xref: /linux/drivers/gpu/drm/amd/amdgpu/amdgpu_gem.c (revision 569d7db70e5dcf13fbf072f10e9096577ac1e565)
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 
112 	bp.size = size;
113 	bp.byte_align = alignment;
114 	bp.type = type;
115 	bp.resv = resv;
116 	bp.preferred_domain = initial_domain;
117 	bp.flags = flags;
118 	bp.domain = initial_domain;
119 	bp.bo_ptr_size = sizeof(struct amdgpu_bo);
120 	bp.xcp_id_plus1 = xcp_id_plus1;
121 
122 	r = amdgpu_bo_create_user(adev, &bp, &ubo);
123 	if (r)
124 		return r;
125 
126 	bo = &ubo->bo;
127 	*obj = &bo->tbo.base;
128 	(*obj)->funcs = &amdgpu_gem_object_funcs;
129 
130 	return 0;
131 }
132 
133 void amdgpu_gem_force_release(struct amdgpu_device *adev)
134 {
135 	struct drm_device *ddev = adev_to_drm(adev);
136 	struct drm_file *file;
137 
138 	mutex_lock(&ddev->filelist_mutex);
139 
140 	list_for_each_entry(file, &ddev->filelist, lhead) {
141 		struct drm_gem_object *gobj;
142 		int handle;
143 
144 		WARN_ONCE(1, "Still active user space clients!\n");
145 		spin_lock(&file->table_lock);
146 		idr_for_each_entry(&file->object_idr, gobj, handle) {
147 			WARN_ONCE(1, "And also active allocations!\n");
148 			drm_gem_object_put(gobj);
149 		}
150 		idr_destroy(&file->object_idr);
151 		spin_unlock(&file->table_lock);
152 	}
153 
154 	mutex_unlock(&ddev->filelist_mutex);
155 }
156 
157 /*
158  * Call from drm_gem_handle_create which appear in both new and open ioctl
159  * case.
160  */
161 static int amdgpu_gem_object_open(struct drm_gem_object *obj,
162 				  struct drm_file *file_priv)
163 {
164 	struct amdgpu_bo *abo = gem_to_amdgpu_bo(obj);
165 	struct amdgpu_device *adev = amdgpu_ttm_adev(abo->tbo.bdev);
166 	struct amdgpu_fpriv *fpriv = file_priv->driver_priv;
167 	struct amdgpu_vm *vm = &fpriv->vm;
168 	struct amdgpu_bo_va *bo_va;
169 	struct mm_struct *mm;
170 	int r;
171 
172 	mm = amdgpu_ttm_tt_get_usermm(abo->tbo.ttm);
173 	if (mm && mm != current->mm)
174 		return -EPERM;
175 
176 	if (abo->flags & AMDGPU_GEM_CREATE_VM_ALWAYS_VALID &&
177 	    !amdgpu_vm_is_bo_always_valid(vm, abo))
178 		return -EPERM;
179 
180 	r = amdgpu_bo_reserve(abo, false);
181 	if (r)
182 		return r;
183 
184 	bo_va = amdgpu_vm_bo_find(vm, abo);
185 	if (!bo_va)
186 		bo_va = amdgpu_vm_bo_add(adev, vm, abo);
187 	else
188 		++bo_va->ref_count;
189 	amdgpu_bo_unreserve(abo);
190 
191 	/* Validate and add eviction fence to DMABuf imports with dynamic
192 	 * attachment in compute VMs. Re-validation will be done by
193 	 * amdgpu_vm_validate. Fences are on the reservation shared with the
194 	 * export, which is currently required to be validated and fenced
195 	 * already by amdgpu_amdkfd_gpuvm_restore_process_bos.
196 	 *
197 	 * Nested locking below for the case that a GEM object is opened in
198 	 * kfd_mem_export_dmabuf. Since the lock below is only taken for imports,
199 	 * but not for export, this is a different lock class that cannot lead to
200 	 * circular lock dependencies.
