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

   1 // SPDX-License-Identifier: MIT
   2 /*
   3  * Copyright 2014-2018 Advanced Micro Devices, Inc.
   4  *
   5  * Permission is hereby granted, free of charge, to any person obtaining a
   6  * copy of this software and associated documentation files (the "Software"),
   7  * to deal in the Software without restriction, including without limitation
   8  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
   9  * and/or sell copies of the Software, and to permit persons to whom the
  10  * Software is furnished to do so, subject to the following conditions:
  11  *
  12  * The above copyright notice and this permission notice shall be included in
  13  * all copies or substantial portions of the Software.
  14  *
  15  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  16  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  17  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
  18  * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR
  19  * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
  20  * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
  21  * OTHER DEALINGS IN THE SOFTWARE.
  22  */
  23 #include <linux/dma-buf.h>
  24 #include <linux/list.h>
  25 #include <linux/pagemap.h>
  26 #include <linux/sched/mm.h>
  27 #include <linux/sched/task.h>
  28 #include <drm/ttm/ttm_tt.h>
  29
  30 #include <drm/drm_exec.h>
  31
  32 #include "amdgpu_object.h"
  33 #include "amdgpu_gem.h"
  34 #include "amdgpu_vm.h"
  35 #include "amdgpu_hmm.h"
  36 #include "amdgpu_amdkfd.h"
  37 #include "amdgpu_dma_buf.h"
  38 #include <uapi/linux/kfd_ioctl.h>
  39 #include "amdgpu_xgmi.h"
  40 #include "kfd_priv.h"
  41 #include "kfd_smi_events.h"
  42 #include <drm/ttm/ttm_tt.h>
  43
  44 /* Userptr restore delay, just long enough to allow consecutive VM
  45  * changes to accumulate
  46  */
  47 #define AMDGPU_USERPTR_RESTORE_DELAY_MS 1
  48
  49 /*
  50  * Align VRAM availability to 2MB to avoid fragmentation caused by 4K allocations in the tail 2MB
  51  * BO chunk
  52  */
  53 #define VRAM_AVAILABLITY_ALIGN (1 << 21)
  54
  55 /* Impose limit on how much memory KFD can use */
  56 static struct {
  57         uint64_t max_system_mem_limit;
  58         uint64_t max_ttm_mem_limit;
  59         int64_t system_mem_used;
  60         int64_t ttm_mem_used;
  61         spinlock_t mem_limit_lock;
  62 } kfd_mem_limit;
  63
  64 static const char * const domain_bit_to_string[] = {
  65                 "CPU",
  66                 "GTT",
  67                 "VRAM",
  68                 "GDS",
  69                 "GWS",
  70                 "OA"
  71 };
  72
  73 #define domain_string(domain) domain_bit_to_string[ffs(domain)-1]
  74
  75 static void amdgpu_amdkfd_restore_userptr_worker(struct work_struct *work);
  76
  77 static bool kfd_mem_is_attached(struct amdgpu_vm *avm,
  78                 struct kgd_mem *mem)
  79 {
  80         struct kfd_mem_attachment *entry;
  81
  82         list_for_each_entry(entry, &mem->attachments, list)
  83                 if (entry->bo_va->base.vm == avm)
  84                         return true;
  85
  86         return false;
  87 }
  88
  89 /**
  90  * reuse_dmamap() - Check whether adev can share the original
  91  * userptr BO
  92  *
  93  * If both adev and bo_adev are in direct mapping or
  94  * in the same iommu group, they can share the original BO.
  95  *
  96  * @adev: Device to which can or cannot share the original BO
  97  * @bo_adev: Device to which allocated BO belongs to
  98  *
  99  * Return: returns true if adev can share original userptr BO,
 100  * false otherwise.
 101  */
 102 static bool reuse_dmamap(struct amdgpu_device *adev, struct amdgpu_device *bo_adev)
 103 {
 104         return (adev->ram_is_direct_mapped && bo_adev->ram_is_direct_mapped) ||
 105                         (adev->dev->iommu_group == bo_adev->dev->iommu_group);
 106 }
 107
 108 /* Set memory usage limits. Current, limits are
 109  *  System (TTM + userptr) memory - 15/16th System RAM
 110  *  TTM memory - 3/8th System RAM
 111  */
 112 void amdgpu_amdkfd_gpuvm_init_mem_limits(void)
 113 {
 114         struct sysinfo si;
 115         uint64_t mem;
 116
 117         if (kfd_mem_limit.max_system_mem_limit)
 118                 return;
 119
 120         si_meminfo(&si);
 121         mem = si.freeram - si.freehigh;
 122         mem *= si.mem_unit;
 123
 124         spin_lock_init(&kfd_mem_limit.mem_limit_lock);
 125         kfd_mem_limit.max_system_mem_limit = mem - (mem >> 4);
 126         kfd_mem_limit.max_ttm_mem_limit = ttm_tt_pages_limit() << PAGE_SHIFT;
 127         pr_debug("Kernel memory limit %lluM, TTM limit %lluM\n",
 128                 (kfd_mem_limit.max_system_mem_limit >> 20),
 129                 (kfd_mem_limit.max_ttm_mem_limit >> 20));
 130 }
 131
 132 void amdgpu_amdkfd_reserve_system_mem(uint64_t size)
 133 {
 134         kfd_mem_limit.system_mem_used += size;
 135 }
 136
 137 /* Estimate page table size needed to represent a given memory size
 138  *
 139  * With 4KB pages, we need one 8 byte PTE for each 4KB of memory
 140  * (factor 512, >> 9). With 2MB pages, we need one 8 byte PTE for 2MB
 141  * of memory (factor 256K, >> 18). ROCm user mode tries to optimize
 142  * for 2MB pages for TLB efficiency. However, small allocations and
 143  * fragmented system memory still need some 4KB pages. We choose a
 144  * compromise that should work in most cases without reserving too
 145  * much memory for page tables unnecessarily (factor 16K, >> 14).
 146  */
 147
 148 #define ESTIMATE_PT_SIZE(mem_size) max(((mem_size) >> 14), AMDGPU_VM_RESERVED_VRAM)
 149
 150 /**
 151  * amdgpu_amdkfd_reserve_mem_limit() - Decrease available memory by size
 152  * of buffer.
 153  *
 154  * @adev: Device to which allocated BO belongs to
 155  * @size: Size of buffer, in bytes, encapsulated by B0. This should be
 156  * equivalent to amdgpu_bo_size(BO)
 157  * @alloc_flag: Flag used in allocating a BO as noted above
 158  * @xcp_id: xcp_id is used to get xcp from xcp manager, one xcp is
 159  * managed as one compute node in driver for app
 160  *
 161  * Return:
 162  *      returns -ENOMEM in case of error, ZERO otherwise
 163  */
 164 int amdgpu_amdkfd_reserve_mem_limit(struct amdgpu_device *adev,
 165                 uint64_t size, u32 alloc_flag, int8_t xcp_id)
 166 {
 167         uint64_t reserved_for_pt =
 168                 ESTIMATE_PT_SIZE(amdgpu_amdkfd_total_mem_size);
 169         size_t system_mem_needed, ttm_mem_needed, vram_needed;
 170         int ret = 0;
 171         uint64_t vram_size = 0;
 172
 173         system_mem_needed = 0;
 174         ttm_mem_needed = 0;
 175         vram_needed = 0;
 176         if (alloc_flag & KFD_IOC_ALLOC_MEM_FLAGS_GTT) {
 177                 system_mem_needed = size;
 178                 ttm_mem_needed = size;
 179         } else if (alloc_flag & KFD_IOC_ALLOC_MEM_FLAGS_VRAM) {
 180                 /*
 181                  * Conservatively round up the allocation requirement to 2 MB
 182                  * to avoid fragmentation caused by 4K allocations in the tail
 183                  * 2M BO chunk.
 184                  */
 185                 vram_needed = size;
 186                 /*
 187                  * For GFX 9.4.3, get the VRAM size from XCP structs
 188                  */
 189                 if (WARN_ONCE(xcp_id < 0, "invalid XCP ID %d", xcp_id))
 190                         return -EINVAL;
 191
 192                 vram_size = KFD_XCP_MEMORY_SIZE(adev, xcp_id);
 193                 if (adev->gmc.is_app_apu) {
 194                         system_mem_needed = size;
 195                         ttm_mem_needed = size;
 196                 }
 197         } else if (alloc_flag & KFD_IOC_ALLOC_MEM_FLAGS_USERPTR) {
 198                 system_mem_needed = size;
 199         } else if (!(alloc_flag &
 200                                 (KFD_IOC_ALLOC_MEM_FLAGS_DOORBELL |
 201                                  KFD_IOC_ALLOC_MEM_FLAGS_MMIO_REMAP))) {
 202                 pr_err("%s: Invalid BO type %#x\n", __func__, alloc_flag);
 203                 return -ENOMEM;
 204         }
 205
 206         spin_lock(&kfd_mem_limit.mem_limit_lock);
 207
 208         if (kfd_mem_limit.system_mem_used + system_mem_needed >
 209             kfd_mem_limit.max_system_mem_limit)
 210                 pr_debug("Set no_system_mem_limit=1 if using shared memory\n");
 211
 212         if ((kfd_mem_limit.system_mem_used + system_mem_needed >
 213              kfd_mem_limit.max_system_mem_limit && !no_system_mem_limit) ||
 214             (kfd_mem_limit.ttm_mem_used + ttm_mem_needed >
 215              kfd_mem_limit.max_ttm_mem_limit) ||
 216             (adev && xcp_id >= 0 && adev->kfd.vram_used[xcp_id] + vram_needed >
 217              vram_size - reserved_for_pt)) {
 218                 ret = -ENOMEM;
 219                 goto release;
 220         }
 221
 222         /* Update memory accounting by decreasing available system
 223          * memory, TTM memory and GPU memory as computed above
 224          */
 225         WARN_ONCE(vram_needed && !adev,
 226                   "adev reference can't be null when vram is used");
 227         if (adev && xcp_id >= 0) {
 228                 adev->kfd.vram_used[xcp_id] += vram_needed;
 229                 adev->kfd.vram_used_aligned[xcp_id] += adev->gmc.is_app_apu ?
 230                                 vram_needed :
 231                                 ALIGN(vram_needed, VRAM_AVAILABLITY_ALIGN);
 232         }
 233         kfd_mem_limit.system_mem_used += system_mem_needed;
 234         kfd_mem_limit.ttm_mem_used += ttm_mem_needed;
 235
 236 release:
 237         spin_unlock(&kfd_mem_limit.mem_limit_lock);
 238         return ret;
 239 }
 240
 241 void amdgpu_amdkfd_unreserve_mem_limit(struct amdgpu_device *adev,
 242                 uint64_t size, u32 alloc_flag, int8_t xcp_id)
 243 {
 244         spin_lock(&kfd_mem_limit.mem_limit_lock);
 245
 246         if (alloc_flag & KFD_IOC_ALLOC_MEM_FLAGS_GTT) {
 247                 kfd_mem_limit.system_mem_used -= size;
 248                 kfd_mem_limit.ttm_mem_used -= size;
 249         } else if (alloc_flag & KFD_IOC_ALLOC_MEM_FLAGS_VRAM) {
 250                 WARN_ONCE(!adev,
 251                           "adev reference can't be null when alloc mem flags vram is set");
 252                 if (WARN_ONCE(xcp_id < 0, "invalid XCP ID %d", xcp_id))
 253                         goto release;
 254
 255                 if (adev) {
 256                         adev->kfd.vram_used[xcp_id] -= size;
 257                         if (adev->gmc.is_app_apu) {
 258                                 adev->kfd.vram_used_aligned[xcp_id] -= size;
 259                                 kfd_mem_limit.system_mem_used -= size;
 260                                 kfd_mem_limit.ttm_mem_used -= size;
 261                         } else {
 262                                 adev->kfd.vram_used_aligned[xcp_id] -=
 263                                         ALIGN(size, VRAM_AVAILABLITY_ALIGN);
 264                         }
 265                 }
 266         } else if (alloc_flag & KFD_IOC_ALLOC_MEM_FLAGS_USERPTR) {
 267                 kfd_mem_limit.system_mem_used -= size;
 268         } else if (!(alloc_flag &
 269                                 (KFD_IOC_ALLOC_MEM_FLAGS_DOORBELL |
 270                                  KFD_IOC_ALLOC_MEM_FLAGS_MMIO_REMAP))) {
 271                 pr_err("%s: Invalid BO type %#x\n", __func__, alloc_flag);
 272                 goto release;
 273         }
 274         WARN_ONCE(adev && xcp_id >= 0 && adev->kfd.vram_used[xcp_id] < 0,
 275                   "KFD VRAM memory accounting unbalanced for xcp: %d", xcp_id);
 276         WARN_ONCE(kfd_mem_limit.ttm_mem_used < 0,
 277                   "KFD TTM memory accounting unbalanced");
 278         WARN_ONCE(kfd_mem_limit.system_mem_used < 0,
 279                   "KFD system memory accounting unbalanced");
 280
 281 release:
 282         spin_unlock(&kfd_mem_limit.mem_limit_lock);
 283 }
 284
 285 void amdgpu_amdkfd_release_notify(struct amdgpu_bo *bo)
 286 {
 287         struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
 288         u32 alloc_flags = bo->kfd_bo->alloc_flags;
 289         u64 size = amdgpu_bo_size(bo);
 290
 291         amdgpu_amdkfd_unreserve_mem_limit(adev, size, alloc_flags,
 292                                           bo->xcp_id);
 293
 294         kfree(bo->kfd_bo);
 295 }
 296
 297 /**
 298  * create_dmamap_sg_bo() - Creates a amdgpu_bo object to reflect information
 299  * about USERPTR or DOOREBELL or MMIO BO.
 300  *
 301  * @adev: Device for which dmamap BO is being created
 302  * @mem: BO of peer device that is being DMA mapped. Provides parameters
 303  *       in building the dmamap BO
 304  * @bo_out: Output parameter updated with handle of dmamap BO
 305  */
 306 static int
 307 create_dmamap_sg_bo(struct amdgpu_device *adev,
 308                  struct kgd_mem *mem, struct amdgpu_bo **bo_out)
 309 {
 310         struct drm_gem_object *gem_obj;
 311         int ret;
 312         uint64_t flags = 0;
 313
 314         ret = amdgpu_bo_reserve(mem->bo, false);
 315         if (ret)
 316                 return ret;
 317
 318         if (mem->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_USERPTR)
 319                 flags |= mem->bo->flags & (AMDGPU_GEM_CREATE_COHERENT |
 320                                         AMDGPU_GEM_CREATE_UNCACHED);
 321
 322         ret = amdgpu_gem_object_create(adev, mem->bo->tbo.base.size, 1,
 323                         AMDGPU_GEM_DOMAIN_CPU, AMDGPU_GEM_CREATE_PREEMPTIBLE | flags,
 324                         ttm_bo_type_sg, mem->bo->tbo.base.resv, &gem_obj, 0);
 325
 326         amdgpu_bo_unreserve(mem->bo);
 327
 328         if (ret) {
 329                 pr_err("Error in creating DMA mappable SG BO on domain: %d\n", ret);
 330                 return -EINVAL;
 331         }
 332
 333         *bo_out = gem_to_amdgpu_bo(gem_obj);
 334         (*bo_out)->parent = amdgpu_bo_ref(mem->bo);
 335         return ret;
 336 }
 337
 338 /* amdgpu_amdkfd_remove_eviction_fence - Removes eviction fence from BO's
 339  *  reservation object.
 340  *
 341  * @bo: [IN] Remove eviction fence(s) from this BO
 342  * @ef: [IN] This eviction fence is removed if it
 343  *  is present in the shared list.
 344  *
 345  * NOTE: Must be called with BO reserved i.e. bo->tbo.resv->lock held.
 346  */
 347 static int amdgpu_amdkfd_remove_eviction_fence(struct amdgpu_bo *bo,
 348                                         struct amdgpu_amdkfd_fence *ef)
 349 {
 350         struct dma_fence *replacement;
 351
 352         if (!ef)
 353                 return -EINVAL;
 354
 355         /* TODO: Instead of block before we should use the fence of the page
 356          * table update and TLB flush here directly.
