xref: /linux/drivers/gpu/drm/amd/amdkfd/kfd_chardev.c (revision fe7fad476ec8153a8b8767a08114e3e4a58a837e)
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/processor.h>
40 #include "kfd_priv.h"
41 #include "kfd_device_queue_manager.h"
42 #include "kfd_svm.h"
43 #include "amdgpu_amdkfd.h"
44 #include "kfd_smi_events.h"
45 #include "amdgpu_dma_buf.h"
46 #include "kfd_debug.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 struct device *kfd_device;
66 static const struct class kfd_class = {
67 	.name = kfd_dev_name,
68 };
69 
70 static inline struct kfd_process_device *kfd_lock_pdd_by_id(struct kfd_process *p, __u32 gpu_id)
71 {
72 	struct kfd_process_device *pdd;
73 
74 	mutex_lock(&p->mutex);
75 	pdd = kfd_process_device_data_by_id(p, gpu_id);
76 
77 	if (pdd)
78 		return pdd;
79 
80 	mutex_unlock(&p->mutex);
81 	return NULL;
82 }
83 
84 static inline void kfd_unlock_pdd(struct kfd_process_device *pdd)
85 {
86 	mutex_unlock(&pdd->process->mutex);
87 }
88 
89 int kfd_chardev_init(void)
90 {
91 	int err = 0;
92 
93 	kfd_char_dev_major = register_chrdev(0, kfd_dev_name, &kfd_fops);
94 	err = kfd_char_dev_major;
95 	if (err < 0)
96 		goto err_register_chrdev;
97 
98 	err = class_register(&kfd_class);
99 	if (err)
100 		goto err_class_create;
101 
102 	kfd_device = device_create(&kfd_class, NULL,
103 				   MKDEV(kfd_char_dev_major, 0),
104 				   NULL, kfd_dev_name);
105 	err = PTR_ERR(kfd_device);
106 	if (IS_ERR(kfd_device))
107 		goto err_device_create;
108 
109 	return 0;
110 
111 err_device_create:
112 	class_unregister(&kfd_class);
113 err_class_create:
114 	unregister_chrdev(kfd_char_dev_major, kfd_dev_name);
115 err_register_chrdev:
116 	return err;
117 }
118 
119 void kfd_chardev_exit(void)
120 {
121 	device_destroy(&kfd_class, MKDEV(kfd_char_dev_major, 0));
122 	class_unregister(&kfd_class);
123 	unregister_chrdev(kfd_char_dev_major, kfd_dev_name);
124 	kfd_device = NULL;
125 }
126 
127 
128 static int kfd_open(struct inode *inode, struct file *filep)
129 {
130 	struct kfd_process *process;
131 	bool is_32bit_user_mode;
132 
133 	if (iminor(inode) != 0)
134 		return -ENODEV;
135 
136 	is_32bit_user_mode = in_compat_syscall();
137 
138 	if (is_32bit_user_mode) {
139 		dev_warn(kfd_device,
140 			"Process %d (32-bit) failed to open /dev/kfd\n"
141 			"32-bit processes are not supported by amdkfd\n",
142 			current->pid);
143 		return -EPERM;
144 	}
145 
146 	process = kfd_create_process(current);
147 	if (IS_ERR(process))
148 		return PTR_ERR(process);
149 
150 	if (kfd_process_init_cwsr_apu(process, filep)) {
151 		kfd_unref_process(process);
152 		return -EFAULT;
153 	}
154 
155 	/* filep now owns the reference returned by kfd_create_process */
156 	filep->private_data = process;
157 
158 	dev_dbg(kfd_device, "process pid %d opened kfd node, compat mode (32 bit) - %d\n",
159 		process->lead_thread->pid, process->is_32bit_user_mode);
160 
161 	return 0;
162 }
163 
164 static int kfd_release(struct inode *inode, struct file *filep)
165 {
166 	struct kfd_process *process = filep->private_data;
167 
168 	if (process)
169 		kfd_unref_process(process);
170 
171 	return 0;
172 }
173 
174 static int kfd_ioctl_get_version(struct file *filep, struct kfd_process *p,
175 					void *data)
176 {
177 	struct kfd_ioctl_get_version_args *args = data;
178 
179 	args->major_version = KFD_IOCTL_MAJOR_VERSION;
180 	args->minor_version = KFD_IOCTL_MINOR_VERSION;
181 
182 	return 0;
183 }
184 
185 static int set_queue_properties_from_user(struct queue_properties *q_properties,
186 				struct kfd_ioctl_create_queue_args *args)
187 {
188 	/*
189 	 * Repurpose queue percentage to accommodate new features:
190 	 * bit 0-7: queue percentage
191 	 * bit 8-15: pm4_target_xcc
192 	 */
193 	if ((args->queue_percentage & 0xFF) > KFD_MAX_QUEUE_PERCENTAGE) {
194 		pr_err("Queue percentage must be between 0 to KFD_MAX_QUEUE_PERCENTAGE\n");
195 		return -EINVAL;
196 	}
197 
198 	if (args->queue_priority > KFD_MAX_QUEUE_PRIORITY) {
199 		pr_err("Queue priority must be between 0 to KFD_MAX_QUEUE_PRIORITY\n");
200 		return -EINVAL;
201 	}
202 
203 	if ((args->ring_base_address) &&
204 		(!access_ok((const void __user *) args->ring_base_address,
205 			sizeof(uint64_t)))) {
206 		pr_err("Can't access ring base address\n");
207 		return -EFAULT;
208 	}
209 
210 	if (!is_power_of_2(args->ring_size) && (args->ring_size != 0)) {
211 		pr_err("Ring size must be a power of 2 or 0\n");
212 		return -EINVAL;
213 	}
214 
215 	if (!access_ok((const void __user *) args->read_pointer_address,
216 			sizeof(uint32_t))) {
217 		pr_err("Can't access read pointer\n");
218 		return -EFAULT;
219 	}
220 
221 	if (!access_ok((const void __user *) args->write_pointer_address,
222 			sizeof(uint32_t))) {
223 		pr_err("Can't access write pointer\n");
224 		return -EFAULT;
225 	}
226 
227 	if (args->eop_buffer_address &&
228 		!access_ok((const void __user *) args->eop_buffer_address,
229 			sizeof(uint32_t))) {
230 		pr_debug("Can't access eop buffer");
231 		return -EFAULT;
232 	}
233 
234 	if (args->ctx_save_restore_address &&
235 		!access_ok((const void __user *) args->ctx_save_restore_address,
236 			sizeof(uint32_t))) {
237 		pr_debug("Can't access ctx save restore buffer");
238 		return -EFAULT;
239 	}
240 
241 	q_properties->is_interop = false;
242 	q_properties->is_gws = false;
243 	q_properties->queue_percent = args->queue_percentage & 0xFF;
244 	/* bit 8-15 are repurposed to be PM4 target XCC */
245 	q_properties->pm4_target_xcc = (args->queue_percentage >> 8) & 0xFF;
246 	q_properties->priority = args->queue_priority;
247 	q_properties->queue_address = args->ring_base_address;
248 	q_properties->queue_size = args->ring_size;
249 	q_properties->read_ptr = (void __user *)args->read_pointer_address;
250 	q_properties->write_ptr = (void __user *)args->write_pointer_address;
251 	q_properties->eop_ring_buffer_address = args->eop_buffer_address;
252 	q_properties->eop_ring_buffer_size = args->eop_buffer_size;
253 	q_properties->ctx_save_restore_area_address =
254 			args->ctx_save_restore_address;
255 	q_properties->ctx_save_restore_area_size = args->ctx_save_restore_size;
256 	q_properties->ctl_stack_size = args->ctl_stack_size;
257 	q_properties->sdma_engine_id = args->sdma_engine_id;
258 	if (args->queue_type == KFD_IOC_QUEUE_TYPE_COMPUTE ||
259 		args->queue_type == KFD_IOC_QUEUE_TYPE_COMPUTE_AQL)
260 		q_properties->type = KFD_QUEUE_TYPE_COMPUTE;
261 	else if (args->queue_type == KFD_IOC_QUEUE_TYPE_SDMA)
262 		q_properties->type = KFD_QUEUE_TYPE_SDMA;
263 	else if (args->queue_type == KFD_IOC_QUEUE_TYPE_SDMA_XGMI)
264 		q_properties->type = KFD_QUEUE_TYPE_SDMA_XGMI;
265 	else if (args->queue_type == KFD_IOC_QUEUE_TYPE_SDMA_BY_ENG_ID)
266 		q_properties->type = KFD_QUEUE_TYPE_SDMA_BY_ENG_ID;
267 	else
268 		return -ENOTSUPP;
269 
270 	if (args->queue_type == KFD_IOC_QUEUE_TYPE_COMPUTE_AQL)
271 		q_properties->format = KFD_QUEUE_FORMAT_AQL;
272 	else
273 		q_properties->format = KFD_QUEUE_FORMAT_PM4;
274 
275 	pr_debug("Queue Percentage: %d, %d\n",
276 			q_properties->queue_percent, args->queue_percentage);
277 
278 	pr_debug("Queue Priority: %d, %d\n",
279 			q_properties->priority, args->queue_priority);
280 
281 	pr_debug("Queue Address: 0x%llX, 0x%llX\n",
282 			q_properties->queue_address, args->ring_base_address);
283 
284 	pr_debug("Queue Size: 0x%llX, %u\n",
285 			q_properties->queue_size, args->ring_size);
286 
287 	pr_debug("Queue r/w Pointers: %px, %px\n",
288 			q_properties->read_ptr,
289 			q_properties->write_ptr);
290 
291 	pr_debug("Queue Format: %d\n", q_properties->format);
292 
293 	pr_debug("Queue EOP: 0x%llX\n", q_properties->eop_ring_buffer_address);
294 
295 	pr_debug("Queue CTX save area: 0x%llX\n",
296 			q_properties->ctx_save_restore_area_address);
297 
298 	return 0;
299 }
300 
301 static int kfd_ioctl_create_queue(struct file *filep, struct kfd_process *p,
302 					void *data)
303 {
304 	struct kfd_ioctl_create_queue_args *args = data;
305 	struct kfd_node *dev;
306 	int err = 0;
307 	unsigned int queue_id;
308 	struct kfd_process_device *pdd;
309 	struct queue_properties q_properties;
310 	uint32_t doorbell_offset_in_process = 0;
311 
312 	memset(&q_properties, 0, sizeof(struct queue_properties));
313 
314 	pr_debug("Creating queue ioctl\n");
315 
316 	err = set_queue_properties_from_user(&q_properties, args);
317 	if (err)
318 		return err;
319 
320 	pr_debug("Looking for gpu id 0x%x\n", args->gpu_id);
321 
322 	mutex_lock(&p->mutex);
323 
324 	pdd = kfd_process_device_data_by_id(p, args->gpu_id);
325 	if (!pdd) {
326 		pr_debug("Could not find gpu id 0x%x\n", args->gpu_id);
327 		err = -EINVAL;
328 		goto err_pdd;
329 	}
330 	dev = pdd->dev;
331 
332 	pdd = kfd_bind_process_to_device(dev, p);
333 	if (IS_ERR(pdd)) {
334 		err = -ESRCH;
335 		goto err_bind_process;
336 	}
337 
338 	if (q_properties.type == KFD_QUEUE_TYPE_SDMA_BY_ENG_ID) {
339 		int max_sdma_eng_id = kfd_get_num_sdma_engines(dev) +
340 				      kfd_get_num_xgmi_sdma_engines(dev) - 1;
341 
342 		if (q_properties.sdma_engine_id > max_sdma_eng_id) {
343 			err = -EINVAL;
344 			pr_err("sdma_engine_id %i exceeds maximum id of %i\n",
345 			       q_properties.sdma_engine_id, max_sdma_eng_id);
346 			goto err_sdma_engine_id;
347 		}
348 	}
349 
350 	if (!pdd->qpd.proc_doorbells) {
351 		err = kfd_alloc_process_doorbells(dev->kfd, pdd);
352 		if (err) {
353 			pr_debug("failed to allocate process doorbells\n");
354 			goto err_bind_process;
355 		}
356 	}
357 
358 	err = kfd_queue_acquire_buffers(pdd, &q_properties);
359 	if (err) {
360 		pr_debug("failed to acquire user queue buffers\n");
361 		goto err_acquire_queue_buf;
362 	}
363 
364 	pr_debug("Creating queue for process pid %d on gpu 0x%x\n",
365 			p->lead_thread->pid,
366 			dev->id);
367 
368 	err = pqm_create_queue(&p->pqm, dev, &q_properties, &queue_id,
369 			NULL, NULL, NULL, &doorbell_offset_in_process);
370 	if (err != 0)
371 		goto err_create_queue;
372 
373 	args->queue_id = queue_id;
374 
375 
376 	/* Return gpu_id as doorbell offset for mmap usage */
377 	args->doorbell_offset = KFD_MMAP_TYPE_DOORBELL;
378 	args->doorbell_offset |= KFD_MMAP_GPU_ID(args->gpu_id);
379 	if (KFD_IS_SOC15(dev))
380 		/* On SOC15 ASICs, include the doorbell offset within the
381 		 * process doorbell frame, which is 2 pages.
382 		 */
383 		args->doorbell_offset |= doorbell_offset_in_process;
384 
385 	mutex_unlock(&p->mutex);
386 
387 	pr_debug("Queue id %d was created successfully\n", args->queue_id);
388 
389 	pr_debug("Ring buffer address == 0x%016llX\n",
390 			args->ring_base_address);
391 
392 	pr_debug("Read ptr address    == 0x%016llX\n",
393 			args->read_pointer_address);
394 
395 	pr_debug("Write ptr address   == 0x%016llX\n",
396 			args->write_pointer_address);
397 
398 	kfd_dbg_ev_raise(KFD_EC_MASK(EC_QUEUE_NEW), p, dev, queue_id, false, NULL, 0);
399 	return 0;
400 
401 err_create_queue:
402 	kfd_queue_unref_bo_vas(pdd, &q_properties);
403 	kfd_queue_release_buffers(pdd, &q_properties);
404 err_acquire_queue_buf:
405 err_sdma_engine_id:
406 err_bind_process:
407 err_pdd:
408 	mutex_unlock(&p->mutex);
409 	return err;
410 }
411 
412 static int kfd_ioctl_destroy_queue(struct file *filp, struct kfd_process *p,
413 					void *data)
414 {
415 	int retval;
416 	struct kfd_ioctl_destroy_queue_args *args = data;
417 
418 	pr_debug("Destroying queue id %d for process pid %d\n",
419 				args->queue_id,
420 				p->lead_thread->pid);
421 
422 	mutex_lock(&p->mutex);
423 
424 	retval = pqm_destroy_queue(&p->pqm, args->queue_id);
425 
426 	mutex_unlock(&p->mutex);
427 	return retval;
428 }
429 
430 static int kfd_ioctl_update_queue(struct file *filp, struct kfd_process *p,
431 					void *data)
432 {
433 	int retval;
434 	struct kfd_ioctl_update_queue_args *args = data;
435 	struct queue_properties properties;
436 
437 	/*
438 	 * Repurpose queue percentage to accommodate new features:
439 	 * bit 0-7: queue percentage
440 	 * bit 8-15: pm4_target_xcc
441 	 */
442 	if ((args->queue_percentage & 0xFF) > KFD_MAX_QUEUE_PERCENTAGE) {
443 		pr_err("Queue percentage must be between 0 to KFD_MAX_QUEUE_PERCENTAGE\n");
444 		return -EINVAL;
445 	}
446 
447 	if (args->queue_priority > KFD_MAX_QUEUE_PRIORITY) {
448 		pr_err("Queue priority must be between 0 to KFD_MAX_QUEUE_PRIORITY\n");
449 		return -EINVAL;
450 	}
451 
452 	if ((args->ring_base_address) &&
453 		(!access_ok((const void __user *) args->ring_base_address,
454 			sizeof(uint64_t)))) {
455 		pr_err("Can't access ring base address\n");
456 		return -EFAULT;
457 	}
458 
459 	if (!is_power_of_2(args->ring_size) && (args->ring_size != 0)) {
460 		pr_err("Ring size must be a power of 2 or 0\n");
461 		return -EINVAL;
462 	}
463 
464 	properties.queue_address = args->ring_base_address;
465 	properties.queue_size = args->ring_size;
466 	properties.queue_percent = args->queue_percentage & 0xFF;
467 	/* bit 8-15 are repurposed to be PM4 target XCC */
468 	properties.pm4_target_xcc = (args->queue_percentage >> 8) & 0xFF;
469 	properties.priority = args->queue_priority;
470 
471 	pr_debug("Updating queue id %d for process pid %d\n",
472 			args->queue_id, p->lead_thread->pid);
473 
474 	mutex_lock(&p->mutex);
475 
476 	retval = pqm_update_queue_properties(&p->pqm, args->queue_id, &properties);
477 
478 	mutex_unlock(&p->mutex);
479 
480 	return retval;
481 }
482 
483 static int kfd_ioctl_set_cu_mask(struct file *filp, struct kfd_process *p,
484 					void *data)
485 {
486 	int retval;
487 	const int max_num_cus = 1024;
488 	struct kfd_ioctl_set_cu_mask_args *args = data;
489 	struct mqd_update_info minfo = {0};
490 	uint32_t __user *cu_mask_ptr = (uint32_t __user *)args->cu_mask_ptr;
491 	size_t cu_mask_size = sizeof(uint32_t) * (args->num_cu_mask / 32);
492 
493 	if ((args->num_cu_mask % 32) != 0) {
494 		pr_debug("num_cu_mask 0x%x must be a multiple of 32",
495 				args->num_cu_mask);
496 		return -EINVAL;
497 	}
498 
499 	minfo.cu_mask.count = args->num_cu_mask;
500 	if (minfo.cu_mask.count == 0) {
501 		pr_debug("CU mask cannot be 0");
502 		return -EINVAL;
503 	}
504 
505 	/* To prevent an unreasonably large CU mask size, set an arbitrary
506 	 * limit of max_num_cus bits.  We can then just drop any CU mask bits
507 	 * past max_num_cus bits and just use the first max_num_cus bits.
