drivers/gpu/drm/xe/xe_vm.c

   1 // SPDX-License-Identifier: MIT
   2 /*
   3  * Copyright © 2021 Intel Corporation
   4  */
   5
   6 #include "xe_vm.h"
   7
   8 #include <linux/dma-fence-array.h>
   9
  10 #include <drm/drm_exec.h>
  11 #include <drm/drm_print.h>
  12 #include <drm/ttm/ttm_execbuf_util.h>
  13 #include <drm/ttm/ttm_tt.h>
  14 #include <drm/xe_drm.h>
  15 #include <linux/delay.h>
  16 #include <linux/kthread.h>
  17 #include <linux/mm.h>
  18 #include <linux/swap.h>
  19
  20 #include "xe_assert.h"
  21 #include "xe_bo.h"
  22 #include "xe_device.h"
  23 #include "xe_drm_client.h"
  24 #include "xe_exec_queue.h"
  25 #include "xe_gt.h"
  26 #include "xe_gt_pagefault.h"
  27 #include "xe_gt_tlb_invalidation.h"
  28 #include "xe_migrate.h"
  29 #include "xe_pm.h"
  30 #include "xe_preempt_fence.h"
  31 #include "xe_pt.h"
  32 #include "xe_res_cursor.h"
  33 #include "xe_sync.h"
  34 #include "xe_trace.h"
  35 #include "generated/xe_wa_oob.h"
  36 #include "xe_wa.h"
  37
  38 #define TEST_VM_ASYNC_OPS_ERROR
  39
  40 static struct drm_gem_object *xe_vm_obj(struct xe_vm *vm)
  41 {
  42         return vm->gpuvm.r_obj;
  43 }
  44
  45 /**
  46  * xe_vma_userptr_check_repin() - Advisory check for repin needed
  47  * @vma: The userptr vma
  48  *
  49  * Check if the userptr vma has been invalidated since last successful
  50  * repin. The check is advisory only and can the function can be called
  51  * without the vm->userptr.notifier_lock held. There is no guarantee that the
  52  * vma userptr will remain valid after a lockless check, so typically
  53  * the call needs to be followed by a proper check under the notifier_lock.
  54  *
  55  * Return: 0 if userptr vma is valid, -EAGAIN otherwise; repin recommended.
  56  */
  57 int xe_vma_userptr_check_repin(struct xe_vma *vma)
  58 {
  59         return mmu_interval_check_retry(&vma->userptr.notifier,
  60                                         vma->userptr.notifier_seq) ?
  61                 -EAGAIN : 0;
  62 }
  63
  64 int xe_vma_userptr_pin_pages(struct xe_vma *vma)
  65 {
  66         struct xe_vm *vm = xe_vma_vm(vma);
  67         struct xe_device *xe = vm->xe;
  68         const unsigned long num_pages = xe_vma_size(vma) >> PAGE_SHIFT;
  69         struct page **pages;
  70         bool in_kthread = !current->mm;
  71         unsigned long notifier_seq;
  72         int pinned, ret, i;
  73         bool read_only = xe_vma_read_only(vma);
  74
  75         lockdep_assert_held(&vm->lock);
  76         xe_assert(xe, xe_vma_is_userptr(vma));
  77 retry:
  78         if (vma->gpuva.flags & XE_VMA_DESTROYED)
  79                 return 0;
  80
  81         notifier_seq = mmu_interval_read_begin(&vma->userptr.notifier);
  82         if (notifier_seq == vma->userptr.notifier_seq)
  83                 return 0;
  84
  85         pages = kvmalloc_array(num_pages, sizeof(*pages), GFP_KERNEL);
  86         if (!pages)
  87                 return -ENOMEM;
  88
  89         if (vma->userptr.sg) {
  90                 dma_unmap_sgtable(xe->drm.dev,
  91                                   vma->userptr.sg,
  92                                   read_only ? DMA_TO_DEVICE :
  93                                   DMA_BIDIRECTIONAL, 0);
  94                 sg_free_table(vma->userptr.sg);
  95                 vma->userptr.sg = NULL;
  96         }
  97
  98         pinned = ret = 0;
  99         if (in_kthread) {
 100                 if (!mmget_not_zero(vma->userptr.notifier.mm)) {
 101                         ret = -EFAULT;
 102                         goto mm_closed;
 103                 }
 104                 kthread_use_mm(vma->userptr.notifier.mm);
 105         }
 106
 107         while (pinned < num_pages) {
 108                 ret = get_user_pages_fast(xe_vma_userptr(vma) +
 109                                           pinned * PAGE_SIZE,
 110                                           num_pages - pinned,
 111                                           read_only ? 0 : FOLL_WRITE,
 112                                           &pages[pinned]);
 113                 if (ret < 0) {
 114                         if (in_kthread)
 115                                 ret = 0;
 116                         break;
 117                 }
 118
 119                 pinned += ret;
 120                 ret = 0;
 121         }
 122
 123         if (in_kthread) {
 124                 kthread_unuse_mm(vma->userptr.notifier.mm);
 125                 mmput(vma->userptr.notifier.mm);
 126         }
 127 mm_closed:
 128         if (ret)
 129                 goto out;
 130
 131         ret = sg_alloc_table_from_pages_segment(&vma->userptr.sgt, pages,
 132                                                 pinned, 0,
 133                                                 (u64)pinned << PAGE_SHIFT,
 134                                                 xe_sg_segment_size(xe->drm.dev),
 135                                                 GFP_KERNEL);
 136         if (ret) {
 137                 vma->userptr.sg = NULL;
 138                 goto out;
 139         }
 140         vma->userptr.sg = &vma->userptr.sgt;
 141
 142         ret = dma_map_sgtable(xe->drm.dev, vma->userptr.sg,
 143                               read_only ? DMA_TO_DEVICE :
 144                               DMA_BIDIRECTIONAL,
 145                               DMA_ATTR_SKIP_CPU_SYNC |
 146                               DMA_ATTR_NO_KERNEL_MAPPING);
 147         if (ret) {
 148                 sg_free_table(vma->userptr.sg);
 149                 vma->userptr.sg = NULL;
 150                 goto out;
 151         }
 152
 153         for (i = 0; i < pinned; ++i) {
 154                 if (!read_only) {
 155                         lock_page(pages[i]);
 156                         set_page_dirty(pages[i]);
 157                         unlock_page(pages[i]);
 158                 }
 159
 160                 mark_page_accessed(pages[i]);
 161         }
 162
 163 out:
 164         release_pages(pages, pinned);
 165         kvfree(pages);
 166
 167         if (!(ret < 0)) {
 168                 vma->userptr.notifier_seq = notifier_seq;
 169                 if (xe_vma_userptr_check_repin(vma) == -EAGAIN)
 170                         goto retry;
 171         }
 172
 173         return ret < 0 ? ret : 0;
 174 }
 175
 176 static bool preempt_fences_waiting(struct xe_vm *vm)
 177 {
 178         struct xe_exec_queue *q;
 179
 180         lockdep_assert_held(&vm->lock);
 181         xe_vm_assert_held(vm);
 182
 183         list_for_each_entry(q, &vm->preempt.exec_queues, compute.link) {
 184                 if (!q->compute.pfence ||
 185                     (q->compute.pfence && test_bit(DMA_FENCE_FLAG_ENABLE_SIGNAL_BIT,
 186                                                    &q->compute.pfence->flags))) {
 187                         return true;
 188                 }
 189         }
 190
 191         return false;
 192 }
 193
 194 static void free_preempt_fences(struct list_head *list)
 195 {
 196         struct list_head *link, *next;
 197
 198         list_for_each_safe(link, next, list)
 199                 xe_preempt_fence_free(to_preempt_fence_from_link(link));
 200 }
 201
 202 static int alloc_preempt_fences(struct xe_vm *vm, struct list_head *list,
 203                                 unsigned int *count)
 204 {
 205         lockdep_assert_held(&vm->lock);
 206         xe_vm_assert_held(vm);
 207
 208         if (*count >= vm->preempt.num_exec_queues)
 209                 return 0;
 210
 211         for (; *count < vm->preempt.num_exec_queues; ++(*count)) {
 212                 struct xe_preempt_fence *pfence = xe_preempt_fence_alloc();
 213
 214                 if (IS_ERR(pfence))
 215                         return PTR_ERR(pfence);
 216
 217                 list_move_tail(xe_preempt_fence_link(pfence), list);
 218         }
 219
 220         return 0;
 221 }
 222
 223 static int wait_for_existing_preempt_fences(struct xe_vm *vm)
 224 {
 225         struct xe_exec_queue *q;
 226
 227         xe_vm_assert_held(vm);
 228
 229         list_for_each_entry(q, &vm->preempt.exec_queues, compute.link) {
 230                 if (q->compute.pfence) {
 231                         long timeout = dma_fence_wait(q->compute.pfence, false);
 232
 233                         if (timeout < 0)
 234                                 return -ETIME;
 235                         dma_fence_put(q->compute.pfence);
 236                         q->compute.pfence = NULL;
 237                 }
 238         }
 239
 240         return 0;
 241 }
 242
 243 static bool xe_vm_is_idle(struct xe_vm *vm)
 244 {
 245         struct xe_exec_queue *q;
 246
 247         xe_vm_assert_held(vm);
 248         list_for_each_entry(q, &vm->preempt.exec_queues, compute.link) {
 249                 if (!xe_exec_queue_is_idle(q))
 250                         return false;
 251         }
 252
 253         return true;
 254 }
 255
 256 static void arm_preempt_fences(struct xe_vm *vm, struct list_head *list)
 257 {
 258         struct list_head *link;
 259         struct xe_exec_queue *q;
 260
 261         list_for_each_entry(q, &vm->preempt.exec_queues, compute.link) {
 262                 struct dma_fence *fence;
 263
 264                 link = list->next;
 265                 xe_assert(vm->xe, link != list);
 266
 267                 fence = xe_preempt_fence_arm(to_preempt_fence_from_link(link),
 268                                              q, q->compute.context,
 269                                              ++q->compute.seqno);
 270                 dma_fence_put(q->compute.pfence);
 271                 q->compute.pfence = fence;
 272         }
 273 }
 274
 275 static int add_preempt_fences(struct xe_vm *vm, struct xe_bo *bo)
 276 {
 277         struct xe_exec_queue *q;
 278         int err;
 279
 280         err = xe_bo_lock(bo, true);
 281         if (err)
 282                 return err;
 283
 284         err = dma_resv_reserve_fences(bo->ttm.base.resv, vm->preempt.num_exec_queues);
 285         if (err)
 286                 goto out_unlock;
 287
 288         list_for_each_entry(q, &vm->preempt.exec_queues, compute.link)
 289                 if (q->compute.pfence) {
 290                         dma_resv_add_fence(bo->ttm.base.resv,
 291                                            q->compute.pfence,
 292                                            DMA_RESV_USAGE_BOOKKEEP);
 293                 }
 294
 295 out_unlock:
 296         xe_bo_unlock(bo);
 297         return err;
 298 }
 299
 300 /**
 301  * xe_vm_fence_all_extobjs() - Add a fence to vm's external objects' resv
 302  * @vm: The vm.
 303  * @fence: The fence to add.
 304  * @usage: The resv usage for the fence.
 305  *
 306  * Loops over all of the vm's external object bindings and adds a @fence
 307  * with the given @usage to all of the external object's reservation
 308  * objects.
 309  */
 310 void xe_vm_fence_all_extobjs(struct xe_vm *vm, struct dma_fence *fence,
 311                              enum dma_resv_usage usage)
 312 {
 313         struct xe_vma *vma;
 314
 315         list_for_each_entry(vma, &vm->extobj.list, extobj.link)
 316                 dma_resv_add_fence(xe_vma_bo(vma)->ttm.base.resv, fence, usage);
 317 }
 318
 319 static void resume_and_reinstall_preempt_fences(struct xe_vm *vm)
 320 {
 321         struct xe_exec_queue *q;
 322
 323         lockdep_assert_held(&vm->lock);
 324         xe_vm_assert_held(vm);
 325
 326         list_for_each_entry(q, &vm->preempt.exec_queues, compute.link) {
 327                 q->ops->resume(q);
 328
 329                 dma_resv_add_fence(xe_vm_resv(vm), q->compute.pfence,
 330                                    DMA_RESV_USAGE_BOOKKEEP);
 331                 xe_vm_fence_all_extobjs(vm, q->compute.pfence,
 332                                         DMA_RESV_USAGE_BOOKKEEP);
 333         }
 334 }
 335
 336 int xe_vm_add_compute_exec_queue(struct xe_vm *vm, struct xe_exec_queue *q)
 337 {
 338         struct drm_exec exec;
 339         struct dma_fence *pfence;
 340         int err;
 341         bool wait;
 342
 343         xe_assert(vm->xe, xe_vm_in_compute_mode(vm));
 344
 345         down_write(&vm->lock);
 346         drm_exec_init(&exec, DRM_EXEC_INTERRUPTIBLE_WAIT);
 347         drm_exec_until_all_locked(&exec) {
 348                 err = xe_vm_lock_dma_resv(vm, &exec, 1, true);
 349                 drm_exec_retry_on_contention(&exec);
 350                 if (err)
 351                         goto out_unlock;
 352         }
 353
 354         pfence = xe_preempt_fence_create(q, q->compute.context,
 355                                          ++q->compute.seqno);
 356         if (!pfence) {
 357                 err = -ENOMEM;
 358                 goto out_unlock;
 359         }
 360
 361         list_add(&q->compute.link, &vm->preempt.exec_queues);
 362         ++vm->preempt.num_exec_queues;
 363         q->compute.pfence = pfence;
 364
 365         down_read(&vm->userptr.notifier_lock);
 366
 367         dma_resv_add_fence(xe_vm_resv(vm), pfence,
 368                            DMA_RESV_USAGE_BOOKKEEP);
 369
 370         xe_vm_fence_all_extobjs(vm, pfence, DMA_RESV_USAGE_BOOKKEEP);
 371
 372         /*
 373          * Check to see if a preemption on VM is in flight or userptr
 374          * invalidation, if so trigger this preempt fence to sync state with
 375          * other preempt fences on the VM.
 376          */
 377         wait = __xe_vm_userptr_needs_repin(vm) || preempt_fences_waiting(vm);
 378         if (wait)
 379                 dma_fence_enable_sw_signaling(pfence);
 380
 381         up_read(&vm->userptr.notifier_lock);
 382
 383 out_unlock:
 384         drm_exec_fini(&exec);
 385         up_write(&vm->lock);
 386
 387         return err;
 388 }
 389
 390 /**
 391  * xe_vm_remove_compute_exec_queue() - Remove compute exec queue from VM
 392  * @vm: The VM.
 393  * @q: The exec_queue
 394  */
 395 void xe_vm_remove_compute_exec_queue(struct xe_vm *vm, struct xe_exec_queue *q)
 396 {
 397         if (!xe_vm_in_compute_mode(vm))
 398                 return;
 399
 400         down_write(&vm->lock);
 401         list_del(&q->compute.link);
 402         --vm->preempt.num_exec_queues;
 403         if (q->compute.pfence) {
 404                 dma_fence_enable_sw_signaling(q->compute.pfence);
 405                 dma_fence_put(q->compute.pfence);
 406                 q->compute.pfence = NULL;
 407         }
 408         up_write(&vm->lock);
 409 }
 410
 411 /**
 412  * __xe_vm_userptr_needs_repin() - Check whether the VM does have userptrs
 413  * that need repinning.
 414  * @vm: The VM.
 415  *
 416  * This function checks for whether the VM has userptrs that need repinning,
 417  * and provides a release-type barrier on the userptr.notifier_lock after
 418  * checking.
 419  *
 420  * Return: 0 if there are no userptrs needing repinning, -EAGAIN if there are.
 421  */
 422 int __xe_vm_userptr_needs_repin(struct xe_vm *vm)
 423 {
 424         lockdep_assert_held_read(&vm->userptr.notifier_lock);
 425
 426         return (list_empty(&vm->userptr.repin_list) &&
 427                 list_empty(&vm->userptr.invalidated)) ? 0 : -EAGAIN;
 428 }
 429
 430 /**
 431  * xe_vm_lock_dma_resv() - Lock the vm dma_resv object and the dma_resv
 432  * objects of the vm's external buffer objects.
 433  * @vm: The vm.
 434  * @exec: Pointer to a struct drm_exec locking context.
 435  * @num_shared: Number of dma-fence slots to reserve in the locked objects.
 436  * @lock_vm: Lock also the vm's dma_resv.