201 	 */
202 	if (!vm->is_compute_context || !vm->process_info)
203 		return 0;
204 	if (!obj->import_attach ||
205 	    !dma_buf_is_dynamic(obj->import_attach->dmabuf))
206 		return 0;
207 	mutex_lock_nested(&vm->process_info->lock, 1);
208 	if (!WARN_ON(!vm->process_info->eviction_fence)) {
209 		r = amdgpu_amdkfd_bo_validate_and_fence(abo, AMDGPU_GEM_DOMAIN_GTT,
210 							&vm->process_info->eviction_fence->base);
211 		if (r) {
212 			struct amdgpu_task_info *ti = amdgpu_vm_get_task_info_vm(vm);
213 
214 			dev_warn(adev->dev, "validate_and_fence failed: %d\n", r);
215 			if (ti) {
216 				dev_warn(adev->dev, "pid %d\n", ti->pid);
217 				amdgpu_vm_put_task_info(ti);
218 			}
219 		}
220 	}
221 	mutex_unlock(&vm->process_info->lock);
222 
223 	return r;
224 }
225 
226 static void amdgpu_gem_object_close(struct drm_gem_object *obj,
227 				    struct drm_file *file_priv)
228 {
229 	struct amdgpu_bo *bo = gem_to_amdgpu_bo(obj);
230 	struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
231 	struct amdgpu_fpriv *fpriv = file_priv->driver_priv;
232 	struct amdgpu_vm *vm = &fpriv->vm;
233 
234 	struct dma_fence *fence = NULL;
235 	struct amdgpu_bo_va *bo_va;
236 	struct drm_exec exec;
237 	long r;
238 
239 	drm_exec_init(&exec, DRM_EXEC_IGNORE_DUPLICATES, 0);
240 	drm_exec_until_all_locked(&exec) {
241 		r = drm_exec_prepare_obj(&exec, &bo->tbo.base, 1);
242 		drm_exec_retry_on_contention(&exec);
243 		if (unlikely(r))
244 			goto out_unlock;
245 
246 		r = amdgpu_vm_lock_pd(vm, &exec, 0);
247 		drm_exec_retry_on_contention(&exec);
248 		if (unlikely(r))
249 			goto out_unlock;
250 	}
251 
252 	bo_va = amdgpu_vm_bo_find(vm, bo);
253 	if (!bo_va || --bo_va->ref_count)
254 		goto out_unlock;
255 
256 	amdgpu_vm_bo_del(adev, bo_va);
257 	if (!amdgpu_vm_ready(vm))
258 		goto out_unlock;
259 
260 	r = amdgpu_vm_clear_freed(adev, vm, &fence);
261 	if (unlikely(r < 0))
262 		dev_err(adev->dev, "failed to clear page "
263 			"tables on GEM object close (%ld)\n", r);
264 	if (r || !fence)
265 		goto out_unlock;
266 
267 	amdgpu_bo_fence(bo, fence, true);
268 	dma_fence_put(fence);
269 
270 out_unlock:
271 	if (r)
272 		dev_err(adev->dev, "leaking bo va (%ld)\n", r);
273 	drm_exec_fini(&exec);
274 }
275 
276 static int amdgpu_gem_object_mmap(struct drm_gem_object *obj, struct vm_area_struct *vma)
277 {
278 	struct amdgpu_bo *bo = gem_to_amdgpu_bo(obj);
279 
280 	if (amdgpu_ttm_tt_get_usermm(bo->tbo.ttm))
281 		return -EPERM;
282 	if (bo->flags & AMDGPU_GEM_CREATE_NO_CPU_ACCESS)
283 		return -EPERM;
284 
285 	/* Workaround for Thunk bug creating PROT_NONE,MAP_PRIVATE mappings
286 	 * for debugger access to invisible VRAM. Should have used MAP_SHARED
287 	 * instead. Clearing VM_MAYWRITE prevents the mapping from ever
288 	 * becoming writable and makes is_cow_mapping(vm_flags) false.
289 	 */
290 	if (is_cow_mapping(vma->vm_flags) &&
291 	    !(vma->vm_flags & VM_ACCESS_FLAGS))
292 		vm_flags_clear(vma, VM_MAYWRITE);
293 
294 	return drm_gem_ttm_mmap(obj, vma);
295 }
296 
297 static const struct drm_gem_object_funcs amdgpu_gem_object_funcs = {
298 	.free = amdgpu_gem_object_free,
299 	.open = amdgpu_gem_object_open,
300 	.close = amdgpu_gem_object_close,
301 	.export = amdgpu_gem_prime_export,
302 	.vmap = drm_gem_ttm_vmap,
303 	.vunmap = drm_gem_ttm_vunmap,
304 	.mmap = amdgpu_gem_object_mmap,
305 	.vm_ops = &amdgpu_gem_vm_ops,
306 };
307 
308 /*
309  * GEM ioctls.