508 	 */
509 	if (minfo.cu_mask.count > max_num_cus) {
510 		pr_debug("CU mask cannot be greater than 1024 bits");
511 		minfo.cu_mask.count = max_num_cus;
512 		cu_mask_size = sizeof(uint32_t) * (max_num_cus/32);
513 	}
514 
515 	minfo.cu_mask.ptr = kzalloc(cu_mask_size, GFP_KERNEL);
516 	if (!minfo.cu_mask.ptr)
517 		return -ENOMEM;
518 
519 	retval = copy_from_user(minfo.cu_mask.ptr, cu_mask_ptr, cu_mask_size);
520 	if (retval) {
521 		pr_debug("Could not copy CU mask from userspace");
522 		retval = -EFAULT;
523 		goto out;
524 	}
525 
526 	mutex_lock(&p->mutex);
527 
528 	retval = pqm_update_mqd(&p->pqm, args->queue_id, &minfo);
529 
530 	mutex_unlock(&p->mutex);
531 
532 out:
533 	kfree(minfo.cu_mask.ptr);
534 	return retval;
535 }
536 
537 static int kfd_ioctl_get_queue_wave_state(struct file *filep,
538 					  struct kfd_process *p, void *data)
539 {
540 	struct kfd_ioctl_get_queue_wave_state_args *args = data;
541 	int r;
542 
543 	mutex_lock(&p->mutex);
544 
545 	r = pqm_get_wave_state(&p->pqm, args->queue_id,
546 			       (void __user *)args->ctl_stack_address,
547 			       &args->ctl_stack_used_size,
548 			       &args->save_area_used_size);
549 
550 	mutex_unlock(&p->mutex);
551 
552 	return r;
553 }
554 
555 static int kfd_ioctl_set_memory_policy(struct file *filep,
556 					struct kfd_process *p, void *data)
557 {
558 	struct kfd_ioctl_set_memory_policy_args *args = data;
559 	int err = 0;
560 	struct kfd_process_device *pdd;
561 	enum cache_policy default_policy, alternate_policy;
562 
563 	if (args->default_policy != KFD_IOC_CACHE_POLICY_COHERENT
564 	    && args->default_policy != KFD_IOC_CACHE_POLICY_NONCOHERENT) {
565 		return -EINVAL;
566 	}
567 
568 	if (args->alternate_policy != KFD_IOC_CACHE_POLICY_COHERENT
569 	    && args->alternate_policy != KFD_IOC_CACHE_POLICY_NONCOHERENT) {
570 		return -EINVAL;
571 	}
572 
573 	mutex_lock(&p->mutex);
574 	pdd = kfd_process_device_data_by_id(p, args->gpu_id);
575 	if (!pdd) {
576 		pr_debug("Could not find gpu id 0x%x\n", args->gpu_id);
577 		err = -EINVAL;
578 		goto err_pdd;
579 	}
580 
581 	pdd = kfd_bind_process_to_device(pdd->dev, p);
582 	if (IS_ERR(pdd)) {
583 		err = -ESRCH;
584 		goto out;
585 	}
586 
587 	default_policy = (args->default_policy == KFD_IOC_CACHE_POLICY_COHERENT)
588 			 ? cache_policy_coherent : cache_policy_noncoherent;
589 
590 	alternate_policy =
591 		(args->alternate_policy == KFD_IOC_CACHE_POLICY_COHERENT)
592 		   ? cache_policy_coherent : cache_policy_noncoherent;
593 
594 	if (!pdd->dev->dqm->ops.set_cache_memory_policy(pdd->dev->dqm,
595 				&pdd->qpd,
596 				default_policy,
597 				alternate_policy,
598 				(void __user *)args->alternate_aperture_base,
599 				args->alternate_aperture_size))
600 		err = -EINVAL;
601 
602 out:
603 err_pdd:
604 	mutex_unlock(&p->mutex);
605 
606 	return err;
607 }
608 
609 static int kfd_ioctl_set_trap_handler(struct file *filep,
610 					struct kfd_process *p, void *data)
611 {
612 	struct kfd_ioctl_set_trap_handler_args *args = data;
613 	int err = 0;
614 	struct kfd_process_device *pdd;
615 
616 	mutex_lock(&p->mutex);
617 
618 	pdd = kfd_process_device_data_by_id(p, args->gpu_id);
619 	if (!pdd) {
620 		err = -EINVAL;
621 		goto err_pdd;
622 	}
623 
624 	pdd = kfd_bind_process_to_device(pdd->dev, p);
625 	if (IS_ERR(pdd)) {
626 		err = -ESRCH;
627 		goto out;
628 	}
629 
630 	kfd_process_set_trap_handler(&pdd->qpd, args->tba_addr, args->tma_addr);
631 
632 out:
633 err_pdd:
634 	mutex_unlock(&p->mutex);
635 
636 	return err;
637 }
638 
639 static int kfd_ioctl_dbg_register(struct file *filep,
640 				struct kfd_process *p, void *data)
641 {
642 	return -EPERM;
643 }
644 
645 static int kfd_ioctl_dbg_unregister(struct file *filep,
646 				struct kfd_process *p, void *data)
647 {
648 	return -EPERM;
649 }
650 
651 static int kfd_ioctl_dbg_address_watch(struct file *filep,
652 					struct kfd_process *p, void *data)
653 {
654 	return -EPERM;
655 }
656 
657 /* Parse and generate fixed size data structure for wave control */
658 static int kfd_ioctl_dbg_wave_control(struct file *filep,
659 					struct kfd_process *p, void *data)
660 {
661 	return -EPERM;
662 }
663 
664 static int kfd_ioctl_get_clock_counters(struct file *filep,
665 				struct kfd_process *p, void *data)
666 {
667 	struct kfd_ioctl_get_clock_counters_args *args = data;
668 	struct kfd_process_device *pdd;
669 
670 	mutex_lock(&p->mutex);
671 	pdd = kfd_process_device_data_by_id(p, args->gpu_id);
672 	mutex_unlock(&p->mutex);
673 	if (pdd)
674 		/* Reading GPU clock counter from KGD */
675 		args->gpu_clock_counter = amdgpu_amdkfd_get_gpu_clock_counter(pdd->dev->adev);
676 	else
677 		/* Node without GPU resource */
678 		args->gpu_clock_counter = 0;
679 
680 	/* No access to rdtsc. Using raw monotonic time */
681 	args->cpu_clock_counter = ktime_get_raw_ns();
682 	args->system_clock_counter = ktime_get_boottime_ns();
683 
684 	/* Since the counter is in nano-seconds we use 1GHz frequency */
685 	args->system_clock_freq = 1000000000;
686 
687 	return 0;
688 }
689 
690 
691 static int kfd_ioctl_get_process_apertures(struct file *filp,
692 				struct kfd_process *p, void *data)
693 {
694 	struct kfd_ioctl_get_process_apertures_args *args = data;
695 	struct kfd_process_device_apertures *pAperture;
696 	int i;
697 
698 	dev_dbg(kfd_device, "get apertures for process pid %d", p->lead_thread->pid);
699 
700 	args->num_of_nodes = 0;
701 
702 	mutex_lock(&p->mutex);
703 	/* Run over all pdd of the process */
704 	for (i = 0; i < p->n_pdds; i++) {
705 		struct kfd_process_device *pdd = p->pdds[i];
706 
707 		pAperture =
708 			&args->process_apertures[args->num_of_nodes];
709 		pAperture->gpu_id = pdd->dev->id;
710 		pAperture->lds_base = pdd->lds_base;
711 		pAperture->lds_limit = pdd->lds_limit;
712 		pAperture->gpuvm_base = pdd->gpuvm_base;
713 		pAperture->gpuvm_limit = pdd->gpuvm_limit;
714 		pAperture->scratch_base = pdd->scratch_base;
715 		pAperture->scratch_limit = pdd->scratch_limit;
716 
717 		dev_dbg(kfd_device,
718 			"node id %u\n", args->num_of_nodes);
719 		dev_dbg(kfd_device,
720 			"gpu id %u\n", pdd->dev->id);
721 		dev_dbg(kfd_device,
722 			"lds_base %llX\n", pdd->lds_base);
723 		dev_dbg(kfd_device,
724 			"lds_limit %llX\n", pdd->lds_limit);
725 		dev_dbg(kfd_device,
726 			"gpuvm_base %llX\n", pdd->gpuvm_base);
727 		dev_dbg(kfd_device,
728 			"gpuvm_limit %llX\n", pdd->gpuvm_limit);
729 		dev_dbg(kfd_device,
730 			"scratch_base %llX\n", pdd->scratch_base);
731 		dev_dbg(kfd_device,
732 			"scratch_limit %llX\n", pdd->scratch_limit);
733 
734 		if (++args->num_of_nodes >= NUM_OF_SUPPORTED_GPUS)
735 			break;
736 	}
737 	mutex_unlock(&p->mutex);
738 
739 	return 0;
740 }
741 
742 static int kfd_ioctl_get_process_apertures_new(struct file *filp,
743 				struct kfd_process *p, void *data)
744 {
745 	struct kfd_ioctl_get_process_apertures_new_args *args = data;
746 	struct kfd_process_device_apertures *pa;
747 	int ret;
748 	int i;
749 
750 	dev_dbg(kfd_device, "get apertures for process pid %d",
751 			p->lead_thread->pid);
752 
753 	if (args->num_of_nodes == 0) {
754 		/* Return number of nodes, so that user space can alloacate
755 		 * sufficient memory
756 		 */
757 		mutex_lock(&p->mutex);
758 		args->num_of_nodes = p->n_pdds;
759 		goto out_unlock;
760 	}
761 
762 	/* Fill in process-aperture information for all available
763 	 * nodes, but not more than args->num_of_nodes as that is
764 	 * the amount of memory allocated by user
765 	 */
766 	pa = kcalloc(args->num_of_nodes, sizeof(struct kfd_process_device_apertures),
767 		     GFP_KERNEL);
768 	if (!pa)
769 		return -ENOMEM;
770 
771 	mutex_lock(&p->mutex);
772 
773 	if (!p->n_pdds) {
774 		args->num_of_nodes = 0;
775 		kfree(pa);
776 		goto out_unlock;
777 	}
778 
779 	/* Run over all pdd of the process */
780 	for (i = 0; i < min(p->n_pdds, args->num_of_nodes); i++) {
781 		struct kfd_process_device *pdd = p->pdds[i];
782 
783 		pa[i].gpu_id = pdd->dev->id;
784 		pa[i].lds_base = pdd->lds_base;
785 		pa[i].lds_limit = pdd->lds_limit;
786 		pa[i].gpuvm_base = pdd->gpuvm_base;
787 		pa[i].gpuvm_limit = pdd->gpuvm_limit;
788 		pa[i].scratch_base = pdd->scratch_base;
789 		pa[i].scratch_limit = pdd->scratch_limit;
790 
791 		dev_dbg(kfd_device,
792 			"gpu id %u\n", pdd->dev->id);
793 		dev_dbg(kfd_device,
794 			"lds_base %llX\n", pdd->lds_base);
795 		dev_dbg(kfd_device,
796 			"lds_limit %llX\n", pdd->lds_limit);
797 		dev_dbg(kfd_device,
798 			"gpuvm_base %llX\n", pdd->gpuvm_base);
799 		dev_dbg(kfd_device,
800 			"gpuvm_limit %llX\n", pdd->gpuvm_limit);
801 		dev_dbg(kfd_device,
802 			"scratch_base %llX\n", pdd->scratch_base);
803 		dev_dbg(kfd_device,
804 			"scratch_limit %llX\n", pdd->scratch_limit);
805 	}
806 	mutex_unlock(&p->mutex);
807 
808 	args->num_of_nodes = i;
809 	ret = copy_to_user(
810 			(void __user *)args->kfd_process_device_apertures_ptr,
811 			pa,
812 			(i * sizeof(struct kfd_process_device_apertures)));
813 	kfree(pa);
814 	return ret ? -EFAULT : 0;
815 
816 out_unlock:
817 	mutex_unlock(&p->mutex);
818 	return 0;
819 }
820 
821 static int kfd_ioctl_create_event(struct file *filp, struct kfd_process *p,
822 					void *data)
823 {
824 	struct kfd_ioctl_create_event_args *args = data;
825 	int err;
826 
827 	/* For dGPUs the event page is allocated in user mode. The
828 	 * handle is passed to KFD with the first call to this IOCTL
829 	 * through the event_page_offset field.