 437  *
 438  * Locks the vm dma-resv objects and all the dma-resv objects of the
 439  * buffer objects on the vm external object list.
 440  *
 441  * Return: 0 on success, Negative error code on error. In particular if
 442  * @intr is set to true, -EINTR or -ERESTARTSYS may be returned.
 443  */
 444 int xe_vm_lock_dma_resv(struct xe_vm *vm, struct drm_exec *exec,
 445                         unsigned int num_shared, bool lock_vm)
 446 {
 447         struct xe_vma *vma, *next;
 448         int err = 0;
 449
 450         lockdep_assert_held(&vm->lock);
 451
 452         if (lock_vm) {
 453                 err = drm_exec_prepare_obj(exec, xe_vm_obj(vm), num_shared);
 454                 if (err)
 455                         return err;
 456         }
 457
 458         list_for_each_entry(vma, &vm->extobj.list, extobj.link) {
 459                 err = drm_exec_prepare_obj(exec, &xe_vma_bo(vma)->ttm.base, num_shared);
 460                 if (err)
 461                         return err;
 462         }
 463
 464         spin_lock(&vm->notifier.list_lock);
 465         list_for_each_entry_safe(vma, next, &vm->notifier.rebind_list,
 466                                  notifier.rebind_link) {
 467                 xe_bo_assert_held(xe_vma_bo(vma));
 468
 469                 list_del_init(&vma->notifier.rebind_link);
 470                 if (vma->tile_present && !(vma->gpuva.flags & XE_VMA_DESTROYED))
 471                         list_move_tail(&vma->combined_links.rebind,
 472                                        &vm->rebind_list);
 473         }
 474         spin_unlock(&vm->notifier.list_lock);
 475
 476         return 0;
 477 }
 478
 479 #define XE_VM_REBIND_RETRY_TIMEOUT_MS 1000
 480
 481 static void xe_vm_kill(struct xe_vm *vm)
 482 {
 483         struct xe_exec_queue *q;
 484
 485         lockdep_assert_held(&vm->lock);
 486
 487         xe_vm_lock(vm, false);
 488         vm->flags |= XE_VM_FLAG_BANNED;
 489         trace_xe_vm_kill(vm);
 490
 491         list_for_each_entry(q, &vm->preempt.exec_queues, compute.link)
 492                 q->ops->kill(q);
 493         xe_vm_unlock(vm);
 494
 495         /* TODO: Inform user the VM is banned */
 496 }
 497
 498 /**
 499  * xe_vm_validate_should_retry() - Whether to retry after a validate error.
 500  * @exec: The drm_exec object used for locking before validation.
 501  * @err: The error returned from ttm_bo_validate().
 502  * @end: A ktime_t cookie that should be set to 0 before first use and
 503  * that should be reused on subsequent calls.
 504  *
 505  * With multiple active VMs, under memory pressure, it is possible that
 506  * ttm_bo_validate() run into -EDEADLK and in such case returns -ENOMEM.
 507  * Until ttm properly handles locking in such scenarios, best thing the
 508  * driver can do is retry with a timeout. Check if that is necessary, and
 509  * if so unlock the drm_exec's objects while keeping the ticket to prepare
 510  * for a rerun.
 511  *
 512  * Return: true if a retry after drm_exec_init() is recommended;
 513  * false otherwise.
 514  */
 515 bool xe_vm_validate_should_retry(struct drm_exec *exec, int err, ktime_t *end)
 516 {
 517         ktime_t cur;
 518
 519         if (err != -ENOMEM)
 520                 return false;
 521
 522         cur = ktime_get();
 523         *end = *end ? : ktime_add_ms(cur, XE_VM_REBIND_RETRY_TIMEOUT_MS);
 524         if (!ktime_before(cur, *end))
 525                 return false;
 526
 527         /*
 528          * We would like to keep the ticket here with
 529          * drm_exec_unlock_all(), but WW mutex asserts currently
 530          * stop us from that. In any case this function could go away
 531          * with proper TTM -EDEADLK handling.
 532          */
 533         drm_exec_fini(exec);
 534
 535         msleep(20);
 536         return true;
 537 }
 538
 539 static int xe_preempt_work_begin(struct drm_exec *exec, struct xe_vm *vm,
 540                                  bool *done)
 541 {
 542         struct xe_vma *vma;
 543         int err;
 544
 545         /*
 546          * 1 fence for each preempt fence plus a fence for each tile from a
 547          * possible rebind
 548          */
 549         err = drm_exec_prepare_obj(exec, xe_vm_obj(vm),
 550                                    vm->preempt.num_exec_queues +
 551                                    vm->xe->info.tile_count);
 552         if (err)
 553                 return err;
 554
 555         if (xe_vm_is_idle(vm)) {
 556                 vm->preempt.rebind_deactivated = true;
 557                 *done = true;
 558                 return 0;
 559         }
 560
 561         if (!preempt_fences_waiting(vm)) {
 562                 *done = true;
 563                 return 0;
 564         }
 565
 566         err = xe_vm_lock_dma_resv(vm, exec, vm->preempt.num_exec_queues, false);
 567         if (err)
 568                 return err;
 569
 570         err = wait_for_existing_preempt_fences(vm);
 571         if (err)
 572                 return err;
 573
 574         list_for_each_entry(vma, &vm->rebind_list, combined_links.rebind) {
 575                 if (xe_vma_has_no_bo(vma) ||
 576                     vma->gpuva.flags & XE_VMA_DESTROYED)
 577                         continue;
 578
 579                 err = xe_bo_validate(xe_vma_bo(vma), vm, false);
 580                 if (err)
 581                         break;
 582         }
 583
 584         return err;
 585 }
 586
 587 static void preempt_rebind_work_func(struct work_struct *w)
 588 {
 589         struct xe_vm *vm = container_of(w, struct xe_vm, preempt.rebind_work);
 590         struct drm_exec exec;
 591         struct dma_fence *rebind_fence;
 592         unsigned int fence_count = 0;
 593         LIST_HEAD(preempt_fences);
 594         ktime_t end = 0;
 595         int err = 0;
 596         long wait;
 597         int __maybe_unused tries = 0;
 598
 599         xe_assert(vm->xe, xe_vm_in_compute_mode(vm));
 600         trace_xe_vm_rebind_worker_enter(vm);
 601
 602         down_write(&vm->lock);
 603
 604         if (xe_vm_is_closed_or_banned(vm)) {
 605                 up_write(&vm->lock);
 606                 trace_xe_vm_rebind_worker_exit(vm);
 607                 return;
 608         }
 609
 610 retry:
 611         if (xe_vm_userptr_check_repin(vm)) {
 612                 err = xe_vm_userptr_pin(vm);
 613                 if (err)
 614                         goto out_unlock_outer;
 615         }
 616
 617         drm_exec_init(&exec, DRM_EXEC_INTERRUPTIBLE_WAIT);
 618
 619         drm_exec_until_all_locked(&exec) {
 620                 bool done = false;
 621
 622                 err = xe_preempt_work_begin(&exec, vm, &done);
 623                 drm_exec_retry_on_contention(&exec);
 624                 if (err && xe_vm_validate_should_retry(&exec, err, &end)) {
 625                         err = -EAGAIN;
 626                         goto out_unlock_outer;
 627                 }
 628                 if (err || done)
 629                         goto out_unlock;
 630         }
 631
 632         err = alloc_preempt_fences(vm, &preempt_fences, &fence_count);
 633         if (err)
 634                 goto out_unlock;
 635
 636         rebind_fence = xe_vm_rebind(vm, true);
 637         if (IS_ERR(rebind_fence)) {
 638                 err = PTR_ERR(rebind_fence);
 639                 goto out_unlock;
 640         }
 641
 642         if (rebind_fence) {
 643                 dma_fence_wait(rebind_fence, false);
 644                 dma_fence_put(rebind_fence);
 645         }
 646
 647         /* Wait on munmap style VM unbinds */
 648         wait = dma_resv_wait_timeout(xe_vm_resv(vm),
 649                                      DMA_RESV_USAGE_KERNEL,
 650                                      false, MAX_SCHEDULE_TIMEOUT);
 651         if (wait <= 0) {
 652                 err = -ETIME;
 653                 goto out_unlock;
 654         }
 655
 656 #define retry_required(__tries, __vm) \
 657         (IS_ENABLED(CONFIG_DRM_XE_USERPTR_INVAL_INJECT) ? \
 658         (!(__tries)++ || __xe_vm_userptr_needs_repin(__vm)) : \
 659         __xe_vm_userptr_needs_repin(__vm))
 660
 661         down_read(&vm->userptr.notifier_lock);
 662         if (retry_required(tries, vm)) {
 663                 up_read(&vm->userptr.notifier_lock);
 664                 err = -EAGAIN;
 665                 goto out_unlock;
 666         }
 667
 668 #undef retry_required
 669
 670         spin_lock(&vm->xe->ttm.lru_lock);
 671         ttm_lru_bulk_move_tail(&vm->lru_bulk_move);
 672         spin_unlock(&vm->xe->ttm.lru_lock);
 673
 674         /* Point of no return. */
 675         arm_preempt_fences(vm, &preempt_fences);
 676         resume_and_reinstall_preempt_fences(vm);
 677         up_read(&vm->userptr.notifier_lock);
 678
 679 out_unlock:
 680         drm_exec_fini(&exec);
 681 out_unlock_outer:
 682         if (err == -EAGAIN) {
 683                 trace_xe_vm_rebind_worker_retry(vm);
 684                 goto retry;
 685         }
 686
 687         if (err) {
 688                 drm_warn(&vm->xe->drm, "VM worker error: %d\n", err);
 689                 xe_vm_kill(vm);
 690         }
 691         up_write(&vm->lock);
 692
 693         free_preempt_fences(&preempt_fences);
 694
 695         trace_xe_vm_rebind_worker_exit(vm);
 696 }
 697
 698 static bool vma_userptr_invalidate(struct mmu_interval_notifier *mni,
 699                                    const struct mmu_notifier_range *range,
 700                                    unsigned long cur_seq)
 701 {
 702         struct xe_vma *vma = container_of(mni, struct xe_vma, userptr.notifier);
 703         struct xe_vm *vm = xe_vma_vm(vma);
 704         struct dma_resv_iter cursor;
 705         struct dma_fence *fence;
 706         long err;
 707
 708         xe_assert(vm->xe, xe_vma_is_userptr(vma));
 709         trace_xe_vma_userptr_invalidate(vma);
 710
 711         if (!mmu_notifier_range_blockable(range))
 712                 return false;
 713
 714         down_write(&vm->userptr.notifier_lock);
 715         mmu_interval_set_seq(mni, cur_seq);
 716
 717         /* No need to stop gpu access if the userptr is not yet bound. */
 718         if (!vma->userptr.initial_bind) {
 719                 up_write(&vm->userptr.notifier_lock);
 720                 return true;
 721         }
 722
 723         /*
 724          * Tell exec and rebind worker they need to repin and rebind this
 725          * userptr.
 726          */
 727         if (!xe_vm_in_fault_mode(vm) &&
 728             !(vma->gpuva.flags & XE_VMA_DESTROYED) && vma->tile_present) {
 729                 spin_lock(&vm->userptr.invalidated_lock);
 730                 list_move_tail(&vma->userptr.invalidate_link,
 731                                &vm->userptr.invalidated);
 732                 spin_unlock(&vm->userptr.invalidated_lock);
 733         }
 734
 735         up_write(&vm->userptr.notifier_lock);
 736
 737         /*
 738          * Preempt fences turn into schedule disables, pipeline these.
 739          * Note that even in fault mode, we need to wait for binds and
 740          * unbinds to complete, and those are attached as BOOKMARK fences
 741          * to the vm.
 742          */
 743         dma_resv_iter_begin(&cursor, xe_vm_resv(vm),
 744                             DMA_RESV_USAGE_BOOKKEEP);
 745         dma_resv_for_each_fence_unlocked(&cursor, fence)
 746                 dma_fence_enable_sw_signaling(fence);
 747         dma_resv_iter_end(&cursor);
 748
 749         err = dma_resv_wait_timeout(xe_vm_resv(vm),
 750                                     DMA_RESV_USAGE_BOOKKEEP,
 751                                     false, MAX_SCHEDULE_TIMEOUT);
 752         XE_WARN_ON(err <= 0);
 753
 754         if (xe_vm_in_fault_mode(vm)) {
 755                 err = xe_vm_invalidate_vma(vma);
 756                 XE_WARN_ON(err);
 757         }
 758
 759         trace_xe_vma_userptr_invalidate_complete(vma);
 760
 761         return true;
 762 }
 763
 764 static const struct mmu_interval_notifier_ops vma_userptr_notifier_ops = {
 765         .invalidate = vma_userptr_invalidate,
 766 };
 767
 768 int xe_vm_userptr_pin(struct xe_vm *vm)
 769 {
 770         struct xe_vma *vma, *next;
 771         int err = 0;
 772         LIST_HEAD(tmp_evict);
 773
 774         lockdep_assert_held_write(&vm->lock);
 775
 776         /* Collect invalidated userptrs */
 777         spin_lock(&vm->userptr.invalidated_lock);
 778         list_for_each_entry_safe(vma, next, &vm->userptr.invalidated,
 779                                  userptr.invalidate_link) {
 780                 list_del_init(&vma->userptr.invalidate_link);
 781                 if (list_empty(&vma->combined_links.userptr))
 782                         list_move_tail(&vma->combined_links.userptr,
 783                                        &vm->userptr.repin_list);
 784         }
 785         spin_unlock(&vm->userptr.invalidated_lock);
 786
 787         /* Pin and move to temporary list */
 788         list_for_each_entry_safe(vma, next, &vm->userptr.repin_list,
 789                                  combined_links.userptr) {
 790                 err = xe_vma_userptr_pin_pages(vma);
 791                 if (err < 0)
 792                         goto out_err;
 793
 794                 list_move_tail(&vma->combined_links.userptr, &tmp_evict);
 795         }
 796
 797         /* Take lock and move to rebind_list for rebinding. */
 798         err = dma_resv_lock_interruptible(xe_vm_resv(vm), NULL);
 799         if (err)
 800                 goto out_err;
 801
 802         list_for_each_entry_safe(vma, next, &tmp_evict, combined_links.userptr)
 803                 list_move_tail(&vma->combined_links.rebind, &vm->rebind_list);
 804
 805         dma_resv_unlock(xe_vm_resv(vm));
 806
 807         return 0;
 808
 809 out_err:
 810         list_splice_tail(&tmp_evict, &vm->userptr.repin_list);
 811
 812         return err;
 813 }
 814
 815 /**
 816  * xe_vm_userptr_check_repin() - Check whether the VM might have userptrs
 817  * that need repinning.
 818  * @vm: The VM.
 819  *
 820  * This function does an advisory check for whether the VM has userptrs that
 821  * need repinning.
 822  *
 823  * Return: 0 if there are no indications of userptrs needing repinning,
 824  * -EAGAIN if there are.
