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 #include <linux/dma-buf.h>
  23 #include <linux/list.h>
  24 #include <linux/pagemap.h>
  25 #include <linux/sched/mm.h>
  26 #include <linux/sched/task.h>
  27
  28 #include "amdgpu_object.h"
  29 #include "amdgpu_gem.h"
  30 #include "amdgpu_vm.h"
  31 #include "amdgpu_amdkfd.h"
  32 #include "amdgpu_dma_buf.h"
  33 #include <uapi/linux/kfd_ioctl.h>
  34 #include "amdgpu_xgmi.h"
  35
  36 /* Userptr restore delay, just long enough to allow consecutive VM
  37  * changes to accumulate
  38  */
  39 #define AMDGPU_USERPTR_RESTORE_DELAY_MS 1
  40
  41 /* Impose limit on how much memory KFD can use */
  42 static struct {
  43         uint64_t max_system_mem_limit;
  44         uint64_t max_ttm_mem_limit;
  45         int64_t system_mem_used;
  46         int64_t ttm_mem_used;
  47         spinlock_t mem_limit_lock;
  48 } kfd_mem_limit;
  49
  50 static const char * const domain_bit_to_string[] = {
  51                 "CPU",
  52                 "GTT",
  53                 "VRAM",
  54                 "GDS",
  55                 "GWS",
  56                 "OA"
  57 };
  58
  59 #define domain_string(domain) domain_bit_to_string[ffs(domain)-1]
  60
  61 static void amdgpu_amdkfd_restore_userptr_worker(struct work_struct *work);
  62
  63
  64 static inline struct amdgpu_device *get_amdgpu_device(struct kgd_dev *kgd)
  65 {
  66         return (struct amdgpu_device *)kgd;
  67 }
  68
  69 static bool kfd_mem_is_attached(struct amdgpu_vm *avm,
  70                 struct kgd_mem *mem)
  71 {
  72         struct kfd_mem_attachment *entry;
  73
  74         list_for_each_entry(entry, &mem->attachments, list)
  75                 if (entry->bo_va->base.vm == avm)
  76                         return true;
  77
  78         return false;
  79 }
  80
  81 /* Set memory usage limits. Current, limits are
  82  *  System (TTM + userptr) memory - 15/16th System RAM
  83  *  TTM memory - 3/8th System RAM
  84  */
  85 void amdgpu_amdkfd_gpuvm_init_mem_limits(void)
  86 {
  87         struct sysinfo si;
  88         uint64_t mem;
  89
  90         si_meminfo(&si);
  91         mem = si.freeram - si.freehigh;
  92         mem *= si.mem_unit;
  93
  94         spin_lock_init(&kfd_mem_limit.mem_limit_lock);
  95         kfd_mem_limit.max_system_mem_limit = mem - (mem >> 4);
  96         kfd_mem_limit.max_ttm_mem_limit = (mem >> 1) - (mem >> 3);
  97         pr_debug("Kernel memory limit %lluM, TTM limit %lluM\n",
  98                 (kfd_mem_limit.max_system_mem_limit >> 20),
  99                 (kfd_mem_limit.max_ttm_mem_limit >> 20));
 100 }
 101
 102 void amdgpu_amdkfd_reserve_system_mem(uint64_t size)
 103 {
 104         kfd_mem_limit.system_mem_used += size;
 105 }
 106
 107 /* Estimate page table size needed to represent a given memory size
 108  *
 109  * With 4KB pages, we need one 8 byte PTE for each 4KB of memory
 110  * (factor 512, >> 9). With 2MB pages, we need one 8 byte PTE for 2MB
 111  * of memory (factor 256K, >> 18). ROCm user mode tries to optimize
 112  * for 2MB pages for TLB efficiency. However, small allocations and
 113  * fragmented system memory still need some 4KB pages. We choose a
 114  * compromise that should work in most cases without reserving too
 115  * much memory for page tables unnecessarily (factor 16K, >> 14).
 116  */
 117 #define ESTIMATE_PT_SIZE(mem_size) ((mem_size) >> 14)
 118
 119 static size_t amdgpu_amdkfd_acc_size(uint64_t size)
 120 {
 121         size >>= PAGE_SHIFT;
 122         size *= sizeof(dma_addr_t) + sizeof(void *);
 123
 124         return __roundup_pow_of_two(sizeof(struct amdgpu_bo)) +
 125                 __roundup_pow_of_two(sizeof(struct ttm_tt)) +
 126                 PAGE_ALIGN(size);
 127 }
 128
 129 static int amdgpu_amdkfd_reserve_mem_limit(struct amdgpu_device *adev,
 130                 uint64_t size, u32 domain, bool sg)
 131 {
 132         uint64_t reserved_for_pt =
 133                 ESTIMATE_PT_SIZE(amdgpu_amdkfd_total_mem_size);
 134         size_t acc_size, system_mem_needed, ttm_mem_needed, vram_needed;
 135         int ret = 0;
 136
 137         acc_size = amdgpu_amdkfd_acc_size(size);
 138
 139         vram_needed = 0;
 140         if (domain == AMDGPU_GEM_DOMAIN_GTT) {
 141                 /* TTM GTT memory */
 142                 system_mem_needed = acc_size + size;
 143                 ttm_mem_needed = acc_size + size;
 144         } else if (domain == AMDGPU_GEM_DOMAIN_CPU && !sg) {
 145                 /* Userptr */
 146                 system_mem_needed = acc_size + size;
 147                 ttm_mem_needed = acc_size;
 148         } else {
 149                 /* VRAM and SG */
 150                 system_mem_needed = acc_size;
 151                 ttm_mem_needed = acc_size;
 152                 if (domain == AMDGPU_GEM_DOMAIN_VRAM)
 153                         vram_needed = size;
 154         }
 155
 156         spin_lock(&kfd_mem_limit.mem_limit_lock);
 157
 158         if (kfd_mem_limit.system_mem_used + system_mem_needed >
 159             kfd_mem_limit.max_system_mem_limit)
 160                 pr_debug("Set no_system_mem_limit=1 if using shared memory\n");
 161
 162         if ((kfd_mem_limit.system_mem_used + system_mem_needed >
 163              kfd_mem_limit.max_system_mem_limit && !no_system_mem_limit) ||
 164             (kfd_mem_limit.ttm_mem_used + ttm_mem_needed >
 165              kfd_mem_limit.max_ttm_mem_limit) ||
 166             (adev->kfd.vram_used + vram_needed >
 167              adev->gmc.real_vram_size - reserved_for_pt)) {
 168                 ret = -ENOMEM;
 169         } else {
 170                 kfd_mem_limit.system_mem_used += system_mem_needed;
 171                 kfd_mem_limit.ttm_mem_used += ttm_mem_needed;
 172                 adev->kfd.vram_used += vram_needed;
 173         }
 174
 175         spin_unlock(&kfd_mem_limit.mem_limit_lock);
 176         return ret;
 177 }
 178
 179 static void unreserve_mem_limit(struct amdgpu_device *adev,
 180                 uint64_t size, u32 domain, bool sg)
 181 {
 182         size_t acc_size;
 183
 184         acc_size = amdgpu_amdkfd_acc_size(size);
 185
 186         spin_lock(&kfd_mem_limit.mem_limit_lock);
 187         if (domain == AMDGPU_GEM_DOMAIN_GTT) {
 188                 kfd_mem_limit.system_mem_used -= (acc_size + size);
 189                 kfd_mem_limit.ttm_mem_used -= (acc_size + size);
 190         } else if (domain == AMDGPU_GEM_DOMAIN_CPU && !sg) {
 191                 kfd_mem_limit.system_mem_used -= (acc_size + size);
 192                 kfd_mem_limit.ttm_mem_used -= acc_size;
 193         } else {
 194                 kfd_mem_limit.system_mem_used -= acc_size;
 195                 kfd_mem_limit.ttm_mem_used -= acc_size;
 196                 if (domain == AMDGPU_GEM_DOMAIN_VRAM) {
 197                         adev->kfd.vram_used -= size;
 198                         WARN_ONCE(adev->kfd.vram_used < 0,
 199                                   "kfd VRAM memory accounting unbalanced");
 200                 }
 201         }
 202         WARN_ONCE(kfd_mem_limit.system_mem_used < 0,
 203                   "kfd system memory accounting unbalanced");
 204         WARN_ONCE(kfd_mem_limit.ttm_mem_used < 0,
 205                   "kfd TTM memory accounting unbalanced");
 206
 207         spin_unlock(&kfd_mem_limit.mem_limit_lock);
 208 }
 209
 210 void amdgpu_amdkfd_unreserve_memory_limit(struct amdgpu_bo *bo)
 211 {
 212         struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
 213         u32 domain = bo->preferred_domains;
 214         bool sg = (bo->preferred_domains == AMDGPU_GEM_DOMAIN_CPU);
 215
 216         if (bo->flags & AMDGPU_AMDKFD_CREATE_USERPTR_BO) {
 217                 domain = AMDGPU_GEM_DOMAIN_CPU;
 218                 sg = false;
 219         }
 220
 221         unreserve_mem_limit(adev, amdgpu_bo_size(bo), domain, sg);
 222 }
 223
 224
 225 /* amdgpu_amdkfd_remove_eviction_fence - Removes eviction fence from BO's
 226  *  reservation object.
 227  *
 228  * @bo: [IN] Remove eviction fence(s) from this BO
 229  * @ef: [IN] This eviction fence is removed if it
 230  *  is present in the shared list.
 231  *
 232  * NOTE: Must be called with BO reserved i.e. bo->tbo.resv->lock held.
 233  */
 234 static int amdgpu_amdkfd_remove_eviction_fence(struct amdgpu_bo *bo,
 235                                         struct amdgpu_amdkfd_fence *ef)
 236 {
 237         struct dma_resv *resv = bo->tbo.base.resv;
 238         struct dma_resv_list *old, *new;
 239         unsigned int i, j, k;
 240
 241         if (!ef)
 242                 return -EINVAL;
 243
 244         old = dma_resv_shared_list(resv);
 245         if (!old)
 246                 return 0;
 247
 248         new = kmalloc(struct_size(new, shared, old->shared_max), GFP_KERNEL);
 249         if (!new)
 250                 return -ENOMEM;
 251
 252         /* Go through all the shared fences in the resevation object and sort
 253          * the interesting ones to the end of the list.
 254          */
 255         for (i = 0, j = old->shared_count, k = 0; i < old->shared_count; ++i) {
 256                 struct dma_fence *f;
 257
 258                 f = rcu_dereference_protected(old->shared[i],
 259                                               dma_resv_held(resv));
 260
 261                 if (f->context == ef->base.context)
 262                         RCU_INIT_POINTER(new->shared[--j], f);
 263                 else
 264                         RCU_INIT_POINTER(new->shared[k++], f);
 265         }
 266         new->shared_max = old->shared_max;
 267         new->shared_count = k;
 268
 269         /* Install the new fence list, seqcount provides the barriers */
 270         write_seqcount_begin(&resv->seq);
 271         RCU_INIT_POINTER(resv->fence, new);
 272         write_seqcount_end(&resv->seq);
 273
 274         /* Drop the references to the removed fences or move them to ef_list */
 275         for (i = j; i < old->shared_count; ++i) {
 276                 struct dma_fence *f;
 277
 278                 f = rcu_dereference_protected(new->shared[i],
 279                                               dma_resv_held(resv));
 280                 dma_fence_put(f);
 281         }
 282         kfree_rcu(old, rcu);
 283
 284         return 0;
 285 }
 286
 287 int amdgpu_amdkfd_remove_fence_on_pt_pd_bos(struct amdgpu_bo *bo)
 288 {
 289         struct amdgpu_bo *root = bo;
 290         struct amdgpu_vm_bo_base *vm_bo;
 291         struct amdgpu_vm *vm;
 292         struct amdkfd_process_info *info;
 293         struct amdgpu_amdkfd_fence *ef;
 294         int ret;
 295
 296         /* we can always get vm_bo from root PD bo.*/
 297         while (root->parent)
 298                 root = root->parent;
 299
 300         vm_bo = root->vm_bo;
 301         if (!vm_bo)
 302                 return 0;
 303
 304         vm = vm_bo->vm;
 305         if (!vm)
 306                 return 0;
 307
 308         info = vm->process_info;
 309         if (!info || !info->eviction_fence)
 310                 return 0;
 311
 312         ef = container_of(dma_fence_get(&info->eviction_fence->base),
 313                         struct amdgpu_amdkfd_fence, base);
 314
 315         BUG_ON(!dma_resv_trylock(bo->tbo.base.resv));
 316         ret = amdgpu_amdkfd_remove_eviction_fence(bo, ef);
 317         dma_resv_unlock(bo->tbo.base.resv);
 318
 319         dma_fence_put(&ef->base);
 320         return ret;
 321 }
 322
 323 static int amdgpu_amdkfd_bo_validate(struct amdgpu_bo *bo, uint32_t domain,
 324                                      bool wait)
 325 {
 326         struct ttm_operation_ctx ctx = { false, false };
 327         int ret;
 328
 329         if (WARN(amdgpu_ttm_tt_get_usermm(bo->tbo.ttm),
 330                  "Called with userptr BO"))
 331                 return -EINVAL;
 332
 333         amdgpu_bo_placement_from_domain(bo, domain);
 334
 335         ret = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
 336         if (ret)
 337                 goto validate_fail;
 338         if (wait)
 339                 amdgpu_bo_sync_wait(bo, AMDGPU_FENCE_OWNER_KFD, false);
 340
 341 validate_fail:
 342         return ret;
 343 }
 344
 345 static int amdgpu_amdkfd_validate_vm_bo(void *_unused, struct amdgpu_bo *bo)
 346 {
 347         return amdgpu_amdkfd_bo_validate(bo, bo->allowed_domains, false);
 348 }
 349
 350 /* vm_validate_pt_pd_bos - Validate page table and directory BOs
 351  *
 352  * Page directories are not updated here because huge page handling
 353  * during page table updates can invalidate page directory entries
 354  * again. Page directories are only updated after updating page
 355  * tables.