 357          */
 358         replacement = dma_fence_get_stub();
 359         dma_resv_replace_fences(bo->tbo.base.resv, ef->base.context,
 360                                 replacement, DMA_RESV_USAGE_BOOKKEEP);
 361         dma_fence_put(replacement);
 362         return 0;
 363 }
 364
 365 int amdgpu_amdkfd_remove_fence_on_pt_pd_bos(struct amdgpu_bo *bo)
 366 {
 367         struct amdgpu_bo *root = bo;
 368         struct amdgpu_vm_bo_base *vm_bo;
 369         struct amdgpu_vm *vm;
 370         struct amdkfd_process_info *info;
 371         struct amdgpu_amdkfd_fence *ef;
 372         int ret;
 373
 374         /* we can always get vm_bo from root PD bo.*/
 375         while (root->parent)
 376                 root = root->parent;
 377
 378         vm_bo = root->vm_bo;
 379         if (!vm_bo)
 380                 return 0;
 381
 382         vm = vm_bo->vm;
 383         if (!vm)
 384                 return 0;
 385
 386         info = vm->process_info;
 387         if (!info || !info->eviction_fence)
 388                 return 0;
 389
 390         ef = container_of(dma_fence_get(&info->eviction_fence->base),
 391                         struct amdgpu_amdkfd_fence, base);
 392
 393         BUG_ON(!dma_resv_trylock(bo->tbo.base.resv));
 394         ret = amdgpu_amdkfd_remove_eviction_fence(bo, ef);
 395         dma_resv_unlock(bo->tbo.base.resv);
 396
 397         dma_fence_put(&ef->base);
 398         return ret;
 399 }
 400
 401 static int amdgpu_amdkfd_bo_validate(struct amdgpu_bo *bo, uint32_t domain,
 402                                      bool wait)
 403 {
 404         struct ttm_operation_ctx ctx = { false, false };
 405         int ret;
 406
 407         if (WARN(amdgpu_ttm_tt_get_usermm(bo->tbo.ttm),
 408                  "Called with userptr BO"))
 409                 return -EINVAL;
 410
 411         amdgpu_bo_placement_from_domain(bo, domain);
 412
 413         ret = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
 414         if (ret)
 415                 goto validate_fail;
 416         if (wait)
 417                 amdgpu_bo_sync_wait(bo, AMDGPU_FENCE_OWNER_KFD, false);
 418
 419 validate_fail:
 420         return ret;
 421 }
 422
 423 static int amdgpu_amdkfd_validate_vm_bo(void *_unused, struct amdgpu_bo *bo)
 424 {
 425         return amdgpu_amdkfd_bo_validate(bo, bo->allowed_domains, false);
 426 }
 427
 428 /* vm_validate_pt_pd_bos - Validate page table and directory BOs
 429  *
 430  * Page directories are not updated here because huge page handling
 431  * during page table updates can invalidate page directory entries
 432  * again. Page directories are only updated after updating page
 433  * tables.
 434  */
 435 static int vm_validate_pt_pd_bos(struct amdgpu_vm *vm)
 436 {
 437         struct amdgpu_bo *pd = vm->root.bo;
 438         struct amdgpu_device *adev = amdgpu_ttm_adev(pd->tbo.bdev);
 439         int ret;
 440
 441         ret = amdgpu_vm_validate_pt_bos(adev, vm, amdgpu_amdkfd_validate_vm_bo, NULL);
 442         if (ret) {
 443                 pr_err("failed to validate PT BOs\n");
 444                 return ret;
 445         }
 446
 447         vm->pd_phys_addr = amdgpu_gmc_pd_addr(vm->root.bo);
 448
 449         return 0;
 450 }
 451
 452 static int vm_update_pds(struct amdgpu_vm *vm, struct amdgpu_sync *sync)
 453 {
 454         struct amdgpu_bo *pd = vm->root.bo;
 455         struct amdgpu_device *adev = amdgpu_ttm_adev(pd->tbo.bdev);
 456         int ret;
 457
 458         ret = amdgpu_vm_update_pdes(adev, vm, false);
 459         if (ret)
 460                 return ret;
 461
 462         return amdgpu_sync_fence(sync, vm->last_update);
 463 }
 464
 465 static uint64_t get_pte_flags(struct amdgpu_device *adev, struct kgd_mem *mem)
 466 {
 467         uint32_t mapping_flags = AMDGPU_VM_PAGE_READABLE |
 468                                  AMDGPU_VM_MTYPE_DEFAULT;
 469
 470         if (mem->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_WRITABLE)
 471                 mapping_flags |= AMDGPU_VM_PAGE_WRITEABLE;
 472         if (mem->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_EXECUTABLE)
 473                 mapping_flags |= AMDGPU_VM_PAGE_EXECUTABLE;
 474
 475         return amdgpu_gem_va_map_flags(adev, mapping_flags);
 476 }
 477
 478 /**
 479  * create_sg_table() - Create an sg_table for a contiguous DMA addr range
 480  * @addr: The starting address to point to
 481  * @size: Size of memory area in bytes being pointed to
 482  *
 483  * Allocates an instance of sg_table and initializes it to point to memory
 484  * area specified by input parameters. The address used to build is assumed
 485  * to be DMA mapped, if needed.
 486  *
 487  * DOORBELL or MMIO BOs use only one scatterlist node in their sg_table
 488  * because they are physically contiguous.
 489  *
 490  * Return: Initialized instance of SG Table or NULL
 491  */
 492 static struct sg_table *create_sg_table(uint64_t addr, uint32_t size)
 493 {
 494         struct sg_table *sg = kmalloc(sizeof(*sg), GFP_KERNEL);
 495
 496         if (!sg)
 497                 return NULL;
 498         if (sg_alloc_table(sg, 1, GFP_KERNEL)) {
 499                 kfree(sg);
 500                 return NULL;
 501         }
 502         sg_dma_address(sg->sgl) = addr;
 503         sg->sgl->length = size;
 504 #ifdef CONFIG_NEED_SG_DMA_LENGTH
 505         sg->sgl->dma_length = size;
 506 #endif
 507         return sg;
 508 }
 509
 510 static int
 511 kfd_mem_dmamap_userptr(struct kgd_mem *mem,
 512                        struct kfd_mem_attachment *attachment)
 513 {
 514         enum dma_data_direction direction =
 515                 mem->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_WRITABLE ?
 516                 DMA_BIDIRECTIONAL : DMA_TO_DEVICE;
 517         struct ttm_operation_ctx ctx = {.interruptible = true};
 518         struct amdgpu_bo *bo = attachment->bo_va->base.bo;
 519         struct amdgpu_device *adev = attachment->adev;
 520         struct ttm_tt *src_ttm = mem->bo->tbo.ttm;
 521         struct ttm_tt *ttm = bo->tbo.ttm;
 522         int ret;
 523
 524         if (WARN_ON(ttm->num_pages != src_ttm->num_pages))
 525                 return -EINVAL;
 526
 527         ttm->sg = kmalloc(sizeof(*ttm->sg), GFP_KERNEL);
 528         if (unlikely(!ttm->sg))
 529                 return -ENOMEM;
 530
 531         /* Same sequence as in amdgpu_ttm_tt_pin_userptr */
 532         ret = sg_alloc_table_from_pages(ttm->sg, src_ttm->pages,
 533                                         ttm->num_pages, 0,
 534                                         (u64)ttm->num_pages << PAGE_SHIFT,
 535                                         GFP_KERNEL);
 536         if (unlikely(ret))
 537                 goto free_sg;
 538
 539         ret = dma_map_sgtable(adev->dev, ttm->sg, direction, 0);
 540         if (unlikely(ret))
 541                 goto release_sg;
 542
 543         amdgpu_bo_placement_from_domain(bo, AMDGPU_GEM_DOMAIN_GTT);
 544         ret = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
 545         if (ret)
 546                 goto unmap_sg;
 547
 548         return 0;
 549
 550 unmap_sg:
 551         dma_unmap_sgtable(adev->dev, ttm->sg, direction, 0);
 552 release_sg:
 553         pr_err("DMA map userptr failed: %d\n", ret);
 554         sg_free_table(ttm->sg);
 555 free_sg:
 556         kfree(ttm->sg);
 557         ttm->sg = NULL;
 558         return ret;
 559 }
 560
 561 static int
 562 kfd_mem_dmamap_dmabuf(struct kfd_mem_attachment *attachment)
 563 {
 564         struct ttm_operation_ctx ctx = {.interruptible = true};
 565         struct amdgpu_bo *bo = attachment->bo_va->base.bo;
 566         int ret;
 567
 568         amdgpu_bo_placement_from_domain(bo, AMDGPU_GEM_DOMAIN_CPU);
 569         ret = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
 570         if (ret)
 571                 return ret;
 572
 573         amdgpu_bo_placement_from_domain(bo, AMDGPU_GEM_DOMAIN_GTT);
 574         return ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
 575 }
 576
 577 /**
 578  * kfd_mem_dmamap_sg_bo() - Create DMA mapped sg_table to access DOORBELL or MMIO BO
 579  * @mem: SG BO of the DOORBELL or MMIO resource on the owning device
 580  * @attachment: Virtual address attachment of the BO on accessing device
 581  *
 582  * An access request from the device that owns DOORBELL does not require DMA mapping.
 583  * This is because the request doesn't go through PCIe root complex i.e. it instead
 584  * loops back. The need to DMA map arises only when accessing peer device's DOORBELL
 585  *
 586  * In contrast, all access requests for MMIO need to be DMA mapped without regard to
 587  * device ownership. This is because access requests for MMIO go through PCIe root
 588  * complex.
 589  *
 590  * This is accomplished in two steps:
 591  *   - Obtain DMA mapped address of DOORBELL or MMIO memory that could be used
 592  *         in updating requesting device's page table
 593  *   - Signal TTM to mark memory pointed to by requesting device's BO as GPU
 594  *         accessible. This allows an update of requesting device's page table
 595  *         with entries associated with DOOREBELL or MMIO memory
 596  *
 597  * This method is invoked in the following contexts:
 598  *   - Mapping of DOORBELL or MMIO BO of same or peer device
 599  *   - Validating an evicted DOOREBELL or MMIO BO on device seeking access
 600  *
 601  * Return: ZERO if successful, NON-ZERO otherwise
 602  */
 603 static int
 604 kfd_mem_dmamap_sg_bo(struct kgd_mem *mem,
 605                      struct kfd_mem_attachment *attachment)
 606 {
 607         struct ttm_operation_ctx ctx = {.interruptible = true};
 608         struct amdgpu_bo *bo = attachment->bo_va->base.bo;
 609         struct amdgpu_device *adev = attachment->adev;
 610         struct ttm_tt *ttm = bo->tbo.ttm;
 611         enum dma_data_direction dir;
 612         dma_addr_t dma_addr;
 613         bool mmio;
 614         int ret;
 615
 616         /* Expect SG Table of dmapmap BO to be NULL */
 617         mmio = (mem->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_MMIO_REMAP);
 618         if (unlikely(ttm->sg)) {
 619                 pr_err("SG Table of %d BO for peer device is UNEXPECTEDLY NON-NULL", mmio);
 620                 return -EINVAL;
 621         }
 622
 623         dir = mem->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_WRITABLE ?
 624                         DMA_BIDIRECTIONAL : DMA_TO_DEVICE;
 625         dma_addr = mem->bo->tbo.sg->sgl->dma_address;
 626         pr_debug("%d BO size: %d\n", mmio, mem->bo->tbo.sg->sgl->length);
 627         pr_debug("%d BO address before DMA mapping: %llx\n", mmio, dma_addr);
 628         dma_addr = dma_map_resource(adev->dev, dma_addr,
 629                         mem->bo->tbo.sg->sgl->length, dir, DMA_ATTR_SKIP_CPU_SYNC);
 630         ret = dma_mapping_error(adev->dev, dma_addr);
 631         if (unlikely(ret))
 632                 return ret;
 633         pr_debug("%d BO address after DMA mapping: %llx\n", mmio, dma_addr);
 634
 635         ttm->sg = create_sg_table(dma_addr, mem->bo->tbo.sg->sgl->length);
 636         if (unlikely(!ttm->sg)) {
 637                 ret = -ENOMEM;
 638                 goto unmap_sg;
 639         }
 640
 641         amdgpu_bo_placement_from_domain(bo, AMDGPU_GEM_DOMAIN_GTT);
 642         ret = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
 643         if (unlikely(ret))
 644                 goto free_sg;
 645
 646         return ret;
 647
 648 free_sg:
 649         sg_free_table(ttm->sg);
 650         kfree(ttm->sg);
 651         ttm->sg = NULL;
 652 unmap_sg:
 653         dma_unmap_resource(adev->dev, dma_addr, mem->bo->tbo.sg->sgl->length,
 654                            dir, DMA_ATTR_SKIP_CPU_SYNC);
 655         return ret;
 656 }
 657
 658 static int
 659 kfd_mem_dmamap_attachment(struct kgd_mem *mem,
 660                           struct kfd_mem_attachment *attachment)
 661 {
 662         switch (attachment->type) {
 663         case KFD_MEM_ATT_SHARED:
 664                 return 0;
 665         case KFD_MEM_ATT_USERPTR:
 666                 return kfd_mem_dmamap_userptr(mem, attachment);
 667         case KFD_MEM_ATT_DMABUF:
 668                 return kfd_mem_dmamap_dmabuf(attachment);
 669         case KFD_MEM_ATT_SG:
 670                 return kfd_mem_dmamap_sg_bo(mem, attachment);
 671         default:
 672                 WARN_ON_ONCE(1);
 673         }
 674         return -EINVAL;
 675 }
 676
 677 static void
 678 kfd_mem_dmaunmap_userptr(struct kgd_mem *mem,
 679                          struct kfd_mem_attachment *attachment)
 680 {
 681         enum dma_data_direction direction =
 682                 mem->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_WRITABLE ?
 683                 DMA_BIDIRECTIONAL : DMA_TO_DEVICE;
 684         struct ttm_operation_ctx ctx = {.interruptible = false};
 685         struct amdgpu_bo *bo = attachment->bo_va->base.bo;
 686         struct amdgpu_device *adev = attachment->adev;
 687         struct ttm_tt *ttm = bo->tbo.ttm;
 688
 689         if (unlikely(!ttm->sg))
 690                 return;
 691
 692         amdgpu_bo_placement_from_domain(bo, AMDGPU_GEM_DOMAIN_CPU);
 693         ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
 694
 695         dma_unmap_sgtable(adev->dev, ttm->sg, direction, 0);
 696         sg_free_table(ttm->sg);
 697         kfree(ttm->sg);
 698         ttm->sg = NULL;
 699 }
 700
 701 static void
 702 kfd_mem_dmaunmap_dmabuf(struct kfd_mem_attachment *attachment)
 703 {
 704         /* This is a no-op. We don't want to trigger eviction fences when
 705          * unmapping DMABufs. Therefore the invalidation (moving to system
 706          * domain) is done in kfd_mem_dmamap_dmabuf.