310  */
311 int amdgpu_gem_create_ioctl(struct drm_device *dev, void *data,
312 			    struct drm_file *filp)
313 {
314 	struct amdgpu_device *adev = drm_to_adev(dev);
315 	struct amdgpu_fpriv *fpriv = filp->driver_priv;
316 	struct amdgpu_vm *vm = &fpriv->vm;
317 	union drm_amdgpu_gem_create *args = data;
318 	uint64_t flags = args->in.domain_flags;
319 	uint64_t size = args->in.bo_size;
320 	struct dma_resv *resv = NULL;
321 	struct drm_gem_object *gobj;
322 	uint32_t handle, initial_domain;
323 	int r;
324 
325 	/* reject DOORBELLs until userspace code to use it is available */
326 	if (args->in.domains & AMDGPU_GEM_DOMAIN_DOORBELL)
327 		return -EINVAL;
328 
329 	/* reject invalid gem flags */
330 	if (flags & ~(AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED |
331 		      AMDGPU_GEM_CREATE_NO_CPU_ACCESS |
332 		      AMDGPU_GEM_CREATE_CPU_GTT_USWC |
333 		      AMDGPU_GEM_CREATE_VRAM_CLEARED |
334 		      AMDGPU_GEM_CREATE_VM_ALWAYS_VALID |
335 		      AMDGPU_GEM_CREATE_EXPLICIT_SYNC |
336 		      AMDGPU_GEM_CREATE_ENCRYPTED |
337 		      AMDGPU_GEM_CREATE_GFX12_DCC |
338 		      AMDGPU_GEM_CREATE_DISCARDABLE))
339 		return -EINVAL;
340 
341 	/* reject invalid gem domains */
342 	if (args->in.domains & ~AMDGPU_GEM_DOMAIN_MASK)
343 		return -EINVAL;
344 
345 	if ((flags & AMDGPU_GEM_CREATE_GFX12_DCC) &&
346 	    ((amdgpu_ip_version(adev, GC_HWIP, 0) < IP_VERSION(12, 0, 0)) ||
347 	     !(args->in.domains & AMDGPU_GEM_DOMAIN_VRAM)))
348 		return -EINVAL;
349 
350 	if (!amdgpu_is_tmz(adev) && (flags & AMDGPU_GEM_CREATE_ENCRYPTED)) {
351 		DRM_NOTE_ONCE("Cannot allocate secure buffer since TMZ is disabled\n");
352 		return -EINVAL;
353 	}
354 
355 	/* create a gem object to contain this object in */
356 	if (args->in.domains & (AMDGPU_GEM_DOMAIN_GDS |
357 	    AMDGPU_GEM_DOMAIN_GWS | AMDGPU_GEM_DOMAIN_OA)) {
358 		if (flags & AMDGPU_GEM_CREATE_VM_ALWAYS_VALID) {
359 			/* if gds bo is created from user space, it must be
360 			 * passed to bo list
361 			 */
362 			DRM_ERROR("GDS bo cannot be per-vm-bo\n");
363 			return -EINVAL;
364 		}
365 		flags |= AMDGPU_GEM_CREATE_NO_CPU_ACCESS;
366 	}
367 
368 	if (flags & AMDGPU_GEM_CREATE_VM_ALWAYS_VALID) {
369 		r = amdgpu_bo_reserve(vm->root.bo, false);
370 		if (r)
371 			return r;
372 
373 		resv = vm->root.bo->tbo.base.resv;
374 	}
375 
376 	initial_domain = (u32)(0xffffffff & args->in.domains);
377 retry:
378 	r = amdgpu_gem_object_create(adev, size, args->in.alignment,
379 				     initial_domain,
380 				     flags, ttm_bo_type_device, resv, &gobj, fpriv->xcp_id + 1);
381 	if (r && r != -ERESTARTSYS) {
382 		if (flags & AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED) {
383 			flags &= ~AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED;
384 			goto retry;
385 		}
386 
387 		if (initial_domain == AMDGPU_GEM_DOMAIN_VRAM) {
388 			initial_domain |= AMDGPU_GEM_DOMAIN_GTT;
389 			goto retry;
390 		}
391 		DRM_DEBUG("Failed to allocate GEM object (%llu, %d, %llu, %d)\n",
392 				size, initial_domain, args->in.alignment, r);
393 	}
394 
395 	if (flags & AMDGPU_GEM_CREATE_VM_ALWAYS_VALID) {
396 		if (!r) {
397 			struct amdgpu_bo *abo = gem_to_amdgpu_bo(gobj);
398 
399 			abo->parent = amdgpu_bo_ref(vm->root.bo);
400 		}
401 		amdgpu_bo_unreserve(vm->root.bo);
402 	}
403 	if (r)
404 		return r;
405 
406 	r = drm_gem_handle_create(filp, gobj, &handle);
407 	/* drop reference from allocate - handle holds it now */
408 	drm_gem_object_put(gobj);
409 	if (r)
410 		return r;
411 
412 	memset(args, 0, sizeof(*args));
413 	args->out.handle = handle;
414 	return 0;