 356  */
 357 static int vm_validate_pt_pd_bos(struct amdgpu_vm *vm)
 358 {
 359         struct amdgpu_bo *pd = vm->root.bo;
 360         struct amdgpu_device *adev = amdgpu_ttm_adev(pd->tbo.bdev);
 361         int ret;
 362
 363         ret = amdgpu_vm_validate_pt_bos(adev, vm, amdgpu_amdkfd_validate_vm_bo, NULL);
 364         if (ret) {
 365                 pr_err("failed to validate PT BOs\n");
 366                 return ret;
 367         }
 368
 369         ret = amdgpu_amdkfd_validate_vm_bo(NULL, pd);
 370         if (ret) {
 371                 pr_err("failed to validate PD\n");
 372                 return ret;
 373         }
 374
 375         vm->pd_phys_addr = amdgpu_gmc_pd_addr(vm->root.bo);
 376
 377         if (vm->use_cpu_for_update) {
 378                 ret = amdgpu_bo_kmap(pd, NULL);
 379                 if (ret) {
 380                         pr_err("failed to kmap PD, ret=%d\n", ret);
 381                         return ret;
 382                 }
 383         }
 384
 385         return 0;
 386 }
 387
 388 static int vm_update_pds(struct amdgpu_vm *vm, struct amdgpu_sync *sync)
 389 {
 390         struct amdgpu_bo *pd = vm->root.bo;
 391         struct amdgpu_device *adev = amdgpu_ttm_adev(pd->tbo.bdev);
 392         int ret;
 393
 394         ret = amdgpu_vm_update_pdes(adev, vm, false);
 395         if (ret)
 396                 return ret;
 397
 398         return amdgpu_sync_fence(sync, vm->last_update);
 399 }
 400
 401 static uint64_t get_pte_flags(struct amdgpu_device *adev, struct kgd_mem *mem)
 402 {
 403         struct amdgpu_device *bo_adev = amdgpu_ttm_adev(mem->bo->tbo.bdev);
 404         bool coherent = mem->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_COHERENT;
 405         bool uncached = mem->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_UNCACHED;
 406         uint32_t mapping_flags;
 407         uint64_t pte_flags;
 408         bool snoop = false;
 409
 410         mapping_flags = AMDGPU_VM_PAGE_READABLE;
 411         if (mem->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_WRITABLE)
 412                 mapping_flags |= AMDGPU_VM_PAGE_WRITEABLE;
 413         if (mem->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_EXECUTABLE)
 414                 mapping_flags |= AMDGPU_VM_PAGE_EXECUTABLE;
 415
 416         switch (adev->asic_type) {
 417         case CHIP_ARCTURUS:
 418                 if (mem->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_VRAM) {
 419                         if (bo_adev == adev)
 420                                 mapping_flags |= coherent ?
 421                                         AMDGPU_VM_MTYPE_CC : AMDGPU_VM_MTYPE_RW;
 422                         else
 423                                 mapping_flags |= coherent ?
 424                                         AMDGPU_VM_MTYPE_UC : AMDGPU_VM_MTYPE_NC;
 425                 } else {
 426                         mapping_flags |= coherent ?
 427                                 AMDGPU_VM_MTYPE_UC : AMDGPU_VM_MTYPE_NC;
 428                 }
 429                 break;
 430         case CHIP_ALDEBARAN:
 431                 if (coherent && uncached) {
 432                         if (adev->gmc.xgmi.connected_to_cpu ||
 433                                 !(mem->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_VRAM))
 434                                 snoop = true;
 435                         mapping_flags |= AMDGPU_VM_MTYPE_UC;
 436                 } else if (mem->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_VRAM) {
 437                         if (bo_adev == adev) {
 438                                 mapping_flags |= coherent ?
 439                                         AMDGPU_VM_MTYPE_CC : AMDGPU_VM_MTYPE_RW;
 440                                 if (adev->gmc.xgmi.connected_to_cpu)
 441                                         snoop = true;
 442                         } else {
 443                                 mapping_flags |= coherent ?
 444                                         AMDGPU_VM_MTYPE_UC : AMDGPU_VM_MTYPE_NC;
 445                                 if (amdgpu_xgmi_same_hive(adev, bo_adev))
 446                                         snoop = true;
 447                         }
 448                 } else {
 449                         snoop = true;
 450                         mapping_flags |= coherent ?
 451                                 AMDGPU_VM_MTYPE_UC : AMDGPU_VM_MTYPE_NC;
 452                 }
 453                 break;
 454         default:
 455                 mapping_flags |= coherent ?
 456                         AMDGPU_VM_MTYPE_UC : AMDGPU_VM_MTYPE_NC;
 457         }
 458
 459         pte_flags = amdgpu_gem_va_map_flags(adev, mapping_flags);
 460         pte_flags |= snoop ? AMDGPU_PTE_SNOOPED : 0;
 461
 462         return pte_flags;
 463 }
 464
 465 static int
 466 kfd_mem_dmamap_userptr(struct kgd_mem *mem,
 467                        struct kfd_mem_attachment *attachment)
 468 {
 469         enum dma_data_direction direction =
 470                 mem->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_WRITABLE ?
 471                 DMA_BIDIRECTIONAL : DMA_TO_DEVICE;
 472         struct ttm_operation_ctx ctx = {.interruptible = true};
 473         struct amdgpu_bo *bo = attachment->bo_va->base.bo;
 474         struct amdgpu_device *adev = attachment->adev;
 475         struct ttm_tt *src_ttm = mem->bo->tbo.ttm;
 476         struct ttm_tt *ttm = bo->tbo.ttm;
 477         int ret;
 478
 479         ttm->sg = kmalloc(sizeof(*ttm->sg), GFP_KERNEL);
 480         if (unlikely(!ttm->sg))
 481                 return -ENOMEM;
 482
 483         if (WARN_ON(ttm->num_pages != src_ttm->num_pages))
 484                 return -EINVAL;
 485
 486         /* Same sequence as in amdgpu_ttm_tt_pin_userptr */
 487         ret = sg_alloc_table_from_pages(ttm->sg, src_ttm->pages,
 488                                         ttm->num_pages, 0,
 489                                         (u64)ttm->num_pages << PAGE_SHIFT,
 490                                         GFP_KERNEL);
 491         if (unlikely(ret))
 492                 goto free_sg;
 493
 494         ret = dma_map_sgtable(adev->dev, ttm->sg, direction, 0);
 495         if (unlikely(ret))
 496                 goto release_sg;
 497
 498         drm_prime_sg_to_dma_addr_array(ttm->sg, ttm->dma_address,
 499                                        ttm->num_pages);
 500
 501         amdgpu_bo_placement_from_domain(bo, AMDGPU_GEM_DOMAIN_GTT);
 502         ret = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
 503         if (ret)
 504                 goto unmap_sg;
 505
 506         return 0;
 507
 508 unmap_sg:
 509         dma_unmap_sgtable(adev->dev, ttm->sg, direction, 0);
 510 release_sg:
 511         pr_err("DMA map userptr failed: %d\n", ret);
 512         sg_free_table(ttm->sg);
 513 free_sg:
 514         kfree(ttm->sg);
 515         ttm->sg = NULL;
 516         return ret;
 517 }
 518
 519 static int
 520 kfd_mem_dmamap_dmabuf(struct kfd_mem_attachment *attachment)
 521 {
 522         struct ttm_operation_ctx ctx = {.interruptible = true};
 523         struct amdgpu_bo *bo = attachment->bo_va->base.bo;
 524
 525         amdgpu_bo_placement_from_domain(bo, AMDGPU_GEM_DOMAIN_GTT);
 526         return ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
 527 }
 528
 529 static int
 530 kfd_mem_dmamap_attachment(struct kgd_mem *mem,
 531                           struct kfd_mem_attachment *attachment)
 532 {
 533         switch (attachment->type) {
 534         case KFD_MEM_ATT_SHARED:
 535                 return 0;
 536         case KFD_MEM_ATT_USERPTR:
 537                 return kfd_mem_dmamap_userptr(mem, attachment);
 538         case KFD_MEM_ATT_DMABUF:
 539                 return kfd_mem_dmamap_dmabuf(attachment);
 540         default:
 541                 WARN_ON_ONCE(1);
 542         }
 543         return -EINVAL;
 544 }
 545
 546 static void
 547 kfd_mem_dmaunmap_userptr(struct kgd_mem *mem,
 548                          struct kfd_mem_attachment *attachment)
 549 {
 550         enum dma_data_direction direction =
 551                 mem->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_WRITABLE ?
 552                 DMA_BIDIRECTIONAL : DMA_TO_DEVICE;
 553         struct ttm_operation_ctx ctx = {.interruptible = false};
 554         struct amdgpu_bo *bo = attachment->bo_va->base.bo;
 555         struct amdgpu_device *adev = attachment->adev;
 556         struct ttm_tt *ttm = bo->tbo.ttm;
 557
 558         if (unlikely(!ttm->sg))
 559                 return;
 560
 561         amdgpu_bo_placement_from_domain(bo, AMDGPU_GEM_DOMAIN_CPU);
 562         ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
 563
 564         dma_unmap_sgtable(adev->dev, ttm->sg, direction, 0);
 565         sg_free_table(ttm->sg);
 566         ttm->sg = NULL;
 567 }
 568
 569 static void
 570 kfd_mem_dmaunmap_dmabuf(struct kfd_mem_attachment *attachment)
 571 {
 572         struct ttm_operation_ctx ctx = {.interruptible = true};
 573         struct amdgpu_bo *bo = attachment->bo_va->base.bo;
 574
 575         amdgpu_bo_placement_from_domain(bo, AMDGPU_GEM_DOMAIN_CPU);
 576         ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
 577 }
 578
 579 static void
 580 kfd_mem_dmaunmap_attachment(struct kgd_mem *mem,
 581                             struct kfd_mem_attachment *attachment)
 582 {
 583         switch (attachment->type) {
 584         case KFD_MEM_ATT_SHARED:
 585                 break;
 586         case KFD_MEM_ATT_USERPTR:
 587                 kfd_mem_dmaunmap_userptr(mem, attachment);
 588                 break;
 589         case KFD_MEM_ATT_DMABUF:
 590                 kfd_mem_dmaunmap_dmabuf(attachment);
 591                 break;
 592         default:
 593                 WARN_ON_ONCE(1);
 594         }
 595 }
 596
 597 static int
 598 kfd_mem_attach_userptr(struct amdgpu_device *adev, struct kgd_mem *mem,
 599                        struct amdgpu_bo **bo)
 600 {
 601         unsigned long bo_size = mem->bo->tbo.base.size;
 602         struct drm_gem_object *gobj;
 603         int ret;
 604
 605         ret = amdgpu_bo_reserve(mem->bo, false);
 606         if (ret)
 607                 return ret;
 608
 609         ret = amdgpu_gem_object_create(adev, bo_size, 1,
 610                                        AMDGPU_GEM_DOMAIN_CPU,
 611                                        AMDGPU_GEM_CREATE_PREEMPTIBLE,
 612                                        ttm_bo_type_sg, mem->bo->tbo.base.resv,
 613                                        &gobj);
 614         amdgpu_bo_unreserve(mem->bo);
 615         if (ret)
 616                 return ret;
 617
 618         *bo = gem_to_amdgpu_bo(gobj);
 619         (*bo)->parent = amdgpu_bo_ref(mem->bo);
 620
 621         return 0;
 622 }
 623
 624 static int
 625 kfd_mem_attach_dmabuf(struct amdgpu_device *adev, struct kgd_mem *mem,
 626                       struct amdgpu_bo **bo)
 627 {
 628         struct drm_gem_object *gobj;
 629         int ret;
 630
 631         if (!mem->dmabuf) {
 632                 mem->dmabuf = amdgpu_gem_prime_export(&mem->bo->tbo.base,
 633                         mem->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_WRITABLE ?
