xref: /linux/drivers/gpu/drm/amd/amdgpu/amdgpu_gem.c (revision b56bc81078e96d5af984b929da5433c2cf776206)
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 #include <drm/drm_syncobj.h>
40 
41 #include "amdgpu.h"
42 #include "amdgpu_display.h"
43 #include "amdgpu_dma_buf.h"
44 #include "amdgpu_hmm.h"
45 #include "amdgpu_xgmi.h"
46 #include "amdgpu_vm.h"
47 
48 static int
49 amdgpu_gem_add_input_fence(struct drm_file *filp,
50 			   uint64_t syncobj_handles_array,
51 			   uint32_t num_syncobj_handles)
52 {
53 	struct dma_fence *fence;
54 	uint32_t *syncobj_handles;
55 	int ret, i;
56 
57 	if (!num_syncobj_handles)
58 		return 0;
59 
60 	syncobj_handles = memdup_user(u64_to_user_ptr(syncobj_handles_array),
61 				      size_mul(sizeof(uint32_t), num_syncobj_handles));
62 	if (IS_ERR(syncobj_handles))
63 		return PTR_ERR(syncobj_handles);
64 
65 	for (i = 0; i < num_syncobj_handles; i++) {
66 
67 		if (!syncobj_handles[i]) {
68 			ret = -EINVAL;
69 			goto free_memdup;
70 		}
71 
72 		ret = drm_syncobj_find_fence(filp, syncobj_handles[i], 0, 0, &fence);
73 		if (ret)
74 			goto free_memdup;
75 
76 		dma_fence_wait(fence, false);
77 
78 		/* TODO: optimize async handling */
79 		dma_fence_put(fence);
80 	}
81 
82 free_memdup:
83 	kfree(syncobj_handles);
84 	return ret;
85 }
86 
87 static int
88 amdgpu_gem_update_timeline_node(struct drm_file *filp,
89 				uint32_t syncobj_handle,
90 				uint64_t point,
91 				struct drm_syncobj **syncobj,
92 				struct dma_fence_chain **chain)
93 {
94 	if (!syncobj_handle)
95 		return 0;
96 
97 	/* Find the sync object */
98 	*syncobj = drm_syncobj_find(filp, syncobj_handle);
99 	if (!*syncobj)
100 		return -ENOENT;
101 
102 	if (!point)
103 		return 0;
104 
105 	/* Allocate the chain node */
106 	*chain = dma_fence_chain_alloc();
107 	if (!*chain) {
108 		drm_syncobj_put(*syncobj);
109 		return -ENOMEM;
110 	}
111 
112 	return 0;
113 }
114 
115 static void
116 amdgpu_gem_update_bo_mapping(struct drm_file *filp,
117 			     struct amdgpu_bo_va *bo_va,
118 			     uint32_t operation,
119 			     uint64_t point,
120 			     struct dma_fence *fence,
121 			     struct drm_syncobj *syncobj,
122 			     struct dma_fence_chain *chain)
123 {
124 	struct amdgpu_bo *bo = bo_va ? bo_va->base.bo : NULL;
125 	struct amdgpu_fpriv *fpriv = filp->driver_priv;
126 	struct amdgpu_vm *vm = &fpriv->vm;
127 	struct dma_fence *last_update;
128 
129 	if (!syncobj)
130 		return;
131 
132 	/* Find the last update fence */
133 	switch (operation) {
134 	case AMDGPU_VA_OP_MAP:
135 	case AMDGPU_VA_OP_REPLACE:
136 		if (bo && (bo->tbo.base.resv == vm->root.bo->tbo.base.resv))
137 			last_update = vm->last_update;
138 		else
139 			last_update = bo_va->last_pt_update;
140 		break;
141 	case AMDGPU_VA_OP_UNMAP:
142 	case AMDGPU_VA_OP_CLEAR:
143 		last_update = fence;
144 		break;
145 	default:
146 		return;
147 	}
148 
149 	/* Add fence to timeline */
150 	if (!point)
151 		drm_syncobj_replace_fence(syncobj, last_update);
152 	else
153 		drm_syncobj_add_point(syncobj, chain, last_update, point);
154 }
155 
156 static vm_fault_t amdgpu_gem_fault(struct vm_fault *vmf)
157 {
158 	struct ttm_buffer_object *bo = vmf->vma->vm_private_data;
159 	struct drm_device *ddev = bo->base.dev;
160 	vm_fault_t ret;
161 	int idx;
162 
163 	ret = ttm_bo_vm_reserve(bo, vmf);
164 	if (ret)
165 		return ret;
166 
167 	if (drm_dev_enter(ddev, &idx)) {
168 		ret = amdgpu_bo_fault_reserve_notify(bo);
169 		if (ret) {
170 			drm_dev_exit(idx);
171 			goto unlock;
172 		}
173 
174 		ret = ttm_bo_vm_fault_reserved(vmf, vmf->vma->vm_page_prot,
175 					       TTM_BO_VM_NUM_PREFAULT);
176 
177 		drm_dev_exit(idx);
178 	} else {
179 		ret = ttm_bo_vm_dummy_page(vmf, vmf->vma->vm_page_prot);
180 	}
181 	if (ret == VM_FAULT_RETRY && !(vmf->flags & FAULT_FLAG_RETRY_NOWAIT))
182 		return ret;
183 
184 unlock:
185 	dma_resv_unlock(bo->base.resv);
186 	return ret;
187 }
188 
189 static const struct vm_operations_struct amdgpu_gem_vm_ops = {
190 	.fault = amdgpu_gem_fault,
191 	.open = ttm_bo_vm_open,
192 	.close = ttm_bo_vm_close,
193 	.access = ttm_bo_vm_access
194 };
195 
196 static void amdgpu_gem_object_free(struct drm_gem_object *gobj)
197 {
198 	struct amdgpu_bo *aobj = gem_to_amdgpu_bo(gobj);
199 
200 	amdgpu_hmm_unregister(aobj);
201 	ttm_bo_put(&aobj->tbo);
202 }
203 
204 int amdgpu_gem_object_create(struct amdgpu_device *adev, unsigned long size,
205 			     int alignment, u32 initial_domain,
206 			     u64 flags, enum ttm_bo_type type,
207 			     struct dma_resv *resv,
208 			     struct drm_gem_object **obj, int8_t xcp_id_plus1)
209 {
210 	struct amdgpu_bo *bo;
211 	struct amdgpu_bo_user *ubo;
212 	struct amdgpu_bo_param bp;
213 	int r;
214 
215 	memset(&bp, 0, sizeof(bp));
216 	*obj = NULL;
217 	flags |= AMDGPU_GEM_CREATE_VRAM_WIPE_ON_RELEASE;
218 
219 	bp.size = size;
220 	bp.byte_align = alignment;
221 	bp.type = type;
222 	bp.resv = resv;
223 	bp.preferred_domain = initial_domain;
224 	bp.flags = flags;
225 	bp.domain = initial_domain;
226 	bp.bo_ptr_size = sizeof(struct amdgpu_bo);
227 	bp.xcp_id_plus1 = xcp_id_plus1;
228 
229 	r = amdgpu_bo_create_user(adev, &bp, &ubo);
230 	if (r)
231 		return r;
232 
233 	bo = &ubo->bo;
234 	*obj = &bo->tbo.base;
235 
236 	return 0;
237 }
238 
239 void amdgpu_gem_force_release(struct amdgpu_device *adev)
240 {
241 	struct drm_device *ddev = adev_to_drm(adev);
242 	struct drm_file *file;
243 
244 	mutex_lock(&ddev->filelist_mutex);
245 
246 	list_for_each_entry(file, &ddev->filelist, lhead) {
247 		struct drm_gem_object *gobj;
248 		int handle;
249 
250 		WARN_ONCE(1, "Still active user space clients!\n");
251 		spin_lock(&file->table_lock);
252 		idr_for_each_entry(&file->object_idr, gobj, handle) {
253 			WARN_ONCE(1, "And also active allocations!\n");
254 			drm_gem_object_put(gobj);
255 		}
256 		idr_destroy(&file->object_idr);
257 		spin_unlock(&file->table_lock);
258 	}
259 
260 	mutex_unlock(&ddev->filelist_mutex);
261 }
262 
263 /*
264  * Call from drm_gem_handle_create which appear in both new and open ioctl
265  * case.
