drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c

   1 /*
   2  * Copyright 2014-2018 Advanced Micro Devices, Inc.
   3  *
   4  * Permission is hereby granted, free of charge, to any person obtaining a
   5  * copy of this software and associated documentation files (the "Software"),
   6  * to deal in the Software without restriction, including without limitation
   7  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
   8  * and/or sell copies of the Software, and to permit persons to whom the
   9  * Software is furnished to do so, subject to the following conditions:
  10  *
  11  * The above copyright notice and this permission notice shall be included in
  12  * all copies or substantial portions of the Software.
  13  *
  14  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  15  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  16  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
  17  * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR
  18  * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
  19  * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
  20  * OTHER DEALINGS IN THE SOFTWARE.
  21  */
  22
  23 #define pr_fmt(fmt) "kfd2kgd: " fmt
  24
  25 #include <linux/dma-buf.h>
  26 #include <linux/list.h>
  27 #include <linux/pagemap.h>
  28 #include <linux/sched/mm.h>
  29 #include <linux/sched/task.h>
  30
  31 #include "amdgpu_object.h"
  32 #include "amdgpu_vm.h"
  33 #include "amdgpu_amdkfd.h"
  34 #include "amdgpu_dma_buf.h"
  35
  36 /* Special VM and GART address alignment needed for VI pre-Fiji due to
  37  * a HW bug.
  38  */
  39 #define VI_BO_SIZE_ALIGN (0x8000)
  40
  41 /* BO flag to indicate a KFD userptr BO */
  42 #define AMDGPU_AMDKFD_USERPTR_BO (1ULL << 63)
  43
  44 /* Userptr restore delay, just long enough to allow consecutive VM
  45  * changes to accumulate
  46  */
  47 #define AMDGPU_USERPTR_RESTORE_DELAY_MS 1
  48
  49 /* Impose limit on how much memory KFD can use */
  50 static struct {
  51         uint64_t max_system_mem_limit;
  52         uint64_t max_ttm_mem_limit;
  53         int64_t system_mem_used;
  54         int64_t ttm_mem_used;
  55         spinlock_t mem_limit_lock;
  56 } kfd_mem_limit;
  57
  58 /* Struct used for amdgpu_amdkfd_bo_validate */
  59 struct amdgpu_vm_parser {
  60         uint32_t        domain;
  61         bool            wait;
  62 };
  63
  64 static const char * const domain_bit_to_string[] = {
  65                 "CPU",
  66                 "GTT",
  67                 "VRAM",
  68                 "GDS",
  69                 "GWS",
  70                 "OA"
  71 };
  72
  73 #define domain_string(domain) domain_bit_to_string[ffs(domain)-1]
  74
  75 static void amdgpu_amdkfd_restore_userptr_worker(struct work_struct *work);
  76
  77
  78 static inline struct amdgpu_device *get_amdgpu_device(struct kgd_dev *kgd)
  79 {
  80         return (struct amdgpu_device *)kgd;
  81 }
  82
  83 static bool check_if_add_bo_to_vm(struct amdgpu_vm *avm,
  84                 struct kgd_mem *mem)
  85 {
  86         struct kfd_bo_va_list *entry;
  87
  88         list_for_each_entry(entry, &mem->bo_va_list, bo_list)
  89                 if (entry->bo_va->base.vm == avm)
  90                         return false;
  91
  92         return true;
  93 }
  94
  95 /* Set memory usage limits. Current, limits are
  96  *  System (TTM + userptr) memory - 3/4th System RAM
  97  *  TTM memory - 3/8th System RAM
  98  */
  99 void amdgpu_amdkfd_gpuvm_init_mem_limits(void)
 100 {
 101         struct sysinfo si;
 102         uint64_t mem;
 103
 104         si_meminfo(&si);
 105         mem = si.totalram - si.totalhigh;
 106         mem *= si.mem_unit;
 107
 108         spin_lock_init(&kfd_mem_limit.mem_limit_lock);
 109         kfd_mem_limit.max_system_mem_limit = (mem >> 1) + (mem >> 2);
 110         kfd_mem_limit.max_ttm_mem_limit = (mem >> 1) - (mem >> 3);
 111         pr_debug("Kernel memory limit %lluM, TTM limit %lluM\n",
 112                 (kfd_mem_limit.max_system_mem_limit >> 20),
 113                 (kfd_mem_limit.max_ttm_mem_limit >> 20));
 114 }
 115
 116 static int amdgpu_amdkfd_reserve_mem_limit(struct amdgpu_device *adev,
 117                 uint64_t size, u32 domain, bool sg)
 118 {
 119         size_t acc_size, system_mem_needed, ttm_mem_needed, vram_needed;
 120         uint64_t reserved_for_pt = amdgpu_amdkfd_total_mem_size >> 9;
 121         int ret = 0;
 122
 123         acc_size = ttm_bo_dma_acc_size(&adev->mman.bdev, size,
 124                                        sizeof(struct amdgpu_bo));
 125
 126         vram_needed = 0;
 127         if (domain == AMDGPU_GEM_DOMAIN_GTT) {
 128                 /* TTM GTT memory */
 129                 system_mem_needed = acc_size + size;
 130                 ttm_mem_needed = acc_size + size;
 131         } else if (domain == AMDGPU_GEM_DOMAIN_CPU && !sg) {
 132                 /* Userptr */
 133                 system_mem_needed = acc_size + size;
 134                 ttm_mem_needed = acc_size;
 135         } else {
 136                 /* VRAM and SG */
 137                 system_mem_needed = acc_size;
 138                 ttm_mem_needed = acc_size;
 139                 if (domain == AMDGPU_GEM_DOMAIN_VRAM)
 140                         vram_needed = size;
 141         }
 142
 143         spin_lock(&kfd_mem_limit.mem_limit_lock);
 144
 145         if ((kfd_mem_limit.system_mem_used + system_mem_needed >
 146              kfd_mem_limit.max_system_mem_limit) ||
 147             (kfd_mem_limit.ttm_mem_used + ttm_mem_needed >
 148              kfd_mem_limit.max_ttm_mem_limit) ||
 149             (adev->kfd.vram_used + vram_needed >
 150              adev->gmc.real_vram_size - reserved_for_pt)) {
 151                 ret = -ENOMEM;
 152         } else {
 153                 kfd_mem_limit.system_mem_used += system_mem_needed;
 154                 kfd_mem_limit.ttm_mem_used += ttm_mem_needed;
 155                 adev->kfd.vram_used += vram_needed;
 156         }
 157
 158         spin_unlock(&kfd_mem_limit.mem_limit_lock);
 159         return ret;
 160 }
 161
 162 static void unreserve_mem_limit(struct amdgpu_device *adev,
 163                 uint64_t size, u32 domain, bool sg)
 164 {
 165         size_t acc_size;
 166
 167         acc_size = ttm_bo_dma_acc_size(&adev->mman.bdev, size,
 168                                        sizeof(struct amdgpu_bo));
 169
 170         spin_lock(&kfd_mem_limit.mem_limit_lock);
 171         if (domain == AMDGPU_GEM_DOMAIN_GTT) {
 172                 kfd_mem_limit.system_mem_used -= (acc_size + size);
 173                 kfd_mem_limit.ttm_mem_used -= (acc_size + size);
 174         } else if (domain == AMDGPU_GEM_DOMAIN_CPU && !sg) {
 175                 kfd_mem_limit.system_mem_used -= (acc_size + size);
 176                 kfd_mem_limit.ttm_mem_used -= acc_size;
 177         } else {
 178                 kfd_mem_limit.system_mem_used -= acc_size;
 179                 kfd_mem_limit.ttm_mem_used -= acc_size;
 180                 if (domain == AMDGPU_GEM_DOMAIN_VRAM) {
 181                         adev->kfd.vram_used -= size;
 182                         WARN_ONCE(adev->kfd.vram_used < 0,
 183                                   "kfd VRAM memory accounting unbalanced");
 184                 }
 185         }
 186         WARN_ONCE(kfd_mem_limit.system_mem_used < 0,
 187                   "kfd system memory accounting unbalanced");
 188         WARN_ONCE(kfd_mem_limit.ttm_mem_used < 0,
 189                   "kfd TTM memory accounting unbalanced");
 190
 191         spin_unlock(&kfd_mem_limit.mem_limit_lock);
 192 }
 193
 194 void amdgpu_amdkfd_unreserve_memory_limit(struct amdgpu_bo *bo)
 195 {
 196         struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
 197         u32 domain = bo->preferred_domains;
 198         bool sg = (bo->preferred_domains == AMDGPU_GEM_DOMAIN_CPU);
 199
 200         if (bo->flags & AMDGPU_AMDKFD_USERPTR_BO) {
 201                 domain = AMDGPU_GEM_DOMAIN_CPU;
 202                 sg = false;
 203         }
 204
 205         unreserve_mem_limit(adev, amdgpu_bo_size(bo), domain, sg);
 206 }
 207
 208
 209 /* amdgpu_amdkfd_remove_eviction_fence - Removes eviction fence from BO's
 210  *  reservation object.
 211  *
 212  * @bo: [IN] Remove eviction fence(s) from this BO
 213  * @ef: [IN] This eviction fence is removed if it
 214  *  is present in the shared list.
 215  *
 216  * NOTE: Must be called with BO reserved i.e. bo->tbo.resv->lock held.
 217  */
 218 static int amdgpu_amdkfd_remove_eviction_fence(struct amdgpu_bo *bo,
 219                                         struct amdgpu_amdkfd_fence *ef)
 220 {
 221         struct dma_resv *resv = bo->tbo.base.resv;
 222         struct dma_resv_list *old, *new;
 223         unsigned int i, j, k;
 224
 225         if (!ef)
 226                 return -EINVAL;
 227
 228         old = dma_resv_get_list(resv);
 229         if (!old)
 230                 return 0;
 231
 232         new = kmalloc(offsetof(typeof(*new), shared[old->shared_max]),
 233                       GFP_KERNEL);
 234         if (!new)
 235                 return -ENOMEM;
 236
 237         /* Go through all the shared fences in the resevation object and sort
 238          * the interesting ones to the end of the list.
 239          */
 240         for (i = 0, j = old->shared_count, k = 0; i < old->shared_count; ++i) {
 241                 struct dma_fence *f;
 242
 243                 f = rcu_dereference_protected(old->shared[i],
 244                                               dma_resv_held(resv));
 245
 246                 if (f->context == ef->base.context)
 247                         RCU_INIT_POINTER(new->shared[--j], f);
 248                 else
 249                         RCU_INIT_POINTER(new->shared[k++], f);
 250         }
 251         new->shared_max = old->shared_max;
 252         new->shared_count = k;
 253
 254         /* Install the new fence list, seqcount provides the barriers */
 255         preempt_disable();
 256         write_seqcount_begin(&resv->seq);
 257         RCU_INIT_POINTER(resv->fence, new);
 258         write_seqcount_end(&resv->seq);
 259         preempt_enable();
 260
 261         /* Drop the references to the removed fences or move them to ef_list */
 262         for (i = j, k = 0; i < old->shared_count; ++i) {
 263                 struct dma_fence *f;
 264
 265                 f = rcu_dereference_protected(new->shared[i],
 266                                               dma_resv_held(resv));
 267                 dma_fence_put(f);
 268         }
 269         kfree_rcu(old, rcu);
 270
 271         return 0;
 272 }
 273
 274 static int amdgpu_amdkfd_bo_validate(struct amdgpu_bo *bo, uint32_t domain,
 275                                      bool wait)
 276 {
 277         struct ttm_operation_ctx ctx = { false, false };
 278         int ret;
 279
 280         if (WARN(amdgpu_ttm_tt_get_usermm(bo->tbo.ttm),
 281                  "Called with userptr BO"))
 282                 return -EINVAL;
 283
 284         amdgpu_bo_placement_from_domain(bo, domain);
 285
 286         ret = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
 287         if (ret)
 288                 goto validate_fail;
 289         if (wait)
 290                 amdgpu_bo_sync_wait(bo, AMDGPU_FENCE_OWNER_KFD, false);
 291
 292 validate_fail:
 293         return ret;
 294 }
 295
 296 static int amdgpu_amdkfd_validate(void *param, struct amdgpu_bo *bo)
 297 {
 298         struct amdgpu_vm_parser *p = param;
 299
 300         return amdgpu_amdkfd_bo_validate(bo, p->domain, p->wait);
 301 }
 302
 303 /* vm_validate_pt_pd_bos - Validate page table and directory BOs
 304  *
 305  * Page directories are not updated here because huge page handling
 306  * during page table updates can invalidate page directory entries
 307  * again. Page directories are only updated after updating page
 308  * tables.
