xref: /linux/drivers/gpu/drm/amd/amdkfd/kfd_chardev.c (revision 3e22193d8cdc0576cd2803da3cac9f4fc9222273)
1 // SPDX-License-Identifier: GPL-2.0 OR MIT
2 /*
3  * Copyright 2014-2022 Advanced Micro Devices, Inc.
4  *
5  * Permission is hereby granted, free of charge, to any person obtaining a
6  * copy of this software and associated documentation files (the "Software"),
7  * to deal in the Software without restriction, including without limitation
8  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
9  * and/or sell copies of the Software, and to permit persons to whom the
10  * Software is furnished to do so, subject to the following conditions:
11  *
12  * The above copyright notice and this permission notice shall be included in
13  * all copies or substantial portions of the Software.
14  *
15  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
18  * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR
19  * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
20  * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
21  * OTHER DEALINGS IN THE SOFTWARE.
22  */
23 
24 #include <linux/device.h>
25 #include <linux/export.h>
26 #include <linux/err.h>
27 #include <linux/fs.h>
28 #include <linux/file.h>
29 #include <linux/sched.h>
30 #include <linux/slab.h>
31 #include <linux/uaccess.h>
32 #include <linux/compat.h>
33 #include <uapi/linux/kfd_ioctl.h>
34 #include <linux/time.h>
35 #include <linux/mm.h>
36 #include <linux/mman.h>
37 #include <linux/ptrace.h>
38 #include <linux/dma-buf.h>
39 #include <linux/fdtable.h>
40 #include <linux/processor.h>
41 #include "kfd_priv.h"
42 #include "kfd_device_queue_manager.h"
43 #include "kfd_svm.h"
44 #include "amdgpu_amdkfd.h"
45 #include "kfd_smi_events.h"
46 #include "amdgpu_dma_buf.h"
47 
48 static long kfd_ioctl(struct file *, unsigned int, unsigned long);
49 static int kfd_open(struct inode *, struct file *);
50 static int kfd_release(struct inode *, struct file *);
51 static int kfd_mmap(struct file *, struct vm_area_struct *);
52 
53 static const char kfd_dev_name[] = "kfd";
54 
55 static const struct file_operations kfd_fops = {
56 	.owner = THIS_MODULE,
57 	.unlocked_ioctl = kfd_ioctl,
58 	.compat_ioctl = compat_ptr_ioctl,
59 	.open = kfd_open,
60 	.release = kfd_release,
61 	.mmap = kfd_mmap,
62 };
63 
64 static int kfd_char_dev_major = -1;
65 static struct class *kfd_class;
66 struct device *kfd_device;
67 
68 static inline struct kfd_process_device *kfd_lock_pdd_by_id(struct kfd_process *p, __u32 gpu_id)
69 {
70 	struct kfd_process_device *pdd;
71 
72 	mutex_lock(&p->mutex);
73 	pdd = kfd_process_device_data_by_id(p, gpu_id);
74 
75 	if (pdd)
76 		return pdd;
77 
78 	mutex_unlock(&p->mutex);
79 	return NULL;
80 }
81 
82 static inline void kfd_unlock_pdd(struct kfd_process_device *pdd)
83 {
84 	mutex_unlock(&pdd->process->mutex);
85 }
86 
87 int kfd_chardev_init(void)
88 {
89 	int err = 0;
90 
91 	kfd_char_dev_major = register_chrdev(0, kfd_dev_name, &kfd_fops);
92 	err = kfd_char_dev_major;
93 	if (err < 0)
94 		goto err_register_chrdev;
95 
96 	kfd_class = class_create(THIS_MODULE, kfd_dev_name);
97 	err = PTR_ERR(kfd_class);
98 	if (IS_ERR(kfd_class))
99 		goto err_class_create;
100 
101 	kfd_device = device_create(kfd_class, NULL,
102 					MKDEV(kfd_char_dev_major, 0),
103 					NULL, kfd_dev_name);
104 	err = PTR_ERR(kfd_device);
105 	if (IS_ERR(kfd_device))
106 		goto err_device_create;
107 
108 	return 0;
109 
110 err_device_create:
111 	class_destroy(kfd_class);
112 err_class_create:
113 	unregister_chrdev(kfd_char_dev_major, kfd_dev_name);
114 err_register_chrdev:
115 	return err;
116 }
117 
118 void kfd_chardev_exit(void)
119 {
120 	device_destroy(kfd_class, MKDEV(kfd_char_dev_major, 0));
121 	class_destroy(kfd_class);
122 	unregister_chrdev(kfd_char_dev_major, kfd_dev_name);
123 	kfd_device = NULL;
124 }
125 
126 
127 static int kfd_open(struct inode *inode, struct file *filep)
128 {
129 	struct kfd_process *process;
130 	bool is_32bit_user_mode;
131 
132 	if (iminor(inode) != 0)
133 		return -ENODEV;
134 
135 	is_32bit_user_mode = in_compat_syscall();
136 
137 	if (is_32bit_user_mode) {
138 		dev_warn(kfd_device,
139 			"Process %d (32-bit) failed to open /dev/kfd\n"
140 			"32-bit processes are not supported by amdkfd\n",
141 			current->pid);
142 		return -EPERM;
143 	}
144 
145 	process = kfd_create_process(filep);
146 	if (IS_ERR(process))
147 		return PTR_ERR(process);
148 
149 	if (kfd_is_locked()) {
150 		dev_dbg(kfd_device, "kfd is locked!\n"
151 				"process %d unreferenced", process->pasid);
152 		kfd_unref_process(process);
153 		return -EAGAIN;
154 	}
155 
156 	/* filep now owns the reference returned by kfd_create_process */
157 	filep->private_data = process;
158 
159 	dev_dbg(kfd_device, "process %d opened, compat mode (32 bit) - %d\n",
160 		process->pasid, process->is_32bit_user_mode);
161 
162 	return 0;
163 }
164 
165 static int kfd_release(struct inode *inode, struct file *filep)
166 {
167 	struct kfd_process *process = filep->private_data;
168 
169 	if (process)
170 		kfd_unref_process(process);
171 
172 	return 0;
173 }
174 
175 static int kfd_ioctl_get_version(struct file *filep, struct kfd_process *p,
176 					void *data)
177 {
178 	struct kfd_ioctl_get_version_args *args = data;
179 
180 	args->major_version = KFD_IOCTL_MAJOR_VERSION;
181 	args->minor_version = KFD_IOCTL_MINOR_VERSION;
182 
183 	return 0;
184 }
185 
186 static int set_queue_properties_from_user(struct queue_properties *q_properties,
187 				struct kfd_ioctl_create_queue_args *args)
188 {
189 	if (args->queue_percentage > KFD_MAX_QUEUE_PERCENTAGE) {
190 		pr_err("Queue percentage must be between 0 to KFD_MAX_QUEUE_PERCENTAGE\n");
191 		return -EINVAL;
192 	}
193 
194 	if (args->queue_priority > KFD_MAX_QUEUE_PRIORITY) {
195 		pr_err("Queue priority must be between 0 to KFD_MAX_QUEUE_PRIORITY\n");
196 		return -EINVAL;
197 	}
198 
199 	if ((args->ring_base_address) &&
200 		(!access_ok((const void __user *) args->ring_base_address,
201 			sizeof(uint64_t)))) {
202 		pr_err("Can't access ring base address\n");
203 		return -EFAULT;
204 	}
205 
206 	if (!is_power_of_2(args->ring_size) && (args->ring_size != 0)) {
207 		pr_err("Ring size must be a power of 2 or 0\n");
208 		return -EINVAL;
209 	}
210 
211 	if (!access_ok((const void __user *) args->read_pointer_address,
212 			sizeof(uint32_t))) {
213 		pr_err("Can't access read pointer\n");
214 		return -EFAULT;
215 	}
216 
217 	if (!access_ok((const void __user *) args->write_pointer_address,
218 			sizeof(uint32_t))) {
219 		pr_err("Can't access write pointer\n");
220 		return -EFAULT;
221 	}
222 
223 	if (args->eop_buffer_address &&
224 		!access_ok((const void __user *) args->eop_buffer_address,
225 			sizeof(uint32_t))) {
226 		pr_debug("Can't access eop buffer");
227 		return -EFAULT;
228 	}
229 
230 	if (args->ctx_save_restore_address &&
231 		!access_ok((const void __user *) args->ctx_save_restore_address,
232 			sizeof(uint32_t))) {
233 		pr_debug("Can't access ctx save restore buffer");
234 		return -EFAULT;
235 	}
236 
237 	q_properties->is_interop = false;
238 	q_properties->is_gws = false;
239 	q_properties->queue_percent = args->queue_percentage;
240 	q_properties->priority = args->queue_priority;
241 	q_properties->queue_address = args->ring_base_address;
242 	q_properties->queue_size = args->ring_size;
243 	q_properties->read_ptr = (uint32_t *) args->read_pointer_address;
244 	q_properties->write_ptr = (uint32_t *) args->write_pointer_address;
245 	q_properties->eop_ring_buffer_address = args->eop_buffer_address;
246 	q_properties->eop_ring_buffer_size = args->eop_buffer_size;
247 	q_properties->ctx_save_restore_area_address =
248 			args->ctx_save_restore_address;
249 	q_properties->ctx_save_restore_area_size = args->ctx_save_restore_size;
250 	q_properties->ctl_stack_size = args->ctl_stack_size;
251 	if (args->queue_type == KFD_IOC_QUEUE_TYPE_COMPUTE ||
252 		args->queue_type == KFD_IOC_QUEUE_TYPE_COMPUTE_AQL)
253 		q_properties->type = KFD_QUEUE_TYPE_COMPUTE;
254 	else if (args->queue_type == KFD_IOC_QUEUE_TYPE_SDMA)
255 		q_properties->type = KFD_QUEUE_TYPE_SDMA;
256 	else if (args->queue_type == KFD_IOC_QUEUE_TYPE_SDMA_XGMI)
257 		q_properties->type = KFD_QUEUE_TYPE_SDMA_XGMI;
258 	else
259 		return -ENOTSUPP;
260 
261 	if (args->queue_type == KFD_IOC_QUEUE_TYPE_COMPUTE_AQL)
262 		q_properties->format = KFD_QUEUE_FORMAT_AQL;
263 	else
264 		q_properties->format = KFD_QUEUE_FORMAT_PM4;
265 
266 	pr_debug("Queue Percentage: %d, %d\n",
267 			q_properties->queue_percent, args->queue_percentage);
268 
269 	pr_debug("Queue Priority: %d, %d\n",
270 			q_properties->priority, args->queue_priority);
271 
272 	pr_debug("Queue Address: 0x%llX, 0x%llX\n",
273 			q_properties->queue_address, args->ring_base_address);
274 
275 	pr_debug("Queue Size: 0x%llX, %u\n",
276 			q_properties->queue_size, args->ring_size);
277 
278 	pr_debug("Queue r/w Pointers: %px, %px\n",
279 			q_properties->read_ptr,
280 			q_properties->write_ptr);
281 
282 	pr_debug("Queue Format: %d\n", q_properties->format);
283 
284 	pr_debug("Queue EOP: 0x%llX\n", q_properties->eop_ring_buffer_address);
285 
286 	pr_debug("Queue CTX save area: 0x%llX\n",
287 			q_properties->ctx_save_restore_area_address);
288 
289 	return 0;
290 }
291 
292 static int kfd_ioctl_create_queue(struct file *filep, struct kfd_process *p,
293 					void *data)
294 {
295 	struct kfd_ioctl_create_queue_args *args = data;
296 	struct kfd_dev *dev;
297 	int err = 0;
298 	unsigned int queue_id;
299 	struct kfd_process_device *pdd;
300 	struct queue_properties q_properties;
301 	uint32_t doorbell_offset_in_process = 0;
302 	struct amdgpu_bo *wptr_bo = NULL;
303 
304 	memset(&q_properties, 0, sizeof(struct queue_properties));
305 
306 	pr_debug("Creating queue ioctl\n");
307 
308 	err = set_queue_properties_from_user(&q_properties, args);
309 	if (err)
310 		return err;
311 
312 	pr_debug("Looking for gpu id 0x%x\n", args->gpu_id);
313 
314 	mutex_lock(&p->mutex);
315 
316 	pdd = kfd_process_device_data_by_id(p, args->gpu_id);
317 	if (!pdd) {
318 		pr_debug("Could not find gpu id 0x%x\n", args->gpu_id);
319 		err = -EINVAL;
320 		goto err_pdd;
321 	}
322 	dev = pdd->dev;
323 
324 	pdd = kfd_bind_process_to_device(dev, p);
325 	if (IS_ERR(pdd)) {
326 		err = -ESRCH;
327 		goto err_bind_process;
328 	}
329 
330 	if (!pdd->doorbell_index &&
331 	    kfd_alloc_process_doorbells(dev, &pdd->doorbell_index) < 0) {
332 		err = -ENOMEM;
333 		goto err_alloc_doorbells;
334 	}
335 
336 	/* Starting with GFX11, wptr BOs must be mapped to GART for MES to determine work
337 	 * on unmapped queues for usermode queue oversubscription (no aggregated doorbell)
338 	 */
339 	if (dev->shared_resources.enable_mes &&
340 			((dev->adev->mes.sched_version & AMDGPU_MES_API_VERSION_MASK)
341 			>> AMDGPU_MES_API_VERSION_SHIFT) >= 2) {
342 		struct amdgpu_bo_va_mapping *wptr_mapping;
343 		struct amdgpu_vm *wptr_vm;
344 
345 		wptr_vm = drm_priv_to_vm(pdd->drm_priv);
346 		err = amdgpu_bo_reserve(wptr_vm->root.bo, false);
347 		if (err)
348 			goto err_wptr_map_gart;
349 
350 		wptr_mapping = amdgpu_vm_bo_lookup_mapping(
351 				wptr_vm, args->write_pointer_address >> PAGE_SHIFT);
352 		amdgpu_bo_unreserve(wptr_vm->root.bo);
353 		if (!wptr_mapping) {
354 			pr_err("Failed to lookup wptr bo\n");
355 			err = -EINVAL;
356 			goto err_wptr_map_gart;
357 		}
358 
359 		wptr_bo = wptr_mapping->bo_va->base.bo;
360 		if (wptr_bo->tbo.base.size > PAGE_SIZE) {
361 			pr_err("Requested GART mapping for wptr bo larger than one page\n");
362 			err = -EINVAL;
363 			goto err_wptr_map_gart;
364 		}
365 
366 		err = amdgpu_amdkfd_map_gtt_bo_to_gart(dev->adev, wptr_bo);
367 		if (err) {
368 			pr_err("Failed to map wptr bo to GART\n");
369 			goto err_wptr_map_gart;
370 		}
371 	}
372 
373 	pr_debug("Creating queue for PASID 0x%x on gpu 0x%x\n",
374 			p->pasid,
375 			dev->id);
376 
377 	err = pqm_create_queue(&p->pqm, dev, filep, &q_properties, &queue_id, wptr_bo,
378 			NULL, NULL, NULL, &doorbell_offset_in_process);
379 	if (err != 0)
380 		goto err_create_queue;
381 
382 	args->queue_id = queue_id;
383 
384 
385 	/* Return gpu_id as doorbell offset for mmap usage */
386 	args->doorbell_offset = KFD_MMAP_TYPE_DOORBELL;
387 	args->doorbell_offset |= KFD_MMAP_GPU_ID(args->gpu_id);
388 	if (KFD_IS_SOC15(dev))
389 		/* On SOC15 ASICs, include the doorbell offset within the
390 		 * process doorbell frame, which is 2 pages.
