xref: /linux/drivers/gpu/drm/amd/amdkfd/kfd_chardev.c (revision 6f17ab9a63e670bd62a287f95e3982f99eafd77e)
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/err.h>
26 #include <linux/fs.h>
27 #include <linux/file.h>
28 #include <linux/sched.h>
29 #include <linux/slab.h>
30 #include <linux/uaccess.h>
31 #include <linux/compat.h>
32 #include <uapi/linux/kfd_ioctl.h>
33 #include <linux/time.h>
34 #include <linux/mm.h>
35 #include <linux/mman.h>
36 #include <linux/ptrace.h>
37 #include <linux/dma-buf.h>
38 #include <linux/processor.h>
39 #include "kfd_priv.h"
40 #include "kfd_device_queue_manager.h"
41 #include "kfd_svm.h"
42 #include "amdgpu_amdkfd.h"
43 #include "kfd_smi_events.h"
44 #include "amdgpu_dma_buf.h"
45 #include "kfd_debug.h"
46 
47 static long kfd_ioctl(struct file *, unsigned int, unsigned long);
48 static int kfd_open(struct inode *, struct file *);
49 static int kfd_release(struct inode *, struct file *);
50 static int kfd_mmap(struct file *, struct vm_area_struct *);
51 
52 static const char kfd_dev_name[] = "kfd";
53 
54 static const struct file_operations kfd_fops = {
55 	.owner = THIS_MODULE,
56 	.unlocked_ioctl = kfd_ioctl,
57 	.compat_ioctl = compat_ptr_ioctl,
58 	.open = kfd_open,
59 	.release = kfd_release,
60 	.mmap = kfd_mmap,
61 };
62 
63 static int kfd_char_dev_major = -1;
64 struct device *kfd_device;
65 static const struct class kfd_class = {
66 	.name = kfd_dev_name,
67 };
68 
69 static inline struct kfd_process_device *kfd_lock_pdd_by_id(struct kfd_process *p, __u32 gpu_id)
70 {
71 	struct kfd_process_device *pdd;
72 
73 	mutex_lock(&p->mutex);
74 	pdd = kfd_process_device_data_by_id(p, gpu_id);
75 
76 	if (pdd)
77 		return pdd;
78 
79 	mutex_unlock(&p->mutex);
80 	return NULL;
81 }
82 
83 static inline void kfd_unlock_pdd(struct kfd_process_device *pdd)
84 {
85 	mutex_unlock(&pdd->process->mutex);
86 }
87 
88 int kfd_chardev_init(void)
89 {
90 	int err = 0;
91 
92 	kfd_char_dev_major = register_chrdev(0, kfd_dev_name, &kfd_fops);
93 	err = kfd_char_dev_major;
94 	if (err < 0)
95 		goto err_register_chrdev;
96 
97 	err = class_register(&kfd_class);
98 	if (err)
99 		goto err_class_create;
100 
101 	kfd_device = device_create(&kfd_class, NULL,
102 				   MKDEV(kfd_char_dev_major, 0),
103 				   NULL, kfd_dev_name);
104 	err = PTR_ERR(kfd_device);
105 	if (IS_ERR(kfd_device))
106 		goto err_device_create;
107 
108 	return 0;
109 
110 err_device_create:
111 	class_unregister(&kfd_class);
112 err_class_create:
113 	unregister_chrdev(kfd_char_dev_major, kfd_dev_name);
114 err_register_chrdev:
115 	return err;
116 }
117 
118 void kfd_chardev_exit(void)
119 {
120 	device_destroy(&kfd_class, MKDEV(kfd_char_dev_major, 0));
121 	class_unregister(&kfd_class);
122 	unregister_chrdev(kfd_char_dev_major, kfd_dev_name);
123 	kfd_device = NULL;
124 }
125 
126 
127 static int kfd_open(struct inode *inode, struct file *filep)
128 {
129 	struct kfd_process *process;
130 	bool is_32bit_user_mode;
131 
132 	if (iminor(inode) != 0)
133 		return -ENODEV;
134 
135 	is_32bit_user_mode = in_compat_syscall();
136 
137 	if (is_32bit_user_mode) {
138 		dev_warn(kfd_device,
139 			"Process %d (32-bit) failed to open /dev/kfd\n"
140 			"32-bit processes are not supported by amdkfd\n",
141 			current->pid);
142 		return -EPERM;
143 	}
144 
145 	process = kfd_create_process(current);
146 	if (IS_ERR(process))
147 		return PTR_ERR(process);
148 
149 	if (kfd_process_init_cwsr_apu(process, filep)) {
150 		kfd_unref_process(process);
151 		return -EFAULT;
152 	}
153 
154 	/* filep now owns the reference returned by kfd_create_process */
155 	filep->private_data = process;
156 
157 	dev_dbg(kfd_device, "process pid %d opened kfd node, compat mode (32 bit) - %d\n",
158 		process->lead_thread->pid, process->is_32bit_user_mode);
159 
160 	return 0;
161 }
162 
163 static int kfd_release(struct inode *inode, struct file *filep)
164 {
165 	struct kfd_process *process = filep->private_data;
166 
167 	if (process)
168 		kfd_unref_process(process);
169 
170 	return 0;
171 }
172 
173 static int kfd_ioctl_get_version(struct file *filep, struct kfd_process *p,
174 					void *data)
175 {
176 	struct kfd_ioctl_get_version_args *args = data;
177 
178 	args->major_version = KFD_IOCTL_MAJOR_VERSION;
179 	args->minor_version = KFD_IOCTL_MINOR_VERSION;
180 
181 	return 0;
182 }
183 
184 static int set_queue_properties_from_user(struct queue_properties *q_properties,
185 				struct kfd_ioctl_create_queue_args *args)
186 {
187 	/*
188 	 * Repurpose queue percentage to accommodate new features:
189 	 * bit 0-7: queue percentage
190 	 * bit 8-15: pm4_target_xcc
191 	 */
192 	if ((args->queue_percentage & 0xFF) > KFD_MAX_QUEUE_PERCENTAGE) {
193 		pr_err("Queue percentage must be between 0 to KFD_MAX_QUEUE_PERCENTAGE\n");
194 		return -EINVAL;
195 	}
196 
197 	if (args->queue_priority > KFD_MAX_QUEUE_PRIORITY) {
198 		pr_err("Queue priority must be between 0 to KFD_MAX_QUEUE_PRIORITY\n");
199 		return -EINVAL;
200 	}
201 
202 	if ((args->ring_base_address) &&
203 		(!access_ok((const void __user *) args->ring_base_address,
204 			sizeof(uint64_t)))) {
205 		pr_err("Can't access ring base address\n");
206 		return -EFAULT;
207 	}
208 
209 	if (!is_power_of_2(args->ring_size) && (args->ring_size != 0)) {
210 		pr_err("Ring size must be a power of 2 or 0\n");
211 		return -EINVAL;
212 	}
213 
214 	if (args->ring_size < KFD_MIN_QUEUE_RING_SIZE) {
215 		args->ring_size = KFD_MIN_QUEUE_RING_SIZE;
216 		pr_debug("Size lower. clamped to KFD_MIN_QUEUE_RING_SIZE");
217 	}
218 
219 	if (!access_ok((const void __user *) args->read_pointer_address,
220 			sizeof(uint32_t))) {
221 		pr_err("Can't access read pointer\n");
222 		return -EFAULT;
223 	}
224 
225 	if (!access_ok((const void __user *) args->write_pointer_address,
226 			sizeof(uint32_t))) {
227 		pr_err("Can't access write pointer\n");
228 		return -EFAULT;
229 	}
230 
231 	if (args->eop_buffer_address &&
232 		!access_ok((const void __user *) args->eop_buffer_address,
233 			sizeof(uint32_t))) {
234 		pr_debug("Can't access eop buffer");
235 		return -EFAULT;
236 	}
237 
238 	if (args->ctx_save_restore_address &&
239 		!access_ok((const void __user *) args->ctx_save_restore_address,
240 			sizeof(uint32_t))) {
241 		pr_debug("Can't access ctx save restore buffer");
242 		return -EFAULT;
243 	}
244 
245 	q_properties->is_interop = false;
246 	q_properties->is_gws = false;
247 	q_properties->queue_percent = args->queue_percentage & 0xFF;
248 	/* bit 8-15 are repurposed to be PM4 target XCC */
249 	q_properties->pm4_target_xcc = (args->queue_percentage >> 8) & 0xFF;
250 	q_properties->priority = args->queue_priority;
251 	q_properties->queue_address = args->ring_base_address;
252 	q_properties->queue_size = args->ring_size;
253 	q_properties->read_ptr = (void __user *)args->read_pointer_address;
254 	q_properties->write_ptr = (void __user *)args->write_pointer_address;
255 	q_properties->eop_ring_buffer_address = args->eop_buffer_address;
256 	q_properties->eop_ring_buffer_size = args->eop_buffer_size;
257 	q_properties->ctx_save_restore_area_address =
258 			args->ctx_save_restore_address;
259 	q_properties->ctx_save_restore_area_size = args->ctx_save_restore_size;
260 	q_properties->ctl_stack_size = args->ctl_stack_size;
261 	q_properties->sdma_engine_id = args->sdma_engine_id;
262 	if (args->queue_type == KFD_IOC_QUEUE_TYPE_COMPUTE ||
263 		args->queue_type == KFD_IOC_QUEUE_TYPE_COMPUTE_AQL)
264 		q_properties->type = KFD_QUEUE_TYPE_COMPUTE;
265 	else if (args->queue_type == KFD_IOC_QUEUE_TYPE_SDMA)
266 		q_properties->type = KFD_QUEUE_TYPE_SDMA;
267 	else if (args->queue_type == KFD_IOC_QUEUE_TYPE_SDMA_XGMI)
268 		q_properties->type = KFD_QUEUE_TYPE_SDMA_XGMI;
269 	else if (args->queue_type == KFD_IOC_QUEUE_TYPE_SDMA_BY_ENG_ID)
270 		q_properties->type = KFD_QUEUE_TYPE_SDMA_BY_ENG_ID;
271 	else
272 		return -ENOTSUPP;
273 
274 	if (args->queue_type == KFD_IOC_QUEUE_TYPE_COMPUTE_AQL)
275 		q_properties->format = KFD_QUEUE_FORMAT_AQL;
276 	else
277 		q_properties->format = KFD_QUEUE_FORMAT_PM4;
278 
279 	pr_debug("Queue Percentage: %d, %d\n",
280 			q_properties->queue_percent, args->queue_percentage);
281 
282 	pr_debug("Queue Priority: %d, %d\n",
283 			q_properties->priority, args->queue_priority);
284 
285 	pr_debug("Queue Address: 0x%llX, 0x%llX\n",
286 			q_properties->queue_address, args->ring_base_address);
287 
288 	pr_debug("Queue Size: 0x%llX, %u\n",
289 			q_properties->queue_size, args->ring_size);
290 
291 	pr_debug("Queue r/w Pointers: %px, %px\n",
292 			q_properties->read_ptr,
293 			q_properties->write_ptr);
294 
295 	pr_debug("Queue Format: %d\n", q_properties->format);
296 
297 	pr_debug("Queue EOP: 0x%llX\n", q_properties->eop_ring_buffer_address);
298 
299 	pr_debug("Queue CTX save area: 0x%llX\n",
300 			q_properties->ctx_save_restore_area_address);
301 
302 	return 0;
303 }
304 
305 static int kfd_ioctl_create_queue(struct file *filep, struct kfd_process *p,
306 					void *data)
307 {
308 	struct kfd_ioctl_create_queue_args *args = data;
309 	struct kfd_node *dev;
310 	int err = 0;
311 	unsigned int queue_id;
312 	struct kfd_process_device *pdd;
313 	struct queue_properties q_properties;
314 	uint32_t doorbell_offset_in_process = 0;
315 
316 	memset(&q_properties, 0, sizeof(struct queue_properties));
317 
318 	pr_debug("Creating queue ioctl\n");
319 
320 	err = set_queue_properties_from_user(&q_properties, args);
321 	if (err)
322 		return err;
323 
324 	pr_debug("Looking for gpu id 0x%x\n", args->gpu_id);
325 
326 	mutex_lock(&p->mutex);
327 
328 	pdd = kfd_process_device_data_by_id(p, args->gpu_id);
329 	if (!pdd) {
330 		pr_debug("Could not find gpu id 0x%x\n", args->gpu_id);
331 		err = -EINVAL;
332 		goto err_pdd;
333 	}
334 	dev = pdd->dev;
335 
336 	pdd = kfd_bind_process_to_device(dev, p);
337 	if (IS_ERR(pdd)) {
338 		err = -ESRCH;
339 		goto err_bind_process;
340 	}
341 
342 	if (q_properties.type == KFD_QUEUE_TYPE_SDMA_BY_ENG_ID) {
343 		int max_sdma_eng_id = kfd_get_num_sdma_engines(dev) +
344 				      kfd_get_num_xgmi_sdma_engines(dev) - 1;
345 
346 		if (q_properties.sdma_engine_id > max_sdma_eng_id) {
347 			err = -EINVAL;
348 			pr_err("sdma_engine_id %i exceeds maximum id of %i\n",
349 			       q_properties.sdma_engine_id, max_sdma_eng_id);
350 			goto err_sdma_engine_id;
351 		}
352 	}
353 
354 	if (!pdd->qpd.proc_doorbells) {
355 		err = kfd_alloc_process_doorbells(dev->kfd, pdd);
356 		if (err) {
357 			pr_debug("failed to allocate process doorbells\n");
358 			goto err_bind_process;
359 		}
360 	}
361 
362 	err = kfd_queue_acquire_buffers(pdd, &q_properties);
363 	if (err) {
364 		pr_debug("failed to acquire user queue buffers\n");
365 		goto err_acquire_queue_buf;
366 	}
367 
368 	pr_debug("Creating queue for process pid %d on gpu 0x%x\n",
369 			p->lead_thread->pid,
370 			dev->id);
371 
372 	err = pqm_create_queue(&p->pqm, dev, &q_properties, &queue_id,
373 			NULL, NULL, NULL, &doorbell_offset_in_process);
374 	if (err != 0)
375 		goto err_create_queue;
376 
377 	args->queue_id = queue_id;
378 
379 
380 	/* Return gpu_id as doorbell offset for mmap usage */
381 	args->doorbell_offset = KFD_MMAP_TYPE_DOORBELL;
382 	args->doorbell_offset |= KFD_MMAP_GPU_ID(args->gpu_id);
383 	if (KFD_IS_SOC15(dev))
384 		/* On SOC15 ASICs, include the doorbell offset within the
385 		 * process doorbell frame, which is 2 pages.
386 		 */
387 		args->doorbell_offset |= doorbell_offset_in_process;
388 
389 	mutex_unlock(&p->mutex);
390 
391 	pr_debug("Queue id %d was created successfully\n", args->queue_id);
392 
393 	pr_debug("Ring buffer address == 0x%016llX\n",
394 			args->ring_base_address);
395 
396 	pr_debug("Read ptr address    == 0x%016llX\n",
397 			args->read_pointer_address);
398 
399 	pr_debug("Write ptr address   == 0x%016llX\n",
400 			args->write_pointer_address);
401 
402 	kfd_dbg_ev_raise(KFD_EC_MASK(EC_QUEUE_NEW), p, dev, queue_id, false, NULL, 0);
403 	return 0;
404 
405 err_create_queue:
406 	kfd_queue_unref_bo_vas(pdd, &q_properties);
407 	kfd_queue_release_buffers(pdd, &q_properties);
408 err_acquire_queue_buf:
409 err_sdma_engine_id:
410 err_bind_process:
411 err_pdd:
412 	mutex_unlock(&p->mutex);
413 	return err;
414 }
415 
416 static int kfd_ioctl_destroy_queue(struct file *filp, struct kfd_process *p,
417 					void *data)
418 {
419 	int retval;
420 	struct kfd_ioctl_destroy_queue_args *args = data;
421 
422 	pr_debug("Destroying queue id %d for process pid %d\n",
423 				args->queue_id,
424 				p->lead_thread->pid);
425 
426 	mutex_lock(&p->mutex);
427 
428 	retval = pqm_destroy_queue(&p->pqm, args->queue_id);
429 
430 	mutex_unlock(&p->mutex);
431 	return retval;
432 }
433 
434 static int kfd_ioctl_update_queue(struct file *filp, struct kfd_process *p,
435 					void *data)
436 {
437 	int retval;
438 	struct kfd_ioctl_update_queue_args *args = data;
439 	struct queue_properties properties;
440 
441 	/*
442 	 * Repurpose queue percentage to accommodate new features:
443 	 * bit 0-7: queue percentage
444 	 * bit 8-15: pm4_target_xcc
445 	 */
446 	if ((args->queue_percentage & 0xFF) > KFD_MAX_QUEUE_PERCENTAGE) {
447 		pr_err("Queue percentage must be between 0 to KFD_MAX_QUEUE_PERCENTAGE\n");
448 		return -EINVAL;
449 	}
450 
451 	if (args->queue_priority > KFD_MAX_QUEUE_PRIORITY) {
452 		pr_err("Queue priority must be between 0 to KFD_MAX_QUEUE_PRIORITY\n");
453 		return -EINVAL;
454 	}
455 
456 	if ((args->ring_base_address) &&
457 		(!access_ok((const void __user *) args->ring_base_address,
458 			sizeof(uint64_t)))) {
459 		pr_err("Can't access ring base address\n");
460 		return -EFAULT;
461 	}
462 
463 	if (!is_power_of_2(args->ring_size) && (args->ring_size != 0)) {
464 		pr_err("Ring size must be a power of 2 or 0\n");
465 		return -EINVAL;
466 	}
467 
468 	if (args->ring_size < KFD_MIN_QUEUE_RING_SIZE) {
469 		args->ring_size = KFD_MIN_QUEUE_RING_SIZE;
470 		pr_debug("Size lower. clamped to KFD_MIN_QUEUE_RING_SIZE");
471 	}
472 
473 	properties.queue_address = args->ring_base_address;
474 	properties.queue_size = args->ring_size;
475 	properties.queue_percent = args->queue_percentage & 0xFF;
476 	/* bit 8-15 are repurposed to be PM4 target XCC */
477 	properties.pm4_target_xcc = (args->queue_percentage >> 8) & 0xFF;
478 	properties.priority = args->queue_priority;
479 
480 	pr_debug("Updating queue id %d for process pid %d\n",
481 			args->queue_id, p->lead_thread->pid);
482 
483 	mutex_lock(&p->mutex);
484 
485 	retval = pqm_update_queue_properties(&p->pqm, args->queue_id, &properties);
486 
487 	mutex_unlock(&p->mutex);
488 
489 	return retval;
490 }
491 
492 static int kfd_ioctl_set_cu_mask(struct file *filp, struct kfd_process *p,
493 					void *data)
494 {
495 	int retval;
496 	const int max_num_cus = 1024;
497 	struct kfd_ioctl_set_cu_mask_args *args = data;
498 	struct mqd_update_info minfo = {0};
499 	uint32_t __user *cu_mask_ptr = (uint32_t __user *)args->cu_mask_ptr;
500 	size_t cu_mask_size = sizeof(uint32_t) * (args->num_cu_mask / 32);
501 
502 	if ((args->num_cu_mask % 32) != 0) {
503 		pr_debug("num_cu_mask 0x%x must be a multiple of 32",
504 				args->num_cu_mask);
505 		return -EINVAL;
506 	}
507 
508 	minfo.cu_mask.count = args->num_cu_mask;
509 	if (minfo.cu_mask.count == 0) {
510 		pr_debug("CU mask cannot be 0");
511 		return -EINVAL;
512 	}
513 
514 	/* To prevent an unreasonably large CU mask size, set an arbitrary
515 	 * limit of max_num_cus bits.  We can then just drop any CU mask bits
516 	 * past max_num_cus bits and just use the first max_num_cus bits.
