1 /* 2 * Copyright 2008 Advanced Micro Devices, Inc. 3 * Copyright 2008 Red Hat Inc. 4 * Copyright 2009 Jerome Glisse. 5 * 6 * Permission is hereby granted, free of charge, to any person obtaining a 7 * copy of this software and associated documentation files (the "Software"), 8 * to deal in the Software without restriction, including without limitation 9 * the rights to use, copy, modify, merge, publish, distribute, sublicense, 10 * and/or sell copies of the Software, and to permit persons to whom the 11 * Software is furnished to do so, subject to the following conditions: 12 * 13 * The above copyright notice and this permission notice shall be included in 14 * all copies or substantial portions of the Software. 15 * 16 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR 17 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, 18 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL 19 * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR 20 * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, 21 * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR 22 * OTHER DEALINGS IN THE SOFTWARE. 23 * 24 * Authors: Dave Airlie 25 * Alex Deucher 26 * Jerome Glisse 27 */ 28 #include <linux/ktime.h> 29 #include <linux/module.h> 30 #include <linux/pagemap.h> 31 #include <linux/pci.h> 32 #include <linux/dma-buf.h> 33 34 #include <drm/amdgpu_drm.h> 35 #include <drm/drm_drv.h> 36 #include <drm/drm_exec.h> 37 #include <drm/drm_gem_ttm_helper.h> 38 #include <drm/ttm/ttm_tt.h> 39 40 #include "amdgpu.h" 41 #include "amdgpu_display.h" 42 #include "amdgpu_dma_buf.h" 43 #include "amdgpu_hmm.h" 44 #include "amdgpu_xgmi.h" 45 46 static const struct drm_gem_object_funcs amdgpu_gem_object_funcs; 47 48 static vm_fault_t amdgpu_gem_fault(struct vm_fault *vmf) 49 { 50 struct ttm_buffer_object *bo = vmf->vma->vm_private_data; 51 struct drm_device *ddev = bo->base.dev; 52 vm_fault_t ret; 53 int idx; 54 55 ret = ttm_bo_vm_reserve(bo, vmf); 56 if (ret) 57 return ret; 58 59 if (drm_dev_enter(ddev, &idx)) { 60 ret = amdgpu_bo_fault_reserve_notify(bo); 61 if (ret) { 62 drm_dev_exit(idx); 63 goto unlock; 64 } 65 66 ret = ttm_bo_vm_fault_reserved(vmf, vmf->vma->vm_page_prot, 67 TTM_BO_VM_NUM_PREFAULT); 68 69 drm_dev_exit(idx); 70 } else { 71 ret = ttm_bo_vm_dummy_page(vmf, vmf->vma->vm_page_prot); 72 } 73 if (ret == VM_FAULT_RETRY && !(vmf->flags & FAULT_FLAG_RETRY_NOWAIT)) 74 return ret; 75 76 unlock: 77 dma_resv_unlock(bo->base.resv); 78 return ret; 79 } 80 81 static const struct vm_operations_struct amdgpu_gem_vm_ops = { 82 .fault = amdgpu_gem_fault, 83 .open = ttm_bo_vm_open, 84 .close = ttm_bo_vm_close, 85 .access = ttm_bo_vm_access 86 }; 87 88 static void amdgpu_gem_object_free(struct drm_gem_object *gobj) 89 { 90 struct amdgpu_bo *robj = gem_to_amdgpu_bo(gobj); 91 92 if (robj) { 93 amdgpu_hmm_unregister(robj); 94 amdgpu_bo_unref(&robj); 95 } 96 } 97 98 int amdgpu_gem_object_create(struct amdgpu_device *adev, unsigned long size, 99 int alignment, u32 initial_domain, 100 u64 flags, enum ttm_bo_type type, 101 struct dma_resv *resv, 102 struct drm_gem_object **obj, int8_t xcp_id_plus1) 103 { 104 struct amdgpu_bo *bo; 105 struct amdgpu_bo_user *ubo; 106 struct amdgpu_bo_param bp; 107 int r; 108 109 memset(&bp, 0, sizeof(bp)); 110 *obj = NULL; 111 flags |= AMDGPU_GEM_CREATE_VRAM_WIPE_ON_RELEASE; 112 113 bp.size = size; 114 bp.byte_align = alignment; 115 bp.type = type; 116 bp.resv = resv; 117 bp.preferred_domain = initial_domain; 118 bp.flags = flags; 119 bp.domain = initial_domain; 120 bp.bo_ptr_size = sizeof(struct amdgpu_bo); 121 bp.xcp_id_plus1 = xcp_id_plus1; 122 123 r = amdgpu_bo_create_user(adev, &bp, &ubo); 124 if (r) 125 return r; 126 127 bo = &ubo->bo; 128 *obj = &bo->tbo.base; 129 (*obj)->funcs = &amdgpu_gem_object_funcs; 130 131 return 0; 132 } 133 134 void amdgpu_gem_force_release(struct amdgpu_device *adev) 135 { 136 struct drm_device *ddev = adev_to_drm(adev); 137 struct drm_file *file; 138 139 mutex_lock(&ddev->filelist_mutex); 140 141 list_for_each_entry(file, &ddev->filelist, lhead) { 142 struct drm_gem_object *gobj; 143 int handle; 144 145 WARN_ONCE(1, "Still active user space clients!