391 		 */
392 		args->doorbell_offset |= doorbell_offset_in_process;
393 
394 	mutex_unlock(&p->mutex);
395 
396 	pr_debug("Queue id %d was created successfully\n", args->queue_id);
397 
398 	pr_debug("Ring buffer address == 0x%016llX\n",
399 			args->ring_base_address);
400 
401 	pr_debug("Read ptr address    == 0x%016llX\n",
402 			args->read_pointer_address);
403 
404 	pr_debug("Write ptr address   == 0x%016llX\n",
405 			args->write_pointer_address);
406 
407 	return 0;
408 
409 err_create_queue:
410 	if (wptr_bo)
411 		amdgpu_amdkfd_free_gtt_mem(dev->adev, wptr_bo);
412 err_wptr_map_gart:
413 err_alloc_doorbells:
414 err_bind_process:
415 err_pdd:
416 	mutex_unlock(&p->mutex);
417 	return err;
418 }
419 
420 static int kfd_ioctl_destroy_queue(struct file *filp, struct kfd_process *p,
421 					void *data)
422 {
423 	int retval;
424 	struct kfd_ioctl_destroy_queue_args *args = data;
425 
426 	pr_debug("Destroying queue id %d for pasid 0x%x\n",
427 				args->queue_id,
428 				p->pasid);
429 
430 	mutex_lock(&p->mutex);
431 
432 	retval = pqm_destroy_queue(&p->pqm, args->queue_id);
433 
434 	mutex_unlock(&p->mutex);
435 	return retval;
436 }
437 
438 static int kfd_ioctl_update_queue(struct file *filp, struct kfd_process *p,
439 					void *data)
440 {
441 	int retval;
442 	struct kfd_ioctl_update_queue_args *args = data;
443 	struct queue_properties properties;
444 
445 	if (args->queue_percentage > KFD_MAX_QUEUE_PERCENTAGE) {
446 		pr_err("Queue percentage must be between 0 to KFD_MAX_QUEUE_PERCENTAGE\n");
447 		return -EINVAL;
448 	}
449 
450 	if (args->queue_priority > KFD_MAX_QUEUE_PRIORITY) {
451 		pr_err("Queue priority must be between 0 to KFD_MAX_QUEUE_PRIORITY\n");
452 		return -EINVAL;
453 	}
454 
455 	if ((args->ring_base_address) &&
456 		(!access_ok((const void __user *) args->ring_base_address,
457 			sizeof(uint64_t)))) {
458 		pr_err("Can't access ring base address\n");
459 		return -EFAULT;
460 	}
461 
462 	if (!is_power_of_2(args->ring_size) && (args->ring_size != 0)) {
463 		pr_err("Ring size must be a power of 2 or 0\n");
464 		return -EINVAL;
465 	}
466 
467 	properties.queue_address = args->ring_base_address;
468 	properties.queue_size = args->ring_size;
469 	properties.queue_percent = args->queue_percentage;
470 	properties.priority = args->queue_priority;
471 
472 	pr_debug("Updating queue id %d for pasid 0x%x\n",
473 			args->queue_id, p->pasid);
474 
475 	mutex_lock(&p->mutex);
476 
477 	retval = pqm_update_queue_properties(&p->pqm, args->queue_id, &properties);
478 
479 	mutex_unlock(&p->mutex);
480 
481 	return retval;
482 }
483 
484 static int kfd_ioctl_set_cu_mask(struct file *filp, struct kfd_process *p,
485 					void *data)
486 {
487 	int retval;
488 	const int max_num_cus = 1024;
489 	struct kfd_ioctl_set_cu_mask_args *args = data;
490 	struct mqd_update_info minfo = {0};
491 	uint32_t __user *cu_mask_ptr = (uint32_t __user *)args->cu_mask_ptr;
492 	size_t cu_mask_size = sizeof(uint32_t) * (args->num_cu_mask / 32);
493 
494 	if ((args->num_cu_mask % 32) != 0) {
495 		pr_debug("num_cu_mask 0x%x must be a multiple of 32",
496 				args->num_cu_mask);
497 		return -EINVAL;
498 	}
499 
500 	minfo.cu_mask.count = args->num_cu_mask;
501 	if (minfo.cu_mask.count == 0) {
502 		pr_debug("CU mask cannot be 0");
503 		return -EINVAL;
504 	}
505 
506 	/* To prevent an unreasonably large CU mask size, set an arbitrary
507 	 * limit of max_num_cus bits.  We can then just drop any CU mask bits
508 	 * past max_num_cus bits and just use the first max_num_cus bits.
509 	 */
510 	if (minfo.cu_mask.count > max_num_cus) {
511 		pr_debug("CU mask cannot be greater than 1024 bits");
512 		minfo.cu_mask.count = max_num_cus;
513 		cu_mask_size = sizeof(uint32_t) * (max_num_cus/32);
514 	}
515 
516 	minfo.cu_mask.ptr = kzalloc(cu_mask_size, GFP_KERNEL);
517 	if (!minfo.cu_mask.ptr)
518 		return -ENOMEM;
519 
520 	retval = copy_from_user(minfo.cu_mask.ptr, cu_mask_ptr, cu_mask_size);
521 	if (retval) {
522 		pr_debug("Could not copy CU mask from userspace");
523 		retval = -EFAULT;
524 		goto out;
525 	}
526 
527 	minfo.update_flag = UPDATE_FLAG_CU_MASK;
528 
529 	mutex_lock(&p->mutex);
530 
531 	retval = pqm_update_mqd(&p->pqm, args->queue_id, &minfo);
532 
533 	mutex_unlock(&p->mutex);
534 
535 out:
536 	kfree(minfo.cu_mask.ptr);
537 	return retval;
538 }
539 
540 static int kfd_ioctl_get_queue_wave_state(struct file *filep,
541 					  struct kfd_process *p, void *data)
542 {
543 	struct kfd_ioctl_get_queue_wave_state_args *args = data;
544 	int r;
545 
546 	mutex_lock(&p->mutex);
547 
548 	r = pqm_get_wave_state(&p->pqm, args->queue_id,
549 			       (void __user *)args->ctl_stack_address,
550 			       &args->ctl_stack_used_size,
551 			       &args->save_area_used_size);
552 
553 	mutex_unlock(&p->mutex);
554 
555 	return r;
556 }
557 
558 static int kfd_ioctl_set_memory_policy(struct file *filep,
559 					struct kfd_process *p, void *data)
560 {
561 	struct kfd_ioctl_set_memory_policy_args *args = data;
562 	int err = 0;
563 	struct kfd_process_device *pdd;
564 	enum cache_policy default_policy, alternate_policy;
565 
566 	if (args->default_policy != KFD_IOC_CACHE_POLICY_COHERENT
567 	    && args->default_policy != KFD_IOC_CACHE_POLICY_NONCOHERENT) {
568 		return -EINVAL;
569 	}
570 
571 	if (args->alternate_policy != KFD_IOC_CACHE_POLICY_COHERENT
572 	    && args->alternate_policy != KFD_IOC_CACHE_POLICY_NONCOHERENT) {
573 		return -EINVAL;
574 	}
575 
576 	mutex_lock(&p->mutex);
577 	pdd = kfd_process_device_data_by_id(p, args->gpu_id);
578 	if (!pdd) {
579 		pr_debug("Could not find gpu id 0x%x\n", args->gpu_id);
580 		err = -EINVAL;
581 		goto err_pdd;
582 	}
583 
584 	pdd = kfd_bind_process_to_device(pdd->dev, p);
585 	if (IS_ERR(pdd)) {
586 		err = -ESRCH;
587 		goto out;
588 	}
589 
590 	default_policy = (args->default_policy == KFD_IOC_CACHE_POLICY_COHERENT)
591 			 ? cache_policy_coherent : cache_policy_noncoherent;
592 
593 	alternate_policy =
594 		(args->alternate_policy == KFD_IOC_CACHE_POLICY_COHERENT)
595 		   ? cache_policy_coherent : cache_policy_noncoherent;
596 
597 	if (!pdd->dev->dqm->ops.set_cache_memory_policy(pdd->dev->dqm,
598 				&pdd->qpd,
599 				default_policy,
600 				alternate_policy,
601 				(void __user *)args->alternate_aperture_base,
602 				args->alternate_aperture_size))
603 		err = -EINVAL;
604 
605 out:
606 err_pdd:
607 	mutex_unlock(&p->mutex);
608 
609 	return err;
610 }
611 
612 static int kfd_ioctl_set_trap_handler(struct file *filep,
613 					struct kfd_process *p, void *data)
614 {
615 	struct kfd_ioctl_set_trap_handler_args *args = data;
616 	int err = 0;
617 	struct kfd_process_device *pdd;
618 
619 	mutex_lock(&p->mutex);
620 
621 	pdd = kfd_process_device_data_by_id(p, args->gpu_id);
622 	if (!pdd) {
623 		err = -EINVAL;
624 		goto err_pdd;
625 	}
626 
627 	pdd = kfd_bind_process_to_device(pdd->dev, p);
628 	if (IS_ERR(pdd)) {
629 		err = -ESRCH;
630 		goto out;
631 	}
632 
633 	kfd_process_set_trap_handler(&pdd->qpd, args->tba_addr, args->tma_addr);
634 
635 out:
636 err_pdd:
637 	mutex_unlock(&p->mutex);
638 
639 	return err;
640 }
641 
642 static int kfd_ioctl_dbg_register(struct file *filep,
643 				struct kfd_process *p, void *data)
644 {
645 	return -EPERM;
646 }
647 
648 static int kfd_ioctl_dbg_unregister(struct file *filep,
649 				struct kfd_process *p, void *data)
650 {
651 	return -EPERM;
652 }
653 
654 static int kfd_ioctl_dbg_address_watch(struct file *filep,
655 					struct kfd_process *p, void *data)
656 {
657 	return -EPERM;
658 }
659 
660 /* Parse and generate fixed size data structure for wave control */
661 static int kfd_ioctl_dbg_wave_control(struct file *filep,
662 					struct kfd_process *p, void *data)
663 {
664 	return -EPERM;
665 }
666 
667 static int kfd_ioctl_get_clock_counters(struct file *filep,
668 				struct kfd_process *p, void *data)
669 {
670 	struct kfd_ioctl_get_clock_counters_args *args = data;
671 	struct kfd_process_device *pdd;
672 
673 	mutex_lock(&p->mutex);
674 	pdd = kfd_process_device_data_by_id(p, args->gpu_id);
675 	mutex_unlock(&p->mutex);
676 	if (pdd)
677 		/* Reading GPU clock counter from KGD */
678 		args->gpu_clock_counter = amdgpu_amdkfd_get_gpu_clock_counter(pdd->dev->adev);
679 	else
680 		/* Node without GPU resource */
681 		args->gpu_clock_counter = 0;
682 
683 	/* No access to rdtsc. Using raw monotonic time */
684 	args->cpu_clock_counter = ktime_get_raw_ns();
685 	args->system_clock_counter = ktime_get_boottime_ns();
686 
687 	/* Since the counter is in nano-seconds we use 1GHz frequency */
688 	args->system_clock_freq = 1000000000;
689 
690 	return 0;
691 }
692 
693 
694 static int kfd_ioctl_get_process_apertures(struct file *filp,
695 				struct kfd_process *p, void *data)
696 {
697 	struct kfd_ioctl_get_process_apertures_args *args = data;
698 	struct kfd_process_device_apertures *pAperture;
699 	int i;
700 
701 	dev_dbg(kfd_device, "get apertures for PASID 0x%x", p->pasid);
702 
703 	args->num_of_nodes = 0;
704 
705 	mutex_lock(&p->mutex);
706 	/* Run over all pdd of the process */
707 	for (i = 0; i < p->n_pdds; i++) {
708 		struct kfd_process_device *pdd = p->pdds[i];
709 
710 		pAperture =
711 			&args->process_apertures[args->num_of_nodes];
712 		pAperture->gpu_id = pdd->dev->id;
713 		pAperture->lds_base = pdd->lds_base;
714 		pAperture->lds_limit = pdd->lds_limit;
715 		pAperture->gpuvm_base = pdd->gpuvm_base;
716 		pAperture->gpuvm_limit = pdd->gpuvm_limit;
717 		pAperture->scratch_base = pdd->scratch_base;
718 		pAperture->scratch_limit = pdd->scratch_limit;
719 
720 		dev_dbg(kfd_device,
721 			"node id %u\n", args->num_of_nodes);
722 		dev_dbg(kfd_device,
723 			"gpu id %u\n", pdd->dev->id);
724 		dev_dbg(kfd_device,
725 			"lds_base %llX\n", pdd->lds_base);
726 		dev_dbg(kfd_device,
727 			"lds_limit %llX\n", pdd->lds_limit);
728 		dev_dbg(kfd_device,
729 			"gpuvm_base %llX\n", pdd->gpuvm_base);
730 		dev_dbg(kfd_device,
731 			"gpuvm_limit %llX\n", pdd->gpuvm_limit);
732 		dev_dbg(kfd_device,
733 			"scratch_base %llX\n", pdd->scratch_base);
734 		dev_dbg(kfd_device,
735 			"scratch_limit %llX\n", pdd->scratch_limit);
736 
737 		if (++args->num_of_nodes >= NUM_OF_SUPPORTED_GPUS)
738 			break;
739 	}
740 	mutex_unlock(&p->mutex);
741 
742 	return 0;
743 }
744 
745 static int kfd_ioctl_get_process_apertures_new(struct file *filp,
746 				struct kfd_process *p, void *data)
747 {
748 	struct kfd_ioctl_get_process_apertures_new_args *args = data;
749 	struct kfd_process_device_apertures *pa;
750 	int ret;
751 	int i;
752 
753 	dev_dbg(kfd_device, "get apertures for PASID 0x%x", p->pasid);
754 
755 	if (args->num_of_nodes == 0) {
756 		/* Return number of nodes, so that user space can alloacate
757 		 * sufficient memory
758 		 */
759 		mutex_lock(&p->mutex);
760 		args->num_of_nodes = p->n_pdds;
761 		goto out_unlock;
762 	}
763 
764 	/* Fill in process-aperture information for all available
765 	 * nodes, but not more than args->num_of_nodes as that is
766 	 * the amount of memory allocated by user
767 	 */
768 	pa = kzalloc((sizeof(struct kfd_process_device_apertures) *
769 				args->num_of_nodes), GFP_KERNEL);
770 	if (!pa)
771 		return -ENOMEM;
772 
773 	mutex_lock(&p->mutex);
774 
775 	if (!p->n_pdds) {
776 		args->num_of_nodes = 0;
777 		kfree(pa);
778 		goto out_unlock;
779 	}
780 
781 	/* Run over all pdd of the process */
782 	for (i = 0; i < min(p->n_pdds, args->num_of_nodes); i++) {
783 		struct kfd_process_device *pdd = p->pdds[i];
784 
785 		pa[i].gpu_id = pdd->dev->id;
786 		pa[i].lds_base = pdd->lds_base;
787 		pa[i].lds_limit = pdd->lds_limit;
788 		pa[i].gpuvm_base = pdd->gpuvm_base;
789 		pa[i].gpuvm_limit = pdd->gpuvm_limit;
790 		pa[i].scratch_base = pdd->scratch_base;
791 		pa[i].scratch_limit = pdd->scratch_limit;
792 
793 		dev_dbg(kfd_device,
794 			"gpu id %u\n", pdd->dev->id);
795 		dev_dbg(kfd_device,
796 			"lds_base %llX\n", pdd->lds_base);
797 		dev_dbg(kfd_device,
798 			"lds_limit %llX\n", pdd->lds_limit);
799 		dev_dbg(kfd_device,
800 			"gpuvm_base %llX\n", pdd->gpuvm_base);
801 		dev_dbg(kfd_device,
802 			"gpuvm_limit %llX\n", pdd->gpuvm_limit);
803 		dev_dbg(kfd_device,
804 			"scratch_base %llX\n", pdd->scratch_base);
805 		dev_dbg(kfd_device,
806 			"scratch_limit %llX\n", pdd->scratch_limit);
807 	}
808 	mutex_unlock(&p->mutex);
809 
810 	args->num_of_nodes = i;
811 	ret = copy_to_user(
812 			(void __user *)args->kfd_process_device_apertures_ptr,
813 			pa,
814 			(i * sizeof(struct kfd_process_device_apertures)));
815 	kfree(pa);
816 	return ret ? -EFAULT : 0;
817 
818 out_unlock:
819 	mutex_unlock(&p->mutex);
820 	return 0;
821 }
822 
823 static int kfd_ioctl_create_event(struct file *filp, struct kfd_process *p,
824 					void *data)
825 {
826 	struct kfd_ioctl_create_event_args *args = data;
827 	int err;
828 
829 	/* For dGPUs the event page is allocated in user mode. The
830 	 * handle is passed to KFD with the first call to this IOCTL
831 	 * through the event_page_offset field.