 309  */
 310 static int vm_validate_pt_pd_bos(struct amdgpu_vm *vm)
 311 {
 312         struct amdgpu_bo *pd = vm->root.base.bo;
 313         struct amdgpu_device *adev = amdgpu_ttm_adev(pd->tbo.bdev);
 314         struct amdgpu_vm_parser param;
 315         int ret;
 316
 317         param.domain = AMDGPU_GEM_DOMAIN_VRAM;
 318         param.wait = false;
 319
 320         ret = amdgpu_vm_validate_pt_bos(adev, vm, amdgpu_amdkfd_validate,
 321                                         &param);
 322         if (ret) {
 323                 pr_err("amdgpu: failed to validate PT BOs\n");
 324                 return ret;
 325         }
 326
 327         ret = amdgpu_amdkfd_validate(&param, pd);
 328         if (ret) {
 329                 pr_err("amdgpu: failed to validate PD\n");
 330                 return ret;
 331         }
 332
 333         vm->pd_phys_addr = amdgpu_gmc_pd_addr(vm->root.base.bo);
 334
 335         if (vm->use_cpu_for_update) {
 336                 ret = amdgpu_bo_kmap(pd, NULL);
 337                 if (ret) {
 338                         pr_err("amdgpu: failed to kmap PD, ret=%d\n", ret);
 339                         return ret;
 340                 }
 341         }
 342
 343         return 0;
 344 }
 345
 346 static int vm_update_pds(struct amdgpu_vm *vm, struct amdgpu_sync *sync)
 347 {
 348         struct amdgpu_bo *pd = vm->root.base.bo;
 349         struct amdgpu_device *adev = amdgpu_ttm_adev(pd->tbo.bdev);
 350         int ret;
 351
 352         ret = amdgpu_vm_update_directories(adev, vm);
 353         if (ret)
 354                 return ret;
 355
 356         return amdgpu_sync_fence(NULL, sync, vm->last_update, false);
 357 }
 358
 359 /* add_bo_to_vm - Add a BO to a VM
 360  *
 361  * Everything that needs to bo done only once when a BO is first added
 362  * to a VM. It can later be mapped and unmapped many times without
 363  * repeating these steps.
 364  *
 365  * 1. Allocate and initialize BO VA entry data structure
 366  * 2. Add BO to the VM
 367  * 3. Determine ASIC-specific PTE flags
 368  * 4. Alloc page tables and directories if needed
 369  * 4a.  Validate new page tables and directories
 370  */
 371 static int add_bo_to_vm(struct amdgpu_device *adev, struct kgd_mem *mem,
 372                 struct amdgpu_vm *vm, bool is_aql,
 373                 struct kfd_bo_va_list **p_bo_va_entry)
 374 {
 375         int ret;
 376         struct kfd_bo_va_list *bo_va_entry;
 377         struct amdgpu_bo *bo = mem->bo;
 378         uint64_t va = mem->va;
 379         struct list_head *list_bo_va = &mem->bo_va_list;
 380         unsigned long bo_size = bo->tbo.mem.size;
 381
 382         if (!va) {
 383                 pr_err("Invalid VA when adding BO to VM\n");
 384                 return -EINVAL;
 385         }
 386
 387         if (is_aql)
 388                 va += bo_size;
 389
 390         bo_va_entry = kzalloc(sizeof(*bo_va_entry), GFP_KERNEL);
 391         if (!bo_va_entry)
 392                 return -ENOMEM;
 393
 394         pr_debug("\t add VA 0x%llx - 0x%llx to vm %p\n", va,
 395                         va + bo_size, vm);
 396
 397         /* Add BO to VM internal data structures*/
 398         bo_va_entry->bo_va = amdgpu_vm_bo_add(adev, vm, bo);
 399         if (!bo_va_entry->bo_va) {
 400                 ret = -EINVAL;
 401                 pr_err("Failed to add BO object to VM. ret == %d\n",
 402                                 ret);
 403                 goto err_vmadd;
 404         }
 405
 406         bo_va_entry->va = va;
 407         bo_va_entry->pte_flags = amdgpu_gmc_get_pte_flags(adev,
 408                                                          mem->mapping_flags);
 409         bo_va_entry->kgd_dev = (void *)adev;
 410         list_add(&bo_va_entry->bo_list, list_bo_va);
 411
 412         if (p_bo_va_entry)
 413                 *p_bo_va_entry = bo_va_entry;
 414
 415         /* Allocate validate page tables if needed */
 416         ret = vm_validate_pt_pd_bos(vm);
 417         if (ret) {
 418                 pr_err("validate_pt_pd_bos() failed\n");
 419                 goto err_alloc_pts;
 420         }
 421
 422         return 0;
 423
 424 err_alloc_pts:
 425         amdgpu_vm_bo_rmv(adev, bo_va_entry->bo_va);
 426         list_del(&bo_va_entry->bo_list);
 427 err_vmadd:
 428         kfree(bo_va_entry);
 429         return ret;
 430 }
 431
 432 static void remove_bo_from_vm(struct amdgpu_device *adev,
 433                 struct kfd_bo_va_list *entry, unsigned long size)
 434 {
 435         pr_debug("\t remove VA 0x%llx - 0x%llx in entry %p\n",
 436                         entry->va,
 437                         entry->va + size, entry);
 438         amdgpu_vm_bo_rmv(adev, entry->bo_va);
 439         list_del(&entry->bo_list);
 440         kfree(entry);
 441 }
 442
 443 static void add_kgd_mem_to_kfd_bo_list(struct kgd_mem *mem,
 444                                 struct amdkfd_process_info *process_info,
 445                                 bool userptr)
 446 {
 447         struct ttm_validate_buffer *entry = &mem->validate_list;
 448         struct amdgpu_bo *bo = mem->bo;
 449
 450         INIT_LIST_HEAD(&entry->head);
 451         entry->num_shared = 1;
 452         entry->bo = &bo->tbo;
 453         mutex_lock(&process_info->lock);
 454         if (userptr)
 455                 list_add_tail(&entry->head, &process_info->userptr_valid_list);
 456         else
 457                 list_add_tail(&entry->head, &process_info->kfd_bo_list);
 458         mutex_unlock(&process_info->lock);
 459 }
 460
 461 static void remove_kgd_mem_from_kfd_bo_list(struct kgd_mem *mem,
 462                 struct amdkfd_process_info *process_info)
 463 {
 464         struct ttm_validate_buffer *bo_list_entry;
 465
 466         bo_list_entry = &mem->validate_list;
 467         mutex_lock(&process_info->lock);
 468         list_del(&bo_list_entry->head);
 469         mutex_unlock(&process_info->lock);
 470 }
 471
 472 /* Initializes user pages. It registers the MMU notifier and validates
 473  * the userptr BO in the GTT domain.
 474  *
 475  * The BO must already be on the userptr_valid_list. Otherwise an
 476  * eviction and restore may happen that leaves the new BO unmapped
 477  * with the user mode queues running.
 478  *
 479  * Takes the process_info->lock to protect against concurrent restore
 480  * workers.
 481  *
 482  * Returns 0 for success, negative errno for errors.
 483  */
 484 static int init_user_pages(struct kgd_mem *mem, struct mm_struct *mm,
 485                            uint64_t user_addr)
 486 {
 487         struct amdkfd_process_info *process_info = mem->process_info;
 488         struct amdgpu_bo *bo = mem->bo;
 489         struct ttm_operation_ctx ctx = { true, false };
 490         int ret = 0;
 491
 492         mutex_lock(&process_info->lock);
 493
 494         ret = amdgpu_ttm_tt_set_userptr(bo->tbo.ttm, user_addr, 0);
 495         if (ret) {
 496                 pr_err("%s: Failed to set userptr: %d\n", __func__, ret);
 497                 goto out;
 498         }
 499
 500         ret = amdgpu_mn_register(bo, user_addr);
 501         if (ret) {
 502                 pr_err("%s: Failed to register MMU notifier: %d\n",
 503                        __func__, ret);
 504                 goto out;
 505         }
 506
 507         ret = amdgpu_ttm_tt_get_user_pages(bo, bo->tbo.ttm->pages);
 508         if (ret) {
 509                 pr_err("%s: Failed to get user pages: %d\n", __func__, ret);
 510                 goto unregister_out;
 511         }
 512
 513         ret = amdgpu_bo_reserve(bo, true);
 514         if (ret) {
 515                 pr_err("%s: Failed to reserve BO\n", __func__);
 516                 goto release_out;
 517         }
 518         amdgpu_bo_placement_from_domain(bo, mem->domain);
 519         ret = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
 520         if (ret)
 521                 pr_err("%s: failed to validate BO\n", __func__);
 522         amdgpu_bo_unreserve(bo);
 523
 524 release_out:
 525         amdgpu_ttm_tt_get_user_pages_done(bo->tbo.ttm);
 526 unregister_out:
 527         if (ret)
 528                 amdgpu_mn_unregister(bo);
 529 out:
 530         mutex_unlock(&process_info->lock);
 531         return ret;
 532 }
 533
 534 /* Reserving a BO and its page table BOs must happen atomically to
 535  * avoid deadlocks. Some operations update multiple VMs at once. Track
 536  * all the reservation info in a context structure. Optionally a sync
 537  * object can track VM updates.
 538  */
 539 struct bo_vm_reservation_context {
 540         struct amdgpu_bo_list_entry kfd_bo; /* BO list entry for the KFD BO */
 541         unsigned int n_vms;                 /* Number of VMs reserved       */
 542         struct amdgpu_bo_list_entry *vm_pd; /* Array of VM BO list entries  */
 543         struct ww_acquire_ctx ticket;       /* Reservation ticket           */
 544         struct list_head list, duplicates;  /* BO lists                     */
 545         struct amdgpu_sync *sync;           /* Pointer to sync object       */
 546         bool reserved;                      /* Whether BOs are reserved     */
 547 };
 548
 549 enum bo_vm_match {
 550         BO_VM_NOT_MAPPED = 0,   /* Match VMs where a BO is not mapped */
 551         BO_VM_MAPPED,           /* Match VMs where a BO is mapped     */
 552         BO_VM_ALL,              /* Match all VMs a BO was added to    */
 553 };
 554
 555 /**
 556  * reserve_bo_and_vm - reserve a BO and a VM unconditionally.