266  */
267 static int amdgpu_gem_object_open(struct drm_gem_object *obj,
268 				  struct drm_file *file_priv)
269 {
270 	struct amdgpu_bo *abo = gem_to_amdgpu_bo(obj);
271 	struct amdgpu_device *adev = amdgpu_ttm_adev(abo->tbo.bdev);
272 	struct amdgpu_fpriv *fpriv = file_priv->driver_priv;
273 	struct amdgpu_vm *vm = &fpriv->vm;
274 	struct amdgpu_bo_va *bo_va;
275 	struct mm_struct *mm;
276 	int r;
277 
278 	mm = amdgpu_ttm_tt_get_usermm(abo->tbo.ttm);
279 	if (mm && mm != current->mm)
280 		return -EPERM;
281 
282 	if (abo->flags & AMDGPU_GEM_CREATE_VM_ALWAYS_VALID &&
283 	    !amdgpu_vm_is_bo_always_valid(vm, abo))
284 		return -EPERM;
285 
286 	r = amdgpu_bo_reserve(abo, false);
287 	if (r)
288 		return r;
289 
290 	amdgpu_vm_bo_update_shared(abo);
291 	bo_va = amdgpu_vm_bo_find(vm, abo);
292 	if (!bo_va)
293 		bo_va = amdgpu_vm_bo_add(adev, vm, abo);
294 	else
295 		++bo_va->ref_count;
296 
297 	/* attach gfx eviction fence */
298 	r = amdgpu_eviction_fence_attach(&fpriv->evf_mgr, abo);
299 	if (r) {
300 		DRM_DEBUG_DRIVER("Failed to attach eviction fence to BO\n");
301 		amdgpu_bo_unreserve(abo);
302 		return r;
303 	}
304 
305 	amdgpu_bo_unreserve(abo);
306 
307 	/* Validate and add eviction fence to DMABuf imports with dynamic
308 	 * attachment in compute VMs. Re-validation will be done by
309 	 * amdgpu_vm_validate. Fences are on the reservation shared with the
310 	 * export, which is currently required to be validated and fenced
311 	 * already by amdgpu_amdkfd_gpuvm_restore_process_bos.
312 	 *
313 	 * Nested locking below for the case that a GEM object is opened in
314 	 * kfd_mem_export_dmabuf. Since the lock below is only taken for imports,
315 	 * but not for export, this is a different lock class that cannot lead to
316 	 * circular lock dependencies.
317 	 */
318 	if (!vm->is_compute_context || !vm->process_info)
319 		return 0;
320 	if (!drm_gem_is_imported(obj) ||
321 	    !dma_buf_is_dynamic(obj->import_attach->dmabuf))
322 		return 0;
323 	mutex_lock_nested(&vm->process_info->lock, 1);
324 	if (!WARN_ON(!vm->process_info->eviction_fence)) {
325 		r = amdgpu_amdkfd_bo_validate_and_fence(abo, AMDGPU_GEM_DOMAIN_GTT,
326 							&vm->process_info->eviction_fence->base);
327 		if (r) {
328 			struct amdgpu_task_info *ti = amdgpu_vm_get_task_info_vm(vm);
329 
330 			dev_warn(adev->dev, "validate_and_fence failed: %d\n", r);
331 			if (ti) {
332 				dev_warn(adev->dev, "pid %d\n", ti->task.pid);
333 				amdgpu_vm_put_task_info(ti);
334 			}
335 		}
336 	}
337 	mutex_unlock(&vm->process_info->lock);
338 
339 	return r;
340 }
341 
342 static void amdgpu_gem_object_close(struct drm_gem_object *obj,
343 				    struct drm_file *file_priv)
344 {
345 	struct amdgpu_bo *bo = gem_to_amdgpu_bo(obj);
346 	struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
347 	struct amdgpu_fpriv *fpriv = file_priv->driver_priv;
348 	struct amdgpu_vm *vm = &fpriv->vm;
349 
350 	struct dma_fence *fence = NULL;
351 	struct amdgpu_bo_va *bo_va;
352 	struct drm_exec exec;
353 	long r;
354 
355 	drm_exec_init(&exec, DRM_EXEC_IGNORE_DUPLICATES, 0);
356 	drm_exec_until_all_locked(&exec) {
357 		r = drm_exec_prepare_obj(&exec, &bo->tbo.base, 1);
358 		drm_exec_retry_on_contention(&exec);
359 		if (unlikely(r))
360 			goto out_unlock;
361 
362 		r = amdgpu_vm_lock_pd(vm, &exec, 0);
363 		drm_exec_retry_on_contention(&exec);
364 		if (unlikely(r))
365 			goto out_unlock;
366 	}
367 
368 	if (!amdgpu_vm_is_bo_always_valid(vm, bo))
369 		amdgpu_eviction_fence_detach(&fpriv->evf_mgr, bo);
370 
371 	bo_va = amdgpu_vm_bo_find(vm, bo);
372 	if (!bo_va || --bo_va->ref_count)
373 		goto out_unlock;
374 
375 	amdgpu_vm_bo_del(adev, bo_va);
376 	amdgpu_vm_bo_update_shared(bo);
377 	if (!amdgpu_vm_ready(vm))
378 		goto out_unlock;
379 
380 	r = amdgpu_vm_clear_freed(adev, vm, &fence);
381 	if (unlikely(r < 0))
382 		dev_err(adev->dev, "failed to clear page "
383 			"tables on GEM object close (%ld)\n", r);
384 	if (r || !fence)
385 		goto out_unlock;
386 
387 	amdgpu_bo_fence(bo, fence, true);
388 	dma_fence_put(fence);
389 
390 out_unlock:
391 	if (r)
392 		dev_err(adev->dev, "leaking bo va (%ld)\n", r);
393 	drm_exec_fini(&exec);
394 }
395 
396 static int amdgpu_gem_object_mmap(struct drm_gem_object *obj, struct vm_area_struct *vma)
397 {
398 	struct amdgpu_bo *bo = gem_to_amdgpu_bo(obj);
399 
400 	if (amdgpu_ttm_tt_get_usermm(bo->tbo.ttm))
401 		return -EPERM;
402 	if (bo->flags & AMDGPU_GEM_CREATE_NO_CPU_ACCESS)
403 		return -EPERM;
404 
405 	/* Workaround for Thunk bug creating PROT_NONE,MAP_PRIVATE mappings
406 	 * for debugger access to invisible VRAM. Should have used MAP_SHARED
407 	 * instead. Clearing VM_MAYWRITE prevents the mapping from ever
408 	 * becoming writable and makes is_cow_mapping(vm_flags) false.