 707          */
 708 }
 709
 710 /**
 711  * kfd_mem_dmaunmap_sg_bo() - Free DMA mapped sg_table of DOORBELL or MMIO BO
 712  * @mem: SG BO of the DOORBELL or MMIO resource on the owning device
 713  * @attachment: Virtual address attachment of the BO on accessing device
 714  *
 715  * The method performs following steps:
 716  *   - Signal TTM to mark memory pointed to by BO as GPU inaccessible
 717  *   - Free SG Table that is used to encapsulate DMA mapped memory of
 718  *          peer device's DOORBELL or MMIO memory
 719  *
 720  * This method is invoked in the following contexts:
 721  *     UNMapping of DOORBELL or MMIO BO on a device having access to its memory
 722  *     Eviction of DOOREBELL or MMIO BO on device having access to its memory
 723  *
 724  * Return: void
 725  */
 726 static void
 727 kfd_mem_dmaunmap_sg_bo(struct kgd_mem *mem,
 728                        struct kfd_mem_attachment *attachment)
 729 {
 730         struct ttm_operation_ctx ctx = {.interruptible = true};
 731         struct amdgpu_bo *bo = attachment->bo_va->base.bo;
 732         struct amdgpu_device *adev = attachment->adev;
 733         struct ttm_tt *ttm = bo->tbo.ttm;
 734         enum dma_data_direction dir;
 735
 736         if (unlikely(!ttm->sg)) {
 737                 pr_err("SG Table of BO is UNEXPECTEDLY NULL");
 738                 return;
 739         }
 740
 741         amdgpu_bo_placement_from_domain(bo, AMDGPU_GEM_DOMAIN_CPU);
 742         ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
 743
 744         dir = mem->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_WRITABLE ?
 745                                 DMA_BIDIRECTIONAL : DMA_TO_DEVICE;
 746         dma_unmap_resource(adev->dev, ttm->sg->sgl->dma_address,
 747                         ttm->sg->sgl->length, dir, DMA_ATTR_SKIP_CPU_SYNC);
 748         sg_free_table(ttm->sg);
 749         kfree(ttm->sg);
 750         ttm->sg = NULL;
 751         bo->tbo.sg = NULL;
 752 }
 753
 754 static void
 755 kfd_mem_dmaunmap_attachment(struct kgd_mem *mem,
 756                             struct kfd_mem_attachment *attachment)
 757 {
 758         switch (attachment->type) {
 759         case KFD_MEM_ATT_SHARED:
 760                 break;
 761         case KFD_MEM_ATT_USERPTR:
 762                 kfd_mem_dmaunmap_userptr(mem, attachment);
 763                 break;
 764         case KFD_MEM_ATT_DMABUF:
 765                 kfd_mem_dmaunmap_dmabuf(attachment);
 766                 break;
 767         case KFD_MEM_ATT_SG:
 768                 kfd_mem_dmaunmap_sg_bo(mem, attachment);
 769                 break;
 770         default:
 771                 WARN_ON_ONCE(1);
 772         }
 773 }
 774
 775 static int kfd_mem_export_dmabuf(struct kgd_mem *mem)
 776 {
 777         if (!mem->dmabuf) {
 778                 struct dma_buf *ret = amdgpu_gem_prime_export(
 779                         &mem->bo->tbo.base,
 780                         mem->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_WRITABLE ?
 781                                 DRM_RDWR : 0);
 782                 if (IS_ERR(ret))
 783                         return PTR_ERR(ret);
 784                 mem->dmabuf = ret;
 785         }
 786
 787         return 0;
 788 }
 789
 790 static int
 791 kfd_mem_attach_dmabuf(struct amdgpu_device *adev, struct kgd_mem *mem,
 792                       struct amdgpu_bo **bo)
 793 {
 794         struct drm_gem_object *gobj;
 795         int ret;
 796
 797         ret = kfd_mem_export_dmabuf(mem);
 798         if (ret)
 799                 return ret;
 800
 801         gobj = amdgpu_gem_prime_import(adev_to_drm(adev), mem->dmabuf);
 802         if (IS_ERR(gobj))
 803                 return PTR_ERR(gobj);
 804
 805         *bo = gem_to_amdgpu_bo(gobj);
 806         (*bo)->flags |= AMDGPU_GEM_CREATE_PREEMPTIBLE;
 807
 808         return 0;
 809 }
 810
 811 /* kfd_mem_attach - Add a BO to a VM
 812  *
 813  * Everything that needs to bo done only once when a BO is first added
 814  * to a VM. It can later be mapped and unmapped many times without
 815  * repeating these steps.
 816  *
 817  * 0. Create BO for DMA mapping, if needed
 818  * 1. Allocate and initialize BO VA entry data structure
 819  * 2. Add BO to the VM
 820  * 3. Determine ASIC-specific PTE flags
 821  * 4. Alloc page tables and directories if needed
 822  * 4a.  Validate new page tables and directories
 823  */
 824 static int kfd_mem_attach(struct amdgpu_device *adev, struct kgd_mem *mem,
 825                 struct amdgpu_vm *vm, bool is_aql)
 826 {
 827         struct amdgpu_device *bo_adev = amdgpu_ttm_adev(mem->bo->tbo.bdev);
 828         unsigned long bo_size = mem->bo->tbo.base.size;
 829         uint64_t va = mem->va;
 830         struct kfd_mem_attachment *attachment[2] = {NULL, NULL};
 831         struct amdgpu_bo *bo[2] = {NULL, NULL};
 832         bool same_hive = false;
 833         int i, ret;
 834
 835         if (!va) {
 836                 pr_err("Invalid VA when adding BO to VM\n");
 837                 return -EINVAL;
 838         }
 839
 840         /* Determine access to VRAM, MMIO and DOORBELL BOs of peer devices
 841          *
 842          * The access path of MMIO and DOORBELL BOs of is always over PCIe.
 843          * In contrast the access path of VRAM BOs depens upon the type of
 844          * link that connects the peer device. Access over PCIe is allowed
 845          * if peer device has large BAR. In contrast, access over xGMI is
 846          * allowed for both small and large BAR configurations of peer device
 847          */
 848         if ((adev != bo_adev && !adev->gmc.is_app_apu) &&
 849             ((mem->domain == AMDGPU_GEM_DOMAIN_VRAM) ||
 850              (mem->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_DOORBELL) ||
 851              (mem->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_MMIO_REMAP))) {
 852                 if (mem->domain == AMDGPU_GEM_DOMAIN_VRAM)
 853                         same_hive = amdgpu_xgmi_same_hive(adev, bo_adev);
 854                 if (!same_hive && !amdgpu_device_is_peer_accessible(bo_adev, adev))
 855                         return -EINVAL;
 856         }
 857
 858         for (i = 0; i <= is_aql; i++) {
 859                 attachment[i] = kzalloc(sizeof(*attachment[i]), GFP_KERNEL);
 860                 if (unlikely(!attachment[i])) {
 861                         ret = -ENOMEM;
 862                         goto unwind;
 863                 }
 864
 865                 pr_debug("\t add VA 0x%llx - 0x%llx to vm %p\n", va,
 866                          va + bo_size, vm);
 867
 868                 if ((adev == bo_adev && !(mem->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_MMIO_REMAP)) ||
 869                     (amdgpu_ttm_tt_get_usermm(mem->bo->tbo.ttm) && reuse_dmamap(adev, bo_adev)) ||
 870                         same_hive) {
 871                         /* Mappings on the local GPU, or VRAM mappings in the
 872                          * local hive, or userptr mapping can reuse dma map
 873                          * address space share the original BO
 874                          */
 875                         attachment[i]->type = KFD_MEM_ATT_SHARED;
 876                         bo[i] = mem->bo;
 877                         drm_gem_object_get(&bo[i]->tbo.base);
 878                 } else if (i > 0) {
 879                         /* Multiple mappings on the same GPU share the BO */
 880                         attachment[i]->type = KFD_MEM_ATT_SHARED;
 881                         bo[i] = bo[0];
 882                         drm_gem_object_get(&bo[i]->tbo.base);
 883                 } else if (amdgpu_ttm_tt_get_usermm(mem->bo->tbo.ttm)) {
 884                         /* Create an SG BO to DMA-map userptrs on other GPUs */
 885                         attachment[i]->type = KFD_MEM_ATT_USERPTR;
 886                         ret = create_dmamap_sg_bo(adev, mem, &bo[i]);
 887                         if (ret)
 888                                 goto unwind;
 889                 /* Handle DOORBELL BOs of peer devices and MMIO BOs of local and peer devices */
 890                 } else if (mem->bo->tbo.type == ttm_bo_type_sg) {
 891                         WARN_ONCE(!(mem->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_DOORBELL ||
 892                                     mem->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_MMIO_REMAP),
 893                                   "Handing invalid SG BO in ATTACH request");
 894                         attachment[i]->type = KFD_MEM_ATT_SG;
 895                         ret = create_dmamap_sg_bo(adev, mem, &bo[i]);
 896                         if (ret)
 897                                 goto unwind;
 898                 /* Enable acces to GTT and VRAM BOs of peer devices */
 899                 } else if (mem->domain == AMDGPU_GEM_DOMAIN_GTT ||
 900                            mem->domain == AMDGPU_GEM_DOMAIN_VRAM) {
 901                         attachment[i]->type = KFD_MEM_ATT_DMABUF;
 902                         ret = kfd_mem_attach_dmabuf(adev, mem, &bo[i]);
 903                         if (ret)
 904                                 goto unwind;
 905                         pr_debug("Employ DMABUF mechanism to enable peer GPU access\n");
 906                 } else {
 907                         WARN_ONCE(true, "Handling invalid ATTACH request");
 908                         ret = -EINVAL;
 909                         goto unwind;
 910                 }
 911
 912                 /* Add BO to VM internal data structures */
 913                 ret = amdgpu_bo_reserve(bo[i], false);
 914                 if (ret) {
 915                         pr_debug("Unable to reserve BO during memory attach");
 916                         goto unwind;
 917                 }
 918                 attachment[i]->bo_va = amdgpu_vm_bo_add(adev, vm, bo[i]);
 919                 amdgpu_bo_unreserve(bo[i]);
 920                 if (unlikely(!attachment[i]->bo_va)) {
 921                         ret = -ENOMEM;
 922                         pr_err("Failed to add BO object to VM. ret == %d\n",
 923                                ret);
 924                         goto unwind;
 925                 }
 926                 attachment[i]->va = va;
 927                 attachment[i]->pte_flags = get_pte_flags(adev, mem);
 928                 attachment[i]->adev = adev;
 929                 list_add(&attachment[i]->list, &mem->attachments);
 930
 931                 va += bo_size;
 932         }
 933
 934         return 0;
 935
 936 unwind:
 937         for (; i >= 0; i--) {
 938                 if (!attachment[i])
 939                         continue;
 940                 if (attachment[i]->bo_va) {
 941                         amdgpu_bo_reserve(bo[i], true);
 942                         amdgpu_vm_bo_del(adev, attachment[i]->bo_va);
 943                         amdgpu_bo_unreserve(bo[i]);
 944                         list_del(&attachment[i]->list);
 945                 }
 946                 if (bo[i])
 947                         drm_gem_object_put(&bo[i]->tbo.base);
 948                 kfree(attachment[i]);
 949         }
 950         return ret;
 951 }
 952
 953 static void kfd_mem_detach(struct kfd_mem_attachment *attachment)
 954 {
 955         struct amdgpu_bo *bo = attachment->bo_va->base.bo;
 956
 957         pr_debug("\t remove VA 0x%llx in entry %p\n",
 958                         attachment->va, attachment);
 959         amdgpu_vm_bo_del(attachment->adev, attachment->bo_va);
 960         drm_gem_object_put(&bo->tbo.base);
 961         list_del(&attachment->list);
 962         kfree(attachment);
 963 }
 964
 965 static void add_kgd_mem_to_kfd_bo_list(struct kgd_mem *mem,
 966                                 struct amdkfd_process_info *process_info,
 967                                 bool userptr)
 968 {
 969         mutex_lock(&process_info->lock);
 970         if (userptr)
 971                 list_add_tail(&mem->validate_list,
 972                               &process_info->userptr_valid_list);
 973         else
 974                 list_add_tail(&mem->validate_list, &process_info->kfd_bo_list);
 975         mutex_unlock(&process_info->lock);
 976 }
 977
 978 static void remove_kgd_mem_from_kfd_bo_list(struct kgd_mem *mem,
 979                 struct amdkfd_process_info *process_info)
 980 {
 981         mutex_lock(&process_info->lock);
 982         list_del(&mem->validate_list);
 983         mutex_unlock(&process_info->lock);
 984 }
 985
 986 /* Initializes user pages. It registers the MMU notifier and validates
 987  * the userptr BO in the GTT domain.
 988  *
 989  * The BO must already be on the userptr_valid_list. Otherwise an
 990  * eviction and restore may happen that leaves the new BO unmapped
 991  * with the user mode queues running.
 992  *
 993  * Takes the process_info->lock to protect against concurrent restore
 994  * workers.
 995  *
 996  * Returns 0 for success, negative errno for errors.
 997  */
 998 static int init_user_pages(struct kgd_mem *mem, uint64_t user_addr,
 999                            bool criu_resume)
1000 {
1001         struct amdkfd_process_info *process_info = mem->process_info;
1002         struct amdgpu_bo *bo = mem->bo;
1003         struct ttm_operation_ctx ctx = { true, false };
1004         struct hmm_range *range;
1005         int ret = 0;
1006
1007         mutex_lock(&process_info->lock);
1008
1009         ret = amdgpu_ttm_tt_set_userptr(&bo->tbo, user_addr, 0);
1010         if (ret) {
1011                 pr_err("%s: Failed to set userptr: %d\n", __func__, ret);
1012                 goto out;
1013         }
1014
1015         ret = amdgpu_hmm_register(bo, user_addr);
1016         if (ret) {
1017                 pr_err("%s: Failed to register MMU notifier: %d\n",
1018                        __func__, ret);
1019                 goto out;
1020         }
1021
1022         if (criu_resume) {
1023                 /*
1024                  * During a CRIU restore operation, the userptr buffer objects
1025                  * will be validated in the restore_userptr_work worker at a
1026                  * later stage when it is scheduled by another ioctl called by
1027                  * CRIU master process for the target pid for restore.
1028                  */
1029                 mutex_lock(&process_info->notifier_lock);
1030                 mem->invalid++;
1031                 mutex_unlock(&process_info->notifier_lock);
1032                 mutex_unlock(&process_info->lock);
1033                 return 0;
1034         }
1035
1036         ret = amdgpu_ttm_tt_get_user_pages(bo, bo->tbo.ttm->pages, &range);
1037         if (ret) {
1038                 pr_err("%s: Failed to get user pages: %d\n", __func__, ret);
1039                 goto unregister_out;
1040         }
1041
1042         ret = amdgpu_bo_reserve(bo, true);
1043         if (ret) {
1044                 pr_err("%s: Failed to reserve BO\n", __func__);
1045                 goto release_out;
1046         }
1047         amdgpu_bo_placement_from_domain(bo, mem->domain);
1048         ret = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
1049         if (ret)
1050                 pr_err("%s: failed to validate BO\n", __func__);
1051         amdgpu_bo_unreserve(bo);
1052
1053 release_out:
1054         amdgpu_ttm_tt_get_user_pages_done(bo->tbo.ttm, range);
1055 unregister_out:
1056         if (ret)
1057                 amdgpu_hmm_unregister(bo);
1058 out:
1059         mutex_unlock(&process_info->lock);
1060         return ret;
1061 }
1062
1063 /* Reserving a BO and its page table BOs must happen atomically to
1064  * avoid deadlocks. Some operations update multiple VMs at once. Track
1065  * all the reservation info in a context structure. Optionally a sync
1066  * object can track VM updates.
1067  */
1068 struct bo_vm_reservation_context {
1069         /* DRM execution context for the reservation */
1070         struct drm_exec exec;
1071         /* Number of VMs reserved */
1072         unsigned int n_vms;
1073         /* Pointer to sync object */
1074         struct amdgpu_sync *sync;
1075 };
1076
1077 enum bo_vm_match {
1078         BO_VM_NOT_MAPPED = 0,   /* Match VMs where a BO is not mapped */
1079         BO_VM_MAPPED,           /* Match VMs where a BO is mapped     */
1080         BO_VM_ALL,              /* Match all VMs a BO was added to    */
1081 };
1082
1083 /**
1084  * reserve_bo_and_vm - reserve a BO and a VM unconditionally.
1085  * @mem: KFD BO structure.
1086  * @vm: the VM to reserve.
1087  * @ctx: the struct that will be used in unreserve_bo_and_vms().