\n"); 146 spin_lock(&file->table_lock); 147 idr_for_each_entry(&file->object_idr, gobj, handle) { 148 WARN_ONCE(1, "And also active allocations!\n"); 149 drm_gem_object_put(gobj); 150 } 151 idr_destroy(&file->object_idr); 152 spin_unlock(&file->table_lock); 153 } 154 155 mutex_unlock(&ddev->filelist_mutex); 156 } 157 158 /* 159 * Call from drm_gem_handle_create which appear in both new and open ioctl 160 * case. 161 */ 162 static int amdgpu_gem_object_open(struct drm_gem_object *obj, 163 struct drm_file *file_priv) 164 { 165 struct amdgpu_bo *abo = gem_to_amdgpu_bo(obj); 166 struct amdgpu_device *adev = amdgpu_ttm_adev(abo->tbo.bdev); 167 struct amdgpu_fpriv *fpriv = file_priv->driver_priv; 168 struct amdgpu_vm *vm = &fpriv->vm; 169 struct amdgpu_bo_va *bo_va; 170 struct mm_struct *mm; 171 int r; 172 173 mm = amdgpu_ttm_tt_get_usermm(abo->tbo.ttm); 174 if (mm && mm != current->mm) 175 return -EPERM; 176 177 if (abo->flags & AMDGPU_GEM_CREATE_VM_ALWAYS_VALID && 178 !amdgpu_vm_is_bo_always_valid(vm, abo)) 179 return -EPERM; 180 181 r = amdgpu_bo_reserve(abo, false); 182 if (r) 183 return r; 184 185 bo_va = amdgpu_vm_bo_find(vm, abo); 186 if (!bo_va) 187 bo_va = amdgpu_vm_bo_add(adev, vm, abo); 188 else 189 ++bo_va->ref_count; 190 amdgpu_bo_unreserve(abo); 191 192 /* Validate and add eviction fence to DMABuf imports with dynamic 193 * attachment in compute VMs. Re-validation will be done by 194 * amdgpu_vm_validate. Fences are on the reservation shared with the 195 * export, which is currently required to be validated and fenced 196 * already by amdgpu_amdkfd_gpuvm_restore_process_bos. 197 * 198 * Nested locking below for the case that a GEM object is opened in 199 * kfd_mem_export_dmabuf. Since the lock below is only taken for imports, 200 * but not for export, this is a different lock class that cannot lead to 201 * circular lock dependencies. 202 */ 203 if (!vm->is_compute_context || !vm->process_info) 204 return 0; 205 if (!obj->import_attach || 206 !dma_buf_is_dynamic(obj->import_attach->dmabuf)) 207 return 0; 208 mutex_lock_nested(&vm->process_info->lock, 1); 209 if (!WARN_ON(!vm->process_info->eviction_fence)) { 210 r = amdgpu_amdkfd_bo_validate_and_fence(abo, AMDGPU_GEM_DOMAIN_GTT, 211 &vm->process_info->eviction_fence->base); 212 if (r) { 213 struct amdgpu_task_info *ti = amdgpu_vm_get_task_info_vm(vm); 214 215 dev_warn(adev->dev, "validate_and_fence failed: %d\n", r); 216 if (ti) { 217 dev_warn(adev->dev, "pid %d\n", ti->pid); 218 amdgpu_vm_put_task_info(ti); 219 } 220 } 221 } 222 mutex_unlock(&vm->process_info->lock); 223 224 return r; 225 } 226 227 static void amdgpu_gem_object_close(struct drm_gem_object *obj, 228 struct drm_file *file_priv) 229 { 230 struct amdgpu_bo *bo = gem_to_amdgpu_bo(obj); 231 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev); 232 struct amdgpu_fpriv *fpriv = file_priv->driver_priv; 233 struct amdgpu_vm *vm = &fpriv->vm; 234 235 struct dma_fence *fence = NULL; 236 struct amdgpu_bo_va *bo_va; 237 struct drm_exec exec; 238 long r; 239 240 drm_exec_init(&exec, DRM_EXEC_IGNORE_DUPLICATES, 0); 241 drm_exec_until_all_locked(&exec) { 242 r = drm_exec_prepare_obj(&exec, &bo->tbo.base, 1); 243 drm_exec_retry_on_contention(&exec); 244 if (unlikely(r)) 245 goto out_unlock; 246 247 r = amdgpu_vm_lock_pd(vm, &exec, 0); 248 drm_exec_retry_on_contention(&exec); 249 