 634                                 DRM_RDWR : 0);
 635                 if (IS_ERR(mem->dmabuf)) {
 636                         ret = PTR_ERR(mem->dmabuf);
 637                         mem->dmabuf = NULL;
 638                         return ret;
 639                 }
 640         }
 641
 642         gobj = amdgpu_gem_prime_import(adev_to_drm(adev), mem->dmabuf);
 643         if (IS_ERR(gobj))
 644                 return PTR_ERR(gobj);
 645
 646         /* Import takes an extra reference on the dmabuf. Drop it now to
 647          * avoid leaking it. We only need the one reference in
 648          * kgd_mem->dmabuf.
 649          */
 650         dma_buf_put(mem->dmabuf);
 651
 652         *bo = gem_to_amdgpu_bo(gobj);
 653         (*bo)->flags |= AMDGPU_GEM_CREATE_PREEMPTIBLE;
 654         (*bo)->parent = amdgpu_bo_ref(mem->bo);
 655
 656         return 0;
 657 }
 658
 659 /* kfd_mem_attach - Add a BO to a VM
 660  *
 661  * Everything that needs to bo done only once when a BO is first added
 662  * to a VM. It can later be mapped and unmapped many times without
 663  * repeating these steps.
 664  *
 665  * 0. Create BO for DMA mapping, if needed
 666  * 1. Allocate and initialize BO VA entry data structure
 667  * 2. Add BO to the VM
 668  * 3. Determine ASIC-specific PTE flags
 669  * 4. Alloc page tables and directories if needed
 670  * 4a.  Validate new page tables and directories
 671  */
 672 static int kfd_mem_attach(struct amdgpu_device *adev, struct kgd_mem *mem,
 673                 struct amdgpu_vm *vm, bool is_aql)
 674 {
 675         struct amdgpu_device *bo_adev = amdgpu_ttm_adev(mem->bo->tbo.bdev);
 676         unsigned long bo_size = mem->bo->tbo.base.size;
 677         uint64_t va = mem->va;
 678         struct kfd_mem_attachment *attachment[2] = {NULL, NULL};
 679         struct amdgpu_bo *bo[2] = {NULL, NULL};
 680         int i, ret;
 681
 682         if (!va) {
 683                 pr_err("Invalid VA when adding BO to VM\n");
 684                 return -EINVAL;
 685         }
 686
 687         for (i = 0; i <= is_aql; i++) {
 688                 attachment[i] = kzalloc(sizeof(*attachment[i]), GFP_KERNEL);
 689                 if (unlikely(!attachment[i])) {
 690                         ret = -ENOMEM;
 691                         goto unwind;
 692                 }
 693
 694                 pr_debug("\t add VA 0x%llx - 0x%llx to vm %p\n", va,
 695                          va + bo_size, vm);
 696
 697                 if (adev == bo_adev || (mem->domain == AMDGPU_GEM_DOMAIN_VRAM &&
 698                                         amdgpu_xgmi_same_hive(adev, bo_adev))) {
 699                         /* Mappings on the local GPU and VRAM mappings in the
 700                          * local hive share the original BO
 701                          */
 702                         attachment[i]->type = KFD_MEM_ATT_SHARED;
 703                         bo[i] = mem->bo;
 704                         drm_gem_object_get(&bo[i]->tbo.base);
 705                 } else if (i > 0) {
 706                         /* Multiple mappings on the same GPU share the BO */
 707                         attachment[i]->type = KFD_MEM_ATT_SHARED;
 708                         bo[i] = bo[0];
 709                         drm_gem_object_get(&bo[i]->tbo.base);
 710                 } else if (amdgpu_ttm_tt_get_usermm(mem->bo->tbo.ttm)) {
 711                         /* Create an SG BO to DMA-map userptrs on other GPUs */
 712                         attachment[i]->type = KFD_MEM_ATT_USERPTR;
 713                         ret = kfd_mem_attach_userptr(adev, mem, &bo[i]);
 714                         if (ret)
 715                                 goto unwind;
 716                 } else if (mem->domain == AMDGPU_GEM_DOMAIN_GTT &&
 717                            mem->bo->tbo.type != ttm_bo_type_sg) {
 718                         /* GTT BOs use DMA-mapping ability of dynamic-attach
 719                          * DMA bufs. TODO: The same should work for VRAM on
 720                          * large-BAR GPUs.
 721                          */
 722                         attachment[i]->type = KFD_MEM_ATT_DMABUF;
 723                         ret = kfd_mem_attach_dmabuf(adev, mem, &bo[i]);
 724                         if (ret)
 725                                 goto unwind;
 726                 } else {
 727                         /* FIXME: Need to DMA-map other BO types:
 728                          * large-BAR VRAM, doorbells, MMIO remap
 729                          */
 730                         attachment[i]->type = KFD_MEM_ATT_SHARED;
 731                         bo[i] = mem->bo;
 732                         drm_gem_object_get(&bo[i]->tbo.base);
 733                 }
 734
 735                 /* Add BO to VM internal data structures */
 736                 attachment[i]->bo_va = amdgpu_vm_bo_add(adev, vm, bo[i]);
 737                 if (unlikely(!attachment[i]->bo_va)) {
 738                         ret = -ENOMEM;
 739                         pr_err("Failed to add BO object to VM. ret == %d\n",
 740                                ret);
 741                         goto unwind;
 742                 }
 743
 744                 attachment[i]->va = va;
 745                 attachment[i]->pte_flags = get_pte_flags(adev, mem);
 746                 attachment[i]->adev = adev;
 747                 list_add(&attachment[i]->list, &mem->attachments);
 748
 749                 va += bo_size;
 750         }
 751
 752         return 0;
 753
 754 unwind:
 755         for (; i >= 0; i--) {
 756                 if (!attachment[i])
 757                         continue;
 758                 if (attachment[i]->bo_va) {
 759                         amdgpu_vm_bo_rmv(adev, attachment[i]->bo_va);
 760                         list_del(&attachment[i]->list);
 761                 }
 762                 if (bo[i])
 763                         drm_gem_object_put(&bo[i]->tbo.base);
 764                 kfree(attachment[i]);
 765         }
 766         return ret;
 767 }
 768
 769 static void kfd_mem_detach(struct kfd_mem_attachment *attachment)
 770 {
 771         struct amdgpu_bo *bo = attachment->bo_va->base.bo;
 772
 773         pr_debug("\t remove VA 0x%llx in entry %p\n",
 774                         attachment->va, attachment);
 775         amdgpu_vm_bo_rmv(attachment->adev, attachment->bo_va);
 776         drm_gem_object_put(&bo->tbo.base);
 777         list_del(&attachment->list);
 778         kfree(attachment);
 779 }
 780
 781 static void add_kgd_mem_to_kfd_bo_list(struct kgd_mem *mem,
 782                                 struct amdkfd_process_info *process_info,
 783                                 bool userptr)
 784 {
 785         struct ttm_validate_buffer *entry = &mem->validate_list;
 786         struct amdgpu_bo *bo = mem->bo;
 787
 788         INIT_LIST_HEAD(&entry->head);
 789         entry->num_shared = 1;
 790         entry->bo = &bo->tbo;
 791         mutex_lock(&process_info->lock);
 792         if (userptr)
 793                 list_add_tail(&entry->head, &process_info->userptr_valid_list);
 794         else
 795                 list_add_tail(&entry->head, &process_info->kfd_bo_list);
 796         mutex_unlock(&process_info->lock);
 797 }
 798
 799 static void remove_kgd_mem_from_kfd_bo_list(struct kgd_mem *mem,
 800                 struct amdkfd_process_info *process_info)
 801 {
 802         struct ttm_validate_buffer *bo_list_entry;
 803
 804         bo_list_entry = &mem->validate_list;
 805         mutex_lock(&process_info->lock);
 806         list_del(&bo_list_entry->head);
 807         mutex_unlock(&process_info->lock);
 808 }
 809
 810 /* Initializes user pages. It registers the MMU notifier and validates
 811  * the userptr BO in the GTT domain.
 812  *
 813  * The BO must already be on the userptr_valid_list. Otherwise an
 814  * eviction and restore may happen that leaves the new BO unmapped
 815  * with the user mode queues running.
 816  *
 817  * Takes the process_info->lock to protect against concurrent restore
 818  * workers.
 819  *
 820  * Returns 0 for success, negative errno for errors.
 821  */
 822 static int init_user_pages(struct kgd_mem *mem, uint64_t user_addr)
 823 {
 824         struct amdkfd_process_info *process_info = mem->process_info;
 825         struct amdgpu_bo *bo = mem->bo;
 826         struct ttm_operation_ctx ctx = { true, false };
 827         int ret = 0;
 828
 829         mutex_lock(&process_info->lock);
 830
 831         ret = amdgpu_ttm_tt_set_userptr(&bo->tbo, user_addr, 0);
 832         if (ret) {
 833                 pr_err("%s: Failed to set userptr: %d\n", __func__, ret);
 834                 goto out;
 835         }
 836
 837         ret = amdgpu_mn_register(bo, user_addr);
 838         if (ret) {
 839                 pr_err("%s: Failed to register MMU notifier: %d\n",
 840                        __func__, ret);
 841                 goto out;
 842         }
 843
 844         ret = amdgpu_ttm_tt_get_user_pages(bo, bo->tbo.ttm->pages);
 845         if (ret) {
 846                 pr_err("%s: Failed to get user pages: %d\n", __func__, ret);
 847                 goto unregister_out;
 848         }
 849
 850         ret = amdgpu_bo_reserve(bo, true);
 851         if (ret) {
 852                 pr_err("%s: Failed to reserve BO\n", __func__);
 853                 goto release_out;
 854         }
 855         amdgpu_bo_placement_from_domain(bo, mem->domain);
 856         ret = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
 857         if (ret)
 858                 pr_err("%s: failed to validate BO\n", __func__);
 859         amdgpu_bo_unreserve(bo);
 860
 861 release_out:
 862         amdgpu_ttm_tt_get_user_pages_done(bo->tbo.ttm);
 863 unregister_out:
 864         if (ret)
 865                 amdgpu_mn_unregister(bo);
 866 out:
 867         mutex_unlock(&process_info->lock);
 868         return ret;
 869 }
 870
 871 /* Reserving a BO and its page table BOs must happen atomically to
 872  * avoid deadlocks. Some operations update multiple VMs at once. Track
 873  * all the reservation info in a context structure. Optionally a sync
 874  * object can track VM updates.
 875  */
 876 struct bo_vm_reservation_context {
 877         struct amdgpu_bo_list_entry kfd_bo; /* BO list entry for the KFD BO */
 878         unsigned int n_vms;                 /* Number of VMs reserved       */
 879         struct amdgpu_bo_list_entry *vm_pd; /* Array of VM BO list entries  */
 880         struct ww_acquire_ctx ticket;       /* Reservation ticket           */
 881         struct list_head list, duplicates;  /* BO lists                     */
 882         struct amdgpu_sync *sync;           /* Pointer to sync object       */
 883         bool reserved;                      /* Whether BOs are reserved     */
 884 };
 885
 886 enum bo_vm_match {
 887         BO_VM_NOT_MAPPED = 0,   /* Match VMs where a BO is not mapped */
 888         BO_VM_MAPPED,           /* Match VMs where a BO is mapped     */
 889         BO_VM_ALL,              /* Match all VMs a BO was added to    */
 890 };
 891
 892 /**
 893  * reserve_bo_and_vm - reserve a BO and a VM unconditionally.
 894  * @mem: KFD BO structure.
 895  * @vm: the VM to reserve.
 896  * @ctx: the struct that will be used in unreserve_bo_and_vms().
 897  */
 898 static int reserve_bo_and_vm(struct kgd_mem *mem,
 899                               struct amdgpu_vm *vm,
 900                               struct bo_vm_reservation_context *ctx)
 901 {
 902         struct amdgpu_bo *bo = mem->bo;
 903         int ret;
 904
 905         WARN_ON(!vm);
 906
 907         ctx->reserved = false;
 908         ctx->n_vms = 1;
 909         ctx->sync = &mem->sync;
 910
 911         INIT_LIST_HEAD(&ctx->list);
 912         INIT_LIST_HEAD(&ctx->duplicates);
 913
 914         ctx->vm_pd = kcalloc(ctx->n_vms, sizeof(*ctx->vm_pd), GFP_KERNEL);
 915         if (!ctx->vm_pd)
 916                 return -ENOMEM;
 917
 918         ctx->kfd_bo.priority = 0;
 919         ctx->kfd_bo.tv.bo = &bo->tbo;
 920         ctx->kfd_bo.tv.num_shared = 1;
 921         list_add(&ctx->kfd_bo.tv.head, &ctx->list);
 922
 923         amdgpu_vm_get_pd_bo(vm, &ctx->list, &ctx->vm_pd[0]);
 924
 925         ret = ttm_eu_reserve_buffers(&ctx->ticket, &ctx->list,
 926                                      false, &ctx->duplicates);
 927         if (ret) {
 928                 pr_err("Failed to reserve buffers in ttm.\n");
 929                 kfree(ctx->vm_pd);
 930                 ctx->vm_pd = NULL;
 931                 return ret;
 932         }
 933
 934         ctx->reserved = true;
 935         return 0;
 936 }
 937
 938 /**
 939  * reserve_bo_and_cond_vms - reserve a BO and some VMs conditionally
 940  * @mem: KFD BO structure.