832 	 */
833 	if (args->event_page_offset) {
834 		mutex_lock(&p->mutex);
835 		err = kfd_kmap_event_page(p, args->event_page_offset);
836 		mutex_unlock(&p->mutex);
837 		if (err)
838 			return err;
839 	}
840 
841 	err = kfd_event_create(filp, p, args->event_type,
842 				args->auto_reset != 0, args->node_id,
843 				&args->event_id, &args->event_trigger_data,
844 				&args->event_page_offset,
845 				&args->event_slot_index);
846 
847 	pr_debug("Created event (id:0x%08x) (%s)\n", args->event_id, __func__);
848 	return err;
849 }
850 
851 static int kfd_ioctl_destroy_event(struct file *filp, struct kfd_process *p,
852 					void *data)
853 {
854 	struct kfd_ioctl_destroy_event_args *args = data;
855 
856 	return kfd_event_destroy(p, args->event_id);
857 }
858 
859 static int kfd_ioctl_set_event(struct file *filp, struct kfd_process *p,
860 				void *data)
861 {
862 	struct kfd_ioctl_set_event_args *args = data;
863 
864 	return kfd_set_event(p, args->event_id);
865 }
866 
867 static int kfd_ioctl_reset_event(struct file *filp, struct kfd_process *p,
868 				void *data)
869 {
870 	struct kfd_ioctl_reset_event_args *args = data;
871 
872 	return kfd_reset_event(p, args->event_id);
873 }
874 
875 static int kfd_ioctl_wait_events(struct file *filp, struct kfd_process *p,
876 				void *data)
877 {
878 	struct kfd_ioctl_wait_events_args *args = data;
879 
880 	return kfd_wait_on_events(p, args->num_events,
881 			(void __user *)args->events_ptr,
882 			(args->wait_for_all != 0),
883 			&args->timeout, &args->wait_result);
884 }
885 static int kfd_ioctl_set_scratch_backing_va(struct file *filep,
886 					struct kfd_process *p, void *data)
887 {
888 	struct kfd_ioctl_set_scratch_backing_va_args *args = data;
889 	struct kfd_process_device *pdd;
890 	struct kfd_dev *dev;
891 	long err;
892 
893 	mutex_lock(&p->mutex);
894 	pdd = kfd_process_device_data_by_id(p, args->gpu_id);
895 	if (!pdd) {
896 		err = -EINVAL;
897 		goto err_pdd;
898 	}
899 	dev = pdd->dev;
900 
901 	pdd = kfd_bind_process_to_device(dev, p);
902 	if (IS_ERR(pdd)) {
903 		err = PTR_ERR(pdd);
904 		goto bind_process_to_device_fail;
905 	}
906 
907 	pdd->qpd.sh_hidden_private_base = args->va_addr;
908 
909 	mutex_unlock(&p->mutex);
910 
911 	if (dev->dqm->sched_policy == KFD_SCHED_POLICY_NO_HWS &&
912 	    pdd->qpd.vmid != 0 && dev->kfd2kgd->set_scratch_backing_va)
913 		dev->kfd2kgd->set_scratch_backing_va(
914 			dev->adev, args->va_addr, pdd->qpd.vmid);
915 
916 	return 0;
917 
918 bind_process_to_device_fail:
919 err_pdd:
920 	mutex_unlock(&p->mutex);
921 	return err;
922 }
923 
924 static int kfd_ioctl_get_tile_config(struct file *filep,
925 		struct kfd_process *p, void *data)
926 {
927 	struct kfd_ioctl_get_tile_config_args *args = data;
928 	struct kfd_process_device *pdd;
929 	struct tile_config config;
930 	int err = 0;
931 
932 	mutex_lock(&p->mutex);
933 	pdd = kfd_process_device_data_by_id(p, args->gpu_id);
934 	mutex_unlock(&p->mutex);
935 	if (!pdd)
936 		return -EINVAL;
937 
938 	amdgpu_amdkfd_get_tile_config(pdd->dev->adev, &config);
939 
940 	args->gb_addr_config = config.gb_addr_config;
941 	args->num_banks = config.num_banks;
942 	args->num_ranks = config.num_ranks;
943 
944 	if (args->num_tile_configs > config.num_tile_configs)
945 		args->num_tile_configs = config.num_tile_configs;
946 	err = copy_to_user((void __user *)args->tile_config_ptr,
947 			config.tile_config_ptr,
948 			args->num_tile_configs * sizeof(uint32_t));
949 	if (err) {
950 		args->num_tile_configs = 0;
951 		return -EFAULT;
952 	}
953 
954 	if (args->num_macro_tile_configs > config.num_macro_tile_configs)
955 		args->num_macro_tile_configs =
956 				config.num_macro_tile_configs;
957 	err = copy_to_user((void __user *)args->macro_tile_config_ptr,
958 			config.macro_tile_config_ptr,
959 			args->num_macro_tile_configs * sizeof(uint32_t));
960 	if (err) {
961 		args->num_macro_tile_configs = 0;
962 		return -EFAULT;
963 	}
964 
965 	return 0;
966 }
967 
968 static int kfd_ioctl_acquire_vm(struct file *filep, struct kfd_process *p,
969 				void *data)
970 {
971 	struct kfd_ioctl_acquire_vm_args *args = data;
972 	struct kfd_process_device *pdd;
973 	struct file *drm_file;
974 	int ret;
975 
976 	drm_file = fget(args->drm_fd);
977 	if (!drm_file)
978 		return -EINVAL;
979 
980 	mutex_lock(&p->mutex);
981 	pdd = kfd_process_device_data_by_id(p, args->gpu_id);
982 	if (!pdd) {
983 		ret = -EINVAL;
984 		goto err_pdd;
985 	}
986 
987 	if (pdd->drm_file) {
988 		ret = pdd->drm_file == drm_file ? 0 : -EBUSY;
989 		goto err_drm_file;
990 	}
991 
992 	ret = kfd_process_device_init_vm(pdd, drm_file);
993 	if (ret)
994 		goto err_unlock;
995 
996 	/* On success, the PDD keeps the drm_file reference */
997 	mutex_unlock(&p->mutex);
998 
999 	return 0;
1000 
1001 err_unlock:
1002 err_pdd:
1003 err_drm_file:
1004 	mutex_unlock(&p->mutex);
1005 	fput(drm_file);
1006 	return ret;
1007 }
1008 
1009 bool kfd_dev_is_large_bar(struct kfd_dev *dev)
1010 {
1011 	if (debug_largebar) {
1012 		pr_debug("Simulate large-bar allocation on non large-bar machine\n");
1013 		return true;
1014 	}
1015 
1016 	if (dev->use_iommu_v2)
1017 		return false;
1018 
1019 	if (dev->local_mem_info.local_mem_size_private == 0 &&
1020 			dev->local_mem_info.local_mem_size_public > 0)
1021 		return true;
1022 	return false;
1023 }
1024 
1025 static int kfd_ioctl_get_available_memory(struct file *filep,
1026 					  struct kfd_process *p, void *data)
1027 {
1028 	struct kfd_ioctl_get_available_memory_args *args = data;
1029 	struct kfd_process_device *pdd = kfd_lock_pdd_by_id(p, args->gpu_id);
1030 
1031 	if (!pdd)
1032 		return -EINVAL;
1033 	args->available = amdgpu_amdkfd_get_available_memory(pdd->dev->adev);
1034 	kfd_unlock_pdd(pdd);
1035 	return 0;
1036 }
1037 
1038 static int kfd_ioctl_alloc_memory_of_gpu(struct file *filep,
1039 					struct kfd_process *p, void *data)
1040 {
1041 	struct kfd_ioctl_alloc_memory_of_gpu_args *args = data;
1042 	struct kfd_process_device *pdd;
1043 	void *mem;
1044 	struct kfd_dev *dev;
1045 	int idr_handle;
1046 	long err;
1047 	uint64_t offset = args->mmap_offset;
1048 	uint32_t flags = args->flags;
1049 
1050 	if (args->size == 0)
1051 		return -EINVAL;
1052 
1053 #if IS_ENABLED(CONFIG_HSA_AMD_SVM)
1054 	/* Flush pending deferred work to avoid racing with deferred actions
1055 	 * from previous memory map changes (e.g. munmap).
1056 	 */
1057 	svm_range_list_lock_and_flush_work(&p->svms, current->mm);
1058 	mutex_lock(&p->svms.lock);
1059 	mmap_write_unlock(current->mm);
1060 	if (interval_tree_iter_first(&p->svms.objects,
1061 				     args->va_addr >> PAGE_SHIFT,
1062 				     (args->va_addr + args->size - 1) >> PAGE_SHIFT)) {
1063 		pr_err("Address: 0x%llx already allocated by SVM\n",
1064 			args->va_addr);
1065 		mutex_unlock(&p->svms.lock);
1066 		return -EADDRINUSE;
1067 	}
1068 
1069 	/* When register user buffer check if it has been registered by svm by
1070 	 * buffer cpu virtual address.
1071 	 */
1072 	if ((flags & KFD_IOC_ALLOC_MEM_FLAGS_USERPTR) &&
1073 	    interval_tree_iter_first(&p->svms.objects,
1074 				     args->mmap_offset >> PAGE_SHIFT,
1075 				     (args->mmap_offset  + args->size - 1) >> PAGE_SHIFT)) {
1076 		pr_err("User Buffer Address: 0x%llx already allocated by SVM\n",
1077 			args->mmap_offset);
1078 		mutex_unlock(&p->svms.lock);
1079 		return -EADDRINUSE;
1080 	}
1081 
1082 	mutex_unlock(&p->svms.lock);
1083 #endif
1084 	mutex_lock(&p->mutex);
1085 	pdd = kfd_process_device_data_by_id(p, args->gpu_id);
1086 	if (!pdd) {
1087 		err = -EINVAL;
1088 		goto err_pdd;
1089 	}
1090 
1091 	dev = pdd->dev;
1092 
1093 	if ((flags & KFD_IOC_ALLOC_MEM_FLAGS_PUBLIC) &&
1094 		(flags & KFD_IOC_ALLOC_MEM_FLAGS_VRAM) &&
1095 		!kfd_dev_is_large_bar(dev)) {
1096 		pr_err("Alloc host visible vram on small bar is not allowed\n");
1097 		err = -EINVAL;
1098 		goto err_large_bar;
1099 	}
1100 
1101 	pdd = kfd_bind_process_to_device(dev, p);
1102 	if (IS_ERR(pdd)) {
1103 		err = PTR_ERR(pdd);
1104 		goto err_unlock;
1105 	}
1106 
1107 	if (flags & KFD_IOC_ALLOC_MEM_FLAGS_DOORBELL) {
1108 		if (args->size != kfd_doorbell_process_slice(dev)) {
1109 			err = -EINVAL;
1110 			goto err_unlock;
1111 		}
1112 		offset = kfd_get_process_doorbells(pdd);
1113 		if (!offset) {
1114 			err = -ENOMEM;
1115 			goto err_unlock;
1116 		}
1117 	} else if (flags & KFD_IOC_ALLOC_MEM_FLAGS_MMIO_REMAP) {
1118 		if (args->size != PAGE_SIZE) {
1119 			err = -EINVAL;
1120 			goto err_unlock;
1121 		}
1122 		offset = dev->adev->rmmio_remap.bus_addr;
1123 		if (!offset) {
1124 			err = -ENOMEM;
1125 			goto err_unlock;
1126 		}
1127 	}
1128 
1129 	err = amdgpu_amdkfd_gpuvm_alloc_memory_of_gpu(
1130 		dev->adev, args->va_addr, args->size,
1131 		pdd->drm_priv, (struct kgd_mem **) &mem, &offset,
1132 		flags, false);
1133 
1134 	if (err)
1135 		goto err_unlock;
1136 
1137 	idr_handle = kfd_process_device_create_obj_handle(pdd, mem);
1138 	if (idr_handle < 0) {
1139 		err = -EFAULT;
1140 		goto err_free;
1141 	}
1142 
1143 	/* Update the VRAM usage count */
1144 	if (flags & KFD_IOC_ALLOC_MEM_FLAGS_VRAM) {
1145 		uint64_t size = args->size;
1146 
1147 		if (flags & KFD_IOC_ALLOC_MEM_FLAGS_AQL_QUEUE_MEM)
1148 			size >>= 1;
1149 		WRITE_ONCE(pdd->vram_usage, pdd->vram_usage + PAGE_ALIGN(size));
1150 	}
1151 
1152 	mutex_unlock(&p->mutex);
1153 
1154 	args->handle = MAKE_HANDLE(args->gpu_id, idr_handle);
1155 	args->mmap_offset = offset;
1156 
1157 	/* MMIO is mapped through kfd device
1158 	 * Generate a kfd mmap offset
1159 	 */
1160 	if (flags & KFD_IOC_ALLOC_MEM_FLAGS_MMIO_REMAP)
1161 		args->mmap_offset = KFD_MMAP_TYPE_MMIO
1162 					| KFD_MMAP_GPU_ID(args->gpu_id);
1163 
1164 	return 0;
1165 
1166 err_free:
1167 	amdgpu_amdkfd_gpuvm_free_memory_of_gpu(dev->adev, (struct kgd_mem *)mem,
1168 					       pdd->drm_priv, NULL);
1169 err_unlock:
1170 err_pdd:
1171 err_large_bar:
1172 	mutex_unlock(&p->mutex);
1173 	return err;
1174 }
1175 
1176 static int kfd_ioctl_free_memory_of_gpu(struct file *filep,
1177 					struct kfd_process *p, void *data)
1178 {
1179 	struct kfd_ioctl_free_memory_of_gpu_args *args = data;
1180 	struct kfd_process_device *pdd;
1181 	void *mem;
1182 	int ret;
1183 	uint64_t size = 0;
1184 
1185 	mutex_lock(&p->mutex);
1186 	/*
1187 	 * Safeguard to prevent user space from freeing signal BO.
