernel
  54 * sets up their pages tables accordingly when they submit their
  55 * command buffers and a VMID is assigned.
  56 * Cayman/Trinity support up to 8 active VMs at any given time;
  57 * SI supports 16.
  58 */
  59
  60#define START(node) ((node)->start)
  61#define LAST(node) ((node)->last)
  62
  63INTERVAL_TREE_DEFINE(struct amdgpu_bo_va_mapping, rb, uint64_t, __subtree_last,
  64		     START, LAST, static, amdgpu_vm_it)
  65
  66#undef START
  67#undef LAST
  68
  69/**
  70 * struct amdgpu_prt_cb - Helper to disable partial resident texture feature from a fence callback
  71 */
  72struct amdgpu_prt_cb {
  73
  74	/**
  75	 * @adev: amdgpu device
  76	 */
  77	struct amdgpu_device *adev;
  78
  79	/**
  80	 * @cb: callback
  81	 */
  82	struct dma_fence_cb cb;
  83};
  84
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
  85/**
  86 * amdgpu_vm_level_shift - return the addr shift for each level
  87 *
  88 * @adev: amdgpu_device pointer
  89 * @level: VMPT level
  90 *
  91 * Returns:
  92 * The number of bits the pfn needs to be right shifted for a level.
  93 */
  94static unsigned amdgpu_vm_level_shift(struct amdgpu_device *adev,
  95				      unsigned level)
  96{
  97	unsigned shift = 0xff;
  98
  99	switch (level) {
 100	case AMDGPU_VM_PDB2:
 101	case AMDGPU_VM_PDB1:
 102	case AMDGPU_VM_PDB0:
 103		shift = 9 * (AMDGPU_VM_PDB0 - level) +
 104			adev->vm_manager.block_size;
 105		break;
 106	case AMDGPU_VM_PTB:
 107		shift = 0;
 108		break;
 109	default:
 110		dev_err(adev->dev, "the level%d isn't supported.\n", level);
 111	}
 112
 113	return shift;
 114}
 115
 116/**
 117 * amdgpu_vm_num_entries - return the number of entries in a PD/PT
 118 *
 119 * @adev: amdgpu_device pointer
 120 * @level: VMPT level
 121 *
 122 * Returns:
 123 * The number of entries in a page directory or page table.
 124 */
 125static unsigned amdgpu_vm_num_entries(struct amdgpu_device *adev,
 126				      unsigned level)
 127{
 128	unsigned shift = amdgpu_vm_level_shift(adev,
 129					       adev->vm_manager.root_level);
 130
 131	if (level == adev->vm_manager.root_level)
 132		/* For the root directory */
 133		return round_up(adev->vm_manager.max_pfn, 1ULL << shift) >> shift;
 
 134	else if (level != AMDGPU_VM_PTB)
 135		/* Everything in between */
 136		return 512;
 137	else
 138		/* For the page tables on the leaves */
 139		return AMDGPU_VM_PTE_COUNT(adev);
 140}
 141
 142/**
 143 * amdgpu_vm_num_ats_entries - return the number of ATS entries in the root PD
 144 *
 145 * @adev: amdgpu_device pointer
 146 *
 147 * Returns:
 148 * The number of entries in the root page directory which needs the ATS setting.
 149 */
 150static unsigned amdgpu_vm_num_ats_entries(struct amdgpu_device *adev)
 151{
 152	unsigned shift;
 153
 154	shift = amdgpu_vm_level_shift(adev, adev->vm_manager.root_level);
 155	return AMDGPU_GMC_HOLE_START >> (shift + AMDGPU_GPU_PAGE_SHIFT);
 156}
 157
 158/**
 159 * amdgpu_vm_entries_mask - the mask to get the entry number of a PD/PT
 160 *
 161 * @adev: amdgpu_device pointer
 162 * @level: VMPT level
 163 *
 164 * Returns:
 165 * The mask to extract the entry number of a PD/PT from an address.
 166 */
 167static uint32_t amdgpu_vm_entries_mask(struct amdgpu_device *adev,
 168				       unsigned int level)
 169{
 170	if (level <= adev->vm_manager.root_level)
 171		return 0xffffffff;
 172	else if (level != AMDGPU_VM_PTB)
 173		return 0x1ff;
 174	else
 175		return AMDGPU_VM_PTE_COUNT(adev) - 1;
 176}
 177
 178/**
 179 * amdgpu_vm_bo_size - returns the size of the BOs in bytes
 180 *
 181 * @adev: amdgpu_device pointer
 182 * @level: VMPT level
 183 *
 184 * Returns:
 185 * The size of the BO for a page directory or page table in bytes.
 186 */
 187static unsigned amdgpu_vm_bo_size(struct amdgpu_device *adev, unsigned level)
 188{
 189	return AMDGPU_GPU_PAGE_ALIGN(amdgpu_vm_num_entries(adev, level) * 8);
 190}
 191
 192/**
 193 * amdgpu_vm_bo_evicted - vm_bo is evicted
 194 *
 195 * @vm_bo: vm_bo which is evicted
 196 *
 197 * State for PDs/PTs and per VM BOs which are not at the location they should
 198 * be.
 199 */
 200static void amdgpu_vm_bo_evicted(struct amdgpu_vm_bo_base *vm_bo)
 201{
 202	struct amdgpu_vm *vm = vm_bo->vm;
 203	struct amdgpu_bo *bo = vm_bo->bo;
 204
 205	vm_bo->moved = true;
 206	if (bo->tbo.type == ttm_bo_type_kernel)
 207		list_move(&vm_bo->vm_status, &vm->evicted);
 208	else
 209		list_move_tail(&vm_bo->vm_status, &vm->evicted);
 210}
 211
 212/**
 213 * amdgpu_vm_bo_relocated - vm_bo is reloacted
 214 *
 215 * @vm_bo: vm_bo which is relocated
 216 *
 217 * State for PDs/PTs which needs to update their parent PD.
 218 */
 219static void amdgpu_vm_bo_relocated(struct amdgpu_vm_bo_base *vm_bo)
 220{
 221	list_move(&vm_bo->vm_status, &vm_bo->vm->relocated);
 222}
 223
 224/**
 225 * amdgpu_vm_bo_moved - vm_bo is moved
 226 *
 227 * @vm_bo: vm_bo which is moved
 228 *
 229 * State for per VM BOs which are moved, but that change is not yet reflected
 230 * in the page tables.
 231 */
 232static void amdgpu_vm_bo_moved(struct amdgpu_vm_bo_base *vm_bo)
 233{
 234	list_move(&vm_bo->vm_status, &vm_bo->vm->moved);
 235}
 236
 237/**
 238 * amdgpu_vm_bo_idle - vm_bo is idle
 239 *
 240 * @vm_bo: vm_bo which is now idle
 241 *
 242 * State for PDs/PTs and per VM BOs which have gone through the state machine
 243 * and are now idle.
 244 */
 245static void amdgpu_vm_bo_idle(struct amdgpu_vm_bo_base *vm_bo)
 246{
 247	list_move(&vm_bo->vm_status, &vm_bo->vm->idle);
 248	vm_bo->moved = false;
 249}
 250
 251/**
 252 * amdgpu_vm_bo_invalidated - vm_bo is invalidated
 253 *
 254 * @vm_bo: vm_bo which is now invalidated
 255 *
 256 * State for normal BOs which are invalidated and that change not yet reflected
 257 * in the PTs.
 258 */
 259static void amdgpu_vm_bo_invalidated(struct amdgpu_vm_bo_base *vm_bo)
 260{
 261	spin_lock(&vm_bo->vm->invalidated_lock);
 262	list_move(&vm_bo->vm_status, &vm_bo->vm->invalidated);
 263	spin_unlock(&vm_bo->vm->invalidated_lock);
 264}
 265
 266/**
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 267 * amdgpu_vm_bo_done - vm_bo is done
 268 *
 269 * @vm_bo: vm_bo which is now done
 270 *
 271 * State for normal BOs which are invalidated and that change has been updated
 272 * in the PTs.
 273 */
 274static void amdgpu_vm_bo_done(struct amdgpu_vm_bo_base *vm_bo)
 275{
 276	spin_lock(&vm_bo->vm->invalidated_lock);
 277	list_del_init(&vm_bo->vm_status);
 278	spin_unlock(&vm_bo->vm->invalidated_lock);
 279}
 280
 281/**
 282 * amdgpu_vm_bo_base_init - Adds bo to the list of bos associated with the vm
 283 *
 284 * @base: base structure for tracking BO usage in a VM
 285 * @vm: vm to which bo is to be added
 286 * @bo: amdgpu buffer object
 287 *
 288 * Initialize a bo_va_base structure and add it to the appropriate lists
 289 *
 290 */
 291static void amdgpu_vm_bo_base_init(struct amdgpu_vm_bo_base *base,
 292				   struct amdgpu_vm *vm,
 293				   struct amdgpu_bo *bo)
 294{
 295	base->vm = vm;
 296	base->bo = bo;
 297	base->next = NULL;
 298	INIT_LIST_HEAD(&base->vm_status);
 299
 300	if (!bo)
 301		return;
 302	base->next = bo->vm_bo;
 303	bo->vm_bo = base;
 304
 305	if (bo->tbo.base.resv != vm->root.base.bo->tbo.base.resv)
 306		return;
 307
 308	vm->bulk_moveable = false;
 309	if (bo->tbo.type == ttm_bo_type_kernel && bo->parent)
 310		amdgpu_vm_bo_relocated(base);
 311	else
 312		amdgpu_vm_bo_idle(base);
 313
 314	if (bo->preferred_domains &
 315	    amdgpu_mem_type_to_domain(bo->tbo.mem.mem_type))
 316		return;
 317
 318	/*
 319	 * we checked all the prerequisites, but it looks like this per vm bo
 320	 * is currently evicted. add the bo to the evicted list to make sure it
 321	 * is validated on next vm use to avoid fault.
 322	 * */
 323	amdgpu_vm_bo_evicted(base);
 324}
 325
 326/**
 327 * amdgpu_vm_pt_parent - get the parent page directory
 328 *
 329 * @pt: child page table
 330 *
 331 * Helper to get the parent entry for the child page table. NULL if we are at
 332 * the root page directory.
 333 */
 334static struct amdgpu_vm_pt *amdgpu_vm_pt_parent(struct amdgpu_vm_pt *pt)
 335{
 336	struct amdgpu_bo *parent = pt->base.bo->parent;
 337
 338	if (!parent)
 339		return NULL;
 340
 341	return container_of(parent->vm_bo, struct amdgpu_vm_pt, base);
 342}
 343
 344/**
 345 * amdgpu_vm_pt_cursor - state for for_each_amdgpu_vm_pt
 346 */
 347struct amdgpu_vm_pt_cursor {
 348	uint64_t pfn;
 349	struct amdgpu_vm_pt *parent;
 350	struct amdgpu_vm_pt *entry;
 351	unsigned level;
 352};
 353
 354/**
 355 * amdgpu_vm_pt_start - start PD/PT walk
 356 *
 357 * @adev: amdgpu_device pointer
 358 * @vm: amdgpu_vm structure
 359 * @start: start address of the walk
 360 * @cursor: state to initialize
 361 *
 362 * Initialize a amdgpu_vm_pt_cursor to start a walk.
 363 */
 364static void amdgpu_vm_pt_start(struct amdgpu_device *adev,
 365			       struct amdgpu_vm *vm, uint64_t start,
 366			       struct amdgpu_vm_pt_cursor *cursor)
 367{
 368	cursor->pfn = start;
 369	cursor->parent = NULL;
 370	cursor->entry = &vm->root;
 371	cursor->level = adev->vm_manager.root_level;
 372}
 373
 374/**
 375 * amdgpu_vm_pt_descendant - go to child node
 376 *
 377 * @adev: amdgpu_device pointer
 378 * @cursor: current state
 379 *
 380 * Walk to the child node of the current node.
 381 * Returns:
 382 * True if the walk was possible, false otherwise.
 383 */
 384static bool amdgpu_vm_pt_descendant(struct amdgpu_device *adev,
 385				    struct amdgpu_vm_pt_cursor *cursor)
 386{
 387	unsigned mask, shift, idx;
 388
 389	if (!cursor->entry->entries)
 
 390		return false;
 391
 392	BUG_ON(!cursor->entry->base.bo);
 393	mask = amdgpu_vm_entries_mask(adev, cursor->level);
 394	shift = amdgpu_vm_level_shift(adev, cursor->level);
 395
 396	++cursor->level;
 397	idx = (cursor->pfn >> shift) & mask;
 398	cursor->parent = cursor->entry;
 399	cursor->entry = &cursor->entry->entries[idx];
 400	return true;
 401}
 402
 403/**
 404 * amdgpu_vm_pt_sibling - go to sibling node
 405 *
 406 * @adev: amdgpu_device pointer
 407 * @cursor: current state
 408 *
 409 * Walk to the sibling node of the current node.
 410 * Returns:
 411 * True if the walk was possible, false otherwise.
 412 */
 413static bool amdgpu_vm_pt_sibling(struct amdgpu_device *adev,
 414				 struct amdgpu_vm_pt_cursor *cursor)
 415{
 416	unsigned shift, num_entries;
 417
 418	/* Root doesn't have a sibling */
 419	if (!cursor->parent)
 420		return false;
 421
 422	/* Go to our parents and see if we got a sibling */
 423	shift = amdgpu_vm_level_shift(adev, cursor->level - 1);
 424	num_entries = amdgpu_vm_num_entries(adev, cursor->level - 1);
 425
 426	if (cursor->entry == &cursor->parent->entries[num_entries - 1])
 427		return false;
 428
 429	cursor->pfn += 1ULL << shift;
 430	cursor->pfn &= ~((1ULL << shift) - 1);
 431	++cursor->entry;
 432	return true;
 433}
 434
 435/**
 436 * amdgpu_vm_pt_ancestor - go to parent node
 437 *
 438 * @cursor: current state
 439 *
 440 * Walk to the parent node of the current node.
 441 * Returns:
 442 * True if the walk was possible, false otherwise.
 443 */
 444static bool amdgpu_vm_pt_ancestor(struct amdgpu_vm_pt_cursor *cursor)
 445{
 446	if (!cursor->parent)
 447		return false;
 448
 449	--cursor->level;
 450	cursor->entry = cursor->parent;
 451	cursor->parent = amdgpu_vm_pt_parent(cursor->parent);
 452	return true;
 453}
 454
 455/**
 456 * amdgpu_vm_pt_next - get next PD/PT in hieratchy
 457 *
 458 * @adev: amdgpu_device pointer
 459 * @cursor: current state
 460 *
 461 * Walk the PD/PT tree to the next node.
 462 */
 463static void amdgpu_vm_pt_next(struct amdgpu_device *adev,
 464			      struct amdgpu_vm_pt_cursor *cursor)
 465{
 466	/* First try a newborn child */
 467	if (amdgpu_vm_pt_descendant(adev, cursor))
 468		return;
 469
 470	/* If that didn't worked try to find a sibling */
 471	while (!amdgpu_vm_pt_sibling(adev, cursor)) {
 472		/* No sibling, go to our parents and grandparents */
 473		if (!amdgpu_vm_pt_ancestor(cursor)) {
 474			cursor->pfn = ~0ll;
 475			return;
 476		}
 477	}
 478}
 479
 480/**
 481 * amdgpu_vm_pt_first_dfs - start a deep first search
 482 *
 483 * @adev: amdgpu_device structure
 484 * @vm: amdgpu_vm structure
 