517 	 */
518 	if (minfo.cu_mask.count > max_num_cus) {
519 		pr_debug("CU mask cannot be greater than 1024 bits");
520 		minfo.cu_mask.count = max_num_cus;
521 		cu_mask_size = sizeof(uint32_t) * (max_num_cus/32);
522 	}
523 
524 	minfo.cu_mask.ptr = kzalloc(cu_mask_size, GFP_KERNEL);
525 	if (!minfo.cu_mask.ptr)
526 		return -ENOMEM;
527 
528 	retval = copy_from_user(minfo.cu_mask.ptr, cu_mask_ptr, cu_mask_size);
529 	if (retval) {
530 		pr_debug("Could not copy CU mask from userspace");
531 		retval = -EFAULT;
532 		goto out;
533 	}
534 
535 	mutex_lock(&p->mutex);
536 
537 	retval = pqm_update_mqd(&p->pqm, args->queue_id, &minfo);
538 
539 	mutex_unlock(&p->mutex);
540 
541 out:
542 	kfree(minfo.cu_mask.ptr);
543 	return retval;
544 }
545 
546 static int kfd_ioctl_get_queue_wave_state(struct file *filep,
547 					  struct kfd_process *p, void *data)
548 {
549 	struct kfd_ioctl_get_queue_wave_state_args *args = data;
550 	int r;
551 
552 	mutex_lock(&p->mutex);
553 
554 	r = pqm_get_wave_state(&p->pqm, args->queue_id,
555 			       (void __user *)args->ctl_stack_address,
556 			       &args->ctl_stack_used_size,
557 			       &args->save_area_used_size);
558 
559 	mutex_unlock(&p->mutex);
560 
561 	return r;
562 }
563 
564 static int kfd_ioctl_set_memory_policy(struct file *filep,
565 					struct kfd_process *p, void *data)
566 {
567 	struct kfd_ioctl_set_memory_policy_args *args = data;
568 	int err = 0;
569 	struct kfd_process_device *pdd;
570 	enum cache_policy default_policy, alternate_policy;
571 
572 	if (args->default_policy != KFD_IOC_CACHE_POLICY_COHERENT
573 	    && args->default_policy != KFD_IOC_CACHE_POLICY_NONCOHERENT) {
574 		return -EINVAL;
575 	}
576 
577 	if (args->alternate_policy != KFD_IOC_CACHE_POLICY_COHERENT
578 	    && args->alternate_policy != KFD_IOC_CACHE_POLICY_NONCOHERENT) {
579 		return -EINVAL;
580 	}
581 
582 	mutex_lock(&p->mutex);
583 	pdd = kfd_process_device_data_by_id(p, args->gpu_id);
584 	if (!pdd) {
585 		pr_debug("Could not find gpu id 0x%x\n", args->gpu_id);
586 		err = -EINVAL;
587 		goto err_pdd;
588 	}
589 
590 	pdd = kfd_bind_process_to_device(pdd->dev, p);
591 	if (IS_ERR(pdd)) {
592 		err = -ESRCH;
593 		goto out;
594 	}
595 
596 	default_policy = (args->default_policy == KFD_IOC_CACHE_POLICY_COHERENT)
597 			 ? cache_policy_coherent : cache_policy_noncoherent;
598 
599 	alternate_policy =
600 		(args->alternate_policy == KFD_IOC_CACHE_POLICY_COHERENT)
601 		   ? cache_policy_coherent : cache_policy_noncoherent;
602 
603 	if (!pdd->dev->dqm->ops.set_cache_memory_policy(pdd->dev->dqm,
604 				&pdd->qpd,
605 				default_policy,
606 				alternate_policy,
607 				(void __user *)args->alternate_aperture_base,
608 				args->alternate_aperture_size,
609 				args->misc_process_flag))
610 		err = -EINVAL;
611 
612 out:
613 err_pdd:
614 	mutex_unlock(&p->mutex);
615 
616 	return err;
617 }
618 
619 static int kfd_ioctl_set_trap_handler(struct file *filep,
620 					struct kfd_process *p, void *data)
621 {
622 	struct kfd_ioctl_set_trap_handler_args *args = data;
623 	int err = 0;
624 	struct kfd_process_device *pdd;
625 
626 	mutex_lock(&p->mutex);
627 
628 	pdd = kfd_process_device_data_by_id(p, args->gpu_id);
629 	if (!pdd) {
630 		err = -EINVAL;
631 		goto err_pdd;
632 	}
633 
634 	pdd = kfd_bind_process_to_device(pdd->dev, p);
635 	if (IS_ERR(pdd)) {
636 		err = -ESRCH;
637 		goto out;
638 	}
639 
640 	kfd_process_set_trap_handler(&pdd->qpd, args->tba_addr, args->tma_addr);
641 
642 out:
643 err_pdd:
644 	mutex_unlock(&p->mutex);
645 
646 	return err;
647 }
648 
649 static int kfd_ioctl_dbg_register(struct file *filep,
650 				struct kfd_process *p, void *data)
651 {
652 	return -EPERM;
653 }
654 
655 static int kfd_ioctl_dbg_unregister(struct file *filep,
656 				struct kfd_process *p, void *data)
657 {
658 	return -EPERM;
659 }
660 
661 static int kfd_ioctl_dbg_address_watch(struct file *filep,
662 					struct kfd_process *p, void *data)
663 {
664 	return -EPERM;
665 }
666 
667 /* Parse and generate fixed size data structure for wave control */
668 static int kfd_ioctl_dbg_wave_control(struct file *filep,
669 					struct kfd_process *p, void *data)
670 {
671 	return -EPERM;
672 }
673 
674 static int kfd_ioctl_get_clock_counters(struct file *filep,
675 				struct kfd_process *p, void *data)
676 {
677 	struct kfd_ioctl_get_clock_counters_args *args = data;
678 	struct kfd_process_device *pdd;
679 
680 	mutex_lock(&p->mutex);
681 	pdd = kfd_process_device_data_by_id(p, args->gpu_id);
682 	mutex_unlock(&p->mutex);
683 	if (pdd)
684 		/* Reading GPU clock counter from KGD */
685 		args->gpu_clock_counter = amdgpu_amdkfd_get_gpu_clock_counter(pdd->dev->adev);
686 	else
687 		/* Node without GPU resource */
688 		args->gpu_clock_counter = 0;
689 
690 	/* No access to rdtsc. Using raw monotonic time */
691 	args->cpu_clock_counter = ktime_get_raw_ns();
692 	args->system_clock_counter = ktime_get_boottime_ns();
693 
694 	/* Since the counter is in nano-seconds we use 1GHz frequency */
695 	args->system_clock_freq = 1000000000;
696 
697 	return 0;
698 }
699 
700 
701 static int kfd_ioctl_get_process_apertures(struct file *filp,
702 				struct kfd_process *p, void *data)
703 {
704 	struct kfd_ioctl_get_process_apertures_args *args = data;
705 	struct kfd_process_device_apertures *pAperture;
706 	int i;
707 
708 	dev_dbg(kfd_device, "get apertures for process pid %d", p->lead_thread->pid);
709 
710 	args->num_of_nodes = 0;
711 
712 	mutex_lock(&p->mutex);
713 	/* Run over all pdd of the process */
714 	for (i = 0; i < p->n_pdds; i++) {
715 		struct kfd_process_device *pdd = p->pdds[i];
716 
717 		pAperture =
718 			&args->process_apertures[args->num_of_nodes];
719 		pAperture->gpu_id = pdd->dev->id;
720 		pAperture->lds_base = pdd->lds_base;
721 		pAperture->lds_limit = pdd->lds_limit;
722 		pAperture->gpuvm_base = pdd->gpuvm_base;
723 		pAperture->gpuvm_limit = pdd->gpuvm_limit;
724 		pAperture->scratch_base = pdd->scratch_base;
725 		pAperture->scratch_limit = pdd->scratch_limit;
726 
727 		dev_dbg(kfd_device,
728 			"node id %u\n", args->num_of_nodes);
729 		dev_dbg(kfd_device,
730 			"gpu id %u\n", pdd->dev->id);
731 		dev_dbg(kfd_device,
732 			"lds_base %llX\n", pdd->lds_base);
733 		dev_dbg(kfd_device,
734 			"lds_limit %llX\n", pdd->lds_limit);
735 		dev_dbg(kfd_device,
736 			"gpuvm_base %llX\n", pdd->gpuvm_base);
737 		dev_dbg(kfd_device,
738 			"gpuvm_limit %llX\n", pdd->gpuvm_limit);
739 		dev_dbg(kfd_device,
740 			"scratch_base %llX\n", pdd->scratch_base);
741 		dev_dbg(kfd_device,
742 			"scratch_limit %llX\n", pdd->scratch_limit);
743 
744 		if (++args->num_of_nodes >= NUM_OF_SUPPORTED_GPUS)
745 			break;
746 	}
747 	mutex_unlock(&p->mutex);
748 
749 	return 0;
750 }
751 
752 static int kfd_ioctl_get_process_apertures_new(struct file *filp,
753 				struct kfd_process *p, void *data)
754 {
755 	struct kfd_ioctl_get_process_apertures_new_args *args = data;
756 	struct kfd_process_device_apertures *pa;
757 	int ret;
758 	int i;
759 
760 	dev_dbg(kfd_device, "get apertures for process pid %d",
761 			p->lead_thread->pid);
762 
763 	if (args->num_of_nodes == 0) {
764 		/* Return number of nodes, so that user space can alloacate
765 		 * sufficient memory
766 		 */
767 		mutex_lock(&p->mutex);
768 		args->num_of_nodes = p->n_pdds;
769 		goto out_unlock;
770 	}
771 
772 	/* Fill in process-aperture information for all available
773 	 * nodes, but not more than args->num_of_nodes as that is
774 	 * the amount of memory allocated by user
775 	 */
776 	pa = kcalloc(args->num_of_nodes, sizeof(struct kfd_process_device_apertures),
777 		     GFP_KERNEL);
778 	if (!pa)
779 		return -ENOMEM;
780 
781 	mutex_lock(&p->mutex);
782 
783 	if (!p->n_pdds) {
784 		args->num_of_nodes = 0;
785 		kfree(pa);
786 		goto out_unlock;
787 	}
788 
789 	/* Run over all pdd of the process */
790 	for (i = 0; i < min(p->n_pdds, args->num_of_nodes); i++) {
791 		struct kfd_process_device *pdd = p->pdds[i];
792 
793 		pa[i].gpu_id = pdd->dev->id;
794 		pa[i].lds_base = pdd->lds_base;
795 		pa[i].lds_limit = pdd->lds_limit;
796 		pa[i].gpuvm_base = pdd->gpuvm_base;
797 		pa[i].gpuvm_limit = pdd->gpuvm_limit;
798 		pa[i].scratch_base = pdd->scratch_base;
799 		pa[i].scratch_limit = pdd->scratch_limit;
800 
801 		dev_dbg(kfd_device,
802 			"gpu id %u\n", pdd->dev->id);
803 		dev_dbg(kfd_device,
804 			"lds_base %llX\n", pdd->lds_base);
805 		dev_dbg(kfd_device,
806 			"lds_limit %llX\n", pdd->lds_limit);
807 		dev_dbg(kfd_device,
808 			"gpuvm_base %llX\n", pdd->gpuvm_base);
809 		dev_dbg(kfd_device,
810 			"gpuvm_limit %llX\n", pdd->gpuvm_limit);
811 		dev_dbg(kfd_device,
812 			"scratch_base %llX\n", pdd->scratch_base);
813 		dev_dbg(kfd_device,
814 			"scratch_limit %llX\n", pdd->scratch_limit);
815 	}
816 	mutex_unlock(&p->mutex);
817 
818 	args->num_of_nodes = i;
819 	ret = copy_to_user(
820 			(void __user *)args->kfd_process_device_apertures_ptr,
821 			pa,
822 			(i * sizeof(struct kfd_process_device_apertures)));
823 	kfree(pa);
824 	return ret ? -EFAULT : 0;
825 
826 out_unlock:
827 	mutex_unlock(&p->mutex);
828 	return 0;
829 }
830 
831 static int kfd_ioctl_create_event(struct file *filp, struct kfd_process *p,
832 					void *data)
833 {
834 	struct kfd_ioctl_create_event_args *args = data;
835 	int err;
836 
837 	/* For dGPUs the event page is allocated in user mode. The
838 	 * handle is passed to KFD with the first call to this IOCTL
839 	 * through the event_page_offset field.