1188 	 * It will be freed at process termination.
1189 	 */
1190 	if (p->signal_handle && (p->signal_handle == args->handle)) {
1191 		pr_err("Free signal BO is not allowed\n");
1192 		ret = -EPERM;
1193 		goto err_unlock;
1194 	}
1195 
1196 	pdd = kfd_process_device_data_by_id(p, GET_GPU_ID(args->handle));
1197 	if (!pdd) {
1198 		pr_err("Process device data doesn't exist\n");
1199 		ret = -EINVAL;
1200 		goto err_pdd;
1201 	}
1202 
1203 	mem = kfd_process_device_translate_handle(
1204 		pdd, GET_IDR_HANDLE(args->handle));
1205 	if (!mem) {
1206 		ret = -EINVAL;
1207 		goto err_unlock;
1208 	}
1209 
1210 	ret = amdgpu_amdkfd_gpuvm_free_memory_of_gpu(pdd->dev->adev,
1211 				(struct kgd_mem *)mem, pdd->drm_priv, &size);
1212 
1213 	/* If freeing the buffer failed, leave the handle in place for
1214 	 * clean-up during process tear-down.
1215 	 */
1216 	if (!ret)
1217 		kfd_process_device_remove_obj_handle(
1218 			pdd, GET_IDR_HANDLE(args->handle));
1219 
1220 	WRITE_ONCE(pdd->vram_usage, pdd->vram_usage - size);
1221 
1222 err_unlock:
1223 err_pdd:
1224 	mutex_unlock(&p->mutex);
1225 	return ret;
1226 }
1227 
1228 static int kfd_ioctl_map_memory_to_gpu(struct file *filep,
1229 					struct kfd_process *p, void *data)
1230 {
1231 	struct kfd_ioctl_map_memory_to_gpu_args *args = data;
1232 	struct kfd_process_device *pdd, *peer_pdd;
1233 	void *mem;
1234 	struct kfd_dev *dev;
1235 	long err = 0;
1236 	int i;
1237 	uint32_t *devices_arr = NULL;
1238 
1239 	if (!args->n_devices) {
1240 		pr_debug("Device IDs array empty\n");
1241 		return -EINVAL;
1242 	}
1243 	if (args->n_success > args->n_devices) {
1244 		pr_debug("n_success exceeds n_devices\n");
1245 		return -EINVAL;
1246 	}
1247 
1248 	devices_arr = kmalloc_array(args->n_devices, sizeof(*devices_arr),
1249 				    GFP_KERNEL);
1250 	if (!devices_arr)
1251 		return -ENOMEM;
1252 
1253 	err = copy_from_user(devices_arr,
1254 			     (void __user *)args->device_ids_array_ptr,
1255 			     args->n_devices * sizeof(*devices_arr));
1256 	if (err != 0) {
1257 		err = -EFAULT;
1258 		goto copy_from_user_failed;
1259 	}
1260 
1261 	mutex_lock(&p->mutex);
1262 	pdd = kfd_process_device_data_by_id(p, GET_GPU_ID(args->handle));
1263 	if (!pdd) {
1264 		err = -EINVAL;
1265 		goto get_process_device_data_failed;
1266 	}
1267 	dev = pdd->dev;
1268 
1269 	pdd = kfd_bind_process_to_device(dev, p);
1270 	if (IS_ERR(pdd)) {
1271 		err = PTR_ERR(pdd);
1272 		goto bind_process_to_device_failed;
1273 	}
1274 
1275 	mem = kfd_process_device_translate_handle(pdd,
1276 						GET_IDR_HANDLE(args->handle));
1277 	if (!mem) {
1278 		err = -ENOMEM;
1279 		goto get_mem_obj_from_handle_failed;
1280 	}
1281 
1282 	for (i = args->n_success; i < args->n_devices; i++) {
1283 		peer_pdd = kfd_process_device_data_by_id(p, devices_arr[i]);
1284 		if (!peer_pdd) {
1285 			pr_debug("Getting device by id failed for 0x%x\n",
1286 				 devices_arr[i]);
1287 			err = -EINVAL;
1288 			goto get_mem_obj_from_handle_failed;
1289 		}
1290 
1291 		peer_pdd = kfd_bind_process_to_device(peer_pdd->dev, p);
1292 		if (IS_ERR(peer_pdd)) {
1293 			err = PTR_ERR(peer_pdd);
1294 			goto get_mem_obj_from_handle_failed;
1295 		}
1296 
1297 		err = amdgpu_amdkfd_gpuvm_map_memory_to_gpu(
1298 			peer_pdd->dev->adev, (struct kgd_mem *)mem,
1299 			peer_pdd->drm_priv);
1300 		if (err) {
1301 			struct pci_dev *pdev = peer_pdd->dev->adev->pdev;
1302 
1303 			dev_err(dev->adev->dev,
1304 			       "Failed to map peer:%04x:%02x:%02x.%d mem_domain:%d\n",
1305 			       pci_domain_nr(pdev->bus),
1306 			       pdev->bus->number,
1307 			       PCI_SLOT(pdev->devfn),
1308 			       PCI_FUNC(pdev->devfn),
1309 			       ((struct kgd_mem *)mem)->domain);
1310 			goto map_memory_to_gpu_failed;
1311 		}
1312 		args->n_success = i+1;
1313 	}
1314 
1315 	mutex_unlock(&p->mutex);
1316 
1317 	err = amdgpu_amdkfd_gpuvm_sync_memory(dev->adev, (struct kgd_mem *) mem, true);
1318 	if (err) {
1319 		pr_debug("Sync memory failed, wait interrupted by user signal\n");
1320 		goto sync_memory_failed;
1321 	}
1322 
1323 	/* Flush TLBs after waiting for the page table updates to complete */
1324 	for (i = 0; i < args->n_devices; i++) {
1325 		peer_pdd = kfd_process_device_data_by_id(p, devices_arr[i]);
1326 		if (WARN_ON_ONCE(!peer_pdd))
1327 			continue;
1328 		kfd_flush_tlb(peer_pdd, TLB_FLUSH_LEGACY);
1329 	}
1330 	kfree(devices_arr);
1331 
1332 	return err;
1333 
1334 get_process_device_data_failed:
1335 bind_process_to_device_failed:
1336 get_mem_obj_from_handle_failed:
1337 map_memory_to_gpu_failed:
1338 	mutex_unlock(&p->mutex);
1339 copy_from_user_failed:
1340 sync_memory_failed:
1341 	kfree(devices_arr);
1342 
1343 	return err;
1344 }
1345 
1346 static int kfd_ioctl_unmap_memory_from_gpu(struct file *filep,
1347 					struct kfd_process *p, void *data)
1348 {
1349 	struct kfd_ioctl_unmap_memory_from_gpu_args *args = data;
1350 	struct kfd_process_device *pdd, *peer_pdd;
1351 	void *mem;
1352 	long err = 0;
1353 	uint32_t *devices_arr = NULL, i;
1354 
1355 	if (!args->n_devices) {
1356 		pr_debug("Device IDs array empty\n");
1357 		return -EINVAL;
1358 	}
1359 	if (args->n_success > args->n_devices) {
1360 		pr_debug("n_success exceeds n_devices\n");
1361 		return -EINVAL;
1362 	}
1363 
1364 	devices_arr = kmalloc_array(args->n_devices, sizeof(*devices_arr),
1365 				    GFP_KERNEL);
1366 	if (!devices_arr)
1367 		return -ENOMEM;
1368 
1369 	err = copy_from_user(devices_arr,
1370 			     (void __user *)args->device_ids_array_ptr,
1371 			     args->n_devices * sizeof(*devices_arr));
1372 	if (err != 0) {
1373 		err = -EFAULT;
1374 		goto copy_from_user_failed;
1375 	}
1376 
1377 	mutex_lock(&p->mutex);
1378 	pdd = kfd_process_device_data_by_id(p, GET_GPU_ID(args->handle));
1379 	if (!pdd) {
1380 		err = -EINVAL;
1381 		goto bind_process_to_device_failed;
1382 	}
1383 
1384 	mem = kfd_process_device_translate_handle(pdd,
1385 						GET_IDR_HANDLE(args->handle));
1386 	if (!mem) {
1387 		err = -ENOMEM;
1388 		goto get_mem_obj_from_handle_failed;
1389 	}
1390 
1391 	for (i = args->n_success; i < args->n_devices; i++) {
1392 		peer_pdd = kfd_process_device_data_by_id(p, devices_arr[i]);
1393 		if (!peer_pdd) {
1394 			err = -EINVAL;
1395 			goto get_mem_obj_from_handle_failed;
1396 		}
1397 		err = amdgpu_amdkfd_gpuvm_unmap_memory_from_gpu(
1398 			peer_pdd->dev->adev, (struct kgd_mem *)mem, peer_pdd->drm_priv);
1399 		if (err) {
1400 			pr_err("Failed to unmap from gpu %d/%d\n",
1401 			       i, args->n_devices);
1402 			goto unmap_memory_from_gpu_failed;
1403 		}
1404 		args->n_success = i+1;
1405 	}
1406 	mutex_unlock(&p->mutex);
1407 
1408 	if (kfd_flush_tlb_after_unmap(pdd->dev)) {
1409 		err = amdgpu_amdkfd_gpuvm_sync_memory(pdd->dev->adev,
1410 				(struct kgd_mem *) mem, true);
1411 		if (err) {
1412 			pr_debug("Sync memory failed, wait interrupted by user signal\n");
1413 			goto sync_memory_failed;
1414 		}
1415 
1416 		/* Flush TLBs after waiting for the page table updates to complete */
1417 		for (i = 0; i < args->n_devices; i++) {
1418 			peer_pdd = kfd_process_device_data_by_id(p, devices_arr[i]);
1419 			if (WARN_ON_ONCE(!peer_pdd))
1420 				continue;
1421 			kfd_flush_tlb(peer_pdd, TLB_FLUSH_HEAVYWEIGHT);
1422 		}
1423 	}
1424 	kfree(devices_arr);
1425 
1426 	return 0;
1427 
1428 bind_process_to_device_failed:
1429 get_mem_obj_from_handle_failed:
1430 unmap_memory_from_gpu_failed:
1431 	mutex_unlock(&p->mutex);
1432 copy_from_user_failed:
1433 sync_memory_failed:
1434 	kfree(devices_arr);
1435 	return err;
1436 }
1437 
1438 static int kfd_ioctl_alloc_queue_gws(struct file *filep,
1439 		struct kfd_process *p, void *data)
1440 {
1441 	int retval;
1442 	struct kfd_ioctl_alloc_queue_gws_args *args = data;
1443 	struct queue *q;
1444 	struct kfd_dev *dev;
1445 
1446 	mutex_lock(&p->mutex);
1447 	q = pqm_get_user_queue(&p->pqm, args->queue_id);
1448 
1449 	if (q) {
1450 		dev = q->device;
1451 	} else {
1452 		retval = -EINVAL;
1453 		goto out_unlock;
1454 	}
1455 
1456 	if (!dev->gws) {
1457 		retval = -ENODEV;
1458 		goto out_unlock;
1459 	}
1460 
1461 	if (dev->dqm->sched_policy == KFD_SCHED_POLICY_NO_HWS) {
1462 		retval = -ENODEV;
1463 		goto out_unlock;
1464 	}
1465 
1466 	retval = pqm_set_gws(&p->pqm, args->queue_id, args->num_gws ? dev->gws : NULL);
1467 	mutex_unlock(&p->mutex);
1468 
1469 	args->first_gws = 0;
1470 	return retval;
1471 
1472 out_unlock:
1473 	mutex_unlock(&p->mutex);
1474 	return retval;
1475 }
1476 
1477 static int kfd_ioctl_get_dmabuf_info(struct file *filep,
1478 		struct kfd_process *p, void *data)
1479 {
1480 	struct kfd_ioctl_get_dmabuf_info_args *args = data;
1481 	struct kfd_dev *dev = NULL;
1482 	struct amdgpu_device *dmabuf_adev;
1483 	void *metadata_buffer = NULL;
1484 	uint32_t flags;
1485 	unsigned int i;
1486 	int r;
1487 
1488 	/* Find a KFD GPU device that supports the get_dmabuf_info query */
1489 	for (i = 0; kfd_topology_enum_kfd_devices(i, &dev) == 0; i++)
1490 		if (dev)
1491 			break;
1492 	if (!dev)
1493 		return -EINVAL;
1494 
1495 	if (args->metadata_ptr) {
1496 		metadata_buffer = kzalloc(args->metadata_size, GFP_KERNEL);
1497 		if (!metadata_buffer)
1498 			return -ENOMEM;
1499 	}
1500 
1501 	/* Get dmabuf info from KGD */
1502 	r = amdgpu_amdkfd_get_dmabuf_info(dev->adev, args->dmabuf_fd,
1503 					  &dmabuf_adev, &args->size,
1504 					  metadata_buffer, args->metadata_size,
1505 					  &args->metadata_size, &flags);
1506 	if (r)
1507 		goto exit;
1508 
1509 	/* Reverse-lookup gpu_id from kgd pointer */
1510 	dev = kfd_device_by_adev(dmabuf_adev);
1511 	if (!dev) {
1512 		r = -EINVAL;
1513 		goto exit;
1514 	}
1515 	args->gpu_id = dev->id;
1516 	args->flags = flags;
1517 
1518 	/* Copy metadata buffer to user mode */
1519 	if (metadata_buffer) {
1520 		r = copy_to_user((void __user *)args->metadata_ptr,
1521 				 metadata_buffer, args->metadata_size);
1522 		if (r != 0)
1523 			r = -EFAULT;
1524 	}
1525 
1526 exit:
1527 	kfree(metadata_buffer);
1528 
1529 	return r;
1530 }
1531 
1532 static int kfd_ioctl_import_dmabuf(struct file *filep,
1533 				   struct kfd_process *p, void *data)
1534 {
1535 	struct kfd_ioctl_import_dmabuf_args *args = data;
1536 	struct kfd_process_device *pdd;
1537 	struct dma_buf *dmabuf;
1538 	int idr_handle;
1539 	uint64_t size;
1540 	void *mem;
1541 	int r;
1542 
1543 	dmabuf = dma_buf_get(args->dmabuf_fd);
1544 	if (IS_ERR(dmabuf))
1545 		return PTR_ERR(dmabuf);
1546 
1547 	mutex_lock(&p->mutex);
1548 	pdd = kfd_process_device_data_by_id(p, args->gpu_id);
1549 	if (!pdd) {
1550 		r = -EINVAL;
1551 		goto err_unlock;
1552 	}
1553 
1554 	pdd = kfd_bind_process_to_device(pdd->dev, p);
1555 	if (IS_ERR(pdd)) {
1556 		r = PTR_ERR(pdd);
1557 		goto err_unlock;
1558 	}
1559 
1560 	r = amdgpu_amdkfd_gpuvm_import_dmabuf(pdd->dev->adev, dmabuf,
1561 					      args->va_addr, pdd->drm_priv,
1562 					      (struct kgd_mem **)&mem, &size,
1563 					      NULL);
1564 	if (r)
1565 		goto err_unlock;
1566 
1567 	idr_handle = kfd_process_device_create_obj_handle(pdd, mem);
1568 	if (idr_handle < 0) {
1569 		r = -EFAULT;
1570 		goto err_free;
1571 	}
1572 
1573 	mutex_unlock(&p->mutex);
1574 	dma_buf_put(dmabuf);
1575 
1576 	args->handle = MAKE_HANDLE(args->gpu_id, idr_handle);
1577 
1578 	return 0;
1579 
1580 err_free:
1581 	amdgpu_amdkfd_gpuvm_free_memory_of_gpu(pdd->dev->adev, (struct kgd_mem *)mem,
1582 					       pdd->drm_priv, NULL);
1583 err_unlock:
1584 	mutex_unlock(&p->mutex);
1585 	dma_buf_put(dmabuf);
1586 	return r;
1587 }
1588 
1589 static int kfd_ioctl_export_dmabuf(struct file *filep,
1590 				   struct kfd_process *p, void *data)
1591 {
1592 	struct kfd_ioctl_export_dmabuf_args *args = data;
1593 	struct kfd_process_device *pdd;
1594 	struct dma_buf *dmabuf;
1595 	struct kfd_dev *dev;
1596 	void *mem;
1597 	int ret = 0;
1598 
1599 	dev = kfd_device_by_id(GET_GPU_ID(args->handle));
1600 	if (!dev)
1601 		return -EINVAL;
1602 
1603 	mutex_lock(&p->mutex);
1604 
1605 	pdd = kfd_get_process_device_data(dev, p);
1606 	if (!pdd) {
1607 		ret = -EINVAL;
1608 		goto err_unlock;
1609 	}
1610 
1611 	mem = kfd_process_device_translate_handle(pdd,
1612 						GET_IDR_HANDLE(args->handle));
1613 	if (!mem) {
1614 		ret = -EINVAL;
1615 		goto err_unlock;
1616 	}
1617 
1618 	ret = amdgpu_amdkfd_gpuvm_export_dmabuf(mem, &dmabuf);
1619 	mutex_unlock(&p->mutex);
1620 	if (ret)
1621 		goto err_out;
1622 
1623 	ret = dma_buf_fd(dmabuf, args->flags);
1624 	if (ret < 0) {
1625 		dma_buf_put(dmabuf);
1626 		goto err_out;
1627 	}
1628 	/* dma_buf_fd assigns the reference count to the fd, no need to
1629 	 * put the reference here.