 557  * @mem: KFD BO structure.
 558  * @vm: the VM to reserve.
 559  * @ctx: the struct that will be used in unreserve_bo_and_vms().
 560  */
 561 static int reserve_bo_and_vm(struct kgd_mem *mem,
 562                               struct amdgpu_vm *vm,
 563                               struct bo_vm_reservation_context *ctx)
 564 {
 565         struct amdgpu_bo *bo = mem->bo;
 566         int ret;
 567
 568         WARN_ON(!vm);
 569
 570         ctx->reserved = false;
 571         ctx->n_vms = 1;
 572         ctx->sync = &mem->sync;
 573
 574         INIT_LIST_HEAD(&ctx->list);
 575         INIT_LIST_HEAD(&ctx->duplicates);
 576
 577         ctx->vm_pd = kcalloc(ctx->n_vms, sizeof(*ctx->vm_pd), GFP_KERNEL);
 578         if (!ctx->vm_pd)
 579                 return -ENOMEM;
 580
 581         ctx->kfd_bo.priority = 0;
 582         ctx->kfd_bo.tv.bo = &bo->tbo;
 583         ctx->kfd_bo.tv.num_shared = 1;
 584         list_add(&ctx->kfd_bo.tv.head, &ctx->list);
 585
 586         amdgpu_vm_get_pd_bo(vm, &ctx->list, &ctx->vm_pd[0]);
 587
 588         ret = ttm_eu_reserve_buffers(&ctx->ticket, &ctx->list,
 589                                      false, &ctx->duplicates, true);
 590         if (!ret)
 591                 ctx->reserved = true;
 592         else {
 593                 pr_err("Failed to reserve buffers in ttm\n");
 594                 kfree(ctx->vm_pd);
 595                 ctx->vm_pd = NULL;
 596         }
 597
 598         return ret;
 599 }
 600
 601 /**
 602  * reserve_bo_and_cond_vms - reserve a BO and some VMs conditionally
 603  * @mem: KFD BO structure.
 604  * @vm: the VM to reserve. If NULL, then all VMs associated with the BO
 605  * is used. Otherwise, a single VM associated with the BO.
 606  * @map_type: the mapping status that will be used to filter the VMs.
 607  * @ctx: the struct that will be used in unreserve_bo_and_vms().
 608  *
 609  * Returns 0 for success, negative for failure.
 610  */
 611 static int reserve_bo_and_cond_vms(struct kgd_mem *mem,
 612                                 struct amdgpu_vm *vm, enum bo_vm_match map_type,
 613                                 struct bo_vm_reservation_context *ctx)
 614 {
 615         struct amdgpu_bo *bo = mem->bo;
 616         struct kfd_bo_va_list *entry;
 617         unsigned int i;
 618         int ret;
 619
 620         ctx->reserved = false;
 621         ctx->n_vms = 0;
 622         ctx->vm_pd = NULL;
 623         ctx->sync = &mem->sync;
 624
 625         INIT_LIST_HEAD(&ctx->list);
 626         INIT_LIST_HEAD(&ctx->duplicates);
 627
 628         list_for_each_entry(entry, &mem->bo_va_list, bo_list) {
 629                 if ((vm && vm != entry->bo_va->base.vm) ||
 630                         (entry->is_mapped != map_type
 631                         && map_type != BO_VM_ALL))
 632                         continue;
 633
 634                 ctx->n_vms++;
 635         }
 636
 637         if (ctx->n_vms != 0) {
 638                 ctx->vm_pd = kcalloc(ctx->n_vms, sizeof(*ctx->vm_pd),
 639                                      GFP_KERNEL);
 640                 if (!ctx->vm_pd)
 641                         return -ENOMEM;
 642         }
 643
 644         ctx->kfd_bo.priority = 0;
 645         ctx->kfd_bo.tv.bo = &bo->tbo;
 646         ctx->kfd_bo.tv.num_shared = 1;
 647         list_add(&ctx->kfd_bo.tv.head, &ctx->list);
 648
 649         i = 0;
 650         list_for_each_entry(entry, &mem->bo_va_list, bo_list) {
 651                 if ((vm && vm != entry->bo_va->base.vm) ||
 652                         (entry->is_mapped != map_type
 653                         && map_type != BO_VM_ALL))
 654                         continue;
 655
 656                 amdgpu_vm_get_pd_bo(entry->bo_va->base.vm, &ctx->list,
 657                                 &ctx->vm_pd[i]);
 658                 i++;
 659         }
 660
 661         ret = ttm_eu_reserve_buffers(&ctx->ticket, &ctx->list,
 662                                      false, &ctx->duplicates, true);
 663         if (!ret)
 664                 ctx->reserved = true;
 665         else
 666                 pr_err("Failed to reserve buffers in ttm.\n");
 667
 668         if (ret) {
 669                 kfree(ctx->vm_pd);
 670                 ctx->vm_pd = NULL;
 671         }
 672
 673         return ret;
 674 }
 675
 676 /**
 677  * unreserve_bo_and_vms - Unreserve BO and VMs from a reservation context
 678  * @ctx: Reservation context to unreserve
 679  * @wait: Optionally wait for a sync object representing pending VM updates
 680  * @intr: Whether the wait is interruptible
 681  *
 682  * Also frees any resources allocated in
 683  * reserve_bo_and_(cond_)vm(s). Returns the status from
 684  * amdgpu_sync_wait.
 685  */
 686 static int unreserve_bo_and_vms(struct bo_vm_reservation_context *ctx,
 687                                  bool wait, bool intr)
 688 {
 689         int ret = 0;
 690
 691         if (wait)
 692                 ret = amdgpu_sync_wait(ctx->sync, intr);
 693
 694         if (ctx->reserved)
 695                 ttm_eu_backoff_reservation(&ctx->ticket, &ctx->list);
 696         kfree(ctx->vm_pd);
 697
 698         ctx->sync = NULL;
 699
 700         ctx->reserved = false;
 701         ctx->vm_pd = NULL;
 702
 703         return ret;
 704 }
 705
 706 static int unmap_bo_from_gpuvm(struct amdgpu_device *adev,
 707                                 struct kfd_bo_va_list *entry,
 708                                 struct amdgpu_sync *sync)
 709 {
 710         struct amdgpu_bo_va *bo_va = entry->bo_va;
 711         struct amdgpu_vm *vm = bo_va->base.vm;
 712
 713         amdgpu_vm_bo_unmap(adev, bo_va, entry->va);
 714
 715         amdgpu_vm_clear_freed(adev, vm, &bo_va->last_pt_update);
 716
 717         amdgpu_sync_fence(NULL, sync, bo_va->last_pt_update, false);
 718
 719         return 0;
 720 }
 721
 722 static int update_gpuvm_pte(struct amdgpu_device *adev,
 723                 struct kfd_bo_va_list *entry,
 724                 struct amdgpu_sync *sync)
 725 {
 726         int ret;
 727         struct amdgpu_bo_va *bo_va = entry->bo_va;
 728
 729         /* Update the page tables  */
 730         ret = amdgpu_vm_bo_update(adev, bo_va, false);
 731         if (ret) {
 732                 pr_err("amdgpu_vm_bo_update failed\n");
 733                 return ret;
 734         }
 735
 736         return amdgpu_sync_fence(NULL, sync, bo_va->last_pt_update, false);
 737 }
 738
 739 static int map_bo_to_gpuvm(struct amdgpu_device *adev,
 740                 struct kfd_bo_va_list *entry, struct amdgpu_sync *sync,
 741                 bool no_update_pte)
 742 {
 743         int ret;
 744
 745         /* Set virtual address for the allocation */
 746         ret = amdgpu_vm_bo_map(adev, entry->bo_va, entry->va, 0,
 747                                amdgpu_bo_size(entry->bo_va->base.bo),
 748                                entry->pte_flags);
 749         if (ret) {
 750                 pr_err("Failed to map VA 0x%llx in vm. ret %d\n",
 751                                 entry->va, ret);
 752                 return ret;
 753         }
 754
 755         if (no_update_pte)
 756                 return 0;
 757
 758         ret = update_gpuvm_pte(adev, entry, sync);
 759         if (ret) {
 760                 pr_err("update_gpuvm_pte() failed\n");
 761                 goto update_gpuvm_pte_failed;
 762         }
 763
 764         return 0;
 765
 766 update_gpuvm_pte_failed:
 767         unmap_bo_from_gpuvm(adev, entry, sync);
 768         return ret;
 769 }
 770
 771 static struct sg_table *create_doorbell_sg(uint64_t addr, uint32_t size)
 772 {
 773         struct sg_table *sg = kmalloc(sizeof(*sg), GFP_KERNEL);
 774
 775         if (!sg)
 776                 return NULL;
 777         if (sg_alloc_table(sg, 1, GFP_KERNEL)) {
 778                 kfree(sg);
 779                 return NULL;
 780         }
 781         sg->sgl->dma_address = addr;
 782         sg->sgl->length = size;
 783 #ifdef CONFIG_NEED_SG_DMA_LENGTH
 784         sg->sgl->dma_length = size;
 785 #endif
 786         return sg;
 787 }
 788
 789 static int process_validate_vms(struct amdkfd_process_info *process_info)
 790 {
 791         struct amdgpu_vm *peer_vm;
 792         int ret;
 793
 794         list_for_each_entry(peer_vm, &process_info->vm_list_head,
 795                             vm_list_node) {
 796                 ret = vm_validate_pt_pd_bos(peer_vm);
 797                 if (ret)
 798                         return ret;
 799         }
 800
 801         return 0;
 802 }
 803
 804 static int process_sync_pds_resv(struct amdkfd_process_info *process_info,
 805                                  struct amdgpu_sync *sync)
 806 {
 807         struct amdgpu_vm *peer_vm;
 808         int ret;
 809
 810         list_for_each_entry(peer_vm, &process_info->vm_list_head,
 811                             vm_list_node) {
 812                 struct amdgpu_bo *pd = peer_vm->root.base.bo;
 813
 814                 ret = amdgpu_sync_resv(NULL,
 815                                         sync, pd->tbo.base.resv,
 816                                         AMDGPU_FENCE_OWNER_KFD, false);
 817                 if (ret)
 818                         return ret;
 819         }
 820
 821         return 0;
 822 }
 823
 824 static int process_update_pds(struct amdkfd_process_info *process_info,
 825                               struct amdgpu_sync *sync)
 826 {
 827         struct amdgpu_vm *peer_vm;
 828         int ret;
 829
 830         list_for_each_entry(peer_vm, &process_info->vm_list_head,
 831                             vm_list_node) {
 832                 ret = vm_update_pds(peer_vm, sync);
 833                 if (ret)
 834                         return ret;
 835         }
 836
 837         return 0;
 838 }
 839
 840 static int init_kfd_vm(struct amdgpu_vm *vm, void **process_info,
 841                        struct dma_fence **ef)
 842 {
 843         struct amdkfd_process_info *info = NULL;
 844         int ret;
 845
 846         if (!*process_info) {
 847                 info = kzalloc(sizeof(*info), GFP_KERNEL);
 848                 if (!info)
 849                         return -ENOMEM;
 850
 851                 mutex_init(&info->lock);
 852                 INIT_LIST_HEAD(&info->vm_list_head);
 853                 INIT_LIST_HEAD(&info->kfd_bo_list);
 854                 INIT_LIST_HEAD(&info->userptr_valid_list);
 855                 INIT_LIST_HEAD(&info->userptr_inval_list);
 856
 857                 info->eviction_fence =
 858                         amdgpu_amdkfd_fence_create(dma_fence_context_alloc(1),
 859                                                    current->mm);
 860                 if (!info->eviction_fence) {
 861                         pr_err("Failed to create eviction fence\n");