409 	 */
410 	if (is_cow_mapping(vma->vm_flags) &&
411 	    !(vma->vm_flags & VM_ACCESS_FLAGS))
412 		vm_flags_clear(vma, VM_MAYWRITE);
413 
414 	return drm_gem_ttm_mmap(obj, vma);
415 }
416 
417 const struct drm_gem_object_funcs amdgpu_gem_object_funcs = {
418 	.free = amdgpu_gem_object_free,
419 	.open = amdgpu_gem_object_open,
420 	.close = amdgpu_gem_object_close,
421 	.export = amdgpu_gem_prime_export,
422 	.vmap = drm_gem_ttm_vmap,
423 	.vunmap = drm_gem_ttm_vunmap,
424 	.mmap = amdgpu_gem_object_mmap,
425 	.vm_ops = &amdgpu_gem_vm_ops,
426 };
427 
428 /*
429  * GEM ioctls.
430  */
431 int amdgpu_gem_create_ioctl(struct drm_device *dev, void *data,
432 			    struct drm_file *filp)
433 {
434 	struct amdgpu_device *adev = drm_to_adev(dev);
435 	struct amdgpu_fpriv *fpriv = filp->driver_priv;
436 	struct amdgpu_vm *vm = &fpriv->vm;
437 	union drm_amdgpu_gem_create *args = data;
438 	uint64_t flags = args->in.domain_flags;
439 	uint64_t size = args->in.bo_size;
440 	struct dma_resv *resv = NULL;
441 	struct drm_gem_object *gobj;
442 	uint32_t handle, initial_domain;
443 	int r;
444 
445 	/* reject invalid gem flags */
446 	if (flags & ~AMDGPU_GEM_CREATE_SETTABLE_MASK)
447 		return -EINVAL;
448 
449 	/* reject invalid gem domains */
450 	if (args->in.domains & ~AMDGPU_GEM_DOMAIN_MASK)
451 		return -EINVAL;
452 
453 	if (!amdgpu_is_tmz(adev) && (flags & AMDGPU_GEM_CREATE_ENCRYPTED)) {
454 		DRM_NOTE_ONCE("Cannot allocate secure buffer since TMZ is disabled\n");
455 		return -EINVAL;
456 	}
457 
458 	/* always clear VRAM */
459 	flags |= AMDGPU_GEM_CREATE_VRAM_CLEARED;
460 
461 	if (args->in.domains & AMDGPU_GEM_DOMAIN_MMIO_REMAP)
462 		return -EINVAL;
463 
464 	/* create a gem object to contain this object in */
465 	if (args->in.domains & (AMDGPU_GEM_DOMAIN_GDS |
466 	    AMDGPU_GEM_DOMAIN_GWS | AMDGPU_GEM_DOMAIN_OA)) {
467 		if (flags & AMDGPU_GEM_CREATE_VM_ALWAYS_VALID) {
468 			/* if gds bo is created from user space, it must be
469 			 * passed to bo list
470 			 */
471 			DRM_ERROR("GDS bo cannot be per-vm-bo\n");
472 			return -EINVAL;
473 		}
474 		flags |= AMDGPU_GEM_CREATE_NO_CPU_ACCESS;
475 	}
476 
477 	if (flags & AMDGPU_GEM_CREATE_VM_ALWAYS_VALID) {
478 		r = amdgpu_bo_reserve(vm->root.bo, false);
479 		if (r)
480 			return r;
481 
482 		resv = vm->root.bo->tbo.base.resv;
483 	}
484 
485 	initial_domain = (u32)(0xffffffff & args->in.domains);
486 retry:
487 	r = amdgpu_gem_object_create(adev, size, args->in.alignment,
488 				     initial_domain,
489 				     flags, ttm_bo_type_device, resv, &gobj, fpriv->xcp_id + 1);
490 	if (r && r != -ERESTARTSYS) {
491 		if (flags & AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED) {
492 			flags &= ~AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED;
493 			goto retry;
494 		}
495 
496 		if (initial_domain == AMDGPU_GEM_DOMAIN_VRAM) {
497 			initial_domain |= AMDGPU_GEM_DOMAIN_GTT;
498 			goto retry;
499 		}
500 		DRM_DEBUG("Failed to allocate GEM object (%llu, %d, %llu, %d)\n",
501 				size, initial_domain, args->in.alignment, r);
502 	}
503 
504 	if (flags & AMDGPU_GEM_CREATE_VM_ALWAYS_VALID) {
505 		if (!r) {
506 			struct amdgpu_bo *abo = gem_to_amdgpu_bo(gobj);
507 
508 			abo->parent = amdgpu_bo_ref(vm->root.bo);
509 		}
510 		amdgpu_bo_unreserve(vm->root.bo);
511 	}
512 	if (r)
513 		return r;
514 
515 	r = drm_gem_handle_create(filp, gobj, &handle);
516 	/* drop reference from allocate - handle holds it now */
517 	drm_gem_object_put(gobj);
518 	if (r)
519 		return r;
520 
521 	memset(args, 0, sizeof(*args));
522 	args->out.handle = handle;