 485 * @cursor: state to initialize
 486 *
 487 * Starts a deep first traversal of the PD/PT tree.
 488 */
 489static void amdgpu_vm_pt_first_dfs(struct amdgpu_device *adev,
 490				   struct amdgpu_vm *vm,
 491				   struct amdgpu_vm_pt_cursor *start,
 492				   struct amdgpu_vm_pt_cursor *cursor)
 493{
 494	if (start)
 495		*cursor = *start;
 496	else
 497		amdgpu_vm_pt_start(adev, vm, 0, cursor);
 498	while (amdgpu_vm_pt_descendant(adev, cursor));
 499}
 500
 501/**
 502 * amdgpu_vm_pt_continue_dfs - check if the deep first search should continue
 503 *
 504 * @start: starting point for the search
 505 * @entry: current entry
 506 *
 507 * Returns:
 508 * True when the search should continue, false otherwise.
 509 */
 510static bool amdgpu_vm_pt_continue_dfs(struct amdgpu_vm_pt_cursor *start,
 511				      struct amdgpu_vm_pt *entry)
 512{
 513	return entry && (!start || entry != start->entry);
 514}
 515
 516/**
 517 * amdgpu_vm_pt_next_dfs - get the next node for a deep first search
 518 *
 519 * @adev: amdgpu_device structure
 520 * @cursor: current state
 521 *
 522 * Move the cursor to the next node in a deep first search.
 523 */
 524static void amdgpu_vm_pt_next_dfs(struct amdgpu_device *adev,
 525				  struct amdgpu_vm_pt_cursor *cursor)
 526{
 527	if (!cursor->entry)
 528		return;
 529
 530	if (!cursor->parent)
 531		cursor->entry = NULL;
 532	else if (amdgpu_vm_pt_sibling(adev, cursor))
 533		while (amdgpu_vm_pt_descendant(adev, cursor));
 534	else
 535		amdgpu_vm_pt_ancestor(cursor);
 536}
 537
 538/**
 539 * for_each_amdgpu_vm_pt_dfs_safe - safe deep first search of all PDs/PTs
 540 */
 541#define for_each_amdgpu_vm_pt_dfs_safe(adev, vm, start, cursor, entry)		\
 542	for (amdgpu_vm_pt_first_dfs((adev), (vm), (start), &(cursor)),		\
 543	     (entry) = (cursor).entry, amdgpu_vm_pt_next_dfs((adev), &(cursor));\
 544	     amdgpu_vm_pt_continue_dfs((start), (entry));			\
 545	     (entry) = (cursor).entry, amdgpu_vm_pt_next_dfs((adev), &(cursor)))
 546
 547/**
 548 * amdgpu_vm_get_pd_bo - add the VM PD to a validation list
 549 *
 550 * @vm: vm providing the BOs
 551 * @validated: head of validation list
 552 * @entry: entry to add
 553 *
 554 * Add the page directory to the list of BOs to
 555 * validate for command submission.
 556 */
 557void amdgpu_vm_get_pd_bo(struct amdgpu_vm *vm,
 558			 struct list_head *validated,
 559			 struct amdgpu_bo_list_entry *entry)
 560{
 561	entry->priority = 0;
 562	entry->tv.bo = &vm->root.base.bo->tbo;
 563	/* One for the VM updates, one for TTM and one for the CS job */
 564	entry->tv.num_shared = 3;
 565	entry->user_pages = NULL;
 566	list_add(&entry->tv.head, validated);
 567}
 568
 
 
 
 
 
 
 
 
 569void amdgpu_vm_del_from_lru_notify(struct ttm_buffer_object *bo)
 570{
 571	struct amdgpu_bo *abo;
 572	struct amdgpu_vm_bo_base *bo_base;
 573
 574	if (!amdgpu_bo_is_amdgpu_bo(bo))
 575		return;
 576
 577	if (bo->mem.placement & TTM_PL_FLAG_NO_EVICT)
 578		return;
 579
 580	abo = ttm_to_amdgpu_bo(bo);
 581	if (!abo->parent)
 582		return;
 583	for (bo_base = abo->vm_bo; bo_base; bo_base = bo_base->next) {
 584		struct amdgpu_vm *vm = bo_base->vm;
 585
 586		if (abo->tbo.base.resv == vm->root.base.bo->tbo.base.resv)
 587			vm->bulk_moveable = false;
 588	}
 589
 590}
 591/**
 592 * amdgpu_vm_move_to_lru_tail - move all BOs to the end of LRU
 593 *
 594 * @adev: amdgpu device pointer
 595 * @vm: vm providing the BOs
 596 *
 597 * Move all BOs to the end of LRU and remember their positions to put them
 598 * together.
 599 */
 600void amdgpu_vm_move_to_lru_tail(struct amdgpu_device *adev,
 601				struct amdgpu_vm *vm)
 602{
 603	struct ttm_bo_global *glob = adev->mman.bdev.glob;
 604	struct amdgpu_vm_bo_base *bo_base;
 605
 606	if (vm->bulk_moveable) {
 607		spin_lock(&glob->lru_lock);
 608		ttm_bo_bulk_move_lru_tail(&vm->lru_bulk_move);
 609		spin_unlock(&glob->lru_lock);
 610		return;
 611	}
 612
 613	memset(&vm->lru_bulk_move, 0, sizeof(vm->lru_bulk_move));
 614
 615	spin_lock(&glob->lru_lock);
 616	list_for_each_entry(bo_base, &vm->idle, vm_status) {
 617		struct amdgpu_bo *bo = bo_base->bo;
 
 618
 619		if (!bo->parent)
 620			continue;
 621
 622		ttm_bo_move_to_lru_tail(&bo->tbo, &vm->lru_bulk_move);
 623		if (bo->shadow)
 624			ttm_bo_move_to_lru_tail(&bo->shadow->tbo,
 
 
 625						&vm->lru_bulk_move);
 626	}
 627	spin_unlock(&glob->lru_lock);
 628
 629	vm->bulk_moveable = true;
 630}
 631
 632/**
 633 * amdgpu_vm_validate_pt_bos - validate the page table BOs
 634 *
 635 * @adev: amdgpu device pointer
 636 * @vm: vm providing the BOs
 637 * @validate: callback to do the validation
 638 * @param: parameter for the validation callback
 639 *
 640 * Validate the page table BOs on command submission if neccessary.
 641 *
 642 * Returns:
 643 * Validation result.
 644 */
 645int amdgpu_vm_validate_pt_bos(struct amdgpu_device *adev, struct amdgpu_vm *vm,
 646			      int (*validate)(void *p, struct amdgpu_bo *bo),
 647			      void *param)
 648{
 649	struct amdgpu_vm_bo_base *bo_base, *tmp;
 650	int r = 0;
 651
 652	vm->bulk_moveable &= list_empty(&vm->evicted);
 653
 654	list_for_each_entry_safe(bo_base, tmp, &vm->evicted, vm_status) {
 655		struct amdgpu_bo *bo = bo_base->bo;
 
 656
 657		r = validate(param, bo);
 658		if (r)
 659			break;
 
 
 
 
 
 660
 661		if (bo->tbo.type != ttm_bo_type_kernel) {
 662			amdgpu_vm_bo_moved(bo_base);
 663		} else {
 664			vm->update_funcs->map_table(bo);
 665			if (bo->parent)
 666				amdgpu_vm_bo_relocated(bo_base);
 667			else
 668				amdgpu_vm_bo_idle(bo_base);
 669		}
 670	}
 671
 672	return r;
 
 
 
 
 673}
 674
 675/**
 676 * amdgpu_vm_ready - check VM is ready for updates
 677 *
 678 * @vm: VM to check
 679 *
 680 * Check if all VM PDs/PTs are ready for updates
 681 *
 682 * Returns:
 683 * True if eviction list is empty.
 684 */
 685bool amdgpu_vm_ready(struct amdgpu_vm *vm)
 686{
 687	return list_empty(&vm->evicted);
 688}
 689
 690/**
 691 * amdgpu_vm_clear_bo - initially clear the PDs/PTs
 692 *
 693 * @adev: amdgpu_device pointer
 694 * @vm: VM to clear BO from
 695 * @bo: BO to clear
 
 696 *
 697 * Root PD needs to be reserved when calling this.
 698 *
 699 * Returns:
 700 * 0 on success, errno otherwise.
 701 */
 702static int amdgpu_vm_clear_bo(struct amdgpu_device *adev,
 703			      struct amdgpu_vm *vm,
 704			      struct amdgpu_bo *bo)
 
 705{
 706	struct ttm_operation_ctx ctx = { true, false };
 707	unsigned level = adev->vm_manager.root_level;
 708	struct amdgpu_vm_update_params params;
 709	struct amdgpu_bo *ancestor = bo;
 
 710	unsigned entries, ats_entries;
 711	uint64_t addr;
 712	int r;
 713
 714	/* Figure out our place in the hierarchy */
 715	if (ancestor->parent) {
 716		++level;
 717		while (ancestor->parent->parent) {
 718			++level;
 719			ancestor = ancestor->parent;
 720		}
 721	}
 722
 723	entries = amdgpu_bo_size(bo) / 8;
 724	if (!vm->pte_support_ats) {
 725		ats_entries = 0;
 726
 727	} else if (!bo->parent) {
 728		ats_entries = amdgpu_vm_num_ats_entries(adev);
 729		ats_entries = min(ats_entries, entries);
 730		entries -= ats_entries;
 731
 732	} else {
 733		struct amdgpu_vm_pt *pt;
 734
 735		pt = container_of(ancestor->vm_bo, struct amdgpu_vm_pt, base);
 736		ats_entries = amdgpu_vm_num_ats_entries(adev);
 737		if ((pt - vm->root.entries) >= ats_entries) {
 738			ats_entries = 0;
 739		} else {
 740			ats_entries = entries;
 741			entries = 0;
 742		}
 743	}
 744
 745	r = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
 746	if (r)
 747		return r;
 748
 749	if (bo->shadow) {
 750		r = ttm_bo_validate(&bo->shadow->tbo, &bo->shadow->placement,
 751				    &ctx);
 
 752		if (r)
 753			return r;
 754	}
 755
 756	r = vm->update_funcs->map_table(bo);
 757	if (r)
 758		return r;
 759
 760	memset(&params, 0, sizeof(params));
 761	params.adev = adev;
 762	params.vm = vm;
 
 763
 764	r = vm->update_funcs->prepare(&params, AMDGPU_FENCE_OWNER_KFD, NULL);
 765	if (r)
 766		return r;
 767
 768	addr = 0;
 769	if (ats_entries) {
 770		uint64_t value = 0, flags;
 771
 772		flags = AMDGPU_PTE_DEFAULT_ATC;
 773		if (level != AMDGPU_VM_PTB) {
 774			/* Handle leaf PDEs as PTEs */
 775			flags |= AMDGPU_PDE_PTE;
 776			amdgpu_gmc_get_vm_pde(adev, level, &value, &flags);
 777		}
 778
 779		r = vm->update_funcs->update(&params, bo, addr, 0, ats_entries,
 780					     value, flags);
 781		if (r)
 782			return r;
 783
 784		addr += ats_entries * 8;
 785	}
 786
 787	if (entries) {
 788		uint64_t value = 0, flags = 0;
 789
 790		if (adev->asic_type >= CHIP_VEGA10) {
 791			if (level != AMDGPU_VM_PTB) {
 792				/* Handle leaf PDEs as PTEs */
 793				flags |= AMDGPU_PDE_PTE;
 794				amdgpu_gmc_get_vm_pde(adev, level,
 795						      &value, &flags);
 796			} else {
 797				/* Workaround for fault priority problem on GMC9 */
 798				flags = AMDGPU_PTE_EXECUTABLE;
 799			}
 800		}
 801
 802		r = vm->update_funcs->update(&params, bo, addr, 0, entries,
 803					     value, flags);
 804		if (r)
 805			return r;
 806	}
 807
 808	return vm->update_funcs->commit(&params, NULL);
 809}
 810
 811/**
 812 * amdgpu_vm_bo_param - fill in parameters for PD/PT allocation
 813 *
 814 * @adev: amdgpu_device pointer
 815 * @vm: requesting vm
 816 * @bp: resulting BO allocation parameters
 
 
 817 */
 818static void amdgpu_vm_bo_param(struct amdgpu_device *adev, struct amdgpu_vm *vm,
 819			       int level, struct amdgpu_bo_param *bp)
 
 
 820{
 821	memset(bp, 0, sizeof(*bp));
 
 
 
 
 822
 823	bp->size = amdgpu_vm_bo_size(adev, level);
 824	bp->byte_align = AMDGPU_GPU_PAGE_SIZE;
 825	bp->domain = AMDGPU_GEM_DOMAIN_VRAM;
 826	bp->domain = amdgpu_bo_get_preferred_pin_domain(adev, bp->domain);
 827	bp->flags = AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS |
 
 
 828		AMDGPU_GEM_CREATE_CPU_GTT_USWC;
 
 
 
 
 
 
 
 
 829	if (vm->use_cpu_for_update)
 830		bp->flags |= AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED;
 831	else if (!vm->root.base.bo || vm->root.base.bo->shadow)
 832		bp->flags |= AMDGPU_GEM_CREATE_SHADOW;
 833	bp->type = ttm_bo_type_kernel;
 834	if (vm->root.base.bo)
 835		bp->resv = vm->root.base.bo->tbo.base.resv;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 836}
 837
 838/**
 839 * amdgpu_vm_alloc_pts - Allocate a specific page table
 840 *
 841 * @adev: amdgpu_device pointer
 842 * @vm: VM to allocate page tables for
 843 * @cursor: Which page table to allocate
 
 844 *
 845 * Make sure a specific page table or directory is allocated.
 846 *
 847 * Returns:
 848 * 1 if page table needed to be allocated, 0 if page table was already
 849 * allocated, negative errno if an error occurred.
 850 */
 851static int amdgpu_vm_alloc_pts(struct amdgpu_device *adev,
 852			       struct amdgpu_vm *vm,
 853			       struct amdgpu_vm_pt_cursor *cursor)
 