if (unlikely(r)) 250 goto out_unlock; 251 } 252 253 bo_va = amdgpu_vm_bo_find(vm, bo); 254 if (!bo_va || --bo_va->ref_count) 255 goto out_unlock; 256 257 amdgpu_vm_bo_del(adev, bo_va); 258 if (!amdgpu_vm_ready(vm)) 259 goto out_unlock; 260 261 r = amdgpu_vm_clear_freed(adev, vm, &fence); 262 if (unlikely(r < 0)) 263 dev_err(adev->dev, "failed to clear page " 264 "tables on GEM object close (%ld)\n", r); 265 if (r || !fence) 266 goto out_unlock; 267 268 amdgpu_bo_fence(bo, fence, true); 269 dma_fence_put(fence); 270 271 out_unlock: 272 if (r) 273 dev_err(adev->dev, "leaking bo va (%ld)\n", r); 274 drm_exec_fini(&exec); 275 } 276 277 static int amdgpu_gem_object_mmap(struct drm_gem_object *obj, struct vm_area_struct *vma) 278 { 279 struct amdgpu_bo *bo = gem_to_amdgpu_bo(obj); 280 281 if (amdgpu_ttm_tt_get_usermm(bo->tbo.ttm)) 282 return -EPERM; 283 if (bo->flags & AMDGPU_GEM_CREATE_NO_CPU_ACCESS) 284 return -EPERM; 285 286 /* Workaround for Thunk bug creating PROT_NONE,MAP_PRIVATE mappings 287 * for debugger access to invisible VRAM. Should have used MAP_SHARED 288 * instead. Clearing VM_MAYWRITE prevents the mapping from ever 289 * becoming writable and makes is_cow_mapping(vm_flags) false. 290 */ 291 if (is_cow_mapping(vma->vm_flags) && 292 !(vma->vm_flags & VM_ACCESS_FLAGS)) 293 vm_flags_clear(vma, VM_MAYWRITE); 294 295 return drm_gem_ttm_mmap(obj, vma); 296 } 297 298 static const struct drm_gem_object_funcs amdgpu_gem_object_funcs = { 299 .free = amdgpu_gem_object_free, 300 .open = amdgpu_gem_object_open, 301 .close = amdgpu_gem_object_close, 302 .export = amdgpu_gem_prime_export, 303 .vmap = drm_gem_ttm_vmap, 304 .vunmap = drm_gem_ttm_vunmap, 305 .mmap = amdgpu_gem_object_mmap, 306 .vm_ops = &amdgpu_gem_vm_ops, 307 }; 308 309 /* 310 * GEM ioctls. 311 */ 312 int amdgpu_gem_create_ioctl(struct drm_device *dev, void *data, 313 struct drm_file *filp) 314 { 315 struct amdgpu_device *adev = drm_to_adev(dev); 316 struct amdgpu_fpriv *fpriv = filp->driver_priv; 317 struct amdgpu_vm *vm = &fpriv->vm; 318 union drm_amdgpu_gem_create *args = data; 319 uint64_t flags = args->in.domain_flags; 320 uint64_t size = args->in.bo_size; 321 struct dma_resv *resv = NULL; 322 struct drm_gem_object *gobj; 323 uint32_t handle, initial_domain; 324 int r; 325 326 /* reject DOORBELLs until userspace code to use it is available */ 327 if (args->in.domains & AMDGPU_GEM_DOMAIN_DOORBELL) 328 return -EINVAL; 329 330 /* reject invalid gem flags */ 331 if (flags & ~(AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED | 332 AMDGPU_GEM_CREATE_NO_CPU_ACCESS | 333 AMDGPU_GEM_CREATE_CPU_GTT_USWC | 334 AMDGPU_GEM_CREATE_VRAM_CLEARED | 335 AMDGPU_GEM_CREATE_VM_ALWAYS_VALID | 336 AMDGPU_GEM_CREATE_EXPLICIT_SYNC | 337 AMDGPU_GEM_CREATE_ENCRYPTED | 338 AMDGPU_GEM_CREATE_GFX12_DCC | 339 AMDGPU_GEM_CREATE_DISCARDABLE)) 340 return -EINVAL; 341 342 /* reject invalid gem domains */ 343 if (args->in.domains & ~AMDGPU_GEM_DOMAIN_MASK) 344 return -EINVAL; 345 346 if (!amdgpu_is_tmz(adev) && (flags & AMDGPU_GEM_CREATE_ENCRYPTED)) { 347 DRM_NOTE_ONCE("Cannot allocate secure buffer since TMZ is disabled\n"); 348 return -EINVAL; 349 } 350 351 /* create a gem object to contain this object in */ 352 if (args->in.domains & (AMDGPU_GEM_DOMAIN_GDS | 353 AMDGPU_GEM_DOMAIN_GWS | AMDGPU_GEM_DOMAIN_OA)) { 354 if (flags & AMDGPU_GEM_CREATE_VM_ALWAYS_VALID) { 355 /* if gds bo is created from user space, it must be 356 * passed to bo list 357 */ 358 DRM_ERROR("GDS bo cannot be per-vm-bo\n"); 359 return -EINVAL; 360 } 361 flags |= AMDGPU_GEM_CREATE_NO_CPU_ACCESS; 362 } 363 364 if (flags & AMDGPU_GEM_CREATE_VM_ALWAYS_VALID) { 365 r = amdgpu_bo_reserve(vm->root.bo, false); 366 if (r) 367 return r; 368 369 resv = vm->root.bo->tbo.base.resv; 370 } 