415 }
416 
417 int amdgpu_gem_userptr_ioctl(struct drm_device *dev, void *data,
418 			     struct drm_file *filp)
419 {
420 	struct ttm_operation_ctx ctx = { true, false };
421 	struct amdgpu_device *adev = drm_to_adev(dev);
422 	struct drm_amdgpu_gem_userptr *args = data;
423 	struct amdgpu_fpriv *fpriv = filp->driver_priv;
424 	struct drm_gem_object *gobj;
425 	struct hmm_range *range;
426 	struct amdgpu_bo *bo;
427 	uint32_t handle;
428 	int r;
429 
430 	args->addr = untagged_addr(args->addr);
431 
432 	if (offset_in_page(args->addr | args->size))
433 		return -EINVAL;
434 
435 	/* reject unknown flag values */
436 	if (args->flags & ~(AMDGPU_GEM_USERPTR_READONLY |
437 	    AMDGPU_GEM_USERPTR_ANONONLY | AMDGPU_GEM_USERPTR_VALIDATE |
438 	    AMDGPU_GEM_USERPTR_REGISTER))
439 		return -EINVAL;
440 
441 	if (!(args->flags & AMDGPU_GEM_USERPTR_READONLY) &&
442 	     !(args->flags & AMDGPU_GEM_USERPTR_REGISTER)) {
443 
444 		/* if we want to write to it we must install a MMU notifier */
445 		return -EACCES;
446 	}
447 
448 	/* create a gem object to contain this object in */
449 	r = amdgpu_gem_object_create(adev, args->size, 0, AMDGPU_GEM_DOMAIN_CPU,
450 				     0, ttm_bo_type_device, NULL, &gobj, fpriv->xcp_id + 1);
451 	if (r)
452 		return r;
453 
454 	bo = gem_to_amdgpu_bo(gobj);
455 	bo->preferred_domains = AMDGPU_GEM_DOMAIN_GTT;
456 	bo->allowed_domains = AMDGPU_GEM_DOMAIN_GTT;
457 	r = amdgpu_ttm_tt_set_userptr(&bo->tbo, args->addr, args->flags);
458 	if (r)
459 		goto release_object;
460 
461 	r = amdgpu_hmm_register(bo, args->addr);
462 	if (r)
463 		goto release_object;
464 
465 	if (args->flags & AMDGPU_GEM_USERPTR_VALIDATE) {
466 		r = amdgpu_ttm_tt_get_user_pages(bo, bo->tbo.ttm->pages,
467 						 &range);
468 		if (r)
469 			goto release_object;
470 
471 		r = amdgpu_bo_reserve(bo, true);
472 		if (r)
473 			goto user_pages_done;
474 
475 		amdgpu_bo_placement_from_domain(bo, AMDGPU_GEM_DOMAIN_GTT);
476 		r = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
477 		amdgpu_bo_unreserve(bo);
478 		if (r)
479 			goto user_pages_done;
480 	}
481 
482 	r = drm_gem_handle_create(filp, gobj, &handle);
483 	if (r)
484 		goto user_pages_done;
485 
486 	args->handle = handle;
487 
488 user_pages_done:
489 	if (args->flags & AMDGPU_GEM_USERPTR_VALIDATE)
490 		amdgpu_ttm_tt_get_user_pages_done(bo->tbo.ttm, range);
491 
492 release_object:
493 	drm_gem_object_put(gobj);
494 
495 	return r;
496 }
497 
498 int amdgpu_mode_dumb_mmap(struct drm_file *filp,
499 			  struct drm_device *dev,
500 			  uint32_t handle, uint64_t *offset_p)
501 {
502 	struct drm_gem_object *gobj;
503 	struct amdgpu_bo *robj;
504 
505 	gobj = drm_gem_object_lookup(filp, handle);
506 	if (!gobj)
507 		return -ENOENT;
508 
509 	robj = gem_to_amdgpu_bo(gobj);
510 	if (amdgpu_ttm_tt_get_usermm(robj->tbo.ttm) ||
511 	    (robj->flags & AMDGPU_GEM_CREATE_NO_CPU_ACCESS)) {
512 		drm_gem_object_put(gobj);
513 		return -EPERM;
514 	}
515 	*offset_p = amdgpu_bo_mmap_offset(robj);
516 	drm_gem_object_put(gobj);
517 	return 0;
518 }
519 
520 int amdgpu_gem_mmap_ioctl(struct drm_device *dev, void *data,
521 			  struct drm_file *filp)
522 {
523 	union drm_amdgpu_gem_mmap *args = data;
524 	uint32_t handle = args->in.handle;
525 
526 	memset(args, 0, sizeof(*args));
527 	return amdgpu_mode_dumb_mmap(filp, dev, handle, &args->out.addr_ptr);
528 }
529 
530 /**
531  * amdgpu_gem_timeout - calculate jiffies timeout from absolute value
532  *
533  * @timeout_ns: timeout in ns
534  *
535  * Calculate the timeout in jiffies from an absolute timeout in ns.