 854{
 855	struct amdgpu_vm_pt *entry = cursor->entry;
 856	struct amdgpu_bo_param bp;
 857	struct amdgpu_bo *pt;
 858	int r;
 859
 860	if (cursor->level < AMDGPU_VM_PTB && !entry->entries) {
 861		unsigned num_entries;
 862
 863		num_entries = amdgpu_vm_num_entries(adev, cursor->level);
 864		entry->entries = kvmalloc_array(num_entries,
 865						sizeof(*entry->entries),
 866						GFP_KERNEL | __GFP_ZERO);
 867		if (!entry->entries)
 868			return -ENOMEM;
 869	}
 870
 871	if (entry->base.bo)
 872		return 0;
 873
 874	amdgpu_vm_bo_param(adev, vm, cursor->level, &bp);
 875
 876	r = amdgpu_bo_create(adev, &bp, &pt);
 877	if (r)
 878		return r;
 879
 880	/* Keep a reference to the root directory to avoid
 881	 * freeing them up in the wrong order.
 882	 */
 883	pt->parent = amdgpu_bo_ref(cursor->parent->base.bo);
 884	amdgpu_vm_bo_base_init(&entry->base, vm, pt);
 885
 886	r = amdgpu_vm_clear_bo(adev, vm, pt);
 887	if (r)
 888		goto error_free_pt;
 889
 890	return 0;
 891
 892error_free_pt:
 893	amdgpu_bo_unref(&pt->shadow);
 894	amdgpu_bo_unref(&pt);
 895	return r;
 896}
 897
 898/**
 899 * amdgpu_vm_free_table - fre one PD/PT
 900 *
 901 * @entry: PDE to free
 902 */
 903static void amdgpu_vm_free_table(struct amdgpu_vm_pt *entry)
 904{
 905	if (entry->base.bo) {
 906		entry->base.bo->vm_bo = NULL;
 907		list_del(&entry->base.vm_status);
 908		amdgpu_bo_unref(&entry->base.bo->shadow);
 909		amdgpu_bo_unref(&entry->base.bo);
 910	}
 911	kvfree(entry->entries);
 912	entry->entries = NULL;
 
 913}
 914
 915/**
 916 * amdgpu_vm_free_pts - free PD/PT levels
 917 *
 918 * @adev: amdgpu device structure
 919 * @vm: amdgpu vm structure
 920 * @start: optional cursor where to start freeing PDs/PTs
 921 *
 922 * Free the page directory or page table level and all sub levels.
 923 */
 924static void amdgpu_vm_free_pts(struct amdgpu_device *adev,
 925			       struct amdgpu_vm *vm,
 926			       struct amdgpu_vm_pt_cursor *start)
 927{
 928	struct amdgpu_vm_pt_cursor cursor;
 929	struct amdgpu_vm_pt *entry;
 930
 931	vm->bulk_moveable = false;
 932
 933	for_each_amdgpu_vm_pt_dfs_safe(adev, vm, start, cursor, entry)
 934		amdgpu_vm_free_table(entry);
 935
 936	if (start)
 937		amdgpu_vm_free_table(start->entry);
 938}
 939
 940/**
 941 * amdgpu_vm_check_compute_bug - check whether asic has compute vm bug
 942 *
 943 * @adev: amdgpu_device pointer
 944 */
 945void amdgpu_vm_check_compute_bug(struct amdgpu_device *adev)
 946{
 947	const struct amdgpu_ip_block *ip_block;
 948	bool has_compute_vm_bug;
 949	struct amdgpu_ring *ring;
 950	int i;
 951
 952	has_compute_vm_bug = false;
 953
 954	ip_block = amdgpu_device_ip_get_ip_block(adev, AMD_IP_BLOCK_TYPE_GFX);
 955	if (ip_block) {
 956		/* Compute has a VM bug for GFX version < 7.
 957		   Compute has a VM bug for GFX 8 MEC firmware version < 673.*/
 958		if (ip_block->version->major <= 7)
 959			has_compute_vm_bug = true;
 960		else if (ip_block->version->major == 8)
 961			if (adev->gfx.mec_fw_version < 673)
 962				has_compute_vm_bug = true;
 963	}
 964
 965	for (i = 0; i < adev->num_rings; i++) {
 966		ring = adev->rings[i];
 967		if (ring->funcs->type == AMDGPU_RING_TYPE_COMPUTE)
 968			/* only compute rings */
 969			ring->has_compute_vm_bug = has_compute_vm_bug;
 970		else
 971			ring->has_compute_vm_bug = false;
 972	}
 973}
 974
 975/**
 976 * amdgpu_vm_need_pipeline_sync - Check if pipe sync is needed for job.
 977 *
 978 * @ring: ring on which the job will be submitted
 979 * @job: job to submit
 980 *
 981 * Returns:
 982 * True if sync is needed.
 983 */
 984bool amdgpu_vm_need_pipeline_sync(struct amdgpu_ring *ring,
 985				  struct amdgpu_job *job)
 986{
 987	struct amdgpu_device *adev = ring->adev;
 988	unsigned vmhub = ring->funcs->vmhub;
 989	struct amdgpu_vmid_mgr *id_mgr = &adev->vm_manager.id_mgr[vmhub];
 990	struct amdgpu_vmid *id;
 991	bool gds_switch_needed;
 992	bool vm_flush_needed = job->vm_needs_flush || ring->has_compute_vm_bug;
 993
 994	if (job->vmid == 0)
 995		return false;
 996	id = &id_mgr->ids[job->vmid];
 997	gds_switch_needed = ring->funcs->emit_gds_switch && (
 998		id->gds_base != job->gds_base ||
 999		id->gds_size != job->gds_size ||
1000		id->gws_base != job->gws_base ||
1001		id->gws_size != job->gws_size ||
1002		id->oa_base != job->oa_base ||
1003		id->oa_size != job->oa_size);
1004
1005	if (amdgpu_vmid_had_gpu_reset(adev, id))
1006		return true;
1007
1008	return vm_flush_needed || gds_switch_needed;
1009}
1010
1011/**
1012 * amdgpu_vm_flush - hardware flush the vm
1013 *
1014 * @ring: ring to use for flush
1015 * @job:  related job
1016 * @need_pipe_sync: is pipe sync needed
1017 *
1018 * Emit a VM flush when it is necessary.
1019 *
1020 * Returns:
1021 * 0 on success, errno otherwise.
1022 */
1023int amdgpu_vm_flush(struct amdgpu_ring *ring, struct amdgpu_job *job, bool need_pipe_sync)
 
1024{
1025	struct amdgpu_device *adev = ring->adev;
1026	unsigned vmhub = ring->funcs->vmhub;
1027	struct amdgpu_vmid_mgr *id_mgr = &adev->vm_manager.id_mgr[vmhub];
1028	struct amdgpu_vmid *id = &id_mgr->ids[job->vmid];
1029	bool gds_switch_needed = ring->funcs->emit_gds_switch && (
1030		id->gds_base != job->gds_base ||
1031		id->gds_size != job->gds_size ||
1032		id->gws_base != job->gws_base ||
1033		id->gws_size != job->gws_size ||
1034		id->oa_base != job->oa_base ||
1035		id->oa_size != job->oa_size);
1036	bool vm_flush_needed = job->vm_needs_flush;
1037	bool pasid_mapping_needed = id->pasid != job->pasid ||
1038		!id->pasid_mapping ||
1039		!dma_fence_is_signaled(id->pasid_mapping);
1040	struct dma_fence *fence = NULL;
 
1041	unsigned patch_offset = 0;
 
1042	int r;
1043
 
 
 
1044	if (amdgpu_vmid_had_gpu_reset(adev, id)) {
1045		gds_switch_needed = true;
1046		vm_flush_needed = true;
1047		pasid_mapping_needed = true;
1048	}
1049
 
 
 
 
 
 
1050	gds_switch_needed &= !!ring->funcs->emit_gds_switch;
1051	vm_flush_needed &= !!ring->funcs->emit_vm_flush  &&
1052			job->vm_pd_addr != AMDGPU_BO_INVALID_OFFSET;
1053	pasid_mapping_needed &= adev->gmc.gmc_funcs->emit_pasid_mapping &&
1054		ring->funcs->emit_wreg;
1055
1056	if (!vm_flush_needed && !gds_switch_needed && !need_pipe_sync)
1057		return 0;
1058
1059	if (ring->funcs->init_cond_exec)
1060		patch_offset = amdgpu_ring_init_cond_exec(ring);
1061
1062	if (need_pipe_sync)
1063		amdgpu_ring_emit_pipeline_sync(ring);
1064
1065	if (vm_flush_needed) {
1066		trace_amdgpu_vm_flush(ring, job->vmid, job->vm_pd_addr);
1067		amdgpu_ring_emit_vm_flush(ring, job->vmid, job->vm_pd_addr);
1068	}
1069
1070	if (pasid_mapping_needed)
1071		amdgpu_gmc_emit_pasid_mapping(ring, job->vmid, job->pasid);
1072
1073	if (vm_flush_needed || pasid_mapping_needed) {
1074		r = amdgpu_fence_emit(ring, &fence, 0);
1075		if (r)
1076			return r;
1077	}
1078
1079	if (vm_flush_needed) {
1080		mutex_lock(&id_mgr->lock);
1081		dma_fence_put(id->last_flush);
1082		id->last_flush = dma_fence_get(fence);
1083		id->current_gpu_reset_count =
1084			atomic_read(&adev->gpu_reset_counter);
1085		mutex_unlock(&id_mgr->lock);
1086	}
1087
1088	if (pasid_mapping_needed) {
 
1089		id->pasid = job->pasid;
1090		dma_fence_put(id->pasid_mapping);
1091		id->pasid_mapping = dma_fence_get(fence);
 
1092	}
1093	dma_fence_put(fence);
1094
1095	if (ring->funcs->emit_gds_switch && gds_switch_needed) {
1096		id->gds_base = job->gds_base;
1097		id->gds_size = job->gds_size;
1098		id->gws_base = job->gws_base;
1099		id->gws_size = job->gws_size;
1100		id->oa_base = job->oa_base;
1101		id->oa_size = job->oa_size;
1102		amdgpu_ring_emit_gds_switch(ring, job->vmid, job->gds_base,
1103					    job->gds_size, job->gws_base,
1104					    job->gws_size, job->oa_base,
1105					    job->oa_size);
1106	}
1107
1108	if (ring->funcs->patch_cond_exec)
1109		amdgpu_ring_patch_cond_exec(ring, patch_offset);
1110
1111	/* the double SWITCH_BUFFER here *cannot* be skipped by COND_EXEC */
1112	if (ring->funcs->emit_switch_buffer) {
1113		amdgpu_ring_emit_switch_buffer(ring);
1114		amdgpu_ring_emit_switch_buffer(ring);
1115	}
1116	return 0;
1117}
1118
1119/**
1120 * amdgpu_vm_bo_find - find the bo_va for a specific vm & bo
1121 *
1122 * @vm: requested vm
1123 * @bo: requested buffer object
1124 *
1125 * Find @bo inside the requested vm.
1126 * Search inside the @bos vm list for the requested vm
1127 * Returns the found bo_va or NULL if none is found
1128 *
1129 * Object has to be reserved!
1130 *
1131 * Returns:
1132 * Found bo_va or NULL.
1133 */
1134struct amdgpu_bo_va *amdgpu_vm_bo_find(struct amdgpu_vm *vm,
1135				       struct amdgpu_bo *bo)
1136{
1137	struct amdgpu_vm_bo_base *base;
1138
1139	for (base = bo->vm_bo; base; base = base->next) {
1140		if (base->vm != vm)
1141			continue;
1142
1143		return container_of(base, struct amdgpu_bo_va, base);
1144	}
1145	return NULL;
1146}
1147
1148/**
1149 * amdgpu_vm_map_gart - Resolve gart mapping of addr
1150 *
1151 * @pages_addr: optional DMA address to use for lookup
1152 * @addr: the unmapped addr
1153 *
1154 * Look up the physical address of the page that the pte resolves
1155 * to.
1156 *
1157 * Returns:
1158 * The pointer for the page table entry.
1159 */
1160uint64_t amdgpu_vm_map_gart(const dma_addr_t *pages_addr, uint64_t addr)
1161{
1162	uint64_t result;
1163
1164	/* page table offset */
1165	result = pages_addr[addr >> PAGE_SHIFT];
1166
1167	/* in case cpu page size != gpu page size*/
1168	result |= addr & (~PAGE_MASK);
1169
1170	result &= 0xFFFFFFFFFFFFF000ULL;
1171
1172	return result;
1173}
1174
1175/*
1176 * amdgpu_vm_update_pde - update a single level in the hierarchy
1177 *
1178 * @param: parameters for the update
1179 * @vm: requested vm
1180 * @entry: entry to update
1181 *
1182 * Makes sure the requested entry in parent is up to date.
1183 */
1184static int amdgpu_vm_update_pde(struct amdgpu_vm_update_params *params,
1185				struct amdgpu_vm *vm,
1186				struct amdgpu_vm_pt *entry)
1187{
1188	struct amdgpu_vm_pt *parent = amdgpu_vm_pt_parent(entry);
1189	struct amdgpu_bo *bo = parent->base.bo, *pbo;
1190	uint64_t pde, pt, flags;
1191	unsigned level;
1192
1193	for (level = 0, pbo = bo->parent; pbo; ++level)
1194		pbo = pbo->parent;
1195
1196	level += params->adev->vm_manager.root_level;
1197	amdgpu_gmc_get_pde_for_bo(entry->base.bo, level, &pt, &flags);
1198	pde = (entry - parent->entries) * 8;
1199	return vm->update_funcs->update(params, bo, pde, pt, 1, 0, flags);
 
1200}
1201
1202/*
1203 * amdgpu_vm_invalidate_pds - mark all PDs as invalid
1204 *
1205 * @adev: amdgpu_device pointer
1206 * @vm: related vm
1207 *
1208 * Mark all PD level as invalid after an error.
1209 */
1210static void amdgpu_vm_invalidate_pds(struct amdgpu_device *adev,
1211				     struct amdgpu_vm *vm)
1212{
1213	struct amdgpu_vm_pt_cursor cursor;
1214	struct amdgpu_vm_pt *entry;
1215
1216	for_each_amdgpu_vm_pt_dfs_safe(adev, vm, NULL, cursor, entry)
1217		if (entry->base.bo && !entry->base.moved)
1218			amdgpu_vm_bo_relocated(&entry->base);
1219}
1220
1221/*
1222 * amdgpu_vm_update_directories - make sure that all directories are valid
1223 *
1224 * @adev: amdgpu_device pointer
1225 * @vm: requested vm
 
1226 *
1227 * Makes sure all directories are up to date.
1228 *
1229 * Returns:
1230 * 0 for success, error for failure.
1231 */
1232int amdgpu_vm_update_directories(struct amdgpu_device *adev,
1233				 struct amdgpu_vm *vm)
1234{
1235	struct amdgpu_vm_update_params params;
1236	int r;
1237
1238	if (list_empty(&vm->relocated))
1239		return 0;
1240
1241	memset(&params, 0, sizeof(params));
1242	params.adev = adev;
1243	params.vm = vm;
 