 825  */
 826 int xe_vm_userptr_check_repin(struct xe_vm *vm)
 827 {
 828         return (list_empty_careful(&vm->userptr.repin_list) &&
 829                 list_empty_careful(&vm->userptr.invalidated)) ? 0 : -EAGAIN;
 830 }
 831
 832 static struct dma_fence *
 833 xe_vm_bind_vma(struct xe_vma *vma, struct xe_exec_queue *q,
 834                struct xe_sync_entry *syncs, u32 num_syncs,
 835                bool first_op, bool last_op);
 836
 837 struct dma_fence *xe_vm_rebind(struct xe_vm *vm, bool rebind_worker)
 838 {
 839         struct dma_fence *fence = NULL;
 840         struct xe_vma *vma, *next;
 841
 842         lockdep_assert_held(&vm->lock);
 843         if (xe_vm_no_dma_fences(vm) && !rebind_worker)
 844                 return NULL;
 845
 846         xe_vm_assert_held(vm);
 847         list_for_each_entry_safe(vma, next, &vm->rebind_list,
 848                                  combined_links.rebind) {
 849                 xe_assert(vm->xe, vma->tile_present);
 850
 851                 list_del_init(&vma->combined_links.rebind);
 852                 dma_fence_put(fence);
 853                 if (rebind_worker)
 854                         trace_xe_vma_rebind_worker(vma);
 855                 else
 856                         trace_xe_vma_rebind_exec(vma);
 857                 fence = xe_vm_bind_vma(vma, NULL, NULL, 0, false, false);
 858                 if (IS_ERR(fence))
 859                         return fence;
 860         }
 861
 862         return fence;
 863 }
 864
 865 static struct xe_vma *xe_vma_create(struct xe_vm *vm,
 866                                     struct xe_bo *bo,
 867                                     u64 bo_offset_or_userptr,
 868                                     u64 start, u64 end,
 869                                     bool read_only,
 870                                     bool is_null,
 871                                     u8 tile_mask)
 872 {
 873         struct xe_vma *vma;
 874         struct xe_tile *tile;
 875         u8 id;
 876
 877         xe_assert(vm->xe, start < end);
 878         xe_assert(vm->xe, end < vm->size);
 879
 880         if (!bo && !is_null)    /* userptr */
 881                 vma = kzalloc(sizeof(*vma), GFP_KERNEL);
 882         else
 883                 vma = kzalloc(sizeof(*vma) - sizeof(struct xe_userptr),
 884                               GFP_KERNEL);
 885         if (!vma) {
 886                 vma = ERR_PTR(-ENOMEM);
 887                 return vma;
 888         }
 889
 890         INIT_LIST_HEAD(&vma->combined_links.rebind);
 891         INIT_LIST_HEAD(&vma->notifier.rebind_link);
 892         INIT_LIST_HEAD(&vma->extobj.link);
 893
 894         INIT_LIST_HEAD(&vma->gpuva.gem.entry);
 895         vma->gpuva.vm = &vm->gpuvm;
 896         vma->gpuva.va.addr = start;
 897         vma->gpuva.va.range = end - start + 1;
 898         if (read_only)
 899                 vma->gpuva.flags |= XE_VMA_READ_ONLY;
 900         if (is_null)
 901                 vma->gpuva.flags |= DRM_GPUVA_SPARSE;
 902
 903         if (tile_mask) {
 904                 vma->tile_mask = tile_mask;
 905         } else {
 906                 for_each_tile(tile, vm->xe, id)
 907                         vma->tile_mask |= 0x1 << id;
 908         }
 909
 910         if (GRAPHICS_VER(vm->xe) >= 20 || vm->xe->info.platform == XE_PVC)
 911                 vma->gpuva.flags |= XE_VMA_ATOMIC_PTE_BIT;
 912
 913         if (bo) {
 914                 struct drm_gpuvm_bo *vm_bo;
 915
 916                 xe_bo_assert_held(bo);
 917
 918                 vm_bo = drm_gpuvm_bo_obtain(vma->gpuva.vm, &bo->ttm.base);
 919                 if (IS_ERR(vm_bo)) {
 920                         kfree(vma);
 921                         return ERR_CAST(vm_bo);
 922                 }
 923
 924                 drm_gem_object_get(&bo->ttm.base);
 925                 vma->gpuva.gem.obj = &bo->ttm.base;
 926                 vma->gpuva.gem.offset = bo_offset_or_userptr;
 927                 drm_gpuva_link(&vma->gpuva, vm_bo);
 928                 drm_gpuvm_bo_put(vm_bo);
 929         } else /* userptr or null */ {
 930                 if (!is_null) {
 931                         u64 size = end - start + 1;
 932                         int err;
 933
 934                         INIT_LIST_HEAD(&vma->userptr.invalidate_link);
 935                         vma->gpuva.gem.offset = bo_offset_or_userptr;
 936
 937                         err = mmu_interval_notifier_insert(&vma->userptr.notifier,
 938                                                            current->mm,
 939                                                            xe_vma_userptr(vma), size,
 940                                                            &vma_userptr_notifier_ops);
 941                         if (err) {
 942                                 kfree(vma);
 943                                 vma = ERR_PTR(err);
 944                                 return vma;
 945                         }
 946
 947                         vma->userptr.notifier_seq = LONG_MAX;
 948                 }
 949
 950                 xe_vm_get(vm);
 951         }
 952
 953         return vma;
 954 }
 955
 956 static bool vm_remove_extobj(struct xe_vma *vma)
 957 {
 958         if (!list_empty(&vma->extobj.link)) {
 959                 xe_vma_vm(vma)->extobj.entries--;
 960                 list_del_init(&vma->extobj.link);
 961                 return true;
 962         }
 963         return false;
 964 }
 965
 966 static void xe_vma_destroy_late(struct xe_vma *vma)
 967 {
 968         struct xe_vm *vm = xe_vma_vm(vma);
 969         struct xe_device *xe = vm->xe;
 970         bool read_only = xe_vma_read_only(vma);
 971
 972         if (xe_vma_is_userptr(vma)) {
 973                 if (vma->userptr.sg) {
 974                         dma_unmap_sgtable(xe->drm.dev,
 975                                           vma->userptr.sg,
 976                                           read_only ? DMA_TO_DEVICE :
 977                                           DMA_BIDIRECTIONAL, 0);
 978                         sg_free_table(vma->userptr.sg);
 979                         vma->userptr.sg = NULL;
 980                 }
 981
 982                 /*
 983                  * Since userptr pages are not pinned, we can't remove
 984                  * the notifer until we're sure the GPU is not accessing
 985                  * them anymore
 986                  */
 987                 mmu_interval_notifier_remove(&vma->userptr.notifier);
 988                 xe_vm_put(vm);
 989         } else if (xe_vma_is_null(vma)) {
 990                 xe_vm_put(vm);
 991         } else {
 992                 xe_bo_put(xe_vma_bo(vma));
 993         }
 994
 995         kfree(vma);
 996 }
 997
 998 static void vma_destroy_work_func(struct work_struct *w)
 999 {
1000         struct xe_vma *vma =
1001                 container_of(w, struct xe_vma, destroy_work);
1002
1003         xe_vma_destroy_late(vma);
1004 }
1005
1006 static struct xe_vma *
1007 bo_has_vm_references_locked(struct xe_bo *bo, struct xe_vm *vm,
1008                             struct xe_vma *ignore)
1009 {
1010         struct drm_gpuvm_bo *vm_bo;
1011         struct drm_gpuva *va;
1012         struct drm_gem_object *obj = &bo->ttm.base;
1013
1014         xe_bo_assert_held(bo);
1015
1016         drm_gem_for_each_gpuvm_bo(vm_bo, obj) {
1017                 drm_gpuvm_bo_for_each_va(va, vm_bo) {
1018                         struct xe_vma *vma = gpuva_to_vma(va);
1019
1020                         if (vma != ignore && xe_vma_vm(vma) == vm)
1021                                 return vma;
1022                 }
1023         }
1024
1025         return NULL;
1026 }
1027
1028 static bool bo_has_vm_references(struct xe_bo *bo, struct xe_vm *vm,
1029                                  struct xe_vma *ignore)
1030 {
1031         bool ret;
1032
1033         xe_bo_lock(bo, false);
1034         ret = !!bo_has_vm_references_locked(bo, vm, ignore);
1035         xe_bo_unlock(bo);
1036
1037         return ret;
1038 }
1039
1040 static void __vm_insert_extobj(struct xe_vm *vm, struct xe_vma *vma)
1041 {
1042         lockdep_assert_held_write(&vm->lock);
1043
1044         list_add(&vma->extobj.link, &vm->extobj.list);
1045         vm->extobj.entries++;
1046 }
1047
1048 static void vm_insert_extobj(struct xe_vm *vm, struct xe_vma *vma)
1049 {
1050         struct xe_bo *bo = xe_vma_bo(vma);
1051
1052         lockdep_assert_held_write(&vm->lock);
1053
1054         if (bo_has_vm_references(bo, vm, vma))
1055                 return;
1056
1057         __vm_insert_extobj(vm, vma);
1058 }
1059
1060 static void vma_destroy_cb(struct dma_fence *fence,
1061                            struct dma_fence_cb *cb)
1062 {
1063         struct xe_vma *vma = container_of(cb, struct xe_vma, destroy_cb);
1064
1065         INIT_WORK(&vma->destroy_work, vma_destroy_work_func);
1066         queue_work(system_unbound_wq, &vma->destroy_work);
1067 }
1068
1069 static void xe_vma_destroy(struct xe_vma *vma, struct dma_fence *fence)
1070 {
1071         struct xe_vm *vm = xe_vma_vm(vma);
1072
1073         lockdep_assert_held_write(&vm->lock);
1074         xe_assert(vm->xe, list_empty(&vma->combined_links.destroy));
1075
1076         if (xe_vma_is_userptr(vma)) {
1077                 xe_assert(vm->xe, vma->gpuva.flags & XE_VMA_DESTROYED);
1078
1079                 spin_lock(&vm->userptr.invalidated_lock);
1080                 list_del(&vma->userptr.invalidate_link);
1081                 spin_unlock(&vm->userptr.invalidated_lock);
1082         } else if (!xe_vma_is_null(vma)) {
1083                 xe_bo_assert_held(xe_vma_bo(vma));
1084
1085                 spin_lock(&vm->notifier.list_lock);
1086                 list_del(&vma->notifier.rebind_link);
1087                 spin_unlock(&vm->notifier.list_lock);
1088
1089                 drm_gpuva_unlink(&vma->gpuva);
1090
1091                 if (!xe_vma_bo(vma)->vm && vm_remove_extobj(vma)) {
1092                         struct xe_vma *other;
1093
1094                         other = bo_has_vm_references_locked(xe_vma_bo(vma), vm, NULL);
1095
1096                         if (other)
1097                                 __vm_insert_extobj(vm, other);
1098                 }
1099         }
1100
1101         xe_vm_assert_held(vm);
1102         if (fence) {
1103                 int ret = dma_fence_add_callback(fence, &vma->destroy_cb,
1104                                                  vma_destroy_cb);
1105
1106                 if (ret) {
1107                         XE_WARN_ON(ret != -ENOENT);
1108                         xe_vma_destroy_late(vma);
1109                 }
1110         } else {
1111                 xe_vma_destroy_late(vma);
1112         }
1113 }
1114
1115 /**
1116  * xe_vm_prepare_vma() - drm_exec utility to lock a vma
1117  * @exec: The drm_exec object we're currently locking for.
1118  * @vma: The vma for witch we want to lock the vm resv and any attached
1119  * object's resv.
1120  * @num_shared: The number of dma-fence slots to pre-allocate in the
1121  * objects' reservation objects.
1122  *
1123  * Return: 0 on success, negative error code on error. In particular
1124  * may return -EDEADLK on WW transaction contention and -EINTR if
1125  * an interruptible wait is terminated by a signal.
1126  */
1127 int xe_vm_prepare_vma(struct drm_exec *exec, struct xe_vma *vma,
1128                       unsigned int num_shared)
1129 {
1130         struct xe_vm *vm = xe_vma_vm(vma);
1131         struct xe_bo *bo = xe_vma_bo(vma);
1132         int err;
1133
1134         XE_WARN_ON(!vm);
1135         err = drm_exec_prepare_obj(exec, xe_vm_obj(vm), num_shared);
1136         if (!err && bo && !bo->vm)
1137                 err = drm_exec_prepare_obj(exec, &bo->ttm.base, num_shared);
1138
1139         return err;
1140 }
1141
1142 static void xe_vma_destroy_unlocked(struct xe_vma *vma)
1143 {
1144         struct drm_exec exec;
1145         int err;
1146
1147         drm_exec_init(&exec, 0);
1148         drm_exec_until_all_locked(&exec) {
1149                 err = xe_vm_prepare_vma(&exec, vma, 0);
1150                 drm_exec_retry_on_contention(&exec);
1151                 if (XE_WARN_ON(err))
1152                         break;
1153         }
1154
1155         xe_vma_destroy(vma, NULL);
1156
1157         drm_exec_fini(&exec);
1158 }
1159
1160 struct xe_vma *
1161 xe_vm_find_overlapping_vma(struct xe_vm *vm, u64 start, u64 range)
1162 {
1163         struct drm_gpuva *gpuva;
1164
1165         lockdep_assert_held(&vm->lock);
1166
1167         if (xe_vm_is_closed_or_banned(vm))
1168                 return NULL;
1169
1170         xe_assert(vm->xe, start + range <= vm->size);
1171
1172         gpuva = drm_gpuva_find_first(&vm->gpuvm, start, range);
1173
1174         return gpuva ? gpuva_to_vma(gpuva) : NULL;
1175 }
1176
1177 static int xe_vm_insert_vma(struct xe_vm *vm, struct xe_vma *vma)
1178 {
1179         int err;
1180
1181         xe_assert(vm->xe, xe_vma_vm(vma) == vm);
1182         lockdep_assert_held(&vm->lock);
1183
1184         err = drm_gpuva_insert(&vm->gpuvm, &vma->gpuva);
1185         XE_WARN_ON(err);        /* Shouldn't be possible */
1186
1187         return err;
1188 }
1189
1190 static void xe_vm_remove_vma(struct xe_vm *vm, struct xe_vma *vma)
1191 {
1192         xe_assert(vm->xe, xe_vma_vm(vma) == vm);
1193         lockdep_assert_held(&vm->lock);
1194
1195         drm_gpuva_remove(&vma->gpuva);
1196         if (vm->usm.last_fault_vma == vma)
1197                 vm->usm.last_fault_vma = NULL;
1198 }
1199
1200 static struct drm_gpuva_op *xe_vm_op_alloc(void)
1201 {
1202         struct xe_vma_op *op;
1203
1204         op = kzalloc(sizeof(*op), GFP_KERNEL);
1205
1206         if (unlikely(!op))
1207                 return NULL;
1208
1209         return &op->base;
1210 }
1211
1212 static void xe_vm_free(struct drm_gpuvm *gpuvm);
1213
1214 static struct drm_gpuvm_ops gpuvm_ops = {
1215         .op_alloc = xe_vm_op_alloc,
1216         .vm_free = xe_vm_free,
1217 };
1218
1219 static u64 pde_encode_pat_index(struct xe_device *xe, u16 pat_index)
1220 {
1221         u64 pte = 0;
1222
1223         if (pat_index & BIT(0))
1224                 pte |= XE_PPGTT_PTE_PAT0;
1225
1226         if (pat_index & BIT(1))
1227                 pte |= XE_PPGTT_PTE_PAT1;
1228
1229         return pte;
1230 }
1231
1232 static u64 pte_encode_pat_index(struct xe_device *xe, u16 pat_index,
1233                                 u32 pt_level)
1234 {
1235         u64 pte = 0;
1236
1237         if (pat_index & BIT(0))
1238                 pte |= XE_PPGTT_PTE_PAT0;
1239
1240         if (pat_index & BIT(1))
1241                 pte |= XE_PPGTT_PTE_PAT1;
1242
1243         if (pat_index & BIT(2)) {
1244                 if (pt_level)
1245                         pte |= XE_PPGTT_PDE_PDPE_PAT2;
1246                 else
1247                         pte |= XE_PPGTT_PTE_PAT2;
1248         }
1249
1250         if (pat_index & BIT(3))
1251                 pte |= XELPG_PPGTT_PTE_PAT3;
1252
1253         if (pat_index & (BIT(4)))
1254                 pte |= XE2_PPGTT_PTE_PAT4;
1255
1256         return pte;
1257 }
1258
1259 static u64 pte_encode_ps(u32 pt_level)
1260 {
1261         XE_WARN_ON(pt_level > 2);
1262
1263         if (pt_level == 1)
1264                 return XE_PDE_PS_2M;
1265         else if (pt_level == 2)
1266                 return XE_PDPE_PS_1G;
1267
1268         return 0;
1269 }
1270
1271 static u64 xelp_pde_encode_bo(struct xe_bo *bo, u64 bo_offset,
1272                               const u16 pat_index)
1273 {
1274         struct xe_device *xe = xe_bo_device(bo);
1275         u64 pde;
1276
1277         pde = xe_bo_addr(bo, bo_offset, XE_PAGE_SIZE);
1278         pde |= XE_PAGE_PRESENT | XE_PAGE_RW;
1279         pde |= pde_encode_pat_index(xe, pat_index);
1280
1281         return pde;
1282 }
1283
1284 static u64 xelp_pte_encode_bo(struct xe_bo *bo, u64 bo_offset,
1285                               u16 pat_index, u32 pt_level)
1286 {
1287         struct xe_device *xe = xe_bo_device(bo);
1288         u64 pte;
1289
1290         pte = xe_bo_addr(bo, bo_offset, XE_PAGE_SIZE);
1291         pte |= XE_PAGE_PRESENT | XE_PAGE_RW;
1292         pte |= pte_encode_pat_index(xe, pat_index, pt_level);
1293         pte |= pte_encode_ps(pt_level);
1294
1295         if (xe_bo_is_vram(bo) || xe_bo_is_stolen_devmem(bo))
1296                 pte |= XE_PPGTT_PTE_DM;
1297
1298         return pte;