840 	 */
841 	if (args->event_page_offset) {
842 		mutex_lock(&p->mutex);
843 		err = kfd_kmap_event_page(p, args->event_page_offset);
844 		mutex_unlock(&p->mutex);
845 		if (err)
846 			return err;
847 	}
848 
849 	err = kfd_event_create(filp, p, args->event_type,
850 				args->auto_reset != 0, args->node_id,
851 				&args->event_id, &args->event_trigger_data,
852 				&args->event_page_offset,
853 				&args->event_slot_index);
854 
855 	pr_debug("Created event (id:0x%08x) (%s)\n", args->event_id, __func__);
856 	return err;
857 }
858 
859 static int kfd_ioctl_destroy_event(struct file *filp, struct kfd_process *p,
860 					void *data)
861 {
862 	struct kfd_ioctl_destroy_event_args *args = data;
863 
864 	return kfd_event_destroy(p, args->event_id);
865 }
866 
867 static int kfd_ioctl_set_event(struct file *filp, struct kfd_process *p,
868 				void *data)
869 {
870 	struct kfd_ioctl_set_event_args *args = data;
871 
872 	return kfd_set_event(p, args->event_id);
873 }
874 
875 static int kfd_ioctl_reset_event(struct file *filp, struct kfd_process *p,
876 				void *data)
877 {
878 	struct kfd_ioctl_reset_event_args *args = data;
879 
880 	return kfd_reset_event(p, args->event_id);
881 }
882 
883 static int kfd_ioctl_wait_events(struct file *filp, struct kfd_process *p,
884 				void *data)
885 {
886 	struct kfd_ioctl_wait_events_args *args = data;
887 
888 	return kfd_wait_on_events(p, args->num_events,
889 			(void __user *)args->events_ptr,
890 			(args->wait_for_all != 0),
891 			&args->timeout, &args->wait_result);
892 }
893 static int kfd_ioctl_set_scratch_backing_va(struct file *filep,
894 					struct kfd_process *p, void *data)
895 {
896 	struct kfd_ioctl_set_scratch_backing_va_args *args = data;
897 	struct kfd_process_device *pdd;
898 	struct kfd_node *dev;
899 	long err;
900 
901 	mutex_lock(&p->mutex);
902 	pdd = kfd_process_device_data_by_id(p, args->gpu_id);
903 	if (!pdd) {
904 		err = -EINVAL;
905 		goto err_pdd;
906 	}
907 	dev = pdd->dev;
908 
909 	pdd = kfd_bind_process_to_device(dev, p);
910 	if (IS_ERR(pdd)) {
911 		err = PTR_ERR(pdd);
912 		goto bind_process_to_device_fail;
913 	}
914 
915 	pdd->qpd.sh_hidden_private_base = args->va_addr;
916 
917 	mutex_unlock(&p->mutex);
918 
919 	if (dev->dqm->sched_policy == KFD_SCHED_POLICY_NO_HWS &&
920 	    pdd->qpd.vmid != 0 && dev->kfd2kgd->set_scratch_backing_va)
921 		dev->kfd2kgd->set_scratch_backing_va(
922 			dev->adev, args->va_addr, pdd->qpd.vmid);
923 
924 	return 0;
925 
926 bind_process_to_device_fail:
927 err_pdd:
928 	mutex_unlock(&p->mutex);
929 	return err;
930 }
931 
932 static int kfd_ioctl_get_tile_config(struct file *filep,
933 		struct kfd_process *p, void *data)
934 {
935 	struct kfd_ioctl_get_tile_config_args *args = data;
936 	struct kfd_process_device *pdd;
937 	struct tile_config config;
938 	int err = 0;
939 
940 	mutex_lock(&p->mutex);
941 	pdd = kfd_process_device_data_by_id(p, args->gpu_id);
942 	mutex_unlock(&p->mutex);
943 	if (!pdd)
944 		return -EINVAL;
945 
946 	amdgpu_amdkfd_get_tile_config(pdd->dev->adev, &config);
947 
948 	args->gb_addr_config = config.gb_addr_config;
949 	args->num_banks = config.num_banks;
950 	args->num_ranks = config.num_ranks;
951 
952 	if (args->num_tile_configs > config.num_tile_configs)
953 		args->num_tile_configs = config.num_tile_configs;
954 	err = copy_to_user((void __user *)args->tile_config_ptr,
955 			config.tile_config_ptr,
956 			args->num_tile_configs * sizeof(uint32_t));
957 	if (err) {
958 		args->num_tile_configs = 0;
959 		return -EFAULT;
960 	}
961 
962 	if (args->num_macro_tile_configs > config.num_macro_tile_configs)
963 		args->num_macro_tile_configs =
964 				config.num_macro_tile_configs;
965 	err = copy_to_user((void __user *)args->macro_tile_config_ptr,
966 			config.macro_tile_config_ptr,
967 			args->num_macro_tile_configs * sizeof(uint32_t));
968 	if (err) {
969 		args->num_macro_tile_configs = 0;
970 		return -EFAULT;
971 	}
972 
973 	return 0;
974 }
975 
976 static int kfd_ioctl_acquire_vm(struct file *filep, struct kfd_process *p,
977 				void *data)
978 {
979 	struct kfd_ioctl_acquire_vm_args *args = data;
980 	struct kfd_process_device *pdd;
981 	struct file *drm_file;
982 	int ret;
983 
984 	drm_file = fget(args->drm_fd);
985 	if (!drm_file)
986 		return -EINVAL;
987 
988 	mutex_lock(&p->mutex);
989 	pdd = kfd_process_device_data_by_id(p, args->gpu_id);
990 	if (!pdd) {
991 		ret = -EINVAL;
992 		goto err_pdd;
993 	}
994 
995 	if (pdd->drm_file) {
996 		ret = pdd->drm_file == drm_file ? 0 : -EBUSY;
997 		goto err_drm_file;
998 	}
999 
1000 	ret = kfd_process_device_init_vm(pdd, drm_file);
1001 	if (ret)
1002 		goto err_unlock;
1003 
1004 	/* On success, the PDD keeps the drm_file reference */
1005 	mutex_unlock(&p->mutex);
1006 
1007 	return 0;
1008 
1009 err_unlock:
1010 err_pdd:
1011 err_drm_file:
1012 	mutex_unlock(&p->mutex);
1013 	fput(drm_file);
1014 	return ret;
1015 }
1016 
1017 bool kfd_dev_is_large_bar(struct kfd_node *dev)
1018 {
1019 	if (dev->kfd->adev->debug_largebar) {
1020 		pr_debug("Simulate large-bar allocation on non large-bar machine\n");
1021 		return true;
1022 	}
1023 
1024 	if (dev->local_mem_info.local_mem_size_private == 0 &&
1025 	    dev->local_mem_info.local_mem_size_public > 0)
1026 		return true;
1027 
1028 	if (dev->local_mem_info.local_mem_size_public == 0 &&
1029 	    dev->kfd->adev->gmc.is_app_apu) {
1030 		pr_debug("APP APU, Consider like a large bar system\n");
1031 		return true;
1032 	}
1033 
1034 	return false;
1035 }
1036 
1037 static int kfd_ioctl_get_available_memory(struct file *filep,
1038 					  struct kfd_process *p, void *data)
1039 {
1040 	struct kfd_ioctl_get_available_memory_args *args = data;
1041 	struct kfd_process_device *pdd = kfd_lock_pdd_by_id(p, args->gpu_id);
1042 
1043 	if (!pdd)
1044 		return -EINVAL;
1045 	args->available = amdgpu_amdkfd_get_available_memory(pdd->dev->adev,
1046 							pdd->dev->node_id);
1047 	kfd_unlock_pdd(pdd);
1048 	return 0;
1049 }
1050 
1051 static int kfd_ioctl_alloc_memory_of_gpu(struct file *filep,
1052 					struct kfd_process *p, void *data)
1053 {
1054 	struct kfd_ioctl_alloc_memory_of_gpu_args *args = data;
1055 	struct kfd_process_device *pdd;
1056 	void *mem;
1057 	struct kfd_node *dev;
1058 	int idr_handle;
1059 	long err;
1060 	uint64_t offset = args->mmap_offset;
1061 	uint32_t flags = args->flags;
1062 
1063 	if (args->size == 0)
1064 		return -EINVAL;
1065 
1066 #if IS_ENABLED(CONFIG_HSA_AMD_SVM)
1067 	/* Flush pending deferred work to avoid racing with deferred actions
1068 	 * from previous memory map changes (e.g. munmap).
1069 	 */
1070 	svm_range_list_lock_and_flush_work(&p->svms, current->mm);
1071 	mutex_lock(&p->svms.lock);
1072 	mmap_write_unlock(current->mm);
1073 
1074 	/* Skip a special case that allocates VRAM without VA,
1075 	 * VA will be invalid of 0.
1076 	 */
1077 	if (!(!args->va_addr && (flags & KFD_IOC_ALLOC_MEM_FLAGS_VRAM)) &&
1078 	    interval_tree_iter_first(&p->svms.objects,
1079 				     args->va_addr >> PAGE_SHIFT,
1080 				     (args->va_addr + args->size - 1) >> PAGE_SHIFT)) {
1081 		pr_err("Address: 0x%llx already allocated by SVM\n",
1082 			args->va_addr);
1083 		mutex_unlock(&p->svms.lock);
1084 		return -EADDRINUSE;
1085 	}
1086 
1087 	/* When register user buffer check if it has been registered by svm by
1088 	 * buffer cpu virtual address.
1089 	 */
1090 	if ((flags & KFD_IOC_ALLOC_MEM_FLAGS_USERPTR) &&
1091 	    interval_tree_iter_first(&p->svms.objects,
1092 				     args->mmap_offset >> PAGE_SHIFT,
1093 				     (args->mmap_offset  + args->size - 1) >> PAGE_SHIFT)) {
1094 		pr_err("User Buffer Address: 0x%llx already allocated by SVM\n",
1095 			args->mmap_offset);
1096 		mutex_unlock(&p->svms.lock);
1097 		return -EADDRINUSE;
1098 	}
1099 
1100 	mutex_unlock(&p->svms.lock);
1101 #endif
1102 	mutex_lock(&p->mutex);
1103 	pdd = kfd_process_device_data_by_id(p, args->gpu_id);
1104 	if (!pdd) {
1105 		err = -EINVAL;
1106 		goto err_pdd;
1107 	}
1108 
1109 	dev = pdd->dev;
1110 
1111 	if ((flags & KFD_IOC_ALLOC_MEM_FLAGS_PUBLIC) &&
1112 		(flags & KFD_IOC_ALLOC_MEM_FLAGS_VRAM) &&
1113 		!kfd_dev_is_large_bar(dev)) {
1114 		pr_err("Alloc host visible vram on small bar is not allowed\n");
1115 		err = -EINVAL;
1116 		goto err_large_bar;
1117 	}
1118 
1119 	pdd = kfd_bind_process_to_device(dev, p);
1120 	if (IS_ERR(pdd)) {
1121 		err = PTR_ERR(pdd);
1122 		goto err_unlock;
1123 	}
1124 
1125 	if (flags & KFD_IOC_ALLOC_MEM_FLAGS_DOORBELL) {
1126 		if (args->size != kfd_doorbell_process_slice(dev->kfd)) {
1127 			err = -EINVAL;
1128 			goto err_unlock;
1129 		}
1130 		offset = kfd_get_process_doorbells(pdd);
1131 		if (!offset) {
1132 			err = -ENOMEM;
1133 			goto err_unlock;
1134 		}
1135 	} else if (flags & KFD_IOC_ALLOC_MEM_FLAGS_MMIO_REMAP) {
1136 		if (args->size != PAGE_SIZE) {
1137 			err = -EINVAL;
1138 			goto err_unlock;
1139 		}
1140 		offset = dev->adev->rmmio_remap.bus_addr;
1141 		if (!offset || (PAGE_SIZE > 4096)) {
1142 			err = -ENOMEM;
1143 			goto err_unlock;
1144 		}
1145 	}
1146 
1147 	err = amdgpu_amdkfd_gpuvm_alloc_memory_of_gpu(
1148 		dev->adev, args->va_addr, args->size,
1149 		pdd->drm_priv, (struct kgd_mem **) &mem, &offset,
1150 		flags, false);
1151 
1152 	if (err)
1153 		goto err_unlock;
1154 
1155 	idr_handle = kfd_process_device_create_obj_handle(pdd, mem);
1156 	if (idr_handle < 0) {
1157 		err = -EFAULT;
1158 		goto err_free;
1159 	}
1160 
1161 	/* Update the VRAM usage count */
1162 	if (flags & KFD_IOC_ALLOC_MEM_FLAGS_VRAM) {
1163 		uint64_t size = args->size;
1164 
1165 		if (flags & KFD_IOC_ALLOC_MEM_FLAGS_AQL_QUEUE_MEM)
1166 			size >>= 1;
1167 		atomic64_add(PAGE_ALIGN(size), &pdd->vram_usage);
1168 	}
1169 
1170 	mutex_unlock(&p->mutex);
1171 
1172 	args->handle = MAKE_HANDLE(args->gpu_id, idr_handle);
1173 	args->mmap_offset = offset;
1174 
1175 	/* MMIO is mapped through kfd device
1176 	 * Generate a kfd mmap offset
1177 	 */
1178 	if (flags & KFD_IOC_ALLOC_MEM_FLAGS_MMIO_REMAP)
1179 		args->mmap_offset = KFD_MMAP_TYPE_MMIO
1180 					| KFD_MMAP_GPU_ID(args->gpu_id);
1181 
1182 	return 0;
1183 
1184 err_free:
1185 	amdgpu_amdkfd_gpuvm_free_memory_of_gpu(dev->adev, (struct kgd_mem *)mem,
1186 					       pdd->drm_priv, NULL);
1187 err_unlock:
1188 err_pdd:
1189 err_large_bar:
1190 	mutex_unlock(&p->mutex);
1191 	return err;
1192 }
1193 
1194 static int kfd_ioctl_free_memory_of_gpu(struct file *filep,
1195 					struct kfd_process *p, void *data)
1196 {
1197 	struct kfd_ioctl_free_memory_of_gpu_args *args = data;
1198 	struct kfd_process_device *pdd;
1199 	void *mem;
1200 	int ret;
1201 	uint64_t size = 0;
1202 
1203 	mutex_lock(&p->mutex);
1204 	/*
1205 	 * Safeguard to prevent user space from freeing signal BO.
1206 	 * It will be freed at process termination.
1207 	 */
1208 	if (p->signal_handle && (p->signal_handle == args->handle)) {
1209 		pr_err("Free signal BO is not allowed\n");
1210 		ret = -EPERM;
1211 		goto err_unlock;
1212 	}
1213 
1214 	pdd = kfd_process_device_data_by_id(p, GET_GPU_ID(args->handle));
1215 	if (!pdd) {
1216 		pr_err("Process device data doesn't exist\n");
1217 		ret = -EINVAL;
1218 		goto err_pdd;
1219 	}
1220 
1221 	mem = kfd_process_device_translate_handle(
1222 		pdd, GET_IDR_HANDLE(args->handle));
1223 	if (!mem) {
1224 		ret = -EINVAL;
1225 		goto err_unlock;
1226 	}
1227 
1228 	ret = amdgpu_amdkfd_gpuvm_free_memory_of_gpu(pdd->dev->adev,
1229 				(struct kgd_mem *)mem, pdd->drm_priv, &size);
1230 
1231 	/* If freeing the buffer failed, leave the handle in place for
1232 	 * clean-up during process tear-down.