830 	 */
831 	if (args->event_page_offset) {
832 		mutex_lock(&p->mutex);
833 		err = kfd_kmap_event_page(p, args->event_page_offset);
834 		mutex_unlock(&p->mutex);
835 		if (err)
836 			return err;
837 	}
838 
839 	err = kfd_event_create(filp, p, args->event_type,
840 				args->auto_reset != 0, args->node_id,
841 				&args->event_id, &args->event_trigger_data,
842 				&args->event_page_offset,
843 				&args->event_slot_index);
844 
845 	pr_debug("Created event (id:0x%08x) (%s)\n", args->event_id, __func__);
846 	return err;
847 }
848 
849 static int kfd_ioctl_destroy_event(struct file *filp, struct kfd_process *p,
850 					void *data)
851 {
852 	struct kfd_ioctl_destroy_event_args *args = data;
853 
854 	return kfd_event_destroy(p, args->event_id);
855 }
856 
857 static int kfd_ioctl_set_event(struct file *filp, struct kfd_process *p,
858 				void *data)
859 {
860 	struct kfd_ioctl_set_event_args *args = data;
861 
862 	return kfd_set_event(p, args->event_id);
863 }
864 
865 static int kfd_ioctl_reset_event(struct file *filp, struct kfd_process *p,
866 				void *data)
867 {
868 	struct kfd_ioctl_reset_event_args *args = data;
869 
870 	return kfd_reset_event(p, args->event_id);
871 }
872 
873 static int kfd_ioctl_wait_events(struct file *filp, struct kfd_process *p,
874 				void *data)
875 {
876 	struct kfd_ioctl_wait_events_args *args = data;
877 
878 	return kfd_wait_on_events(p, args->num_events,
879 			(void __user *)args->events_ptr,
880 			(args->wait_for_all != 0),
881 			&args->timeout, &args->wait_result);
882 }
883 static int kfd_ioctl_set_scratch_backing_va(struct file *filep,
884 					struct kfd_process *p, void *data)
885 {
886 	struct kfd_ioctl_set_scratch_backing_va_args *args = data;
887 	struct kfd_process_device *pdd;
888 	struct kfd_node *dev;
889 	long err;
890 
891 	mutex_lock(&p->mutex);
892 	pdd = kfd_process_device_data_by_id(p, args->gpu_id);
893 	if (!pdd) {
894 		err = -EINVAL;
895 		goto err_pdd;
896 	}
897 	dev = pdd->dev;
898 
899 	pdd = kfd_bind_process_to_device(dev, p);
900 	if (IS_ERR(pdd)) {
901 		err = PTR_ERR(pdd);
902 		goto bind_process_to_device_fail;
903 	}
904 
905 	pdd->qpd.sh_hidden_private_base = args->va_addr;
906 
907 	mutex_unlock(&p->mutex);
908 
909 	if (dev->dqm->sched_policy == KFD_SCHED_POLICY_NO_HWS &&
910 	    pdd->qpd.vmid != 0 && dev->kfd2kgd->set_scratch_backing_va)
911 		dev->kfd2kgd->set_scratch_backing_va(
912 			dev->adev, args->va_addr, pdd->qpd.vmid);
913 
914 	return 0;
915 
916 bind_process_to_device_fail:
917 err_pdd:
918 	mutex_unlock(&p->mutex);
919 	return err;
920 }
921 
922 static int kfd_ioctl_get_tile_config(struct file *filep,
923 		struct kfd_process *p, void *data)
924 {
925 	struct kfd_ioctl_get_tile_config_args *args = data;
926 	struct kfd_process_device *pdd;
927 	struct tile_config config;
928 	int err = 0;
929 
930 	mutex_lock(&p->mutex);
931 	pdd = kfd_process_device_data_by_id(p, args->gpu_id);
932 	mutex_unlock(&p->mutex);
933 	if (!pdd)
934 		return -EINVAL;
935 
936 	amdgpu_amdkfd_get_tile_config(pdd->dev->adev, &config);
937 
938 	args->gb_addr_config = config.gb_addr_config;
939 	args->num_banks = config.num_banks;
940 	args->num_ranks = config.num_ranks;
941 
942 	if (args->num_tile_configs > config.num_tile_configs)
943 		args->num_tile_configs = config.num_tile_configs;
944 	err = copy_to_user((void __user *)args->tile_config_ptr,
945 			config.tile_config_ptr,
946 			args->num_tile_configs * sizeof(uint32_t));
947 	if (err) {
948 		args->num_tile_configs = 0;
949 		return -EFAULT;
950 	}
951 
952 	if (args->num_macro_tile_configs > config.num_macro_tile_configs)
953 		args->num_macro_tile_configs =
954 				config.num_macro_tile_configs;
955 	err = copy_to_user((void __user *)args->macro_tile_config_ptr,
956 			config.macro_tile_config_ptr,
957 			args->num_macro_tile_configs * sizeof(uint32_t));
958 	if (err) {
959 		args->num_macro_tile_configs = 0;
960 		return -EFAULT;
961 	}
962 
963 	return 0;
964 }
965 
966 static int kfd_ioctl_acquire_vm(struct file *filep, struct kfd_process *p,
967 				void *data)
968 {
969 	struct kfd_ioctl_acquire_vm_args *args = data;
970 	struct kfd_process_device *pdd;
971 	struct file *drm_file;
972 	int ret;
973 
974 	drm_file = fget(args->drm_fd);
975 	if (!drm_file)
976 		return -EINVAL;
977 
978 	mutex_lock(&p->mutex);
979 	pdd = kfd_process_device_data_by_id(p, args->gpu_id);
980 	if (!pdd) {
981 		ret = -EINVAL;
982 		goto err_pdd;
983 	}
984 
985 	if (pdd->drm_file) {
986 		ret = pdd->drm_file == drm_file ? 0 : -EBUSY;
987 		goto err_drm_file;
988 	}
989 
990 	ret = kfd_process_device_init_vm(pdd, drm_file);
991 	if (ret)
992 		goto err_unlock;
993 
994 	/* On success, the PDD keeps the drm_file reference */
995 	mutex_unlock(&p->mutex);
996 
997 	return 0;
998 
999 err_unlock:
1000 err_pdd:
1001 err_drm_file:
1002 	mutex_unlock(&p->mutex);
1003 	fput(drm_file);
1004 	return ret;
1005 }
1006 
1007 bool kfd_dev_is_large_bar(struct kfd_node *dev)
1008 {
1009 	if (dev->kfd->adev->debug_largebar) {
1010 		pr_debug("Simulate large-bar allocation on non large-bar machine\n");
1011 		return true;
1012 	}
1013 
1014 	if (dev->local_mem_info.local_mem_size_private == 0 &&
1015 	    dev->local_mem_info.local_mem_size_public > 0)
1016 		return true;
1017 
1018 	if (dev->local_mem_info.local_mem_size_public == 0 &&
1019 	    dev->kfd->adev->gmc.is_app_apu) {
1020 		pr_debug("APP APU, Consider like a large bar system\n");
1021 		return true;
1022 	}
1023 
1024 	return false;
1025 }
1026 
1027 static int kfd_ioctl_get_available_memory(struct file *filep,
1028 					  struct kfd_process *p, void *data)
1029 {
1030 	struct kfd_ioctl_get_available_memory_args *args = data;
1031 	struct kfd_process_device *pdd = kfd_lock_pdd_by_id(p, args->gpu_id);
1032 
1033 	if (!pdd)
1034 		return -EINVAL;
1035 	args->available = amdgpu_amdkfd_get_available_memory(pdd->dev->adev,
1036 							pdd->dev->node_id);
1037 	kfd_unlock_pdd(pdd);
1038 	return 0;
1039 }
1040 
1041 static int kfd_ioctl_alloc_memory_of_gpu(struct file *filep,
1042 					struct kfd_process *p, void *data)
1043 {
1044 	struct kfd_ioctl_alloc_memory_of_gpu_args *args = data;
1045 	struct kfd_process_device *pdd;
1046 	void *mem;
1047 	struct kfd_node *dev;
1048 	int idr_handle;
1049 	long err;
1050 	uint64_t offset = args->mmap_offset;
1051 	uint32_t flags = args->flags;
1052 
1053 	if (args->size == 0)
1054 		return -EINVAL;
1055 
1056 #if IS_ENABLED(CONFIG_HSA_AMD_SVM)
1057 	/* Flush pending deferred work to avoid racing with deferred actions
1058 	 * from previous memory map changes (e.g. munmap).
1059 	 */
1060 	svm_range_list_lock_and_flush_work(&p->svms, current->mm);
1061 	mutex_lock(&p->svms.lock);
1062 	mmap_write_unlock(current->mm);
1063 	if (interval_tree_iter_first(&p->svms.objects,
1064 				     args->va_addr >> PAGE_SHIFT,
1065 				     (args->va_addr + args->size - 1) >> PAGE_SHIFT)) {
1066 		pr_err("Address: 0x%llx already allocated by SVM\n",
1067 			args->va_addr);
1068 		mutex_unlock(&p->svms.lock);
1069 		return -EADDRINUSE;
1070 	}
1071 
1072 	/* When register user buffer check if it has been registered by svm by
1073 	 * buffer cpu virtual address.
1074 	 */
1075 	if ((flags & KFD_IOC_ALLOC_MEM_FLAGS_USERPTR) &&
1076 	    interval_tree_iter_first(&p->svms.objects,
1077 				     args->mmap_offset >> PAGE_SHIFT,
1078 				     (args->mmap_offset  + args->size - 1) >> PAGE_SHIFT)) {
1079 		pr_err("User Buffer Address: 0x%llx already allocated by SVM\n",
1080 			args->mmap_offset);
1081 		mutex_unlock(&p->svms.lock);
1082 		return -EADDRINUSE;
1083 	}
1084 
1085 	mutex_unlock(&p->svms.lock);
1086 #endif
1087 	mutex_lock(&p->mutex);
1088 	pdd = kfd_process_device_data_by_id(p, args->gpu_id);
1089 	if (!pdd) {
1090 		err = -EINVAL;
1091 		goto err_pdd;
1092 	}
1093 
1094 	dev = pdd->dev;
1095 
1096 	if ((flags & KFD_IOC_ALLOC_MEM_FLAGS_PUBLIC) &&
1097 		(flags & KFD_IOC_ALLOC_MEM_FLAGS_VRAM) &&
1098 		!kfd_dev_is_large_bar(dev)) {
1099 		pr_err("Alloc host visible vram on small bar is not allowed\n");
1100 		err = -EINVAL;
1101 		goto err_large_bar;
1102 	}
1103 
1104 	pdd = kfd_bind_process_to_device(dev, p);
1105 	if (IS_ERR(pdd)) {
1106 		err = PTR_ERR(pdd);
1107 		goto err_unlock;
1108 	}
1109 
1110 	if (flags & KFD_IOC_ALLOC_MEM_FLAGS_DOORBELL) {
1111 		if (args->size != kfd_doorbell_process_slice(dev->kfd)) {
1112 			err = -EINVAL;
1113 			goto err_unlock;
1114 		}
1115 		offset = kfd_get_process_doorbells(pdd);
1116 		if (!offset) {
1117 			err = -ENOMEM;
1118 			goto err_unlock;
1119 		}
1120 	} else if (flags & KFD_IOC_ALLOC_MEM_FLAGS_MMIO_REMAP) {
1121 		if (args->size != PAGE_SIZE) {
1122 			err = -EINVAL;
1123 			goto err_unlock;
1124 		}
1125 		offset = dev->adev->rmmio_remap.bus_addr;
1126 		if (!offset || (PAGE_SIZE > 4096)) {
1127 			err = -ENOMEM;
1128 			goto err_unlock;
1129 		}
1130 	}
1131 
1132 	err = amdgpu_amdkfd_gpuvm_alloc_memory_of_gpu(
1133 		dev->adev, args->va_addr, args->size,
1134 		pdd->drm_priv, (struct kgd_mem **) &mem, &offset,
1135 		flags, false);
1136 
1137 	if (err)
1138 		goto err_unlock;
1139 
1140 	idr_handle = kfd_process_device_create_obj_handle(pdd, mem);
1141 	if (idr_handle < 0) {
1142 		err = -EFAULT;
1143 		goto err_free;
1144 	}
1145 
1146 	/* Update the VRAM usage count */
1147 	if (flags & KFD_IOC_ALLOC_MEM_FLAGS_VRAM) {
1148 		uint64_t size = args->size;
1149 
1150 		if (flags & KFD_IOC_ALLOC_MEM_FLAGS_AQL_QUEUE_MEM)
1151 			size >>= 1;
1152 		atomic64_add(PAGE_ALIGN(size), &pdd->vram_usage);
1153 	}
1154 
1155 	mutex_unlock(&p->mutex);
1156 
1157 	args->handle = MAKE_HANDLE(args->gpu_id, idr_handle);
1158 	args->mmap_offset = offset;
1159 
1160 	/* MMIO is mapped through kfd device
1161 	 * Generate a kfd mmap offset
1162 	 */
1163 	if (flags & KFD_IOC_ALLOC_MEM_FLAGS_MMIO_REMAP)
1164 		args->mmap_offset = KFD_MMAP_TYPE_MMIO
1165 					| KFD_MMAP_GPU_ID(args->gpu_id);
1166 
1167 	return 0;
1168 
1169 err_free:
1170 	amdgpu_amdkfd_gpuvm_free_memory_of_gpu(dev->adev, (struct kgd_mem *)mem,
1171 					       pdd->drm_priv, NULL);
1172 err_unlock:
1173 err_pdd:
1174 err_large_bar:
1175 	mutex_unlock(&p->mutex);
1176 	return err;
1177 }
1178 
1179 static int kfd_ioctl_free_memory_of_gpu(struct file *filep,
1180 					struct kfd_process *p, void *data)
1181 {
1182 	struct kfd_ioctl_free_memory_of_gpu_args *args = data;
1183 	struct kfd_process_device *pdd;
1184 	void *mem;
1185 	int ret;
1186 	uint64_t size = 0;
1187 
1188 	mutex_lock(&p->mutex);
1189 	/*
1190 	 * Safeguard to prevent user space from freeing signal BO.
1191 	 * It will be freed at process termination.
1192 	 */
1193 	if (p->signal_handle && (p->signal_handle == args->handle)) {
1194 		pr_err("Free signal BO is not allowed\n");
1195 		ret = -EPERM;
1196 		goto err_unlock;
1197 	}
1198 
1199 	pdd = kfd_process_device_data_by_id(p, GET_GPU_ID(args->handle));
1200 	if (!pdd) {
1201 		pr_err("Process device data doesn't exist\n");
1202 		ret = -EINVAL;
1203 		goto err_pdd;
1204 	}
1205 
1206 	mem = kfd_process_device_translate_handle(
1207 		pdd, GET_IDR_HANDLE(args->handle));
1208 	if (!mem) {
1209 		ret = -EINVAL;
1210 		goto err_unlock;
1211 	}
1212 
1213 	ret = amdgpu_amdkfd_gpuvm_free_memory_of_gpu(pdd->dev->adev,
1214 				(struct kgd_mem *)mem, pdd->drm_priv, &size);
1215 
1216 	/* If freeing the buffer failed, leave the handle in place for
1217 	 * clean-up during process tear-down.
1218 	 */
1219 	if (!ret)
1220 		kfd_process_device_remove_obj_handle(
1221 			pdd, GET_IDR_HANDLE(args->handle));
1222 
1223 	atomic64_sub(size, &pdd->vram_usage);
1224 
1225 err_unlock:
1226 err_pdd:
1227 	mutex_unlock(&p->mutex);
1228 	return ret;
1229 }
1230 
1231 static int kfd_ioctl_map_memory_to_gpu(struct file *filep,
1232 					struct kfd_process *p, void *data)
1233 {
1234 	struct kfd_ioctl_map_memory_to_gpu_args *args = data;
1235 	struct kfd_process_device *pdd, *peer_pdd;
1236 	void *mem;
1237 	struct kfd_node *dev;
1238 	long err = 0;
1239 	int i;
1240 	uint32_t *devices_arr = NULL;
1241 
1242 	if (!args->n_devices) {
1243 		pr_debug("Device IDs array empty\n");
1244 		return -EINVAL;
1245 	}
1246 	if (args->n_success > args->n_devices) {
1247 		pr_debug("n_success exceeds n_devices\n");
1248 		return -EINVAL;
1249 	}
1250 
1251 	devices_arr = kmalloc_array(args->n_devices, sizeof(*devices_arr),
1252 				    GFP_KERNEL);
1253 	if (!devices_arr)
1254 		return -ENOMEM;
1255 
1256 	err = copy_from_user(devices_arr,
1257 			     (void __user *)args->device_ids_array_ptr,
1258 			     args->n_devices * sizeof(*devices_arr));
1259 	if (err != 0) {
1260 		err = -EFAULT;
1261 		goto copy_from_user_failed;
1262 	}
1263 
1264 	mutex_lock(&p->mutex);
1265 	pdd = kfd_process_device_data_by_id(p, GET_GPU_ID(args->handle));
1266 	if (!pdd) {
1267 		err = -EINVAL;
1268 		goto get_process_device_data_failed;
1269 	}
1270 	dev = pdd->dev;
1271 
1272 	pdd = kfd_bind_process_to_device(dev, p);
1273 	if (IS_ERR(pdd)) {
1274 		err = PTR_ERR(pdd);
1275 		goto bind_process_to_device_failed;
1276 	}
1277 
1278 	mem = kfd_process_device_translate_handle(pdd,
1279 						GET_IDR_HANDLE(args->handle));
1280 	if (!mem) {
1281 		err = -ENOMEM;
1282 		goto get_mem_obj_from_handle_failed;
1283 	}
1284 
1285 	for (i = args->n_success; i < args->n_devices; i++) {
1286 		peer_pdd = kfd_process_device_data_by_id(p, devices_arr[i]);
1287 		if (!peer_pdd) {
1288 			pr_debug("Getting device by id failed for 0x%x\n",
1289 				 devices_arr[i]);
1290 			err = -EINVAL;
1291 			goto get_mem_obj_from_handle_failed;
1292 		}
1293 
1294 		peer_pdd = kfd_bind_process_to_device(peer_pdd->dev, p);
1295 		if (IS_ERR(peer_pdd)) {
1296 			err = PTR_ERR(peer_pdd);
1297 			goto get_mem_obj_from_handle_failed;
1298 		}
1299 
1300 		err = amdgpu_amdkfd_gpuvm_map_memory_to_gpu(
1301 			peer_pdd->dev->adev, (struct kgd_mem *)mem,
1302 			peer_pdd->drm_priv);
1303 		if (err) {
1304 			struct pci_dev *pdev = peer_pdd->dev->adev->pdev;