1299 }
1300
1301 static u64 xelp_pte_encode_vma(u64 pte, struct xe_vma *vma,
1302                                u16 pat_index, u32 pt_level)
1303 {
1304         struct xe_device *xe = xe_vma_vm(vma)->xe;
1305
1306         pte |= XE_PAGE_PRESENT;
1307
1308         if (likely(!xe_vma_read_only(vma)))
1309                 pte |= XE_PAGE_RW;
1310
1311         pte |= pte_encode_pat_index(xe, pat_index, pt_level);
1312         pte |= pte_encode_ps(pt_level);
1313
1314         if (unlikely(xe_vma_is_null(vma)))
1315                 pte |= XE_PTE_NULL;
1316
1317         return pte;
1318 }
1319
1320 static u64 xelp_pte_encode_addr(struct xe_device *xe, u64 addr,
1321                                 u16 pat_index,
1322                                 u32 pt_level, bool devmem, u64 flags)
1323 {
1324         u64 pte;
1325
1326         /* Avoid passing random bits directly as flags */
1327         xe_assert(xe, !(flags & ~XE_PTE_PS64));
1328
1329         pte = addr;
1330         pte |= XE_PAGE_PRESENT | XE_PAGE_RW;
1331         pte |= pte_encode_pat_index(xe, pat_index, pt_level);
1332         pte |= pte_encode_ps(pt_level);
1333
1334         if (devmem)
1335                 pte |= XE_PPGTT_PTE_DM;
1336
1337         pte |= flags;
1338
1339         return pte;
1340 }
1341
1342 static const struct xe_pt_ops xelp_pt_ops = {
1343         .pte_encode_bo = xelp_pte_encode_bo,
1344         .pte_encode_vma = xelp_pte_encode_vma,
1345         .pte_encode_addr = xelp_pte_encode_addr,
1346         .pde_encode_bo = xelp_pde_encode_bo,
1347 };
1348
1349 static void vm_destroy_work_func(struct work_struct *w);
1350
1351 struct xe_vm *xe_vm_create(struct xe_device *xe, u32 flags)
1352 {
1353         struct drm_gem_object *vm_resv_obj;
1354         struct xe_vm *vm;
1355         int err, number_tiles = 0;
1356         struct xe_tile *tile;
1357         u8 id;
1358
1359         vm = kzalloc(sizeof(*vm), GFP_KERNEL);
1360         if (!vm)
1361                 return ERR_PTR(-ENOMEM);
1362
1363         vm->xe = xe;
1364
1365         vm->size = 1ull << xe->info.va_bits;
1366
1367         vm->flags = flags;
1368
1369         init_rwsem(&vm->lock);
1370
1371         INIT_LIST_HEAD(&vm->rebind_list);
1372
1373         INIT_LIST_HEAD(&vm->userptr.repin_list);
1374         INIT_LIST_HEAD(&vm->userptr.invalidated);
1375         init_rwsem(&vm->userptr.notifier_lock);
1376         spin_lock_init(&vm->userptr.invalidated_lock);
1377
1378         INIT_LIST_HEAD(&vm->notifier.rebind_list);
1379         spin_lock_init(&vm->notifier.list_lock);
1380
1381         INIT_WORK(&vm->destroy_work, vm_destroy_work_func);
1382
1383         INIT_LIST_HEAD(&vm->preempt.exec_queues);
1384         vm->preempt.min_run_period_ms = 10;     /* FIXME: Wire up to uAPI */
1385
1386         for_each_tile(tile, xe, id)
1387                 xe_range_fence_tree_init(&vm->rftree[id]);
1388
1389         INIT_LIST_HEAD(&vm->extobj.list);
1390
1391         vm->pt_ops = &xelp_pt_ops;
1392
1393         if (!(flags & XE_VM_FLAG_MIGRATION))
1394                 xe_device_mem_access_get(xe);
1395
1396         vm_resv_obj = drm_gpuvm_resv_object_alloc(&xe->drm);
1397         if (!vm_resv_obj) {
1398                 err = -ENOMEM;
1399                 goto err_no_resv;
1400         }
1401
1402         drm_gpuvm_init(&vm->gpuvm, "Xe VM", 0, &xe->drm, vm_resv_obj,
1403                        0, vm->size, 0, 0, &gpuvm_ops);
1404
1405         drm_gem_object_put(vm_resv_obj);
1406
1407         err = dma_resv_lock_interruptible(xe_vm_resv(vm), NULL);
1408         if (err)
1409                 goto err_close;
1410
1411         if (IS_DGFX(xe) && xe->info.vram_flags & XE_VRAM_FLAGS_NEED64K)
1412                 vm->flags |= XE_VM_FLAG_64K;
1413
1414         for_each_tile(tile, xe, id) {
1415                 if (flags & XE_VM_FLAG_MIGRATION &&
1416                     tile->id != XE_VM_FLAG_TILE_ID(flags))
1417                         continue;
1418
1419                 vm->pt_root[id] = xe_pt_create(vm, tile, xe->info.vm_max_level);
1420                 if (IS_ERR(vm->pt_root[id])) {
1421                         err = PTR_ERR(vm->pt_root[id]);
1422                         vm->pt_root[id] = NULL;
1423                         goto err_unlock_close;
1424                 }
1425         }
1426
1427         if (flags & XE_VM_FLAG_SCRATCH_PAGE) {
1428                 for_each_tile(tile, xe, id) {
1429                         if (!vm->pt_root[id])
1430                                 continue;
1431
1432                         err = xe_pt_create_scratch(xe, tile, vm);
1433                         if (err)
1434                                 goto err_unlock_close;
1435                 }
1436                 vm->batch_invalidate_tlb = true;
1437         }
1438
1439         if (flags & XE_VM_FLAG_COMPUTE_MODE) {
1440                 INIT_WORK(&vm->preempt.rebind_work, preempt_rebind_work_func);
1441                 vm->flags |= XE_VM_FLAG_COMPUTE_MODE;
1442                 vm->batch_invalidate_tlb = false;
1443         }
1444
1445         /* Fill pt_root after allocating scratch tables */
1446         for_each_tile(tile, xe, id) {
1447                 if (!vm->pt_root[id])
1448                         continue;
1449
1450                 xe_pt_populate_empty(tile, vm, vm->pt_root[id]);
1451         }
1452         dma_resv_unlock(xe_vm_resv(vm));
1453
1454         /* Kernel migration VM shouldn't have a circular loop.. */
1455         if (!(flags & XE_VM_FLAG_MIGRATION)) {
1456                 for_each_tile(tile, xe, id) {
1457                         struct xe_gt *gt = tile->primary_gt;
1458                         struct xe_vm *migrate_vm;
1459                         struct xe_exec_queue *q;
1460                         u32 create_flags = EXEC_QUEUE_FLAG_VM |
1461                                 ((flags & XE_VM_FLAG_ASYNC_DEFAULT) ?
1462                                 EXEC_QUEUE_FLAG_VM_ASYNC : 0);
1463
1464                         if (!vm->pt_root[id])
1465                                 continue;
1466
1467                         migrate_vm = xe_migrate_get_vm(tile->migrate);
1468                         q = xe_exec_queue_create_class(xe, gt, migrate_vm,
1469                                                        XE_ENGINE_CLASS_COPY,
1470                                                        create_flags);
1471                         xe_vm_put(migrate_vm);
1472                         if (IS_ERR(q)) {
1473                                 err = PTR_ERR(q);
1474                                 goto err_close;
1475                         }
1476                         vm->q[id] = q;
1477                         number_tiles++;
1478                 }
1479         }
1480
1481         if (number_tiles > 1)
1482                 vm->composite_fence_ctx = dma_fence_context_alloc(1);
1483
1484         mutex_lock(&xe->usm.lock);
1485         if (flags & XE_VM_FLAG_FAULT_MODE)
1486                 xe->usm.num_vm_in_fault_mode++;
1487         else if (!(flags & XE_VM_FLAG_MIGRATION))
1488                 xe->usm.num_vm_in_non_fault_mode++;
1489         mutex_unlock(&xe->usm.lock);
1490
1491         trace_xe_vm_create(vm);
1492
1493         return vm;
1494
1495 err_unlock_close:
1496         dma_resv_unlock(xe_vm_resv(vm));
1497 err_close:
1498         xe_vm_close_and_put(vm);
1499         return ERR_PTR(err);
1500
1501 err_no_resv:
1502         for_each_tile(tile, xe, id)
1503                 xe_range_fence_tree_fini(&vm->rftree[id]);
1504         kfree(vm);
1505         if (!(flags & XE_VM_FLAG_MIGRATION))
1506                 xe_device_mem_access_put(xe);
1507         return ERR_PTR(err);
1508 }
1509
1510 static void xe_vm_close(struct xe_vm *vm)
1511 {
1512         down_write(&vm->lock);
1513         vm->size = 0;
1514         up_write(&vm->lock);
1515 }
1516
1517 void xe_vm_close_and_put(struct xe_vm *vm)
1518 {
1519         LIST_HEAD(contested);
1520         struct xe_device *xe = vm->xe;
1521         struct xe_tile *tile;
1522         struct xe_vma *vma, *next_vma;
1523         struct drm_gpuva *gpuva, *next;
1524         u8 id;
1525
1526         xe_assert(xe, !vm->preempt.num_exec_queues);
1527
1528         xe_vm_close(vm);
1529         if (xe_vm_in_compute_mode(vm))
1530                 flush_work(&vm->preempt.rebind_work);
1531
1532         down_write(&vm->lock);
1533         for_each_tile(tile, xe, id) {
1534                 if (vm->q[id])
1535                         xe_exec_queue_last_fence_put(vm->q[id], vm);
1536         }
1537         up_write(&vm->lock);
1538
1539         for_each_tile(tile, xe, id) {
1540                 if (vm->q[id]) {
1541                         xe_exec_queue_kill(vm->q[id]);
1542                         xe_exec_queue_put(vm->q[id]);
1543                         vm->q[id] = NULL;
1544                 }
1545         }
1546
1547         down_write(&vm->lock);
1548         xe_vm_lock(vm, false);
1549         drm_gpuvm_for_each_va_safe(gpuva, next, &vm->gpuvm) {
1550                 vma = gpuva_to_vma(gpuva);
1551
1552                 if (xe_vma_has_no_bo(vma)) {
1553                         down_read(&vm->userptr.notifier_lock);
1554                         vma->gpuva.flags |= XE_VMA_DESTROYED;
1555                         up_read(&vm->userptr.notifier_lock);
1556                 }
1557
1558                 xe_vm_remove_vma(vm, vma);
1559
1560                 /* easy case, remove from VMA? */
1561                 if (xe_vma_has_no_bo(vma) || xe_vma_bo(vma)->vm) {
1562                         list_del_init(&vma->combined_links.rebind);
1563                         xe_vma_destroy(vma, NULL);
1564                         continue;
1565                 }
1566
1567                 list_move_tail(&vma->combined_links.destroy, &contested);
1568                 vma->gpuva.flags |= XE_VMA_DESTROYED;
1569         }
1570
1571         /*
1572          * All vm operations will add shared fences to resv.
1573          * The only exception is eviction for a shared object,
1574          * but even so, the unbind when evicted would still
1575          * install a fence to resv. Hence it's safe to
1576          * destroy the pagetables immediately.
1577          */
1578         for_each_tile(tile, xe, id) {
1579                 if (vm->scratch_bo[id]) {
1580                         u32 i;
1581
1582                         xe_bo_unpin(vm->scratch_bo[id]);
1583                         xe_bo_put(vm->scratch_bo[id]);
1584                         for (i = 0; i < vm->pt_root[id]->level; i++)
1585                                 xe_pt_destroy(vm->scratch_pt[id][i], vm->flags,
1586                                               NULL);
1587                 }
1588                 if (vm->pt_root[id]) {
1589                         xe_pt_destroy(vm->pt_root[id], vm->flags, NULL);
1590                         vm->pt_root[id] = NULL;
1591                 }
1592         }
1593         xe_vm_unlock(vm);
1594
1595         /*
1596          * VM is now dead, cannot re-add nodes to vm->vmas if it's NULL
1597          * Since we hold a refcount to the bo, we can remove and free
1598          * the members safely without locking.
1599          */
1600         list_for_each_entry_safe(vma, next_vma, &contested,
1601                                  combined_links.destroy) {
1602                 list_del_init(&vma->combined_links.destroy);
1603                 xe_vma_destroy_unlocked(vma);
1604         }
1605
1606         xe_assert(xe, list_empty(&vm->extobj.list));
1607         up_write(&vm->lock);
1608
1609         mutex_lock(&xe->usm.lock);
1610         if (vm->flags & XE_VM_FLAG_FAULT_MODE)
1611                 xe->usm.num_vm_in_fault_mode--;
1612         else if (!(vm->flags & XE_VM_FLAG_MIGRATION))
1613                 xe->usm.num_vm_in_non_fault_mode--;
1614         mutex_unlock(&xe->usm.lock);
1615
1616         for_each_tile(tile, xe, id)
1617                 xe_range_fence_tree_fini(&vm->rftree[id]);
1618
1619         xe_vm_put(vm);
1620 }
1621
1622 static void vm_destroy_work_func(struct work_struct *w)
1623 {
1624         struct xe_vm *vm =
1625                 container_of(w, struct xe_vm, destroy_work);
1626         struct xe_device *xe = vm->xe;
1627         struct xe_tile *tile;
1628         u8 id;
1629         void *lookup;
1630
1631         /* xe_vm_close_and_put was not called? */
1632         xe_assert(xe, !vm->size);
1633
1634         if (!(vm->flags & XE_VM_FLAG_MIGRATION)) {
1635                 xe_device_mem_access_put(xe);
1636
1637                 if (xe->info.has_asid) {
1638                         mutex_lock(&xe->usm.lock);
1639                         lookup = xa_erase(&xe->usm.asid_to_vm, vm->usm.asid);
1640                         xe_assert(xe, lookup == vm);
1641                         mutex_unlock(&xe->usm.lock);
1642                 }
1643         }
1644
1645         for_each_tile(tile, xe, id)
1646                 XE_WARN_ON(vm->pt_root[id]);
1647
1648         trace_xe_vm_free(vm);
1649         dma_fence_put(vm->rebind_fence);
1650         kfree(vm);
1651 }
1652
1653 static void xe_vm_free(struct drm_gpuvm *gpuvm)
1654 {
1655         struct xe_vm *vm = container_of(gpuvm, struct xe_vm, gpuvm);
1656
1657         /* To destroy the VM we need to be able to sleep */
1658         queue_work(system_unbound_wq, &vm->destroy_work);
1659 }
1660
1661 struct xe_vm *xe_vm_lookup(struct xe_file *xef, u32 id)
1662 {
1663         struct xe_vm *vm;
1664
1665         mutex_lock(&xef->vm.lock);
1666         vm = xa_load(&xef->vm.xa, id);
1667         if (vm)
1668                 xe_vm_get(vm);
1669         mutex_unlock(&xef->vm.lock);
1670
1671         return vm;
1672 }
1673
1674 u64 xe_vm_pdp4_descriptor(struct xe_vm *vm, struct xe_tile *tile)
1675 {
1676         return vm->pt_ops->pde_encode_bo(vm->pt_root[tile->id]->bo, 0,
1677                                          tile_to_xe(tile)->pat.idx[XE_CACHE_WB]);
1678 }
1679
1680 static struct xe_exec_queue *
1681 to_wait_exec_queue(struct xe_vm *vm, struct xe_exec_queue *q)
1682 {
1683         return q ? q : vm->q[0];
1684 }
1685
1686 static struct dma_fence *
1687 xe_vm_unbind_vma(struct xe_vma *vma, struct xe_exec_queue *q,
1688                  struct xe_sync_entry *syncs, u32 num_syncs,
1689                  bool first_op, bool last_op)
1690 {
1691         struct xe_vm *vm = xe_vma_vm(vma);
1692         struct xe_exec_queue *wait_exec_queue = to_wait_exec_queue(vm, q);
1693         struct xe_tile *tile;
1694         struct dma_fence *fence = NULL;
1695         struct dma_fence **fences = NULL;
1696         struct dma_fence_array *cf = NULL;
1697         int cur_fence = 0, i;
1698         int number_tiles = hweight8(vma->tile_present);
1699         int err;
1700         u8 id;
1701
1702         trace_xe_vma_unbind(vma);
1703
1704         if (number_tiles > 1) {
1705                 fences = kmalloc_array(number_tiles, sizeof(*fences),
1706                                        GFP_KERNEL);
1707                 if (!fences)
1708                         return ERR_PTR(-ENOMEM);
1709         }
1710
1711         for_each_tile(tile, vm->xe, id) {
1712                 if (!(vma->tile_present & BIT(id)))
1713                         goto next;
1714
1715                 fence = __xe_pt_unbind_vma(tile, vma, q ? q : vm->q[id],
1716                                            first_op ? syncs : NULL,
1717                                            first_op ? num_syncs : 0);
1718                 if (IS_ERR(fence)) {
1719                         err = PTR_ERR(fence);
1720                         goto err_fences;
1721                 }
1722
1723                 if (fences)
1724                         fences[cur_fence++] = fence;
1725
1726 next:
1727                 if (q && vm->pt_root[id] && !list_empty(&q->multi_gt_list))
1728                         q = list_next_entry(q, multi_gt_list);
1729         }
1730
1731         if (fences) {
1732                 cf = dma_fence_array_create(number_tiles, fences,
1733                                             vm->composite_fence_ctx,
1734                                             vm->composite_fence_seqno++,
1735                                             false);
1736                 if (!cf) {
1737                         --vm->composite_fence_seqno;
1738                         err = -ENOMEM;
1739                         goto err_fences;
1740                 }
1741         }
1742
1743         fence = cf ? &cf->base : !fence ?