523 	return 0;
524 }
525 
526 int amdgpu_gem_userptr_ioctl(struct drm_device *dev, void *data,
527 			     struct drm_file *filp)
528 {
529 	struct ttm_operation_ctx ctx = { true, false };
530 	struct amdgpu_device *adev = drm_to_adev(dev);
531 	struct drm_amdgpu_gem_userptr *args = data;
532 	struct amdgpu_fpriv *fpriv = filp->driver_priv;
533 	struct drm_gem_object *gobj;
534 	struct hmm_range *range;
535 	struct amdgpu_bo *bo;
536 	uint32_t handle;
537 	int r;
538 
539 	args->addr = untagged_addr(args->addr);
540 
541 	if (offset_in_page(args->addr | args->size))
542 		return -EINVAL;
543 
544 	/* reject unknown flag values */
545 	if (args->flags & ~(AMDGPU_GEM_USERPTR_READONLY |
546 	    AMDGPU_GEM_USERPTR_ANONONLY | AMDGPU_GEM_USERPTR_VALIDATE |
547 	    AMDGPU_GEM_USERPTR_REGISTER))
548 		return -EINVAL;
549 
550 	if (!(args->flags & AMDGPU_GEM_USERPTR_READONLY) &&
551 	     !(args->flags & AMDGPU_GEM_USERPTR_REGISTER)) {
552 
553 		/* if we want to write to it we must install a MMU notifier */
554 		return -EACCES;
555 	}
556 
557 	/* create a gem object to contain this object in */
558 	r = amdgpu_gem_object_create(adev, args->size, 0, AMDGPU_GEM_DOMAIN_CPU,
559 				     0, ttm_bo_type_device, NULL, &gobj, fpriv->xcp_id + 1);
560 	if (r)
561 		return r;
562 
563 	bo = gem_to_amdgpu_bo(gobj);
564 	bo->preferred_domains = AMDGPU_GEM_DOMAIN_GTT;
565 	bo->allowed_domains = AMDGPU_GEM_DOMAIN_GTT;
566 	r = amdgpu_ttm_tt_set_userptr(&bo->tbo, args->addr, args->flags);
567 	if (r)
568 		goto release_object;
569 
570 	r = amdgpu_hmm_register(bo, args->addr);
571 	if (r)
572 		goto release_object;
573 
574 	if (args->flags & AMDGPU_GEM_USERPTR_VALIDATE) {
575 		r = amdgpu_ttm_tt_get_user_pages(bo, bo->tbo.ttm->pages,
576 						 &range);
577 		if (r)
578 			goto release_object;
579 
580 		r = amdgpu_bo_reserve(bo, true);
581 		if (r)
582 			goto user_pages_done;
583 
584 		amdgpu_bo_placement_from_domain(bo, AMDGPU_GEM_DOMAIN_GTT);
585 		r = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
586 		amdgpu_bo_unreserve(bo);
587 		if (r)
588 			goto user_pages_done;
589 	}
590 
591 	r = drm_gem_handle_create(filp, gobj, &handle);
592 	if (r)
593 		goto user_pages_done;
594 
595 	args->handle = handle;
596 
597 user_pages_done:
598 	if (args->flags & AMDGPU_GEM_USERPTR_VALIDATE)
599 		amdgpu_ttm_tt_get_user_pages_done(bo->tbo.ttm, range);
600 
601 release_object:
602 	drm_gem_object_put(gobj);
603 
604 	return r;
605 }
606 
607 int amdgpu_mode_dumb_mmap(struct drm_file *filp,
608 			  struct drm_device *dev,
609 			  uint32_t handle, uint64_t *offset_p)
610 {
611 	struct drm_gem_object *gobj;
612 	struct amdgpu_bo *robj;
613 
614 	gobj = drm_gem_object_lookup(filp, handle);
615 	if (!gobj)
616 		return -ENOENT;
617 
618 	robj = gem_to_amdgpu_bo(gobj);
619 	if (amdgpu_ttm_tt_get_usermm(robj->tbo.ttm) ||
620 	    (robj->flags & AMDGPU_GEM_CREATE_NO_CPU_ACCESS)) {
621 		drm_gem_object_put(gobj);
622 		return -EPERM;
623 	}
624 	*offset_p = amdgpu_bo_mmap_offset(robj);
625 	drm_gem_object_put(gobj);
626 	return 0;
627 }
628 
629 int amdgpu_gem_mmap_ioctl(struct drm_device *dev, void *data,
630 			  struct drm_file *filp)
631 {
632 	union drm_amdgpu_gem_mmap *args = data;
633 	uint32_t handle = args->in.handle;
634 
635 	memset(args, 0, sizeof(*args));
636 	return amdgpu_mode_dumb_mmap(filp, dev, handle, &args->out.addr_ptr);
637 }
638 
639 /**
640  * amdgpu_gem_timeout - calculate jiffies timeout from absolute value
641  *
642  * @timeout_ns: timeout in ns
643  *
644  * Calculate the timeout in jiffies from an absolute timeout in ns.