1088  */
1089 static int reserve_bo_and_vm(struct kgd_mem *mem,
1090                               struct amdgpu_vm *vm,
1091                               struct bo_vm_reservation_context *ctx)
1092 {
1093         struct amdgpu_bo *bo = mem->bo;
1094         int ret;
1095
1096         WARN_ON(!vm);
1097
1098         ctx->n_vms = 1;
1099         ctx->sync = &mem->sync;
1100         drm_exec_init(&ctx->exec, DRM_EXEC_INTERRUPTIBLE_WAIT);
1101         drm_exec_until_all_locked(&ctx->exec) {
1102                 ret = amdgpu_vm_lock_pd(vm, &ctx->exec, 2);
1103                 drm_exec_retry_on_contention(&ctx->exec);
1104                 if (unlikely(ret))
1105                         goto error;
1106
1107                 ret = drm_exec_lock_obj(&ctx->exec, &bo->tbo.base);
1108                 drm_exec_retry_on_contention(&ctx->exec);
1109                 if (unlikely(ret))
1110                         goto error;
1111         }
1112         return 0;
1113
1114 error:
1115         pr_err("Failed to reserve buffers in ttm.\n");
1116         drm_exec_fini(&ctx->exec);
1117         return ret;
1118 }
1119
1120 /**
1121  * reserve_bo_and_cond_vms - reserve a BO and some VMs conditionally
1122  * @mem: KFD BO structure.
1123  * @vm: the VM to reserve. If NULL, then all VMs associated with the BO
1124  * is used. Otherwise, a single VM associated with the BO.
1125  * @map_type: the mapping status that will be used to filter the VMs.
1126  * @ctx: the struct that will be used in unreserve_bo_and_vms().
1127  *
1128  * Returns 0 for success, negative for failure.
1129  */
1130 static int reserve_bo_and_cond_vms(struct kgd_mem *mem,
1131                                 struct amdgpu_vm *vm, enum bo_vm_match map_type,
1132                                 struct bo_vm_reservation_context *ctx)
1133 {
1134         struct kfd_mem_attachment *entry;
1135         struct amdgpu_bo *bo = mem->bo;
1136         int ret;
1137
1138         ctx->sync = &mem->sync;
1139         drm_exec_init(&ctx->exec, DRM_EXEC_INTERRUPTIBLE_WAIT);
1140         drm_exec_until_all_locked(&ctx->exec) {
1141                 ctx->n_vms = 0;
1142                 list_for_each_entry(entry, &mem->attachments, list) {
1143                         if ((vm && vm != entry->bo_va->base.vm) ||
1144                                 (entry->is_mapped != map_type
1145                                 && map_type != BO_VM_ALL))
1146                                 continue;
1147
1148                         ret = amdgpu_vm_lock_pd(entry->bo_va->base.vm,
1149                                                 &ctx->exec, 2);
1150                         drm_exec_retry_on_contention(&ctx->exec);
1151                         if (unlikely(ret))
1152                                 goto error;
1153                         ++ctx->n_vms;
1154                 }
1155
1156                 ret = drm_exec_prepare_obj(&ctx->exec, &bo->tbo.base, 1);
1157                 drm_exec_retry_on_contention(&ctx->exec);
1158                 if (unlikely(ret))
1159                         goto error;
1160         }
1161         return 0;
1162
1163 error:
1164         pr_err("Failed to reserve buffers in ttm.\n");
1165         drm_exec_fini(&ctx->exec);
1166         return ret;
1167 }
1168
1169 /**
1170  * unreserve_bo_and_vms - Unreserve BO and VMs from a reservation context
1171  * @ctx: Reservation context to unreserve
1172  * @wait: Optionally wait for a sync object representing pending VM updates
1173  * @intr: Whether the wait is interruptible
1174  *
1175  * Also frees any resources allocated in
1176  * reserve_bo_and_(cond_)vm(s). Returns the status from
1177  * amdgpu_sync_wait.
1178  */
1179 static int unreserve_bo_and_vms(struct bo_vm_reservation_context *ctx,
1180                                  bool wait, bool intr)
1181 {
1182         int ret = 0;
1183
1184         if (wait)
1185                 ret = amdgpu_sync_wait(ctx->sync, intr);
1186
1187         drm_exec_fini(&ctx->exec);
1188         ctx->sync = NULL;
1189         return ret;
1190 }
1191
1192 static void unmap_bo_from_gpuvm(struct kgd_mem *mem,
1193                                 struct kfd_mem_attachment *entry,
1194                                 struct amdgpu_sync *sync)
1195 {
1196         struct amdgpu_bo_va *bo_va = entry->bo_va;
1197         struct amdgpu_device *adev = entry->adev;
1198         struct amdgpu_vm *vm = bo_va->base.vm;
1199
1200         amdgpu_vm_bo_unmap(adev, bo_va, entry->va);
1201
1202         amdgpu_vm_clear_freed(adev, vm, &bo_va->last_pt_update);
1203
1204         amdgpu_sync_fence(sync, bo_va->last_pt_update);
1205
1206         kfd_mem_dmaunmap_attachment(mem, entry);
1207 }
1208
1209 static int update_gpuvm_pte(struct kgd_mem *mem,
1210                             struct kfd_mem_attachment *entry,
1211                             struct amdgpu_sync *sync)
1212 {
1213         struct amdgpu_bo_va *bo_va = entry->bo_va;
1214         struct amdgpu_device *adev = entry->adev;
1215         int ret;
1216
1217         ret = kfd_mem_dmamap_attachment(mem, entry);
1218         if (ret)
1219                 return ret;
1220
1221         /* Update the page tables  */
1222         ret = amdgpu_vm_bo_update(adev, bo_va, false);
1223         if (ret) {
1224                 pr_err("amdgpu_vm_bo_update failed\n");
1225                 return ret;
1226         }
1227
1228         return amdgpu_sync_fence(sync, bo_va->last_pt_update);
1229 }
1230
1231 static int map_bo_to_gpuvm(struct kgd_mem *mem,
1232                            struct kfd_mem_attachment *entry,
1233                            struct amdgpu_sync *sync,
1234                            bool no_update_pte)
1235 {
1236         int ret;
1237
1238         /* Set virtual address for the allocation */
1239         ret = amdgpu_vm_bo_map(entry->adev, entry->bo_va, entry->va, 0,
1240                                amdgpu_bo_size(entry->bo_va->base.bo),
1241                                entry->pte_flags);
1242         if (ret) {
1243                 pr_err("Failed to map VA 0x%llx in vm. ret %d\n",
1244                                 entry->va, ret);
1245                 return ret;
1246         }
1247
1248         if (no_update_pte)
1249                 return 0;
1250
1251         ret = update_gpuvm_pte(mem, entry, sync);
1252         if (ret) {
1253                 pr_err("update_gpuvm_pte() failed\n");
1254                 goto update_gpuvm_pte_failed;
1255         }
1256
1257         return 0;
1258
1259 update_gpuvm_pte_failed:
1260         unmap_bo_from_gpuvm(mem, entry, sync);
1261         return ret;
1262 }
1263
1264 static int process_validate_vms(struct amdkfd_process_info *process_info)
1265 {
1266         struct amdgpu_vm *peer_vm;
1267         int ret;
1268
1269         list_for_each_entry(peer_vm, &process_info->vm_list_head,
1270                             vm_list_node) {
1271                 ret = vm_validate_pt_pd_bos(peer_vm);
1272                 if (ret)
1273                         return ret;
1274         }
1275
1276         return 0;
1277 }
1278
1279 static int process_sync_pds_resv(struct amdkfd_process_info *process_info,
1280                                  struct amdgpu_sync *sync)
1281 {
1282         struct amdgpu_vm *peer_vm;
1283         int ret;
1284
1285         list_for_each_entry(peer_vm, &process_info->vm_list_head,
1286                             vm_list_node) {
1287                 struct amdgpu_bo *pd = peer_vm->root.bo;
1288
1289                 ret = amdgpu_sync_resv(NULL, sync, pd->tbo.base.resv,
1290                                        AMDGPU_SYNC_NE_OWNER,
1291                                        AMDGPU_FENCE_OWNER_KFD);
1292                 if (ret)
1293                         return ret;
1294         }
1295
1296         return 0;
1297 }
1298
1299 static int process_update_pds(struct amdkfd_process_info *process_info,
1300                               struct amdgpu_sync *sync)
1301 {
1302         struct amdgpu_vm *peer_vm;
1303         int ret;
1304
1305         list_for_each_entry(peer_vm, &process_info->vm_list_head,
1306                             vm_list_node) {
1307                 ret = vm_update_pds(peer_vm, sync);
1308                 if (ret)
1309                         return ret;
1310         }
1311
1312         return 0;
1313 }
1314
1315 static int init_kfd_vm(struct amdgpu_vm *vm, void **process_info,
1316                        struct dma_fence **ef)
1317 {
1318         struct amdkfd_process_info *info = NULL;
1319         int ret;
1320
1321         if (!*process_info) {
1322                 info = kzalloc(sizeof(*info), GFP_KERNEL);
1323                 if (!info)
1324                         return -ENOMEM;
1325
1326                 mutex_init(&info->lock);
1327                 mutex_init(&info->notifier_lock);
1328                 INIT_LIST_HEAD(&info->vm_list_head);
1329                 INIT_LIST_HEAD(&info->kfd_bo_list);
1330                 INIT_LIST_HEAD(&info->userptr_valid_list);
1331                 INIT_LIST_HEAD(&info->userptr_inval_list);
1332
1333                 info->eviction_fence =
1334                         amdgpu_amdkfd_fence_create(dma_fence_context_alloc(1),
1335                                                    current->mm,
1336                                                    NULL);
1337                 if (!info->eviction_fence) {
1338                         pr_err("Failed to create eviction fence\n");
1339                         ret = -ENOMEM;
1340                         goto create_evict_fence_fail;
1341                 }
1342
1343                 info->pid = get_task_pid(current->group_leader, PIDTYPE_PID);
1344                 INIT_DELAYED_WORK(&info->restore_userptr_work,
1345                                   amdgpu_amdkfd_restore_userptr_worker);
1346
1347                 *process_info = info;
1348                 *ef = dma_fence_get(&info->eviction_fence->base);
1349         }
1350
1351         vm->process_info = *process_info;
1352
1353         /* Validate page directory and attach eviction fence */
1354         ret = amdgpu_bo_reserve(vm->root.bo, true);
1355         if (ret)
1356                 goto reserve_pd_fail;
1357         ret = vm_validate_pt_pd_bos(vm);
1358         if (ret) {
1359                 pr_err("validate_pt_pd_bos() failed\n");
1360                 goto validate_pd_fail;
1361         }
1362         ret = amdgpu_bo_sync_wait(vm->root.bo,
1363                                   AMDGPU_FENCE_OWNER_KFD, false);
1364         if (ret)
1365                 goto wait_pd_fail;
1366         ret = dma_resv_reserve_fences(vm->root.bo->tbo.base.resv, 1);
1367         if (ret)
1368                 goto reserve_shared_fail;
1369         dma_resv_add_fence(vm->root.bo->tbo.base.resv,
1370                            &vm->process_info->eviction_fence->base,
1371                            DMA_RESV_USAGE_BOOKKEEP);
1372         amdgpu_bo_unreserve(vm->root.bo);
1373
1374         /* Update process info */
1375         mutex_lock(&vm->process_info->lock);
1376         list_add_tail(&vm->vm_list_node,
1377                         &(vm->process_info->vm_list_head));
1378         vm->process_info->n_vms++;
1379         mutex_unlock(&vm->process_info->lock);
1380
1381         return 0;
1382
1383 reserve_shared_fail:
1384 wait_pd_fail:
1385 validate_pd_fail:
1386         amdgpu_bo_unreserve(vm->root.bo);
1387 reserve_pd_fail:
1388         vm->process_info = NULL;
1389         if (info) {
1390                 /* Two fence references: one in info and one in *ef */
1391                 dma_fence_put(&info->eviction_fence->base);
1392                 dma_fence_put(*ef);
1393                 *ef = NULL;
1394                 *process_info = NULL;
1395                 put_pid(info->pid);
1396 create_evict_fence_fail:
1397                 mutex_destroy(&info->lock);
1398                 mutex_destroy(&info->notifier_lock);
1399                 kfree(info);
1400         }
1401         return ret;
1402 }
1403
1404 /**
1405  * amdgpu_amdkfd_gpuvm_pin_bo() - Pins a BO using following criteria
1406  * @bo: Handle of buffer object being pinned
1407  * @domain: Domain into which BO should be pinned
1408  *
1409  *   - USERPTR BOs are UNPINNABLE and will return error
1410  *   - All other BO types (GTT, VRAM, MMIO and DOORBELL) will have their
1411  *     PIN count incremented. It is valid to PIN a BO multiple times
1412  *
1413  * Return: ZERO if successful in pinning, Non-Zero in case of error.
1414  */
1415 static int amdgpu_amdkfd_gpuvm_pin_bo(struct amdgpu_bo *bo, u32 domain)
1416 {
1417         int ret = 0;
1418
1419         ret = amdgpu_bo_reserve(bo, false);
1420         if (unlikely(ret))
1421                 return ret;
1422
1423         ret = amdgpu_bo_pin_restricted(bo, domain, 0, 0);
1424         if (ret)
1425                 pr_err("Error in Pinning BO to domain: %d\n", domain);
1426
1427         amdgpu_bo_sync_wait(bo, AMDGPU_FENCE_OWNER_KFD, false);
1428         amdgpu_bo_unreserve(bo);
1429
1430         return ret;
1431 }
1432
1433 /**
1434  * amdgpu_amdkfd_gpuvm_unpin_bo() - Unpins BO using following criteria
1435  * @bo: Handle of buffer object being unpinned
1436  *
1437  *   - Is a illegal request for USERPTR BOs and is ignored
1438  *   - All other BO types (GTT, VRAM, MMIO and DOORBELL) will have their
1439  *     PIN count decremented. Calls to UNPIN must balance calls to PIN
1440  */
1441 static void amdgpu_amdkfd_gpuvm_unpin_bo(struct amdgpu_bo *bo)
1442 {
1443         int ret = 0;
1444
1445         ret = amdgpu_bo_reserve(bo, false);
1446         if (unlikely(ret))
1447                 return;
1448
1449         amdgpu_bo_unpin(bo);
1450         amdgpu_bo_unreserve(bo);
1451 }
1452
1453 int amdgpu_amdkfd_gpuvm_set_vm_pasid(struct amdgpu_device *adev,
1454                                      struct amdgpu_vm *avm, u32 pasid)
1455
1456 {
1457         int ret;
1458
1459         /* Free the original amdgpu allocated pasid,
1460          * will be replaced with kfd allocated pasid.