1744                 xe_exec_queue_last_fence_get(wait_exec_queue, vm) : fence;
1745         if (last_op) {
1746                 for (i = 0; i < num_syncs; i++)
1747                         xe_sync_entry_signal(&syncs[i], NULL, fence);
1748         }
1749
1750         return fence;
1751
1752 err_fences:
1753         if (fences) {
1754                 while (cur_fence)
1755                         dma_fence_put(fences[--cur_fence]);
1756                 kfree(fences);
1757         }
1758
1759         return ERR_PTR(err);
1760 }
1761
1762 static struct dma_fence *
1763 xe_vm_bind_vma(struct xe_vma *vma, struct xe_exec_queue *q,
1764                struct xe_sync_entry *syncs, u32 num_syncs,
1765                bool first_op, bool last_op)
1766 {
1767         struct xe_tile *tile;
1768         struct dma_fence *fence;
1769         struct dma_fence **fences = NULL;
1770         struct dma_fence_array *cf = NULL;
1771         struct xe_vm *vm = xe_vma_vm(vma);
1772         int cur_fence = 0, i;
1773         int number_tiles = hweight8(vma->tile_mask);
1774         int err;
1775         u8 id;
1776
1777         trace_xe_vma_bind(vma);
1778
1779         if (number_tiles > 1) {
1780                 fences = kmalloc_array(number_tiles, sizeof(*fences),
1781                                        GFP_KERNEL);
1782                 if (!fences)
1783                         return ERR_PTR(-ENOMEM);
1784         }
1785
1786         for_each_tile(tile, vm->xe, id) {
1787                 if (!(vma->tile_mask & BIT(id)))
1788                         goto next;
1789
1790                 fence = __xe_pt_bind_vma(tile, vma, q ? q : vm->q[id],
1791                                          first_op ? syncs : NULL,
1792                                          first_op ? num_syncs : 0,
1793                                          vma->tile_present & BIT(id));
1794                 if (IS_ERR(fence)) {
1795                         err = PTR_ERR(fence);
1796                         goto err_fences;
1797                 }
1798
1799                 if (fences)
1800                         fences[cur_fence++] = fence;
1801
1802 next:
1803                 if (q && vm->pt_root[id] && !list_empty(&q->multi_gt_list))
1804                         q = list_next_entry(q, multi_gt_list);
1805         }
1806
1807         if (fences) {
1808                 cf = dma_fence_array_create(number_tiles, fences,
1809                                             vm->composite_fence_ctx,
1810                                             vm->composite_fence_seqno++,
1811                                             false);
1812                 if (!cf) {
1813                         --vm->composite_fence_seqno;
1814                         err = -ENOMEM;
1815                         goto err_fences;
1816                 }
1817         }
1818
1819         if (last_op) {
1820                 for (i = 0; i < num_syncs; i++)
1821                         xe_sync_entry_signal(&syncs[i], NULL,
1822                                              cf ? &cf->base : fence);
1823         }
1824
1825         return cf ? &cf->base : fence;
1826
1827 err_fences:
1828         if (fences) {
1829                 while (cur_fence)
1830                         dma_fence_put(fences[--cur_fence]);
1831                 kfree(fences);
1832         }
1833
1834         return ERR_PTR(err);
1835 }
1836
1837 static bool xe_vm_sync_mode(struct xe_vm *vm, struct xe_exec_queue *q)
1838 {
1839         return q ? !(q->flags & EXEC_QUEUE_FLAG_VM_ASYNC) :
1840                 !(vm->flags & XE_VM_FLAG_ASYNC_DEFAULT);
1841 }
1842
1843 static int __xe_vm_bind(struct xe_vm *vm, struct xe_vma *vma,
1844                         struct xe_exec_queue *q, struct xe_sync_entry *syncs,
1845                         u32 num_syncs, bool immediate, bool first_op,
1846                         bool last_op)
1847 {
1848         struct dma_fence *fence;
1849         struct xe_exec_queue *wait_exec_queue = to_wait_exec_queue(vm, q);
1850
1851         xe_vm_assert_held(vm);
1852
1853         if (immediate) {
1854                 fence = xe_vm_bind_vma(vma, q, syncs, num_syncs, first_op,
1855                                        last_op);
1856                 if (IS_ERR(fence))
1857                         return PTR_ERR(fence);
1858         } else {
1859                 int i;
1860
1861                 xe_assert(vm->xe, xe_vm_in_fault_mode(vm));
1862
1863                 fence = xe_exec_queue_last_fence_get(wait_exec_queue, vm);
1864                 if (last_op) {
1865                         for (i = 0; i < num_syncs; i++)
1866                                 xe_sync_entry_signal(&syncs[i], NULL, fence);
1867                 }
1868         }
1869
1870         if (last_op)
1871                 xe_exec_queue_last_fence_set(wait_exec_queue, vm, fence);
1872         if (last_op && xe_vm_sync_mode(vm, q))
1873                 dma_fence_wait(fence, true);
1874         dma_fence_put(fence);
1875
1876         return 0;
1877 }
1878
1879 static int xe_vm_bind(struct xe_vm *vm, struct xe_vma *vma, struct xe_exec_queue *q,
1880                       struct xe_bo *bo, struct xe_sync_entry *syncs,
1881                       u32 num_syncs, bool immediate, bool first_op,
1882                       bool last_op)
1883 {
1884         int err;
1885
1886         xe_vm_assert_held(vm);
1887         xe_bo_assert_held(bo);
1888
1889         if (bo && immediate) {
1890                 err = xe_bo_validate(bo, vm, true);
1891                 if (err)
1892                         return err;
1893         }
1894
1895         return __xe_vm_bind(vm, vma, q, syncs, num_syncs, immediate, first_op,
1896                             last_op);
1897 }
1898
1899 static int xe_vm_unbind(struct xe_vm *vm, struct xe_vma *vma,
1900                         struct xe_exec_queue *q, struct xe_sync_entry *syncs,
1901                         u32 num_syncs, bool first_op, bool last_op)
1902 {
1903         struct dma_fence *fence;
1904         struct xe_exec_queue *wait_exec_queue = to_wait_exec_queue(vm, q);
1905
1906         xe_vm_assert_held(vm);
1907         xe_bo_assert_held(xe_vma_bo(vma));
1908
1909         fence = xe_vm_unbind_vma(vma, q, syncs, num_syncs, first_op, last_op);
1910         if (IS_ERR(fence))
1911                 return PTR_ERR(fence);
1912
1913         xe_vma_destroy(vma, fence);
1914         if (last_op)
1915                 xe_exec_queue_last_fence_set(wait_exec_queue, vm, fence);
1916         if (last_op && xe_vm_sync_mode(vm, q))
1917                 dma_fence_wait(fence, true);
1918         dma_fence_put(fence);
1919
1920         return 0;
1921 }
1922
1923 #define ALL_DRM_XE_VM_CREATE_FLAGS (DRM_XE_VM_CREATE_FLAG_SCRATCH_PAGE | \
1924                                     DRM_XE_VM_CREATE_FLAG_COMPUTE_MODE | \
1925                                     DRM_XE_VM_CREATE_FLAG_ASYNC_DEFAULT | \
1926                                     DRM_XE_VM_CREATE_FLAG_FAULT_MODE)
1927
1928 int xe_vm_create_ioctl(struct drm_device *dev, void *data,
1929                        struct drm_file *file)
1930 {
1931         struct xe_device *xe = to_xe_device(dev);
1932         struct xe_file *xef = to_xe_file(file);
1933         struct drm_xe_vm_create *args = data;
1934         struct xe_tile *tile;
1935         struct xe_vm *vm;
1936         u32 id, asid;
1937         int err;
1938         u32 flags = 0;
1939
1940         if (XE_IOCTL_DBG(xe, args->extensions))
1941                 return -EINVAL;
1942
1943         if (XE_WA(xe_root_mmio_gt(xe), 14016763929))
1944                 args->flags |= DRM_XE_VM_CREATE_FLAG_SCRATCH_PAGE;
1945
1946         if (XE_IOCTL_DBG(xe, args->flags & DRM_XE_VM_CREATE_FLAG_FAULT_MODE &&
1947                          !xe->info.supports_usm))
1948                 return -EINVAL;
1949
1950         if (XE_IOCTL_DBG(xe, args->reserved[0] || args->reserved[1]))
1951                 return -EINVAL;
1952
1953         if (XE_IOCTL_DBG(xe, args->flags & ~ALL_DRM_XE_VM_CREATE_FLAGS))
1954                 return -EINVAL;
1955
1956         if (XE_IOCTL_DBG(xe, args->flags & DRM_XE_VM_CREATE_FLAG_SCRATCH_PAGE &&
1957                          args->flags & DRM_XE_VM_CREATE_FLAG_FAULT_MODE))
1958                 return -EINVAL;
1959
1960         if (XE_IOCTL_DBG(xe, args->flags & DRM_XE_VM_CREATE_FLAG_COMPUTE_MODE &&
1961                          args->flags & DRM_XE_VM_CREATE_FLAG_FAULT_MODE))
1962                 return -EINVAL;
1963
1964         if (XE_IOCTL_DBG(xe, args->flags & DRM_XE_VM_CREATE_FLAG_FAULT_MODE &&
1965                          xe_device_in_non_fault_mode(xe)))
1966                 return -EINVAL;
1967
1968         if (XE_IOCTL_DBG(xe, !(args->flags & DRM_XE_VM_CREATE_FLAG_FAULT_MODE) &&
1969                          xe_device_in_fault_mode(xe)))
1970                 return -EINVAL;
1971
1972         if (XE_IOCTL_DBG(xe, args->extensions))
1973                 return -EINVAL;
1974
1975         if (args->flags & DRM_XE_VM_CREATE_FLAG_SCRATCH_PAGE)
1976                 flags |= XE_VM_FLAG_SCRATCH_PAGE;
1977         if (args->flags & DRM_XE_VM_CREATE_FLAG_COMPUTE_MODE)
1978                 flags |= XE_VM_FLAG_COMPUTE_MODE;
1979         if (args->flags & DRM_XE_VM_CREATE_FLAG_ASYNC_DEFAULT)
1980                 flags |= XE_VM_FLAG_ASYNC_DEFAULT;
1981         if (args->flags & DRM_XE_VM_CREATE_FLAG_FAULT_MODE)
1982                 flags |= XE_VM_FLAG_FAULT_MODE;
1983
1984         vm = xe_vm_create(xe, flags);
1985         if (IS_ERR(vm))
1986                 return PTR_ERR(vm);
1987
1988         mutex_lock(&xef->vm.lock);
1989         err = xa_alloc(&xef->vm.xa, &id, vm, xa_limit_32b, GFP_KERNEL);
1990         mutex_unlock(&xef->vm.lock);
1991         if (err) {
1992                 xe_vm_close_and_put(vm);
1993                 return err;
1994         }
1995
1996         if (xe->info.has_asid) {
1997                 mutex_lock(&xe->usm.lock);
1998                 err = xa_alloc_cyclic(&xe->usm.asid_to_vm, &asid, vm,
1999                                       XA_LIMIT(0, XE_MAX_ASID - 1),
2000                                       &xe->usm.next_asid, GFP_KERNEL);
2001                 mutex_unlock(&xe->usm.lock);
2002                 if (err) {
2003                         xe_vm_close_and_put(vm);
2004                         return err;
2005                 }
2006                 vm->usm.asid = asid;
2007         }
2008
2009         args->vm_id = id;
2010         vm->xef = xef;
2011
2012         /* Record BO memory for VM pagetable created against client */
2013         for_each_tile(tile, xe, id)
2014                 if (vm->pt_root[id])
2015                         xe_drm_client_add_bo(vm->xef->client, vm->pt_root[id]->bo);
2016
2017 #if IS_ENABLED(CONFIG_DRM_XE_DEBUG_MEM)
2018         /* Warning: Security issue - never enable by default */
2019         args->reserved[0] = xe_bo_main_addr(vm->pt_root[0]->bo, XE_PAGE_SIZE);
2020 #endif
2021
2022         return 0;
2023 }
2024
2025 int xe_vm_destroy_ioctl(struct drm_device *dev, void *data,
2026                         struct drm_file *file)
2027 {
2028         struct xe_device *xe = to_xe_device(dev);
2029         struct xe_file *xef = to_xe_file(file);
2030         struct drm_xe_vm_destroy *args = data;
2031         struct xe_vm *vm;
2032         int err = 0;
2033
2034         if (XE_IOCTL_DBG(xe, args->pad) ||
2035             XE_IOCTL_DBG(xe, args->reserved[0] || args->reserved[1]))
2036                 return -EINVAL;
2037
2038         mutex_lock(&xef->vm.lock);
2039         vm = xa_load(&xef->vm.xa, args->vm_id);
2040         if (XE_IOCTL_DBG(xe, !vm))
2041                 err = -ENOENT;
2042         else if (XE_IOCTL_DBG(xe, vm->preempt.num_exec_queues))
2043                 err = -EBUSY;
2044         else
2045                 xa_erase(&xef->vm.xa, args->vm_id);
2046         mutex_unlock(&xef->vm.lock);
2047
2048         if (!err)
2049                 xe_vm_close_and_put(vm);
2050
2051         return err;
2052 }
2053
2054 static const u32 region_to_mem_type[] = {
2055         XE_PL_TT,
2056         XE_PL_VRAM0,
2057         XE_PL_VRAM1,
2058 };
2059
2060 static int xe_vm_prefetch(struct xe_vm *vm, struct xe_vma *vma,
2061                           struct xe_exec_queue *q, u32 region,
2062                           struct xe_sync_entry *syncs, u32 num_syncs,
2063                           bool first_op, bool last_op)
2064 {
2065         struct xe_exec_queue *wait_exec_queue = to_wait_exec_queue(vm, q);
2066         int err;
2067
2068         xe_assert(vm->xe, region <= ARRAY_SIZE(region_to_mem_type));
2069
2070         if (!xe_vma_has_no_bo(vma)) {
2071                 err = xe_bo_migrate(xe_vma_bo(vma), region_to_mem_type[region]);
2072                 if (err)
2073                         return err;
2074         }
2075
2076         if (vma->tile_mask != (vma->tile_present & ~vma->usm.tile_invalidated)) {
2077                 return xe_vm_bind(vm, vma, q, xe_vma_bo(vma), syncs, num_syncs,
2078                                   true, first_op, last_op);
2079         } else {
2080                 int i;
2081
2082                 /* Nothing to do, signal fences now */
2083                 if (last_op) {
2084                         for (i = 0; i < num_syncs; i++) {
2085                                 struct dma_fence *fence =
2086                                         xe_exec_queue_last_fence_get(wait_exec_queue, vm);
2087
2088                                 xe_sync_entry_signal(&syncs[i], NULL, fence);
2089                         }
2090                 }
2091
2092                 return 0;
2093         }
2094 }
2095
2096 static void prep_vma_destroy(struct xe_vm *vm, struct xe_vma *vma,
2097                              bool post_commit)
2098 {
2099         down_read(&vm->userptr.notifier_lock);
2100         vma->gpuva.flags |= XE_VMA_DESTROYED;
2101         up_read(&vm->userptr.notifier_lock);
2102         if (post_commit)
2103                 xe_vm_remove_vma(vm, vma);
2104 }
2105
2106 #undef ULL
2107 #define ULL     unsigned long long
2108
2109 #if IS_ENABLED(CONFIG_DRM_XE_DEBUG_VM)
2110 static void print_op(struct xe_device *xe, struct drm_gpuva_op *op)
2111 {
2112         struct xe_vma *vma;
2113
2114         switch (op->op) {
2115         case DRM_GPUVA_OP_MAP:
2116                 vm_dbg(&xe->drm, "MAP: addr=0x%016llx, range=0x%016llx",
2117                        (ULL)op->map.va.addr, (ULL)op->map.va.range);
2118                 break;
2119         case DRM_GPUVA_OP_REMAP:
2120                 vma = gpuva_to_vma(op->remap.unmap->va);
2121                 vm_dbg(&xe->drm, "REMAP:UNMAP: addr=0x%016llx, range=0x%016llx, keep=%d",
2122                        (ULL)xe_vma_start(vma), (ULL)xe_vma_size(vma),
2123                        op->remap.unmap->keep ? 1 : 0);
2124                 if (op->remap.prev)
2125                         vm_dbg(&xe->drm,
2126                                "REMAP:PREV: addr=0x%016llx, range=0x%016llx",
2127                                (ULL)op->remap.prev->va.addr,
2128                                (ULL)op->remap.prev->va.range);
2129                 if (op->remap.next)
2130                         vm_dbg(&xe->drm,
2131                                "REMAP:NEXT: addr=0x%016llx, range=0x%016llx",
2132                                (ULL)op->remap.next->va.addr,
2133                                (ULL)op->remap.next->va.range);
2134                 break;
2135         case DRM_GPUVA_OP_UNMAP:
2136                 vma = gpuva_to_vma(op->unmap.va);
2137                 vm_dbg(&xe->drm, "UNMAP: addr=0x%016llx, range=0x%016llx, keep=%d",
2138                        (ULL)xe_vma_start(vma), (ULL)xe_vma_size(vma),
2139                        op->unmap.keep ? 1 : 0);
2140                 break;
2141         case DRM_GPUVA_OP_PREFETCH:
2142                 vma = gpuva_to_vma(op->prefetch.va);
2143                 vm_dbg(&xe->drm, "PREFETCH: addr=0x%016llx, range=0x%016llx",
2144                        (ULL)xe_vma_start(vma), (ULL)xe_vma_size(vma));
2145                 break;
2146         default:
2147                 drm_warn(&xe->drm, "NOT POSSIBLE");
2148         }
2149 }
2150 #else
2151 static void print_op(struct xe_device *xe, struct drm_gpuva_op *op)
2152 {
2153 }
2154 #endif
2155
2156 /*
2157  * Create operations list from IOCTL arguments, setup operations fields so parse
2158  * and commit steps are decoupled from IOCTL arguments. This step can fail.