1630 	 */
1631 	args->dmabuf_fd = ret;
1632 
1633 	return 0;
1634 
1635 err_unlock:
1636 	mutex_unlock(&p->mutex);
1637 err_out:
1638 	return ret;
1639 }
1640 
1641 /* Handle requests for watching SMI events */
1642 static int kfd_ioctl_smi_events(struct file *filep,
1643 				struct kfd_process *p, void *data)
1644 {
1645 	struct kfd_ioctl_smi_events_args *args = data;
1646 	struct kfd_process_device *pdd;
1647 
1648 	mutex_lock(&p->mutex);
1649 
1650 	pdd = kfd_process_device_data_by_id(p, args->gpuid);
1651 	mutex_unlock(&p->mutex);
1652 	if (!pdd)
1653 		return -EINVAL;
1654 
1655 	return kfd_smi_event_open(pdd->dev, &args->anon_fd);
1656 }
1657 
1658 #if IS_ENABLED(CONFIG_HSA_AMD_SVM)
1659 
1660 static int kfd_ioctl_set_xnack_mode(struct file *filep,
1661 				    struct kfd_process *p, void *data)
1662 {
1663 	struct kfd_ioctl_set_xnack_mode_args *args = data;
1664 	int r = 0;
1665 
1666 	mutex_lock(&p->mutex);
1667 	if (args->xnack_enabled >= 0) {
1668 		if (!list_empty(&p->pqm.queues)) {
1669 			pr_debug("Process has user queues running\n");
1670 			r = -EBUSY;
1671 			goto out_unlock;
1672 		}
1673 
1674 		if (p->xnack_enabled == args->xnack_enabled)
1675 			goto out_unlock;
1676 
1677 		if (args->xnack_enabled && !kfd_process_xnack_mode(p, true)) {
1678 			r = -EPERM;
1679 			goto out_unlock;
1680 		}
1681 
1682 		r = svm_range_switch_xnack_reserve_mem(p, args->xnack_enabled);
1683 	} else {
1684 		args->xnack_enabled = p->xnack_enabled;
1685 	}
1686 
1687 out_unlock:
1688 	mutex_unlock(&p->mutex);
1689 
1690 	return r;
1691 }
1692 
1693 static int kfd_ioctl_svm(struct file *filep, struct kfd_process *p, void *data)
1694 {
1695 	struct kfd_ioctl_svm_args *args = data;
1696 	int r = 0;
1697 
1698 	pr_debug("start 0x%llx size 0x%llx op 0x%x nattr 0x%x\n",
1699 		 args->start_addr, args->size, args->op, args->nattr);
1700 
1701 	if ((args->start_addr & ~PAGE_MASK) || (args->size & ~PAGE_MASK))
1702 		return -EINVAL;
1703 	if (!args->start_addr || !args->size)
1704 		return -EINVAL;
1705 
1706 	r = svm_ioctl(p, args->op, args->start_addr, args->size, args->nattr,
1707 		      args->attrs);
1708 
1709 	return r;
1710 }
1711 #else
1712 static int kfd_ioctl_set_xnack_mode(struct file *filep,
1713 				    struct kfd_process *p, void *data)
1714 {
1715 	return -EPERM;
1716 }
1717 static int kfd_ioctl_svm(struct file *filep, struct kfd_process *p, void *data)
1718 {
1719 	return -EPERM;
1720 }
1721 #endif
1722 
1723 static int criu_checkpoint_process(struct kfd_process *p,
1724 			     uint8_t __user *user_priv_data,
1725 			     uint64_t *priv_offset)
1726 {
1727 	struct kfd_criu_process_priv_data process_priv;
1728 	int ret;
1729 
1730 	memset(&process_priv, 0, sizeof(process_priv));
1731 
1732 	process_priv.version = KFD_CRIU_PRIV_VERSION;
1733 	/* For CR, we don't consider negative xnack mode which is used for
1734 	 * querying without changing it, here 0 simply means disabled and 1
1735 	 * means enabled so retry for finding a valid PTE.
1736 	 */
1737 	process_priv.xnack_mode = p->xnack_enabled ? 1 : 0;
1738 
1739 	ret = copy_to_user(user_priv_data + *priv_offset,
1740 				&process_priv, sizeof(process_priv));
1741 
1742 	if (ret) {
1743 		pr_err("Failed to copy process information to user\n");
1744 		ret = -EFAULT;
1745 	}
1746 
1747 	*priv_offset += sizeof(process_priv);
1748 	return ret;
1749 }
1750 
1751 static int criu_checkpoint_devices(struct kfd_process *p,
1752 			     uint32_t num_devices,
1753 			     uint8_t __user *user_addr,
1754 			     uint8_t __user *user_priv_data,
1755 			     uint64_t *priv_offset)
1756 {
1757 	struct kfd_criu_device_priv_data *device_priv = NULL;
1758 	struct kfd_criu_device_bucket *device_buckets = NULL;
1759 	int ret = 0, i;
1760 
1761 	device_buckets = kvzalloc(num_devices * sizeof(*device_buckets), GFP_KERNEL);
1762 	if (!device_buckets) {
1763 		ret = -ENOMEM;
1764 		goto exit;
1765 	}
1766 
1767 	device_priv = kvzalloc(num_devices * sizeof(*device_priv), GFP_KERNEL);
1768 	if (!device_priv) {
1769 		ret = -ENOMEM;
1770 		goto exit;
1771 	}
1772 
1773 	for (i = 0; i < num_devices; i++) {
1774 		struct kfd_process_device *pdd = p->pdds[i];
1775 
1776 		device_buckets[i].user_gpu_id = pdd->user_gpu_id;
1777 		device_buckets[i].actual_gpu_id = pdd->dev->id;
1778 
1779 		/*
1780 		 * priv_data does not contain useful information for now and is reserved for
1781 		 * future use, so we do not set its contents.
1782 		 */
1783 	}
1784 
1785 	ret = copy_to_user(user_addr, device_buckets, num_devices * sizeof(*device_buckets));
1786 	if (ret) {
1787 		pr_err("Failed to copy device information to user\n");
1788 		ret = -EFAULT;
1789 		goto exit;
1790 	}
1791 
1792 	ret = copy_to_user(user_priv_data + *priv_offset,
1793 			   device_priv,
1794 			   num_devices * sizeof(*device_priv));
1795 	if (ret) {
1796 		pr_err("Failed to copy device information to user\n");
1797 		ret = -EFAULT;
1798 	}
1799 	*priv_offset += num_devices * sizeof(*device_priv);
1800 
1801 exit:
1802 	kvfree(device_buckets);
1803 	kvfree(device_priv);
1804 	return ret;
1805 }
1806 
1807 static uint32_t get_process_num_bos(struct kfd_process *p)
1808 {
1809 	uint32_t num_of_bos = 0;
1810 	int i;
1811 
1812 	/* Run over all PDDs of the process */
1813 	for (i = 0; i < p->n_pdds; i++) {
1814 		struct kfd_process_device *pdd = p->pdds[i];
1815 		void *mem;
1816 		int id;
1817 
1818 		idr_for_each_entry(&pdd->alloc_idr, mem, id) {
1819 			struct kgd_mem *kgd_mem = (struct kgd_mem *)mem;
1820 
1821 			if ((uint64_t)kgd_mem->va > pdd->gpuvm_base)
1822 				num_of_bos++;
1823 		}
1824 	}
1825 	return num_of_bos;
1826 }
1827 
1828 static int criu_get_prime_handle(struct drm_gem_object *gobj, int flags,
1829 				      u32 *shared_fd)
1830 {
1831 	struct dma_buf *dmabuf;
1832 	int ret;
1833 
1834 	dmabuf = amdgpu_gem_prime_export(gobj, flags);
1835 	if (IS_ERR(dmabuf)) {
1836 		ret = PTR_ERR(dmabuf);
1837 		pr_err("dmabuf export failed for the BO\n");
1838 		return ret;
1839 	}
1840 
1841 	ret = dma_buf_fd(dmabuf, flags);
1842 	if (ret < 0) {
1843 		pr_err("dmabuf create fd failed, ret:%d\n", ret);
1844 		goto out_free_dmabuf;
1845 	}
1846 
1847 	*shared_fd = ret;
1848 	return 0;
1849 
1850 out_free_dmabuf:
1851 	dma_buf_put(dmabuf);
1852 	return ret;
1853 }
1854 
1855 static int criu_checkpoint_bos(struct kfd_process *p,
1856 			       uint32_t num_bos,
1857 			       uint8_t __user *user_bos,
1858 			       uint8_t __user *user_priv_data,
1859 			       uint64_t *priv_offset)
1860 {
1861 	struct kfd_criu_bo_bucket *bo_buckets;
1862 	struct kfd_criu_bo_priv_data *bo_privs;
1863 	int ret = 0, pdd_index, bo_index = 0, id;
1864 	void *mem;
1865 
1866 	bo_buckets = kvzalloc(num_bos * sizeof(*bo_buckets), GFP_KERNEL);
1867 	if (!bo_buckets)
1868 		return -ENOMEM;
1869 
1870 	bo_privs = kvzalloc(num_bos * sizeof(*bo_privs), GFP_KERNEL);
1871 	if (!bo_privs) {
1872 		ret = -ENOMEM;
1873 		goto exit;
1874 	}
1875 
1876 	for (pdd_index = 0; pdd_index < p->n_pdds; pdd_index++) {
1877 		struct kfd_process_device *pdd = p->pdds[pdd_index];
1878 		struct amdgpu_bo *dumper_bo;
1879 		struct kgd_mem *kgd_mem;
1880 
1881 		idr_for_each_entry(&pdd->alloc_idr, mem, id) {
1882 			struct kfd_criu_bo_bucket *bo_bucket;
1883 			struct kfd_criu_bo_priv_data *bo_priv;
1884 			int i, dev_idx = 0;
1885 
1886 			if (!mem) {
1887 				ret = -ENOMEM;
1888 				goto exit;
1889 			}
1890 
1891 			kgd_mem = (struct kgd_mem *)mem;
1892 			dumper_bo = kgd_mem->bo;
1893 
1894 			if ((uint64_t)kgd_mem->va <= pdd->gpuvm_base)
1895 				continue;
1896 
1897 			bo_bucket = &bo_buckets[bo_index];
1898 			bo_priv = &bo_privs[bo_index];
1899 
1900 			bo_bucket->gpu_id = pdd->user_gpu_id;
1901 			bo_bucket->addr = (uint64_t)kgd_mem->va;
1902 			bo_bucket->size = amdgpu_bo_size(dumper_bo);
1903 			bo_bucket->alloc_flags = (uint32_t)kgd_mem->alloc_flags;
1904 			bo_priv->idr_handle = id;
1905 
1906 			if (bo_bucket->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_USERPTR) {
1907 				ret = amdgpu_ttm_tt_get_userptr(&dumper_bo->tbo,
1908 								&bo_priv->user_addr);
1909 				if (ret) {
1910 					pr_err("Failed to obtain user address for user-pointer bo\n");
1911 					goto exit;
1912 				}
1913 			}
1914 			if (bo_bucket->alloc_flags
1915 			    & (KFD_IOC_ALLOC_MEM_FLAGS_VRAM | KFD_IOC_ALLOC_MEM_FLAGS_GTT)) {
1916 				ret = criu_get_prime_handle(&dumper_bo->tbo.base,
1917 						bo_bucket->alloc_flags &
1918 						KFD_IOC_ALLOC_MEM_FLAGS_WRITABLE ? DRM_RDWR : 0,
1919 						&bo_bucket->dmabuf_fd);
1920 				if (ret)
1921 					goto exit;
1922 			} else {
1923 				bo_bucket->dmabuf_fd = KFD_INVALID_FD;
1924 			}
1925 
1926 			if (bo_bucket->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_DOORBELL)
1927 				bo_bucket->offset = KFD_MMAP_TYPE_DOORBELL |
1928 					KFD_MMAP_GPU_ID(pdd->dev->id);
1929 			else if (bo_bucket->alloc_flags &
1930 				KFD_IOC_ALLOC_MEM_FLAGS_MMIO_REMAP)
1931 				bo_bucket->offset = KFD_MMAP_TYPE_MMIO |
1932 					KFD_MMAP_GPU_ID(pdd->dev->id);
1933 			else
1934 				bo_bucket->offset = amdgpu_bo_mmap_offset(dumper_bo);
1935 
1936 			for (i = 0; i < p->n_pdds; i++) {
1937 				if (amdgpu_amdkfd_bo_mapped_to_dev(p->pdds[i]->dev->adev, kgd_mem))
1938 					bo_priv->mapped_gpuids[dev_idx++] = p->pdds[i]->user_gpu_id;
1939 			}
1940 
1941 			pr_debug("bo_size = 0x%llx, bo_addr = 0x%llx bo_offset = 0x%llx\n"
1942 					"gpu_id = 0x%x alloc_flags = 0x%x idr_handle = 0x%x",
1943 					bo_bucket->size,
1944 					bo_bucket->addr,
1945 					bo_bucket->offset,