1461          */
1462         if (avm->pasid) {
1463                 amdgpu_pasid_free(avm->pasid);
1464                 amdgpu_vm_set_pasid(adev, avm, 0);
1465         }
1466
1467         ret = amdgpu_vm_set_pasid(adev, avm, pasid);
1468         if (ret)
1469                 return ret;
1470
1471         return 0;
1472 }
1473
1474 int amdgpu_amdkfd_gpuvm_acquire_process_vm(struct amdgpu_device *adev,
1475                                            struct amdgpu_vm *avm,
1476                                            void **process_info,
1477                                            struct dma_fence **ef)
1478 {
1479         int ret;
1480
1481         /* Already a compute VM? */
1482         if (avm->process_info)
1483                 return -EINVAL;
1484
1485         /* Convert VM into a compute VM */
1486         ret = amdgpu_vm_make_compute(adev, avm);
1487         if (ret)
1488                 return ret;
1489
1490         /* Initialize KFD part of the VM and process info */
1491         ret = init_kfd_vm(avm, process_info, ef);
1492         if (ret)
1493                 return ret;
1494
1495         amdgpu_vm_set_task_info(avm);
1496
1497         return 0;
1498 }
1499
1500 void amdgpu_amdkfd_gpuvm_destroy_cb(struct amdgpu_device *adev,
1501                                     struct amdgpu_vm *vm)
1502 {
1503         struct amdkfd_process_info *process_info = vm->process_info;
1504
1505         if (!process_info)
1506                 return;
1507
1508         /* Update process info */
1509         mutex_lock(&process_info->lock);
1510         process_info->n_vms--;
1511         list_del(&vm->vm_list_node);
1512         mutex_unlock(&process_info->lock);
1513
1514         vm->process_info = NULL;
1515
1516         /* Release per-process resources when last compute VM is destroyed */
1517         if (!process_info->n_vms) {
1518                 WARN_ON(!list_empty(&process_info->kfd_bo_list));
1519                 WARN_ON(!list_empty(&process_info->userptr_valid_list));
1520                 WARN_ON(!list_empty(&process_info->userptr_inval_list));
1521
1522                 dma_fence_put(&process_info->eviction_fence->base);
1523                 cancel_delayed_work_sync(&process_info->restore_userptr_work);
1524                 put_pid(process_info->pid);
1525                 mutex_destroy(&process_info->lock);
1526                 mutex_destroy(&process_info->notifier_lock);
1527                 kfree(process_info);
1528         }
1529 }
1530
1531 void amdgpu_amdkfd_gpuvm_release_process_vm(struct amdgpu_device *adev,
1532                                             void *drm_priv)
1533 {
1534         struct amdgpu_vm *avm;
1535
1536         if (WARN_ON(!adev || !drm_priv))
1537                 return;
1538
1539         avm = drm_priv_to_vm(drm_priv);
1540
1541         pr_debug("Releasing process vm %p\n", avm);
1542
1543         /* The original pasid of amdgpu vm has already been
1544          * released during making a amdgpu vm to a compute vm
1545          * The current pasid is managed by kfd and will be
1546          * released on kfd process destroy. Set amdgpu pasid
1547          * to 0 to avoid duplicate release.
1548          */
1549         amdgpu_vm_release_compute(adev, avm);
1550 }
1551
1552 uint64_t amdgpu_amdkfd_gpuvm_get_process_page_dir(void *drm_priv)
1553 {
1554         struct amdgpu_vm *avm = drm_priv_to_vm(drm_priv);
1555         struct amdgpu_bo *pd = avm->root.bo;
1556         struct amdgpu_device *adev = amdgpu_ttm_adev(pd->tbo.bdev);
1557
1558         if (adev->asic_type < CHIP_VEGA10)
1559                 return avm->pd_phys_addr >> AMDGPU_GPU_PAGE_SHIFT;
1560         return avm->pd_phys_addr;
1561 }
1562
1563 void amdgpu_amdkfd_block_mmu_notifications(void *p)
1564 {
1565         struct amdkfd_process_info *pinfo = (struct amdkfd_process_info *)p;
1566
1567         mutex_lock(&pinfo->lock);
1568         WRITE_ONCE(pinfo->block_mmu_notifications, true);
1569         mutex_unlock(&pinfo->lock);
1570 }
1571
1572 int amdgpu_amdkfd_criu_resume(void *p)
1573 {
1574         int ret = 0;
1575         struct amdkfd_process_info *pinfo = (struct amdkfd_process_info *)p;
1576
1577         mutex_lock(&pinfo->lock);
1578         pr_debug("scheduling work\n");
1579         mutex_lock(&pinfo->notifier_lock);
1580         pinfo->evicted_bos++;
1581         mutex_unlock(&pinfo->notifier_lock);
1582         if (!READ_ONCE(pinfo->block_mmu_notifications)) {
1583                 ret = -EINVAL;
1584                 goto out_unlock;
1585         }
1586         WRITE_ONCE(pinfo->block_mmu_notifications, false);
1587         schedule_delayed_work(&pinfo->restore_userptr_work, 0);
1588
1589 out_unlock:
1590         mutex_unlock(&pinfo->lock);
1591         return ret;
1592 }
1593
1594 size_t amdgpu_amdkfd_get_available_memory(struct amdgpu_device *adev,
1595                                           uint8_t xcp_id)
1596 {
1597         uint64_t reserved_for_pt =
1598                 ESTIMATE_PT_SIZE(amdgpu_amdkfd_total_mem_size);
1599         ssize_t available;
1600         uint64_t vram_available, system_mem_available, ttm_mem_available;
1601
1602         spin_lock(&kfd_mem_limit.mem_limit_lock);
1603         vram_available = KFD_XCP_MEMORY_SIZE(adev, xcp_id)
1604                 - adev->kfd.vram_used_aligned[xcp_id]
1605                 - atomic64_read(&adev->vram_pin_size)
1606                 - reserved_for_pt;
1607
1608         if (adev->gmc.is_app_apu) {
1609                 system_mem_available = no_system_mem_limit ?
1610                                         kfd_mem_limit.max_system_mem_limit :
1611                                         kfd_mem_limit.max_system_mem_limit -
1612                                         kfd_mem_limit.system_mem_used;
1613
1614                 ttm_mem_available = kfd_mem_limit.max_ttm_mem_limit -
1615                                 kfd_mem_limit.ttm_mem_used;
1616
1617                 available = min3(system_mem_available, ttm_mem_available,
1618                                  vram_available);
1619                 available = ALIGN_DOWN(available, PAGE_SIZE);
1620         } else {
1621                 available = ALIGN_DOWN(vram_available, VRAM_AVAILABLITY_ALIGN);
1622         }
1623
1624         spin_unlock(&kfd_mem_limit.mem_limit_lock);
1625
1626         if (available < 0)
1627                 available = 0;
1628
1629         return available;
1630 }
1631
1632 int amdgpu_amdkfd_gpuvm_alloc_memory_of_gpu(
1633                 struct amdgpu_device *adev, uint64_t va, uint64_t size,
1634                 void *drm_priv, struct kgd_mem **mem,
1635                 uint64_t *offset, uint32_t flags, bool criu_resume)
1636 {
1637         struct amdgpu_vm *avm = drm_priv_to_vm(drm_priv);
1638         struct amdgpu_fpriv *fpriv = container_of(avm, struct amdgpu_fpriv, vm);
1639         enum ttm_bo_type bo_type = ttm_bo_type_device;
1640         struct sg_table *sg = NULL;
1641         uint64_t user_addr = 0;
1642         struct amdgpu_bo *bo;
1643         struct drm_gem_object *gobj = NULL;
1644         u32 domain, alloc_domain;
1645         uint64_t aligned_size;
1646         int8_t xcp_id = -1;
1647         u64 alloc_flags;
1648         int ret;
1649
1650         /*
1651          * Check on which domain to allocate BO
1652          */
1653         if (flags & KFD_IOC_ALLOC_MEM_FLAGS_VRAM) {
1654                 domain = alloc_domain = AMDGPU_GEM_DOMAIN_VRAM;
1655
1656                 if (adev->gmc.is_app_apu) {
1657                         domain = AMDGPU_GEM_DOMAIN_GTT;
1658                         alloc_domain = AMDGPU_GEM_DOMAIN_GTT;
1659                         alloc_flags = 0;
1660                 } else {
1661                         alloc_flags = AMDGPU_GEM_CREATE_VRAM_WIPE_ON_RELEASE;
1662                         alloc_flags |= (flags & KFD_IOC_ALLOC_MEM_FLAGS_PUBLIC) ?
1663                         AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED : 0;
1664                 }
1665                 xcp_id = fpriv->xcp_id == ~0 ? 0 : fpriv->xcp_id;
1666         } else if (flags & KFD_IOC_ALLOC_MEM_FLAGS_GTT) {
1667                 domain = alloc_domain = AMDGPU_GEM_DOMAIN_GTT;
1668                 alloc_flags = 0;
1669         } else {
1670                 domain = AMDGPU_GEM_DOMAIN_GTT;
1671                 alloc_domain = AMDGPU_GEM_DOMAIN_CPU;
1672                 alloc_flags = AMDGPU_GEM_CREATE_PREEMPTIBLE;
1673
1674                 if (flags & KFD_IOC_ALLOC_MEM_FLAGS_USERPTR) {
1675                         if (!offset || !*offset)
1676                                 return -EINVAL;
1677                         user_addr = untagged_addr(*offset);
1678                 } else if (flags & (KFD_IOC_ALLOC_MEM_FLAGS_DOORBELL |
1679                                     KFD_IOC_ALLOC_MEM_FLAGS_MMIO_REMAP)) {
1680                         bo_type = ttm_bo_type_sg;
1681                         if (size > UINT_MAX)
1682                                 return -EINVAL;
1683                         sg = create_sg_table(*offset, size);
1684                         if (!sg)
1685                                 return -ENOMEM;
1686                 } else {
1687                         return -EINVAL;
1688                 }
1689         }
1690
1691         if (flags & KFD_IOC_ALLOC_MEM_FLAGS_COHERENT)
1692                 alloc_flags |= AMDGPU_GEM_CREATE_COHERENT;
1693         if (flags & KFD_IOC_ALLOC_MEM_FLAGS_UNCACHED)
1694                 alloc_flags |= AMDGPU_GEM_CREATE_UNCACHED;
1695
1696         *mem = kzalloc(sizeof(struct kgd_mem), GFP_KERNEL);
1697         if (!*mem) {
1698                 ret = -ENOMEM;
1699                 goto err;
1700         }
1701         INIT_LIST_HEAD(&(*mem)->attachments);
1702         mutex_init(&(*mem)->lock);
1703         (*mem)->aql_queue = !!(flags & KFD_IOC_ALLOC_MEM_FLAGS_AQL_QUEUE_MEM);
1704
1705         /* Workaround for AQL queue wraparound bug. Map the same
1706          * memory twice. That means we only actually allocate half
1707          * the memory.
1708          */
1709         if ((*mem)->aql_queue)
1710                 size >>= 1;
1711         aligned_size = PAGE_ALIGN(size);
1712
1713         (*mem)->alloc_flags = flags;
1714
1715         amdgpu_sync_create(&(*mem)->sync);
1716
1717         ret = amdgpu_amdkfd_reserve_mem_limit(adev, aligned_size, flags,
1718                                               xcp_id);
1719         if (ret) {
1720                 pr_debug("Insufficient memory\n");
1721                 goto err_reserve_limit;
1722         }
1723
1724         pr_debug("\tcreate BO VA 0x%llx size 0x%llx domain %s xcp_id %d\n",
1725                  va, (*mem)->aql_queue ? size << 1 : size,
1726                  domain_string(alloc_domain), xcp_id);
1727
1728         ret = amdgpu_gem_object_create(adev, aligned_size, 1, alloc_domain, alloc_flags,
1729                                        bo_type, NULL, &gobj, xcp_id + 1);
1730         if (ret) {
1731                 pr_debug("Failed to create BO on domain %s. ret %d\n",
1732                          domain_string(alloc_domain), ret);
1733                 goto err_bo_create;
1734         }
1735         ret = drm_vma_node_allow(&gobj->vma_node, drm_priv);
1736         if (ret) {
1737                 pr_debug("Failed to allow vma node access. ret %d\n", ret);
1738                 goto err_node_allow;
1739         }
1740         bo = gem_to_amdgpu_bo(gobj);
1741         if (bo_type == ttm_bo_type_sg) {
1742                 bo->tbo.sg = sg;
1743                 bo->tbo.ttm->sg = sg;
1744         }
1745         bo->kfd_bo = *mem;
1746         (*mem)->bo = bo;
1747         if (user_addr)
1748                 bo->flags |= AMDGPU_AMDKFD_CREATE_USERPTR_BO;
1749
1750         (*mem)->va = va;
1751         (*mem)->domain = domain;
1752         (*mem)->mapped_to_gpu_memory = 0;
1753         (*mem)->process_info = avm->process_info;
1754
1755         add_kgd_mem_to_kfd_bo_list(*mem, avm->process_info, user_addr);
1756
1757         if (user_addr) {
1758                 pr_debug("creating userptr BO for user_addr = %llx\n", user_addr);
1759                 ret = init_user_pages(*mem, user_addr, criu_resume);
1760                 if (ret)
1761                         goto allocate_init_user_pages_failed;
1762         } else  if (flags & (KFD_IOC_ALLOC_MEM_FLAGS_DOORBELL |
1763                                 KFD_IOC_ALLOC_MEM_FLAGS_MMIO_REMAP)) {
1764                 ret = amdgpu_amdkfd_gpuvm_pin_bo(bo, AMDGPU_GEM_DOMAIN_GTT);
1765                 if (ret) {
1766                         pr_err("Pinning MMIO/DOORBELL BO during ALLOC FAILED\n");
1767                         goto err_pin_bo;
1768                 }
1769                 bo->allowed_domains = AMDGPU_GEM_DOMAIN_GTT;
1770                 bo->preferred_domains = AMDGPU_GEM_DOMAIN_GTT;
1771         }
1772
1773         if (offset)
1774                 *offset = amdgpu_bo_mmap_offset(bo);
1775
1776         return 0;
1777
1778 allocate_init_user_pages_failed:
1779 err_pin_bo:
1780         remove_kgd_mem_from_kfd_bo_list(*mem, avm->process_info);
1781         drm_vma_node_revoke(&gobj->vma_node, drm_priv);
1782 err_node_allow:
1783         /* Don't unreserve system mem limit twice */
1784         goto err_reserve_limit;
1785 err_bo_create:
1786         amdgpu_amdkfd_unreserve_mem_limit(adev, aligned_size, flags, xcp_id);
1787 err_reserve_limit:
1788         mutex_destroy(&(*mem)->lock);
1789         if (gobj)
1790                 drm_gem_object_put(gobj);
1791         else
1792                 kfree(*mem);
1793 err:
1794         if (sg) {
1795                 sg_free_table(sg);
1796                 kfree(sg);
1797         }
1798         return ret;
1799 }
1800
1801 int amdgpu_amdkfd_gpuvm_free_memory_of_gpu(
1802                 struct amdgpu_device *adev, struct kgd_mem *mem, void *drm_priv,
1803                 uint64_t *size)
1804 {
1805         struct amdkfd_process_info *process_info = mem->process_info;
1806         unsigned long bo_size = mem->bo->tbo.base.size;
1807         bool use_release_notifier = (mem->bo->kfd_bo == mem);
1808         struct kfd_mem_attachment *entry, *tmp;
1809         struct bo_vm_reservation_context ctx;
1810         unsigned int mapped_to_gpu_memory;
1811         int ret;
1812         bool is_imported = false;
1813
1814         mutex_lock(&mem->lock);
1815
1816         /* Unpin MMIO/DOORBELL BO's that were pinned during allocation */
1817         if (mem->alloc_flags &
1818             (KFD_IOC_ALLOC_MEM_FLAGS_DOORBELL |
1819              KFD_IOC_ALLOC_MEM_FLAGS_MMIO_REMAP)) {
1820                 amdgpu_amdkfd_gpuvm_unpin_bo(mem->bo);
1821         }
1822
1823         mapped_to_gpu_memory = mem->mapped_to_gpu_memory;
1824         is_imported = mem->is_imported;
1825         mutex_unlock(&mem->lock);
1826         /* lock is not needed after this, since mem is unused and will
1827          * be freed anyway
1828          */
1829
1830         if (mapped_to_gpu_memory > 0) {
1831                 pr_debug("BO VA 0x%llx size 0x%lx is still mapped.\n",
1832                                 mem->va, bo_size);
1833                 return -EBUSY;
1834         }
1835
1836         /* Make sure restore workers don't access the BO any more */
1837         mutex_lock(&process_info->lock);
1838         list_del(&mem->validate_list);
1839         mutex_unlock(&process_info->lock);
1840
1841         /* Cleanup user pages and MMU notifiers */
1842         if (amdgpu_ttm_tt_get_usermm(mem->bo->tbo.ttm)) {
1843                 amdgpu_hmm_unregister(mem->bo);
1844                 mutex_lock(&process_info->notifier_lock);
1845                 amdgpu_ttm_tt_discard_user_pages(mem->bo->tbo.ttm, mem->range);
1846                 mutex_unlock(&process_info->notifier_lock);
1847         }
1848
1849         ret = reserve_bo_and_cond_vms(mem, NULL, BO_VM_ALL, &ctx);
1850         if (unlikely(ret))
1851                 return ret;
1852
1853         /* The eviction fence should be removed by the last unmap.