371 372 initial_domain = (u32)(0xffffffff & args->in.domains); 373 retry: 374 r = amdgpu_gem_object_create(adev, size, args->in.alignment, 375 initial_domain, 376 flags, ttm_bo_type_device, resv, &gobj, fpriv->xcp_id + 1); 377 if (r && r != -ERESTARTSYS) { 378 if (flags & AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED) { 379 flags &= ~AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED; 380 goto retry; 381 } 382 383 if (initial_domain == AMDGPU_GEM_DOMAIN_VRAM) { 384 initial_domain |= AMDGPU_GEM_DOMAIN_GTT; 385 goto retry; 386 } 387 DRM_DEBUG("Failed to allocate GEM object (%llu, %d, %llu, %d)\n", 388 size, initial_domain, args->in.alignment, r); 389 } 390 391 if (flags & AMDGPU_GEM_CREATE_VM_ALWAYS_VALID) { 392 if (!r) { 393 struct amdgpu_bo *abo = gem_to_amdgpu_bo(gobj); 394 395 abo->parent = amdgpu_bo_ref(vm->root.bo); 396 } 397 amdgpu_bo_unreserve(vm->root.bo); 398 } 399 if (r) 400 return r; 401 402 r = drm_gem_handle_create(filp, gobj, &handle); 403 /* drop reference from allocate - handle holds it now */ 404 drm_gem_object_put(gobj); 405 if (r) 406 return r; 407 408 memset(args, 0, sizeof(*args)); 409 args->out.handle = handle; 410 return 0; 411 } 412 413 int amdgpu_gem_userptr_ioctl(struct drm_device *dev, void *data, 414 struct drm_file *filp) 415 { 416 struct ttm_operation_ctx ctx = { true, false }; 417 struct amdgpu_device *adev = drm_to_adev(dev); 418 struct drm_amdgpu_gem_userptr *args = data; 419 struct amdgpu_fpriv *fpriv = filp->driver_priv; 420 struct drm_gem_object *gobj; 421 struct hmm_range *range; 422 struct amdgpu_bo *bo; 423 uint32_t handle; 424 int r; 425 426 args->addr = untagged_addr(args->addr); 427 428 if (offset_in_page(args->addr | args->size)) 429 return -EINVAL; 430 431 /* reject unknown flag values */ 432 if (args->flags & ~(AMDGPU_GEM_USERPTR_READONLY | 433 AMDGPU_GEM_USERPTR_ANONONLY | AMDGPU_GEM_USERPTR_VALIDATE | 434 AMDGPU_GEM_USERPTR_REGISTER)) 435 return -EINVAL; 436 437 if (!(args->flags & AMDGPU_GEM_USERPTR_READONLY) && 438 !(args->flags & AMDGPU_GEM_USERPTR_REGISTER)) { 439 440 /* if we want to write to it we must install a MMU notifier */ 441 return -EACCES; 442 } 443 444 /* create a gem object to contain this object in */ 445 r = amdgpu_gem_object_create(adev, args->size, 0, AMDGPU_GEM_DOMAIN_CPU, 446 0, ttm_bo_type_device, NULL, &gobj, fpriv->xcp_id + 1); 447 if (r) 448 return r; 449 450 bo = gem_to_amdgpu_bo(gobj); 451 bo->preferred_domains = AMDGPU_GEM_DOMAIN_GTT; 452 bo->allowed_domains = AMDGPU_GEM_DOMAIN_GTT; 453 r = amdgpu_ttm_tt_set_userptr(&bo->tbo, args->addr, args->flags); 454 if (r) 455 goto release_object; 456 457 r = amdgpu_hmm_register(bo, args->addr); 458 if (r) 459 goto release_object; 460 461 if (args->flags & AMDGPU_GEM_USERPTR_VALIDATE) { 462 r = amdgpu_ttm_tt_get_user_pages(bo, bo->tbo.ttm->pages, 463 &range); 464 if (r) 465 goto release_object; 466 467 r = amdgpu_bo_reserve(bo, true); 468 if (r) 469 goto user_pages_done; 470 471 amdgpu_bo_placement_from_domain(bo, AMDGPU_GEM_DOMAIN_GTT); 472 r = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx); 473 amdgpu_bo_unreserve(bo); 474 if (r) 475 goto user_pages_done; 476 } 477 478 r = drm_gem_handle_create(filp, gobj, &handle); 479 if (r) 480 goto user_pages_done; 481 482 args->handle = handle; 483 484 user_pages_done: 485 if (args->flags & AMDGPU_GEM_USERPTR_VALIDATE) 486 amdgpu_ttm_tt_get_user_pages_done(bo->tbo.ttm, range); 487 488 release_object: 489 drm_gem_object_put(gobj); 490 491 return r; 492 } 493 494 int amdgpu_mode_dumb_mmap(struct drm_file *filp, 495 struct drm_device *dev, 496 uint32_t handle, uint64_t *offset_p) 497 { 498 struct drm_gem_object *gobj; 499 struct amdgpu_bo *robj; 500 501 gobj = drm_gem_object_lookup(filp, handle); 502 if (!gobj) 