1946 					bo_bucket->gpu_id,
1947 					bo_bucket->alloc_flags,
1948 					bo_priv->idr_handle);
1949 			bo_index++;
1950 		}
1951 	}
1952 
1953 	ret = copy_to_user(user_bos, bo_buckets, num_bos * sizeof(*bo_buckets));
1954 	if (ret) {
1955 		pr_err("Failed to copy BO information to user\n");
1956 		ret = -EFAULT;
1957 		goto exit;
1958 	}
1959 
1960 	ret = copy_to_user(user_priv_data + *priv_offset, bo_privs, num_bos * sizeof(*bo_privs));
1961 	if (ret) {
1962 		pr_err("Failed to copy BO priv information to user\n");
1963 		ret = -EFAULT;
1964 		goto exit;
1965 	}
1966 
1967 	*priv_offset += num_bos * sizeof(*bo_privs);
1968 
1969 exit:
1970 	while (ret && bo_index--) {
1971 		if (bo_buckets[bo_index].alloc_flags
1972 		    & (KFD_IOC_ALLOC_MEM_FLAGS_VRAM | KFD_IOC_ALLOC_MEM_FLAGS_GTT))
1973 			close_fd(bo_buckets[bo_index].dmabuf_fd);
1974 	}
1975 
1976 	kvfree(bo_buckets);
1977 	kvfree(bo_privs);
1978 	return ret;
1979 }
1980 
1981 static int criu_get_process_object_info(struct kfd_process *p,
1982 					uint32_t *num_devices,
1983 					uint32_t *num_bos,
1984 					uint32_t *num_objects,
1985 					uint64_t *objs_priv_size)
1986 {
1987 	uint64_t queues_priv_data_size, svm_priv_data_size, priv_size;
1988 	uint32_t num_queues, num_events, num_svm_ranges;
1989 	int ret;
1990 
1991 	*num_devices = p->n_pdds;
1992 	*num_bos = get_process_num_bos(p);
1993 
1994 	ret = kfd_process_get_queue_info(p, &num_queues, &queues_priv_data_size);
1995 	if (ret)
1996 		return ret;
1997 
1998 	num_events = kfd_get_num_events(p);
1999 
2000 	ret = svm_range_get_info(p, &num_svm_ranges, &svm_priv_data_size);
2001 	if (ret)
2002 		return ret;
2003 
2004 	*num_objects = num_queues + num_events + num_svm_ranges;
2005 
2006 	if (objs_priv_size) {
2007 		priv_size = sizeof(struct kfd_criu_process_priv_data);
2008 		priv_size += *num_devices * sizeof(struct kfd_criu_device_priv_data);
2009 		priv_size += *num_bos * sizeof(struct kfd_criu_bo_priv_data);
2010 		priv_size += queues_priv_data_size;
2011 		priv_size += num_events * sizeof(struct kfd_criu_event_priv_data);
2012 		priv_size += svm_priv_data_size;
2013 		*objs_priv_size = priv_size;
2014 	}
2015 	return 0;
2016 }
2017 
2018 static int criu_checkpoint(struct file *filep,
2019 			   struct kfd_process *p,
2020 			   struct kfd_ioctl_criu_args *args)
2021 {
2022 	int ret;
2023 	uint32_t num_devices, num_bos, num_objects;
2024 	uint64_t priv_size, priv_offset = 0, bo_priv_offset;
2025 
2026 	if (!args->devices || !args->bos || !args->priv_data)
2027 		return -EINVAL;
2028 
2029 	mutex_lock(&p->mutex);
2030 
2031 	if (!p->n_pdds) {
2032 		pr_err("No pdd for given process\n");
2033 		ret = -ENODEV;
2034 		goto exit_unlock;
2035 	}
2036 
2037 	/* Confirm all process queues are evicted */
2038 	if (!p->queues_paused) {
2039 		pr_err("Cannot dump process when queues are not in evicted state\n");
2040 		/* CRIU plugin did not call op PROCESS_INFO before checkpointing */
2041 		ret = -EINVAL;
2042 		goto exit_unlock;
2043 	}
2044 
2045 	ret = criu_get_process_object_info(p, &num_devices, &num_bos, &num_objects, &priv_size);
2046 	if (ret)
2047 		goto exit_unlock;
2048 
2049 	if (num_devices != args->num_devices ||
2050 	    num_bos != args->num_bos ||
2051 	    num_objects != args->num_objects ||
2052 	    priv_size != args->priv_data_size) {
2053 
2054 		ret = -EINVAL;
2055 		goto exit_unlock;
2056 	}
2057 
2058 	/* each function will store private data inside priv_data and adjust priv_offset */
2059 	ret = criu_checkpoint_process(p, (uint8_t __user *)args->priv_data, &priv_offset);
2060 	if (ret)
2061 		goto exit_unlock;
2062 
2063 	ret = criu_checkpoint_devices(p, num_devices, (uint8_t __user *)args->devices,
2064 				(uint8_t __user *)args->priv_data, &priv_offset);
2065 	if (ret)
2066 		goto exit_unlock;
2067 
2068 	/* Leave room for BOs in the private data. They need to be restored
2069 	 * before events, but we checkpoint them last to simplify the error
2070 	 * handling.
2071 	 */
2072 	bo_priv_offset = priv_offset;
2073 	priv_offset += num_bos * sizeof(struct kfd_criu_bo_priv_data);
2074 
2075 	if (num_objects) {
2076 		ret = kfd_criu_checkpoint_queues(p, (uint8_t __user *)args->priv_data,
2077 						 &priv_offset);
2078 		if (ret)
2079 			goto exit_unlock;
2080 
2081 		ret = kfd_criu_checkpoint_events(p, (uint8_t __user *)args->priv_data,
2082 						 &priv_offset);
2083 		if (ret)
2084 			goto exit_unlock;
2085 
2086 		ret = kfd_criu_checkpoint_svm(p, (uint8_t __user *)args->priv_data, &priv_offset);
2087 		if (ret)
2088 			goto exit_unlock;
2089 	}
2090 
2091 	/* This must be the last thing in this function that can fail.
2092 	 * Otherwise we leak dmabuf file descriptors.
2093 	 */
2094 	ret = criu_checkpoint_bos(p, num_bos, (uint8_t __user *)args->bos,
2095 			   (uint8_t __user *)args->priv_data, &bo_priv_offset);
2096 
2097 exit_unlock:
2098 	mutex_unlock(&p->mutex);
2099 	if (ret)
2100 		pr_err("Failed to dump CRIU ret:%d\n", ret);
2101 	else
2102 		pr_debug("CRIU dump ret:%d\n", ret);
2103 
2104 	return ret;
2105 }
2106 
2107 static int criu_restore_process(struct kfd_process *p,
2108 				struct kfd_ioctl_criu_args *args,
2109 				uint64_t *priv_offset,
2110 				uint64_t max_priv_data_size)
2111 {
2112 	int ret = 0;
2113 	struct kfd_criu_process_priv_data process_priv;
2114 
2115 	if (*priv_offset + sizeof(process_priv) > max_priv_data_size)
2116 		return -EINVAL;
2117 
2118 	ret = copy_from_user(&process_priv,
2119 				(void __user *)(args->priv_data + *priv_offset),
2120 				sizeof(process_priv));
2121 	if (ret) {
2122 		pr_err("Failed to copy process private information from user\n");
2123 		ret = -EFAULT;
2124 		goto exit;
2125 	}
2126 	*priv_offset += sizeof(process_priv);
2127 
2128 	if (process_priv.version != KFD_CRIU_PRIV_VERSION) {
2129 		pr_err("Invalid CRIU API version (checkpointed:%d current:%d)\n",
2130 			process_priv.version, KFD_CRIU_PRIV_VERSION);
2131 		return -EINVAL;
2132 	}
2133 
2134 	pr_debug("Setting XNACK mode\n");
2135 	if (process_priv.xnack_mode && !kfd_process_xnack_mode(p, true)) {
2136 		pr_err("xnack mode cannot be set\n");
2137 		ret = -EPERM;
2138 		goto exit;
2139 	} else {
2140 		pr_debug("set xnack mode: %d\n", process_priv.xnack_mode);
2141 		p->xnack_enabled = process_priv.xnack_mode;
2142 	}
2143 
2144 exit:
2145 	return ret;
2146 }
2147 
2148 static int criu_restore_devices(struct kfd_process *p,
2149 				struct kfd_ioctl_criu_args *args,
2150 				uint64_t *priv_offset,
2151 				uint64_t max_priv_data_size)
2152 {
2153 	struct kfd_criu_device_bucket *device_buckets;
2154 	struct kfd_criu_device_priv_data *device_privs;
2155 	int ret = 0;
2156 	uint32_t i;
2157 
2158 	if (args->num_devices != p->n_pdds)
2159 		return -EINVAL;
2160 
2161 	if (*priv_offset + (args->num_devices * sizeof(*device_privs)) > max_priv_data_size)
2162 		return -EINVAL;
2163 
2164 	device_buckets = kmalloc_array(args->num_devices, sizeof(*device_buckets), GFP_KERNEL);
2165 	if (!device_buckets)
2166 		return -ENOMEM;
2167 
2168 	ret = copy_from_user(device_buckets, (void __user *)args->devices,
2169 				args->num_devices * sizeof(*device_buckets));
2170 	if (ret) {
2171 		pr_err("Failed to copy devices buckets from user\n");
2172 		ret = -EFAULT;
2173 		goto exit;
2174 	}
2175 
2176 	for (i = 0; i < args->num_devices; i++) {
2177 		struct kfd_dev *dev;
2178 		struct kfd_process_device *pdd;
2179 		struct file *drm_file;
2180 
2181 		/* device private data is not currently used */
2182 
2183 		if (!device_buckets[i].user_gpu_id) {
2184 			pr_err("Invalid user gpu_id\n");
2185 			ret = -EINVAL;
2186 			goto exit;
2187 		}
2188 
2189 		dev = kfd_device_by_id(device_buckets[i].actual_gpu_id);
2190 		if (!dev) {
2191 			pr_err("Failed to find device with gpu_id = %x\n",
2192 				device_buckets[i].actual_gpu_id);
2193 			ret = -EINVAL;
2194 			goto exit;
2195 		}
2196 
2197 		pdd = kfd_get_process_device_data(dev, p);
2198 		if (!pdd) {
2199 			pr_err("Failed to get pdd for gpu_id = %x\n",
2200 					device_buckets[i].actual_gpu_id);
2201 			ret = -EINVAL;
2202 			goto exit;
2203 		}
2204 		pdd->user_gpu_id = device_buckets[i].user_gpu_id;
2205 
2206 		drm_file = fget(device_buckets[i].drm_fd);
2207 		if (!drm_file) {
2208 			pr_err("Invalid render node file descriptor sent from plugin (%d)\n",
2209 				device_buckets[i].drm_fd);
2210 			ret = -EINVAL;
2211 			goto exit;
2212 		}
2213 
2214 		if (pdd->drm_file) {
2215 			ret = -EINVAL;
2216 			goto exit;
2217 		}
2218 
2219 		/* create the vm using render nodes for kfd pdd */
2220 		if (kfd_process_device_init_vm(pdd, drm_file)) {
2221 			pr_err("could not init vm for given pdd\n");
2222 			/* On success, the PDD keeps the drm_file reference */
2223 			fput(drm_file);
2224 			ret = -EINVAL;
2225 			goto exit;
2226 		}
2227 		/*
2228 		 * pdd now already has the vm bound to render node so below api won't create a new
2229 		 * exclusive kfd mapping but use existing one with renderDXXX but is still needed
2230 		 * for iommu v2 binding  and runtime pm.
2231 		 */
2232 		pdd = kfd_bind_process_to_device(dev, p);
2233 		if (IS_ERR(pdd)) {
2234 			ret = PTR_ERR(pdd);
2235 			goto exit;
2236 		}
2237 
2238 		if (!pdd->doorbell_index &&
2239 		    kfd_alloc_process_doorbells(pdd->dev, &pdd->doorbell_index) < 0) {
2240 			ret = -ENOMEM;
2241 			goto exit;
2242 		}
2243 	}
2244 
2245 	/*
2246 	 * We are not copying device private data from user as we are not using the data for now,
2247 	 * but we still adjust for its private data.