1854          * TODO: Log an error condition if the bo still has the eviction fence
1855          * attached
1856          */
1857         amdgpu_amdkfd_remove_eviction_fence(mem->bo,
1858                                         process_info->eviction_fence);
1859         pr_debug("Release VA 0x%llx - 0x%llx\n", mem->va,
1860                 mem->va + bo_size * (1 + mem->aql_queue));
1861
1862         /* Remove from VM internal data structures */
1863         list_for_each_entry_safe(entry, tmp, &mem->attachments, list)
1864                 kfd_mem_detach(entry);
1865
1866         ret = unreserve_bo_and_vms(&ctx, false, false);
1867
1868         /* Free the sync object */
1869         amdgpu_sync_free(&mem->sync);
1870
1871         /* If the SG is not NULL, it's one we created for a doorbell or mmio
1872          * remap BO. We need to free it.
1873          */
1874         if (mem->bo->tbo.sg) {
1875                 sg_free_table(mem->bo->tbo.sg);
1876                 kfree(mem->bo->tbo.sg);
1877         }
1878
1879         /* Update the size of the BO being freed if it was allocated from
1880          * VRAM and is not imported. For APP APU VRAM allocations are done
1881          * in GTT domain
1882          */
1883         if (size) {
1884                 if (!is_imported &&
1885                    (mem->bo->preferred_domains == AMDGPU_GEM_DOMAIN_VRAM ||
1886                    (adev->gmc.is_app_apu &&
1887                     mem->bo->preferred_domains == AMDGPU_GEM_DOMAIN_GTT)))
1888                         *size = bo_size;
1889                 else
1890                         *size = 0;
1891         }
1892
1893         /* Free the BO*/
1894         drm_vma_node_revoke(&mem->bo->tbo.base.vma_node, drm_priv);
1895         if (mem->dmabuf)
1896                 dma_buf_put(mem->dmabuf);
1897         mutex_destroy(&mem->lock);
1898
1899         /* If this releases the last reference, it will end up calling
1900          * amdgpu_amdkfd_release_notify and kfree the mem struct. That's why
1901          * this needs to be the last call here.
1902          */
1903         drm_gem_object_put(&mem->bo->tbo.base);
1904
1905         /*
1906          * For kgd_mem allocated in amdgpu_amdkfd_gpuvm_import_dmabuf(),
1907          * explicitly free it here.
1908          */
1909         if (!use_release_notifier)
1910                 kfree(mem);
1911
1912         return ret;
1913 }
1914
1915 int amdgpu_amdkfd_gpuvm_map_memory_to_gpu(
1916                 struct amdgpu_device *adev, struct kgd_mem *mem,
1917                 void *drm_priv)
1918 {
1919         struct amdgpu_vm *avm = drm_priv_to_vm(drm_priv);
1920         int ret;
1921         struct amdgpu_bo *bo;
1922         uint32_t domain;
1923         struct kfd_mem_attachment *entry;
1924         struct bo_vm_reservation_context ctx;
1925         unsigned long bo_size;
1926         bool is_invalid_userptr = false;
1927
1928         bo = mem->bo;
1929         if (!bo) {
1930                 pr_err("Invalid BO when mapping memory to GPU\n");
1931                 return -EINVAL;
1932         }
1933
1934         /* Make sure restore is not running concurrently. Since we
1935          * don't map invalid userptr BOs, we rely on the next restore
1936          * worker to do the mapping
1937          */
1938         mutex_lock(&mem->process_info->lock);
1939
1940         /* Lock notifier lock. If we find an invalid userptr BO, we can be
1941          * sure that the MMU notifier is no longer running
1942          * concurrently and the queues are actually stopped
1943          */
1944         if (amdgpu_ttm_tt_get_usermm(bo->tbo.ttm)) {
1945                 mutex_lock(&mem->process_info->notifier_lock);
1946                 is_invalid_userptr = !!mem->invalid;
1947                 mutex_unlock(&mem->process_info->notifier_lock);
1948         }
1949
1950         mutex_lock(&mem->lock);
1951
1952         domain = mem->domain;
1953         bo_size = bo->tbo.base.size;
1954
1955         pr_debug("Map VA 0x%llx - 0x%llx to vm %p domain %s\n",
1956                         mem->va,
1957                         mem->va + bo_size * (1 + mem->aql_queue),
1958                         avm, domain_string(domain));
1959
1960         if (!kfd_mem_is_attached(avm, mem)) {
1961                 ret = kfd_mem_attach(adev, mem, avm, mem->aql_queue);
1962                 if (ret)
1963                         goto out;
1964         }
1965
1966         ret = reserve_bo_and_vm(mem, avm, &ctx);
1967         if (unlikely(ret))
1968                 goto out;
1969
1970         /* Userptr can be marked as "not invalid", but not actually be
1971          * validated yet (still in the system domain). In that case
1972          * the queues are still stopped and we can leave mapping for
1973          * the next restore worker
1974          */
1975         if (amdgpu_ttm_tt_get_usermm(bo->tbo.ttm) &&
1976             bo->tbo.resource->mem_type == TTM_PL_SYSTEM)
1977                 is_invalid_userptr = true;
1978
1979         ret = vm_validate_pt_pd_bos(avm);
1980         if (unlikely(ret))
1981                 goto out_unreserve;
1982
1983         if (mem->mapped_to_gpu_memory == 0 &&
1984             !amdgpu_ttm_tt_get_usermm(bo->tbo.ttm)) {
1985                 /* Validate BO only once. The eviction fence gets added to BO
1986                  * the first time it is mapped. Validate will wait for all
1987                  * background evictions to complete.
1988                  */
1989                 ret = amdgpu_amdkfd_bo_validate(bo, domain, true);
1990                 if (ret) {
1991                         pr_debug("Validate failed\n");
1992                         goto out_unreserve;
1993                 }
1994         }
1995
1996         list_for_each_entry(entry, &mem->attachments, list) {
1997                 if (entry->bo_va->base.vm != avm || entry->is_mapped)
1998                         continue;
1999
2000                 pr_debug("\t map VA 0x%llx - 0x%llx in entry %p\n",
2001                          entry->va, entry->va + bo_size, entry);
2002
2003                 ret = map_bo_to_gpuvm(mem, entry, ctx.sync,
2004                                       is_invalid_userptr);
2005                 if (ret) {
2006                         pr_err("Failed to map bo to gpuvm\n");
2007                         goto out_unreserve;
2008                 }
2009
2010                 ret = vm_update_pds(avm, ctx.sync);
2011                 if (ret) {
2012                         pr_err("Failed to update page directories\n");
2013                         goto out_unreserve;
2014                 }
2015
2016                 entry->is_mapped = true;
2017                 mem->mapped_to_gpu_memory++;
2018                 pr_debug("\t INC mapping count %d\n",
2019                          mem->mapped_to_gpu_memory);
2020         }
2021
2022         if (!amdgpu_ttm_tt_get_usermm(bo->tbo.ttm) && !bo->tbo.pin_count)
2023                 dma_resv_add_fence(bo->tbo.base.resv,
2024                                    &avm->process_info->eviction_fence->base,
2025                                    DMA_RESV_USAGE_BOOKKEEP);
2026         ret = unreserve_bo_and_vms(&ctx, false, false);
2027
2028         goto out;
2029
2030 out_unreserve:
2031         unreserve_bo_and_vms(&ctx, false, false);
2032 out:
2033         mutex_unlock(&mem->process_info->lock);
2034         mutex_unlock(&mem->lock);
2035         return ret;
2036 }
2037
2038 int amdgpu_amdkfd_gpuvm_unmap_memory_from_gpu(
2039                 struct amdgpu_device *adev, struct kgd_mem *mem, void *drm_priv)
2040 {
2041         struct amdgpu_vm *avm = drm_priv_to_vm(drm_priv);
2042         struct amdkfd_process_info *process_info = avm->process_info;
2043         unsigned long bo_size = mem->bo->tbo.base.size;
2044         struct kfd_mem_attachment *entry;
2045         struct bo_vm_reservation_context ctx;
2046         int ret;
2047
2048         mutex_lock(&mem->lock);
2049
2050         ret = reserve_bo_and_cond_vms(mem, avm, BO_VM_MAPPED, &ctx);
2051         if (unlikely(ret))
2052                 goto out;
2053         /* If no VMs were reserved, it means the BO wasn't actually mapped */
2054         if (ctx.n_vms == 0) {
2055                 ret = -EINVAL;
2056                 goto unreserve_out;
2057         }
2058
2059         ret = vm_validate_pt_pd_bos(avm);
2060         if (unlikely(ret))
2061                 goto unreserve_out;
2062
2063         pr_debug("Unmap VA 0x%llx - 0x%llx from vm %p\n",
2064                 mem->va,
2065                 mem->va + bo_size * (1 + mem->aql_queue),
2066                 avm);
2067
2068         list_for_each_entry(entry, &mem->attachments, list) {
2069                 if (entry->bo_va->base.vm != avm || !entry->is_mapped)
2070                         continue;
2071
2072                 pr_debug("\t unmap VA 0x%llx - 0x%llx from entry %p\n",
2073                          entry->va, entry->va + bo_size, entry);
2074
2075                 unmap_bo_from_gpuvm(mem, entry, ctx.sync);
2076                 entry->is_mapped = false;
2077
2078                 mem->mapped_to_gpu_memory--;
2079                 pr_debug("\t DEC mapping count %d\n",
2080                          mem->mapped_to_gpu_memory);
2081         }
2082
2083         /* If BO is unmapped from all VMs, unfence it. It can be evicted if
2084          * required.
2085          */
2086         if (mem->mapped_to_gpu_memory == 0 &&
2087             !amdgpu_ttm_tt_get_usermm(mem->bo->tbo.ttm) &&
2088             !mem->bo->tbo.pin_count)
2089                 amdgpu_amdkfd_remove_eviction_fence(mem->bo,
2090                                                 process_info->eviction_fence);
2091
2092 unreserve_out:
2093         unreserve_bo_and_vms(&ctx, false, false);
2094 out:
2095         mutex_unlock(&mem->lock);
2096         return ret;
2097 }
2098
2099 int amdgpu_amdkfd_gpuvm_sync_memory(
2100                 struct amdgpu_device *adev, struct kgd_mem *mem, bool intr)
2101 {
2102         struct amdgpu_sync sync;
2103         int ret;
2104
2105         amdgpu_sync_create(&sync);
2106
2107         mutex_lock(&mem->lock);
2108         amdgpu_sync_clone(&mem->sync, &sync);
2109         mutex_unlock(&mem->lock);
2110
2111         ret = amdgpu_sync_wait(&sync, intr);
2112         amdgpu_sync_free(&sync);
2113         return ret;
2114 }
2115
2116 /**
2117  * amdgpu_amdkfd_map_gtt_bo_to_gart - Map BO to GART and increment reference count
2118  * @adev: Device to which allocated BO belongs
2119  * @bo: Buffer object to be mapped
2120  *
2121  * Before return, bo reference count is incremented. To release the reference and unpin/
2122  * unmap the BO, call amdgpu_amdkfd_free_gtt_mem.
2123  */
2124 int amdgpu_amdkfd_map_gtt_bo_to_gart(struct amdgpu_device *adev, struct amdgpu_bo *bo)
2125 {
2126         int ret;
2127
2128         ret = amdgpu_bo_reserve(bo, true);
2129         if (ret) {
2130                 pr_err("Failed to reserve bo. ret %d\n", ret);
2131                 goto err_reserve_bo_failed;
2132         }
2133
2134         ret = amdgpu_bo_pin(bo, AMDGPU_GEM_DOMAIN_GTT);
2135         if (ret) {
2136                 pr_err("Failed to pin bo. ret %d\n", ret);
2137                 goto err_pin_bo_failed;
2138         }
2139
2140         ret = amdgpu_ttm_alloc_gart(&bo->tbo);
2141         if (ret) {
2142                 pr_err("Failed to bind bo to GART. ret %d\n", ret);
2143                 goto err_map_bo_gart_failed;
2144         }
2145
2146         amdgpu_amdkfd_remove_eviction_fence(
2147                 bo, bo->vm_bo->vm->process_info->eviction_fence);
2148
2149         amdgpu_bo_unreserve(bo);
2150
2151         bo = amdgpu_bo_ref(bo);
2152
2153         return 0;
2154
2155 err_map_bo_gart_failed:
2156         amdgpu_bo_unpin(bo);
2157 err_pin_bo_failed:
2158         amdgpu_bo_unreserve(bo);
2159 err_reserve_bo_failed:
2160
2161         return ret;
2162 }
2163
2164 /** amdgpu_amdkfd_gpuvm_map_gtt_bo_to_kernel() - Map a GTT BO for kernel CPU access
2165  *
2166  * @mem: Buffer object to be mapped for CPU access
2167  * @kptr[out]: pointer in kernel CPU address space
2168  * @size[out]: size of the buffer
2169  *
2170  * Pins the BO and maps it for kernel CPU access. The eviction fence is removed
2171  * from the BO, since pinned BOs cannot be evicted. The bo must remain on the
2172  * validate_list, so the GPU mapping can be restored after a page table was
2173  * evicted.
2174  *
2175  * Return: 0 on success, error code on failure
2176  */
2177 int amdgpu_amdkfd_gpuvm_map_gtt_bo_to_kernel(struct kgd_mem *mem,
2178                                              void **kptr, uint64_t *size)
2179 {
2180         int ret;
2181         struct amdgpu_bo *bo = mem->bo;
2182
2183         if (amdgpu_ttm_tt_get_usermm(bo->tbo.ttm)) {
2184                 pr_err("userptr can't be mapped to kernel\n");
2185                 return -EINVAL;
2186         }
2187
2188         mutex_lock(&mem->process_info->lock);
2189
2190         ret = amdgpu_bo_reserve(bo, true);
2191         if (ret) {
2192                 pr_err("Failed to reserve bo. ret %d\n", ret);
2193                 goto bo_reserve_failed;
2194         }
2195
2196         ret = amdgpu_bo_pin(bo, AMDGPU_GEM_DOMAIN_GTT);
2197         if (ret) {
2198                 pr_err("Failed to pin bo. ret %d\n", ret);
2199                 goto pin_failed;
2200         }
2201
2202         ret = amdgpu_bo_kmap(bo, kptr);
2203         if (ret) {
2204                 pr_err("Failed to map bo to kernel. ret %d\n", ret);
2205                 goto kmap_failed;
2206         }
2207
2208         amdgpu_amdkfd_remove_eviction_fence(
2209                 bo, mem->process_info->eviction_fence);
2210
2211         if (size)
2212                 *size = amdgpu_bo_size(bo);
2213
2214         amdgpu_bo_unreserve(bo);
2215
2216         mutex_unlock(&mem->process_info->lock);
2217         return 0;
2218
2219 kmap_failed:
2220         amdgpu_bo_unpin(bo);
2221 pin_failed:
2222         amdgpu_bo_unreserve(bo);
2223 bo_reserve_failed:
2224         mutex_unlock(&mem->process_info->lock);
2225
2226         return ret;
2227 }
2228
2229 /** amdgpu_amdkfd_gpuvm_map_gtt_bo_to_kernel() - Unmap a GTT BO for kernel CPU access
2230  *
2231  * @mem: Buffer object to be unmapped for CPU access
2232  *
2233  * Removes the kernel CPU mapping and unpins the BO. It does not restore the
2234  * eviction fence, so this function should only be used for cleanup before the
2235  * BO is destroyed.