536  */
537 unsigned long amdgpu_gem_timeout(uint64_t timeout_ns)
538 {
539 	unsigned long timeout_jiffies;
540 	ktime_t timeout;
541 
542 	/* clamp timeout if it's to large */
543 	if (((int64_t)timeout_ns) < 0)
544 		return MAX_SCHEDULE_TIMEOUT;
545 
546 	timeout = ktime_sub(ns_to_ktime(timeout_ns), ktime_get());
547 	if (ktime_to_ns(timeout) < 0)
548 		return 0;
549 
550 	timeout_jiffies = nsecs_to_jiffies(ktime_to_ns(timeout));
551 	/*  clamp timeout to avoid unsigned-> signed overflow */
552 	if (timeout_jiffies > MAX_SCHEDULE_TIMEOUT)
553 		return MAX_SCHEDULE_TIMEOUT - 1;
554 
555 	return timeout_jiffies;
556 }
557 
558 int amdgpu_gem_wait_idle_ioctl(struct drm_device *dev, void *data,
559 			      struct drm_file *filp)
560 {
561 	union drm_amdgpu_gem_wait_idle *args = data;
562 	struct drm_gem_object *gobj;
563 	struct amdgpu_bo *robj;
564 	uint32_t handle = args->in.handle;
565 	unsigned long timeout = amdgpu_gem_timeout(args->in.timeout);
566 	int r = 0;
567 	long ret;
568 
569 	gobj = drm_gem_object_lookup(filp, handle);
570 	if (!gobj)
571 		return -ENOENT;
572 
573 	robj = gem_to_amdgpu_bo(gobj);
574 	ret = dma_resv_wait_timeout(robj->tbo.base.resv, DMA_RESV_USAGE_READ,
575 				    true, timeout);
576 
577 	/* ret == 0 means not signaled,
578 	 * ret > 0 means signaled
579 	 * ret < 0 means interrupted before timeout
580 	 */
581 	if (ret >= 0) {
582 		memset(args, 0, sizeof(*args));
583 		args->out.status = (ret == 0);
584 	} else
585 		r = ret;
586 
587 	drm_gem_object_put(gobj);
588 	return r;
589 }
590 
591 int amdgpu_gem_metadata_ioctl(struct drm_device *dev, void *data,
592 				struct drm_file *filp)
593 {
594 	struct drm_amdgpu_gem_metadata *args = data;
595 	struct drm_gem_object *gobj;
596 	struct amdgpu_bo *robj;
597 	int r = -1;
598 
599 	DRM_DEBUG("%d\n", args->handle);
600 	gobj = drm_gem_object_lookup(filp, args->handle);
601 	if (gobj == NULL)
602 		return -ENOENT;
603 	robj = gem_to_amdgpu_bo(gobj);
604 
605 	r = amdgpu_bo_reserve(robj, false);
606 	if (unlikely(r != 0))
607 		goto out;
608 
609 	if (args->op == AMDGPU_GEM_METADATA_OP_GET_METADATA) {
610 		amdgpu_bo_get_tiling_flags(robj, &args->data.tiling_info);
611 		r = amdgpu_bo_get_metadata(robj, args->data.data,
612 					   sizeof(args->data.data),
613 					   &args->data.data_size_bytes,
614 					   &args->data.flags);
615 	} else if (args->op == AMDGPU_GEM_METADATA_OP_SET_METADATA) {
616 		if (args->data.data_size_bytes > sizeof(args->data.data)) {
617 			r = -EINVAL;
618 			goto unreserve;
619 		}
620 		r = amdgpu_bo_set_tiling_flags(robj, args->data.tiling_info);
621 		if (!r)
622 			r = amdgpu_bo_set_metadata(robj, args->data.data,
623 						   args->data.data_size_bytes,
624 						   args->data.flags);
625 	}
626 
627 unreserve:
628 	amdgpu_bo_unreserve(robj);
629 out:
630 	drm_gem_object_put(gobj);
631 	return r;
632 }
633 
634 /**
635  * amdgpu_gem_va_update_vm -update the bo_va in its VM
636  *
637  * @adev: amdgpu_device pointer
638  * @vm: vm to update
639  * @bo_va: bo_va to update
640  * @operation: map, unmap or clear
641  *
642  * Update the bo_va directly after setting its address. Errors are not
643  * vital here, so they are not reported back to userspace.
644  */
645 static void amdgpu_gem_va_update_vm(struct amdgpu_device *adev,
646 				    struct amdgpu_vm *vm,
647 				    struct amdgpu_bo_va *bo_va,
648 				    uint32_t operation)
649 {
650 	int r;
651 
652 	if (!amdgpu_vm_ready(vm))
653 		return;
654 
655 	r = amdgpu_vm_clear_freed(adev, vm, NULL);
656 	if (r)
657 		goto error;
658 
659 	if (operation == AMDGPU_VA_OP_MAP ||
660 	    operation == AMDGPU_VA_OP_REPLACE) {
661 		r = amdgpu_vm_bo_update(adev, bo_va, false);
662 		if (r)
663 			goto error;
664 	}
665 
666 	r = amdgpu_vm_update_pdes(adev, vm, false);
667 
668 error:
669 	if (r && r != -ERESTARTSYS)
670 		DRM_ERROR("Couldn't update BO_VA (%d)\n", r);
671 }
672 
673 /**
674  * amdgpu_gem_va_map_flags - map GEM UAPI flags into hardware flags
675  *
676  * @adev: amdgpu_device pointer
677  * @flags: GEM UAPI flags
678  *
679  * Returns the GEM UAPI flags mapped into hardware for the ASIC.