645  */
646 unsigned long amdgpu_gem_timeout(uint64_t timeout_ns)
647 {
648 	unsigned long timeout_jiffies;
649 	ktime_t timeout;
650 
651 	/* clamp timeout if it's to large */
652 	if (((int64_t)timeout_ns) < 0)
653 		return MAX_SCHEDULE_TIMEOUT;
654 
655 	timeout = ktime_sub(ns_to_ktime(timeout_ns), ktime_get());
656 	if (ktime_to_ns(timeout) < 0)
657 		return 0;
658 
659 	timeout_jiffies = nsecs_to_jiffies(ktime_to_ns(timeout));
660 	/*  clamp timeout to avoid unsigned-> signed overflow */
661 	if (timeout_jiffies > MAX_SCHEDULE_TIMEOUT)
662 		return MAX_SCHEDULE_TIMEOUT - 1;
663 
664 	return timeout_jiffies;
665 }
666 
667 int amdgpu_gem_wait_idle_ioctl(struct drm_device *dev, void *data,
668 			      struct drm_file *filp)
669 {
670 	union drm_amdgpu_gem_wait_idle *args = data;
671 	struct drm_gem_object *gobj;
672 	struct amdgpu_bo *robj;
673 	uint32_t handle = args->in.handle;
674 	unsigned long timeout = amdgpu_gem_timeout(args->in.timeout);
675 	int r = 0;
676 	long ret;
677 
678 	gobj = drm_gem_object_lookup(filp, handle);
679 	if (!gobj)
680 		return -ENOENT;
681 
682 	robj = gem_to_amdgpu_bo(gobj);
683 	ret = dma_resv_wait_timeout(robj->tbo.base.resv, DMA_RESV_USAGE_READ,
684 				    true, timeout);
685 
686 	/* ret == 0 means not signaled,
687 	 * ret > 0 means signaled
688 	 * ret < 0 means interrupted before timeout
689 	 */
690 	if (ret >= 0) {
691 		memset(args, 0, sizeof(*args));
692 		args->out.status = (ret == 0);
693 	} else
694 		r = ret;
695 
696 	drm_gem_object_put(gobj);
697 	return r;
698 }
699 
700 int amdgpu_gem_metadata_ioctl(struct drm_device *dev, void *data,
701 				struct drm_file *filp)
702 {
703 	struct drm_amdgpu_gem_metadata *args = data;
704 	struct drm_gem_object *gobj;
705 	struct amdgpu_bo *robj;
706 	int r = -1;
707 
708 	DRM_DEBUG("%d\n", args->handle);
709 	gobj = drm_gem_object_lookup(filp, args->handle);
710 	if (gobj == NULL)
711 		return -ENOENT;
712 	robj = gem_to_amdgpu_bo(gobj);
713 
714 	r = amdgpu_bo_reserve(robj, false);
715 	if (unlikely(r != 0))
716 		goto out;
717 
718 	if (args->op == AMDGPU_GEM_METADATA_OP_GET_METADATA) {
719 		amdgpu_bo_get_tiling_flags(robj, &args->data.tiling_info);
720 		r = amdgpu_bo_get_metadata(robj, args->data.data,
721 					   sizeof(args->data.data),
722 					   &args->data.data_size_bytes,
723 					   &args->data.flags);
724 	} else if (args->op == AMDGPU_GEM_METADATA_OP_SET_METADATA) {
725 		if (args->data.data_size_bytes > sizeof(args->data.data)) {
726 			r = -EINVAL;
727 			goto unreserve;
728 		}
729 		r = amdgpu_bo_set_tiling_flags(robj, args->data.tiling_info);
730 		if (!r)
731 			r = amdgpu_bo_set_metadata(robj, args->data.data,
732 						   args->data.data_size_bytes,
733 						   args->data.flags);
734 	}
735 
736 unreserve:
737 	amdgpu_bo_unreserve(robj);
738 out:
739 	drm_gem_object_put(gobj);
740 	return r;
741 }
742 
743 /**
744  * amdgpu_gem_va_update_vm -update the bo_va in its VM
745  *
746  * @adev: amdgpu_device pointer
747  * @vm: vm to update
748  * @bo_va: bo_va to update
749  * @operation: map, unmap or clear
750  *
751  * Update the bo_va directly after setting its address. Errors are not
752  * vital here, so they are not reported back to userspace.
753  *
754  * Returns resulting fence if freed BO(s) got cleared from the PT.
755  * otherwise stub fence in case of error.
756  */
757 static struct dma_fence *
758 amdgpu_gem_va_update_vm(struct amdgpu_device *adev,
759 			struct amdgpu_vm *vm,
760 			struct amdgpu_bo_va *bo_va,
761 			uint32_t operation)
762 {
763 	struct dma_fence *fence = dma_fence_get_stub();
764 	int r;
765 
766 	if (!amdgpu_vm_ready(vm))
767 		return fence;
768 
769 	r = amdgpu_vm_clear_freed(adev, vm, &fence);
770 	if (r)
771 		goto error;
772 
773 	if (operation == AMDGPU_VA_OP_MAP ||
774 	    operation == AMDGPU_VA_OP_REPLACE) {
775 		r = amdgpu_vm_bo_update(adev, bo_va, false);
776 		if (r)
777 			goto error;
778 	}
779 
780 	r = amdgpu_vm_update_pdes(adev, vm, false);
781 
782 error:
783 	if (r && r != -ERESTARTSYS)
784 		DRM_ERROR("Couldn't update BO_VA (%d)\n", r);
785 
786 	return fence;
787 }
788 
789 int amdgpu_gem_va_ioctl(struct drm_device *dev, void *data,
790 			  struct drm_file *filp)
791 {
792 	const uint32_t valid_flags = AMDGPU_VM_DELAY_UPDATE |
793 		AMDGPU_VM_PAGE_READABLE | AMDGPU_VM_PAGE_WRITEABLE |
794 		AMDGPU_VM_PAGE_EXECUTABLE | AMDGPU_VM_MTYPE_MASK |
795 		AMDGPU_VM_PAGE_NOALLOC;
796 	const uint32_t prt_flags = AMDGPU_VM_DELAY_UPDATE |
797 		AMDGPU_VM_PAGE_PRT;
798 
799 	struct drm_amdgpu_gem_va *args = data;
800 	struct drm_gem_object *gobj;
801 	struct amdgpu_device *adev = drm_to_adev(dev);
802 	struct amdgpu_fpriv *fpriv = filp->driver_priv;
803 	struct amdgpu_bo *abo;
804 	struct amdgpu_bo_va *bo_va;