1244
1245	r = vm->update_funcs->prepare(&params, AMDGPU_FENCE_OWNER_VM, NULL);
1246	if (r)
1247		return r;
1248
1249	while (!list_empty(&vm->relocated)) {
1250		struct amdgpu_vm_pt *entry;
1251
1252		entry = list_first_entry(&vm->relocated, struct amdgpu_vm_pt,
1253					 base.vm_status);
1254		amdgpu_vm_bo_idle(&entry->base);
 
1255
1256		r = amdgpu_vm_update_pde(&params, vm, entry);
1257		if (r)
1258			goto error;
1259	}
1260
1261	r = vm->update_funcs->commit(&params, &vm->last_update);
1262	if (r)
1263		goto error;
1264	return 0;
1265
1266error:
1267	amdgpu_vm_invalidate_pds(adev, vm);
1268	return r;
1269}
1270
1271/**
1272 * amdgpu_vm_update_flags - figure out flags for PTE updates
1273 *
1274 * Make sure to set the right flags for the PTEs at the desired level.
1275 */
1276static void amdgpu_vm_update_flags(struct amdgpu_vm_update_params *params,
1277				   struct amdgpu_bo *bo, unsigned level,
1278				   uint64_t pe, uint64_t addr,
1279				   unsigned count, uint32_t incr,
1280				   uint64_t flags)
1281
1282{
1283	if (level != AMDGPU_VM_PTB) {
1284		flags |= AMDGPU_PDE_PTE;
1285		amdgpu_gmc_get_vm_pde(params->adev, level, &addr, &flags);
1286
1287	} else if (params->adev->asic_type >= CHIP_VEGA10 &&
1288		   !(flags & AMDGPU_PTE_VALID) &&
1289		   !(flags & AMDGPU_PTE_PRT)) {
1290
1291		/* Workaround for fault priority problem on GMC9 */
1292		flags |= AMDGPU_PTE_EXECUTABLE;
1293	}
1294
1295	params->vm->update_funcs->update(params, bo, pe, addr, count, incr,
1296					 flags);
1297}
1298
1299/**
1300 * amdgpu_vm_fragment - get fragment for PTEs
1301 *
1302 * @params: see amdgpu_vm_update_params definition
1303 * @start: first PTE to handle
1304 * @end: last PTE to handle
1305 * @flags: hw mapping flags
1306 * @frag: resulting fragment size
1307 * @frag_end: end of this fragment
1308 *
1309 * Returns the first possible fragment for the start and end address.
1310 */
1311static void amdgpu_vm_fragment(struct amdgpu_vm_update_params *params,
1312			       uint64_t start, uint64_t end, uint64_t flags,
1313			       unsigned int *frag, uint64_t *frag_end)
1314{
1315	/**
1316	 * The MC L1 TLB supports variable sized pages, based on a fragment
1317	 * field in the PTE. When this field is set to a non-zero value, page
1318	 * granularity is increased from 4KB to (1 << (12 + frag)). The PTE
1319	 * flags are considered valid for all PTEs within the fragment range
1320	 * and corresponding mappings are assumed to be physically contiguous.
1321	 *
1322	 * The L1 TLB can store a single PTE for the whole fragment,
1323	 * significantly increasing the space available for translation
1324	 * caching. This leads to large improvements in throughput when the
1325	 * TLB is under pressure.
1326	 *
1327	 * The L2 TLB distributes small and large fragments into two
1328	 * asymmetric partitions. The large fragment cache is significantly
1329	 * larger. Thus, we try to use large fragments wherever possible.
1330	 * Userspace can support this by aligning virtual base address and
1331	 * allocation size to the fragment size.
1332	 *
1333	 * Starting with Vega10 the fragment size only controls the L1. The L2
1334	 * is now directly feed with small/huge/giant pages from the walker.
1335	 */
1336	unsigned max_frag;
1337
1338	if (params->adev->asic_type < CHIP_VEGA10)
1339		max_frag = params->adev->vm_manager.fragment_size;
1340	else
1341		max_frag = 31;
1342
1343	/* system pages are non continuously */
1344	if (params->pages_addr) {
1345		*frag = 0;
1346		*frag_end = end;
1347		return;
1348	}
1349
1350	/* This intentionally wraps around if no bit is set */
1351	*frag = min((unsigned)ffs(start) - 1, (unsigned)fls64(end - start) - 1);
1352	if (*frag >= max_frag) {
1353		*frag = max_frag;
1354		*frag_end = end & ~((1ULL << max_frag) - 1);
1355	} else {
1356		*frag_end = start + (1 << *frag);
1357	}
1358}
1359
1360/**
1361 * amdgpu_vm_update_ptes - make sure that page tables are valid
1362 *
1363 * @params: see amdgpu_vm_update_params definition
1364 * @start: start of GPU address range
1365 * @end: end of GPU address range
1366 * @dst: destination address to map to, the next dst inside the function
1367 * @flags: mapping flags
1368 *
1369 * Update the page tables in the range @start - @end.
1370 *
1371 * Returns:
1372 * 0 for success, -EINVAL for failure.
1373 */
1374static int amdgpu_vm_update_ptes(struct amdgpu_vm_update_params *params,
1375				 uint64_t start, uint64_t end,
1376				 uint64_t dst, uint64_t flags)
1377{
1378	struct amdgpu_device *adev = params->adev;
1379	struct amdgpu_vm_pt_cursor cursor;
1380	uint64_t frag_start = start, frag_end;
1381	unsigned int frag;
1382	int r;
1383
1384	/* figure out the initial fragment */
1385	amdgpu_vm_fragment(params, frag_start, end, flags, &frag, &frag_end);
1386
1387	/* walk over the address space and update the PTs */
1388	amdgpu_vm_pt_start(adev, params->vm, start, &cursor);
1389	while (cursor.pfn < end) {
1390		unsigned shift, parent_shift, mask;
1391		uint64_t incr, entry_end, pe_start;
1392		struct amdgpu_bo *pt;
1393
1394		r = amdgpu_vm_alloc_pts(params->adev, params->vm, &cursor);
1395		if (r)
1396			return r;
1397
1398		pt = cursor.entry->base.bo;
1399
1400		/* The root level can't be a huge page */
1401		if (cursor.level == adev->vm_manager.root_level) {
1402			if (!amdgpu_vm_pt_descendant(adev, &cursor))
1403				return -ENOENT;
1404			continue;
1405		}
1406
1407		shift = amdgpu_vm_level_shift(adev, cursor.level);
1408		parent_shift = amdgpu_vm_level_shift(adev, cursor.level - 1);
1409		if (adev->asic_type < CHIP_VEGA10 &&
1410		    (flags & AMDGPU_PTE_VALID)) {
 
 
 
 
1411			/* No huge page support before GMC v9 */
1412			if (cursor.level != AMDGPU_VM_PTB) {
1413				if (!amdgpu_vm_pt_descendant(adev, &cursor))
1414					return -ENOENT;
1415				continue;
1416			}
1417		} else if (frag < shift) {
1418			/* We can't use this level when the fragment size is
1419			 * smaller than the address shift. Go to the next
1420			 * child entry and try again.
1421			 */
1422			if (!amdgpu_vm_pt_descendant(adev, &cursor))
1423				return -ENOENT;
1424			continue;
1425		} else if (frag >= parent_shift &&
1426			   cursor.level - 1 != adev->vm_manager.root_level) {
1427			/* If the fragment size is even larger than the parent
1428			 * shift we should go up one level and check it again
1429			 * unless one level up is the root level.
1430			 */
1431			if (!amdgpu_vm_pt_ancestor(&cursor))
 
 
 
 
 
 
 
 
1432				return -ENOENT;
1433			continue;
 
 
 
 
 
 
 
 
 
 
1434		}
1435
1436		/* Looks good so far, calculate parameters for the update */
1437		incr = (uint64_t)AMDGPU_GPU_PAGE_SIZE << shift;
1438		mask = amdgpu_vm_entries_mask(adev, cursor.level);
1439		pe_start = ((cursor.pfn >> shift) & mask) * 8;
1440		entry_end = (uint64_t)(mask + 1) << shift;
1441		entry_end += cursor.pfn & ~(entry_end - 1);
1442		entry_end = min(entry_end, end);
1443
1444		do {
 
1445			uint64_t upd_end = min(entry_end, frag_end);
1446			unsigned nptes = (upd_end - frag_start) >> shift;
 
 
 
 
 
 
1447
1448			amdgpu_vm_update_flags(params, pt, cursor.level,
1449					       pe_start, dst, nptes, incr,
1450					       flags | AMDGPU_PTE_FRAG(frag));
 
 
 
 
1451
1452			pe_start += nptes * 8;
1453			dst += (uint64_t)nptes * AMDGPU_GPU_PAGE_SIZE << shift;
1454
1455			frag_start = upd_end;
1456			if (frag_start >= frag_end) {
1457				/* figure out the next fragment */
1458				amdgpu_vm_fragment(params, frag_start, end,
1459						   flags, &frag, &frag_end);
1460				if (frag < shift)
1461					break;
1462			}
1463		} while (frag_start < entry_end);
1464
1465		if (amdgpu_vm_pt_descendant(adev, &cursor)) {
1466			/* Free all child entries */
 
 
 
 
 
1467			while (cursor.pfn < frag_start) {
1468				amdgpu_vm_free_pts(adev, params->vm, &cursor);
 
 
 
 
1469				amdgpu_vm_pt_next(adev, &cursor);
1470			}
1471
1472		} else if (frag >= shift) {
1473			/* or just move on to the next on the same level. */
1474			amdgpu_vm_pt_next(adev, &cursor);
1475		}
1476	}
1477
1478	return 0;
1479}
1480
1481/**
1482 * amdgpu_vm_bo_update_mapping - update a mapping in the vm page table
1483 *
1484 * @adev: amdgpu_device pointer
1485 * @exclusive: fence we need to sync to
1486 * @pages_addr: DMA addresses to use for mapping
1487 * @vm: requested vm
 
 
 
1488 * @start: start of mapped range
1489 * @last: last mapped entry
1490 * @flags: flags for the entries
1491 * @addr: addr to set the area to
 
 
1492 * @fence: optional resulting fence
 
1493 *
1494 * Fill in the page table entries between @start and @last.
1495 *
1496 * Returns:
1497 * 0 for success, -EINVAL for failure.
1498 */
1499static int amdgpu_vm_bo_update_mapping(struct amdgpu_device *adev,
1500				       struct dma_fence *exclusive,
1501				       dma_addr_t *pages_addr,
1502				       struct amdgpu_vm *vm,
1503				       uint64_t start, uint64_t last,
1504				       uint64_t flags, uint64_t addr,
1505				       struct dma_fence **fence)
 
 
 
1506{
1507	struct amdgpu_vm_update_params params;
1508	void *owner = AMDGPU_FENCE_OWNER_VM;
1509	int r;
 
 
 
 
1510
1511	memset(&params, 0, sizeof(params));
1512	params.adev = adev;
1513	params.vm = vm;
 
1514	params.pages_addr = pages_addr;
 
1515
1516	/* sync to everything except eviction fences on unmapping */
 
 
1517	if (!(flags & AMDGPU_PTE_VALID))
1518		owner = AMDGPU_FENCE_OWNER_KFD;
1519
1520	r = vm->update_funcs->prepare(&params, owner, exclusive);
1521	if (r)
1522		return r;
1523
1524	r = amdgpu_vm_update_ptes(&params, start, last + 1, addr, flags);
1525	if (r)
1526		return r;
1527
1528	return vm->update_funcs->commit(&params, fence);
1529}
1530
1531/**
1532 * amdgpu_vm_bo_split_mapping - split a mapping into smaller chunks
1533 *
1534 * @adev: amdgpu_device pointer
1535 * @exclusive: fence we need to sync to
1536 * @pages_addr: DMA addresses to use for mapping
1537 * @vm: requested vm
1538 * @mapping: mapped range and flags to use for the update
1539 * @flags: HW flags for the mapping
1540 * @bo_adev: amdgpu_device pointer that bo actually been allocated
1541 * @nodes: array of drm_mm_nodes with the MC addresses
1542 * @fence: optional resulting fence
1543 *
1544 * Split the mapping into smaller chunks so that each update fits
1545 * into a SDMA IB.
1546 *
1547 * Returns:
1548 * 0 for success, -EINVAL for failure.
1549 */
1550static int amdgpu_vm_bo_split_mapping(struct amdgpu_device *adev,
1551				      struct dma_fence *exclusive,
1552				      dma_addr_t *pages_addr,
1553				      struct amdgpu_vm *vm,
1554				      struct amdgpu_bo_va_mapping *mapping,
1555				      uint64_t flags,
1556				      struct amdgpu_device *bo_adev,
1557				      struct drm_mm_node *nodes,
1558				      struct dma_fence **fence)
1559{
1560	unsigned min_linear_pages = 1 << adev->vm_manager.fragment_size;
1561	uint64_t pfn, start = mapping->start;
1562	int r;
1563
1564	/* normally,bo_va->flags only contians READABLE and WIRTEABLE bit go here
1565	 * but in case of something, we filter the flags in first place
1566	 */
1567	if (!(mapping->flags & AMDGPU_PTE_READABLE))
1568		flags &= ~AMDGPU_PTE_READABLE;
1569	if (!(mapping->flags & AMDGPU_PTE_WRITEABLE))
1570		flags &= ~AMDGPU_PTE_WRITEABLE;
1571
1572	flags &= ~AMDGPU_PTE_EXECUTABLE;
1573	flags |= mapping->flags & AMDGPU_PTE_EXECUTABLE;
1574
1575	if (adev->asic_type >= CHIP_NAVI10) {
1576		flags &= ~AMDGPU_PTE_MTYPE_NV10_MASK;
1577		flags |= (mapping->flags & AMDGPU_PTE_MTYPE_NV10_MASK);
1578	} else {
1579		flags &= ~AMDGPU_PTE_MTYPE_VG10_MASK;
1580		flags |= (mapping->flags & AMDGPU_PTE_MTYPE_VG10_MASK);
1581	}
1582
1583	if ((mapping->flags & AMDGPU_PTE_PRT) &&
1584	    (adev->asic_type >= CHIP_VEGA10)) {
1585		flags |= AMDGPU_PTE_PRT;
1586		if (adev->asic_type >= CHIP_NAVI10) {
1587			flags |= AMDGPU_PTE_SNOOPED;
1588			flags |= AMDGPU_PTE_LOG;
1589			flags |= AMDGPU_PTE_SYSTEM;
1590		}
1591		flags &= ~AMDGPU_PTE_VALID;
1592	}
1593
1594	trace_amdgpu_vm_bo_update(mapping);
 