 862                         ret = -ENOMEM;
 863                         goto create_evict_fence_fail;
 864                 }
 865
 866                 info->pid = get_task_pid(current->group_leader, PIDTYPE_PID);
 867                 atomic_set(&info->evicted_bos, 0);
 868                 INIT_DELAYED_WORK(&info->restore_userptr_work,
 869                                   amdgpu_amdkfd_restore_userptr_worker);
 870
 871                 *process_info = info;
 872                 *ef = dma_fence_get(&info->eviction_fence->base);
 873         }
 874
 875         vm->process_info = *process_info;
 876
 877         /* Validate page directory and attach eviction fence */
 878         ret = amdgpu_bo_reserve(vm->root.base.bo, true);
 879         if (ret)
 880                 goto reserve_pd_fail;
 881         ret = vm_validate_pt_pd_bos(vm);
 882         if (ret) {
 883                 pr_err("validate_pt_pd_bos() failed\n");
 884                 goto validate_pd_fail;
 885         }
 886         ret = amdgpu_bo_sync_wait(vm->root.base.bo,
 887                                   AMDGPU_FENCE_OWNER_KFD, false);
 888         if (ret)
 889                 goto wait_pd_fail;
 890         ret = dma_resv_reserve_shared(vm->root.base.bo->tbo.base.resv, 1);
 891         if (ret)
 892                 goto reserve_shared_fail;
 893         amdgpu_bo_fence(vm->root.base.bo,
 894                         &vm->process_info->eviction_fence->base, true);
 895         amdgpu_bo_unreserve(vm->root.base.bo);
 896
 897         /* Update process info */
 898         mutex_lock(&vm->process_info->lock);
 899         list_add_tail(&vm->vm_list_node,
 900                         &(vm->process_info->vm_list_head));
 901         vm->process_info->n_vms++;
 902         mutex_unlock(&vm->process_info->lock);
 903
 904         return 0;
 905
 906 reserve_shared_fail:
 907 wait_pd_fail:
 908 validate_pd_fail:
 909         amdgpu_bo_unreserve(vm->root.base.bo);
 910 reserve_pd_fail:
 911         vm->process_info = NULL;
 912         if (info) {
 913                 /* Two fence references: one in info and one in *ef */
 914                 dma_fence_put(&info->eviction_fence->base);
 915                 dma_fence_put(*ef);
 916                 *ef = NULL;
 917                 *process_info = NULL;
 918                 put_pid(info->pid);
 919 create_evict_fence_fail:
 920                 mutex_destroy(&info->lock);
 921                 kfree(info);
 922         }
 923         return ret;
 924 }
 925
 926 int amdgpu_amdkfd_gpuvm_create_process_vm(struct kgd_dev *kgd, unsigned int pasid,
 927                                           void **vm, void **process_info,
 928                                           struct dma_fence **ef)
 929 {
 930         struct amdgpu_device *adev = get_amdgpu_device(kgd);
 931         struct amdgpu_vm *new_vm;
 932         int ret;
 933
 934         new_vm = kzalloc(sizeof(*new_vm), GFP_KERNEL);
 935         if (!new_vm)
 936                 return -ENOMEM;
 937
 938         /* Initialize AMDGPU part of the VM */
 939         ret = amdgpu_vm_init(adev, new_vm, AMDGPU_VM_CONTEXT_COMPUTE, pasid);
 940         if (ret) {
 941                 pr_err("Failed init vm ret %d\n", ret);
 942                 goto amdgpu_vm_init_fail;
 943         }
 944
 945         /* Initialize KFD part of the VM and process info */
 946         ret = init_kfd_vm(new_vm, process_info, ef);
 947         if (ret)
 948                 goto init_kfd_vm_fail;
 949
 950         *vm = (void *) new_vm;
 951
 952         return 0;
 953
 954 init_kfd_vm_fail:
 955         amdgpu_vm_fini(adev, new_vm);
 956 amdgpu_vm_init_fail:
 957         kfree(new_vm);
 958         return ret;
 959 }
 960
 961 int amdgpu_amdkfd_gpuvm_acquire_process_vm(struct kgd_dev *kgd,
 962                                            struct file *filp, unsigned int pasid,
 963                                            void **vm, void **process_info,
 964                                            struct dma_fence **ef)
 965 {
 966         struct amdgpu_device *adev = get_amdgpu_device(kgd);
 967         struct drm_file *drm_priv = filp->private_data;
 968         struct amdgpu_fpriv *drv_priv = drm_priv->driver_priv;
 969         struct amdgpu_vm *avm = &drv_priv->vm;
 970         int ret;
 971
 972         /* Already a compute VM? */
 973         if (avm->process_info)
 974                 return -EINVAL;
 975
 976         /* Convert VM into a compute VM */
 977         ret = amdgpu_vm_make_compute(adev, avm, pasid);
 978         if (ret)
 979                 return ret;
 980
 981         /* Initialize KFD part of the VM and process info */
 982         ret = init_kfd_vm(avm, process_info, ef);
 983         if (ret)
 984                 return ret;
 985
 986         *vm = (void *)avm;
 987
 988         return 0;
 989 }
 990
 991 void amdgpu_amdkfd_gpuvm_destroy_cb(struct amdgpu_device *adev,
 992                                     struct amdgpu_vm *vm)
 993 {
 994         struct amdkfd_process_info *process_info = vm->process_info;
 995         struct amdgpu_bo *pd = vm->root.base.bo;
 996
 997         if (!process_info)
 998                 return;
 999
1000         /* Release eviction fence from PD */
1001         amdgpu_bo_reserve(pd, false);
1002         amdgpu_bo_fence(pd, NULL, false);
1003         amdgpu_bo_unreserve(pd);
1004
1005         /* Update process info */
1006         mutex_lock(&process_info->lock);
1007         process_info->n_vms--;
1008         list_del(&vm->vm_list_node);
1009         mutex_unlock(&process_info->lock);
1010
1011         /* Release per-process resources when last compute VM is destroyed */
1012         if (!process_info->n_vms) {
1013                 WARN_ON(!list_empty(&process_info->kfd_bo_list));
1014                 WARN_ON(!list_empty(&process_info->userptr_valid_list));
1015                 WARN_ON(!list_empty(&process_info->userptr_inval_list));
1016
1017                 dma_fence_put(&process_info->eviction_fence->base);
1018                 cancel_delayed_work_sync(&process_info->restore_userptr_work);
1019                 put_pid(process_info->pid);
1020                 mutex_destroy(&process_info->lock);
1021                 kfree(process_info);
1022         }
1023 }
1024
1025 void amdgpu_amdkfd_gpuvm_destroy_process_vm(struct kgd_dev *kgd, void *vm)
1026 {
1027         struct amdgpu_device *adev = get_amdgpu_device(kgd);
1028         struct amdgpu_vm *avm = (struct amdgpu_vm *)vm;
1029
1030         if (WARN_ON(!kgd || !vm))
1031                 return;
1032
1033         pr_debug("Destroying process vm %p\n", vm);
1034
1035         /* Release the VM context */
1036         amdgpu_vm_fini(adev, avm);
1037         kfree(vm);
1038 }
1039
1040 void amdgpu_amdkfd_gpuvm_release_process_vm(struct kgd_dev *kgd, void *vm)
1041 {
1042         struct amdgpu_device *adev = get_amdgpu_device(kgd);
1043         struct amdgpu_vm *avm = (struct amdgpu_vm *)vm;
1044
1045         if (WARN_ON(!kgd || !vm))
1046                 return;
1047
1048         pr_debug("Releasing process vm %p\n", vm);
1049
1050         /* The original pasid of amdgpu vm has already been
1051          * released during making a amdgpu vm to a compute vm
1052          * The current pasid is managed by kfd and will be
1053          * released on kfd process destroy. Set amdgpu pasid
1054          * to 0 to avoid duplicate release.
1055          */
1056         amdgpu_vm_release_compute(adev, avm);
1057 }
1058
1059 uint64_t amdgpu_amdkfd_gpuvm_get_process_page_dir(void *vm)
1060 {
1061         struct amdgpu_vm *avm = (struct amdgpu_vm *)vm;
1062         struct amdgpu_bo *pd = avm->root.base.bo;
1063         struct amdgpu_device *adev = amdgpu_ttm_adev(pd->tbo.bdev);
1064
1065         if (adev->asic_type < CHIP_VEGA10)
1066                 return avm->pd_phys_addr >> AMDGPU_GPU_PAGE_SHIFT;
1067         return avm->pd_phys_addr;
1068 }
1069
1070 int amdgpu_amdkfd_gpuvm_alloc_memory_of_gpu(
1071                 struct kgd_dev *kgd, uint64_t va, uint64_t size,
1072                 void *vm, struct kgd_mem **mem,
1073                 uint64_t *offset, uint32_t flags)
1074 {
1075         struct amdgpu_device *adev = get_amdgpu_device(kgd);
1076         struct amdgpu_vm *avm = (struct amdgpu_vm *)vm;
1077         enum ttm_bo_type bo_type = ttm_bo_type_device;
1078         struct sg_table *sg = NULL;
1079         uint64_t user_addr = 0;
1080         struct amdgpu_bo *bo;
1081         struct amdgpu_bo_param bp;
1082         int byte_align;
1083         u32 domain, alloc_domain;
1084         u64 alloc_flags;
1085         uint32_t mapping_flags;
1086         int ret;
1087
1088         /*
1089          * Check on which domain to allocate BO
1090          */
1091         if (flags & ALLOC_MEM_FLAGS_VRAM) {
1092                 domain = alloc_domain = AMDGPU_GEM_DOMAIN_VRAM;
1093                 alloc_flags = AMDGPU_GEM_CREATE_VRAM_WIPE_ON_RELEASE;
1094                 alloc_flags |= (flags & ALLOC_MEM_FLAGS_PUBLIC) ?
1095                         AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED :
1096                         AMDGPU_GEM_CREATE_NO_CPU_ACCESS;
1097         } else if (flags & ALLOC_MEM_FLAGS_GTT) {
1098                 domain = alloc_domain = AMDGPU_GEM_DOMAIN_GTT;
1099                 alloc_flags = 0;
1100         } else if (flags & ALLOC_MEM_FLAGS_USERPTR) {
1101                 domain = AMDGPU_GEM_DOMAIN_GTT;
1102                 alloc_domain = AMDGPU_GEM_DOMAIN_CPU;
1103                 alloc_flags = 0;
1104                 if (!offset || !*offset)
1105                         return -EINVAL;
1106                 user_addr = *offset;
1107         } else if (flags & (ALLOC_MEM_FLAGS_DOORBELL |
1108                         ALLOC_MEM_FLAGS_MMIO_REMAP)) {
1109                 domain = AMDGPU_GEM_DOMAIN_GTT;
1110                 alloc_domain = AMDGPU_GEM_DOMAIN_CPU;
1111                 bo_type = ttm_bo_type_sg;
1112                 alloc_flags = 0;
1113                 if (size > UINT_MAX)
1114                         return -EINVAL;
1115                 sg = create_doorbell_sg(*offset, size);
1116                 if (!sg)
1117                         return -ENOMEM;
1118         } else {
1119                 return -EINVAL;
1120         }
1121
1122         *mem = kzalloc(sizeof(struct kgd_mem), GFP_KERNEL);
1123         if (!*mem) {
1124                 ret = -ENOMEM;
1125                 goto err;
1126         }
1127         INIT_LIST_HEAD(&(*mem)->bo_va_list);
1128         mutex_init(&(*mem)->lock);
1129         (*mem)->aql_queue = !!(flags & ALLOC_MEM_FLAGS_AQL_QUEUE_MEM);
1130
1131         /* Workaround for AQL queue wraparound bug. Map the same
1132          * memory twice. That means we only actually allocate half
1133          * the memory.