1233 	 */
1234 	if (!ret)
1235 		kfd_process_device_remove_obj_handle(
1236 			pdd, GET_IDR_HANDLE(args->handle));
1237 
1238 	atomic64_sub(size, &pdd->vram_usage);
1239 
1240 err_unlock:
1241 err_pdd:
1242 	mutex_unlock(&p->mutex);
1243 	return ret;
1244 }
1245 
1246 static int kfd_ioctl_map_memory_to_gpu(struct file *filep,
1247 					struct kfd_process *p, void *data)
1248 {
1249 	struct kfd_ioctl_map_memory_to_gpu_args *args = data;
1250 	struct kfd_process_device *pdd, *peer_pdd;
1251 	void *mem;
1252 	struct kfd_node *dev;
1253 	long err = 0;
1254 	int i;
1255 	uint32_t *devices_arr = NULL;
1256 
1257 	if (!args->n_devices) {
1258 		pr_debug("Device IDs array empty\n");
1259 		return -EINVAL;
1260 	}
1261 	if (args->n_success > args->n_devices) {
1262 		pr_debug("n_success exceeds n_devices\n");
1263 		return -EINVAL;
1264 	}
1265 
1266 	devices_arr = kmalloc_array(args->n_devices, sizeof(*devices_arr),
1267 				    GFP_KERNEL);
1268 	if (!devices_arr)
1269 		return -ENOMEM;
1270 
1271 	err = copy_from_user(devices_arr,
1272 			     (void __user *)args->device_ids_array_ptr,
1273 			     args->n_devices * sizeof(*devices_arr));
1274 	if (err != 0) {
1275 		err = -EFAULT;
1276 		goto copy_from_user_failed;
1277 	}
1278 
1279 	mutex_lock(&p->mutex);
1280 	pdd = kfd_process_device_data_by_id(p, GET_GPU_ID(args->handle));
1281 	if (!pdd) {
1282 		err = -EINVAL;
1283 		goto get_process_device_data_failed;
1284 	}
1285 	dev = pdd->dev;
1286 
1287 	pdd = kfd_bind_process_to_device(dev, p);
1288 	if (IS_ERR(pdd)) {
1289 		err = PTR_ERR(pdd);
1290 		goto bind_process_to_device_failed;
1291 	}
1292 
1293 	mem = kfd_process_device_translate_handle(pdd,
1294 						GET_IDR_HANDLE(args->handle));
1295 	if (!mem) {
1296 		err = -ENOMEM;
1297 		goto get_mem_obj_from_handle_failed;
1298 	}
1299 
1300 	for (i = args->n_success; i < args->n_devices; i++) {
1301 		peer_pdd = kfd_process_device_data_by_id(p, devices_arr[i]);
1302 		if (!peer_pdd) {
1303 			pr_debug("Getting device by id failed for 0x%x\n",
1304 				 devices_arr[i]);
1305 			err = -EINVAL;
1306 			goto get_mem_obj_from_handle_failed;
1307 		}
1308 
1309 		peer_pdd = kfd_bind_process_to_device(peer_pdd->dev, p);
1310 		if (IS_ERR(peer_pdd)) {
1311 			err = PTR_ERR(peer_pdd);
1312 			goto get_mem_obj_from_handle_failed;
1313 		}
1314 
1315 		err = amdgpu_amdkfd_gpuvm_map_memory_to_gpu(
1316 			peer_pdd->dev->adev, (struct kgd_mem *)mem,
1317 			peer_pdd->drm_priv);
1318 		if (err) {
1319 			struct pci_dev *pdev = peer_pdd->dev->adev->pdev;
1320 
1321 			dev_err(dev->adev->dev,
1322 			       "Failed to map peer:%04x:%02x:%02x.%d mem_domain:%d\n",
1323 			       pci_domain_nr(pdev->bus),
1324 			       pdev->bus->number,
1325 			       PCI_SLOT(pdev->devfn),
1326 			       PCI_FUNC(pdev->devfn),
1327 			       ((struct kgd_mem *)mem)->domain);
1328 			goto map_memory_to_gpu_failed;
1329 		}
1330 		args->n_success = i+1;
1331 	}
1332 
1333 	err = amdgpu_amdkfd_gpuvm_sync_memory(dev->adev, (struct kgd_mem *) mem, true);
1334 	if (err) {
1335 		pr_debug("Sync memory failed, wait interrupted by user signal\n");
1336 		goto sync_memory_failed;
1337 	}
1338 
1339 	mutex_unlock(&p->mutex);
1340 
1341 	/* Flush TLBs after waiting for the page table updates to complete */
1342 	for (i = 0; i < args->n_devices; i++) {
1343 		peer_pdd = kfd_process_device_data_by_id(p, devices_arr[i]);
1344 		if (WARN_ON_ONCE(!peer_pdd))
1345 			continue;
1346 		kfd_flush_tlb(peer_pdd, TLB_FLUSH_LEGACY);
1347 	}
1348 	kfree(devices_arr);
1349 
1350 	return err;
1351 
1352 get_process_device_data_failed:
1353 bind_process_to_device_failed:
1354 get_mem_obj_from_handle_failed:
1355 map_memory_to_gpu_failed:
1356 sync_memory_failed:
1357 	mutex_unlock(&p->mutex);
1358 copy_from_user_failed:
1359 	kfree(devices_arr);
1360 
1361 	return err;
1362 }
1363 
1364 static int kfd_ioctl_unmap_memory_from_gpu(struct file *filep,
1365 					struct kfd_process *p, void *data)
1366 {
1367 	struct kfd_ioctl_unmap_memory_from_gpu_args *args = data;
1368 	struct kfd_process_device *pdd, *peer_pdd;
1369 	void *mem;
1370 	long err = 0;
1371 	uint32_t *devices_arr = NULL, i;
1372 	bool flush_tlb;
1373 
1374 	if (!args->n_devices) {
1375 		pr_debug("Device IDs array empty\n");
1376 		return -EINVAL;
1377 	}
1378 	if (args->n_success > args->n_devices) {
1379 		pr_debug("n_success exceeds n_devices\n");
1380 		return -EINVAL;
1381 	}
1382 
1383 	devices_arr = kmalloc_array(args->n_devices, sizeof(*devices_arr),
1384 				    GFP_KERNEL);
1385 	if (!devices_arr)
1386 		return -ENOMEM;
1387 
1388 	err = copy_from_user(devices_arr,
1389 			     (void __user *)args->device_ids_array_ptr,
1390 			     args->n_devices * sizeof(*devices_arr));
1391 	if (err != 0) {
1392 		err = -EFAULT;
1393 		goto copy_from_user_failed;
1394 	}
1395 
1396 	mutex_lock(&p->mutex);
1397 	pdd = kfd_process_device_data_by_id(p, GET_GPU_ID(args->handle));
1398 	if (!pdd) {
1399 		err = -EINVAL;
1400 		goto bind_process_to_device_failed;
1401 	}
1402 
1403 	mem = kfd_process_device_translate_handle(pdd,
1404 						GET_IDR_HANDLE(args->handle));
1405 	if (!mem) {
1406 		err = -ENOMEM;
1407 		goto get_mem_obj_from_handle_failed;
1408 	}
1409 
1410 	for (i = args->n_success; i < args->n_devices; i++) {
1411 		peer_pdd = kfd_process_device_data_by_id(p, devices_arr[i]);
1412 		if (!peer_pdd) {
1413 			err = -EINVAL;
1414 			goto get_mem_obj_from_handle_failed;
1415 		}
1416 		err = amdgpu_amdkfd_gpuvm_unmap_memory_from_gpu(
1417 			peer_pdd->dev->adev, (struct kgd_mem *)mem, peer_pdd->drm_priv);
1418 		if (err) {
1419 			pr_debug("Failed to unmap from gpu %d/%d\n", i, args->n_devices);
1420 			goto unmap_memory_from_gpu_failed;
1421 		}
1422 		args->n_success = i+1;
1423 	}
1424 
1425 	flush_tlb = kfd_flush_tlb_after_unmap(pdd->dev->kfd);
1426 	if (flush_tlb) {
1427 		err = amdgpu_amdkfd_gpuvm_sync_memory(pdd->dev->adev,
1428 				(struct kgd_mem *) mem, true);
1429 		if (err) {
1430 			pr_debug("Sync memory failed, wait interrupted by user signal\n");
1431 			goto sync_memory_failed;
1432 		}
1433 	}
1434 
1435 	/* Flush TLBs after waiting for the page table updates to complete */
1436 	for (i = 0; i < args->n_devices; i++) {
1437 		peer_pdd = kfd_process_device_data_by_id(p, devices_arr[i]);
1438 		if (WARN_ON_ONCE(!peer_pdd))
1439 			continue;
1440 		if (flush_tlb)
1441 			kfd_flush_tlb(peer_pdd, TLB_FLUSH_HEAVYWEIGHT);
1442 
1443 		/* Remove dma mapping after tlb flush to avoid IO_PAGE_FAULT */
1444 		err = amdgpu_amdkfd_gpuvm_dmaunmap_mem(mem, peer_pdd->drm_priv);
1445 		if (err)
1446 			goto sync_memory_failed;
1447 	}
1448 
1449 	mutex_unlock(&p->mutex);
1450 
1451 	kfree(devices_arr);
1452 
1453 	return 0;
1454 
1455 bind_process_to_device_failed:
1456 get_mem_obj_from_handle_failed:
1457 unmap_memory_from_gpu_failed:
1458 sync_memory_failed:
1459 	mutex_unlock(&p->mutex);
1460 copy_from_user_failed:
1461 	kfree(devices_arr);
1462 	return err;
1463 }
1464 
1465 static int kfd_ioctl_alloc_queue_gws(struct file *filep,
1466 		struct kfd_process *p, void *data)
1467 {
1468 	int retval;
1469 	struct kfd_ioctl_alloc_queue_gws_args *args = data;
1470 	struct queue *q;
1471 	struct kfd_node *dev;
1472 
1473 	mutex_lock(&p->mutex);
1474 	q = pqm_get_user_queue(&p->pqm, args->queue_id);
1475 
1476 	if (q) {
1477 		dev = q->device;
1478 	} else {
1479 		retval = -EINVAL;
1480 		goto out_unlock;
1481 	}
1482 
1483 	if (!dev->gws) {
1484 		retval = -ENODEV;
1485 		goto out_unlock;
1486 	}
1487 
1488 	if (dev->dqm->sched_policy == KFD_SCHED_POLICY_NO_HWS) {
1489 		retval = -ENODEV;
1490 		goto out_unlock;
1491 	}
1492 
1493 	if (p->debug_trap_enabled && (!kfd_dbg_has_gws_support(dev) ||
1494 				      kfd_dbg_has_cwsr_workaround(dev))) {
1495 		retval = -EBUSY;
1496 		goto out_unlock;
1497 	}
1498 
1499 	retval = pqm_set_gws(&p->pqm, args->queue_id, args->num_gws ? dev->gws : NULL);
1500 	mutex_unlock(&p->mutex);
1501 
1502 	args->first_gws = 0;
1503 	return retval;
1504 
1505 out_unlock:
1506 	mutex_unlock(&p->mutex);
1507 	return retval;
1508 }
1509 
1510 static int kfd_ioctl_get_dmabuf_info(struct file *filep,
1511 		struct kfd_process *p, void *data)
1512 {
1513 	struct kfd_ioctl_get_dmabuf_info_args *args = data;
1514 	struct kfd_node *dev = NULL;
1515 	struct amdgpu_device *dmabuf_adev;
1516 	void *metadata_buffer = NULL;
1517 	uint32_t flags;
1518 	int8_t xcp_id;
1519 	unsigned int i;
1520 	int r;
1521 
1522 	/* Find a KFD GPU device that supports the get_dmabuf_info query */
1523 	for (i = 0; kfd_topology_enum_kfd_devices(i, &dev) == 0; i++)
1524 		if (dev && !kfd_devcgroup_check_permission(dev))
1525 			break;
1526 	if (!dev)
1527 		return -EINVAL;
1528 
1529 	if (args->metadata_ptr) {
1530 		metadata_buffer = kzalloc(args->metadata_size, GFP_KERNEL);
1531 		if (!metadata_buffer)
1532 			return -ENOMEM;
1533 	}
1534 
1535 	/* Get dmabuf info from KGD */
1536 	r = amdgpu_amdkfd_get_dmabuf_info(dev->adev, args->dmabuf_fd,
1537 					  &dmabuf_adev, &args->size,
1538 					  metadata_buffer, args->metadata_size,
1539 					  &args->metadata_size, &flags, &xcp_id);
1540 	if (r)
1541 		goto exit;
1542 
1543 	if (xcp_id >= 0)
1544 		args->gpu_id = dmabuf_adev->kfd.dev->nodes[xcp_id]->id;
1545 	else
1546 		args->gpu_id = dev->id;
1547 	args->flags = flags;
1548 
1549 	/* Copy metadata buffer to user mode */
1550 	if (metadata_buffer) {
1551 		r = copy_to_user((void __user *)args->metadata_ptr,
1552 				 metadata_buffer, args->metadata_size);
1553 		if (r != 0)
1554 			r = -EFAULT;
1555 	}
1556 
1557 exit:
1558 	kfree(metadata_buffer);
1559 
1560 	return r;
1561 }
1562 
1563 static int kfd_ioctl_import_dmabuf(struct file *filep,
1564 				   struct kfd_process *p, void *data)
1565 {
1566 	struct kfd_ioctl_import_dmabuf_args *args = data;
1567 	struct kfd_process_device *pdd;
1568 	int idr_handle;
1569 	uint64_t size;
1570 	void *mem;
1571 	int r;
1572 
1573 	mutex_lock(&p->mutex);
1574 	pdd = kfd_process_device_data_by_id(p, args->gpu_id);
1575 	if (!pdd) {
1576 		r = -EINVAL;
1577 		goto err_unlock;
1578 	}
1579 
1580 	pdd = kfd_bind_process_to_device(pdd->dev, p);
1581 	if (IS_ERR(pdd)) {
1582 		r = PTR_ERR(pdd);
1583 		goto err_unlock;
1584 	}
1585 
1586 	r = amdgpu_amdkfd_gpuvm_import_dmabuf_fd(pdd->dev->adev, args->dmabuf_fd,
1587 						 args->va_addr, pdd->drm_priv,
1588 						 (struct kgd_mem **)&mem, &size,
1589 						 NULL);
1590 	if (r)
1591 		goto err_unlock;
1592 
1593 	idr_handle = kfd_process_device_create_obj_handle(pdd, mem);
1594 	if (idr_handle < 0) {
1595 		r = -EFAULT;
1596 		goto err_free;
1597 	}
1598 
1599 	mutex_unlock(&p->mutex);
1600 
1601 	args->handle = MAKE_HANDLE(args->gpu_id, idr_handle);
1602 
1603 	return 0;
1604 
1605 err_free:
1606 	amdgpu_amdkfd_gpuvm_free_memory_of_gpu(pdd->dev->adev, (struct kgd_mem *)mem,
1607 					       pdd->drm_priv, NULL);
1608 err_unlock:
1609 	mutex_unlock(&p->mutex);
1610 	return r;
1611 }
1612 
1613 static int kfd_ioctl_export_dmabuf(struct file *filep,
1614 				   struct kfd_process *p, void *data)
1615 {
1616 	struct kfd_ioctl_export_dmabuf_args *args = data;
1617 	struct kfd_process_device *pdd;
1618 	struct dma_buf *dmabuf;
1619 	struct kfd_node *dev;
1620 	void *mem;
1621 	int ret = 0;
1622 
1623 	dev = kfd_device_by_id(GET_GPU_ID(args->handle));
1624 	if (!dev)
1625 		return -EINVAL;
1626 
1627 	mutex_lock(&p->mutex);
1628 
1629 	pdd = kfd_get_process_device_data(dev, p);
1630 	if (!pdd) {
1631 		ret = -EINVAL;
1632 		goto err_unlock;
1633 	}
1634 
1635 	mem = kfd_process_device_translate_handle(pdd,
1636 						GET_IDR_HANDLE(args->handle));
1637 	if (!mem) {
1638 		ret = -EINVAL;
1639 		goto err_unlock;
1640 	}
1641 
1642 	ret = amdgpu_amdkfd_gpuvm_export_dmabuf(mem, &dmabuf);
1643 	mutex_unlock(&p->mutex);
1644 	if (ret)
1645 		goto err_out;
1646 
1647 	ret = dma_buf_fd(dmabuf, args->flags);
1648 	if (ret < 0) {
1649 		dma_buf_put(dmabuf);
1650 		goto err_out;
1651 	}
1652 	/* dma_buf_fd assigns the reference count to the fd, no need to
1653 	 * put the reference here.
1654 	 */
1655 	args->dmabuf_fd = ret;
1656 
1657 	return 0;
1658 
1659 err_unlock:
1660 	mutex_unlock(&p->mutex);
1661 err_out:
1662 	return ret;
1663 }
1664 
1665 /* Handle requests for watching SMI events */
1666 static int kfd_ioctl_smi_events(struct file *filep,
1667 				struct kfd_process *p, void *data)
1668 {
1669 	struct kfd_ioctl_smi_events_args *args = data;
1670 	struct kfd_process_device *pdd;
1671 
1672 	mutex_lock(&p->mutex);
1673 
1674 	pdd = kfd_process_device_data_by_id(p, args->gpuid);
1675 	mutex_unlock(&p->mutex);
1676 	if (!pdd)
1677 		return -EINVAL;
1678 
1679 	return kfd_smi_event_open(pdd->dev, &args->anon_fd);
1680 }
1681 
1682 #if IS_ENABLED(CONFIG_HSA_AMD_SVM)
1683 
1684 static int kfd_ioctl_set_xnack_mode(struct file *filep,
1685 				    struct kfd_process *p, void *data)
1686 {
1687 	struct kfd_ioctl_set_xnack_mode_args *args = data;
1688 	int r = 0;
1689 
1690 	mutex_lock(&p->mutex);
1691 	if (args->xnack_enabled >= 0) {
1692 		if (!list_empty(&p->pqm.queues)) {
1693 			pr_debug("Process has user queues running\n");
1694 			r = -EBUSY;
1695 			goto out_unlock;
1696 		}
1697 
1698 		if (p->xnack_enabled == args->xnack_enabled)
1699 			goto out_unlock;
1700 
1701 		if (args->xnack_enabled && !kfd_process_xnack_mode(p, true)) {
1702 			r = -EPERM;
1703 			goto out_unlock;
1704 		}
1705 
1706 		r = svm_range_switch_xnack_reserve_mem(p, args->xnack_enabled);
1707 	} else {
1708 		args->xnack_enabled = p->xnack_enabled;
1709 	}
1710 
1711 out_unlock:
1712 	mutex_unlock(&p->mutex);
1713 
1714 	return r;
1715 }
1716 
1717 static int kfd_ioctl_svm(struct file *filep, struct kfd_process *p, void *data)
1718 {
1719 	struct kfd_ioctl_svm_args *args = data;
1720 	int r = 0;
1721 
1722 	pr_debug("start 0x%llx size 0x%llx op 0x%x nattr 0x%x\n",
1723 		 args->start_addr, args->size, args->op, args->nattr);
1724 
1725 	if ((args->start_addr & ~PAGE_MASK) || (args->size & ~PAGE_MASK))
1726 		return -EINVAL;
1727 	if (!args->start_addr || !args->size)
1728 		return -EINVAL;
1729 
1730 	r = svm_ioctl(p, args->op, args->start_addr, args->size, args->nattr,
1731 		      args->attrs);
1732 
1733 	return r;
1734 }
1735 #else
1736 static int kfd_ioctl_set_xnack_mode(struct file *filep,
1737 				    struct kfd_process *p, void *data)
1738 {
1739 	return -EPERM;
1740 }
1741 static int kfd_ioctl_svm(struct file *filep, struct kfd_process *p, void *data)
1742 {
1743 	return -EPERM;
1744 }
1745 #endif
1746 
1747 static int criu_checkpoint_process(struct kfd_process *p,
1748 			     uint8_t __user *user_priv_data,
1749 			     uint64_t *priv_offset)
1750 {
1751 	struct kfd_criu_process_priv_data process_priv;
1752 	int ret;
1753 
1754 	memset(&process_priv, 0, sizeof(process_priv));
1755 
1756 	process_priv.version = KFD_CRIU_PRIV_VERSION;
1757 	/* For CR, we don't consider negative xnack mode which is used for
1758 	 * querying without changing it, here 0 simply means disabled and 1
1759 	 * means enabled so retry for finding a valid PTE.
1760 	 */
1761 	process_priv.xnack_mode = p->xnack_enabled ? 1 : 0;
1762 
1763 	ret = copy_to_user(user_priv_data + *priv_offset,
1764 				&process_priv, sizeof(process_priv));
1765 
1766 	if (ret) {
1767 		pr_err("Failed to copy process information to user\n");
1768 		ret = -EFAULT;
1769 	}
1770 
1771 	*priv_offset += sizeof(process_priv);
1772 	return ret;
1773 }
1774 
1775 static int criu_checkpoint_devices(struct kfd_process *p,
1776 			     uint32_t num_devices,
1777 			     uint8_t __user *user_addr,
1778 			     uint8_t __user *user_priv_data,
1779 			     uint64_t *priv_offset)
1780 {
1781 	struct kfd_criu_device_priv_data *device_priv = NULL;
1782 	struct kfd_criu_device_bucket *device_buckets = NULL;
1783 	int ret = 0, i;
1784 
1785 	device_buckets = kvzalloc(num_devices * sizeof(*device_buckets), GFP_KERNEL);
1786 	if (!device_buckets) {
1787 		ret = -ENOMEM;
1788 		goto exit;
1789 	}
1790 
1791 	device_priv = kvzalloc(num_devices * sizeof(*device_priv), GFP_KERNEL);
1792 	if (!device_priv) {
1793 		ret = -ENOMEM;
1794 		goto exit;
1795 	}
1796 
1797 	for (i = 0; i < num_devices; i++) {
1798 		struct kfd_process_device *pdd = p->pdds[i];
1799 
1800 		device_buckets[i].user_gpu_id = pdd->user_gpu_id;
1801 		device_buckets[i].actual_gpu_id = pdd->dev->id;
1802 
1803 		/*
1804 		 * priv_data does not contain useful information for now and is reserved for
1805 		 * future use, so we do not set its contents.