1305 
1306 			dev_err(dev->adev->dev,
1307 			       "Failed to map peer:%04x:%02x:%02x.%d mem_domain:%d\n",
1308 			       pci_domain_nr(pdev->bus),
1309 			       pdev->bus->number,
1310 			       PCI_SLOT(pdev->devfn),
1311 			       PCI_FUNC(pdev->devfn),
1312 			       ((struct kgd_mem *)mem)->domain);
1313 			goto map_memory_to_gpu_failed;
1314 		}
1315 		args->n_success = i+1;
1316 	}
1317 
1318 	err = amdgpu_amdkfd_gpuvm_sync_memory(dev->adev, (struct kgd_mem *) mem, true);
1319 	if (err) {
1320 		pr_debug("Sync memory failed, wait interrupted by user signal\n");
1321 		goto sync_memory_failed;
1322 	}
1323 
1324 	mutex_unlock(&p->mutex);
1325 
1326 	/* Flush TLBs after waiting for the page table updates to complete */
1327 	for (i = 0; i < args->n_devices; i++) {
1328 		peer_pdd = kfd_process_device_data_by_id(p, devices_arr[i]);
1329 		if (WARN_ON_ONCE(!peer_pdd))
1330 			continue;
1331 		kfd_flush_tlb(peer_pdd, TLB_FLUSH_LEGACY);
1332 	}
1333 	kfree(devices_arr);
1334 
1335 	return err;
1336 
1337 get_process_device_data_failed:
1338 bind_process_to_device_failed:
1339 get_mem_obj_from_handle_failed:
1340 map_memory_to_gpu_failed:
1341 sync_memory_failed:
1342 	mutex_unlock(&p->mutex);
1343 copy_from_user_failed:
1344 	kfree(devices_arr);
1345 
1346 	return err;
1347 }
1348 
1349 static int kfd_ioctl_unmap_memory_from_gpu(struct file *filep,
1350 					struct kfd_process *p, void *data)
1351 {
1352 	struct kfd_ioctl_unmap_memory_from_gpu_args *args = data;
1353 	struct kfd_process_device *pdd, *peer_pdd;
1354 	void *mem;
1355 	long err = 0;
1356 	uint32_t *devices_arr = NULL, i;
1357 	bool flush_tlb;
1358 
1359 	if (!args->n_devices) {
1360 		pr_debug("Device IDs array empty\n");
1361 		return -EINVAL;
1362 	}
1363 	if (args->n_success > args->n_devices) {
1364 		pr_debug("n_success exceeds n_devices\n");
1365 		return -EINVAL;
1366 	}
1367 
1368 	devices_arr = kmalloc_array(args->n_devices, sizeof(*devices_arr),
1369 				    GFP_KERNEL);
1370 	if (!devices_arr)
1371 		return -ENOMEM;
1372 
1373 	err = copy_from_user(devices_arr,
1374 			     (void __user *)args->device_ids_array_ptr,
1375 			     args->n_devices * sizeof(*devices_arr));
1376 	if (err != 0) {
1377 		err = -EFAULT;
1378 		goto copy_from_user_failed;
1379 	}
1380 
1381 	mutex_lock(&p->mutex);
1382 	pdd = kfd_process_device_data_by_id(p, GET_GPU_ID(args->handle));
1383 	if (!pdd) {
1384 		err = -EINVAL;
1385 		goto bind_process_to_device_failed;
1386 	}
1387 
1388 	mem = kfd_process_device_translate_handle(pdd,
1389 						GET_IDR_HANDLE(args->handle));
1390 	if (!mem) {
1391 		err = -ENOMEM;
1392 		goto get_mem_obj_from_handle_failed;
1393 	}
1394 
1395 	for (i = args->n_success; i < args->n_devices; i++) {
1396 		peer_pdd = kfd_process_device_data_by_id(p, devices_arr[i]);
1397 		if (!peer_pdd) {
1398 			err = -EINVAL;
1399 			goto get_mem_obj_from_handle_failed;
1400 		}
1401 		err = amdgpu_amdkfd_gpuvm_unmap_memory_from_gpu(
1402 			peer_pdd->dev->adev, (struct kgd_mem *)mem, peer_pdd->drm_priv);
1403 		if (err) {
1404 			pr_debug("Failed to unmap from gpu %d/%d\n", i, args->n_devices);
1405 			goto unmap_memory_from_gpu_failed;
1406 		}
1407 		args->n_success = i+1;
1408 	}
1409 
1410 	flush_tlb = kfd_flush_tlb_after_unmap(pdd->dev->kfd);
1411 	if (flush_tlb) {
1412 		err = amdgpu_amdkfd_gpuvm_sync_memory(pdd->dev->adev,
1413 				(struct kgd_mem *) mem, true);
1414 		if (err) {
1415 			pr_debug("Sync memory failed, wait interrupted by user signal\n");
1416 			goto sync_memory_failed;
1417 		}
1418 	}
1419 
1420 	/* Flush TLBs after waiting for the page table updates to complete */
1421 	for (i = 0; i < args->n_devices; i++) {
1422 		peer_pdd = kfd_process_device_data_by_id(p, devices_arr[i]);
1423 		if (WARN_ON_ONCE(!peer_pdd))
1424 			continue;
1425 		if (flush_tlb)
1426 			kfd_flush_tlb(peer_pdd, TLB_FLUSH_HEAVYWEIGHT);
1427 
1428 		/* Remove dma mapping after tlb flush to avoid IO_PAGE_FAULT */
1429 		err = amdgpu_amdkfd_gpuvm_dmaunmap_mem(mem, peer_pdd->drm_priv);
1430 		if (err)
1431 			goto sync_memory_failed;
1432 	}
1433 
1434 	mutex_unlock(&p->mutex);
1435 
1436 	kfree(devices_arr);
1437 
1438 	return 0;
1439 
1440 bind_process_to_device_failed:
1441 get_mem_obj_from_handle_failed:
1442 unmap_memory_from_gpu_failed:
1443 sync_memory_failed:
1444 	mutex_unlock(&p->mutex);
1445 copy_from_user_failed:
1446 	kfree(devices_arr);
1447 	return err;
1448 }
1449 
1450 static int kfd_ioctl_alloc_queue_gws(struct file *filep,
1451 		struct kfd_process *p, void *data)
1452 {
1453 	int retval;
1454 	struct kfd_ioctl_alloc_queue_gws_args *args = data;
1455 	struct queue *q;
1456 	struct kfd_node *dev;
1457 
1458 	mutex_lock(&p->mutex);
1459 	q = pqm_get_user_queue(&p->pqm, args->queue_id);
1460 
1461 	if (q) {
1462 		dev = q->device;
1463 	} else {
1464 		retval = -EINVAL;
1465 		goto out_unlock;
1466 	}
1467 
1468 	if (!dev->gws) {
1469 		retval = -ENODEV;
1470 		goto out_unlock;
1471 	}
1472 
1473 	if (dev->dqm->sched_policy == KFD_SCHED_POLICY_NO_HWS) {
1474 		retval = -ENODEV;
1475 		goto out_unlock;
1476 	}
1477 
1478 	if (p->debug_trap_enabled && (!kfd_dbg_has_gws_support(dev) ||
1479 				      kfd_dbg_has_cwsr_workaround(dev))) {
1480 		retval = -EBUSY;
1481 		goto out_unlock;
1482 	}
1483 
1484 	retval = pqm_set_gws(&p->pqm, args->queue_id, args->num_gws ? dev->gws : NULL);
1485 	mutex_unlock(&p->mutex);
1486 
1487 	args->first_gws = 0;
1488 	return retval;
1489 
1490 out_unlock:
1491 	mutex_unlock(&p->mutex);
1492 	return retval;
1493 }
1494 
1495 static int kfd_ioctl_get_dmabuf_info(struct file *filep,
1496 		struct kfd_process *p, void *data)
1497 {
1498 	struct kfd_ioctl_get_dmabuf_info_args *args = data;
1499 	struct kfd_node *dev = NULL;
1500 	struct amdgpu_device *dmabuf_adev;
1501 	void *metadata_buffer = NULL;
1502 	uint32_t flags;
1503 	int8_t xcp_id;
1504 	unsigned int i;
1505 	int r;
1506 
1507 	/* Find a KFD GPU device that supports the get_dmabuf_info query */
1508 	for (i = 0; kfd_topology_enum_kfd_devices(i, &dev) == 0; i++)
1509 		if (dev && !kfd_devcgroup_check_permission(dev))
1510 			break;
1511 	if (!dev)
1512 		return -EINVAL;
1513 
1514 	if (args->metadata_ptr) {
1515 		metadata_buffer = kzalloc(args->metadata_size, GFP_KERNEL);
1516 		if (!metadata_buffer)
1517 			return -ENOMEM;
1518 	}
1519 
1520 	/* Get dmabuf info from KGD */
1521 	r = amdgpu_amdkfd_get_dmabuf_info(dev->adev, args->dmabuf_fd,
1522 					  &dmabuf_adev, &args->size,
1523 					  metadata_buffer, args->metadata_size,
1524 					  &args->metadata_size, &flags, &xcp_id);
1525 	if (r)
1526 		goto exit;
1527 
1528 	if (xcp_id >= 0)
1529 		args->gpu_id = dmabuf_adev->kfd.dev->nodes[xcp_id]->id;
1530 	else
1531 		args->gpu_id = dev->id;
1532 	args->flags = flags;
1533 
1534 	/* Copy metadata buffer to user mode */
1535 	if (metadata_buffer) {
1536 		r = copy_to_user((void __user *)args->metadata_ptr,
1537 				 metadata_buffer, args->metadata_size);
1538 		if (r != 0)
1539 			r = -EFAULT;
1540 	}
1541 
1542 exit:
1543 	kfree(metadata_buffer);
1544 
1545 	return r;
1546 }
1547 
1548 static int kfd_ioctl_import_dmabuf(struct file *filep,
1549 				   struct kfd_process *p, void *data)
1550 {
1551 	struct kfd_ioctl_import_dmabuf_args *args = data;
1552 	struct kfd_process_device *pdd;
1553 	int idr_handle;
1554 	uint64_t size;
1555 	void *mem;
1556 	int r;
1557 
1558 	mutex_lock(&p->mutex);
1559 	pdd = kfd_process_device_data_by_id(p, args->gpu_id);
1560 	if (!pdd) {
1561 		r = -EINVAL;
1562 		goto err_unlock;
1563 	}
1564 
1565 	pdd = kfd_bind_process_to_device(pdd->dev, p);
1566 	if (IS_ERR(pdd)) {
1567 		r = PTR_ERR(pdd);
1568 		goto err_unlock;
1569 	}
1570 
1571 	r = amdgpu_amdkfd_gpuvm_import_dmabuf_fd(pdd->dev->adev, args->dmabuf_fd,
1572 						 args->va_addr, pdd->drm_priv,
1573 						 (struct kgd_mem **)&mem, &size,
1574 						 NULL);
1575 	if (r)
1576 		goto err_unlock;
1577 
1578 	idr_handle = kfd_process_device_create_obj_handle(pdd, mem);
1579 	if (idr_handle < 0) {
1580 		r = -EFAULT;
1581 		goto err_free;
1582 	}
1583 
1584 	mutex_unlock(&p->mutex);
1585 
1586 	args->handle = MAKE_HANDLE(args->gpu_id, idr_handle);
1587 
1588 	return 0;
1589 
1590 err_free:
1591 	amdgpu_amdkfd_gpuvm_free_memory_of_gpu(pdd->dev->adev, (struct kgd_mem *)mem,
1592 					       pdd->drm_priv, NULL);
1593 err_unlock:
1594 	mutex_unlock(&p->mutex);
1595 	return r;
1596 }
1597 
1598 static int kfd_ioctl_export_dmabuf(struct file *filep,
1599 				   struct kfd_process *p, void *data)
1600 {
1601 	struct kfd_ioctl_export_dmabuf_args *args = data;
1602 	struct kfd_process_device *pdd;
1603 	struct dma_buf *dmabuf;
1604 	struct kfd_node *dev;
1605 	void *mem;
1606 	int ret = 0;
1607 
1608 	dev = kfd_device_by_id(GET_GPU_ID(args->handle));
1609 	if (!dev)
1610 		return -EINVAL;
1611 
1612 	mutex_lock(&p->mutex);
1613 
1614 	pdd = kfd_get_process_device_data(dev, p);
1615 	if (!pdd) {
1616 		ret = -EINVAL;
1617 		goto err_unlock;
1618 	}
1619 
1620 	mem = kfd_process_device_translate_handle(pdd,
1621 						GET_IDR_HANDLE(args->handle));
1622 	if (!mem) {
1623 		ret = -EINVAL;
1624 		goto err_unlock;
1625 	}
1626 
1627 	ret = amdgpu_amdkfd_gpuvm_export_dmabuf(mem, &dmabuf);
1628 	mutex_unlock(&p->mutex);
1629 	if (ret)
1630 		goto err_out;
1631 
1632 	ret = dma_buf_fd(dmabuf, args->flags);
1633 	if (ret < 0) {
1634 		dma_buf_put(dmabuf);
1635 		goto err_out;
1636 	}
1637 	/* dma_buf_fd assigns the reference count to the fd, no need to
1638 	 * put the reference here.
1639 	 */
1640 	args->dmabuf_fd = ret;
1641 
1642 	return 0;
1643 
1644 err_unlock:
1645 	mutex_unlock(&p->mutex);
1646 err_out:
1647 	return ret;
1648 }
1649 
1650 /* Handle requests for watching SMI events */
1651 static int kfd_ioctl_smi_events(struct file *filep,
1652 				struct kfd_process *p, void *data)
1653 {
1654 	struct kfd_ioctl_smi_events_args *args = data;
1655 	struct kfd_process_device *pdd;
1656 
1657 	mutex_lock(&p->mutex);
1658 
1659 	pdd = kfd_process_device_data_by_id(p, args->gpuid);
1660 	mutex_unlock(&p->mutex);
1661 	if (!pdd)
1662 		return -EINVAL;
1663 
1664 	return kfd_smi_event_open(pdd->dev, &args->anon_fd);
1665 }
1666 
1667 #if IS_ENABLED(CONFIG_HSA_AMD_SVM)
1668 
1669 static int kfd_ioctl_set_xnack_mode(struct file *filep,
1670 				    struct kfd_process *p, void *data)
1671 {
1672 	struct kfd_ioctl_set_xnack_mode_args *args = data;
1673 	int r = 0;
1674 
1675 	mutex_lock(&p->mutex);
1676 	if (args->xnack_enabled >= 0) {
1677 		if (!list_empty(&p->pqm.queues)) {
1678 			pr_debug("Process has user queues running\n");
1679 			r = -EBUSY;
1680 			goto out_unlock;
1681 		}
1682 
1683 		if (p->xnack_enabled == args->xnack_enabled)
1684 			goto out_unlock;
1685 
1686 		if (args->xnack_enabled && !kfd_process_xnack_mode(p, true)) {
1687 			r = -EPERM;
1688 			goto out_unlock;
1689 		}
1690 
1691 		r = svm_range_switch_xnack_reserve_mem(p, args->xnack_enabled);
1692 	} else {
1693 		args->xnack_enabled = p->xnack_enabled;
1694 	}
1695 
1696 out_unlock:
1697 	mutex_unlock(&p->mutex);
1698 
1699 	return r;
1700 }
1701 
1702 static int kfd_ioctl_svm(struct file *filep, struct kfd_process *p, void *data)
1703 {
1704 	struct kfd_ioctl_svm_args *args = data;
1705 	int r = 0;
1706 
1707 	pr_debug("start 0x%llx size 0x%llx op 0x%x nattr 0x%x\n",
1708 		 args->start_addr, args->size, args->op, args->nattr);
1709 
1710 	if ((args->start_addr & ~PAGE_MASK) || (args->size & ~PAGE_MASK))
1711 		return -EINVAL;
1712 	if (!args->start_addr || !args->size)
1713 		return -EINVAL;
1714 
1715 	r = svm_ioctl(p, args->op, args->start_addr, args->size, args->nattr,
1716 		      args->attrs);
1717 
1718 	return r;
1719 }
1720 #else
1721 static int kfd_ioctl_set_xnack_mode(struct file *filep,
1722 				    struct kfd_process *p, void *data)
1723 {
1724 	return -EPERM;
1725 }
1726 static int kfd_ioctl_svm(struct file *filep, struct kfd_process *p, void *data)
1727 {
1728 	return -EPERM;
1729 }
1730 #endif
1731 
1732 static int criu_checkpoint_process(struct kfd_process *p,
1733 			     uint8_t __user *user_priv_data,
1734 			     uint64_t *priv_offset)
1735 {
1736 	struct kfd_criu_process_priv_data process_priv;
1737 	int ret;
1738 
1739 	memset(&process_priv, 0, sizeof(process_priv));
1740 
1741 	process_priv.version = KFD_CRIU_PRIV_VERSION;
1742 	/* For CR, we don't consider negative xnack mode which is used for
1743 	 * querying without changing it, here 0 simply means disabled and 1
1744 	 * means enabled so retry for finding a valid PTE.
1745 	 */
1746 	process_priv.xnack_mode = p->xnack_enabled ? 1 : 0;
1747 
1748 	ret = copy_to_user(user_priv_data + *priv_offset,
1749 				&process_priv, sizeof(process_priv));
1750 
1751 	if (ret) {
1752 		pr_err("Failed to copy process information to user\n");
1753 		ret = -EFAULT;
1754 	}
1755 
1756 	*priv_offset += sizeof(process_priv);
1757 	return ret;
1758 }
1759 
1760 static int criu_checkpoint_devices(struct kfd_process *p,
1761 			     uint32_t num_devices,
1762 			     uint8_t __user *user_addr,
1763 			     uint8_t __user *user_priv_data,
1764 			     uint64_t *priv_offset)
1765 {
1766 	struct kfd_criu_device_priv_data *device_priv = NULL;
1767 	struct kfd_criu_device_bucket *device_buckets = NULL;
1768 	int ret = 0, i;
1769 
1770 	device_buckets = kvzalloc(num_devices * sizeof(*device_buckets), GFP_KERNEL);
1771 	if (!device_buckets) {
1772 		ret = -ENOMEM;
1773 		goto exit;
1774 	}
1775 
1776 	device_priv = kvzalloc(num_devices * sizeof(*device_priv), GFP_KERNEL);
1777 	if (!device_priv) {
1778 		ret = -ENOMEM;
1779 		goto exit;
1780 	}
1781 
1782 	for (i = 0; i < num_devices; i++) {
1783 		struct kfd_process_device *pdd = p->pdds[i];
1784 
1785 		device_buckets[i].user_gpu_id = pdd->user_gpu_id;
1786 		device_buckets[i].actual_gpu_id = pdd->dev->id;
1787 
1788 		/*
1789 		 * priv_data does not contain useful information for now and is reserved for
1790 		 * future use, so we do not set its contents.