805 	struct drm_syncobj *timeline_syncobj = NULL;
806 	struct dma_fence_chain *timeline_chain = NULL;
807 	struct dma_fence *fence;
808 	struct drm_exec exec;
809 	uint64_t vm_size;
810 	int r = 0;
811 
812 	if (args->va_address < AMDGPU_VA_RESERVED_BOTTOM) {
813 		dev_dbg(dev->dev,
814 			"va_address 0x%llx is in reserved area 0x%llx\n",
815 			args->va_address, AMDGPU_VA_RESERVED_BOTTOM);
816 		return -EINVAL;
817 	}
818 
819 	if (args->va_address >= AMDGPU_GMC_HOLE_START &&
820 	    args->va_address < AMDGPU_GMC_HOLE_END) {
821 		dev_dbg(dev->dev,
822 			"va_address 0x%llx is in VA hole 0x%llx-0x%llx\n",
823 			args->va_address, AMDGPU_GMC_HOLE_START,
824 			AMDGPU_GMC_HOLE_END);
825 		return -EINVAL;
826 	}
827 
828 	args->va_address &= AMDGPU_GMC_HOLE_MASK;
829 
830 	vm_size = adev->vm_manager.max_pfn * AMDGPU_GPU_PAGE_SIZE;
831 	vm_size -= AMDGPU_VA_RESERVED_TOP;
832 	if (args->va_address + args->map_size > vm_size) {
833 		dev_dbg(dev->dev,
834 			"va_address 0x%llx is in top reserved area 0x%llx\n",
835 			args->va_address + args->map_size, vm_size);
836 		return -EINVAL;
837 	}
838 
839 	if ((args->flags & ~valid_flags) && (args->flags & ~prt_flags)) {
840 		dev_dbg(dev->dev, "invalid flags combination 0x%08X\n",
841 			args->flags);
842 		return -EINVAL;
843 	}
844 
845 	switch (args->operation) {
846 	case AMDGPU_VA_OP_MAP:
847 	case AMDGPU_VA_OP_UNMAP:
848 	case AMDGPU_VA_OP_CLEAR:
849 	case AMDGPU_VA_OP_REPLACE:
850 		break;
851 	default:
852 		dev_dbg(dev->dev, "unsupported operation %d\n",
853 			args->operation);
854 		return -EINVAL;
855 	}
856 
857 	if ((args->operation != AMDGPU_VA_OP_CLEAR) &&
858 	    !(args->flags & AMDGPU_VM_PAGE_PRT)) {
859 		gobj = drm_gem_object_lookup(filp, args->handle);
860 		if (gobj == NULL)
861 			return -ENOENT;
862 		abo = gem_to_amdgpu_bo(gobj);
863 	} else {
864 		gobj = NULL;
865 		abo = NULL;
866 	}
867 
868 	r = amdgpu_gem_add_input_fence(filp,
869 				       args->input_fence_syncobj_handles,
870 				       args->num_syncobj_handles);
871 	if (r)
872 		goto error_put_gobj;
873 
874 	drm_exec_init(&exec, DRM_EXEC_INTERRUPTIBLE_WAIT |
875 		      DRM_EXEC_IGNORE_DUPLICATES, 0);
876 	drm_exec_until_all_locked(&exec) {
877 		if (gobj) {
878 			r = drm_exec_lock_obj(&exec, gobj);
879 			drm_exec_retry_on_contention(&exec);
880 			if (unlikely(r))
881 				goto error;
882 		}
883 
884 		r = amdgpu_vm_lock_pd(&fpriv->vm, &exec, 2);
885 		drm_exec_retry_on_contention(&exec);
886 		if (unlikely(r))
887 			goto error;
888 	}
889 
890 	if (abo) {
891 		bo_va = amdgpu_vm_bo_find(&fpriv->vm, abo);
892 		if (!bo_va) {
893 			r = -ENOENT;
894 			goto error;
895 		}
896 	} else if (args->operation != AMDGPU_VA_OP_CLEAR) {
897 		bo_va = fpriv->prt_va;
898 	} else {
899 		bo_va = NULL;
900 	}
901 
902 	r = amdgpu_gem_update_timeline_node(filp,
903 					    args->vm_timeline_syncobj_out,
904 					    args->vm_timeline_point,
905 					    &timeline_syncobj,
906 					    &timeline_chain);
907 	if (r)
908 		goto error;
909 
910 	switch (args->operation) {
911 	case AMDGPU_VA_OP_MAP:
912 		r = amdgpu_vm_bo_map(adev, bo_va, args->va_address,
913 				     args->offset_in_bo, args->map_size,
914 				     args->flags);
915 		break;
916 	case AMDGPU_VA_OP_UNMAP:
917 		r = amdgpu_vm_bo_unmap(adev, bo_va, args->va_address);
918 		break;
919 
920 	case AMDGPU_VA_OP_CLEAR:
921 		r = amdgpu_vm_bo_clear_mappings(adev, &fpriv->vm,
922 						args->va_address,
923 						args->map_size);
924 		break;
925 	case AMDGPU_VA_OP_REPLACE:
926 		r = amdgpu_vm_bo_replace_map(adev, bo_va, args->va_address,
927 					     args->offset_in_bo, args->map_size,
928 					     args->flags);
929 		break;
930 	default:
931 		break;
932 	}
933 	if (!r && !(args->flags & AMDGPU_VM_DELAY_UPDATE) && !adev->debug_vm) {
934 		fence = amdgpu_gem_va_update_vm(adev, &fpriv->vm, bo_va,
935 						args->operation);
936 
937 		if (timeline_syncobj)
938 			amdgpu_gem_update_bo_mapping(filp, bo_va,
939 					     args->operation,
940 					     args->vm_timeline_point,
941 					     fence, timeline_syncobj,
942 					     timeline_chain);
943 		else
944 			dma_fence_put(fence);
945 
946 	}
947 
948 error:
949 	drm_exec_fini(&exec);
950 error_put_gobj:
951 	drm_gem_object_put(gobj);
952 	return r;
953 }
954 
955 int amdgpu_gem_op_ioctl(struct drm_device *dev, void *data,
956 			struct drm_file *filp)
957 {
958 	struct drm_amdgpu_gem_op *args = data;
959 	struct drm_gem_object *gobj;
960 	struct amdgpu_vm_bo_base *base;
961 	struct amdgpu_bo *robj;
962 	struct drm_exec exec;
963 	struct amdgpu_fpriv *fpriv = filp->driver_priv;
964 	int r;
965 
966 	if (args->padding)
967 		return -EINVAL;
968 
969 	gobj = drm_gem_object_lookup(filp, args->handle);
970 	if (!gobj)
971 		return -ENOENT;
972 