503 return -ENOENT; 504 505 robj = gem_to_amdgpu_bo(gobj); 506 if (amdgpu_ttm_tt_get_usermm(robj->tbo.ttm) || 507 (robj->flags & AMDGPU_GEM_CREATE_NO_CPU_ACCESS)) { 508 drm_gem_object_put(gobj); 509 return -EPERM; 510 } 511 *offset_p = amdgpu_bo_mmap_offset(robj); 512 drm_gem_object_put(gobj); 513 return 0; 514 } 515 516 int amdgpu_gem_mmap_ioctl(struct drm_device *dev, void *data, 517 struct drm_file *filp) 518 { 519 union drm_amdgpu_gem_mmap *args = data; 520 uint32_t handle = args->in.handle; 521 522 memset(args, 0, sizeof(*args)); 523 return amdgpu_mode_dumb_mmap(filp, dev, handle, &args->out.addr_ptr); 524 } 525 526 /** 527 * amdgpu_gem_timeout - calculate jiffies timeout from absolute value 528 * 529 * @timeout_ns: timeout in ns 530 * 531 * Calculate the timeout in jiffies from an absolute timeout in ns. 532 */ 533 unsigned long amdgpu_gem_timeout(uint64_t timeout_ns) 534 { 535 unsigned long timeout_jiffies; 536 ktime_t timeout; 537 538 /* clamp timeout if it's to large */ 539 if (((int64_t)timeout_ns) < 0) 540 return MAX_SCHEDULE_TIMEOUT; 541 542 timeout = ktime_sub(ns_to_ktime(timeout_ns), ktime_get()); 543 if (ktime_to_ns(timeout) < 0) 544 return 0; 545 546 timeout_jiffies = nsecs_to_jiffies(ktime_to_ns(timeout)); 547 /* clamp timeout to avoid unsigned-> signed overflow */ 548 if (timeout_jiffies > MAX_SCHEDULE_TIMEOUT) 549 return MAX_SCHEDULE_TIMEOUT - 1; 550 551 return timeout_jiffies; 552 } 553 554 int amdgpu_gem_wait_idle_ioctl(struct drm_device *dev, void *data, 555 struct drm_file *filp) 556 { 557 union drm_amdgpu_gem_wait_idle *args = data; 558 struct drm_gem_object *gobj; 559 struct amdgpu_bo *robj; 560 uint32_t handle = args->in.handle; 561 unsigned long timeout = amdgpu_gem_timeout(args->in.timeout); 562 int r = 0; 563 long ret; 564 565 gobj = drm_gem_object_lookup(filp, handle); 566 if (!gobj) 567 return -ENOENT; 568 569 robj = gem_to_amdgpu_bo(gobj); 570 ret = dma_resv_wait_timeout(robj->tbo.base.resv, DMA_RESV_USAGE_READ, 571 true, timeout); 572 573 /* ret == 0 means not signaled, 574 * ret > 0 means signaled 575 * ret < 0 means interrupted before timeout 576 */ 577 if (ret >= 0) { 578 memset(args, 0, sizeof(*args)); 579 args->out.status = (ret == 0); 580 } else 581 r = ret; 582 583 drm_gem_object_put(gobj); 584 return r; 585 } 586 587 int amdgpu_gem_metadata_ioctl(struct drm_device *dev, void *data, 588 struct drm_file *filp) 589 { 590 struct drm_amdgpu_gem_metadata *args = data; 591 struct drm_gem_object *gobj; 592 struct amdgpu_bo *robj; 593 int r = -1; 594 595 DRM_DEBUG("%d\n", args->handle); 596 gobj = drm_gem_object_lookup(filp, args->handle); 597 if (gobj == NULL) 598 return -ENOENT; 599 robj = gem_to_amdgpu_bo(gobj); 600 601 r = amdgpu_bo_reserve(robj, false); 602 if (unlikely(r != 0)) 603 goto out; 604 605 if (args->op == AMDGPU_GEM_METADATA_OP_GET_METADATA) { 606 amdgpu_bo_get_tiling_flags(robj, &args->data.tiling_info); 607 r = amdgpu_bo_get_metadata(robj, args->data.data, 608 sizeof(args->data.data), 609 &args->data.data_size_bytes, 610 &args->data.flags); 611 } else if (args->op == AMDGPU_GEM_METADATA_OP_SET_METADATA) { 612 if (args->data.data_size_bytes > sizeof(args->data.data)) { 613 r = -EINVAL; 614 goto unreserve; 615 } 616 r = amdgpu_bo_set_tiling_flags(robj, args->data.tiling_info); 617 if (!r) 618 r = amdgpu_bo_set_metadata(robj, args->data.data, 619 args->data.data_size_bytes, 620 args->data.flags); 621 } 622 623 unreserve: 624 amdgpu_bo_unreserve(robj); 625 out: 626 drm_gem_object_put(gobj); 627 return r; 628 } 629 630 /** 631 * amdgpu_gem_va_update_vm -update the bo_va in its VM 632 * 633 * @adev: amdgpu_device pointer 634 * @vm: vm to update 635 * @bo_va: bo_va to update 636 * @operation: map, unmap or clear 637 * 638 * Update the bo_va directly after setting its address. Errors are not 639 * vital here, so they are not reported back to userspace. 