1806 		 */
1807 	}
1808 
1809 	ret = copy_to_user(user_addr, device_buckets, num_devices * sizeof(*device_buckets));
1810 	if (ret) {
1811 		pr_err("Failed to copy device information to user\n");
1812 		ret = -EFAULT;
1813 		goto exit;
1814 	}
1815 
1816 	ret = copy_to_user(user_priv_data + *priv_offset,
1817 			   device_priv,
1818 			   num_devices * sizeof(*device_priv));
1819 	if (ret) {
1820 		pr_err("Failed to copy device information to user\n");
1821 		ret = -EFAULT;
1822 	}
1823 	*priv_offset += num_devices * sizeof(*device_priv);
1824 
1825 exit:
1826 	kvfree(device_buckets);
1827 	kvfree(device_priv);
1828 	return ret;
1829 }
1830 
1831 static uint32_t get_process_num_bos(struct kfd_process *p)
1832 {
1833 	uint32_t num_of_bos = 0;
1834 	int i;
1835 
1836 	/* Run over all PDDs of the process */
1837 	for (i = 0; i < p->n_pdds; i++) {
1838 		struct kfd_process_device *pdd = p->pdds[i];
1839 		void *mem;
1840 		int id;
1841 
1842 		idr_for_each_entry(&pdd->alloc_idr, mem, id) {
1843 			struct kgd_mem *kgd_mem = (struct kgd_mem *)mem;
1844 
1845 			if (!kgd_mem->va || kgd_mem->va > pdd->gpuvm_base)
1846 				num_of_bos++;
1847 		}
1848 	}
1849 	return num_of_bos;
1850 }
1851 
1852 static int criu_get_prime_handle(struct kgd_mem *mem,
1853 				 int flags, u32 *shared_fd,
1854 				 struct file **file)
1855 {
1856 	struct dma_buf *dmabuf;
1857 	int ret;
1858 
1859 	ret = amdgpu_amdkfd_gpuvm_export_dmabuf(mem, &dmabuf);
1860 	if (ret) {
1861 		pr_err("dmabuf export failed for the BO\n");
1862 		return ret;
1863 	}
1864 
1865 	ret = get_unused_fd_flags(flags);
1866 	if (ret < 0) {
1867 		pr_err("dmabuf create fd failed, ret:%d\n", ret);
1868 		goto out_free_dmabuf;
1869 	}
1870 
1871 	*shared_fd = ret;
1872 	*file = dmabuf->file;
1873 	return 0;
1874 
1875 out_free_dmabuf:
1876 	dma_buf_put(dmabuf);
1877 	return ret;
1878 }
1879 
1880 static void commit_files(struct file **files,
1881 			 struct kfd_criu_bo_bucket *bo_buckets,
1882 			 unsigned int count,
1883 			 int err)
1884 {
1885 	while (count--) {
1886 		struct file *file = files[count];
1887 
1888 		if (!file)
1889 			continue;
1890 		if (err) {
1891 			fput(file);
1892 			put_unused_fd(bo_buckets[count].dmabuf_fd);
1893 		} else {
1894 			fd_install(bo_buckets[count].dmabuf_fd, file);
1895 		}
1896 	}
1897 }
1898 
1899 static int criu_checkpoint_bos(struct kfd_process *p,
1900 			       uint32_t num_bos,
1901 			       uint8_t __user *user_bos,
1902 			       uint8_t __user *user_priv_data,
1903 			       uint64_t *priv_offset)
1904 {
1905 	struct kfd_criu_bo_bucket *bo_buckets;
1906 	struct kfd_criu_bo_priv_data *bo_privs;
1907 	struct file **files = NULL;
1908 	int ret = 0, pdd_index, bo_index = 0, id;
1909 	void *mem;
1910 
1911 	bo_buckets = kvzalloc(num_bos * sizeof(*bo_buckets), GFP_KERNEL);
1912 	if (!bo_buckets)
1913 		return -ENOMEM;
1914 
1915 	bo_privs = kvzalloc(num_bos * sizeof(*bo_privs), GFP_KERNEL);
1916 	if (!bo_privs) {
1917 		ret = -ENOMEM;
1918 		goto exit;
1919 	}
1920 
1921 	files = kvzalloc(num_bos * sizeof(struct file *), GFP_KERNEL);
1922 	if (!files) {
1923 		ret = -ENOMEM;
1924 		goto exit;
1925 	}
1926 
1927 	for (pdd_index = 0; pdd_index < p->n_pdds; pdd_index++) {
1928 		struct kfd_process_device *pdd = p->pdds[pdd_index];
1929 		struct amdgpu_bo *dumper_bo;
1930 		struct kgd_mem *kgd_mem;
1931 
1932 		idr_for_each_entry(&pdd->alloc_idr, mem, id) {
1933 			struct kfd_criu_bo_bucket *bo_bucket;
1934 			struct kfd_criu_bo_priv_data *bo_priv;
1935 			int i, dev_idx = 0;
1936 
1937 			kgd_mem = (struct kgd_mem *)mem;
1938 			dumper_bo = kgd_mem->bo;
1939 
1940 			/* Skip checkpointing BOs that are used for Trap handler
1941 			 * code and state. Currently, these BOs have a VA that
1942 			 * is less GPUVM Base
1943 			 */
1944 			if (kgd_mem->va && kgd_mem->va <= pdd->gpuvm_base)
1945 				continue;
1946 
1947 			bo_bucket = &bo_buckets[bo_index];
1948 			bo_priv = &bo_privs[bo_index];
1949 
1950 			bo_bucket->gpu_id = pdd->user_gpu_id;
1951 			bo_bucket->addr = (uint64_t)kgd_mem->va;
1952 			bo_bucket->size = amdgpu_bo_size(dumper_bo);
1953 			bo_bucket->alloc_flags = (uint32_t)kgd_mem->alloc_flags;
1954 			bo_priv->idr_handle = id;
1955 
1956 			if (bo_bucket->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_USERPTR) {
1957 				ret = amdgpu_ttm_tt_get_userptr(&dumper_bo->tbo,
1958 								&bo_priv->user_addr);
1959 				if (ret) {
1960 					pr_err("Failed to obtain user address for user-pointer bo\n");
1961 					goto exit;
1962 				}
1963 			}
1964 			if (bo_bucket->alloc_flags
1965 			    & (KFD_IOC_ALLOC_MEM_FLAGS_VRAM | KFD_IOC_ALLOC_MEM_FLAGS_GTT)) {
1966 				ret = criu_get_prime_handle(kgd_mem,
1967 						bo_bucket->alloc_flags &
1968 						KFD_IOC_ALLOC_MEM_FLAGS_WRITABLE ? DRM_RDWR : 0,
1969 						&bo_bucket->dmabuf_fd, &files[bo_index]);
1970 				if (ret)
1971 					goto exit;
1972 			} else {
1973 				bo_bucket->dmabuf_fd = KFD_INVALID_FD;
1974 			}
1975 
1976 			if (bo_bucket->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_DOORBELL)
1977 				bo_bucket->offset = KFD_MMAP_TYPE_DOORBELL |
1978 					KFD_MMAP_GPU_ID(pdd->dev->id);
1979 			else if (bo_bucket->alloc_flags &
1980 				KFD_IOC_ALLOC_MEM_FLAGS_MMIO_REMAP)
1981 				bo_bucket->offset = KFD_MMAP_TYPE_MMIO |
1982 					KFD_MMAP_GPU_ID(pdd->dev->id);
1983 			else
1984 				bo_bucket->offset = amdgpu_bo_mmap_offset(dumper_bo);
1985 
1986 			for (i = 0; i < p->n_pdds; i++) {
1987 				if (amdgpu_amdkfd_bo_mapped_to_dev(p->pdds[i]->drm_priv, kgd_mem))
1988 					bo_priv->mapped_gpuids[dev_idx++] = p->pdds[i]->user_gpu_id;
1989 			}
1990 
1991 			pr_debug("bo_size = 0x%llx, bo_addr = 0x%llx bo_offset = 0x%llx\n"
1992 					"gpu_id = 0x%x alloc_flags = 0x%x idr_handle = 0x%x",
1993 					bo_bucket->size,
1994 					bo_bucket->addr,
1995 					bo_bucket->offset,
1996 					bo_bucket->gpu_id,
1997 					bo_bucket->alloc_flags,
1998 					bo_priv->idr_handle);
1999 			bo_index++;
2000 		}
2001 	}
2002 
2003 	ret = copy_to_user(user_bos, bo_buckets, num_bos * sizeof(*bo_buckets));
2004 	if (ret) {
2005 		pr_err("Failed to copy BO information to user\n");
2006 		ret = -EFAULT;
2007 		goto exit;
2008 	}
2009 
2010 	ret = copy_to_user(user_priv_data + *priv_offset, bo_privs, num_bos * sizeof(*bo_privs));
2011 	if (ret) {
2012 		pr_err("Failed to copy BO priv information to user\n");
2013 		ret = -EFAULT;
2014 		goto exit;
2015 	}
2016 
2017 	*priv_offset += num_bos * sizeof(*bo_privs);
2018 
2019 exit:
2020 	commit_files(files, bo_buckets, bo_index, ret);
2021 	kvfree(files);
2022 	kvfree(bo_buckets);
2023 	kvfree(bo_privs);
2024 	return ret;
2025 }
2026 
2027 static int criu_get_process_object_info(struct kfd_process *p,
2028 					uint32_t *num_devices,
2029 					uint32_t *num_bos,
2030 					uint32_t *num_objects,
2031 					uint64_t *objs_priv_size)
2032 {
2033 	uint64_t queues_priv_data_size, svm_priv_data_size, priv_size;
2034 	uint32_t num_queues, num_events, num_svm_ranges;
2035 	int ret;
2036 
2037 	*num_devices = p->n_pdds;
2038 	*num_bos = get_process_num_bos(p);
2039 
2040 	ret = kfd_process_get_queue_info(p, &num_queues, &queues_priv_data_size);
2041 	if (ret)
2042 		return ret;
2043 
2044 	num_events = kfd_get_num_events(p);
2045 
2046 	svm_range_get_info(p, &num_svm_ranges, &svm_priv_data_size);
2047 
2048 	*num_objects = num_queues + num_events + num_svm_ranges;
2049 
2050 	if (objs_priv_size) {
2051 		priv_size = sizeof(struct kfd_criu_process_priv_data);
2052 		priv_size += *num_devices * sizeof(struct kfd_criu_device_priv_data);
2053 		priv_size += *num_bos * sizeof(struct kfd_criu_bo_priv_data);
2054 		priv_size += queues_priv_data_size;
2055 		priv_size += num_events * sizeof(struct kfd_criu_event_priv_data);
2056 		priv_size += svm_priv_data_size;
2057 		*objs_priv_size = priv_size;
2058 	}
2059 	return 0;
2060 }
2061 
2062 static int criu_checkpoint(struct file *filep,
2063 			   struct kfd_process *p,
2064 			   struct kfd_ioctl_criu_args *args)
2065 {
2066 	int ret;
2067 	uint32_t num_devices, num_bos, num_objects;
2068 	uint64_t priv_size, priv_offset = 0, bo_priv_offset;
2069 
2070 	if (!args->devices || !args->bos || !args->priv_data)
2071 		return -EINVAL;
2072 
2073 	mutex_lock(&p->mutex);
2074 
2075 	if (!p->n_pdds) {
2076 		pr_err("No pdd for given process\n");
2077 		ret = -ENODEV;
2078 		goto exit_unlock;
2079 	}
2080 
2081 	/* Confirm all process queues are evicted */
2082 	if (!p->queues_paused) {
2083 		pr_err("Cannot dump process when queues are not in evicted state\n");
2084 		/* CRIU plugin did not call op PROCESS_INFO before checkpointing */
2085 		ret = -EINVAL;
2086 		goto exit_unlock;
2087 	}
2088 
2089 	ret = criu_get_process_object_info(p, &num_devices, &num_bos, &num_objects, &priv_size);
2090 	if (ret)
2091 		goto exit_unlock;
2092 
2093 	if (num_devices != args->num_devices ||
2094 	    num_bos != args->num_bos ||
2095 	    num_objects != args->num_objects ||
2096 	    priv_size != args->priv_data_size) {
2097 
2098 		ret = -EINVAL;
2099 		goto exit_unlock;
2100 	}
2101 
2102 	/* each function will store private data inside priv_data and adjust priv_offset */
2103 	ret = criu_checkpoint_process(p, (uint8_t __user *)args->priv_data, &priv_offset);
2104 	if (ret)
2105 		goto exit_unlock;
2106 
2107 	ret = criu_checkpoint_devices(p, num_devices, (uint8_t __user *)args->devices,
2108 				(uint8_t __user *)args->priv_data, &priv_offset);
2109 	if (ret)
2110 		goto exit_unlock;
2111 
2112 	/* Leave room for BOs in the private data. They need to be restored
2113 	 * before events, but we checkpoint them last to simplify the error
2114 	 * handling.
2115 	 */
2116 	bo_priv_offset = priv_offset;
2117 	priv_offset += num_bos * sizeof(struct kfd_criu_bo_priv_data);
2118 
2119 	if (num_objects) {
2120 		ret = kfd_criu_checkpoint_queues(p, (uint8_t __user *)args->priv_data,
2121 						 &priv_offset);
2122 		if (ret)
2123 			goto exit_unlock;
2124 
2125 		ret = kfd_criu_checkpoint_events(p, (uint8_t __user *)args->priv_data,
2126 						 &priv_offset);
2127 		if (ret)
2128 			goto exit_unlock;
2129 
2130 		ret = kfd_criu_checkpoint_svm(p, (uint8_t __user *)args->priv_data, &priv_offset);
2131 		if (ret)
2132 			goto exit_unlock;
2133 	}
2134 
2135 	/* This must be the last thing in this function that can fail.
2136 	 * Otherwise we leak dmabuf file descriptors.
2137 	 */
2138 	ret = criu_checkpoint_bos(p, num_bos, (uint8_t __user *)args->bos,
2139 			   (uint8_t __user *)args->priv_data, &bo_priv_offset);
2140 
2141 exit_unlock:
2142 	mutex_unlock(&p->mutex);
2143 	if (ret)
2144 		pr_err("Failed to dump CRIU ret:%d\n", ret);
2145 	else
2146 		pr_debug("CRIU dump ret:%d\n", ret);
2147 
2148 	return ret;
2149 }
2150 
2151 static int criu_restore_process(struct kfd_process *p,
2152 				struct kfd_ioctl_criu_args *args,
2153 				uint64_t *priv_offset,
2154 				uint64_t max_priv_data_size)
2155 {
2156 	int ret = 0;
2157 	struct kfd_criu_process_priv_data process_priv;
2158 
2159 	if (*priv_offset + sizeof(process_priv) > max_priv_data_size)
2160 		return -EINVAL;
2161 
2162 	ret = copy_from_user(&process_priv,
2163 				(void __user *)(args->priv_data + *priv_offset),
2164 				sizeof(process_priv));
2165 	if (ret) {
2166 		pr_err("Failed to copy process private information from user\n");
2167 		ret = -EFAULT;
2168 		goto exit;
2169 	}
2170 	*priv_offset += sizeof(process_priv);
2171 
2172 	if (process_priv.version != KFD_CRIU_PRIV_VERSION) {
2173 		pr_err("Invalid CRIU API version (checkpointed:%d current:%d)\n",
2174 			process_priv.version, KFD_CRIU_PRIV_VERSION);
2175 		return -EINVAL;
2176 	}
2177 
2178 	pr_debug("Setting XNACK mode\n");
2179 	if (process_priv.xnack_mode && !kfd_process_xnack_mode(p, true)) {
2180 		pr_err("xnack mode cannot be set\n");
2181 		ret = -EPERM;
2182 		goto exit;
2183 	} else {
2184 		pr_debug("set xnack mode: %d\n", process_priv.xnack_mode);
2185 		p->xnack_enabled = process_priv.xnack_mode;
2186 	}
2187 
2188 exit:
2189 	return ret;
2190 }
2191 
2192 static int criu_restore_devices(struct kfd_process *p,
2193 				struct kfd_ioctl_criu_args *args,
2194 				uint64_t *priv_offset,
2195 				uint64_t max_priv_data_size)
2196 {
2197 	struct kfd_criu_device_bucket *device_buckets;
2198 	struct kfd_criu_device_priv_data *device_privs;
2199 	int ret = 0;
2200 	uint32_t i;
2201 
2202 	if (args->num_devices != p->n_pdds)
2203 		return -EINVAL;
2204 
2205 	if (*priv_offset + (args->num_devices * sizeof(*device_privs)) > max_priv_data_size)
2206 		return -EINVAL;
2207 
2208 	device_buckets = kmalloc_array(args->num_devices, sizeof(*device_buckets), GFP_KERNEL);
2209 	if (!device_buckets)
2210 		return -ENOMEM;
2211 
2212 	ret = copy_from_user(device_buckets, (void __user *)args->devices,
2213 				args->num_devices * sizeof(*device_buckets));
2214 	if (ret) {
2215 		pr_err("Failed to copy devices buckets from user\n");
2216 		ret = -EFAULT;
2217 		goto exit;
2218 	}
2219 
2220 	for (i = 0; i < args->num_devices; i++) {
2221 		struct kfd_node *dev;
2222 		struct kfd_process_device *pdd;
2223 		struct file *drm_file;
2224 
2225 		/* device private data is not currently used */
2226 
2227 		if (!device_buckets[i].user_gpu_id) {
2228 			pr_err("Invalid user gpu_id\n");
2229 			ret = -EINVAL;
2230 			goto exit;
2231 		}
2232 
2233 		dev = kfd_device_by_id(device_buckets[i].actual_gpu_id);
2234 		if (!dev) {
2235 			pr_err("Failed to find device with gpu_id = %x\n",
2236 				device_buckets[i].actual_gpu_id);
2237 			ret = -EINVAL;
2238 			goto exit;
2239 		}
2240 
2241 		pdd = kfd_get_process_device_data(dev, p);
2242 		if (!pdd) {
2243 			pr_err("Failed to get pdd for gpu_id = %x\n",
2244 					device_buckets[i].actual_gpu_id);
2245 			ret = -EINVAL;
2246 			goto exit;
2247 		}
2248 		pdd->user_gpu_id = device_buckets[i].user_gpu_id;
2249 
2250 		drm_file = fget(device_buckets[i].drm_fd);
2251 		if (!drm_file) {
2252 			pr_err("Invalid render node file descriptor sent from plugin (%d)\n",
2253 				device_buckets[i].drm_fd);
2254 			ret = -EINVAL;
2255 			goto exit;
2256 		}
2257 
2258 		if (pdd->drm_file) {
2259 			ret = -EINVAL;
2260 			goto exit;
2261 		}
2262 
2263 		/* create the vm using render nodes for kfd pdd */
2264 		if (kfd_process_device_init_vm(pdd, drm_file)) {
2265 			pr_err("could not init vm for given pdd\n");
2266 			/* On success, the PDD keeps the drm_file reference */
2267 			fput(drm_file);
2268 			ret = -EINVAL;
2269 			goto exit;
2270 		}
2271 		/*
2272 		 * pdd now already has the vm bound to render node so below api won't create a new
2273 		 * exclusive kfd mapping but use existing one with renderDXXX but is still needed
2274 		 * for iommu v2 binding  and runtime pm.