1791 		 */
1792 	}
1793 
1794 	ret = copy_to_user(user_addr, device_buckets, num_devices * sizeof(*device_buckets));
1795 	if (ret) {
1796 		pr_err("Failed to copy device information to user\n");
1797 		ret = -EFAULT;
1798 		goto exit;
1799 	}
1800 
1801 	ret = copy_to_user(user_priv_data + *priv_offset,
1802 			   device_priv,
1803 			   num_devices * sizeof(*device_priv));
1804 	if (ret) {
1805 		pr_err("Failed to copy device information to user\n");
1806 		ret = -EFAULT;
1807 	}
1808 	*priv_offset += num_devices * sizeof(*device_priv);
1809 
1810 exit:
1811 	kvfree(device_buckets);
1812 	kvfree(device_priv);
1813 	return ret;
1814 }
1815 
1816 static uint32_t get_process_num_bos(struct kfd_process *p)
1817 {
1818 	uint32_t num_of_bos = 0;
1819 	int i;
1820 
1821 	/* Run over all PDDs of the process */
1822 	for (i = 0; i < p->n_pdds; i++) {
1823 		struct kfd_process_device *pdd = p->pdds[i];
1824 		void *mem;
1825 		int id;
1826 
1827 		idr_for_each_entry(&pdd->alloc_idr, mem, id) {
1828 			struct kgd_mem *kgd_mem = (struct kgd_mem *)mem;
1829 
1830 			if (!kgd_mem->va || kgd_mem->va > pdd->gpuvm_base)
1831 				num_of_bos++;
1832 		}
1833 	}
1834 	return num_of_bos;
1835 }
1836 
1837 static int criu_get_prime_handle(struct kgd_mem *mem,
1838 				 int flags, u32 *shared_fd,
1839 				 struct file **file)
1840 {
1841 	struct dma_buf *dmabuf;
1842 	int ret;
1843 
1844 	ret = amdgpu_amdkfd_gpuvm_export_dmabuf(mem, &dmabuf);
1845 	if (ret) {
1846 		pr_err("dmabuf export failed for the BO\n");
1847 		return ret;
1848 	}
1849 
1850 	ret = get_unused_fd_flags(flags);
1851 	if (ret < 0) {
1852 		pr_err("dmabuf create fd failed, ret:%d\n", ret);
1853 		goto out_free_dmabuf;
1854 	}
1855 
1856 	*shared_fd = ret;
1857 	*file = dmabuf->file;
1858 	return 0;
1859 
1860 out_free_dmabuf:
1861 	dma_buf_put(dmabuf);
1862 	return ret;
1863 }
1864 
1865 static void commit_files(struct file **files,
1866 			 struct kfd_criu_bo_bucket *bo_buckets,
1867 			 unsigned int count,
1868 			 int err)
1869 {
1870 	while (count--) {
1871 		struct file *file = files[count];
1872 
1873 		if (!file)
1874 			continue;
1875 		if (err) {
1876 			fput(file);
1877 			put_unused_fd(bo_buckets[count].dmabuf_fd);
1878 		} else {
1879 			fd_install(bo_buckets[count].dmabuf_fd, file);
1880 		}
1881 	}
1882 }
1883 
1884 static int criu_checkpoint_bos(struct kfd_process *p,
1885 			       uint32_t num_bos,
1886 			       uint8_t __user *user_bos,
1887 			       uint8_t __user *user_priv_data,
1888 			       uint64_t *priv_offset)
1889 {
1890 	struct kfd_criu_bo_bucket *bo_buckets;
1891 	struct kfd_criu_bo_priv_data *bo_privs;
1892 	struct file **files = NULL;
1893 	int ret = 0, pdd_index, bo_index = 0, id;
1894 	void *mem;
1895 
1896 	bo_buckets = kvzalloc(num_bos * sizeof(*bo_buckets), GFP_KERNEL);
1897 	if (!bo_buckets)
1898 		return -ENOMEM;
1899 
1900 	bo_privs = kvzalloc(num_bos * sizeof(*bo_privs), GFP_KERNEL);
1901 	if (!bo_privs) {
1902 		ret = -ENOMEM;
1903 		goto exit;
1904 	}
1905 
1906 	files = kvzalloc(num_bos * sizeof(struct file *), GFP_KERNEL);
1907 	if (!files) {
1908 		ret = -ENOMEM;
1909 		goto exit;
1910 	}
1911 
1912 	for (pdd_index = 0; pdd_index < p->n_pdds; pdd_index++) {
1913 		struct kfd_process_device *pdd = p->pdds[pdd_index];
1914 		struct amdgpu_bo *dumper_bo;
1915 		struct kgd_mem *kgd_mem;
1916 
1917 		idr_for_each_entry(&pdd->alloc_idr, mem, id) {
1918 			struct kfd_criu_bo_bucket *bo_bucket;
1919 			struct kfd_criu_bo_priv_data *bo_priv;
1920 			int i, dev_idx = 0;
1921 
1922 			kgd_mem = (struct kgd_mem *)mem;
1923 			dumper_bo = kgd_mem->bo;
1924 
1925 			/* Skip checkpointing BOs that are used for Trap handler
1926 			 * code and state. Currently, these BOs have a VA that
1927 			 * is less GPUVM Base
1928 			 */
1929 			if (kgd_mem->va && kgd_mem->va <= pdd->gpuvm_base)
1930 				continue;
1931 
1932 			bo_bucket = &bo_buckets[bo_index];
1933 			bo_priv = &bo_privs[bo_index];
1934 
1935 			bo_bucket->gpu_id = pdd->user_gpu_id;
1936 			bo_bucket->addr = (uint64_t)kgd_mem->va;
1937 			bo_bucket->size = amdgpu_bo_size(dumper_bo);
1938 			bo_bucket->alloc_flags = (uint32_t)kgd_mem->alloc_flags;
1939 			bo_priv->idr_handle = id;
1940 
1941 			if (bo_bucket->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_USERPTR) {
1942 				ret = amdgpu_ttm_tt_get_userptr(&dumper_bo->tbo,
1943 								&bo_priv->user_addr);
1944 				if (ret) {
1945 					pr_err("Failed to obtain user address for user-pointer bo\n");
1946 					goto exit;
1947 				}
1948 			}
1949 			if (bo_bucket->alloc_flags
1950 			    & (KFD_IOC_ALLOC_MEM_FLAGS_VRAM | KFD_IOC_ALLOC_MEM_FLAGS_GTT)) {
1951 				ret = criu_get_prime_handle(kgd_mem,
1952 						bo_bucket->alloc_flags &
1953 						KFD_IOC_ALLOC_MEM_FLAGS_WRITABLE ? DRM_RDWR : 0,
1954 						&bo_bucket->dmabuf_fd, &files[bo_index]);
1955 				if (ret)
1956 					goto exit;
1957 			} else {
1958 				bo_bucket->dmabuf_fd = KFD_INVALID_FD;
1959 			}
1960 
1961 			if (bo_bucket->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_DOORBELL)
1962 				bo_bucket->offset = KFD_MMAP_TYPE_DOORBELL |
1963 					KFD_MMAP_GPU_ID(pdd->dev->id);
1964 			else if (bo_bucket->alloc_flags &
1965 				KFD_IOC_ALLOC_MEM_FLAGS_MMIO_REMAP)
1966 				bo_bucket->offset = KFD_MMAP_TYPE_MMIO |
1967 					KFD_MMAP_GPU_ID(pdd->dev->id);
1968 			else
1969 				bo_bucket->offset = amdgpu_bo_mmap_offset(dumper_bo);
1970 
1971 			for (i = 0; i < p->n_pdds; i++) {
1972 				if (amdgpu_amdkfd_bo_mapped_to_dev(p->pdds[i]->drm_priv, kgd_mem))
1973 					bo_priv->mapped_gpuids[dev_idx++] = p->pdds[i]->user_gpu_id;
1974 			}
1975 
1976 			pr_debug("bo_size = 0x%llx, bo_addr = 0x%llx bo_offset = 0x%llx\n"
1977 					"gpu_id = 0x%x alloc_flags = 0x%x idr_handle = 0x%x",
1978 					bo_bucket->size,
1979 					bo_bucket->addr,
1980 					bo_bucket->offset,
1981 					bo_bucket->gpu_id,
1982 					bo_bucket->alloc_flags,
1983 					bo_priv->idr_handle);
1984 			bo_index++;
1985 		}
1986 	}
1987 
1988 	ret = copy_to_user(user_bos, bo_buckets, num_bos * sizeof(*bo_buckets));
1989 	if (ret) {
1990 		pr_err("Failed to copy BO information to user\n");
1991 		ret = -EFAULT;
1992 		goto exit;
1993 	}
1994 
1995 	ret = copy_to_user(user_priv_data + *priv_offset, bo_privs, num_bos * sizeof(*bo_privs));
1996 	if (ret) {
1997 		pr_err("Failed to copy BO priv information to user\n");
1998 		ret = -EFAULT;
1999 		goto exit;
2000 	}
2001 
2002 	*priv_offset += num_bos * sizeof(*bo_privs);
2003 
2004 exit:
2005 	commit_files(files, bo_buckets, bo_index, ret);
2006 	kvfree(files);
2007 	kvfree(bo_buckets);
2008 	kvfree(bo_privs);
2009 	return ret;
2010 }
2011 
2012 static int criu_get_process_object_info(struct kfd_process *p,
2013 					uint32_t *num_devices,
2014 					uint32_t *num_bos,
2015 					uint32_t *num_objects,
2016 					uint64_t *objs_priv_size)
2017 {
2018 	uint64_t queues_priv_data_size, svm_priv_data_size, priv_size;
2019 	uint32_t num_queues, num_events, num_svm_ranges;
2020 	int ret;
2021 
2022 	*num_devices = p->n_pdds;
2023 	*num_bos = get_process_num_bos(p);
2024 
2025 	ret = kfd_process_get_queue_info(p, &num_queues, &queues_priv_data_size);
2026 	if (ret)
2027 		return ret;
2028 
2029 	num_events = kfd_get_num_events(p);
2030 
2031 	ret = svm_range_get_info(p, &num_svm_ranges, &svm_priv_data_size);
2032 	if (ret)
2033 		return ret;
2034 
2035 	*num_objects = num_queues + num_events + num_svm_ranges;
2036 
2037 	if (objs_priv_size) {
2038 		priv_size = sizeof(struct kfd_criu_process_priv_data);
2039 		priv_size += *num_devices * sizeof(struct kfd_criu_device_priv_data);
2040 		priv_size += *num_bos * sizeof(struct kfd_criu_bo_priv_data);
2041 		priv_size += queues_priv_data_size;
2042 		priv_size += num_events * sizeof(struct kfd_criu_event_priv_data);
2043 		priv_size += svm_priv_data_size;
2044 		*objs_priv_size = priv_size;
2045 	}
2046 	return 0;
2047 }
2048 
2049 static int criu_checkpoint(struct file *filep,
2050 			   struct kfd_process *p,
2051 			   struct kfd_ioctl_criu_args *args)
2052 {
2053 	int ret;
2054 	uint32_t num_devices, num_bos, num_objects;
2055 	uint64_t priv_size, priv_offset = 0, bo_priv_offset;
2056 
2057 	if (!args->devices || !args->bos || !args->priv_data)
2058 		return -EINVAL;
2059 
2060 	mutex_lock(&p->mutex);
2061 
2062 	if (!p->n_pdds) {
2063 		pr_err("No pdd for given process\n");
2064 		ret = -ENODEV;
2065 		goto exit_unlock;
2066 	}
2067 
2068 	/* Confirm all process queues are evicted */
2069 	if (!p->queues_paused) {
2070 		pr_err("Cannot dump process when queues are not in evicted state\n");
2071 		/* CRIU plugin did not call op PROCESS_INFO before checkpointing */
2072 		ret = -EINVAL;
2073 		goto exit_unlock;
2074 	}
2075 
2076 	ret = criu_get_process_object_info(p, &num_devices, &num_bos, &num_objects, &priv_size);
2077 	if (ret)
2078 		goto exit_unlock;
2079 
2080 	if (num_devices != args->num_devices ||
2081 	    num_bos != args->num_bos ||
2082 	    num_objects != args->num_objects ||
2083 	    priv_size != args->priv_data_size) {
2084 
2085 		ret = -EINVAL;
2086 		goto exit_unlock;
2087 	}
2088 
2089 	/* each function will store private data inside priv_data and adjust priv_offset */
2090 	ret = criu_checkpoint_process(p, (uint8_t __user *)args->priv_data, &priv_offset);
2091 	if (ret)
2092 		goto exit_unlock;
2093 
2094 	ret = criu_checkpoint_devices(p, num_devices, (uint8_t __user *)args->devices,
2095 				(uint8_t __user *)args->priv_data, &priv_offset);
2096 	if (ret)
2097 		goto exit_unlock;
2098 
2099 	/* Leave room for BOs in the private data. They need to be restored
2100 	 * before events, but we checkpoint them last to simplify the error
2101 	 * handling.
2102 	 */
2103 	bo_priv_offset = priv_offset;
2104 	priv_offset += num_bos * sizeof(struct kfd_criu_bo_priv_data);
2105 
2106 	if (num_objects) {
2107 		ret = kfd_criu_checkpoint_queues(p, (uint8_t __user *)args->priv_data,
2108 						 &priv_offset);
2109 		if (ret)
2110 			goto exit_unlock;
2111 
2112 		ret = kfd_criu_checkpoint_events(p, (uint8_t __user *)args->priv_data,
2113 						 &priv_offset);
2114 		if (ret)
2115 			goto exit_unlock;
2116 
2117 		ret = kfd_criu_checkpoint_svm(p, (uint8_t __user *)args->priv_data, &priv_offset);
2118 		if (ret)
2119 			goto exit_unlock;
2120 	}
2121 
2122 	/* This must be the last thing in this function that can fail.
2123 	 * Otherwise we leak dmabuf file descriptors.
2124 	 */
2125 	ret = criu_checkpoint_bos(p, num_bos, (uint8_t __user *)args->bos,
2126 			   (uint8_t __user *)args->priv_data, &bo_priv_offset);
2127 
2128 exit_unlock:
2129 	mutex_unlock(&p->mutex);
2130 	if (ret)
2131 		pr_err("Failed to dump CRIU ret:%d\n", ret);
2132 	else
2133 		pr_debug("CRIU dump ret:%d\n", ret);
2134 
2135 	return ret;
2136 }
2137 
2138 static int criu_restore_process(struct kfd_process *p,
2139 				struct kfd_ioctl_criu_args *args,
2140 				uint64_t *priv_offset,
2141 				uint64_t max_priv_data_size)
2142 {
2143 	int ret = 0;
2144 	struct kfd_criu_process_priv_data process_priv;
2145 
2146 	if (*priv_offset + sizeof(process_priv) > max_priv_data_size)
2147 		return -EINVAL;
2148 
2149 	ret = copy_from_user(&process_priv,
2150 				(void __user *)(args->priv_data + *priv_offset),
2151 				sizeof(process_priv));
2152 	if (ret) {
2153 		pr_err("Failed to copy process private information from user\n");
2154 		ret = -EFAULT;
2155 		goto exit;
2156 	}
2157 	*priv_offset += sizeof(process_priv);
2158 
2159 	if (process_priv.version != KFD_CRIU_PRIV_VERSION) {
2160 		pr_err("Invalid CRIU API version (checkpointed:%d current:%d)\n",
2161 			process_priv.version, KFD_CRIU_PRIV_VERSION);
2162 		return -EINVAL;
2163 	}
2164 
2165 	pr_debug("Setting XNACK mode\n");
2166 	if (process_priv.xnack_mode && !kfd_process_xnack_mode(p, true)) {
2167 		pr_err("xnack mode cannot be set\n");
2168 		ret = -EPERM;
2169 		goto exit;
2170 	} else {
2171 		pr_debug("set xnack mode: %d\n", process_priv.xnack_mode);
2172 		p->xnack_enabled = process_priv.xnack_mode;
2173 	}
2174 
2175 exit:
2176 	return ret;
2177 }
2178 
2179 static int criu_restore_devices(struct kfd_process *p,
2180 				struct kfd_ioctl_criu_args *args,
2181 				uint64_t *priv_offset,
2182 				uint64_t max_priv_data_size)
2183 {
2184 	struct kfd_criu_device_bucket *device_buckets;
2185 	struct kfd_criu_device_priv_data *device_privs;
2186 	int ret = 0;
2187 	uint32_t i;
2188 
2189 	if (args->num_devices != p->n_pdds)
2190 		return -EINVAL;
2191 
2192 	if (*priv_offset + (args->num_devices * sizeof(*device_privs)) > max_priv_data_size)
2193 		return -EINVAL;
2194 
2195 	device_buckets = kmalloc_array(args->num_devices, sizeof(*device_buckets), GFP_KERNEL);
2196 	if (!device_buckets)
2197 		return -ENOMEM;
2198 
2199 	ret = copy_from_user(device_buckets, (void __user *)args->devices,
2200 				args->num_devices * sizeof(*device_buckets));
2201 	if (ret) {
2202 		pr_err("Failed to copy devices buckets from user\n");
2203 		ret = -EFAULT;
2204 		goto exit;
2205 	}
2206 
2207 	for (i = 0; i < args->num_devices; i++) {
2208 		struct kfd_node *dev;
2209 		struct kfd_process_device *pdd;
2210 		struct file *drm_file;
2211 
2212 		/* device private data is not currently used */
2213 
2214 		if (!device_buckets[i].user_gpu_id) {
2215 			pr_err("Invalid user gpu_id\n");
2216 			ret = -EINVAL;
2217 			goto exit;
2218 		}
2219 
2220 		dev = kfd_device_by_id(device_buckets[i].actual_gpu_id);
2221 		if (!dev) {
2222 			pr_err("Failed to find device with gpu_id = %x\n",
2223 				device_buckets[i].actual_gpu_id);
2224 			ret = -EINVAL;
2225 			goto exit;
2226 		}
2227 
2228 		pdd = kfd_get_process_device_data(dev, p);
2229 		if (!pdd) {
2230 			pr_err("Failed to get pdd for gpu_id = %x\n",
2231 					device_buckets[i].actual_gpu_id);
2232 			ret = -EINVAL;
2233 			goto exit;
2234 		}
2235 		pdd->user_gpu_id = device_buckets[i].user_gpu_id;
2236 
2237 		drm_file = fget(device_buckets[i].drm_fd);
2238 		if (!drm_file) {
2239 			pr_err("Invalid render node file descriptor sent from plugin (%d)\n",
2240 				device_buckets[i].drm_fd);
2241 			ret = -EINVAL;
2242 			goto exit;
2243 		}
2244 
2245 		if (pdd->drm_file) {
2246 			ret = -EINVAL;
2247 			goto exit;
2248 		}
2249 
2250 		/* create the vm using render nodes for kfd pdd */
2251 		if (kfd_process_device_init_vm(pdd, drm_file)) {
2252 			pr_err("could not init vm for given pdd\n");
2253 			/* On success, the PDD keeps the drm_file reference */
2254 			fput(drm_file);
2255 			ret = -EINVAL;
2256 			goto exit;
2257 		}
2258 		/*
2259 		 * pdd now already has the vm bound to render node so below api won't create a new
2260 		 * exclusive kfd mapping but use existing one with renderDXXX but is still needed
2261 		 * for iommu v2 binding  and runtime pm.