2159  */
2160 static struct drm_gpuva_ops *
2161 vm_bind_ioctl_ops_create(struct xe_vm *vm, struct xe_bo *bo,
2162                          u64 bo_offset_or_userptr, u64 addr, u64 range,
2163                          u32 operation, u32 flags, u8 tile_mask,
2164                          u32 prefetch_region)
2165 {
2166         struct drm_gem_object *obj = bo ? &bo->ttm.base : NULL;
2167         struct drm_gpuva_ops *ops;
2168         struct drm_gpuva_op *__op;
2169         struct xe_vma_op *op;
2170         struct drm_gpuvm_bo *vm_bo;
2171         int err;
2172
2173         lockdep_assert_held_write(&vm->lock);
2174
2175         vm_dbg(&vm->xe->drm,
2176                "op=%d, addr=0x%016llx, range=0x%016llx, bo_offset_or_userptr=0x%016llx",
2177                operation, (ULL)addr, (ULL)range,
2178                (ULL)bo_offset_or_userptr);
2179
2180         switch (operation) {
2181         case DRM_XE_VM_BIND_OP_MAP:
2182         case DRM_XE_VM_BIND_OP_MAP_USERPTR:
2183                 ops = drm_gpuvm_sm_map_ops_create(&vm->gpuvm, addr, range,
2184                                                   obj, bo_offset_or_userptr);
2185                 break;
2186         case DRM_XE_VM_BIND_OP_UNMAP:
2187                 ops = drm_gpuvm_sm_unmap_ops_create(&vm->gpuvm, addr, range);
2188                 break;
2189         case DRM_XE_VM_BIND_OP_PREFETCH:
2190                 ops = drm_gpuvm_prefetch_ops_create(&vm->gpuvm, addr, range);
2191                 break;
2192         case DRM_XE_VM_BIND_OP_UNMAP_ALL:
2193                 xe_assert(vm->xe, bo);
2194
2195                 err = xe_bo_lock(bo, true);
2196                 if (err)
2197                         return ERR_PTR(err);
2198
2199                 vm_bo = drm_gpuvm_bo_find(&vm->gpuvm, obj);
2200                 if (!vm_bo)
2201                         break;
2202
2203                 ops = drm_gpuvm_bo_unmap_ops_create(vm_bo);
2204                 drm_gpuvm_bo_put(vm_bo);
2205                 xe_bo_unlock(bo);
2206                 break;
2207         default:
2208                 drm_warn(&vm->xe->drm, "NOT POSSIBLE");
2209                 ops = ERR_PTR(-EINVAL);
2210         }
2211         if (IS_ERR(ops))
2212                 return ops;
2213
2214 #ifdef TEST_VM_ASYNC_OPS_ERROR
2215         if (operation & FORCE_ASYNC_OP_ERROR) {
2216                 op = list_first_entry_or_null(&ops->list, struct xe_vma_op,
2217                                               base.entry);
2218                 if (op)
2219                         op->inject_error = true;
2220         }
2221 #endif
2222
2223         drm_gpuva_for_each_op(__op, ops) {
2224                 struct xe_vma_op *op = gpuva_op_to_vma_op(__op);
2225
2226                 op->tile_mask = tile_mask;
2227                 if (__op->op == DRM_GPUVA_OP_MAP) {
2228                         op->map.immediate =
2229                                 flags & DRM_XE_VM_BIND_FLAG_IMMEDIATE;
2230                         op->map.read_only =
2231                                 flags & DRM_XE_VM_BIND_FLAG_READONLY;
2232                         op->map.is_null = flags & DRM_XE_VM_BIND_FLAG_NULL;
2233                 } else if (__op->op == DRM_GPUVA_OP_PREFETCH) {
2234                         op->prefetch.region = prefetch_region;
2235                 }
2236
2237                 print_op(vm->xe, __op);
2238         }
2239
2240         return ops;
2241 }
2242
2243 static struct xe_vma *new_vma(struct xe_vm *vm, struct drm_gpuva_op_map *op,
2244                               u8 tile_mask, bool read_only, bool is_null)
2245 {
2246         struct xe_bo *bo = op->gem.obj ? gem_to_xe_bo(op->gem.obj) : NULL;
2247         struct xe_vma *vma;
2248         int err;
2249
2250         lockdep_assert_held_write(&vm->lock);
2251
2252         if (bo) {
2253                 err = xe_bo_lock(bo, true);
2254                 if (err)
2255                         return ERR_PTR(err);
2256         }
2257         vma = xe_vma_create(vm, bo, op->gem.offset,
2258                             op->va.addr, op->va.addr +
2259                             op->va.range - 1, read_only, is_null,
2260                             tile_mask);
2261         if (bo)
2262                 xe_bo_unlock(bo);
2263
2264         if (xe_vma_is_userptr(vma)) {
2265                 err = xe_vma_userptr_pin_pages(vma);
2266                 if (err) {
2267                         prep_vma_destroy(vm, vma, false);
2268                         xe_vma_destroy_unlocked(vma);
2269                         return ERR_PTR(err);
2270                 }
2271         } else if (!xe_vma_has_no_bo(vma) && !bo->vm) {
2272                 vm_insert_extobj(vm, vma);
2273                 err = add_preempt_fences(vm, bo);
2274                 if (err) {
2275                         prep_vma_destroy(vm, vma, false);
2276                         xe_vma_destroy_unlocked(vma);
2277                         return ERR_PTR(err);
2278                 }
2279         }
2280
2281         return vma;
2282 }
2283
2284 static u64 xe_vma_max_pte_size(struct xe_vma *vma)
2285 {
2286         if (vma->gpuva.flags & XE_VMA_PTE_1G)
2287                 return SZ_1G;
2288         else if (vma->gpuva.flags & XE_VMA_PTE_2M)
2289                 return SZ_2M;
2290
2291         return SZ_4K;
2292 }
2293
2294 static u64 xe_vma_set_pte_size(struct xe_vma *vma, u64 size)
2295 {
2296         switch (size) {
2297         case SZ_1G:
2298                 vma->gpuva.flags |= XE_VMA_PTE_1G;
2299                 break;
2300         case SZ_2M:
2301                 vma->gpuva.flags |= XE_VMA_PTE_2M;
2302                 break;
2303         }
2304
2305         return SZ_4K;
2306 }
2307
2308 static int xe_vma_op_commit(struct xe_vm *vm, struct xe_vma_op *op)
2309 {
2310         int err = 0;
2311
2312         lockdep_assert_held_write(&vm->lock);
2313
2314         switch (op->base.op) {
2315         case DRM_GPUVA_OP_MAP:
2316                 err |= xe_vm_insert_vma(vm, op->map.vma);
2317                 if (!err)
2318                         op->flags |= XE_VMA_OP_COMMITTED;
2319                 break;
2320         case DRM_GPUVA_OP_REMAP:
2321         {
2322                 u8 tile_present =
2323                         gpuva_to_vma(op->base.remap.unmap->va)->tile_present;
2324
2325                 prep_vma_destroy(vm, gpuva_to_vma(op->base.remap.unmap->va),
2326                                  true);
2327                 op->flags |= XE_VMA_OP_COMMITTED;
2328
2329                 if (op->remap.prev) {
2330                         err |= xe_vm_insert_vma(vm, op->remap.prev);
2331                         if (!err)
2332                                 op->flags |= XE_VMA_OP_PREV_COMMITTED;
2333                         if (!err && op->remap.skip_prev) {
2334                                 op->remap.prev->tile_present =
2335                                         tile_present;
2336                                 op->remap.prev = NULL;
2337                         }
2338                 }
2339                 if (op->remap.next) {
2340                         err |= xe_vm_insert_vma(vm, op->remap.next);
2341                         if (!err)
2342                                 op->flags |= XE_VMA_OP_NEXT_COMMITTED;
2343                         if (!err && op->remap.skip_next) {
2344                                 op->remap.next->tile_present =
2345                                         tile_present;
2346                                 op->remap.next = NULL;
2347                         }
2348                 }
2349
2350                 /* Adjust for partial unbind after removin VMA from VM */
2351                 if (!err) {
2352                         op->base.remap.unmap->va->va.addr = op->remap.start;
2353                         op->base.remap.unmap->va->va.range = op->remap.range;
2354                 }
2355                 break;
2356         }
2357         case DRM_GPUVA_OP_UNMAP:
2358                 prep_vma_destroy(vm, gpuva_to_vma(op->base.unmap.va), true);
2359                 op->flags |= XE_VMA_OP_COMMITTED;
2360                 break;
2361         case DRM_GPUVA_OP_PREFETCH:
2362                 op->flags |= XE_VMA_OP_COMMITTED;
2363                 break;
2364         default:
2365                 drm_warn(&vm->xe->drm, "NOT POSSIBLE");
2366         }
2367
2368         return err;
2369 }
2370
2371
2372 static int vm_bind_ioctl_ops_parse(struct xe_vm *vm, struct xe_exec_queue *q,
2373                                    struct drm_gpuva_ops *ops,
2374                                    struct xe_sync_entry *syncs, u32 num_syncs,
2375                                    struct list_head *ops_list, bool last,
2376                                    bool async)
2377 {
2378         struct xe_vma_op *last_op = NULL;
2379         struct drm_gpuva_op *__op;
2380         int err = 0;
2381
2382         lockdep_assert_held_write(&vm->lock);
2383
2384         drm_gpuva_for_each_op(__op, ops) {
2385                 struct xe_vma_op *op = gpuva_op_to_vma_op(__op);
2386                 bool first = list_empty(ops_list);
2387
2388                 INIT_LIST_HEAD(&op->link);
2389                 list_add_tail(&op->link, ops_list);
2390
2391                 if (first) {
2392                         op->flags |= XE_VMA_OP_FIRST;
2393                         op->num_syncs = num_syncs;
2394                         op->syncs = syncs;
2395                 }
2396
2397                 op->q = q;
2398
2399                 switch (op->base.op) {
2400                 case DRM_GPUVA_OP_MAP:
2401                 {
2402                         struct xe_vma *vma;
2403
2404                         vma = new_vma(vm, &op->base.map,
2405                                       op->tile_mask, op->map.read_only,
2406                                       op->map.is_null);
2407                         if (IS_ERR(vma))
2408                                 return PTR_ERR(vma);
2409
2410                         op->map.vma = vma;
2411                         break;
2412                 }
2413                 case DRM_GPUVA_OP_REMAP:
2414                 {
2415                         struct xe_vma *old =
2416                                 gpuva_to_vma(op->base.remap.unmap->va);
2417
2418                         op->remap.start = xe_vma_start(old);
2419                         op->remap.range = xe_vma_size(old);
2420
2421                         if (op->base.remap.prev) {
2422                                 struct xe_vma *vma;
2423                                 bool read_only =
2424                                         op->base.remap.unmap->va->flags &
2425                                         XE_VMA_READ_ONLY;
2426                                 bool is_null =
2427                                         op->base.remap.unmap->va->flags &
2428                                         DRM_GPUVA_SPARSE;
2429
2430                                 vma = new_vma(vm, op->base.remap.prev,
2431                                               op->tile_mask, read_only,
2432                                               is_null);
2433                                 if (IS_ERR(vma))
2434                                         return PTR_ERR(vma);
2435
2436                                 op->remap.prev = vma;
2437
2438                                 /*
2439                                  * Userptr creates a new SG mapping so
2440                                  * we must also rebind.
2441                                  */
2442                                 op->remap.skip_prev = !xe_vma_is_userptr(old) &&
2443                                         IS_ALIGNED(xe_vma_end(vma),
2444                                                    xe_vma_max_pte_size(old));
2445                                 if (op->remap.skip_prev) {
2446                                         xe_vma_set_pte_size(vma, xe_vma_max_pte_size(old));
2447                                         op->remap.range -=
2448                                                 xe_vma_end(vma) -
2449                                                 xe_vma_start(old);
2450                                         op->remap.start = xe_vma_end(vma);
2451                                 }
2452                         }
2453
2454                         if (op->base.remap.next) {
2455                                 struct xe_vma *vma;
2456                                 bool read_only =
2457                                         op->base.remap.unmap->va->flags &
2458                                         XE_VMA_READ_ONLY;
2459
2460                                 bool is_null =
2461                                         op->base.remap.unmap->va->flags &
2462                                         DRM_GPUVA_SPARSE;
2463
2464                                 vma = new_vma(vm, op->base.remap.next,
2465                                               op->tile_mask, read_only,
2466                                               is_null);
2467                                 if (IS_ERR(vma))
2468                                         return PTR_ERR(vma);
2469
2470                                 op->remap.next = vma;
2471
2472                                 /*
2473                                  * Userptr creates a new SG mapping so
2474                                  * we must also rebind.