2275 		 */
2276 		pdd = kfd_bind_process_to_device(dev, p);
2277 		if (IS_ERR(pdd)) {
2278 			ret = PTR_ERR(pdd);
2279 			goto exit;
2280 		}
2281 
2282 		if (!pdd->qpd.proc_doorbells) {
2283 			ret = kfd_alloc_process_doorbells(dev->kfd, pdd);
2284 			if (ret)
2285 				goto exit;
2286 		}
2287 	}
2288 
2289 	/*
2290 	 * We are not copying device private data from user as we are not using the data for now,
2291 	 * but we still adjust for its private data.
2292 	 */
2293 	*priv_offset += args->num_devices * sizeof(*device_privs);
2294 
2295 exit:
2296 	kfree(device_buckets);
2297 	return ret;
2298 }
2299 
2300 static int criu_restore_memory_of_gpu(struct kfd_process_device *pdd,
2301 				      struct kfd_criu_bo_bucket *bo_bucket,
2302 				      struct kfd_criu_bo_priv_data *bo_priv,
2303 				      struct kgd_mem **kgd_mem)
2304 {
2305 	int idr_handle;
2306 	int ret;
2307 	const bool criu_resume = true;
2308 	u64 offset;
2309 
2310 	if (bo_bucket->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_DOORBELL) {
2311 		if (bo_bucket->size !=
2312 				kfd_doorbell_process_slice(pdd->dev->kfd))
2313 			return -EINVAL;
2314 
2315 		offset = kfd_get_process_doorbells(pdd);
2316 		if (!offset)
2317 			return -ENOMEM;
2318 	} else if (bo_bucket->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_MMIO_REMAP) {
2319 		/* MMIO BOs need remapped bus address */
2320 		if (bo_bucket->size != PAGE_SIZE) {
2321 			pr_err("Invalid page size\n");
2322 			return -EINVAL;
2323 		}
2324 		offset = pdd->dev->adev->rmmio_remap.bus_addr;
2325 		if (!offset || (PAGE_SIZE > 4096)) {
2326 			pr_err("amdgpu_amdkfd_get_mmio_remap_phys_addr failed\n");
2327 			return -ENOMEM;
2328 		}
2329 	} else if (bo_bucket->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_USERPTR) {
2330 		offset = bo_priv->user_addr;
2331 	}
2332 	/* Create the BO */
2333 	ret = amdgpu_amdkfd_gpuvm_alloc_memory_of_gpu(pdd->dev->adev, bo_bucket->addr,
2334 						      bo_bucket->size, pdd->drm_priv, kgd_mem,
2335 						      &offset, bo_bucket->alloc_flags, criu_resume);
2336 	if (ret) {
2337 		pr_err("Could not create the BO\n");
2338 		return ret;
2339 	}
2340 	pr_debug("New BO created: size:0x%llx addr:0x%llx offset:0x%llx\n",
2341 		 bo_bucket->size, bo_bucket->addr, offset);
2342 
2343 	/* Restore previous IDR handle */
2344 	pr_debug("Restoring old IDR handle for the BO");
2345 	idr_handle = idr_alloc(&pdd->alloc_idr, *kgd_mem, bo_priv->idr_handle,
2346 			       bo_priv->idr_handle + 1, GFP_KERNEL);
2347 
2348 	if (idr_handle < 0) {
2349 		pr_err("Could not allocate idr\n");
2350 		amdgpu_amdkfd_gpuvm_free_memory_of_gpu(pdd->dev->adev, *kgd_mem, pdd->drm_priv,
2351 						       NULL);
2352 		return -ENOMEM;
2353 	}
2354 
2355 	if (bo_bucket->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_DOORBELL)
2356 		bo_bucket->restored_offset = KFD_MMAP_TYPE_DOORBELL | KFD_MMAP_GPU_ID(pdd->dev->id);
2357 	if (bo_bucket->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_MMIO_REMAP) {
2358 		bo_bucket->restored_offset = KFD_MMAP_TYPE_MMIO | KFD_MMAP_GPU_ID(pdd->dev->id);
2359 	} else if (bo_bucket->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_GTT) {
2360 		bo_bucket->restored_offset = offset;
2361 	} else if (bo_bucket->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_VRAM) {
2362 		bo_bucket->restored_offset = offset;
2363 		/* Update the VRAM usage count */
2364 		atomic64_add(bo_bucket->size, &pdd->vram_usage);
2365 	}
2366 	return 0;
2367 }
2368 
2369 static int criu_restore_bo(struct kfd_process *p,
2370 			   struct kfd_criu_bo_bucket *bo_bucket,
2371 			   struct kfd_criu_bo_priv_data *bo_priv,
2372 			   struct file **file)
2373 {
2374 	struct kfd_process_device *pdd;
2375 	struct kgd_mem *kgd_mem;
2376 	int ret;
2377 	int j;
2378 
2379 	pr_debug("Restoring BO size:0x%llx addr:0x%llx gpu_id:0x%x flags:0x%x idr_handle:0x%x\n",
2380 		 bo_bucket->size, bo_bucket->addr, bo_bucket->gpu_id, bo_bucket->alloc_flags,
2381 		 bo_priv->idr_handle);
2382 
2383 	pdd = kfd_process_device_data_by_id(p, bo_bucket->gpu_id);
2384 	if (!pdd) {
2385 		pr_err("Failed to get pdd\n");
2386 		return -ENODEV;
2387 	}
2388 
2389 	ret = criu_restore_memory_of_gpu(pdd, bo_bucket, bo_priv, &kgd_mem);
2390 	if (ret)
2391 		return ret;
2392 
2393 	/* now map these BOs to GPU/s */
2394 	for (j = 0; j < p->n_pdds; j++) {
2395 		struct kfd_node *peer;
2396 		struct kfd_process_device *peer_pdd;
2397 
2398 		if (!bo_priv->mapped_gpuids[j])
2399 			break;
2400 
2401 		peer_pdd = kfd_process_device_data_by_id(p, bo_priv->mapped_gpuids[j]);
2402 		if (!peer_pdd)
2403 			return -EINVAL;
2404 
2405 		peer = peer_pdd->dev;
2406 
2407 		peer_pdd = kfd_bind_process_to_device(peer, p);
2408 		if (IS_ERR(peer_pdd))
2409 			return PTR_ERR(peer_pdd);
2410 
2411 		ret = amdgpu_amdkfd_gpuvm_map_memory_to_gpu(peer->adev, kgd_mem,
2412 							    peer_pdd->drm_priv);
2413 		if (ret) {
2414 			pr_err("Failed to map to gpu %d/%d\n", j, p->n_pdds);
2415 			return ret;
2416 		}
2417 	}
2418 
2419 	pr_debug("map memory was successful for the BO\n");
2420 	/* create the dmabuf object and export the bo */
2421 	if (bo_bucket->alloc_flags
2422 	    & (KFD_IOC_ALLOC_MEM_FLAGS_VRAM | KFD_IOC_ALLOC_MEM_FLAGS_GTT)) {
2423 		ret = criu_get_prime_handle(kgd_mem, DRM_RDWR,
2424 					    &bo_bucket->dmabuf_fd, file);
2425 		if (ret)
2426 			return ret;
2427 	} else {
2428 		bo_bucket->dmabuf_fd = KFD_INVALID_FD;
2429 	}
2430 
2431 	return 0;
2432 }
2433 
2434 static int criu_restore_bos(struct kfd_process *p,
2435 			    struct kfd_ioctl_criu_args *args,
2436 			    uint64_t *priv_offset,
2437 			    uint64_t max_priv_data_size)
2438 {
2439 	struct kfd_criu_bo_bucket *bo_buckets = NULL;
2440 	struct kfd_criu_bo_priv_data *bo_privs = NULL;
2441 	struct file **files = NULL;
2442 	int ret = 0;
2443 	uint32_t i = 0;
2444 
2445 	if (*priv_offset + (args->num_bos * sizeof(*bo_privs)) > max_priv_data_size)
2446 		return -EINVAL;
2447 
2448 	/* Prevent MMU notifications until stage-4 IOCTL (CRIU_RESUME) is received */
2449 	amdgpu_amdkfd_block_mmu_notifications(p->kgd_process_info);
2450 
2451 	bo_buckets = kvmalloc_array(args->num_bos, sizeof(*bo_buckets), GFP_KERNEL);
2452 	if (!bo_buckets)
2453 		return -ENOMEM;
2454 
2455 	files = kvzalloc(args->num_bos * sizeof(struct file *), GFP_KERNEL);
2456 	if (!files) {
2457 		ret = -ENOMEM;
2458 		goto exit;
2459 	}
2460 
2461 	ret = copy_from_user(bo_buckets, (void __user *)args->bos,
2462 			     args->num_bos * sizeof(*bo_buckets));
2463 	if (ret) {
2464 		pr_err("Failed to copy BOs information from user\n");
2465 		ret = -EFAULT;
2466 		goto exit;
2467 	}
2468 
2469 	bo_privs = kvmalloc_array(args->num_bos, sizeof(*bo_privs), GFP_KERNEL);
2470 	if (!bo_privs) {
2471 		ret = -ENOMEM;
2472 		goto exit;
2473 	}
2474 
2475 	ret = copy_from_user(bo_privs, (void __user *)args->priv_data + *priv_offset,
2476 			     args->num_bos * sizeof(*bo_privs));
2477 	if (ret) {
2478 		pr_err("Failed to copy BOs information from user\n");
2479 		ret = -EFAULT;
2480 		goto exit;
2481 	}
2482 	*priv_offset += args->num_bos * sizeof(*bo_privs);
2483 
2484 	/* Create and map new BOs */
2485 	for (; i < args->num_bos; i++) {
2486 		ret = criu_restore_bo(p, &bo_buckets[i], &bo_privs[i], &files[i]);
2487 		if (ret) {
2488 			pr_debug("Failed to restore BO[%d] ret%d\n", i, ret);
2489 			goto exit;
2490 		}
2491 	} /* done */
2492 
2493 	/* Copy only the buckets back so user can read bo_buckets[N].restored_offset */
2494 	ret = copy_to_user((void __user *)args->bos,
2495 				bo_buckets,
2496 				(args->num_bos * sizeof(*bo_buckets)));
2497 	if (ret)
2498 		ret = -EFAULT;
2499 
2500 exit:
2501 	commit_files(files, bo_buckets, i, ret);
2502 	kvfree(files);
2503 	kvfree(bo_buckets);
2504 	kvfree(bo_privs);
2505 	return ret;
2506 }
2507 
2508 static int criu_restore_objects(struct file *filep,
2509 				struct kfd_process *p,
2510 				struct kfd_ioctl_criu_args *args,
2511 				uint64_t *priv_offset,
2512 				uint64_t max_priv_data_size)
2513 {
2514 	int ret = 0;
2515 	uint32_t i;
2516 
2517 	BUILD_BUG_ON(offsetof(struct kfd_criu_queue_priv_data, object_type));
2518 	BUILD_BUG_ON(offsetof(struct kfd_criu_event_priv_data, object_type));
2519 	BUILD_BUG_ON(offsetof(struct kfd_criu_svm_range_priv_data, object_type));
2520 
2521 	for (i = 0; i < args->num_objects; i++) {
2522 		uint32_t object_type;
2523 
2524 		if (*priv_offset + sizeof(object_type) > max_priv_data_size) {
2525 			pr_err("Invalid private data size\n");
2526 			return -EINVAL;
2527 		}
2528 
2529 		ret = get_user(object_type, (uint32_t __user *)(args->priv_data + *priv_offset));
2530 		if (ret) {
2531 			pr_err("Failed to copy private information from user\n");
2532 			goto exit;
2533 		}
2534 
2535 		switch (object_type) {
2536 		case KFD_CRIU_OBJECT_TYPE_QUEUE:
2537 			ret = kfd_criu_restore_queue(p, (uint8_t __user *)args->priv_data,
2538 						     priv_offset, max_priv_data_size);
2539 			if (ret)
2540 				goto exit;
2541 			break;
2542 		case KFD_CRIU_OBJECT_TYPE_EVENT:
2543 			ret = kfd_criu_restore_event(filep, p, (uint8_t __user *)args->priv_data,
2544 						     priv_offset, max_priv_data_size);
2545 			if (ret)
2546 				goto exit;
2547 			break;
2548 		case KFD_CRIU_OBJECT_TYPE_SVM_RANGE:
2549 			ret = kfd_criu_restore_svm(p, (uint8_t __user *)args->priv_data,
2550 						     priv_offset, max_priv_data_size);
2551 			if (ret)
2552 				goto exit;
2553 			break;
2554 		default:
2555 			pr_err("Invalid object type:%u at index:%d\n", object_type, i);
2556 			ret = -EINVAL;
2557 			goto exit;
2558 		}
2559 	}
2560 exit:
2561 	return ret;
2562 }
2563 
2564 static int criu_restore(struct file *filep,
2565 			struct kfd_process *p,
2566 			struct kfd_ioctl_criu_args *args)
2567 {
2568 	uint64_t priv_offset = 0;
2569 	int ret = 0;
2570 
2571 	pr_debug("CRIU restore (num_devices:%u num_bos:%u num_objects:%u priv_data_size:%llu)\n",
2572 		 args->num_devices, args->num_bos, args->num_objects, args->priv_data_size);
2573 
2574 	if ((args->num_bos > 0 && !args->bos) || !args->devices || !args->priv_data ||
2575 	    !args->priv_data_size || !args->num_devices)
2576 		return -EINVAL;
2577 
2578 	mutex_lock(&p->mutex);
2579 
2580 	/*
2581 	 * Set the process to evicted state to avoid running any new queues before all the memory
2582 	 * mappings are ready.