 941  * @vm: the VM to reserve. If NULL, then all VMs associated with the BO
 942  * is used. Otherwise, a single VM associated with the BO.
 943  * @map_type: the mapping status that will be used to filter the VMs.
 944  * @ctx: the struct that will be used in unreserve_bo_and_vms().
 945  *
 946  * Returns 0 for success, negative for failure.
 947  */
 948 static int reserve_bo_and_cond_vms(struct kgd_mem *mem,
 949                                 struct amdgpu_vm *vm, enum bo_vm_match map_type,
 950                                 struct bo_vm_reservation_context *ctx)
 951 {
 952         struct amdgpu_bo *bo = mem->bo;
 953         struct kfd_mem_attachment *entry;
 954         unsigned int i;
 955         int ret;
 956
 957         ctx->reserved = false;
 958         ctx->n_vms = 0;
 959         ctx->vm_pd = NULL;
 960         ctx->sync = &mem->sync;
 961
 962         INIT_LIST_HEAD(&ctx->list);
 963         INIT_LIST_HEAD(&ctx->duplicates);
 964
 965         list_for_each_entry(entry, &mem->attachments, list) {
 966                 if ((vm && vm != entry->bo_va->base.vm) ||
 967                         (entry->is_mapped != map_type
 968                         && map_type != BO_VM_ALL))
 969                         continue;
 970
 971                 ctx->n_vms++;
 972         }
 973
 974         if (ctx->n_vms != 0) {
 975                 ctx->vm_pd = kcalloc(ctx->n_vms, sizeof(*ctx->vm_pd),
 976                                      GFP_KERNEL);
 977                 if (!ctx->vm_pd)
 978                         return -ENOMEM;
 979         }
 980
 981         ctx->kfd_bo.priority = 0;
 982         ctx->kfd_bo.tv.bo = &bo->tbo;
 983         ctx->kfd_bo.tv.num_shared = 1;
 984         list_add(&ctx->kfd_bo.tv.head, &ctx->list);
 985
 986         i = 0;
 987         list_for_each_entry(entry, &mem->attachments, list) {
 988                 if ((vm && vm != entry->bo_va->base.vm) ||
 989                         (entry->is_mapped != map_type
 990                         && map_type != BO_VM_ALL))
 991                         continue;
 992
 993                 amdgpu_vm_get_pd_bo(entry->bo_va->base.vm, &ctx->list,
 994                                 &ctx->vm_pd[i]);
 995                 i++;
 996         }
 997
 998         ret = ttm_eu_reserve_buffers(&ctx->ticket, &ctx->list,
 999                                      false, &ctx->duplicates);
1000         if (ret) {
1001                 pr_err("Failed to reserve buffers in ttm.\n");
1002                 kfree(ctx->vm_pd);
1003                 ctx->vm_pd = NULL;
1004                 return ret;
1005         }
1006
1007         ctx->reserved = true;
1008         return 0;
1009 }
1010
1011 /**
1012  * unreserve_bo_and_vms - Unreserve BO and VMs from a reservation context
1013  * @ctx: Reservation context to unreserve
1014  * @wait: Optionally wait for a sync object representing pending VM updates
1015  * @intr: Whether the wait is interruptible
1016  *
1017  * Also frees any resources allocated in
1018  * reserve_bo_and_(cond_)vm(s). Returns the status from
1019  * amdgpu_sync_wait.
1020  */
1021 static int unreserve_bo_and_vms(struct bo_vm_reservation_context *ctx,
1022                                  bool wait, bool intr)
1023 {
1024         int ret = 0;
1025
1026         if (wait)
1027                 ret = amdgpu_sync_wait(ctx->sync, intr);
1028
1029         if (ctx->reserved)
1030                 ttm_eu_backoff_reservation(&ctx->ticket, &ctx->list);
1031         kfree(ctx->vm_pd);
1032
1033         ctx->sync = NULL;
1034
1035         ctx->reserved = false;
1036         ctx->vm_pd = NULL;
1037
1038         return ret;
1039 }
1040
1041 static void unmap_bo_from_gpuvm(struct kgd_mem *mem,
1042                                 struct kfd_mem_attachment *entry,
1043                                 struct amdgpu_sync *sync)
1044 {
1045         struct amdgpu_bo_va *bo_va = entry->bo_va;
1046         struct amdgpu_device *adev = entry->adev;
1047         struct amdgpu_vm *vm = bo_va->base.vm;
1048
1049         amdgpu_vm_bo_unmap(adev, bo_va, entry->va);
1050
1051         amdgpu_vm_clear_freed(adev, vm, &bo_va->last_pt_update);
1052
1053         amdgpu_sync_fence(sync, bo_va->last_pt_update);
1054
1055         kfd_mem_dmaunmap_attachment(mem, entry);
1056 }
1057
1058 static int update_gpuvm_pte(struct kgd_mem *mem,
1059                             struct kfd_mem_attachment *entry,
1060                             struct amdgpu_sync *sync)
1061 {
1062         struct amdgpu_bo_va *bo_va = entry->bo_va;
1063         struct amdgpu_device *adev = entry->adev;
1064         int ret;
1065
1066         ret = kfd_mem_dmamap_attachment(mem, entry);
1067         if (ret)
1068                 return ret;
1069
1070         /* Update the page tables  */
1071         ret = amdgpu_vm_bo_update(adev, bo_va, false);
1072         if (ret) {
1073                 pr_err("amdgpu_vm_bo_update failed\n");
1074                 return ret;
1075         }
1076
1077         return amdgpu_sync_fence(sync, bo_va->last_pt_update);
1078 }
1079
1080 static int map_bo_to_gpuvm(struct kgd_mem *mem,
1081                            struct kfd_mem_attachment *entry,
1082                            struct amdgpu_sync *sync,
1083                            bool no_update_pte)
1084 {
1085         int ret;
1086
1087         /* Set virtual address for the allocation */
1088         ret = amdgpu_vm_bo_map(entry->adev, entry->bo_va, entry->va, 0,
1089                                amdgpu_bo_size(entry->bo_va->base.bo),
1090                                entry->pte_flags);
1091         if (ret) {
1092                 pr_err("Failed to map VA 0x%llx in vm. ret %d\n",
1093                                 entry->va, ret);
1094                 return ret;
1095         }
1096
1097         if (no_update_pte)
1098                 return 0;
1099
1100         ret = update_gpuvm_pte(mem, entry, sync);
1101         if (ret) {
1102                 pr_err("update_gpuvm_pte() failed\n");
1103                 goto update_gpuvm_pte_failed;
1104         }
1105
1106         return 0;
1107
1108 update_gpuvm_pte_failed:
1109         unmap_bo_from_gpuvm(mem, entry, sync);
1110         return ret;
1111 }
1112
1113 static struct sg_table *create_doorbell_sg(uint64_t addr, uint32_t size)
1114 {
1115         struct sg_table *sg = kmalloc(sizeof(*sg), GFP_KERNEL);
1116
1117         if (!sg)
1118                 return NULL;
1119         if (sg_alloc_table(sg, 1, GFP_KERNEL)) {
1120                 kfree(sg);
1121                 return NULL;
1122         }
1123         sg->sgl->dma_address = addr;
1124         sg->sgl->length = size;
1125 #ifdef CONFIG_NEED_SG_DMA_LENGTH
1126         sg->sgl->dma_length = size;
1127 #endif
1128         return sg;
1129 }
1130
1131 static int process_validate_vms(struct amdkfd_process_info *process_info)
1132 {
1133         struct amdgpu_vm *peer_vm;
1134         int ret;
1135
1136         list_for_each_entry(peer_vm, &process_info->vm_list_head,
1137                             vm_list_node) {
1138                 ret = vm_validate_pt_pd_bos(peer_vm);
1139                 if (ret)
1140                         return ret;
1141         }
1142
1143         return 0;
1144 }
1145
1146 static int process_sync_pds_resv(struct amdkfd_process_info *process_info,
1147                                  struct amdgpu_sync *sync)
1148 {
1149         struct amdgpu_vm *peer_vm;
1150         int ret;
1151
1152         list_for_each_entry(peer_vm, &process_info->vm_list_head,
1153                             vm_list_node) {
1154                 struct amdgpu_bo *pd = peer_vm->root.bo;
1155
1156                 ret = amdgpu_sync_resv(NULL, sync, pd->tbo.base.resv,
1157                                        AMDGPU_SYNC_NE_OWNER,
1158                                        AMDGPU_FENCE_OWNER_KFD);
1159                 if (ret)
1160                         return ret;
1161         }
1162
1163         return 0;
1164 }
1165
1166 static int process_update_pds(struct amdkfd_process_info *process_info,
1167                               struct amdgpu_sync *sync)
1168 {
1169         struct amdgpu_vm *peer_vm;
1170         int ret;
1171
1172         list_for_each_entry(peer_vm, &process_info->vm_list_head,
1173                             vm_list_node) {
1174                 ret = vm_update_pds(peer_vm, sync);
1175                 if (ret)
1176                         return ret;
1177         }
1178
1179         return 0;
1180 }
1181
1182 static int init_kfd_vm(struct amdgpu_vm *vm, void **process_info,
1183                        struct dma_fence **ef)
1184 {
1185         struct amdkfd_process_info *info = NULL;
1186         int ret;
1187
1188         if (!*process_info) {
1189                 info = kzalloc(sizeof(*info), GFP_KERNEL);
1190                 if (!info)
1191                         return -ENOMEM;
1192
1193                 mutex_init(&info->lock);
1194                 INIT_LIST_HEAD(&info->vm_list_head);
1195                 INIT_LIST_HEAD(&info->kfd_bo_list);
1196                 INIT_LIST_HEAD(&info->userptr_valid_list);
1197                 INIT_LIST_HEAD(&info->userptr_inval_list);
1198
1199                 info->eviction_fence =
1200                         amdgpu_amdkfd_fence_create(dma_fence_context_alloc(1),
1201                                                    current->mm,
1202                                                    NULL);
1203                 if (!info->eviction_fence) {
1204                         pr_err("Failed to create eviction fence\n");
1205                         ret = -ENOMEM;
1206                         goto create_evict_fence_fail;
1207                 }
1208
1209                 info->pid = get_task_pid(current->group_leader, PIDTYPE_PID);
1210                 atomic_set(&info->evicted_bos, 0);
1211                 INIT_DELAYED_WORK(&info->restore_userptr_work,
1212                                   amdgpu_amdkfd_restore_userptr_worker);
1213
1214                 *process_info = info;
1215                 *ef = dma_fence_get(&info->eviction_fence->base);
1216         }
1217
1218         vm->process_info = *process_info;
1219
1220         /* Validate page directory and attach eviction fence */
1221         ret = amdgpu_bo_reserve(vm->root.bo, true);
1222         if (ret)
1223                 goto reserve_pd_fail;
1224         ret = vm_validate_pt_pd_bos(vm);
1225         if (ret) {
1226                 pr_err("validate_pt_pd_bos() failed\n");
1227                 goto validate_pd_fail;
1228         }
1229         ret = amdgpu_bo_sync_wait(vm->root.bo,
1230                                   AMDGPU_FENCE_OWNER_KFD, false);
1231         if (ret)
1232                 goto wait_pd_fail;
1233         ret = dma_resv_reserve_shared(vm->root.bo->tbo.base.resv, 1);
1234         if (ret)
1235                 goto reserve_shared_fail;
1236         amdgpu_bo_fence(vm->root.bo,
1237                         &vm->process_info->eviction_fence->base, true);
1238         amdgpu_bo_unreserve(vm->root.bo);
1239
1240         /* Update process info */
1241         mutex_lock(&vm->process_info->lock);
1242         list_add_tail(&vm->vm_list_node,
1243                         &(vm->process_info->vm_list_head));
1244         vm->process_info->n_vms++;
1245         mutex_unlock(&vm->process_info->lock);
1246
1247         return 0;
1248
1249 reserve_shared_fail:
1250 wait_pd_fail:
1251 validate_pd_fail:
1252         amdgpu_bo_unreserve(vm->root.bo);
1253 reserve_pd_fail:
1254         vm->process_info = NULL;
1255         if (info) {
1256                 /* Two fence references: one in info and one in *ef */
1257                 dma_fence_put(&info->eviction_fence->base);
1258                 dma_fence_put(*ef);
1259                 *ef = NULL;
1260                 *process_info = NULL;
1261                 put_pid(info->pid);
1262 create_evict_fence_fail:
1263                 mutex_destroy(&info->lock);
1264                 kfree(info);
1265         }
1266         return ret;
1267 }
1268
1269 int amdgpu_amdkfd_gpuvm_acquire_process_vm(struct kgd_dev *kgd,
1270                                            struct file *filp, u32 pasid,
1271                                            void **process_info,
1272                                            struct dma_fence **ef)
1273 {
1274         struct amdgpu_device *adev = get_amdgpu_device(kgd);
1275         struct amdgpu_fpriv *drv_priv;
1276         struct amdgpu_vm *avm;
1277         int ret;
1278
1279         ret = amdgpu_file_to_fpriv(filp, &drv_priv);
1280         if (ret)
1281                 return ret;
1282         avm = &drv_priv->vm;
1283
1284         /* Already a compute VM? */
1285         if (avm->process_info)
1286                 return -EINVAL;
1287
1288         /* Convert VM into a compute VM */
1289         ret = amdgpu_vm_make_compute(adev, avm, pasid);
1290         if (ret)
1291                 return ret;
1292
1293         /* Initialize KFD part of the VM and process info */
1294         ret = init_kfd_vm(avm, process_info, ef);
1295         if (ret)
1296                 return ret;
1297
1298         amdgpu_vm_set_task_info(avm);
1299
1300         return 0;
1301 }
1302
1303 void amdgpu_amdkfd_gpuvm_destroy_cb(struct amdgpu_device *adev,
1304                                     struct amdgpu_vm *vm)
1305 {
1306         struct amdkfd_process_info *process_info = vm->process_info;
1307         struct amdgpu_bo *pd = vm->root.bo;
1308
1309         if (!process_info)
1310                 return;
1311
1312         /* Release eviction fence from PD */
1313         amdgpu_bo_reserve(pd, false);
1314         amdgpu_bo_fence(pd, NULL, false);
1315         amdgpu_bo_unreserve(pd);
1316
1317         /* Update process info */
1318         mutex_lock(&process_info->lock);
1319         process_info->n_vms--;
1320         list_del(&vm->vm_list_node);
1321         mutex_unlock(&process_info->lock);
1322
1323         vm->process_info = NULL;
1324
1325         /* Release per-process resources when last compute VM is destroyed */
1326         if (!process_info->n_vms) {
1327                 WARN_ON(!list_empty(&process_info->kfd_bo_list));
1328                 WARN_ON(!list_empty(&process_info->userptr_valid_list));
1329                 WARN_ON(!list_empty(&process_info->userptr_inval_list));
1330
1331                 dma_fence_put(&process_info->eviction_fence->base);
1332                 cancel_delayed_work_sync(&process_info->restore_userptr_work);
1333                 put_pid(process_info->pid);
1334                 mutex_destroy(&process_info->lock);
1335                 kfree(process_info);
1336         }
1337 }
1338
1339 void amdgpu_amdkfd_gpuvm_release_process_vm(struct kgd_dev *kgd, void *drm_priv)
1340 {
1341         struct amdgpu_device *adev = get_amdgpu_device(kgd);
1342         struct amdgpu_vm *avm;
1343
1344         if (WARN_ON(!kgd || !drm_priv))
1345                 return;
1346
1347         avm = drm_priv_to_vm(drm_priv);
1348
1349         pr_debug("Releasing process vm %p\n", avm);
1350
1351         /* The original pasid of amdgpu vm has already been
1352          * released during making a amdgpu vm to a compute vm
1353          * The current pasid is managed by kfd and will be
1354          * released on kfd process destroy. Set amdgpu pasid
1355          * to 0 to avoid duplicate release.