640 */ 641 static void amdgpu_gem_va_update_vm(struct amdgpu_device *adev, 642 struct amdgpu_vm *vm, 643 struct amdgpu_bo_va *bo_va, 644 uint32_t operation) 645 { 646 int r; 647 648 if (!amdgpu_vm_ready(vm)) 649 return; 650 651 r = amdgpu_vm_clear_freed(adev, vm, NULL); 652 if (r) 653 goto error; 654 655 if (operation == AMDGPU_VA_OP_MAP || 656 operation == AMDGPU_VA_OP_REPLACE) { 657 r = amdgpu_vm_bo_update(adev, bo_va, false); 658 if (r) 659 goto error; 660 } 661 662 r = amdgpu_vm_update_pdes(adev, vm, false); 663 664 error: 665 if (r && r != -ERESTARTSYS) 666 DRM_ERROR("Couldn't update BO_VA (%d)\n", r); 667 } 668 669 /** 670 * amdgpu_gem_va_map_flags - map GEM UAPI flags into hardware flags 671 * 672 * @adev: amdgpu_device pointer 673 * @flags: GEM UAPI flags 674 * 675 * Returns the GEM UAPI flags mapped into hardware for the ASIC. 676 */ 677 uint64_t amdgpu_gem_va_map_flags(struct amdgpu_device *adev, uint32_t flags) 678 { 679 uint64_t pte_flag = 0; 680 681 if (flags & AMDGPU_VM_PAGE_EXECUTABLE) 682 pte_flag |= AMDGPU_PTE_EXECUTABLE; 683 if (flags & AMDGPU_VM_PAGE_READABLE) 684 pte_flag |= AMDGPU_PTE_READABLE; 685 if (flags & AMDGPU_VM_PAGE_WRITEABLE) 686 pte_flag |= AMDGPU_PTE_WRITEABLE; 687 if (flags & AMDGPU_VM_PAGE_PRT) 688 pte_flag |= AMDGPU_PTE_PRT_FLAG(adev); 689 if (flags & AMDGPU_VM_PAGE_NOALLOC) 690 pte_flag |= AMDGPU_PTE_NOALLOC; 691 692 if (adev->gmc.gmc_funcs->map_mtype) 693 pte_flag |= amdgpu_gmc_map_mtype(adev, 694 flags & AMDGPU_VM_MTYPE_MASK); 695 696 return pte_flag; 697 } 698 699 int amdgpu_gem_va_ioctl(struct drm_device *dev, void *data, 700 struct drm_file *filp) 701 { 702 const uint32_t valid_flags = AMDGPU_VM_DELAY_UPDATE | 703 AMDGPU_VM_PAGE_READABLE | AMDGPU_VM_PAGE_WRITEABLE | 704 AMDGPU_VM_PAGE_EXECUTABLE | AMDGPU_VM_MTYPE_MASK | 705 AMDGPU_VM_PAGE_NOALLOC; 706 const uint32_t prt_flags = AMDGPU_VM_DELAY_UPDATE | 707 AMDGPU_VM_PAGE_PRT; 708 709 struct drm_amdgpu_gem_va *args = data; 710 struct drm_gem_object *gobj; 711 struct amdgpu_device *adev = drm_to_adev(dev); 712 struct amdgpu_fpriv *fpriv = filp->driver_priv; 713 struct amdgpu_bo *abo; 714 struct amdgpu_bo_va *bo_va; 715 struct drm_exec exec; 716 uint64_t va_flags; 717 uint64_t vm_size; 718 int r = 0; 719 720 if (args->va_address < AMDGPU_VA_RESERVED_BOTTOM) { 721 dev_dbg(dev->dev, 722 "va_address 0x%llx is in reserved area 0x%llx\n", 723 args->va_address, AMDGPU_VA_RESERVED_BOTTOM); 724 return -EINVAL; 725 } 726 727 if (args->va_address >= AMDGPU_GMC_HOLE_START && 728 args->va_address < AMDGPU_GMC_HOLE_END) { 729 dev_dbg(dev->dev, 730 "va_address 0x%llx is in VA hole 0x%llx-0x%llx\n", 731 args->va_address, AMDGPU_GMC_HOLE_START, 732 AMDGPU_GMC_HOLE_END); 733 return -EINVAL; 734 } 735 736 args->va_address &= AMDGPU_GMC_HOLE_MASK; 737 738 vm_size = adev->vm_manager.max_pfn * AMDGPU_GPU_PAGE_SIZE; 739 vm_size -= AMDGPU_VA_RESERVED_TOP; 740 if (args->va_address + args->map_size > vm_size) { 741 dev_dbg(dev->dev, 742 "va_address 0x%llx is in top reserved area 0x%llx\n", 743 args->va_address + args->map_size, vm_size); 744 return -EINVAL; 745 } 746 747 if ((args->flags & ~valid_flags) && (args->flags & ~prt_flags)) { 748 dev_dbg(dev->dev, "invalid flags combination 0x%08X\n", 749 args->flags); 750 return -EINVAL; 751 } 752 753 switch (args->operation) { 754 case AMDGPU_VA_OP_MAP: 755 case AMDGPU_VA_OP_UNMAP: 756 case AMDGPU_VA_OP_CLEAR: 757 case AMDGPU_VA_OP_REPLACE: 758 break; 759 default: 760 dev_dbg(dev->dev, "unsupported operation %d\n", 761 args->operation); 762 return -EINVAL; 763 } 764 765 if ((args->operation != AMDGPU_VA_OP_CLEAR) && 766 !