2248 	 */
2249 	*priv_offset += args->num_devices * sizeof(*device_privs);
2250 
2251 exit:
2252 	kfree(device_buckets);
2253 	return ret;
2254 }
2255 
2256 static int criu_restore_memory_of_gpu(struct kfd_process_device *pdd,
2257 				      struct kfd_criu_bo_bucket *bo_bucket,
2258 				      struct kfd_criu_bo_priv_data *bo_priv,
2259 				      struct kgd_mem **kgd_mem)
2260 {
2261 	int idr_handle;
2262 	int ret;
2263 	const bool criu_resume = true;
2264 	u64 offset;
2265 
2266 	if (bo_bucket->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_DOORBELL) {
2267 		if (bo_bucket->size != kfd_doorbell_process_slice(pdd->dev))
2268 			return -EINVAL;
2269 
2270 		offset = kfd_get_process_doorbells(pdd);
2271 		if (!offset)
2272 			return -ENOMEM;
2273 	} else if (bo_bucket->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_MMIO_REMAP) {
2274 		/* MMIO BOs need remapped bus address */
2275 		if (bo_bucket->size != PAGE_SIZE) {
2276 			pr_err("Invalid page size\n");
2277 			return -EINVAL;
2278 		}
2279 		offset = pdd->dev->adev->rmmio_remap.bus_addr;
2280 		if (!offset) {
2281 			pr_err("amdgpu_amdkfd_get_mmio_remap_phys_addr failed\n");
2282 			return -ENOMEM;
2283 		}
2284 	} else if (bo_bucket->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_USERPTR) {
2285 		offset = bo_priv->user_addr;
2286 	}
2287 	/* Create the BO */
2288 	ret = amdgpu_amdkfd_gpuvm_alloc_memory_of_gpu(pdd->dev->adev, bo_bucket->addr,
2289 						      bo_bucket->size, pdd->drm_priv, kgd_mem,
2290 						      &offset, bo_bucket->alloc_flags, criu_resume);
2291 	if (ret) {
2292 		pr_err("Could not create the BO\n");
2293 		return ret;
2294 	}
2295 	pr_debug("New BO created: size:0x%llx addr:0x%llx offset:0x%llx\n",
2296 		 bo_bucket->size, bo_bucket->addr, offset);
2297 
2298 	/* Restore previous IDR handle */
2299 	pr_debug("Restoring old IDR handle for the BO");
2300 	idr_handle = idr_alloc(&pdd->alloc_idr, *kgd_mem, bo_priv->idr_handle,
2301 			       bo_priv->idr_handle + 1, GFP_KERNEL);
2302 
2303 	if (idr_handle < 0) {
2304 		pr_err("Could not allocate idr\n");
2305 		amdgpu_amdkfd_gpuvm_free_memory_of_gpu(pdd->dev->adev, *kgd_mem, pdd->drm_priv,
2306 						       NULL);
2307 		return -ENOMEM;
2308 	}
2309 
2310 	if (bo_bucket->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_DOORBELL)
2311 		bo_bucket->restored_offset = KFD_MMAP_TYPE_DOORBELL | KFD_MMAP_GPU_ID(pdd->dev->id);
2312 	if (bo_bucket->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_MMIO_REMAP) {
2313 		bo_bucket->restored_offset = KFD_MMAP_TYPE_MMIO | KFD_MMAP_GPU_ID(pdd->dev->id);
2314 	} else if (bo_bucket->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_GTT) {
2315 		bo_bucket->restored_offset = offset;
2316 	} else if (bo_bucket->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_VRAM) {
2317 		bo_bucket->restored_offset = offset;
2318 		/* Update the VRAM usage count */
2319 		WRITE_ONCE(pdd->vram_usage, pdd->vram_usage + bo_bucket->size);
2320 	}
2321 	return 0;
2322 }
2323 
2324 static int criu_restore_bo(struct kfd_process *p,
2325 			   struct kfd_criu_bo_bucket *bo_bucket,
2326 			   struct kfd_criu_bo_priv_data *bo_priv)
2327 {
2328 	struct kfd_process_device *pdd;
2329 	struct kgd_mem *kgd_mem;
2330 	int ret;
2331 	int j;
2332 
2333 	pr_debug("Restoring BO size:0x%llx addr:0x%llx gpu_id:0x%x flags:0x%x idr_handle:0x%x\n",
2334 		 bo_bucket->size, bo_bucket->addr, bo_bucket->gpu_id, bo_bucket->alloc_flags,
2335 		 bo_priv->idr_handle);
2336 
2337 	pdd = kfd_process_device_data_by_id(p, bo_bucket->gpu_id);
2338 	if (!pdd) {
2339 		pr_err("Failed to get pdd\n");
2340 		return -ENODEV;
2341 	}
2342 
2343 	ret = criu_restore_memory_of_gpu(pdd, bo_bucket, bo_priv, &kgd_mem);
2344 	if (ret)
2345 		return ret;
2346 
2347 	/* now map these BOs to GPU/s */
2348 	for (j = 0; j < p->n_pdds; j++) {
2349 		struct kfd_dev *peer;
2350 		struct kfd_process_device *peer_pdd;
2351 
2352 		if (!bo_priv->mapped_gpuids[j])
2353 			break;
2354 
2355 		peer_pdd = kfd_process_device_data_by_id(p, bo_priv->mapped_gpuids[j]);
2356 		if (!peer_pdd)
2357 			return -EINVAL;
2358 
2359 		peer = peer_pdd->dev;
2360 
2361 		peer_pdd = kfd_bind_process_to_device(peer, p);
2362 		if (IS_ERR(peer_pdd))
2363 			return PTR_ERR(peer_pdd);
2364 
2365 		ret = amdgpu_amdkfd_gpuvm_map_memory_to_gpu(peer->adev, kgd_mem,
2366 							    peer_pdd->drm_priv);
2367 		if (ret) {
2368 			pr_err("Failed to map to gpu %d/%d\n", j, p->n_pdds);
2369 			return ret;
2370 		}
2371 	}
2372 
2373 	pr_debug("map memory was successful for the BO\n");
2374 	/* create the dmabuf object and export the bo */
2375 	if (bo_bucket->alloc_flags
2376 	    & (KFD_IOC_ALLOC_MEM_FLAGS_VRAM | KFD_IOC_ALLOC_MEM_FLAGS_GTT)) {
2377 		ret = criu_get_prime_handle(&kgd_mem->bo->tbo.base, DRM_RDWR,
2378 					    &bo_bucket->dmabuf_fd);
2379 		if (ret)
2380 			return ret;
2381 	} else {
2382 		bo_bucket->dmabuf_fd = KFD_INVALID_FD;
2383 	}
2384 
2385 	return 0;
2386 }
2387 
2388 static int criu_restore_bos(struct kfd_process *p,
2389 			    struct kfd_ioctl_criu_args *args,
2390 			    uint64_t *priv_offset,
2391 			    uint64_t max_priv_data_size)
2392 {
2393 	struct kfd_criu_bo_bucket *bo_buckets = NULL;
2394 	struct kfd_criu_bo_priv_data *bo_privs = NULL;
2395 	int ret = 0;
2396 	uint32_t i = 0;
2397 
2398 	if (*priv_offset + (args->num_bos * sizeof(*bo_privs)) > max_priv_data_size)
2399 		return -EINVAL;
2400 
2401 	/* Prevent MMU notifications until stage-4 IOCTL (CRIU_RESUME) is received */
2402 	amdgpu_amdkfd_block_mmu_notifications(p->kgd_process_info);
2403 
2404 	bo_buckets = kvmalloc_array(args->num_bos, sizeof(*bo_buckets), GFP_KERNEL);
2405 	if (!bo_buckets)
2406 		return -ENOMEM;
2407 
2408 	ret = copy_from_user(bo_buckets, (void __user *)args->bos,
2409 			     args->num_bos * sizeof(*bo_buckets));
2410 	if (ret) {
2411 		pr_err("Failed to copy BOs information from user\n");
2412 		ret = -EFAULT;
2413 		goto exit;
2414 	}
2415 
2416 	bo_privs = kvmalloc_array(args->num_bos, sizeof(*bo_privs), GFP_KERNEL);
2417 	if (!bo_privs) {
2418 		ret = -ENOMEM;
2419 		goto exit;
2420 	}
2421 
2422 	ret = copy_from_user(bo_privs, (void __user *)args->priv_data + *priv_offset,
2423 			     args->num_bos * sizeof(*bo_privs));
2424 	if (ret) {
2425 		pr_err("Failed to copy BOs information from user\n");
2426 		ret = -EFAULT;
2427 		goto exit;
2428 	}
2429 	*priv_offset += args->num_bos * sizeof(*bo_privs);
2430 
2431 	/* Create and map new BOs */
2432 	for (; i < args->num_bos; i++) {
2433 		ret = criu_restore_bo(p, &bo_buckets[i], &bo_privs[i]);
2434 		if (ret) {
2435 			pr_debug("Failed to restore BO[%d] ret%d\n", i, ret);
2436 			goto exit;
2437 		}
2438 	} /* done */
2439 
2440 	/* Copy only the buckets back so user can read bo_buckets[N].restored_offset */
2441 	ret = copy_to_user((void __user *)args->bos,
2442 				bo_buckets,
2443 				(args->num_bos * sizeof(*bo_buckets)));
2444 	if (ret)
2445 		ret = -EFAULT;
2446 
2447 exit:
2448 	while (ret && i--) {
2449 		if (bo_buckets[i].alloc_flags
2450 		   & (KFD_IOC_ALLOC_MEM_FLAGS_VRAM | KFD_IOC_ALLOC_MEM_FLAGS_GTT))
2451 			close_fd(bo_buckets[i].dmabuf_fd);
2452 	}
2453 	kvfree(bo_buckets);
2454 	kvfree(bo_privs);
2455 	return ret;
2456 }
2457 
2458 static int criu_restore_objects(struct file *filep,
2459 				struct kfd_process *p,
2460 				struct kfd_ioctl_criu_args *args,
2461 				uint64_t *priv_offset,
2462 				uint64_t max_priv_data_size)
2463 {
2464 	int ret = 0;
2465 	uint32_t i;
2466 
2467 	BUILD_BUG_ON(offsetof(struct kfd_criu_queue_priv_data, object_type));
2468 	BUILD_BUG_ON(offsetof(struct kfd_criu_event_priv_data, object_type));
2469 	BUILD_BUG_ON(offsetof(struct kfd_criu_svm_range_priv_data, object_type));
2470 
2471 	for (i = 0; i < args->num_objects; i++) {
2472 		uint32_t object_type;
2473 
2474 		if (*priv_offset + sizeof(object_type) > max_priv_data_size) {
2475 			pr_err("Invalid private data size\n");
2476 			return -EINVAL;
2477 		}
2478 
2479 		ret = get_user(object_type, (uint32_t __user *)(args->priv_data + *priv_offset));
2480 		if (ret) {
2481 			pr_err("Failed to copy private information from user\n");
2482 			goto exit;
2483 		}
2484 
2485 		switch (object_type) {
2486 		case KFD_CRIU_OBJECT_TYPE_QUEUE:
2487 			ret = kfd_criu_restore_queue(p, (uint8_t __user *)args->priv_data,
2488 						     priv_offset, max_priv_data_size);
2489 			if (ret)
2490 				goto exit;
2491 			break;
2492 		case KFD_CRIU_OBJECT_TYPE_EVENT:
2493 			ret = kfd_criu_restore_event(filep, p, (uint8_t __user *)args->priv_data,
2494 						     priv_offset, max_priv_data_size);
2495 			if (ret)
2496 				goto exit;
2497 			break;
2498 		case KFD_CRIU_OBJECT_TYPE_SVM_RANGE:
2499 			ret = kfd_criu_restore_svm(p, (uint8_t __user *)args->priv_data,
2500 						     priv_offset, max_priv_data_size);
2501 			if (ret)
2502 				goto exit;
2503 			break;
2504 		default:
2505 			pr_err("Invalid object type:%u at index:%d\n", object_type, i);
2506 			ret = -EINVAL;
2507 			goto exit;
2508 		}
2509 	}
2510 exit:
2511 	return ret;
2512 }
2513 
2514 static int criu_restore(struct file *filep,
2515 			struct kfd_process *p,
2516 			struct kfd_ioctl_criu_args *args)
2517 {
2518 	uint64_t priv_offset = 0;
2519 	int ret = 0;
2520 
2521 	pr_debug("CRIU restore (num_devices:%u num_bos:%u num_objects:%u priv_data_size:%llu)\n",
2522 		 args->num_devices, args->num_bos, args->num_objects, args->priv_data_size);
2523 
2524 	if (!args->bos || !args->devices || !args->priv_data || !args->priv_data_size ||
2525 	    !args->num_devices || !args->num_bos)
2526 		return -EINVAL;
2527 
2528 	mutex_lock(&p->mutex);
2529 
2530 	/*
2531 	 * Set the process to evicted state to avoid running any new queues before all the memory
2532 	 * mappings are ready.