1595
1596	pfn = mapping->offset >> PAGE_SHIFT;
1597	if (nodes) {
1598		while (pfn >= nodes->size) {
1599			pfn -= nodes->size;
1600			++nodes;
1601		}
1602	}
1603
1604	do {
1605		dma_addr_t *dma_addr = NULL;
1606		uint64_t max_entries;
1607		uint64_t addr, last;
1608
1609		if (nodes) {
1610			addr = nodes->start << PAGE_SHIFT;
1611			max_entries = (nodes->size - pfn) *
1612				AMDGPU_GPU_PAGES_IN_CPU_PAGE;
1613		} else {
1614			addr = 0;
1615			max_entries = S64_MAX;
1616		}
1617
 
1618		if (pages_addr) {
1619			uint64_t count;
1620
1621			for (count = 1;
1622			     count < max_entries / AMDGPU_GPU_PAGES_IN_CPU_PAGE;
1623			     ++count) {
1624				uint64_t idx = pfn + count;
1625
1626				if (pages_addr[idx] !=
1627				    (pages_addr[idx - 1] + PAGE_SIZE))
1628					break;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1629			}
1630
1631			if (count < min_linear_pages) {
1632				addr = pfn << PAGE_SHIFT;
1633				dma_addr = pages_addr;
1634			} else {
1635				addr = pages_addr[pfn];
1636				max_entries = count * AMDGPU_GPU_PAGES_IN_CPU_PAGE;
1637			}
1638
1639		} else if (flags & AMDGPU_PTE_VALID) {
1640			addr += bo_adev->vm_manager.vram_base_offset;
1641			addr += pfn << PAGE_SHIFT;
 
 
1642		}
1643
1644		last = min((uint64_t)mapping->last, start + max_entries - 1);
1645		r = amdgpu_vm_bo_update_mapping(adev, exclusive, dma_addr, vm,
1646						start, last, flags, addr,
1647						fence);
1648		if (r)
1649			return r;
 
 
 
 
1650
1651		pfn += (last - start + 1) / AMDGPU_GPU_PAGES_IN_CPU_PAGE;
1652		if (nodes && nodes->size == pfn) {
1653			pfn = 0;
1654			++nodes;
1655		}
1656		start = last + 1;
1657
1658	} while (unlikely(start != mapping->last + 1));
 
1659
1660	return 0;
 
 
 
1661}
1662
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1663/**
1664 * amdgpu_vm_bo_update - update all BO mappings in the vm page table
1665 *
1666 * @adev: amdgpu_device pointer
1667 * @bo_va: requested BO and VM object
1668 * @clear: if true clear the entries
1669 *
1670 * Fill in the page table entries for @bo_va.
1671 *
1672 * Returns:
1673 * 0 for success, -EINVAL for failure.
1674 */
1675int amdgpu_vm_bo_update(struct amdgpu_device *adev,
1676			struct amdgpu_bo_va *bo_va,
1677			bool clear)
1678{
1679	struct amdgpu_bo *bo = bo_va->base.bo;
1680	struct amdgpu_vm *vm = bo_va->base.vm;
1681	struct amdgpu_bo_va_mapping *mapping;
1682	dma_addr_t *pages_addr = NULL;
1683	struct ttm_mem_reg *mem;
1684	struct drm_mm_node *nodes;
1685	struct dma_fence *exclusive, **last_update;
1686	uint64_t flags;
1687	struct amdgpu_device *bo_adev = adev;
1688	int r;
1689
1690	if (clear || !bo) {
1691		mem = NULL;
1692		nodes = NULL;
1693		exclusive = NULL;
1694	} else {
1695		struct ttm_dma_tt *ttm;
1696
1697		mem = &bo->tbo.mem;
1698		nodes = mem->mm_node;
1699		if (mem->mem_type == TTM_PL_TT) {
1700			ttm = container_of(bo->tbo.ttm, struct ttm_dma_tt, ttm);
1701			pages_addr = ttm->dma_address;
1702		}
1703		exclusive = dma_resv_get_excl(bo->tbo.base.resv);
 
 
 
 
 
 
1704	}
1705
1706	if (bo) {
1707		flags = amdgpu_ttm_tt_pte_flags(adev, bo->tbo.ttm, mem);
 
 
 
 
1708		bo_adev = amdgpu_ttm_adev(bo->tbo.bdev);
1709	} else {
1710		flags = 0x0;
1711	}
1712
1713	if (clear || (bo && bo->tbo.base.resv == vm->root.base.bo->tbo.base.resv))
 
1714		last_update = &vm->last_update;
1715	else
1716		last_update = &bo_va->last_pt_update;
1717
1718	if (!clear && bo_va->base.moved) {
1719		bo_va->base.moved = false;
1720		list_splice_init(&bo_va->valids, &bo_va->invalids);
1721
1722	} else if (bo_va->cleared != clear) {
1723		list_splice_init(&bo_va->valids, &bo_va->invalids);
1724	}
1725
1726	list_for_each_entry(mapping, &bo_va->invalids, list) {
1727		r = amdgpu_vm_bo_split_mapping(adev, exclusive, pages_addr, vm,
1728					       mapping, flags, bo_adev, nodes,
1729					       last_update);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1730		if (r)
1731			return r;
1732	}
1733
1734	if (vm->use_cpu_for_update) {
1735		/* Flush HDP */
1736		mb();
1737		amdgpu_asic_flush_hdp(adev, NULL);
1738	}
1739
1740	/* If the BO is not in its preferred location add it back to
1741	 * the evicted list so that it gets validated again on the
1742	 * next command submission.
1743	 */
1744	if (bo && bo->tbo.base.resv == vm->root.base.bo->tbo.base.resv) {
1745		uint32_t mem_type = bo->tbo.mem.mem_type;
1746
1747		if (!(bo->preferred_domains & amdgpu_mem_type_to_domain(mem_type)))
 
1748			amdgpu_vm_bo_evicted(&bo_va->base);
1749		else
1750			amdgpu_vm_bo_idle(&bo_va->base);
1751	} else {
1752		amdgpu_vm_bo_done(&bo_va->base);
1753	}
1754
1755	list_splice_init(&bo_va->invalids, &bo_va->valids);
1756	bo_va->cleared = clear;
1757
1758	if (trace_amdgpu_vm_bo_mapping_enabled()) {
1759		list_for_each_entry(mapping, &bo_va->valids, list)
1760			trace_amdgpu_vm_bo_mapping(mapping);
1761	}
1762
1763	return 0;
1764}
1765
1766/**
1767 * amdgpu_vm_update_prt_state - update the global PRT state
1768 *
1769 * @adev: amdgpu_device pointer
1770 */
1771static void amdgpu_vm_update_prt_state(struct amdgpu_device *adev)
1772{
1773	unsigned long flags;
1774	bool enable;
1775
1776	spin_lock_irqsave(&adev->vm_manager.prt_lock, flags);
1777	enable = !!atomic_read(&adev->vm_manager.num_prt_users);
1778	adev->gmc.gmc_funcs->set_prt(adev, enable);
1779	spin_unlock_irqrestore(&adev->vm_manager.prt_lock, flags);
1780}
1781
1782/**
1783 * amdgpu_vm_prt_get - add a PRT user
1784 *
1785 * @adev: amdgpu_device pointer
1786 */
1787static void amdgpu_vm_prt_get(struct amdgpu_device *adev)
1788{
1789	if (!adev->gmc.gmc_funcs->set_prt)
1790		return;
1791
1792	if (atomic_inc_return(&adev->vm_manager.num_prt_users) == 1)
1793		amdgpu_vm_update_prt_state(adev);
1794}
1795
1796/**
1797 * amdgpu_vm_prt_put - drop a PRT user
1798 *
1799 * @adev: amdgpu_device pointer
1800 */
1801static void amdgpu_vm_prt_put(struct amdgpu_device *adev)
1802{
1803	if (atomic_dec_return(&adev->vm_manager.num_prt_users) == 0)
1804		amdgpu_vm_update_prt_state(adev);
1805}
1806
1807/**
1808 * amdgpu_vm_prt_cb - callback for updating the PRT status
1809 *
1810 * @fence: fence for the callback
1811 * @_cb: the callback function
1812 */
1813static void amdgpu_vm_prt_cb(struct dma_fence *fence, struct dma_fence_cb *_cb)
1814{
1815	struct amdgpu_prt_cb *cb = container_of(_cb, struct amdgpu_prt_cb, cb);
1816
1817	amdgpu_vm_prt_put(cb->adev);
1818	kfree(cb);
1819}
1820
1821/**
1822 * amdgpu_vm_add_prt_cb - add callback for updating the PRT status
1823 *
1824 * @adev: amdgpu_device pointer
1825 * @fence: fence for the callback
1826 */
1827static void amdgpu_vm_add_prt_cb(struct amdgpu_device *adev,
1828				 struct dma_fence *fence)
1829{
1830	struct amdgpu_prt_cb *cb;
1831
1832	if (!adev->gmc.gmc_funcs->set_prt)
1833		return;
1834
1835	cb = kmalloc(sizeof(struct amdgpu_prt_cb), GFP_KERNEL);
1836	if (!cb) {
1837		/* Last resort when we are OOM */
1838		if (fence)
1839			dma_fence_wait(fence, false);
1840
1841		amdgpu_vm_prt_put(adev);
1842	} else {
1843		cb->adev = adev;
1844		if (!fence || dma_fence_add_callback(fence, &cb->cb,
1845						     amdgpu_vm_prt_cb))
1846			amdgpu_vm_prt_cb(fence, &cb->cb);
1847	}
1848}
1849
1850/**
1851 * amdgpu_vm_free_mapping - free a mapping
1852 *
1853 * @adev: amdgpu_device pointer
1854 * @vm: requested vm
1855 * @mapping: mapping to be freed
1856 * @fence: fence of the unmap operation
1857 *
1858 * Free a mapping and make sure we decrease the PRT usage count if applicable.
1859 */
1860static void amdgpu_vm_free_mapping(struct amdgpu_device *adev,
1861				   struct amdgpu_vm *vm,
1862				   struct amdgpu_bo_va_mapping *mapping,
1863				   struct dma_fence *fence)
1864{
1865	if (mapping->flags & AMDGPU_PTE_PRT)
1866		amdgpu_vm_add_prt_cb(adev, fence);
1867	kfree(mapping);
1868}
1869
1870/**
1871 * amdgpu_vm_prt_fini - finish all prt mappings
1872 *
1873 * @adev: amdgpu_device pointer
1874 * @vm: requested vm
1875 *
1876 * Register a cleanup callback to disable PRT support after VM dies.
1877 */
1878static void amdgpu_vm_prt_fini(struct amdgpu_device *adev, struct amdgpu_vm *vm)
1879{
1880	struct dma_resv *resv = vm->root.base.bo->tbo.base.resv;
1881	struct dma_fence *excl, **shared;
1882	unsigned i, shared_count;
1883	int r;
1884
1885	r = dma_resv_get_fences_rcu(resv, &excl,
1886					      &shared_count, &shared);
1887	if (r) {
1888		/* Not enough memory to grab the fence list, as last resort
1889		 * block for all the fences to complete.
1890		 */
1891		dma_resv_wait_timeout_rcu(resv, true, false,
1892						    MAX_SCHEDULE_TIMEOUT);
1893		return;
1894	}
1895
1896	/* Add a callback for each fence in the reservation object */
1897	amdgpu_vm_prt_get(adev);
1898	amdgpu_vm_add_prt_cb(adev, excl);
1899
1900	for (i = 0; i < shared_count; ++i) {
1901		amdgpu_vm_prt_get(adev);
1902		amdgpu_vm_add_prt_cb(adev, shared[i]);
1903	}
1904
1905	kfree(shared);
1906}
1907
1908/**
1909 * amdgpu_vm_clear_freed - clear freed BOs in the PT
1910 *
1911 * @adev: amdgpu_device pointer
1912 * @vm: requested vm
1913 * @fence: optional resulting fence (unchanged if no work needed to be done
1914 * or if an error occurred)
1915 *
1916 * Make sure all freed BOs are cleared in the PT.
1917 * PTs have to be reserved and mutex must be locked!
1918 *
1919 * Returns:
1920 * 0 for success.
1921 *
1922 */
1923int amdgpu_vm_clear_freed(struct amdgpu_device *adev,
1924			  struct amdgpu_vm *vm,
1925			  struct dma_fence **fence)
1926{
 
1927	struct amdgpu_bo_va_mapping *mapping;
1928	uint64_t init_pte_value = 0;
1929	struct dma_fence *f = NULL;
1930	int r;
1931
1932	while (!list_empty(&vm->freed)) {
1933		mapping = list_first_entry(&vm->freed,
1934			struct amdgpu_bo_va_mapping, list);
1935		list_del(&mapping->list);
1936
1937		if (vm->pte_support_ats &&
1938		    mapping->start < AMDGPU_GMC_HOLE_START)
1939			init_pte_value = AMDGPU_PTE_DEFAULT_ATC;
1940
1941		r = amdgpu_vm_bo_update_mapping(adev, NULL, NULL, vm,
1942						mapping->start, mapping->last,
1943						init_pte_value, 0, &f);
 
1944		amdgpu_vm_free_mapping(adev, vm, mapping, f);
1945		if (r) {
1946			dma_fence_put(f);
1947			return r;
1948		}
1949	}
1950
1951	if (fence && f) {
1952		dma_fence_put(*fence);
1953		*fence = f;
1954	} else {
1955		dma_fence_put(f);
1956	}
1957
1958	return 0;
1959
1960}
1961
1962/**
1963 * amdgpu_vm_handle_moved - handle moved BOs in the PT
1964 *
1965 * @adev: amdgpu_device pointer
1966 * @vm: requested vm
1967 *
1968 * Make sure all BOs which are moved are updated in the PTs.
1969 *
1970 * Returns:
1971 * 0 for success.
1972 *
1973 * PTs have to be reserved!
1974 */
1975int amdgpu_vm_handle_moved(struct amdgpu_device *adev,
1976			   struct amdgpu_vm *vm)
1977{
1978	struct amdgpu_bo_va *bo_va, *tmp;
1979	struct dma_resv *resv;
1980	bool clear;
1981	int r;
1982
1983	list_for_each_entry_safe(bo_va, tmp, &vm->moved, base.vm_status) {
1984		/* Per VM BOs never need to bo cleared in the page tables */
1985		r = amdgpu_vm_bo_update(adev, bo_va, false);
1986		if (r)
1987			return r;
1988	}
1989
1990	spin_lock(&vm->invalidated_lock);
1991	while (!list_empty(&vm->invalidated)) {
1992		bo_va = list_first_entry(&vm->invalidated, struct amdgpu_bo_va,
1993					 base.vm_status);
1994		resv = bo_va->base.bo->tbo.base.resv;
1995		spin_unlock(&vm->invalidated_lock);
1996
1997		/* Try to reserve the BO to avoid clearing its ptes */
1998		if (!amdgpu_vm_debug && dma_resv_trylock(resv))
1999			clear = false;
2000		/* Somebody else is using the BO right now */
2001		else
2002			clear = true;
2003
2004		r = amdgpu_vm_bo_update(adev, bo_va, clear);
2005		if (r)
2006			return r;
2007
2008		if (!clear)
2009			dma_resv_unlock(resv);
2010		spin_lock(&vm->invalidated_lock);
2011	}
2012	spin_unlock(&vm->invalidated_lock);
2013
2014	return 0;
2015}
2016
2017/**
2018 * amdgpu_vm_bo_add - add a bo to a specific vm
2019 *
2020 * @adev: amdgpu_device pointer
2021 * @vm: requested vm
2022 * @bo: amdgpu buffer object
2023 *
2024 * Add @bo into the requested vm.
2025 * Add @bo to the list of bos associated with the vm
2026 *
2027 * Returns:
2028 * Newly added bo_va or NULL for failure
2029 *
2030 * Object has to be reserved!
2031 */
2032struct amdgpu_bo_va *amdgpu_vm_bo_add(struct amdgpu_device *adev,
2033				      struct amdgpu_vm *vm,
2034				      struct amdgpu_bo *bo)
2035{
2036	struct amdgpu_bo_va *bo_va;
2037
2038	bo_va = kzalloc(sizeof(struct amdgpu_bo_va), GFP_KERNEL);
2039	if (bo_va == NULL) {
2040		return NULL;
2041	}
2042	amdgpu_vm_bo_base_init(&bo_va->base, vm, bo);
2043
2044	bo_va->ref_count = 1;
2045	INIT_LIST_HEAD(&bo_va->valids);
2046	INIT_LIST_HEAD(&bo_va->invalids);
2047
2048	if (bo && amdgpu_xgmi_same_hive(adev, amdgpu_ttm_adev(bo->tbo.bdev)) &&
2049	    (bo->preferred_domains & AMDGPU_GEM_DOMAIN_VRAM)) {
 