(args->flags & AMDGPU_VM_PAGE_PRT)) { 767 gobj = drm_gem_object_lookup(filp, args->handle); 768 if (gobj == NULL) 769 return -ENOENT; 770 abo = gem_to_amdgpu_bo(gobj); 771 } else { 772 gobj = NULL; 773 abo = NULL; 774 } 775 776 drm_exec_init(&exec, DRM_EXEC_INTERRUPTIBLE_WAIT | 777 DRM_EXEC_IGNORE_DUPLICATES, 0); 778 drm_exec_until_all_locked(&exec) { 779 if (gobj) { 780 r = drm_exec_lock_obj(&exec, gobj); 781 drm_exec_retry_on_contention(&exec); 782 if (unlikely(r)) 783 goto error; 784 } 785 786 r = amdgpu_vm_lock_pd(&fpriv->vm, &exec, 2); 787 drm_exec_retry_on_contention(&exec); 788 if (unlikely(r)) 789 goto error; 790 } 791 792 if (abo) { 793 bo_va = amdgpu_vm_bo_find(&fpriv->vm, abo); 794 if (!bo_va) { 795 r = -ENOENT; 796 goto error; 797 } 798 } else if (args->operation != AMDGPU_VA_OP_CLEAR) { 799 bo_va = fpriv->prt_va; 800 } else { 801 bo_va = NULL; 802 } 803 804 switch (args->operation) { 805 case AMDGPU_VA_OP_MAP: 806 va_flags = amdgpu_gem_va_map_flags(adev, args->flags); 807 r = amdgpu_vm_bo_map(adev, bo_va, args->va_address, 808 args->offset_in_bo, args->map_size, 809 va_flags); 810 break; 811 case AMDGPU_VA_OP_UNMAP: 812 r = amdgpu_vm_bo_unmap(adev, bo_va, args->va_address); 813 break; 814 815 case AMDGPU_VA_OP_CLEAR: 816 r = amdgpu_vm_bo_clear_mappings(adev, &fpriv->vm, 817 args->va_address, 818 args->map_size); 819 break; 820 case AMDGPU_VA_OP_REPLACE: 821 va_flags = amdgpu_gem_va_map_flags(adev, args->flags); 822 r = amdgpu_vm_bo_replace_map(adev, bo_va, args->va_address, 823 args->offset_in_bo, args->map_size, 824 va_flags); 825 break; 826 default: 827 break; 828 } 829 if (!r && !(args->flags & AMDGPU_VM_DELAY_UPDATE) && !adev->debug_vm) 830 amdgpu_gem_va_update_vm(adev, &fpriv->vm, bo_va, 831 args->operation); 832 833 error: 834 drm_exec_fini(&exec); 835 drm_gem_object_put(gobj); 836 return r; 837 } 838 839 int amdgpu_gem_op_ioctl(struct drm_device *dev, void *data, 840 struct drm_file *filp) 841 { 842 struct amdgpu_device *adev = drm_to_adev(dev); 843 struct drm_amdgpu_gem_op *args = data; 844 struct drm_gem_object *gobj; 845 struct amdgpu_vm_bo_base *base; 846 struct amdgpu_bo *robj; 847 int r; 848 849 gobj = drm_gem_object_lookup(filp, args->handle); 850 if (!gobj) 851 return -ENOENT; 852 853 robj = gem_to_amdgpu_bo(gobj); 854 855 r = amdgpu_bo_reserve(robj, false); 856 if (unlikely(r)) 857 goto out; 858 859 switch (args->op) { 860 case AMDGPU_GEM_OP_GET_GEM_CREATE_INFO: { 861 struct drm_amdgpu_gem_create_in info; 862 void __user *out = u64_to_user_ptr(args->value); 863 864 info.bo_size = robj->tbo.base.size; 865 info.alignment = robj->tbo.page_alignment << PAGE_SHIFT; 866 info.domains = robj->preferred_domains; 867 info.domain_flags = robj->flags; 868 amdgpu_bo_unreserve(robj); 869 if (copy_to_user(out, &info, sizeof(info))) 870 r = -EFAULT; 871 break; 872 } 873 case AMDGPU_GEM_OP_SET_PLACEMENT: 874 if (robj->tbo.base.import_attach && 875 args->value & AMDGPU_GEM_DOMAIN_VRAM) { 876 r = -EINVAL; 877 amdgpu_bo_unreserve(robj); 878 break; 879 } 880 if (amdgpu_ttm_tt_get_usermm(robj->tbo.ttm)) { 881 r = -EPERM; 882 amdgpu_bo_unreserve(robj); 883 break; 884 } 885 for (base = robj->vm_bo; base; base = base->next) 886 if (amdgpu_xgmi_same_hive(amdgpu_ttm_adev(robj->tbo.bdev), 887 amdgpu_ttm_adev(base->vm->root.bo->tbo.bdev))) { 888 r = -EINVAL; 889 amdgpu_bo_unreserve(robj); 890 goto out; 891 } 892 893 894 robj->preferred_domains = args->value & (AMDGPU_GEM_DOMAIN_VRAM | 895 AMDGPU_GEM_DOMAIN_GTT | 896 AMDGPU_GEM_DOMAIN_CPU); 897 robj->allowed_domains = robj->preferred_domains; 898 if (robj->allowed_domains == AMDGPU_GEM_DOMAIN_VRAM) 899 robj->allowed_domains |= AMDGPU_GEM_DOMAIN_GTT; 900 901 if (robj->flags & AMDGPU_GEM_CREATE_VM_ALWAYS_VALID) 902 amdgpu_vm_bo_invalidate(adev, robj, true); 903 904 amdgpu_bo_unreserve(robj); 