2583 	 */
2584 	ret = kfd_process_evict_queues(p, KFD_QUEUE_EVICTION_CRIU_RESTORE);
2585 	if (ret)
2586 		goto exit_unlock;
2587 
2588 	/* Each function will adjust priv_offset based on how many bytes they consumed */
2589 	ret = criu_restore_process(p, args, &priv_offset, args->priv_data_size);
2590 	if (ret)
2591 		goto exit_unlock;
2592 
2593 	ret = criu_restore_devices(p, args, &priv_offset, args->priv_data_size);
2594 	if (ret)
2595 		goto exit_unlock;
2596 
2597 	ret = criu_restore_bos(p, args, &priv_offset, args->priv_data_size);
2598 	if (ret)
2599 		goto exit_unlock;
2600 
2601 	ret = criu_restore_objects(filep, p, args, &priv_offset, args->priv_data_size);
2602 	if (ret)
2603 		goto exit_unlock;
2604 
2605 	if (priv_offset != args->priv_data_size) {
2606 		pr_err("Invalid private data size\n");
2607 		ret = -EINVAL;
2608 	}
2609 
2610 exit_unlock:
2611 	mutex_unlock(&p->mutex);
2612 	if (ret)
2613 		pr_err("Failed to restore CRIU ret:%d\n", ret);
2614 	else
2615 		pr_debug("CRIU restore successful\n");
2616 
2617 	return ret;
2618 }
2619 
2620 static int criu_unpause(struct file *filep,
2621 			struct kfd_process *p,
2622 			struct kfd_ioctl_criu_args *args)
2623 {
2624 	int ret;
2625 
2626 	mutex_lock(&p->mutex);
2627 
2628 	if (!p->queues_paused) {
2629 		mutex_unlock(&p->mutex);
2630 		return -EINVAL;
2631 	}
2632 
2633 	ret = kfd_process_restore_queues(p);
2634 	if (ret)
2635 		pr_err("Failed to unpause queues ret:%d\n", ret);
2636 	else
2637 		p->queues_paused = false;
2638 
2639 	mutex_unlock(&p->mutex);
2640 
2641 	return ret;
2642 }
2643 
2644 static int criu_resume(struct file *filep,
2645 			struct kfd_process *p,
2646 			struct kfd_ioctl_criu_args *args)
2647 {
2648 	struct kfd_process *target = NULL;
2649 	struct pid *pid = NULL;
2650 	int ret = 0;
2651 
2652 	pr_debug("Inside %s, target pid for criu restore: %d\n", __func__,
2653 		 args->pid);
2654 
2655 	pid = find_get_pid(args->pid);
2656 	if (!pid) {
2657 		pr_err("Cannot find pid info for %i\n", args->pid);
2658 		return -ESRCH;
2659 	}
2660 
2661 	pr_debug("calling kfd_lookup_process_by_pid\n");
2662 	target = kfd_lookup_process_by_pid(pid);
2663 
2664 	put_pid(pid);
2665 
2666 	if (!target) {
2667 		pr_debug("Cannot find process info for %i\n", args->pid);
2668 		return -ESRCH;
2669 	}
2670 
2671 	mutex_lock(&target->mutex);
2672 	ret = kfd_criu_resume_svm(target);
2673 	if (ret) {
2674 		pr_err("kfd_criu_resume_svm failed for %i\n", args->pid);
2675 		goto exit;
2676 	}
2677 
2678 	ret =  amdgpu_amdkfd_criu_resume(target->kgd_process_info);
2679 	if (ret)
2680 		pr_err("amdgpu_amdkfd_criu_resume failed for %i\n", args->pid);
2681 
2682 exit:
2683 	mutex_unlock(&target->mutex);
2684 
2685 	kfd_unref_process(target);
2686 	return ret;
2687 }
2688 
2689 static int criu_process_info(struct file *filep,
2690 				struct kfd_process *p,
2691 				struct kfd_ioctl_criu_args *args)
2692 {
2693 	int ret = 0;
2694 
2695 	mutex_lock(&p->mutex);
2696 
2697 	if (!p->n_pdds) {
2698 		pr_err("No pdd for given process\n");
2699 		ret = -ENODEV;
2700 		goto err_unlock;
2701 	}
2702 
2703 	ret = kfd_process_evict_queues(p, KFD_QUEUE_EVICTION_CRIU_CHECKPOINT);
2704 	if (ret)
2705 		goto err_unlock;
2706 
2707 	p->queues_paused = true;
2708 
2709 	args->pid = task_pid_nr_ns(p->lead_thread,
2710 					task_active_pid_ns(p->lead_thread));
2711 
2712 	ret = criu_get_process_object_info(p, &args->num_devices, &args->num_bos,
2713 					   &args->num_objects, &args->priv_data_size);
2714 	if (ret)
2715 		goto err_unlock;
2716 
2717 	dev_dbg(kfd_device, "Num of devices:%u bos:%u objects:%u priv_data_size:%lld\n",
2718 				args->num_devices, args->num_bos, args->num_objects,
2719 				args->priv_data_size);
2720 
2721 err_unlock:
2722 	if (ret) {
2723 		kfd_process_restore_queues(p);
2724 		p->queues_paused = false;
2725 	}
2726 	mutex_unlock(&p->mutex);
2727 	return ret;
2728 }
2729 
2730 static int kfd_ioctl_criu(struct file *filep, struct kfd_process *p, void *data)
2731 {
2732 	struct kfd_ioctl_criu_args *args = data;
2733 	int ret;
2734 
2735 	dev_dbg(kfd_device, "CRIU operation: %d\n", args->op);
2736 	switch (args->op) {
2737 	case KFD_CRIU_OP_PROCESS_INFO:
2738 		ret = criu_process_info(filep, p, args);
2739 		break;
2740 	case KFD_CRIU_OP_CHECKPOINT:
2741 		ret = criu_checkpoint(filep, p, args);
2742 		break;
2743 	case KFD_CRIU_OP_UNPAUSE:
2744 		ret = criu_unpause(filep, p, args);
2745 		break;
2746 	case KFD_CRIU_OP_RESTORE:
2747 		ret = criu_restore(filep, p, args);
2748 		break;
2749 	case KFD_CRIU_OP_RESUME:
2750 		ret = criu_resume(filep, p, args);
2751 		break;
2752 	default:
2753 		dev_dbg(kfd_device, "Unsupported CRIU operation:%d\n", args->op);
2754 		ret = -EINVAL;
2755 		break;
2756 	}
2757 
2758 	if (ret)
2759 		dev_dbg(kfd_device, "CRIU operation:%d err:%d\n", args->op, ret);
2760 
2761 	return ret;
2762 }
2763 
2764 static int runtime_enable(struct kfd_process *p, uint64_t r_debug,
2765 			bool enable_ttmp_setup)
2766 {
2767 	int i = 0, ret = 0;
2768 
2769 	if (p->is_runtime_retry)
2770 		goto retry;
2771 
2772 	if (p->runtime_info.runtime_state != DEBUG_RUNTIME_STATE_DISABLED)
2773 		return -EBUSY;
2774 
2775 	for (i = 0; i < p->n_pdds; i++) {
2776 		struct kfd_process_device *pdd = p->pdds[i];
2777 
2778 		if (pdd->qpd.queue_count)
2779 			return -EEXIST;
2780 
2781 		/*
2782 		 * Setup TTMPs by default.
2783 		 * Note that this call must remain here for MES ADD QUEUE to
2784 		 * skip_process_ctx_clear unconditionally as the first call to
2785 		 * SET_SHADER_DEBUGGER clears any stale process context data
2786 		 * saved in MES.
2787 		 */
2788 		if (pdd->dev->kfd->shared_resources.enable_mes)
2789 			kfd_dbg_set_mes_debug_mode(pdd, !kfd_dbg_has_cwsr_workaround(pdd->dev));
2790 	}
2791 
2792 	p->runtime_info.runtime_state = DEBUG_RUNTIME_STATE_ENABLED;
2793 	p->runtime_info.r_debug = r_debug;
2794 	p->runtime_info.ttmp_setup = enable_ttmp_setup;
2795 
2796 	if (p->runtime_info.ttmp_setup) {
2797 		for (i = 0; i < p->n_pdds; i++) {
2798 			struct kfd_process_device *pdd = p->pdds[i];
2799 
2800 			if (!kfd_dbg_is_rlc_restore_supported(pdd->dev)) {
2801 				amdgpu_gfx_off_ctrl(pdd->dev->adev, false);
2802 				pdd->dev->kfd2kgd->enable_debug_trap(
2803 						pdd->dev->adev,
2804 						true,
2805 						pdd->dev->vm_info.last_vmid_kfd);
2806 			} else if (kfd_dbg_is_per_vmid_supported(pdd->dev)) {
2807 				pdd->spi_dbg_override = pdd->dev->kfd2kgd->enable_debug_trap(
2808 						pdd->dev->adev,
2809 						false,
2810 						0);
2811 			}
2812 		}
2813 	}
2814 
2815 retry:
2816 	if (p->debug_trap_enabled) {
2817 		if (!p->is_runtime_retry) {
2818 			kfd_dbg_trap_activate(p);
2819 			kfd_dbg_ev_raise(KFD_EC_MASK(EC_PROCESS_RUNTIME),
2820 					p, NULL, 0, false, NULL, 0);
2821 		}
2822 
2823 		mutex_unlock(&p->mutex);
2824 		ret = down_interruptible(&p->runtime_enable_sema);
2825 		mutex_lock(&p->mutex);
2826 
2827 		p->is_runtime_retry = !!ret;
2828 	}
2829 
2830 	return ret;
2831 }
2832 
2833 static int runtime_disable(struct kfd_process *p)
2834 {
2835 	int i = 0, ret;
2836 	bool was_enabled = p->runtime_info.runtime_state == DEBUG_RUNTIME_STATE_ENABLED;
2837 
2838 	p->runtime_info.runtime_state = DEBUG_RUNTIME_STATE_DISABLED;
2839 	p->runtime_info.r_debug = 0;
2840 
2841 	if (p->debug_trap_enabled) {
2842 		if (was_enabled)
2843 			kfd_dbg_trap_deactivate(p, false, 0);
2844 
2845 		if (!p->is_runtime_retry)
2846 			kfd_dbg_ev_raise(KFD_EC_MASK(EC_PROCESS_RUNTIME),
2847 					p, NULL, 0, false, NULL, 0);
2848 
2849 		mutex_unlock(&p->mutex);
2850 		ret = down_interruptible(&p->runtime_enable_sema);
2851 		mutex_lock(&p->mutex);
2852 
2853 		p->is_runtime_retry = !!ret;
2854 		if (ret)
2855 			return ret;
2856 	}
2857 
2858 	if (was_enabled && p->runtime_info.ttmp_setup) {
2859 		for (i = 0; i < p->n_pdds; i++) {
2860 			struct kfd_process_device *pdd = p->pdds[i];
2861 
2862 			if (!kfd_dbg_is_rlc_restore_supported(pdd->dev))
2863 				amdgpu_gfx_off_ctrl(pdd->dev->adev, true);
2864 		}
2865 	}
2866 
2867 	p->runtime_info.ttmp_setup = false;
2868 
2869 	/* disable ttmp setup */
2870 	for (i = 0; i < p->n_pdds; i++) {
2871 		struct kfd_process_device *pdd = p->pdds[i];
2872 
2873 		if (kfd_dbg_is_per_vmid_supported(pdd->dev)) {
2874 			pdd->spi_dbg_override =
2875 					pdd->dev->kfd2kgd->disable_debug_trap(
2876 					pdd->dev->adev,
2877 					false,
2878 					pdd->dev->vm_info.last_vmid_kfd);
2879 
2880 			if (!pdd->dev->kfd->shared_resources.enable_mes)
2881 				debug_refresh_runlist(pdd->dev->dqm);
2882 			else
2883 				kfd_dbg_set_mes_debug_mode(pdd,
2884 							   !kfd_dbg_has_cwsr_workaround(pdd->dev));
2885 		}
2886 	}
2887 
2888 	return 0;
2889 }
2890 
2891 static int kfd_ioctl_runtime_enable(struct file *filep, struct kfd_process *p, void *data)
2892 {
2893 	struct kfd_ioctl_runtime_enable_args *args = data;
2894 	int r;
2895 
2896 	mutex_lock(&p->mutex);
2897 
2898 	if (args->mode_mask & KFD_RUNTIME_ENABLE_MODE_ENABLE_MASK)
2899 		r = runtime_enable(p, args->r_debug,
2900 				!!(args->mode_mask & KFD_RUNTIME_ENABLE_MODE_TTMP_SAVE_MASK));
2901 	else
2902 		r = runtime_disable(p);
2903 
2904 	mutex_unlock(&p->mutex);
2905 
2906 	return r;
2907 }
2908 
2909 static int kfd_ioctl_set_debug_trap(struct file *filep, struct kfd_process *p, void *data)
2910 {
2911 	struct kfd_ioctl_dbg_trap_args *args = data;
2912 	struct task_struct *thread = NULL;
2913 	struct mm_struct *mm = NULL;
2914 	struct pid *pid = NULL;
2915 	struct kfd_process *target = NULL;
2916 	struct kfd_process_device *pdd = NULL;
2917 	int r = 0;
2918 
2919 	if (sched_policy == KFD_SCHED_POLICY_NO_HWS) {
2920 		pr_err("Debugging does not support sched_policy %i", sched_policy);
2921 		return -EINVAL;
2922 	}
2923 
2924 	pid = find_get_pid(args->pid);
2925 	if (!pid) {
2926 		pr_debug("Cannot find pid info for %i\n", args->pid);
2927 		r = -ESRCH;
2928 		goto out;
2929 	}
2930 
2931 	thread = get_pid_task(pid, PIDTYPE_PID);
2932 	if (!thread) {
2933 		r = -ESRCH;
2934 		goto out;
2935 	}
2936 
2937 	mm = get_task_mm(thread);
2938 	if (!mm) {
2939 		r = -ESRCH;
2940 		goto out;
2941 	}
2942 
2943 	if (args->op == KFD_IOC_DBG_TRAP_ENABLE) {
2944 		bool create_process;
2945 
2946 		rcu_read_lock();
2947 		create_process = thread && thread != current && ptrace_parent(thread) == current;
2948 		rcu_read_unlock();
2949 
2950 		target = create_process ? kfd_create_process(thread) :
2951 					kfd_lookup_process_by_pid(pid);
2952 	} else {
2953 		target = kfd_lookup_process_by_pid(pid);
2954 	}
2955 
2956 	if (IS_ERR_OR_NULL(target)) {
2957 		pr_debug("Cannot find process PID %i to debug\n", args->pid);
2958 		r = target ? PTR_ERR(target) : -ESRCH;
2959 		target = NULL;
2960 		goto out;
2961 	}
2962 
2963 	/* Check if target is still PTRACED. */
2964 	rcu_read_lock();
2965 	if (target != p && args->op != KFD_IOC_DBG_TRAP_DISABLE
2966 				&& ptrace_parent(target->lead_thread) != current) {
2967 		pr_err("PID %i is not PTRACED and cannot be debugged\n", args->pid);
2968 		r = -EPERM;
2969 	}
2970 	rcu_read_unlock();
2971 
2972 	if (r)
2973 		goto out;
2974 
2975 	mutex_lock(&target->mutex);
2976 
2977 	if (args->op != KFD_IOC_DBG_TRAP_ENABLE && !target->debug_trap_enabled) {
2978 		pr_err("PID %i not debug enabled for op %i\n", args->pid, args->op);
2979 		r = -EINVAL;
2980 		goto unlock_out;
2981 	}
2982 
2983 	if (target->runtime_info.runtime_state != DEBUG_RUNTIME_STATE_ENABLED &&
2984 			(args->op == KFD_IOC_DBG_TRAP_SET_WAVE_LAUNCH_OVERRIDE ||
2985 			 args->op == KFD_IOC_DBG_TRAP_SET_WAVE_LAUNCH_MODE ||
2986 			 args->op == KFD_IOC_DBG_TRAP_SUSPEND_QUEUES ||
2987 			 args->op == KFD_IOC_DBG_TRAP_RESUME_QUEUES ||
2988 			 args->op == KFD_IOC_DBG_TRAP_SET_NODE_ADDRESS_WATCH ||
2989 			 args->op == KFD_IOC_DBG_TRAP_CLEAR_NODE_ADDRESS_WATCH ||
2990 			 args->op == KFD_IOC_DBG_TRAP_SET_FLAGS)) {
2991 		r = -EPERM;
2992 		goto unlock_out;
2993 	}
2994 
2995 	if (args->op == KFD_IOC_DBG_TRAP_SET_NODE_ADDRESS_WATCH ||
2996 	    args->op == KFD_IOC_DBG_TRAP_CLEAR_NODE_ADDRESS_WATCH) {
2997 		int user_gpu_id = kfd_process_get_user_gpu_id(target,
2998 				args->op == KFD_IOC_DBG_TRAP_SET_NODE_ADDRESS_WATCH ?