1134          */
1135         if ((*mem)->aql_queue)
1136                 size = size >> 1;
1137
1138         /* Workaround for TLB bug on older VI chips */
1139         byte_align = (adev->family == AMDGPU_FAMILY_VI &&
1140                         adev->asic_type != CHIP_FIJI &&
1141                         adev->asic_type != CHIP_POLARIS10 &&
1142                         adev->asic_type != CHIP_POLARIS11 &&
1143                         adev->asic_type != CHIP_POLARIS12 &&
1144                         adev->asic_type != CHIP_VEGAM) ?
1145                         VI_BO_SIZE_ALIGN : 1;
1146
1147         mapping_flags = AMDGPU_VM_PAGE_READABLE;
1148         if (flags & ALLOC_MEM_FLAGS_WRITABLE)
1149                 mapping_flags |= AMDGPU_VM_PAGE_WRITEABLE;
1150         if (flags & ALLOC_MEM_FLAGS_EXECUTABLE)
1151                 mapping_flags |= AMDGPU_VM_PAGE_EXECUTABLE;
1152         if (flags & ALLOC_MEM_FLAGS_COHERENT)
1153                 mapping_flags |= AMDGPU_VM_MTYPE_UC;
1154         else
1155                 mapping_flags |= AMDGPU_VM_MTYPE_NC;
1156         (*mem)->mapping_flags = mapping_flags;
1157
1158         amdgpu_sync_create(&(*mem)->sync);
1159
1160         ret = amdgpu_amdkfd_reserve_mem_limit(adev, size, alloc_domain, !!sg);
1161         if (ret) {
1162                 pr_debug("Insufficient system memory\n");
1163                 goto err_reserve_limit;
1164         }
1165
1166         pr_debug("\tcreate BO VA 0x%llx size 0x%llx domain %s\n",
1167                         va, size, domain_string(alloc_domain));
1168
1169         memset(&bp, 0, sizeof(bp));
1170         bp.size = size;
1171         bp.byte_align = byte_align;
1172         bp.domain = alloc_domain;
1173         bp.flags = alloc_flags;
1174         bp.type = bo_type;
1175         bp.resv = NULL;
1176         ret = amdgpu_bo_create(adev, &bp, &bo);
1177         if (ret) {
1178                 pr_debug("Failed to create BO on domain %s. ret %d\n",
1179                                 domain_string(alloc_domain), ret);
1180                 goto err_bo_create;
1181         }
1182         if (bo_type == ttm_bo_type_sg) {
1183                 bo->tbo.sg = sg;
1184                 bo->tbo.ttm->sg = sg;
1185         }
1186         bo->kfd_bo = *mem;
1187         (*mem)->bo = bo;
1188         if (user_addr)
1189                 bo->flags |= AMDGPU_AMDKFD_USERPTR_BO;
1190
1191         (*mem)->va = va;
1192         (*mem)->domain = domain;
1193         (*mem)->mapped_to_gpu_memory = 0;
1194         (*mem)->process_info = avm->process_info;
1195         add_kgd_mem_to_kfd_bo_list(*mem, avm->process_info, user_addr);
1196
1197         if (user_addr) {
1198                 ret = init_user_pages(*mem, current->mm, user_addr);
1199                 if (ret)
1200                         goto allocate_init_user_pages_failed;
1201         }
1202
1203         if (offset)
1204                 *offset = amdgpu_bo_mmap_offset(bo);
1205
1206         return 0;
1207
1208 allocate_init_user_pages_failed:
1209         remove_kgd_mem_from_kfd_bo_list(*mem, avm->process_info);
1210         amdgpu_bo_unref(&bo);
1211         /* Don't unreserve system mem limit twice */
1212         goto err_reserve_limit;
1213 err_bo_create:
1214         unreserve_mem_limit(adev, size, alloc_domain, !!sg);
1215 err_reserve_limit:
1216         mutex_destroy(&(*mem)->lock);
1217         kfree(*mem);
1218 err:
1219         if (sg) {
1220                 sg_free_table(sg);
1221                 kfree(sg);
1222         }
1223         return ret;
1224 }
1225
1226 int amdgpu_amdkfd_gpuvm_free_memory_of_gpu(
1227                 struct kgd_dev *kgd, struct kgd_mem *mem)
1228 {
1229         struct amdkfd_process_info *process_info = mem->process_info;
1230         unsigned long bo_size = mem->bo->tbo.mem.size;
1231         struct kfd_bo_va_list *entry, *tmp;
1232         struct bo_vm_reservation_context ctx;
1233         struct ttm_validate_buffer *bo_list_entry;
1234         int ret;
1235
1236         mutex_lock(&mem->lock);
1237
1238         if (mem->mapped_to_gpu_memory > 0) {
1239                 pr_debug("BO VA 0x%llx size 0x%lx is still mapped.\n",
1240                                 mem->va, bo_size);
1241                 mutex_unlock(&mem->lock);
1242                 return -EBUSY;
1243         }
1244
1245         mutex_unlock(&mem->lock);
1246         /* lock is not needed after this, since mem is unused and will
1247          * be freed anyway
1248          */
1249
1250         /* No more MMU notifiers */
1251         amdgpu_mn_unregister(mem->bo);
1252
1253         /* Make sure restore workers don't access the BO any more */
1254         bo_list_entry = &mem->validate_list;
1255         mutex_lock(&process_info->lock);
1256         list_del(&bo_list_entry->head);
1257         mutex_unlock(&process_info->lock);
1258
1259         ret = reserve_bo_and_cond_vms(mem, NULL, BO_VM_ALL, &ctx);
1260         if (unlikely(ret))
1261                 return ret;
1262
1263         /* The eviction fence should be removed by the last unmap.
1264          * TODO: Log an error condition if the bo still has the eviction fence
1265          * attached
1266          */
1267         amdgpu_amdkfd_remove_eviction_fence(mem->bo,
1268                                         process_info->eviction_fence);
1269         pr_debug("Release VA 0x%llx - 0x%llx\n", mem->va,
1270                 mem->va + bo_size * (1 + mem->aql_queue));
1271
1272         /* Remove from VM internal data structures */
1273         list_for_each_entry_safe(entry, tmp, &mem->bo_va_list, bo_list)
1274                 remove_bo_from_vm((struct amdgpu_device *)entry->kgd_dev,
1275                                 entry, bo_size);
1276
1277         ret = unreserve_bo_and_vms(&ctx, false, false);
1278
1279         /* Free the sync object */
1280         amdgpu_sync_free(&mem->sync);
1281
1282         /* If the SG is not NULL, it's one we created for a doorbell or mmio
1283          * remap BO. We need to free it.
1284          */
1285         if (mem->bo->tbo.sg) {
1286                 sg_free_table(mem->bo->tbo.sg);
1287                 kfree(mem->bo->tbo.sg);
1288         }
1289
1290         /* Free the BO*/
1291         amdgpu_bo_unref(&mem->bo);
1292         mutex_destroy(&mem->lock);
1293         kfree(mem);
1294
1295         return ret;
1296 }
1297
1298 int amdgpu_amdkfd_gpuvm_map_memory_to_gpu(
1299                 struct kgd_dev *kgd, struct kgd_mem *mem, void *vm)
1300 {
1301         struct amdgpu_device *adev = get_amdgpu_device(kgd);
1302         struct amdgpu_vm *avm = (struct amdgpu_vm *)vm;
1303         int ret;
1304         struct amdgpu_bo *bo;
1305         uint32_t domain;
1306         struct kfd_bo_va_list *entry;
1307         struct bo_vm_reservation_context ctx;
1308         struct kfd_bo_va_list *bo_va_entry = NULL;
1309         struct kfd_bo_va_list *bo_va_entry_aql = NULL;
1310         unsigned long bo_size;
1311         bool is_invalid_userptr = false;
1312
1313         bo = mem->bo;
1314         if (!bo) {
1315                 pr_err("Invalid BO when mapping memory to GPU\n");
1316                 return -EINVAL;
1317         }
1318
1319         /* Make sure restore is not running concurrently. Since we
1320          * don't map invalid userptr BOs, we rely on the next restore
1321          * worker to do the mapping
1322          */
1323         mutex_lock(&mem->process_info->lock);
1324
1325         /* Lock mmap-sem. If we find an invalid userptr BO, we can be
1326          * sure that the MMU notifier is no longer running
1327          * concurrently and the queues are actually stopped
1328          */
1329         if (amdgpu_ttm_tt_get_usermm(bo->tbo.ttm)) {
1330                 down_write(&current->mm->mmap_sem);
1331                 is_invalid_userptr = atomic_read(&mem->invalid);
1332                 up_write(&current->mm->mmap_sem);
1333         }
1334
1335         mutex_lock(&mem->lock);
1336
1337         domain = mem->domain;
1338         bo_size = bo->tbo.mem.size;
1339
1340         pr_debug("Map VA 0x%llx - 0x%llx to vm %p domain %s\n",
1341                         mem->va,
1342                         mem->va + bo_size * (1 + mem->aql_queue),
1343                         vm, domain_string(domain));
1344
1345         ret = reserve_bo_and_vm(mem, vm, &ctx);
1346         if (unlikely(ret))
1347                 goto out;
1348
1349         /* Userptr can be marked as "not invalid", but not actually be
1350          * validated yet (still in the system domain). In that case
1351          * the queues are still stopped and we can leave mapping for
1352          * the next restore worker
1353          */
1354         if (amdgpu_ttm_tt_get_usermm(bo->tbo.ttm) &&
1355             bo->tbo.mem.mem_type == TTM_PL_SYSTEM)
1356                 is_invalid_userptr = true;
1357
1358         if (check_if_add_bo_to_vm(avm, mem)) {
1359                 ret = add_bo_to_vm(adev, mem, avm, false,
1360                                 &bo_va_entry);
1361                 if (ret)
1362                         goto add_bo_to_vm_failed;
1363                 if (mem->aql_queue) {
1364                         ret = add_bo_to_vm(adev, mem, avm,
1365                                         true, &bo_va_entry_aql);
1366                         if (ret)
1367                                 goto add_bo_to_vm_failed_aql;
1368                 }
1369         } else {
1370                 ret = vm_validate_pt_pd_bos(avm);
1371                 if (unlikely(ret))
1372                         goto add_bo_to_vm_failed;
1373         }
1374
1375         if (mem->mapped_to_gpu_memory == 0 &&
1376             !amdgpu_ttm_tt_get_usermm(bo->tbo.ttm)) {
1377                 /* Validate BO only once. The eviction fence gets added to BO
1378                  * the first time it is mapped. Validate will wait for all
1379                  * background evictions to complete.