905 break; 906 default: 907 amdgpu_bo_unreserve(robj); 908 r = -EINVAL; 909 } 910 911 out: 912 drm_gem_object_put(gobj); 913 return r; 914 } 915 916 static int amdgpu_gem_align_pitch(struct amdgpu_device *adev, 917 int width, 918 int cpp, 919 bool tiled) 920 { 921 int aligned = width; 922 int pitch_mask = 0; 923 924 switch (cpp) { 925 case 1: 926 pitch_mask = 255; 927 break; 928 case 2: 929 pitch_mask = 127; 930 break; 931 case 3: 932 case 4: 933 pitch_mask = 63; 934 break; 935 } 936 937 aligned += pitch_mask; 938 aligned &= ~pitch_mask; 939 return aligned * cpp; 940 } 941 942 int amdgpu_mode_dumb_create(struct drm_file *file_priv, 943 struct drm_device *dev, 944 struct drm_mode_create_dumb *args) 945 { 946 struct amdgpu_device *adev = drm_to_adev(dev); 947 struct amdgpu_fpriv *fpriv = file_priv->driver_priv; 948 struct drm_gem_object *gobj; 949 uint32_t handle; 950 u64 flags = AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED | 951 AMDGPU_GEM_CREATE_CPU_GTT_USWC | 952 AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS; 953 u32 domain; 954 int r; 955 956 /* 957 * The buffer returned from this function should be cleared, but 958 * it can only be done if the ring is enabled or we'll fail to 959 * create the buffer. 960 */ 961 if (adev->mman.buffer_funcs_enabled) 962 flags |= AMDGPU_GEM_CREATE_VRAM_CLEARED; 963 964 args->pitch = amdgpu_gem_align_pitch(adev, args->width, 965 DIV_ROUND_UP(args->bpp, 8), 0); 966 args->size = (u64)args->pitch * args->height; 967 args->size = ALIGN(args->size, PAGE_SIZE); 968 domain = amdgpu_bo_get_preferred_domain(adev, 969 amdgpu_display_supported_domains(adev, flags)); 970 r = amdgpu_gem_object_create(adev, args->size, 0, domain, flags, 971 ttm_bo_type_device, NULL, &gobj, fpriv->xcp_id + 1); 972 if (r) 973 return -ENOMEM; 974 975 r = drm_gem_handle_create(file_priv, gobj, &handle); 976 /* drop reference from allocate - handle holds it now */ 977 drm_gem_object_put(gobj); 978 if (r) 979 return r; 980 981 args->handle = handle; 982 return 0; 983 } 984 985 #if defined(CONFIG_DEBUG_FS) 986 static int amdgpu_debugfs_gem_info_show(struct seq_file *m, void *unused) 987 { 988 struct amdgpu_device *adev = m->private; 989 struct drm_device *dev = adev_to_drm(adev); 990 struct drm_file *file; 991 int r; 992 993 r = mutex_lock_interruptible(&dev->filelist_mutex); 994 if (r) 995 return r; 996 997 list_for_each_entry(file, &dev->filelist, lhead) { 998 struct task_struct *task; 999 struct drm_gem_object *gobj; 1000 struct pid *pid; 1001 int id; 1002 1003 /* 1004 * Although we have a valid reference on file->pid, that does 1005 * not guarantee that the task_struct who called get_pid() is 1006 * still alive (e.g. get_pid(current) => fork() => exit()). 1007 * Therefore, we need to protect this ->comm access using RCU. 1008 */ 1009 rcu_read_lock(); 1010 pid = rcu_dereference(file->pid); 1011 task = pid_task(pid, PIDTYPE_TGID); 1012 seq_printf(m, "pid %8d command %s:\n", pid_nr(pid), 1013 task ? task->comm : "<unknown>"); 1014 rcu_read_unlock(); 1015 1016 spin_lock(&file->table_lock); 1017 idr_for_each_entry(&file->object_idr, gobj, id) { 1018 struct amdgpu_bo *bo = gem_to_amdgpu_bo(gobj); 1019 1020 amdgpu_bo_print_info(id, bo, m); 1021 } 1022 spin_unlock(&file->table_lock); 1023 } 1024 1025 mutex_unlock(&dev->filelist_mutex); 1026 return 0; 1027 } 1028 1029 DEFINE_SHOW_ATTRIBUTE(amdgpu_debugfs_gem_info); 1030 1031 #endif 1032 1033 void amdgpu_debugfs_gem_init(struct amdgpu_device *adev) 1034 { 1035 #if defined(CONFIG_DEBUG_FS) 1036 struct drm_minor *minor = adev_to_drm(adev)->primary; 1037 struct dentry *root = minor->debugfs_root; 1038 1039 debugfs_create_file("amdgpu_gem_info", 0444, root, adev, 1040 &amdgpu_debugfs_gem_info_fops); 1041 #endif 1042 } 1043