1356          */
1357         amdgpu_vm_release_compute(adev, avm);
1358 }
1359
1360 uint64_t amdgpu_amdkfd_gpuvm_get_process_page_dir(void *drm_priv)
1361 {
1362         struct amdgpu_vm *avm = drm_priv_to_vm(drm_priv);
1363         struct amdgpu_bo *pd = avm->root.bo;
1364         struct amdgpu_device *adev = amdgpu_ttm_adev(pd->tbo.bdev);
1365
1366         if (adev->asic_type < CHIP_VEGA10)
1367                 return avm->pd_phys_addr >> AMDGPU_GPU_PAGE_SHIFT;
1368         return avm->pd_phys_addr;
1369 }
1370
1371 int amdgpu_amdkfd_gpuvm_alloc_memory_of_gpu(
1372                 struct kgd_dev *kgd, uint64_t va, uint64_t size,
1373                 void *drm_priv, struct kgd_mem **mem,
1374                 uint64_t *offset, uint32_t flags)
1375 {
1376         struct amdgpu_device *adev = get_amdgpu_device(kgd);
1377         struct amdgpu_vm *avm = drm_priv_to_vm(drm_priv);
1378         enum ttm_bo_type bo_type = ttm_bo_type_device;
1379         struct sg_table *sg = NULL;
1380         uint64_t user_addr = 0;
1381         struct amdgpu_bo *bo;
1382         struct drm_gem_object *gobj;
1383         u32 domain, alloc_domain;
1384         u64 alloc_flags;
1385         int ret;
1386
1387         /*
1388          * Check on which domain to allocate BO
1389          */
1390         if (flags & KFD_IOC_ALLOC_MEM_FLAGS_VRAM) {
1391                 domain = alloc_domain = AMDGPU_GEM_DOMAIN_VRAM;
1392                 alloc_flags = AMDGPU_GEM_CREATE_VRAM_WIPE_ON_RELEASE;
1393                 alloc_flags |= (flags & KFD_IOC_ALLOC_MEM_FLAGS_PUBLIC) ?
1394                         AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED : 0;
1395         } else if (flags & KFD_IOC_ALLOC_MEM_FLAGS_GTT) {
1396                 domain = alloc_domain = AMDGPU_GEM_DOMAIN_GTT;
1397                 alloc_flags = 0;
1398         } else if (flags & KFD_IOC_ALLOC_MEM_FLAGS_USERPTR) {
1399                 domain = AMDGPU_GEM_DOMAIN_GTT;
1400                 alloc_domain = AMDGPU_GEM_DOMAIN_CPU;
1401                 alloc_flags = AMDGPU_GEM_CREATE_PREEMPTIBLE;
1402                 if (!offset || !*offset)
1403                         return -EINVAL;
1404                 user_addr = untagged_addr(*offset);
1405         } else if (flags & (KFD_IOC_ALLOC_MEM_FLAGS_DOORBELL |
1406                         KFD_IOC_ALLOC_MEM_FLAGS_MMIO_REMAP)) {
1407                 domain = AMDGPU_GEM_DOMAIN_GTT;
1408                 alloc_domain = AMDGPU_GEM_DOMAIN_CPU;
1409                 bo_type = ttm_bo_type_sg;
1410                 alloc_flags = 0;
1411                 if (size > UINT_MAX)
1412                         return -EINVAL;
1413                 sg = create_doorbell_sg(*offset, size);
1414                 if (!sg)
1415                         return -ENOMEM;
1416         } else {
1417                 return -EINVAL;
1418         }
1419
1420         *mem = kzalloc(sizeof(struct kgd_mem), GFP_KERNEL);
1421         if (!*mem) {
1422                 ret = -ENOMEM;
1423                 goto err;
1424         }
1425         INIT_LIST_HEAD(&(*mem)->attachments);
1426         mutex_init(&(*mem)->lock);
1427         (*mem)->aql_queue = !!(flags & KFD_IOC_ALLOC_MEM_FLAGS_AQL_QUEUE_MEM);
1428
1429         /* Workaround for AQL queue wraparound bug. Map the same
1430          * memory twice. That means we only actually allocate half
1431          * the memory.
1432          */
1433         if ((*mem)->aql_queue)
1434                 size = size >> 1;
1435
1436         (*mem)->alloc_flags = flags;
1437
1438         amdgpu_sync_create(&(*mem)->sync);
1439
1440         ret = amdgpu_amdkfd_reserve_mem_limit(adev, size, alloc_domain, !!sg);
1441         if (ret) {
1442                 pr_debug("Insufficient memory\n");
1443                 goto err_reserve_limit;
1444         }
1445
1446         pr_debug("\tcreate BO VA 0x%llx size 0x%llx domain %s\n",
1447                         va, size, domain_string(alloc_domain));
1448
1449         ret = amdgpu_gem_object_create(adev, size, 1, alloc_domain, alloc_flags,
1450                                        bo_type, NULL, &gobj);
1451         if (ret) {
1452                 pr_debug("Failed to create BO on domain %s. ret %d\n",
1453                          domain_string(alloc_domain), ret);
1454                 goto err_bo_create;
1455         }
1456         ret = drm_vma_node_allow(&gobj->vma_node, drm_priv);
1457         if (ret) {
1458                 pr_debug("Failed to allow vma node access. ret %d\n", ret);
1459                 goto err_node_allow;
1460         }
1461         bo = gem_to_amdgpu_bo(gobj);
1462         if (bo_type == ttm_bo_type_sg) {
1463                 bo->tbo.sg = sg;
1464                 bo->tbo.ttm->sg = sg;
1465         }
1466         bo->kfd_bo = *mem;
1467         (*mem)->bo = bo;
1468         if (user_addr)
1469                 bo->flags |= AMDGPU_AMDKFD_CREATE_USERPTR_BO;
1470
1471         (*mem)->va = va;
1472         (*mem)->domain = domain;
1473         (*mem)->mapped_to_gpu_memory = 0;
1474         (*mem)->process_info = avm->process_info;
1475         add_kgd_mem_to_kfd_bo_list(*mem, avm->process_info, user_addr);
1476
1477         if (user_addr) {
1478                 ret = init_user_pages(*mem, user_addr);
1479                 if (ret)
1480                         goto allocate_init_user_pages_failed;
1481         }
1482
1483         if (offset)
1484                 *offset = amdgpu_bo_mmap_offset(bo);
1485
1486         return 0;
1487
1488 allocate_init_user_pages_failed:
1489         remove_kgd_mem_from_kfd_bo_list(*mem, avm->process_info);
1490         drm_vma_node_revoke(&gobj->vma_node, drm_priv);
1491 err_node_allow:
1492         amdgpu_bo_unref(&bo);
1493         /* Don't unreserve system mem limit twice */
1494         goto err_reserve_limit;
1495 err_bo_create:
1496         unreserve_mem_limit(adev, size, alloc_domain, !!sg);
1497 err_reserve_limit:
1498         mutex_destroy(&(*mem)->lock);
1499         kfree(*mem);
1500 err:
1501         if (sg) {
1502                 sg_free_table(sg);
1503                 kfree(sg);
1504         }
1505         return ret;
1506 }
1507
1508 int amdgpu_amdkfd_gpuvm_free_memory_of_gpu(
1509                 struct kgd_dev *kgd, struct kgd_mem *mem, void *drm_priv,
1510                 uint64_t *size)
1511 {
1512         struct amdkfd_process_info *process_info = mem->process_info;
1513         unsigned long bo_size = mem->bo->tbo.base.size;
1514         struct kfd_mem_attachment *entry, *tmp;
1515         struct bo_vm_reservation_context ctx;
1516         struct ttm_validate_buffer *bo_list_entry;
1517         unsigned int mapped_to_gpu_memory;
1518         int ret;
1519         bool is_imported = false;
1520
1521         mutex_lock(&mem->lock);
1522         mapped_to_gpu_memory = mem->mapped_to_gpu_memory;
1523         is_imported = mem->is_imported;
1524         mutex_unlock(&mem->lock);
1525         /* lock is not needed after this, since mem is unused and will
1526          * be freed anyway
1527          */
1528
1529         if (mapped_to_gpu_memory > 0) {
1530                 pr_debug("BO VA 0x%llx size 0x%lx is still mapped.\n",
1531                                 mem->va, bo_size);
1532                 return -EBUSY;
1533         }
1534
1535         /* Make sure restore workers don't access the BO any more */
1536         bo_list_entry = &mem->validate_list;
1537         mutex_lock(&process_info->lock);
1538         list_del(&bo_list_entry->head);
1539         mutex_unlock(&process_info->lock);
1540
1541         /* No more MMU notifiers */
1542         amdgpu_mn_unregister(mem->bo);
1543
1544         ret = reserve_bo_and_cond_vms(mem, NULL, BO_VM_ALL, &ctx);
1545         if (unlikely(ret))
1546                 return ret;
1547
1548         /* The eviction fence should be removed by the last unmap.
1549          * TODO: Log an error condition if the bo still has the eviction fence
1550          * attached
1551          */
1552         amdgpu_amdkfd_remove_eviction_fence(mem->bo,
1553                                         process_info->eviction_fence);
1554         pr_debug("Release VA 0x%llx - 0x%llx\n", mem->va,
1555                 mem->va + bo_size * (1 + mem->aql_queue));
1556
1557         ret = unreserve_bo_and_vms(&ctx, false, false);
1558
1559         /* Remove from VM internal data structures */
1560         list_for_each_entry_safe(entry, tmp, &mem->attachments, list)
1561                 kfd_mem_detach(entry);
1562
1563         /* Free the sync object */
1564         amdgpu_sync_free(&mem->sync);
1565
1566         /* If the SG is not NULL, it's one we created for a doorbell or mmio
1567          * remap BO. We need to free it.
1568          */
1569         if (mem->bo->tbo.sg) {
1570                 sg_free_table(mem->bo->tbo.sg);
1571                 kfree(mem->bo->tbo.sg);
1572         }
1573
1574         /* Update the size of the BO being freed if it was allocated from
1575          * VRAM and is not imported.