 
2050		bo_va->is_xgmi = true;
2051		mutex_lock(&adev->vm_manager.lock_pstate);
2052		/* Power up XGMI if it can be potentially used */
2053		if (++adev->vm_manager.xgmi_map_counter == 1)
2054			amdgpu_xgmi_set_pstate(adev, 1);
2055		mutex_unlock(&adev->vm_manager.lock_pstate);
2056	}
2057
2058	return bo_va;
2059}
2060
2061
2062/**
2063 * amdgpu_vm_bo_insert_mapping - insert a new mapping
2064 *
2065 * @adev: amdgpu_device pointer
2066 * @bo_va: bo_va to store the address
2067 * @mapping: the mapping to insert
2068 *
2069 * Insert a new mapping into all structures.
2070 */
2071static void amdgpu_vm_bo_insert_map(struct amdgpu_device *adev,
2072				    struct amdgpu_bo_va *bo_va,
2073				    struct amdgpu_bo_va_mapping *mapping)
2074{
2075	struct amdgpu_vm *vm = bo_va->base.vm;
2076	struct amdgpu_bo *bo = bo_va->base.bo;
2077
2078	mapping->bo_va = bo_va;
2079	list_add(&mapping->list, &bo_va->invalids);
2080	amdgpu_vm_it_insert(mapping, &vm->va);
2081
2082	if (mapping->flags & AMDGPU_PTE_PRT)
2083		amdgpu_vm_prt_get(adev);
2084
2085	if (bo && bo->tbo.base.resv == vm->root.base.bo->tbo.base.resv &&
2086	    !bo_va->base.moved) {
2087		list_move(&bo_va->base.vm_status, &vm->moved);
2088	}
2089	trace_amdgpu_vm_bo_map(bo_va, mapping);
2090}
2091
2092/**
2093 * amdgpu_vm_bo_map - map bo inside a vm
2094 *
2095 * @adev: amdgpu_device pointer
2096 * @bo_va: bo_va to store the address
2097 * @saddr: where to map the BO
2098 * @offset: requested offset in the BO
2099 * @size: BO size in bytes
2100 * @flags: attributes of pages (read/write/valid/etc.)
2101 *
2102 * Add a mapping of the BO at the specefied addr into the VM.
2103 *
2104 * Returns:
2105 * 0 for success, error for failure.
2106 *
2107 * Object has to be reserved and unreserved outside!
2108 */
2109int amdgpu_vm_bo_map(struct amdgpu_device *adev,
2110		     struct amdgpu_bo_va *bo_va,
2111		     uint64_t saddr, uint64_t offset,
2112		     uint64_t size, uint64_t flags)
2113{
2114	struct amdgpu_bo_va_mapping *mapping, *tmp;
2115	struct amdgpu_bo *bo = bo_va->base.bo;
2116	struct amdgpu_vm *vm = bo_va->base.vm;
2117	uint64_t eaddr;
2118
2119	/* validate the parameters */
2120	if (saddr & AMDGPU_GPU_PAGE_MASK || offset & AMDGPU_GPU_PAGE_MASK ||
2121	    size == 0 || size & AMDGPU_GPU_PAGE_MASK)
2122		return -EINVAL;
2123
2124	/* make sure object fit at this offset */
2125	eaddr = saddr + size - 1;
2126	if (saddr >= eaddr ||
2127	    (bo && offset + size > amdgpu_bo_size(bo)))
 
2128		return -EINVAL;
2129
2130	saddr /= AMDGPU_GPU_PAGE_SIZE;
2131	eaddr /= AMDGPU_GPU_PAGE_SIZE;
2132
2133	tmp = amdgpu_vm_it_iter_first(&vm->va, saddr, eaddr);
2134	if (tmp) {
2135		/* bo and tmp overlap, invalid addr */
2136		dev_err(adev->dev, "bo %p va 0x%010Lx-0x%010Lx conflict with "
2137			"0x%010Lx-0x%010Lx\n", bo, saddr, eaddr,
2138			tmp->start, tmp->last + 1);
2139		return -EINVAL;
2140	}
2141
2142	mapping = kmalloc(sizeof(*mapping), GFP_KERNEL);
2143	if (!mapping)
2144		return -ENOMEM;
2145
2146	mapping->start = saddr;
2147	mapping->last = eaddr;
2148	mapping->offset = offset;
2149	mapping->flags = flags;
2150
2151	amdgpu_vm_bo_insert_map(adev, bo_va, mapping);
2152
2153	return 0;
2154}
2155
2156/**
2157 * amdgpu_vm_bo_replace_map - map bo inside a vm, replacing existing mappings
2158 *
2159 * @adev: amdgpu_device pointer
2160 * @bo_va: bo_va to store the address
2161 * @saddr: where to map the BO
2162 * @offset: requested offset in the BO
2163 * @size: BO size in bytes
2164 * @flags: attributes of pages (read/write/valid/etc.)
2165 *
2166 * Add a mapping of the BO at the specefied addr into the VM. Replace existing
2167 * mappings as we do so.
2168 *
2169 * Returns:
2170 * 0 for success, error for failure.
2171 *
2172 * Object has to be reserved and unreserved outside!
2173 */
2174int amdgpu_vm_bo_replace_map(struct amdgpu_device *adev,
2175			     struct amdgpu_bo_va *bo_va,
2176			     uint64_t saddr, uint64_t offset,
2177			     uint64_t size, uint64_t flags)
2178{
2179	struct amdgpu_bo_va_mapping *mapping;
2180	struct amdgpu_bo *bo = bo_va->base.bo;
2181	uint64_t eaddr;
2182	int r;
2183
2184	/* validate the parameters */
2185	if (saddr & AMDGPU_GPU_PAGE_MASK || offset & AMDGPU_GPU_PAGE_MASK ||
2186	    size == 0 || size & AMDGPU_GPU_PAGE_MASK)
2187		return -EINVAL;
2188
2189	/* make sure object fit at this offset */
2190	eaddr = saddr + size - 1;
2191	if (saddr >= eaddr ||
2192	    (bo && offset + size > amdgpu_bo_size(bo)))
 
2193		return -EINVAL;
2194
2195	/* Allocate all the needed memory */
2196	mapping = kmalloc(sizeof(*mapping), GFP_KERNEL);
2197	if (!mapping)
2198		return -ENOMEM;
2199
2200	r = amdgpu_vm_bo_clear_mappings(adev, bo_va->base.vm, saddr, size);
2201	if (r) {
2202		kfree(mapping);
2203		return r;
2204	}
2205
2206	saddr /= AMDGPU_GPU_PAGE_SIZE;
2207	eaddr /= AMDGPU_GPU_PAGE_SIZE;
2208
2209	mapping->start = saddr;
2210	mapping->last = eaddr;
2211	mapping->offset = offset;
2212	mapping->flags = flags;
2213
2214	amdgpu_vm_bo_insert_map(adev, bo_va, mapping);
2215
2216	return 0;
2217}
2218
2219/**
2220 * amdgpu_vm_bo_unmap - remove bo mapping from vm
2221 *
2222 * @adev: amdgpu_device pointer
2223 * @bo_va: bo_va to remove the address from
2224 * @saddr: where to the BO is mapped
2225 *
2226 * Remove a mapping of the BO at the specefied addr from the VM.
2227 *
2228 * Returns:
2229 * 0 for success, error for failure.
2230 *
2231 * Object has to be reserved and unreserved outside!
2232 */
2233int amdgpu_vm_bo_unmap(struct amdgpu_device *adev,
2234		       struct amdgpu_bo_va *bo_va,
2235		       uint64_t saddr)
2236{
2237	struct amdgpu_bo_va_mapping *mapping;
2238	struct amdgpu_vm *vm = bo_va->base.vm;
2239	bool valid = true;
2240
2241	saddr /= AMDGPU_GPU_PAGE_SIZE;
2242
2243	list_for_each_entry(mapping, &bo_va->valids, list) {
2244		if (mapping->start == saddr)
2245			break;
2246	}
2247
2248	if (&mapping->list == &bo_va->valids) {
2249		valid = false;
2250
2251		list_for_each_entry(mapping, &bo_va->invalids, list) {
2252			if (mapping->start == saddr)
2253				break;
2254		}
2255
2256		if (&mapping->list == &bo_va->invalids)
2257			return -ENOENT;
2258	}
2259
2260	list_del(&mapping->list);
2261	amdgpu_vm_it_remove(mapping, &vm->va);
2262	mapping->bo_va = NULL;
2263	trace_amdgpu_vm_bo_unmap(bo_va, mapping);
2264
2265	if (valid)
2266		list_add(&mapping->list, &vm->freed);
2267	else
2268		amdgpu_vm_free_mapping(adev, vm, mapping,
2269				       bo_va->last_pt_update);
2270
2271	return 0;
2272}
2273
2274/**
2275 * amdgpu_vm_bo_clear_mappings - remove all mappings in a specific range
2276 *
2277 * @adev: amdgpu_device pointer
2278 * @vm: VM structure to use
2279 * @saddr: start of the range
2280 * @size: size of the range
2281 *
2282 * Remove all mappings in a range, split them as appropriate.
2283 *
2284 * Returns:
2285 * 0 for success, error for failure.
2286 */
2287int amdgpu_vm_bo_clear_mappings(struct amdgpu_device *adev,
2288				struct amdgpu_vm *vm,
2289				uint64_t saddr, uint64_t size)
2290{
2291	struct amdgpu_bo_va_mapping *before, *after, *tmp, *next;
2292	LIST_HEAD(removed);
2293	uint64_t eaddr;
2294
2295	eaddr = saddr + size - 1;
2296	saddr /= AMDGPU_GPU_PAGE_SIZE;
2297	eaddr /= AMDGPU_GPU_PAGE_SIZE;
2298
2299	/* Allocate all the needed memory */
2300	before = kzalloc(sizeof(*before), GFP_KERNEL);
2301	if (!before)
2302		return -ENOMEM;
2303	INIT_LIST_HEAD(&before->list);
2304
2305	after = kzalloc(sizeof(*after), GFP_KERNEL);
2306	if (!after) {
2307		kfree(before);
2308		return -ENOMEM;
2309	}
2310	INIT_LIST_HEAD(&after->list);
2311
2312	/* Now gather all removed mappings */
2313	tmp = amdgpu_vm_it_iter_first(&vm->va, saddr, eaddr);
2314	while (tmp) {
2315		/* Remember mapping split at the start */
2316		if (tmp->start < saddr) {
2317			before->start = tmp->start;
2318			before->last = saddr - 1;
2319			before->offset = tmp->offset;
2320			before->flags = tmp->flags;
2321			before->bo_va = tmp->bo_va;
2322			list_add(&before->list, &tmp->bo_va->invalids);
2323		}
2324
2325		/* Remember mapping split at the end */
2326		if (tmp->last > eaddr) {
2327			after->start = eaddr + 1;
2328			after->last = tmp->last;
2329			after->offset = tmp->offset;
2330			after->offset += after->start - tmp->start;
2331			after->flags = tmp->flags;
2332			after->bo_va = tmp->bo_va;
2333			list_add(&after->list, &tmp->bo_va->invalids);
2334		}
2335
2336		list_del(&tmp->list);
2337		list_add(&tmp->list, &removed);
2338
2339		tmp = amdgpu_vm_it_iter_next(tmp, saddr, eaddr);
2340	}
2341
2342	/* And free them up */
2343	list_for_each_entry_safe(tmp, next, &removed, list) {
2344		amdgpu_vm_it_remove(tmp, &vm->va);
2345		list_del(&tmp->list);
2346
2347		if (tmp->start < saddr)
2348		    tmp->start = saddr;
2349		if (tmp->last > eaddr)
2350		    tmp->last = eaddr;
2351
2352		tmp->bo_va = NULL;
2353		list_add(&tmp->list, &vm->freed);
2354		trace_amdgpu_vm_bo_unmap(NULL, tmp);
2355	}
2356
2357	/* Insert partial mapping before the range */
2358	if (!list_empty(&before->list)) {
2359		amdgpu_vm_it_insert(before, &vm->va);
2360		if (before->flags & AMDGPU_PTE_PRT)
2361			amdgpu_vm_prt_get(adev);
2362	} else {
2363		kfree(before);
2364	}
2365
2366	/* Insert partial mapping after the range */
2367	if (!list_empty(&after->list)) {
2368		amdgpu_vm_it_insert(after, &vm->va);
2369		if (after->flags & AMDGPU_PTE_PRT)
2370			amdgpu_vm_prt_get(adev);
2371	} else {
2372		kfree(after);
2373	}
2374
2375	return 0;
2376}
2377
2378/**
2379 * amdgpu_vm_bo_lookup_mapping - find mapping by address
2380 *
2381 * @vm: the requested VM
2382 * @addr: the address
2383 *
2384 * Find a mapping by it's address.
2385 *
2386 * Returns:
2387 * The amdgpu_bo_va_mapping matching for addr or NULL
2388 *
2389 */
2390struct amdgpu_bo_va_mapping *amdgpu_vm_bo_lookup_mapping(struct amdgpu_vm *vm,
2391							 uint64_t addr)
2392{
2393	return amdgpu_vm_it_iter_first(&vm->va, addr, addr);
2394}
2395
2396/**
2397 * amdgpu_vm_bo_trace_cs - trace all reserved mappings
2398 *
2399 * @vm: the requested vm
2400 * @ticket: CS ticket
2401 *
2402 * Trace all mappings of BOs reserved during a command submission.
2403 */
2404void amdgpu_vm_bo_trace_cs(struct amdgpu_vm *vm, struct ww_acquire_ctx *ticket)
2405{
2406	struct amdgpu_bo_va_mapping *mapping;
2407
2408	if (!trace_amdgpu_vm_bo_cs_enabled())
2409		return;
2410
2411	for (mapping = amdgpu_vm_it_iter_first(&vm->va, 0, U64_MAX); mapping;
2412	     mapping = amdgpu_vm_it_iter_next(mapping, 0, U64_MAX)) {
2413		if (mapping->bo_va && mapping->bo_va->base.bo) {
2414			struct amdgpu_bo *bo;
2415
2416			bo = mapping->bo_va->base.bo;
2417			if (dma_resv_locking_ctx(bo->tbo.base.resv) !=
2418			    ticket)
2419				continue;
2420		}
2421
2422		trace_amdgpu_vm_bo_cs(mapping);
2423	}
2424}
2425
2426/**
2427 * amdgpu_vm_bo_rmv - remove a bo to a specific vm
2428 *
2429 * @adev: amdgpu_device pointer
2430 * @bo_va: requested bo_va
2431 *
2432 * Remove @bo_va->bo from the requested vm.
2433 *
2434 * Object have to be reserved!
2435 */
2436void amdgpu_vm_bo_rmv(struct amdgpu_device *adev,
2437		      struct amdgpu_bo_va *bo_va)
2438{
2439	struct amdgpu_bo_va_mapping *mapping, *next;
2440	struct amdgpu_bo *bo = bo_va->base.bo;
2441	struct amdgpu_vm *vm = bo_va->base.vm;
2442	struct amdgpu_vm_bo_base **base;
2443
2444	if (bo) {
2445		if (bo->tbo.base.resv == vm->root.base.bo->tbo.base.resv)
2446			vm->bulk_moveable = false;
2447
2448		for (base = &bo_va->base.bo->vm_bo; *base;
2449		     base = &(*base)->next) {
2450			if (*base != &bo_va->base)
2451				continue;
2452
2453			*base = bo_va->base.next;
2454			break;
2455		}
2456	}
2457
2458	spin_lock(&vm->invalidated_lock);
2459	list_del(&bo_va->base.vm_status);
2460	spin_unlock(&vm->invalidated_lock);
2461
2462	list_for_each_entry_safe(mapping, next, &bo_va->valids, list) {
2463		list_del(&mapping->list);
2464		amdgpu_vm_it_remove(mapping, &vm->va);
2465		mapping->bo_va = NULL;
2466		trace_amdgpu_vm_bo_unmap(bo_va, mapping);
2467		list_add(&mapping->list, &vm->freed);
2468	}
2469	list_for_each_entry_safe(mapping, next, &bo_va->invalids, list) {
2470		list_del(&mapping->list);
2471		amdgpu_vm_it_remove(mapping, &vm->va);
2472		amdgpu_vm_free_mapping(adev, vm, mapping,
2473				       bo_va->last_pt_update);
2474	}
2475
2476	dma_fence_put(bo_va->last_pt_update);
2477
2478	if (bo && bo_va->is_xgmi) {
2479		mutex_lock(&adev->vm_manager.lock_pstate);
2480		if (--adev->vm_manager.xgmi_map_counter == 0)
2481			amdgpu_xgmi_set_pstate(adev, 0);
2482		mutex_unlock(&adev->vm_manager.lock_pstate);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2483	}
2484
2485	kfree(bo_va);
 