2262 		 */
2263 		pdd = kfd_bind_process_to_device(dev, p);
2264 		if (IS_ERR(pdd)) {
2265 			ret = PTR_ERR(pdd);
2266 			goto exit;
2267 		}
2268 
2269 		if (!pdd->qpd.proc_doorbells) {
2270 			ret = kfd_alloc_process_doorbells(dev->kfd, pdd);
2271 			if (ret)
2272 				goto exit;
2273 		}
2274 	}
2275 
2276 	/*
2277 	 * We are not copying device private data from user as we are not using the data for now,
2278 	 * but we still adjust for its private data.
2279 	 */
2280 	*priv_offset += args->num_devices * sizeof(*device_privs);
2281 
2282 exit:
2283 	kfree(device_buckets);
2284 	return ret;
2285 }
2286 
2287 static int criu_restore_memory_of_gpu(struct kfd_process_device *pdd,
2288 				      struct kfd_criu_bo_bucket *bo_bucket,
2289 				      struct kfd_criu_bo_priv_data *bo_priv,
2290 				      struct kgd_mem **kgd_mem)
2291 {
2292 	int idr_handle;
2293 	int ret;
2294 	const bool criu_resume = true;
2295 	u64 offset;
2296 
2297 	if (bo_bucket->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_DOORBELL) {
2298 		if (bo_bucket->size !=
2299 				kfd_doorbell_process_slice(pdd->dev->kfd))
2300 			return -EINVAL;
2301 
2302 		offset = kfd_get_process_doorbells(pdd);
2303 		if (!offset)
2304 			return -ENOMEM;
2305 	} else if (bo_bucket->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_MMIO_REMAP) {
2306 		/* MMIO BOs need remapped bus address */
2307 		if (bo_bucket->size != PAGE_SIZE) {
2308 			pr_err("Invalid page size\n");
2309 			return -EINVAL;
2310 		}
2311 		offset = pdd->dev->adev->rmmio_remap.bus_addr;
2312 		if (!offset || (PAGE_SIZE > 4096)) {
2313 			pr_err("amdgpu_amdkfd_get_mmio_remap_phys_addr failed\n");
2314 			return -ENOMEM;
2315 		}
2316 	} else if (bo_bucket->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_USERPTR) {
2317 		offset = bo_priv->user_addr;
2318 	}
2319 	/* Create the BO */
2320 	ret = amdgpu_amdkfd_gpuvm_alloc_memory_of_gpu(pdd->dev->adev, bo_bucket->addr,
2321 						      bo_bucket->size, pdd->drm_priv, kgd_mem,
2322 						      &offset, bo_bucket->alloc_flags, criu_resume);
2323 	if (ret) {
2324 		pr_err("Could not create the BO\n");
2325 		return ret;
2326 	}
2327 	pr_debug("New BO created: size:0x%llx addr:0x%llx offset:0x%llx\n",
2328 		 bo_bucket->size, bo_bucket->addr, offset);
2329 
2330 	/* Restore previous IDR handle */
2331 	pr_debug("Restoring old IDR handle for the BO");
2332 	idr_handle = idr_alloc(&pdd->alloc_idr, *kgd_mem, bo_priv->idr_handle,
2333 			       bo_priv->idr_handle + 1, GFP_KERNEL);
2334 
2335 	if (idr_handle < 0) {
2336 		pr_err("Could not allocate idr\n");
2337 		amdgpu_amdkfd_gpuvm_free_memory_of_gpu(pdd->dev->adev, *kgd_mem, pdd->drm_priv,
2338 						       NULL);
2339 		return -ENOMEM;
2340 	}
2341 
2342 	if (bo_bucket->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_DOORBELL)
2343 		bo_bucket->restored_offset = KFD_MMAP_TYPE_DOORBELL | KFD_MMAP_GPU_ID(pdd->dev->id);
2344 	if (bo_bucket->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_MMIO_REMAP) {
2345 		bo_bucket->restored_offset = KFD_MMAP_TYPE_MMIO | KFD_MMAP_GPU_ID(pdd->dev->id);
2346 	} else if (bo_bucket->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_GTT) {
2347 		bo_bucket->restored_offset = offset;
2348 	} else if (bo_bucket->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_VRAM) {
2349 		bo_bucket->restored_offset = offset;
2350 		/* Update the VRAM usage count */
2351 		atomic64_add(bo_bucket->size, &pdd->vram_usage);
2352 	}
2353 	return 0;
2354 }
2355 
2356 static int criu_restore_bo(struct kfd_process *p,
2357 			   struct kfd_criu_bo_bucket *bo_bucket,
2358 			   struct kfd_criu_bo_priv_data *bo_priv,
2359 			   struct file **file)
2360 {
2361 	struct kfd_process_device *pdd;
2362 	struct kgd_mem *kgd_mem;
2363 	int ret;
2364 	int j;
2365 
2366 	pr_debug("Restoring BO size:0x%llx addr:0x%llx gpu_id:0x%x flags:0x%x idr_handle:0x%x\n",
2367 		 bo_bucket->size, bo_bucket->addr, bo_bucket->gpu_id, bo_bucket->alloc_flags,
2368 		 bo_priv->idr_handle);
2369 
2370 	pdd = kfd_process_device_data_by_id(p, bo_bucket->gpu_id);
2371 	if (!pdd) {
2372 		pr_err("Failed to get pdd\n");
2373 		return -ENODEV;
2374 	}
2375 
2376 	ret = criu_restore_memory_of_gpu(pdd, bo_bucket, bo_priv, &kgd_mem);
2377 	if (ret)
2378 		return ret;
2379 
2380 	/* now map these BOs to GPU/s */
2381 	for (j = 0; j < p->n_pdds; j++) {
2382 		struct kfd_node *peer;
2383 		struct kfd_process_device *peer_pdd;
2384 
2385 		if (!bo_priv->mapped_gpuids[j])
2386 			break;
2387 
2388 		peer_pdd = kfd_process_device_data_by_id(p, bo_priv->mapped_gpuids[j]);
2389 		if (!peer_pdd)
2390 			return -EINVAL;
2391 
2392 		peer = peer_pdd->dev;
2393 
2394 		peer_pdd = kfd_bind_process_to_device(peer, p);
2395 		if (IS_ERR(peer_pdd))
2396 			return PTR_ERR(peer_pdd);
2397 
2398 		ret = amdgpu_amdkfd_gpuvm_map_memory_to_gpu(peer->adev, kgd_mem,
2399 							    peer_pdd->drm_priv);
2400 		if (ret) {
2401 			pr_err("Failed to map to gpu %d/%d\n", j, p->n_pdds);
2402 			return ret;
2403 		}
2404 	}
2405 
2406 	pr_debug("map memory was successful for the BO\n");
2407 	/* create the dmabuf object and export the bo */
2408 	if (bo_bucket->alloc_flags
2409 	    & (KFD_IOC_ALLOC_MEM_FLAGS_VRAM | KFD_IOC_ALLOC_MEM_FLAGS_GTT)) {
2410 		ret = criu_get_prime_handle(kgd_mem, DRM_RDWR,
2411 					    &bo_bucket->dmabuf_fd, file);
2412 		if (ret)
2413 			return ret;
2414 	} else {
2415 		bo_bucket->dmabuf_fd = KFD_INVALID_FD;
2416 	}
2417 
2418 	return 0;
2419 }
2420 
2421 static int criu_restore_bos(struct kfd_process *p,
2422 			    struct kfd_ioctl_criu_args *args,
2423 			    uint64_t *priv_offset,
2424 			    uint64_t max_priv_data_size)
2425 {
2426 	struct kfd_criu_bo_bucket *bo_buckets = NULL;
2427 	struct kfd_criu_bo_priv_data *bo_privs = NULL;
2428 	struct file **files = NULL;
2429 	int ret = 0;
2430 	uint32_t i = 0;
2431 
2432 	if (*priv_offset + (args->num_bos * sizeof(*bo_privs)) > max_priv_data_size)
2433 		return -EINVAL;
2434 
2435 	/* Prevent MMU notifications until stage-4 IOCTL (CRIU_RESUME) is received */
2436 	amdgpu_amdkfd_block_mmu_notifications(p->kgd_process_info);
2437 
2438 	bo_buckets = kvmalloc_array(args->num_bos, sizeof(*bo_buckets), GFP_KERNEL);
2439 	if (!bo_buckets)
2440 		return -ENOMEM;
2441 
2442 	files = kvzalloc(args->num_bos * sizeof(struct file *), GFP_KERNEL);
2443 	if (!files) {
2444 		ret = -ENOMEM;
2445 		goto exit;
2446 	}
2447 
2448 	ret = copy_from_user(bo_buckets, (void __user *)args->bos,
2449 			     args->num_bos * sizeof(*bo_buckets));
2450 	if (ret) {
2451 		pr_err("Failed to copy BOs information from user\n");
2452 		ret = -EFAULT;
2453 		goto exit;
2454 	}
2455 
2456 	bo_privs = kvmalloc_array(args->num_bos, sizeof(*bo_privs), GFP_KERNEL);
2457 	if (!bo_privs) {
2458 		ret = -ENOMEM;
2459 		goto exit;
2460 	}
2461 
2462 	ret = copy_from_user(bo_privs, (void __user *)args->priv_data + *priv_offset,
2463 			     args->num_bos * sizeof(*bo_privs));
2464 	if (ret) {
2465 		pr_err("Failed to copy BOs information from user\n");
2466 		ret = -EFAULT;
2467 		goto exit;
2468 	}
2469 	*priv_offset += args->num_bos * sizeof(*bo_privs);
2470 
2471 	/* Create and map new BOs */
2472 	for (; i < args->num_bos; i++) {
2473 		ret = criu_restore_bo(p, &bo_buckets[i], &bo_privs[i], &files[i]);
2474 		if (ret) {
2475 			pr_debug("Failed to restore BO[%d] ret%d\n", i, ret);
2476 			goto exit;
2477 		}
2478 	} /* done */
2479 
2480 	/* Copy only the buckets back so user can read bo_buckets[N].restored_offset */
2481 	ret = copy_to_user((void __user *)args->bos,
2482 				bo_buckets,
2483 				(args->num_bos * sizeof(*bo_buckets)));
2484 	if (ret)
2485 		ret = -EFAULT;
2486 
2487 exit:
2488 	commit_files(files, bo_buckets, i, ret);
2489 	kvfree(files);
2490 	kvfree(bo_buckets);
2491 	kvfree(bo_privs);
2492 	return ret;
2493 }
2494 
2495 static int criu_restore_objects(struct file *filep,
2496 				struct kfd_process *p,
2497 				struct kfd_ioctl_criu_args *args,
2498 				uint64_t *priv_offset,
2499 				uint64_t max_priv_data_size)
2500 {
2501 	int ret = 0;
2502 	uint32_t i;
2503 
2504 	BUILD_BUG_ON(offsetof(struct kfd_criu_queue_priv_data, object_type));
2505 	BUILD_BUG_ON(offsetof(struct kfd_criu_event_priv_data, object_type));
2506 	BUILD_BUG_ON(offsetof(struct kfd_criu_svm_range_priv_data, object_type));
2507 
2508 	for (i = 0; i < args->num_objects; i++) {
2509 		uint32_t object_type;
2510 
2511 		if (*priv_offset + sizeof(object_type) > max_priv_data_size) {
2512 			pr_err("Invalid private data size\n");
2513 			return -EINVAL;
2514 		}
2515 
2516 		ret = get_user(object_type, (uint32_t __user *)(args->priv_data + *priv_offset));
2517 		if (ret) {
2518 			pr_err("Failed to copy private information from user\n");
2519 			goto exit;
2520 		}
2521 
2522 		switch (object_type) {
2523 		case KFD_CRIU_OBJECT_TYPE_QUEUE:
2524 			ret = kfd_criu_restore_queue(p, (uint8_t __user *)args->priv_data,
2525 						     priv_offset, max_priv_data_size);
2526 			if (ret)
2527 				goto exit;
2528 			break;
2529 		case KFD_CRIU_OBJECT_TYPE_EVENT:
2530 			ret = kfd_criu_restore_event(filep, p, (uint8_t __user *)args->priv_data,
2531 						     priv_offset, max_priv_data_size);
2532 			if (ret)
2533 				goto exit;
2534 			break;
2535 		case KFD_CRIU_OBJECT_TYPE_SVM_RANGE:
2536 			ret = kfd_criu_restore_svm(p, (uint8_t __user *)args->priv_data,
2537 						     priv_offset, max_priv_data_size);
2538 			if (ret)
2539 				goto exit;
2540 			break;
2541 		default:
2542 			pr_err("Invalid object type:%u at index:%d\n", object_type, i);
2543 			ret = -EINVAL;
2544 			goto exit;
2545 		}
2546 	}
2547 exit:
2548 	return ret;
2549 }
2550 
2551 static int criu_restore(struct file *filep,
2552 			struct kfd_process *p,
2553 			struct kfd_ioctl_criu_args *args)
2554 {
2555 	uint64_t priv_offset = 0;
2556 	int ret = 0;
2557 
2558 	pr_debug("CRIU restore (num_devices:%u num_bos:%u num_objects:%u priv_data_size:%llu)\n",
2559 		 args->num_devices, args->num_bos, args->num_objects, args->priv_data_size);
2560 
2561 	if (!args->bos || !args->devices || !args->priv_data || !args->priv_data_size ||
2562 	    !args->num_devices || !args->num_bos)
2563 		return -EINVAL;
2564 
2565 	mutex_lock(&p->mutex);
2566 
2567 	/*
2568 	 * Set the process to evicted state to avoid running any new queues before all the memory
2569 	 * mappings are ready.