680  */
681 uint64_t amdgpu_gem_va_map_flags(struct amdgpu_device *adev, uint32_t flags)
682 {
683 	uint64_t pte_flag = 0;
684 
685 	if (flags & AMDGPU_VM_PAGE_EXECUTABLE)
686 		pte_flag |= AMDGPU_PTE_EXECUTABLE;
687 	if (flags & AMDGPU_VM_PAGE_READABLE)
688 		pte_flag |= AMDGPU_PTE_READABLE;
689 	if (flags & AMDGPU_VM_PAGE_WRITEABLE)
690 		pte_flag |= AMDGPU_PTE_WRITEABLE;
691 	if (flags & AMDGPU_VM_PAGE_PRT)
692 		pte_flag |= AMDGPU_PTE_PRT_FLAG(adev);
693 	if (flags & AMDGPU_VM_PAGE_NOALLOC)
694 		pte_flag |= AMDGPU_PTE_NOALLOC;
695 
696 	if (adev->gmc.gmc_funcs->map_mtype)
697 		pte_flag |= amdgpu_gmc_map_mtype(adev,
698 						 flags & AMDGPU_VM_MTYPE_MASK);
699 
700 	return pte_flag;
701 }
702 
703 int amdgpu_gem_va_ioctl(struct drm_device *dev, void *data,
704 			  struct drm_file *filp)
705 {
706 	const uint32_t valid_flags = AMDGPU_VM_DELAY_UPDATE |
707 		AMDGPU_VM_PAGE_READABLE | AMDGPU_VM_PAGE_WRITEABLE |
708 		AMDGPU_VM_PAGE_EXECUTABLE | AMDGPU_VM_MTYPE_MASK |
709 		AMDGPU_VM_PAGE_NOALLOC;
710 	const uint32_t prt_flags = AMDGPU_VM_DELAY_UPDATE |
711 		AMDGPU_VM_PAGE_PRT;
712 
713 	struct drm_amdgpu_gem_va *args = data;
714 	struct drm_gem_object *gobj;
715 	struct amdgpu_device *adev = drm_to_adev(dev);
716 	struct amdgpu_fpriv *fpriv = filp->driver_priv;
717 	struct amdgpu_bo *abo;
718 	struct amdgpu_bo_va *bo_va;
719 	struct drm_exec exec;
720 	uint64_t va_flags;
721 	uint64_t vm_size;
722 	int r = 0;
723 
724 	if (args->va_address < AMDGPU_VA_RESERVED_BOTTOM) {
725 		dev_dbg(dev->dev,
726 			"va_address 0x%llx is in reserved area 0x%llx\n",
727 			args->va_address, AMDGPU_VA_RESERVED_BOTTOM);
728 		return -EINVAL;
729 	}
730 
731 	if (args->va_address >= AMDGPU_GMC_HOLE_START &&
732 	    args->va_address < AMDGPU_GMC_HOLE_END) {
733 		dev_dbg(dev->dev,
734 			"va_address 0x%llx is in VA hole 0x%llx-0x%llx\n",
735 			args->va_address, AMDGPU_GMC_HOLE_START,
736 			AMDGPU_GMC_HOLE_END);
737 		return -EINVAL;
738 	}
739 
740 	args->va_address &= AMDGPU_GMC_HOLE_MASK;
741 
742 	vm_size = adev->vm_manager.max_pfn * AMDGPU_GPU_PAGE_SIZE;
743 	vm_size -= AMDGPU_VA_RESERVED_TOP;
744 	if (args->va_address + args->map_size > vm_size) {
745 		dev_dbg(dev->dev,
746 			"va_address 0x%llx is in top reserved area 0x%llx\n",
747 			args->va_address + args->map_size, vm_size);
748 		return -EINVAL;
749 	}
750 
751 	if ((args->flags & ~valid_flags) && (args->flags & ~prt_flags)) {
752 		dev_dbg(dev->dev, "invalid flags combination 0x%08X\n",
753 			args->flags);
754 		return -EINVAL;
755 	}
756 
757 	switch (args->operation) {
758 	case AMDGPU_VA_OP_MAP:
759 	case AMDGPU_VA_OP_UNMAP:
760 	case AMDGPU_VA_OP_CLEAR:
761 	case AMDGPU_VA_OP_REPLACE:
762 		break;
763 	default:
764 		dev_dbg(dev->dev, "unsupported operation %d\n",
765 			args->operation);
766 		return -EINVAL;
767 	}
768 
769 	if ((args->operation != AMDGPU_VA_OP_CLEAR) &&
770 	    !(args->flags & AMDGPU_VM_PAGE_PRT)) {
771 		gobj = drm_gem_object_lookup(filp, args->handle);
772 		if (gobj == NULL)
773 			return -ENOENT;
774 		abo = gem_to_amdgpu_bo(gobj);
775 	} else {
776 		gobj = NULL;
777 		abo = NULL;
778 	}
779 
780 	drm_exec_init(&exec, DRM_EXEC_INTERRUPTIBLE_WAIT |
781 		      DRM_EXEC_IGNORE_DUPLICATES, 0);
782 	drm_exec_until_all_locked(&exec) {
783 		if (gobj) {
784 			r = drm_exec_lock_obj(&exec, gobj);
785 			drm_exec_retry_on_contention(&exec);
786 			if (unlikely(r))
787 				goto error;
788 		}
789 
790 		r = amdgpu_vm_lock_pd(&fpriv->vm, &exec, 2);
791 		drm_exec_retry_on_contention(&exec);
792 		if (unlikely(r))
793 			goto error;
794 	}
795 
796 	if (abo) {
797 		bo_va = amdgpu_vm_bo_find(&fpriv->vm, abo);
798 		if (!bo_va) {
799 			r = -ENOENT;
800 			goto error;
801 		}
802 	} else if (args->operation != AMDGPU_VA_OP_CLEAR) {
803 		bo_va = fpriv->prt_va;
804 	} else {
805 		bo_va = NULL;
806 	}
807 
808 	switch (args->operation) {
809 	case AMDGPU_VA_OP_MAP:
810 		va_flags = amdgpu_gem_va_map_flags(adev, args->flags);
811 		r = amdgpu_vm_bo_map(adev, bo_va, args->va_address,
812 				     args->offset_in_bo, args->map_size,
813 				     va_flags);
814 		break;
815 	case AMDGPU_VA_OP_UNMAP:
816 		r = amdgpu_vm_bo_unmap(adev, bo_va, args->va_address);