 
2486}
2487
2488/**
2489 * amdgpu_vm_bo_invalidate - mark the bo as invalid
2490 *
2491 * @adev: amdgpu_device pointer
2492 * @bo: amdgpu buffer object
2493 * @evicted: is the BO evicted
2494 *
2495 * Mark @bo as invalid.
2496 */
2497void amdgpu_vm_bo_invalidate(struct amdgpu_device *adev,
2498			     struct amdgpu_bo *bo, bool evicted)
2499{
2500	struct amdgpu_vm_bo_base *bo_base;
2501
2502	/* shadow bo doesn't have bo base, its validation needs its parent */
2503	if (bo->parent && bo->parent->shadow == bo)
2504		bo = bo->parent;
2505
2506	for (bo_base = bo->vm_bo; bo_base; bo_base = bo_base->next) {
2507		struct amdgpu_vm *vm = bo_base->vm;
2508
2509		if (evicted && bo->tbo.base.resv == vm->root.base.bo->tbo.base.resv) {
2510			amdgpu_vm_bo_evicted(bo_base);
2511			continue;
2512		}
2513
2514		if (bo_base->moved)
2515			continue;
2516		bo_base->moved = true;
2517
2518		if (bo->tbo.type == ttm_bo_type_kernel)
2519			amdgpu_vm_bo_relocated(bo_base);
2520		else if (bo->tbo.base.resv == vm->root.base.bo->tbo.base.resv)
2521			amdgpu_vm_bo_moved(bo_base);
2522		else
2523			amdgpu_vm_bo_invalidated(bo_base);
2524	}
2525}
2526
2527/**
2528 * amdgpu_vm_get_block_size - calculate VM page table size as power of two
2529 *
2530 * @vm_size: VM size
2531 *
2532 * Returns:
2533 * VM page table as power of two
2534 */
2535static uint32_t amdgpu_vm_get_block_size(uint64_t vm_size)
2536{
2537	/* Total bits covered by PD + PTs */
2538	unsigned bits = ilog2(vm_size) + 18;
2539
2540	/* Make sure the PD is 4K in size up to 8GB address space.
2541	   Above that split equal between PD and PTs */
2542	if (vm_size <= 8)
2543		return (bits - 9);
2544	else
2545		return ((bits + 3) / 2);
2546}
2547
2548/**
2549 * amdgpu_vm_adjust_size - adjust vm size, block size and fragment size
2550 *
2551 * @adev: amdgpu_device pointer
2552 * @min_vm_size: the minimum vm size in GB if it's set auto
2553 * @fragment_size_default: Default PTE fragment size
2554 * @max_level: max VMPT level
2555 * @max_bits: max address space size in bits
2556 *
2557 */
2558void amdgpu_vm_adjust_size(struct amdgpu_device *adev, uint32_t min_vm_size,
2559			   uint32_t fragment_size_default, unsigned max_level,
2560			   unsigned max_bits)
2561{
2562	unsigned int max_size = 1 << (max_bits - 30);
2563	unsigned int vm_size;
2564	uint64_t tmp;
2565
2566	/* adjust vm size first */
2567	if (amdgpu_vm_size != -1) {
2568		vm_size = amdgpu_vm_size;
2569		if (vm_size > max_size) {
2570			dev_warn(adev->dev, "VM size (%d) too large, max is %u GB\n",
2571				 amdgpu_vm_size, max_size);
2572			vm_size = max_size;
2573		}
2574	} else {
2575		struct sysinfo si;
2576		unsigned int phys_ram_gb;
2577
2578		/* Optimal VM size depends on the amount of physical
2579		 * RAM available. Underlying requirements and
2580		 * assumptions:
2581		 *
2582		 *  - Need to map system memory and VRAM from all GPUs
2583		 *     - VRAM from other GPUs not known here
2584		 *     - Assume VRAM <= system memory
2585		 *  - On GFX8 and older, VM space can be segmented for
2586		 *    different MTYPEs
2587		 *  - Need to allow room for fragmentation, guard pages etc.
2588		 *
2589		 * This adds up to a rough guess of system memory x3.
2590		 * Round up to power of two to maximize the available
2591		 * VM size with the given page table size.
2592		 */
2593		si_meminfo(&si);
2594		phys_ram_gb = ((uint64_t)si.totalram * si.mem_unit +
2595			       (1 << 30) - 1) >> 30;
2596		vm_size = roundup_pow_of_two(
2597			min(max(phys_ram_gb * 3, min_vm_size), max_size));
2598	}
2599
2600	adev->vm_manager.max_pfn = (uint64_t)vm_size << 18;
2601
2602	tmp = roundup_pow_of_two(adev->vm_manager.max_pfn);
2603	if (amdgpu_vm_block_size != -1)
2604		tmp >>= amdgpu_vm_block_size - 9;
2605	tmp = DIV_ROUND_UP(fls64(tmp) - 1, 9) - 1;
2606	adev->vm_manager.num_level = min(max_level, (unsigned)tmp);
2607	switch (adev->vm_manager.num_level) {
2608	case 3:
2609		adev->vm_manager.root_level = AMDGPU_VM_PDB2;
2610		break;
2611	case 2:
2612		adev->vm_manager.root_level = AMDGPU_VM_PDB1;
2613		break;
2614	case 1:
2615		adev->vm_manager.root_level = AMDGPU_VM_PDB0;
2616		break;
2617	default:
2618		dev_err(adev->dev, "VMPT only supports 2~4+1 levels\n");
2619	}
2620	/* block size depends on vm size and hw setup*/
2621	if (amdgpu_vm_block_size != -1)
2622		adev->vm_manager.block_size =
2623			min((unsigned)amdgpu_vm_block_size, max_bits
2624			    - AMDGPU_GPU_PAGE_SHIFT
2625			    - 9 * adev->vm_manager.num_level);
2626	else if (adev->vm_manager.num_level > 1)
2627		adev->vm_manager.block_size = 9;
2628	else
2629		adev->vm_manager.block_size = amdgpu_vm_get_block_size(tmp);
2630
2631	if (amdgpu_vm_fragment_size == -1)
2632		adev->vm_manager.fragment_size = fragment_size_default;
2633	else
2634		adev->vm_manager.fragment_size = amdgpu_vm_fragment_size;
2635
2636	DRM_INFO("vm size is %u GB, %u levels, block size is %u-bit, fragment size is %u-bit\n",
2637		 vm_size, adev->vm_manager.num_level + 1,
2638		 adev->vm_manager.block_size,
2639		 adev->vm_manager.fragment_size);
2640}
2641
2642/**
2643 * amdgpu_vm_wait_idle - wait for the VM to become idle
2644 *
2645 * @vm: VM object to wait for
2646 * @timeout: timeout to wait for VM to become idle
2647 */
2648long amdgpu_vm_wait_idle(struct amdgpu_vm *vm, long timeout)
2649{
2650	return dma_resv_wait_timeout_rcu(vm->root.base.bo->tbo.base.resv,
2651						   true, true, timeout);
 
 
 
 
2652}
2653
2654/**
2655 * amdgpu_vm_init - initialize a vm instance
2656 *
2657 * @adev: amdgpu_device pointer
2658 * @vm: requested vm
2659 * @vm_context: Indicates if it GFX or Compute context
2660 * @pasid: Process address space identifier
2661 *
2662 * Init @vm fields.
2663 *
2664 * Returns:
2665 * 0 for success, error for failure.
2666 */
2667int amdgpu_vm_init(struct amdgpu_device *adev, struct amdgpu_vm *vm,
2668		   int vm_context, unsigned int pasid)
2669{
2670	struct amdgpu_bo_param bp;
2671	struct amdgpu_bo *root;
2672	int r, i;
2673
2674	vm->va = RB_ROOT_CACHED;
2675	for (i = 0; i < AMDGPU_MAX_VMHUBS; i++)
2676		vm->reserved_vmid[i] = NULL;
2677	INIT_LIST_HEAD(&vm->evicted);
2678	INIT_LIST_HEAD(&vm->relocated);
2679	INIT_LIST_HEAD(&vm->moved);
2680	INIT_LIST_HEAD(&vm->idle);
2681	INIT_LIST_HEAD(&vm->invalidated);
2682	spin_lock_init(&vm->invalidated_lock);
2683	INIT_LIST_HEAD(&vm->freed);
 
2684
2685	/* create scheduler entity for page table updates */
2686	r = drm_sched_entity_init(&vm->entity, adev->vm_manager.vm_pte_rqs,
2687				  adev->vm_manager.vm_pte_num_rqs, NULL);
 
2688	if (r)
2689		return r;
2690
 
 
 
 
 
 
2691	vm->pte_support_ats = false;
 
2692
2693	if (vm_context == AMDGPU_VM_CONTEXT_COMPUTE) {
2694		vm->use_cpu_for_update = !!(adev->vm_manager.vm_update_mode &
2695						AMDGPU_VM_USE_CPU_FOR_COMPUTE);
2696
2697		if (adev->asic_type == CHIP_RAVEN)
2698			vm->pte_support_ats = true;
2699	} else {
2700		vm->use_cpu_for_update = !!(adev->vm_manager.vm_update_mode &
2701						AMDGPU_VM_USE_CPU_FOR_GFX);
2702	}
2703	DRM_DEBUG_DRIVER("VM update mode is %s\n",
2704			 vm->use_cpu_for_update ? "CPU" : "SDMA");
2705	WARN_ONCE((vm->use_cpu_for_update && !amdgpu_gmc_vram_full_visible(&adev->gmc)),
 
2706		  "CPU update of VM recommended only for large BAR system\n");
2707
2708	if (vm->use_cpu_for_update)
2709		vm->update_funcs = &amdgpu_vm_cpu_funcs;
2710	else
2711		vm->update_funcs = &amdgpu_vm_sdma_funcs;
2712	vm->last_update = NULL;
 
2713
2714	amdgpu_vm_bo_param(adev, vm, adev->vm_manager.root_level, &bp);
2715	if (vm_context == AMDGPU_VM_CONTEXT_COMPUTE)
2716		bp.flags &= ~AMDGPU_GEM_CREATE_SHADOW;
2717	r = amdgpu_bo_create(adev, &bp, &root);
 
2718	if (r)
2719		goto error_free_sched_entity;
2720
2721	r = amdgpu_bo_reserve(root, true);
2722	if (r)
2723		goto error_free_root;
2724
2725	r = dma_resv_reserve_shared(root->tbo.base.resv, 1);
2726	if (r)
2727		goto error_unreserve;
2728
2729	amdgpu_vm_bo_base_init(&vm->root.base, vm, root);
2730
2731	r = amdgpu_vm_clear_bo(adev, vm, root);
2732	if (r)
2733		goto error_unreserve;
2734
2735	amdgpu_bo_unreserve(vm->root.base.bo);
2736
2737	if (pasid) {
2738		unsigned long flags;
2739
2740		spin_lock_irqsave(&adev->vm_manager.pasid_lock, flags);
2741		r = idr_alloc(&adev->vm_manager.pasid_idr, vm, pasid, pasid + 1,
2742			      GFP_ATOMIC);
2743		spin_unlock_irqrestore(&adev->vm_manager.pasid_lock, flags);
2744		if (r < 0)
2745			goto error_free_root;
2746
2747		vm->pasid = pasid;
2748	}
2749
2750	INIT_KFIFO(vm->faults);
2751
2752	return 0;
2753
2754error_unreserve:
2755	amdgpu_bo_unreserve(vm->root.base.bo);
2756
2757error_free_root:
2758	amdgpu_bo_unref(&vm->root.base.bo->shadow);
2759	amdgpu_bo_unref(&vm->root.base.bo);
2760	vm->root.base.bo = NULL;
 