2236  */
2237 void amdgpu_amdkfd_gpuvm_unmap_gtt_bo_from_kernel(struct kgd_mem *mem)
2238 {
2239         struct amdgpu_bo *bo = mem->bo;
2240
2241         amdgpu_bo_reserve(bo, true);
2242         amdgpu_bo_kunmap(bo);
2243         amdgpu_bo_unpin(bo);
2244         amdgpu_bo_unreserve(bo);
2245 }
2246
2247 int amdgpu_amdkfd_gpuvm_get_vm_fault_info(struct amdgpu_device *adev,
2248                                           struct kfd_vm_fault_info *mem)
2249 {
2250         if (atomic_read(&adev->gmc.vm_fault_info_updated) == 1) {
2251                 *mem = *adev->gmc.vm_fault_info;
2252                 mb(); /* make sure read happened */
2253                 atomic_set(&adev->gmc.vm_fault_info_updated, 0);
2254         }
2255         return 0;
2256 }
2257
2258 int amdgpu_amdkfd_gpuvm_import_dmabuf(struct amdgpu_device *adev,
2259                                       struct dma_buf *dma_buf,
2260                                       uint64_t va, void *drm_priv,
2261                                       struct kgd_mem **mem, uint64_t *size,
2262                                       uint64_t *mmap_offset)
2263 {
2264         struct amdgpu_vm *avm = drm_priv_to_vm(drm_priv);
2265         struct drm_gem_object *obj;
2266         struct amdgpu_bo *bo;
2267         int ret;
2268
2269         obj = amdgpu_gem_prime_import(adev_to_drm(adev), dma_buf);
2270         if (IS_ERR(obj))
2271                 return PTR_ERR(obj);
2272
2273         bo = gem_to_amdgpu_bo(obj);
2274         if (!(bo->preferred_domains & (AMDGPU_GEM_DOMAIN_VRAM |
2275                                     AMDGPU_GEM_DOMAIN_GTT))) {
2276                 /* Only VRAM and GTT BOs are supported */
2277                 ret = -EINVAL;
2278                 goto err_put_obj;
2279         }
2280
2281         *mem = kzalloc(sizeof(struct kgd_mem), GFP_KERNEL);
2282         if (!*mem) {
2283                 ret = -ENOMEM;
2284                 goto err_put_obj;
2285         }
2286
2287         ret = drm_vma_node_allow(&obj->vma_node, drm_priv);
2288         if (ret)
2289                 goto err_free_mem;
2290
2291         if (size)
2292                 *size = amdgpu_bo_size(bo);
2293
2294         if (mmap_offset)
2295                 *mmap_offset = amdgpu_bo_mmap_offset(bo);
2296
2297         INIT_LIST_HEAD(&(*mem)->attachments);
2298         mutex_init(&(*mem)->lock);
2299
2300         (*mem)->alloc_flags =
2301                 ((bo->preferred_domains & AMDGPU_GEM_DOMAIN_VRAM) ?
2302                 KFD_IOC_ALLOC_MEM_FLAGS_VRAM : KFD_IOC_ALLOC_MEM_FLAGS_GTT)
2303                 | KFD_IOC_ALLOC_MEM_FLAGS_WRITABLE
2304                 | KFD_IOC_ALLOC_MEM_FLAGS_EXECUTABLE;
2305
2306         get_dma_buf(dma_buf);
2307         (*mem)->dmabuf = dma_buf;
2308         (*mem)->bo = bo;
2309         (*mem)->va = va;
2310         (*mem)->domain = (bo->preferred_domains & AMDGPU_GEM_DOMAIN_VRAM) && !adev->gmc.is_app_apu ?
2311                 AMDGPU_GEM_DOMAIN_VRAM : AMDGPU_GEM_DOMAIN_GTT;
2312
2313         (*mem)->mapped_to_gpu_memory = 0;
2314         (*mem)->process_info = avm->process_info;
2315         add_kgd_mem_to_kfd_bo_list(*mem, avm->process_info, false);
2316         amdgpu_sync_create(&(*mem)->sync);
2317         (*mem)->is_imported = true;
2318
2319         return 0;
2320
2321 err_free_mem:
2322         kfree(*mem);
2323 err_put_obj:
2324         drm_gem_object_put(obj);
2325         return ret;
2326 }
2327
2328 int amdgpu_amdkfd_gpuvm_export_dmabuf(struct kgd_mem *mem,
2329                                       struct dma_buf **dma_buf)
2330 {
2331         int ret;
2332
2333         mutex_lock(&mem->lock);
2334         ret = kfd_mem_export_dmabuf(mem);
2335         if (ret)
2336                 goto out;
2337
2338         get_dma_buf(mem->dmabuf);
2339         *dma_buf = mem->dmabuf;
2340 out:
2341         mutex_unlock(&mem->lock);
2342         return ret;
2343 }
2344
2345 /* Evict a userptr BO by stopping the queues if necessary
2346  *
2347  * Runs in MMU notifier, may be in RECLAIM_FS context. This means it
2348  * cannot do any memory allocations, and cannot take any locks that
2349  * are held elsewhere while allocating memory.
2350  *
2351  * It doesn't do anything to the BO itself. The real work happens in
2352  * restore, where we get updated page addresses. This function only
2353  * ensures that GPU access to the BO is stopped.
2354  */
2355 int amdgpu_amdkfd_evict_userptr(struct mmu_interval_notifier *mni,
2356                                 unsigned long cur_seq, struct kgd_mem *mem)
2357 {
2358         struct amdkfd_process_info *process_info = mem->process_info;
2359         int r = 0;
2360
2361         /* Do not process MMU notifications during CRIU restore until
2362          * KFD_CRIU_OP_RESUME IOCTL is received
2363          */
2364         if (READ_ONCE(process_info->block_mmu_notifications))
2365                 return 0;
2366
2367         mutex_lock(&process_info->notifier_lock);
2368         mmu_interval_set_seq(mni, cur_seq);
2369
2370         mem->invalid++;
2371         if (++process_info->evicted_bos == 1) {
2372                 /* First eviction, stop the queues */
2373                 r = kgd2kfd_quiesce_mm(mni->mm,
2374                                        KFD_QUEUE_EVICTION_TRIGGER_USERPTR);
2375                 if (r)
2376                         pr_err("Failed to quiesce KFD\n");
2377                 schedule_delayed_work(&process_info->restore_userptr_work,
2378                         msecs_to_jiffies(AMDGPU_USERPTR_RESTORE_DELAY_MS));
2379         }
2380         mutex_unlock(&process_info->notifier_lock);
2381
2382         return r;
2383 }
2384
2385 /* Update invalid userptr BOs
2386  *
2387  * Moves invalidated (evicted) userptr BOs from userptr_valid_list to
2388  * userptr_inval_list and updates user pages for all BOs that have
2389  * been invalidated since their last update.
2390  */
2391 static int update_invalid_user_pages(struct amdkfd_process_info *process_info,
2392                                      struct mm_struct *mm)
2393 {
2394         struct kgd_mem *mem, *tmp_mem;
2395         struct amdgpu_bo *bo;
2396         struct ttm_operation_ctx ctx = { false, false };
2397         uint32_t invalid;
2398         int ret = 0;
2399
2400         mutex_lock(&process_info->notifier_lock);
2401
2402         /* Move all invalidated BOs to the userptr_inval_list */
2403         list_for_each_entry_safe(mem, tmp_mem,
2404                                  &process_info->userptr_valid_list,
2405                                  validate_list)
2406                 if (mem->invalid)
2407                         list_move_tail(&mem->validate_list,
2408                                        &process_info->userptr_inval_list);
2409
2410         /* Go through userptr_inval_list and update any invalid user_pages */
2411         list_for_each_entry(mem, &process_info->userptr_inval_list,
2412                             validate_list) {
2413                 invalid = mem->invalid;
2414                 if (!invalid)
2415                         /* BO hasn't been invalidated since the last
2416                          * revalidation attempt. Keep its page list.
2417                          */
2418                         continue;
2419
2420                 bo = mem->bo;
2421
2422                 amdgpu_ttm_tt_discard_user_pages(bo->tbo.ttm, mem->range);
2423                 mem->range = NULL;
2424
2425                 /* BO reservations and getting user pages (hmm_range_fault)
2426                  * must happen outside the notifier lock
2427                  */
2428                 mutex_unlock(&process_info->notifier_lock);
2429
2430                 /* Move the BO to system (CPU) domain if necessary to unmap
2431                  * and free the SG table
2432                  */
2433                 if (bo->tbo.resource->mem_type != TTM_PL_SYSTEM) {
2434                         if (amdgpu_bo_reserve(bo, true))
2435                                 return -EAGAIN;
2436                         amdgpu_bo_placement_from_domain(bo, AMDGPU_GEM_DOMAIN_CPU);
2437                         ret = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
2438                         amdgpu_bo_unreserve(bo);
2439                         if (ret) {
2440                                 pr_err("%s: Failed to invalidate userptr BO\n",
2441                                        __func__);
2442                                 return -EAGAIN;
2443                         }
2444                 }
2445
2446                 /* Get updated user pages */
2447                 ret = amdgpu_ttm_tt_get_user_pages(bo, bo->tbo.ttm->pages,
2448                                                    &mem->range);
2449                 if (ret) {
2450                         pr_debug("Failed %d to get user pages\n", ret);
2451
2452                         /* Return -EFAULT bad address error as success. It will
2453                          * fail later with a VM fault if the GPU tries to access
2454                          * it. Better than hanging indefinitely with stalled
2455                          * user mode queues.
2456                          *
2457                          * Return other error -EBUSY or -ENOMEM to retry restore
2458                          */
2459                         if (ret != -EFAULT)
2460                                 return ret;
2461
2462                         ret = 0;
2463                 }
2464
2465                 mutex_lock(&process_info->notifier_lock);
2466
2467                 /* Mark the BO as valid unless it was invalidated
2468                  * again concurrently.
2469                  */
2470                 if (mem->invalid != invalid) {
2471                         ret = -EAGAIN;
2472                         goto unlock_out;
2473                 }
2474                  /* set mem valid if mem has hmm range associated */
2475                 if (mem->range)
2476                         mem->invalid = 0;
2477         }
2478
2479 unlock_out:
2480         mutex_unlock(&process_info->notifier_lock);
2481
2482         return ret;
2483 }
2484
2485 /* Validate invalid userptr BOs
2486  *
2487  * Validates BOs on the userptr_inval_list. Also updates GPUVM page tables
2488  * with new page addresses and waits for the page table updates to complete.
2489  */
2490 static int validate_invalid_user_pages(struct amdkfd_process_info *process_info)
2491 {
2492         struct ttm_operation_ctx ctx = { false, false };
2493         struct amdgpu_sync sync;
2494         struct drm_exec exec;
2495
2496         struct amdgpu_vm *peer_vm;
2497         struct kgd_mem *mem, *tmp_mem;
2498         struct amdgpu_bo *bo;
2499         int ret;
2500
2501         amdgpu_sync_create(&sync);
2502
2503         drm_exec_init(&exec, 0);
2504         /* Reserve all BOs and page tables for validation */
2505         drm_exec_until_all_locked(&exec) {
2506                 /* Reserve all the page directories */
2507                 list_for_each_entry(peer_vm, &process_info->vm_list_head,
2508                                     vm_list_node) {
2509                         ret = amdgpu_vm_lock_pd(peer_vm, &exec, 2);
2510                         drm_exec_retry_on_contention(&exec);
2511                         if (unlikely(ret))
2512                                 goto unreserve_out;
2513                 }
2514
2515                 /* Reserve the userptr_inval_list entries to resv_list */
2516                 list_for_each_entry(mem, &process_info->userptr_inval_list,
2517                                     validate_list) {
2518                         struct drm_gem_object *gobj;
2519
2520                         gobj = &mem->bo->tbo.base;
2521                         ret = drm_exec_prepare_obj(&exec, gobj, 1);
2522                         drm_exec_retry_on_contention(&exec);
2523                         if (unlikely(ret))
2524                                 goto unreserve_out;
2525                 }
2526         }
2527
2528         ret = process_validate_vms(process_info);
2529         if (ret)
2530                 goto unreserve_out;
2531
2532         /* Validate BOs and update GPUVM page tables */
2533         list_for_each_entry_safe(mem, tmp_mem,
2534                                  &process_info->userptr_inval_list,
2535                                  validate_list) {
2536                 struct kfd_mem_attachment *attachment;
2537
2538                 bo = mem->bo;
2539
2540                 /* Validate the BO if we got user pages */
2541                 if (bo->tbo.ttm->pages[0]) {
2542                         amdgpu_bo_placement_from_domain(bo, mem->domain);
2543                         ret = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
2544                         if (ret) {
2545                                 pr_err("%s: failed to validate BO\n", __func__);
2546                                 goto unreserve_out;
2547                         }
2548                 }
2549
2550                 /* Update mapping. If the BO was not validated
2551                  * (because we couldn't get user pages), this will
2552                  * clear the page table entries, which will result in
2553                  * VM faults if the GPU tries to access the invalid
2554                  * memory.
2555                  */
2556                 list_for_each_entry(attachment, &mem->attachments, list) {
2557                         if (!attachment->is_mapped)
2558                                 continue;
2559
2560                         kfd_mem_dmaunmap_attachment(mem, attachment);
2561                         ret = update_gpuvm_pte(mem, attachment, &sync);
2562                         if (ret) {
2563                                 pr_err("%s: update PTE failed\n", __func__);
2564                                 /* make sure this gets validated again */
2565                                 mutex_lock(&process_info->notifier_lock);
2566                                 mem->invalid++;
2567                                 mutex_unlock(&process_info->notifier_lock);
2568                                 goto unreserve_out;
2569                         }
2570                 }
2571         }
2572
2573         /* Update page directories */
2574         ret = process_update_pds(process_info, &sync);
2575
2576 unreserve_out:
2577         drm_exec_fini(&exec);
2578         amdgpu_sync_wait(&sync, false);
2579         amdgpu_sync_free(&sync);
2580
2581         return ret;
2582 }
2583
2584 /* Confirm that all user pages are valid while holding the notifier lock
2585  *
2586  * Moves valid BOs from the userptr_inval_list back to userptr_val_list.
2587  */
2588 static int confirm_valid_user_pages_locked(struct amdkfd_process_info *process_info)
2589 {
2590         struct kgd_mem *mem, *tmp_mem;
2591         int ret = 0;
2592
2593         list_for_each_entry_safe(mem, tmp_mem,
2594                                  &process_info->userptr_inval_list,
2595                                  validate_list) {
2596                 bool valid;
2597
2598                 /* keep mem without hmm range at userptr_inval_list */
2599                 if (!mem->range)
2600                          continue;
2601
2602                 /* Only check mem with hmm range associated */
2603                 valid = amdgpu_ttm_tt_get_user_pages_done(
2604                                         mem->bo->tbo.ttm, mem->range);
2605
2606                 mem->range = NULL;
2607                 if (!valid) {
2608                         WARN(!mem->invalid, "Invalid BO not marked invalid");
2609                         ret = -EAGAIN;
2610                         continue;
2611                 }
2612
2613                 if (mem->invalid) {
2614                         WARN(1, "Valid BO is marked invalid");
2615                         ret = -EAGAIN;
2616                         continue;
2617                 }
2618
2619                 list_move_tail(&mem->validate_list,
2620                                &process_info->userptr_valid_list);
2621         }
2622
2623         return ret;
2624 }
2625
2626 /* Worker callback to restore evicted userptr BOs
2627  *
2628  * Tries to update and validate all userptr BOs. If successful and no
2629  * concurrent evictions happened, the queues are restarted. Otherwise,
2630  * reschedule for another attempt later.
2631  */
2632 static void amdgpu_amdkfd_restore_userptr_worker(struct work_struct *work)
2633 {
2634         struct delayed_work *dwork = to_delayed_work(work);
2635         struct amdkfd_process_info *process_info =
2636                 container_of(dwork, struct amdkfd_process_info,
2637                              restore_userptr_work);
2638         struct task_struct *usertask;
2639         struct mm_struct *mm;
2640         uint32_t evicted_bos;
2641
2642         mutex_lock(&process_info->notifier_lock);
2643         evicted_bos = process_info->evicted_bos;
2644         mutex_unlock(&process_info->notifier_lock);
2645         if (!evicted_bos)
2646                 return;
2647
2648         /* Reference task and mm in case of concurrent process termination */
2649         usertask = get_pid_task(process_info->pid, PIDTYPE_PID);
2650         if (!usertask)
2651                 return;
2652         mm = get_task_mm(usertask);
2653         if (!mm) {
2654                 put_task_struct(usertask);
2655                 return;
2656         }
2657
2658         mutex_lock(&process_info->lock);
2659
2660         if (update_invalid_user_pages(process_info, mm))
2661                 goto unlock_out;
2662         /* userptr_inval_list can be empty if all evicted userptr BOs
2663          * have been freed. In that case there is nothing to validate
2664          * and we can just restart the queues.