2570 	 */
2571 	ret = kfd_process_evict_queues(p, KFD_QUEUE_EVICTION_CRIU_RESTORE);
2572 	if (ret)
2573 		goto exit_unlock;
2574 
2575 	/* Each function will adjust priv_offset based on how many bytes they consumed */
2576 	ret = criu_restore_process(p, args, &priv_offset, args->priv_data_size);
2577 	if (ret)
2578 		goto exit_unlock;
2579 
2580 	ret = criu_restore_devices(p, args, &priv_offset, args->priv_data_size);
2581 	if (ret)
2582 		goto exit_unlock;
2583 
2584 	ret = criu_restore_bos(p, args, &priv_offset, args->priv_data_size);
2585 	if (ret)
2586 		goto exit_unlock;
2587 
2588 	ret = criu_restore_objects(filep, p, args, &priv_offset, args->priv_data_size);
2589 	if (ret)
2590 		goto exit_unlock;
2591 
2592 	if (priv_offset != args->priv_data_size) {
2593 		pr_err("Invalid private data size\n");
2594 		ret = -EINVAL;
2595 	}
2596 
2597 exit_unlock:
2598 	mutex_unlock(&p->mutex);
2599 	if (ret)
2600 		pr_err("Failed to restore CRIU ret:%d\n", ret);
2601 	else
2602 		pr_debug("CRIU restore successful\n");
2603 
2604 	return ret;
2605 }
2606 
2607 static int criu_unpause(struct file *filep,
2608 			struct kfd_process *p,
2609 			struct kfd_ioctl_criu_args *args)
2610 {
2611 	int ret;
2612 
2613 	mutex_lock(&p->mutex);
2614 
2615 	if (!p->queues_paused) {
2616 		mutex_unlock(&p->mutex);
2617 		return -EINVAL;
2618 	}
2619 
2620 	ret = kfd_process_restore_queues(p);
2621 	if (ret)
2622 		pr_err("Failed to unpause queues ret:%d\n", ret);
2623 	else
2624 		p->queues_paused = false;
2625 
2626 	mutex_unlock(&p->mutex);
2627 
2628 	return ret;
2629 }
2630 
2631 static int criu_resume(struct file *filep,
2632 			struct kfd_process *p,
2633 			struct kfd_ioctl_criu_args *args)
2634 {
2635 	struct kfd_process *target = NULL;
2636 	struct pid *pid = NULL;
2637 	int ret = 0;
2638 
2639 	pr_debug("Inside %s, target pid for criu restore: %d\n", __func__,
2640 		 args->pid);
2641 
2642 	pid = find_get_pid(args->pid);
2643 	if (!pid) {
2644 		pr_err("Cannot find pid info for %i\n", args->pid);
2645 		return -ESRCH;
2646 	}
2647 
2648 	pr_debug("calling kfd_lookup_process_by_pid\n");
2649 	target = kfd_lookup_process_by_pid(pid);
2650 
2651 	put_pid(pid);
2652 
2653 	if (!target) {
2654 		pr_debug("Cannot find process info for %i\n", args->pid);
2655 		return -ESRCH;
2656 	}
2657 
2658 	mutex_lock(&target->mutex);
2659 	ret = kfd_criu_resume_svm(target);
2660 	if (ret) {
2661 		pr_err("kfd_criu_resume_svm failed for %i\n", args->pid);
2662 		goto exit;
2663 	}
2664 
2665 	ret =  amdgpu_amdkfd_criu_resume(target->kgd_process_info);
2666 	if (ret)
2667 		pr_err("amdgpu_amdkfd_criu_resume failed for %i\n", args->pid);
2668 
2669 exit:
2670 	mutex_unlock(&target->mutex);
2671 
2672 	kfd_unref_process(target);
2673 	return ret;
2674 }
2675 
2676 static int criu_process_info(struct file *filep,
2677 				struct kfd_process *p,
2678 				struct kfd_ioctl_criu_args *args)
2679 {
2680 	int ret = 0;
2681 
2682 	mutex_lock(&p->mutex);
2683 
2684 	if (!p->n_pdds) {
2685 		pr_err("No pdd for given process\n");
2686 		ret = -ENODEV;
2687 		goto err_unlock;
2688 	}
2689 
2690 	ret = kfd_process_evict_queues(p, KFD_QUEUE_EVICTION_CRIU_CHECKPOINT);
2691 	if (ret)
2692 		goto err_unlock;
2693 
2694 	p->queues_paused = true;
2695 
2696 	args->pid = task_pid_nr_ns(p->lead_thread,
2697 					task_active_pid_ns(p->lead_thread));
2698 
2699 	ret = criu_get_process_object_info(p, &args->num_devices, &args->num_bos,
2700 					   &args->num_objects, &args->priv_data_size);
2701 	if (ret)
2702 		goto err_unlock;
2703 
2704 	dev_dbg(kfd_device, "Num of devices:%u bos:%u objects:%u priv_data_size:%lld\n",
2705 				args->num_devices, args->num_bos, args->num_objects,
2706 				args->priv_data_size);
2707 
2708 err_unlock:
2709 	if (ret) {
2710 		kfd_process_restore_queues(p);
2711 		p->queues_paused = false;
2712 	}
2713 	mutex_unlock(&p->mutex);
2714 	return ret;
2715 }
2716 
2717 static int kfd_ioctl_criu(struct file *filep, struct kfd_process *p, void *data)
2718 {
2719 	struct kfd_ioctl_criu_args *args = data;
2720 	int ret;
2721 
2722 	dev_dbg(kfd_device, "CRIU operation: %d\n", args->op);
2723 	switch (args->op) {
2724 	case KFD_CRIU_OP_PROCESS_INFO:
2725 		ret = criu_process_info(filep, p, args);
2726 		break;
2727 	case KFD_CRIU_OP_CHECKPOINT:
2728 		ret = criu_checkpoint(filep, p, args);
2729 		break;
2730 	case KFD_CRIU_OP_UNPAUSE:
2731 		ret = criu_unpause(filep, p, args);
2732 		break;
2733 	case KFD_CRIU_OP_RESTORE:
2734 		ret = criu_restore(filep, p, args);
2735 		break;
2736 	case KFD_CRIU_OP_RESUME:
2737 		ret = criu_resume(filep, p, args);
2738 		break;
2739 	default:
2740 		dev_dbg(kfd_device, "Unsupported CRIU operation:%d\n", args->op);
2741 		ret = -EINVAL;
2742 		break;
2743 	}
2744 
2745 	if (ret)
2746 		dev_dbg(kfd_device, "CRIU operation:%d err:%d\n", args->op, ret);
2747 
2748 	return ret;
2749 }
2750 
2751 static int runtime_enable(struct kfd_process *p, uint64_t r_debug,
2752 			bool enable_ttmp_setup)
2753 {
2754 	int i = 0, ret = 0;
2755 
2756 	if (p->is_runtime_retry)
2757 		goto retry;
2758 
2759 	if (p->runtime_info.runtime_state != DEBUG_RUNTIME_STATE_DISABLED)
2760 		return -EBUSY;
2761 
2762 	for (i = 0; i < p->n_pdds; i++) {
2763 		struct kfd_process_device *pdd = p->pdds[i];
2764 
2765 		if (pdd->qpd.queue_count)
2766 			return -EEXIST;
2767 
2768 		/*
2769 		 * Setup TTMPs by default.
2770 		 * Note that this call must remain here for MES ADD QUEUE to
2771 		 * skip_process_ctx_clear unconditionally as the first call to
2772 		 * SET_SHADER_DEBUGGER clears any stale process context data
2773 		 * saved in MES.
2774 		 */
2775 		if (pdd->dev->kfd->shared_resources.enable_mes)
2776 			kfd_dbg_set_mes_debug_mode(pdd, !kfd_dbg_has_cwsr_workaround(pdd->dev));
2777 	}
2778 
2779 	p->runtime_info.runtime_state = DEBUG_RUNTIME_STATE_ENABLED;
2780 	p->runtime_info.r_debug = r_debug;
2781 	p->runtime_info.ttmp_setup = enable_ttmp_setup;
2782 
2783 	if (p->runtime_info.ttmp_setup) {
2784 		for (i = 0; i < p->n_pdds; i++) {
2785 			struct kfd_process_device *pdd = p->pdds[i];
2786 
2787 			if (!kfd_dbg_is_rlc_restore_supported(pdd->dev)) {
2788 				amdgpu_gfx_off_ctrl(pdd->dev->adev, false);
2789 				pdd->dev->kfd2kgd->enable_debug_trap(
2790 						pdd->dev->adev,
2791 						true,
2792 						pdd->dev->vm_info.last_vmid_kfd);
2793 			} else if (kfd_dbg_is_per_vmid_supported(pdd->dev)) {
2794 				pdd->spi_dbg_override = pdd->dev->kfd2kgd->enable_debug_trap(
2795 						pdd->dev->adev,
2796 						false,
2797 						0);
2798 			}
2799 		}
2800 	}
2801 
2802 retry:
2803 	if (p->debug_trap_enabled) {
2804 		if (!p->is_runtime_retry) {
2805 			kfd_dbg_trap_activate(p);
2806 			kfd_dbg_ev_raise(KFD_EC_MASK(EC_PROCESS_RUNTIME),
2807 					p, NULL, 0, false, NULL, 0);
2808 		}
2809 
2810 		mutex_unlock(&p->mutex);
2811 		ret = down_interruptible(&p->runtime_enable_sema);
2812 		mutex_lock(&p->mutex);
2813 
2814 		p->is_runtime_retry = !!ret;
2815 	}
2816 
2817 	return ret;
2818 }
2819 
2820 static int runtime_disable(struct kfd_process *p)
2821 {
2822 	int i = 0, ret;
2823 	bool was_enabled = p->runtime_info.runtime_state == DEBUG_RUNTIME_STATE_ENABLED;
2824 
2825 	p->runtime_info.runtime_state = DEBUG_RUNTIME_STATE_DISABLED;
2826 	p->runtime_info.r_debug = 0;
2827 
2828 	if (p->debug_trap_enabled) {
2829 		if (was_enabled)
2830 			kfd_dbg_trap_deactivate(p, false, 0);
2831 
2832 		if (!p->is_runtime_retry)
2833 			kfd_dbg_ev_raise(KFD_EC_MASK(EC_PROCESS_RUNTIME),
2834 					p, NULL, 0, false, NULL, 0);
2835 
2836 		mutex_unlock(&p->mutex);
2837 		ret = down_interruptible(&p->runtime_enable_sema);
2838 		mutex_lock(&p->mutex);
2839 
2840 		p->is_runtime_retry = !!ret;
2841 		if (ret)
2842 			return ret;
2843 	}
2844 
2845 	if (was_enabled && p->runtime_info.ttmp_setup) {
2846 		for (i = 0; i < p->n_pdds; i++) {
2847 			struct kfd_process_device *pdd = p->pdds[i];
2848 
2849 			if (!kfd_dbg_is_rlc_restore_supported(pdd->dev))
2850 				amdgpu_gfx_off_ctrl(pdd->dev->adev, true);
2851 		}
2852 	}
2853 
2854 	p->runtime_info.ttmp_setup = false;
2855 
2856 	/* disable ttmp setup */
2857 	for (i = 0; i < p->n_pdds; i++) {
2858 		struct kfd_process_device *pdd = p->pdds[i];
2859 
2860 		if (kfd_dbg_is_per_vmid_supported(pdd->dev)) {
2861 			pdd->spi_dbg_override =
2862 					pdd->dev->kfd2kgd->disable_debug_trap(
2863 					pdd->dev->adev,
2864 					false,
2865 					pdd->dev->vm_info.last_vmid_kfd);
2866 
2867 			if (!pdd->dev->kfd->shared_resources.enable_mes)
2868 				debug_refresh_runlist(pdd->dev->dqm);
2869 			else
2870 				kfd_dbg_set_mes_debug_mode(pdd,
2871 							   !kfd_dbg_has_cwsr_workaround(pdd->dev));
2872 		}
2873 	}
2874 
2875 	return 0;
2876 }
2877 
2878 static int kfd_ioctl_runtime_enable(struct file *filep, struct kfd_process *p, void *data)
2879 {
2880 	struct kfd_ioctl_runtime_enable_args *args = data;
2881 	int r;
2882 
2883 	mutex_lock(&p->mutex);
2884 
2885 	if (args->mode_mask & KFD_RUNTIME_ENABLE_MODE_ENABLE_MASK)
2886 		r = runtime_enable(p, args->r_debug,
2887 				!!(args->mode_mask & KFD_RUNTIME_ENABLE_MODE_TTMP_SAVE_MASK));
2888 	else
2889 		r = runtime_disable(p);
2890 
2891 	mutex_unlock(&p->mutex);
2892 
2893 	return r;
2894 }
2895 
2896 static int kfd_ioctl_set_debug_trap(struct file *filep, struct kfd_process *p, void *data)
2897 {
2898 	struct kfd_ioctl_dbg_trap_args *args = data;
2899 	struct task_struct *thread = NULL;
2900 	struct mm_struct *mm = NULL;
2901 	struct pid *pid = NULL;
2902 	struct kfd_process *target = NULL;
2903 	struct kfd_process_device *pdd = NULL;
2904 	int r = 0;
2905 
2906 	if (sched_policy == KFD_SCHED_POLICY_NO_HWS) {
2907 		pr_err("Debugging does not support sched_policy %i", sched_policy);
2908 		return -EINVAL;
2909 	}
2910 
2911 	pid = find_get_pid(args->pid);
2912 	if (!pid) {
2913 		pr_debug("Cannot find pid info for %i\n", args->pid);
2914 		r = -ESRCH;
2915 		goto out;
2916 	}
2917 
2918 	thread = get_pid_task(pid, PIDTYPE_PID);
2919 	if (!thread) {
2920 		r = -ESRCH;
2921 		goto out;
2922 	}
2923 
2924 	mm = get_task_mm(thread);
2925 	if (!mm) {
2926 		r = -ESRCH;
2927 		goto out;
2928 	}
2929 
2930 	if (args->op == KFD_IOC_DBG_TRAP_ENABLE) {
2931 		bool create_process;
2932 
2933 		rcu_read_lock();
2934 		create_process = thread && thread != current && ptrace_parent(thread) == current;
2935 		rcu_read_unlock();
2936 
2937 		target = create_process ? kfd_create_process(thread) :
2938 					kfd_lookup_process_by_pid(pid);
2939 	} else {
2940 		target = kfd_lookup_process_by_pid(pid);
2941 	}
2942 
2943 	if (IS_ERR_OR_NULL(target)) {
2944 		pr_debug("Cannot find process PID %i to debug\n", args->pid);
2945 		r = target ? PTR_ERR(target) : -ESRCH;
2946 		target = NULL;
2947 		goto out;
2948 	}
2949 
2950 	/* Check if target is still PTRACED. */
2951 	rcu_read_lock();
2952 	if (target != p && args->op != KFD_IOC_DBG_TRAP_DISABLE
2953 				&& ptrace_parent(target->lead_thread) != current) {
2954 		pr_err("PID %i is not PTRACED and cannot be debugged\n", args->pid);
2955 		r = -EPERM;
2956 	}
2957 	rcu_read_unlock();
2958 
2959 	if (r)
2960 		goto out;
2961 
2962 	mutex_lock(&target->mutex);
2963 
2964 	if (args->op != KFD_IOC_DBG_TRAP_ENABLE && !target->debug_trap_enabled) {
2965 		pr_err("PID %i not debug enabled for op %i\n", args->pid, args->op);
2966 		r = -EINVAL;
2967 		goto unlock_out;
2968 	}
2969 
2970 	if (target->runtime_info.runtime_state != DEBUG_RUNTIME_STATE_ENABLED &&
2971 			(args->op == KFD_IOC_DBG_TRAP_SET_WAVE_LAUNCH_OVERRIDE ||
2972 			 args->op == KFD_IOC_DBG_TRAP_SET_WAVE_LAUNCH_MODE ||
2973 			 args->op == KFD_IOC_DBG_TRAP_SUSPEND_QUEUES ||
2974 			 args->op == KFD_IOC_DBG_TRAP_RESUME_QUEUES ||
2975 			 args->op == KFD_IOC_DBG_TRAP_SET_NODE_ADDRESS_WATCH ||
2976 			 args->op == KFD_IOC_DBG_TRAP_CLEAR_NODE_ADDRESS_WATCH ||
2977 			 args->op == KFD_IOC_DBG_TRAP_SET_FLAGS)) {
2978 		r = -EPERM;
2979 		goto unlock_out;
2980 	}
2981 
2982 	if (args->op == KFD_IOC_DBG_TRAP_SET_NODE_ADDRESS_WATCH ||
2983 	    args->op == KFD_IOC_DBG_TRAP_CLEAR_NODE_ADDRESS_WATCH) {
2984 		int user_gpu_id = kfd_process_get_user_gpu_id(target,
2985 				args->op == KFD_IOC_DBG_TRAP_SET_NODE_ADDRESS_WATCH ?