2475                                  */
2476                                 op->remap.skip_next = !xe_vma_is_userptr(old) &&
2477                                         IS_ALIGNED(xe_vma_start(vma),
2478                                                    xe_vma_max_pte_size(old));
2479                                 if (op->remap.skip_next) {
2480                                         xe_vma_set_pte_size(vma, xe_vma_max_pte_size(old));
2481                                         op->remap.range -=
2482                                                 xe_vma_end(old) -
2483                                                 xe_vma_start(vma);
2484                                 }
2485                         }
2486                         break;
2487                 }
2488                 case DRM_GPUVA_OP_UNMAP:
2489                 case DRM_GPUVA_OP_PREFETCH:
2490                         /* Nothing to do */
2491                         break;
2492                 default:
2493                         drm_warn(&vm->xe->drm, "NOT POSSIBLE");
2494                 }
2495
2496                 last_op = op;
2497
2498                 err = xe_vma_op_commit(vm, op);
2499                 if (err)
2500                         return err;
2501         }
2502
2503         /* FIXME: Unhandled corner case */
2504         XE_WARN_ON(!last_op && last && !list_empty(ops_list));
2505
2506         if (!last_op)
2507                 return 0;
2508
2509         last_op->ops = ops;
2510         if (last) {
2511                 last_op->flags |= XE_VMA_OP_LAST;
2512                 last_op->num_syncs = num_syncs;
2513                 last_op->syncs = syncs;
2514         }
2515
2516         return 0;
2517 }
2518
2519 static int op_execute(struct drm_exec *exec, struct xe_vm *vm,
2520                       struct xe_vma *vma, struct xe_vma_op *op)
2521 {
2522         int err;
2523
2524         lockdep_assert_held_write(&vm->lock);
2525
2526         err = xe_vm_prepare_vma(exec, vma, 1);
2527         if (err)
2528                 return err;
2529
2530         xe_vm_assert_held(vm);
2531         xe_bo_assert_held(xe_vma_bo(vma));
2532
2533         switch (op->base.op) {
2534         case DRM_GPUVA_OP_MAP:
2535                 err = xe_vm_bind(vm, vma, op->q, xe_vma_bo(vma),
2536                                  op->syncs, op->num_syncs,
2537                                  op->map.immediate || !xe_vm_in_fault_mode(vm),
2538                                  op->flags & XE_VMA_OP_FIRST,
2539                                  op->flags & XE_VMA_OP_LAST);
2540                 break;
2541         case DRM_GPUVA_OP_REMAP:
2542         {
2543                 bool prev = !!op->remap.prev;
2544                 bool next = !!op->remap.next;
2545
2546                 if (!op->remap.unmap_done) {
2547                         if (prev || next)
2548                                 vma->gpuva.flags |= XE_VMA_FIRST_REBIND;
2549                         err = xe_vm_unbind(vm, vma, op->q, op->syncs,
2550                                            op->num_syncs,
2551                                            op->flags & XE_VMA_OP_FIRST,
2552                                            op->flags & XE_VMA_OP_LAST &&
2553                                            !prev && !next);
2554                         if (err)
2555                                 break;
2556                         op->remap.unmap_done = true;
2557                 }
2558
2559                 if (prev) {
2560                         op->remap.prev->gpuva.flags |= XE_VMA_LAST_REBIND;
2561                         err = xe_vm_bind(vm, op->remap.prev, op->q,
2562                                          xe_vma_bo(op->remap.prev), op->syncs,
2563                                          op->num_syncs, true, false,
2564                                          op->flags & XE_VMA_OP_LAST && !next);
2565                         op->remap.prev->gpuva.flags &= ~XE_VMA_LAST_REBIND;
2566                         if (err)
2567                                 break;
2568                         op->remap.prev = NULL;
2569                 }
2570
2571                 if (next) {
2572                         op->remap.next->gpuva.flags |= XE_VMA_LAST_REBIND;
2573                         err = xe_vm_bind(vm, op->remap.next, op->q,
2574                                          xe_vma_bo(op->remap.next),
2575                                          op->syncs, op->num_syncs,
2576                                          true, false,
2577                                          op->flags & XE_VMA_OP_LAST);
2578                         op->remap.next->gpuva.flags &= ~XE_VMA_LAST_REBIND;
2579                         if (err)
2580                                 break;
2581                         op->remap.next = NULL;
2582                 }
2583
2584                 break;
2585         }
2586         case DRM_GPUVA_OP_UNMAP:
2587                 err = xe_vm_unbind(vm, vma, op->q, op->syncs,
2588                                    op->num_syncs, op->flags & XE_VMA_OP_FIRST,
2589                                    op->flags & XE_VMA_OP_LAST);
2590                 break;
2591         case DRM_GPUVA_OP_PREFETCH:
2592                 err = xe_vm_prefetch(vm, vma, op->q, op->prefetch.region,
2593                                      op->syncs, op->num_syncs,
2594                                      op->flags & XE_VMA_OP_FIRST,
2595                                      op->flags & XE_VMA_OP_LAST);
2596                 break;
2597         default:
2598                 drm_warn(&vm->xe->drm, "NOT POSSIBLE");
2599         }
2600
2601         if (err)
2602                 trace_xe_vma_fail(vma);
2603
2604         return err;
2605 }
2606
2607 static int __xe_vma_op_execute(struct xe_vm *vm, struct xe_vma *vma,
2608                                struct xe_vma_op *op)
2609 {
2610         struct drm_exec exec;
2611         int err;
2612
2613 retry_userptr:
2614         drm_exec_init(&exec, DRM_EXEC_INTERRUPTIBLE_WAIT);
2615         drm_exec_until_all_locked(&exec) {
2616                 err = op_execute(&exec, vm, vma, op);
2617                 drm_exec_retry_on_contention(&exec);
2618                 if (err)
2619                         break;
2620         }
2621         drm_exec_fini(&exec);
2622
2623         if (err == -EAGAIN && xe_vma_is_userptr(vma)) {
2624                 lockdep_assert_held_write(&vm->lock);
2625                 err = xe_vma_userptr_pin_pages(vma);
2626                 if (!err)
2627                         goto retry_userptr;
2628
2629                 trace_xe_vma_fail(vma);
2630         }
2631
2632         return err;
2633 }
2634
2635 static int xe_vma_op_execute(struct xe_vm *vm, struct xe_vma_op *op)
2636 {
2637         int ret = 0;
2638
2639         lockdep_assert_held_write(&vm->lock);
2640
2641 #ifdef TEST_VM_ASYNC_OPS_ERROR
2642         if (op->inject_error) {
2643                 op->inject_error = false;
2644                 return -ENOMEM;
2645         }
2646 #endif
2647
2648         switch (op->base.op) {
2649         case DRM_GPUVA_OP_MAP:
2650                 ret = __xe_vma_op_execute(vm, op->map.vma, op);
2651                 break;
2652         case DRM_GPUVA_OP_REMAP:
2653         {
2654                 struct xe_vma *vma;
2655
2656                 if (!op->remap.unmap_done)
2657                         vma = gpuva_to_vma(op->base.remap.unmap->va);
2658                 else if (op->remap.prev)
2659                         vma = op->remap.prev;
2660                 else
2661                         vma = op->remap.next;
2662
2663                 ret = __xe_vma_op_execute(vm, vma, op);
2664                 break;
2665         }
2666         case DRM_GPUVA_OP_UNMAP:
2667                 ret = __xe_vma_op_execute(vm, gpuva_to_vma(op->base.unmap.va),
2668                                           op);
2669                 break;
2670         case DRM_GPUVA_OP_PREFETCH:
2671                 ret = __xe_vma_op_execute(vm,
2672                                           gpuva_to_vma(op->base.prefetch.va),
2673                                           op);
2674                 break;
2675         default:
2676                 drm_warn(&vm->xe->drm, "NOT POSSIBLE");
2677         }
2678
2679         return ret;
2680 }
2681
2682 static void xe_vma_op_cleanup(struct xe_vm *vm, struct xe_vma_op *op)
2683 {
2684         bool last = op->flags & XE_VMA_OP_LAST;
2685
2686         if (last) {
2687                 while (op->num_syncs--)
2688                         xe_sync_entry_cleanup(&op->syncs[op->num_syncs]);
2689                 kfree(op->syncs);
2690                 if (op->q)
2691                         xe_exec_queue_put(op->q);
2692         }
2693         if (!list_empty(&op->link))
2694                 list_del(&op->link);
2695         if (op->ops)
2696                 drm_gpuva_ops_free(&vm->gpuvm, op->ops);
2697         if (last)
2698                 xe_vm_put(vm);
2699 }
2700
2701 static void xe_vma_op_unwind(struct xe_vm *vm, struct xe_vma_op *op,
2702                              bool post_commit, bool prev_post_commit,
2703                              bool next_post_commit)
2704 {
2705         lockdep_assert_held_write(&vm->lock);
2706
2707         switch (op->base.op) {
2708         case DRM_GPUVA_OP_MAP:
2709                 if (op->map.vma) {
2710                         prep_vma_destroy(vm, op->map.vma, post_commit);
2711                         xe_vma_destroy_unlocked(op->map.vma);
2712                 }
2713                 break;
2714         case DRM_GPUVA_OP_UNMAP:
2715         {
2716                 struct xe_vma *vma = gpuva_to_vma(op->base.unmap.va);
2717
2718                 if (vma) {
2719                         down_read(&vm->userptr.notifier_lock);
2720                         vma->gpuva.flags &= ~XE_VMA_DESTROYED;
2721                         up_read(&vm->userptr.notifier_lock);
2722                         if (post_commit)
2723                                 xe_vm_insert_vma(vm, vma);
2724                 }
2725                 break;
2726         }
2727         case DRM_GPUVA_OP_REMAP:
2728         {
2729                 struct xe_vma *vma = gpuva_to_vma(op->base.remap.unmap->va);
2730
2731                 if (op->remap.prev) {
2732                         prep_vma_destroy(vm, op->remap.prev, prev_post_commit);
2733                         xe_vma_destroy_unlocked(op->remap.prev);
2734                 }
2735                 if (op->remap.next) {
2736                         prep_vma_destroy(vm, op->remap.next, next_post_commit);
2737                         xe_vma_destroy_unlocked(op->remap.next);
2738                 }
2739                 if (vma) {
2740                         down_read(&vm->userptr.notifier_lock);
2741                         vma->gpuva.flags &= ~XE_VMA_DESTROYED;
2742                         up_read(&vm->userptr.notifier_lock);
2743                         if (post_commit)
2744                                 xe_vm_insert_vma(vm, vma);
2745                 }
2746                 break;
2747         }
2748         case DRM_GPUVA_OP_PREFETCH:
2749                 /* Nothing to do */
2750                 break;
2751         default:
2752                 drm_warn(&vm->xe->drm, "NOT POSSIBLE");
2753         }
2754 }
2755
2756 static void vm_bind_ioctl_ops_unwind(struct xe_vm *vm,
2757                                      struct drm_gpuva_ops **ops,
2758                                      int num_ops_list)
2759 {
2760         int i;
2761
2762         for (i = num_ops_list - 1; i; ++i) {
2763                 struct drm_gpuva_ops *__ops = ops[i];
2764                 struct drm_gpuva_op *__op;
2765
2766                 if (!__ops)
2767                         continue;
2768
2769                 drm_gpuva_for_each_op_reverse(__op, __ops) {
2770                         struct xe_vma_op *op = gpuva_op_to_vma_op(__op);
2771
2772                         xe_vma_op_unwind(vm, op,
2773                                          op->flags & XE_VMA_OP_COMMITTED,
2774                                          op->flags & XE_VMA_OP_PREV_COMMITTED,
2775                                          op->flags & XE_VMA_OP_NEXT_COMMITTED);
2776                 }
2777
2778                 drm_gpuva_ops_free(&vm->gpuvm, __ops);
2779         }
2780 }
2781
2782 static int vm_bind_ioctl_ops_execute(struct xe_vm *vm,
2783                                      struct list_head *ops_list)
2784 {
2785         struct xe_vma_op *op, *next;
2786         int err;
2787
2788         lockdep_assert_held_write(&vm->lock);
2789
2790         list_for_each_entry_safe(op, next, ops_list, link) {
2791                 err = xe_vma_op_execute(vm, op);
2792                 if (err) {
2793                         drm_warn(&vm->xe->drm, "VM op(%d) failed with %d",
2794                                  op->base.op, err);
2795                         /*
2796                          * FIXME: Killing VM rather than proper error handling
2797                          */
2798                         xe_vm_kill(vm);
2799                         return -ENOSPC;
2800                 }
2801                 xe_vma_op_cleanup(vm, op);
2802         }
2803
2804         return 0;
2805 }
2806
2807 #ifdef TEST_VM_ASYNC_OPS_ERROR
2808 #define SUPPORTED_FLAGS \
2809         (FORCE_ASYNC_OP_ERROR | DRM_XE_VM_BIND_FLAG_ASYNC | \
2810          DRM_XE_VM_BIND_FLAG_READONLY | DRM_XE_VM_BIND_FLAG_IMMEDIATE | \
2811          DRM_XE_VM_BIND_FLAG_NULL | 0xffff)
2812 #else
2813 #define SUPPORTED_FLAGS \
2814         (DRM_XE_VM_BIND_FLAG_ASYNC | DRM_XE_VM_BIND_FLAG_READONLY | \
2815          DRM_XE_VM_BIND_FLAG_IMMEDIATE | DRM_XE_VM_BIND_FLAG_NULL | \
2816          0xffff)
2817 #endif
2818 #define XE_64K_PAGE_MASK 0xffffull
2819
2820 #define MAX_BINDS       512     /* FIXME: Picking random upper limit */
2821
2822 static int vm_bind_ioctl_check_args(struct xe_device *xe,
2823                                     struct drm_xe_vm_bind *args,
2824                                     struct drm_xe_vm_bind_op **bind_ops,
2825                                     bool *async)
2826 {
2827         int err;
2828         int i;
2829
2830         if (XE_IOCTL_DBG(xe, args->extensions) ||
2831             XE_IOCTL_DBG(xe, !args->num_binds) ||
2832             XE_IOCTL_DBG(xe, args->num_binds > MAX_BINDS))
2833                 return -EINVAL;
2834
2835         if (args->num_binds > 1) {
2836                 u64 __user *bind_user =
2837                         u64_to_user_ptr(args->vector_of_binds);
2838
2839                 *bind_ops = kmalloc(sizeof(struct drm_xe_vm_bind_op) *
2840                                     args->num_binds, GFP_KERNEL);
2841                 if (!*bind_ops)
2842                         return -ENOMEM;
2843
2844                 err = __copy_from_user(*bind_ops, bind_user,
2845                                        sizeof(struct drm_xe_vm_bind_op) *
2846                                        args->num_binds);
2847                 if (XE_IOCTL_DBG(xe, err)) {
2848                         err = -EFAULT;
2849                         goto free_bind_ops;
2850                 }
2851         } else {
2852                 *bind_ops = &args->bind;
2853         }
2854
2855         for (i = 0; i < args->num_binds; ++i) {
2856                 u64 range = (*bind_ops)[i].range;
2857                 u64 addr = (*bind_ops)[i].addr;
2858                 u32 op = (*bind_ops)[i].op;
2859                 u32 flags = (*bind_ops)[i].flags;
2860                 u32 obj = (*bind_ops)[i].obj;
2861                 u64 obj_offset = (*bind_ops)[i].obj_offset;
2862                 u32 prefetch_region = (*bind_ops)[i].prefetch_mem_region_instance;
2863                 bool is_null = flags & DRM_XE_VM_BIND_FLAG_NULL;
2864
2865                 if (i == 0) {
2866                         *async = !!(flags & DRM_XE_VM_BIND_FLAG_ASYNC);
2867                         if (XE_IOCTL_DBG(xe, !*async && args->num_syncs)) {
2868                                 err = -EINVAL;
2869                                 goto free_bind_ops;
2870                         }
2871                 } else if (XE_IOCTL_DBG(xe, *async !=
2872                                         !!(flags & DRM_XE_VM_BIND_FLAG_ASYNC))) {
2873                         err = -EINVAL;
2874                         goto free_bind_ops;
2875                 }
2876
2877                 if (XE_IOCTL_DBG(xe, op > DRM_XE_VM_BIND_OP_PREFETCH) ||
2878                     XE_IOCTL_DBG(xe, flags & ~SUPPORTED_FLAGS) ||
2879                     XE_IOCTL_DBG(xe, obj && is_null) ||
2880                     XE_IOCTL_DBG(xe, obj_offset && is_null) ||