973 	robj = gem_to_amdgpu_bo(gobj);
974 
975 	drm_exec_init(&exec, DRM_EXEC_INTERRUPTIBLE_WAIT |
976 			  DRM_EXEC_IGNORE_DUPLICATES, 0);
977 	drm_exec_until_all_locked(&exec) {
978 		r = drm_exec_lock_obj(&exec, gobj);
979 		drm_exec_retry_on_contention(&exec);
980 		if (r)
981 			goto out_exec;
982 
983 		if (args->op == AMDGPU_GEM_OP_GET_MAPPING_INFO) {
984 			r = amdgpu_vm_lock_pd(&fpriv->vm, &exec, 0);
985 			drm_exec_retry_on_contention(&exec);
986 			if (r)
987 				goto out_exec;
988 		}
989 	}
990 
991 	switch (args->op) {
992 	case AMDGPU_GEM_OP_GET_GEM_CREATE_INFO: {
993 		struct drm_amdgpu_gem_create_in info;
994 		void __user *out = u64_to_user_ptr(args->value);
995 
996 		info.bo_size = robj->tbo.base.size;
997 		info.alignment = robj->tbo.page_alignment << PAGE_SHIFT;
998 		info.domains = robj->preferred_domains;
999 		info.domain_flags = robj->flags;
1000 		drm_exec_fini(&exec);
1001 		if (copy_to_user(out, &info, sizeof(info)))
1002 			r = -EFAULT;
1003 		break;
1004 	}
1005 	case AMDGPU_GEM_OP_SET_PLACEMENT:
1006 		if (drm_gem_is_imported(&robj->tbo.base) &&
1007 		    args->value & AMDGPU_GEM_DOMAIN_VRAM) {
1008 			r = -EINVAL;
1009 			goto out_exec;
1010 		}
1011 		if (amdgpu_ttm_tt_get_usermm(robj->tbo.ttm)) {
1012 			r = -EPERM;
1013 			goto out_exec;
1014 		}
1015 		for (base = robj->vm_bo; base; base = base->next)
1016 			if (amdgpu_xgmi_same_hive(amdgpu_ttm_adev(robj->tbo.bdev),
1017 				amdgpu_ttm_adev(base->vm->root.bo->tbo.bdev))) {
1018 				r = -EINVAL;
1019 				goto out_exec;
1020 			}
1021 
1022 
1023 		robj->preferred_domains = args->value & (AMDGPU_GEM_DOMAIN_VRAM |
1024 							AMDGPU_GEM_DOMAIN_GTT |
1025 							AMDGPU_GEM_DOMAIN_CPU);
1026 		robj->allowed_domains = robj->preferred_domains;
1027 		if (robj->allowed_domains == AMDGPU_GEM_DOMAIN_VRAM)
1028 			robj->allowed_domains |= AMDGPU_GEM_DOMAIN_GTT;
1029 
1030 		if (robj->flags & AMDGPU_GEM_CREATE_VM_ALWAYS_VALID)
1031 			amdgpu_vm_bo_invalidate(robj, true);
1032 		drm_exec_fini(&exec);
1033 		break;
1034 	case AMDGPU_GEM_OP_GET_MAPPING_INFO: {
1035 		struct amdgpu_bo_va *bo_va = amdgpu_vm_bo_find(&fpriv->vm, robj);
1036 		struct drm_amdgpu_gem_vm_entry *vm_entries;
1037 		struct amdgpu_bo_va_mapping *mapping;
1038 		int num_mappings = 0;
1039 		/*
1040 		 * num_entries is set as an input to the size of the user-allocated array of
1041 		 * drm_amdgpu_gem_vm_entry stored at args->value.
1042 		 * num_entries is sent back as output as the number of mappings the bo has.
1043 		 * If that number is larger than the size of the array, the ioctl must
1044 		 * be retried.
1045 		 */
1046 		vm_entries = kvcalloc(args->num_entries, sizeof(*vm_entries), GFP_KERNEL);
1047 		if (!vm_entries)
1048 			return -ENOMEM;
1049 
1050 		amdgpu_vm_bo_va_for_each_valid_mapping(bo_va, mapping) {
1051 			if (num_mappings < args->num_entries) {
1052 				vm_entries[num_mappings].addr = mapping->start * AMDGPU_GPU_PAGE_SIZE;
1053 				vm_entries[num_mappings].size = (mapping->last - mapping->start + 1) * AMDGPU_GPU_PAGE_SIZE;
1054 				vm_entries[num_mappings].offset = mapping->offset;
1055 				vm_entries[num_mappings].flags = mapping->flags;
1056 			}
1057 			num_mappings += 1;
1058 		}
1059 
1060 		amdgpu_vm_bo_va_for_each_invalid_mapping(bo_va, mapping) {
1061 			if (num_mappings < args->num_entries) {
1062 				vm_entries[num_mappings].addr = mapping->start * AMDGPU_GPU_PAGE_SIZE;
1063 				vm_entries[num_mappings].size = (mapping->last - mapping->start + 1) * AMDGPU_GPU_PAGE_SIZE;
1064 				vm_entries[num_mappings].offset = mapping->offset;
1065 				vm_entries[num_mappings].flags = mapping->flags;
1066 			}
1067 			num_mappings += 1;
1068 		}
1069 
1070 		drm_exec_fini(&exec);
1071 
1072 		if (num_mappings > 0 && num_mappings <= args->num_entries)
1073 			if (copy_to_user(u64_to_user_ptr(args->value), vm_entries, num_mappings * sizeof(*vm_entries)))
1074 				r = -EFAULT;
1075 
1076 		args->num_entries = num_mappings;
1077 
1078 		kvfree(vm_entries);
1079 		break;
1080 	}
1081 	default:
1082 		drm_exec_fini(&exec);
1083 		r = -EINVAL;
1084 	}
1085 
1086 	drm_gem_object_put(gobj);
1087 	return r;
1088 out_exec:
1089 	drm_exec_fini(&exec);
1090 	drm_gem_object_put(gobj);
1091 	return r;
1092 }
1093 
1094 /**
1095  * amdgpu_gem_list_handles_ioctl - get information about a process' buffer objects
1096  *
1097  * @dev: drm device pointer
1098  * @data: drm_amdgpu_gem_list_handles
1099  * @filp: drm file pointer
1100  *
1101  * num_entries is set as an input to the size of the entries array.
1102  * num_entries is sent back as output as the number of bos in the process.
1103  * If that number is larger than the size of the array, the ioctl must
1104  * be retried.
1105  *
1106  * Returns:
1107  * 0 for success, -errno for errors.