1576          */
1577         if (size) {
1578                 if ((mem->bo->preferred_domains == AMDGPU_GEM_DOMAIN_VRAM) &&
1579                     (!is_imported))
1580                         *size = bo_size;
1581                 else
1582                         *size = 0;
1583         }
1584
1585         /* Free the BO*/
1586         drm_vma_node_revoke(&mem->bo->tbo.base.vma_node, drm_priv);
1587         if (mem->dmabuf)
1588                 dma_buf_put(mem->dmabuf);
1589         drm_gem_object_put(&mem->bo->tbo.base);
1590         mutex_destroy(&mem->lock);
1591         kfree(mem);
1592
1593         return ret;
1594 }
1595
1596 int amdgpu_amdkfd_gpuvm_map_memory_to_gpu(
1597                 struct kgd_dev *kgd, struct kgd_mem *mem, void *drm_priv)
1598 {
1599         struct amdgpu_device *adev = get_amdgpu_device(kgd);
1600         struct amdgpu_vm *avm = drm_priv_to_vm(drm_priv);
1601         int ret;
1602         struct amdgpu_bo *bo;
1603         uint32_t domain;
1604         struct kfd_mem_attachment *entry;
1605         struct bo_vm_reservation_context ctx;
1606         unsigned long bo_size;
1607         bool is_invalid_userptr = false;
1608
1609         bo = mem->bo;
1610         if (!bo) {
1611                 pr_err("Invalid BO when mapping memory to GPU\n");
1612                 return -EINVAL;
1613         }
1614
1615         /* Make sure restore is not running concurrently. Since we
1616          * don't map invalid userptr BOs, we rely on the next restore
1617          * worker to do the mapping
1618          */
1619         mutex_lock(&mem->process_info->lock);
1620
1621         /* Lock mmap-sem. If we find an invalid userptr BO, we can be
1622          * sure that the MMU notifier is no longer running
1623          * concurrently and the queues are actually stopped
1624          */
1625         if (amdgpu_ttm_tt_get_usermm(bo->tbo.ttm)) {
1626                 mmap_write_lock(current->mm);
1627                 is_invalid_userptr = atomic_read(&mem->invalid);
1628                 mmap_write_unlock(current->mm);
1629         }
1630
1631         mutex_lock(&mem->lock);
1632
1633         domain = mem->domain;
1634         bo_size = bo->tbo.base.size;
1635
1636         pr_debug("Map VA 0x%llx - 0x%llx to vm %p domain %s\n",
1637                         mem->va,
1638                         mem->va + bo_size * (1 + mem->aql_queue),
1639                         avm, domain_string(domain));
1640
1641         if (!kfd_mem_is_attached(avm, mem)) {
1642                 ret = kfd_mem_attach(adev, mem, avm, mem->aql_queue);
1643                 if (ret)
1644                         goto out;
1645         }
1646
1647         ret = reserve_bo_and_vm(mem, avm, &ctx);
1648         if (unlikely(ret))
1649                 goto out;
1650
1651         /* Userptr can be marked as "not invalid", but not actually be
1652          * validated yet (still in the system domain). In that case
1653          * the queues are still stopped and we can leave mapping for
1654          * the next restore worker
1655          */
1656         if (amdgpu_ttm_tt_get_usermm(bo->tbo.ttm) &&
1657             bo->tbo.resource->mem_type == TTM_PL_SYSTEM)
1658                 is_invalid_userptr = true;
1659
1660         ret = vm_validate_pt_pd_bos(avm);
1661         if (unlikely(ret))
1662                 goto out_unreserve;
1663
1664         if (mem->mapped_to_gpu_memory == 0 &&
1665             !amdgpu_ttm_tt_get_usermm(bo->tbo.ttm)) {
1666                 /* Validate BO only once. The eviction fence gets added to BO
1667                  * the first time it is mapped. Validate will wait for all
1668                  * background evictions to complete.
1669                  */
1670                 ret = amdgpu_amdkfd_bo_validate(bo, domain, true);
1671                 if (ret) {
1672                         pr_debug("Validate failed\n");
1673                         goto out_unreserve;
1674                 }
1675         }
1676
1677         list_for_each_entry(entry, &mem->attachments, list) {
1678                 if (entry->bo_va->base.vm != avm || entry->is_mapped)
1679                         continue;
1680
1681                 pr_debug("\t map VA 0x%llx - 0x%llx in entry %p\n",
1682                          entry->va, entry->va + bo_size, entry);
1683
1684                 ret = map_bo_to_gpuvm(mem, entry, ctx.sync,
1685                                       is_invalid_userptr);
1686                 if (ret) {
1687                         pr_err("Failed to map bo to gpuvm\n");
1688                         goto out_unreserve;
1689                 }
1690
1691                 ret = vm_update_pds(avm, ctx.sync);
1692                 if (ret) {
1693                         pr_err("Failed to update page directories\n");
1694                         goto out_unreserve;
1695                 }
1696
1697                 entry->is_mapped = true;
1698                 mem->mapped_to_gpu_memory++;
1699                 pr_debug("\t INC mapping count %d\n",
1700                          mem->mapped_to_gpu_memory);
1701         }
1702
1703         if (!amdgpu_ttm_tt_get_usermm(bo->tbo.ttm) && !bo->tbo.pin_count)
1704                 amdgpu_bo_fence(bo,
1705                                 &avm->process_info->eviction_fence->base,
1706                                 true);
1707         ret = unreserve_bo_and_vms(&ctx, false, false);
1708
1709         goto out;
1710
1711 out_unreserve:
1712         unreserve_bo_and_vms(&ctx, false, false);
1713 out:
1714         mutex_unlock(&mem->process_info->lock);
1715         mutex_unlock(&mem->lock);
1716         return ret;
1717 }
1718
1719 int amdgpu_amdkfd_gpuvm_unmap_memory_from_gpu(
1720                 struct kgd_dev *kgd, struct kgd_mem *mem, void *drm_priv)
1721 {
1722         struct amdgpu_vm *avm = drm_priv_to_vm(drm_priv);
1723         struct amdkfd_process_info *process_info = avm->process_info;
1724         unsigned long bo_size = mem->bo->tbo.base.size;
1725         struct kfd_mem_attachment *entry;
1726         struct bo_vm_reservation_context ctx;
1727         int ret;
1728
1729         mutex_lock(&mem->lock);
1730
1731         ret = reserve_bo_and_cond_vms(mem, avm, BO_VM_MAPPED, &ctx);
1732         if (unlikely(ret))
1733                 goto out;
1734         /* If no VMs were reserved, it means the BO wasn't actually mapped */
1735         if (ctx.n_vms == 0) {
1736                 ret = -EINVAL;
1737                 goto unreserve_out;
1738         }
1739
1740         ret = vm_validate_pt_pd_bos(avm);
1741         if (unlikely(ret))
1742                 goto unreserve_out;
1743
1744         pr_debug("Unmap VA 0x%llx - 0x%llx from vm %p\n",
1745                 mem->va,
1746                 mem->va + bo_size * (1 + mem->aql_queue),
1747                 avm);
1748
1749         list_for_each_entry(entry, &mem->attachments, list) {
1750                 if (entry->bo_va->base.vm != avm || !entry->is_mapped)
1751                         continue;
1752
1753                 pr_debug("\t unmap VA 0x%llx - 0x%llx from entry %p\n",
1754                          entry->va, entry->va + bo_size, entry);
1755
1756                 unmap_bo_from_gpuvm(mem, entry, ctx.sync);
1757                 entry->is_mapped = false;
1758
1759                 mem->mapped_to_gpu_memory--;
1760                 pr_debug("\t DEC mapping count %d\n",
1761                          mem->mapped_to_gpu_memory);
1762         }
1763
1764         /* If BO is unmapped from all VMs, unfence it. It can be evicted if
1765          * required.
1766          */
1767         if (mem->mapped_to_gpu_memory == 0 &&
1768             !amdgpu_ttm_tt_get_usermm(mem->bo->tbo.ttm) &&
1769             !mem->bo->tbo.pin_count)
1770                 amdgpu_amdkfd_remove_eviction_fence(mem->bo,
1771                                                 process_info->eviction_fence);
1772
1773 unreserve_out:
1774         unreserve_bo_and_vms(&ctx, false, false);
1775 out:
1776         mutex_unlock(&mem->lock);
1777         return ret;
1778 }
1779
1780 int amdgpu_amdkfd_gpuvm_sync_memory(
1781                 struct kgd_dev *kgd, struct kgd_mem *mem, bool intr)
1782 {
1783         struct amdgpu_sync sync;
1784         int ret;
1785
1786         amdgpu_sync_create(&sync);
1787
1788         mutex_lock(&mem->lock);
1789         amdgpu_sync_clone(&mem->sync, &sync);
1790         mutex_unlock(&mem->lock);
1791
1792         ret = amdgpu_sync_wait(&sync, intr);
1793         amdgpu_sync_free(&sync);
1794         return ret;
1795 }
1796
1797 int amdgpu_amdkfd_gpuvm_map_gtt_bo_to_kernel(struct kgd_dev *kgd,
1798                 struct kgd_mem *mem, void **kptr, uint64_t *size)
1799 {
1800         int ret;
1801         struct amdgpu_bo *bo = mem->bo;
1802
1803         if (amdgpu_ttm_tt_get_usermm(bo->tbo.ttm)) {
1804                 pr_err("userptr can't be mapped to kernel\n");
1805                 return -EINVAL;
1806         }
1807
1808         /* delete kgd_mem from kfd_bo_list to avoid re-validating
1809          * this BO in BO's restoring after eviction.
1810          */
1811         mutex_lock(&mem->process_info->lock);
1812
1813         ret = amdgpu_bo_reserve(bo, true);
1814         if (ret) {
1815                 pr_err("Failed to reserve bo. ret %d\n", ret);
1816                 goto bo_reserve_failed;
1817         }
1818
1819         ret = amdgpu_bo_pin(bo, AMDGPU_GEM_DOMAIN_GTT);
1820         if (ret) {
1821                 pr_err("Failed to pin bo. ret %d\n", ret);
1822                 goto pin_failed;
1823         }
1824
1825         ret = amdgpu_bo_kmap(bo, kptr);
1826         if (ret) {
1827                 pr_err("Failed to map bo to kernel. ret %d\n", ret);
1828                 goto kmap_failed;
1829         }
1830
1831         amdgpu_amdkfd_remove_eviction_fence(
1832                 bo, mem->process_info->eviction_fence);
1833         list_del_init(&mem->validate_list.head);
1834
1835         if (size)
1836                 *size = amdgpu_bo_size(bo);
1837
1838         amdgpu_bo_unreserve(bo);
1839
1840         mutex_unlock(&mem->process_info->lock);
1841         return 0;
1842
1843 kmap_failed:
1844         amdgpu_bo_unpin(bo);
1845 pin_failed:
1846         amdgpu_bo_unreserve(bo);
1847 bo_reserve_failed:
1848         mutex_unlock(&mem->process_info->lock);
1849
1850         return ret;
1851 }
1852
1853 int amdgpu_amdkfd_gpuvm_get_vm_fault_info(struct kgd_dev *kgd,
1854                                               struct kfd_vm_fault_info *mem)
1855 {
1856         struct amdgpu_device *adev;
1857
1858         adev = (struct amdgpu_device *)kgd;
1859         if (atomic_read(&adev->gmc.vm_fault_info_updated) == 1) {
1860                 *mem = *adev->gmc.vm_fault_info;
1861                 mb();
1862                 atomic_set(&adev->gmc.vm_fault_info_updated, 0);
1863         }
1864         return 0;
1865 }
1866
1867 int amdgpu_amdkfd_gpuvm_import_dmabuf(struct kgd_dev *kgd,
1868                                       struct dma_buf *dma_buf,
1869                                       uint64_t va, void *drm_priv,
1870                                       struct kgd_mem **mem, uint64_t *size,
1871                                       uint64_t *mmap_offset)
1872 {
1873         struct amdgpu_device *adev = (struct amdgpu_device *)kgd;
1874         struct amdgpu_vm *avm = drm_priv_to_vm(drm_priv);
1875         struct drm_gem_object *obj;
1876         struct amdgpu_bo *bo;
1877         int ret;
1878
1879         if (dma_buf->ops != &amdgpu_dmabuf_ops)
1880                 /* Can't handle non-graphics buffers */
1881                 return -EINVAL;
1882
1883         obj = dma_buf->priv;
1884         if (drm_to_adev(obj->dev) != adev)
1885                 /* Can't handle buffers from other devices */
1886                 return -EINVAL;
1887
1888         bo = gem_to_amdgpu_bo(obj);
1889         if (!(bo->preferred_domains & (AMDGPU_GEM_DOMAIN_VRAM |
1890                                     AMDGPU_GEM_DOMAIN_GTT)))
1891                 /* Only VRAM and GTT BOs are supported */
1892                 return -EINVAL;
1893
1894         *mem = kzalloc(sizeof(struct kgd_mem), GFP_KERNEL);
1895         if (!*mem)
1896                 return -ENOMEM;
1897
1898         ret = drm_vma_node_allow(&obj->vma_node, drm_priv);
1899         if (ret) {
1900                 kfree(mem);
1901                 return ret;
1902         }
1903
1904         if (size)
1905                 *size = amdgpu_bo_size(bo);
1906
1907         if (mmap_offset)
1908                 *mmap_offset = amdgpu_bo_mmap_offset(bo);
1909
1910         INIT_LIST_HEAD(&(*mem)->attachments);
1911         mutex_init(&(*mem)->lock);
1912
1913         (*mem)->alloc_flags =
1914                 ((bo->preferred_domains & AMDGPU_GEM_DOMAIN_VRAM) ?
