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