817 		break;
818 
819 	case AMDGPU_VA_OP_CLEAR:
820 		r = amdgpu_vm_bo_clear_mappings(adev, &fpriv->vm,
821 						args->va_address,
822 						args->map_size);
823 		break;
824 	case AMDGPU_VA_OP_REPLACE:
825 		va_flags = amdgpu_gem_va_map_flags(adev, args->flags);
826 		r = amdgpu_vm_bo_replace_map(adev, bo_va, args->va_address,
827 					     args->offset_in_bo, args->map_size,
828 					     va_flags);
829 		break;
830 	default:
831 		break;
832 	}
833 	if (!r && !(args->flags & AMDGPU_VM_DELAY_UPDATE) && !adev->debug_vm)
834 		amdgpu_gem_va_update_vm(adev, &fpriv->vm, bo_va,
835 					args->operation);
836 
837 error:
838 	drm_exec_fini(&exec);
839 	drm_gem_object_put(gobj);
840 	return r;
841 }
842 
843 int amdgpu_gem_op_ioctl(struct drm_device *dev, void *data,
844 			struct drm_file *filp)
845 {
846 	struct amdgpu_device *adev = drm_to_adev(dev);
847 	struct drm_amdgpu_gem_op *args = data;
848 	struct drm_gem_object *gobj;
849 	struct amdgpu_vm_bo_base *base;
850 	struct amdgpu_bo *robj;
851 	int r;
852 
853 	gobj = drm_gem_object_lookup(filp, args->handle);
854 	if (!gobj)
855 		return -ENOENT;
856 
857 	robj = gem_to_amdgpu_bo(gobj);
858 
859 	r = amdgpu_bo_reserve(robj, false);
860 	if (unlikely(r))
861 		goto out;
862 
863 	switch (args->op) {
864 	case AMDGPU_GEM_OP_GET_GEM_CREATE_INFO: {
865 		struct drm_amdgpu_gem_create_in info;
866 		void __user *out = u64_to_user_ptr(args->value);
867 
868 		info.bo_size = robj->tbo.base.size;
869 		info.alignment = robj->tbo.page_alignment << PAGE_SHIFT;
870 		info.domains = robj->preferred_domains;
871 		info.domain_flags = robj->flags;
872 		amdgpu_bo_unreserve(robj);
873 		if (copy_to_user(out, &info, sizeof(info)))
874 			r = -EFAULT;
875 		break;
876 	}
877 	case AMDGPU_GEM_OP_SET_PLACEMENT:
878 		if (robj->tbo.base.import_attach &&
879 		    args->value & AMDGPU_GEM_DOMAIN_VRAM) {
880 			r = -EINVAL;
881 			amdgpu_bo_unreserve(robj);
882 			break;
883 		}
884 		if (amdgpu_ttm_tt_get_usermm(robj->tbo.ttm)) {
885 			r = -EPERM;
886 			amdgpu_bo_unreserve(robj);
887 			break;
888 		}
889 		for (base = robj->vm_bo; base; base = base->next)
890 			if (amdgpu_xgmi_same_hive(amdgpu_ttm_adev(robj->tbo.bdev),
891 				amdgpu_ttm_adev(base->vm->root.bo->tbo.bdev))) {
892 				r = -EINVAL;
893 				amdgpu_bo_unreserve(robj);
894 				goto out;
895 			}
896 
897 
898 		robj->preferred_domains = args->value & (AMDGPU_GEM_DOMAIN_VRAM |
899 							AMDGPU_GEM_DOMAIN_GTT |
900 							AMDGPU_GEM_DOMAIN_CPU);
901 		robj->allowed_domains = robj->preferred_domains;
902 		if (robj->allowed_domains == AMDGPU_GEM_DOMAIN_VRAM)
903 			robj->allowed_domains |= AMDGPU_GEM_DOMAIN_GTT;
904 
905 		if (robj->flags & AMDGPU_GEM_CREATE_VM_ALWAYS_VALID)
906 			amdgpu_vm_bo_invalidate(adev, robj, true);
907 
908 		amdgpu_bo_unreserve(robj);
909 		break;
910 	default:
911 		amdgpu_bo_unreserve(robj);
912 		r = -EINVAL;
913 	}
914 
915 out:
916 	drm_gem_object_put(gobj);
917 	return r;
918 }
919 
920 static int amdgpu_gem_align_pitch(struct amdgpu_device *adev,
921 				  int width,
922 				  int cpp,
923 				  bool tiled)
924 {
925 	int aligned = width;
926 	int pitch_mask = 0;
927 
928 	switch (cpp) {
929 	case 1:
930 		pitch_mask = 255;
931 		break;
932 	case 2:
933 		pitch_mask = 127;
934 		break;
935 	case 3:
936 	case 4:
937 		pitch_mask = 63;
938 		break;
939 	}
940 
941 	aligned += pitch_mask;
942 	aligned &= ~pitch_mask;
943 	return aligned * cpp;
944 }
945 
946 int amdgpu_mode_dumb_create(struct drm_file *file_priv,
947 			    struct drm_device *dev,
948 			    struct drm_mode_create_dumb *args)
949 {
950 	struct amdgpu_device *adev = drm_to_adev(dev);
951 	struct amdgpu_fpriv *fpriv = file_priv->driver_priv;
952 	struct drm_gem_object *gobj;
953 	uint32_t handle;
954 	u64 flags = AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED |
955 		    AMDGPU_GEM_CREATE_CPU_GTT_USWC |
956 		    AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS;
957 	u32 domain;
958 	int r;
959 
960 	/*
961 	 * The buffer returned from this function should be cleared, but
962 	 * it can only be done if the ring is enabled or we'll fail to
963 	 * create the buffer.