1915                 KFD_IOC_ALLOC_MEM_FLAGS_VRAM : KFD_IOC_ALLOC_MEM_FLAGS_GTT)
1916                 | KFD_IOC_ALLOC_MEM_FLAGS_WRITABLE
1917                 | KFD_IOC_ALLOC_MEM_FLAGS_EXECUTABLE;
1918
1919         drm_gem_object_get(&bo->tbo.base);
1920         (*mem)->bo = bo;
1921         (*mem)->va = va;
1922         (*mem)->domain = (bo->preferred_domains & AMDGPU_GEM_DOMAIN_VRAM) ?
1923                 AMDGPU_GEM_DOMAIN_VRAM : AMDGPU_GEM_DOMAIN_GTT;
1924         (*mem)->mapped_to_gpu_memory = 0;
1925         (*mem)->process_info = avm->process_info;
1926         add_kgd_mem_to_kfd_bo_list(*mem, avm->process_info, false);
1927         amdgpu_sync_create(&(*mem)->sync);
1928         (*mem)->is_imported = true;
1929
1930         return 0;
1931 }
1932
1933 /* Evict a userptr BO by stopping the queues if necessary
1934  *
1935  * Runs in MMU notifier, may be in RECLAIM_FS context. This means it
1936  * cannot do any memory allocations, and cannot take any locks that
1937  * are held elsewhere while allocating memory. Therefore this is as
1938  * simple as possible, using atomic counters.
1939  *
1940  * It doesn't do anything to the BO itself. The real work happens in
1941  * restore, where we get updated page addresses. This function only
1942  * ensures that GPU access to the BO is stopped.
1943  */
1944 int amdgpu_amdkfd_evict_userptr(struct kgd_mem *mem,
1945                                 struct mm_struct *mm)
1946 {
1947         struct amdkfd_process_info *process_info = mem->process_info;
1948         int evicted_bos;
1949         int r = 0;
1950
1951         atomic_inc(&mem->invalid);
1952         evicted_bos = atomic_inc_return(&process_info->evicted_bos);
1953         if (evicted_bos == 1) {
1954                 /* First eviction, stop the queues */
1955                 r = kgd2kfd_quiesce_mm(mm);
1956                 if (r)
1957                         pr_err("Failed to quiesce KFD\n");
1958                 schedule_delayed_work(&process_info->restore_userptr_work,
1959                         msecs_to_jiffies(AMDGPU_USERPTR_RESTORE_DELAY_MS));
1960         }
1961
1962         return r;
1963 }
1964
1965 /* Update invalid userptr BOs
1966  *
1967  * Moves invalidated (evicted) userptr BOs from userptr_valid_list to
1968  * userptr_inval_list and updates user pages for all BOs that have
1969  * been invalidated since their last update.
1970  */
1971 static int update_invalid_user_pages(struct amdkfd_process_info *process_info,
1972                                      struct mm_struct *mm)
1973 {
1974         struct kgd_mem *mem, *tmp_mem;
1975         struct amdgpu_bo *bo;
1976         struct ttm_operation_ctx ctx = { false, false };
1977         int invalid, ret;
1978
1979         /* Move all invalidated BOs to the userptr_inval_list and
1980          * release their user pages by migration to the CPU domain
1981          */
1982         list_for_each_entry_safe(mem, tmp_mem,
1983                                  &process_info->userptr_valid_list,
1984                                  validate_list.head) {
1985                 if (!atomic_read(&mem->invalid))
1986                         continue; /* BO is still valid */
1987
1988                 bo = mem->bo;
1989
1990                 if (amdgpu_bo_reserve(bo, true))
1991                         return -EAGAIN;
1992                 amdgpu_bo_placement_from_domain(bo, AMDGPU_GEM_DOMAIN_CPU);
1993                 ret = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
1994                 amdgpu_bo_unreserve(bo);
1995                 if (ret) {
1996                         pr_err("%s: Failed to invalidate userptr BO\n",
1997                                __func__);
1998                         return -EAGAIN;
1999                 }
2000
2001                 list_move_tail(&mem->validate_list.head,
2002                                &process_info->userptr_inval_list);
2003         }
2004
2005         if (list_empty(&process_info->userptr_inval_list))
2006                 return 0; /* All evicted userptr BOs were freed */
2007
2008         /* Go through userptr_inval_list and update any invalid user_pages */
2009         list_for_each_entry(mem, &process_info->userptr_inval_list,
2010                             validate_list.head) {
2011                 invalid = atomic_read(&mem->invalid);
2012                 if (!invalid)
2013                         /* BO hasn't been invalidated since the last
2014                          * revalidation attempt. Keep its BO list.
2015                          */
2016                         continue;
2017
2018                 bo = mem->bo;
2019
2020                 /* Get updated user pages */
2021                 ret = amdgpu_ttm_tt_get_user_pages(bo, bo->tbo.ttm->pages);
2022                 if (ret) {
2023                         pr_debug("%s: Failed to get user pages: %d\n",
2024                                 __func__, ret);
2025
2026                         /* Return error -EBUSY or -ENOMEM, retry restore */
2027                         return ret;
2028                 }
2029
2030                 /*
2031                  * FIXME: Cannot ignore the return code, must hold
2032                  * notifier_lock
2033                  */
2034                 amdgpu_ttm_tt_get_user_pages_done(bo->tbo.ttm);
2035
2036                 /* Mark the BO as valid unless it was invalidated
2037                  * again concurrently.
2038                  */
2039                 if (atomic_cmpxchg(&mem->invalid, invalid, 0) != invalid)
2040                         return -EAGAIN;
2041         }
2042
2043         return 0;
2044 }
2045
2046 /* Validate invalid userptr BOs
2047  *
2048  * Validates BOs on the userptr_inval_list, and moves them back to the
2049  * userptr_valid_list. Also updates GPUVM page tables with new page
2050  * addresses and waits for the page table updates to complete.
2051  */
2052 static int validate_invalid_user_pages(struct amdkfd_process_info *process_info)
2053 {
2054         struct amdgpu_bo_list_entry *pd_bo_list_entries;
2055         struct list_head resv_list, duplicates;
2056         struct ww_acquire_ctx ticket;
2057         struct amdgpu_sync sync;
2058
2059         struct amdgpu_vm *peer_vm;
2060         struct kgd_mem *mem, *tmp_mem;
2061         struct amdgpu_bo *bo;
2062         struct ttm_operation_ctx ctx = { false, false };
2063         int i, ret;
2064
2065         pd_bo_list_entries = kcalloc(process_info->n_vms,
2066                                      sizeof(struct amdgpu_bo_list_entry),
2067                                      GFP_KERNEL);
2068         if (!pd_bo_list_entries) {
2069                 pr_err("%s: Failed to allocate PD BO list entries\n", __func__);
2070                 ret = -ENOMEM;
2071                 goto out_no_mem;
2072         }
2073
2074         INIT_LIST_HEAD(&resv_list);
2075         INIT_LIST_HEAD(&duplicates);
2076
2077         /* Get all the page directory BOs that need to be reserved */
2078         i = 0;
2079         list_for_each_entry(peer_vm, &process_info->vm_list_head,
2080                             vm_list_node)
2081                 amdgpu_vm_get_pd_bo(peer_vm, &resv_list,
2082                                     &pd_bo_list_entries[i++]);
2083         /* Add the userptr_inval_list entries to resv_list */
2084         list_for_each_entry(mem, &process_info->userptr_inval_list,
2085                             validate_list.head) {
2086                 list_add_tail(&mem->resv_list.head, &resv_list);
2087                 mem->resv_list.bo = mem->validate_list.bo;
2088                 mem->resv_list.num_shared = mem->validate_list.num_shared;
2089         }
2090
2091         /* Reserve all BOs and page tables for validation */
2092         ret = ttm_eu_reserve_buffers(&ticket, &resv_list, false, &duplicates);
2093         WARN(!list_empty(&duplicates), "Duplicates should be empty");
2094         if (ret)
2095                 goto out_free;
2096
2097         amdgpu_sync_create(&sync);
2098
2099         ret = process_validate_vms(process_info);
2100         if (ret)
2101                 goto unreserve_out;
2102
2103         /* Validate BOs and update GPUVM page tables */
2104         list_for_each_entry_safe(mem, tmp_mem,
2105                                  &process_info->userptr_inval_list,
2106                                  validate_list.head) {
2107                 struct kfd_mem_attachment *attachment;
2108
2109                 bo = mem->bo;
2110
2111                 /* Validate the BO if we got user pages */
2112                 if (bo->tbo.ttm->pages[0]) {
2113                         amdgpu_bo_placement_from_domain(bo, mem->domain);
2114                         ret = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
2115                         if (ret) {
2116                                 pr_err("%s: failed to validate BO\n", __func__);
2117                                 goto unreserve_out;
2118                         }
2119                 }
2120
2121                 list_move_tail(&mem->validate_list.head,
2122                                &process_info->userptr_valid_list);
2123
2124                 /* Update mapping. If the BO was not validated
2125                  * (because we couldn't get user pages), this will
2126                  * clear the page table entries, which will result in
2127                  * VM faults if the GPU tries to access the invalid
2128                  * memory.
2129                  */
2130                 list_for_each_entry(attachment, &mem->attachments, list) {
2131                         if (!attachment->is_mapped)
2132                                 continue;
2133
2134                         kfd_mem_dmaunmap_attachment(mem, attachment);
2135                         ret = update_gpuvm_pte(mem, attachment, &sync);
2136                         if (ret) {
2137                                 pr_err("%s: update PTE failed\n", __func__);
2138                                 /* make sure this gets validated again */
2139                                 atomic_inc(&mem->invalid);
2140                                 goto unreserve_out;
2141                         }
2142                 }
2143         }
2144
2145         /* Update page directories */
2146         ret = process_update_pds(process_info, &sync);
2147
2148 unreserve_out:
2149         ttm_eu_backoff_reservation(&ticket, &resv_list);
2150         amdgpu_sync_wait(&sync, false);
2151         amdgpu_sync_free(&sync);
2152 out_free:
2153         kfree(pd_bo_list_entries);
2154 out_no_mem:
2155
2156         return ret;
2157 }
2158
2159 /* Worker callback to restore evicted userptr BOs
2160  *
2161  * Tries to update and validate all userptr BOs. If successful and no
2162  * concurrent evictions happened, the queues are restarted. Otherwise,
2163  * reschedule for another attempt later.
2164  */
2165 static void amdgpu_amdkfd_restore_userptr_worker(struct work_struct *work)
2166 {
2167         struct delayed_work *dwork = to_delayed_work(work);
2168         struct amdkfd_process_info *process_info =
2169                 container_of(dwork, struct amdkfd_process_info,
2170                              restore_userptr_work);
2171         struct task_struct *usertask;
2172         struct mm_struct *mm;
2173         int evicted_bos;
2174
2175         evicted_bos = atomic_read(&process_info->evicted_bos);
2176         if (!evicted_bos)
2177                 return;
2178
2179         /* Reference task and mm in case of concurrent process termination */
2180         usertask = get_pid_task(process_info->pid, PIDTYPE_PID);
2181         if (!usertask)
2182                 return;
2183         mm = get_task_mm(usertask);
2184         if (!mm) {
2185                 put_task_struct(usertask);
2186                 return;
2187         }
2188
2189         mutex_lock(&process_info->lock);
2190
2191         if (update_invalid_user_pages(process_info, mm))
2192                 goto unlock_out;
2193         /* userptr_inval_list can be empty if all evicted userptr BOs
2194          * have been freed. In that case there is nothing to validate
2195          * and we can just restart the queues.
2196          */
2197         if (!list_empty(&process_info->userptr_inval_list)) {
2198                 if (atomic_read(&process_info->evicted_bos) != evicted_bos)
2199                         goto unlock_out; /* Concurrent eviction, try again */
2200
2201                 if (validate_invalid_user_pages(process_info))
2202                         goto unlock_out;
2203         }
2204         /* Final check for concurrent evicton and atomic update. If
2205          * another eviction happens after successful update, it will
2206          * be a first eviction that calls quiesce_mm. The eviction
2207          * reference counting inside KFD will handle this case.
2208          */
2209         if (atomic_cmpxchg(&process_info->evicted_bos, evicted_bos, 0) !=
2210             evicted_bos)
2211                 goto unlock_out;
2212         evicted_bos = 0;
2213         if (kgd2kfd_resume_mm(mm)) {
2214                 pr_err("%s: Failed to resume KFD\n", __func__);
2215                 /* No recovery from this failure. Probably the CP is
2216                  * hanging. No point trying again.