2533 	 */
2534 	ret = kfd_process_evict_queues(p, KFD_QUEUE_EVICTION_CRIU_RESTORE);
2535 	if (ret)
2536 		goto exit_unlock;
2537 
2538 	/* Each function will adjust priv_offset based on how many bytes they consumed */
2539 	ret = criu_restore_process(p, args, &priv_offset, args->priv_data_size);
2540 	if (ret)
2541 		goto exit_unlock;
2542 
2543 	ret = criu_restore_devices(p, args, &priv_offset, args->priv_data_size);
2544 	if (ret)
2545 		goto exit_unlock;
2546 
2547 	ret = criu_restore_bos(p, args, &priv_offset, args->priv_data_size);
2548 	if (ret)
2549 		goto exit_unlock;
2550 
2551 	ret = criu_restore_objects(filep, p, args, &priv_offset, args->priv_data_size);
2552 	if (ret)
2553 		goto exit_unlock;
2554 
2555 	if (priv_offset != args->priv_data_size) {
2556 		pr_err("Invalid private data size\n");
2557 		ret = -EINVAL;
2558 	}
2559 
2560 exit_unlock:
2561 	mutex_unlock(&p->mutex);
2562 	if (ret)
2563 		pr_err("Failed to restore CRIU ret:%d\n", ret);
2564 	else
2565 		pr_debug("CRIU restore successful\n");
2566 
2567 	return ret;
2568 }
2569 
2570 static int criu_unpause(struct file *filep,
2571 			struct kfd_process *p,
2572 			struct kfd_ioctl_criu_args *args)
2573 {
2574 	int ret;
2575 
2576 	mutex_lock(&p->mutex);
2577 
2578 	if (!p->queues_paused) {
2579 		mutex_unlock(&p->mutex);
2580 		return -EINVAL;
2581 	}
2582 
2583 	ret = kfd_process_restore_queues(p);
2584 	if (ret)
2585 		pr_err("Failed to unpause queues ret:%d\n", ret);
2586 	else
2587 		p->queues_paused = false;
2588 
2589 	mutex_unlock(&p->mutex);
2590 
2591 	return ret;
2592 }
2593 
2594 static int criu_resume(struct file *filep,
2595 			struct kfd_process *p,
2596 			struct kfd_ioctl_criu_args *args)
2597 {
2598 	struct kfd_process *target = NULL;
2599 	struct pid *pid = NULL;
2600 	int ret = 0;
2601 
2602 	pr_debug("Inside %s, target pid for criu restore: %d\n", __func__,
2603 		 args->pid);
2604 
2605 	pid = find_get_pid(args->pid);
2606 	if (!pid) {
2607 		pr_err("Cannot find pid info for %i\n", args->pid);
2608 		return -ESRCH;
2609 	}
2610 
2611 	pr_debug("calling kfd_lookup_process_by_pid\n");
2612 	target = kfd_lookup_process_by_pid(pid);
2613 
2614 	put_pid(pid);
2615 
2616 	if (!target) {
2617 		pr_debug("Cannot find process info for %i\n", args->pid);
2618 		return -ESRCH;
2619 	}
2620 
2621 	mutex_lock(&target->mutex);
2622 	ret = kfd_criu_resume_svm(target);
2623 	if (ret) {
2624 		pr_err("kfd_criu_resume_svm failed for %i\n", args->pid);
2625 		goto exit;
2626 	}
2627 
2628 	ret =  amdgpu_amdkfd_criu_resume(target->kgd_process_info);
2629 	if (ret)
2630 		pr_err("amdgpu_amdkfd_criu_resume failed for %i\n", args->pid);
2631 
2632 exit:
2633 	mutex_unlock(&target->mutex);
2634 
2635 	kfd_unref_process(target);
2636 	return ret;
2637 }
2638 
2639 static int criu_process_info(struct file *filep,
2640 				struct kfd_process *p,
2641 				struct kfd_ioctl_criu_args *args)
2642 {
2643 	int ret = 0;
2644 
2645 	mutex_lock(&p->mutex);
2646 
2647 	if (!p->n_pdds) {
2648 		pr_err("No pdd for given process\n");
2649 		ret = -ENODEV;
2650 		goto err_unlock;
2651 	}
2652 
2653 	ret = kfd_process_evict_queues(p, KFD_QUEUE_EVICTION_CRIU_CHECKPOINT);
2654 	if (ret)
2655 		goto err_unlock;
2656 
2657 	p->queues_paused = true;
2658 
2659 	args->pid = task_pid_nr_ns(p->lead_thread,
2660 					task_active_pid_ns(p->lead_thread));
2661 
2662 	ret = criu_get_process_object_info(p, &args->num_devices, &args->num_bos,
2663 					   &args->num_objects, &args->priv_data_size);
2664 	if (ret)
2665 		goto err_unlock;
2666 
2667 	dev_dbg(kfd_device, "Num of devices:%u bos:%u objects:%u priv_data_size:%lld\n",
2668 				args->num_devices, args->num_bos, args->num_objects,
2669 				args->priv_data_size);
2670 
2671 err_unlock:
2672 	if (ret) {
2673 		kfd_process_restore_queues(p);
2674 		p->queues_paused = false;
2675 	}
2676 	mutex_unlock(&p->mutex);
2677 	return ret;
2678 }
2679 
2680 static int kfd_ioctl_criu(struct file *filep, struct kfd_process *p, void *data)
2681 {
2682 	struct kfd_ioctl_criu_args *args = data;
2683 	int ret;
2684 
2685 	dev_dbg(kfd_device, "CRIU operation: %d\n", args->op);
2686 	switch (args->op) {
2687 	case KFD_CRIU_OP_PROCESS_INFO:
2688 		ret = criu_process_info(filep, p, args);
2689 		break;
2690 	case KFD_CRIU_OP_CHECKPOINT:
2691 		ret = criu_checkpoint(filep, p, args);
2692 		break;
2693 	case KFD_CRIU_OP_UNPAUSE:
2694 		ret = criu_unpause(filep, p, args);
2695 		break;
2696 	case KFD_CRIU_OP_RESTORE:
2697 		ret = criu_restore(filep, p, args);
2698 		break;
2699 	case KFD_CRIU_OP_RESUME:
2700 		ret = criu_resume(filep, p, args);
2701 		break;
2702 	default:
2703 		dev_dbg(kfd_device, "Unsupported CRIU operation:%d\n", args->op);
2704 		ret = -EINVAL;
2705 		break;
2706 	}
2707 
2708 	if (ret)
2709 		dev_dbg(kfd_device, "CRIU operation:%d err:%d\n", args->op, ret);
2710 
2711 	return ret;
2712 }
2713 
2714 #define AMDKFD_IOCTL_DEF(ioctl, _func, _flags) \
2715 	[_IOC_NR(ioctl)] = {.cmd = ioctl, .func = _func, .flags = _flags, \
2716 			    .cmd_drv = 0, .name = #ioctl}
2717 
2718 /** Ioctl table */
2719 static const struct amdkfd_ioctl_desc amdkfd_ioctls[] = {
2720 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_GET_VERSION,
2721 			kfd_ioctl_get_version, 0),
2722 
2723 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_CREATE_QUEUE,
2724 			kfd_ioctl_create_queue, 0),
2725 
2726 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_DESTROY_QUEUE,
2727 			kfd_ioctl_destroy_queue, 0),
2728 
2729 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_SET_MEMORY_POLICY,
2730 			kfd_ioctl_set_memory_policy, 0),
2731 
2732 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_GET_CLOCK_COUNTERS,
2733 			kfd_ioctl_get_clock_counters, 0),
2734 
2735 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_GET_PROCESS_APERTURES,
2736 			kfd_ioctl_get_process_apertures, 0),
2737 
2738 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_UPDATE_QUEUE,
2739 			kfd_ioctl_update_queue, 0),
2740 
2741 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_CREATE_EVENT,
2742 			kfd_ioctl_create_event, 0),
2743 
2744 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_DESTROY_EVENT,
2745 			kfd_ioctl_destroy_event, 0),
2746 
2747 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_SET_EVENT,
2748 			kfd_ioctl_set_event, 0),
2749 
2750 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_RESET_EVENT,
2751 			kfd_ioctl_reset_event, 0),
2752 
2753 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_WAIT_EVENTS,
2754 			kfd_ioctl_wait_events, 0),
2755 
2756 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_DBG_REGISTER_DEPRECATED,
2757 			kfd_ioctl_dbg_register, 0),
2758 
2759 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_DBG_UNREGISTER_DEPRECATED,
2760 			kfd_ioctl_dbg_unregister, 0),
2761 
2762 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_DBG_ADDRESS_WATCH_DEPRECATED,
2763 			kfd_ioctl_dbg_address_watch, 0),
2764 
2765 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_DBG_WAVE_CONTROL_DEPRECATED,
2766 			kfd_ioctl_dbg_wave_control, 0),
2767 
2768 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_SET_SCRATCH_BACKING_VA,
2769 			kfd_ioctl_set_scratch_backing_va, 0),
2770 
2771 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_GET_TILE_CONFIG,
2772 			kfd_ioctl_get_tile_config, 0),
2773 
2774 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_SET_TRAP_HANDLER,
2775 			kfd_ioctl_set_trap_handler, 0),
2776 
2777 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_GET_PROCESS_APERTURES_NEW,
2778 			kfd_ioctl_get_process_apertures_new, 0),
2779 
2780 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_ACQUIRE_VM,
2781 			kfd_ioctl_acquire_vm, 0),
2782 
2783 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_ALLOC_MEMORY_OF_GPU,
2784 			kfd_ioctl_alloc_memory_of_gpu, 0),
2785 
2786 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_FREE_MEMORY_OF_GPU,
2787 			kfd_ioctl_free_memory_of_gpu, 0),
2788 
2789 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_MAP_MEMORY_TO_GPU,
2790 			kfd_ioctl_map_memory_to_gpu, 0),
2791 
2792 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_UNMAP_MEMORY_FROM_GPU,
2793 			kfd_ioctl_unmap_memory_from_gpu, 0),
2794 
2795 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_SET_CU_MASK,
2796 			kfd_ioctl_set_cu_mask, 0),
2797 
2798 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_GET_QUEUE_WAVE_STATE,
2799 			kfd_ioctl_get_queue_wave_state, 0),
2800 
2801 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_GET_DMABUF_INFO,
2802 				kfd_ioctl_get_dmabuf_info, 0),
2803 
2804 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_IMPORT_DMABUF,
2805 				kfd_ioctl_import_dmabuf, 0),
2806 
2807 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_ALLOC_QUEUE_GWS,
2808 			kfd_ioctl_alloc_queue_gws, 0),
2809 
2810 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_SMI_EVENTS,
2811 			kfd_ioctl_smi_events, 0),
2812 
2813 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_SVM, kfd_ioctl_svm, 0),
2814 
2815 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_SET_XNACK_MODE,
2816 			kfd_ioctl_set_xnack_mode, 0),
2817 
2818 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_CRIU_OP,
2819 			kfd_ioctl_criu, KFD_IOC_FLAG_CHECKPOINT_RESTORE),
2820 
2821 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_AVAILABLE_MEMORY,
2822 			kfd_ioctl_get_available_memory, 0),
2823 
2824 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_EXPORT_DMABUF,
2825 				kfd_ioctl_export_dmabuf, 0),
2826 };
2827 
2828 #define AMDKFD_CORE_IOCTL_COUNT	ARRAY_SIZE(amdkfd_ioctls)
2829 
2830 static long kfd_ioctl(struct file *filep, unsigned int cmd, unsigned long arg)
2831 {
2832 	struct kfd_process *process;
2833 	amdkfd_ioctl_t *func;
2834 	const struct amdkfd_ioctl_desc *ioctl = NULL;
2835 	unsigned int nr = _IOC_NR(cmd);
2836 	char stack_kdata[128];
2837 	char *kdata = NULL;
2838 	unsigned int usize, asize;
2839 	int retcode = -EINVAL;
2840 	bool ptrace_attached = false;
2841 
2842 	if (nr >= AMDKFD_CORE_IOCTL_COUNT)
2843 		goto err_i1;
2844 
2845 	if ((nr >= AMDKFD_COMMAND_START) && (nr < AMDKFD_COMMAND_END)) {
2846 		u32 amdkfd_size;
2847 
2848 		ioctl = &amdkfd_ioctls[nr];
2849 
2850 		amdkfd_size = _IOC_SIZE(ioctl->cmd);
2851 		usize = asize = _IOC_SIZE(cmd);
2852 		if (amdkfd_size > asize)
2853 			asize = amdkfd_size;
2854 
2855 		cmd = ioctl->cmd;
2856 	} else
2857 		goto err_i1;
2858 
2859 	dev_dbg(kfd_device, "ioctl cmd 0x%x (#0x%x), arg 0x%lx\n", cmd, nr, arg);
2860 
2861 	/* Get the process struct from the filep. Only the process
2862 	 * that opened /dev/kfd can use the file descriptor. Child
2863 	 * processes need to create their own KFD device context.
2864 	 */
2865 	process = filep->private_data;
2866 
2867 	rcu_read_lock();
2868 	if ((ioctl->flags & KFD_IOC_FLAG_CHECKPOINT_RESTORE) &&
2869 	    ptrace_parent(process->lead_thread) == current)
2870 		ptrace_attached = true;
2871 	rcu_read_unlock();
2872 
2873 	if (process->lead_thread != current->group_leader
2874 	    && !ptrace_attached) {
2875 		dev_dbg(kfd_device, "Using KFD FD in wrong process\n");
2876 		retcode = -EBADF;
2877 		goto err_i1;
2878 	}
2879 
2880 	/* Do not trust userspace, use our own definition */
2881 	func = ioctl->func;
2882 
2883 	if (unlikely(!func)) {
2884 		dev_dbg(kfd_device, "no function\n");
2885 		retcode = -EINVAL;
2886 		goto err_i1;
2887 	}
2888 
2889 	/*
2890 	 * Versions of docker shipped in Ubuntu 18.xx and 20.xx do not support
2891 	 * CAP_CHECKPOINT_RESTORE, so we also allow access if CAP_SYS_ADMIN as CAP_SYS_ADMIN is a
2892 	 * more priviledged access.
2893 	 */
2894 	if (unlikely(ioctl->flags & KFD_IOC_FLAG_CHECKPOINT_RESTORE)) {
2895 		if (!capable(CAP_CHECKPOINT_RESTORE) &&
2896 						!capable(CAP_SYS_ADMIN)) {
2897 			retcode = -EACCES;
2898 			goto err_i1;
2899 		}
2900 	}
2901 
2902 	if (cmd & (IOC_IN | IOC_OUT)) {
2903 		if (asize <= sizeof(stack_kdata)) {
2904 			kdata = stack_kdata;
2905 		} else {
2906 			kdata = kmalloc(asize, GFP_KERNEL);
2907 			if (!kdata) {
2908 				retcode = -ENOMEM;
2909 				goto err_i1;
2910 			}
2911 		}
2912 		if (asize > usize)
2913 			memset(kdata + usize, 0, asize - usize);
2914 	}
2915 
2916 	if (cmd & IOC_IN) {
2917 		if (copy_from_user(kdata, (void __user *)arg, usize) != 0) {
2918 			retcode = -EFAULT;
2919 			goto err_i1;
2920 		}
2921 	} else if (cmd & IOC_OUT) {
2922 		memset(kdata, 0, usize);
2923 	}
2924 
2925 	retcode = func(filep, process, kdata);
2926 
2927 	if (cmd & IOC_OUT)
2928 		if (copy_to_user((void __user *)arg, kdata, usize) != 0)
2929 			retcode = -EFAULT;
2930 
2931 err_i1:
2932 	if (!ioctl)
2933 		dev_dbg(kfd_device, "invalid ioctl: pid=%d, cmd=0x%02x, nr=0x%02x\n",
2934 			  task_pid_nr(current), cmd, nr);
2935 
2936 	if (kdata != stack_kdata)
2937 		kfree(kdata);
2938 
2939 	if (retcode)
2940 		dev_dbg(kfd_device, "ioctl cmd (#0x%x), arg 0x%lx, ret = %d\n",
2941 				nr, arg, retcode);
2942 
2943 	return retcode;
2944 }
2945 
2946 static int kfd_mmio_mmap(struct kfd_dev *dev, struct kfd_process *process,
2947 		      struct vm_area_struct *vma)
2948 {
2949 	phys_addr_t address;
2950 
2951 	if (vma->vm_end - vma->vm_start != PAGE_SIZE)
2952 		return -EINVAL;
2953 
2954 	address = dev->adev->rmmio_remap.bus_addr;
2955 
2956 	vma->vm_flags |= VM_IO | VM_DONTCOPY | VM_DONTEXPAND | VM_NORESERVE |
2957 				VM_DONTDUMP | VM_PFNMAP;
2958 
2959 	vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot);
2960 
2961 	pr_debug("pasid 0x%x mapping mmio page\n"
2962 		 "     target user address == 0x%08llX\n"
2963 		 "     physical address    == 0x%08llX\n"
2964 		 "     vm_flags            == 0x%04lX\n"
2965 		 "     size                == 0x%04lX\n",
2966 		 process->pasid, (unsigned long long) vma->vm_start,
2967 		 address, vma->vm_flags, PAGE_SIZE);
2968 
2969 	return io_remap_pfn_range(vma,
2970 				vma->vm_start,
2971 				address >> PAGE_SHIFT,
2972 				PAGE_SIZE,
2973 				vma->vm_page_prot);
2974 }
2975 
2976 
2977 static int kfd_mmap(struct file *filp, struct vm_area_struct *vma)
2978 {
2979 	struct kfd_process *process;
2980 	struct kfd_dev *dev = NULL;
2981 	unsigned long mmap_offset;
2982 	unsigned int gpu_id;
2983 
2984 	process = kfd_get_process(current);
2985 	if (IS_ERR(process))
2986 		return PTR_ERR(process);
2987 
2988 	mmap_offset = vma->vm_pgoff << PAGE_SHIFT;
2989 	gpu_id = KFD_MMAP_GET_GPU_ID(mmap_offset);
2990 	if (gpu_id)
2991 		dev = kfd_device_by_id(gpu_id);
2992 
2993 	switch (mmap_offset & KFD_MMAP_TYPE_MASK) {
2994 	case KFD_MMAP_TYPE_DOORBELL:
2995 		if (!dev)
2996 			return -ENODEV;
2997 		return kfd_doorbell_mmap(dev, process, vma);
2998 
2999 	case KFD_MMAP_TYPE_EVENTS:
3000 		return kfd_event_mmap(process, vma);
3001 
3002 	case KFD_MMAP_TYPE_RESERVED_MEM:
3003 		if (!dev)
3004 			return -ENODEV;
3005 		return kfd_reserved_mem_mmap(dev, process, vma);
3006 	case KFD_MMAP_TYPE_MMIO:
3007 		if (!dev)
3008 			return -ENODEV;
3009 		return kfd_mmio_mmap(dev, process, vma);
3010 	}
3011 
3012 	return -EFAULT;
3013 }
3014