964 	 */
965 	if (adev->mman.buffer_funcs_enabled)
966 		flags |= AMDGPU_GEM_CREATE_VRAM_CLEARED;
967 
968 	args->pitch = amdgpu_gem_align_pitch(adev, args->width,
969 					     DIV_ROUND_UP(args->bpp, 8), 0);
970 	args->size = (u64)args->pitch * args->height;
971 	args->size = ALIGN(args->size, PAGE_SIZE);
972 	domain = amdgpu_bo_get_preferred_domain(adev,
973 				amdgpu_display_supported_domains(adev, flags));
974 	r = amdgpu_gem_object_create(adev, args->size, 0, domain, flags,
975 				     ttm_bo_type_device, NULL, &gobj, fpriv->xcp_id + 1);
976 	if (r)
977 		return -ENOMEM;
978 
979 	r = drm_gem_handle_create(file_priv, gobj, &handle);
980 	/* drop reference from allocate - handle holds it now */
981 	drm_gem_object_put(gobj);
982 	if (r)
983 		return r;
984 
985 	args->handle = handle;
986 	return 0;
987 }
988 
989 #if defined(CONFIG_DEBUG_FS)
990 static int amdgpu_debugfs_gem_info_show(struct seq_file *m, void *unused)
991 {
992 	struct amdgpu_device *adev = m->private;
993 	struct drm_device *dev = adev_to_drm(adev);
994 	struct drm_file *file;
995 	int r;
996 
997 	r = mutex_lock_interruptible(&dev->filelist_mutex);
998 	if (r)
999 		return r;
1000 
1001 	list_for_each_entry(file, &dev->filelist, lhead) {
1002 		struct task_struct *task;
1003 		struct drm_gem_object *gobj;
1004 		struct pid *pid;
1005 		int id;
1006 
1007 		/*
1008 		 * Although we have a valid reference on file->pid, that does
1009 		 * not guarantee that the task_struct who called get_pid() is
1010 		 * still alive (e.g. get_pid(current) => fork() => exit()).
1011 		 * Therefore, we need to protect this ->comm access using RCU.
1012 		 */
1013 		rcu_read_lock();
1014 		pid = rcu_dereference(file->pid);
1015 		task = pid_task(pid, PIDTYPE_TGID);
1016 		seq_printf(m, "pid %8d command %s:\n", pid_nr(pid),
1017 			   task ? task->comm : "<unknown>");
1018 		rcu_read_unlock();
1019 
1020 		spin_lock(&file->table_lock);
1021 		idr_for_each_entry(&file->object_idr, gobj, id) {
1022 			struct amdgpu_bo *bo = gem_to_amdgpu_bo(gobj);
1023 
1024 			amdgpu_bo_print_info(id, bo, m);
1025 		}
1026 		spin_unlock(&file->table_lock);
1027 	}
1028 
1029 	mutex_unlock(&dev->filelist_mutex);
1030 	return 0;
1031 }
1032 
1033 DEFINE_SHOW_ATTRIBUTE(amdgpu_debugfs_gem_info);
1034 
1035 #endif
1036 
1037 void amdgpu_debugfs_gem_init(struct amdgpu_device *adev)
1038 {
1039 #if defined(CONFIG_DEBUG_FS)
1040 	struct drm_minor *minor = adev_to_drm(adev)->primary;
1041 	struct dentry *root = minor->debugfs_root;
1042 
1043 	debugfs_create_file("amdgpu_gem_info", 0444, root, adev,
1044 			    &amdgpu_debugfs_gem_info_fops);
1045 #endif
1046 }
1047