2999 					args->set_node_address_watch.gpu_id :
3000 					args->clear_node_address_watch.gpu_id);
3001 
3002 		pdd = kfd_process_device_data_by_id(target, user_gpu_id);
3003 		if (user_gpu_id == -EINVAL || !pdd) {
3004 			r = -ENODEV;
3005 			goto unlock_out;
3006 		}
3007 	}
3008 
3009 	switch (args->op) {
3010 	case KFD_IOC_DBG_TRAP_ENABLE:
3011 		if (target != p)
3012 			target->debugger_process = p;
3013 
3014 		r = kfd_dbg_trap_enable(target,
3015 					args->enable.dbg_fd,
3016 					(void __user *)args->enable.rinfo_ptr,
3017 					&args->enable.rinfo_size);
3018 		if (!r)
3019 			target->exception_enable_mask = args->enable.exception_mask;
3020 
3021 		break;
3022 	case KFD_IOC_DBG_TRAP_DISABLE:
3023 		r = kfd_dbg_trap_disable(target);
3024 		break;
3025 	case KFD_IOC_DBG_TRAP_SEND_RUNTIME_EVENT:
3026 		r = kfd_dbg_send_exception_to_runtime(target,
3027 				args->send_runtime_event.gpu_id,
3028 				args->send_runtime_event.queue_id,
3029 				args->send_runtime_event.exception_mask);
3030 		break;
3031 	case KFD_IOC_DBG_TRAP_SET_EXCEPTIONS_ENABLED:
3032 		kfd_dbg_set_enabled_debug_exception_mask(target,
3033 				args->set_exceptions_enabled.exception_mask);
3034 		break;
3035 	case KFD_IOC_DBG_TRAP_SET_WAVE_LAUNCH_OVERRIDE:
3036 		r = kfd_dbg_trap_set_wave_launch_override(target,
3037 				args->launch_override.override_mode,
3038 				args->launch_override.enable_mask,
3039 				args->launch_override.support_request_mask,
3040 				&args->launch_override.enable_mask,
3041 				&args->launch_override.support_request_mask);
3042 		break;
3043 	case KFD_IOC_DBG_TRAP_SET_WAVE_LAUNCH_MODE:
3044 		r = kfd_dbg_trap_set_wave_launch_mode(target,
3045 				args->launch_mode.launch_mode);
3046 		break;
3047 	case KFD_IOC_DBG_TRAP_SUSPEND_QUEUES:
3048 		r = suspend_queues(target,
3049 				args->suspend_queues.num_queues,
3050 				args->suspend_queues.grace_period,
3051 				args->suspend_queues.exception_mask,
3052 				(uint32_t *)args->suspend_queues.queue_array_ptr);
3053 
3054 		break;
3055 	case KFD_IOC_DBG_TRAP_RESUME_QUEUES:
3056 		r = resume_queues(target, args->resume_queues.num_queues,
3057 				(uint32_t *)args->resume_queues.queue_array_ptr);
3058 		break;
3059 	case KFD_IOC_DBG_TRAP_SET_NODE_ADDRESS_WATCH:
3060 		r = kfd_dbg_trap_set_dev_address_watch(pdd,
3061 				args->set_node_address_watch.address,
3062 				args->set_node_address_watch.mask,
3063 				&args->set_node_address_watch.id,
3064 				args->set_node_address_watch.mode);
3065 		break;
3066 	case KFD_IOC_DBG_TRAP_CLEAR_NODE_ADDRESS_WATCH:
3067 		r = kfd_dbg_trap_clear_dev_address_watch(pdd,
3068 				args->clear_node_address_watch.id);
3069 		break;
3070 	case KFD_IOC_DBG_TRAP_SET_FLAGS:
3071 		r = kfd_dbg_trap_set_flags(target, &args->set_flags.flags);
3072 		break;
3073 	case KFD_IOC_DBG_TRAP_QUERY_DEBUG_EVENT:
3074 		r = kfd_dbg_ev_query_debug_event(target,
3075 				&args->query_debug_event.queue_id,
3076 				&args->query_debug_event.gpu_id,
3077 				args->query_debug_event.exception_mask,
3078 				&args->query_debug_event.exception_mask);
3079 		break;
3080 	case KFD_IOC_DBG_TRAP_QUERY_EXCEPTION_INFO:
3081 		r = kfd_dbg_trap_query_exception_info(target,
3082 				args->query_exception_info.source_id,
3083 				args->query_exception_info.exception_code,
3084 				args->query_exception_info.clear_exception,
3085 				(void __user *)args->query_exception_info.info_ptr,
3086 				&args->query_exception_info.info_size);
3087 		break;
3088 	case KFD_IOC_DBG_TRAP_GET_QUEUE_SNAPSHOT:
3089 		r = pqm_get_queue_snapshot(&target->pqm,
3090 				args->queue_snapshot.exception_mask,
3091 				(void __user *)args->queue_snapshot.snapshot_buf_ptr,
3092 				&args->queue_snapshot.num_queues,
3093 				&args->queue_snapshot.entry_size);
3094 		break;
3095 	case KFD_IOC_DBG_TRAP_GET_DEVICE_SNAPSHOT:
3096 		r = kfd_dbg_trap_device_snapshot(target,
3097 				args->device_snapshot.exception_mask,
3098 				(void __user *)args->device_snapshot.snapshot_buf_ptr,
3099 				&args->device_snapshot.num_devices,
3100 				&args->device_snapshot.entry_size);
3101 		break;
3102 	default:
3103 		pr_err("Invalid option: %i\n", args->op);
3104 		r = -EINVAL;
3105 	}
3106 
3107 unlock_out:
3108 	mutex_unlock(&target->mutex);
3109 
3110 out:
3111 	if (thread)
3112 		put_task_struct(thread);
3113 
3114 	if (mm)
3115 		mmput(mm);
3116 
3117 	if (pid)
3118 		put_pid(pid);
3119 
3120 	if (target)
3121 		kfd_unref_process(target);
3122 
3123 	return r;
3124 }
3125 
3126 #define AMDKFD_IOCTL_DEF(ioctl, _func, _flags) \
3127 	[_IOC_NR(ioctl)] = {.cmd = ioctl, .func = _func, .flags = _flags, \
3128 			    .cmd_drv = 0, .name = #ioctl}
3129 
3130 /** Ioctl table */
3131 static const struct amdkfd_ioctl_desc amdkfd_ioctls[] = {
3132 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_GET_VERSION,
3133 			kfd_ioctl_get_version, 0),
3134 
3135 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_CREATE_QUEUE,
3136 			kfd_ioctl_create_queue, 0),
3137 
3138 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_DESTROY_QUEUE,
3139 			kfd_ioctl_destroy_queue, 0),
3140 
3141 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_SET_MEMORY_POLICY,
3142 			kfd_ioctl_set_memory_policy, 0),
3143 
3144 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_GET_CLOCK_COUNTERS,
3145 			kfd_ioctl_get_clock_counters, 0),
3146 
3147 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_GET_PROCESS_APERTURES,
3148 			kfd_ioctl_get_process_apertures, 0),
3149 
3150 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_UPDATE_QUEUE,
3151 			kfd_ioctl_update_queue, 0),
3152 
3153 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_CREATE_EVENT,
3154 			kfd_ioctl_create_event, 0),
3155 
3156 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_DESTROY_EVENT,
3157 			kfd_ioctl_destroy_event, 0),
3158 
3159 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_SET_EVENT,
3160 			kfd_ioctl_set_event, 0),
3161 
3162 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_RESET_EVENT,
3163 			kfd_ioctl_reset_event, 0),
3164 
3165 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_WAIT_EVENTS,
3166 			kfd_ioctl_wait_events, 0),
3167 
3168 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_DBG_REGISTER_DEPRECATED,
3169 			kfd_ioctl_dbg_register, 0),
3170 
3171 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_DBG_UNREGISTER_DEPRECATED,
3172 			kfd_ioctl_dbg_unregister, 0),
3173 
3174 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_DBG_ADDRESS_WATCH_DEPRECATED,
3175 			kfd_ioctl_dbg_address_watch, 0),
3176 
3177 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_DBG_WAVE_CONTROL_DEPRECATED,
3178 			kfd_ioctl_dbg_wave_control, 0),
3179 
3180 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_SET_SCRATCH_BACKING_VA,
3181 			kfd_ioctl_set_scratch_backing_va, 0),
3182 
3183 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_GET_TILE_CONFIG,
3184 			kfd_ioctl_get_tile_config, 0),
3185 
3186 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_SET_TRAP_HANDLER,
3187 			kfd_ioctl_set_trap_handler, 0),
3188 
3189 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_GET_PROCESS_APERTURES_NEW,
3190 			kfd_ioctl_get_process_apertures_new, 0),
3191 
3192 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_ACQUIRE_VM,
3193 			kfd_ioctl_acquire_vm, 0),
3194 
3195 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_ALLOC_MEMORY_OF_GPU,
3196 			kfd_ioctl_alloc_memory_of_gpu, 0),
3197 
3198 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_FREE_MEMORY_OF_GPU,
3199 			kfd_ioctl_free_memory_of_gpu, 0),
3200 
3201 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_MAP_MEMORY_TO_GPU,
3202 			kfd_ioctl_map_memory_to_gpu, 0),
3203 
3204 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_UNMAP_MEMORY_FROM_GPU,
3205 			kfd_ioctl_unmap_memory_from_gpu, 0),
3206 
3207 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_SET_CU_MASK,
3208 			kfd_ioctl_set_cu_mask, 0),
3209 
3210 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_GET_QUEUE_WAVE_STATE,
3211 			kfd_ioctl_get_queue_wave_state, 0),
3212 
3213 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_GET_DMABUF_INFO,
3214 				kfd_ioctl_get_dmabuf_info, 0),
3215 
3216 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_IMPORT_DMABUF,
3217 				kfd_ioctl_import_dmabuf, 0),
3218 
3219 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_ALLOC_QUEUE_GWS,
3220 			kfd_ioctl_alloc_queue_gws, 0),
3221 
3222 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_SMI_EVENTS,
3223 			kfd_ioctl_smi_events, 0),
3224 
3225 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_SVM, kfd_ioctl_svm, 0),
3226 
3227 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_SET_XNACK_MODE,
3228 			kfd_ioctl_set_xnack_mode, 0),
3229 
3230 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_CRIU_OP,
3231 			kfd_ioctl_criu, KFD_IOC_FLAG_CHECKPOINT_RESTORE),
3232 
3233 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_AVAILABLE_MEMORY,
3234 			kfd_ioctl_get_available_memory, 0),
3235 
3236 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_EXPORT_DMABUF,
3237 				kfd_ioctl_export_dmabuf, 0),
3238 
3239 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_RUNTIME_ENABLE,
3240 			kfd_ioctl_runtime_enable, 0),
3241 
3242 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_DBG_TRAP,
3243 			kfd_ioctl_set_debug_trap, 0),
3244 };
3245 
3246 #define AMDKFD_CORE_IOCTL_COUNT	ARRAY_SIZE(amdkfd_ioctls)
3247 
3248 static long kfd_ioctl(struct file *filep, unsigned int cmd, unsigned long arg)
3249 {
3250 	struct kfd_process *process;
3251 	amdkfd_ioctl_t *func;
3252 	const struct amdkfd_ioctl_desc *ioctl = NULL;
3253 	unsigned int nr = _IOC_NR(cmd);
3254 	char stack_kdata[128];
3255 	char *kdata = NULL;
3256 	unsigned int usize, asize;
3257 	int retcode = -EINVAL;
3258 	bool ptrace_attached = false;
3259 
3260 	if (nr >= AMDKFD_CORE_IOCTL_COUNT) {
3261 		retcode = -ENOTTY;
3262 		goto err_i1;
3263 	}
3264 
3265 	if ((nr >= AMDKFD_COMMAND_START) && (nr < AMDKFD_COMMAND_END)) {
3266 		u32 amdkfd_size;
3267 
3268 		ioctl = &amdkfd_ioctls[nr];
3269 
3270 		amdkfd_size = _IOC_SIZE(ioctl->cmd);
3271 		usize = asize = _IOC_SIZE(cmd);
3272 		if (amdkfd_size > asize)
3273 			asize = amdkfd_size;
3274 
3275 		cmd = ioctl->cmd;
3276 	} else {
3277 		retcode = -ENOTTY;
3278 		goto err_i1;
3279 	}
3280 
3281 	dev_dbg(kfd_device, "ioctl cmd 0x%x (#0x%x), arg 0x%lx\n", cmd, nr, arg);
3282 
3283 	/* Get the process struct from the filep. Only the process
3284 	 * that opened /dev/kfd can use the file descriptor. Child
3285 	 * processes need to create their own KFD device context.
3286 	 */
3287 	process = filep->private_data;
3288 
3289 	rcu_read_lock();
3290 	if ((ioctl->flags & KFD_IOC_FLAG_CHECKPOINT_RESTORE) &&
3291 	    ptrace_parent(process->lead_thread) == current)
3292 		ptrace_attached = true;
3293 	rcu_read_unlock();
3294 
3295 	if (process->lead_thread != current->group_leader
3296 	    && !ptrace_attached) {
3297 		dev_dbg(kfd_device, "Using KFD FD in wrong process\n");
3298 		retcode = -EBADF;
3299 		goto err_i1;
3300 	}
3301 
3302 	/* Do not trust userspace, use our own definition */
3303 	func = ioctl->func;
3304 
3305 	if (unlikely(!func)) {
3306 		dev_dbg(kfd_device, "no function\n");
3307 		retcode = -EINVAL;
3308 		goto err_i1;
3309 	}
3310 
3311 	/*
3312 	 * Versions of docker shipped in Ubuntu 18.xx and 20.xx do not support
3313 	 * CAP_CHECKPOINT_RESTORE, so we also allow access if CAP_SYS_ADMIN as CAP_SYS_ADMIN is a
3314 	 * more priviledged access.
3315 	 */
3316 	if (unlikely(ioctl->flags & KFD_IOC_FLAG_CHECKPOINT_RESTORE)) {
3317 		if (!capable(CAP_CHECKPOINT_RESTORE) &&
3318 						!capable(CAP_SYS_ADMIN)) {
3319 			retcode = -EACCES;
3320 			goto err_i1;
3321 		}
3322 	}
3323 
3324 	if (cmd & (IOC_IN | IOC_OUT)) {
3325 		if (asize <= sizeof(stack_kdata)) {
3326 			kdata = stack_kdata;
3327 		} else {
3328 			kdata = kmalloc(asize, GFP_KERNEL);
3329 			if (!kdata) {
3330 				retcode = -ENOMEM;
3331 				goto err_i1;
3332 			}
3333 		}
3334 		if (asize > usize)
3335 			memset(kdata + usize, 0, asize - usize);
3336 	}
3337 
3338 	if (cmd & IOC_IN) {
3339 		if (copy_from_user(kdata, (void __user *)arg, usize) != 0) {
3340 			retcode = -EFAULT;
3341 			goto err_i1;
3342 		}
3343 	} else if (cmd & IOC_OUT) {
3344 		memset(kdata, 0, usize);
3345 	}
3346 
3347 	retcode = func(filep, process, kdata);
3348 
3349 	if (cmd & IOC_OUT)
3350 		if (copy_to_user((void __user *)arg, kdata, usize) != 0)
3351 			retcode = -EFAULT;
3352 
3353 err_i1:
3354 	if (!ioctl)
3355 		dev_dbg(kfd_device, "invalid ioctl: pid=%d, cmd=0x%02x, nr=0x%02x\n",
3356 			  task_pid_nr(current), cmd, nr);
3357 
3358 	if (kdata != stack_kdata)
3359 		kfree(kdata);
3360 
3361 	if (retcode)
3362 		dev_dbg(kfd_device, "ioctl cmd (#0x%x), arg 0x%lx, ret = %d\n",
3363 				nr, arg, retcode);
3364 
3365 	return retcode;
3366 }
3367 
3368 static int kfd_mmio_mmap(struct kfd_node *dev, struct kfd_process *process,
3369 		      struct vm_area_struct *vma)
3370 {
3371 	phys_addr_t address;
3372 
3373 	if (vma->vm_end - vma->vm_start != PAGE_SIZE)
3374 		return -EINVAL;
3375 
3376 	if (PAGE_SIZE > 4096)
3377 		return -EINVAL;
3378 
3379 	address = dev->adev->rmmio_remap.bus_addr;
3380 
3381 	vm_flags_set(vma, VM_IO | VM_DONTCOPY | VM_DONTEXPAND | VM_NORESERVE |
3382 				VM_DONTDUMP | VM_PFNMAP);
3383 
3384 	vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot);
3385 
3386 	pr_debug("process pid %d mapping mmio page\n"
3387 		 "     target user address == 0x%08llX\n"
3388 		 "     physical address    == 0x%08llX\n"
3389 		 "     vm_flags            == 0x%04lX\n"
3390 		 "     size                == 0x%04lX\n",
3391 		 process->lead_thread->pid, (unsigned long long) vma->vm_start,
3392 		 address, vma->vm_flags, PAGE_SIZE);
3393 
3394 	return io_remap_pfn_range(vma,
3395 				vma->vm_start,
3396 				address >> PAGE_SHIFT,
3397 				PAGE_SIZE,
3398 				vma->vm_page_prot);
3399 }
3400 
3401 
3402 static int kfd_mmap(struct file *filp, struct vm_area_struct *vma)
3403 {
3404 	struct kfd_process *process;
3405 	struct kfd_node *dev = NULL;
3406 	unsigned long mmap_offset;
3407 	unsigned int gpu_id;
3408 
3409 	process = kfd_get_process(current);
3410 	if (IS_ERR(process))
3411 		return PTR_ERR(process);
3412 
3413 	mmap_offset = vma->vm_pgoff << PAGE_SHIFT;
3414 	gpu_id = KFD_MMAP_GET_GPU_ID(mmap_offset);
3415 	if (gpu_id)
3416 		dev = kfd_device_by_id(gpu_id);
3417 
3418 	switch (mmap_offset & KFD_MMAP_TYPE_MASK) {
3419 	case KFD_MMAP_TYPE_DOORBELL:
3420 		if (!dev)
3421 			return -ENODEV;
3422 		return kfd_doorbell_mmap(dev, process, vma);
3423 
3424 	case KFD_MMAP_TYPE_EVENTS:
3425 		return kfd_event_mmap(process, vma);
3426 
3427 	case KFD_MMAP_TYPE_RESERVED_MEM:
3428 		if (!dev)
3429 			return -ENODEV;
3430 		return kfd_reserved_mem_mmap(dev, process, vma);
3431 	case KFD_MMAP_TYPE_MMIO:
3432 		if (!dev)
3433 			return -ENODEV;
3434 		return kfd_mmio_mmap(dev, process, vma);
3435 	}
3436 
3437 	return -EFAULT;
3438 }
3439