1108  */
1109 int amdgpu_gem_list_handles_ioctl(struct drm_device *dev, void *data,
1110 				  struct drm_file *filp)
1111 {
1112 	struct drm_amdgpu_gem_list_handles *args = data;
1113 	struct drm_amdgpu_gem_list_handles_entry *bo_entries;
1114 	struct drm_gem_object *gobj;
1115 	int id, ret = 0;
1116 	int bo_index = 0;
1117 	int num_bos = 0;
1118 
1119 	spin_lock(&filp->table_lock);
1120 	idr_for_each_entry(&filp->object_idr, gobj, id)
1121 		num_bos += 1;
1122 	spin_unlock(&filp->table_lock);
1123 
1124 	if (args->num_entries < num_bos) {
1125 		args->num_entries = num_bos;
1126 		return 0;
1127 	}
1128 
1129 	if (num_bos == 0) {
1130 		args->num_entries = 0;
1131 		return 0;
1132 	}
1133 
1134 	bo_entries = kvcalloc(num_bos, sizeof(*bo_entries), GFP_KERNEL);
1135 	if (!bo_entries)
1136 		return -ENOMEM;
1137 
1138 	spin_lock(&filp->table_lock);
1139 	idr_for_each_entry(&filp->object_idr, gobj, id) {
1140 		struct amdgpu_bo *bo = gem_to_amdgpu_bo(gobj);
1141 		struct drm_amdgpu_gem_list_handles_entry *bo_entry;
1142 
1143 		if (bo_index >= num_bos) {
1144 			ret = -EAGAIN;
1145 			break;
1146 		}
1147 
1148 		bo_entry = &bo_entries[bo_index];
1149 
1150 		bo_entry->size = amdgpu_bo_size(bo);
1151 		bo_entry->alloc_flags = bo->flags & AMDGPU_GEM_CREATE_SETTABLE_MASK;
1152 		bo_entry->preferred_domains = bo->preferred_domains;
1153 		bo_entry->gem_handle = id;
1154 		bo_entry->alignment = bo->tbo.page_alignment;
1155 
1156 		if (bo->tbo.base.import_attach)
1157 			bo_entry->flags |= AMDGPU_GEM_LIST_HANDLES_FLAG_IS_IMPORT;
1158 
1159 		bo_index += 1;
1160 	}
1161 	spin_unlock(&filp->table_lock);
1162 
1163 	args->num_entries = bo_index;
1164 
1165 	if (!ret)
1166 		if (copy_to_user(u64_to_user_ptr(args->entries), bo_entries, num_bos * sizeof(*bo_entries)))
1167 			ret = -EFAULT;
1168 
1169 	kvfree(bo_entries);
1170 
1171 	return ret;
1172 }
1173 
1174 static int amdgpu_gem_align_pitch(struct amdgpu_device *adev,
1175 				  int width,
1176 				  int cpp,
1177 				  bool tiled)
1178 {
1179 	int aligned = width;
1180 	int pitch_mask = 0;
1181 
1182 	switch (cpp) {
1183 	case 1:
1184 		pitch_mask = 255;
1185 		break;
1186 	case 2:
1187 		pitch_mask = 127;
1188 		break;
1189 	case 3:
1190 	case 4:
1191 		pitch_mask = 63;
1192 		break;
1193 	}
1194 
1195 	aligned += pitch_mask;
1196 	aligned &= ~pitch_mask;
1197 	return aligned * cpp;
1198 }
1199 
1200 int amdgpu_mode_dumb_create(struct drm_file *file_priv,
1201 			    struct drm_device *dev,
1202 			    struct drm_mode_create_dumb *args)
1203 {
1204 	struct amdgpu_device *adev = drm_to_adev(dev);
1205 	struct amdgpu_fpriv *fpriv = file_priv->driver_priv;
1206 	struct drm_gem_object *gobj;
1207 	uint32_t handle;
1208 	u64 flags = AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED |
1209 		    AMDGPU_GEM_CREATE_CPU_GTT_USWC |
1210 		    AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS;
1211 	u32 domain;
1212 	int r;
1213 
1214 	/*
1215 	 * The buffer returned from this function should be cleared, but
1216 	 * it can only be done if the ring is enabled or we'll fail to
1217 	 * create the buffer.
1218 	 */
1219 	if (adev->mman.buffer_funcs_enabled)
1220 		flags |= AMDGPU_GEM_CREATE_VRAM_CLEARED;
1221 
1222 	args->pitch = amdgpu_gem_align_pitch(adev, args->width,
1223 					     DIV_ROUND_UP(args->bpp, 8), 0);
1224 	args->size = (u64)args->pitch * args->height;
1225 	args->size = ALIGN(args->size, PAGE_SIZE);
1226 	domain = amdgpu_bo_get_preferred_domain(adev,
1227 				amdgpu_display_supported_domains(adev, flags));
1228 	r = amdgpu_gem_object_create(adev, args->size, 0, domain, flags,
1229 				     ttm_bo_type_device, NULL, &gobj, fpriv->xcp_id + 1);
1230 	if (r)
1231 		return -ENOMEM;
1232 
1233 	r = drm_gem_handle_create(file_priv, gobj, &handle);
1234 	/* drop reference from allocate - handle holds it now */
1235 	drm_gem_object_put(gobj);
1236 	if (r)
1237 		return r;
1238 
1239 	args->handle = handle;
1240 	return 0;
1241 }
1242 
1243 #if defined(CONFIG_DEBUG_FS)
1244 static int amdgpu_debugfs_gem_info_show(struct seq_file *m, void *unused)
1245 {
1246 	struct amdgpu_device *adev = m->private;
1247 	struct drm_device *dev = adev_to_drm(adev);
1248 	struct drm_file *file;
1249 	int r;
1250 
1251 	r = mutex_lock_interruptible(&dev->filelist_mutex);
1252 	if (r)
1253 		return r;
1254 
1255 	list_for_each_entry(file, &dev->filelist, lhead) {
1256 		struct task_struct *task;
1257 		struct drm_gem_object *gobj;
1258 		struct pid *pid;
1259 		int id;
1260 
1261 		/*
1262 		 * Although we have a valid reference on file->pid, that does
1263 		 * not guarantee that the task_struct who called get_pid() is
1264 		 * still alive (e.g. get_pid(current) => fork() => exit()).
1265 		 * Therefore, we need to protect this ->comm access using RCU.
1266 		 */
1267 		rcu_read_lock();
1268 		pid = rcu_dereference(file->pid);
1269 		task = pid_task(pid, PIDTYPE_TGID);
1270 		seq_printf(m, "pid %8d command %s:\n", pid_nr(pid),
1271 			   task ? task->comm : "<unknown>");
1272 		rcu_read_unlock();
1273 
1274 		spin_lock(&file->table_lock);
1275 		idr_for_each_entry(&file->object_idr, gobj, id) {
1276 			struct amdgpu_bo *bo = gem_to_amdgpu_bo(gobj);
1277 
1278 			amdgpu_bo_print_info(id, bo, m);
1279 		}
1280 		spin_unlock(&file->table_lock);
1281 	}
1282 
1283 	mutex_unlock(&dev->filelist_mutex);
1284 	return 0;
1285 }
1286 
1287 DEFINE_SHOW_ATTRIBUTE(amdgpu_debugfs_gem_info);
1288 
1289 #endif
1290 
1291 void amdgpu_debugfs_gem_init(struct amdgpu_device *adev)
1292 {
1293 #if defined(CONFIG_DEBUG_FS)
1294 	struct drm_minor *minor = adev_to_drm(adev)->primary;
1295 	struct dentry *root = minor->debugfs_root;
1296 
1297 	debugfs_create_file("amdgpu_gem_info", 0444, root, adev,
1298 			    &amdgpu_debugfs_gem_info_fops);
1299 #endif
1300 }
1301