2217                  */
2218         }
2219
2220 unlock_out:
2221         mutex_unlock(&process_info->lock);
2222         mmput(mm);
2223         put_task_struct(usertask);
2224
2225         /* If validation failed, reschedule another attempt */
2226         if (evicted_bos)
2227                 schedule_delayed_work(&process_info->restore_userptr_work,
2228                         msecs_to_jiffies(AMDGPU_USERPTR_RESTORE_DELAY_MS));
2229 }
2230
2231 /** amdgpu_amdkfd_gpuvm_restore_process_bos - Restore all BOs for the given
2232  *   KFD process identified by process_info
2233  *
2234  * @process_info: amdkfd_process_info of the KFD process
2235  *
2236  * After memory eviction, restore thread calls this function. The function
2237  * should be called when the Process is still valid. BO restore involves -
2238  *
2239  * 1.  Release old eviction fence and create new one
2240  * 2.  Get two copies of PD BO list from all the VMs. Keep one copy as pd_list.
2241  * 3   Use the second PD list and kfd_bo_list to create a list (ctx.list) of
2242  *     BOs that need to be reserved.
2243  * 4.  Reserve all the BOs
2244  * 5.  Validate of PD and PT BOs.
2245  * 6.  Validate all KFD BOs using kfd_bo_list and Map them and add new fence
2246  * 7.  Add fence to all PD and PT BOs.
2247  * 8.  Unreserve all BOs
2248  */
2249 int amdgpu_amdkfd_gpuvm_restore_process_bos(void *info, struct dma_fence **ef)
2250 {
2251         struct amdgpu_bo_list_entry *pd_bo_list;
2252         struct amdkfd_process_info *process_info = info;
2253         struct amdgpu_vm *peer_vm;
2254         struct kgd_mem *mem;
2255         struct bo_vm_reservation_context ctx;
2256         struct amdgpu_amdkfd_fence *new_fence;
2257         int ret = 0, i;
2258         struct list_head duplicate_save;
2259         struct amdgpu_sync sync_obj;
2260         unsigned long failed_size = 0;
2261         unsigned long total_size = 0;
2262
2263         INIT_LIST_HEAD(&duplicate_save);
2264         INIT_LIST_HEAD(&ctx.list);
2265         INIT_LIST_HEAD(&ctx.duplicates);
2266
2267         pd_bo_list = kcalloc(process_info->n_vms,
2268                              sizeof(struct amdgpu_bo_list_entry),
2269                              GFP_KERNEL);
2270         if (!pd_bo_list)
2271                 return -ENOMEM;
2272
2273         i = 0;
2274         mutex_lock(&process_info->lock);
2275         list_for_each_entry(peer_vm, &process_info->vm_list_head,
2276                         vm_list_node)
2277                 amdgpu_vm_get_pd_bo(peer_vm, &ctx.list, &pd_bo_list[i++]);
2278
2279         /* Reserve all BOs and page tables/directory. Add all BOs from
2280          * kfd_bo_list to ctx.list
2281          */
2282         list_for_each_entry(mem, &process_info->kfd_bo_list,
2283                             validate_list.head) {
2284
2285                 list_add_tail(&mem->resv_list.head, &ctx.list);
2286                 mem->resv_list.bo = mem->validate_list.bo;
2287                 mem->resv_list.num_shared = mem->validate_list.num_shared;
2288         }
2289
2290         ret = ttm_eu_reserve_buffers(&ctx.ticket, &ctx.list,
2291                                      false, &duplicate_save);
2292         if (ret) {
2293                 pr_debug("Memory eviction: TTM Reserve Failed. Try again\n");
2294                 goto ttm_reserve_fail;
2295         }
2296
2297         amdgpu_sync_create(&sync_obj);
2298
2299         /* Validate PDs and PTs */
2300         ret = process_validate_vms(process_info);
2301         if (ret)
2302                 goto validate_map_fail;
2303
2304         ret = process_sync_pds_resv(process_info, &sync_obj);
2305         if (ret) {
2306                 pr_debug("Memory eviction: Failed to sync to PD BO moving fence. Try again\n");
2307                 goto validate_map_fail;
2308         }
2309
2310         /* Validate BOs and map them to GPUVM (update VM page tables). */
2311         list_for_each_entry(mem, &process_info->kfd_bo_list,
2312                             validate_list.head) {
2313
2314                 struct amdgpu_bo *bo = mem->bo;
2315                 uint32_t domain = mem->domain;
2316                 struct kfd_mem_attachment *attachment;
2317
2318                 total_size += amdgpu_bo_size(bo);
2319
2320                 ret = amdgpu_amdkfd_bo_validate(bo, domain, false);
2321                 if (ret) {
2322                         pr_debug("Memory eviction: Validate BOs failed\n");
2323                         failed_size += amdgpu_bo_size(bo);
2324                         ret = amdgpu_amdkfd_bo_validate(bo,
2325                                                 AMDGPU_GEM_DOMAIN_GTT, false);
2326                         if (ret) {
2327                                 pr_debug("Memory eviction: Try again\n");
2328                                 goto validate_map_fail;
2329                         }
2330                 }
2331                 ret = amdgpu_sync_fence(&sync_obj, bo->tbo.moving);
2332                 if (ret) {
2333                         pr_debug("Memory eviction: Sync BO fence failed. Try again\n");
2334                         goto validate_map_fail;
2335                 }
2336                 list_for_each_entry(attachment, &mem->attachments, list) {
2337                         if (!attachment->is_mapped)
2338                                 continue;
2339
2340                         kfd_mem_dmaunmap_attachment(mem, attachment);
2341                         ret = update_gpuvm_pte(mem, attachment, &sync_obj);
2342                         if (ret) {
2343                                 pr_debug("Memory eviction: update PTE failed. Try again\n");
2344                                 goto validate_map_fail;
2345                         }
2346                 }
2347         }
2348
2349         if (failed_size)
2350                 pr_debug("0x%lx/0x%lx in system\n", failed_size, total_size);
2351
2352         /* Update page directories */
2353         ret = process_update_pds(process_info, &sync_obj);
2354         if (ret) {
2355                 pr_debug("Memory eviction: update PDs failed. Try again\n");
2356                 goto validate_map_fail;
2357         }
2358
2359         /* Wait for validate and PT updates to finish */
2360         amdgpu_sync_wait(&sync_obj, false);
2361
2362         /* Release old eviction fence and create new one, because fence only
2363          * goes from unsignaled to signaled, fence cannot be reused.
2364          * Use context and mm from the old fence.
2365          */
2366         new_fence = amdgpu_amdkfd_fence_create(
2367                                 process_info->eviction_fence->base.context,
2368                                 process_info->eviction_fence->mm,
2369                                 NULL);
2370         if (!new_fence) {
2371                 pr_err("Failed to create eviction fence\n");
2372                 ret = -ENOMEM;
2373                 goto validate_map_fail;
2374         }
2375         dma_fence_put(&process_info->eviction_fence->base);
2376         process_info->eviction_fence = new_fence;
2377         *ef = dma_fence_get(&new_fence->base);
2378
2379         /* Attach new eviction fence to all BOs */
2380         list_for_each_entry(mem, &process_info->kfd_bo_list,
2381                 validate_list.head)
2382                 amdgpu_bo_fence(mem->bo,
2383                         &process_info->eviction_fence->base, true);
2384
2385         /* Attach eviction fence to PD / PT BOs */
2386         list_for_each_entry(peer_vm, &process_info->vm_list_head,
2387                             vm_list_node) {
2388                 struct amdgpu_bo *bo = peer_vm->root.bo;
2389
2390                 amdgpu_bo_fence(bo, &process_info->eviction_fence->base, true);
2391         }
2392
2393 validate_map_fail:
2394         ttm_eu_backoff_reservation(&ctx.ticket, &ctx.list);
2395         amdgpu_sync_free(&sync_obj);
2396 ttm_reserve_fail:
2397         mutex_unlock(&process_info->lock);
2398         kfree(pd_bo_list);
2399         return ret;
2400 }
2401
2402 int amdgpu_amdkfd_add_gws_to_process(void *info, void *gws, struct kgd_mem **mem)
2403 {
2404         struct amdkfd_process_info *process_info = (struct amdkfd_process_info *)info;
2405         struct amdgpu_bo *gws_bo = (struct amdgpu_bo *)gws;
2406         int ret;
2407
2408         if (!info || !gws)
2409                 return -EINVAL;
2410
2411         *mem = kzalloc(sizeof(struct kgd_mem), GFP_KERNEL);
2412         if (!*mem)
2413                 return -ENOMEM;
2414
2415         mutex_init(&(*mem)->lock);
2416         INIT_LIST_HEAD(&(*mem)->attachments);
2417         (*mem)->bo = amdgpu_bo_ref(gws_bo);
2418         (*mem)->domain = AMDGPU_GEM_DOMAIN_GWS;
2419         (*mem)->process_info = process_info;
2420         add_kgd_mem_to_kfd_bo_list(*mem, process_info, false);
2421         amdgpu_sync_create(&(*mem)->sync);
2422
2423
2424         /* Validate gws bo the first time it is added to process */
2425         mutex_lock(&(*mem)->process_info->lock);
2426         ret = amdgpu_bo_reserve(gws_bo, false);
2427         if (unlikely(ret)) {
2428                 pr_err("Reserve gws bo failed %d\n", ret);
2429                 goto bo_reservation_failure;
2430         }
2431
2432         ret = amdgpu_amdkfd_bo_validate(gws_bo, AMDGPU_GEM_DOMAIN_GWS, true);
2433         if (ret) {
2434                 pr_err("GWS BO validate failed %d\n", ret);
2435                 goto bo_validation_failure;
2436         }
2437         /* GWS resource is shared b/t amdgpu and amdkfd
2438          * Add process eviction fence to bo so they can
2439          * evict each other.
2440          */
2441         ret = dma_resv_reserve_shared(gws_bo->tbo.base.resv, 1);
2442         if (ret)
2443                 goto reserve_shared_fail;
2444         amdgpu_bo_fence(gws_bo, &process_info->eviction_fence->base, true);
2445         amdgpu_bo_unreserve(gws_bo);
2446         mutex_unlock(&(*mem)->process_info->lock);
2447
2448         return ret;
2449
2450 reserve_shared_fail:
2451 bo_validation_failure:
2452         amdgpu_bo_unreserve(gws_bo);
2453 bo_reservation_failure:
2454         mutex_unlock(&(*mem)->process_info->lock);
2455         amdgpu_sync_free(&(*mem)->sync);
2456         remove_kgd_mem_from_kfd_bo_list(*mem, process_info);
2457         amdgpu_bo_unref(&gws_bo);
2458         mutex_destroy(&(*mem)->lock);
2459         kfree(*mem);
2460         *mem = NULL;
2461         return ret;
2462 }
2463
2464 int amdgpu_amdkfd_remove_gws_from_process(void *info, void *mem)
2465 {
2466         int ret;
2467         struct amdkfd_process_info *process_info = (struct amdkfd_process_info *)info;
2468         struct kgd_mem *kgd_mem = (struct kgd_mem *)mem;
2469         struct amdgpu_bo *gws_bo = kgd_mem->bo;
2470
2471         /* Remove BO from process's validate list so restore worker won't touch
2472          * it anymore
2473          */
2474         remove_kgd_mem_from_kfd_bo_list(kgd_mem, process_info);
2475
2476         ret = amdgpu_bo_reserve(gws_bo, false);
2477         if (unlikely(ret)) {
2478                 pr_err("Reserve gws bo failed %d\n", ret);
2479                 //TODO add BO back to validate_list?
2480                 return ret;
2481         }
2482         amdgpu_amdkfd_remove_eviction_fence(gws_bo,
2483                         process_info->eviction_fence);
2484         amdgpu_bo_unreserve(gws_bo);
2485         amdgpu_sync_free(&kgd_mem->sync);
2486         amdgpu_bo_unref(&gws_bo);
2487         mutex_destroy(&kgd_mem->lock);
2488         kfree(mem);
2489         return 0;
2490 }
2491
2492 /* Returns GPU-specific tiling mode information */
2493 int amdgpu_amdkfd_get_tile_config(struct kgd_dev *kgd,
2494                                 struct tile_config *config)
2495 {
2496         struct amdgpu_device *adev = (struct amdgpu_device *)kgd;
2497
2498         config->gb_addr_config = adev->gfx.config.gb_addr_config;
2499         config->tile_config_ptr = adev->gfx.config.tile_mode_array;
2500         config->num_tile_configs =
2501                         ARRAY_SIZE(adev->gfx.config.tile_mode_array);
2502         config->macro_tile_config_ptr =
2503                         adev->gfx.config.macrotile_mode_array;
2504         config->num_macro_tile_configs =
2505                         ARRAY_SIZE(adev->gfx.config.macrotile_mode_array);
2506
2507         /* Those values are not set from GFX9 onwards */
2508         config->num_banks = adev->gfx.config.num_banks;
2509         config->num_ranks = adev->gfx.config.num_ranks;
2510
2511         return 0;
2512 }