2986 					args->set_node_address_watch.gpu_id :
2987 					args->clear_node_address_watch.gpu_id);
2988 
2989 		pdd = kfd_process_device_data_by_id(target, user_gpu_id);
2990 		if (user_gpu_id == -EINVAL || !pdd) {
2991 			r = -ENODEV;
2992 			goto unlock_out;
2993 		}
2994 	}
2995 
2996 	switch (args->op) {
2997 	case KFD_IOC_DBG_TRAP_ENABLE:
2998 		if (target != p)
2999 			target->debugger_process = p;
3000 
3001 		r = kfd_dbg_trap_enable(target,
3002 					args->enable.dbg_fd,
3003 					(void __user *)args->enable.rinfo_ptr,
3004 					&args->enable.rinfo_size);
3005 		if (!r)
3006 			target->exception_enable_mask = args->enable.exception_mask;
3007 
3008 		break;
3009 	case KFD_IOC_DBG_TRAP_DISABLE:
3010 		r = kfd_dbg_trap_disable(target);
3011 		break;
3012 	case KFD_IOC_DBG_TRAP_SEND_RUNTIME_EVENT:
3013 		r = kfd_dbg_send_exception_to_runtime(target,
3014 				args->send_runtime_event.gpu_id,
3015 				args->send_runtime_event.queue_id,
3016 				args->send_runtime_event.exception_mask);
3017 		break;
3018 	case KFD_IOC_DBG_TRAP_SET_EXCEPTIONS_ENABLED:
3019 		kfd_dbg_set_enabled_debug_exception_mask(target,
3020 				args->set_exceptions_enabled.exception_mask);
3021 		break;
3022 	case KFD_IOC_DBG_TRAP_SET_WAVE_LAUNCH_OVERRIDE:
3023 		r = kfd_dbg_trap_set_wave_launch_override(target,
3024 				args->launch_override.override_mode,
3025 				args->launch_override.enable_mask,
3026 				args->launch_override.support_request_mask,
3027 				&args->launch_override.enable_mask,
3028 				&args->launch_override.support_request_mask);
3029 		break;
3030 	case KFD_IOC_DBG_TRAP_SET_WAVE_LAUNCH_MODE:
3031 		r = kfd_dbg_trap_set_wave_launch_mode(target,
3032 				args->launch_mode.launch_mode);
3033 		break;
3034 	case KFD_IOC_DBG_TRAP_SUSPEND_QUEUES:
3035 		r = suspend_queues(target,
3036 				args->suspend_queues.num_queues,
3037 				args->suspend_queues.grace_period,
3038 				args->suspend_queues.exception_mask,
3039 				(uint32_t *)args->suspend_queues.queue_array_ptr);
3040 
3041 		break;
3042 	case KFD_IOC_DBG_TRAP_RESUME_QUEUES:
3043 		r = resume_queues(target, args->resume_queues.num_queues,
3044 				(uint32_t *)args->resume_queues.queue_array_ptr);
3045 		break;
3046 	case KFD_IOC_DBG_TRAP_SET_NODE_ADDRESS_WATCH:
3047 		r = kfd_dbg_trap_set_dev_address_watch(pdd,
3048 				args->set_node_address_watch.address,
3049 				args->set_node_address_watch.mask,
3050 				&args->set_node_address_watch.id,
3051 				args->set_node_address_watch.mode);
3052 		break;
3053 	case KFD_IOC_DBG_TRAP_CLEAR_NODE_ADDRESS_WATCH:
3054 		r = kfd_dbg_trap_clear_dev_address_watch(pdd,
3055 				args->clear_node_address_watch.id);
3056 		break;
3057 	case KFD_IOC_DBG_TRAP_SET_FLAGS:
3058 		r = kfd_dbg_trap_set_flags(target, &args->set_flags.flags);
3059 		break;
3060 	case KFD_IOC_DBG_TRAP_QUERY_DEBUG_EVENT:
3061 		r = kfd_dbg_ev_query_debug_event(target,
3062 				&args->query_debug_event.queue_id,
3063 				&args->query_debug_event.gpu_id,
3064 				args->query_debug_event.exception_mask,
3065 				&args->query_debug_event.exception_mask);
3066 		break;
3067 	case KFD_IOC_DBG_TRAP_QUERY_EXCEPTION_INFO:
3068 		r = kfd_dbg_trap_query_exception_info(target,
3069 				args->query_exception_info.source_id,
3070 				args->query_exception_info.exception_code,
3071 				args->query_exception_info.clear_exception,
3072 				(void __user *)args->query_exception_info.info_ptr,
3073 				&args->query_exception_info.info_size);
3074 		break;
3075 	case KFD_IOC_DBG_TRAP_GET_QUEUE_SNAPSHOT:
3076 		r = pqm_get_queue_snapshot(&target->pqm,
3077 				args->queue_snapshot.exception_mask,
3078 				(void __user *)args->queue_snapshot.snapshot_buf_ptr,
3079 				&args->queue_snapshot.num_queues,
3080 				&args->queue_snapshot.entry_size);
3081 		break;
3082 	case KFD_IOC_DBG_TRAP_GET_DEVICE_SNAPSHOT:
3083 		r = kfd_dbg_trap_device_snapshot(target,
3084 				args->device_snapshot.exception_mask,
3085 				(void __user *)args->device_snapshot.snapshot_buf_ptr,
3086 				&args->device_snapshot.num_devices,
3087 				&args->device_snapshot.entry_size);
3088 		break;
3089 	default:
3090 		pr_err("Invalid option: %i\n", args->op);
3091 		r = -EINVAL;
3092 	}
3093 
3094 unlock_out:
3095 	mutex_unlock(&target->mutex);
3096 
3097 out:
3098 	if (thread)
3099 		put_task_struct(thread);
3100 
3101 	if (mm)
3102 		mmput(mm);
3103 
3104 	if (pid)
3105 		put_pid(pid);
3106 
3107 	if (target)
3108 		kfd_unref_process(target);
3109 
3110 	return r;
3111 }
3112 
3113 #define AMDKFD_IOCTL_DEF(ioctl, _func, _flags) \
3114 	[_IOC_NR(ioctl)] = {.cmd = ioctl, .func = _func, .flags = _flags, \
3115 			    .cmd_drv = 0, .name = #ioctl}
3116 
3117 /** Ioctl table */
3118 static const struct amdkfd_ioctl_desc amdkfd_ioctls[] = {
3119 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_GET_VERSION,
3120 			kfd_ioctl_get_version, 0),
3121 
3122 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_CREATE_QUEUE,
3123 			kfd_ioctl_create_queue, 0),
3124 
3125 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_DESTROY_QUEUE,
3126 			kfd_ioctl_destroy_queue, 0),
3127 
3128 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_SET_MEMORY_POLICY,
3129 			kfd_ioctl_set_memory_policy, 0),
3130 
3131 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_GET_CLOCK_COUNTERS,
3132 			kfd_ioctl_get_clock_counters, 0),
3133 
3134 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_GET_PROCESS_APERTURES,
3135 			kfd_ioctl_get_process_apertures, 0),
3136 
3137 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_UPDATE_QUEUE,
3138 			kfd_ioctl_update_queue, 0),
3139 
3140 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_CREATE_EVENT,
3141 			kfd_ioctl_create_event, 0),
3142 
3143 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_DESTROY_EVENT,
3144 			kfd_ioctl_destroy_event, 0),
3145 
3146 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_SET_EVENT,
3147 			kfd_ioctl_set_event, 0),
3148 
3149 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_RESET_EVENT,
3150 			kfd_ioctl_reset_event, 0),
3151 
3152 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_WAIT_EVENTS,
3153 			kfd_ioctl_wait_events, 0),
3154 
3155 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_DBG_REGISTER_DEPRECATED,
3156 			kfd_ioctl_dbg_register, 0),
3157 
3158 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_DBG_UNREGISTER_DEPRECATED,
3159 			kfd_ioctl_dbg_unregister, 0),
3160 
3161 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_DBG_ADDRESS_WATCH_DEPRECATED,
3162 			kfd_ioctl_dbg_address_watch, 0),
3163 
3164 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_DBG_WAVE_CONTROL_DEPRECATED,
3165 			kfd_ioctl_dbg_wave_control, 0),
3166 
3167 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_SET_SCRATCH_BACKING_VA,
3168 			kfd_ioctl_set_scratch_backing_va, 0),
3169 
3170 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_GET_TILE_CONFIG,
3171 			kfd_ioctl_get_tile_config, 0),
3172 
3173 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_SET_TRAP_HANDLER,
3174 			kfd_ioctl_set_trap_handler, 0),
3175 
3176 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_GET_PROCESS_APERTURES_NEW,
3177 			kfd_ioctl_get_process_apertures_new, 0),
3178 
3179 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_ACQUIRE_VM,
3180 			kfd_ioctl_acquire_vm, 0),
3181 
3182 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_ALLOC_MEMORY_OF_GPU,
3183 			kfd_ioctl_alloc_memory_of_gpu, 0),
3184 
3185 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_FREE_MEMORY_OF_GPU,
3186 			kfd_ioctl_free_memory_of_gpu, 0),
3187 
3188 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_MAP_MEMORY_TO_GPU,
3189 			kfd_ioctl_map_memory_to_gpu, 0),
3190 
3191 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_UNMAP_MEMORY_FROM_GPU,
3192 			kfd_ioctl_unmap_memory_from_gpu, 0),
3193 
3194 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_SET_CU_MASK,
3195 			kfd_ioctl_set_cu_mask, 0),
3196 
3197 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_GET_QUEUE_WAVE_STATE,
3198 			kfd_ioctl_get_queue_wave_state, 0),
3199 
3200 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_GET_DMABUF_INFO,
3201 				kfd_ioctl_get_dmabuf_info, 0),
3202 
3203 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_IMPORT_DMABUF,
3204 				kfd_ioctl_import_dmabuf, 0),
3205 
3206 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_ALLOC_QUEUE_GWS,
3207 			kfd_ioctl_alloc_queue_gws, 0),
3208 
3209 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_SMI_EVENTS,
3210 			kfd_ioctl_smi_events, 0),
3211 
3212 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_SVM, kfd_ioctl_svm, 0),
3213 
3214 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_SET_XNACK_MODE,
3215 			kfd_ioctl_set_xnack_mode, 0),
3216 
3217 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_CRIU_OP,
3218 			kfd_ioctl_criu, KFD_IOC_FLAG_CHECKPOINT_RESTORE),
3219 
3220 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_AVAILABLE_MEMORY,
3221 			kfd_ioctl_get_available_memory, 0),
3222 
3223 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_EXPORT_DMABUF,
3224 				kfd_ioctl_export_dmabuf, 0),
3225 
3226 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_RUNTIME_ENABLE,
3227 			kfd_ioctl_runtime_enable, 0),
3228 
3229 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_DBG_TRAP,
3230 			kfd_ioctl_set_debug_trap, 0),
3231 };
3232 
3233 #define AMDKFD_CORE_IOCTL_COUNT	ARRAY_SIZE(amdkfd_ioctls)
3234 
3235 static long kfd_ioctl(struct file *filep, unsigned int cmd, unsigned long arg)
3236 {
3237 	struct kfd_process *process;
3238 	amdkfd_ioctl_t *func;
3239 	const struct amdkfd_ioctl_desc *ioctl = NULL;
3240 	unsigned int nr = _IOC_NR(cmd);
3241 	char stack_kdata[128];
3242 	char *kdata = NULL;
3243 	unsigned int usize, asize;
3244 	int retcode = -EINVAL;
3245 	bool ptrace_attached = false;
3246 
3247 	if (nr >= AMDKFD_CORE_IOCTL_COUNT)
3248 		goto err_i1;
3249 
3250 	if ((nr >= AMDKFD_COMMAND_START) && (nr < AMDKFD_COMMAND_END)) {
3251 		u32 amdkfd_size;
3252 
3253 		ioctl = &amdkfd_ioctls[nr];
3254 
3255 		amdkfd_size = _IOC_SIZE(ioctl->cmd);
3256 		usize = asize = _IOC_SIZE(cmd);
3257 		if (amdkfd_size > asize)
3258 			asize = amdkfd_size;
3259 
3260 		cmd = ioctl->cmd;
3261 	} else
3262 		goto err_i1;
3263 
3264 	dev_dbg(kfd_device, "ioctl cmd 0x%x (#0x%x), arg 0x%lx\n", cmd, nr, arg);
3265 
3266 	/* Get the process struct from the filep. Only the process
3267 	 * that opened /dev/kfd can use the file descriptor. Child
3268 	 * processes need to create their own KFD device context.
3269 	 */
3270 	process = filep->private_data;
3271 
3272 	rcu_read_lock();
3273 	if ((ioctl->flags & KFD_IOC_FLAG_CHECKPOINT_RESTORE) &&
3274 	    ptrace_parent(process->lead_thread) == current)
3275 		ptrace_attached = true;
3276 	rcu_read_unlock();
3277 
3278 	if (process->lead_thread != current->group_leader
3279 	    && !ptrace_attached) {
3280 		dev_dbg(kfd_device, "Using KFD FD in wrong process\n");
3281 		retcode = -EBADF;
3282 		goto err_i1;
3283 	}
3284 
3285 	/* Do not trust userspace, use our own definition */
3286 	func = ioctl->func;
3287 
3288 	if (unlikely(!func)) {
3289 		dev_dbg(kfd_device, "no function\n");
3290 		retcode = -EINVAL;
3291 		goto err_i1;
3292 	}
3293 
3294 	/*
3295 	 * Versions of docker shipped in Ubuntu 18.xx and 20.xx do not support
3296 	 * CAP_CHECKPOINT_RESTORE, so we also allow access if CAP_SYS_ADMIN as CAP_SYS_ADMIN is a
3297 	 * more priviledged access.
3298 	 */
3299 	if (unlikely(ioctl->flags & KFD_IOC_FLAG_CHECKPOINT_RESTORE)) {
3300 		if (!capable(CAP_CHECKPOINT_RESTORE) &&
3301 						!capable(CAP_SYS_ADMIN)) {
3302 			retcode = -EACCES;
3303 			goto err_i1;
3304 		}
3305 	}
3306 
3307 	if (cmd & (IOC_IN | IOC_OUT)) {
3308 		if (asize <= sizeof(stack_kdata)) {
3309 			kdata = stack_kdata;
3310 		} else {
3311 			kdata = kmalloc(asize, GFP_KERNEL);
3312 			if (!kdata) {
3313 				retcode = -ENOMEM;
3314 				goto err_i1;
3315 			}
3316 		}
3317 		if (asize > usize)
3318 			memset(kdata + usize, 0, asize - usize);
3319 	}
3320 
3321 	if (cmd & IOC_IN) {
3322 		if (copy_from_user(kdata, (void __user *)arg, usize) != 0) {
3323 			retcode = -EFAULT;
3324 			goto err_i1;
3325 		}
3326 	} else if (cmd & IOC_OUT) {
3327 		memset(kdata, 0, usize);
3328 	}
3329 
3330 	retcode = func(filep, process, kdata);
3331 
3332 	if (cmd & IOC_OUT)
3333 		if (copy_to_user((void __user *)arg, kdata, usize) != 0)
3334 			retcode = -EFAULT;
3335 
3336 err_i1:
3337 	if (!ioctl)
3338 		dev_dbg(kfd_device, "invalid ioctl: pid=%d, cmd=0x%02x, nr=0x%02x\n",
3339 			  task_pid_nr(current), cmd, nr);
3340 
3341 	if (kdata != stack_kdata)
3342 		kfree(kdata);
3343 
3344 	if (retcode)
3345 		dev_dbg(kfd_device, "ioctl cmd (#0x%x), arg 0x%lx, ret = %d\n",
3346 				nr, arg, retcode);
3347 
3348 	return retcode;
3349 }
3350 
3351 static int kfd_mmio_mmap(struct kfd_node *dev, struct kfd_process *process,
3352 		      struct vm_area_struct *vma)
3353 {
3354 	phys_addr_t address;
3355 
3356 	if (vma->vm_end - vma->vm_start != PAGE_SIZE)
3357 		return -EINVAL;
3358 
3359 	if (PAGE_SIZE > 4096)
3360 		return -EINVAL;
3361 
3362 	address = dev->adev->rmmio_remap.bus_addr;
3363 
3364 	vm_flags_set(vma, VM_IO | VM_DONTCOPY | VM_DONTEXPAND | VM_NORESERVE |
3365 				VM_DONTDUMP | VM_PFNMAP);
3366 
3367 	vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot);
3368 
3369 	pr_debug("process pid %d mapping mmio page\n"
3370 		 "     target user address == 0x%08llX\n"
3371 		 "     physical address    == 0x%08llX\n"
3372 		 "     vm_flags            == 0x%04lX\n"
3373 		 "     size                == 0x%04lX\n",
3374 		 process->lead_thread->pid, (unsigned long long) vma->vm_start,
3375 		 address, vma->vm_flags, PAGE_SIZE);
3376 
3377 	return io_remap_pfn_range(vma,
3378 				vma->vm_start,
3379 				address >> PAGE_SHIFT,
3380 				PAGE_SIZE,
3381 				vma->vm_page_prot);
3382 }
3383 
3384 
3385 static int kfd_mmap(struct file *filp, struct vm_area_struct *vma)
3386 {
3387 	struct kfd_process *process;
3388 	struct kfd_node *dev = NULL;
3389 	unsigned long mmap_offset;
3390 	unsigned int gpu_id;
3391 
3392 	process = kfd_get_process(current);
3393 	if (IS_ERR(process))
3394 		return PTR_ERR(process);
3395 
3396 	mmap_offset = vma->vm_pgoff << PAGE_SHIFT;
3397 	gpu_id = KFD_MMAP_GET_GPU_ID(mmap_offset);
3398 	if (gpu_id)
3399 		dev = kfd_device_by_id(gpu_id);
3400 
3401 	switch (mmap_offset & KFD_MMAP_TYPE_MASK) {
3402 	case KFD_MMAP_TYPE_DOORBELL:
3403 		if (!dev)
3404 			return -ENODEV;
3405 		return kfd_doorbell_mmap(dev, process, vma);
3406 
3407 	case KFD_MMAP_TYPE_EVENTS:
3408 		return kfd_event_mmap(process, vma);
3409 
3410 	case KFD_MMAP_TYPE_RESERVED_MEM:
3411 		if (!dev)
3412 			return -ENODEV;
3413 		return kfd_reserved_mem_mmap(dev, process, vma);
3414 	case KFD_MMAP_TYPE_MMIO:
3415 		if (!dev)
3416 			return -ENODEV;
3417 		return kfd_mmio_mmap(dev, process, vma);
3418 	}
3419 
3420 	return -EFAULT;
3421 }
3422