1380                  */
1381                 ret = amdgpu_amdkfd_bo_validate(bo, domain, true);
1382                 if (ret) {
1383                         pr_debug("Validate failed\n");
1384                         goto map_bo_to_gpuvm_failed;
1385                 }
1386         }
1387
1388         list_for_each_entry(entry, &mem->bo_va_list, bo_list) {
1389                 if (entry->bo_va->base.vm == vm && !entry->is_mapped) {
1390                         pr_debug("\t map VA 0x%llx - 0x%llx in entry %p\n",
1391                                         entry->va, entry->va + bo_size,
1392                                         entry);
1393
1394                         ret = map_bo_to_gpuvm(adev, entry, ctx.sync,
1395                                               is_invalid_userptr);
1396                         if (ret) {
1397                                 pr_err("Failed to map bo to gpuvm\n");
1398                                 goto map_bo_to_gpuvm_failed;
1399                         }
1400
1401                         ret = vm_update_pds(vm, ctx.sync);
1402                         if (ret) {
1403                                 pr_err("Failed to update page directories\n");
1404                                 goto map_bo_to_gpuvm_failed;
1405                         }
1406
1407                         entry->is_mapped = true;
1408                         mem->mapped_to_gpu_memory++;
1409                         pr_debug("\t INC mapping count %d\n",
1410                                         mem->mapped_to_gpu_memory);
1411                 }
1412         }
1413
1414         if (!amdgpu_ttm_tt_get_usermm(bo->tbo.ttm) && !bo->pin_count)
1415                 amdgpu_bo_fence(bo,
1416                                 &avm->process_info->eviction_fence->base,
1417                                 true);
1418         ret = unreserve_bo_and_vms(&ctx, false, false);
1419
1420         goto out;
1421
1422 map_bo_to_gpuvm_failed:
1423         if (bo_va_entry_aql)
1424                 remove_bo_from_vm(adev, bo_va_entry_aql, bo_size);
1425 add_bo_to_vm_failed_aql:
1426         if (bo_va_entry)
1427                 remove_bo_from_vm(adev, bo_va_entry, bo_size);
1428 add_bo_to_vm_failed:
1429         unreserve_bo_and_vms(&ctx, false, false);
1430 out:
1431         mutex_unlock(&mem->process_info->lock);
1432         mutex_unlock(&mem->lock);
1433         return ret;
1434 }
1435
1436 int amdgpu_amdkfd_gpuvm_unmap_memory_from_gpu(
1437                 struct kgd_dev *kgd, struct kgd_mem *mem, void *vm)
1438 {
1439         struct amdgpu_device *adev = get_amdgpu_device(kgd);
1440         struct amdkfd_process_info *process_info =
1441                 ((struct amdgpu_vm *)vm)->process_info;
1442         unsigned long bo_size = mem->bo->tbo.mem.size;
1443         struct kfd_bo_va_list *entry;
1444         struct bo_vm_reservation_context ctx;
1445         int ret;
1446
1447         mutex_lock(&mem->lock);
1448
1449         ret = reserve_bo_and_cond_vms(mem, vm, BO_VM_MAPPED, &ctx);
1450         if (unlikely(ret))
1451                 goto out;
1452         /* If no VMs were reserved, it means the BO wasn't actually mapped */
1453         if (ctx.n_vms == 0) {
1454                 ret = -EINVAL;
1455                 goto unreserve_out;
1456         }
1457
1458         ret = vm_validate_pt_pd_bos((struct amdgpu_vm *)vm);
1459         if (unlikely(ret))
1460                 goto unreserve_out;
1461
1462         pr_debug("Unmap VA 0x%llx - 0x%llx from vm %p\n",
1463                 mem->va,
1464                 mem->va + bo_size * (1 + mem->aql_queue),
1465                 vm);
1466
1467         list_for_each_entry(entry, &mem->bo_va_list, bo_list) {
1468                 if (entry->bo_va->base.vm == vm && entry->is_mapped) {
1469                         pr_debug("\t unmap VA 0x%llx - 0x%llx from entry %p\n",
1470                                         entry->va,
1471                                         entry->va + bo_size,
1472                                         entry);
1473
1474                         ret = unmap_bo_from_gpuvm(adev, entry, ctx.sync);
1475                         if (ret == 0) {
1476                                 entry->is_mapped = false;
1477                         } else {
1478                                 pr_err("failed to unmap VA 0x%llx\n",
1479                                                 mem->va);
1480                                 goto unreserve_out;
1481                         }
1482
1483                         mem->mapped_to_gpu_memory--;
1484                         pr_debug("\t DEC mapping count %d\n",
1485                                         mem->mapped_to_gpu_memory);
1486                 }
1487         }
1488
1489         /* If BO is unmapped from all VMs, unfence it. It can be evicted if
1490          * required.
1491          */
1492         if (mem->mapped_to_gpu_memory == 0 &&
1493             !amdgpu_ttm_tt_get_usermm(mem->bo->tbo.ttm) && !mem->bo->pin_count)
1494                 amdgpu_amdkfd_remove_eviction_fence(mem->bo,
1495                                                 process_info->eviction_fence);
1496
1497 unreserve_out:
1498         unreserve_bo_and_vms(&ctx, false, false);
1499 out:
1500         mutex_unlock(&mem->lock);
1501         return ret;
1502 }
1503
1504 int amdgpu_amdkfd_gpuvm_sync_memory(
1505                 struct kgd_dev *kgd, struct kgd_mem *mem, bool intr)
1506 {
1507         struct amdgpu_sync sync;
1508         int ret;
1509
1510         amdgpu_sync_create(&sync);
1511
1512         mutex_lock(&mem->lock);
1513         amdgpu_sync_clone(&mem->sync, &sync);
1514         mutex_unlock(&mem->lock);
1515
1516         ret = amdgpu_sync_wait(&sync, intr);
1517         amdgpu_sync_free(&sync);
1518         return ret;
1519 }
1520
1521 int amdgpu_amdkfd_gpuvm_map_gtt_bo_to_kernel(struct kgd_dev *kgd,
1522                 struct kgd_mem *mem, void **kptr, uint64_t *size)
1523 {
1524         int ret;
1525         struct amdgpu_bo *bo = mem->bo;
1526
1527         if (amdgpu_ttm_tt_get_usermm(bo->tbo.ttm)) {
1528                 pr_err("userptr can't be mapped to kernel\n");
1529                 return -EINVAL;
1530         }
1531
1532         /* delete kgd_mem from kfd_bo_list to avoid re-validating
1533          * this BO in BO's restoring after eviction.
1534          */
1535         mutex_lock(&mem->process_info->lock);
1536
1537         ret = amdgpu_bo_reserve(bo, true);
1538         if (ret) {
1539                 pr_err("Failed to reserve bo. ret %d\n", ret);
1540                 goto bo_reserve_failed;
1541         }
1542
1543         ret = amdgpu_bo_pin(bo, AMDGPU_GEM_DOMAIN_GTT);
1544         if (ret) {
1545                 pr_err("Failed to pin bo. ret %d\n", ret);
1546                 goto pin_failed;
1547         }
1548
1549         ret = amdgpu_bo_kmap(bo, kptr);
1550         if (ret) {
1551                 pr_err("Failed to map bo to kernel. ret %d\n", ret);
1552                 goto kmap_failed;
1553         }
1554
1555         amdgpu_amdkfd_remove_eviction_fence(
1556                 bo, mem->process_info->eviction_fence);
1557         list_del_init(&mem->validate_list.head);
1558
1559         if (size)
1560                 *size = amdgpu_bo_size(bo);
1561
1562         amdgpu_bo_unreserve(bo);
1563
1564         mutex_unlock(&mem->process_info->lock);
1565         return 0;
1566
1567 kmap_failed:
1568         amdgpu_bo_unpin(bo);
1569 pin_failed:
1570         amdgpu_bo_unreserve(bo);
1571 bo_reserve_failed:
1572         mutex_unlock(&mem->process_info->lock);
1573
1574         return ret;
1575 }
1576
1577 int amdgpu_amdkfd_gpuvm_get_vm_fault_info(struct kgd_dev *kgd,
1578                                               struct kfd_vm_fault_info *mem)
1579 {
1580         struct amdgpu_device *adev;
1581
1582         adev = (struct amdgpu_device *)kgd;
1583         if (atomic_read(&adev->gmc.vm_fault_info_updated) == 1) {
1584                 *mem = *adev->gmc.vm_fault_info;
1585                 mb();
1586                 atomic_set(&adev->gmc.vm_fault_info_updated, 0);
1587         }
1588         return 0;
1589 }
1590
1591 int amdgpu_amdkfd_gpuvm_import_dmabuf(struct kgd_dev *kgd,
1592                                       struct dma_buf *dma_buf,
1593                                       uint64_t va, void *vm,
1594                                       struct kgd_mem **mem, uint64_t *size,
1595                                       uint64_t *mmap_offset)
1596 {
1597         struct amdgpu_device *adev = (struct amdgpu_device *)kgd;
1598         struct drm_gem_object *obj;
1599         struct amdgpu_bo *bo;
1600         struct amdgpu_vm *avm = (struct amdgpu_vm *)vm;
1601
1602         if (dma_buf->ops != &amdgpu_dmabuf_ops)
1603                 /* Can't handle non-graphics buffers */
1604                 return -EINVAL;
1605
1606         obj = dma_buf->priv;
1607         if (obj->dev->dev_private != adev)
1608                 /* Can't handle buffers from other devices */
1609                 return -EINVAL;
1610
1611         bo = gem_to_amdgpu_bo(obj);
1612         if (!(bo->preferred_domains & (AMDGPU_GEM_DOMAIN_VRAM |
1613                                     AMDGPU_GEM_DOMAIN_GTT)))
1614                 /* Only VRAM and GTT BOs are supported */
1615                 return -EINVAL;
1616
1617         *mem = kzalloc(sizeof(struct kgd_mem), GFP_KERNEL);
1618         if (!*mem)
1619                 return -ENOMEM;
1620
1621         if (size)
1622                 *size = amdgpu_bo_size(bo);
1623
1624         if (mmap_offset)
1625                 *mmap_offset = amdgpu_bo_mmap_offset(bo);
1626
1627         INIT_LIST_HEAD(&(*mem)->bo_va_list);
1628         mutex_init(&(*mem)->lock);
1629         (*mem)->mapping_flags =
1630                 AMDGPU_VM_PAGE_READABLE | AMDGPU_VM_PAGE_WRITEABLE |
1631                 AMDGPU_VM_PAGE_EXECUTABLE | AMDGPU_VM_MTYPE_NC;
1632
1633         (*mem)->bo = amdgpu_bo_ref(bo);
1634         (*mem)->va = va;
1635         (*mem)->domain = (bo->preferred_domains & AMDGPU_GEM_DOMAIN_VRAM) ?
1636                 AMDGPU_GEM_DOMAIN_VRAM : AMDGPU_GEM_DOMAIN_GTT;
1637         (*mem)->mapped_to_gpu_memory = 0;
1638         (*mem)->process_info = avm->process_info;
1639         add_kgd_mem_to_kfd_bo_list(*mem, avm->process_info, false);
1640         amdgpu_sync_create(&(*mem)->sync);
1641
1642         return 0;
1643 }
1644
1645 /* Evict a userptr BO by stopping the queues if necessary
1646  *
1647  * Runs in MMU notifier, may be in RECLAIM_FS context. This means it
1648  * cannot do any memory allocations, and cannot take any locks that
1649  * are held elsewhere while allocating memory. Therefore this is as
1650  * simple as possible, using atomic counters.