2881                     XE_IOCTL_DBG(xe, op != DRM_XE_VM_BIND_OP_MAP &&
2882                                  is_null) ||
2883                     XE_IOCTL_DBG(xe, !obj &&
2884                                  op == DRM_XE_VM_BIND_OP_MAP &&
2885                                  !is_null) ||
2886                     XE_IOCTL_DBG(xe, !obj &&
2887                                  op == DRM_XE_VM_BIND_OP_UNMAP_ALL) ||
2888                     XE_IOCTL_DBG(xe, addr &&
2889                                  op == DRM_XE_VM_BIND_OP_UNMAP_ALL) ||
2890                     XE_IOCTL_DBG(xe, range &&
2891                                  op == DRM_XE_VM_BIND_OP_UNMAP_ALL) ||
2892                     XE_IOCTL_DBG(xe, obj &&
2893                                  op == DRM_XE_VM_BIND_OP_MAP_USERPTR) ||
2894                     XE_IOCTL_DBG(xe, obj &&
2895                                  op == DRM_XE_VM_BIND_OP_PREFETCH) ||
2896                     XE_IOCTL_DBG(xe, prefetch_region &&
2897                                  op != DRM_XE_VM_BIND_OP_PREFETCH) ||
2898                     XE_IOCTL_DBG(xe, !(BIT(prefetch_region) &
2899                                        xe->info.mem_region_mask)) ||
2900                     XE_IOCTL_DBG(xe, obj &&
2901                                  op == DRM_XE_VM_BIND_OP_UNMAP)) {
2902                         err = -EINVAL;
2903                         goto free_bind_ops;
2904                 }
2905
2906                 if (XE_IOCTL_DBG(xe, obj_offset & ~PAGE_MASK) ||
2907                     XE_IOCTL_DBG(xe, addr & ~PAGE_MASK) ||
2908                     XE_IOCTL_DBG(xe, range & ~PAGE_MASK) ||
2909                     XE_IOCTL_DBG(xe, !range &&
2910                                  op != DRM_XE_VM_BIND_OP_UNMAP_ALL)) {
2911                         err = -EINVAL;
2912                         goto free_bind_ops;
2913                 }
2914         }
2915
2916         return 0;
2917
2918 free_bind_ops:
2919         if (args->num_binds > 1)
2920                 kfree(*bind_ops);
2921         return err;
2922 }
2923
2924 int xe_vm_bind_ioctl(struct drm_device *dev, void *data, struct drm_file *file)
2925 {
2926         struct xe_device *xe = to_xe_device(dev);
2927         struct xe_file *xef = to_xe_file(file);
2928         struct drm_xe_vm_bind *args = data;
2929         struct drm_xe_sync __user *syncs_user;
2930         struct xe_bo **bos = NULL;
2931         struct drm_gpuva_ops **ops = NULL;
2932         struct xe_vm *vm;
2933         struct xe_exec_queue *q = NULL;
2934         u32 num_syncs;
2935         struct xe_sync_entry *syncs = NULL;
2936         struct drm_xe_vm_bind_op *bind_ops;
2937         LIST_HEAD(ops_list);
2938         bool async;
2939         int err;
2940         int i;
2941
2942         err = vm_bind_ioctl_check_args(xe, args, &bind_ops, &async);
2943         if (err)
2944                 return err;
2945
2946         if (args->exec_queue_id) {
2947                 q = xe_exec_queue_lookup(xef, args->exec_queue_id);
2948                 if (XE_IOCTL_DBG(xe, !q)) {
2949                         err = -ENOENT;
2950                         goto free_objs;
2951                 }
2952
2953                 if (XE_IOCTL_DBG(xe, !(q->flags & EXEC_QUEUE_FLAG_VM))) {
2954                         err = -EINVAL;
2955                         goto put_exec_queue;
2956                 }
2957
2958                 if (XE_IOCTL_DBG(xe, async !=
2959                                  !!(q->flags & EXEC_QUEUE_FLAG_VM_ASYNC))) {
2960                         err = -EINVAL;
2961                         goto put_exec_queue;
2962                 }
2963         }
2964
2965         vm = xe_vm_lookup(xef, args->vm_id);
2966         if (XE_IOCTL_DBG(xe, !vm)) {
2967                 err = -EINVAL;
2968                 goto put_exec_queue;
2969         }
2970
2971         if (!args->exec_queue_id) {
2972                 if (XE_IOCTL_DBG(xe, async !=
2973                                  !!(vm->flags & XE_VM_FLAG_ASYNC_DEFAULT))) {
2974                         err = -EINVAL;
2975                         goto put_vm;
2976                 }
2977         }
2978
2979         err = down_write_killable(&vm->lock);
2980         if (err)
2981                 goto put_vm;
2982
2983         if (XE_IOCTL_DBG(xe, xe_vm_is_closed_or_banned(vm))) {
2984                 err = -ENOENT;
2985                 goto release_vm_lock;
2986         }
2987
2988         for (i = 0; i < args->num_binds; ++i) {
2989                 u64 range = bind_ops[i].range;
2990                 u64 addr = bind_ops[i].addr;
2991
2992                 if (XE_IOCTL_DBG(xe, range > vm->size) ||
2993                     XE_IOCTL_DBG(xe, addr > vm->size - range)) {
2994                         err = -EINVAL;
2995                         goto release_vm_lock;
2996                 }
2997
2998                 if (bind_ops[i].tile_mask) {
2999                         u64 valid_tiles = BIT(xe->info.tile_count) - 1;
3000
3001                         if (XE_IOCTL_DBG(xe, bind_ops[i].tile_mask &
3002                                          ~valid_tiles)) {
3003                                 err = -EINVAL;
3004                                 goto release_vm_lock;
3005                         }
3006                 }
3007         }
3008
3009         bos = kzalloc(sizeof(*bos) * args->num_binds, GFP_KERNEL);
3010         if (!bos) {
3011                 err = -ENOMEM;
3012                 goto release_vm_lock;
3013         }
3014
3015         ops = kzalloc(sizeof(*ops) * args->num_binds, GFP_KERNEL);
3016         if (!ops) {
3017                 err = -ENOMEM;
3018                 goto release_vm_lock;
3019         }
3020
3021         for (i = 0; i < args->num_binds; ++i) {
3022                 struct drm_gem_object *gem_obj;
3023                 u64 range = bind_ops[i].range;
3024                 u64 addr = bind_ops[i].addr;
3025                 u32 obj = bind_ops[i].obj;
3026                 u64 obj_offset = bind_ops[i].obj_offset;
3027
3028                 if (!obj)
3029                         continue;
3030
3031                 gem_obj = drm_gem_object_lookup(file, obj);
3032                 if (XE_IOCTL_DBG(xe, !gem_obj)) {
3033                         err = -ENOENT;
3034                         goto put_obj;
3035                 }
3036                 bos[i] = gem_to_xe_bo(gem_obj);
3037
3038                 if (XE_IOCTL_DBG(xe, range > bos[i]->size) ||
3039                     XE_IOCTL_DBG(xe, obj_offset >
3040                                  bos[i]->size - range)) {
3041                         err = -EINVAL;
3042                         goto put_obj;
3043                 }
3044
3045                 if (bos[i]->flags & XE_BO_INTERNAL_64K) {
3046                         if (XE_IOCTL_DBG(xe, obj_offset &
3047                                          XE_64K_PAGE_MASK) ||
3048                             XE_IOCTL_DBG(xe, addr & XE_64K_PAGE_MASK) ||
3049                             XE_IOCTL_DBG(xe, range & XE_64K_PAGE_MASK)) {
3050                                 err = -EINVAL;
3051                                 goto put_obj;
3052                         }
3053                 }
3054         }
3055
3056         if (args->num_syncs) {
3057                 syncs = kcalloc(args->num_syncs, sizeof(*syncs), GFP_KERNEL);
3058                 if (!syncs) {
3059                         err = -ENOMEM;
3060                         goto put_obj;
3061                 }
3062         }
3063
3064         syncs_user = u64_to_user_ptr(args->syncs);
3065         for (num_syncs = 0; num_syncs < args->num_syncs; num_syncs++) {
3066                 err = xe_sync_entry_parse(xe, xef, &syncs[num_syncs],
3067                                           &syncs_user[num_syncs], false,
3068                                           xe_vm_no_dma_fences(vm));
3069                 if (err)
3070                         goto free_syncs;
3071         }
3072
3073         for (i = 0; i < args->num_binds; ++i) {
3074                 u64 range = bind_ops[i].range;
3075                 u64 addr = bind_ops[i].addr;
3076                 u32 op = bind_ops[i].op;
3077                 u32 flags = bind_ops[i].flags;
3078                 u64 obj_offset = bind_ops[i].obj_offset;
3079                 u8 tile_mask = bind_ops[i].tile_mask;
3080                 u32 prefetch_region = bind_ops[i].prefetch_mem_region_instance;
3081
3082                 ops[i] = vm_bind_ioctl_ops_create(vm, bos[i], obj_offset,
3083                                                   addr, range, op, flags,
3084                                                   tile_mask, prefetch_region);
3085                 if (IS_ERR(ops[i])) {
3086                         err = PTR_ERR(ops[i]);
3087                         ops[i] = NULL;
3088                         goto unwind_ops;
3089                 }
3090
3091                 err = vm_bind_ioctl_ops_parse(vm, q, ops[i], syncs, num_syncs,
3092                                               &ops_list,
3093                                               i == args->num_binds - 1,
3094                                               async);
3095                 if (err)
3096                         goto unwind_ops;
3097         }
3098
3099         /* Nothing to do */
3100         if (list_empty(&ops_list)) {
3101                 err = -ENODATA;
3102                 goto unwind_ops;
3103         }
3104
3105         xe_vm_get(vm);
3106         if (q)
3107                 xe_exec_queue_get(q);
3108
3109         err = vm_bind_ioctl_ops_execute(vm, &ops_list);
3110
3111         up_write(&vm->lock);
3112
3113         if (q)
3114                 xe_exec_queue_put(q);
3115         xe_vm_put(vm);
3116
3117         for (i = 0; bos && i < args->num_binds; ++i)
3118                 xe_bo_put(bos[i]);
3119
3120         kfree(bos);
3121         kfree(ops);
3122         if (args->num_binds > 1)
3123                 kfree(bind_ops);
3124
3125         return err;
3126
3127 unwind_ops:
3128         vm_bind_ioctl_ops_unwind(vm, ops, args->num_binds);
3129 free_syncs:
3130         for (i = 0; err == -ENODATA && i < num_syncs; i++) {
3131                 struct dma_fence *fence =
3132                         xe_exec_queue_last_fence_get(to_wait_exec_queue(vm, q), vm);
3133
3134                 xe_sync_entry_signal(&syncs[i], NULL, fence);
3135         }
3136         while (num_syncs--)
3137                 xe_sync_entry_cleanup(&syncs[num_syncs]);
3138
3139         kfree(syncs);
3140 put_obj:
3141         for (i = 0; i < args->num_binds; ++i)
3142                 xe_bo_put(bos[i]);
3143 release_vm_lock:
3144         up_write(&vm->lock);
3145 put_vm:
3146         xe_vm_put(vm);
3147 put_exec_queue:
3148         if (q)
3149                 xe_exec_queue_put(q);
3150 free_objs:
3151         kfree(bos);
3152         kfree(ops);
3153         if (args->num_binds > 1)
3154                 kfree(bind_ops);
3155         return err == -ENODATA ? 0 : err;
3156 }
3157
3158 /**
3159  * xe_vm_lock() - Lock the vm's dma_resv object
3160  * @vm: The struct xe_vm whose lock is to be locked
3161  * @intr: Whether to perform any wait interruptible
3162  *
3163  * Return: 0 on success, -EINTR if @intr is true and the wait for a
3164  * contended lock was interrupted. If @intr is false, the function
3165  * always returns 0.
3166  */
3167 int xe_vm_lock(struct xe_vm *vm, bool intr)
3168 {
3169         if (intr)
3170                 return dma_resv_lock_interruptible(xe_vm_resv(vm), NULL);
3171
3172         return dma_resv_lock(xe_vm_resv(vm), NULL);
3173 }
3174
3175 /**
3176  * xe_vm_unlock() - Unlock the vm's dma_resv object
3177  * @vm: The struct xe_vm whose lock is to be released.
3178  *
3179  * Unlock a buffer object lock that was locked by xe_vm_lock().
3180  */
3181 void xe_vm_unlock(struct xe_vm *vm)
3182 {
3183         dma_resv_unlock(xe_vm_resv(vm));
3184 }
3185
3186 /**
3187  * xe_vm_invalidate_vma - invalidate GPU mappings for VMA without a lock
3188  * @vma: VMA to invalidate
3189  *
3190  * Walks a list of page tables leaves which it memset the entries owned by this
3191  * VMA to zero, invalidates the TLBs, and block until TLBs invalidation is
3192  * complete.
3193  *
3194  * Returns 0 for success, negative error code otherwise.
3195  */
3196 int xe_vm_invalidate_vma(struct xe_vma *vma)
3197 {
3198         struct xe_device *xe = xe_vma_vm(vma)->xe;
3199         struct xe_tile *tile;
3200         u32 tile_needs_invalidate = 0;
3201         int seqno[XE_MAX_TILES_PER_DEVICE];
3202         u8 id;
3203         int ret;
3204
3205         xe_assert(xe, xe_vm_in_fault_mode(xe_vma_vm(vma)));
3206         xe_assert(xe, !xe_vma_is_null(vma));
3207         trace_xe_vma_usm_invalidate(vma);
3208
3209         /* Check that we don't race with page-table updates */
3210         if (IS_ENABLED(CONFIG_PROVE_LOCKING)) {
3211                 if (xe_vma_is_userptr(vma)) {
3212                         WARN_ON_ONCE(!mmu_interval_check_retry
3213                                      (&vma->userptr.notifier,
3214                                       vma->userptr.notifier_seq));
3215                         WARN_ON_ONCE(!dma_resv_test_signaled(xe_vm_resv(xe_vma_vm(vma)),
3216                                                              DMA_RESV_USAGE_BOOKKEEP));
3217
3218                 } else {
3219                         xe_bo_assert_held(xe_vma_bo(vma));
3220                 }
3221         }
3222
3223         for_each_tile(tile, xe, id) {
3224                 if (xe_pt_zap_ptes(tile, vma)) {
3225                         tile_needs_invalidate |= BIT(id);
3226                         xe_device_wmb(xe);
3227                         /*
3228                          * FIXME: We potentially need to invalidate multiple
3229                          * GTs within the tile
3230                          */
3231                         seqno[id] = xe_gt_tlb_invalidation_vma(tile->primary_gt, NULL, vma);
3232                         if (seqno[id] < 0)
3233                                 return seqno[id];
3234                 }
3235         }
3236
3237         for_each_tile(tile, xe, id) {
3238                 if (tile_needs_invalidate & BIT(id)) {
3239                         ret = xe_gt_tlb_invalidation_wait(tile->primary_gt, seqno[id]);
3240                         if (ret < 0)
3241                                 return ret;
3242                 }
3243         }
3244
3245         vma->usm.tile_invalidated = vma->tile_mask;
3246
3247         return 0;
3248 }
3249
3250 int xe_analyze_vm(struct drm_printer *p, struct xe_vm *vm, int gt_id)
3251 {
3252         struct drm_gpuva *gpuva;
3253         bool is_vram;
3254         uint64_t addr;
3255
3256         if (!down_read_trylock(&vm->lock)) {
3257                 drm_printf(p, " Failed to acquire VM lock to dump capture");
3258                 return 0;
3259         }
3260         if (vm->pt_root[gt_id]) {
3261                 addr = xe_bo_addr(vm->pt_root[gt_id]->bo, 0, XE_PAGE_SIZE);
3262                 is_vram = xe_bo_is_vram(vm->pt_root[gt_id]->bo);
3263                 drm_printf(p, " VM root: A:0x%llx %s\n", addr,
3264                            is_vram ? "VRAM" : "SYS");
3265         }
3266
3267         drm_gpuvm_for_each_va(gpuva, &vm->gpuvm) {
3268                 struct xe_vma *vma = gpuva_to_vma(gpuva);
3269                 bool is_userptr = xe_vma_is_userptr(vma);
3270                 bool is_null = xe_vma_is_null(vma);
3271
3272                 if (is_null) {
3273                         addr = 0;
3274                 } else if (is_userptr) {
3275                         struct xe_res_cursor cur;
3276
3277                         if (vma->userptr.sg) {
3278                                 xe_res_first_sg(vma->userptr.sg, 0, XE_PAGE_SIZE,
3279                                                 &cur);
3280                                 addr = xe_res_dma(&cur);
3281                         } else {
3282                                 addr = 0;
3283                         }
3284                 } else {
3285                         addr = __xe_bo_addr(xe_vma_bo(vma), 0, XE_PAGE_SIZE);
3286                         is_vram = xe_bo_is_vram(xe_vma_bo(vma));
3287                 }
3288                 drm_printf(p, " [%016llx-%016llx] S:0x%016llx A:%016llx %s\n",
3289                            xe_vma_start(vma), xe_vma_end(vma) - 1,
3290                            xe_vma_size(vma),
3291                            addr, is_null ? "NULL" : is_userptr ? "USR" :
3292                            is_vram ? "VRAM" : "SYS");
3293         }
3294         up_read(&vm->lock);
3295
3296         return 0;
3297 }