 
 
 
2761
2762error_free_sched_entity:
2763	drm_sched_entity_destroy(&vm->entity);
2764
2765	return r;
2766}
2767
2768/**
2769 * amdgpu_vm_check_clean_reserved - check if a VM is clean
2770 *
2771 * @adev: amdgpu_device pointer
2772 * @vm: the VM to check
2773 *
2774 * check all entries of the root PD, if any subsequent PDs are allocated,
2775 * it means there are page table creating and filling, and is no a clean
2776 * VM
2777 *
2778 * Returns:
2779 *	0 if this VM is clean
2780 */
2781static int amdgpu_vm_check_clean_reserved(struct amdgpu_device *adev,
2782	struct amdgpu_vm *vm)
2783{
2784	enum amdgpu_vm_level root = adev->vm_manager.root_level;
2785	unsigned int entries = amdgpu_vm_num_entries(adev, root);
2786	unsigned int i = 0;
2787
2788	if (!(vm->root.entries))
2789		return 0;
2790
2791	for (i = 0; i < entries; i++) {
2792		if (vm->root.entries[i].base.bo)
2793			return -EINVAL;
2794	}
2795
2796	return 0;
2797}
2798
2799/**
2800 * amdgpu_vm_make_compute - Turn a GFX VM into a compute VM
2801 *
2802 * @adev: amdgpu_device pointer
2803 * @vm: requested vm
 
2804 *
2805 * This only works on GFX VMs that don't have any BOs added and no
2806 * page tables allocated yet.
2807 *
2808 * Changes the following VM parameters:
2809 * - use_cpu_for_update
2810 * - pte_supports_ats
2811 * - pasid (old PASID is released, because compute manages its own PASIDs)
2812 *
2813 * Reinitializes the page directory to reflect the changed ATS
2814 * setting.
2815 *
2816 * Returns:
2817 * 0 for success, -errno for errors.
2818 */
2819int amdgpu_vm_make_compute(struct amdgpu_device *adev, struct amdgpu_vm *vm, unsigned int pasid)
 
2820{
2821	bool pte_support_ats = (adev->asic_type == CHIP_RAVEN);
2822	int r;
2823
2824	r = amdgpu_bo_reserve(vm->root.base.bo, true);
2825	if (r)
2826		return r;
2827
2828	/* Sanity checks */
2829	r = amdgpu_vm_check_clean_reserved(adev, vm);
2830	if (r)
2831		goto unreserve_bo;
2832
2833	if (pasid) {
2834		unsigned long flags;
2835
2836		spin_lock_irqsave(&adev->vm_manager.pasid_lock, flags);
2837		r = idr_alloc(&adev->vm_manager.pasid_idr, vm, pasid, pasid + 1,
2838			      GFP_ATOMIC);
2839		spin_unlock_irqrestore(&adev->vm_manager.pasid_lock, flags);
2840
2841		if (r == -ENOSPC)
2842			goto unreserve_bo;
2843		r = 0;
2844	}
2845
2846	/* Check if PD needs to be reinitialized and do it before
2847	 * changing any other state, in case it fails.
2848	 */
2849	if (pte_support_ats != vm->pte_support_ats) {
2850		vm->pte_support_ats = pte_support_ats;
2851		r = amdgpu_vm_clear_bo(adev, vm, vm->root.base.bo);
 
 
2852		if (r)
2853			goto free_idr;
2854	}
2855
2856	/* Update VM state */
2857	vm->use_cpu_for_update = !!(adev->vm_manager.vm_update_mode &
2858				    AMDGPU_VM_USE_CPU_FOR_COMPUTE);
2859	DRM_DEBUG_DRIVER("VM update mode is %s\n",
2860			 vm->use_cpu_for_update ? "CPU" : "SDMA");
2861	WARN_ONCE((vm->use_cpu_for_update && !amdgpu_gmc_vram_full_visible(&adev->gmc)),
 
2862		  "CPU update of VM recommended only for large BAR system\n");
2863
2864	if (vm->use_cpu_for_update)
 
 
 
 
 
 
2865		vm->update_funcs = &amdgpu_vm_cpu_funcs;
2866	else
2867		vm->update_funcs = &amdgpu_vm_sdma_funcs;
 
2868	dma_fence_put(vm->last_update);
2869	vm->last_update = NULL;
 
2870
2871	if (vm->pasid) {
2872		unsigned long flags;
2873
2874		spin_lock_irqsave(&adev->vm_manager.pasid_lock, flags);
2875		idr_remove(&adev->vm_manager.pasid_idr, vm->pasid);
2876		spin_unlock_irqrestore(&adev->vm_manager.pasid_lock, flags);
2877
2878		/* Free the original amdgpu allocated pasid
2879		 * Will be replaced with kfd allocated pasid
2880		 */
2881		amdgpu_pasid_free(vm->pasid);
2882		vm->pasid = 0;
2883	}
2884
2885	/* Free the shadow bo for compute VM */
2886	amdgpu_bo_unref(&vm->root.base.bo->shadow);
2887
2888	if (pasid)
2889		vm->pasid = pasid;
2890
2891	goto unreserve_bo;
2892
2893free_idr:
2894	if (pasid) {
2895		unsigned long flags;
2896
2897		spin_lock_irqsave(&adev->vm_manager.pasid_lock, flags);
2898		idr_remove(&adev->vm_manager.pasid_idr, pasid);
2899		spin_unlock_irqrestore(&adev->vm_manager.pasid_lock, flags);
2900	}
2901unreserve_bo:
2902	amdgpu_bo_unreserve(vm->root.base.bo);
2903	return r;
2904}
2905
2906/**
2907 * amdgpu_vm_release_compute - release a compute vm
2908 * @adev: amdgpu_device pointer
2909 * @vm: a vm turned into compute vm by calling amdgpu_vm_make_compute
2910 *
2911 * This is a correspondant of amdgpu_vm_make_compute. It decouples compute
2912 * pasid from vm. Compute should stop use of vm after this call.
2913 */
2914void amdgpu_vm_release_compute(struct amdgpu_device *adev, struct amdgpu_vm *vm)
2915{
2916	if (vm->pasid) {
2917		unsigned long flags;
2918
2919		spin_lock_irqsave(&adev->vm_manager.pasid_lock, flags);
2920		idr_remove(&adev->vm_manager.pasid_idr, vm->pasid);
2921		spin_unlock_irqrestore(&adev->vm_manager.pasid_lock, flags);
2922	}
2923	vm->pasid = 0;
 
2924}
2925
2926/**
2927 * amdgpu_vm_fini - tear down a vm instance
2928 *
2929 * @adev: amdgpu_device pointer
2930 * @vm: requested vm
2931 *
2932 * Tear down @vm.
2933 * Unbind the VM and remove all bos from the vm bo list
2934 */
2935void amdgpu_vm_fini(struct amdgpu_device *adev, struct amdgpu_vm *vm)
2936{
2937	struct amdgpu_bo_va_mapping *mapping, *tmp;
2938	bool prt_fini_needed = !!adev->gmc.gmc_funcs->set_prt;
2939	struct amdgpu_bo *root;
2940	int i, r;
2941
2942	amdgpu_amdkfd_gpuvm_destroy_cb(adev, vm);
2943
 
 
2944	if (vm->pasid) {
2945		unsigned long flags;
2946
2947		spin_lock_irqsave(&adev->vm_manager.pasid_lock, flags);
2948		idr_remove(&adev->vm_manager.pasid_idr, vm->pasid);
2949		spin_unlock_irqrestore(&adev->vm_manager.pasid_lock, flags);
 
2950	}
2951
2952	drm_sched_entity_destroy(&vm->entity);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2953
2954	if (!RB_EMPTY_ROOT(&vm->va.rb_root)) {
2955		dev_err(adev->dev, "still active bo inside vm\n");
2956	}
2957	rbtree_postorder_for_each_entry_safe(mapping, tmp,
2958					     &vm->va.rb_root, rb) {
2959		/* Don't remove the mapping here, we don't want to trigger a
2960		 * rebalance and the tree is about to be destroyed anyway.
2961		 */
2962		list_del(&mapping->list);
2963		kfree(mapping);
2964	}
2965	list_for_each_entry_safe(mapping, tmp, &vm->freed, list) {
2966		if (mapping->flags & AMDGPU_PTE_PRT && prt_fini_needed) {
2967			amdgpu_vm_prt_fini(adev, vm);
2968			prt_fini_needed = false;
2969		}
2970
2971		list_del(&mapping->list);
2972		amdgpu_vm_free_mapping(adev, vm, mapping, NULL);
2973	}
2974
2975	root = amdgpu_bo_ref(vm->root.base.bo);
2976	r = amdgpu_bo_reserve(root, true);
2977	if (r) {
2978		dev_err(adev->dev, "Leaking page tables because BO reservation failed\n");
2979	} else {
2980		amdgpu_vm_free_pts(adev, vm, NULL);
2981		amdgpu_bo_unreserve(root);
2982	}
2983	amdgpu_bo_unref(&root);
2984	WARN_ON(vm->root.base.bo);
2985	dma_fence_put(vm->last_update);
2986	for (i = 0; i < AMDGPU_MAX_VMHUBS; i++)
2987		amdgpu_vmid_free_reserved(adev, vm, i);
2988}
2989
2990/**
2991 * amdgpu_vm_manager_init - init the VM manager
2992 *
2993 * @adev: amdgpu_device pointer
2994 *
2995 * Initialize the VM manager structures
2996 */
2997void amdgpu_vm_manager_init(struct amdgpu_device *adev)
2998{
2999	unsigned i;
3000
 
 
 
 
 
 
3001	amdgpu_vmid_mgr_init(adev);
3002
3003	adev->vm_manager.fence_context =
3004		dma_fence_context_alloc(AMDGPU_MAX_RINGS);
3005	for (i = 0; i < AMDGPU_MAX_RINGS; ++i)
3006		adev->vm_manager.seqno[i] = 0;
3007
3008	spin_lock_init(&adev->vm_manager.prt_lock);
3009	atomic_set(&adev->vm_manager.num_prt_users, 0);
3010
3011	/* If not overridden by the user, by default, only in large BAR systems
3012	 * Compute VM tables will be updated by CPU
3013	 */
3014#ifdef CONFIG_X86_64
3015	if (amdgpu_vm_update_mode == -1) {
3016		if (amdgpu_gmc_vram_full_visible(&adev->gmc))
3017			adev->vm_manager.vm_update_mode =
3018				AMDGPU_VM_USE_CPU_FOR_COMPUTE;
3019		else
3020			adev->vm_manager.vm_update_mode = 0;
3021	} else
3022		adev->vm_manager.vm_update_mode = amdgpu_vm_update_mode;
3023#else
3024	adev->vm_manager.vm_update_mode = 0;
3025#endif
3026
3027	idr_init(&adev->vm_manager.pasid_idr);
3028	spin_lock_init(&adev->vm_manager.pasid_lock);
3029
3030	adev->vm_manager.xgmi_map_counter = 0;
3031	mutex_init(&adev->vm_manager.lock_pstate);
3032}
3033
3034/**
3035 * amdgpu_vm_manager_fini - cleanup VM manager
3036 *
3037 * @adev: amdgpu_device pointer
3038 *
3039 * Cleanup the VM manager and free resources.
3040 */
3041void amdgpu_vm_manager_fini(struct amdgpu_device *adev)
3042{
3043	WARN_ON(!idr_is_empty(&adev->vm_manager.pasid_idr));
3044	idr_destroy(&adev->vm_manager.pasid_idr);
3045
3046	amdgpu_vmid_mgr_fini(adev);
3047}
3048
3049/**
3050 * amdgpu_vm_ioctl - Manages VMID reservation for vm hubs.
3051 *
3052 * @dev: drm device pointer
3053 * @data: drm_amdgpu_vm
3054 * @filp: drm file pointer
3055 *
3056 * Returns:
3057 * 0 for success, -errno for errors.
3058 */
3059int amdgpu_vm_ioctl(struct drm_device *dev, void *data, struct drm_file *filp)
3060{
3061	union drm_amdgpu_vm *args = data;
3062	struct amdgpu_device *adev = dev->dev_private;
3063	struct amdgpu_fpriv *fpriv = filp->driver_priv;
 
3064	int r;
3065
3066	switch (args->in.op) {
3067	case AMDGPU_VM_OP_RESERVE_VMID:
3068		/* current, we only have requirement to reserve vmid from gfxhub */
3069		r = amdgpu_vmid_alloc_reserved(adev, &fpriv->vm, AMDGPU_GFXHUB_0);
 
3070		if (r)
3071			return r;
3072		break;
3073	case AMDGPU_VM_OP_UNRESERVE_VMID:
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3074		amdgpu_vmid_free_reserved(adev, &fpriv->vm, AMDGPU_GFXHUB_0);
3075		break;
3076	default:
3077		return -EINVAL;
3078	}
3079
3080	return 0;
3081}
3082
3083/**
3084 * amdgpu_vm_get_task_info - Extracts task info for a PASID.
3085 *
3086 * @adev: drm device pointer
3087 * @pasid: PASID identifier for VM
3088 * @task_info: task_info to fill.
3089 */
3090void amdgpu_vm_get_task_info(struct amdgpu_device *adev, unsigned int pasid,
3091			 struct amdgpu_task_info *task_info)
3092{
3093	struct amdgpu_vm *vm;
3094	unsigned long flags;
3095
3096	spin_lock_irqsave(&adev->vm_manager.pasid_lock, flags);
3097
3098	vm = idr_find(&adev->vm_manager.pasid_idr, pasid);
3099	if (vm)
3100		*task_info = vm->task_info;
3101
3102	spin_unlock_irqrestore(&adev->vm_manager.pasid_lock, flags);
3103}
3104
3105/**
3106 * amdgpu_vm_set_task_info - Sets VMs task info.
3107 *
3108 * @vm: vm for which to set the info
3109 */
3110void amdgpu_vm_set_task_info(struct amdgpu_vm *vm)
3111{
3112	if (!vm->task_info.pid) {
3113		vm->task_info.pid = current->pid;
3114		get_task_comm(vm->task_info.task_name, current);
3115
3116		if (current->group_leader->mm == current->mm) {
3117			vm->task_info.tgid = current->group_leader->pid;
3118			get_task_comm(vm->task_info.process_name, current->group_leader);
3119		}
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3120	}
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3121}