1651  *
1652  * It doesn't do anything to the BO itself. The real work happens in
1653  * restore, where we get updated page addresses. This function only
1654  * ensures that GPU access to the BO is stopped.
1655  */
1656 int amdgpu_amdkfd_evict_userptr(struct kgd_mem *mem,
1657                                 struct mm_struct *mm)
1658 {
1659         struct amdkfd_process_info *process_info = mem->process_info;
1660         int invalid, evicted_bos;
1661         int r = 0;
1662
1663         invalid = atomic_inc_return(&mem->invalid);
1664         evicted_bos = atomic_inc_return(&process_info->evicted_bos);
1665         if (evicted_bos == 1) {
1666                 /* First eviction, stop the queues */
1667                 r = kgd2kfd_quiesce_mm(mm);
1668                 if (r)
1669                         pr_err("Failed to quiesce KFD\n");
1670                 schedule_delayed_work(&process_info->restore_userptr_work,
1671                         msecs_to_jiffies(AMDGPU_USERPTR_RESTORE_DELAY_MS));
1672         }
1673
1674         return r;
1675 }
1676
1677 /* Update invalid userptr BOs
1678  *
1679  * Moves invalidated (evicted) userptr BOs from userptr_valid_list to
1680  * userptr_inval_list and updates user pages for all BOs that have
1681  * been invalidated since their last update.
1682  */
1683 static int update_invalid_user_pages(struct amdkfd_process_info *process_info,
1684                                      struct mm_struct *mm)
1685 {
1686         struct kgd_mem *mem, *tmp_mem;
1687         struct amdgpu_bo *bo;
1688         struct ttm_operation_ctx ctx = { false, false };
1689         int invalid, ret;
1690
1691         /* Move all invalidated BOs to the userptr_inval_list and
1692          * release their user pages by migration to the CPU domain
1693          */
1694         list_for_each_entry_safe(mem, tmp_mem,
1695                                  &process_info->userptr_valid_list,
1696                                  validate_list.head) {
1697                 if (!atomic_read(&mem->invalid))
1698                         continue; /* BO is still valid */
1699
1700                 bo = mem->bo;
1701
1702                 if (amdgpu_bo_reserve(bo, true))
1703                         return -EAGAIN;
1704                 amdgpu_bo_placement_from_domain(bo, AMDGPU_GEM_DOMAIN_CPU);
1705                 ret = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
1706                 amdgpu_bo_unreserve(bo);
1707                 if (ret) {
1708                         pr_err("%s: Failed to invalidate userptr BO\n",
1709                                __func__);
1710                         return -EAGAIN;
1711                 }
1712
1713                 list_move_tail(&mem->validate_list.head,
1714                                &process_info->userptr_inval_list);
1715         }
1716
1717         if (list_empty(&process_info->userptr_inval_list))
1718                 return 0; /* All evicted userptr BOs were freed */
1719
1720         /* Go through userptr_inval_list and update any invalid user_pages */
1721         list_for_each_entry(mem, &process_info->userptr_inval_list,
1722                             validate_list.head) {
1723                 invalid = atomic_read(&mem->invalid);
1724                 if (!invalid)
1725                         /* BO hasn't been invalidated since the last
1726                          * revalidation attempt. Keep its BO list.
1727                          */
1728                         continue;
1729
1730                 bo = mem->bo;
1731
1732                 /* Get updated user pages */
1733                 ret = amdgpu_ttm_tt_get_user_pages(bo, bo->tbo.ttm->pages);
1734                 if (ret) {
1735                         pr_debug("%s: Failed to get user pages: %d\n",
1736                                 __func__, ret);
1737
1738                         /* Return error -EBUSY or -ENOMEM, retry restore */
1739                         return ret;
1740                 }
1741
1742                 amdgpu_ttm_tt_get_user_pages_done(bo->tbo.ttm);
1743
1744                 /* Mark the BO as valid unless it was invalidated
1745                  * again concurrently.
1746                  */
1747                 if (atomic_cmpxchg(&mem->invalid, invalid, 0) != invalid)
1748                         return -EAGAIN;
1749         }
1750
1751         return 0;
1752 }
1753
1754 /* Validate invalid userptr BOs
1755  *
1756  * Validates BOs on the userptr_inval_list, and moves them back to the
1757  * userptr_valid_list. Also updates GPUVM page tables with new page
1758  * addresses and waits for the page table updates to complete.
1759  */
1760 static int validate_invalid_user_pages(struct amdkfd_process_info *process_info)
1761 {
1762         struct amdgpu_bo_list_entry *pd_bo_list_entries;
1763         struct list_head resv_list, duplicates;
1764         struct ww_acquire_ctx ticket;
1765         struct amdgpu_sync sync;
1766
1767         struct amdgpu_vm *peer_vm;
1768         struct kgd_mem *mem, *tmp_mem;
1769         struct amdgpu_bo *bo;
1770         struct ttm_operation_ctx ctx = { false, false };
1771         int i, ret;
1772
1773         pd_bo_list_entries = kcalloc(process_info->n_vms,
1774                                      sizeof(struct amdgpu_bo_list_entry),
1775                                      GFP_KERNEL);
1776         if (!pd_bo_list_entries) {
1777                 pr_err("%s: Failed to allocate PD BO list entries\n", __func__);
1778                 ret = -ENOMEM;
1779                 goto out_no_mem;
1780         }
1781
1782         INIT_LIST_HEAD(&resv_list);
1783         INIT_LIST_HEAD(&duplicates);
1784
1785         /* Get all the page directory BOs that need to be reserved */
1786         i = 0;
1787         list_for_each_entry(peer_vm, &process_info->vm_list_head,
1788                             vm_list_node)
1789                 amdgpu_vm_get_pd_bo(peer_vm, &resv_list,
1790                                     &pd_bo_list_entries[i++]);
1791         /* Add the userptr_inval_list entries to resv_list */
1792         list_for_each_entry(mem, &process_info->userptr_inval_list,
1793                             validate_list.head) {
1794                 list_add_tail(&mem->resv_list.head, &resv_list);
1795                 mem->resv_list.bo = mem->validate_list.bo;
1796                 mem->resv_list.num_shared = mem->validate_list.num_shared;
1797         }
1798
1799         /* Reserve all BOs and page tables for validation */
1800         ret = ttm_eu_reserve_buffers(&ticket, &resv_list, false, &duplicates,
1801                                      true);
1802         WARN(!list_empty(&duplicates), "Duplicates should be empty");
1803         if (ret)
1804                 goto out_free;
1805
1806         amdgpu_sync_create(&sync);
1807
1808         ret = process_validate_vms(process_info);
1809         if (ret)
1810                 goto unreserve_out;
1811
1812         /* Validate BOs and update GPUVM page tables */
1813         list_for_each_entry_safe(mem, tmp_mem,
1814                                  &process_info->userptr_inval_list,
1815                                  validate_list.head) {
1816                 struct kfd_bo_va_list *bo_va_entry;
1817
1818                 bo = mem->bo;
1819
1820                 /* Validate the BO if we got user pages */
1821                 if (bo->tbo.ttm->pages[0]) {
1822                         amdgpu_bo_placement_from_domain(bo, mem->domain);
1823                         ret = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
1824                         if (ret) {
1825                                 pr_err("%s: failed to validate BO\n", __func__);
1826                                 goto unreserve_out;
1827                         }
1828                 }
1829
1830                 list_move_tail(&mem->validate_list.head,
1831                                &process_info->userptr_valid_list);
1832
1833                 /* Update mapping. If the BO was not validated
1834                  * (because we couldn't get user pages), this will
1835                  * clear the page table entries, which will result in
1836                  * VM faults if the GPU tries to access the invalid
1837                  * memory.
1838                  */
1839                 list_for_each_entry(bo_va_entry, &mem->bo_va_list, bo_list) {
1840                         if (!bo_va_entry->is_mapped)
1841                                 continue;
1842
1843                         ret = update_gpuvm_pte((struct amdgpu_device *)
1844                                                bo_va_entry->kgd_dev,
1845                                                bo_va_entry, &sync);
1846                         if (ret) {
1847                                 pr_err("%s: update PTE failed\n", __func__);
1848                                 /* make sure this gets validated again */
1849                                 atomic_inc(&mem->invalid);
1850                                 goto unreserve_out;
1851                         }
1852                 }
1853         }
1854
1855         /* Update page directories */
1856         ret = process_update_pds(process_info, &sync);
1857
1858 unreserve_out:
1859         ttm_eu_backoff_reservation(&ticket, &resv_list);
1860         amdgpu_sync_wait(&sync, false);
1861         amdgpu_sync_free(&sync);
1862 out_free:
1863         kfree(pd_bo_list_entries);
1864 out_no_mem:
1865
1866         return ret;
1867 }
1868
1869 /* Worker callback to restore evicted userptr BOs
1870  *
1871  * Tries to update and validate all userptr BOs. If successful and no
1872  * concurrent evictions happened, the queues are restarted. Otherwise,
1873  * reschedule for another attempt later.
1874  */
1875 static void amdgpu_amdkfd_restore_userptr_worker(struct work_struct *work)
1876 {
1877         struct delayed_work *dwork = to_delayed_work(work);
1878         struct amdkfd_process_info *process_info =
1879                 container_of(dwork, struct amdkfd_process_info,
1880                              restore_userptr_work);
1881         struct task_struct *usertask;
1882         struct mm_struct *mm;
1883         int evicted_bos;
1884
1885         evicted_bos = atomic_read(&process_info->evicted_bos);
1886         if (!evicted_bos)
1887                 return;
1888
1889         /* Reference task and mm in case of concurrent process termination */
1890         usertask = get_pid_task(process_info->pid, PIDTYPE_PID);
1891         if (!usertask)
1892                 return;
1893         mm = get_task_mm(usertask);
1894         if (!mm) {
1895                 put_task_struct(usertask);
1896                 return;
1897         }
1898
1899         mutex_lock(&process_info->lock);
1900
1901         if (update_invalid_user_pages(process_info, mm))
1902                 goto unlock_out;
1903         /* userptr_inval_list can be empty if all evicted userptr BOs
1904          * have been freed. In that case there is nothing to validate
1905          * and we can just restart the queues.
1906          */
1907         if (!list_empty(&process_info->userptr_inval_list)) {
1908                 if (atomic_read(&process_info->evicted_bos) != evicted_bos)
1909                         goto unlock_out; /* Concurrent eviction, try again */
1910
1911                 if (validate_invalid_user_pages(process_info))
1912                         goto unlock_out;
1913         }
1914         /* Final check for concurrent evicton and atomic update. If
1915          * another eviction happens after successful update, it will
1916          * be a first eviction that calls quiesce_mm. The eviction
1917          * reference counting inside KFD will handle this case.
1918          */
1919         if (atomic_cmpxchg(&process_info->evicted_bos, evicted_bos, 0) !=
1920             evicted_bos)
1921                 goto unlock_out;
1922         evicted_bos = 0;
1923         if (kgd2kfd_resume_mm(mm)) {
1924                 pr_err("%s: Failed to resume KFD\n", __func__);
1925                 /* No recovery from this failure. Probably the CP is
1926                  * hanging. No point trying again.