2665          */
2666         if (!list_empty(&process_info->userptr_inval_list)) {
2667                 if (validate_invalid_user_pages(process_info))
2668                         goto unlock_out;
2669         }
2670         /* Final check for concurrent evicton and atomic update. If
2671          * another eviction happens after successful update, it will
2672          * be a first eviction that calls quiesce_mm. The eviction
2673          * reference counting inside KFD will handle this case.
2674          */
2675         mutex_lock(&process_info->notifier_lock);
2676         if (process_info->evicted_bos != evicted_bos)
2677                 goto unlock_notifier_out;
2678
2679         if (confirm_valid_user_pages_locked(process_info)) {
2680                 WARN(1, "User pages unexpectedly invalid");
2681                 goto unlock_notifier_out;
2682         }
2683
2684         process_info->evicted_bos = evicted_bos = 0;
2685
2686         if (kgd2kfd_resume_mm(mm)) {
2687                 pr_err("%s: Failed to resume KFD\n", __func__);
2688                 /* No recovery from this failure. Probably the CP is
2689                  * hanging. No point trying again.
2690                  */
2691         }
2692
2693 unlock_notifier_out:
2694         mutex_unlock(&process_info->notifier_lock);
2695 unlock_out:
2696         mutex_unlock(&process_info->lock);
2697
2698         /* If validation failed, reschedule another attempt */
2699         if (evicted_bos) {
2700                 schedule_delayed_work(&process_info->restore_userptr_work,
2701                         msecs_to_jiffies(AMDGPU_USERPTR_RESTORE_DELAY_MS));
2702
2703                 kfd_smi_event_queue_restore_rescheduled(mm);
2704         }
2705         mmput(mm);
2706         put_task_struct(usertask);
2707 }
2708
2709 /** amdgpu_amdkfd_gpuvm_restore_process_bos - Restore all BOs for the given
2710  *   KFD process identified by process_info
2711  *
2712  * @process_info: amdkfd_process_info of the KFD process
2713  *
2714  * After memory eviction, restore thread calls this function. The function
2715  * should be called when the Process is still valid. BO restore involves -
2716  *
2717  * 1.  Release old eviction fence and create new one
2718  * 2.  Get two copies of PD BO list from all the VMs. Keep one copy as pd_list.
2719  * 3   Use the second PD list and kfd_bo_list to create a list (ctx.list) of
2720  *     BOs that need to be reserved.
2721  * 4.  Reserve all the BOs
2722  * 5.  Validate of PD and PT BOs.
2723  * 6.  Validate all KFD BOs using kfd_bo_list and Map them and add new fence
2724  * 7.  Add fence to all PD and PT BOs.
2725  * 8.  Unreserve all BOs
2726  */
2727 int amdgpu_amdkfd_gpuvm_restore_process_bos(void *info, struct dma_fence **ef)
2728 {
2729         struct amdkfd_process_info *process_info = info;
2730         struct amdgpu_vm *peer_vm;
2731         struct kgd_mem *mem;
2732         struct amdgpu_amdkfd_fence *new_fence;
2733         struct list_head duplicate_save;
2734         struct amdgpu_sync sync_obj;
2735         unsigned long failed_size = 0;
2736         unsigned long total_size = 0;
2737         struct drm_exec exec;
2738         int ret;
2739
2740         INIT_LIST_HEAD(&duplicate_save);
2741
2742         mutex_lock(&process_info->lock);
2743
2744         drm_exec_init(&exec, 0);
2745         drm_exec_until_all_locked(&exec) {
2746                 list_for_each_entry(peer_vm, &process_info->vm_list_head,
2747                                     vm_list_node) {
2748                         ret = amdgpu_vm_lock_pd(peer_vm, &exec, 2);
2749                         drm_exec_retry_on_contention(&exec);
2750                         if (unlikely(ret))
2751                                 goto ttm_reserve_fail;
2752                 }
2753
2754                 /* Reserve all BOs and page tables/directory. Add all BOs from
2755                  * kfd_bo_list to ctx.list
2756                  */
2757                 list_for_each_entry(mem, &process_info->kfd_bo_list,
2758                                     validate_list) {
2759                         struct drm_gem_object *gobj;
2760
2761                         gobj = &mem->bo->tbo.base;
2762                         ret = drm_exec_prepare_obj(&exec, gobj, 1);
2763                         drm_exec_retry_on_contention(&exec);
2764                         if (unlikely(ret))
2765                                 goto ttm_reserve_fail;
2766                 }
2767         }
2768
2769         amdgpu_sync_create(&sync_obj);
2770
2771         /* Validate PDs and PTs */
2772         ret = process_validate_vms(process_info);
2773         if (ret)
2774                 goto validate_map_fail;
2775
2776         ret = process_sync_pds_resv(process_info, &sync_obj);
2777         if (ret) {
2778                 pr_debug("Memory eviction: Failed to sync to PD BO moving fence. Try again\n");
2779                 goto validate_map_fail;
2780         }
2781
2782         /* Validate BOs and map them to GPUVM (update VM page tables). */
2783         list_for_each_entry(mem, &process_info->kfd_bo_list,
2784                             validate_list) {
2785
2786                 struct amdgpu_bo *bo = mem->bo;
2787                 uint32_t domain = mem->domain;
2788                 struct kfd_mem_attachment *attachment;
2789                 struct dma_resv_iter cursor;
2790                 struct dma_fence *fence;
2791
2792                 total_size += amdgpu_bo_size(bo);
2793
2794                 ret = amdgpu_amdkfd_bo_validate(bo, domain, false);
2795                 if (ret) {
2796                         pr_debug("Memory eviction: Validate BOs failed\n");
2797                         failed_size += amdgpu_bo_size(bo);
2798                         ret = amdgpu_amdkfd_bo_validate(bo,
2799                                                 AMDGPU_GEM_DOMAIN_GTT, false);
2800                         if (ret) {
2801                                 pr_debug("Memory eviction: Try again\n");
2802                                 goto validate_map_fail;
2803                         }
2804                 }
2805                 dma_resv_for_each_fence(&cursor, bo->tbo.base.resv,
2806                                         DMA_RESV_USAGE_KERNEL, fence) {
2807                         ret = amdgpu_sync_fence(&sync_obj, fence);
2808                         if (ret) {
2809                                 pr_debug("Memory eviction: Sync BO fence failed. Try again\n");
2810                                 goto validate_map_fail;
2811                         }
2812                 }
2813                 list_for_each_entry(attachment, &mem->attachments, list) {
2814                         if (!attachment->is_mapped)
2815                                 continue;
2816
2817                         if (attachment->bo_va->base.bo->tbo.pin_count)
2818                                 continue;
2819
2820                         kfd_mem_dmaunmap_attachment(mem, attachment);
2821                         ret = update_gpuvm_pte(mem, attachment, &sync_obj);
2822                         if (ret) {
2823                                 pr_debug("Memory eviction: update PTE failed. Try again\n");
2824                                 goto validate_map_fail;
2825                         }
2826                 }
2827         }
2828
2829         if (failed_size)
2830                 pr_debug("0x%lx/0x%lx in system\n", failed_size, total_size);
2831
2832         /* Update page directories */
2833         ret = process_update_pds(process_info, &sync_obj);
2834         if (ret) {
2835                 pr_debug("Memory eviction: update PDs failed. Try again\n");
2836                 goto validate_map_fail;
2837         }
2838
2839         /* Wait for validate and PT updates to finish */
2840         amdgpu_sync_wait(&sync_obj, false);
2841
2842         /* Release old eviction fence and create new one, because fence only
2843          * goes from unsignaled to signaled, fence cannot be reused.
2844          * Use context and mm from the old fence.
2845          */
2846         new_fence = amdgpu_amdkfd_fence_create(
2847                                 process_info->eviction_fence->base.context,
2848                                 process_info->eviction_fence->mm,
2849                                 NULL);
2850         if (!new_fence) {
2851                 pr_err("Failed to create eviction fence\n");
2852                 ret = -ENOMEM;
2853                 goto validate_map_fail;
2854         }
2855         dma_fence_put(&process_info->eviction_fence->base);
2856         process_info->eviction_fence = new_fence;
2857         *ef = dma_fence_get(&new_fence->base);
2858
2859         /* Attach new eviction fence to all BOs except pinned ones */
2860         list_for_each_entry(mem, &process_info->kfd_bo_list, validate_list) {
2861                 if (mem->bo->tbo.pin_count)
2862                         continue;
2863
2864                 dma_resv_add_fence(mem->bo->tbo.base.resv,
2865                                    &process_info->eviction_fence->base,
2866                                    DMA_RESV_USAGE_BOOKKEEP);
2867         }
2868         /* Attach eviction fence to PD / PT BOs */
2869         list_for_each_entry(peer_vm, &process_info->vm_list_head,
2870                             vm_list_node) {
2871                 struct amdgpu_bo *bo = peer_vm->root.bo;
2872
2873                 dma_resv_add_fence(bo->tbo.base.resv,
2874                                    &process_info->eviction_fence->base,
2875                                    DMA_RESV_USAGE_BOOKKEEP);
2876         }
2877
2878 validate_map_fail:
2879         amdgpu_sync_free(&sync_obj);
2880 ttm_reserve_fail:
2881         drm_exec_fini(&exec);
2882         mutex_unlock(&process_info->lock);
2883         return ret;
2884 }
2885
2886 int amdgpu_amdkfd_add_gws_to_process(void *info, void *gws, struct kgd_mem **mem)
2887 {
2888         struct amdkfd_process_info *process_info = (struct amdkfd_process_info *)info;
2889         struct amdgpu_bo *gws_bo = (struct amdgpu_bo *)gws;
2890         int ret;
2891
2892         if (!info || !gws)
2893                 return -EINVAL;
2894
2895         *mem = kzalloc(sizeof(struct kgd_mem), GFP_KERNEL);
2896         if (!*mem)
2897                 return -ENOMEM;
2898
2899         mutex_init(&(*mem)->lock);
2900         INIT_LIST_HEAD(&(*mem)->attachments);
2901         (*mem)->bo = amdgpu_bo_ref(gws_bo);
2902         (*mem)->domain = AMDGPU_GEM_DOMAIN_GWS;
2903         (*mem)->process_info = process_info;
2904         add_kgd_mem_to_kfd_bo_list(*mem, process_info, false);
2905         amdgpu_sync_create(&(*mem)->sync);
2906
2907
2908         /* Validate gws bo the first time it is added to process */
2909         mutex_lock(&(*mem)->process_info->lock);
2910         ret = amdgpu_bo_reserve(gws_bo, false);
2911         if (unlikely(ret)) {
2912                 pr_err("Reserve gws bo failed %d\n", ret);
2913                 goto bo_reservation_failure;
2914         }
2915
2916         ret = amdgpu_amdkfd_bo_validate(gws_bo, AMDGPU_GEM_DOMAIN_GWS, true);
2917         if (ret) {
2918                 pr_err("GWS BO validate failed %d\n", ret);
2919                 goto bo_validation_failure;
2920         }
2921         /* GWS resource is shared b/t amdgpu and amdkfd
2922          * Add process eviction fence to bo so they can
2923          * evict each other.
2924          */
2925         ret = dma_resv_reserve_fences(gws_bo->tbo.base.resv, 1);
2926         if (ret)
2927                 goto reserve_shared_fail;
2928         dma_resv_add_fence(gws_bo->tbo.base.resv,
2929                            &process_info->eviction_fence->base,
2930                            DMA_RESV_USAGE_BOOKKEEP);
2931         amdgpu_bo_unreserve(gws_bo);
2932         mutex_unlock(&(*mem)->process_info->lock);
2933
2934         return ret;
2935
2936 reserve_shared_fail:
2937 bo_validation_failure:
2938         amdgpu_bo_unreserve(gws_bo);
2939 bo_reservation_failure:
2940         mutex_unlock(&(*mem)->process_info->lock);
2941         amdgpu_sync_free(&(*mem)->sync);
2942         remove_kgd_mem_from_kfd_bo_list(*mem, process_info);
2943         amdgpu_bo_unref(&gws_bo);
2944         mutex_destroy(&(*mem)->lock);
2945         kfree(*mem);
2946         *mem = NULL;
2947         return ret;
2948 }
2949
2950 int amdgpu_amdkfd_remove_gws_from_process(void *info, void *mem)
2951 {
2952         int ret;
2953         struct amdkfd_process_info *process_info = (struct amdkfd_process_info *)info;
2954         struct kgd_mem *kgd_mem = (struct kgd_mem *)mem;
2955         struct amdgpu_bo *gws_bo = kgd_mem->bo;
2956
2957         /* Remove BO from process's validate list so restore worker won't touch
2958          * it anymore
2959          */
2960         remove_kgd_mem_from_kfd_bo_list(kgd_mem, process_info);
2961
2962         ret = amdgpu_bo_reserve(gws_bo, false);
2963         if (unlikely(ret)) {
2964                 pr_err("Reserve gws bo failed %d\n", ret);
2965                 //TODO add BO back to validate_list?
2966                 return ret;
2967         }
2968         amdgpu_amdkfd_remove_eviction_fence(gws_bo,
2969                         process_info->eviction_fence);
2970         amdgpu_bo_unreserve(gws_bo);
2971         amdgpu_sync_free(&kgd_mem->sync);
2972         amdgpu_bo_unref(&gws_bo);
2973         mutex_destroy(&kgd_mem->lock);
2974         kfree(mem);
2975         return 0;
2976 }
2977
2978 /* Returns GPU-specific tiling mode information */
2979 int amdgpu_amdkfd_get_tile_config(struct amdgpu_device *adev,
2980                                 struct tile_config *config)
2981 {
2982         config->gb_addr_config = adev->gfx.config.gb_addr_config;
2983         config->tile_config_ptr = adev->gfx.config.tile_mode_array;
2984         config->num_tile_configs =
2985                         ARRAY_SIZE(adev->gfx.config.tile_mode_array);
2986         config->macro_tile_config_ptr =
2987                         adev->gfx.config.macrotile_mode_array;
2988         config->num_macro_tile_configs =
2989                         ARRAY_SIZE(adev->gfx.config.macrotile_mode_array);
2990
2991         /* Those values are not set from GFX9 onwards */
2992         config->num_banks = adev->gfx.config.num_banks;
2993         config->num_ranks = adev->gfx.config.num_ranks;
2994
2995         return 0;
2996 }
2997
2998 bool amdgpu_amdkfd_bo_mapped_to_dev(struct amdgpu_device *adev, struct kgd_mem *mem)
2999 {
3000         struct kfd_mem_attachment *entry;
3001
3002         list_for_each_entry(entry, &mem->attachments, list) {
3003                 if (entry->is_mapped && entry->adev == adev)
3004                         return true;
3005         }
3006         return false;
3007 }
3008
3009 #if defined(CONFIG_DEBUG_FS)
3010
3011 int kfd_debugfs_kfd_mem_limits(struct seq_file *m, void *data)
3012 {
3013
3014         spin_lock(&kfd_mem_limit.mem_limit_lock);
3015         seq_printf(m, "System mem used %lldM out of %lluM\n",
3016                   (kfd_mem_limit.system_mem_used >> 20),
3017                   (kfd_mem_limit.max_system_mem_limit >> 20));
3018         seq_printf(m, "TTM mem used %lldM out of %lluM\n",
3019                   (kfd_mem_limit.ttm_mem_used >> 20),
3020                   (kfd_mem_limit.max_ttm_mem_limit >> 20));
3021         spin_unlock(&kfd_mem_limit.mem_limit_lock);
3022
3023         return 0;
3024 }
3025
3026 #endif