1927                  */
1928         }
1929
1930 unlock_out:
1931         mutex_unlock(&process_info->lock);
1932         mmput(mm);
1933         put_task_struct(usertask);
1934
1935         /* If validation failed, reschedule another attempt */
1936         if (evicted_bos)
1937                 schedule_delayed_work(&process_info->restore_userptr_work,
1938                         msecs_to_jiffies(AMDGPU_USERPTR_RESTORE_DELAY_MS));
1939 }
1940
1941 /** amdgpu_amdkfd_gpuvm_restore_process_bos - Restore all BOs for the given
1942  *   KFD process identified by process_info
1943  *
1944  * @process_info: amdkfd_process_info of the KFD process
1945  *
1946  * After memory eviction, restore thread calls this function. The function
1947  * should be called when the Process is still valid. BO restore involves -
1948  *
1949  * 1.  Release old eviction fence and create new one
1950  * 2.  Get two copies of PD BO list from all the VMs. Keep one copy as pd_list.
1951  * 3   Use the second PD list and kfd_bo_list to create a list (ctx.list) of
1952  *     BOs that need to be reserved.
1953  * 4.  Reserve all the BOs
1954  * 5.  Validate of PD and PT BOs.
1955  * 6.  Validate all KFD BOs using kfd_bo_list and Map them and add new fence
1956  * 7.  Add fence to all PD and PT BOs.
1957  * 8.  Unreserve all BOs
1958  */
1959 int amdgpu_amdkfd_gpuvm_restore_process_bos(void *info, struct dma_fence **ef)
1960 {
1961         struct amdgpu_bo_list_entry *pd_bo_list;
1962         struct amdkfd_process_info *process_info = info;
1963         struct amdgpu_vm *peer_vm;
1964         struct kgd_mem *mem;
1965         struct bo_vm_reservation_context ctx;
1966         struct amdgpu_amdkfd_fence *new_fence;
1967         int ret = 0, i;
1968         struct list_head duplicate_save;
1969         struct amdgpu_sync sync_obj;
1970
1971         INIT_LIST_HEAD(&duplicate_save);
1972         INIT_LIST_HEAD(&ctx.list);
1973         INIT_LIST_HEAD(&ctx.duplicates);
1974
1975         pd_bo_list = kcalloc(process_info->n_vms,
1976                              sizeof(struct amdgpu_bo_list_entry),
1977                              GFP_KERNEL);
1978         if (!pd_bo_list)
1979                 return -ENOMEM;
1980
1981         i = 0;
1982         mutex_lock(&process_info->lock);
1983         list_for_each_entry(peer_vm, &process_info->vm_list_head,
1984                         vm_list_node)
1985                 amdgpu_vm_get_pd_bo(peer_vm, &ctx.list, &pd_bo_list[i++]);
1986
1987         /* Reserve all BOs and page tables/directory. Add all BOs from
1988          * kfd_bo_list to ctx.list
1989          */
1990         list_for_each_entry(mem, &process_info->kfd_bo_list,
1991                             validate_list.head) {
1992
1993                 list_add_tail(&mem->resv_list.head, &ctx.list);
1994                 mem->resv_list.bo = mem->validate_list.bo;
1995                 mem->resv_list.num_shared = mem->validate_list.num_shared;
1996         }
1997
1998         ret = ttm_eu_reserve_buffers(&ctx.ticket, &ctx.list,
1999                                      false, &duplicate_save, true);
2000         if (ret) {
2001                 pr_debug("Memory eviction: TTM Reserve Failed. Try again\n");
2002                 goto ttm_reserve_fail;
2003         }
2004
2005         amdgpu_sync_create(&sync_obj);
2006
2007         /* Validate PDs and PTs */
2008         ret = process_validate_vms(process_info);
2009         if (ret)
2010                 goto validate_map_fail;
2011
2012         ret = process_sync_pds_resv(process_info, &sync_obj);
2013         if (ret) {
2014                 pr_debug("Memory eviction: Failed to sync to PD BO moving fence. Try again\n");
2015                 goto validate_map_fail;
2016         }
2017
2018         /* Validate BOs and map them to GPUVM (update VM page tables). */
2019         list_for_each_entry(mem, &process_info->kfd_bo_list,
2020                             validate_list.head) {
2021
2022                 struct amdgpu_bo *bo = mem->bo;
2023                 uint32_t domain = mem->domain;
2024                 struct kfd_bo_va_list *bo_va_entry;
2025
2026                 ret = amdgpu_amdkfd_bo_validate(bo, domain, false);
2027                 if (ret) {
2028                         pr_debug("Memory eviction: Validate BOs failed. Try again\n");
2029                         goto validate_map_fail;
2030                 }
2031                 ret = amdgpu_sync_fence(NULL, &sync_obj, bo->tbo.moving, false);
2032                 if (ret) {
2033                         pr_debug("Memory eviction: Sync BO fence failed. Try again\n");
2034                         goto validate_map_fail;
2035                 }
2036                 list_for_each_entry(bo_va_entry, &mem->bo_va_list,
2037                                     bo_list) {
2038                         ret = update_gpuvm_pte((struct amdgpu_device *)
2039                                               bo_va_entry->kgd_dev,
2040                                               bo_va_entry,
2041                                               &sync_obj);
2042                         if (ret) {
2043                                 pr_debug("Memory eviction: update PTE failed. Try again\n");
2044                                 goto validate_map_fail;
2045                         }
2046                 }
2047         }
2048
2049         /* Update page directories */
2050         ret = process_update_pds(process_info, &sync_obj);
2051         if (ret) {
2052                 pr_debug("Memory eviction: update PDs failed. Try again\n");
2053                 goto validate_map_fail;
2054         }
2055
2056         /* Wait for validate and PT updates to finish */
2057         amdgpu_sync_wait(&sync_obj, false);
2058
2059         /* Release old eviction fence and create new one, because fence only
2060          * goes from unsignaled to signaled, fence cannot be reused.
2061          * Use context and mm from the old fence.
2062          */
2063         new_fence = amdgpu_amdkfd_fence_create(
2064                                 process_info->eviction_fence->base.context,
2065                                 process_info->eviction_fence->mm);
2066         if (!new_fence) {
2067                 pr_err("Failed to create eviction fence\n");
2068                 ret = -ENOMEM;
2069                 goto validate_map_fail;
2070         }
2071         dma_fence_put(&process_info->eviction_fence->base);
2072         process_info->eviction_fence = new_fence;
2073         *ef = dma_fence_get(&new_fence->base);
2074
2075         /* Attach new eviction fence to all BOs */
2076         list_for_each_entry(mem, &process_info->kfd_bo_list,
2077                 validate_list.head)
2078                 amdgpu_bo_fence(mem->bo,
2079                         &process_info->eviction_fence->base, true);
2080
2081         /* Attach eviction fence to PD / PT BOs */
2082         list_for_each_entry(peer_vm, &process_info->vm_list_head,
2083                             vm_list_node) {
2084                 struct amdgpu_bo *bo = peer_vm->root.base.bo;
2085
2086                 amdgpu_bo_fence(bo, &process_info->eviction_fence->base, true);
2087         }
2088
2089 validate_map_fail:
2090         ttm_eu_backoff_reservation(&ctx.ticket, &ctx.list);
2091         amdgpu_sync_free(&sync_obj);
2092 ttm_reserve_fail:
2093         mutex_unlock(&process_info->lock);
2094         kfree(pd_bo_list);
2095         return ret;
2096 }
2097
2098 int amdgpu_amdkfd_add_gws_to_process(void *info, void *gws, struct kgd_mem **mem)
2099 {
2100         struct amdkfd_process_info *process_info = (struct amdkfd_process_info *)info;
2101         struct amdgpu_bo *gws_bo = (struct amdgpu_bo *)gws;
2102         int ret;
2103
2104         if (!info || !gws)
2105                 return -EINVAL;
2106
2107         *mem = kzalloc(sizeof(struct kgd_mem), GFP_KERNEL);
2108         if (!*mem)
2109                 return -ENOMEM;
2110
2111         mutex_init(&(*mem)->lock);
2112         (*mem)->bo = amdgpu_bo_ref(gws_bo);
2113         (*mem)->domain = AMDGPU_GEM_DOMAIN_GWS;
2114         (*mem)->process_info = process_info;
2115         add_kgd_mem_to_kfd_bo_list(*mem, process_info, false);
2116         amdgpu_sync_create(&(*mem)->sync);
2117
2118
2119         /* Validate gws bo the first time it is added to process */
2120         mutex_lock(&(*mem)->process_info->lock);
2121         ret = amdgpu_bo_reserve(gws_bo, false);
2122         if (unlikely(ret)) {
2123                 pr_err("Reserve gws bo failed %d\n", ret);
2124                 goto bo_reservation_failure;
2125         }
2126
2127         ret = amdgpu_amdkfd_bo_validate(gws_bo, AMDGPU_GEM_DOMAIN_GWS, true);
2128         if (ret) {
2129                 pr_err("GWS BO validate failed %d\n", ret);
2130                 goto bo_validation_failure;
2131         }
2132         /* GWS resource is shared b/t amdgpu and amdkfd
2133          * Add process eviction fence to bo so they can
2134          * evict each other.
2135          */
2136         ret = dma_resv_reserve_shared(gws_bo->tbo.base.resv, 1);
2137         if (ret)
2138                 goto reserve_shared_fail;
2139         amdgpu_bo_fence(gws_bo, &process_info->eviction_fence->base, true);
2140         amdgpu_bo_unreserve(gws_bo);
2141         mutex_unlock(&(*mem)->process_info->lock);
2142
2143         return ret;
2144
2145 reserve_shared_fail:
2146 bo_validation_failure:
2147         amdgpu_bo_unreserve(gws_bo);
2148 bo_reservation_failure:
2149         mutex_unlock(&(*mem)->process_info->lock);
2150         amdgpu_sync_free(&(*mem)->sync);
2151         remove_kgd_mem_from_kfd_bo_list(*mem, process_info);
2152         amdgpu_bo_unref(&gws_bo);
2153         mutex_destroy(&(*mem)->lock);
2154         kfree(*mem);
2155         *mem = NULL;
2156         return ret;
2157 }
2158
2159 int amdgpu_amdkfd_remove_gws_from_process(void *info, void *mem)
2160 {
2161         int ret;
2162         struct amdkfd_process_info *process_info = (struct amdkfd_process_info *)info;
2163         struct kgd_mem *kgd_mem = (struct kgd_mem *)mem;
2164         struct amdgpu_bo *gws_bo = kgd_mem->bo;
2165
2166         /* Remove BO from process's validate list so restore worker won't touch
2167          * it anymore
2168          */
2169         remove_kgd_mem_from_kfd_bo_list(kgd_mem, process_info);
2170
2171         ret = amdgpu_bo_reserve(gws_bo, false);
2172         if (unlikely(ret)) {
2173                 pr_err("Reserve gws bo failed %d\n", ret);
2174                 //TODO add BO back to validate_list?
2175                 return ret;
2176         }
2177         amdgpu_amdkfd_remove_eviction_fence(gws_bo,
2178                         process_info->eviction_fence);
2179         amdgpu_bo_unreserve(gws_bo);
2180         amdgpu_sync_free(&kgd_mem->sync);
2181         amdgpu_bo_unref(&gws_bo);
2182         mutex_destroy(&kgd_mem->lock);
2183         kfree(mem);
2184         return 0;
2185 }